1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
5 */
6 #include <uapi/linux/btf.h>
7 #include <linux/bpf-cgroup.h>
8 #include <linux/kernel.h>
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/bpf.h>
12 #include <linux/btf.h>
13 #include <linux/bpf_verifier.h>
14 #include <linux/filter.h>
15 #include <net/netlink.h>
16 #include <linux/file.h>
17 #include <linux/vmalloc.h>
18 #include <linux/stringify.h>
19 #include <linux/bsearch.h>
20 #include <linux/sort.h>
21 #include <linux/perf_event.h>
22 #include <linux/ctype.h>
23 #include <linux/error-injection.h>
24 #include <linux/bpf_lsm.h>
25 #include <linux/btf_ids.h>
26 #include <linux/poison.h>
27 #include <linux/module.h>
28 #include <linux/cpumask.h>
29 #include <net/xdp.h>
30
31 #include "disasm.h"
32
33 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
34 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
35 [_id] = & _name ## _verifier_ops,
36 #define BPF_MAP_TYPE(_id, _ops)
37 #define BPF_LINK_TYPE(_id, _name)
38 #include <linux/bpf_types.h>
39 #undef BPF_PROG_TYPE
40 #undef BPF_MAP_TYPE
41 #undef BPF_LINK_TYPE
42 };
43
44 /* bpf_check() is a static code analyzer that walks eBPF program
45 * instruction by instruction and updates register/stack state.
46 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
47 *
48 * The first pass is depth-first-search to check that the program is a DAG.
49 * It rejects the following programs:
50 * - larger than BPF_MAXINSNS insns
51 * - if loop is present (detected via back-edge)
52 * - unreachable insns exist (shouldn't be a forest. program = one function)
53 * - out of bounds or malformed jumps
54 * The second pass is all possible path descent from the 1st insn.
55 * Since it's analyzing all paths through the program, the length of the
56 * analysis is limited to 64k insn, which may be hit even if total number of
57 * insn is less then 4K, but there are too many branches that change stack/regs.
58 * Number of 'branches to be analyzed' is limited to 1k
59 *
60 * On entry to each instruction, each register has a type, and the instruction
61 * changes the types of the registers depending on instruction semantics.
62 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
63 * copied to R1.
64 *
65 * All registers are 64-bit.
66 * R0 - return register
67 * R1-R5 argument passing registers
68 * R6-R9 callee saved registers
69 * R10 - frame pointer read-only
70 *
71 * At the start of BPF program the register R1 contains a pointer to bpf_context
72 * and has type PTR_TO_CTX.
73 *
74 * Verifier tracks arithmetic operations on pointers in case:
75 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
76 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
77 * 1st insn copies R10 (which has FRAME_PTR) type into R1
78 * and 2nd arithmetic instruction is pattern matched to recognize
79 * that it wants to construct a pointer to some element within stack.
80 * So after 2nd insn, the register R1 has type PTR_TO_STACK
81 * (and -20 constant is saved for further stack bounds checking).
82 * Meaning that this reg is a pointer to stack plus known immediate constant.
83 *
84 * Most of the time the registers have SCALAR_VALUE type, which
85 * means the register has some value, but it's not a valid pointer.
86 * (like pointer plus pointer becomes SCALAR_VALUE type)
87 *
88 * When verifier sees load or store instructions the type of base register
89 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
90 * four pointer types recognized by check_mem_access() function.
91 *
92 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
93 * and the range of [ptr, ptr + map's value_size) is accessible.
94 *
95 * registers used to pass values to function calls are checked against
96 * function argument constraints.
97 *
98 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
99 * It means that the register type passed to this function must be
100 * PTR_TO_STACK and it will be used inside the function as
101 * 'pointer to map element key'
102 *
103 * For example the argument constraints for bpf_map_lookup_elem():
104 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
105 * .arg1_type = ARG_CONST_MAP_PTR,
106 * .arg2_type = ARG_PTR_TO_MAP_KEY,
107 *
108 * ret_type says that this function returns 'pointer to map elem value or null'
109 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
110 * 2nd argument should be a pointer to stack, which will be used inside
111 * the helper function as a pointer to map element key.
112 *
113 * On the kernel side the helper function looks like:
114 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
115 * {
116 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
117 * void *key = (void *) (unsigned long) r2;
118 * void *value;
119 *
120 * here kernel can access 'key' and 'map' pointers safely, knowing that
121 * [key, key + map->key_size) bytes are valid and were initialized on
122 * the stack of eBPF program.
123 * }
124 *
125 * Corresponding eBPF program may look like:
126 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
127 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
128 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
129 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
130 * here verifier looks at prototype of map_lookup_elem() and sees:
131 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
132 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
133 *
134 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
135 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
136 * and were initialized prior to this call.
137 * If it's ok, then verifier allows this BPF_CALL insn and looks at
138 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
139 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
140 * returns either pointer to map value or NULL.
141 *
142 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
143 * insn, the register holding that pointer in the true branch changes state to
144 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
145 * branch. See check_cond_jmp_op().
146 *
147 * After the call R0 is set to return type of the function and registers R1-R5
148 * are set to NOT_INIT to indicate that they are no longer readable.
149 *
150 * The following reference types represent a potential reference to a kernel
151 * resource which, after first being allocated, must be checked and freed by
152 * the BPF program:
153 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
154 *
155 * When the verifier sees a helper call return a reference type, it allocates a
156 * pointer id for the reference and stores it in the current function state.
157 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
158 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
159 * passes through a NULL-check conditional. For the branch wherein the state is
160 * changed to CONST_IMM, the verifier releases the reference.
161 *
162 * For each helper function that allocates a reference, such as
163 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
164 * bpf_sk_release(). When a reference type passes into the release function,
165 * the verifier also releases the reference. If any unchecked or unreleased
166 * reference remains at the end of the program, the verifier rejects it.
167 */
168
169 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
170 struct bpf_verifier_stack_elem {
171 /* verifer state is 'st'
172 * before processing instruction 'insn_idx'
173 * and after processing instruction 'prev_insn_idx'
174 */
175 struct bpf_verifier_state st;
176 int insn_idx;
177 int prev_insn_idx;
178 struct bpf_verifier_stack_elem *next;
179 /* length of verifier log at the time this state was pushed on stack */
180 u32 log_pos;
181 };
182
183 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
184 #define BPF_COMPLEXITY_LIMIT_STATES 64
185
186 #define BPF_MAP_KEY_POISON (1ULL << 63)
187 #define BPF_MAP_KEY_SEEN (1ULL << 62)
188
189 #define BPF_MAP_PTR_UNPRIV 1UL
190 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
191 POISON_POINTER_DELTA))
192 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
193
194 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx);
195 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
196 static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
197 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env);
198 static int ref_set_non_owning(struct bpf_verifier_env *env,
199 struct bpf_reg_state *reg);
200 static void specialize_kfunc(struct bpf_verifier_env *env,
201 u32 func_id, u16 offset, unsigned long *addr);
202 static bool is_trusted_reg(const struct bpf_reg_state *reg);
203
bpf_map_ptr_poisoned(const struct bpf_insn_aux_data * aux)204 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
205 {
206 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
207 }
208
bpf_map_ptr_unpriv(const struct bpf_insn_aux_data * aux)209 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
210 {
211 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
212 }
213
bpf_map_ptr_store(struct bpf_insn_aux_data * aux,const struct bpf_map * map,bool unpriv)214 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
215 const struct bpf_map *map, bool unpriv)
216 {
217 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
218 unpriv |= bpf_map_ptr_unpriv(aux);
219 aux->map_ptr_state = (unsigned long)map |
220 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
221 }
222
bpf_map_key_poisoned(const struct bpf_insn_aux_data * aux)223 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
224 {
225 return aux->map_key_state & BPF_MAP_KEY_POISON;
226 }
227
bpf_map_key_unseen(const struct bpf_insn_aux_data * aux)228 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
229 {
230 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
231 }
232
bpf_map_key_immediate(const struct bpf_insn_aux_data * aux)233 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
234 {
235 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
236 }
237
bpf_map_key_store(struct bpf_insn_aux_data * aux,u64 state)238 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
239 {
240 bool poisoned = bpf_map_key_poisoned(aux);
241
242 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
243 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
244 }
245
bpf_helper_call(const struct bpf_insn * insn)246 static bool bpf_helper_call(const struct bpf_insn *insn)
247 {
248 return insn->code == (BPF_JMP | BPF_CALL) &&
249 insn->src_reg == 0;
250 }
251
bpf_pseudo_call(const struct bpf_insn * insn)252 static bool bpf_pseudo_call(const struct bpf_insn *insn)
253 {
254 return insn->code == (BPF_JMP | BPF_CALL) &&
255 insn->src_reg == BPF_PSEUDO_CALL;
256 }
257
bpf_pseudo_kfunc_call(const struct bpf_insn * insn)258 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
259 {
260 return insn->code == (BPF_JMP | BPF_CALL) &&
261 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
262 }
263
264 struct bpf_call_arg_meta {
265 struct bpf_map *map_ptr;
266 bool raw_mode;
267 bool pkt_access;
268 u8 release_regno;
269 int regno;
270 int access_size;
271 int mem_size;
272 u64 msize_max_value;
273 int ref_obj_id;
274 int dynptr_id;
275 int map_uid;
276 int func_id;
277 struct btf *btf;
278 u32 btf_id;
279 struct btf *ret_btf;
280 u32 ret_btf_id;
281 u32 subprogno;
282 struct btf_field *kptr_field;
283 };
284
285 struct bpf_kfunc_call_arg_meta {
286 /* In parameters */
287 struct btf *btf;
288 u32 func_id;
289 u32 kfunc_flags;
290 const struct btf_type *func_proto;
291 const char *func_name;
292 /* Out parameters */
293 u32 ref_obj_id;
294 u8 release_regno;
295 bool r0_rdonly;
296 u32 ret_btf_id;
297 u64 r0_size;
298 u32 subprogno;
299 struct {
300 u64 value;
301 bool found;
302 } arg_constant;
303
304 /* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling,
305 * generally to pass info about user-defined local kptr types to later
306 * verification logic
307 * bpf_obj_drop
308 * Record the local kptr type to be drop'd
309 * bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type)
310 * Record the local kptr type to be refcount_incr'd and use
311 * arg_owning_ref to determine whether refcount_acquire should be
312 * fallible
313 */
314 struct btf *arg_btf;
315 u32 arg_btf_id;
316 bool arg_owning_ref;
317
318 struct {
319 struct btf_field *field;
320 } arg_list_head;
321 struct {
322 struct btf_field *field;
323 } arg_rbtree_root;
324 struct {
325 enum bpf_dynptr_type type;
326 u32 id;
327 u32 ref_obj_id;
328 } initialized_dynptr;
329 struct {
330 u8 spi;
331 u8 frameno;
332 } iter;
333 u64 mem_size;
334 };
335
336 struct btf *btf_vmlinux;
337
338 static DEFINE_MUTEX(bpf_verifier_lock);
339
340 static const struct bpf_line_info *
find_linfo(const struct bpf_verifier_env * env,u32 insn_off)341 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
342 {
343 const struct bpf_line_info *linfo;
344 const struct bpf_prog *prog;
345 u32 i, nr_linfo;
346
347 prog = env->prog;
348 nr_linfo = prog->aux->nr_linfo;
349
350 if (!nr_linfo || insn_off >= prog->len)
351 return NULL;
352
353 linfo = prog->aux->linfo;
354 for (i = 1; i < nr_linfo; i++)
355 if (insn_off < linfo[i].insn_off)
356 break;
357
358 return &linfo[i - 1];
359 }
360
verbose(void * private_data,const char * fmt,...)361 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
362 {
363 struct bpf_verifier_env *env = private_data;
364 va_list args;
365
366 if (!bpf_verifier_log_needed(&env->log))
367 return;
368
369 va_start(args, fmt);
370 bpf_verifier_vlog(&env->log, fmt, args);
371 va_end(args);
372 }
373
ltrim(const char * s)374 static const char *ltrim(const char *s)
375 {
376 while (isspace(*s))
377 s++;
378
379 return s;
380 }
381
verbose_linfo(struct bpf_verifier_env * env,u32 insn_off,const char * prefix_fmt,...)382 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
383 u32 insn_off,
384 const char *prefix_fmt, ...)
385 {
386 const struct bpf_line_info *linfo;
387
388 if (!bpf_verifier_log_needed(&env->log))
389 return;
390
391 linfo = find_linfo(env, insn_off);
392 if (!linfo || linfo == env->prev_linfo)
393 return;
394
395 if (prefix_fmt) {
396 va_list args;
397
398 va_start(args, prefix_fmt);
399 bpf_verifier_vlog(&env->log, prefix_fmt, args);
400 va_end(args);
401 }
402
403 verbose(env, "%s\n",
404 ltrim(btf_name_by_offset(env->prog->aux->btf,
405 linfo->line_off)));
406
407 env->prev_linfo = linfo;
408 }
409
verbose_invalid_scalar(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct tnum * range,const char * ctx,const char * reg_name)410 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
411 struct bpf_reg_state *reg,
412 struct tnum *range, const char *ctx,
413 const char *reg_name)
414 {
415 char tn_buf[48];
416
417 verbose(env, "At %s the register %s ", ctx, reg_name);
418 if (!tnum_is_unknown(reg->var_off)) {
419 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
420 verbose(env, "has value %s", tn_buf);
421 } else {
422 verbose(env, "has unknown scalar value");
423 }
424 tnum_strn(tn_buf, sizeof(tn_buf), *range);
425 verbose(env, " should have been in %s\n", tn_buf);
426 }
427
type_is_pkt_pointer(enum bpf_reg_type type)428 static bool type_is_pkt_pointer(enum bpf_reg_type type)
429 {
430 type = base_type(type);
431 return type == PTR_TO_PACKET ||
432 type == PTR_TO_PACKET_META;
433 }
434
type_is_sk_pointer(enum bpf_reg_type type)435 static bool type_is_sk_pointer(enum bpf_reg_type type)
436 {
437 return type == PTR_TO_SOCKET ||
438 type == PTR_TO_SOCK_COMMON ||
439 type == PTR_TO_TCP_SOCK ||
440 type == PTR_TO_XDP_SOCK;
441 }
442
type_may_be_null(u32 type)443 static bool type_may_be_null(u32 type)
444 {
445 return type & PTR_MAYBE_NULL;
446 }
447
reg_not_null(const struct bpf_reg_state * reg)448 static bool reg_not_null(const struct bpf_reg_state *reg)
449 {
450 enum bpf_reg_type type;
451
452 type = reg->type;
453 if (type_may_be_null(type))
454 return false;
455
456 type = base_type(type);
457 return type == PTR_TO_SOCKET ||
458 type == PTR_TO_TCP_SOCK ||
459 type == PTR_TO_MAP_VALUE ||
460 type == PTR_TO_MAP_KEY ||
461 type == PTR_TO_SOCK_COMMON ||
462 (type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
463 type == PTR_TO_MEM;
464 }
465
type_is_ptr_alloc_obj(u32 type)466 static bool type_is_ptr_alloc_obj(u32 type)
467 {
468 return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC;
469 }
470
type_is_non_owning_ref(u32 type)471 static bool type_is_non_owning_ref(u32 type)
472 {
473 return type_is_ptr_alloc_obj(type) && type_flag(type) & NON_OWN_REF;
474 }
475
reg_btf_record(const struct bpf_reg_state * reg)476 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
477 {
478 struct btf_record *rec = NULL;
479 struct btf_struct_meta *meta;
480
481 if (reg->type == PTR_TO_MAP_VALUE) {
482 rec = reg->map_ptr->record;
483 } else if (type_is_ptr_alloc_obj(reg->type)) {
484 meta = btf_find_struct_meta(reg->btf, reg->btf_id);
485 if (meta)
486 rec = meta->record;
487 }
488 return rec;
489 }
490
subprog_is_global(const struct bpf_verifier_env * env,int subprog)491 static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog)
492 {
493 struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux;
494
495 return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
496 }
497
reg_may_point_to_spin_lock(const struct bpf_reg_state * reg)498 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
499 {
500 return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK);
501 }
502
type_is_rdonly_mem(u32 type)503 static bool type_is_rdonly_mem(u32 type)
504 {
505 return type & MEM_RDONLY;
506 }
507
is_acquire_function(enum bpf_func_id func_id,const struct bpf_map * map)508 static bool is_acquire_function(enum bpf_func_id func_id,
509 const struct bpf_map *map)
510 {
511 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
512
513 if (func_id == BPF_FUNC_sk_lookup_tcp ||
514 func_id == BPF_FUNC_sk_lookup_udp ||
515 func_id == BPF_FUNC_skc_lookup_tcp ||
516 func_id == BPF_FUNC_ringbuf_reserve ||
517 func_id == BPF_FUNC_kptr_xchg)
518 return true;
519
520 if (func_id == BPF_FUNC_map_lookup_elem &&
521 (map_type == BPF_MAP_TYPE_SOCKMAP ||
522 map_type == BPF_MAP_TYPE_SOCKHASH))
523 return true;
524
525 return false;
526 }
527
is_ptr_cast_function(enum bpf_func_id func_id)528 static bool is_ptr_cast_function(enum bpf_func_id func_id)
529 {
530 return func_id == BPF_FUNC_tcp_sock ||
531 func_id == BPF_FUNC_sk_fullsock ||
532 func_id == BPF_FUNC_skc_to_tcp_sock ||
533 func_id == BPF_FUNC_skc_to_tcp6_sock ||
534 func_id == BPF_FUNC_skc_to_udp6_sock ||
535 func_id == BPF_FUNC_skc_to_mptcp_sock ||
536 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
537 func_id == BPF_FUNC_skc_to_tcp_request_sock;
538 }
539
is_dynptr_ref_function(enum bpf_func_id func_id)540 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
541 {
542 return func_id == BPF_FUNC_dynptr_data;
543 }
544
545 static bool is_sync_callback_calling_kfunc(u32 btf_id);
546
is_sync_callback_calling_function(enum bpf_func_id func_id)547 static bool is_sync_callback_calling_function(enum bpf_func_id func_id)
548 {
549 return func_id == BPF_FUNC_for_each_map_elem ||
550 func_id == BPF_FUNC_find_vma ||
551 func_id == BPF_FUNC_loop ||
552 func_id == BPF_FUNC_user_ringbuf_drain;
553 }
554
is_async_callback_calling_function(enum bpf_func_id func_id)555 static bool is_async_callback_calling_function(enum bpf_func_id func_id)
556 {
557 return func_id == BPF_FUNC_timer_set_callback;
558 }
559
is_callback_calling_function(enum bpf_func_id func_id)560 static bool is_callback_calling_function(enum bpf_func_id func_id)
561 {
562 return is_sync_callback_calling_function(func_id) ||
563 is_async_callback_calling_function(func_id);
564 }
565
is_sync_callback_calling_insn(struct bpf_insn * insn)566 static bool is_sync_callback_calling_insn(struct bpf_insn *insn)
567 {
568 return (bpf_helper_call(insn) && is_sync_callback_calling_function(insn->imm)) ||
569 (bpf_pseudo_kfunc_call(insn) && is_sync_callback_calling_kfunc(insn->imm));
570 }
571
is_storage_get_function(enum bpf_func_id func_id)572 static bool is_storage_get_function(enum bpf_func_id func_id)
573 {
574 return func_id == BPF_FUNC_sk_storage_get ||
575 func_id == BPF_FUNC_inode_storage_get ||
576 func_id == BPF_FUNC_task_storage_get ||
577 func_id == BPF_FUNC_cgrp_storage_get;
578 }
579
helper_multiple_ref_obj_use(enum bpf_func_id func_id,const struct bpf_map * map)580 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
581 const struct bpf_map *map)
582 {
583 int ref_obj_uses = 0;
584
585 if (is_ptr_cast_function(func_id))
586 ref_obj_uses++;
587 if (is_acquire_function(func_id, map))
588 ref_obj_uses++;
589 if (is_dynptr_ref_function(func_id))
590 ref_obj_uses++;
591
592 return ref_obj_uses > 1;
593 }
594
is_cmpxchg_insn(const struct bpf_insn * insn)595 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
596 {
597 return BPF_CLASS(insn->code) == BPF_STX &&
598 BPF_MODE(insn->code) == BPF_ATOMIC &&
599 insn->imm == BPF_CMPXCHG;
600 }
601
602 /* string representation of 'enum bpf_reg_type'
603 *
604 * Note that reg_type_str() can not appear more than once in a single verbose()
605 * statement.
606 */
reg_type_str(struct bpf_verifier_env * env,enum bpf_reg_type type)607 static const char *reg_type_str(struct bpf_verifier_env *env,
608 enum bpf_reg_type type)
609 {
610 char postfix[16] = {0}, prefix[64] = {0};
611 static const char * const str[] = {
612 [NOT_INIT] = "?",
613 [SCALAR_VALUE] = "scalar",
614 [PTR_TO_CTX] = "ctx",
615 [CONST_PTR_TO_MAP] = "map_ptr",
616 [PTR_TO_MAP_VALUE] = "map_value",
617 [PTR_TO_STACK] = "fp",
618 [PTR_TO_PACKET] = "pkt",
619 [PTR_TO_PACKET_META] = "pkt_meta",
620 [PTR_TO_PACKET_END] = "pkt_end",
621 [PTR_TO_FLOW_KEYS] = "flow_keys",
622 [PTR_TO_SOCKET] = "sock",
623 [PTR_TO_SOCK_COMMON] = "sock_common",
624 [PTR_TO_TCP_SOCK] = "tcp_sock",
625 [PTR_TO_TP_BUFFER] = "tp_buffer",
626 [PTR_TO_XDP_SOCK] = "xdp_sock",
627 [PTR_TO_BTF_ID] = "ptr_",
628 [PTR_TO_MEM] = "mem",
629 [PTR_TO_BUF] = "buf",
630 [PTR_TO_FUNC] = "func",
631 [PTR_TO_MAP_KEY] = "map_key",
632 [CONST_PTR_TO_DYNPTR] = "dynptr_ptr",
633 };
634
635 if (type & PTR_MAYBE_NULL) {
636 if (base_type(type) == PTR_TO_BTF_ID)
637 strncpy(postfix, "or_null_", 16);
638 else
639 strncpy(postfix, "_or_null", 16);
640 }
641
642 snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s",
643 type & MEM_RDONLY ? "rdonly_" : "",
644 type & MEM_RINGBUF ? "ringbuf_" : "",
645 type & MEM_USER ? "user_" : "",
646 type & MEM_PERCPU ? "percpu_" : "",
647 type & MEM_RCU ? "rcu_" : "",
648 type & PTR_UNTRUSTED ? "untrusted_" : "",
649 type & PTR_TRUSTED ? "trusted_" : ""
650 );
651
652 snprintf(env->tmp_str_buf, TMP_STR_BUF_LEN, "%s%s%s",
653 prefix, str[base_type(type)], postfix);
654 return env->tmp_str_buf;
655 }
656
657 static char slot_type_char[] = {
658 [STACK_INVALID] = '?',
659 [STACK_SPILL] = 'r',
660 [STACK_MISC] = 'm',
661 [STACK_ZERO] = '0',
662 [STACK_DYNPTR] = 'd',
663 [STACK_ITER] = 'i',
664 };
665
print_liveness(struct bpf_verifier_env * env,enum bpf_reg_liveness live)666 static void print_liveness(struct bpf_verifier_env *env,
667 enum bpf_reg_liveness live)
668 {
669 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
670 verbose(env, "_");
671 if (live & REG_LIVE_READ)
672 verbose(env, "r");
673 if (live & REG_LIVE_WRITTEN)
674 verbose(env, "w");
675 if (live & REG_LIVE_DONE)
676 verbose(env, "D");
677 }
678
__get_spi(s32 off)679 static int __get_spi(s32 off)
680 {
681 return (-off - 1) / BPF_REG_SIZE;
682 }
683
func(struct bpf_verifier_env * env,const struct bpf_reg_state * reg)684 static struct bpf_func_state *func(struct bpf_verifier_env *env,
685 const struct bpf_reg_state *reg)
686 {
687 struct bpf_verifier_state *cur = env->cur_state;
688
689 return cur->frame[reg->frameno];
690 }
691
is_spi_bounds_valid(struct bpf_func_state * state,int spi,int nr_slots)692 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
693 {
694 int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
695
696 /* We need to check that slots between [spi - nr_slots + 1, spi] are
697 * within [0, allocated_stack).
698 *
699 * Please note that the spi grows downwards. For example, a dynptr
700 * takes the size of two stack slots; the first slot will be at
701 * spi and the second slot will be at spi - 1.
702 */
703 return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
704 }
705
stack_slot_obj_get_spi(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * obj_kind,int nr_slots)706 static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
707 const char *obj_kind, int nr_slots)
708 {
709 int off, spi;
710
711 if (!tnum_is_const(reg->var_off)) {
712 verbose(env, "%s has to be at a constant offset\n", obj_kind);
713 return -EINVAL;
714 }
715
716 off = reg->off + reg->var_off.value;
717 if (off % BPF_REG_SIZE) {
718 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
719 return -EINVAL;
720 }
721
722 spi = __get_spi(off);
723 if (spi + 1 < nr_slots) {
724 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
725 return -EINVAL;
726 }
727
728 if (!is_spi_bounds_valid(func(env, reg), spi, nr_slots))
729 return -ERANGE;
730 return spi;
731 }
732
dynptr_get_spi(struct bpf_verifier_env * env,struct bpf_reg_state * reg)733 static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
734 {
735 return stack_slot_obj_get_spi(env, reg, "dynptr", BPF_DYNPTR_NR_SLOTS);
736 }
737
iter_get_spi(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int nr_slots)738 static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots)
739 {
740 return stack_slot_obj_get_spi(env, reg, "iter", nr_slots);
741 }
742
btf_type_name(const struct btf * btf,u32 id)743 static const char *btf_type_name(const struct btf *btf, u32 id)
744 {
745 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
746 }
747
dynptr_type_str(enum bpf_dynptr_type type)748 static const char *dynptr_type_str(enum bpf_dynptr_type type)
749 {
750 switch (type) {
751 case BPF_DYNPTR_TYPE_LOCAL:
752 return "local";
753 case BPF_DYNPTR_TYPE_RINGBUF:
754 return "ringbuf";
755 case BPF_DYNPTR_TYPE_SKB:
756 return "skb";
757 case BPF_DYNPTR_TYPE_XDP:
758 return "xdp";
759 case BPF_DYNPTR_TYPE_INVALID:
760 return "<invalid>";
761 default:
762 WARN_ONCE(1, "unknown dynptr type %d\n", type);
763 return "<unknown>";
764 }
765 }
766
iter_type_str(const struct btf * btf,u32 btf_id)767 static const char *iter_type_str(const struct btf *btf, u32 btf_id)
768 {
769 if (!btf || btf_id == 0)
770 return "<invalid>";
771
772 /* we already validated that type is valid and has conforming name */
773 return btf_type_name(btf, btf_id) + sizeof(ITER_PREFIX) - 1;
774 }
775
iter_state_str(enum bpf_iter_state state)776 static const char *iter_state_str(enum bpf_iter_state state)
777 {
778 switch (state) {
779 case BPF_ITER_STATE_ACTIVE:
780 return "active";
781 case BPF_ITER_STATE_DRAINED:
782 return "drained";
783 case BPF_ITER_STATE_INVALID:
784 return "<invalid>";
785 default:
786 WARN_ONCE(1, "unknown iter state %d\n", state);
787 return "<unknown>";
788 }
789 }
790
mark_reg_scratched(struct bpf_verifier_env * env,u32 regno)791 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
792 {
793 env->scratched_regs |= 1U << regno;
794 }
795
mark_stack_slot_scratched(struct bpf_verifier_env * env,u32 spi)796 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
797 {
798 env->scratched_stack_slots |= 1ULL << spi;
799 }
800
reg_scratched(const struct bpf_verifier_env * env,u32 regno)801 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
802 {
803 return (env->scratched_regs >> regno) & 1;
804 }
805
stack_slot_scratched(const struct bpf_verifier_env * env,u64 regno)806 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
807 {
808 return (env->scratched_stack_slots >> regno) & 1;
809 }
810
verifier_state_scratched(const struct bpf_verifier_env * env)811 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
812 {
813 return env->scratched_regs || env->scratched_stack_slots;
814 }
815
mark_verifier_state_clean(struct bpf_verifier_env * env)816 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
817 {
818 env->scratched_regs = 0U;
819 env->scratched_stack_slots = 0ULL;
820 }
821
822 /* Used for printing the entire verifier state. */
mark_verifier_state_scratched(struct bpf_verifier_env * env)823 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
824 {
825 env->scratched_regs = ~0U;
826 env->scratched_stack_slots = ~0ULL;
827 }
828
arg_to_dynptr_type(enum bpf_arg_type arg_type)829 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
830 {
831 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
832 case DYNPTR_TYPE_LOCAL:
833 return BPF_DYNPTR_TYPE_LOCAL;
834 case DYNPTR_TYPE_RINGBUF:
835 return BPF_DYNPTR_TYPE_RINGBUF;
836 case DYNPTR_TYPE_SKB:
837 return BPF_DYNPTR_TYPE_SKB;
838 case DYNPTR_TYPE_XDP:
839 return BPF_DYNPTR_TYPE_XDP;
840 default:
841 return BPF_DYNPTR_TYPE_INVALID;
842 }
843 }
844
get_dynptr_type_flag(enum bpf_dynptr_type type)845 static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type)
846 {
847 switch (type) {
848 case BPF_DYNPTR_TYPE_LOCAL:
849 return DYNPTR_TYPE_LOCAL;
850 case BPF_DYNPTR_TYPE_RINGBUF:
851 return DYNPTR_TYPE_RINGBUF;
852 case BPF_DYNPTR_TYPE_SKB:
853 return DYNPTR_TYPE_SKB;
854 case BPF_DYNPTR_TYPE_XDP:
855 return DYNPTR_TYPE_XDP;
856 default:
857 return 0;
858 }
859 }
860
dynptr_type_refcounted(enum bpf_dynptr_type type)861 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
862 {
863 return type == BPF_DYNPTR_TYPE_RINGBUF;
864 }
865
866 static void __mark_dynptr_reg(struct bpf_reg_state *reg,
867 enum bpf_dynptr_type type,
868 bool first_slot, int dynptr_id);
869
870 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
871 struct bpf_reg_state *reg);
872
mark_dynptr_stack_regs(struct bpf_verifier_env * env,struct bpf_reg_state * sreg1,struct bpf_reg_state * sreg2,enum bpf_dynptr_type type)873 static void mark_dynptr_stack_regs(struct bpf_verifier_env *env,
874 struct bpf_reg_state *sreg1,
875 struct bpf_reg_state *sreg2,
876 enum bpf_dynptr_type type)
877 {
878 int id = ++env->id_gen;
879
880 __mark_dynptr_reg(sreg1, type, true, id);
881 __mark_dynptr_reg(sreg2, type, false, id);
882 }
883
mark_dynptr_cb_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,enum bpf_dynptr_type type)884 static void mark_dynptr_cb_reg(struct bpf_verifier_env *env,
885 struct bpf_reg_state *reg,
886 enum bpf_dynptr_type type)
887 {
888 __mark_dynptr_reg(reg, type, true, ++env->id_gen);
889 }
890
891 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
892 struct bpf_func_state *state, int spi);
893
mark_stack_slots_dynptr(struct bpf_verifier_env * env,struct bpf_reg_state * reg,enum bpf_arg_type arg_type,int insn_idx,int clone_ref_obj_id)894 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
895 enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id)
896 {
897 struct bpf_func_state *state = func(env, reg);
898 enum bpf_dynptr_type type;
899 int spi, i, err;
900
901 spi = dynptr_get_spi(env, reg);
902 if (spi < 0)
903 return spi;
904
905 /* We cannot assume both spi and spi - 1 belong to the same dynptr,
906 * hence we need to call destroy_if_dynptr_stack_slot twice for both,
907 * to ensure that for the following example:
908 * [d1][d1][d2][d2]
909 * spi 3 2 1 0
910 * So marking spi = 2 should lead to destruction of both d1 and d2. In
911 * case they do belong to same dynptr, second call won't see slot_type
912 * as STACK_DYNPTR and will simply skip destruction.
913 */
914 err = destroy_if_dynptr_stack_slot(env, state, spi);
915 if (err)
916 return err;
917 err = destroy_if_dynptr_stack_slot(env, state, spi - 1);
918 if (err)
919 return err;
920
921 for (i = 0; i < BPF_REG_SIZE; i++) {
922 state->stack[spi].slot_type[i] = STACK_DYNPTR;
923 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
924 }
925
926 type = arg_to_dynptr_type(arg_type);
927 if (type == BPF_DYNPTR_TYPE_INVALID)
928 return -EINVAL;
929
930 mark_dynptr_stack_regs(env, &state->stack[spi].spilled_ptr,
931 &state->stack[spi - 1].spilled_ptr, type);
932
933 if (dynptr_type_refcounted(type)) {
934 /* The id is used to track proper releasing */
935 int id;
936
937 if (clone_ref_obj_id)
938 id = clone_ref_obj_id;
939 else
940 id = acquire_reference_state(env, insn_idx);
941
942 if (id < 0)
943 return id;
944
945 state->stack[spi].spilled_ptr.ref_obj_id = id;
946 state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
947 }
948
949 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
950 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
951
952 return 0;
953 }
954
invalidate_dynptr(struct bpf_verifier_env * env,struct bpf_func_state * state,int spi)955 static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi)
956 {
957 int i;
958
959 for (i = 0; i < BPF_REG_SIZE; i++) {
960 state->stack[spi].slot_type[i] = STACK_INVALID;
961 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
962 }
963
964 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
965 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
966
967 /* Why do we need to set REG_LIVE_WRITTEN for STACK_INVALID slot?
968 *
969 * While we don't allow reading STACK_INVALID, it is still possible to
970 * do <8 byte writes marking some but not all slots as STACK_MISC. Then,
971 * helpers or insns can do partial read of that part without failing,
972 * but check_stack_range_initialized, check_stack_read_var_off, and
973 * check_stack_read_fixed_off will do mark_reg_read for all 8-bytes of
974 * the slot conservatively. Hence we need to prevent those liveness
975 * marking walks.
976 *
977 * This was not a problem before because STACK_INVALID is only set by
978 * default (where the default reg state has its reg->parent as NULL), or
979 * in clean_live_states after REG_LIVE_DONE (at which point
980 * mark_reg_read won't walk reg->parent chain), but not randomly during
981 * verifier state exploration (like we did above). Hence, for our case
982 * parentage chain will still be live (i.e. reg->parent may be
983 * non-NULL), while earlier reg->parent was NULL, so we need
984 * REG_LIVE_WRITTEN to screen off read marker propagation when it is
985 * done later on reads or by mark_dynptr_read as well to unnecessary
986 * mark registers in verifier state.
987 */
988 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
989 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
990 }
991
unmark_stack_slots_dynptr(struct bpf_verifier_env * env,struct bpf_reg_state * reg)992 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
993 {
994 struct bpf_func_state *state = func(env, reg);
995 int spi, ref_obj_id, i;
996
997 spi = dynptr_get_spi(env, reg);
998 if (spi < 0)
999 return spi;
1000
1001 if (!dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
1002 invalidate_dynptr(env, state, spi);
1003 return 0;
1004 }
1005
1006 ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id;
1007
1008 /* If the dynptr has a ref_obj_id, then we need to invalidate
1009 * two things:
1010 *
1011 * 1) Any dynptrs with a matching ref_obj_id (clones)
1012 * 2) Any slices derived from this dynptr.
1013 */
1014
1015 /* Invalidate any slices associated with this dynptr */
1016 WARN_ON_ONCE(release_reference(env, ref_obj_id));
1017
1018 /* Invalidate any dynptr clones */
1019 for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1020 if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
1021 continue;
1022
1023 /* it should always be the case that if the ref obj id
1024 * matches then the stack slot also belongs to a
1025 * dynptr
1026 */
1027 if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
1028 verbose(env, "verifier internal error: misconfigured ref_obj_id\n");
1029 return -EFAULT;
1030 }
1031 if (state->stack[i].spilled_ptr.dynptr.first_slot)
1032 invalidate_dynptr(env, state, i);
1033 }
1034
1035 return 0;
1036 }
1037
1038 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1039 struct bpf_reg_state *reg);
1040
mark_reg_invalid(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)1041 static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1042 {
1043 if (!env->allow_ptr_leaks)
1044 __mark_reg_not_init(env, reg);
1045 else
1046 __mark_reg_unknown(env, reg);
1047 }
1048
destroy_if_dynptr_stack_slot(struct bpf_verifier_env * env,struct bpf_func_state * state,int spi)1049 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
1050 struct bpf_func_state *state, int spi)
1051 {
1052 struct bpf_func_state *fstate;
1053 struct bpf_reg_state *dreg;
1054 int i, dynptr_id;
1055
1056 /* We always ensure that STACK_DYNPTR is never set partially,
1057 * hence just checking for slot_type[0] is enough. This is
1058 * different for STACK_SPILL, where it may be only set for
1059 * 1 byte, so code has to use is_spilled_reg.
1060 */
1061 if (state->stack[spi].slot_type[0] != STACK_DYNPTR)
1062 return 0;
1063
1064 /* Reposition spi to first slot */
1065 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1066 spi = spi + 1;
1067
1068 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
1069 verbose(env, "cannot overwrite referenced dynptr\n");
1070 return -EINVAL;
1071 }
1072
1073 mark_stack_slot_scratched(env, spi);
1074 mark_stack_slot_scratched(env, spi - 1);
1075
1076 /* Writing partially to one dynptr stack slot destroys both. */
1077 for (i = 0; i < BPF_REG_SIZE; i++) {
1078 state->stack[spi].slot_type[i] = STACK_INVALID;
1079 state->stack[spi - 1].slot_type[i] = STACK_INVALID;
1080 }
1081
1082 dynptr_id = state->stack[spi].spilled_ptr.id;
1083 /* Invalidate any slices associated with this dynptr */
1084 bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({
1085 /* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */
1086 if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM)
1087 continue;
1088 if (dreg->dynptr_id == dynptr_id)
1089 mark_reg_invalid(env, dreg);
1090 }));
1091
1092 /* Do not release reference state, we are destroying dynptr on stack,
1093 * not using some helper to release it. Just reset register.
1094 */
1095 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
1096 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
1097
1098 /* Same reason as unmark_stack_slots_dynptr above */
1099 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1100 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
1101
1102 return 0;
1103 }
1104
is_dynptr_reg_valid_uninit(struct bpf_verifier_env * env,struct bpf_reg_state * reg)1105 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1106 {
1107 int spi;
1108
1109 if (reg->type == CONST_PTR_TO_DYNPTR)
1110 return false;
1111
1112 spi = dynptr_get_spi(env, reg);
1113
1114 /* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an
1115 * error because this just means the stack state hasn't been updated yet.
1116 * We will do check_mem_access to check and update stack bounds later.
1117 */
1118 if (spi < 0 && spi != -ERANGE)
1119 return false;
1120
1121 /* We don't need to check if the stack slots are marked by previous
1122 * dynptr initializations because we allow overwriting existing unreferenced
1123 * STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls
1124 * destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are
1125 * touching are completely destructed before we reinitialize them for a new
1126 * one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early
1127 * instead of delaying it until the end where the user will get "Unreleased
1128 * reference" error.
1129 */
1130 return true;
1131 }
1132
is_dynptr_reg_valid_init(struct bpf_verifier_env * env,struct bpf_reg_state * reg)1133 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1134 {
1135 struct bpf_func_state *state = func(env, reg);
1136 int i, spi;
1137
1138 /* This already represents first slot of initialized bpf_dynptr.
1139 *
1140 * CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
1141 * check_func_arg_reg_off's logic, so we don't need to check its
1142 * offset and alignment.
1143 */
1144 if (reg->type == CONST_PTR_TO_DYNPTR)
1145 return true;
1146
1147 spi = dynptr_get_spi(env, reg);
1148 if (spi < 0)
1149 return false;
1150 if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
1151 return false;
1152
1153 for (i = 0; i < BPF_REG_SIZE; i++) {
1154 if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
1155 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
1156 return false;
1157 }
1158
1159 return true;
1160 }
1161
is_dynptr_type_expected(struct bpf_verifier_env * env,struct bpf_reg_state * reg,enum bpf_arg_type arg_type)1162 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1163 enum bpf_arg_type arg_type)
1164 {
1165 struct bpf_func_state *state = func(env, reg);
1166 enum bpf_dynptr_type dynptr_type;
1167 int spi;
1168
1169 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */
1170 if (arg_type == ARG_PTR_TO_DYNPTR)
1171 return true;
1172
1173 dynptr_type = arg_to_dynptr_type(arg_type);
1174 if (reg->type == CONST_PTR_TO_DYNPTR) {
1175 return reg->dynptr.type == dynptr_type;
1176 } else {
1177 spi = dynptr_get_spi(env, reg);
1178 if (spi < 0)
1179 return false;
1180 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
1181 }
1182 }
1183
1184 static void __mark_reg_known_zero(struct bpf_reg_state *reg);
1185
mark_stack_slots_iter(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int insn_idx,struct btf * btf,u32 btf_id,int nr_slots)1186 static int mark_stack_slots_iter(struct bpf_verifier_env *env,
1187 struct bpf_reg_state *reg, int insn_idx,
1188 struct btf *btf, u32 btf_id, int nr_slots)
1189 {
1190 struct bpf_func_state *state = func(env, reg);
1191 int spi, i, j, id;
1192
1193 spi = iter_get_spi(env, reg, nr_slots);
1194 if (spi < 0)
1195 return spi;
1196
1197 id = acquire_reference_state(env, insn_idx);
1198 if (id < 0)
1199 return id;
1200
1201 for (i = 0; i < nr_slots; i++) {
1202 struct bpf_stack_state *slot = &state->stack[spi - i];
1203 struct bpf_reg_state *st = &slot->spilled_ptr;
1204
1205 __mark_reg_known_zero(st);
1206 st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1207 st->live |= REG_LIVE_WRITTEN;
1208 st->ref_obj_id = i == 0 ? id : 0;
1209 st->iter.btf = btf;
1210 st->iter.btf_id = btf_id;
1211 st->iter.state = BPF_ITER_STATE_ACTIVE;
1212 st->iter.depth = 0;
1213
1214 for (j = 0; j < BPF_REG_SIZE; j++)
1215 slot->slot_type[j] = STACK_ITER;
1216
1217 mark_stack_slot_scratched(env, spi - i);
1218 }
1219
1220 return 0;
1221 }
1222
unmark_stack_slots_iter(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int nr_slots)1223 static int unmark_stack_slots_iter(struct bpf_verifier_env *env,
1224 struct bpf_reg_state *reg, int nr_slots)
1225 {
1226 struct bpf_func_state *state = func(env, reg);
1227 int spi, i, j;
1228
1229 spi = iter_get_spi(env, reg, nr_slots);
1230 if (spi < 0)
1231 return spi;
1232
1233 for (i = 0; i < nr_slots; i++) {
1234 struct bpf_stack_state *slot = &state->stack[spi - i];
1235 struct bpf_reg_state *st = &slot->spilled_ptr;
1236
1237 if (i == 0)
1238 WARN_ON_ONCE(release_reference(env, st->ref_obj_id));
1239
1240 __mark_reg_not_init(env, st);
1241
1242 /* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */
1243 st->live |= REG_LIVE_WRITTEN;
1244
1245 for (j = 0; j < BPF_REG_SIZE; j++)
1246 slot->slot_type[j] = STACK_INVALID;
1247
1248 mark_stack_slot_scratched(env, spi - i);
1249 }
1250
1251 return 0;
1252 }
1253
is_iter_reg_valid_uninit(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int nr_slots)1254 static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env,
1255 struct bpf_reg_state *reg, int nr_slots)
1256 {
1257 struct bpf_func_state *state = func(env, reg);
1258 int spi, i, j;
1259
1260 /* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1261 * will do check_mem_access to check and update stack bounds later, so
1262 * return true for that case.
1263 */
1264 spi = iter_get_spi(env, reg, nr_slots);
1265 if (spi == -ERANGE)
1266 return true;
1267 if (spi < 0)
1268 return false;
1269
1270 for (i = 0; i < nr_slots; i++) {
1271 struct bpf_stack_state *slot = &state->stack[spi - i];
1272
1273 for (j = 0; j < BPF_REG_SIZE; j++)
1274 if (slot->slot_type[j] == STACK_ITER)
1275 return false;
1276 }
1277
1278 return true;
1279 }
1280
is_iter_reg_valid_init(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct btf * btf,u32 btf_id,int nr_slots)1281 static bool is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1282 struct btf *btf, u32 btf_id, int nr_slots)
1283 {
1284 struct bpf_func_state *state = func(env, reg);
1285 int spi, i, j;
1286
1287 spi = iter_get_spi(env, reg, nr_slots);
1288 if (spi < 0)
1289 return false;
1290
1291 for (i = 0; i < nr_slots; i++) {
1292 struct bpf_stack_state *slot = &state->stack[spi - i];
1293 struct bpf_reg_state *st = &slot->spilled_ptr;
1294
1295 /* only main (first) slot has ref_obj_id set */
1296 if (i == 0 && !st->ref_obj_id)
1297 return false;
1298 if (i != 0 && st->ref_obj_id)
1299 return false;
1300 if (st->iter.btf != btf || st->iter.btf_id != btf_id)
1301 return false;
1302
1303 for (j = 0; j < BPF_REG_SIZE; j++)
1304 if (slot->slot_type[j] != STACK_ITER)
1305 return false;
1306 }
1307
1308 return true;
1309 }
1310
1311 /* Check if given stack slot is "special":
1312 * - spilled register state (STACK_SPILL);
1313 * - dynptr state (STACK_DYNPTR);
1314 * - iter state (STACK_ITER).
1315 */
is_stack_slot_special(const struct bpf_stack_state * stack)1316 static bool is_stack_slot_special(const struct bpf_stack_state *stack)
1317 {
1318 enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1];
1319
1320 switch (type) {
1321 case STACK_SPILL:
1322 case STACK_DYNPTR:
1323 case STACK_ITER:
1324 return true;
1325 case STACK_INVALID:
1326 case STACK_MISC:
1327 case STACK_ZERO:
1328 return false;
1329 default:
1330 WARN_ONCE(1, "unknown stack slot type %d\n", type);
1331 return true;
1332 }
1333 }
1334
1335 /* The reg state of a pointer or a bounded scalar was saved when
1336 * it was spilled to the stack.
1337 */
is_spilled_reg(const struct bpf_stack_state * stack)1338 static bool is_spilled_reg(const struct bpf_stack_state *stack)
1339 {
1340 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
1341 }
1342
is_spilled_scalar_reg(const struct bpf_stack_state * stack)1343 static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack)
1344 {
1345 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL &&
1346 stack->spilled_ptr.type == SCALAR_VALUE;
1347 }
1348
scrub_spilled_slot(u8 * stype)1349 static void scrub_spilled_slot(u8 *stype)
1350 {
1351 if (*stype != STACK_INVALID)
1352 *stype = STACK_MISC;
1353 }
1354
print_verifier_state(struct bpf_verifier_env * env,const struct bpf_func_state * state,bool print_all)1355 static void print_verifier_state(struct bpf_verifier_env *env,
1356 const struct bpf_func_state *state,
1357 bool print_all)
1358 {
1359 const struct bpf_reg_state *reg;
1360 enum bpf_reg_type t;
1361 int i;
1362
1363 if (state->frameno)
1364 verbose(env, " frame%d:", state->frameno);
1365 for (i = 0; i < MAX_BPF_REG; i++) {
1366 reg = &state->regs[i];
1367 t = reg->type;
1368 if (t == NOT_INIT)
1369 continue;
1370 if (!print_all && !reg_scratched(env, i))
1371 continue;
1372 verbose(env, " R%d", i);
1373 print_liveness(env, reg->live);
1374 verbose(env, "=");
1375 if (t == SCALAR_VALUE && reg->precise)
1376 verbose(env, "P");
1377 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
1378 tnum_is_const(reg->var_off)) {
1379 /* reg->off should be 0 for SCALAR_VALUE */
1380 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1381 verbose(env, "%lld", reg->var_off.value + reg->off);
1382 } else {
1383 const char *sep = "";
1384
1385 verbose(env, "%s", reg_type_str(env, t));
1386 if (base_type(t) == PTR_TO_BTF_ID)
1387 verbose(env, "%s", btf_type_name(reg->btf, reg->btf_id));
1388 verbose(env, "(");
1389 /*
1390 * _a stands for append, was shortened to avoid multiline statements below.
1391 * This macro is used to output a comma separated list of attributes.
1392 */
1393 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; })
1394
1395 if (reg->id)
1396 verbose_a("id=%d", reg->id);
1397 if (reg->ref_obj_id)
1398 verbose_a("ref_obj_id=%d", reg->ref_obj_id);
1399 if (type_is_non_owning_ref(reg->type))
1400 verbose_a("%s", "non_own_ref");
1401 if (t != SCALAR_VALUE)
1402 verbose_a("off=%d", reg->off);
1403 if (type_is_pkt_pointer(t))
1404 verbose_a("r=%d", reg->range);
1405 else if (base_type(t) == CONST_PTR_TO_MAP ||
1406 base_type(t) == PTR_TO_MAP_KEY ||
1407 base_type(t) == PTR_TO_MAP_VALUE)
1408 verbose_a("ks=%d,vs=%d",
1409 reg->map_ptr->key_size,
1410 reg->map_ptr->value_size);
1411 if (tnum_is_const(reg->var_off)) {
1412 /* Typically an immediate SCALAR_VALUE, but
1413 * could be a pointer whose offset is too big
1414 * for reg->off
1415 */
1416 verbose_a("imm=%llx", reg->var_off.value);
1417 } else {
1418 if (reg->smin_value != reg->umin_value &&
1419 reg->smin_value != S64_MIN)
1420 verbose_a("smin=%lld", (long long)reg->smin_value);
1421 if (reg->smax_value != reg->umax_value &&
1422 reg->smax_value != S64_MAX)
1423 verbose_a("smax=%lld", (long long)reg->smax_value);
1424 if (reg->umin_value != 0)
1425 verbose_a("umin=%llu", (unsigned long long)reg->umin_value);
1426 if (reg->umax_value != U64_MAX)
1427 verbose_a("umax=%llu", (unsigned long long)reg->umax_value);
1428 if (!tnum_is_unknown(reg->var_off)) {
1429 char tn_buf[48];
1430
1431 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1432 verbose_a("var_off=%s", tn_buf);
1433 }
1434 if (reg->s32_min_value != reg->smin_value &&
1435 reg->s32_min_value != S32_MIN)
1436 verbose_a("s32_min=%d", (int)(reg->s32_min_value));
1437 if (reg->s32_max_value != reg->smax_value &&
1438 reg->s32_max_value != S32_MAX)
1439 verbose_a("s32_max=%d", (int)(reg->s32_max_value));
1440 if (reg->u32_min_value != reg->umin_value &&
1441 reg->u32_min_value != U32_MIN)
1442 verbose_a("u32_min=%d", (int)(reg->u32_min_value));
1443 if (reg->u32_max_value != reg->umax_value &&
1444 reg->u32_max_value != U32_MAX)
1445 verbose_a("u32_max=%d", (int)(reg->u32_max_value));
1446 }
1447 #undef verbose_a
1448
1449 verbose(env, ")");
1450 }
1451 }
1452 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1453 char types_buf[BPF_REG_SIZE + 1];
1454 bool valid = false;
1455 int j;
1456
1457 for (j = 0; j < BPF_REG_SIZE; j++) {
1458 if (state->stack[i].slot_type[j] != STACK_INVALID)
1459 valid = true;
1460 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1461 }
1462 types_buf[BPF_REG_SIZE] = 0;
1463 if (!valid)
1464 continue;
1465 if (!print_all && !stack_slot_scratched(env, i))
1466 continue;
1467 switch (state->stack[i].slot_type[BPF_REG_SIZE - 1]) {
1468 case STACK_SPILL:
1469 reg = &state->stack[i].spilled_ptr;
1470 t = reg->type;
1471
1472 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1473 print_liveness(env, reg->live);
1474 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t));
1475 if (t == SCALAR_VALUE && reg->precise)
1476 verbose(env, "P");
1477 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
1478 verbose(env, "%lld", reg->var_off.value + reg->off);
1479 break;
1480 case STACK_DYNPTR:
1481 i += BPF_DYNPTR_NR_SLOTS - 1;
1482 reg = &state->stack[i].spilled_ptr;
1483
1484 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1485 print_liveness(env, reg->live);
1486 verbose(env, "=dynptr_%s", dynptr_type_str(reg->dynptr.type));
1487 if (reg->ref_obj_id)
1488 verbose(env, "(ref_id=%d)", reg->ref_obj_id);
1489 break;
1490 case STACK_ITER:
1491 /* only main slot has ref_obj_id set; skip others */
1492 reg = &state->stack[i].spilled_ptr;
1493 if (!reg->ref_obj_id)
1494 continue;
1495
1496 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1497 print_liveness(env, reg->live);
1498 verbose(env, "=iter_%s(ref_id=%d,state=%s,depth=%u)",
1499 iter_type_str(reg->iter.btf, reg->iter.btf_id),
1500 reg->ref_obj_id, iter_state_str(reg->iter.state),
1501 reg->iter.depth);
1502 break;
1503 case STACK_MISC:
1504 case STACK_ZERO:
1505 default:
1506 reg = &state->stack[i].spilled_ptr;
1507
1508 for (j = 0; j < BPF_REG_SIZE; j++)
1509 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]];
1510 types_buf[BPF_REG_SIZE] = 0;
1511
1512 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
1513 print_liveness(env, reg->live);
1514 verbose(env, "=%s", types_buf);
1515 break;
1516 }
1517 }
1518 if (state->acquired_refs && state->refs[0].id) {
1519 verbose(env, " refs=%d", state->refs[0].id);
1520 for (i = 1; i < state->acquired_refs; i++)
1521 if (state->refs[i].id)
1522 verbose(env, ",%d", state->refs[i].id);
1523 }
1524 if (state->in_callback_fn)
1525 verbose(env, " cb");
1526 if (state->in_async_callback_fn)
1527 verbose(env, " async_cb");
1528 verbose(env, "\n");
1529 if (!print_all)
1530 mark_verifier_state_clean(env);
1531 }
1532
vlog_alignment(u32 pos)1533 static inline u32 vlog_alignment(u32 pos)
1534 {
1535 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
1536 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
1537 }
1538
print_insn_state(struct bpf_verifier_env * env,const struct bpf_func_state * state)1539 static void print_insn_state(struct bpf_verifier_env *env,
1540 const struct bpf_func_state *state)
1541 {
1542 if (env->prev_log_pos && env->prev_log_pos == env->log.end_pos) {
1543 /* remove new line character */
1544 bpf_vlog_reset(&env->log, env->prev_log_pos - 1);
1545 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_pos), ' ');
1546 } else {
1547 verbose(env, "%d:", env->insn_idx);
1548 }
1549 print_verifier_state(env, state, false);
1550 }
1551
1552 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1553 * small to hold src. This is different from krealloc since we don't want to preserve
1554 * the contents of dst.
1555 *
1556 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1557 * not be allocated.
1558 */
copy_array(void * dst,const void * src,size_t n,size_t size,gfp_t flags)1559 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1560 {
1561 size_t alloc_bytes;
1562 void *orig = dst;
1563 size_t bytes;
1564
1565 if (ZERO_OR_NULL_PTR(src))
1566 goto out;
1567
1568 if (unlikely(check_mul_overflow(n, size, &bytes)))
1569 return NULL;
1570
1571 alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1572 dst = krealloc(orig, alloc_bytes, flags);
1573 if (!dst) {
1574 kfree(orig);
1575 return NULL;
1576 }
1577
1578 memcpy(dst, src, bytes);
1579 out:
1580 return dst ? dst : ZERO_SIZE_PTR;
1581 }
1582
1583 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1584 * small to hold new_n items. new items are zeroed out if the array grows.
1585 *
1586 * Contrary to krealloc_array, does not free arr if new_n is zero.
1587 */
realloc_array(void * arr,size_t old_n,size_t new_n,size_t size)1588 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1589 {
1590 size_t alloc_size;
1591 void *new_arr;
1592
1593 if (!new_n || old_n == new_n)
1594 goto out;
1595
1596 alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1597 new_arr = krealloc(arr, alloc_size, GFP_KERNEL);
1598 if (!new_arr) {
1599 kfree(arr);
1600 return NULL;
1601 }
1602 arr = new_arr;
1603
1604 if (new_n > old_n)
1605 memset(arr + old_n * size, 0, (new_n - old_n) * size);
1606
1607 out:
1608 return arr ? arr : ZERO_SIZE_PTR;
1609 }
1610
copy_reference_state(struct bpf_func_state * dst,const struct bpf_func_state * src)1611 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1612 {
1613 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1614 sizeof(struct bpf_reference_state), GFP_KERNEL);
1615 if (!dst->refs)
1616 return -ENOMEM;
1617
1618 dst->acquired_refs = src->acquired_refs;
1619 return 0;
1620 }
1621
copy_stack_state(struct bpf_func_state * dst,const struct bpf_func_state * src)1622 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1623 {
1624 size_t n = src->allocated_stack / BPF_REG_SIZE;
1625
1626 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1627 GFP_KERNEL);
1628 if (!dst->stack)
1629 return -ENOMEM;
1630
1631 dst->allocated_stack = src->allocated_stack;
1632 return 0;
1633 }
1634
resize_reference_state(struct bpf_func_state * state,size_t n)1635 static int resize_reference_state(struct bpf_func_state *state, size_t n)
1636 {
1637 state->refs = realloc_array(state->refs, state->acquired_refs, n,
1638 sizeof(struct bpf_reference_state));
1639 if (!state->refs)
1640 return -ENOMEM;
1641
1642 state->acquired_refs = n;
1643 return 0;
1644 }
1645
1646 /* Possibly update state->allocated_stack to be at least size bytes. Also
1647 * possibly update the function's high-water mark in its bpf_subprog_info.
1648 */
grow_stack_state(struct bpf_verifier_env * env,struct bpf_func_state * state,int size)1649 static int grow_stack_state(struct bpf_verifier_env *env, struct bpf_func_state *state, int size)
1650 {
1651 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
1652
1653 if (old_n >= n)
1654 return 0;
1655
1656 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1657 if (!state->stack)
1658 return -ENOMEM;
1659
1660 state->allocated_stack = size;
1661
1662 /* update known max for given subprogram */
1663 if (env->subprog_info[state->subprogno].stack_depth < size)
1664 env->subprog_info[state->subprogno].stack_depth = size;
1665
1666 return 0;
1667 }
1668
1669 /* Acquire a pointer id from the env and update the state->refs to include
1670 * this new pointer reference.
1671 * On success, returns a valid pointer id to associate with the register
1672 * On failure, returns a negative errno.
1673 */
acquire_reference_state(struct bpf_verifier_env * env,int insn_idx)1674 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1675 {
1676 struct bpf_func_state *state = cur_func(env);
1677 int new_ofs = state->acquired_refs;
1678 int id, err;
1679
1680 err = resize_reference_state(state, state->acquired_refs + 1);
1681 if (err)
1682 return err;
1683 id = ++env->id_gen;
1684 state->refs[new_ofs].id = id;
1685 state->refs[new_ofs].insn_idx = insn_idx;
1686 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
1687
1688 return id;
1689 }
1690
1691 /* release function corresponding to acquire_reference_state(). Idempotent. */
release_reference_state(struct bpf_func_state * state,int ptr_id)1692 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
1693 {
1694 int i, last_idx;
1695
1696 last_idx = state->acquired_refs - 1;
1697 for (i = 0; i < state->acquired_refs; i++) {
1698 if (state->refs[i].id == ptr_id) {
1699 /* Cannot release caller references in callbacks */
1700 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
1701 return -EINVAL;
1702 if (last_idx && i != last_idx)
1703 memcpy(&state->refs[i], &state->refs[last_idx],
1704 sizeof(*state->refs));
1705 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1706 state->acquired_refs--;
1707 return 0;
1708 }
1709 }
1710 return -EINVAL;
1711 }
1712
free_func_state(struct bpf_func_state * state)1713 static void free_func_state(struct bpf_func_state *state)
1714 {
1715 if (!state)
1716 return;
1717 kfree(state->refs);
1718 kfree(state->stack);
1719 kfree(state);
1720 }
1721
clear_jmp_history(struct bpf_verifier_state * state)1722 static void clear_jmp_history(struct bpf_verifier_state *state)
1723 {
1724 kfree(state->jmp_history);
1725 state->jmp_history = NULL;
1726 state->jmp_history_cnt = 0;
1727 }
1728
free_verifier_state(struct bpf_verifier_state * state,bool free_self)1729 static void free_verifier_state(struct bpf_verifier_state *state,
1730 bool free_self)
1731 {
1732 int i;
1733
1734 for (i = 0; i <= state->curframe; i++) {
1735 free_func_state(state->frame[i]);
1736 state->frame[i] = NULL;
1737 }
1738 clear_jmp_history(state);
1739 if (free_self)
1740 kfree(state);
1741 }
1742
1743 /* copy verifier state from src to dst growing dst stack space
1744 * when necessary to accommodate larger src stack
1745 */
copy_func_state(struct bpf_func_state * dst,const struct bpf_func_state * src)1746 static int copy_func_state(struct bpf_func_state *dst,
1747 const struct bpf_func_state *src)
1748 {
1749 int err;
1750
1751 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
1752 err = copy_reference_state(dst, src);
1753 if (err)
1754 return err;
1755 return copy_stack_state(dst, src);
1756 }
1757
copy_verifier_state(struct bpf_verifier_state * dst_state,const struct bpf_verifier_state * src)1758 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1759 const struct bpf_verifier_state *src)
1760 {
1761 struct bpf_func_state *dst;
1762 int i, err;
1763
1764 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1765 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1766 GFP_USER);
1767 if (!dst_state->jmp_history)
1768 return -ENOMEM;
1769 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1770
1771 /* if dst has more stack frames then src frame, free them */
1772 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1773 free_func_state(dst_state->frame[i]);
1774 dst_state->frame[i] = NULL;
1775 }
1776 dst_state->speculative = src->speculative;
1777 dst_state->active_rcu_lock = src->active_rcu_lock;
1778 dst_state->curframe = src->curframe;
1779 dst_state->active_lock.ptr = src->active_lock.ptr;
1780 dst_state->active_lock.id = src->active_lock.id;
1781 dst_state->branches = src->branches;
1782 dst_state->parent = src->parent;
1783 dst_state->first_insn_idx = src->first_insn_idx;
1784 dst_state->last_insn_idx = src->last_insn_idx;
1785 dst_state->dfs_depth = src->dfs_depth;
1786 dst_state->callback_unroll_depth = src->callback_unroll_depth;
1787 dst_state->used_as_loop_entry = src->used_as_loop_entry;
1788 for (i = 0; i <= src->curframe; i++) {
1789 dst = dst_state->frame[i];
1790 if (!dst) {
1791 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1792 if (!dst)
1793 return -ENOMEM;
1794 dst_state->frame[i] = dst;
1795 }
1796 err = copy_func_state(dst, src->frame[i]);
1797 if (err)
1798 return err;
1799 }
1800 return 0;
1801 }
1802
state_htab_size(struct bpf_verifier_env * env)1803 static u32 state_htab_size(struct bpf_verifier_env *env)
1804 {
1805 return env->prog->len;
1806 }
1807
explored_state(struct bpf_verifier_env * env,int idx)1808 static struct bpf_verifier_state_list **explored_state(struct bpf_verifier_env *env, int idx)
1809 {
1810 struct bpf_verifier_state *cur = env->cur_state;
1811 struct bpf_func_state *state = cur->frame[cur->curframe];
1812
1813 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
1814 }
1815
same_callsites(struct bpf_verifier_state * a,struct bpf_verifier_state * b)1816 static bool same_callsites(struct bpf_verifier_state *a, struct bpf_verifier_state *b)
1817 {
1818 int fr;
1819
1820 if (a->curframe != b->curframe)
1821 return false;
1822
1823 for (fr = a->curframe; fr >= 0; fr--)
1824 if (a->frame[fr]->callsite != b->frame[fr]->callsite)
1825 return false;
1826
1827 return true;
1828 }
1829
1830 /* Open coded iterators allow back-edges in the state graph in order to
1831 * check unbounded loops that iterators.
1832 *
1833 * In is_state_visited() it is necessary to know if explored states are
1834 * part of some loops in order to decide whether non-exact states
1835 * comparison could be used:
1836 * - non-exact states comparison establishes sub-state relation and uses
1837 * read and precision marks to do so, these marks are propagated from
1838 * children states and thus are not guaranteed to be final in a loop;
1839 * - exact states comparison just checks if current and explored states
1840 * are identical (and thus form a back-edge).
1841 *
1842 * Paper "A New Algorithm for Identifying Loops in Decompilation"
1843 * by Tao Wei, Jian Mao, Wei Zou and Yu Chen [1] presents a convenient
1844 * algorithm for loop structure detection and gives an overview of
1845 * relevant terminology. It also has helpful illustrations.
1846 *
1847 * [1] https://api.semanticscholar.org/CorpusID:15784067
1848 *
1849 * We use a similar algorithm but because loop nested structure is
1850 * irrelevant for verifier ours is significantly simpler and resembles
1851 * strongly connected components algorithm from Sedgewick's textbook.
1852 *
1853 * Define topmost loop entry as a first node of the loop traversed in a
1854 * depth first search starting from initial state. The goal of the loop
1855 * tracking algorithm is to associate topmost loop entries with states
1856 * derived from these entries.
1857 *
1858 * For each step in the DFS states traversal algorithm needs to identify
1859 * the following situations:
1860 *
1861 * initial initial initial
1862 * | | |
1863 * V V V
1864 * ... ... .---------> hdr
1865 * | | | |
1866 * V V | V
1867 * cur .-> succ | .------...
1868 * | | | | | |
1869 * V | V | V V
1870 * succ '-- cur | ... ...
1871 * | | |
1872 * | V V
1873 * | succ <- cur
1874 * | |
1875 * | V
1876 * | ...
1877 * | |
1878 * '----'
1879 *
1880 * (A) successor state of cur (B) successor state of cur or it's entry
1881 * not yet traversed are in current DFS path, thus cur and succ
1882 * are members of the same outermost loop
1883 *
1884 * initial initial
1885 * | |
1886 * V V
1887 * ... ...
1888 * | |
1889 * V V
1890 * .------... .------...
1891 * | | | |
1892 * V V V V
1893 * .-> hdr ... ... ...
1894 * | | | | |
1895 * | V V V V
1896 * | succ <- cur succ <- cur
1897 * | | |
1898 * | V V
1899 * | ... ...
1900 * | | |
1901 * '----' exit
1902 *
1903 * (C) successor state of cur is a part of some loop but this loop
1904 * does not include cur or successor state is not in a loop at all.
1905 *
1906 * Algorithm could be described as the following python code:
1907 *
1908 * traversed = set() # Set of traversed nodes
1909 * entries = {} # Mapping from node to loop entry
1910 * depths = {} # Depth level assigned to graph node
1911 * path = set() # Current DFS path
1912 *
1913 * # Find outermost loop entry known for n
1914 * def get_loop_entry(n):
1915 * h = entries.get(n, None)
1916 * while h in entries and entries[h] != h:
1917 * h = entries[h]
1918 * return h
1919 *
1920 * # Update n's loop entry if h's outermost entry comes
1921 * # before n's outermost entry in current DFS path.
1922 * def update_loop_entry(n, h):
1923 * n1 = get_loop_entry(n) or n
1924 * h1 = get_loop_entry(h) or h
1925 * if h1 in path and depths[h1] <= depths[n1]:
1926 * entries[n] = h1
1927 *
1928 * def dfs(n, depth):
1929 * traversed.add(n)
1930 * path.add(n)
1931 * depths[n] = depth
1932 * for succ in G.successors(n):
1933 * if succ not in traversed:
1934 * # Case A: explore succ and update cur's loop entry
1935 * # only if succ's entry is in current DFS path.
1936 * dfs(succ, depth + 1)
1937 * h = get_loop_entry(succ)
1938 * update_loop_entry(n, h)
1939 * else:
1940 * # Case B or C depending on `h1 in path` check in update_loop_entry().
1941 * update_loop_entry(n, succ)
1942 * path.remove(n)
1943 *
1944 * To adapt this algorithm for use with verifier:
1945 * - use st->branch == 0 as a signal that DFS of succ had been finished
1946 * and cur's loop entry has to be updated (case A), handle this in
1947 * update_branch_counts();
1948 * - use st->branch > 0 as a signal that st is in the current DFS path;
1949 * - handle cases B and C in is_state_visited();
1950 * - update topmost loop entry for intermediate states in get_loop_entry().
1951 */
get_loop_entry(struct bpf_verifier_state * st)1952 static struct bpf_verifier_state *get_loop_entry(struct bpf_verifier_state *st)
1953 {
1954 struct bpf_verifier_state *topmost = st->loop_entry, *old;
1955
1956 while (topmost && topmost->loop_entry && topmost != topmost->loop_entry)
1957 topmost = topmost->loop_entry;
1958 /* Update loop entries for intermediate states to avoid this
1959 * traversal in future get_loop_entry() calls.
1960 */
1961 while (st && st->loop_entry != topmost) {
1962 old = st->loop_entry;
1963 st->loop_entry = topmost;
1964 st = old;
1965 }
1966 return topmost;
1967 }
1968
update_loop_entry(struct bpf_verifier_state * cur,struct bpf_verifier_state * hdr)1969 static void update_loop_entry(struct bpf_verifier_state *cur, struct bpf_verifier_state *hdr)
1970 {
1971 struct bpf_verifier_state *cur1, *hdr1;
1972
1973 cur1 = get_loop_entry(cur) ?: cur;
1974 hdr1 = get_loop_entry(hdr) ?: hdr;
1975 /* The head1->branches check decides between cases B and C in
1976 * comment for get_loop_entry(). If hdr1->branches == 0 then
1977 * head's topmost loop entry is not in current DFS path,
1978 * hence 'cur' and 'hdr' are not in the same loop and there is
1979 * no need to update cur->loop_entry.
1980 */
1981 if (hdr1->branches && hdr1->dfs_depth <= cur1->dfs_depth) {
1982 cur->loop_entry = hdr;
1983 hdr->used_as_loop_entry = true;
1984 }
1985 }
1986
update_branch_counts(struct bpf_verifier_env * env,struct bpf_verifier_state * st)1987 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1988 {
1989 while (st) {
1990 u32 br = --st->branches;
1991
1992 /* br == 0 signals that DFS exploration for 'st' is finished,
1993 * thus it is necessary to update parent's loop entry if it
1994 * turned out that st is a part of some loop.
1995 * This is a part of 'case A' in get_loop_entry() comment.
1996 */
1997 if (br == 0 && st->parent && st->loop_entry)
1998 update_loop_entry(st->parent, st->loop_entry);
1999
2000 /* WARN_ON(br > 1) technically makes sense here,
2001 * but see comment in push_stack(), hence:
2002 */
2003 WARN_ONCE((int)br < 0,
2004 "BUG update_branch_counts:branches_to_explore=%d\n",
2005 br);
2006 if (br)
2007 break;
2008 st = st->parent;
2009 }
2010 }
2011
pop_stack(struct bpf_verifier_env * env,int * prev_insn_idx,int * insn_idx,bool pop_log)2012 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
2013 int *insn_idx, bool pop_log)
2014 {
2015 struct bpf_verifier_state *cur = env->cur_state;
2016 struct bpf_verifier_stack_elem *elem, *head = env->head;
2017 int err;
2018
2019 if (env->head == NULL)
2020 return -ENOENT;
2021
2022 if (cur) {
2023 err = copy_verifier_state(cur, &head->st);
2024 if (err)
2025 return err;
2026 }
2027 if (pop_log)
2028 bpf_vlog_reset(&env->log, head->log_pos);
2029 if (insn_idx)
2030 *insn_idx = head->insn_idx;
2031 if (prev_insn_idx)
2032 *prev_insn_idx = head->prev_insn_idx;
2033 elem = head->next;
2034 free_verifier_state(&head->st, false);
2035 kfree(head);
2036 env->head = elem;
2037 env->stack_size--;
2038 return 0;
2039 }
2040
push_stack(struct bpf_verifier_env * env,int insn_idx,int prev_insn_idx,bool speculative)2041 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
2042 int insn_idx, int prev_insn_idx,
2043 bool speculative)
2044 {
2045 struct bpf_verifier_state *cur = env->cur_state;
2046 struct bpf_verifier_stack_elem *elem;
2047 int err;
2048
2049 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
2050 if (!elem)
2051 goto err;
2052
2053 elem->insn_idx = insn_idx;
2054 elem->prev_insn_idx = prev_insn_idx;
2055 elem->next = env->head;
2056 elem->log_pos = env->log.end_pos;
2057 env->head = elem;
2058 env->stack_size++;
2059 err = copy_verifier_state(&elem->st, cur);
2060 if (err)
2061 goto err;
2062 elem->st.speculative |= speculative;
2063 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2064 verbose(env, "The sequence of %d jumps is too complex.\n",
2065 env->stack_size);
2066 goto err;
2067 }
2068 if (elem->st.parent) {
2069 ++elem->st.parent->branches;
2070 /* WARN_ON(branches > 2) technically makes sense here,
2071 * but
2072 * 1. speculative states will bump 'branches' for non-branch
2073 * instructions
2074 * 2. is_state_visited() heuristics may decide not to create
2075 * a new state for a sequence of branches and all such current
2076 * and cloned states will be pointing to a single parent state
2077 * which might have large 'branches' count.
2078 */
2079 }
2080 return &elem->st;
2081 err:
2082 free_verifier_state(env->cur_state, true);
2083 env->cur_state = NULL;
2084 /* pop all elements and return */
2085 while (!pop_stack(env, NULL, NULL, false));
2086 return NULL;
2087 }
2088
2089 #define CALLER_SAVED_REGS 6
2090 static const int caller_saved[CALLER_SAVED_REGS] = {
2091 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
2092 };
2093
2094 /* This helper doesn't clear reg->id */
___mark_reg_known(struct bpf_reg_state * reg,u64 imm)2095 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
2096 {
2097 reg->var_off = tnum_const(imm);
2098 reg->smin_value = (s64)imm;
2099 reg->smax_value = (s64)imm;
2100 reg->umin_value = imm;
2101 reg->umax_value = imm;
2102
2103 reg->s32_min_value = (s32)imm;
2104 reg->s32_max_value = (s32)imm;
2105 reg->u32_min_value = (u32)imm;
2106 reg->u32_max_value = (u32)imm;
2107 }
2108
2109 /* Mark the unknown part of a register (variable offset or scalar value) as
2110 * known to have the value @imm.
2111 */
__mark_reg_known(struct bpf_reg_state * reg,u64 imm)2112 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
2113 {
2114 /* Clear off and union(map_ptr, range) */
2115 memset(((u8 *)reg) + sizeof(reg->type), 0,
2116 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
2117 reg->id = 0;
2118 reg->ref_obj_id = 0;
2119 ___mark_reg_known(reg, imm);
2120 }
2121
__mark_reg32_known(struct bpf_reg_state * reg,u64 imm)2122 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
2123 {
2124 reg->var_off = tnum_const_subreg(reg->var_off, imm);
2125 reg->s32_min_value = (s32)imm;
2126 reg->s32_max_value = (s32)imm;
2127 reg->u32_min_value = (u32)imm;
2128 reg->u32_max_value = (u32)imm;
2129 }
2130
2131 /* Mark the 'variable offset' part of a register as zero. This should be
2132 * used only on registers holding a pointer type.
2133 */
__mark_reg_known_zero(struct bpf_reg_state * reg)2134 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
2135 {
2136 __mark_reg_known(reg, 0);
2137 }
2138
__mark_reg_const_zero(struct bpf_reg_state * reg)2139 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
2140 {
2141 __mark_reg_known(reg, 0);
2142 reg->type = SCALAR_VALUE;
2143 }
2144
mark_reg_known_zero(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)2145 static void mark_reg_known_zero(struct bpf_verifier_env *env,
2146 struct bpf_reg_state *regs, u32 regno)
2147 {
2148 if (WARN_ON(regno >= MAX_BPF_REG)) {
2149 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
2150 /* Something bad happened, let's kill all regs */
2151 for (regno = 0; regno < MAX_BPF_REG; regno++)
2152 __mark_reg_not_init(env, regs + regno);
2153 return;
2154 }
2155 __mark_reg_known_zero(regs + regno);
2156 }
2157
__mark_dynptr_reg(struct bpf_reg_state * reg,enum bpf_dynptr_type type,bool first_slot,int dynptr_id)2158 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
2159 bool first_slot, int dynptr_id)
2160 {
2161 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
2162 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
2163 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
2164 */
2165 __mark_reg_known_zero(reg);
2166 reg->type = CONST_PTR_TO_DYNPTR;
2167 /* Give each dynptr a unique id to uniquely associate slices to it. */
2168 reg->id = dynptr_id;
2169 reg->dynptr.type = type;
2170 reg->dynptr.first_slot = first_slot;
2171 }
2172
mark_ptr_not_null_reg(struct bpf_reg_state * reg)2173 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
2174 {
2175 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
2176 const struct bpf_map *map = reg->map_ptr;
2177
2178 if (map->inner_map_meta) {
2179 reg->type = CONST_PTR_TO_MAP;
2180 reg->map_ptr = map->inner_map_meta;
2181 /* transfer reg's id which is unique for every map_lookup_elem
2182 * as UID of the inner map.
2183 */
2184 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
2185 reg->map_uid = reg->id;
2186 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
2187 reg->type = PTR_TO_XDP_SOCK;
2188 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
2189 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
2190 reg->type = PTR_TO_SOCKET;
2191 } else {
2192 reg->type = PTR_TO_MAP_VALUE;
2193 }
2194 return;
2195 }
2196
2197 reg->type &= ~PTR_MAYBE_NULL;
2198 }
2199
mark_reg_graph_node(struct bpf_reg_state * regs,u32 regno,struct btf_field_graph_root * ds_head)2200 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
2201 struct btf_field_graph_root *ds_head)
2202 {
2203 __mark_reg_known_zero(®s[regno]);
2204 regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
2205 regs[regno].btf = ds_head->btf;
2206 regs[regno].btf_id = ds_head->value_btf_id;
2207 regs[regno].off = ds_head->node_offset;
2208 }
2209
reg_is_pkt_pointer(const struct bpf_reg_state * reg)2210 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
2211 {
2212 return type_is_pkt_pointer(reg->type);
2213 }
2214
reg_is_pkt_pointer_any(const struct bpf_reg_state * reg)2215 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2216 {
2217 return reg_is_pkt_pointer(reg) ||
2218 reg->type == PTR_TO_PACKET_END;
2219 }
2220
reg_is_dynptr_slice_pkt(const struct bpf_reg_state * reg)2221 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2222 {
2223 return base_type(reg->type) == PTR_TO_MEM &&
2224 (reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP);
2225 }
2226
2227 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
reg_is_init_pkt_pointer(const struct bpf_reg_state * reg,enum bpf_reg_type which)2228 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2229 enum bpf_reg_type which)
2230 {
2231 /* The register can already have a range from prior markings.
2232 * This is fine as long as it hasn't been advanced from its
2233 * origin.
2234 */
2235 return reg->type == which &&
2236 reg->id == 0 &&
2237 reg->off == 0 &&
2238 tnum_equals_const(reg->var_off, 0);
2239 }
2240
2241 /* Reset the min/max bounds of a register */
__mark_reg_unbounded(struct bpf_reg_state * reg)2242 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2243 {
2244 reg->smin_value = S64_MIN;
2245 reg->smax_value = S64_MAX;
2246 reg->umin_value = 0;
2247 reg->umax_value = U64_MAX;
2248
2249 reg->s32_min_value = S32_MIN;
2250 reg->s32_max_value = S32_MAX;
2251 reg->u32_min_value = 0;
2252 reg->u32_max_value = U32_MAX;
2253 }
2254
__mark_reg64_unbounded(struct bpf_reg_state * reg)2255 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2256 {
2257 reg->smin_value = S64_MIN;
2258 reg->smax_value = S64_MAX;
2259 reg->umin_value = 0;
2260 reg->umax_value = U64_MAX;
2261 }
2262
__mark_reg32_unbounded(struct bpf_reg_state * reg)2263 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2264 {
2265 reg->s32_min_value = S32_MIN;
2266 reg->s32_max_value = S32_MAX;
2267 reg->u32_min_value = 0;
2268 reg->u32_max_value = U32_MAX;
2269 }
2270
__update_reg32_bounds(struct bpf_reg_state * reg)2271 static void __update_reg32_bounds(struct bpf_reg_state *reg)
2272 {
2273 struct tnum var32_off = tnum_subreg(reg->var_off);
2274
2275 /* min signed is max(sign bit) | min(other bits) */
2276 reg->s32_min_value = max_t(s32, reg->s32_min_value,
2277 var32_off.value | (var32_off.mask & S32_MIN));
2278 /* max signed is min(sign bit) | max(other bits) */
2279 reg->s32_max_value = min_t(s32, reg->s32_max_value,
2280 var32_off.value | (var32_off.mask & S32_MAX));
2281 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2282 reg->u32_max_value = min(reg->u32_max_value,
2283 (u32)(var32_off.value | var32_off.mask));
2284 }
2285
__update_reg64_bounds(struct bpf_reg_state * reg)2286 static void __update_reg64_bounds(struct bpf_reg_state *reg)
2287 {
2288 /* min signed is max(sign bit) | min(other bits) */
2289 reg->smin_value = max_t(s64, reg->smin_value,
2290 reg->var_off.value | (reg->var_off.mask & S64_MIN));
2291 /* max signed is min(sign bit) | max(other bits) */
2292 reg->smax_value = min_t(s64, reg->smax_value,
2293 reg->var_off.value | (reg->var_off.mask & S64_MAX));
2294 reg->umin_value = max(reg->umin_value, reg->var_off.value);
2295 reg->umax_value = min(reg->umax_value,
2296 reg->var_off.value | reg->var_off.mask);
2297 }
2298
__update_reg_bounds(struct bpf_reg_state * reg)2299 static void __update_reg_bounds(struct bpf_reg_state *reg)
2300 {
2301 __update_reg32_bounds(reg);
2302 __update_reg64_bounds(reg);
2303 }
2304
2305 /* Uses signed min/max values to inform unsigned, and vice-versa */
__reg32_deduce_bounds(struct bpf_reg_state * reg)2306 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2307 {
2308 /* Learn sign from signed bounds.
2309 * If we cannot cross the sign boundary, then signed and unsigned bounds
2310 * are the same, so combine. This works even in the negative case, e.g.
2311 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2312 */
2313 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
2314 reg->s32_min_value = reg->u32_min_value =
2315 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2316 reg->s32_max_value = reg->u32_max_value =
2317 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2318 return;
2319 }
2320 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2321 * boundary, so we must be careful.
2322 */
2323 if ((s32)reg->u32_max_value >= 0) {
2324 /* Positive. We can't learn anything from the smin, but smax
2325 * is positive, hence safe.
2326 */
2327 reg->s32_min_value = reg->u32_min_value;
2328 reg->s32_max_value = reg->u32_max_value =
2329 min_t(u32, reg->s32_max_value, reg->u32_max_value);
2330 } else if ((s32)reg->u32_min_value < 0) {
2331 /* Negative. We can't learn anything from the smax, but smin
2332 * is negative, hence safe.
2333 */
2334 reg->s32_min_value = reg->u32_min_value =
2335 max_t(u32, reg->s32_min_value, reg->u32_min_value);
2336 reg->s32_max_value = reg->u32_max_value;
2337 }
2338 }
2339
__reg64_deduce_bounds(struct bpf_reg_state * reg)2340 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2341 {
2342 /* Learn sign from signed bounds.
2343 * If we cannot cross the sign boundary, then signed and unsigned bounds
2344 * are the same, so combine. This works even in the negative case, e.g.
2345 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2346 */
2347 if (reg->smin_value >= 0 || reg->smax_value < 0) {
2348 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2349 reg->umin_value);
2350 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2351 reg->umax_value);
2352 return;
2353 }
2354 /* Learn sign from unsigned bounds. Signed bounds cross the sign
2355 * boundary, so we must be careful.
2356 */
2357 if ((s64)reg->umax_value >= 0) {
2358 /* Positive. We can't learn anything from the smin, but smax
2359 * is positive, hence safe.
2360 */
2361 reg->smin_value = reg->umin_value;
2362 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
2363 reg->umax_value);
2364 } else if ((s64)reg->umin_value < 0) {
2365 /* Negative. We can't learn anything from the smax, but smin
2366 * is negative, hence safe.
2367 */
2368 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
2369 reg->umin_value);
2370 reg->smax_value = reg->umax_value;
2371 }
2372 }
2373
__reg_deduce_bounds(struct bpf_reg_state * reg)2374 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2375 {
2376 __reg32_deduce_bounds(reg);
2377 __reg64_deduce_bounds(reg);
2378 }
2379
2380 /* Attempts to improve var_off based on unsigned min/max information */
__reg_bound_offset(struct bpf_reg_state * reg)2381 static void __reg_bound_offset(struct bpf_reg_state *reg)
2382 {
2383 struct tnum var64_off = tnum_intersect(reg->var_off,
2384 tnum_range(reg->umin_value,
2385 reg->umax_value));
2386 struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off),
2387 tnum_range(reg->u32_min_value,
2388 reg->u32_max_value));
2389
2390 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
2391 }
2392
reg_bounds_sync(struct bpf_reg_state * reg)2393 static void reg_bounds_sync(struct bpf_reg_state *reg)
2394 {
2395 /* We might have learned new bounds from the var_off. */
2396 __update_reg_bounds(reg);
2397 /* We might have learned something about the sign bit. */
2398 __reg_deduce_bounds(reg);
2399 /* We might have learned some bits from the bounds. */
2400 __reg_bound_offset(reg);
2401 /* Intersecting with the old var_off might have improved our bounds
2402 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2403 * then new var_off is (0; 0x7f...fc) which improves our umax.
2404 */
2405 __update_reg_bounds(reg);
2406 }
2407
__reg32_bound_s64(s32 a)2408 static bool __reg32_bound_s64(s32 a)
2409 {
2410 return a >= 0 && a <= S32_MAX;
2411 }
2412
__reg_assign_32_into_64(struct bpf_reg_state * reg)2413 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2414 {
2415 reg->umin_value = reg->u32_min_value;
2416 reg->umax_value = reg->u32_max_value;
2417
2418 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2419 * be positive otherwise set to worse case bounds and refine later
2420 * from tnum.
2421 */
2422 if (__reg32_bound_s64(reg->s32_min_value) &&
2423 __reg32_bound_s64(reg->s32_max_value)) {
2424 reg->smin_value = reg->s32_min_value;
2425 reg->smax_value = reg->s32_max_value;
2426 } else {
2427 reg->smin_value = 0;
2428 reg->smax_value = U32_MAX;
2429 }
2430 }
2431
__reg_combine_32_into_64(struct bpf_reg_state * reg)2432 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
2433 {
2434 /* special case when 64-bit register has upper 32-bit register
2435 * zeroed. Typically happens after zext or <<32, >>32 sequence
2436 * allowing us to use 32-bit bounds directly,
2437 */
2438 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
2439 __reg_assign_32_into_64(reg);
2440 } else {
2441 /* Otherwise the best we can do is push lower 32bit known and
2442 * unknown bits into register (var_off set from jmp logic)
2443 * then learn as much as possible from the 64-bit tnum
2444 * known and unknown bits. The previous smin/smax bounds are
2445 * invalid here because of jmp32 compare so mark them unknown
2446 * so they do not impact tnum bounds calculation.
2447 */
2448 __mark_reg64_unbounded(reg);
2449 }
2450 reg_bounds_sync(reg);
2451 }
2452
__reg64_bound_s32(s64 a)2453 static bool __reg64_bound_s32(s64 a)
2454 {
2455 return a >= S32_MIN && a <= S32_MAX;
2456 }
2457
__reg64_bound_u32(u64 a)2458 static bool __reg64_bound_u32(u64 a)
2459 {
2460 return a >= U32_MIN && a <= U32_MAX;
2461 }
2462
__reg_combine_64_into_32(struct bpf_reg_state * reg)2463 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
2464 {
2465 __mark_reg32_unbounded(reg);
2466 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
2467 reg->s32_min_value = (s32)reg->smin_value;
2468 reg->s32_max_value = (s32)reg->smax_value;
2469 }
2470 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
2471 reg->u32_min_value = (u32)reg->umin_value;
2472 reg->u32_max_value = (u32)reg->umax_value;
2473 }
2474 reg_bounds_sync(reg);
2475 }
2476
2477 /* Mark a register as having a completely unknown (scalar) value. */
__mark_reg_unknown(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)2478 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2479 struct bpf_reg_state *reg)
2480 {
2481 /*
2482 * Clear type, off, and union(map_ptr, range) and
2483 * padding between 'type' and union
2484 */
2485 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2486 reg->type = SCALAR_VALUE;
2487 reg->id = 0;
2488 reg->ref_obj_id = 0;
2489 reg->var_off = tnum_unknown;
2490 reg->frameno = 0;
2491 reg->precise = !env->bpf_capable;
2492 __mark_reg_unbounded(reg);
2493 }
2494
mark_reg_unknown(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)2495 static void mark_reg_unknown(struct bpf_verifier_env *env,
2496 struct bpf_reg_state *regs, u32 regno)
2497 {
2498 if (WARN_ON(regno >= MAX_BPF_REG)) {
2499 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
2500 /* Something bad happened, let's kill all regs except FP */
2501 for (regno = 0; regno < BPF_REG_FP; regno++)
2502 __mark_reg_not_init(env, regs + regno);
2503 return;
2504 }
2505 __mark_reg_unknown(env, regs + regno);
2506 }
2507
__mark_reg_not_init(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)2508 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2509 struct bpf_reg_state *reg)
2510 {
2511 __mark_reg_unknown(env, reg);
2512 reg->type = NOT_INIT;
2513 }
2514
mark_reg_not_init(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)2515 static void mark_reg_not_init(struct bpf_verifier_env *env,
2516 struct bpf_reg_state *regs, u32 regno)
2517 {
2518 if (WARN_ON(regno >= MAX_BPF_REG)) {
2519 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
2520 /* Something bad happened, let's kill all regs except FP */
2521 for (regno = 0; regno < BPF_REG_FP; regno++)
2522 __mark_reg_not_init(env, regs + regno);
2523 return;
2524 }
2525 __mark_reg_not_init(env, regs + regno);
2526 }
2527
mark_btf_ld_reg(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,enum bpf_reg_type reg_type,struct btf * btf,u32 btf_id,enum bpf_type_flag flag)2528 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
2529 struct bpf_reg_state *regs, u32 regno,
2530 enum bpf_reg_type reg_type,
2531 struct btf *btf, u32 btf_id,
2532 enum bpf_type_flag flag)
2533 {
2534 if (reg_type == SCALAR_VALUE) {
2535 mark_reg_unknown(env, regs, regno);
2536 return;
2537 }
2538 mark_reg_known_zero(env, regs, regno);
2539 regs[regno].type = PTR_TO_BTF_ID | flag;
2540 regs[regno].btf = btf;
2541 regs[regno].btf_id = btf_id;
2542 if (type_may_be_null(flag))
2543 regs[regno].id = ++env->id_gen;
2544 }
2545
2546 #define DEF_NOT_SUBREG (0)
init_reg_state(struct bpf_verifier_env * env,struct bpf_func_state * state)2547 static void init_reg_state(struct bpf_verifier_env *env,
2548 struct bpf_func_state *state)
2549 {
2550 struct bpf_reg_state *regs = state->regs;
2551 int i;
2552
2553 for (i = 0; i < MAX_BPF_REG; i++) {
2554 mark_reg_not_init(env, regs, i);
2555 regs[i].live = REG_LIVE_NONE;
2556 regs[i].parent = NULL;
2557 regs[i].subreg_def = DEF_NOT_SUBREG;
2558 }
2559
2560 /* frame pointer */
2561 regs[BPF_REG_FP].type = PTR_TO_STACK;
2562 mark_reg_known_zero(env, regs, BPF_REG_FP);
2563 regs[BPF_REG_FP].frameno = state->frameno;
2564 }
2565
2566 #define BPF_MAIN_FUNC (-1)
init_func_state(struct bpf_verifier_env * env,struct bpf_func_state * state,int callsite,int frameno,int subprogno)2567 static void init_func_state(struct bpf_verifier_env *env,
2568 struct bpf_func_state *state,
2569 int callsite, int frameno, int subprogno)
2570 {
2571 state->callsite = callsite;
2572 state->frameno = frameno;
2573 state->subprogno = subprogno;
2574 state->callback_ret_range = tnum_range(0, 0);
2575 init_reg_state(env, state);
2576 mark_verifier_state_scratched(env);
2577 }
2578
2579 /* Similar to push_stack(), but for async callbacks */
push_async_cb(struct bpf_verifier_env * env,int insn_idx,int prev_insn_idx,int subprog)2580 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2581 int insn_idx, int prev_insn_idx,
2582 int subprog)
2583 {
2584 struct bpf_verifier_stack_elem *elem;
2585 struct bpf_func_state *frame;
2586
2587 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
2588 if (!elem)
2589 goto err;
2590
2591 elem->insn_idx = insn_idx;
2592 elem->prev_insn_idx = prev_insn_idx;
2593 elem->next = env->head;
2594 elem->log_pos = env->log.end_pos;
2595 env->head = elem;
2596 env->stack_size++;
2597 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2598 verbose(env,
2599 "The sequence of %d jumps is too complex for async cb.\n",
2600 env->stack_size);
2601 goto err;
2602 }
2603 /* Unlike push_stack() do not copy_verifier_state().
2604 * The caller state doesn't matter.
2605 * This is async callback. It starts in a fresh stack.
2606 * Initialize it similar to do_check_common().
2607 */
2608 elem->st.branches = 1;
2609 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
2610 if (!frame)
2611 goto err;
2612 init_func_state(env, frame,
2613 BPF_MAIN_FUNC /* callsite */,
2614 0 /* frameno within this callchain */,
2615 subprog /* subprog number within this prog */);
2616 elem->st.frame[0] = frame;
2617 return &elem->st;
2618 err:
2619 free_verifier_state(env->cur_state, true);
2620 env->cur_state = NULL;
2621 /* pop all elements and return */
2622 while (!pop_stack(env, NULL, NULL, false));
2623 return NULL;
2624 }
2625
2626
2627 enum reg_arg_type {
2628 SRC_OP, /* register is used as source operand */
2629 DST_OP, /* register is used as destination operand */
2630 DST_OP_NO_MARK /* same as above, check only, don't mark */
2631 };
2632
cmp_subprogs(const void * a,const void * b)2633 static int cmp_subprogs(const void *a, const void *b)
2634 {
2635 return ((struct bpf_subprog_info *)a)->start -
2636 ((struct bpf_subprog_info *)b)->start;
2637 }
2638
find_subprog(struct bpf_verifier_env * env,int off)2639 static int find_subprog(struct bpf_verifier_env *env, int off)
2640 {
2641 struct bpf_subprog_info *p;
2642
2643 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
2644 sizeof(env->subprog_info[0]), cmp_subprogs);
2645 if (!p)
2646 return -ENOENT;
2647 return p - env->subprog_info;
2648
2649 }
2650
add_subprog(struct bpf_verifier_env * env,int off)2651 static int add_subprog(struct bpf_verifier_env *env, int off)
2652 {
2653 int insn_cnt = env->prog->len;
2654 int ret;
2655
2656 if (off >= insn_cnt || off < 0) {
2657 verbose(env, "call to invalid destination\n");
2658 return -EINVAL;
2659 }
2660 ret = find_subprog(env, off);
2661 if (ret >= 0)
2662 return ret;
2663 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
2664 verbose(env, "too many subprograms\n");
2665 return -E2BIG;
2666 }
2667 /* determine subprog starts. The end is one before the next starts */
2668 env->subprog_info[env->subprog_cnt++].start = off;
2669 sort(env->subprog_info, env->subprog_cnt,
2670 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
2671 return env->subprog_cnt - 1;
2672 }
2673
2674 #define MAX_KFUNC_DESCS 256
2675 #define MAX_KFUNC_BTFS 256
2676
2677 struct bpf_kfunc_desc {
2678 struct btf_func_model func_model;
2679 u32 func_id;
2680 s32 imm;
2681 u16 offset;
2682 unsigned long addr;
2683 };
2684
2685 struct bpf_kfunc_btf {
2686 struct btf *btf;
2687 struct module *module;
2688 u16 offset;
2689 };
2690
2691 struct bpf_kfunc_desc_tab {
2692 /* Sorted by func_id (BTF ID) and offset (fd_array offset) during
2693 * verification. JITs do lookups by bpf_insn, where func_id may not be
2694 * available, therefore at the end of verification do_misc_fixups()
2695 * sorts this by imm and offset.
2696 */
2697 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
2698 u32 nr_descs;
2699 };
2700
2701 struct bpf_kfunc_btf_tab {
2702 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
2703 u32 nr_descs;
2704 };
2705
kfunc_desc_cmp_by_id_off(const void * a,const void * b)2706 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
2707 {
2708 const struct bpf_kfunc_desc *d0 = a;
2709 const struct bpf_kfunc_desc *d1 = b;
2710
2711 /* func_id is not greater than BTF_MAX_TYPE */
2712 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
2713 }
2714
kfunc_btf_cmp_by_off(const void * a,const void * b)2715 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
2716 {
2717 const struct bpf_kfunc_btf *d0 = a;
2718 const struct bpf_kfunc_btf *d1 = b;
2719
2720 return d0->offset - d1->offset;
2721 }
2722
2723 static const struct bpf_kfunc_desc *
find_kfunc_desc(const struct bpf_prog * prog,u32 func_id,u16 offset)2724 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
2725 {
2726 struct bpf_kfunc_desc desc = {
2727 .func_id = func_id,
2728 .offset = offset,
2729 };
2730 struct bpf_kfunc_desc_tab *tab;
2731
2732 tab = prog->aux->kfunc_tab;
2733 return bsearch(&desc, tab->descs, tab->nr_descs,
2734 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
2735 }
2736
bpf_get_kfunc_addr(const struct bpf_prog * prog,u32 func_id,u16 btf_fd_idx,u8 ** func_addr)2737 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
2738 u16 btf_fd_idx, u8 **func_addr)
2739 {
2740 const struct bpf_kfunc_desc *desc;
2741
2742 desc = find_kfunc_desc(prog, func_id, btf_fd_idx);
2743 if (!desc)
2744 return -EFAULT;
2745
2746 *func_addr = (u8 *)desc->addr;
2747 return 0;
2748 }
2749
__find_kfunc_desc_btf(struct bpf_verifier_env * env,s16 offset)2750 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
2751 s16 offset)
2752 {
2753 struct bpf_kfunc_btf kf_btf = { .offset = offset };
2754 struct bpf_kfunc_btf_tab *tab;
2755 struct bpf_kfunc_btf *b;
2756 struct module *mod;
2757 struct btf *btf;
2758 int btf_fd;
2759
2760 tab = env->prog->aux->kfunc_btf_tab;
2761 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
2762 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
2763 if (!b) {
2764 if (tab->nr_descs == MAX_KFUNC_BTFS) {
2765 verbose(env, "too many different module BTFs\n");
2766 return ERR_PTR(-E2BIG);
2767 }
2768
2769 if (bpfptr_is_null(env->fd_array)) {
2770 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
2771 return ERR_PTR(-EPROTO);
2772 }
2773
2774 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
2775 offset * sizeof(btf_fd),
2776 sizeof(btf_fd)))
2777 return ERR_PTR(-EFAULT);
2778
2779 btf = btf_get_by_fd(btf_fd);
2780 if (IS_ERR(btf)) {
2781 verbose(env, "invalid module BTF fd specified\n");
2782 return btf;
2783 }
2784
2785 if (!btf_is_module(btf)) {
2786 verbose(env, "BTF fd for kfunc is not a module BTF\n");
2787 btf_put(btf);
2788 return ERR_PTR(-EINVAL);
2789 }
2790
2791 mod = btf_try_get_module(btf);
2792 if (!mod) {
2793 btf_put(btf);
2794 return ERR_PTR(-ENXIO);
2795 }
2796
2797 b = &tab->descs[tab->nr_descs++];
2798 b->btf = btf;
2799 b->module = mod;
2800 b->offset = offset;
2801
2802 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2803 kfunc_btf_cmp_by_off, NULL);
2804 }
2805 return b->btf;
2806 }
2807
bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab * tab)2808 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
2809 {
2810 if (!tab)
2811 return;
2812
2813 while (tab->nr_descs--) {
2814 module_put(tab->descs[tab->nr_descs].module);
2815 btf_put(tab->descs[tab->nr_descs].btf);
2816 }
2817 kfree(tab);
2818 }
2819
find_kfunc_desc_btf(struct bpf_verifier_env * env,s16 offset)2820 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
2821 {
2822 if (offset) {
2823 if (offset < 0) {
2824 /* In the future, this can be allowed to increase limit
2825 * of fd index into fd_array, interpreted as u16.
2826 */
2827 verbose(env, "negative offset disallowed for kernel module function call\n");
2828 return ERR_PTR(-EINVAL);
2829 }
2830
2831 return __find_kfunc_desc_btf(env, offset);
2832 }
2833 return btf_vmlinux ?: ERR_PTR(-ENOENT);
2834 }
2835
add_kfunc_call(struct bpf_verifier_env * env,u32 func_id,s16 offset)2836 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
2837 {
2838 const struct btf_type *func, *func_proto;
2839 struct bpf_kfunc_btf_tab *btf_tab;
2840 struct bpf_kfunc_desc_tab *tab;
2841 struct bpf_prog_aux *prog_aux;
2842 struct bpf_kfunc_desc *desc;
2843 const char *func_name;
2844 struct btf *desc_btf;
2845 unsigned long call_imm;
2846 unsigned long addr;
2847 int err;
2848
2849 prog_aux = env->prog->aux;
2850 tab = prog_aux->kfunc_tab;
2851 btf_tab = prog_aux->kfunc_btf_tab;
2852 if (!tab) {
2853 if (!btf_vmlinux) {
2854 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
2855 return -ENOTSUPP;
2856 }
2857
2858 if (!env->prog->jit_requested) {
2859 verbose(env, "JIT is required for calling kernel function\n");
2860 return -ENOTSUPP;
2861 }
2862
2863 if (!bpf_jit_supports_kfunc_call()) {
2864 verbose(env, "JIT does not support calling kernel function\n");
2865 return -ENOTSUPP;
2866 }
2867
2868 if (!env->prog->gpl_compatible) {
2869 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
2870 return -EINVAL;
2871 }
2872
2873 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
2874 if (!tab)
2875 return -ENOMEM;
2876 prog_aux->kfunc_tab = tab;
2877 }
2878
2879 /* func_id == 0 is always invalid, but instead of returning an error, be
2880 * conservative and wait until the code elimination pass before returning
2881 * error, so that invalid calls that get pruned out can be in BPF programs
2882 * loaded from userspace. It is also required that offset be untouched
2883 * for such calls.
2884 */
2885 if (!func_id && !offset)
2886 return 0;
2887
2888 if (!btf_tab && offset) {
2889 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
2890 if (!btf_tab)
2891 return -ENOMEM;
2892 prog_aux->kfunc_btf_tab = btf_tab;
2893 }
2894
2895 desc_btf = find_kfunc_desc_btf(env, offset);
2896 if (IS_ERR(desc_btf)) {
2897 verbose(env, "failed to find BTF for kernel function\n");
2898 return PTR_ERR(desc_btf);
2899 }
2900
2901 if (find_kfunc_desc(env->prog, func_id, offset))
2902 return 0;
2903
2904 if (tab->nr_descs == MAX_KFUNC_DESCS) {
2905 verbose(env, "too many different kernel function calls\n");
2906 return -E2BIG;
2907 }
2908
2909 func = btf_type_by_id(desc_btf, func_id);
2910 if (!func || !btf_type_is_func(func)) {
2911 verbose(env, "kernel btf_id %u is not a function\n",
2912 func_id);
2913 return -EINVAL;
2914 }
2915 func_proto = btf_type_by_id(desc_btf, func->type);
2916 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
2917 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
2918 func_id);
2919 return -EINVAL;
2920 }
2921
2922 func_name = btf_name_by_offset(desc_btf, func->name_off);
2923 addr = kallsyms_lookup_name(func_name);
2924 if (!addr) {
2925 verbose(env, "cannot find address for kernel function %s\n",
2926 func_name);
2927 return -EINVAL;
2928 }
2929 specialize_kfunc(env, func_id, offset, &addr);
2930
2931 if (bpf_jit_supports_far_kfunc_call()) {
2932 call_imm = func_id;
2933 } else {
2934 call_imm = BPF_CALL_IMM(addr);
2935 /* Check whether the relative offset overflows desc->imm */
2936 if ((unsigned long)(s32)call_imm != call_imm) {
2937 verbose(env, "address of kernel function %s is out of range\n",
2938 func_name);
2939 return -EINVAL;
2940 }
2941 }
2942
2943 if (bpf_dev_bound_kfunc_id(func_id)) {
2944 err = bpf_dev_bound_kfunc_check(&env->log, prog_aux);
2945 if (err)
2946 return err;
2947 }
2948
2949 desc = &tab->descs[tab->nr_descs++];
2950 desc->func_id = func_id;
2951 desc->imm = call_imm;
2952 desc->offset = offset;
2953 desc->addr = addr;
2954 err = btf_distill_func_proto(&env->log, desc_btf,
2955 func_proto, func_name,
2956 &desc->func_model);
2957 if (!err)
2958 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2959 kfunc_desc_cmp_by_id_off, NULL);
2960 return err;
2961 }
2962
kfunc_desc_cmp_by_imm_off(const void * a,const void * b)2963 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
2964 {
2965 const struct bpf_kfunc_desc *d0 = a;
2966 const struct bpf_kfunc_desc *d1 = b;
2967
2968 if (d0->imm != d1->imm)
2969 return d0->imm < d1->imm ? -1 : 1;
2970 if (d0->offset != d1->offset)
2971 return d0->offset < d1->offset ? -1 : 1;
2972 return 0;
2973 }
2974
sort_kfunc_descs_by_imm_off(struct bpf_prog * prog)2975 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog)
2976 {
2977 struct bpf_kfunc_desc_tab *tab;
2978
2979 tab = prog->aux->kfunc_tab;
2980 if (!tab)
2981 return;
2982
2983 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
2984 kfunc_desc_cmp_by_imm_off, NULL);
2985 }
2986
bpf_prog_has_kfunc_call(const struct bpf_prog * prog)2987 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
2988 {
2989 return !!prog->aux->kfunc_tab;
2990 }
2991
2992 const struct btf_func_model *
bpf_jit_find_kfunc_model(const struct bpf_prog * prog,const struct bpf_insn * insn)2993 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
2994 const struct bpf_insn *insn)
2995 {
2996 const struct bpf_kfunc_desc desc = {
2997 .imm = insn->imm,
2998 .offset = insn->off,
2999 };
3000 const struct bpf_kfunc_desc *res;
3001 struct bpf_kfunc_desc_tab *tab;
3002
3003 tab = prog->aux->kfunc_tab;
3004 res = bsearch(&desc, tab->descs, tab->nr_descs,
3005 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off);
3006
3007 return res ? &res->func_model : NULL;
3008 }
3009
add_subprog_and_kfunc(struct bpf_verifier_env * env)3010 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
3011 {
3012 struct bpf_subprog_info *subprog = env->subprog_info;
3013 struct bpf_insn *insn = env->prog->insnsi;
3014 int i, ret, insn_cnt = env->prog->len;
3015
3016 /* Add entry function. */
3017 ret = add_subprog(env, 0);
3018 if (ret)
3019 return ret;
3020
3021 for (i = 0; i < insn_cnt; i++, insn++) {
3022 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
3023 !bpf_pseudo_kfunc_call(insn))
3024 continue;
3025
3026 if (!env->bpf_capable) {
3027 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
3028 return -EPERM;
3029 }
3030
3031 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
3032 ret = add_subprog(env, i + insn->imm + 1);
3033 else
3034 ret = add_kfunc_call(env, insn->imm, insn->off);
3035
3036 if (ret < 0)
3037 return ret;
3038 }
3039
3040 /* Add a fake 'exit' subprog which could simplify subprog iteration
3041 * logic. 'subprog_cnt' should not be increased.
3042 */
3043 subprog[env->subprog_cnt].start = insn_cnt;
3044
3045 if (env->log.level & BPF_LOG_LEVEL2)
3046 for (i = 0; i < env->subprog_cnt; i++)
3047 verbose(env, "func#%d @%d\n", i, subprog[i].start);
3048
3049 return 0;
3050 }
3051
check_subprogs(struct bpf_verifier_env * env)3052 static int check_subprogs(struct bpf_verifier_env *env)
3053 {
3054 int i, subprog_start, subprog_end, off, cur_subprog = 0;
3055 struct bpf_subprog_info *subprog = env->subprog_info;
3056 struct bpf_insn *insn = env->prog->insnsi;
3057 int insn_cnt = env->prog->len;
3058
3059 /* now check that all jumps are within the same subprog */
3060 subprog_start = subprog[cur_subprog].start;
3061 subprog_end = subprog[cur_subprog + 1].start;
3062 for (i = 0; i < insn_cnt; i++) {
3063 u8 code = insn[i].code;
3064
3065 if (code == (BPF_JMP | BPF_CALL) &&
3066 insn[i].src_reg == 0 &&
3067 insn[i].imm == BPF_FUNC_tail_call) {
3068 subprog[cur_subprog].has_tail_call = true;
3069 subprog[cur_subprog].tail_call_reachable = true;
3070 }
3071 if (BPF_CLASS(code) == BPF_LD &&
3072 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
3073 subprog[cur_subprog].has_ld_abs = true;
3074 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
3075 goto next;
3076 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
3077 goto next;
3078 if (code == (BPF_JMP32 | BPF_JA))
3079 off = i + insn[i].imm + 1;
3080 else
3081 off = i + insn[i].off + 1;
3082 if (off < subprog_start || off >= subprog_end) {
3083 verbose(env, "jump out of range from insn %d to %d\n", i, off);
3084 return -EINVAL;
3085 }
3086 next:
3087 if (i == subprog_end - 1) {
3088 /* to avoid fall-through from one subprog into another
3089 * the last insn of the subprog should be either exit
3090 * or unconditional jump back
3091 */
3092 if (code != (BPF_JMP | BPF_EXIT) &&
3093 code != (BPF_JMP32 | BPF_JA) &&
3094 code != (BPF_JMP | BPF_JA)) {
3095 verbose(env, "last insn is not an exit or jmp\n");
3096 return -EINVAL;
3097 }
3098 subprog_start = subprog_end;
3099 cur_subprog++;
3100 if (cur_subprog < env->subprog_cnt)
3101 subprog_end = subprog[cur_subprog + 1].start;
3102 }
3103 }
3104 return 0;
3105 }
3106
3107 /* Parentage chain of this register (or stack slot) should take care of all
3108 * issues like callee-saved registers, stack slot allocation time, etc.
3109 */
mark_reg_read(struct bpf_verifier_env * env,const struct bpf_reg_state * state,struct bpf_reg_state * parent,u8 flag)3110 static int mark_reg_read(struct bpf_verifier_env *env,
3111 const struct bpf_reg_state *state,
3112 struct bpf_reg_state *parent, u8 flag)
3113 {
3114 bool writes = parent == state->parent; /* Observe write marks */
3115 int cnt = 0;
3116
3117 while (parent) {
3118 /* if read wasn't screened by an earlier write ... */
3119 if (writes && state->live & REG_LIVE_WRITTEN)
3120 break;
3121 if (parent->live & REG_LIVE_DONE) {
3122 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
3123 reg_type_str(env, parent->type),
3124 parent->var_off.value, parent->off);
3125 return -EFAULT;
3126 }
3127 /* The first condition is more likely to be true than the
3128 * second, checked it first.
3129 */
3130 if ((parent->live & REG_LIVE_READ) == flag ||
3131 parent->live & REG_LIVE_READ64)
3132 /* The parentage chain never changes and
3133 * this parent was already marked as LIVE_READ.
3134 * There is no need to keep walking the chain again and
3135 * keep re-marking all parents as LIVE_READ.
3136 * This case happens when the same register is read
3137 * multiple times without writes into it in-between.
3138 * Also, if parent has the stronger REG_LIVE_READ64 set,
3139 * then no need to set the weak REG_LIVE_READ32.
3140 */
3141 break;
3142 /* ... then we depend on parent's value */
3143 parent->live |= flag;
3144 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
3145 if (flag == REG_LIVE_READ64)
3146 parent->live &= ~REG_LIVE_READ32;
3147 state = parent;
3148 parent = state->parent;
3149 writes = true;
3150 cnt++;
3151 }
3152
3153 if (env->longest_mark_read_walk < cnt)
3154 env->longest_mark_read_walk = cnt;
3155 return 0;
3156 }
3157
mark_dynptr_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg)3158 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
3159 {
3160 struct bpf_func_state *state = func(env, reg);
3161 int spi, ret;
3162
3163 /* For CONST_PTR_TO_DYNPTR, it must have already been done by
3164 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
3165 * check_kfunc_call.
3166 */
3167 if (reg->type == CONST_PTR_TO_DYNPTR)
3168 return 0;
3169 spi = dynptr_get_spi(env, reg);
3170 if (spi < 0)
3171 return spi;
3172 /* Caller ensures dynptr is valid and initialized, which means spi is in
3173 * bounds and spi is the first dynptr slot. Simply mark stack slot as
3174 * read.
3175 */
3176 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr,
3177 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64);
3178 if (ret)
3179 return ret;
3180 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr,
3181 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64);
3182 }
3183
mark_iter_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int spi,int nr_slots)3184 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
3185 int spi, int nr_slots)
3186 {
3187 struct bpf_func_state *state = func(env, reg);
3188 int err, i;
3189
3190 for (i = 0; i < nr_slots; i++) {
3191 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
3192
3193 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
3194 if (err)
3195 return err;
3196
3197 mark_stack_slot_scratched(env, spi - i);
3198 }
3199
3200 return 0;
3201 }
3202
3203 /* This function is supposed to be used by the following 32-bit optimization
3204 * code only. It returns TRUE if the source or destination register operates
3205 * on 64-bit, otherwise return FALSE.
3206 */
is_reg64(struct bpf_verifier_env * env,struct bpf_insn * insn,u32 regno,struct bpf_reg_state * reg,enum reg_arg_type t)3207 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
3208 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
3209 {
3210 u8 code, class, op;
3211
3212 code = insn->code;
3213 class = BPF_CLASS(code);
3214 op = BPF_OP(code);
3215 if (class == BPF_JMP) {
3216 /* BPF_EXIT for "main" will reach here. Return TRUE
3217 * conservatively.
3218 */
3219 if (op == BPF_EXIT)
3220 return true;
3221 if (op == BPF_CALL) {
3222 /* BPF to BPF call will reach here because of marking
3223 * caller saved clobber with DST_OP_NO_MARK for which we
3224 * don't care the register def because they are anyway
3225 * marked as NOT_INIT already.
3226 */
3227 if (insn->src_reg == BPF_PSEUDO_CALL)
3228 return false;
3229 /* Helper call will reach here because of arg type
3230 * check, conservatively return TRUE.
3231 */
3232 if (t == SRC_OP)
3233 return true;
3234
3235 return false;
3236 }
3237 }
3238
3239 if (class == BPF_ALU64 && op == BPF_END && (insn->imm == 16 || insn->imm == 32))
3240 return false;
3241
3242 if (class == BPF_ALU64 || class == BPF_JMP ||
3243 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3244 return true;
3245
3246 if (class == BPF_ALU || class == BPF_JMP32)
3247 return false;
3248
3249 if (class == BPF_LDX) {
3250 if (t != SRC_OP)
3251 return BPF_SIZE(code) == BPF_DW;
3252 /* LDX source must be ptr. */
3253 return true;
3254 }
3255
3256 if (class == BPF_STX) {
3257 /* BPF_STX (including atomic variants) has multiple source
3258 * operands, one of which is a ptr. Check whether the caller is
3259 * asking about it.
3260 */
3261 if (t == SRC_OP && reg->type != SCALAR_VALUE)
3262 return true;
3263 return BPF_SIZE(code) == BPF_DW;
3264 }
3265
3266 if (class == BPF_LD) {
3267 u8 mode = BPF_MODE(code);
3268
3269 /* LD_IMM64 */
3270 if (mode == BPF_IMM)
3271 return true;
3272
3273 /* Both LD_IND and LD_ABS return 32-bit data. */
3274 if (t != SRC_OP)
3275 return false;
3276
3277 /* Implicit ctx ptr. */
3278 if (regno == BPF_REG_6)
3279 return true;
3280
3281 /* Explicit source could be any width. */
3282 return true;
3283 }
3284
3285 if (class == BPF_ST)
3286 /* The only source register for BPF_ST is a ptr. */
3287 return true;
3288
3289 /* Conservatively return true at default. */
3290 return true;
3291 }
3292
3293 /* Return the regno defined by the insn, or -1. */
insn_def_regno(const struct bpf_insn * insn)3294 static int insn_def_regno(const struct bpf_insn *insn)
3295 {
3296 switch (BPF_CLASS(insn->code)) {
3297 case BPF_JMP:
3298 case BPF_JMP32:
3299 case BPF_ST:
3300 return -1;
3301 case BPF_STX:
3302 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
3303 (insn->imm & BPF_FETCH)) {
3304 if (insn->imm == BPF_CMPXCHG)
3305 return BPF_REG_0;
3306 else
3307 return insn->src_reg;
3308 } else {
3309 return -1;
3310 }
3311 default:
3312 return insn->dst_reg;
3313 }
3314 }
3315
3316 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
insn_has_def32(struct bpf_verifier_env * env,struct bpf_insn * insn)3317 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3318 {
3319 int dst_reg = insn_def_regno(insn);
3320
3321 if (dst_reg == -1)
3322 return false;
3323
3324 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3325 }
3326
mark_insn_zext(struct bpf_verifier_env * env,struct bpf_reg_state * reg)3327 static void mark_insn_zext(struct bpf_verifier_env *env,
3328 struct bpf_reg_state *reg)
3329 {
3330 s32 def_idx = reg->subreg_def;
3331
3332 if (def_idx == DEF_NOT_SUBREG)
3333 return;
3334
3335 env->insn_aux_data[def_idx - 1].zext_dst = true;
3336 /* The dst will be zero extended, so won't be sub-register anymore. */
3337 reg->subreg_def = DEF_NOT_SUBREG;
3338 }
3339
__check_reg_arg(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,enum reg_arg_type t)3340 static int __check_reg_arg(struct bpf_verifier_env *env, struct bpf_reg_state *regs, u32 regno,
3341 enum reg_arg_type t)
3342 {
3343 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3344 struct bpf_reg_state *reg;
3345 bool rw64;
3346
3347 if (regno >= MAX_BPF_REG) {
3348 verbose(env, "R%d is invalid\n", regno);
3349 return -EINVAL;
3350 }
3351
3352 mark_reg_scratched(env, regno);
3353
3354 reg = ®s[regno];
3355 rw64 = is_reg64(env, insn, regno, reg, t);
3356 if (t == SRC_OP) {
3357 /* check whether register used as source operand can be read */
3358 if (reg->type == NOT_INIT) {
3359 verbose(env, "R%d !read_ok\n", regno);
3360 return -EACCES;
3361 }
3362 /* We don't need to worry about FP liveness because it's read-only */
3363 if (regno == BPF_REG_FP)
3364 return 0;
3365
3366 if (rw64)
3367 mark_insn_zext(env, reg);
3368
3369 return mark_reg_read(env, reg, reg->parent,
3370 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3371 } else {
3372 /* check whether register used as dest operand can be written to */
3373 if (regno == BPF_REG_FP) {
3374 verbose(env, "frame pointer is read only\n");
3375 return -EACCES;
3376 }
3377 reg->live |= REG_LIVE_WRITTEN;
3378 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3379 if (t == DST_OP)
3380 mark_reg_unknown(env, regs, regno);
3381 }
3382 return 0;
3383 }
3384
check_reg_arg(struct bpf_verifier_env * env,u32 regno,enum reg_arg_type t)3385 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3386 enum reg_arg_type t)
3387 {
3388 struct bpf_verifier_state *vstate = env->cur_state;
3389 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3390
3391 return __check_reg_arg(env, state->regs, regno, t);
3392 }
3393
mark_jmp_point(struct bpf_verifier_env * env,int idx)3394 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3395 {
3396 env->insn_aux_data[idx].jmp_point = true;
3397 }
3398
is_jmp_point(struct bpf_verifier_env * env,int insn_idx)3399 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3400 {
3401 return env->insn_aux_data[insn_idx].jmp_point;
3402 }
3403
3404 /* for any branch, call, exit record the history of jmps in the given state */
push_jmp_history(struct bpf_verifier_env * env,struct bpf_verifier_state * cur)3405 static int push_jmp_history(struct bpf_verifier_env *env,
3406 struct bpf_verifier_state *cur)
3407 {
3408 u32 cnt = cur->jmp_history_cnt;
3409 struct bpf_idx_pair *p;
3410 size_t alloc_size;
3411
3412 if (!is_jmp_point(env, env->insn_idx))
3413 return 0;
3414
3415 cnt++;
3416 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3417 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
3418 if (!p)
3419 return -ENOMEM;
3420 p[cnt - 1].idx = env->insn_idx;
3421 p[cnt - 1].prev_idx = env->prev_insn_idx;
3422 cur->jmp_history = p;
3423 cur->jmp_history_cnt = cnt;
3424 return 0;
3425 }
3426
3427 /* Backtrack one insn at a time. If idx is not at the top of recorded
3428 * history then previous instruction came from straight line execution.
3429 * Return -ENOENT if we exhausted all instructions within given state.
3430 *
3431 * It's legal to have a bit of a looping with the same starting and ending
3432 * insn index within the same state, e.g.: 3->4->5->3, so just because current
3433 * instruction index is the same as state's first_idx doesn't mean we are
3434 * done. If there is still some jump history left, we should keep going. We
3435 * need to take into account that we might have a jump history between given
3436 * state's parent and itself, due to checkpointing. In this case, we'll have
3437 * history entry recording a jump from last instruction of parent state and
3438 * first instruction of given state.
3439 */
get_prev_insn_idx(struct bpf_verifier_state * st,int i,u32 * history)3440 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3441 u32 *history)
3442 {
3443 u32 cnt = *history;
3444
3445 if (i == st->first_insn_idx) {
3446 if (cnt == 0)
3447 return -ENOENT;
3448 if (cnt == 1 && st->jmp_history[0].idx == i)
3449 return -ENOENT;
3450 }
3451
3452 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3453 i = st->jmp_history[cnt - 1].prev_idx;
3454 (*history)--;
3455 } else {
3456 i--;
3457 }
3458 return i;
3459 }
3460
disasm_kfunc_name(void * data,const struct bpf_insn * insn)3461 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3462 {
3463 const struct btf_type *func;
3464 struct btf *desc_btf;
3465
3466 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3467 return NULL;
3468
3469 desc_btf = find_kfunc_desc_btf(data, insn->off);
3470 if (IS_ERR(desc_btf))
3471 return "<error>";
3472
3473 func = btf_type_by_id(desc_btf, insn->imm);
3474 return btf_name_by_offset(desc_btf, func->name_off);
3475 }
3476
bt_init(struct backtrack_state * bt,u32 frame)3477 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3478 {
3479 bt->frame = frame;
3480 }
3481
bt_reset(struct backtrack_state * bt)3482 static inline void bt_reset(struct backtrack_state *bt)
3483 {
3484 struct bpf_verifier_env *env = bt->env;
3485
3486 memset(bt, 0, sizeof(*bt));
3487 bt->env = env;
3488 }
3489
bt_empty(struct backtrack_state * bt)3490 static inline u32 bt_empty(struct backtrack_state *bt)
3491 {
3492 u64 mask = 0;
3493 int i;
3494
3495 for (i = 0; i <= bt->frame; i++)
3496 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3497
3498 return mask == 0;
3499 }
3500
bt_subprog_enter(struct backtrack_state * bt)3501 static inline int bt_subprog_enter(struct backtrack_state *bt)
3502 {
3503 if (bt->frame == MAX_CALL_FRAMES - 1) {
3504 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3505 WARN_ONCE(1, "verifier backtracking bug");
3506 return -EFAULT;
3507 }
3508 bt->frame++;
3509 return 0;
3510 }
3511
bt_subprog_exit(struct backtrack_state * bt)3512 static inline int bt_subprog_exit(struct backtrack_state *bt)
3513 {
3514 if (bt->frame == 0) {
3515 verbose(bt->env, "BUG subprog exit from frame 0\n");
3516 WARN_ONCE(1, "verifier backtracking bug");
3517 return -EFAULT;
3518 }
3519 bt->frame--;
3520 return 0;
3521 }
3522
bt_set_frame_reg(struct backtrack_state * bt,u32 frame,u32 reg)3523 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3524 {
3525 bt->reg_masks[frame] |= 1 << reg;
3526 }
3527
bt_clear_frame_reg(struct backtrack_state * bt,u32 frame,u32 reg)3528 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3529 {
3530 bt->reg_masks[frame] &= ~(1 << reg);
3531 }
3532
bt_set_reg(struct backtrack_state * bt,u32 reg)3533 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3534 {
3535 bt_set_frame_reg(bt, bt->frame, reg);
3536 }
3537
bt_clear_reg(struct backtrack_state * bt,u32 reg)3538 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3539 {
3540 bt_clear_frame_reg(bt, bt->frame, reg);
3541 }
3542
bt_set_frame_slot(struct backtrack_state * bt,u32 frame,u32 slot)3543 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3544 {
3545 bt->stack_masks[frame] |= 1ull << slot;
3546 }
3547
bt_clear_frame_slot(struct backtrack_state * bt,u32 frame,u32 slot)3548 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3549 {
3550 bt->stack_masks[frame] &= ~(1ull << slot);
3551 }
3552
bt_set_slot(struct backtrack_state * bt,u32 slot)3553 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3554 {
3555 bt_set_frame_slot(bt, bt->frame, slot);
3556 }
3557
bt_clear_slot(struct backtrack_state * bt,u32 slot)3558 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3559 {
3560 bt_clear_frame_slot(bt, bt->frame, slot);
3561 }
3562
bt_frame_reg_mask(struct backtrack_state * bt,u32 frame)3563 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3564 {
3565 return bt->reg_masks[frame];
3566 }
3567
bt_reg_mask(struct backtrack_state * bt)3568 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3569 {
3570 return bt->reg_masks[bt->frame];
3571 }
3572
bt_frame_stack_mask(struct backtrack_state * bt,u32 frame)3573 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3574 {
3575 return bt->stack_masks[frame];
3576 }
3577
bt_stack_mask(struct backtrack_state * bt)3578 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3579 {
3580 return bt->stack_masks[bt->frame];
3581 }
3582
bt_is_reg_set(struct backtrack_state * bt,u32 reg)3583 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3584 {
3585 return bt->reg_masks[bt->frame] & (1 << reg);
3586 }
3587
bt_is_slot_set(struct backtrack_state * bt,u32 slot)3588 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3589 {
3590 return bt->stack_masks[bt->frame] & (1ull << slot);
3591 }
3592
3593 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
fmt_reg_mask(char * buf,ssize_t buf_sz,u32 reg_mask)3594 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3595 {
3596 DECLARE_BITMAP(mask, 64);
3597 bool first = true;
3598 int i, n;
3599
3600 buf[0] = '\0';
3601
3602 bitmap_from_u64(mask, reg_mask);
3603 for_each_set_bit(i, mask, 32) {
3604 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3605 first = false;
3606 buf += n;
3607 buf_sz -= n;
3608 if (buf_sz < 0)
3609 break;
3610 }
3611 }
3612 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
fmt_stack_mask(char * buf,ssize_t buf_sz,u64 stack_mask)3613 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3614 {
3615 DECLARE_BITMAP(mask, 64);
3616 bool first = true;
3617 int i, n;
3618
3619 buf[0] = '\0';
3620
3621 bitmap_from_u64(mask, stack_mask);
3622 for_each_set_bit(i, mask, 64) {
3623 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3624 first = false;
3625 buf += n;
3626 buf_sz -= n;
3627 if (buf_sz < 0)
3628 break;
3629 }
3630 }
3631
3632 static bool calls_callback(struct bpf_verifier_env *env, int insn_idx);
3633
3634 /* For given verifier state backtrack_insn() is called from the last insn to
3635 * the first insn. Its purpose is to compute a bitmask of registers and
3636 * stack slots that needs precision in the parent verifier state.
3637 *
3638 * @idx is an index of the instruction we are currently processing;
3639 * @subseq_idx is an index of the subsequent instruction that:
3640 * - *would be* executed next, if jump history is viewed in forward order;
3641 * - *was* processed previously during backtracking.
3642 */
backtrack_insn(struct bpf_verifier_env * env,int idx,int subseq_idx,struct backtrack_state * bt)3643 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3644 struct backtrack_state *bt)
3645 {
3646 const struct bpf_insn_cbs cbs = {
3647 .cb_call = disasm_kfunc_name,
3648 .cb_print = verbose,
3649 .private_data = env,
3650 };
3651 struct bpf_insn *insn = env->prog->insnsi + idx;
3652 u8 class = BPF_CLASS(insn->code);
3653 u8 opcode = BPF_OP(insn->code);
3654 u8 mode = BPF_MODE(insn->code);
3655 u32 dreg = insn->dst_reg;
3656 u32 sreg = insn->src_reg;
3657 u32 spi, i;
3658
3659 if (insn->code == 0)
3660 return 0;
3661 if (env->log.level & BPF_LOG_LEVEL2) {
3662 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3663 verbose(env, "mark_precise: frame%d: regs=%s ",
3664 bt->frame, env->tmp_str_buf);
3665 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3666 verbose(env, "stack=%s before ", env->tmp_str_buf);
3667 verbose(env, "%d: ", idx);
3668 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3669 }
3670
3671 if (class == BPF_ALU || class == BPF_ALU64) {
3672 if (!bt_is_reg_set(bt, dreg))
3673 return 0;
3674 if (opcode == BPF_END || opcode == BPF_NEG) {
3675 /* sreg is reserved and unused
3676 * dreg still need precision before this insn
3677 */
3678 return 0;
3679 } else if (opcode == BPF_MOV) {
3680 if (BPF_SRC(insn->code) == BPF_X) {
3681 /* dreg = sreg or dreg = (s8, s16, s32)sreg
3682 * dreg needs precision after this insn
3683 * sreg needs precision before this insn
3684 */
3685 bt_clear_reg(bt, dreg);
3686 if (sreg != BPF_REG_FP)
3687 bt_set_reg(bt, sreg);
3688 } else {
3689 /* dreg = K
3690 * dreg needs precision after this insn.
3691 * Corresponding register is already marked
3692 * as precise=true in this verifier state.
3693 * No further markings in parent are necessary
3694 */
3695 bt_clear_reg(bt, dreg);
3696 }
3697 } else {
3698 if (BPF_SRC(insn->code) == BPF_X) {
3699 /* dreg += sreg
3700 * both dreg and sreg need precision
3701 * before this insn
3702 */
3703 if (sreg != BPF_REG_FP)
3704 bt_set_reg(bt, sreg);
3705 } /* else dreg += K
3706 * dreg still needs precision before this insn
3707 */
3708 }
3709 } else if (class == BPF_LDX) {
3710 if (!bt_is_reg_set(bt, dreg))
3711 return 0;
3712 bt_clear_reg(bt, dreg);
3713
3714 /* scalars can only be spilled into stack w/o losing precision.
3715 * Load from any other memory can be zero extended.
3716 * The desire to keep that precision is already indicated
3717 * by 'precise' mark in corresponding register of this state.
3718 * No further tracking necessary.
3719 */
3720 if (insn->src_reg != BPF_REG_FP)
3721 return 0;
3722
3723 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3724 * that [fp - off] slot contains scalar that needs to be
3725 * tracked with precision
3726 */
3727 spi = (-insn->off - 1) / BPF_REG_SIZE;
3728 if (spi >= 64) {
3729 verbose(env, "BUG spi %d\n", spi);
3730 WARN_ONCE(1, "verifier backtracking bug");
3731 return -EFAULT;
3732 }
3733 bt_set_slot(bt, spi);
3734 } else if (class == BPF_STX || class == BPF_ST) {
3735 if (bt_is_reg_set(bt, dreg))
3736 /* stx & st shouldn't be using _scalar_ dst_reg
3737 * to access memory. It means backtracking
3738 * encountered a case of pointer subtraction.
3739 */
3740 return -ENOTSUPP;
3741 /* scalars can only be spilled into stack */
3742 if (insn->dst_reg != BPF_REG_FP)
3743 return 0;
3744 spi = (-insn->off - 1) / BPF_REG_SIZE;
3745 if (spi >= 64) {
3746 verbose(env, "BUG spi %d\n", spi);
3747 WARN_ONCE(1, "verifier backtracking bug");
3748 return -EFAULT;
3749 }
3750 if (!bt_is_slot_set(bt, spi))
3751 return 0;
3752 bt_clear_slot(bt, spi);
3753 if (class == BPF_STX)
3754 bt_set_reg(bt, sreg);
3755 } else if (class == BPF_JMP || class == BPF_JMP32) {
3756 if (bpf_pseudo_call(insn)) {
3757 int subprog_insn_idx, subprog;
3758
3759 subprog_insn_idx = idx + insn->imm + 1;
3760 subprog = find_subprog(env, subprog_insn_idx);
3761 if (subprog < 0)
3762 return -EFAULT;
3763
3764 if (subprog_is_global(env, subprog)) {
3765 /* check that jump history doesn't have any
3766 * extra instructions from subprog; the next
3767 * instruction after call to global subprog
3768 * should be literally next instruction in
3769 * caller program
3770 */
3771 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3772 /* r1-r5 are invalidated after subprog call,
3773 * so for global func call it shouldn't be set
3774 * anymore
3775 */
3776 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3777 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3778 WARN_ONCE(1, "verifier backtracking bug");
3779 return -EFAULT;
3780 }
3781 /* global subprog always sets R0 */
3782 bt_clear_reg(bt, BPF_REG_0);
3783 return 0;
3784 } else {
3785 /* static subprog call instruction, which
3786 * means that we are exiting current subprog,
3787 * so only r1-r5 could be still requested as
3788 * precise, r0 and r6-r10 or any stack slot in
3789 * the current frame should be zero by now
3790 */
3791 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3792 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3793 WARN_ONCE(1, "verifier backtracking bug");
3794 return -EFAULT;
3795 }
3796 /* we don't track register spills perfectly,
3797 * so fallback to force-precise instead of failing */
3798 if (bt_stack_mask(bt) != 0)
3799 return -ENOTSUPP;
3800 /* propagate r1-r5 to the caller */
3801 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3802 if (bt_is_reg_set(bt, i)) {
3803 bt_clear_reg(bt, i);
3804 bt_set_frame_reg(bt, bt->frame - 1, i);
3805 }
3806 }
3807 if (bt_subprog_exit(bt))
3808 return -EFAULT;
3809 return 0;
3810 }
3811 } else if (is_sync_callback_calling_insn(insn) && idx != subseq_idx - 1) {
3812 /* exit from callback subprog to callback-calling helper or
3813 * kfunc call. Use idx/subseq_idx check to discern it from
3814 * straight line code backtracking.
3815 * Unlike the subprog call handling above, we shouldn't
3816 * propagate precision of r1-r5 (if any requested), as they are
3817 * not actually arguments passed directly to callback subprogs
3818 */
3819 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3820 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3821 WARN_ONCE(1, "verifier backtracking bug");
3822 return -EFAULT;
3823 }
3824 if (bt_stack_mask(bt) != 0)
3825 return -ENOTSUPP;
3826 /* clear r1-r5 in callback subprog's mask */
3827 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3828 bt_clear_reg(bt, i);
3829 if (bt_subprog_exit(bt))
3830 return -EFAULT;
3831 return 0;
3832 } else if (opcode == BPF_CALL) {
3833 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3834 * catch this error later. Make backtracking conservative
3835 * with ENOTSUPP.
3836 */
3837 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3838 return -ENOTSUPP;
3839 /* regular helper call sets R0 */
3840 bt_clear_reg(bt, BPF_REG_0);
3841 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3842 /* if backtracing was looking for registers R1-R5
3843 * they should have been found already.
3844 */
3845 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3846 WARN_ONCE(1, "verifier backtracking bug");
3847 return -EFAULT;
3848 }
3849 } else if (opcode == BPF_EXIT) {
3850 bool r0_precise;
3851
3852 /* Backtracking to a nested function call, 'idx' is a part of
3853 * the inner frame 'subseq_idx' is a part of the outer frame.
3854 * In case of a regular function call, instructions giving
3855 * precision to registers R1-R5 should have been found already.
3856 * In case of a callback, it is ok to have R1-R5 marked for
3857 * backtracking, as these registers are set by the function
3858 * invoking callback.
3859 */
3860 if (subseq_idx >= 0 && calls_callback(env, subseq_idx))
3861 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3862 bt_clear_reg(bt, i);
3863 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3864 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3865 WARN_ONCE(1, "verifier backtracking bug");
3866 return -EFAULT;
3867 }
3868
3869 /* BPF_EXIT in subprog or callback always returns
3870 * right after the call instruction, so by checking
3871 * whether the instruction at subseq_idx-1 is subprog
3872 * call or not we can distinguish actual exit from
3873 * *subprog* from exit from *callback*. In the former
3874 * case, we need to propagate r0 precision, if
3875 * necessary. In the former we never do that.
3876 */
3877 r0_precise = subseq_idx - 1 >= 0 &&
3878 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3879 bt_is_reg_set(bt, BPF_REG_0);
3880
3881 bt_clear_reg(bt, BPF_REG_0);
3882 if (bt_subprog_enter(bt))
3883 return -EFAULT;
3884
3885 if (r0_precise)
3886 bt_set_reg(bt, BPF_REG_0);
3887 /* r6-r9 and stack slots will stay set in caller frame
3888 * bitmasks until we return back from callee(s)
3889 */
3890 return 0;
3891 } else if (BPF_SRC(insn->code) == BPF_X) {
3892 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3893 return 0;
3894 /* dreg <cond> sreg
3895 * Both dreg and sreg need precision before
3896 * this insn. If only sreg was marked precise
3897 * before it would be equally necessary to
3898 * propagate it to dreg.
3899 */
3900 bt_set_reg(bt, dreg);
3901 bt_set_reg(bt, sreg);
3902 /* else dreg <cond> K
3903 * Only dreg still needs precision before
3904 * this insn, so for the K-based conditional
3905 * there is nothing new to be marked.
3906 */
3907 }
3908 } else if (class == BPF_LD) {
3909 if (!bt_is_reg_set(bt, dreg))
3910 return 0;
3911 bt_clear_reg(bt, dreg);
3912 /* It's ld_imm64 or ld_abs or ld_ind.
3913 * For ld_imm64 no further tracking of precision
3914 * into parent is necessary
3915 */
3916 if (mode == BPF_IND || mode == BPF_ABS)
3917 /* to be analyzed */
3918 return -ENOTSUPP;
3919 }
3920 return 0;
3921 }
3922
3923 /* the scalar precision tracking algorithm:
3924 * . at the start all registers have precise=false.
3925 * . scalar ranges are tracked as normal through alu and jmp insns.
3926 * . once precise value of the scalar register is used in:
3927 * . ptr + scalar alu
3928 * . if (scalar cond K|scalar)
3929 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3930 * backtrack through the verifier states and mark all registers and
3931 * stack slots with spilled constants that these scalar regisers
3932 * should be precise.
3933 * . during state pruning two registers (or spilled stack slots)
3934 * are equivalent if both are not precise.
3935 *
3936 * Note the verifier cannot simply walk register parentage chain,
3937 * since many different registers and stack slots could have been
3938 * used to compute single precise scalar.
3939 *
3940 * The approach of starting with precise=true for all registers and then
3941 * backtrack to mark a register as not precise when the verifier detects
3942 * that program doesn't care about specific value (e.g., when helper
3943 * takes register as ARG_ANYTHING parameter) is not safe.
3944 *
3945 * It's ok to walk single parentage chain of the verifier states.
3946 * It's possible that this backtracking will go all the way till 1st insn.
3947 * All other branches will be explored for needing precision later.
3948 *
3949 * The backtracking needs to deal with cases like:
3950 * R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0)
3951 * r9 -= r8
3952 * r5 = r9
3953 * if r5 > 0x79f goto pc+7
3954 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3955 * r5 += 1
3956 * ...
3957 * call bpf_perf_event_output#25
3958 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3959 *
3960 * and this case:
3961 * r6 = 1
3962 * call foo // uses callee's r6 inside to compute r0
3963 * r0 += r6
3964 * if r0 == 0 goto
3965 *
3966 * to track above reg_mask/stack_mask needs to be independent for each frame.
3967 *
3968 * Also if parent's curframe > frame where backtracking started,
3969 * the verifier need to mark registers in both frames, otherwise callees
3970 * may incorrectly prune callers. This is similar to
3971 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3972 *
3973 * For now backtracking falls back into conservative marking.
3974 */
mark_all_scalars_precise(struct bpf_verifier_env * env,struct bpf_verifier_state * st)3975 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3976 struct bpf_verifier_state *st)
3977 {
3978 struct bpf_func_state *func;
3979 struct bpf_reg_state *reg;
3980 int i, j;
3981
3982 if (env->log.level & BPF_LOG_LEVEL2) {
3983 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3984 st->curframe);
3985 }
3986
3987 /* big hammer: mark all scalars precise in this path.
3988 * pop_stack may still get !precise scalars.
3989 * We also skip current state and go straight to first parent state,
3990 * because precision markings in current non-checkpointed state are
3991 * not needed. See why in the comment in __mark_chain_precision below.
3992 */
3993 for (st = st->parent; st; st = st->parent) {
3994 for (i = 0; i <= st->curframe; i++) {
3995 func = st->frame[i];
3996 for (j = 0; j < BPF_REG_FP; j++) {
3997 reg = &func->regs[j];
3998 if (reg->type != SCALAR_VALUE || reg->precise)
3999 continue;
4000 reg->precise = true;
4001 if (env->log.level & BPF_LOG_LEVEL2) {
4002 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
4003 i, j);
4004 }
4005 }
4006 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4007 if (!is_spilled_reg(&func->stack[j]))
4008 continue;
4009 reg = &func->stack[j].spilled_ptr;
4010 if (reg->type != SCALAR_VALUE || reg->precise)
4011 continue;
4012 reg->precise = true;
4013 if (env->log.level & BPF_LOG_LEVEL2) {
4014 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
4015 i, -(j + 1) * 8);
4016 }
4017 }
4018 }
4019 }
4020 }
4021
mark_all_scalars_imprecise(struct bpf_verifier_env * env,struct bpf_verifier_state * st)4022 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
4023 {
4024 struct bpf_func_state *func;
4025 struct bpf_reg_state *reg;
4026 int i, j;
4027
4028 for (i = 0; i <= st->curframe; i++) {
4029 func = st->frame[i];
4030 for (j = 0; j < BPF_REG_FP; j++) {
4031 reg = &func->regs[j];
4032 if (reg->type != SCALAR_VALUE)
4033 continue;
4034 reg->precise = false;
4035 }
4036 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4037 if (!is_spilled_reg(&func->stack[j]))
4038 continue;
4039 reg = &func->stack[j].spilled_ptr;
4040 if (reg->type != SCALAR_VALUE)
4041 continue;
4042 reg->precise = false;
4043 }
4044 }
4045 }
4046
idset_contains(struct bpf_idset * s,u32 id)4047 static bool idset_contains(struct bpf_idset *s, u32 id)
4048 {
4049 u32 i;
4050
4051 for (i = 0; i < s->count; ++i)
4052 if (s->ids[i] == id)
4053 return true;
4054
4055 return false;
4056 }
4057
idset_push(struct bpf_idset * s,u32 id)4058 static int idset_push(struct bpf_idset *s, u32 id)
4059 {
4060 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
4061 return -EFAULT;
4062 s->ids[s->count++] = id;
4063 return 0;
4064 }
4065
idset_reset(struct bpf_idset * s)4066 static void idset_reset(struct bpf_idset *s)
4067 {
4068 s->count = 0;
4069 }
4070
4071 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
4072 * Mark all registers with these IDs as precise.
4073 */
mark_precise_scalar_ids(struct bpf_verifier_env * env,struct bpf_verifier_state * st)4074 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
4075 {
4076 struct bpf_idset *precise_ids = &env->idset_scratch;
4077 struct backtrack_state *bt = &env->bt;
4078 struct bpf_func_state *func;
4079 struct bpf_reg_state *reg;
4080 DECLARE_BITMAP(mask, 64);
4081 int i, fr;
4082
4083 idset_reset(precise_ids);
4084
4085 for (fr = bt->frame; fr >= 0; fr--) {
4086 func = st->frame[fr];
4087
4088 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4089 for_each_set_bit(i, mask, 32) {
4090 reg = &func->regs[i];
4091 if (!reg->id || reg->type != SCALAR_VALUE)
4092 continue;
4093 if (idset_push(precise_ids, reg->id))
4094 return -EFAULT;
4095 }
4096
4097 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4098 for_each_set_bit(i, mask, 64) {
4099 if (i >= func->allocated_stack / BPF_REG_SIZE)
4100 break;
4101 if (!is_spilled_scalar_reg(&func->stack[i]))
4102 continue;
4103 reg = &func->stack[i].spilled_ptr;
4104 if (!reg->id)
4105 continue;
4106 if (idset_push(precise_ids, reg->id))
4107 return -EFAULT;
4108 }
4109 }
4110
4111 for (fr = 0; fr <= st->curframe; ++fr) {
4112 func = st->frame[fr];
4113
4114 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
4115 reg = &func->regs[i];
4116 if (!reg->id)
4117 continue;
4118 if (!idset_contains(precise_ids, reg->id))
4119 continue;
4120 bt_set_frame_reg(bt, fr, i);
4121 }
4122 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
4123 if (!is_spilled_scalar_reg(&func->stack[i]))
4124 continue;
4125 reg = &func->stack[i].spilled_ptr;
4126 if (!reg->id)
4127 continue;
4128 if (!idset_contains(precise_ids, reg->id))
4129 continue;
4130 bt_set_frame_slot(bt, fr, i);
4131 }
4132 }
4133
4134 return 0;
4135 }
4136
4137 /*
4138 * __mark_chain_precision() backtracks BPF program instruction sequence and
4139 * chain of verifier states making sure that register *regno* (if regno >= 0)
4140 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
4141 * SCALARS, as well as any other registers and slots that contribute to
4142 * a tracked state of given registers/stack slots, depending on specific BPF
4143 * assembly instructions (see backtrack_insns() for exact instruction handling
4144 * logic). This backtracking relies on recorded jmp_history and is able to
4145 * traverse entire chain of parent states. This process ends only when all the
4146 * necessary registers/slots and their transitive dependencies are marked as
4147 * precise.
4148 *
4149 * One important and subtle aspect is that precise marks *do not matter* in
4150 * the currently verified state (current state). It is important to understand
4151 * why this is the case.
4152 *
4153 * First, note that current state is the state that is not yet "checkpointed",
4154 * i.e., it is not yet put into env->explored_states, and it has no children
4155 * states as well. It's ephemeral, and can end up either a) being discarded if
4156 * compatible explored state is found at some point or BPF_EXIT instruction is
4157 * reached or b) checkpointed and put into env->explored_states, branching out
4158 * into one or more children states.
4159 *
4160 * In the former case, precise markings in current state are completely
4161 * ignored by state comparison code (see regsafe() for details). Only
4162 * checkpointed ("old") state precise markings are important, and if old
4163 * state's register/slot is precise, regsafe() assumes current state's
4164 * register/slot as precise and checks value ranges exactly and precisely. If
4165 * states turn out to be compatible, current state's necessary precise
4166 * markings and any required parent states' precise markings are enforced
4167 * after the fact with propagate_precision() logic, after the fact. But it's
4168 * important to realize that in this case, even after marking current state
4169 * registers/slots as precise, we immediately discard current state. So what
4170 * actually matters is any of the precise markings propagated into current
4171 * state's parent states, which are always checkpointed (due to b) case above).
4172 * As such, for scenario a) it doesn't matter if current state has precise
4173 * markings set or not.
4174 *
4175 * Now, for the scenario b), checkpointing and forking into child(ren)
4176 * state(s). Note that before current state gets to checkpointing step, any
4177 * processed instruction always assumes precise SCALAR register/slot
4178 * knowledge: if precise value or range is useful to prune jump branch, BPF
4179 * verifier takes this opportunity enthusiastically. Similarly, when
4180 * register's value is used to calculate offset or memory address, exact
4181 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
4182 * what we mentioned above about state comparison ignoring precise markings
4183 * during state comparison, BPF verifier ignores and also assumes precise
4184 * markings *at will* during instruction verification process. But as verifier
4185 * assumes precision, it also propagates any precision dependencies across
4186 * parent states, which are not yet finalized, so can be further restricted
4187 * based on new knowledge gained from restrictions enforced by their children
4188 * states. This is so that once those parent states are finalized, i.e., when
4189 * they have no more active children state, state comparison logic in
4190 * is_state_visited() would enforce strict and precise SCALAR ranges, if
4191 * required for correctness.
4192 *
4193 * To build a bit more intuition, note also that once a state is checkpointed,
4194 * the path we took to get to that state is not important. This is crucial
4195 * property for state pruning. When state is checkpointed and finalized at
4196 * some instruction index, it can be correctly and safely used to "short
4197 * circuit" any *compatible* state that reaches exactly the same instruction
4198 * index. I.e., if we jumped to that instruction from a completely different
4199 * code path than original finalized state was derived from, it doesn't
4200 * matter, current state can be discarded because from that instruction
4201 * forward having a compatible state will ensure we will safely reach the
4202 * exit. States describe preconditions for further exploration, but completely
4203 * forget the history of how we got here.
4204 *
4205 * This also means that even if we needed precise SCALAR range to get to
4206 * finalized state, but from that point forward *that same* SCALAR register is
4207 * never used in a precise context (i.e., it's precise value is not needed for
4208 * correctness), it's correct and safe to mark such register as "imprecise"
4209 * (i.e., precise marking set to false). This is what we rely on when we do
4210 * not set precise marking in current state. If no child state requires
4211 * precision for any given SCALAR register, it's safe to dictate that it can
4212 * be imprecise. If any child state does require this register to be precise,
4213 * we'll mark it precise later retroactively during precise markings
4214 * propagation from child state to parent states.
4215 *
4216 * Skipping precise marking setting in current state is a mild version of
4217 * relying on the above observation. But we can utilize this property even
4218 * more aggressively by proactively forgetting any precise marking in the
4219 * current state (which we inherited from the parent state), right before we
4220 * checkpoint it and branch off into new child state. This is done by
4221 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
4222 * finalized states which help in short circuiting more future states.
4223 */
__mark_chain_precision(struct bpf_verifier_env * env,int regno)4224 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
4225 {
4226 struct backtrack_state *bt = &env->bt;
4227 struct bpf_verifier_state *st = env->cur_state;
4228 int first_idx = st->first_insn_idx;
4229 int last_idx = env->insn_idx;
4230 int subseq_idx = -1;
4231 struct bpf_func_state *func;
4232 struct bpf_reg_state *reg;
4233 bool skip_first = true;
4234 int i, fr, err;
4235
4236 if (!env->bpf_capable)
4237 return 0;
4238
4239 /* set frame number from which we are starting to backtrack */
4240 bt_init(bt, env->cur_state->curframe);
4241
4242 /* Do sanity checks against current state of register and/or stack
4243 * slot, but don't set precise flag in current state, as precision
4244 * tracking in the current state is unnecessary.
4245 */
4246 func = st->frame[bt->frame];
4247 if (regno >= 0) {
4248 reg = &func->regs[regno];
4249 if (reg->type != SCALAR_VALUE) {
4250 WARN_ONCE(1, "backtracing misuse");
4251 return -EFAULT;
4252 }
4253 bt_set_reg(bt, regno);
4254 }
4255
4256 if (bt_empty(bt))
4257 return 0;
4258
4259 for (;;) {
4260 DECLARE_BITMAP(mask, 64);
4261 u32 history = st->jmp_history_cnt;
4262
4263 if (env->log.level & BPF_LOG_LEVEL2) {
4264 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4265 bt->frame, last_idx, first_idx, subseq_idx);
4266 }
4267
4268 /* If some register with scalar ID is marked as precise,
4269 * make sure that all registers sharing this ID are also precise.
4270 * This is needed to estimate effect of find_equal_scalars().
4271 * Do this at the last instruction of each state,
4272 * bpf_reg_state::id fields are valid for these instructions.
4273 *
4274 * Allows to track precision in situation like below:
4275 *
4276 * r2 = unknown value
4277 * ...
4278 * --- state #0 ---
4279 * ...
4280 * r1 = r2 // r1 and r2 now share the same ID
4281 * ...
4282 * --- state #1 {r1.id = A, r2.id = A} ---
4283 * ...
4284 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4285 * ...
4286 * --- state #2 {r1.id = A, r2.id = A} ---
4287 * r3 = r10
4288 * r3 += r1 // need to mark both r1 and r2
4289 */
4290 if (mark_precise_scalar_ids(env, st))
4291 return -EFAULT;
4292
4293 if (last_idx < 0) {
4294 /* we are at the entry into subprog, which
4295 * is expected for global funcs, but only if
4296 * requested precise registers are R1-R5
4297 * (which are global func's input arguments)
4298 */
4299 if (st->curframe == 0 &&
4300 st->frame[0]->subprogno > 0 &&
4301 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4302 bt_stack_mask(bt) == 0 &&
4303 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4304 bitmap_from_u64(mask, bt_reg_mask(bt));
4305 for_each_set_bit(i, mask, 32) {
4306 reg = &st->frame[0]->regs[i];
4307 bt_clear_reg(bt, i);
4308 if (reg->type == SCALAR_VALUE)
4309 reg->precise = true;
4310 }
4311 return 0;
4312 }
4313
4314 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4315 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4316 WARN_ONCE(1, "verifier backtracking bug");
4317 return -EFAULT;
4318 }
4319
4320 for (i = last_idx;;) {
4321 if (skip_first) {
4322 err = 0;
4323 skip_first = false;
4324 } else {
4325 err = backtrack_insn(env, i, subseq_idx, bt);
4326 }
4327 if (err == -ENOTSUPP) {
4328 mark_all_scalars_precise(env, env->cur_state);
4329 bt_reset(bt);
4330 return 0;
4331 } else if (err) {
4332 return err;
4333 }
4334 if (bt_empty(bt))
4335 /* Found assignment(s) into tracked register in this state.
4336 * Since this state is already marked, just return.
4337 * Nothing to be tracked further in the parent state.
4338 */
4339 return 0;
4340 subseq_idx = i;
4341 i = get_prev_insn_idx(st, i, &history);
4342 if (i == -ENOENT)
4343 break;
4344 if (i >= env->prog->len) {
4345 /* This can happen if backtracking reached insn 0
4346 * and there are still reg_mask or stack_mask
4347 * to backtrack.
4348 * It means the backtracking missed the spot where
4349 * particular register was initialized with a constant.
4350 */
4351 verbose(env, "BUG backtracking idx %d\n", i);
4352 WARN_ONCE(1, "verifier backtracking bug");
4353 return -EFAULT;
4354 }
4355 }
4356 st = st->parent;
4357 if (!st)
4358 break;
4359
4360 for (fr = bt->frame; fr >= 0; fr--) {
4361 func = st->frame[fr];
4362 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4363 for_each_set_bit(i, mask, 32) {
4364 reg = &func->regs[i];
4365 if (reg->type != SCALAR_VALUE) {
4366 bt_clear_frame_reg(bt, fr, i);
4367 continue;
4368 }
4369 if (reg->precise)
4370 bt_clear_frame_reg(bt, fr, i);
4371 else
4372 reg->precise = true;
4373 }
4374
4375 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4376 for_each_set_bit(i, mask, 64) {
4377 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4378 /* the sequence of instructions:
4379 * 2: (bf) r3 = r10
4380 * 3: (7b) *(u64 *)(r3 -8) = r0
4381 * 4: (79) r4 = *(u64 *)(r10 -8)
4382 * doesn't contain jmps. It's backtracked
4383 * as a single block.
4384 * During backtracking insn 3 is not recognized as
4385 * stack access, so at the end of backtracking
4386 * stack slot fp-8 is still marked in stack_mask.
4387 * However the parent state may not have accessed
4388 * fp-8 and it's "unallocated" stack space.
4389 * In such case fallback to conservative.
4390 */
4391 mark_all_scalars_precise(env, env->cur_state);
4392 bt_reset(bt);
4393 return 0;
4394 }
4395
4396 if (!is_spilled_scalar_reg(&func->stack[i])) {
4397 bt_clear_frame_slot(bt, fr, i);
4398 continue;
4399 }
4400 reg = &func->stack[i].spilled_ptr;
4401 if (reg->precise)
4402 bt_clear_frame_slot(bt, fr, i);
4403 else
4404 reg->precise = true;
4405 }
4406 if (env->log.level & BPF_LOG_LEVEL2) {
4407 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4408 bt_frame_reg_mask(bt, fr));
4409 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4410 fr, env->tmp_str_buf);
4411 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4412 bt_frame_stack_mask(bt, fr));
4413 verbose(env, "stack=%s: ", env->tmp_str_buf);
4414 print_verifier_state(env, func, true);
4415 }
4416 }
4417
4418 if (bt_empty(bt))
4419 return 0;
4420
4421 subseq_idx = first_idx;
4422 last_idx = st->last_insn_idx;
4423 first_idx = st->first_insn_idx;
4424 }
4425
4426 /* if we still have requested precise regs or slots, we missed
4427 * something (e.g., stack access through non-r10 register), so
4428 * fallback to marking all precise
4429 */
4430 if (!bt_empty(bt)) {
4431 mark_all_scalars_precise(env, env->cur_state);
4432 bt_reset(bt);
4433 }
4434
4435 return 0;
4436 }
4437
mark_chain_precision(struct bpf_verifier_env * env,int regno)4438 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4439 {
4440 return __mark_chain_precision(env, regno);
4441 }
4442
4443 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4444 * desired reg and stack masks across all relevant frames
4445 */
mark_chain_precision_batch(struct bpf_verifier_env * env)4446 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4447 {
4448 return __mark_chain_precision(env, -1);
4449 }
4450
is_spillable_regtype(enum bpf_reg_type type)4451 static bool is_spillable_regtype(enum bpf_reg_type type)
4452 {
4453 switch (base_type(type)) {
4454 case PTR_TO_MAP_VALUE:
4455 case PTR_TO_STACK:
4456 case PTR_TO_CTX:
4457 case PTR_TO_PACKET:
4458 case PTR_TO_PACKET_META:
4459 case PTR_TO_PACKET_END:
4460 case PTR_TO_FLOW_KEYS:
4461 case CONST_PTR_TO_MAP:
4462 case PTR_TO_SOCKET:
4463 case PTR_TO_SOCK_COMMON:
4464 case PTR_TO_TCP_SOCK:
4465 case PTR_TO_XDP_SOCK:
4466 case PTR_TO_BTF_ID:
4467 case PTR_TO_BUF:
4468 case PTR_TO_MEM:
4469 case PTR_TO_FUNC:
4470 case PTR_TO_MAP_KEY:
4471 return true;
4472 default:
4473 return false;
4474 }
4475 }
4476
4477 /* Does this register contain a constant zero? */
register_is_null(struct bpf_reg_state * reg)4478 static bool register_is_null(struct bpf_reg_state *reg)
4479 {
4480 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4481 }
4482
register_is_const(struct bpf_reg_state * reg)4483 static bool register_is_const(struct bpf_reg_state *reg)
4484 {
4485 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4486 }
4487
__is_scalar_unbounded(struct bpf_reg_state * reg)4488 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4489 {
4490 return tnum_is_unknown(reg->var_off) &&
4491 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4492 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4493 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4494 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4495 }
4496
register_is_bounded(struct bpf_reg_state * reg)4497 static bool register_is_bounded(struct bpf_reg_state *reg)
4498 {
4499 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4500 }
4501
__is_pointer_value(bool allow_ptr_leaks,const struct bpf_reg_state * reg)4502 static bool __is_pointer_value(bool allow_ptr_leaks,
4503 const struct bpf_reg_state *reg)
4504 {
4505 if (allow_ptr_leaks)
4506 return false;
4507
4508 return reg->type != SCALAR_VALUE;
4509 }
4510
4511 /* Copy src state preserving dst->parent and dst->live fields */
copy_register_state(struct bpf_reg_state * dst,const struct bpf_reg_state * src)4512 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4513 {
4514 struct bpf_reg_state *parent = dst->parent;
4515 enum bpf_reg_liveness live = dst->live;
4516
4517 *dst = *src;
4518 dst->parent = parent;
4519 dst->live = live;
4520 }
4521
save_register_state(struct bpf_func_state * state,int spi,struct bpf_reg_state * reg,int size)4522 static void save_register_state(struct bpf_func_state *state,
4523 int spi, struct bpf_reg_state *reg,
4524 int size)
4525 {
4526 int i;
4527
4528 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4529 if (size == BPF_REG_SIZE)
4530 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4531
4532 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4533 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4534
4535 /* size < 8 bytes spill */
4536 for (; i; i--)
4537 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4538 }
4539
is_bpf_st_mem(struct bpf_insn * insn)4540 static bool is_bpf_st_mem(struct bpf_insn *insn)
4541 {
4542 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4543 }
4544
4545 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4546 * stack boundary and alignment are checked in check_mem_access()
4547 */
check_stack_write_fixed_off(struct bpf_verifier_env * env,struct bpf_func_state * state,int off,int size,int value_regno,int insn_idx)4548 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4549 /* stack frame we're writing to */
4550 struct bpf_func_state *state,
4551 int off, int size, int value_regno,
4552 int insn_idx)
4553 {
4554 struct bpf_func_state *cur; /* state of the current function */
4555 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4556 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4557 struct bpf_reg_state *reg = NULL;
4558 u32 dst_reg = insn->dst_reg;
4559
4560 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4561 * so it's aligned access and [off, off + size) are within stack limits
4562 */
4563 if (!env->allow_ptr_leaks &&
4564 is_spilled_reg(&state->stack[spi]) &&
4565 size != BPF_REG_SIZE) {
4566 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4567 return -EACCES;
4568 }
4569
4570 cur = env->cur_state->frame[env->cur_state->curframe];
4571 if (value_regno >= 0)
4572 reg = &cur->regs[value_regno];
4573 if (!env->bypass_spec_v4) {
4574 bool sanitize = reg && is_spillable_regtype(reg->type);
4575
4576 for (i = 0; i < size; i++) {
4577 u8 type = state->stack[spi].slot_type[i];
4578
4579 if (type != STACK_MISC && type != STACK_ZERO) {
4580 sanitize = true;
4581 break;
4582 }
4583 }
4584
4585 if (sanitize)
4586 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4587 }
4588
4589 err = destroy_if_dynptr_stack_slot(env, state, spi);
4590 if (err)
4591 return err;
4592
4593 mark_stack_slot_scratched(env, spi);
4594 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4595 !register_is_null(reg) && env->bpf_capable) {
4596 if (dst_reg != BPF_REG_FP) {
4597 /* The backtracking logic can only recognize explicit
4598 * stack slot address like [fp - 8]. Other spill of
4599 * scalar via different register has to be conservative.
4600 * Backtrack from here and mark all registers as precise
4601 * that contributed into 'reg' being a constant.
4602 */
4603 err = mark_chain_precision(env, value_regno);
4604 if (err)
4605 return err;
4606 }
4607 save_register_state(state, spi, reg, size);
4608 /* Break the relation on a narrowing spill. */
4609 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4610 state->stack[spi].spilled_ptr.id = 0;
4611 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4612 insn->imm != 0 && env->bpf_capable) {
4613 struct bpf_reg_state fake_reg = {};
4614
4615 __mark_reg_known(&fake_reg, insn->imm);
4616 fake_reg.type = SCALAR_VALUE;
4617 save_register_state(state, spi, &fake_reg, size);
4618 } else if (reg && is_spillable_regtype(reg->type)) {
4619 /* register containing pointer is being spilled into stack */
4620 if (size != BPF_REG_SIZE) {
4621 verbose_linfo(env, insn_idx, "; ");
4622 verbose(env, "invalid size of register spill\n");
4623 return -EACCES;
4624 }
4625 if (state != cur && reg->type == PTR_TO_STACK) {
4626 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4627 return -EINVAL;
4628 }
4629 save_register_state(state, spi, reg, size);
4630 } else {
4631 u8 type = STACK_MISC;
4632
4633 /* regular write of data into stack destroys any spilled ptr */
4634 state->stack[spi].spilled_ptr.type = NOT_INIT;
4635 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4636 if (is_stack_slot_special(&state->stack[spi]))
4637 for (i = 0; i < BPF_REG_SIZE; i++)
4638 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4639
4640 /* only mark the slot as written if all 8 bytes were written
4641 * otherwise read propagation may incorrectly stop too soon
4642 * when stack slots are partially written.
4643 * This heuristic means that read propagation will be
4644 * conservative, since it will add reg_live_read marks
4645 * to stack slots all the way to first state when programs
4646 * writes+reads less than 8 bytes
4647 */
4648 if (size == BPF_REG_SIZE)
4649 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4650
4651 /* when we zero initialize stack slots mark them as such */
4652 if ((reg && register_is_null(reg)) ||
4653 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4654 /* backtracking doesn't work for STACK_ZERO yet. */
4655 err = mark_chain_precision(env, value_regno);
4656 if (err)
4657 return err;
4658 type = STACK_ZERO;
4659 }
4660
4661 /* Mark slots affected by this stack write. */
4662 for (i = 0; i < size; i++)
4663 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4664 type;
4665 }
4666 return 0;
4667 }
4668
4669 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4670 * known to contain a variable offset.
4671 * This function checks whether the write is permitted and conservatively
4672 * tracks the effects of the write, considering that each stack slot in the
4673 * dynamic range is potentially written to.
4674 *
4675 * 'off' includes 'regno->off'.
4676 * 'value_regno' can be -1, meaning that an unknown value is being written to
4677 * the stack.
4678 *
4679 * Spilled pointers in range are not marked as written because we don't know
4680 * what's going to be actually written. This means that read propagation for
4681 * future reads cannot be terminated by this write.
4682 *
4683 * For privileged programs, uninitialized stack slots are considered
4684 * initialized by this write (even though we don't know exactly what offsets
4685 * are going to be written to). The idea is that we don't want the verifier to
4686 * reject future reads that access slots written to through variable offsets.
4687 */
check_stack_write_var_off(struct bpf_verifier_env * env,struct bpf_func_state * state,int ptr_regno,int off,int size,int value_regno,int insn_idx)4688 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4689 /* func where register points to */
4690 struct bpf_func_state *state,
4691 int ptr_regno, int off, int size,
4692 int value_regno, int insn_idx)
4693 {
4694 struct bpf_func_state *cur; /* state of the current function */
4695 int min_off, max_off;
4696 int i, err;
4697 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4698 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4699 bool writing_zero = false;
4700 /* set if the fact that we're writing a zero is used to let any
4701 * stack slots remain STACK_ZERO
4702 */
4703 bool zero_used = false;
4704
4705 cur = env->cur_state->frame[env->cur_state->curframe];
4706 ptr_reg = &cur->regs[ptr_regno];
4707 min_off = ptr_reg->smin_value + off;
4708 max_off = ptr_reg->smax_value + off + size;
4709 if (value_regno >= 0)
4710 value_reg = &cur->regs[value_regno];
4711 if ((value_reg && register_is_null(value_reg)) ||
4712 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4713 writing_zero = true;
4714
4715 for (i = min_off; i < max_off; i++) {
4716 int spi;
4717
4718 spi = __get_spi(i);
4719 err = destroy_if_dynptr_stack_slot(env, state, spi);
4720 if (err)
4721 return err;
4722 }
4723
4724 /* Variable offset writes destroy any spilled pointers in range. */
4725 for (i = min_off; i < max_off; i++) {
4726 u8 new_type, *stype;
4727 int slot, spi;
4728
4729 slot = -i - 1;
4730 spi = slot / BPF_REG_SIZE;
4731 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4732 mark_stack_slot_scratched(env, spi);
4733
4734 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4735 /* Reject the write if range we may write to has not
4736 * been initialized beforehand. If we didn't reject
4737 * here, the ptr status would be erased below (even
4738 * though not all slots are actually overwritten),
4739 * possibly opening the door to leaks.
4740 *
4741 * We do however catch STACK_INVALID case below, and
4742 * only allow reading possibly uninitialized memory
4743 * later for CAP_PERFMON, as the write may not happen to
4744 * that slot.
4745 */
4746 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4747 insn_idx, i);
4748 return -EINVAL;
4749 }
4750
4751 /* Erase all spilled pointers. */
4752 state->stack[spi].spilled_ptr.type = NOT_INIT;
4753
4754 /* Update the slot type. */
4755 new_type = STACK_MISC;
4756 if (writing_zero && *stype == STACK_ZERO) {
4757 new_type = STACK_ZERO;
4758 zero_used = true;
4759 }
4760 /* If the slot is STACK_INVALID, we check whether it's OK to
4761 * pretend that it will be initialized by this write. The slot
4762 * might not actually be written to, and so if we mark it as
4763 * initialized future reads might leak uninitialized memory.
4764 * For privileged programs, we will accept such reads to slots
4765 * that may or may not be written because, if we're reject
4766 * them, the error would be too confusing.
4767 */
4768 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4769 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4770 insn_idx, i);
4771 return -EINVAL;
4772 }
4773 *stype = new_type;
4774 }
4775 if (zero_used) {
4776 /* backtracking doesn't work for STACK_ZERO yet. */
4777 err = mark_chain_precision(env, value_regno);
4778 if (err)
4779 return err;
4780 }
4781 return 0;
4782 }
4783
4784 /* When register 'dst_regno' is assigned some values from stack[min_off,
4785 * max_off), we set the register's type according to the types of the
4786 * respective stack slots. If all the stack values are known to be zeros, then
4787 * so is the destination reg. Otherwise, the register is considered to be
4788 * SCALAR. This function does not deal with register filling; the caller must
4789 * ensure that all spilled registers in the stack range have been marked as
4790 * read.
4791 */
mark_reg_stack_read(struct bpf_verifier_env * env,struct bpf_func_state * ptr_state,int min_off,int max_off,int dst_regno)4792 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4793 /* func where src register points to */
4794 struct bpf_func_state *ptr_state,
4795 int min_off, int max_off, int dst_regno)
4796 {
4797 struct bpf_verifier_state *vstate = env->cur_state;
4798 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4799 int i, slot, spi;
4800 u8 *stype;
4801 int zeros = 0;
4802
4803 for (i = min_off; i < max_off; i++) {
4804 slot = -i - 1;
4805 spi = slot / BPF_REG_SIZE;
4806 mark_stack_slot_scratched(env, spi);
4807 stype = ptr_state->stack[spi].slot_type;
4808 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4809 break;
4810 zeros++;
4811 }
4812 if (zeros == max_off - min_off) {
4813 /* any access_size read into register is zero extended,
4814 * so the whole register == const_zero
4815 */
4816 __mark_reg_const_zero(&state->regs[dst_regno]);
4817 /* backtracking doesn't support STACK_ZERO yet,
4818 * so mark it precise here, so that later
4819 * backtracking can stop here.
4820 * Backtracking may not need this if this register
4821 * doesn't participate in pointer adjustment.
4822 * Forward propagation of precise flag is not
4823 * necessary either. This mark is only to stop
4824 * backtracking. Any register that contributed
4825 * to const 0 was marked precise before spill.
4826 */
4827 state->regs[dst_regno].precise = true;
4828 } else {
4829 /* have read misc data from the stack */
4830 mark_reg_unknown(env, state->regs, dst_regno);
4831 }
4832 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4833 }
4834
4835 /* Read the stack at 'off' and put the results into the register indicated by
4836 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4837 * spilled reg.
4838 *
4839 * 'dst_regno' can be -1, meaning that the read value is not going to a
4840 * register.
4841 *
4842 * The access is assumed to be within the current stack bounds.
4843 */
check_stack_read_fixed_off(struct bpf_verifier_env * env,struct bpf_func_state * reg_state,int off,int size,int dst_regno)4844 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4845 /* func where src register points to */
4846 struct bpf_func_state *reg_state,
4847 int off, int size, int dst_regno)
4848 {
4849 struct bpf_verifier_state *vstate = env->cur_state;
4850 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4851 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4852 struct bpf_reg_state *reg;
4853 u8 *stype, type;
4854
4855 stype = reg_state->stack[spi].slot_type;
4856 reg = ®_state->stack[spi].spilled_ptr;
4857
4858 mark_stack_slot_scratched(env, spi);
4859
4860 if (is_spilled_reg(®_state->stack[spi])) {
4861 u8 spill_size = 1;
4862
4863 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4864 spill_size++;
4865
4866 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4867 if (reg->type != SCALAR_VALUE) {
4868 verbose_linfo(env, env->insn_idx, "; ");
4869 verbose(env, "invalid size of register fill\n");
4870 return -EACCES;
4871 }
4872
4873 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4874 if (dst_regno < 0)
4875 return 0;
4876
4877 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4878 /* The earlier check_reg_arg() has decided the
4879 * subreg_def for this insn. Save it first.
4880 */
4881 s32 subreg_def = state->regs[dst_regno].subreg_def;
4882
4883 copy_register_state(&state->regs[dst_regno], reg);
4884 state->regs[dst_regno].subreg_def = subreg_def;
4885 } else {
4886 for (i = 0; i < size; i++) {
4887 type = stype[(slot - i) % BPF_REG_SIZE];
4888 if (type == STACK_SPILL)
4889 continue;
4890 if (type == STACK_MISC)
4891 continue;
4892 if (type == STACK_INVALID && env->allow_uninit_stack)
4893 continue;
4894 verbose(env, "invalid read from stack off %d+%d size %d\n",
4895 off, i, size);
4896 return -EACCES;
4897 }
4898 mark_reg_unknown(env, state->regs, dst_regno);
4899 }
4900 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4901 return 0;
4902 }
4903
4904 if (dst_regno >= 0) {
4905 /* restore register state from stack */
4906 copy_register_state(&state->regs[dst_regno], reg);
4907 /* mark reg as written since spilled pointer state likely
4908 * has its liveness marks cleared by is_state_visited()
4909 * which resets stack/reg liveness for state transitions
4910 */
4911 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4912 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4913 /* If dst_regno==-1, the caller is asking us whether
4914 * it is acceptable to use this value as a SCALAR_VALUE
4915 * (e.g. for XADD).
4916 * We must not allow unprivileged callers to do that
4917 * with spilled pointers.
4918 */
4919 verbose(env, "leaking pointer from stack off %d\n",
4920 off);
4921 return -EACCES;
4922 }
4923 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4924 } else {
4925 for (i = 0; i < size; i++) {
4926 type = stype[(slot - i) % BPF_REG_SIZE];
4927 if (type == STACK_MISC)
4928 continue;
4929 if (type == STACK_ZERO)
4930 continue;
4931 if (type == STACK_INVALID && env->allow_uninit_stack)
4932 continue;
4933 verbose(env, "invalid read from stack off %d+%d size %d\n",
4934 off, i, size);
4935 return -EACCES;
4936 }
4937 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4938 if (dst_regno >= 0)
4939 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4940 }
4941 return 0;
4942 }
4943
4944 enum bpf_access_src {
4945 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4946 ACCESS_HELPER = 2, /* the access is performed by a helper */
4947 };
4948
4949 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4950 int regno, int off, int access_size,
4951 bool zero_size_allowed,
4952 enum bpf_access_src type,
4953 struct bpf_call_arg_meta *meta);
4954
reg_state(struct bpf_verifier_env * env,int regno)4955 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4956 {
4957 return cur_regs(env) + regno;
4958 }
4959
4960 /* Read the stack at 'ptr_regno + off' and put the result into the register
4961 * 'dst_regno'.
4962 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4963 * but not its variable offset.
4964 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4965 *
4966 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4967 * filling registers (i.e. reads of spilled register cannot be detected when
4968 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4969 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4970 * offset; for a fixed offset check_stack_read_fixed_off should be used
4971 * instead.
4972 */
check_stack_read_var_off(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)4973 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4974 int ptr_regno, int off, int size, int dst_regno)
4975 {
4976 /* The state of the source register. */
4977 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4978 struct bpf_func_state *ptr_state = func(env, reg);
4979 int err;
4980 int min_off, max_off;
4981
4982 /* Note that we pass a NULL meta, so raw access will not be permitted.
4983 */
4984 err = check_stack_range_initialized(env, ptr_regno, off, size,
4985 false, ACCESS_DIRECT, NULL);
4986 if (err)
4987 return err;
4988
4989 min_off = reg->smin_value + off;
4990 max_off = reg->smax_value + off;
4991 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4992 return 0;
4993 }
4994
4995 /* check_stack_read dispatches to check_stack_read_fixed_off or
4996 * check_stack_read_var_off.
4997 *
4998 * The caller must ensure that the offset falls within the allocated stack
4999 * bounds.
5000 *
5001 * 'dst_regno' is a register which will receive the value from the stack. It
5002 * can be -1, meaning that the read value is not going to a register.
5003 */
check_stack_read(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)5004 static int check_stack_read(struct bpf_verifier_env *env,
5005 int ptr_regno, int off, int size,
5006 int dst_regno)
5007 {
5008 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
5009 struct bpf_func_state *state = func(env, reg);
5010 int err;
5011 /* Some accesses are only permitted with a static offset. */
5012 bool var_off = !tnum_is_const(reg->var_off);
5013
5014 /* The offset is required to be static when reads don't go to a
5015 * register, in order to not leak pointers (see
5016 * check_stack_read_fixed_off).
5017 */
5018 if (dst_regno < 0 && var_off) {
5019 char tn_buf[48];
5020
5021 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5022 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
5023 tn_buf, off, size);
5024 return -EACCES;
5025 }
5026 /* Variable offset is prohibited for unprivileged mode for simplicity
5027 * since it requires corresponding support in Spectre masking for stack
5028 * ALU. See also retrieve_ptr_limit(). The check in
5029 * check_stack_access_for_ptr_arithmetic() called by
5030 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
5031 * with variable offsets, therefore no check is required here. Further,
5032 * just checking it here would be insufficient as speculative stack
5033 * writes could still lead to unsafe speculative behaviour.
5034 */
5035 if (!var_off) {
5036 off += reg->var_off.value;
5037 err = check_stack_read_fixed_off(env, state, off, size,
5038 dst_regno);
5039 } else {
5040 /* Variable offset stack reads need more conservative handling
5041 * than fixed offset ones. Note that dst_regno >= 0 on this
5042 * branch.
5043 */
5044 err = check_stack_read_var_off(env, ptr_regno, off, size,
5045 dst_regno);
5046 }
5047 return err;
5048 }
5049
5050
5051 /* check_stack_write dispatches to check_stack_write_fixed_off or
5052 * check_stack_write_var_off.
5053 *
5054 * 'ptr_regno' is the register used as a pointer into the stack.
5055 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
5056 * 'value_regno' is the register whose value we're writing to the stack. It can
5057 * be -1, meaning that we're not writing from a register.
5058 *
5059 * The caller must ensure that the offset falls within the maximum stack size.
5060 */
check_stack_write(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int value_regno,int insn_idx)5061 static int check_stack_write(struct bpf_verifier_env *env,
5062 int ptr_regno, int off, int size,
5063 int value_regno, int insn_idx)
5064 {
5065 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
5066 struct bpf_func_state *state = func(env, reg);
5067 int err;
5068
5069 if (tnum_is_const(reg->var_off)) {
5070 off += reg->var_off.value;
5071 err = check_stack_write_fixed_off(env, state, off, size,
5072 value_regno, insn_idx);
5073 } else {
5074 /* Variable offset stack reads need more conservative handling
5075 * than fixed offset ones.
5076 */
5077 err = check_stack_write_var_off(env, state,
5078 ptr_regno, off, size,
5079 value_regno, insn_idx);
5080 }
5081 return err;
5082 }
5083
check_map_access_type(struct bpf_verifier_env * env,u32 regno,int off,int size,enum bpf_access_type type)5084 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
5085 int off, int size, enum bpf_access_type type)
5086 {
5087 struct bpf_reg_state *regs = cur_regs(env);
5088 struct bpf_map *map = regs[regno].map_ptr;
5089 u32 cap = bpf_map_flags_to_cap(map);
5090
5091 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
5092 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
5093 map->value_size, off, size);
5094 return -EACCES;
5095 }
5096
5097 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
5098 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
5099 map->value_size, off, size);
5100 return -EACCES;
5101 }
5102
5103 return 0;
5104 }
5105
5106 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
__check_mem_access(struct bpf_verifier_env * env,int regno,int off,int size,u32 mem_size,bool zero_size_allowed)5107 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
5108 int off, int size, u32 mem_size,
5109 bool zero_size_allowed)
5110 {
5111 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
5112 struct bpf_reg_state *reg;
5113
5114 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
5115 return 0;
5116
5117 reg = &cur_regs(env)[regno];
5118 switch (reg->type) {
5119 case PTR_TO_MAP_KEY:
5120 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
5121 mem_size, off, size);
5122 break;
5123 case PTR_TO_MAP_VALUE:
5124 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
5125 mem_size, off, size);
5126 break;
5127 case PTR_TO_PACKET:
5128 case PTR_TO_PACKET_META:
5129 case PTR_TO_PACKET_END:
5130 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
5131 off, size, regno, reg->id, off, mem_size);
5132 break;
5133 case PTR_TO_MEM:
5134 default:
5135 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
5136 mem_size, off, size);
5137 }
5138
5139 return -EACCES;
5140 }
5141
5142 /* check read/write into a memory region with possible variable offset */
check_mem_region_access(struct bpf_verifier_env * env,u32 regno,int off,int size,u32 mem_size,bool zero_size_allowed)5143 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
5144 int off, int size, u32 mem_size,
5145 bool zero_size_allowed)
5146 {
5147 struct bpf_verifier_state *vstate = env->cur_state;
5148 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5149 struct bpf_reg_state *reg = &state->regs[regno];
5150 int err;
5151
5152 /* We may have adjusted the register pointing to memory region, so we
5153 * need to try adding each of min_value and max_value to off
5154 * to make sure our theoretical access will be safe.
5155 *
5156 * The minimum value is only important with signed
5157 * comparisons where we can't assume the floor of a
5158 * value is 0. If we are using signed variables for our
5159 * index'es we need to make sure that whatever we use
5160 * will have a set floor within our range.
5161 */
5162 if (reg->smin_value < 0 &&
5163 (reg->smin_value == S64_MIN ||
5164 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
5165 reg->smin_value + off < 0)) {
5166 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5167 regno);
5168 return -EACCES;
5169 }
5170 err = __check_mem_access(env, regno, reg->smin_value + off, size,
5171 mem_size, zero_size_allowed);
5172 if (err) {
5173 verbose(env, "R%d min value is outside of the allowed memory range\n",
5174 regno);
5175 return err;
5176 }
5177
5178 /* If we haven't set a max value then we need to bail since we can't be
5179 * sure we won't do bad things.
5180 * If reg->umax_value + off could overflow, treat that as unbounded too.
5181 */
5182 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
5183 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
5184 regno);
5185 return -EACCES;
5186 }
5187 err = __check_mem_access(env, regno, reg->umax_value + off, size,
5188 mem_size, zero_size_allowed);
5189 if (err) {
5190 verbose(env, "R%d max value is outside of the allowed memory range\n",
5191 regno);
5192 return err;
5193 }
5194
5195 return 0;
5196 }
5197
__check_ptr_off_reg(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,bool fixed_off_ok)5198 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
5199 const struct bpf_reg_state *reg, int regno,
5200 bool fixed_off_ok)
5201 {
5202 /* Access to this pointer-typed register or passing it to a helper
5203 * is only allowed in its original, unmodified form.
5204 */
5205
5206 if (reg->off < 0) {
5207 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
5208 reg_type_str(env, reg->type), regno, reg->off);
5209 return -EACCES;
5210 }
5211
5212 if (!fixed_off_ok && reg->off) {
5213 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
5214 reg_type_str(env, reg->type), regno, reg->off);
5215 return -EACCES;
5216 }
5217
5218 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5219 char tn_buf[48];
5220
5221 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5222 verbose(env, "variable %s access var_off=%s disallowed\n",
5223 reg_type_str(env, reg->type), tn_buf);
5224 return -EACCES;
5225 }
5226
5227 return 0;
5228 }
5229
check_ptr_off_reg(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno)5230 int check_ptr_off_reg(struct bpf_verifier_env *env,
5231 const struct bpf_reg_state *reg, int regno)
5232 {
5233 return __check_ptr_off_reg(env, reg, regno, false);
5234 }
5235
map_kptr_match_type(struct bpf_verifier_env * env,struct btf_field * kptr_field,struct bpf_reg_state * reg,u32 regno)5236 static int map_kptr_match_type(struct bpf_verifier_env *env,
5237 struct btf_field *kptr_field,
5238 struct bpf_reg_state *reg, u32 regno)
5239 {
5240 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
5241 int perm_flags;
5242 const char *reg_name = "";
5243
5244 if (btf_is_kernel(reg->btf)) {
5245 perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
5246
5247 /* Only unreferenced case accepts untrusted pointers */
5248 if (kptr_field->type == BPF_KPTR_UNREF)
5249 perm_flags |= PTR_UNTRUSTED;
5250 } else {
5251 perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5252 }
5253
5254 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
5255 goto bad_type;
5256
5257 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5258 reg_name = btf_type_name(reg->btf, reg->btf_id);
5259
5260 /* For ref_ptr case, release function check should ensure we get one
5261 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5262 * normal store of unreferenced kptr, we must ensure var_off is zero.
5263 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5264 * reg->off and reg->ref_obj_id are not needed here.
5265 */
5266 if (__check_ptr_off_reg(env, reg, regno, true))
5267 return -EACCES;
5268
5269 /* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5270 * we also need to take into account the reg->off.
5271 *
5272 * We want to support cases like:
5273 *
5274 * struct foo {
5275 * struct bar br;
5276 * struct baz bz;
5277 * };
5278 *
5279 * struct foo *v;
5280 * v = func(); // PTR_TO_BTF_ID
5281 * val->foo = v; // reg->off is zero, btf and btf_id match type
5282 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5283 * // first member type of struct after comparison fails
5284 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5285 * // to match type
5286 *
5287 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5288 * is zero. We must also ensure that btf_struct_ids_match does not walk
5289 * the struct to match type against first member of struct, i.e. reject
5290 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5291 * strict mode to true for type match.
5292 */
5293 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5294 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5295 kptr_field->type == BPF_KPTR_REF))
5296 goto bad_type;
5297 return 0;
5298 bad_type:
5299 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5300 reg_type_str(env, reg->type), reg_name);
5301 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5302 if (kptr_field->type == BPF_KPTR_UNREF)
5303 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5304 targ_name);
5305 else
5306 verbose(env, "\n");
5307 return -EINVAL;
5308 }
5309
5310 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5311 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5312 */
in_rcu_cs(struct bpf_verifier_env * env)5313 static bool in_rcu_cs(struct bpf_verifier_env *env)
5314 {
5315 return env->cur_state->active_rcu_lock ||
5316 env->cur_state->active_lock.ptr ||
5317 !env->prog->aux->sleepable;
5318 }
5319
5320 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5321 BTF_SET_START(rcu_protected_types)
BTF_ID(struct,prog_test_ref_kfunc)5322 BTF_ID(struct, prog_test_ref_kfunc)
5323 BTF_ID(struct, cgroup)
5324 BTF_ID(struct, bpf_cpumask)
5325 BTF_ID(struct, task_struct)
5326 BTF_SET_END(rcu_protected_types)
5327
5328 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5329 {
5330 if (!btf_is_kernel(btf))
5331 return false;
5332 return btf_id_set_contains(&rcu_protected_types, btf_id);
5333 }
5334
rcu_safe_kptr(const struct btf_field * field)5335 static bool rcu_safe_kptr(const struct btf_field *field)
5336 {
5337 const struct btf_field_kptr *kptr = &field->kptr;
5338
5339 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5340 }
5341
check_map_kptr_access(struct bpf_verifier_env * env,u32 regno,int value_regno,int insn_idx,struct btf_field * kptr_field)5342 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5343 int value_regno, int insn_idx,
5344 struct btf_field *kptr_field)
5345 {
5346 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5347 int class = BPF_CLASS(insn->code);
5348 struct bpf_reg_state *val_reg;
5349
5350 /* Things we already checked for in check_map_access and caller:
5351 * - Reject cases where variable offset may touch kptr
5352 * - size of access (must be BPF_DW)
5353 * - tnum_is_const(reg->var_off)
5354 * - kptr_field->offset == off + reg->var_off.value
5355 */
5356 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5357 if (BPF_MODE(insn->code) != BPF_MEM) {
5358 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5359 return -EACCES;
5360 }
5361
5362 /* We only allow loading referenced kptr, since it will be marked as
5363 * untrusted, similar to unreferenced kptr.
5364 */
5365 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5366 verbose(env, "store to referenced kptr disallowed\n");
5367 return -EACCES;
5368 }
5369
5370 if (class == BPF_LDX) {
5371 val_reg = reg_state(env, value_regno);
5372 /* We can simply mark the value_regno receiving the pointer
5373 * value from map as PTR_TO_BTF_ID, with the correct type.
5374 */
5375 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5376 kptr_field->kptr.btf_id,
5377 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5378 PTR_MAYBE_NULL | MEM_RCU :
5379 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5380 } else if (class == BPF_STX) {
5381 val_reg = reg_state(env, value_regno);
5382 if (!register_is_null(val_reg) &&
5383 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5384 return -EACCES;
5385 } else if (class == BPF_ST) {
5386 if (insn->imm) {
5387 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5388 kptr_field->offset);
5389 return -EACCES;
5390 }
5391 } else {
5392 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5393 return -EACCES;
5394 }
5395 return 0;
5396 }
5397
5398 /* check read/write into a map element with possible variable offset */
check_map_access(struct bpf_verifier_env * env,u32 regno,int off,int size,bool zero_size_allowed,enum bpf_access_src src)5399 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5400 int off, int size, bool zero_size_allowed,
5401 enum bpf_access_src src)
5402 {
5403 struct bpf_verifier_state *vstate = env->cur_state;
5404 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5405 struct bpf_reg_state *reg = &state->regs[regno];
5406 struct bpf_map *map = reg->map_ptr;
5407 struct btf_record *rec;
5408 int err, i;
5409
5410 err = check_mem_region_access(env, regno, off, size, map->value_size,
5411 zero_size_allowed);
5412 if (err)
5413 return err;
5414
5415 if (IS_ERR_OR_NULL(map->record))
5416 return 0;
5417 rec = map->record;
5418 for (i = 0; i < rec->cnt; i++) {
5419 struct btf_field *field = &rec->fields[i];
5420 u32 p = field->offset;
5421
5422 /* If any part of a field can be touched by load/store, reject
5423 * this program. To check that [x1, x2) overlaps with [y1, y2),
5424 * it is sufficient to check x1 < y2 && y1 < x2.
5425 */
5426 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5427 p < reg->umax_value + off + size) {
5428 switch (field->type) {
5429 case BPF_KPTR_UNREF:
5430 case BPF_KPTR_REF:
5431 if (src != ACCESS_DIRECT) {
5432 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5433 return -EACCES;
5434 }
5435 if (!tnum_is_const(reg->var_off)) {
5436 verbose(env, "kptr access cannot have variable offset\n");
5437 return -EACCES;
5438 }
5439 if (p != off + reg->var_off.value) {
5440 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5441 p, off + reg->var_off.value);
5442 return -EACCES;
5443 }
5444 if (size != bpf_size_to_bytes(BPF_DW)) {
5445 verbose(env, "kptr access size must be BPF_DW\n");
5446 return -EACCES;
5447 }
5448 break;
5449 default:
5450 verbose(env, "%s cannot be accessed directly by load/store\n",
5451 btf_field_type_name(field->type));
5452 return -EACCES;
5453 }
5454 }
5455 }
5456 return 0;
5457 }
5458
5459 #define MAX_PACKET_OFF 0xffff
5460
may_access_direct_pkt_data(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_access_type t)5461 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5462 const struct bpf_call_arg_meta *meta,
5463 enum bpf_access_type t)
5464 {
5465 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5466
5467 switch (prog_type) {
5468 /* Program types only with direct read access go here! */
5469 case BPF_PROG_TYPE_LWT_IN:
5470 case BPF_PROG_TYPE_LWT_OUT:
5471 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5472 case BPF_PROG_TYPE_SK_REUSEPORT:
5473 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5474 case BPF_PROG_TYPE_CGROUP_SKB:
5475 if (t == BPF_WRITE)
5476 return false;
5477 fallthrough;
5478
5479 /* Program types with direct read + write access go here! */
5480 case BPF_PROG_TYPE_SCHED_CLS:
5481 case BPF_PROG_TYPE_SCHED_ACT:
5482 case BPF_PROG_TYPE_XDP:
5483 case BPF_PROG_TYPE_LWT_XMIT:
5484 case BPF_PROG_TYPE_SK_SKB:
5485 case BPF_PROG_TYPE_SK_MSG:
5486 if (meta)
5487 return meta->pkt_access;
5488
5489 env->seen_direct_write = true;
5490 return true;
5491
5492 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5493 if (t == BPF_WRITE)
5494 env->seen_direct_write = true;
5495
5496 return true;
5497
5498 default:
5499 return false;
5500 }
5501 }
5502
check_packet_access(struct bpf_verifier_env * env,u32 regno,int off,int size,bool zero_size_allowed)5503 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5504 int size, bool zero_size_allowed)
5505 {
5506 struct bpf_reg_state *regs = cur_regs(env);
5507 struct bpf_reg_state *reg = ®s[regno];
5508 int err;
5509
5510 /* We may have added a variable offset to the packet pointer; but any
5511 * reg->range we have comes after that. We are only checking the fixed
5512 * offset.
5513 */
5514
5515 /* We don't allow negative numbers, because we aren't tracking enough
5516 * detail to prove they're safe.
5517 */
5518 if (reg->smin_value < 0) {
5519 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5520 regno);
5521 return -EACCES;
5522 }
5523
5524 err = reg->range < 0 ? -EINVAL :
5525 __check_mem_access(env, regno, off, size, reg->range,
5526 zero_size_allowed);
5527 if (err) {
5528 verbose(env, "R%d offset is outside of the packet\n", regno);
5529 return err;
5530 }
5531
5532 /* __check_mem_access has made sure "off + size - 1" is within u16.
5533 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5534 * otherwise find_good_pkt_pointers would have refused to set range info
5535 * that __check_mem_access would have rejected this pkt access.
5536 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5537 */
5538 env->prog->aux->max_pkt_offset =
5539 max_t(u32, env->prog->aux->max_pkt_offset,
5540 off + reg->umax_value + size - 1);
5541
5542 return err;
5543 }
5544
5545 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
check_ctx_access(struct bpf_verifier_env * env,int insn_idx,int off,int size,enum bpf_access_type t,enum bpf_reg_type * reg_type,struct btf ** btf,u32 * btf_id)5546 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5547 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5548 struct btf **btf, u32 *btf_id)
5549 {
5550 struct bpf_insn_access_aux info = {
5551 .reg_type = *reg_type,
5552 .log = &env->log,
5553 };
5554
5555 if (env->ops->is_valid_access &&
5556 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5557 /* A non zero info.ctx_field_size indicates that this field is a
5558 * candidate for later verifier transformation to load the whole
5559 * field and then apply a mask when accessed with a narrower
5560 * access than actual ctx access size. A zero info.ctx_field_size
5561 * will only allow for whole field access and rejects any other
5562 * type of narrower access.
5563 */
5564 *reg_type = info.reg_type;
5565
5566 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5567 *btf = info.btf;
5568 *btf_id = info.btf_id;
5569 } else {
5570 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5571 }
5572 /* remember the offset of last byte accessed in ctx */
5573 if (env->prog->aux->max_ctx_offset < off + size)
5574 env->prog->aux->max_ctx_offset = off + size;
5575 return 0;
5576 }
5577
5578 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5579 return -EACCES;
5580 }
5581
check_flow_keys_access(struct bpf_verifier_env * env,int off,int size)5582 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5583 int size)
5584 {
5585 if (size < 0 || off < 0 ||
5586 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5587 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5588 off, size);
5589 return -EACCES;
5590 }
5591 return 0;
5592 }
5593
check_sock_access(struct bpf_verifier_env * env,int insn_idx,u32 regno,int off,int size,enum bpf_access_type t)5594 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5595 u32 regno, int off, int size,
5596 enum bpf_access_type t)
5597 {
5598 struct bpf_reg_state *regs = cur_regs(env);
5599 struct bpf_reg_state *reg = ®s[regno];
5600 struct bpf_insn_access_aux info = {};
5601 bool valid;
5602
5603 if (reg->smin_value < 0) {
5604 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5605 regno);
5606 return -EACCES;
5607 }
5608
5609 switch (reg->type) {
5610 case PTR_TO_SOCK_COMMON:
5611 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5612 break;
5613 case PTR_TO_SOCKET:
5614 valid = bpf_sock_is_valid_access(off, size, t, &info);
5615 break;
5616 case PTR_TO_TCP_SOCK:
5617 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5618 break;
5619 case PTR_TO_XDP_SOCK:
5620 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5621 break;
5622 default:
5623 valid = false;
5624 }
5625
5626
5627 if (valid) {
5628 env->insn_aux_data[insn_idx].ctx_field_size =
5629 info.ctx_field_size;
5630 return 0;
5631 }
5632
5633 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5634 regno, reg_type_str(env, reg->type), off, size);
5635
5636 return -EACCES;
5637 }
5638
is_pointer_value(struct bpf_verifier_env * env,int regno)5639 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5640 {
5641 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5642 }
5643
is_ctx_reg(struct bpf_verifier_env * env,int regno)5644 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5645 {
5646 const struct bpf_reg_state *reg = reg_state(env, regno);
5647
5648 return reg->type == PTR_TO_CTX;
5649 }
5650
is_sk_reg(struct bpf_verifier_env * env,int regno)5651 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5652 {
5653 const struct bpf_reg_state *reg = reg_state(env, regno);
5654
5655 return type_is_sk_pointer(reg->type);
5656 }
5657
is_pkt_reg(struct bpf_verifier_env * env,int regno)5658 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5659 {
5660 const struct bpf_reg_state *reg = reg_state(env, regno);
5661
5662 return type_is_pkt_pointer(reg->type);
5663 }
5664
is_flow_key_reg(struct bpf_verifier_env * env,int regno)5665 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5666 {
5667 const struct bpf_reg_state *reg = reg_state(env, regno);
5668
5669 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5670 return reg->type == PTR_TO_FLOW_KEYS;
5671 }
5672
5673 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
5674 #ifdef CONFIG_NET
5675 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
5676 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5677 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
5678 #endif
5679 [CONST_PTR_TO_MAP] = btf_bpf_map_id,
5680 };
5681
is_trusted_reg(const struct bpf_reg_state * reg)5682 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5683 {
5684 /* A referenced register is always trusted. */
5685 if (reg->ref_obj_id)
5686 return true;
5687
5688 /* Types listed in the reg2btf_ids are always trusted */
5689 if (reg2btf_ids[base_type(reg->type)] &&
5690 !bpf_type_has_unsafe_modifiers(reg->type))
5691 return true;
5692
5693 /* If a register is not referenced, it is trusted if it has the
5694 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5695 * other type modifiers may be safe, but we elect to take an opt-in
5696 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5697 * not.
5698 *
5699 * Eventually, we should make PTR_TRUSTED the single source of truth
5700 * for whether a register is trusted.
5701 */
5702 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5703 !bpf_type_has_unsafe_modifiers(reg->type);
5704 }
5705
is_rcu_reg(const struct bpf_reg_state * reg)5706 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5707 {
5708 return reg->type & MEM_RCU;
5709 }
5710
clear_trusted_flags(enum bpf_type_flag * flag)5711 static void clear_trusted_flags(enum bpf_type_flag *flag)
5712 {
5713 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5714 }
5715
check_pkt_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict)5716 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5717 const struct bpf_reg_state *reg,
5718 int off, int size, bool strict)
5719 {
5720 struct tnum reg_off;
5721 int ip_align;
5722
5723 /* Byte size accesses are always allowed. */
5724 if (!strict || size == 1)
5725 return 0;
5726
5727 /* For platforms that do not have a Kconfig enabling
5728 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5729 * NET_IP_ALIGN is universally set to '2'. And on platforms
5730 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5731 * to this code only in strict mode where we want to emulate
5732 * the NET_IP_ALIGN==2 checking. Therefore use an
5733 * unconditional IP align value of '2'.
5734 */
5735 ip_align = 2;
5736
5737 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5738 if (!tnum_is_aligned(reg_off, size)) {
5739 char tn_buf[48];
5740
5741 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5742 verbose(env,
5743 "misaligned packet access off %d+%s+%d+%d size %d\n",
5744 ip_align, tn_buf, reg->off, off, size);
5745 return -EACCES;
5746 }
5747
5748 return 0;
5749 }
5750
check_generic_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,const char * pointer_desc,int off,int size,bool strict)5751 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5752 const struct bpf_reg_state *reg,
5753 const char *pointer_desc,
5754 int off, int size, bool strict)
5755 {
5756 struct tnum reg_off;
5757
5758 /* Byte size accesses are always allowed. */
5759 if (!strict || size == 1)
5760 return 0;
5761
5762 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5763 if (!tnum_is_aligned(reg_off, size)) {
5764 char tn_buf[48];
5765
5766 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5767 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5768 pointer_desc, tn_buf, reg->off, off, size);
5769 return -EACCES;
5770 }
5771
5772 return 0;
5773 }
5774
check_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict_alignment_once)5775 static int check_ptr_alignment(struct bpf_verifier_env *env,
5776 const struct bpf_reg_state *reg, int off,
5777 int size, bool strict_alignment_once)
5778 {
5779 bool strict = env->strict_alignment || strict_alignment_once;
5780 const char *pointer_desc = "";
5781
5782 switch (reg->type) {
5783 case PTR_TO_PACKET:
5784 case PTR_TO_PACKET_META:
5785 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5786 * right in front, treat it the very same way.
5787 */
5788 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5789 case PTR_TO_FLOW_KEYS:
5790 pointer_desc = "flow keys ";
5791 break;
5792 case PTR_TO_MAP_KEY:
5793 pointer_desc = "key ";
5794 break;
5795 case PTR_TO_MAP_VALUE:
5796 pointer_desc = "value ";
5797 break;
5798 case PTR_TO_CTX:
5799 pointer_desc = "context ";
5800 break;
5801 case PTR_TO_STACK:
5802 pointer_desc = "stack ";
5803 /* The stack spill tracking logic in check_stack_write_fixed_off()
5804 * and check_stack_read_fixed_off() relies on stack accesses being
5805 * aligned.
5806 */
5807 strict = true;
5808 break;
5809 case PTR_TO_SOCKET:
5810 pointer_desc = "sock ";
5811 break;
5812 case PTR_TO_SOCK_COMMON:
5813 pointer_desc = "sock_common ";
5814 break;
5815 case PTR_TO_TCP_SOCK:
5816 pointer_desc = "tcp_sock ";
5817 break;
5818 case PTR_TO_XDP_SOCK:
5819 pointer_desc = "xdp_sock ";
5820 break;
5821 default:
5822 break;
5823 }
5824 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5825 strict);
5826 }
5827
5828 /* starting from main bpf function walk all instructions of the function
5829 * and recursively walk all callees that given function can call.
5830 * Ignore jump and exit insns.
5831 * Since recursion is prevented by check_cfg() this algorithm
5832 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5833 */
check_max_stack_depth_subprog(struct bpf_verifier_env * env,int idx)5834 static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx)
5835 {
5836 struct bpf_subprog_info *subprog = env->subprog_info;
5837 struct bpf_insn *insn = env->prog->insnsi;
5838 int depth = 0, frame = 0, i, subprog_end;
5839 bool tail_call_reachable = false;
5840 int ret_insn[MAX_CALL_FRAMES];
5841 int ret_prog[MAX_CALL_FRAMES];
5842 int j;
5843
5844 i = subprog[idx].start;
5845 process_func:
5846 /* protect against potential stack overflow that might happen when
5847 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5848 * depth for such case down to 256 so that the worst case scenario
5849 * would result in 8k stack size (32 which is tailcall limit * 256 =
5850 * 8k).
5851 *
5852 * To get the idea what might happen, see an example:
5853 * func1 -> sub rsp, 128
5854 * subfunc1 -> sub rsp, 256
5855 * tailcall1 -> add rsp, 256
5856 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5857 * subfunc2 -> sub rsp, 64
5858 * subfunc22 -> sub rsp, 128
5859 * tailcall2 -> add rsp, 128
5860 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5861 *
5862 * tailcall will unwind the current stack frame but it will not get rid
5863 * of caller's stack as shown on the example above.
5864 */
5865 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5866 verbose(env,
5867 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5868 depth);
5869 return -EACCES;
5870 }
5871 /* round up to 32-bytes, since this is granularity
5872 * of interpreter stack size
5873 */
5874 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5875 if (depth > MAX_BPF_STACK) {
5876 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5877 frame + 1, depth);
5878 return -EACCES;
5879 }
5880 continue_func:
5881 subprog_end = subprog[idx + 1].start;
5882 for (; i < subprog_end; i++) {
5883 int next_insn, sidx;
5884
5885 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5886 continue;
5887 /* remember insn and function to return to */
5888 ret_insn[frame] = i + 1;
5889 ret_prog[frame] = idx;
5890
5891 /* find the callee */
5892 next_insn = i + insn[i].imm + 1;
5893 sidx = find_subprog(env, next_insn);
5894 if (sidx < 0) {
5895 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5896 next_insn);
5897 return -EFAULT;
5898 }
5899 if (subprog[sidx].is_async_cb) {
5900 if (subprog[sidx].has_tail_call) {
5901 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5902 return -EFAULT;
5903 }
5904 /* async callbacks don't increase bpf prog stack size unless called directly */
5905 if (!bpf_pseudo_call(insn + i))
5906 continue;
5907 }
5908 i = next_insn;
5909 idx = sidx;
5910
5911 if (subprog[idx].has_tail_call)
5912 tail_call_reachable = true;
5913
5914 frame++;
5915 if (frame >= MAX_CALL_FRAMES) {
5916 verbose(env, "the call stack of %d frames is too deep !\n",
5917 frame);
5918 return -E2BIG;
5919 }
5920 goto process_func;
5921 }
5922 /* if tail call got detected across bpf2bpf calls then mark each of the
5923 * currently present subprog frames as tail call reachable subprogs;
5924 * this info will be utilized by JIT so that we will be preserving the
5925 * tail call counter throughout bpf2bpf calls combined with tailcalls
5926 */
5927 if (tail_call_reachable)
5928 for (j = 0; j < frame; j++)
5929 subprog[ret_prog[j]].tail_call_reachable = true;
5930 if (subprog[0].tail_call_reachable)
5931 env->prog->aux->tail_call_reachable = true;
5932
5933 /* end of for() loop means the last insn of the 'subprog'
5934 * was reached. Doesn't matter whether it was JA or EXIT
5935 */
5936 if (frame == 0)
5937 return 0;
5938 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5939 frame--;
5940 i = ret_insn[frame];
5941 idx = ret_prog[frame];
5942 goto continue_func;
5943 }
5944
check_max_stack_depth(struct bpf_verifier_env * env)5945 static int check_max_stack_depth(struct bpf_verifier_env *env)
5946 {
5947 struct bpf_subprog_info *si = env->subprog_info;
5948 int ret;
5949
5950 for (int i = 0; i < env->subprog_cnt; i++) {
5951 if (!i || si[i].is_async_cb) {
5952 ret = check_max_stack_depth_subprog(env, i);
5953 if (ret < 0)
5954 return ret;
5955 }
5956 continue;
5957 }
5958 return 0;
5959 }
5960
5961 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
get_callee_stack_depth(struct bpf_verifier_env * env,const struct bpf_insn * insn,int idx)5962 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5963 const struct bpf_insn *insn, int idx)
5964 {
5965 int start = idx + insn->imm + 1, subprog;
5966
5967 subprog = find_subprog(env, start);
5968 if (subprog < 0) {
5969 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5970 start);
5971 return -EFAULT;
5972 }
5973 return env->subprog_info[subprog].stack_depth;
5974 }
5975 #endif
5976
__check_buffer_access(struct bpf_verifier_env * env,const char * buf_info,const struct bpf_reg_state * reg,int regno,int off,int size)5977 static int __check_buffer_access(struct bpf_verifier_env *env,
5978 const char *buf_info,
5979 const struct bpf_reg_state *reg,
5980 int regno, int off, int size)
5981 {
5982 if (off < 0) {
5983 verbose(env,
5984 "R%d invalid %s buffer access: off=%d, size=%d\n",
5985 regno, buf_info, off, size);
5986 return -EACCES;
5987 }
5988 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5989 char tn_buf[48];
5990
5991 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5992 verbose(env,
5993 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5994 regno, off, tn_buf);
5995 return -EACCES;
5996 }
5997
5998 return 0;
5999 }
6000
check_tp_buffer_access(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,int off,int size)6001 static int check_tp_buffer_access(struct bpf_verifier_env *env,
6002 const struct bpf_reg_state *reg,
6003 int regno, int off, int size)
6004 {
6005 int err;
6006
6007 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
6008 if (err)
6009 return err;
6010
6011 if (off + size > env->prog->aux->max_tp_access)
6012 env->prog->aux->max_tp_access = off + size;
6013
6014 return 0;
6015 }
6016
check_buffer_access(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,int off,int size,bool zero_size_allowed,u32 * max_access)6017 static int check_buffer_access(struct bpf_verifier_env *env,
6018 const struct bpf_reg_state *reg,
6019 int regno, int off, int size,
6020 bool zero_size_allowed,
6021 u32 *max_access)
6022 {
6023 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
6024 int err;
6025
6026 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
6027 if (err)
6028 return err;
6029
6030 if (off + size > *max_access)
6031 *max_access = off + size;
6032
6033 return 0;
6034 }
6035
6036 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
zext_32_to_64(struct bpf_reg_state * reg)6037 static void zext_32_to_64(struct bpf_reg_state *reg)
6038 {
6039 reg->var_off = tnum_subreg(reg->var_off);
6040 __reg_assign_32_into_64(reg);
6041 }
6042
6043 /* truncate register to smaller size (in bytes)
6044 * must be called with size < BPF_REG_SIZE
6045 */
coerce_reg_to_size(struct bpf_reg_state * reg,int size)6046 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
6047 {
6048 u64 mask;
6049
6050 /* clear high bits in bit representation */
6051 reg->var_off = tnum_cast(reg->var_off, size);
6052
6053 /* fix arithmetic bounds */
6054 mask = ((u64)1 << (size * 8)) - 1;
6055 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
6056 reg->umin_value &= mask;
6057 reg->umax_value &= mask;
6058 } else {
6059 reg->umin_value = 0;
6060 reg->umax_value = mask;
6061 }
6062 reg->smin_value = reg->umin_value;
6063 reg->smax_value = reg->umax_value;
6064
6065 /* If size is smaller than 32bit register the 32bit register
6066 * values are also truncated so we push 64-bit bounds into
6067 * 32-bit bounds. Above were truncated < 32-bits already.
6068 */
6069 if (size >= 4)
6070 return;
6071 __reg_combine_64_into_32(reg);
6072 }
6073
set_sext64_default_val(struct bpf_reg_state * reg,int size)6074 static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
6075 {
6076 if (size == 1) {
6077 reg->smin_value = reg->s32_min_value = S8_MIN;
6078 reg->smax_value = reg->s32_max_value = S8_MAX;
6079 } else if (size == 2) {
6080 reg->smin_value = reg->s32_min_value = S16_MIN;
6081 reg->smax_value = reg->s32_max_value = S16_MAX;
6082 } else {
6083 /* size == 4 */
6084 reg->smin_value = reg->s32_min_value = S32_MIN;
6085 reg->smax_value = reg->s32_max_value = S32_MAX;
6086 }
6087 reg->umin_value = reg->u32_min_value = 0;
6088 reg->umax_value = U64_MAX;
6089 reg->u32_max_value = U32_MAX;
6090 reg->var_off = tnum_unknown;
6091 }
6092
coerce_reg_to_size_sx(struct bpf_reg_state * reg,int size)6093 static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
6094 {
6095 s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
6096 u64 top_smax_value, top_smin_value;
6097 u64 num_bits = size * 8;
6098
6099 if (tnum_is_const(reg->var_off)) {
6100 u64_cval = reg->var_off.value;
6101 if (size == 1)
6102 reg->var_off = tnum_const((s8)u64_cval);
6103 else if (size == 2)
6104 reg->var_off = tnum_const((s16)u64_cval);
6105 else
6106 /* size == 4 */
6107 reg->var_off = tnum_const((s32)u64_cval);
6108
6109 u64_cval = reg->var_off.value;
6110 reg->smax_value = reg->smin_value = u64_cval;
6111 reg->umax_value = reg->umin_value = u64_cval;
6112 reg->s32_max_value = reg->s32_min_value = u64_cval;
6113 reg->u32_max_value = reg->u32_min_value = u64_cval;
6114 return;
6115 }
6116
6117 top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
6118 top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
6119
6120 if (top_smax_value != top_smin_value)
6121 goto out;
6122
6123 /* find the s64_min and s64_min after sign extension */
6124 if (size == 1) {
6125 init_s64_max = (s8)reg->smax_value;
6126 init_s64_min = (s8)reg->smin_value;
6127 } else if (size == 2) {
6128 init_s64_max = (s16)reg->smax_value;
6129 init_s64_min = (s16)reg->smin_value;
6130 } else {
6131 init_s64_max = (s32)reg->smax_value;
6132 init_s64_min = (s32)reg->smin_value;
6133 }
6134
6135 s64_max = max(init_s64_max, init_s64_min);
6136 s64_min = min(init_s64_max, init_s64_min);
6137
6138 /* both of s64_max/s64_min positive or negative */
6139 if ((s64_max >= 0) == (s64_min >= 0)) {
6140 reg->smin_value = reg->s32_min_value = s64_min;
6141 reg->smax_value = reg->s32_max_value = s64_max;
6142 reg->umin_value = reg->u32_min_value = s64_min;
6143 reg->umax_value = reg->u32_max_value = s64_max;
6144 reg->var_off = tnum_range(s64_min, s64_max);
6145 return;
6146 }
6147
6148 out:
6149 set_sext64_default_val(reg, size);
6150 }
6151
set_sext32_default_val(struct bpf_reg_state * reg,int size)6152 static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
6153 {
6154 if (size == 1) {
6155 reg->s32_min_value = S8_MIN;
6156 reg->s32_max_value = S8_MAX;
6157 } else {
6158 /* size == 2 */
6159 reg->s32_min_value = S16_MIN;
6160 reg->s32_max_value = S16_MAX;
6161 }
6162 reg->u32_min_value = 0;
6163 reg->u32_max_value = U32_MAX;
6164 reg->var_off = tnum_subreg(tnum_unknown);
6165 }
6166
coerce_subreg_to_size_sx(struct bpf_reg_state * reg,int size)6167 static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
6168 {
6169 s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
6170 u32 top_smax_value, top_smin_value;
6171 u32 num_bits = size * 8;
6172
6173 if (tnum_is_const(reg->var_off)) {
6174 u32_val = reg->var_off.value;
6175 if (size == 1)
6176 reg->var_off = tnum_const((s8)u32_val);
6177 else
6178 reg->var_off = tnum_const((s16)u32_val);
6179
6180 u32_val = reg->var_off.value;
6181 reg->s32_min_value = reg->s32_max_value = u32_val;
6182 reg->u32_min_value = reg->u32_max_value = u32_val;
6183 return;
6184 }
6185
6186 top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
6187 top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
6188
6189 if (top_smax_value != top_smin_value)
6190 goto out;
6191
6192 /* find the s32_min and s32_min after sign extension */
6193 if (size == 1) {
6194 init_s32_max = (s8)reg->s32_max_value;
6195 init_s32_min = (s8)reg->s32_min_value;
6196 } else {
6197 /* size == 2 */
6198 init_s32_max = (s16)reg->s32_max_value;
6199 init_s32_min = (s16)reg->s32_min_value;
6200 }
6201 s32_max = max(init_s32_max, init_s32_min);
6202 s32_min = min(init_s32_max, init_s32_min);
6203
6204 if ((s32_min >= 0) == (s32_max >= 0)) {
6205 reg->s32_min_value = s32_min;
6206 reg->s32_max_value = s32_max;
6207 reg->u32_min_value = (u32)s32_min;
6208 reg->u32_max_value = (u32)s32_max;
6209 reg->var_off = tnum_subreg(tnum_range(s32_min, s32_max));
6210 return;
6211 }
6212
6213 out:
6214 set_sext32_default_val(reg, size);
6215 }
6216
bpf_map_is_rdonly(const struct bpf_map * map)6217 static bool bpf_map_is_rdonly(const struct bpf_map *map)
6218 {
6219 /* A map is considered read-only if the following condition are true:
6220 *
6221 * 1) BPF program side cannot change any of the map content. The
6222 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
6223 * and was set at map creation time.
6224 * 2) The map value(s) have been initialized from user space by a
6225 * loader and then "frozen", such that no new map update/delete
6226 * operations from syscall side are possible for the rest of
6227 * the map's lifetime from that point onwards.
6228 * 3) Any parallel/pending map update/delete operations from syscall
6229 * side have been completed. Only after that point, it's safe to
6230 * assume that map value(s) are immutable.
6231 */
6232 return (map->map_flags & BPF_F_RDONLY_PROG) &&
6233 READ_ONCE(map->frozen) &&
6234 !bpf_map_write_active(map);
6235 }
6236
bpf_map_direct_read(struct bpf_map * map,int off,int size,u64 * val,bool is_ldsx)6237 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
6238 bool is_ldsx)
6239 {
6240 void *ptr;
6241 u64 addr;
6242 int err;
6243
6244 err = map->ops->map_direct_value_addr(map, &addr, off);
6245 if (err)
6246 return err;
6247 ptr = (void *)(long)addr + off;
6248
6249 switch (size) {
6250 case sizeof(u8):
6251 *val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
6252 break;
6253 case sizeof(u16):
6254 *val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
6255 break;
6256 case sizeof(u32):
6257 *val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
6258 break;
6259 case sizeof(u64):
6260 *val = *(u64 *)ptr;
6261 break;
6262 default:
6263 return -EINVAL;
6264 }
6265 return 0;
6266 }
6267
6268 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
6269 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
6270 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
6271 #define BTF_TYPE_SAFE_TRUSTED_OR_NULL(__type) __PASTE(__type, __safe_trusted_or_null)
6272
6273 /*
6274 * Allow list few fields as RCU trusted or full trusted.
6275 * This logic doesn't allow mix tagging and will be removed once GCC supports
6276 * btf_type_tag.
6277 */
6278
6279 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
BTF_TYPE_SAFE_RCU(struct task_struct)6280 BTF_TYPE_SAFE_RCU(struct task_struct) {
6281 const cpumask_t *cpus_ptr;
6282 struct css_set __rcu *cgroups;
6283 struct task_struct __rcu *real_parent;
6284 struct task_struct *group_leader;
6285 };
6286
BTF_TYPE_SAFE_RCU(struct cgroup)6287 BTF_TYPE_SAFE_RCU(struct cgroup) {
6288 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
6289 struct kernfs_node *kn;
6290 };
6291
BTF_TYPE_SAFE_RCU(struct css_set)6292 BTF_TYPE_SAFE_RCU(struct css_set) {
6293 struct cgroup *dfl_cgrp;
6294 };
6295
6296 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct)6297 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
6298 struct file __rcu *exe_file;
6299 };
6300
6301 /* skb->sk, req->sk are not RCU protected, but we mark them as such
6302 * because bpf prog accessible sockets are SOCK_RCU_FREE.
6303 */
BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff)6304 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
6305 struct sock *sk;
6306 };
6307
BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock)6308 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
6309 struct sock *sk;
6310 };
6311
6312 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta)6313 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
6314 struct seq_file *seq;
6315 };
6316
BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task)6317 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
6318 struct bpf_iter_meta *meta;
6319 struct task_struct *task;
6320 };
6321
BTF_TYPE_SAFE_TRUSTED(struct linux_binprm)6322 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
6323 struct file *file;
6324 };
6325
BTF_TYPE_SAFE_TRUSTED(struct file)6326 BTF_TYPE_SAFE_TRUSTED(struct file) {
6327 struct inode *f_inode;
6328 };
6329
BTF_TYPE_SAFE_TRUSTED(struct dentry)6330 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
6331 /* no negative dentry-s in places where bpf can see it */
6332 struct inode *d_inode;
6333 };
6334
BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket)6335 BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket) {
6336 struct sock *sk;
6337 };
6338
type_is_rcu(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6339 static bool type_is_rcu(struct bpf_verifier_env *env,
6340 struct bpf_reg_state *reg,
6341 const char *field_name, u32 btf_id)
6342 {
6343 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
6344 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
6345 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
6346
6347 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
6348 }
6349
type_is_rcu_or_null(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6350 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
6351 struct bpf_reg_state *reg,
6352 const char *field_name, u32 btf_id)
6353 {
6354 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
6355 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
6356 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
6357
6358 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
6359 }
6360
type_is_trusted(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6361 static bool type_is_trusted(struct bpf_verifier_env *env,
6362 struct bpf_reg_state *reg,
6363 const char *field_name, u32 btf_id)
6364 {
6365 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
6366 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
6367 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
6368 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
6369 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
6370
6371 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
6372 }
6373
type_is_trusted_or_null(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6374 static bool type_is_trusted_or_null(struct bpf_verifier_env *env,
6375 struct bpf_reg_state *reg,
6376 const char *field_name, u32 btf_id)
6377 {
6378 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket));
6379
6380 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id,
6381 "__safe_trusted_or_null");
6382 }
6383
check_ptr_to_btf_access(struct bpf_verifier_env * env,struct bpf_reg_state * regs,int regno,int off,int size,enum bpf_access_type atype,int value_regno)6384 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
6385 struct bpf_reg_state *regs,
6386 int regno, int off, int size,
6387 enum bpf_access_type atype,
6388 int value_regno)
6389 {
6390 struct bpf_reg_state *reg = regs + regno;
6391 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
6392 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
6393 const char *field_name = NULL;
6394 enum bpf_type_flag flag = 0;
6395 u32 btf_id = 0;
6396 int ret;
6397
6398 if (!env->allow_ptr_leaks) {
6399 verbose(env,
6400 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6401 tname);
6402 return -EPERM;
6403 }
6404 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
6405 verbose(env,
6406 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
6407 tname);
6408 return -EINVAL;
6409 }
6410 if (off < 0) {
6411 verbose(env,
6412 "R%d is ptr_%s invalid negative access: off=%d\n",
6413 regno, tname, off);
6414 return -EACCES;
6415 }
6416 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6417 char tn_buf[48];
6418
6419 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6420 verbose(env,
6421 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
6422 regno, tname, off, tn_buf);
6423 return -EACCES;
6424 }
6425
6426 if (reg->type & MEM_USER) {
6427 verbose(env,
6428 "R%d is ptr_%s access user memory: off=%d\n",
6429 regno, tname, off);
6430 return -EACCES;
6431 }
6432
6433 if (reg->type & MEM_PERCPU) {
6434 verbose(env,
6435 "R%d is ptr_%s access percpu memory: off=%d\n",
6436 regno, tname, off);
6437 return -EACCES;
6438 }
6439
6440 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6441 if (!btf_is_kernel(reg->btf)) {
6442 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6443 return -EFAULT;
6444 }
6445 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6446 } else {
6447 /* Writes are permitted with default btf_struct_access for
6448 * program allocated objects (which always have ref_obj_id > 0),
6449 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6450 */
6451 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6452 verbose(env, "only read is supported\n");
6453 return -EACCES;
6454 }
6455
6456 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6457 !reg->ref_obj_id) {
6458 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6459 return -EFAULT;
6460 }
6461
6462 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6463 }
6464
6465 if (ret < 0)
6466 return ret;
6467
6468 if (ret != PTR_TO_BTF_ID) {
6469 /* just mark; */
6470
6471 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6472 /* If this is an untrusted pointer, all pointers formed by walking it
6473 * also inherit the untrusted flag.
6474 */
6475 flag = PTR_UNTRUSTED;
6476
6477 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6478 /* By default any pointer obtained from walking a trusted pointer is no
6479 * longer trusted, unless the field being accessed has explicitly been
6480 * marked as inheriting its parent's state of trust (either full or RCU).
6481 * For example:
6482 * 'cgroups' pointer is untrusted if task->cgroups dereference
6483 * happened in a sleepable program outside of bpf_rcu_read_lock()
6484 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6485 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6486 *
6487 * A regular RCU-protected pointer with __rcu tag can also be deemed
6488 * trusted if we are in an RCU CS. Such pointer can be NULL.
6489 */
6490 if (type_is_trusted(env, reg, field_name, btf_id)) {
6491 flag |= PTR_TRUSTED;
6492 } else if (type_is_trusted_or_null(env, reg, field_name, btf_id)) {
6493 flag |= PTR_TRUSTED | PTR_MAYBE_NULL;
6494 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6495 if (type_is_rcu(env, reg, field_name, btf_id)) {
6496 /* ignore __rcu tag and mark it MEM_RCU */
6497 flag |= MEM_RCU;
6498 } else if (flag & MEM_RCU ||
6499 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6500 /* __rcu tagged pointers can be NULL */
6501 flag |= MEM_RCU | PTR_MAYBE_NULL;
6502
6503 /* We always trust them */
6504 if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
6505 flag & PTR_UNTRUSTED)
6506 flag &= ~PTR_UNTRUSTED;
6507 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6508 /* keep as-is */
6509 } else {
6510 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6511 clear_trusted_flags(&flag);
6512 }
6513 } else {
6514 /*
6515 * If not in RCU CS or MEM_RCU pointer can be NULL then
6516 * aggressively mark as untrusted otherwise such
6517 * pointers will be plain PTR_TO_BTF_ID without flags
6518 * and will be allowed to be passed into helpers for
6519 * compat reasons.
6520 */
6521 flag = PTR_UNTRUSTED;
6522 }
6523 } else {
6524 /* Old compat. Deprecated */
6525 clear_trusted_flags(&flag);
6526 }
6527
6528 if (atype == BPF_READ && value_regno >= 0)
6529 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6530
6531 return 0;
6532 }
6533
check_ptr_to_map_access(struct bpf_verifier_env * env,struct bpf_reg_state * regs,int regno,int off,int size,enum bpf_access_type atype,int value_regno)6534 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6535 struct bpf_reg_state *regs,
6536 int regno, int off, int size,
6537 enum bpf_access_type atype,
6538 int value_regno)
6539 {
6540 struct bpf_reg_state *reg = regs + regno;
6541 struct bpf_map *map = reg->map_ptr;
6542 struct bpf_reg_state map_reg;
6543 enum bpf_type_flag flag = 0;
6544 const struct btf_type *t;
6545 const char *tname;
6546 u32 btf_id;
6547 int ret;
6548
6549 if (!btf_vmlinux) {
6550 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6551 return -ENOTSUPP;
6552 }
6553
6554 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6555 verbose(env, "map_ptr access not supported for map type %d\n",
6556 map->map_type);
6557 return -ENOTSUPP;
6558 }
6559
6560 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6561 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6562
6563 if (!env->allow_ptr_leaks) {
6564 verbose(env,
6565 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6566 tname);
6567 return -EPERM;
6568 }
6569
6570 if (off < 0) {
6571 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6572 regno, tname, off);
6573 return -EACCES;
6574 }
6575
6576 if (atype != BPF_READ) {
6577 verbose(env, "only read from %s is supported\n", tname);
6578 return -EACCES;
6579 }
6580
6581 /* Simulate access to a PTR_TO_BTF_ID */
6582 memset(&map_reg, 0, sizeof(map_reg));
6583 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6584 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6585 if (ret < 0)
6586 return ret;
6587
6588 if (value_regno >= 0)
6589 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6590
6591 return 0;
6592 }
6593
6594 /* Check that the stack access at the given offset is within bounds. The
6595 * maximum valid offset is -1.
6596 *
6597 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6598 * -state->allocated_stack for reads.
6599 */
check_stack_slot_within_bounds(struct bpf_verifier_env * env,s64 off,struct bpf_func_state * state,enum bpf_access_type t)6600 static int check_stack_slot_within_bounds(struct bpf_verifier_env *env,
6601 s64 off,
6602 struct bpf_func_state *state,
6603 enum bpf_access_type t)
6604 {
6605 int min_valid_off;
6606
6607 if (t == BPF_WRITE || env->allow_uninit_stack)
6608 min_valid_off = -MAX_BPF_STACK;
6609 else
6610 min_valid_off = -state->allocated_stack;
6611
6612 if (off < min_valid_off || off > -1)
6613 return -EACCES;
6614 return 0;
6615 }
6616
6617 /* Check that the stack access at 'regno + off' falls within the maximum stack
6618 * bounds.
6619 *
6620 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6621 */
check_stack_access_within_bounds(struct bpf_verifier_env * env,int regno,int off,int access_size,enum bpf_access_src src,enum bpf_access_type type)6622 static int check_stack_access_within_bounds(
6623 struct bpf_verifier_env *env,
6624 int regno, int off, int access_size,
6625 enum bpf_access_src src, enum bpf_access_type type)
6626 {
6627 struct bpf_reg_state *regs = cur_regs(env);
6628 struct bpf_reg_state *reg = regs + regno;
6629 struct bpf_func_state *state = func(env, reg);
6630 s64 min_off, max_off;
6631 int err;
6632 char *err_extra;
6633
6634 if (src == ACCESS_HELPER)
6635 /* We don't know if helpers are reading or writing (or both). */
6636 err_extra = " indirect access to";
6637 else if (type == BPF_READ)
6638 err_extra = " read from";
6639 else
6640 err_extra = " write to";
6641
6642 if (tnum_is_const(reg->var_off)) {
6643 min_off = (s64)reg->var_off.value + off;
6644 max_off = min_off + access_size;
6645 } else {
6646 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6647 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6648 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6649 err_extra, regno);
6650 return -EACCES;
6651 }
6652 min_off = reg->smin_value + off;
6653 max_off = reg->smax_value + off + access_size;
6654 }
6655
6656 err = check_stack_slot_within_bounds(env, min_off, state, type);
6657 if (!err && max_off > 0)
6658 err = -EINVAL; /* out of stack access into non-negative offsets */
6659 if (!err && access_size < 0)
6660 /* access_size should not be negative (or overflow an int); others checks
6661 * along the way should have prevented such an access.
6662 */
6663 err = -EFAULT; /* invalid negative access size; integer overflow? */
6664
6665 if (err) {
6666 if (tnum_is_const(reg->var_off)) {
6667 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6668 err_extra, regno, off, access_size);
6669 } else {
6670 char tn_buf[48];
6671
6672 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6673 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6674 err_extra, regno, tn_buf, access_size);
6675 }
6676 return err;
6677 }
6678
6679 return grow_stack_state(env, state, round_up(-min_off, BPF_REG_SIZE));
6680 }
6681
6682 /* check whether memory at (regno + off) is accessible for t = (read | write)
6683 * if t==write, value_regno is a register which value is stored into memory
6684 * if t==read, value_regno is a register which will receive the value from memory
6685 * if t==write && value_regno==-1, some unknown value is stored into memory
6686 * if t==read && value_regno==-1, don't care what we read from memory
6687 */
check_mem_access(struct bpf_verifier_env * env,int insn_idx,u32 regno,int off,int bpf_size,enum bpf_access_type t,int value_regno,bool strict_alignment_once,bool is_ldsx)6688 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6689 int off, int bpf_size, enum bpf_access_type t,
6690 int value_regno, bool strict_alignment_once, bool is_ldsx)
6691 {
6692 struct bpf_reg_state *regs = cur_regs(env);
6693 struct bpf_reg_state *reg = regs + regno;
6694 int size, err = 0;
6695
6696 size = bpf_size_to_bytes(bpf_size);
6697 if (size < 0)
6698 return size;
6699
6700 /* alignment checks will add in reg->off themselves */
6701 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6702 if (err)
6703 return err;
6704
6705 /* for access checks, reg->off is just part of off */
6706 off += reg->off;
6707
6708 if (reg->type == PTR_TO_MAP_KEY) {
6709 if (t == BPF_WRITE) {
6710 verbose(env, "write to change key R%d not allowed\n", regno);
6711 return -EACCES;
6712 }
6713
6714 err = check_mem_region_access(env, regno, off, size,
6715 reg->map_ptr->key_size, false);
6716 if (err)
6717 return err;
6718 if (value_regno >= 0)
6719 mark_reg_unknown(env, regs, value_regno);
6720 } else if (reg->type == PTR_TO_MAP_VALUE) {
6721 struct btf_field *kptr_field = NULL;
6722
6723 if (t == BPF_WRITE && value_regno >= 0 &&
6724 is_pointer_value(env, value_regno)) {
6725 verbose(env, "R%d leaks addr into map\n", value_regno);
6726 return -EACCES;
6727 }
6728 err = check_map_access_type(env, regno, off, size, t);
6729 if (err)
6730 return err;
6731 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6732 if (err)
6733 return err;
6734 if (tnum_is_const(reg->var_off))
6735 kptr_field = btf_record_find(reg->map_ptr->record,
6736 off + reg->var_off.value, BPF_KPTR);
6737 if (kptr_field) {
6738 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6739 } else if (t == BPF_READ && value_regno >= 0) {
6740 struct bpf_map *map = reg->map_ptr;
6741
6742 /* if map is read-only, track its contents as scalars */
6743 if (tnum_is_const(reg->var_off) &&
6744 bpf_map_is_rdonly(map) &&
6745 map->ops->map_direct_value_addr) {
6746 int map_off = off + reg->var_off.value;
6747 u64 val = 0;
6748
6749 err = bpf_map_direct_read(map, map_off, size,
6750 &val, is_ldsx);
6751 if (err)
6752 return err;
6753
6754 regs[value_regno].type = SCALAR_VALUE;
6755 __mark_reg_known(®s[value_regno], val);
6756 } else {
6757 mark_reg_unknown(env, regs, value_regno);
6758 }
6759 }
6760 } else if (base_type(reg->type) == PTR_TO_MEM) {
6761 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6762
6763 if (type_may_be_null(reg->type)) {
6764 verbose(env, "R%d invalid mem access '%s'\n", regno,
6765 reg_type_str(env, reg->type));
6766 return -EACCES;
6767 }
6768
6769 if (t == BPF_WRITE && rdonly_mem) {
6770 verbose(env, "R%d cannot write into %s\n",
6771 regno, reg_type_str(env, reg->type));
6772 return -EACCES;
6773 }
6774
6775 if (t == BPF_WRITE && value_regno >= 0 &&
6776 is_pointer_value(env, value_regno)) {
6777 verbose(env, "R%d leaks addr into mem\n", value_regno);
6778 return -EACCES;
6779 }
6780
6781 err = check_mem_region_access(env, regno, off, size,
6782 reg->mem_size, false);
6783 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6784 mark_reg_unknown(env, regs, value_regno);
6785 } else if (reg->type == PTR_TO_CTX) {
6786 enum bpf_reg_type reg_type = SCALAR_VALUE;
6787 struct btf *btf = NULL;
6788 u32 btf_id = 0;
6789
6790 if (t == BPF_WRITE && value_regno >= 0 &&
6791 is_pointer_value(env, value_regno)) {
6792 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6793 return -EACCES;
6794 }
6795
6796 err = check_ptr_off_reg(env, reg, regno);
6797 if (err < 0)
6798 return err;
6799
6800 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6801 &btf_id);
6802 if (err)
6803 verbose_linfo(env, insn_idx, "; ");
6804 if (!err && t == BPF_READ && value_regno >= 0) {
6805 /* ctx access returns either a scalar, or a
6806 * PTR_TO_PACKET[_META,_END]. In the latter
6807 * case, we know the offset is zero.
6808 */
6809 if (reg_type == SCALAR_VALUE) {
6810 mark_reg_unknown(env, regs, value_regno);
6811 } else {
6812 mark_reg_known_zero(env, regs,
6813 value_regno);
6814 if (type_may_be_null(reg_type))
6815 regs[value_regno].id = ++env->id_gen;
6816 /* A load of ctx field could have different
6817 * actual load size with the one encoded in the
6818 * insn. When the dst is PTR, it is for sure not
6819 * a sub-register.
6820 */
6821 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6822 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6823 regs[value_regno].btf = btf;
6824 regs[value_regno].btf_id = btf_id;
6825 }
6826 }
6827 regs[value_regno].type = reg_type;
6828 }
6829
6830 } else if (reg->type == PTR_TO_STACK) {
6831 /* Basic bounds checks. */
6832 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6833 if (err)
6834 return err;
6835
6836 if (t == BPF_READ)
6837 err = check_stack_read(env, regno, off, size,
6838 value_regno);
6839 else
6840 err = check_stack_write(env, regno, off, size,
6841 value_regno, insn_idx);
6842 } else if (reg_is_pkt_pointer(reg)) {
6843 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6844 verbose(env, "cannot write into packet\n");
6845 return -EACCES;
6846 }
6847 if (t == BPF_WRITE && value_regno >= 0 &&
6848 is_pointer_value(env, value_regno)) {
6849 verbose(env, "R%d leaks addr into packet\n",
6850 value_regno);
6851 return -EACCES;
6852 }
6853 err = check_packet_access(env, regno, off, size, false);
6854 if (!err && t == BPF_READ && value_regno >= 0)
6855 mark_reg_unknown(env, regs, value_regno);
6856 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6857 if (t == BPF_WRITE && value_regno >= 0 &&
6858 is_pointer_value(env, value_regno)) {
6859 verbose(env, "R%d leaks addr into flow keys\n",
6860 value_regno);
6861 return -EACCES;
6862 }
6863
6864 err = check_flow_keys_access(env, off, size);
6865 if (!err && t == BPF_READ && value_regno >= 0)
6866 mark_reg_unknown(env, regs, value_regno);
6867 } else if (type_is_sk_pointer(reg->type)) {
6868 if (t == BPF_WRITE) {
6869 verbose(env, "R%d cannot write into %s\n",
6870 regno, reg_type_str(env, reg->type));
6871 return -EACCES;
6872 }
6873 err = check_sock_access(env, insn_idx, regno, off, size, t);
6874 if (!err && value_regno >= 0)
6875 mark_reg_unknown(env, regs, value_regno);
6876 } else if (reg->type == PTR_TO_TP_BUFFER) {
6877 err = check_tp_buffer_access(env, reg, regno, off, size);
6878 if (!err && t == BPF_READ && value_regno >= 0)
6879 mark_reg_unknown(env, regs, value_regno);
6880 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6881 !type_may_be_null(reg->type)) {
6882 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6883 value_regno);
6884 } else if (reg->type == CONST_PTR_TO_MAP) {
6885 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6886 value_regno);
6887 } else if (base_type(reg->type) == PTR_TO_BUF) {
6888 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6889 u32 *max_access;
6890
6891 if (rdonly_mem) {
6892 if (t == BPF_WRITE) {
6893 verbose(env, "R%d cannot write into %s\n",
6894 regno, reg_type_str(env, reg->type));
6895 return -EACCES;
6896 }
6897 max_access = &env->prog->aux->max_rdonly_access;
6898 } else {
6899 max_access = &env->prog->aux->max_rdwr_access;
6900 }
6901
6902 err = check_buffer_access(env, reg, regno, off, size, false,
6903 max_access);
6904
6905 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6906 mark_reg_unknown(env, regs, value_regno);
6907 } else {
6908 verbose(env, "R%d invalid mem access '%s'\n", regno,
6909 reg_type_str(env, reg->type));
6910 return -EACCES;
6911 }
6912
6913 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6914 regs[value_regno].type == SCALAR_VALUE) {
6915 if (!is_ldsx)
6916 /* b/h/w load zero-extends, mark upper bits as known 0 */
6917 coerce_reg_to_size(®s[value_regno], size);
6918 else
6919 coerce_reg_to_size_sx(®s[value_regno], size);
6920 }
6921 return err;
6922 }
6923
check_atomic(struct bpf_verifier_env * env,int insn_idx,struct bpf_insn * insn)6924 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6925 {
6926 int load_reg;
6927 int err;
6928
6929 switch (insn->imm) {
6930 case BPF_ADD:
6931 case BPF_ADD | BPF_FETCH:
6932 case BPF_AND:
6933 case BPF_AND | BPF_FETCH:
6934 case BPF_OR:
6935 case BPF_OR | BPF_FETCH:
6936 case BPF_XOR:
6937 case BPF_XOR | BPF_FETCH:
6938 case BPF_XCHG:
6939 case BPF_CMPXCHG:
6940 break;
6941 default:
6942 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6943 return -EINVAL;
6944 }
6945
6946 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6947 verbose(env, "invalid atomic operand size\n");
6948 return -EINVAL;
6949 }
6950
6951 /* check src1 operand */
6952 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6953 if (err)
6954 return err;
6955
6956 /* check src2 operand */
6957 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6958 if (err)
6959 return err;
6960
6961 if (insn->imm == BPF_CMPXCHG) {
6962 /* Check comparison of R0 with memory location */
6963 const u32 aux_reg = BPF_REG_0;
6964
6965 err = check_reg_arg(env, aux_reg, SRC_OP);
6966 if (err)
6967 return err;
6968
6969 if (is_pointer_value(env, aux_reg)) {
6970 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6971 return -EACCES;
6972 }
6973 }
6974
6975 if (is_pointer_value(env, insn->src_reg)) {
6976 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6977 return -EACCES;
6978 }
6979
6980 if (is_ctx_reg(env, insn->dst_reg) ||
6981 is_pkt_reg(env, insn->dst_reg) ||
6982 is_flow_key_reg(env, insn->dst_reg) ||
6983 is_sk_reg(env, insn->dst_reg)) {
6984 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6985 insn->dst_reg,
6986 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6987 return -EACCES;
6988 }
6989
6990 if (insn->imm & BPF_FETCH) {
6991 if (insn->imm == BPF_CMPXCHG)
6992 load_reg = BPF_REG_0;
6993 else
6994 load_reg = insn->src_reg;
6995
6996 /* check and record load of old value */
6997 err = check_reg_arg(env, load_reg, DST_OP);
6998 if (err)
6999 return err;
7000 } else {
7001 /* This instruction accesses a memory location but doesn't
7002 * actually load it into a register.
7003 */
7004 load_reg = -1;
7005 }
7006
7007 /* Check whether we can read the memory, with second call for fetch
7008 * case to simulate the register fill.
7009 */
7010 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
7011 BPF_SIZE(insn->code), BPF_READ, -1, true, false);
7012 if (!err && load_reg >= 0)
7013 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
7014 BPF_SIZE(insn->code), BPF_READ, load_reg,
7015 true, false);
7016 if (err)
7017 return err;
7018
7019 /* Check whether we can write into the same memory. */
7020 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
7021 BPF_SIZE(insn->code), BPF_WRITE, -1, true, false);
7022 if (err)
7023 return err;
7024
7025 return 0;
7026 }
7027
7028 /* When register 'regno' is used to read the stack (either directly or through
7029 * a helper function) make sure that it's within stack boundary and, depending
7030 * on the access type and privileges, that all elements of the stack are
7031 * initialized.
7032 *
7033 * 'off' includes 'regno->off', but not its dynamic part (if any).
7034 *
7035 * All registers that have been spilled on the stack in the slots within the
7036 * read offsets are marked as read.
7037 */
check_stack_range_initialized(struct bpf_verifier_env * env,int regno,int off,int access_size,bool zero_size_allowed,enum bpf_access_src type,struct bpf_call_arg_meta * meta)7038 static int check_stack_range_initialized(
7039 struct bpf_verifier_env *env, int regno, int off,
7040 int access_size, bool zero_size_allowed,
7041 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
7042 {
7043 struct bpf_reg_state *reg = reg_state(env, regno);
7044 struct bpf_func_state *state = func(env, reg);
7045 int err, min_off, max_off, i, j, slot, spi;
7046 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
7047 enum bpf_access_type bounds_check_type;
7048 /* Some accesses can write anything into the stack, others are
7049 * read-only.
7050 */
7051 bool clobber = false;
7052
7053 if (access_size == 0 && !zero_size_allowed) {
7054 verbose(env, "invalid zero-sized read\n");
7055 return -EACCES;
7056 }
7057
7058 if (type == ACCESS_HELPER) {
7059 /* The bounds checks for writes are more permissive than for
7060 * reads. However, if raw_mode is not set, we'll do extra
7061 * checks below.
7062 */
7063 bounds_check_type = BPF_WRITE;
7064 clobber = true;
7065 } else {
7066 bounds_check_type = BPF_READ;
7067 }
7068 err = check_stack_access_within_bounds(env, regno, off, access_size,
7069 type, bounds_check_type);
7070 if (err)
7071 return err;
7072
7073
7074 if (tnum_is_const(reg->var_off)) {
7075 min_off = max_off = reg->var_off.value + off;
7076 } else {
7077 /* Variable offset is prohibited for unprivileged mode for
7078 * simplicity since it requires corresponding support in
7079 * Spectre masking for stack ALU.
7080 * See also retrieve_ptr_limit().
7081 */
7082 if (!env->bypass_spec_v1) {
7083 char tn_buf[48];
7084
7085 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7086 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
7087 regno, err_extra, tn_buf);
7088 return -EACCES;
7089 }
7090 /* Only initialized buffer on stack is allowed to be accessed
7091 * with variable offset. With uninitialized buffer it's hard to
7092 * guarantee that whole memory is marked as initialized on
7093 * helper return since specific bounds are unknown what may
7094 * cause uninitialized stack leaking.
7095 */
7096 if (meta && meta->raw_mode)
7097 meta = NULL;
7098
7099 min_off = reg->smin_value + off;
7100 max_off = reg->smax_value + off;
7101 }
7102
7103 if (meta && meta->raw_mode) {
7104 /* Ensure we won't be overwriting dynptrs when simulating byte
7105 * by byte access in check_helper_call using meta.access_size.
7106 * This would be a problem if we have a helper in the future
7107 * which takes:
7108 *
7109 * helper(uninit_mem, len, dynptr)
7110 *
7111 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
7112 * may end up writing to dynptr itself when touching memory from
7113 * arg 1. This can be relaxed on a case by case basis for known
7114 * safe cases, but reject due to the possibilitiy of aliasing by
7115 * default.
7116 */
7117 for (i = min_off; i < max_off + access_size; i++) {
7118 int stack_off = -i - 1;
7119
7120 spi = __get_spi(i);
7121 /* raw_mode may write past allocated_stack */
7122 if (state->allocated_stack <= stack_off)
7123 continue;
7124 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
7125 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
7126 return -EACCES;
7127 }
7128 }
7129 meta->access_size = access_size;
7130 meta->regno = regno;
7131 return 0;
7132 }
7133
7134 for (i = min_off; i < max_off + access_size; i++) {
7135 u8 *stype;
7136
7137 slot = -i - 1;
7138 spi = slot / BPF_REG_SIZE;
7139 if (state->allocated_stack <= slot) {
7140 verbose(env, "verifier bug: allocated_stack too small");
7141 return -EFAULT;
7142 }
7143
7144 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
7145 if (*stype == STACK_MISC)
7146 goto mark;
7147 if ((*stype == STACK_ZERO) ||
7148 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
7149 if (clobber) {
7150 /* helper can write anything into the stack */
7151 *stype = STACK_MISC;
7152 }
7153 goto mark;
7154 }
7155
7156 if (is_spilled_reg(&state->stack[spi]) &&
7157 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
7158 env->allow_ptr_leaks)) {
7159 if (clobber) {
7160 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
7161 for (j = 0; j < BPF_REG_SIZE; j++)
7162 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
7163 }
7164 goto mark;
7165 }
7166
7167 if (tnum_is_const(reg->var_off)) {
7168 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
7169 err_extra, regno, min_off, i - min_off, access_size);
7170 } else {
7171 char tn_buf[48];
7172
7173 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7174 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
7175 err_extra, regno, tn_buf, i - min_off, access_size);
7176 }
7177 return -EACCES;
7178 mark:
7179 /* reading any byte out of 8-byte 'spill_slot' will cause
7180 * the whole slot to be marked as 'read'
7181 */
7182 mark_reg_read(env, &state->stack[spi].spilled_ptr,
7183 state->stack[spi].spilled_ptr.parent,
7184 REG_LIVE_READ64);
7185 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
7186 * be sure that whether stack slot is written to or not. Hence,
7187 * we must still conservatively propagate reads upwards even if
7188 * helper may write to the entire memory range.
7189 */
7190 }
7191 return 0;
7192 }
7193
check_helper_mem_access(struct bpf_verifier_env * env,int regno,int access_size,bool zero_size_allowed,struct bpf_call_arg_meta * meta)7194 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
7195 int access_size, bool zero_size_allowed,
7196 struct bpf_call_arg_meta *meta)
7197 {
7198 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7199 u32 *max_access;
7200
7201 switch (base_type(reg->type)) {
7202 case PTR_TO_PACKET:
7203 case PTR_TO_PACKET_META:
7204 return check_packet_access(env, regno, reg->off, access_size,
7205 zero_size_allowed);
7206 case PTR_TO_MAP_KEY:
7207 if (meta && meta->raw_mode) {
7208 verbose(env, "R%d cannot write into %s\n", regno,
7209 reg_type_str(env, reg->type));
7210 return -EACCES;
7211 }
7212 return check_mem_region_access(env, regno, reg->off, access_size,
7213 reg->map_ptr->key_size, false);
7214 case PTR_TO_MAP_VALUE:
7215 if (check_map_access_type(env, regno, reg->off, access_size,
7216 meta && meta->raw_mode ? BPF_WRITE :
7217 BPF_READ))
7218 return -EACCES;
7219 return check_map_access(env, regno, reg->off, access_size,
7220 zero_size_allowed, ACCESS_HELPER);
7221 case PTR_TO_MEM:
7222 if (type_is_rdonly_mem(reg->type)) {
7223 if (meta && meta->raw_mode) {
7224 verbose(env, "R%d cannot write into %s\n", regno,
7225 reg_type_str(env, reg->type));
7226 return -EACCES;
7227 }
7228 }
7229 return check_mem_region_access(env, regno, reg->off,
7230 access_size, reg->mem_size,
7231 zero_size_allowed);
7232 case PTR_TO_BUF:
7233 if (type_is_rdonly_mem(reg->type)) {
7234 if (meta && meta->raw_mode) {
7235 verbose(env, "R%d cannot write into %s\n", regno,
7236 reg_type_str(env, reg->type));
7237 return -EACCES;
7238 }
7239
7240 max_access = &env->prog->aux->max_rdonly_access;
7241 } else {
7242 max_access = &env->prog->aux->max_rdwr_access;
7243 }
7244 return check_buffer_access(env, reg, regno, reg->off,
7245 access_size, zero_size_allowed,
7246 max_access);
7247 case PTR_TO_STACK:
7248 return check_stack_range_initialized(
7249 env,
7250 regno, reg->off, access_size,
7251 zero_size_allowed, ACCESS_HELPER, meta);
7252 case PTR_TO_BTF_ID:
7253 return check_ptr_to_btf_access(env, regs, regno, reg->off,
7254 access_size, BPF_READ, -1);
7255 case PTR_TO_CTX:
7256 /* in case the function doesn't know how to access the context,
7257 * (because we are in a program of type SYSCALL for example), we
7258 * can not statically check its size.
7259 * Dynamically check it now.
7260 */
7261 if (!env->ops->convert_ctx_access) {
7262 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
7263 int offset = access_size - 1;
7264
7265 /* Allow zero-byte read from PTR_TO_CTX */
7266 if (access_size == 0)
7267 return zero_size_allowed ? 0 : -EACCES;
7268
7269 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
7270 atype, -1, false, false);
7271 }
7272
7273 fallthrough;
7274 default: /* scalar_value or invalid ptr */
7275 /* Allow zero-byte read from NULL, regardless of pointer type */
7276 if (zero_size_allowed && access_size == 0 &&
7277 register_is_null(reg))
7278 return 0;
7279
7280 verbose(env, "R%d type=%s ", regno,
7281 reg_type_str(env, reg->type));
7282 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
7283 return -EACCES;
7284 }
7285 }
7286
check_mem_size_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,bool zero_size_allowed,struct bpf_call_arg_meta * meta)7287 static int check_mem_size_reg(struct bpf_verifier_env *env,
7288 struct bpf_reg_state *reg, u32 regno,
7289 bool zero_size_allowed,
7290 struct bpf_call_arg_meta *meta)
7291 {
7292 int err;
7293
7294 /* This is used to refine r0 return value bounds for helpers
7295 * that enforce this value as an upper bound on return values.
7296 * See do_refine_retval_range() for helpers that can refine
7297 * the return value. C type of helper is u32 so we pull register
7298 * bound from umax_value however, if negative verifier errors
7299 * out. Only upper bounds can be learned because retval is an
7300 * int type and negative retvals are allowed.
7301 */
7302 meta->msize_max_value = reg->umax_value;
7303
7304 /* The register is SCALAR_VALUE; the access check
7305 * happens using its boundaries.
7306 */
7307 if (!tnum_is_const(reg->var_off))
7308 /* For unprivileged variable accesses, disable raw
7309 * mode so that the program is required to
7310 * initialize all the memory that the helper could
7311 * just partially fill up.
7312 */
7313 meta = NULL;
7314
7315 if (reg->smin_value < 0) {
7316 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
7317 regno);
7318 return -EACCES;
7319 }
7320
7321 if (reg->umin_value == 0) {
7322 err = check_helper_mem_access(env, regno - 1, 0,
7323 zero_size_allowed,
7324 meta);
7325 if (err)
7326 return err;
7327 }
7328
7329 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
7330 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
7331 regno);
7332 return -EACCES;
7333 }
7334 err = check_helper_mem_access(env, regno - 1,
7335 reg->umax_value,
7336 zero_size_allowed, meta);
7337 if (!err)
7338 err = mark_chain_precision(env, regno);
7339 return err;
7340 }
7341
check_mem_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,u32 mem_size)7342 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7343 u32 regno, u32 mem_size)
7344 {
7345 bool may_be_null = type_may_be_null(reg->type);
7346 struct bpf_reg_state saved_reg;
7347 struct bpf_call_arg_meta meta;
7348 int err;
7349
7350 if (register_is_null(reg))
7351 return 0;
7352
7353 memset(&meta, 0, sizeof(meta));
7354 /* Assuming that the register contains a value check if the memory
7355 * access is safe. Temporarily save and restore the register's state as
7356 * the conversion shouldn't be visible to a caller.
7357 */
7358 if (may_be_null) {
7359 saved_reg = *reg;
7360 mark_ptr_not_null_reg(reg);
7361 }
7362
7363 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
7364 /* Check access for BPF_WRITE */
7365 meta.raw_mode = true;
7366 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
7367
7368 if (may_be_null)
7369 *reg = saved_reg;
7370
7371 return err;
7372 }
7373
check_kfunc_mem_size_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno)7374 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7375 u32 regno)
7376 {
7377 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
7378 bool may_be_null = type_may_be_null(mem_reg->type);
7379 struct bpf_reg_state saved_reg;
7380 struct bpf_call_arg_meta meta;
7381 int err;
7382
7383 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
7384
7385 memset(&meta, 0, sizeof(meta));
7386
7387 if (may_be_null) {
7388 saved_reg = *mem_reg;
7389 mark_ptr_not_null_reg(mem_reg);
7390 }
7391
7392 err = check_mem_size_reg(env, reg, regno, true, &meta);
7393 /* Check access for BPF_WRITE */
7394 meta.raw_mode = true;
7395 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
7396
7397 if (may_be_null)
7398 *mem_reg = saved_reg;
7399 return err;
7400 }
7401
7402 /* Implementation details:
7403 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
7404 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
7405 * Two bpf_map_lookups (even with the same key) will have different reg->id.
7406 * Two separate bpf_obj_new will also have different reg->id.
7407 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
7408 * clears reg->id after value_or_null->value transition, since the verifier only
7409 * cares about the range of access to valid map value pointer and doesn't care
7410 * about actual address of the map element.
7411 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
7412 * reg->id > 0 after value_or_null->value transition. By doing so
7413 * two bpf_map_lookups will be considered two different pointers that
7414 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
7415 * returned from bpf_obj_new.
7416 * The verifier allows taking only one bpf_spin_lock at a time to avoid
7417 * dead-locks.
7418 * Since only one bpf_spin_lock is allowed the checks are simpler than
7419 * reg_is_refcounted() logic. The verifier needs to remember only
7420 * one spin_lock instead of array of acquired_refs.
7421 * cur_state->active_lock remembers which map value element or allocated
7422 * object got locked and clears it after bpf_spin_unlock.
7423 */
process_spin_lock(struct bpf_verifier_env * env,int regno,bool is_lock)7424 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
7425 bool is_lock)
7426 {
7427 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7428 struct bpf_verifier_state *cur = env->cur_state;
7429 bool is_const = tnum_is_const(reg->var_off);
7430 u64 val = reg->var_off.value;
7431 struct bpf_map *map = NULL;
7432 struct btf *btf = NULL;
7433 struct btf_record *rec;
7434
7435 if (!is_const) {
7436 verbose(env,
7437 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7438 regno);
7439 return -EINVAL;
7440 }
7441 if (reg->type == PTR_TO_MAP_VALUE) {
7442 map = reg->map_ptr;
7443 if (!map->btf) {
7444 verbose(env,
7445 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7446 map->name);
7447 return -EINVAL;
7448 }
7449 } else {
7450 btf = reg->btf;
7451 }
7452
7453 rec = reg_btf_record(reg);
7454 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7455 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7456 map ? map->name : "kptr");
7457 return -EINVAL;
7458 }
7459 if (rec->spin_lock_off != val + reg->off) {
7460 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7461 val + reg->off, rec->spin_lock_off);
7462 return -EINVAL;
7463 }
7464 if (is_lock) {
7465 if (cur->active_lock.ptr) {
7466 verbose(env,
7467 "Locking two bpf_spin_locks are not allowed\n");
7468 return -EINVAL;
7469 }
7470 if (map)
7471 cur->active_lock.ptr = map;
7472 else
7473 cur->active_lock.ptr = btf;
7474 cur->active_lock.id = reg->id;
7475 } else {
7476 void *ptr;
7477
7478 if (map)
7479 ptr = map;
7480 else
7481 ptr = btf;
7482
7483 if (!cur->active_lock.ptr) {
7484 verbose(env, "bpf_spin_unlock without taking a lock\n");
7485 return -EINVAL;
7486 }
7487 if (cur->active_lock.ptr != ptr ||
7488 cur->active_lock.id != reg->id) {
7489 verbose(env, "bpf_spin_unlock of different lock\n");
7490 return -EINVAL;
7491 }
7492
7493 invalidate_non_owning_refs(env);
7494
7495 cur->active_lock.ptr = NULL;
7496 cur->active_lock.id = 0;
7497 }
7498 return 0;
7499 }
7500
process_timer_func(struct bpf_verifier_env * env,int regno,struct bpf_call_arg_meta * meta)7501 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7502 struct bpf_call_arg_meta *meta)
7503 {
7504 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7505 bool is_const = tnum_is_const(reg->var_off);
7506 struct bpf_map *map = reg->map_ptr;
7507 u64 val = reg->var_off.value;
7508
7509 if (!is_const) {
7510 verbose(env,
7511 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7512 regno);
7513 return -EINVAL;
7514 }
7515 if (!map->btf) {
7516 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7517 map->name);
7518 return -EINVAL;
7519 }
7520 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7521 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7522 return -EINVAL;
7523 }
7524 if (map->record->timer_off != val + reg->off) {
7525 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7526 val + reg->off, map->record->timer_off);
7527 return -EINVAL;
7528 }
7529 if (meta->map_ptr) {
7530 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7531 return -EFAULT;
7532 }
7533 meta->map_uid = reg->map_uid;
7534 meta->map_ptr = map;
7535 return 0;
7536 }
7537
process_kptr_func(struct bpf_verifier_env * env,int regno,struct bpf_call_arg_meta * meta)7538 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7539 struct bpf_call_arg_meta *meta)
7540 {
7541 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7542 struct bpf_map *map_ptr = reg->map_ptr;
7543 struct btf_field *kptr_field;
7544 u32 kptr_off;
7545
7546 if (!tnum_is_const(reg->var_off)) {
7547 verbose(env,
7548 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7549 regno);
7550 return -EINVAL;
7551 }
7552 if (!map_ptr->btf) {
7553 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7554 map_ptr->name);
7555 return -EINVAL;
7556 }
7557 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7558 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7559 return -EINVAL;
7560 }
7561
7562 meta->map_ptr = map_ptr;
7563 kptr_off = reg->off + reg->var_off.value;
7564 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7565 if (!kptr_field) {
7566 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7567 return -EACCES;
7568 }
7569 if (kptr_field->type != BPF_KPTR_REF) {
7570 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7571 return -EACCES;
7572 }
7573 meta->kptr_field = kptr_field;
7574 return 0;
7575 }
7576
7577 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7578 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7579 *
7580 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7581 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7582 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7583 *
7584 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7585 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7586 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7587 * mutate the view of the dynptr and also possibly destroy it. In the latter
7588 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7589 * memory that dynptr points to.
7590 *
7591 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7592 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7593 * readonly dynptr view yet, hence only the first case is tracked and checked.
7594 *
7595 * This is consistent with how C applies the const modifier to a struct object,
7596 * where the pointer itself inside bpf_dynptr becomes const but not what it
7597 * points to.
7598 *
7599 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7600 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7601 */
process_dynptr_func(struct bpf_verifier_env * env,int regno,int insn_idx,enum bpf_arg_type arg_type,int clone_ref_obj_id)7602 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7603 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7604 {
7605 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7606 int err;
7607
7608 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7609 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7610 */
7611 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7612 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7613 return -EFAULT;
7614 }
7615
7616 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7617 * constructing a mutable bpf_dynptr object.
7618 *
7619 * Currently, this is only possible with PTR_TO_STACK
7620 * pointing to a region of at least 16 bytes which doesn't
7621 * contain an existing bpf_dynptr.
7622 *
7623 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7624 * mutated or destroyed. However, the memory it points to
7625 * may be mutated.
7626 *
7627 * None - Points to a initialized dynptr that can be mutated and
7628 * destroyed, including mutation of the memory it points
7629 * to.
7630 */
7631 if (arg_type & MEM_UNINIT) {
7632 int i;
7633
7634 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7635 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7636 return -EINVAL;
7637 }
7638
7639 /* we write BPF_DW bits (8 bytes) at a time */
7640 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7641 err = check_mem_access(env, insn_idx, regno,
7642 i, BPF_DW, BPF_WRITE, -1, false, false);
7643 if (err)
7644 return err;
7645 }
7646
7647 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7648 } else /* MEM_RDONLY and None case from above */ {
7649 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7650 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7651 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7652 return -EINVAL;
7653 }
7654
7655 if (!is_dynptr_reg_valid_init(env, reg)) {
7656 verbose(env,
7657 "Expected an initialized dynptr as arg #%d\n",
7658 regno);
7659 return -EINVAL;
7660 }
7661
7662 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7663 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7664 verbose(env,
7665 "Expected a dynptr of type %s as arg #%d\n",
7666 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7667 return -EINVAL;
7668 }
7669
7670 err = mark_dynptr_read(env, reg);
7671 }
7672 return err;
7673 }
7674
iter_ref_obj_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int spi)7675 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7676 {
7677 struct bpf_func_state *state = func(env, reg);
7678
7679 return state->stack[spi].spilled_ptr.ref_obj_id;
7680 }
7681
is_iter_kfunc(struct bpf_kfunc_call_arg_meta * meta)7682 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7683 {
7684 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7685 }
7686
is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta * meta)7687 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7688 {
7689 return meta->kfunc_flags & KF_ITER_NEW;
7690 }
7691
is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta * meta)7692 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7693 {
7694 return meta->kfunc_flags & KF_ITER_NEXT;
7695 }
7696
is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta * meta)7697 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7698 {
7699 return meta->kfunc_flags & KF_ITER_DESTROY;
7700 }
7701
is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta * meta,int arg)7702 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7703 {
7704 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7705 * kfunc is iter state pointer
7706 */
7707 return arg == 0 && is_iter_kfunc(meta);
7708 }
7709
process_iter_arg(struct bpf_verifier_env * env,int regno,int insn_idx,struct bpf_kfunc_call_arg_meta * meta)7710 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7711 struct bpf_kfunc_call_arg_meta *meta)
7712 {
7713 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7714 const struct btf_type *t;
7715 const struct btf_param *arg;
7716 int spi, err, i, nr_slots;
7717 u32 btf_id;
7718
7719 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7720 arg = &btf_params(meta->func_proto)[0];
7721 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7722 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7723 nr_slots = t->size / BPF_REG_SIZE;
7724
7725 if (is_iter_new_kfunc(meta)) {
7726 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7727 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7728 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7729 iter_type_str(meta->btf, btf_id), regno);
7730 return -EINVAL;
7731 }
7732
7733 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7734 err = check_mem_access(env, insn_idx, regno,
7735 i, BPF_DW, BPF_WRITE, -1, false, false);
7736 if (err)
7737 return err;
7738 }
7739
7740 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7741 if (err)
7742 return err;
7743 } else {
7744 /* iter_next() or iter_destroy() expect initialized iter state*/
7745 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7746 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7747 iter_type_str(meta->btf, btf_id), regno);
7748 return -EINVAL;
7749 }
7750
7751 spi = iter_get_spi(env, reg, nr_slots);
7752 if (spi < 0)
7753 return spi;
7754
7755 err = mark_iter_read(env, reg, spi, nr_slots);
7756 if (err)
7757 return err;
7758
7759 /* remember meta->iter info for process_iter_next_call() */
7760 meta->iter.spi = spi;
7761 meta->iter.frameno = reg->frameno;
7762 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7763
7764 if (is_iter_destroy_kfunc(meta)) {
7765 err = unmark_stack_slots_iter(env, reg, nr_slots);
7766 if (err)
7767 return err;
7768 }
7769 }
7770
7771 return 0;
7772 }
7773
7774 /* Look for a previous loop entry at insn_idx: nearest parent state
7775 * stopped at insn_idx with callsites matching those in cur->frame.
7776 */
find_prev_entry(struct bpf_verifier_env * env,struct bpf_verifier_state * cur,int insn_idx)7777 static struct bpf_verifier_state *find_prev_entry(struct bpf_verifier_env *env,
7778 struct bpf_verifier_state *cur,
7779 int insn_idx)
7780 {
7781 struct bpf_verifier_state_list *sl;
7782 struct bpf_verifier_state *st;
7783
7784 /* Explored states are pushed in stack order, most recent states come first */
7785 sl = *explored_state(env, insn_idx);
7786 for (; sl; sl = sl->next) {
7787 /* If st->branches != 0 state is a part of current DFS verification path,
7788 * hence cur & st for a loop.
7789 */
7790 st = &sl->state;
7791 if (st->insn_idx == insn_idx && st->branches && same_callsites(st, cur) &&
7792 st->dfs_depth < cur->dfs_depth)
7793 return st;
7794 }
7795
7796 return NULL;
7797 }
7798
7799 static void reset_idmap_scratch(struct bpf_verifier_env *env);
7800 static bool regs_exact(const struct bpf_reg_state *rold,
7801 const struct bpf_reg_state *rcur,
7802 struct bpf_idmap *idmap);
7803
maybe_widen_reg(struct bpf_verifier_env * env,struct bpf_reg_state * rold,struct bpf_reg_state * rcur,struct bpf_idmap * idmap)7804 static void maybe_widen_reg(struct bpf_verifier_env *env,
7805 struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
7806 struct bpf_idmap *idmap)
7807 {
7808 if (rold->type != SCALAR_VALUE)
7809 return;
7810 if (rold->type != rcur->type)
7811 return;
7812 if (rold->precise || rcur->precise || regs_exact(rold, rcur, idmap))
7813 return;
7814 __mark_reg_unknown(env, rcur);
7815 }
7816
widen_imprecise_scalars(struct bpf_verifier_env * env,struct bpf_verifier_state * old,struct bpf_verifier_state * cur)7817 static int widen_imprecise_scalars(struct bpf_verifier_env *env,
7818 struct bpf_verifier_state *old,
7819 struct bpf_verifier_state *cur)
7820 {
7821 struct bpf_func_state *fold, *fcur;
7822 int i, fr;
7823
7824 reset_idmap_scratch(env);
7825 for (fr = old->curframe; fr >= 0; fr--) {
7826 fold = old->frame[fr];
7827 fcur = cur->frame[fr];
7828
7829 for (i = 0; i < MAX_BPF_REG; i++)
7830 maybe_widen_reg(env,
7831 &fold->regs[i],
7832 &fcur->regs[i],
7833 &env->idmap_scratch);
7834
7835 for (i = 0; i < fold->allocated_stack / BPF_REG_SIZE; i++) {
7836 if (!is_spilled_reg(&fold->stack[i]) ||
7837 !is_spilled_reg(&fcur->stack[i]))
7838 continue;
7839
7840 maybe_widen_reg(env,
7841 &fold->stack[i].spilled_ptr,
7842 &fcur->stack[i].spilled_ptr,
7843 &env->idmap_scratch);
7844 }
7845 }
7846 return 0;
7847 }
7848
7849 /* process_iter_next_call() is called when verifier gets to iterator's next
7850 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7851 * to it as just "iter_next()" in comments below.
7852 *
7853 * BPF verifier relies on a crucial contract for any iter_next()
7854 * implementation: it should *eventually* return NULL, and once that happens
7855 * it should keep returning NULL. That is, once iterator exhausts elements to
7856 * iterate, it should never reset or spuriously return new elements.
7857 *
7858 * With the assumption of such contract, process_iter_next_call() simulates
7859 * a fork in the verifier state to validate loop logic correctness and safety
7860 * without having to simulate infinite amount of iterations.
7861 *
7862 * In current state, we first assume that iter_next() returned NULL and
7863 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7864 * conditions we should not form an infinite loop and should eventually reach
7865 * exit.
7866 *
7867 * Besides that, we also fork current state and enqueue it for later
7868 * verification. In a forked state we keep iterator state as ACTIVE
7869 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7870 * also bump iteration depth to prevent erroneous infinite loop detection
7871 * later on (see iter_active_depths_differ() comment for details). In this
7872 * state we assume that we'll eventually loop back to another iter_next()
7873 * calls (it could be in exactly same location or in some other instruction,
7874 * it doesn't matter, we don't make any unnecessary assumptions about this,
7875 * everything revolves around iterator state in a stack slot, not which
7876 * instruction is calling iter_next()). When that happens, we either will come
7877 * to iter_next() with equivalent state and can conclude that next iteration
7878 * will proceed in exactly the same way as we just verified, so it's safe to
7879 * assume that loop converges. If not, we'll go on another iteration
7880 * simulation with a different input state, until all possible starting states
7881 * are validated or we reach maximum number of instructions limit.
7882 *
7883 * This way, we will either exhaustively discover all possible input states
7884 * that iterator loop can start with and eventually will converge, or we'll
7885 * effectively regress into bounded loop simulation logic and either reach
7886 * maximum number of instructions if loop is not provably convergent, or there
7887 * is some statically known limit on number of iterations (e.g., if there is
7888 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7889 *
7890 * Iteration convergence logic in is_state_visited() relies on exact
7891 * states comparison, which ignores read and precision marks.
7892 * This is necessary because read and precision marks are not finalized
7893 * while in the loop. Exact comparison might preclude convergence for
7894 * simple programs like below:
7895 *
7896 * i = 0;
7897 * while(iter_next(&it))
7898 * i++;
7899 *
7900 * At each iteration step i++ would produce a new distinct state and
7901 * eventually instruction processing limit would be reached.
7902 *
7903 * To avoid such behavior speculatively forget (widen) range for
7904 * imprecise scalar registers, if those registers were not precise at the
7905 * end of the previous iteration and do not match exactly.
7906 *
7907 * This is a conservative heuristic that allows to verify wide range of programs,
7908 * however it precludes verification of programs that conjure an
7909 * imprecise value on the first loop iteration and use it as precise on a second.
7910 * For example, the following safe program would fail to verify:
7911 *
7912 * struct bpf_num_iter it;
7913 * int arr[10];
7914 * int i = 0, a = 0;
7915 * bpf_iter_num_new(&it, 0, 10);
7916 * while (bpf_iter_num_next(&it)) {
7917 * if (a == 0) {
7918 * a = 1;
7919 * i = 7; // Because i changed verifier would forget
7920 * // it's range on second loop entry.
7921 * } else {
7922 * arr[i] = 42; // This would fail to verify.
7923 * }
7924 * }
7925 * bpf_iter_num_destroy(&it);
7926 */
process_iter_next_call(struct bpf_verifier_env * env,int insn_idx,struct bpf_kfunc_call_arg_meta * meta)7927 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7928 struct bpf_kfunc_call_arg_meta *meta)
7929 {
7930 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st, *prev_st;
7931 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7932 struct bpf_reg_state *cur_iter, *queued_iter;
7933 int iter_frameno = meta->iter.frameno;
7934 int iter_spi = meta->iter.spi;
7935
7936 BTF_TYPE_EMIT(struct bpf_iter);
7937
7938 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7939
7940 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7941 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7942 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7943 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7944 return -EFAULT;
7945 }
7946
7947 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7948 /* Because iter_next() call is a checkpoint is_state_visitied()
7949 * should guarantee parent state with same call sites and insn_idx.
7950 */
7951 if (!cur_st->parent || cur_st->parent->insn_idx != insn_idx ||
7952 !same_callsites(cur_st->parent, cur_st)) {
7953 verbose(env, "bug: bad parent state for iter next call");
7954 return -EFAULT;
7955 }
7956 /* Note cur_st->parent in the call below, it is necessary to skip
7957 * checkpoint created for cur_st by is_state_visited()
7958 * right at this instruction.
7959 */
7960 prev_st = find_prev_entry(env, cur_st->parent, insn_idx);
7961 /* branch out active iter state */
7962 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7963 if (!queued_st)
7964 return -ENOMEM;
7965
7966 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7967 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7968 queued_iter->iter.depth++;
7969 if (prev_st)
7970 widen_imprecise_scalars(env, prev_st, queued_st);
7971
7972 queued_fr = queued_st->frame[queued_st->curframe];
7973 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7974 }
7975
7976 /* switch to DRAINED state, but keep the depth unchanged */
7977 /* mark current iter state as drained and assume returned NULL */
7978 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7979 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7980
7981 return 0;
7982 }
7983
arg_type_is_mem_size(enum bpf_arg_type type)7984 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7985 {
7986 return type == ARG_CONST_SIZE ||
7987 type == ARG_CONST_SIZE_OR_ZERO;
7988 }
7989
arg_type_is_release(enum bpf_arg_type type)7990 static bool arg_type_is_release(enum bpf_arg_type type)
7991 {
7992 return type & OBJ_RELEASE;
7993 }
7994
arg_type_is_dynptr(enum bpf_arg_type type)7995 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7996 {
7997 return base_type(type) == ARG_PTR_TO_DYNPTR;
7998 }
7999
int_ptr_type_to_size(enum bpf_arg_type type)8000 static int int_ptr_type_to_size(enum bpf_arg_type type)
8001 {
8002 if (type == ARG_PTR_TO_INT)
8003 return sizeof(u32);
8004 else if (type == ARG_PTR_TO_LONG)
8005 return sizeof(u64);
8006
8007 return -EINVAL;
8008 }
8009
resolve_map_arg_type(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_arg_type * arg_type)8010 static int resolve_map_arg_type(struct bpf_verifier_env *env,
8011 const struct bpf_call_arg_meta *meta,
8012 enum bpf_arg_type *arg_type)
8013 {
8014 if (!meta->map_ptr) {
8015 /* kernel subsystem misconfigured verifier */
8016 verbose(env, "invalid map_ptr to access map->type\n");
8017 return -EACCES;
8018 }
8019
8020 switch (meta->map_ptr->map_type) {
8021 case BPF_MAP_TYPE_SOCKMAP:
8022 case BPF_MAP_TYPE_SOCKHASH:
8023 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
8024 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
8025 } else {
8026 verbose(env, "invalid arg_type for sockmap/sockhash\n");
8027 return -EINVAL;
8028 }
8029 break;
8030 case BPF_MAP_TYPE_BLOOM_FILTER:
8031 if (meta->func_id == BPF_FUNC_map_peek_elem)
8032 *arg_type = ARG_PTR_TO_MAP_VALUE;
8033 break;
8034 default:
8035 break;
8036 }
8037 return 0;
8038 }
8039
8040 struct bpf_reg_types {
8041 const enum bpf_reg_type types[10];
8042 u32 *btf_id;
8043 };
8044
8045 static const struct bpf_reg_types sock_types = {
8046 .types = {
8047 PTR_TO_SOCK_COMMON,
8048 PTR_TO_SOCKET,
8049 PTR_TO_TCP_SOCK,
8050 PTR_TO_XDP_SOCK,
8051 },
8052 };
8053
8054 #ifdef CONFIG_NET
8055 static const struct bpf_reg_types btf_id_sock_common_types = {
8056 .types = {
8057 PTR_TO_SOCK_COMMON,
8058 PTR_TO_SOCKET,
8059 PTR_TO_TCP_SOCK,
8060 PTR_TO_XDP_SOCK,
8061 PTR_TO_BTF_ID,
8062 PTR_TO_BTF_ID | PTR_TRUSTED,
8063 },
8064 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
8065 };
8066 #endif
8067
8068 static const struct bpf_reg_types mem_types = {
8069 .types = {
8070 PTR_TO_STACK,
8071 PTR_TO_PACKET,
8072 PTR_TO_PACKET_META,
8073 PTR_TO_MAP_KEY,
8074 PTR_TO_MAP_VALUE,
8075 PTR_TO_MEM,
8076 PTR_TO_MEM | MEM_RINGBUF,
8077 PTR_TO_BUF,
8078 PTR_TO_BTF_ID | PTR_TRUSTED,
8079 },
8080 };
8081
8082 static const struct bpf_reg_types int_ptr_types = {
8083 .types = {
8084 PTR_TO_STACK,
8085 PTR_TO_PACKET,
8086 PTR_TO_PACKET_META,
8087 PTR_TO_MAP_KEY,
8088 PTR_TO_MAP_VALUE,
8089 },
8090 };
8091
8092 static const struct bpf_reg_types spin_lock_types = {
8093 .types = {
8094 PTR_TO_MAP_VALUE,
8095 PTR_TO_BTF_ID | MEM_ALLOC,
8096 }
8097 };
8098
8099 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
8100 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
8101 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
8102 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
8103 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
8104 static const struct bpf_reg_types btf_ptr_types = {
8105 .types = {
8106 PTR_TO_BTF_ID,
8107 PTR_TO_BTF_ID | PTR_TRUSTED,
8108 PTR_TO_BTF_ID | MEM_RCU,
8109 },
8110 };
8111 static const struct bpf_reg_types percpu_btf_ptr_types = {
8112 .types = {
8113 PTR_TO_BTF_ID | MEM_PERCPU,
8114 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
8115 }
8116 };
8117 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
8118 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
8119 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
8120 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
8121 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
8122 static const struct bpf_reg_types dynptr_types = {
8123 .types = {
8124 PTR_TO_STACK,
8125 CONST_PTR_TO_DYNPTR,
8126 }
8127 };
8128
8129 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
8130 [ARG_PTR_TO_MAP_KEY] = &mem_types,
8131 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
8132 [ARG_CONST_SIZE] = &scalar_types,
8133 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
8134 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
8135 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
8136 [ARG_PTR_TO_CTX] = &context_types,
8137 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
8138 #ifdef CONFIG_NET
8139 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
8140 #endif
8141 [ARG_PTR_TO_SOCKET] = &fullsock_types,
8142 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
8143 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
8144 [ARG_PTR_TO_MEM] = &mem_types,
8145 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
8146 [ARG_PTR_TO_INT] = &int_ptr_types,
8147 [ARG_PTR_TO_LONG] = &int_ptr_types,
8148 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
8149 [ARG_PTR_TO_FUNC] = &func_ptr_types,
8150 [ARG_PTR_TO_STACK] = &stack_ptr_types,
8151 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
8152 [ARG_PTR_TO_TIMER] = &timer_types,
8153 [ARG_PTR_TO_KPTR] = &kptr_types,
8154 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
8155 };
8156
check_reg_type(struct bpf_verifier_env * env,u32 regno,enum bpf_arg_type arg_type,const u32 * arg_btf_id,struct bpf_call_arg_meta * meta)8157 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
8158 enum bpf_arg_type arg_type,
8159 const u32 *arg_btf_id,
8160 struct bpf_call_arg_meta *meta)
8161 {
8162 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8163 enum bpf_reg_type expected, type = reg->type;
8164 const struct bpf_reg_types *compatible;
8165 int i, j;
8166
8167 compatible = compatible_reg_types[base_type(arg_type)];
8168 if (!compatible) {
8169 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
8170 return -EFAULT;
8171 }
8172
8173 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
8174 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
8175 *
8176 * Same for MAYBE_NULL:
8177 *
8178 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
8179 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
8180 *
8181 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
8182 *
8183 * Therefore we fold these flags depending on the arg_type before comparison.
8184 */
8185 if (arg_type & MEM_RDONLY)
8186 type &= ~MEM_RDONLY;
8187 if (arg_type & PTR_MAYBE_NULL)
8188 type &= ~PTR_MAYBE_NULL;
8189 if (base_type(arg_type) == ARG_PTR_TO_MEM)
8190 type &= ~DYNPTR_TYPE_FLAG_MASK;
8191
8192 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
8193 type &= ~MEM_ALLOC;
8194
8195 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
8196 expected = compatible->types[i];
8197 if (expected == NOT_INIT)
8198 break;
8199
8200 if (type == expected)
8201 goto found;
8202 }
8203
8204 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
8205 for (j = 0; j + 1 < i; j++)
8206 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
8207 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
8208 return -EACCES;
8209
8210 found:
8211 if (base_type(reg->type) != PTR_TO_BTF_ID)
8212 return 0;
8213
8214 if (compatible == &mem_types) {
8215 if (!(arg_type & MEM_RDONLY)) {
8216 verbose(env,
8217 "%s() may write into memory pointed by R%d type=%s\n",
8218 func_id_name(meta->func_id),
8219 regno, reg_type_str(env, reg->type));
8220 return -EACCES;
8221 }
8222 return 0;
8223 }
8224
8225 switch ((int)reg->type) {
8226 case PTR_TO_BTF_ID:
8227 case PTR_TO_BTF_ID | PTR_TRUSTED:
8228 case PTR_TO_BTF_ID | MEM_RCU:
8229 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
8230 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
8231 {
8232 /* For bpf_sk_release, it needs to match against first member
8233 * 'struct sock_common', hence make an exception for it. This
8234 * allows bpf_sk_release to work for multiple socket types.
8235 */
8236 bool strict_type_match = arg_type_is_release(arg_type) &&
8237 meta->func_id != BPF_FUNC_sk_release;
8238
8239 if (type_may_be_null(reg->type) &&
8240 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
8241 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
8242 return -EACCES;
8243 }
8244
8245 if (!arg_btf_id) {
8246 if (!compatible->btf_id) {
8247 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
8248 return -EFAULT;
8249 }
8250 arg_btf_id = compatible->btf_id;
8251 }
8252
8253 if (meta->func_id == BPF_FUNC_kptr_xchg) {
8254 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
8255 return -EACCES;
8256 } else {
8257 if (arg_btf_id == BPF_PTR_POISON) {
8258 verbose(env, "verifier internal error:");
8259 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
8260 regno);
8261 return -EACCES;
8262 }
8263
8264 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
8265 btf_vmlinux, *arg_btf_id,
8266 strict_type_match)) {
8267 verbose(env, "R%d is of type %s but %s is expected\n",
8268 regno, btf_type_name(reg->btf, reg->btf_id),
8269 btf_type_name(btf_vmlinux, *arg_btf_id));
8270 return -EACCES;
8271 }
8272 }
8273 break;
8274 }
8275 case PTR_TO_BTF_ID | MEM_ALLOC:
8276 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
8277 meta->func_id != BPF_FUNC_kptr_xchg) {
8278 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
8279 return -EFAULT;
8280 }
8281 if (meta->func_id == BPF_FUNC_kptr_xchg) {
8282 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
8283 return -EACCES;
8284 }
8285 break;
8286 case PTR_TO_BTF_ID | MEM_PERCPU:
8287 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
8288 /* Handled by helper specific checks */
8289 break;
8290 default:
8291 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
8292 return -EFAULT;
8293 }
8294 return 0;
8295 }
8296
8297 static struct btf_field *
reg_find_field_offset(const struct bpf_reg_state * reg,s32 off,u32 fields)8298 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
8299 {
8300 struct btf_field *field;
8301 struct btf_record *rec;
8302
8303 rec = reg_btf_record(reg);
8304 if (!rec)
8305 return NULL;
8306
8307 field = btf_record_find(rec, off, fields);
8308 if (!field)
8309 return NULL;
8310
8311 return field;
8312 }
8313
check_func_arg_reg_off(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,enum bpf_arg_type arg_type)8314 int check_func_arg_reg_off(struct bpf_verifier_env *env,
8315 const struct bpf_reg_state *reg, int regno,
8316 enum bpf_arg_type arg_type)
8317 {
8318 u32 type = reg->type;
8319
8320 /* When referenced register is passed to release function, its fixed
8321 * offset must be 0.
8322 *
8323 * We will check arg_type_is_release reg has ref_obj_id when storing
8324 * meta->release_regno.
8325 */
8326 if (arg_type_is_release(arg_type)) {
8327 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
8328 * may not directly point to the object being released, but to
8329 * dynptr pointing to such object, which might be at some offset
8330 * on the stack. In that case, we simply to fallback to the
8331 * default handling.
8332 */
8333 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
8334 return 0;
8335
8336 /* Doing check_ptr_off_reg check for the offset will catch this
8337 * because fixed_off_ok is false, but checking here allows us
8338 * to give the user a better error message.
8339 */
8340 if (reg->off) {
8341 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
8342 regno);
8343 return -EINVAL;
8344 }
8345 return __check_ptr_off_reg(env, reg, regno, false);
8346 }
8347
8348 switch (type) {
8349 /* Pointer types where both fixed and variable offset is explicitly allowed: */
8350 case PTR_TO_STACK:
8351 case PTR_TO_PACKET:
8352 case PTR_TO_PACKET_META:
8353 case PTR_TO_MAP_KEY:
8354 case PTR_TO_MAP_VALUE:
8355 case PTR_TO_MEM:
8356 case PTR_TO_MEM | MEM_RDONLY:
8357 case PTR_TO_MEM | MEM_RINGBUF:
8358 case PTR_TO_BUF:
8359 case PTR_TO_BUF | MEM_RDONLY:
8360 case SCALAR_VALUE:
8361 return 0;
8362 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
8363 * fixed offset.
8364 */
8365 case PTR_TO_BTF_ID:
8366 case PTR_TO_BTF_ID | MEM_ALLOC:
8367 case PTR_TO_BTF_ID | PTR_TRUSTED:
8368 case PTR_TO_BTF_ID | MEM_RCU:
8369 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
8370 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
8371 /* When referenced PTR_TO_BTF_ID is passed to release function,
8372 * its fixed offset must be 0. In the other cases, fixed offset
8373 * can be non-zero. This was already checked above. So pass
8374 * fixed_off_ok as true to allow fixed offset for all other
8375 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
8376 * still need to do checks instead of returning.
8377 */
8378 return __check_ptr_off_reg(env, reg, regno, true);
8379 default:
8380 return __check_ptr_off_reg(env, reg, regno, false);
8381 }
8382 }
8383
get_dynptr_arg_reg(struct bpf_verifier_env * env,const struct bpf_func_proto * fn,struct bpf_reg_state * regs)8384 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
8385 const struct bpf_func_proto *fn,
8386 struct bpf_reg_state *regs)
8387 {
8388 struct bpf_reg_state *state = NULL;
8389 int i;
8390
8391 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
8392 if (arg_type_is_dynptr(fn->arg_type[i])) {
8393 if (state) {
8394 verbose(env, "verifier internal error: multiple dynptr args\n");
8395 return NULL;
8396 }
8397 state = ®s[BPF_REG_1 + i];
8398 }
8399
8400 if (!state)
8401 verbose(env, "verifier internal error: no dynptr arg found\n");
8402
8403 return state;
8404 }
8405
dynptr_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg)8406 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8407 {
8408 struct bpf_func_state *state = func(env, reg);
8409 int spi;
8410
8411 if (reg->type == CONST_PTR_TO_DYNPTR)
8412 return reg->id;
8413 spi = dynptr_get_spi(env, reg);
8414 if (spi < 0)
8415 return spi;
8416 return state->stack[spi].spilled_ptr.id;
8417 }
8418
dynptr_ref_obj_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg)8419 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8420 {
8421 struct bpf_func_state *state = func(env, reg);
8422 int spi;
8423
8424 if (reg->type == CONST_PTR_TO_DYNPTR)
8425 return reg->ref_obj_id;
8426 spi = dynptr_get_spi(env, reg);
8427 if (spi < 0)
8428 return spi;
8429 return state->stack[spi].spilled_ptr.ref_obj_id;
8430 }
8431
dynptr_get_type(struct bpf_verifier_env * env,struct bpf_reg_state * reg)8432 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
8433 struct bpf_reg_state *reg)
8434 {
8435 struct bpf_func_state *state = func(env, reg);
8436 int spi;
8437
8438 if (reg->type == CONST_PTR_TO_DYNPTR)
8439 return reg->dynptr.type;
8440
8441 spi = __get_spi(reg->off);
8442 if (spi < 0) {
8443 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
8444 return BPF_DYNPTR_TYPE_INVALID;
8445 }
8446
8447 return state->stack[spi].spilled_ptr.dynptr.type;
8448 }
8449
check_func_arg(struct bpf_verifier_env * env,u32 arg,struct bpf_call_arg_meta * meta,const struct bpf_func_proto * fn,int insn_idx)8450 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
8451 struct bpf_call_arg_meta *meta,
8452 const struct bpf_func_proto *fn,
8453 int insn_idx)
8454 {
8455 u32 regno = BPF_REG_1 + arg;
8456 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8457 enum bpf_arg_type arg_type = fn->arg_type[arg];
8458 enum bpf_reg_type type = reg->type;
8459 u32 *arg_btf_id = NULL;
8460 int err = 0;
8461
8462 if (arg_type == ARG_DONTCARE)
8463 return 0;
8464
8465 err = check_reg_arg(env, regno, SRC_OP);
8466 if (err)
8467 return err;
8468
8469 if (arg_type == ARG_ANYTHING) {
8470 if (is_pointer_value(env, regno)) {
8471 verbose(env, "R%d leaks addr into helper function\n",
8472 regno);
8473 return -EACCES;
8474 }
8475 return 0;
8476 }
8477
8478 if (type_is_pkt_pointer(type) &&
8479 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
8480 verbose(env, "helper access to the packet is not allowed\n");
8481 return -EACCES;
8482 }
8483
8484 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
8485 err = resolve_map_arg_type(env, meta, &arg_type);
8486 if (err)
8487 return err;
8488 }
8489
8490 if (register_is_null(reg) && type_may_be_null(arg_type))
8491 /* A NULL register has a SCALAR_VALUE type, so skip
8492 * type checking.
8493 */
8494 goto skip_type_check;
8495
8496 /* arg_btf_id and arg_size are in a union. */
8497 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
8498 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
8499 arg_btf_id = fn->arg_btf_id[arg];
8500
8501 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
8502 if (err)
8503 return err;
8504
8505 err = check_func_arg_reg_off(env, reg, regno, arg_type);
8506 if (err)
8507 return err;
8508
8509 skip_type_check:
8510 if (arg_type_is_release(arg_type)) {
8511 if (arg_type_is_dynptr(arg_type)) {
8512 struct bpf_func_state *state = func(env, reg);
8513 int spi;
8514
8515 /* Only dynptr created on stack can be released, thus
8516 * the get_spi and stack state checks for spilled_ptr
8517 * should only be done before process_dynptr_func for
8518 * PTR_TO_STACK.
8519 */
8520 if (reg->type == PTR_TO_STACK) {
8521 spi = dynptr_get_spi(env, reg);
8522 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
8523 verbose(env, "arg %d is an unacquired reference\n", regno);
8524 return -EINVAL;
8525 }
8526 } else {
8527 verbose(env, "cannot release unowned const bpf_dynptr\n");
8528 return -EINVAL;
8529 }
8530 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
8531 verbose(env, "R%d must be referenced when passed to release function\n",
8532 regno);
8533 return -EINVAL;
8534 }
8535 if (meta->release_regno) {
8536 verbose(env, "verifier internal error: more than one release argument\n");
8537 return -EFAULT;
8538 }
8539 meta->release_regno = regno;
8540 }
8541
8542 if (reg->ref_obj_id) {
8543 if (meta->ref_obj_id) {
8544 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8545 regno, reg->ref_obj_id,
8546 meta->ref_obj_id);
8547 return -EFAULT;
8548 }
8549 meta->ref_obj_id = reg->ref_obj_id;
8550 }
8551
8552 switch (base_type(arg_type)) {
8553 case ARG_CONST_MAP_PTR:
8554 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8555 if (meta->map_ptr) {
8556 /* Use map_uid (which is unique id of inner map) to reject:
8557 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8558 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8559 * if (inner_map1 && inner_map2) {
8560 * timer = bpf_map_lookup_elem(inner_map1);
8561 * if (timer)
8562 * // mismatch would have been allowed
8563 * bpf_timer_init(timer, inner_map2);
8564 * }
8565 *
8566 * Comparing map_ptr is enough to distinguish normal and outer maps.
8567 */
8568 if (meta->map_ptr != reg->map_ptr ||
8569 meta->map_uid != reg->map_uid) {
8570 verbose(env,
8571 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8572 meta->map_uid, reg->map_uid);
8573 return -EINVAL;
8574 }
8575 }
8576 meta->map_ptr = reg->map_ptr;
8577 meta->map_uid = reg->map_uid;
8578 break;
8579 case ARG_PTR_TO_MAP_KEY:
8580 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8581 * check that [key, key + map->key_size) are within
8582 * stack limits and initialized
8583 */
8584 if (!meta->map_ptr) {
8585 /* in function declaration map_ptr must come before
8586 * map_key, so that it's verified and known before
8587 * we have to check map_key here. Otherwise it means
8588 * that kernel subsystem misconfigured verifier
8589 */
8590 verbose(env, "invalid map_ptr to access map->key\n");
8591 return -EACCES;
8592 }
8593 err = check_helper_mem_access(env, regno,
8594 meta->map_ptr->key_size, false,
8595 NULL);
8596 break;
8597 case ARG_PTR_TO_MAP_VALUE:
8598 if (type_may_be_null(arg_type) && register_is_null(reg))
8599 return 0;
8600
8601 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8602 * check [value, value + map->value_size) validity
8603 */
8604 if (!meta->map_ptr) {
8605 /* kernel subsystem misconfigured verifier */
8606 verbose(env, "invalid map_ptr to access map->value\n");
8607 return -EACCES;
8608 }
8609 meta->raw_mode = arg_type & MEM_UNINIT;
8610 err = check_helper_mem_access(env, regno,
8611 meta->map_ptr->value_size, false,
8612 meta);
8613 break;
8614 case ARG_PTR_TO_PERCPU_BTF_ID:
8615 if (!reg->btf_id) {
8616 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8617 return -EACCES;
8618 }
8619 meta->ret_btf = reg->btf;
8620 meta->ret_btf_id = reg->btf_id;
8621 break;
8622 case ARG_PTR_TO_SPIN_LOCK:
8623 if (in_rbtree_lock_required_cb(env)) {
8624 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8625 return -EACCES;
8626 }
8627 if (meta->func_id == BPF_FUNC_spin_lock) {
8628 err = process_spin_lock(env, regno, true);
8629 if (err)
8630 return err;
8631 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8632 err = process_spin_lock(env, regno, false);
8633 if (err)
8634 return err;
8635 } else {
8636 verbose(env, "verifier internal error\n");
8637 return -EFAULT;
8638 }
8639 break;
8640 case ARG_PTR_TO_TIMER:
8641 err = process_timer_func(env, regno, meta);
8642 if (err)
8643 return err;
8644 break;
8645 case ARG_PTR_TO_FUNC:
8646 meta->subprogno = reg->subprogno;
8647 break;
8648 case ARG_PTR_TO_MEM:
8649 /* The access to this pointer is only checked when we hit the
8650 * next is_mem_size argument below.
8651 */
8652 meta->raw_mode = arg_type & MEM_UNINIT;
8653 if (arg_type & MEM_FIXED_SIZE) {
8654 err = check_helper_mem_access(env, regno,
8655 fn->arg_size[arg], false,
8656 meta);
8657 }
8658 break;
8659 case ARG_CONST_SIZE:
8660 err = check_mem_size_reg(env, reg, regno, false, meta);
8661 break;
8662 case ARG_CONST_SIZE_OR_ZERO:
8663 err = check_mem_size_reg(env, reg, regno, true, meta);
8664 break;
8665 case ARG_PTR_TO_DYNPTR:
8666 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8667 if (err)
8668 return err;
8669 break;
8670 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8671 if (!tnum_is_const(reg->var_off)) {
8672 verbose(env, "R%d is not a known constant'\n",
8673 regno);
8674 return -EACCES;
8675 }
8676 meta->mem_size = reg->var_off.value;
8677 err = mark_chain_precision(env, regno);
8678 if (err)
8679 return err;
8680 break;
8681 case ARG_PTR_TO_INT:
8682 case ARG_PTR_TO_LONG:
8683 {
8684 int size = int_ptr_type_to_size(arg_type);
8685
8686 err = check_helper_mem_access(env, regno, size, false, meta);
8687 if (err)
8688 return err;
8689 err = check_ptr_alignment(env, reg, 0, size, true);
8690 break;
8691 }
8692 case ARG_PTR_TO_CONST_STR:
8693 {
8694 struct bpf_map *map = reg->map_ptr;
8695 int map_off;
8696 u64 map_addr;
8697 char *str_ptr;
8698
8699 if (!bpf_map_is_rdonly(map)) {
8700 verbose(env, "R%d does not point to a readonly map'\n", regno);
8701 return -EACCES;
8702 }
8703
8704 if (!tnum_is_const(reg->var_off)) {
8705 verbose(env, "R%d is not a constant address'\n", regno);
8706 return -EACCES;
8707 }
8708
8709 if (!map->ops->map_direct_value_addr) {
8710 verbose(env, "no direct value access support for this map type\n");
8711 return -EACCES;
8712 }
8713
8714 err = check_map_access(env, regno, reg->off,
8715 map->value_size - reg->off, false,
8716 ACCESS_HELPER);
8717 if (err)
8718 return err;
8719
8720 map_off = reg->off + reg->var_off.value;
8721 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8722 if (err) {
8723 verbose(env, "direct value access on string failed\n");
8724 return err;
8725 }
8726
8727 str_ptr = (char *)(long)(map_addr);
8728 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8729 verbose(env, "string is not zero-terminated\n");
8730 return -EINVAL;
8731 }
8732 break;
8733 }
8734 case ARG_PTR_TO_KPTR:
8735 err = process_kptr_func(env, regno, meta);
8736 if (err)
8737 return err;
8738 break;
8739 }
8740
8741 return err;
8742 }
8743
may_update_sockmap(struct bpf_verifier_env * env,int func_id)8744 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8745 {
8746 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8747 enum bpf_prog_type type = resolve_prog_type(env->prog);
8748
8749 if (func_id != BPF_FUNC_map_update_elem &&
8750 func_id != BPF_FUNC_map_delete_elem)
8751 return false;
8752
8753 /* It's not possible to get access to a locked struct sock in these
8754 * contexts, so updating is safe.
8755 */
8756 switch (type) {
8757 case BPF_PROG_TYPE_TRACING:
8758 if (eatype == BPF_TRACE_ITER)
8759 return true;
8760 break;
8761 case BPF_PROG_TYPE_SOCK_OPS:
8762 /* map_update allowed only via dedicated helpers with event type checks */
8763 if (func_id == BPF_FUNC_map_delete_elem)
8764 return true;
8765 break;
8766 case BPF_PROG_TYPE_SOCKET_FILTER:
8767 case BPF_PROG_TYPE_SCHED_CLS:
8768 case BPF_PROG_TYPE_SCHED_ACT:
8769 case BPF_PROG_TYPE_XDP:
8770 case BPF_PROG_TYPE_SK_REUSEPORT:
8771 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8772 case BPF_PROG_TYPE_SK_LOOKUP:
8773 return true;
8774 default:
8775 break;
8776 }
8777
8778 verbose(env, "cannot update sockmap in this context\n");
8779 return false;
8780 }
8781
allow_tail_call_in_subprogs(struct bpf_verifier_env * env)8782 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8783 {
8784 return env->prog->jit_requested &&
8785 bpf_jit_supports_subprog_tailcalls();
8786 }
8787
check_map_func_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,int func_id)8788 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8789 struct bpf_map *map, int func_id)
8790 {
8791 if (!map)
8792 return 0;
8793
8794 /* We need a two way check, first is from map perspective ... */
8795 switch (map->map_type) {
8796 case BPF_MAP_TYPE_PROG_ARRAY:
8797 if (func_id != BPF_FUNC_tail_call)
8798 goto error;
8799 break;
8800 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8801 if (func_id != BPF_FUNC_perf_event_read &&
8802 func_id != BPF_FUNC_perf_event_output &&
8803 func_id != BPF_FUNC_skb_output &&
8804 func_id != BPF_FUNC_perf_event_read_value &&
8805 func_id != BPF_FUNC_xdp_output)
8806 goto error;
8807 break;
8808 case BPF_MAP_TYPE_RINGBUF:
8809 if (func_id != BPF_FUNC_ringbuf_output &&
8810 func_id != BPF_FUNC_ringbuf_reserve &&
8811 func_id != BPF_FUNC_ringbuf_query &&
8812 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8813 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8814 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8815 goto error;
8816 break;
8817 case BPF_MAP_TYPE_USER_RINGBUF:
8818 if (func_id != BPF_FUNC_user_ringbuf_drain)
8819 goto error;
8820 break;
8821 case BPF_MAP_TYPE_STACK_TRACE:
8822 if (func_id != BPF_FUNC_get_stackid)
8823 goto error;
8824 break;
8825 case BPF_MAP_TYPE_CGROUP_ARRAY:
8826 if (func_id != BPF_FUNC_skb_under_cgroup &&
8827 func_id != BPF_FUNC_current_task_under_cgroup)
8828 goto error;
8829 break;
8830 case BPF_MAP_TYPE_CGROUP_STORAGE:
8831 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8832 if (func_id != BPF_FUNC_get_local_storage)
8833 goto error;
8834 break;
8835 case BPF_MAP_TYPE_DEVMAP:
8836 case BPF_MAP_TYPE_DEVMAP_HASH:
8837 if (func_id != BPF_FUNC_redirect_map &&
8838 func_id != BPF_FUNC_map_lookup_elem)
8839 goto error;
8840 break;
8841 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8842 * appear.
8843 */
8844 case BPF_MAP_TYPE_CPUMAP:
8845 if (func_id != BPF_FUNC_redirect_map)
8846 goto error;
8847 break;
8848 case BPF_MAP_TYPE_XSKMAP:
8849 if (func_id != BPF_FUNC_redirect_map &&
8850 func_id != BPF_FUNC_map_lookup_elem)
8851 goto error;
8852 break;
8853 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8854 case BPF_MAP_TYPE_HASH_OF_MAPS:
8855 if (func_id != BPF_FUNC_map_lookup_elem)
8856 goto error;
8857 break;
8858 case BPF_MAP_TYPE_SOCKMAP:
8859 if (func_id != BPF_FUNC_sk_redirect_map &&
8860 func_id != BPF_FUNC_sock_map_update &&
8861 func_id != BPF_FUNC_msg_redirect_map &&
8862 func_id != BPF_FUNC_sk_select_reuseport &&
8863 func_id != BPF_FUNC_map_lookup_elem &&
8864 !may_update_sockmap(env, func_id))
8865 goto error;
8866 break;
8867 case BPF_MAP_TYPE_SOCKHASH:
8868 if (func_id != BPF_FUNC_sk_redirect_hash &&
8869 func_id != BPF_FUNC_sock_hash_update &&
8870 func_id != BPF_FUNC_msg_redirect_hash &&
8871 func_id != BPF_FUNC_sk_select_reuseport &&
8872 func_id != BPF_FUNC_map_lookup_elem &&
8873 !may_update_sockmap(env, func_id))
8874 goto error;
8875 break;
8876 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8877 if (func_id != BPF_FUNC_sk_select_reuseport)
8878 goto error;
8879 break;
8880 case BPF_MAP_TYPE_QUEUE:
8881 case BPF_MAP_TYPE_STACK:
8882 if (func_id != BPF_FUNC_map_peek_elem &&
8883 func_id != BPF_FUNC_map_pop_elem &&
8884 func_id != BPF_FUNC_map_push_elem)
8885 goto error;
8886 break;
8887 case BPF_MAP_TYPE_SK_STORAGE:
8888 if (func_id != BPF_FUNC_sk_storage_get &&
8889 func_id != BPF_FUNC_sk_storage_delete &&
8890 func_id != BPF_FUNC_kptr_xchg)
8891 goto error;
8892 break;
8893 case BPF_MAP_TYPE_INODE_STORAGE:
8894 if (func_id != BPF_FUNC_inode_storage_get &&
8895 func_id != BPF_FUNC_inode_storage_delete &&
8896 func_id != BPF_FUNC_kptr_xchg)
8897 goto error;
8898 break;
8899 case BPF_MAP_TYPE_TASK_STORAGE:
8900 if (func_id != BPF_FUNC_task_storage_get &&
8901 func_id != BPF_FUNC_task_storage_delete &&
8902 func_id != BPF_FUNC_kptr_xchg)
8903 goto error;
8904 break;
8905 case BPF_MAP_TYPE_CGRP_STORAGE:
8906 if (func_id != BPF_FUNC_cgrp_storage_get &&
8907 func_id != BPF_FUNC_cgrp_storage_delete &&
8908 func_id != BPF_FUNC_kptr_xchg)
8909 goto error;
8910 break;
8911 case BPF_MAP_TYPE_BLOOM_FILTER:
8912 if (func_id != BPF_FUNC_map_peek_elem &&
8913 func_id != BPF_FUNC_map_push_elem)
8914 goto error;
8915 break;
8916 default:
8917 break;
8918 }
8919
8920 /* ... and second from the function itself. */
8921 switch (func_id) {
8922 case BPF_FUNC_tail_call:
8923 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8924 goto error;
8925 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8926 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8927 return -EINVAL;
8928 }
8929 break;
8930 case BPF_FUNC_perf_event_read:
8931 case BPF_FUNC_perf_event_output:
8932 case BPF_FUNC_perf_event_read_value:
8933 case BPF_FUNC_skb_output:
8934 case BPF_FUNC_xdp_output:
8935 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8936 goto error;
8937 break;
8938 case BPF_FUNC_ringbuf_output:
8939 case BPF_FUNC_ringbuf_reserve:
8940 case BPF_FUNC_ringbuf_query:
8941 case BPF_FUNC_ringbuf_reserve_dynptr:
8942 case BPF_FUNC_ringbuf_submit_dynptr:
8943 case BPF_FUNC_ringbuf_discard_dynptr:
8944 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8945 goto error;
8946 break;
8947 case BPF_FUNC_user_ringbuf_drain:
8948 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8949 goto error;
8950 break;
8951 case BPF_FUNC_get_stackid:
8952 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8953 goto error;
8954 break;
8955 case BPF_FUNC_current_task_under_cgroup:
8956 case BPF_FUNC_skb_under_cgroup:
8957 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8958 goto error;
8959 break;
8960 case BPF_FUNC_redirect_map:
8961 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8962 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8963 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8964 map->map_type != BPF_MAP_TYPE_XSKMAP)
8965 goto error;
8966 break;
8967 case BPF_FUNC_sk_redirect_map:
8968 case BPF_FUNC_msg_redirect_map:
8969 case BPF_FUNC_sock_map_update:
8970 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8971 goto error;
8972 break;
8973 case BPF_FUNC_sk_redirect_hash:
8974 case BPF_FUNC_msg_redirect_hash:
8975 case BPF_FUNC_sock_hash_update:
8976 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8977 goto error;
8978 break;
8979 case BPF_FUNC_get_local_storage:
8980 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8981 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8982 goto error;
8983 break;
8984 case BPF_FUNC_sk_select_reuseport:
8985 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8986 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8987 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8988 goto error;
8989 break;
8990 case BPF_FUNC_map_pop_elem:
8991 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8992 map->map_type != BPF_MAP_TYPE_STACK)
8993 goto error;
8994 break;
8995 case BPF_FUNC_map_peek_elem:
8996 case BPF_FUNC_map_push_elem:
8997 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8998 map->map_type != BPF_MAP_TYPE_STACK &&
8999 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
9000 goto error;
9001 break;
9002 case BPF_FUNC_map_lookup_percpu_elem:
9003 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
9004 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
9005 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
9006 goto error;
9007 break;
9008 case BPF_FUNC_sk_storage_get:
9009 case BPF_FUNC_sk_storage_delete:
9010 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
9011 goto error;
9012 break;
9013 case BPF_FUNC_inode_storage_get:
9014 case BPF_FUNC_inode_storage_delete:
9015 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
9016 goto error;
9017 break;
9018 case BPF_FUNC_task_storage_get:
9019 case BPF_FUNC_task_storage_delete:
9020 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
9021 goto error;
9022 break;
9023 case BPF_FUNC_cgrp_storage_get:
9024 case BPF_FUNC_cgrp_storage_delete:
9025 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
9026 goto error;
9027 break;
9028 default:
9029 break;
9030 }
9031
9032 return 0;
9033 error:
9034 verbose(env, "cannot pass map_type %d into func %s#%d\n",
9035 map->map_type, func_id_name(func_id), func_id);
9036 return -EINVAL;
9037 }
9038
check_raw_mode_ok(const struct bpf_func_proto * fn)9039 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
9040 {
9041 int count = 0;
9042
9043 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
9044 count++;
9045 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
9046 count++;
9047 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
9048 count++;
9049 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
9050 count++;
9051 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
9052 count++;
9053
9054 /* We only support one arg being in raw mode at the moment,
9055 * which is sufficient for the helper functions we have
9056 * right now.
9057 */
9058 return count <= 1;
9059 }
9060
check_args_pair_invalid(const struct bpf_func_proto * fn,int arg)9061 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
9062 {
9063 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
9064 bool has_size = fn->arg_size[arg] != 0;
9065 bool is_next_size = false;
9066
9067 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
9068 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
9069
9070 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
9071 return is_next_size;
9072
9073 return has_size == is_next_size || is_next_size == is_fixed;
9074 }
9075
check_arg_pair_ok(const struct bpf_func_proto * fn)9076 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
9077 {
9078 /* bpf_xxx(..., buf, len) call will access 'len'
9079 * bytes from memory 'buf'. Both arg types need
9080 * to be paired, so make sure there's no buggy
9081 * helper function specification.
9082 */
9083 if (arg_type_is_mem_size(fn->arg1_type) ||
9084 check_args_pair_invalid(fn, 0) ||
9085 check_args_pair_invalid(fn, 1) ||
9086 check_args_pair_invalid(fn, 2) ||
9087 check_args_pair_invalid(fn, 3) ||
9088 check_args_pair_invalid(fn, 4))
9089 return false;
9090
9091 return true;
9092 }
9093
check_btf_id_ok(const struct bpf_func_proto * fn)9094 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
9095 {
9096 int i;
9097
9098 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
9099 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
9100 return !!fn->arg_btf_id[i];
9101 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
9102 return fn->arg_btf_id[i] == BPF_PTR_POISON;
9103 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
9104 /* arg_btf_id and arg_size are in a union. */
9105 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
9106 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
9107 return false;
9108 }
9109
9110 return true;
9111 }
9112
check_func_proto(const struct bpf_func_proto * fn,int func_id)9113 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
9114 {
9115 return check_raw_mode_ok(fn) &&
9116 check_arg_pair_ok(fn) &&
9117 check_btf_id_ok(fn) ? 0 : -EINVAL;
9118 }
9119
9120 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
9121 * are now invalid, so turn them into unknown SCALAR_VALUE.
9122 *
9123 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
9124 * since these slices point to packet data.
9125 */
clear_all_pkt_pointers(struct bpf_verifier_env * env)9126 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
9127 {
9128 struct bpf_func_state *state;
9129 struct bpf_reg_state *reg;
9130
9131 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
9132 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
9133 mark_reg_invalid(env, reg);
9134 }));
9135 }
9136
9137 enum {
9138 AT_PKT_END = -1,
9139 BEYOND_PKT_END = -2,
9140 };
9141
mark_pkt_end(struct bpf_verifier_state * vstate,int regn,bool range_open)9142 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
9143 {
9144 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9145 struct bpf_reg_state *reg = &state->regs[regn];
9146
9147 if (reg->type != PTR_TO_PACKET)
9148 /* PTR_TO_PACKET_META is not supported yet */
9149 return;
9150
9151 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
9152 * How far beyond pkt_end it goes is unknown.
9153 * if (!range_open) it's the case of pkt >= pkt_end
9154 * if (range_open) it's the case of pkt > pkt_end
9155 * hence this pointer is at least 1 byte bigger than pkt_end
9156 */
9157 if (range_open)
9158 reg->range = BEYOND_PKT_END;
9159 else
9160 reg->range = AT_PKT_END;
9161 }
9162
9163 /* The pointer with the specified id has released its reference to kernel
9164 * resources. Identify all copies of the same pointer and clear the reference.
9165 */
release_reference(struct bpf_verifier_env * env,int ref_obj_id)9166 static int release_reference(struct bpf_verifier_env *env,
9167 int ref_obj_id)
9168 {
9169 struct bpf_func_state *state;
9170 struct bpf_reg_state *reg;
9171 int err;
9172
9173 err = release_reference_state(cur_func(env), ref_obj_id);
9174 if (err)
9175 return err;
9176
9177 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
9178 if (reg->ref_obj_id == ref_obj_id)
9179 mark_reg_invalid(env, reg);
9180 }));
9181
9182 return 0;
9183 }
9184
invalidate_non_owning_refs(struct bpf_verifier_env * env)9185 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
9186 {
9187 struct bpf_func_state *unused;
9188 struct bpf_reg_state *reg;
9189
9190 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
9191 if (type_is_non_owning_ref(reg->type))
9192 mark_reg_invalid(env, reg);
9193 }));
9194 }
9195
clear_caller_saved_regs(struct bpf_verifier_env * env,struct bpf_reg_state * regs)9196 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
9197 struct bpf_reg_state *regs)
9198 {
9199 int i;
9200
9201 /* after the call registers r0 - r5 were scratched */
9202 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9203 mark_reg_not_init(env, regs, caller_saved[i]);
9204 __check_reg_arg(env, regs, caller_saved[i], DST_OP_NO_MARK);
9205 }
9206 }
9207
9208 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
9209 struct bpf_func_state *caller,
9210 struct bpf_func_state *callee,
9211 int insn_idx);
9212
9213 static int set_callee_state(struct bpf_verifier_env *env,
9214 struct bpf_func_state *caller,
9215 struct bpf_func_state *callee, int insn_idx);
9216
setup_func_entry(struct bpf_verifier_env * env,int subprog,int callsite,set_callee_state_fn set_callee_state_cb,struct bpf_verifier_state * state)9217 static int setup_func_entry(struct bpf_verifier_env *env, int subprog, int callsite,
9218 set_callee_state_fn set_callee_state_cb,
9219 struct bpf_verifier_state *state)
9220 {
9221 struct bpf_func_state *caller, *callee;
9222 int err;
9223
9224 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
9225 verbose(env, "the call stack of %d frames is too deep\n",
9226 state->curframe + 2);
9227 return -E2BIG;
9228 }
9229
9230 if (state->frame[state->curframe + 1]) {
9231 verbose(env, "verifier bug. Frame %d already allocated\n",
9232 state->curframe + 1);
9233 return -EFAULT;
9234 }
9235
9236 caller = state->frame[state->curframe];
9237 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
9238 if (!callee)
9239 return -ENOMEM;
9240 state->frame[state->curframe + 1] = callee;
9241
9242 /* callee cannot access r0, r6 - r9 for reading and has to write
9243 * into its own stack before reading from it.
9244 * callee can read/write into caller's stack
9245 */
9246 init_func_state(env, callee,
9247 /* remember the callsite, it will be used by bpf_exit */
9248 callsite,
9249 state->curframe + 1 /* frameno within this callchain */,
9250 subprog /* subprog number within this prog */);
9251 /* Transfer references to the callee */
9252 err = copy_reference_state(callee, caller);
9253 err = err ?: set_callee_state_cb(env, caller, callee, callsite);
9254 if (err)
9255 goto err_out;
9256
9257 /* only increment it after check_reg_arg() finished */
9258 state->curframe++;
9259
9260 return 0;
9261
9262 err_out:
9263 free_func_state(callee);
9264 state->frame[state->curframe + 1] = NULL;
9265 return err;
9266 }
9267
push_callback_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int insn_idx,int subprog,set_callee_state_fn set_callee_state_cb)9268 static int push_callback_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9269 int insn_idx, int subprog,
9270 set_callee_state_fn set_callee_state_cb)
9271 {
9272 struct bpf_verifier_state *state = env->cur_state, *callback_state;
9273 struct bpf_func_state *caller, *callee;
9274 int err;
9275
9276 caller = state->frame[state->curframe];
9277 err = btf_check_subprog_call(env, subprog, caller->regs);
9278 if (err == -EFAULT)
9279 return err;
9280
9281 /* set_callee_state is used for direct subprog calls, but we are
9282 * interested in validating only BPF helpers that can call subprogs as
9283 * callbacks
9284 */
9285 if (bpf_pseudo_kfunc_call(insn) &&
9286 !is_sync_callback_calling_kfunc(insn->imm)) {
9287 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
9288 func_id_name(insn->imm), insn->imm);
9289 return -EFAULT;
9290 } else if (!bpf_pseudo_kfunc_call(insn) &&
9291 !is_callback_calling_function(insn->imm)) { /* helper */
9292 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
9293 func_id_name(insn->imm), insn->imm);
9294 return -EFAULT;
9295 }
9296
9297 if (insn->code == (BPF_JMP | BPF_CALL) &&
9298 insn->src_reg == 0 &&
9299 insn->imm == BPF_FUNC_timer_set_callback) {
9300 struct bpf_verifier_state *async_cb;
9301
9302 /* there is no real recursion here. timer callbacks are async */
9303 env->subprog_info[subprog].is_async_cb = true;
9304 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
9305 insn_idx, subprog);
9306 if (!async_cb)
9307 return -EFAULT;
9308 callee = async_cb->frame[0];
9309 callee->async_entry_cnt = caller->async_entry_cnt + 1;
9310
9311 /* Convert bpf_timer_set_callback() args into timer callback args */
9312 err = set_callee_state_cb(env, caller, callee, insn_idx);
9313 if (err)
9314 return err;
9315
9316 return 0;
9317 }
9318
9319 /* for callback functions enqueue entry to callback and
9320 * proceed with next instruction within current frame.
9321 */
9322 callback_state = push_stack(env, env->subprog_info[subprog].start, insn_idx, false);
9323 if (!callback_state)
9324 return -ENOMEM;
9325
9326 err = setup_func_entry(env, subprog, insn_idx, set_callee_state_cb,
9327 callback_state);
9328 if (err)
9329 return err;
9330
9331 callback_state->callback_unroll_depth++;
9332 callback_state->frame[callback_state->curframe - 1]->callback_depth++;
9333 caller->callback_depth = 0;
9334 return 0;
9335 }
9336
check_func_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)9337 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9338 int *insn_idx)
9339 {
9340 struct bpf_verifier_state *state = env->cur_state;
9341 struct bpf_func_state *caller;
9342 int err, subprog, target_insn;
9343
9344 target_insn = *insn_idx + insn->imm + 1;
9345 subprog = find_subprog(env, target_insn);
9346 if (subprog < 0) {
9347 verbose(env, "verifier bug. No program starts at insn %d\n", target_insn);
9348 return -EFAULT;
9349 }
9350
9351 caller = state->frame[state->curframe];
9352 err = btf_check_subprog_call(env, subprog, caller->regs);
9353 if (err == -EFAULT)
9354 return err;
9355 if (subprog_is_global(env, subprog)) {
9356 if (err) {
9357 verbose(env, "Caller passes invalid args into func#%d\n", subprog);
9358 return err;
9359 }
9360
9361 if (env->log.level & BPF_LOG_LEVEL)
9362 verbose(env, "Func#%d is global and valid. Skipping.\n", subprog);
9363 clear_caller_saved_regs(env, caller->regs);
9364
9365 /* All global functions return a 64-bit SCALAR_VALUE */
9366 mark_reg_unknown(env, caller->regs, BPF_REG_0);
9367 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9368
9369 /* continue with next insn after call */
9370 return 0;
9371 }
9372
9373 /* for regular function entry setup new frame and continue
9374 * from that frame.
9375 */
9376 err = setup_func_entry(env, subprog, *insn_idx, set_callee_state, state);
9377 if (err)
9378 return err;
9379
9380 clear_caller_saved_regs(env, caller->regs);
9381
9382 /* and go analyze first insn of the callee */
9383 *insn_idx = env->subprog_info[subprog].start - 1;
9384
9385 if (env->log.level & BPF_LOG_LEVEL) {
9386 verbose(env, "caller:\n");
9387 print_verifier_state(env, caller, true);
9388 verbose(env, "callee:\n");
9389 print_verifier_state(env, state->frame[state->curframe], true);
9390 }
9391
9392 return 0;
9393 }
9394
map_set_for_each_callback_args(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee)9395 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
9396 struct bpf_func_state *caller,
9397 struct bpf_func_state *callee)
9398 {
9399 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
9400 * void *callback_ctx, u64 flags);
9401 * callback_fn(struct bpf_map *map, void *key, void *value,
9402 * void *callback_ctx);
9403 */
9404 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9405
9406 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9407 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9408 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9409
9410 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9411 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9412 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9413
9414 /* pointer to stack or null */
9415 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
9416
9417 /* unused */
9418 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9419 return 0;
9420 }
9421
set_callee_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9422 static int set_callee_state(struct bpf_verifier_env *env,
9423 struct bpf_func_state *caller,
9424 struct bpf_func_state *callee, int insn_idx)
9425 {
9426 int i;
9427
9428 /* copy r1 - r5 args that callee can access. The copy includes parent
9429 * pointers, which connects us up to the liveness chain
9430 */
9431 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
9432 callee->regs[i] = caller->regs[i];
9433 return 0;
9434 }
9435
set_map_elem_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9436 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
9437 struct bpf_func_state *caller,
9438 struct bpf_func_state *callee,
9439 int insn_idx)
9440 {
9441 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
9442 struct bpf_map *map;
9443 int err;
9444
9445 if (bpf_map_ptr_poisoned(insn_aux)) {
9446 verbose(env, "tail_call abusing map_ptr\n");
9447 return -EINVAL;
9448 }
9449
9450 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
9451 if (!map->ops->map_set_for_each_callback_args ||
9452 !map->ops->map_for_each_callback) {
9453 verbose(env, "callback function not allowed for map\n");
9454 return -ENOTSUPP;
9455 }
9456
9457 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
9458 if (err)
9459 return err;
9460
9461 callee->in_callback_fn = true;
9462 callee->callback_ret_range = tnum_range(0, 1);
9463 return 0;
9464 }
9465
set_loop_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9466 static int set_loop_callback_state(struct bpf_verifier_env *env,
9467 struct bpf_func_state *caller,
9468 struct bpf_func_state *callee,
9469 int insn_idx)
9470 {
9471 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
9472 * u64 flags);
9473 * callback_fn(u32 index, void *callback_ctx);
9474 */
9475 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
9476 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9477
9478 /* unused */
9479 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9480 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9481 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9482
9483 callee->in_callback_fn = true;
9484 callee->callback_ret_range = tnum_range(0, 1);
9485 return 0;
9486 }
9487
set_timer_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9488 static int set_timer_callback_state(struct bpf_verifier_env *env,
9489 struct bpf_func_state *caller,
9490 struct bpf_func_state *callee,
9491 int insn_idx)
9492 {
9493 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
9494
9495 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
9496 * callback_fn(struct bpf_map *map, void *key, void *value);
9497 */
9498 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
9499 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
9500 callee->regs[BPF_REG_1].map_ptr = map_ptr;
9501
9502 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9503 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9504 callee->regs[BPF_REG_2].map_ptr = map_ptr;
9505
9506 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9507 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9508 callee->regs[BPF_REG_3].map_ptr = map_ptr;
9509
9510 /* unused */
9511 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9512 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9513 callee->in_async_callback_fn = true;
9514 callee->callback_ret_range = tnum_range(0, 1);
9515 return 0;
9516 }
9517
set_find_vma_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9518 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
9519 struct bpf_func_state *caller,
9520 struct bpf_func_state *callee,
9521 int insn_idx)
9522 {
9523 /* bpf_find_vma(struct task_struct *task, u64 addr,
9524 * void *callback_fn, void *callback_ctx, u64 flags)
9525 * (callback_fn)(struct task_struct *task,
9526 * struct vm_area_struct *vma, void *callback_ctx);
9527 */
9528 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9529
9530 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
9531 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9532 callee->regs[BPF_REG_2].btf = btf_vmlinux;
9533 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
9534
9535 /* pointer to stack or null */
9536 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
9537
9538 /* unused */
9539 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9540 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9541 callee->in_callback_fn = true;
9542 callee->callback_ret_range = tnum_range(0, 1);
9543 return 0;
9544 }
9545
set_user_ringbuf_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9546 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
9547 struct bpf_func_state *caller,
9548 struct bpf_func_state *callee,
9549 int insn_idx)
9550 {
9551 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
9552 * callback_ctx, u64 flags);
9553 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
9554 */
9555 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
9556 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
9557 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9558
9559 /* unused */
9560 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9561 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9562 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9563
9564 callee->in_callback_fn = true;
9565 callee->callback_ret_range = tnum_range(0, 1);
9566 return 0;
9567 }
9568
set_rbtree_add_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9569 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
9570 struct bpf_func_state *caller,
9571 struct bpf_func_state *callee,
9572 int insn_idx)
9573 {
9574 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9575 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9576 *
9577 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9578 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9579 * by this point, so look at 'root'
9580 */
9581 struct btf_field *field;
9582
9583 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9584 BPF_RB_ROOT);
9585 if (!field || !field->graph_root.value_btf_id)
9586 return -EFAULT;
9587
9588 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9589 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9590 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9591 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9592
9593 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9594 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9595 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9596 callee->in_callback_fn = true;
9597 callee->callback_ret_range = tnum_range(0, 1);
9598 return 0;
9599 }
9600
9601 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9602
9603 /* Are we currently verifying the callback for a rbtree helper that must
9604 * be called with lock held? If so, no need to complain about unreleased
9605 * lock
9606 */
in_rbtree_lock_required_cb(struct bpf_verifier_env * env)9607 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9608 {
9609 struct bpf_verifier_state *state = env->cur_state;
9610 struct bpf_insn *insn = env->prog->insnsi;
9611 struct bpf_func_state *callee;
9612 int kfunc_btf_id;
9613
9614 if (!state->curframe)
9615 return false;
9616
9617 callee = state->frame[state->curframe];
9618
9619 if (!callee->in_callback_fn)
9620 return false;
9621
9622 kfunc_btf_id = insn[callee->callsite].imm;
9623 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9624 }
9625
prepare_func_exit(struct bpf_verifier_env * env,int * insn_idx)9626 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9627 {
9628 struct bpf_verifier_state *state = env->cur_state, *prev_st;
9629 struct bpf_func_state *caller, *callee;
9630 struct bpf_reg_state *r0;
9631 bool in_callback_fn;
9632 int err;
9633
9634 callee = state->frame[state->curframe];
9635 r0 = &callee->regs[BPF_REG_0];
9636 if (r0->type == PTR_TO_STACK) {
9637 /* technically it's ok to return caller's stack pointer
9638 * (or caller's caller's pointer) back to the caller,
9639 * since these pointers are valid. Only current stack
9640 * pointer will be invalid as soon as function exits,
9641 * but let's be conservative
9642 */
9643 verbose(env, "cannot return stack pointer to the caller\n");
9644 return -EINVAL;
9645 }
9646
9647 caller = state->frame[state->curframe - 1];
9648 if (callee->in_callback_fn) {
9649 /* enforce R0 return value range [0, 1]. */
9650 struct tnum range = callee->callback_ret_range;
9651
9652 if (r0->type != SCALAR_VALUE) {
9653 verbose(env, "R0 not a scalar value\n");
9654 return -EACCES;
9655 }
9656
9657 /* we are going to rely on register's precise value */
9658 err = mark_reg_read(env, r0, r0->parent, REG_LIVE_READ64);
9659 err = err ?: mark_chain_precision(env, BPF_REG_0);
9660 if (err)
9661 return err;
9662
9663 if (!tnum_in(range, r0->var_off)) {
9664 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9665 return -EINVAL;
9666 }
9667 if (!calls_callback(env, callee->callsite)) {
9668 verbose(env, "BUG: in callback at %d, callsite %d !calls_callback\n",
9669 *insn_idx, callee->callsite);
9670 return -EFAULT;
9671 }
9672 } else {
9673 /* return to the caller whatever r0 had in the callee */
9674 caller->regs[BPF_REG_0] = *r0;
9675 }
9676
9677 /* callback_fn frame should have released its own additions to parent's
9678 * reference state at this point, or check_reference_leak would
9679 * complain, hence it must be the same as the caller. There is no need
9680 * to copy it back.
9681 */
9682 if (!callee->in_callback_fn) {
9683 /* Transfer references to the caller */
9684 err = copy_reference_state(caller, callee);
9685 if (err)
9686 return err;
9687 }
9688
9689 /* for callbacks like bpf_loop or bpf_for_each_map_elem go back to callsite,
9690 * there function call logic would reschedule callback visit. If iteration
9691 * converges is_state_visited() would prune that visit eventually.
9692 */
9693 in_callback_fn = callee->in_callback_fn;
9694 if (in_callback_fn)
9695 *insn_idx = callee->callsite;
9696 else
9697 *insn_idx = callee->callsite + 1;
9698
9699 if (env->log.level & BPF_LOG_LEVEL) {
9700 verbose(env, "returning from callee:\n");
9701 print_verifier_state(env, callee, true);
9702 verbose(env, "to caller at %d:\n", *insn_idx);
9703 print_verifier_state(env, caller, true);
9704 }
9705 /* clear everything in the callee */
9706 free_func_state(callee);
9707 state->frame[state->curframe--] = NULL;
9708
9709 /* for callbacks widen imprecise scalars to make programs like below verify:
9710 *
9711 * struct ctx { int i; }
9712 * void cb(int idx, struct ctx *ctx) { ctx->i++; ... }
9713 * ...
9714 * struct ctx = { .i = 0; }
9715 * bpf_loop(100, cb, &ctx, 0);
9716 *
9717 * This is similar to what is done in process_iter_next_call() for open
9718 * coded iterators.
9719 */
9720 prev_st = in_callback_fn ? find_prev_entry(env, state, *insn_idx) : NULL;
9721 if (prev_st) {
9722 err = widen_imprecise_scalars(env, prev_st, state);
9723 if (err)
9724 return err;
9725 }
9726 return 0;
9727 }
9728
do_refine_retval_range(struct bpf_reg_state * regs,int ret_type,int func_id,struct bpf_call_arg_meta * meta)9729 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9730 int func_id,
9731 struct bpf_call_arg_meta *meta)
9732 {
9733 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9734
9735 if (ret_type != RET_INTEGER)
9736 return;
9737
9738 switch (func_id) {
9739 case BPF_FUNC_get_stack:
9740 case BPF_FUNC_get_task_stack:
9741 case BPF_FUNC_probe_read_str:
9742 case BPF_FUNC_probe_read_kernel_str:
9743 case BPF_FUNC_probe_read_user_str:
9744 ret_reg->smax_value = meta->msize_max_value;
9745 ret_reg->s32_max_value = meta->msize_max_value;
9746 ret_reg->smin_value = -MAX_ERRNO;
9747 ret_reg->s32_min_value = -MAX_ERRNO;
9748 reg_bounds_sync(ret_reg);
9749 break;
9750 case BPF_FUNC_get_smp_processor_id:
9751 ret_reg->umax_value = nr_cpu_ids - 1;
9752 ret_reg->u32_max_value = nr_cpu_ids - 1;
9753 ret_reg->smax_value = nr_cpu_ids - 1;
9754 ret_reg->s32_max_value = nr_cpu_ids - 1;
9755 ret_reg->umin_value = 0;
9756 ret_reg->u32_min_value = 0;
9757 ret_reg->smin_value = 0;
9758 ret_reg->s32_min_value = 0;
9759 reg_bounds_sync(ret_reg);
9760 break;
9761 }
9762 }
9763
9764 static int
record_func_map(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)9765 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9766 int func_id, int insn_idx)
9767 {
9768 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9769 struct bpf_map *map = meta->map_ptr;
9770
9771 if (func_id != BPF_FUNC_tail_call &&
9772 func_id != BPF_FUNC_map_lookup_elem &&
9773 func_id != BPF_FUNC_map_update_elem &&
9774 func_id != BPF_FUNC_map_delete_elem &&
9775 func_id != BPF_FUNC_map_push_elem &&
9776 func_id != BPF_FUNC_map_pop_elem &&
9777 func_id != BPF_FUNC_map_peek_elem &&
9778 func_id != BPF_FUNC_for_each_map_elem &&
9779 func_id != BPF_FUNC_redirect_map &&
9780 func_id != BPF_FUNC_map_lookup_percpu_elem)
9781 return 0;
9782
9783 if (map == NULL) {
9784 verbose(env, "kernel subsystem misconfigured verifier\n");
9785 return -EINVAL;
9786 }
9787
9788 /* In case of read-only, some additional restrictions
9789 * need to be applied in order to prevent altering the
9790 * state of the map from program side.
9791 */
9792 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9793 (func_id == BPF_FUNC_map_delete_elem ||
9794 func_id == BPF_FUNC_map_update_elem ||
9795 func_id == BPF_FUNC_map_push_elem ||
9796 func_id == BPF_FUNC_map_pop_elem)) {
9797 verbose(env, "write into map forbidden\n");
9798 return -EACCES;
9799 }
9800
9801 if (!BPF_MAP_PTR(aux->map_ptr_state))
9802 bpf_map_ptr_store(aux, meta->map_ptr,
9803 !meta->map_ptr->bypass_spec_v1);
9804 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9805 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9806 !meta->map_ptr->bypass_spec_v1);
9807 return 0;
9808 }
9809
9810 static int
record_func_key(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)9811 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9812 int func_id, int insn_idx)
9813 {
9814 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9815 struct bpf_reg_state *regs = cur_regs(env), *reg;
9816 struct bpf_map *map = meta->map_ptr;
9817 u64 val, max;
9818 int err;
9819
9820 if (func_id != BPF_FUNC_tail_call)
9821 return 0;
9822 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9823 verbose(env, "kernel subsystem misconfigured verifier\n");
9824 return -EINVAL;
9825 }
9826
9827 reg = ®s[BPF_REG_3];
9828 val = reg->var_off.value;
9829 max = map->max_entries;
9830
9831 if (!(register_is_const(reg) && val < max)) {
9832 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9833 return 0;
9834 }
9835
9836 err = mark_chain_precision(env, BPF_REG_3);
9837 if (err)
9838 return err;
9839 if (bpf_map_key_unseen(aux))
9840 bpf_map_key_store(aux, val);
9841 else if (!bpf_map_key_poisoned(aux) &&
9842 bpf_map_key_immediate(aux) != val)
9843 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9844 return 0;
9845 }
9846
check_reference_leak(struct bpf_verifier_env * env)9847 static int check_reference_leak(struct bpf_verifier_env *env)
9848 {
9849 struct bpf_func_state *state = cur_func(env);
9850 bool refs_lingering = false;
9851 int i;
9852
9853 if (state->frameno && !state->in_callback_fn)
9854 return 0;
9855
9856 for (i = 0; i < state->acquired_refs; i++) {
9857 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9858 continue;
9859 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9860 state->refs[i].id, state->refs[i].insn_idx);
9861 refs_lingering = true;
9862 }
9863 return refs_lingering ? -EINVAL : 0;
9864 }
9865
check_bpf_snprintf_call(struct bpf_verifier_env * env,struct bpf_reg_state * regs)9866 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9867 struct bpf_reg_state *regs)
9868 {
9869 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9870 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9871 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9872 struct bpf_bprintf_data data = {};
9873 int err, fmt_map_off, num_args;
9874 u64 fmt_addr;
9875 char *fmt;
9876
9877 /* data must be an array of u64 */
9878 if (data_len_reg->var_off.value % 8)
9879 return -EINVAL;
9880 num_args = data_len_reg->var_off.value / 8;
9881
9882 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9883 * and map_direct_value_addr is set.
9884 */
9885 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9886 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9887 fmt_map_off);
9888 if (err) {
9889 verbose(env, "verifier bug\n");
9890 return -EFAULT;
9891 }
9892 fmt = (char *)(long)fmt_addr + fmt_map_off;
9893
9894 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9895 * can focus on validating the format specifiers.
9896 */
9897 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9898 if (err < 0)
9899 verbose(env, "Invalid format string\n");
9900
9901 return err;
9902 }
9903
check_get_func_ip(struct bpf_verifier_env * env)9904 static int check_get_func_ip(struct bpf_verifier_env *env)
9905 {
9906 enum bpf_prog_type type = resolve_prog_type(env->prog);
9907 int func_id = BPF_FUNC_get_func_ip;
9908
9909 if (type == BPF_PROG_TYPE_TRACING) {
9910 if (!bpf_prog_has_trampoline(env->prog)) {
9911 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9912 func_id_name(func_id), func_id);
9913 return -ENOTSUPP;
9914 }
9915 return 0;
9916 } else if (type == BPF_PROG_TYPE_KPROBE) {
9917 return 0;
9918 }
9919
9920 verbose(env, "func %s#%d not supported for program type %d\n",
9921 func_id_name(func_id), func_id, type);
9922 return -ENOTSUPP;
9923 }
9924
cur_aux(struct bpf_verifier_env * env)9925 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9926 {
9927 return &env->insn_aux_data[env->insn_idx];
9928 }
9929
loop_flag_is_zero(struct bpf_verifier_env * env)9930 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9931 {
9932 struct bpf_reg_state *regs = cur_regs(env);
9933 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9934 bool reg_is_null = register_is_null(reg);
9935
9936 if (reg_is_null)
9937 mark_chain_precision(env, BPF_REG_4);
9938
9939 return reg_is_null;
9940 }
9941
update_loop_inline_state(struct bpf_verifier_env * env,u32 subprogno)9942 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9943 {
9944 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9945
9946 if (!state->initialized) {
9947 state->initialized = 1;
9948 state->fit_for_inline = loop_flag_is_zero(env);
9949 state->callback_subprogno = subprogno;
9950 return;
9951 }
9952
9953 if (!state->fit_for_inline)
9954 return;
9955
9956 state->fit_for_inline = (loop_flag_is_zero(env) &&
9957 state->callback_subprogno == subprogno);
9958 }
9959
check_helper_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx_p)9960 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9961 int *insn_idx_p)
9962 {
9963 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9964 const struct bpf_func_proto *fn = NULL;
9965 enum bpf_return_type ret_type;
9966 enum bpf_type_flag ret_flag;
9967 struct bpf_reg_state *regs;
9968 struct bpf_call_arg_meta meta;
9969 int insn_idx = *insn_idx_p;
9970 bool changes_data;
9971 int i, err, func_id;
9972
9973 /* find function prototype */
9974 func_id = insn->imm;
9975 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9976 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9977 func_id);
9978 return -EINVAL;
9979 }
9980
9981 if (env->ops->get_func_proto)
9982 fn = env->ops->get_func_proto(func_id, env->prog);
9983 if (!fn) {
9984 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9985 func_id);
9986 return -EINVAL;
9987 }
9988
9989 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9990 if (!env->prog->gpl_compatible && fn->gpl_only) {
9991 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9992 return -EINVAL;
9993 }
9994
9995 if (fn->allowed && !fn->allowed(env->prog)) {
9996 verbose(env, "helper call is not allowed in probe\n");
9997 return -EINVAL;
9998 }
9999
10000 if (!env->prog->aux->sleepable && fn->might_sleep) {
10001 verbose(env, "helper call might sleep in a non-sleepable prog\n");
10002 return -EINVAL;
10003 }
10004
10005 /* With LD_ABS/IND some JITs save/restore skb from r1. */
10006 changes_data = bpf_helper_changes_pkt_data(fn->func);
10007 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
10008 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
10009 func_id_name(func_id), func_id);
10010 return -EINVAL;
10011 }
10012
10013 memset(&meta, 0, sizeof(meta));
10014 meta.pkt_access = fn->pkt_access;
10015
10016 err = check_func_proto(fn, func_id);
10017 if (err) {
10018 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
10019 func_id_name(func_id), func_id);
10020 return err;
10021 }
10022
10023 if (env->cur_state->active_rcu_lock) {
10024 if (fn->might_sleep) {
10025 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
10026 func_id_name(func_id), func_id);
10027 return -EINVAL;
10028 }
10029
10030 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
10031 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
10032 }
10033
10034 meta.func_id = func_id;
10035 /* check args */
10036 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
10037 err = check_func_arg(env, i, &meta, fn, insn_idx);
10038 if (err)
10039 return err;
10040 }
10041
10042 err = record_func_map(env, &meta, func_id, insn_idx);
10043 if (err)
10044 return err;
10045
10046 err = record_func_key(env, &meta, func_id, insn_idx);
10047 if (err)
10048 return err;
10049
10050 /* Mark slots with STACK_MISC in case of raw mode, stack offset
10051 * is inferred from register state.
10052 */
10053 for (i = 0; i < meta.access_size; i++) {
10054 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
10055 BPF_WRITE, -1, false, false);
10056 if (err)
10057 return err;
10058 }
10059
10060 regs = cur_regs(env);
10061
10062 if (meta.release_regno) {
10063 err = -EINVAL;
10064 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
10065 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
10066 * is safe to do directly.
10067 */
10068 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
10069 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
10070 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
10071 return -EFAULT;
10072 }
10073 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
10074 } else if (meta.ref_obj_id) {
10075 err = release_reference(env, meta.ref_obj_id);
10076 } else if (register_is_null(®s[meta.release_regno])) {
10077 /* meta.ref_obj_id can only be 0 if register that is meant to be
10078 * released is NULL, which must be > R0.
10079 */
10080 err = 0;
10081 }
10082 if (err) {
10083 verbose(env, "func %s#%d reference has not been acquired before\n",
10084 func_id_name(func_id), func_id);
10085 return err;
10086 }
10087 }
10088
10089 switch (func_id) {
10090 case BPF_FUNC_tail_call:
10091 err = check_reference_leak(env);
10092 if (err) {
10093 verbose(env, "tail_call would lead to reference leak\n");
10094 return err;
10095 }
10096 break;
10097 case BPF_FUNC_get_local_storage:
10098 /* check that flags argument in get_local_storage(map, flags) is 0,
10099 * this is required because get_local_storage() can't return an error.
10100 */
10101 if (!register_is_null(®s[BPF_REG_2])) {
10102 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
10103 return -EINVAL;
10104 }
10105 break;
10106 case BPF_FUNC_for_each_map_elem:
10107 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10108 set_map_elem_callback_state);
10109 break;
10110 case BPF_FUNC_timer_set_callback:
10111 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10112 set_timer_callback_state);
10113 break;
10114 case BPF_FUNC_find_vma:
10115 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10116 set_find_vma_callback_state);
10117 break;
10118 case BPF_FUNC_snprintf:
10119 err = check_bpf_snprintf_call(env, regs);
10120 break;
10121 case BPF_FUNC_loop:
10122 update_loop_inline_state(env, meta.subprogno);
10123 /* Verifier relies on R1 value to determine if bpf_loop() iteration
10124 * is finished, thus mark it precise.
10125 */
10126 err = mark_chain_precision(env, BPF_REG_1);
10127 if (err)
10128 return err;
10129 if (cur_func(env)->callback_depth < regs[BPF_REG_1].umax_value) {
10130 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10131 set_loop_callback_state);
10132 } else {
10133 cur_func(env)->callback_depth = 0;
10134 if (env->log.level & BPF_LOG_LEVEL2)
10135 verbose(env, "frame%d bpf_loop iteration limit reached\n",
10136 env->cur_state->curframe);
10137 }
10138 break;
10139 case BPF_FUNC_dynptr_from_mem:
10140 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
10141 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
10142 reg_type_str(env, regs[BPF_REG_1].type));
10143 return -EACCES;
10144 }
10145 break;
10146 case BPF_FUNC_set_retval:
10147 if (prog_type == BPF_PROG_TYPE_LSM &&
10148 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
10149 if (!env->prog->aux->attach_func_proto->type) {
10150 /* Make sure programs that attach to void
10151 * hooks don't try to modify return value.
10152 */
10153 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
10154 return -EINVAL;
10155 }
10156 }
10157 break;
10158 case BPF_FUNC_dynptr_data:
10159 {
10160 struct bpf_reg_state *reg;
10161 int id, ref_obj_id;
10162
10163 reg = get_dynptr_arg_reg(env, fn, regs);
10164 if (!reg)
10165 return -EFAULT;
10166
10167
10168 if (meta.dynptr_id) {
10169 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
10170 return -EFAULT;
10171 }
10172 if (meta.ref_obj_id) {
10173 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
10174 return -EFAULT;
10175 }
10176
10177 id = dynptr_id(env, reg);
10178 if (id < 0) {
10179 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
10180 return id;
10181 }
10182
10183 ref_obj_id = dynptr_ref_obj_id(env, reg);
10184 if (ref_obj_id < 0) {
10185 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
10186 return ref_obj_id;
10187 }
10188
10189 meta.dynptr_id = id;
10190 meta.ref_obj_id = ref_obj_id;
10191
10192 break;
10193 }
10194 case BPF_FUNC_dynptr_write:
10195 {
10196 enum bpf_dynptr_type dynptr_type;
10197 struct bpf_reg_state *reg;
10198
10199 reg = get_dynptr_arg_reg(env, fn, regs);
10200 if (!reg)
10201 return -EFAULT;
10202
10203 dynptr_type = dynptr_get_type(env, reg);
10204 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
10205 return -EFAULT;
10206
10207 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
10208 /* this will trigger clear_all_pkt_pointers(), which will
10209 * invalidate all dynptr slices associated with the skb
10210 */
10211 changes_data = true;
10212
10213 break;
10214 }
10215 case BPF_FUNC_user_ringbuf_drain:
10216 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10217 set_user_ringbuf_callback_state);
10218 break;
10219 }
10220
10221 if (err)
10222 return err;
10223
10224 /* reset caller saved regs */
10225 for (i = 0; i < CALLER_SAVED_REGS; i++) {
10226 mark_reg_not_init(env, regs, caller_saved[i]);
10227 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
10228 }
10229
10230 /* helper call returns 64-bit value. */
10231 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
10232
10233 /* update return register (already marked as written above) */
10234 ret_type = fn->ret_type;
10235 ret_flag = type_flag(ret_type);
10236
10237 switch (base_type(ret_type)) {
10238 case RET_INTEGER:
10239 /* sets type to SCALAR_VALUE */
10240 mark_reg_unknown(env, regs, BPF_REG_0);
10241 break;
10242 case RET_VOID:
10243 regs[BPF_REG_0].type = NOT_INIT;
10244 break;
10245 case RET_PTR_TO_MAP_VALUE:
10246 /* There is no offset yet applied, variable or fixed */
10247 mark_reg_known_zero(env, regs, BPF_REG_0);
10248 /* remember map_ptr, so that check_map_access()
10249 * can check 'value_size' boundary of memory access
10250 * to map element returned from bpf_map_lookup_elem()
10251 */
10252 if (meta.map_ptr == NULL) {
10253 verbose(env,
10254 "kernel subsystem misconfigured verifier\n");
10255 return -EINVAL;
10256 }
10257 regs[BPF_REG_0].map_ptr = meta.map_ptr;
10258 regs[BPF_REG_0].map_uid = meta.map_uid;
10259 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
10260 if (!type_may_be_null(ret_type) &&
10261 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
10262 regs[BPF_REG_0].id = ++env->id_gen;
10263 }
10264 break;
10265 case RET_PTR_TO_SOCKET:
10266 mark_reg_known_zero(env, regs, BPF_REG_0);
10267 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
10268 break;
10269 case RET_PTR_TO_SOCK_COMMON:
10270 mark_reg_known_zero(env, regs, BPF_REG_0);
10271 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
10272 break;
10273 case RET_PTR_TO_TCP_SOCK:
10274 mark_reg_known_zero(env, regs, BPF_REG_0);
10275 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
10276 break;
10277 case RET_PTR_TO_MEM:
10278 mark_reg_known_zero(env, regs, BPF_REG_0);
10279 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
10280 regs[BPF_REG_0].mem_size = meta.mem_size;
10281 break;
10282 case RET_PTR_TO_MEM_OR_BTF_ID:
10283 {
10284 const struct btf_type *t;
10285
10286 mark_reg_known_zero(env, regs, BPF_REG_0);
10287 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
10288 if (!btf_type_is_struct(t)) {
10289 u32 tsize;
10290 const struct btf_type *ret;
10291 const char *tname;
10292
10293 /* resolve the type size of ksym. */
10294 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
10295 if (IS_ERR(ret)) {
10296 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
10297 verbose(env, "unable to resolve the size of type '%s': %ld\n",
10298 tname, PTR_ERR(ret));
10299 return -EINVAL;
10300 }
10301 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
10302 regs[BPF_REG_0].mem_size = tsize;
10303 } else {
10304 /* MEM_RDONLY may be carried from ret_flag, but it
10305 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
10306 * it will confuse the check of PTR_TO_BTF_ID in
10307 * check_mem_access().
10308 */
10309 ret_flag &= ~MEM_RDONLY;
10310
10311 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
10312 regs[BPF_REG_0].btf = meta.ret_btf;
10313 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
10314 }
10315 break;
10316 }
10317 case RET_PTR_TO_BTF_ID:
10318 {
10319 struct btf *ret_btf;
10320 int ret_btf_id;
10321
10322 mark_reg_known_zero(env, regs, BPF_REG_0);
10323 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
10324 if (func_id == BPF_FUNC_kptr_xchg) {
10325 ret_btf = meta.kptr_field->kptr.btf;
10326 ret_btf_id = meta.kptr_field->kptr.btf_id;
10327 if (!btf_is_kernel(ret_btf))
10328 regs[BPF_REG_0].type |= MEM_ALLOC;
10329 } else {
10330 if (fn->ret_btf_id == BPF_PTR_POISON) {
10331 verbose(env, "verifier internal error:");
10332 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
10333 func_id_name(func_id));
10334 return -EINVAL;
10335 }
10336 ret_btf = btf_vmlinux;
10337 ret_btf_id = *fn->ret_btf_id;
10338 }
10339 if (ret_btf_id == 0) {
10340 verbose(env, "invalid return type %u of func %s#%d\n",
10341 base_type(ret_type), func_id_name(func_id),
10342 func_id);
10343 return -EINVAL;
10344 }
10345 regs[BPF_REG_0].btf = ret_btf;
10346 regs[BPF_REG_0].btf_id = ret_btf_id;
10347 break;
10348 }
10349 default:
10350 verbose(env, "unknown return type %u of func %s#%d\n",
10351 base_type(ret_type), func_id_name(func_id), func_id);
10352 return -EINVAL;
10353 }
10354
10355 if (type_may_be_null(regs[BPF_REG_0].type))
10356 regs[BPF_REG_0].id = ++env->id_gen;
10357
10358 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
10359 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
10360 func_id_name(func_id), func_id);
10361 return -EFAULT;
10362 }
10363
10364 if (is_dynptr_ref_function(func_id))
10365 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
10366
10367 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
10368 /* For release_reference() */
10369 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
10370 } else if (is_acquire_function(func_id, meta.map_ptr)) {
10371 int id = acquire_reference_state(env, insn_idx);
10372
10373 if (id < 0)
10374 return id;
10375 /* For mark_ptr_or_null_reg() */
10376 regs[BPF_REG_0].id = id;
10377 /* For release_reference() */
10378 regs[BPF_REG_0].ref_obj_id = id;
10379 }
10380
10381 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
10382
10383 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
10384 if (err)
10385 return err;
10386
10387 if ((func_id == BPF_FUNC_get_stack ||
10388 func_id == BPF_FUNC_get_task_stack) &&
10389 !env->prog->has_callchain_buf) {
10390 const char *err_str;
10391
10392 #ifdef CONFIG_PERF_EVENTS
10393 err = get_callchain_buffers(sysctl_perf_event_max_stack);
10394 err_str = "cannot get callchain buffer for func %s#%d\n";
10395 #else
10396 err = -ENOTSUPP;
10397 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
10398 #endif
10399 if (err) {
10400 verbose(env, err_str, func_id_name(func_id), func_id);
10401 return err;
10402 }
10403
10404 env->prog->has_callchain_buf = true;
10405 }
10406
10407 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
10408 env->prog->call_get_stack = true;
10409
10410 if (func_id == BPF_FUNC_get_func_ip) {
10411 if (check_get_func_ip(env))
10412 return -ENOTSUPP;
10413 env->prog->call_get_func_ip = true;
10414 }
10415
10416 if (changes_data)
10417 clear_all_pkt_pointers(env);
10418 return 0;
10419 }
10420
10421 /* mark_btf_func_reg_size() is used when the reg size is determined by
10422 * the BTF func_proto's return value size and argument.
10423 */
mark_btf_func_reg_size(struct bpf_verifier_env * env,u32 regno,size_t reg_size)10424 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
10425 size_t reg_size)
10426 {
10427 struct bpf_reg_state *reg = &cur_regs(env)[regno];
10428
10429 if (regno == BPF_REG_0) {
10430 /* Function return value */
10431 reg->live |= REG_LIVE_WRITTEN;
10432 reg->subreg_def = reg_size == sizeof(u64) ?
10433 DEF_NOT_SUBREG : env->insn_idx + 1;
10434 } else {
10435 /* Function argument */
10436 if (reg_size == sizeof(u64)) {
10437 mark_insn_zext(env, reg);
10438 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
10439 } else {
10440 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
10441 }
10442 }
10443 }
10444
is_kfunc_acquire(struct bpf_kfunc_call_arg_meta * meta)10445 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
10446 {
10447 return meta->kfunc_flags & KF_ACQUIRE;
10448 }
10449
is_kfunc_release(struct bpf_kfunc_call_arg_meta * meta)10450 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
10451 {
10452 return meta->kfunc_flags & KF_RELEASE;
10453 }
10454
is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta * meta)10455 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
10456 {
10457 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
10458 }
10459
is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta * meta)10460 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
10461 {
10462 return meta->kfunc_flags & KF_SLEEPABLE;
10463 }
10464
is_kfunc_destructive(struct bpf_kfunc_call_arg_meta * meta)10465 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
10466 {
10467 return meta->kfunc_flags & KF_DESTRUCTIVE;
10468 }
10469
is_kfunc_rcu(struct bpf_kfunc_call_arg_meta * meta)10470 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
10471 {
10472 return meta->kfunc_flags & KF_RCU;
10473 }
10474
__kfunc_param_match_suffix(const struct btf * btf,const struct btf_param * arg,const char * suffix)10475 static bool __kfunc_param_match_suffix(const struct btf *btf,
10476 const struct btf_param *arg,
10477 const char *suffix)
10478 {
10479 int suffix_len = strlen(suffix), len;
10480 const char *param_name;
10481
10482 /* In the future, this can be ported to use BTF tagging */
10483 param_name = btf_name_by_offset(btf, arg->name_off);
10484 if (str_is_empty(param_name))
10485 return false;
10486 len = strlen(param_name);
10487 if (len < suffix_len)
10488 return false;
10489 param_name += len - suffix_len;
10490 return !strncmp(param_name, suffix, suffix_len);
10491 }
10492
is_kfunc_arg_mem_size(const struct btf * btf,const struct btf_param * arg,const struct bpf_reg_state * reg)10493 static bool is_kfunc_arg_mem_size(const struct btf *btf,
10494 const struct btf_param *arg,
10495 const struct bpf_reg_state *reg)
10496 {
10497 const struct btf_type *t;
10498
10499 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10500 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10501 return false;
10502
10503 return __kfunc_param_match_suffix(btf, arg, "__sz");
10504 }
10505
is_kfunc_arg_const_mem_size(const struct btf * btf,const struct btf_param * arg,const struct bpf_reg_state * reg)10506 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
10507 const struct btf_param *arg,
10508 const struct bpf_reg_state *reg)
10509 {
10510 const struct btf_type *t;
10511
10512 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10513 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10514 return false;
10515
10516 return __kfunc_param_match_suffix(btf, arg, "__szk");
10517 }
10518
is_kfunc_arg_optional(const struct btf * btf,const struct btf_param * arg)10519 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
10520 {
10521 return __kfunc_param_match_suffix(btf, arg, "__opt");
10522 }
10523
is_kfunc_arg_constant(const struct btf * btf,const struct btf_param * arg)10524 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
10525 {
10526 return __kfunc_param_match_suffix(btf, arg, "__k");
10527 }
10528
is_kfunc_arg_ignore(const struct btf * btf,const struct btf_param * arg)10529 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
10530 {
10531 return __kfunc_param_match_suffix(btf, arg, "__ign");
10532 }
10533
is_kfunc_arg_alloc_obj(const struct btf * btf,const struct btf_param * arg)10534 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
10535 {
10536 return __kfunc_param_match_suffix(btf, arg, "__alloc");
10537 }
10538
is_kfunc_arg_uninit(const struct btf * btf,const struct btf_param * arg)10539 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
10540 {
10541 return __kfunc_param_match_suffix(btf, arg, "__uninit");
10542 }
10543
is_kfunc_arg_refcounted_kptr(const struct btf * btf,const struct btf_param * arg)10544 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
10545 {
10546 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
10547 }
10548
is_kfunc_arg_scalar_with_name(const struct btf * btf,const struct btf_param * arg,const char * name)10549 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
10550 const struct btf_param *arg,
10551 const char *name)
10552 {
10553 int len, target_len = strlen(name);
10554 const char *param_name;
10555
10556 param_name = btf_name_by_offset(btf, arg->name_off);
10557 if (str_is_empty(param_name))
10558 return false;
10559 len = strlen(param_name);
10560 if (len != target_len)
10561 return false;
10562 if (strcmp(param_name, name))
10563 return false;
10564
10565 return true;
10566 }
10567
10568 enum {
10569 KF_ARG_DYNPTR_ID,
10570 KF_ARG_LIST_HEAD_ID,
10571 KF_ARG_LIST_NODE_ID,
10572 KF_ARG_RB_ROOT_ID,
10573 KF_ARG_RB_NODE_ID,
10574 };
10575
10576 BTF_ID_LIST(kf_arg_btf_ids)
BTF_ID(struct,bpf_dynptr_kern)10577 BTF_ID(struct, bpf_dynptr_kern)
10578 BTF_ID(struct, bpf_list_head)
10579 BTF_ID(struct, bpf_list_node)
10580 BTF_ID(struct, bpf_rb_root)
10581 BTF_ID(struct, bpf_rb_node)
10582
10583 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
10584 const struct btf_param *arg, int type)
10585 {
10586 const struct btf_type *t;
10587 u32 res_id;
10588
10589 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10590 if (!t)
10591 return false;
10592 if (!btf_type_is_ptr(t))
10593 return false;
10594 t = btf_type_skip_modifiers(btf, t->type, &res_id);
10595 if (!t)
10596 return false;
10597 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
10598 }
10599
is_kfunc_arg_dynptr(const struct btf * btf,const struct btf_param * arg)10600 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
10601 {
10602 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
10603 }
10604
is_kfunc_arg_list_head(const struct btf * btf,const struct btf_param * arg)10605 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
10606 {
10607 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
10608 }
10609
is_kfunc_arg_list_node(const struct btf * btf,const struct btf_param * arg)10610 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
10611 {
10612 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
10613 }
10614
is_kfunc_arg_rbtree_root(const struct btf * btf,const struct btf_param * arg)10615 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
10616 {
10617 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
10618 }
10619
is_kfunc_arg_rbtree_node(const struct btf * btf,const struct btf_param * arg)10620 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
10621 {
10622 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
10623 }
10624
is_kfunc_arg_callback(struct bpf_verifier_env * env,const struct btf * btf,const struct btf_param * arg)10625 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
10626 const struct btf_param *arg)
10627 {
10628 const struct btf_type *t;
10629
10630 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
10631 if (!t)
10632 return false;
10633
10634 return true;
10635 }
10636
10637 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
__btf_type_is_scalar_struct(struct bpf_verifier_env * env,const struct btf * btf,const struct btf_type * t,int rec)10638 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
10639 const struct btf *btf,
10640 const struct btf_type *t, int rec)
10641 {
10642 const struct btf_type *member_type;
10643 const struct btf_member *member;
10644 u32 i;
10645
10646 if (!btf_type_is_struct(t))
10647 return false;
10648
10649 for_each_member(i, t, member) {
10650 const struct btf_array *array;
10651
10652 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10653 if (btf_type_is_struct(member_type)) {
10654 if (rec >= 3) {
10655 verbose(env, "max struct nesting depth exceeded\n");
10656 return false;
10657 }
10658 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10659 return false;
10660 continue;
10661 }
10662 if (btf_type_is_array(member_type)) {
10663 array = btf_array(member_type);
10664 if (!array->nelems)
10665 return false;
10666 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10667 if (!btf_type_is_scalar(member_type))
10668 return false;
10669 continue;
10670 }
10671 if (!btf_type_is_scalar(member_type))
10672 return false;
10673 }
10674 return true;
10675 }
10676
10677 enum kfunc_ptr_arg_type {
10678 KF_ARG_PTR_TO_CTX,
10679 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10680 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10681 KF_ARG_PTR_TO_DYNPTR,
10682 KF_ARG_PTR_TO_ITER,
10683 KF_ARG_PTR_TO_LIST_HEAD,
10684 KF_ARG_PTR_TO_LIST_NODE,
10685 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10686 KF_ARG_PTR_TO_MEM,
10687 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10688 KF_ARG_PTR_TO_CALLBACK,
10689 KF_ARG_PTR_TO_RB_ROOT,
10690 KF_ARG_PTR_TO_RB_NODE,
10691 };
10692
10693 enum special_kfunc_type {
10694 KF_bpf_obj_new_impl,
10695 KF_bpf_obj_drop_impl,
10696 KF_bpf_refcount_acquire_impl,
10697 KF_bpf_list_push_front_impl,
10698 KF_bpf_list_push_back_impl,
10699 KF_bpf_list_pop_front,
10700 KF_bpf_list_pop_back,
10701 KF_bpf_cast_to_kern_ctx,
10702 KF_bpf_rdonly_cast,
10703 KF_bpf_rcu_read_lock,
10704 KF_bpf_rcu_read_unlock,
10705 KF_bpf_rbtree_remove,
10706 KF_bpf_rbtree_add_impl,
10707 KF_bpf_rbtree_first,
10708 KF_bpf_dynptr_from_skb,
10709 KF_bpf_dynptr_from_xdp,
10710 KF_bpf_dynptr_slice,
10711 KF_bpf_dynptr_slice_rdwr,
10712 KF_bpf_dynptr_clone,
10713 };
10714
10715 BTF_SET_START(special_kfunc_set)
BTF_ID(func,bpf_obj_new_impl)10716 BTF_ID(func, bpf_obj_new_impl)
10717 BTF_ID(func, bpf_obj_drop_impl)
10718 BTF_ID(func, bpf_refcount_acquire_impl)
10719 BTF_ID(func, bpf_list_push_front_impl)
10720 BTF_ID(func, bpf_list_push_back_impl)
10721 BTF_ID(func, bpf_list_pop_front)
10722 BTF_ID(func, bpf_list_pop_back)
10723 BTF_ID(func, bpf_cast_to_kern_ctx)
10724 BTF_ID(func, bpf_rdonly_cast)
10725 BTF_ID(func, bpf_rbtree_remove)
10726 BTF_ID(func, bpf_rbtree_add_impl)
10727 BTF_ID(func, bpf_rbtree_first)
10728 BTF_ID(func, bpf_dynptr_from_skb)
10729 BTF_ID(func, bpf_dynptr_from_xdp)
10730 BTF_ID(func, bpf_dynptr_slice)
10731 BTF_ID(func, bpf_dynptr_slice_rdwr)
10732 BTF_ID(func, bpf_dynptr_clone)
10733 BTF_SET_END(special_kfunc_set)
10734
10735 BTF_ID_LIST(special_kfunc_list)
10736 BTF_ID(func, bpf_obj_new_impl)
10737 BTF_ID(func, bpf_obj_drop_impl)
10738 BTF_ID(func, bpf_refcount_acquire_impl)
10739 BTF_ID(func, bpf_list_push_front_impl)
10740 BTF_ID(func, bpf_list_push_back_impl)
10741 BTF_ID(func, bpf_list_pop_front)
10742 BTF_ID(func, bpf_list_pop_back)
10743 BTF_ID(func, bpf_cast_to_kern_ctx)
10744 BTF_ID(func, bpf_rdonly_cast)
10745 BTF_ID(func, bpf_rcu_read_lock)
10746 BTF_ID(func, bpf_rcu_read_unlock)
10747 BTF_ID(func, bpf_rbtree_remove)
10748 BTF_ID(func, bpf_rbtree_add_impl)
10749 BTF_ID(func, bpf_rbtree_first)
10750 BTF_ID(func, bpf_dynptr_from_skb)
10751 BTF_ID(func, bpf_dynptr_from_xdp)
10752 BTF_ID(func, bpf_dynptr_slice)
10753 BTF_ID(func, bpf_dynptr_slice_rdwr)
10754 BTF_ID(func, bpf_dynptr_clone)
10755
10756 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10757 {
10758 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10759 meta->arg_owning_ref) {
10760 return false;
10761 }
10762
10763 return meta->kfunc_flags & KF_RET_NULL;
10764 }
10765
is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta * meta)10766 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10767 {
10768 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10769 }
10770
is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta * meta)10771 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10772 {
10773 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10774 }
10775
10776 static enum kfunc_ptr_arg_type
get_kfunc_ptr_arg_type(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,const struct btf_type * t,const struct btf_type * ref_t,const char * ref_tname,const struct btf_param * args,int argno,int nargs)10777 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10778 struct bpf_kfunc_call_arg_meta *meta,
10779 const struct btf_type *t, const struct btf_type *ref_t,
10780 const char *ref_tname, const struct btf_param *args,
10781 int argno, int nargs)
10782 {
10783 u32 regno = argno + 1;
10784 struct bpf_reg_state *regs = cur_regs(env);
10785 struct bpf_reg_state *reg = ®s[regno];
10786 bool arg_mem_size = false;
10787
10788 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10789 return KF_ARG_PTR_TO_CTX;
10790
10791 /* In this function, we verify the kfunc's BTF as per the argument type,
10792 * leaving the rest of the verification with respect to the register
10793 * type to our caller. When a set of conditions hold in the BTF type of
10794 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10795 */
10796 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10797 return KF_ARG_PTR_TO_CTX;
10798
10799 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10800 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10801
10802 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10803 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10804
10805 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10806 return KF_ARG_PTR_TO_DYNPTR;
10807
10808 if (is_kfunc_arg_iter(meta, argno))
10809 return KF_ARG_PTR_TO_ITER;
10810
10811 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10812 return KF_ARG_PTR_TO_LIST_HEAD;
10813
10814 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10815 return KF_ARG_PTR_TO_LIST_NODE;
10816
10817 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10818 return KF_ARG_PTR_TO_RB_ROOT;
10819
10820 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10821 return KF_ARG_PTR_TO_RB_NODE;
10822
10823 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10824 if (!btf_type_is_struct(ref_t)) {
10825 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10826 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10827 return -EINVAL;
10828 }
10829 return KF_ARG_PTR_TO_BTF_ID;
10830 }
10831
10832 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10833 return KF_ARG_PTR_TO_CALLBACK;
10834
10835
10836 if (argno + 1 < nargs &&
10837 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10838 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10839 arg_mem_size = true;
10840
10841 /* This is the catch all argument type of register types supported by
10842 * check_helper_mem_access. However, we only allow when argument type is
10843 * pointer to scalar, or struct composed (recursively) of scalars. When
10844 * arg_mem_size is true, the pointer can be void *.
10845 */
10846 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10847 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10848 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10849 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10850 return -EINVAL;
10851 }
10852 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10853 }
10854
process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const struct btf_type * ref_t,const char * ref_tname,u32 ref_id,struct bpf_kfunc_call_arg_meta * meta,int argno)10855 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10856 struct bpf_reg_state *reg,
10857 const struct btf_type *ref_t,
10858 const char *ref_tname, u32 ref_id,
10859 struct bpf_kfunc_call_arg_meta *meta,
10860 int argno)
10861 {
10862 const struct btf_type *reg_ref_t;
10863 bool strict_type_match = false;
10864 const struct btf *reg_btf;
10865 const char *reg_ref_tname;
10866 u32 reg_ref_id;
10867
10868 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10869 reg_btf = reg->btf;
10870 reg_ref_id = reg->btf_id;
10871 } else {
10872 reg_btf = btf_vmlinux;
10873 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10874 }
10875
10876 /* Enforce strict type matching for calls to kfuncs that are acquiring
10877 * or releasing a reference, or are no-cast aliases. We do _not_
10878 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10879 * as we want to enable BPF programs to pass types that are bitwise
10880 * equivalent without forcing them to explicitly cast with something
10881 * like bpf_cast_to_kern_ctx().
10882 *
10883 * For example, say we had a type like the following:
10884 *
10885 * struct bpf_cpumask {
10886 * cpumask_t cpumask;
10887 * refcount_t usage;
10888 * };
10889 *
10890 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10891 * to a struct cpumask, so it would be safe to pass a struct
10892 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10893 *
10894 * The philosophy here is similar to how we allow scalars of different
10895 * types to be passed to kfuncs as long as the size is the same. The
10896 * only difference here is that we're simply allowing
10897 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10898 * resolve types.
10899 */
10900 if (is_kfunc_acquire(meta) ||
10901 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10902 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10903 strict_type_match = true;
10904
10905 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10906
10907 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10908 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10909 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10910 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10911 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10912 btf_type_str(reg_ref_t), reg_ref_tname);
10913 return -EINVAL;
10914 }
10915 return 0;
10916 }
10917
ref_set_non_owning(struct bpf_verifier_env * env,struct bpf_reg_state * reg)10918 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10919 {
10920 struct bpf_verifier_state *state = env->cur_state;
10921 struct btf_record *rec = reg_btf_record(reg);
10922
10923 if (!state->active_lock.ptr) {
10924 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10925 return -EFAULT;
10926 }
10927
10928 if (type_flag(reg->type) & NON_OWN_REF) {
10929 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10930 return -EFAULT;
10931 }
10932
10933 reg->type |= NON_OWN_REF;
10934 if (rec->refcount_off >= 0)
10935 reg->type |= MEM_RCU;
10936
10937 return 0;
10938 }
10939
ref_convert_owning_non_owning(struct bpf_verifier_env * env,u32 ref_obj_id)10940 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10941 {
10942 struct bpf_func_state *state, *unused;
10943 struct bpf_reg_state *reg;
10944 int i;
10945
10946 state = cur_func(env);
10947
10948 if (!ref_obj_id) {
10949 verbose(env, "verifier internal error: ref_obj_id is zero for "
10950 "owning -> non-owning conversion\n");
10951 return -EFAULT;
10952 }
10953
10954 for (i = 0; i < state->acquired_refs; i++) {
10955 if (state->refs[i].id != ref_obj_id)
10956 continue;
10957
10958 /* Clear ref_obj_id here so release_reference doesn't clobber
10959 * the whole reg
10960 */
10961 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10962 if (reg->ref_obj_id == ref_obj_id) {
10963 reg->ref_obj_id = 0;
10964 ref_set_non_owning(env, reg);
10965 }
10966 }));
10967 return 0;
10968 }
10969
10970 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10971 return -EFAULT;
10972 }
10973
10974 /* Implementation details:
10975 *
10976 * Each register points to some region of memory, which we define as an
10977 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10978 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10979 * allocation. The lock and the data it protects are colocated in the same
10980 * memory region.
10981 *
10982 * Hence, everytime a register holds a pointer value pointing to such
10983 * allocation, the verifier preserves a unique reg->id for it.
10984 *
10985 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10986 * bpf_spin_lock is called.
10987 *
10988 * To enable this, lock state in the verifier captures two values:
10989 * active_lock.ptr = Register's type specific pointer
10990 * active_lock.id = A unique ID for each register pointer value
10991 *
10992 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10993 * supported register types.
10994 *
10995 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10996 * allocated objects is the reg->btf pointer.
10997 *
10998 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10999 * can establish the provenance of the map value statically for each distinct
11000 * lookup into such maps. They always contain a single map value hence unique
11001 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
11002 *
11003 * So, in case of global variables, they use array maps with max_entries = 1,
11004 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
11005 * into the same map value as max_entries is 1, as described above).
11006 *
11007 * In case of inner map lookups, the inner map pointer has same map_ptr as the
11008 * outer map pointer (in verifier context), but each lookup into an inner map
11009 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
11010 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
11011 * will get different reg->id assigned to each lookup, hence different
11012 * active_lock.id.
11013 *
11014 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
11015 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
11016 * returned from bpf_obj_new. Each allocation receives a new reg->id.
11017 */
check_reg_allocation_locked(struct bpf_verifier_env * env,struct bpf_reg_state * reg)11018 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
11019 {
11020 void *ptr;
11021 u32 id;
11022
11023 switch ((int)reg->type) {
11024 case PTR_TO_MAP_VALUE:
11025 ptr = reg->map_ptr;
11026 break;
11027 case PTR_TO_BTF_ID | MEM_ALLOC:
11028 ptr = reg->btf;
11029 break;
11030 default:
11031 verbose(env, "verifier internal error: unknown reg type for lock check\n");
11032 return -EFAULT;
11033 }
11034 id = reg->id;
11035
11036 if (!env->cur_state->active_lock.ptr)
11037 return -EINVAL;
11038 if (env->cur_state->active_lock.ptr != ptr ||
11039 env->cur_state->active_lock.id != id) {
11040 verbose(env, "held lock and object are not in the same allocation\n");
11041 return -EINVAL;
11042 }
11043 return 0;
11044 }
11045
is_bpf_list_api_kfunc(u32 btf_id)11046 static bool is_bpf_list_api_kfunc(u32 btf_id)
11047 {
11048 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11049 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11050 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11051 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
11052 }
11053
is_bpf_rbtree_api_kfunc(u32 btf_id)11054 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
11055 {
11056 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
11057 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11058 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
11059 }
11060
is_bpf_graph_api_kfunc(u32 btf_id)11061 static bool is_bpf_graph_api_kfunc(u32 btf_id)
11062 {
11063 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
11064 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
11065 }
11066
is_sync_callback_calling_kfunc(u32 btf_id)11067 static bool is_sync_callback_calling_kfunc(u32 btf_id)
11068 {
11069 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
11070 }
11071
is_rbtree_lock_required_kfunc(u32 btf_id)11072 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
11073 {
11074 return is_bpf_rbtree_api_kfunc(btf_id);
11075 }
11076
check_kfunc_is_graph_root_api(struct bpf_verifier_env * env,enum btf_field_type head_field_type,u32 kfunc_btf_id)11077 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
11078 enum btf_field_type head_field_type,
11079 u32 kfunc_btf_id)
11080 {
11081 bool ret;
11082
11083 switch (head_field_type) {
11084 case BPF_LIST_HEAD:
11085 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
11086 break;
11087 case BPF_RB_ROOT:
11088 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
11089 break;
11090 default:
11091 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
11092 btf_field_type_name(head_field_type));
11093 return false;
11094 }
11095
11096 if (!ret)
11097 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
11098 btf_field_type_name(head_field_type));
11099 return ret;
11100 }
11101
check_kfunc_is_graph_node_api(struct bpf_verifier_env * env,enum btf_field_type node_field_type,u32 kfunc_btf_id)11102 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
11103 enum btf_field_type node_field_type,
11104 u32 kfunc_btf_id)
11105 {
11106 bool ret;
11107
11108 switch (node_field_type) {
11109 case BPF_LIST_NODE:
11110 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11111 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
11112 break;
11113 case BPF_RB_NODE:
11114 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11115 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
11116 break;
11117 default:
11118 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
11119 btf_field_type_name(node_field_type));
11120 return false;
11121 }
11122
11123 if (!ret)
11124 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
11125 btf_field_type_name(node_field_type));
11126 return ret;
11127 }
11128
11129 static int
__process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta,enum btf_field_type head_field_type,struct btf_field ** head_field)11130 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
11131 struct bpf_reg_state *reg, u32 regno,
11132 struct bpf_kfunc_call_arg_meta *meta,
11133 enum btf_field_type head_field_type,
11134 struct btf_field **head_field)
11135 {
11136 const char *head_type_name;
11137 struct btf_field *field;
11138 struct btf_record *rec;
11139 u32 head_off;
11140
11141 if (meta->btf != btf_vmlinux) {
11142 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
11143 return -EFAULT;
11144 }
11145
11146 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
11147 return -EFAULT;
11148
11149 head_type_name = btf_field_type_name(head_field_type);
11150 if (!tnum_is_const(reg->var_off)) {
11151 verbose(env,
11152 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
11153 regno, head_type_name);
11154 return -EINVAL;
11155 }
11156
11157 rec = reg_btf_record(reg);
11158 head_off = reg->off + reg->var_off.value;
11159 field = btf_record_find(rec, head_off, head_field_type);
11160 if (!field) {
11161 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
11162 return -EINVAL;
11163 }
11164
11165 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
11166 if (check_reg_allocation_locked(env, reg)) {
11167 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
11168 rec->spin_lock_off, head_type_name);
11169 return -EINVAL;
11170 }
11171
11172 if (*head_field) {
11173 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
11174 return -EFAULT;
11175 }
11176 *head_field = field;
11177 return 0;
11178 }
11179
process_kf_arg_ptr_to_list_head(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)11180 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
11181 struct bpf_reg_state *reg, u32 regno,
11182 struct bpf_kfunc_call_arg_meta *meta)
11183 {
11184 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
11185 &meta->arg_list_head.field);
11186 }
11187
process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)11188 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
11189 struct bpf_reg_state *reg, u32 regno,
11190 struct bpf_kfunc_call_arg_meta *meta)
11191 {
11192 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
11193 &meta->arg_rbtree_root.field);
11194 }
11195
11196 static int
__process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta,enum btf_field_type head_field_type,enum btf_field_type node_field_type,struct btf_field ** node_field)11197 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
11198 struct bpf_reg_state *reg, u32 regno,
11199 struct bpf_kfunc_call_arg_meta *meta,
11200 enum btf_field_type head_field_type,
11201 enum btf_field_type node_field_type,
11202 struct btf_field **node_field)
11203 {
11204 const char *node_type_name;
11205 const struct btf_type *et, *t;
11206 struct btf_field *field;
11207 u32 node_off;
11208
11209 if (meta->btf != btf_vmlinux) {
11210 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
11211 return -EFAULT;
11212 }
11213
11214 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
11215 return -EFAULT;
11216
11217 node_type_name = btf_field_type_name(node_field_type);
11218 if (!tnum_is_const(reg->var_off)) {
11219 verbose(env,
11220 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
11221 regno, node_type_name);
11222 return -EINVAL;
11223 }
11224
11225 node_off = reg->off + reg->var_off.value;
11226 field = reg_find_field_offset(reg, node_off, node_field_type);
11227 if (!field || field->offset != node_off) {
11228 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
11229 return -EINVAL;
11230 }
11231
11232 field = *node_field;
11233
11234 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
11235 t = btf_type_by_id(reg->btf, reg->btf_id);
11236 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
11237 field->graph_root.value_btf_id, true)) {
11238 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
11239 "in struct %s, but arg is at offset=%d in struct %s\n",
11240 btf_field_type_name(head_field_type),
11241 btf_field_type_name(node_field_type),
11242 field->graph_root.node_offset,
11243 btf_name_by_offset(field->graph_root.btf, et->name_off),
11244 node_off, btf_name_by_offset(reg->btf, t->name_off));
11245 return -EINVAL;
11246 }
11247 meta->arg_btf = reg->btf;
11248 meta->arg_btf_id = reg->btf_id;
11249
11250 if (node_off != field->graph_root.node_offset) {
11251 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
11252 node_off, btf_field_type_name(node_field_type),
11253 field->graph_root.node_offset,
11254 btf_name_by_offset(field->graph_root.btf, et->name_off));
11255 return -EINVAL;
11256 }
11257
11258 return 0;
11259 }
11260
process_kf_arg_ptr_to_list_node(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)11261 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
11262 struct bpf_reg_state *reg, u32 regno,
11263 struct bpf_kfunc_call_arg_meta *meta)
11264 {
11265 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
11266 BPF_LIST_HEAD, BPF_LIST_NODE,
11267 &meta->arg_list_head.field);
11268 }
11269
process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)11270 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
11271 struct bpf_reg_state *reg, u32 regno,
11272 struct bpf_kfunc_call_arg_meta *meta)
11273 {
11274 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
11275 BPF_RB_ROOT, BPF_RB_NODE,
11276 &meta->arg_rbtree_root.field);
11277 }
11278
check_kfunc_args(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,int insn_idx)11279 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
11280 int insn_idx)
11281 {
11282 const char *func_name = meta->func_name, *ref_tname;
11283 const struct btf *btf = meta->btf;
11284 const struct btf_param *args;
11285 struct btf_record *rec;
11286 u32 i, nargs;
11287 int ret;
11288
11289 args = (const struct btf_param *)(meta->func_proto + 1);
11290 nargs = btf_type_vlen(meta->func_proto);
11291 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
11292 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
11293 MAX_BPF_FUNC_REG_ARGS);
11294 return -EINVAL;
11295 }
11296
11297 /* Check that BTF function arguments match actual types that the
11298 * verifier sees.
11299 */
11300 for (i = 0; i < nargs; i++) {
11301 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
11302 const struct btf_type *t, *ref_t, *resolve_ret;
11303 enum bpf_arg_type arg_type = ARG_DONTCARE;
11304 u32 regno = i + 1, ref_id, type_size;
11305 bool is_ret_buf_sz = false;
11306 int kf_arg_type;
11307
11308 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
11309
11310 if (is_kfunc_arg_ignore(btf, &args[i]))
11311 continue;
11312
11313 if (btf_type_is_scalar(t)) {
11314 if (reg->type != SCALAR_VALUE) {
11315 verbose(env, "R%d is not a scalar\n", regno);
11316 return -EINVAL;
11317 }
11318
11319 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
11320 if (meta->arg_constant.found) {
11321 verbose(env, "verifier internal error: only one constant argument permitted\n");
11322 return -EFAULT;
11323 }
11324 if (!tnum_is_const(reg->var_off)) {
11325 verbose(env, "R%d must be a known constant\n", regno);
11326 return -EINVAL;
11327 }
11328 ret = mark_chain_precision(env, regno);
11329 if (ret < 0)
11330 return ret;
11331 meta->arg_constant.found = true;
11332 meta->arg_constant.value = reg->var_off.value;
11333 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
11334 meta->r0_rdonly = true;
11335 is_ret_buf_sz = true;
11336 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
11337 is_ret_buf_sz = true;
11338 }
11339
11340 if (is_ret_buf_sz) {
11341 if (meta->r0_size) {
11342 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
11343 return -EINVAL;
11344 }
11345
11346 if (!tnum_is_const(reg->var_off)) {
11347 verbose(env, "R%d is not a const\n", regno);
11348 return -EINVAL;
11349 }
11350
11351 meta->r0_size = reg->var_off.value;
11352 ret = mark_chain_precision(env, regno);
11353 if (ret)
11354 return ret;
11355 }
11356 continue;
11357 }
11358
11359 if (!btf_type_is_ptr(t)) {
11360 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
11361 return -EINVAL;
11362 }
11363
11364 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
11365 (register_is_null(reg) || type_may_be_null(reg->type))) {
11366 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
11367 return -EACCES;
11368 }
11369
11370 if (reg->ref_obj_id) {
11371 if (is_kfunc_release(meta) && meta->ref_obj_id) {
11372 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
11373 regno, reg->ref_obj_id,
11374 meta->ref_obj_id);
11375 return -EFAULT;
11376 }
11377 meta->ref_obj_id = reg->ref_obj_id;
11378 if (is_kfunc_release(meta))
11379 meta->release_regno = regno;
11380 }
11381
11382 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
11383 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
11384
11385 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
11386 if (kf_arg_type < 0)
11387 return kf_arg_type;
11388
11389 switch (kf_arg_type) {
11390 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11391 case KF_ARG_PTR_TO_BTF_ID:
11392 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
11393 break;
11394
11395 if (!is_trusted_reg(reg)) {
11396 if (!is_kfunc_rcu(meta)) {
11397 verbose(env, "R%d must be referenced or trusted\n", regno);
11398 return -EINVAL;
11399 }
11400 if (!is_rcu_reg(reg)) {
11401 verbose(env, "R%d must be a rcu pointer\n", regno);
11402 return -EINVAL;
11403 }
11404 }
11405
11406 fallthrough;
11407 case KF_ARG_PTR_TO_CTX:
11408 /* Trusted arguments have the same offset checks as release arguments */
11409 arg_type |= OBJ_RELEASE;
11410 break;
11411 case KF_ARG_PTR_TO_DYNPTR:
11412 case KF_ARG_PTR_TO_ITER:
11413 case KF_ARG_PTR_TO_LIST_HEAD:
11414 case KF_ARG_PTR_TO_LIST_NODE:
11415 case KF_ARG_PTR_TO_RB_ROOT:
11416 case KF_ARG_PTR_TO_RB_NODE:
11417 case KF_ARG_PTR_TO_MEM:
11418 case KF_ARG_PTR_TO_MEM_SIZE:
11419 case KF_ARG_PTR_TO_CALLBACK:
11420 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11421 /* Trusted by default */
11422 break;
11423 default:
11424 WARN_ON_ONCE(1);
11425 return -EFAULT;
11426 }
11427
11428 if (is_kfunc_release(meta) && reg->ref_obj_id)
11429 arg_type |= OBJ_RELEASE;
11430 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
11431 if (ret < 0)
11432 return ret;
11433
11434 switch (kf_arg_type) {
11435 case KF_ARG_PTR_TO_CTX:
11436 if (reg->type != PTR_TO_CTX) {
11437 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
11438 return -EINVAL;
11439 }
11440
11441 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11442 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
11443 if (ret < 0)
11444 return -EINVAL;
11445 meta->ret_btf_id = ret;
11446 }
11447 break;
11448 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11449 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11450 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11451 return -EINVAL;
11452 }
11453 if (!reg->ref_obj_id) {
11454 verbose(env, "allocated object must be referenced\n");
11455 return -EINVAL;
11456 }
11457 if (meta->btf == btf_vmlinux &&
11458 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11459 meta->arg_btf = reg->btf;
11460 meta->arg_btf_id = reg->btf_id;
11461 }
11462 break;
11463 case KF_ARG_PTR_TO_DYNPTR:
11464 {
11465 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
11466 int clone_ref_obj_id = 0;
11467
11468 if (reg->type != PTR_TO_STACK &&
11469 reg->type != CONST_PTR_TO_DYNPTR) {
11470 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
11471 return -EINVAL;
11472 }
11473
11474 if (reg->type == CONST_PTR_TO_DYNPTR)
11475 dynptr_arg_type |= MEM_RDONLY;
11476
11477 if (is_kfunc_arg_uninit(btf, &args[i]))
11478 dynptr_arg_type |= MEM_UNINIT;
11479
11480 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
11481 dynptr_arg_type |= DYNPTR_TYPE_SKB;
11482 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
11483 dynptr_arg_type |= DYNPTR_TYPE_XDP;
11484 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
11485 (dynptr_arg_type & MEM_UNINIT)) {
11486 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
11487
11488 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
11489 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
11490 return -EFAULT;
11491 }
11492
11493 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
11494 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
11495 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
11496 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
11497 return -EFAULT;
11498 }
11499 }
11500
11501 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
11502 if (ret < 0)
11503 return ret;
11504
11505 if (!(dynptr_arg_type & MEM_UNINIT)) {
11506 int id = dynptr_id(env, reg);
11507
11508 if (id < 0) {
11509 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
11510 return id;
11511 }
11512 meta->initialized_dynptr.id = id;
11513 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
11514 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
11515 }
11516
11517 break;
11518 }
11519 case KF_ARG_PTR_TO_ITER:
11520 ret = process_iter_arg(env, regno, insn_idx, meta);
11521 if (ret < 0)
11522 return ret;
11523 break;
11524 case KF_ARG_PTR_TO_LIST_HEAD:
11525 if (reg->type != PTR_TO_MAP_VALUE &&
11526 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11527 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11528 return -EINVAL;
11529 }
11530 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11531 verbose(env, "allocated object must be referenced\n");
11532 return -EINVAL;
11533 }
11534 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
11535 if (ret < 0)
11536 return ret;
11537 break;
11538 case KF_ARG_PTR_TO_RB_ROOT:
11539 if (reg->type != PTR_TO_MAP_VALUE &&
11540 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11541 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11542 return -EINVAL;
11543 }
11544 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11545 verbose(env, "allocated object must be referenced\n");
11546 return -EINVAL;
11547 }
11548 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
11549 if (ret < 0)
11550 return ret;
11551 break;
11552 case KF_ARG_PTR_TO_LIST_NODE:
11553 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11554 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11555 return -EINVAL;
11556 }
11557 if (!reg->ref_obj_id) {
11558 verbose(env, "allocated object must be referenced\n");
11559 return -EINVAL;
11560 }
11561 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
11562 if (ret < 0)
11563 return ret;
11564 break;
11565 case KF_ARG_PTR_TO_RB_NODE:
11566 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
11567 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
11568 verbose(env, "rbtree_remove node input must be non-owning ref\n");
11569 return -EINVAL;
11570 }
11571 if (in_rbtree_lock_required_cb(env)) {
11572 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
11573 return -EINVAL;
11574 }
11575 } else {
11576 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11577 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11578 return -EINVAL;
11579 }
11580 if (!reg->ref_obj_id) {
11581 verbose(env, "allocated object must be referenced\n");
11582 return -EINVAL;
11583 }
11584 }
11585
11586 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
11587 if (ret < 0)
11588 return ret;
11589 break;
11590 case KF_ARG_PTR_TO_BTF_ID:
11591 /* Only base_type is checked, further checks are done here */
11592 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
11593 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
11594 !reg2btf_ids[base_type(reg->type)]) {
11595 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
11596 verbose(env, "expected %s or socket\n",
11597 reg_type_str(env, base_type(reg->type) |
11598 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
11599 return -EINVAL;
11600 }
11601 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
11602 if (ret < 0)
11603 return ret;
11604 break;
11605 case KF_ARG_PTR_TO_MEM:
11606 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
11607 if (IS_ERR(resolve_ret)) {
11608 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
11609 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
11610 return -EINVAL;
11611 }
11612 ret = check_mem_reg(env, reg, regno, type_size);
11613 if (ret < 0)
11614 return ret;
11615 break;
11616 case KF_ARG_PTR_TO_MEM_SIZE:
11617 {
11618 struct bpf_reg_state *buff_reg = ®s[regno];
11619 const struct btf_param *buff_arg = &args[i];
11620 struct bpf_reg_state *size_reg = ®s[regno + 1];
11621 const struct btf_param *size_arg = &args[i + 1];
11622
11623 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
11624 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
11625 if (ret < 0) {
11626 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
11627 return ret;
11628 }
11629 }
11630
11631 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
11632 if (meta->arg_constant.found) {
11633 verbose(env, "verifier internal error: only one constant argument permitted\n");
11634 return -EFAULT;
11635 }
11636 if (!tnum_is_const(size_reg->var_off)) {
11637 verbose(env, "R%d must be a known constant\n", regno + 1);
11638 return -EINVAL;
11639 }
11640 meta->arg_constant.found = true;
11641 meta->arg_constant.value = size_reg->var_off.value;
11642 }
11643
11644 /* Skip next '__sz' or '__szk' argument */
11645 i++;
11646 break;
11647 }
11648 case KF_ARG_PTR_TO_CALLBACK:
11649 if (reg->type != PTR_TO_FUNC) {
11650 verbose(env, "arg%d expected pointer to func\n", i);
11651 return -EINVAL;
11652 }
11653 meta->subprogno = reg->subprogno;
11654 break;
11655 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11656 if (!type_is_ptr_alloc_obj(reg->type)) {
11657 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11658 return -EINVAL;
11659 }
11660 if (!type_is_non_owning_ref(reg->type))
11661 meta->arg_owning_ref = true;
11662
11663 rec = reg_btf_record(reg);
11664 if (!rec) {
11665 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11666 return -EFAULT;
11667 }
11668
11669 if (rec->refcount_off < 0) {
11670 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11671 return -EINVAL;
11672 }
11673
11674 meta->arg_btf = reg->btf;
11675 meta->arg_btf_id = reg->btf_id;
11676 break;
11677 }
11678 }
11679
11680 if (is_kfunc_release(meta) && !meta->release_regno) {
11681 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11682 func_name);
11683 return -EINVAL;
11684 }
11685
11686 return 0;
11687 }
11688
fetch_kfunc_meta(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_kfunc_call_arg_meta * meta,const char ** kfunc_name)11689 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11690 struct bpf_insn *insn,
11691 struct bpf_kfunc_call_arg_meta *meta,
11692 const char **kfunc_name)
11693 {
11694 const struct btf_type *func, *func_proto;
11695 u32 func_id, *kfunc_flags;
11696 const char *func_name;
11697 struct btf *desc_btf;
11698
11699 if (kfunc_name)
11700 *kfunc_name = NULL;
11701
11702 if (!insn->imm)
11703 return -EINVAL;
11704
11705 desc_btf = find_kfunc_desc_btf(env, insn->off);
11706 if (IS_ERR(desc_btf))
11707 return PTR_ERR(desc_btf);
11708
11709 func_id = insn->imm;
11710 func = btf_type_by_id(desc_btf, func_id);
11711 func_name = btf_name_by_offset(desc_btf, func->name_off);
11712 if (kfunc_name)
11713 *kfunc_name = func_name;
11714 func_proto = btf_type_by_id(desc_btf, func->type);
11715
11716 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11717 if (!kfunc_flags) {
11718 return -EACCES;
11719 }
11720
11721 memset(meta, 0, sizeof(*meta));
11722 meta->btf = desc_btf;
11723 meta->func_id = func_id;
11724 meta->kfunc_flags = *kfunc_flags;
11725 meta->func_proto = func_proto;
11726 meta->func_name = func_name;
11727
11728 return 0;
11729 }
11730
check_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx_p)11731 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11732 int *insn_idx_p)
11733 {
11734 const struct btf_type *t, *ptr_type;
11735 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11736 struct bpf_reg_state *regs = cur_regs(env);
11737 const char *func_name, *ptr_type_name;
11738 bool sleepable, rcu_lock, rcu_unlock;
11739 struct bpf_kfunc_call_arg_meta meta;
11740 struct bpf_insn_aux_data *insn_aux;
11741 int err, insn_idx = *insn_idx_p;
11742 const struct btf_param *args;
11743 const struct btf_type *ret_t;
11744 struct btf *desc_btf;
11745
11746 /* skip for now, but return error when we find this in fixup_kfunc_call */
11747 if (!insn->imm)
11748 return 0;
11749
11750 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11751 if (err == -EACCES && func_name)
11752 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11753 if (err)
11754 return err;
11755 desc_btf = meta.btf;
11756 insn_aux = &env->insn_aux_data[insn_idx];
11757
11758 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11759
11760 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11761 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11762 return -EACCES;
11763 }
11764
11765 sleepable = is_kfunc_sleepable(&meta);
11766 if (sleepable && !env->prog->aux->sleepable) {
11767 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11768 return -EACCES;
11769 }
11770
11771 /* Check the arguments */
11772 err = check_kfunc_args(env, &meta, insn_idx);
11773 if (err < 0)
11774 return err;
11775
11776 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11777 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
11778 set_rbtree_add_callback_state);
11779 if (err) {
11780 verbose(env, "kfunc %s#%d failed callback verification\n",
11781 func_name, meta.func_id);
11782 return err;
11783 }
11784 }
11785
11786 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11787 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11788
11789 if (env->cur_state->active_rcu_lock) {
11790 struct bpf_func_state *state;
11791 struct bpf_reg_state *reg;
11792
11793 if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
11794 verbose(env, "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
11795 return -EACCES;
11796 }
11797
11798 if (rcu_lock) {
11799 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11800 return -EINVAL;
11801 } else if (rcu_unlock) {
11802 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11803 if (reg->type & MEM_RCU) {
11804 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11805 reg->type |= PTR_UNTRUSTED;
11806 }
11807 }));
11808 env->cur_state->active_rcu_lock = false;
11809 } else if (sleepable) {
11810 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11811 return -EACCES;
11812 }
11813 } else if (rcu_lock) {
11814 env->cur_state->active_rcu_lock = true;
11815 } else if (rcu_unlock) {
11816 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11817 return -EINVAL;
11818 }
11819
11820 /* In case of release function, we get register number of refcounted
11821 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11822 */
11823 if (meta.release_regno) {
11824 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11825 if (err) {
11826 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11827 func_name, meta.func_id);
11828 return err;
11829 }
11830 }
11831
11832 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11833 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11834 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11835 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11836 insn_aux->insert_off = regs[BPF_REG_2].off;
11837 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11838 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11839 if (err) {
11840 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11841 func_name, meta.func_id);
11842 return err;
11843 }
11844
11845 err = release_reference(env, release_ref_obj_id);
11846 if (err) {
11847 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11848 func_name, meta.func_id);
11849 return err;
11850 }
11851 }
11852
11853 for (i = 0; i < CALLER_SAVED_REGS; i++)
11854 mark_reg_not_init(env, regs, caller_saved[i]);
11855
11856 /* Check return type */
11857 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11858
11859 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11860 /* Only exception is bpf_obj_new_impl */
11861 if (meta.btf != btf_vmlinux ||
11862 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11863 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11864 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11865 return -EINVAL;
11866 }
11867 }
11868
11869 if (btf_type_is_scalar(t)) {
11870 mark_reg_unknown(env, regs, BPF_REG_0);
11871 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11872 } else if (btf_type_is_ptr(t)) {
11873 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11874
11875 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11876 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11877 struct btf *ret_btf;
11878 u32 ret_btf_id;
11879
11880 if (unlikely(!bpf_global_ma_set))
11881 return -ENOMEM;
11882
11883 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11884 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11885 return -EINVAL;
11886 }
11887
11888 ret_btf = env->prog->aux->btf;
11889 ret_btf_id = meta.arg_constant.value;
11890
11891 /* This may be NULL due to user not supplying a BTF */
11892 if (!ret_btf) {
11893 verbose(env, "bpf_obj_new requires prog BTF\n");
11894 return -EINVAL;
11895 }
11896
11897 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11898 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11899 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11900 return -EINVAL;
11901 }
11902
11903 mark_reg_known_zero(env, regs, BPF_REG_0);
11904 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11905 regs[BPF_REG_0].btf = ret_btf;
11906 regs[BPF_REG_0].btf_id = ret_btf_id;
11907
11908 insn_aux->obj_new_size = ret_t->size;
11909 insn_aux->kptr_struct_meta =
11910 btf_find_struct_meta(ret_btf, ret_btf_id);
11911 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11912 mark_reg_known_zero(env, regs, BPF_REG_0);
11913 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11914 regs[BPF_REG_0].btf = meta.arg_btf;
11915 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11916
11917 insn_aux->kptr_struct_meta =
11918 btf_find_struct_meta(meta.arg_btf,
11919 meta.arg_btf_id);
11920 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11921 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11922 struct btf_field *field = meta.arg_list_head.field;
11923
11924 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11925 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11926 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11927 struct btf_field *field = meta.arg_rbtree_root.field;
11928
11929 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11930 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11931 mark_reg_known_zero(env, regs, BPF_REG_0);
11932 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11933 regs[BPF_REG_0].btf = desc_btf;
11934 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11935 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11936 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11937 if (!ret_t || !btf_type_is_struct(ret_t)) {
11938 verbose(env,
11939 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11940 return -EINVAL;
11941 }
11942
11943 mark_reg_known_zero(env, regs, BPF_REG_0);
11944 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11945 regs[BPF_REG_0].btf = desc_btf;
11946 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11947 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11948 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11949 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11950
11951 mark_reg_known_zero(env, regs, BPF_REG_0);
11952
11953 if (!meta.arg_constant.found) {
11954 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11955 return -EFAULT;
11956 }
11957
11958 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11959
11960 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11961 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11962
11963 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11964 regs[BPF_REG_0].type |= MEM_RDONLY;
11965 } else {
11966 /* this will set env->seen_direct_write to true */
11967 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11968 verbose(env, "the prog does not allow writes to packet data\n");
11969 return -EINVAL;
11970 }
11971 }
11972
11973 if (!meta.initialized_dynptr.id) {
11974 verbose(env, "verifier internal error: no dynptr id\n");
11975 return -EFAULT;
11976 }
11977 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11978
11979 /* we don't need to set BPF_REG_0's ref obj id
11980 * because packet slices are not refcounted (see
11981 * dynptr_type_refcounted)
11982 */
11983 } else {
11984 verbose(env, "kernel function %s unhandled dynamic return type\n",
11985 meta.func_name);
11986 return -EFAULT;
11987 }
11988 } else if (!__btf_type_is_struct(ptr_type)) {
11989 if (!meta.r0_size) {
11990 __u32 sz;
11991
11992 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11993 meta.r0_size = sz;
11994 meta.r0_rdonly = true;
11995 }
11996 }
11997 if (!meta.r0_size) {
11998 ptr_type_name = btf_name_by_offset(desc_btf,
11999 ptr_type->name_off);
12000 verbose(env,
12001 "kernel function %s returns pointer type %s %s is not supported\n",
12002 func_name,
12003 btf_type_str(ptr_type),
12004 ptr_type_name);
12005 return -EINVAL;
12006 }
12007
12008 mark_reg_known_zero(env, regs, BPF_REG_0);
12009 regs[BPF_REG_0].type = PTR_TO_MEM;
12010 regs[BPF_REG_0].mem_size = meta.r0_size;
12011
12012 if (meta.r0_rdonly)
12013 regs[BPF_REG_0].type |= MEM_RDONLY;
12014
12015 /* Ensures we don't access the memory after a release_reference() */
12016 if (meta.ref_obj_id)
12017 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
12018 } else {
12019 mark_reg_known_zero(env, regs, BPF_REG_0);
12020 regs[BPF_REG_0].btf = desc_btf;
12021 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
12022 regs[BPF_REG_0].btf_id = ptr_type_id;
12023 }
12024
12025 if (is_kfunc_ret_null(&meta)) {
12026 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
12027 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
12028 regs[BPF_REG_0].id = ++env->id_gen;
12029 }
12030 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
12031 if (is_kfunc_acquire(&meta)) {
12032 int id = acquire_reference_state(env, insn_idx);
12033
12034 if (id < 0)
12035 return id;
12036 if (is_kfunc_ret_null(&meta))
12037 regs[BPF_REG_0].id = id;
12038 regs[BPF_REG_0].ref_obj_id = id;
12039 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
12040 ref_set_non_owning(env, ®s[BPF_REG_0]);
12041 }
12042
12043 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
12044 regs[BPF_REG_0].id = ++env->id_gen;
12045 } else if (btf_type_is_void(t)) {
12046 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
12047 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
12048 insn_aux->kptr_struct_meta =
12049 btf_find_struct_meta(meta.arg_btf,
12050 meta.arg_btf_id);
12051 }
12052 }
12053 }
12054
12055 nargs = btf_type_vlen(meta.func_proto);
12056 args = (const struct btf_param *)(meta.func_proto + 1);
12057 for (i = 0; i < nargs; i++) {
12058 u32 regno = i + 1;
12059
12060 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
12061 if (btf_type_is_ptr(t))
12062 mark_btf_func_reg_size(env, regno, sizeof(void *));
12063 else
12064 /* scalar. ensured by btf_check_kfunc_arg_match() */
12065 mark_btf_func_reg_size(env, regno, t->size);
12066 }
12067
12068 if (is_iter_next_kfunc(&meta)) {
12069 err = process_iter_next_call(env, insn_idx, &meta);
12070 if (err)
12071 return err;
12072 }
12073
12074 return 0;
12075 }
12076
signed_add_overflows(s64 a,s64 b)12077 static bool signed_add_overflows(s64 a, s64 b)
12078 {
12079 /* Do the add in u64, where overflow is well-defined */
12080 s64 res = (s64)((u64)a + (u64)b);
12081
12082 if (b < 0)
12083 return res > a;
12084 return res < a;
12085 }
12086
signed_add32_overflows(s32 a,s32 b)12087 static bool signed_add32_overflows(s32 a, s32 b)
12088 {
12089 /* Do the add in u32, where overflow is well-defined */
12090 s32 res = (s32)((u32)a + (u32)b);
12091
12092 if (b < 0)
12093 return res > a;
12094 return res < a;
12095 }
12096
signed_sub_overflows(s64 a,s64 b)12097 static bool signed_sub_overflows(s64 a, s64 b)
12098 {
12099 /* Do the sub in u64, where overflow is well-defined */
12100 s64 res = (s64)((u64)a - (u64)b);
12101
12102 if (b < 0)
12103 return res < a;
12104 return res > a;
12105 }
12106
signed_sub32_overflows(s32 a,s32 b)12107 static bool signed_sub32_overflows(s32 a, s32 b)
12108 {
12109 /* Do the sub in u32, where overflow is well-defined */
12110 s32 res = (s32)((u32)a - (u32)b);
12111
12112 if (b < 0)
12113 return res < a;
12114 return res > a;
12115 }
12116
check_reg_sane_offset(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,enum bpf_reg_type type)12117 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
12118 const struct bpf_reg_state *reg,
12119 enum bpf_reg_type type)
12120 {
12121 bool known = tnum_is_const(reg->var_off);
12122 s64 val = reg->var_off.value;
12123 s64 smin = reg->smin_value;
12124
12125 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
12126 verbose(env, "math between %s pointer and %lld is not allowed\n",
12127 reg_type_str(env, type), val);
12128 return false;
12129 }
12130
12131 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
12132 verbose(env, "%s pointer offset %d is not allowed\n",
12133 reg_type_str(env, type), reg->off);
12134 return false;
12135 }
12136
12137 if (smin == S64_MIN) {
12138 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
12139 reg_type_str(env, type));
12140 return false;
12141 }
12142
12143 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
12144 verbose(env, "value %lld makes %s pointer be out of bounds\n",
12145 smin, reg_type_str(env, type));
12146 return false;
12147 }
12148
12149 return true;
12150 }
12151
12152 enum {
12153 REASON_BOUNDS = -1,
12154 REASON_TYPE = -2,
12155 REASON_PATHS = -3,
12156 REASON_LIMIT = -4,
12157 REASON_STACK = -5,
12158 };
12159
retrieve_ptr_limit(const struct bpf_reg_state * ptr_reg,u32 * alu_limit,bool mask_to_left)12160 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
12161 u32 *alu_limit, bool mask_to_left)
12162 {
12163 u32 max = 0, ptr_limit = 0;
12164
12165 switch (ptr_reg->type) {
12166 case PTR_TO_STACK:
12167 /* Offset 0 is out-of-bounds, but acceptable start for the
12168 * left direction, see BPF_REG_FP. Also, unknown scalar
12169 * offset where we would need to deal with min/max bounds is
12170 * currently prohibited for unprivileged.
12171 */
12172 max = MAX_BPF_STACK + mask_to_left;
12173 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
12174 break;
12175 case PTR_TO_MAP_VALUE:
12176 max = ptr_reg->map_ptr->value_size;
12177 ptr_limit = (mask_to_left ?
12178 ptr_reg->smin_value :
12179 ptr_reg->umax_value) + ptr_reg->off;
12180 break;
12181 default:
12182 return REASON_TYPE;
12183 }
12184
12185 if (ptr_limit >= max)
12186 return REASON_LIMIT;
12187 *alu_limit = ptr_limit;
12188 return 0;
12189 }
12190
can_skip_alu_sanitation(const struct bpf_verifier_env * env,const struct bpf_insn * insn)12191 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
12192 const struct bpf_insn *insn)
12193 {
12194 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
12195 }
12196
update_alu_sanitation_state(struct bpf_insn_aux_data * aux,u32 alu_state,u32 alu_limit)12197 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
12198 u32 alu_state, u32 alu_limit)
12199 {
12200 /* If we arrived here from different branches with different
12201 * state or limits to sanitize, then this won't work.
12202 */
12203 if (aux->alu_state &&
12204 (aux->alu_state != alu_state ||
12205 aux->alu_limit != alu_limit))
12206 return REASON_PATHS;
12207
12208 /* Corresponding fixup done in do_misc_fixups(). */
12209 aux->alu_state = alu_state;
12210 aux->alu_limit = alu_limit;
12211 return 0;
12212 }
12213
sanitize_val_alu(struct bpf_verifier_env * env,struct bpf_insn * insn)12214 static int sanitize_val_alu(struct bpf_verifier_env *env,
12215 struct bpf_insn *insn)
12216 {
12217 struct bpf_insn_aux_data *aux = cur_aux(env);
12218
12219 if (can_skip_alu_sanitation(env, insn))
12220 return 0;
12221
12222 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
12223 }
12224
sanitize_needed(u8 opcode)12225 static bool sanitize_needed(u8 opcode)
12226 {
12227 return opcode == BPF_ADD || opcode == BPF_SUB;
12228 }
12229
12230 struct bpf_sanitize_info {
12231 struct bpf_insn_aux_data aux;
12232 bool mask_to_left;
12233 };
12234
12235 static struct bpf_verifier_state *
sanitize_speculative_path(struct bpf_verifier_env * env,const struct bpf_insn * insn,u32 next_idx,u32 curr_idx)12236 sanitize_speculative_path(struct bpf_verifier_env *env,
12237 const struct bpf_insn *insn,
12238 u32 next_idx, u32 curr_idx)
12239 {
12240 struct bpf_verifier_state *branch;
12241 struct bpf_reg_state *regs;
12242
12243 branch = push_stack(env, next_idx, curr_idx, true);
12244 if (branch && insn) {
12245 regs = branch->frame[branch->curframe]->regs;
12246 if (BPF_SRC(insn->code) == BPF_K) {
12247 mark_reg_unknown(env, regs, insn->dst_reg);
12248 } else if (BPF_SRC(insn->code) == BPF_X) {
12249 mark_reg_unknown(env, regs, insn->dst_reg);
12250 mark_reg_unknown(env, regs, insn->src_reg);
12251 }
12252 }
12253 return branch;
12254 }
12255
sanitize_ptr_alu(struct bpf_verifier_env * env,struct bpf_insn * insn,const struct bpf_reg_state * ptr_reg,const struct bpf_reg_state * off_reg,struct bpf_reg_state * dst_reg,struct bpf_sanitize_info * info,const bool commit_window)12256 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
12257 struct bpf_insn *insn,
12258 const struct bpf_reg_state *ptr_reg,
12259 const struct bpf_reg_state *off_reg,
12260 struct bpf_reg_state *dst_reg,
12261 struct bpf_sanitize_info *info,
12262 const bool commit_window)
12263 {
12264 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
12265 struct bpf_verifier_state *vstate = env->cur_state;
12266 bool off_is_imm = tnum_is_const(off_reg->var_off);
12267 bool off_is_neg = off_reg->smin_value < 0;
12268 bool ptr_is_dst_reg = ptr_reg == dst_reg;
12269 u8 opcode = BPF_OP(insn->code);
12270 u32 alu_state, alu_limit;
12271 struct bpf_reg_state tmp;
12272 bool ret;
12273 int err;
12274
12275 if (can_skip_alu_sanitation(env, insn))
12276 return 0;
12277
12278 /* We already marked aux for masking from non-speculative
12279 * paths, thus we got here in the first place. We only care
12280 * to explore bad access from here.
12281 */
12282 if (vstate->speculative)
12283 goto do_sim;
12284
12285 if (!commit_window) {
12286 if (!tnum_is_const(off_reg->var_off) &&
12287 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
12288 return REASON_BOUNDS;
12289
12290 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
12291 (opcode == BPF_SUB && !off_is_neg);
12292 }
12293
12294 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
12295 if (err < 0)
12296 return err;
12297
12298 if (commit_window) {
12299 /* In commit phase we narrow the masking window based on
12300 * the observed pointer move after the simulated operation.
12301 */
12302 alu_state = info->aux.alu_state;
12303 alu_limit = abs(info->aux.alu_limit - alu_limit);
12304 } else {
12305 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
12306 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
12307 alu_state |= ptr_is_dst_reg ?
12308 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
12309
12310 /* Limit pruning on unknown scalars to enable deep search for
12311 * potential masking differences from other program paths.
12312 */
12313 if (!off_is_imm)
12314 env->explore_alu_limits = true;
12315 }
12316
12317 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
12318 if (err < 0)
12319 return err;
12320 do_sim:
12321 /* If we're in commit phase, we're done here given we already
12322 * pushed the truncated dst_reg into the speculative verification
12323 * stack.
12324 *
12325 * Also, when register is a known constant, we rewrite register-based
12326 * operation to immediate-based, and thus do not need masking (and as
12327 * a consequence, do not need to simulate the zero-truncation either).
12328 */
12329 if (commit_window || off_is_imm)
12330 return 0;
12331
12332 /* Simulate and find potential out-of-bounds access under
12333 * speculative execution from truncation as a result of
12334 * masking when off was not within expected range. If off
12335 * sits in dst, then we temporarily need to move ptr there
12336 * to simulate dst (== 0) +/-= ptr. Needed, for example,
12337 * for cases where we use K-based arithmetic in one direction
12338 * and truncated reg-based in the other in order to explore
12339 * bad access.
12340 */
12341 if (!ptr_is_dst_reg) {
12342 tmp = *dst_reg;
12343 copy_register_state(dst_reg, ptr_reg);
12344 }
12345 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
12346 env->insn_idx);
12347 if (!ptr_is_dst_reg && ret)
12348 *dst_reg = tmp;
12349 return !ret ? REASON_STACK : 0;
12350 }
12351
sanitize_mark_insn_seen(struct bpf_verifier_env * env)12352 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
12353 {
12354 struct bpf_verifier_state *vstate = env->cur_state;
12355
12356 /* If we simulate paths under speculation, we don't update the
12357 * insn as 'seen' such that when we verify unreachable paths in
12358 * the non-speculative domain, sanitize_dead_code() can still
12359 * rewrite/sanitize them.
12360 */
12361 if (!vstate->speculative)
12362 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
12363 }
12364
sanitize_err(struct bpf_verifier_env * env,const struct bpf_insn * insn,int reason,const struct bpf_reg_state * off_reg,const struct bpf_reg_state * dst_reg)12365 static int sanitize_err(struct bpf_verifier_env *env,
12366 const struct bpf_insn *insn, int reason,
12367 const struct bpf_reg_state *off_reg,
12368 const struct bpf_reg_state *dst_reg)
12369 {
12370 static const char *err = "pointer arithmetic with it prohibited for !root";
12371 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
12372 u32 dst = insn->dst_reg, src = insn->src_reg;
12373
12374 switch (reason) {
12375 case REASON_BOUNDS:
12376 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
12377 off_reg == dst_reg ? dst : src, err);
12378 break;
12379 case REASON_TYPE:
12380 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
12381 off_reg == dst_reg ? src : dst, err);
12382 break;
12383 case REASON_PATHS:
12384 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
12385 dst, op, err);
12386 break;
12387 case REASON_LIMIT:
12388 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
12389 dst, op, err);
12390 break;
12391 case REASON_STACK:
12392 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
12393 dst, err);
12394 break;
12395 default:
12396 verbose(env, "verifier internal error: unknown reason (%d)\n",
12397 reason);
12398 break;
12399 }
12400
12401 return -EACCES;
12402 }
12403
12404 /* check that stack access falls within stack limits and that 'reg' doesn't
12405 * have a variable offset.
12406 *
12407 * Variable offset is prohibited for unprivileged mode for simplicity since it
12408 * requires corresponding support in Spectre masking for stack ALU. See also
12409 * retrieve_ptr_limit().
12410 *
12411 *
12412 * 'off' includes 'reg->off'.
12413 */
check_stack_access_for_ptr_arithmetic(struct bpf_verifier_env * env,int regno,const struct bpf_reg_state * reg,int off)12414 static int check_stack_access_for_ptr_arithmetic(
12415 struct bpf_verifier_env *env,
12416 int regno,
12417 const struct bpf_reg_state *reg,
12418 int off)
12419 {
12420 if (!tnum_is_const(reg->var_off)) {
12421 char tn_buf[48];
12422
12423 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
12424 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
12425 regno, tn_buf, off);
12426 return -EACCES;
12427 }
12428
12429 if (off >= 0 || off < -MAX_BPF_STACK) {
12430 verbose(env, "R%d stack pointer arithmetic goes out of range, "
12431 "prohibited for !root; off=%d\n", regno, off);
12432 return -EACCES;
12433 }
12434
12435 return 0;
12436 }
12437
sanitize_check_bounds(struct bpf_verifier_env * env,const struct bpf_insn * insn,const struct bpf_reg_state * dst_reg)12438 static int sanitize_check_bounds(struct bpf_verifier_env *env,
12439 const struct bpf_insn *insn,
12440 const struct bpf_reg_state *dst_reg)
12441 {
12442 u32 dst = insn->dst_reg;
12443
12444 /* For unprivileged we require that resulting offset must be in bounds
12445 * in order to be able to sanitize access later on.
12446 */
12447 if (env->bypass_spec_v1)
12448 return 0;
12449
12450 switch (dst_reg->type) {
12451 case PTR_TO_STACK:
12452 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
12453 dst_reg->off + dst_reg->var_off.value))
12454 return -EACCES;
12455 break;
12456 case PTR_TO_MAP_VALUE:
12457 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
12458 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
12459 "prohibited for !root\n", dst);
12460 return -EACCES;
12461 }
12462 break;
12463 default:
12464 break;
12465 }
12466
12467 return 0;
12468 }
12469
12470 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
12471 * Caller should also handle BPF_MOV case separately.
12472 * If we return -EACCES, caller may want to try again treating pointer as a
12473 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
12474 */
adjust_ptr_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn,const struct bpf_reg_state * ptr_reg,const struct bpf_reg_state * off_reg)12475 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
12476 struct bpf_insn *insn,
12477 const struct bpf_reg_state *ptr_reg,
12478 const struct bpf_reg_state *off_reg)
12479 {
12480 struct bpf_verifier_state *vstate = env->cur_state;
12481 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12482 struct bpf_reg_state *regs = state->regs, *dst_reg;
12483 bool known = tnum_is_const(off_reg->var_off);
12484 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
12485 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
12486 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
12487 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
12488 struct bpf_sanitize_info info = {};
12489 u8 opcode = BPF_OP(insn->code);
12490 u32 dst = insn->dst_reg;
12491 int ret;
12492
12493 dst_reg = ®s[dst];
12494
12495 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
12496 smin_val > smax_val || umin_val > umax_val) {
12497 /* Taint dst register if offset had invalid bounds derived from
12498 * e.g. dead branches.
12499 */
12500 __mark_reg_unknown(env, dst_reg);
12501 return 0;
12502 }
12503
12504 if (BPF_CLASS(insn->code) != BPF_ALU64) {
12505 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
12506 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12507 __mark_reg_unknown(env, dst_reg);
12508 return 0;
12509 }
12510
12511 verbose(env,
12512 "R%d 32-bit pointer arithmetic prohibited\n",
12513 dst);
12514 return -EACCES;
12515 }
12516
12517 if (ptr_reg->type & PTR_MAYBE_NULL) {
12518 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
12519 dst, reg_type_str(env, ptr_reg->type));
12520 return -EACCES;
12521 }
12522
12523 switch (base_type(ptr_reg->type)) {
12524 case PTR_TO_FLOW_KEYS:
12525 if (known)
12526 break;
12527 fallthrough;
12528 case CONST_PTR_TO_MAP:
12529 /* smin_val represents the known value */
12530 if (known && smin_val == 0 && opcode == BPF_ADD)
12531 break;
12532 fallthrough;
12533 case PTR_TO_PACKET_END:
12534 case PTR_TO_SOCKET:
12535 case PTR_TO_SOCK_COMMON:
12536 case PTR_TO_TCP_SOCK:
12537 case PTR_TO_XDP_SOCK:
12538 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
12539 dst, reg_type_str(env, ptr_reg->type));
12540 return -EACCES;
12541 default:
12542 break;
12543 }
12544
12545 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
12546 * The id may be overwritten later if we create a new variable offset.
12547 */
12548 dst_reg->type = ptr_reg->type;
12549 dst_reg->id = ptr_reg->id;
12550
12551 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
12552 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
12553 return -EINVAL;
12554
12555 /* pointer types do not carry 32-bit bounds at the moment. */
12556 __mark_reg32_unbounded(dst_reg);
12557
12558 if (sanitize_needed(opcode)) {
12559 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
12560 &info, false);
12561 if (ret < 0)
12562 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12563 }
12564
12565 switch (opcode) {
12566 case BPF_ADD:
12567 /* We can take a fixed offset as long as it doesn't overflow
12568 * the s32 'off' field
12569 */
12570 if (known && (ptr_reg->off + smin_val ==
12571 (s64)(s32)(ptr_reg->off + smin_val))) {
12572 /* pointer += K. Accumulate it into fixed offset */
12573 dst_reg->smin_value = smin_ptr;
12574 dst_reg->smax_value = smax_ptr;
12575 dst_reg->umin_value = umin_ptr;
12576 dst_reg->umax_value = umax_ptr;
12577 dst_reg->var_off = ptr_reg->var_off;
12578 dst_reg->off = ptr_reg->off + smin_val;
12579 dst_reg->raw = ptr_reg->raw;
12580 break;
12581 }
12582 /* A new variable offset is created. Note that off_reg->off
12583 * == 0, since it's a scalar.
12584 * dst_reg gets the pointer type and since some positive
12585 * integer value was added to the pointer, give it a new 'id'
12586 * if it's a PTR_TO_PACKET.
12587 * this creates a new 'base' pointer, off_reg (variable) gets
12588 * added into the variable offset, and we copy the fixed offset
12589 * from ptr_reg.
12590 */
12591 if (signed_add_overflows(smin_ptr, smin_val) ||
12592 signed_add_overflows(smax_ptr, smax_val)) {
12593 dst_reg->smin_value = S64_MIN;
12594 dst_reg->smax_value = S64_MAX;
12595 } else {
12596 dst_reg->smin_value = smin_ptr + smin_val;
12597 dst_reg->smax_value = smax_ptr + smax_val;
12598 }
12599 if (umin_ptr + umin_val < umin_ptr ||
12600 umax_ptr + umax_val < umax_ptr) {
12601 dst_reg->umin_value = 0;
12602 dst_reg->umax_value = U64_MAX;
12603 } else {
12604 dst_reg->umin_value = umin_ptr + umin_val;
12605 dst_reg->umax_value = umax_ptr + umax_val;
12606 }
12607 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
12608 dst_reg->off = ptr_reg->off;
12609 dst_reg->raw = ptr_reg->raw;
12610 if (reg_is_pkt_pointer(ptr_reg)) {
12611 dst_reg->id = ++env->id_gen;
12612 /* something was added to pkt_ptr, set range to zero */
12613 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12614 }
12615 break;
12616 case BPF_SUB:
12617 if (dst_reg == off_reg) {
12618 /* scalar -= pointer. Creates an unknown scalar */
12619 verbose(env, "R%d tried to subtract pointer from scalar\n",
12620 dst);
12621 return -EACCES;
12622 }
12623 /* We don't allow subtraction from FP, because (according to
12624 * test_verifier.c test "invalid fp arithmetic", JITs might not
12625 * be able to deal with it.
12626 */
12627 if (ptr_reg->type == PTR_TO_STACK) {
12628 verbose(env, "R%d subtraction from stack pointer prohibited\n",
12629 dst);
12630 return -EACCES;
12631 }
12632 if (known && (ptr_reg->off - smin_val ==
12633 (s64)(s32)(ptr_reg->off - smin_val))) {
12634 /* pointer -= K. Subtract it from fixed offset */
12635 dst_reg->smin_value = smin_ptr;
12636 dst_reg->smax_value = smax_ptr;
12637 dst_reg->umin_value = umin_ptr;
12638 dst_reg->umax_value = umax_ptr;
12639 dst_reg->var_off = ptr_reg->var_off;
12640 dst_reg->id = ptr_reg->id;
12641 dst_reg->off = ptr_reg->off - smin_val;
12642 dst_reg->raw = ptr_reg->raw;
12643 break;
12644 }
12645 /* A new variable offset is created. If the subtrahend is known
12646 * nonnegative, then any reg->range we had before is still good.
12647 */
12648 if (signed_sub_overflows(smin_ptr, smax_val) ||
12649 signed_sub_overflows(smax_ptr, smin_val)) {
12650 /* Overflow possible, we know nothing */
12651 dst_reg->smin_value = S64_MIN;
12652 dst_reg->smax_value = S64_MAX;
12653 } else {
12654 dst_reg->smin_value = smin_ptr - smax_val;
12655 dst_reg->smax_value = smax_ptr - smin_val;
12656 }
12657 if (umin_ptr < umax_val) {
12658 /* Overflow possible, we know nothing */
12659 dst_reg->umin_value = 0;
12660 dst_reg->umax_value = U64_MAX;
12661 } else {
12662 /* Cannot overflow (as long as bounds are consistent) */
12663 dst_reg->umin_value = umin_ptr - umax_val;
12664 dst_reg->umax_value = umax_ptr - umin_val;
12665 }
12666 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12667 dst_reg->off = ptr_reg->off;
12668 dst_reg->raw = ptr_reg->raw;
12669 if (reg_is_pkt_pointer(ptr_reg)) {
12670 dst_reg->id = ++env->id_gen;
12671 /* something was added to pkt_ptr, set range to zero */
12672 if (smin_val < 0)
12673 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12674 }
12675 break;
12676 case BPF_AND:
12677 case BPF_OR:
12678 case BPF_XOR:
12679 /* bitwise ops on pointers are troublesome, prohibit. */
12680 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12681 dst, bpf_alu_string[opcode >> 4]);
12682 return -EACCES;
12683 default:
12684 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12685 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12686 dst, bpf_alu_string[opcode >> 4]);
12687 return -EACCES;
12688 }
12689
12690 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12691 return -EINVAL;
12692 reg_bounds_sync(dst_reg);
12693 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12694 return -EACCES;
12695 if (sanitize_needed(opcode)) {
12696 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12697 &info, true);
12698 if (ret < 0)
12699 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12700 }
12701
12702 return 0;
12703 }
12704
scalar32_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12705 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12706 struct bpf_reg_state *src_reg)
12707 {
12708 s32 smin_val = src_reg->s32_min_value;
12709 s32 smax_val = src_reg->s32_max_value;
12710 u32 umin_val = src_reg->u32_min_value;
12711 u32 umax_val = src_reg->u32_max_value;
12712
12713 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12714 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12715 dst_reg->s32_min_value = S32_MIN;
12716 dst_reg->s32_max_value = S32_MAX;
12717 } else {
12718 dst_reg->s32_min_value += smin_val;
12719 dst_reg->s32_max_value += smax_val;
12720 }
12721 if (dst_reg->u32_min_value + umin_val < umin_val ||
12722 dst_reg->u32_max_value + umax_val < umax_val) {
12723 dst_reg->u32_min_value = 0;
12724 dst_reg->u32_max_value = U32_MAX;
12725 } else {
12726 dst_reg->u32_min_value += umin_val;
12727 dst_reg->u32_max_value += umax_val;
12728 }
12729 }
12730
scalar_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12731 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12732 struct bpf_reg_state *src_reg)
12733 {
12734 s64 smin_val = src_reg->smin_value;
12735 s64 smax_val = src_reg->smax_value;
12736 u64 umin_val = src_reg->umin_value;
12737 u64 umax_val = src_reg->umax_value;
12738
12739 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12740 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12741 dst_reg->smin_value = S64_MIN;
12742 dst_reg->smax_value = S64_MAX;
12743 } else {
12744 dst_reg->smin_value += smin_val;
12745 dst_reg->smax_value += smax_val;
12746 }
12747 if (dst_reg->umin_value + umin_val < umin_val ||
12748 dst_reg->umax_value + umax_val < umax_val) {
12749 dst_reg->umin_value = 0;
12750 dst_reg->umax_value = U64_MAX;
12751 } else {
12752 dst_reg->umin_value += umin_val;
12753 dst_reg->umax_value += umax_val;
12754 }
12755 }
12756
scalar32_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12757 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12758 struct bpf_reg_state *src_reg)
12759 {
12760 s32 smin_val = src_reg->s32_min_value;
12761 s32 smax_val = src_reg->s32_max_value;
12762 u32 umin_val = src_reg->u32_min_value;
12763 u32 umax_val = src_reg->u32_max_value;
12764
12765 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12766 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12767 /* Overflow possible, we know nothing */
12768 dst_reg->s32_min_value = S32_MIN;
12769 dst_reg->s32_max_value = S32_MAX;
12770 } else {
12771 dst_reg->s32_min_value -= smax_val;
12772 dst_reg->s32_max_value -= smin_val;
12773 }
12774 if (dst_reg->u32_min_value < umax_val) {
12775 /* Overflow possible, we know nothing */
12776 dst_reg->u32_min_value = 0;
12777 dst_reg->u32_max_value = U32_MAX;
12778 } else {
12779 /* Cannot overflow (as long as bounds are consistent) */
12780 dst_reg->u32_min_value -= umax_val;
12781 dst_reg->u32_max_value -= umin_val;
12782 }
12783 }
12784
scalar_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12785 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12786 struct bpf_reg_state *src_reg)
12787 {
12788 s64 smin_val = src_reg->smin_value;
12789 s64 smax_val = src_reg->smax_value;
12790 u64 umin_val = src_reg->umin_value;
12791 u64 umax_val = src_reg->umax_value;
12792
12793 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12794 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12795 /* Overflow possible, we know nothing */
12796 dst_reg->smin_value = S64_MIN;
12797 dst_reg->smax_value = S64_MAX;
12798 } else {
12799 dst_reg->smin_value -= smax_val;
12800 dst_reg->smax_value -= smin_val;
12801 }
12802 if (dst_reg->umin_value < umax_val) {
12803 /* Overflow possible, we know nothing */
12804 dst_reg->umin_value = 0;
12805 dst_reg->umax_value = U64_MAX;
12806 } else {
12807 /* Cannot overflow (as long as bounds are consistent) */
12808 dst_reg->umin_value -= umax_val;
12809 dst_reg->umax_value -= umin_val;
12810 }
12811 }
12812
scalar32_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12813 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12814 struct bpf_reg_state *src_reg)
12815 {
12816 s32 smin_val = src_reg->s32_min_value;
12817 u32 umin_val = src_reg->u32_min_value;
12818 u32 umax_val = src_reg->u32_max_value;
12819
12820 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12821 /* Ain't nobody got time to multiply that sign */
12822 __mark_reg32_unbounded(dst_reg);
12823 return;
12824 }
12825 /* Both values are positive, so we can work with unsigned and
12826 * copy the result to signed (unless it exceeds S32_MAX).
12827 */
12828 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12829 /* Potential overflow, we know nothing */
12830 __mark_reg32_unbounded(dst_reg);
12831 return;
12832 }
12833 dst_reg->u32_min_value *= umin_val;
12834 dst_reg->u32_max_value *= umax_val;
12835 if (dst_reg->u32_max_value > S32_MAX) {
12836 /* Overflow possible, we know nothing */
12837 dst_reg->s32_min_value = S32_MIN;
12838 dst_reg->s32_max_value = S32_MAX;
12839 } else {
12840 dst_reg->s32_min_value = dst_reg->u32_min_value;
12841 dst_reg->s32_max_value = dst_reg->u32_max_value;
12842 }
12843 }
12844
scalar_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12845 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12846 struct bpf_reg_state *src_reg)
12847 {
12848 s64 smin_val = src_reg->smin_value;
12849 u64 umin_val = src_reg->umin_value;
12850 u64 umax_val = src_reg->umax_value;
12851
12852 if (smin_val < 0 || dst_reg->smin_value < 0) {
12853 /* Ain't nobody got time to multiply that sign */
12854 __mark_reg64_unbounded(dst_reg);
12855 return;
12856 }
12857 /* Both values are positive, so we can work with unsigned and
12858 * copy the result to signed (unless it exceeds S64_MAX).
12859 */
12860 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12861 /* Potential overflow, we know nothing */
12862 __mark_reg64_unbounded(dst_reg);
12863 return;
12864 }
12865 dst_reg->umin_value *= umin_val;
12866 dst_reg->umax_value *= umax_val;
12867 if (dst_reg->umax_value > S64_MAX) {
12868 /* Overflow possible, we know nothing */
12869 dst_reg->smin_value = S64_MIN;
12870 dst_reg->smax_value = S64_MAX;
12871 } else {
12872 dst_reg->smin_value = dst_reg->umin_value;
12873 dst_reg->smax_value = dst_reg->umax_value;
12874 }
12875 }
12876
scalar32_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12877 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12878 struct bpf_reg_state *src_reg)
12879 {
12880 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12881 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12882 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12883 s32 smin_val = src_reg->s32_min_value;
12884 u32 umax_val = src_reg->u32_max_value;
12885
12886 if (src_known && dst_known) {
12887 __mark_reg32_known(dst_reg, var32_off.value);
12888 return;
12889 }
12890
12891 /* We get our minimum from the var_off, since that's inherently
12892 * bitwise. Our maximum is the minimum of the operands' maxima.
12893 */
12894 dst_reg->u32_min_value = var32_off.value;
12895 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12896 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12897 /* Lose signed bounds when ANDing negative numbers,
12898 * ain't nobody got time for that.
12899 */
12900 dst_reg->s32_min_value = S32_MIN;
12901 dst_reg->s32_max_value = S32_MAX;
12902 } else {
12903 /* ANDing two positives gives a positive, so safe to
12904 * cast result into s64.
12905 */
12906 dst_reg->s32_min_value = dst_reg->u32_min_value;
12907 dst_reg->s32_max_value = dst_reg->u32_max_value;
12908 }
12909 }
12910
scalar_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12911 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12912 struct bpf_reg_state *src_reg)
12913 {
12914 bool src_known = tnum_is_const(src_reg->var_off);
12915 bool dst_known = tnum_is_const(dst_reg->var_off);
12916 s64 smin_val = src_reg->smin_value;
12917 u64 umax_val = src_reg->umax_value;
12918
12919 if (src_known && dst_known) {
12920 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12921 return;
12922 }
12923
12924 /* We get our minimum from the var_off, since that's inherently
12925 * bitwise. Our maximum is the minimum of the operands' maxima.
12926 */
12927 dst_reg->umin_value = dst_reg->var_off.value;
12928 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12929 if (dst_reg->smin_value < 0 || smin_val < 0) {
12930 /* Lose signed bounds when ANDing negative numbers,
12931 * ain't nobody got time for that.
12932 */
12933 dst_reg->smin_value = S64_MIN;
12934 dst_reg->smax_value = S64_MAX;
12935 } else {
12936 /* ANDing two positives gives a positive, so safe to
12937 * cast result into s64.
12938 */
12939 dst_reg->smin_value = dst_reg->umin_value;
12940 dst_reg->smax_value = dst_reg->umax_value;
12941 }
12942 /* We may learn something more from the var_off */
12943 __update_reg_bounds(dst_reg);
12944 }
12945
scalar32_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12946 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12947 struct bpf_reg_state *src_reg)
12948 {
12949 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12950 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12951 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12952 s32 smin_val = src_reg->s32_min_value;
12953 u32 umin_val = src_reg->u32_min_value;
12954
12955 if (src_known && dst_known) {
12956 __mark_reg32_known(dst_reg, var32_off.value);
12957 return;
12958 }
12959
12960 /* We get our maximum from the var_off, and our minimum is the
12961 * maximum of the operands' minima
12962 */
12963 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12964 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12965 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12966 /* Lose signed bounds when ORing negative numbers,
12967 * ain't nobody got time for that.
12968 */
12969 dst_reg->s32_min_value = S32_MIN;
12970 dst_reg->s32_max_value = S32_MAX;
12971 } else {
12972 /* ORing two positives gives a positive, so safe to
12973 * cast result into s64.
12974 */
12975 dst_reg->s32_min_value = dst_reg->u32_min_value;
12976 dst_reg->s32_max_value = dst_reg->u32_max_value;
12977 }
12978 }
12979
scalar_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12980 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12981 struct bpf_reg_state *src_reg)
12982 {
12983 bool src_known = tnum_is_const(src_reg->var_off);
12984 bool dst_known = tnum_is_const(dst_reg->var_off);
12985 s64 smin_val = src_reg->smin_value;
12986 u64 umin_val = src_reg->umin_value;
12987
12988 if (src_known && dst_known) {
12989 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12990 return;
12991 }
12992
12993 /* We get our maximum from the var_off, and our minimum is the
12994 * maximum of the operands' minima
12995 */
12996 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12997 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12998 if (dst_reg->smin_value < 0 || smin_val < 0) {
12999 /* Lose signed bounds when ORing negative numbers,
13000 * ain't nobody got time for that.
13001 */
13002 dst_reg->smin_value = S64_MIN;
13003 dst_reg->smax_value = S64_MAX;
13004 } else {
13005 /* ORing two positives gives a positive, so safe to
13006 * cast result into s64.
13007 */
13008 dst_reg->smin_value = dst_reg->umin_value;
13009 dst_reg->smax_value = dst_reg->umax_value;
13010 }
13011 /* We may learn something more from the var_off */
13012 __update_reg_bounds(dst_reg);
13013 }
13014
scalar32_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13015 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
13016 struct bpf_reg_state *src_reg)
13017 {
13018 bool src_known = tnum_subreg_is_const(src_reg->var_off);
13019 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
13020 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
13021 s32 smin_val = src_reg->s32_min_value;
13022
13023 if (src_known && dst_known) {
13024 __mark_reg32_known(dst_reg, var32_off.value);
13025 return;
13026 }
13027
13028 /* We get both minimum and maximum from the var32_off. */
13029 dst_reg->u32_min_value = var32_off.value;
13030 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
13031
13032 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
13033 /* XORing two positive sign numbers gives a positive,
13034 * so safe to cast u32 result into s32.
13035 */
13036 dst_reg->s32_min_value = dst_reg->u32_min_value;
13037 dst_reg->s32_max_value = dst_reg->u32_max_value;
13038 } else {
13039 dst_reg->s32_min_value = S32_MIN;
13040 dst_reg->s32_max_value = S32_MAX;
13041 }
13042 }
13043
scalar_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13044 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
13045 struct bpf_reg_state *src_reg)
13046 {
13047 bool src_known = tnum_is_const(src_reg->var_off);
13048 bool dst_known = tnum_is_const(dst_reg->var_off);
13049 s64 smin_val = src_reg->smin_value;
13050
13051 if (src_known && dst_known) {
13052 /* dst_reg->var_off.value has been updated earlier */
13053 __mark_reg_known(dst_reg, dst_reg->var_off.value);
13054 return;
13055 }
13056
13057 /* We get both minimum and maximum from the var_off. */
13058 dst_reg->umin_value = dst_reg->var_off.value;
13059 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
13060
13061 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
13062 /* XORing two positive sign numbers gives a positive,
13063 * so safe to cast u64 result into s64.
13064 */
13065 dst_reg->smin_value = dst_reg->umin_value;
13066 dst_reg->smax_value = dst_reg->umax_value;
13067 } else {
13068 dst_reg->smin_value = S64_MIN;
13069 dst_reg->smax_value = S64_MAX;
13070 }
13071
13072 __update_reg_bounds(dst_reg);
13073 }
13074
__scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)13075 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
13076 u64 umin_val, u64 umax_val)
13077 {
13078 /* We lose all sign bit information (except what we can pick
13079 * up from var_off)
13080 */
13081 dst_reg->s32_min_value = S32_MIN;
13082 dst_reg->s32_max_value = S32_MAX;
13083 /* If we might shift our top bit out, then we know nothing */
13084 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
13085 dst_reg->u32_min_value = 0;
13086 dst_reg->u32_max_value = U32_MAX;
13087 } else {
13088 dst_reg->u32_min_value <<= umin_val;
13089 dst_reg->u32_max_value <<= umax_val;
13090 }
13091 }
13092
scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13093 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
13094 struct bpf_reg_state *src_reg)
13095 {
13096 u32 umax_val = src_reg->u32_max_value;
13097 u32 umin_val = src_reg->u32_min_value;
13098 /* u32 alu operation will zext upper bits */
13099 struct tnum subreg = tnum_subreg(dst_reg->var_off);
13100
13101 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
13102 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
13103 /* Not required but being careful mark reg64 bounds as unknown so
13104 * that we are forced to pick them up from tnum and zext later and
13105 * if some path skips this step we are still safe.
13106 */
13107 __mark_reg64_unbounded(dst_reg);
13108 __update_reg32_bounds(dst_reg);
13109 }
13110
__scalar64_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)13111 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
13112 u64 umin_val, u64 umax_val)
13113 {
13114 /* Special case <<32 because it is a common compiler pattern to sign
13115 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
13116 * positive we know this shift will also be positive so we can track
13117 * bounds correctly. Otherwise we lose all sign bit information except
13118 * what we can pick up from var_off. Perhaps we can generalize this
13119 * later to shifts of any length.
13120 */
13121 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
13122 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
13123 else
13124 dst_reg->smax_value = S64_MAX;
13125
13126 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
13127 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
13128 else
13129 dst_reg->smin_value = S64_MIN;
13130
13131 /* If we might shift our top bit out, then we know nothing */
13132 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
13133 dst_reg->umin_value = 0;
13134 dst_reg->umax_value = U64_MAX;
13135 } else {
13136 dst_reg->umin_value <<= umin_val;
13137 dst_reg->umax_value <<= umax_val;
13138 }
13139 }
13140
scalar_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13141 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
13142 struct bpf_reg_state *src_reg)
13143 {
13144 u64 umax_val = src_reg->umax_value;
13145 u64 umin_val = src_reg->umin_value;
13146
13147 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
13148 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
13149 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
13150
13151 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
13152 /* We may learn something more from the var_off */
13153 __update_reg_bounds(dst_reg);
13154 }
13155
scalar32_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13156 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
13157 struct bpf_reg_state *src_reg)
13158 {
13159 struct tnum subreg = tnum_subreg(dst_reg->var_off);
13160 u32 umax_val = src_reg->u32_max_value;
13161 u32 umin_val = src_reg->u32_min_value;
13162
13163 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
13164 * be negative, then either:
13165 * 1) src_reg might be zero, so the sign bit of the result is
13166 * unknown, so we lose our signed bounds
13167 * 2) it's known negative, thus the unsigned bounds capture the
13168 * signed bounds
13169 * 3) the signed bounds cross zero, so they tell us nothing
13170 * about the result
13171 * If the value in dst_reg is known nonnegative, then again the
13172 * unsigned bounds capture the signed bounds.
13173 * Thus, in all cases it suffices to blow away our signed bounds
13174 * and rely on inferring new ones from the unsigned bounds and
13175 * var_off of the result.
13176 */
13177 dst_reg->s32_min_value = S32_MIN;
13178 dst_reg->s32_max_value = S32_MAX;
13179
13180 dst_reg->var_off = tnum_rshift(subreg, umin_val);
13181 dst_reg->u32_min_value >>= umax_val;
13182 dst_reg->u32_max_value >>= umin_val;
13183
13184 __mark_reg64_unbounded(dst_reg);
13185 __update_reg32_bounds(dst_reg);
13186 }
13187
scalar_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13188 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
13189 struct bpf_reg_state *src_reg)
13190 {
13191 u64 umax_val = src_reg->umax_value;
13192 u64 umin_val = src_reg->umin_value;
13193
13194 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
13195 * be negative, then either:
13196 * 1) src_reg might be zero, so the sign bit of the result is
13197 * unknown, so we lose our signed bounds
13198 * 2) it's known negative, thus the unsigned bounds capture the
13199 * signed bounds
13200 * 3) the signed bounds cross zero, so they tell us nothing
13201 * about the result
13202 * If the value in dst_reg is known nonnegative, then again the
13203 * unsigned bounds capture the signed bounds.
13204 * Thus, in all cases it suffices to blow away our signed bounds
13205 * and rely on inferring new ones from the unsigned bounds and
13206 * var_off of the result.
13207 */
13208 dst_reg->smin_value = S64_MIN;
13209 dst_reg->smax_value = S64_MAX;
13210 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
13211 dst_reg->umin_value >>= umax_val;
13212 dst_reg->umax_value >>= umin_val;
13213
13214 /* Its not easy to operate on alu32 bounds here because it depends
13215 * on bits being shifted in. Take easy way out and mark unbounded
13216 * so we can recalculate later from tnum.
13217 */
13218 __mark_reg32_unbounded(dst_reg);
13219 __update_reg_bounds(dst_reg);
13220 }
13221
scalar32_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13222 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
13223 struct bpf_reg_state *src_reg)
13224 {
13225 u64 umin_val = src_reg->u32_min_value;
13226
13227 /* Upon reaching here, src_known is true and
13228 * umax_val is equal to umin_val.
13229 */
13230 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
13231 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
13232
13233 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
13234
13235 /* blow away the dst_reg umin_value/umax_value and rely on
13236 * dst_reg var_off to refine the result.
13237 */
13238 dst_reg->u32_min_value = 0;
13239 dst_reg->u32_max_value = U32_MAX;
13240
13241 __mark_reg64_unbounded(dst_reg);
13242 __update_reg32_bounds(dst_reg);
13243 }
13244
scalar_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13245 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
13246 struct bpf_reg_state *src_reg)
13247 {
13248 u64 umin_val = src_reg->umin_value;
13249
13250 /* Upon reaching here, src_known is true and umax_val is equal
13251 * to umin_val.
13252 */
13253 dst_reg->smin_value >>= umin_val;
13254 dst_reg->smax_value >>= umin_val;
13255
13256 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
13257
13258 /* blow away the dst_reg umin_value/umax_value and rely on
13259 * dst_reg var_off to refine the result.
13260 */
13261 dst_reg->umin_value = 0;
13262 dst_reg->umax_value = U64_MAX;
13263
13264 /* Its not easy to operate on alu32 bounds here because it depends
13265 * on bits being shifted in from upper 32-bits. Take easy way out
13266 * and mark unbounded so we can recalculate later from tnum.
13267 */
13268 __mark_reg32_unbounded(dst_reg);
13269 __update_reg_bounds(dst_reg);
13270 }
13271
13272 /* WARNING: This function does calculations on 64-bit values, but the actual
13273 * execution may occur on 32-bit values. Therefore, things like bitshifts
13274 * need extra checks in the 32-bit case.
13275 */
adjust_scalar_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_reg_state * dst_reg,struct bpf_reg_state src_reg)13276 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
13277 struct bpf_insn *insn,
13278 struct bpf_reg_state *dst_reg,
13279 struct bpf_reg_state src_reg)
13280 {
13281 struct bpf_reg_state *regs = cur_regs(env);
13282 u8 opcode = BPF_OP(insn->code);
13283 bool src_known;
13284 s64 smin_val, smax_val;
13285 u64 umin_val, umax_val;
13286 s32 s32_min_val, s32_max_val;
13287 u32 u32_min_val, u32_max_val;
13288 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
13289 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
13290 int ret;
13291
13292 smin_val = src_reg.smin_value;
13293 smax_val = src_reg.smax_value;
13294 umin_val = src_reg.umin_value;
13295 umax_val = src_reg.umax_value;
13296
13297 s32_min_val = src_reg.s32_min_value;
13298 s32_max_val = src_reg.s32_max_value;
13299 u32_min_val = src_reg.u32_min_value;
13300 u32_max_val = src_reg.u32_max_value;
13301
13302 if (alu32) {
13303 src_known = tnum_subreg_is_const(src_reg.var_off);
13304 if ((src_known &&
13305 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
13306 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
13307 /* Taint dst register if offset had invalid bounds
13308 * derived from e.g. dead branches.
13309 */
13310 __mark_reg_unknown(env, dst_reg);
13311 return 0;
13312 }
13313 } else {
13314 src_known = tnum_is_const(src_reg.var_off);
13315 if ((src_known &&
13316 (smin_val != smax_val || umin_val != umax_val)) ||
13317 smin_val > smax_val || umin_val > umax_val) {
13318 /* Taint dst register if offset had invalid bounds
13319 * derived from e.g. dead branches.
13320 */
13321 __mark_reg_unknown(env, dst_reg);
13322 return 0;
13323 }
13324 }
13325
13326 if (!src_known &&
13327 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
13328 __mark_reg_unknown(env, dst_reg);
13329 return 0;
13330 }
13331
13332 if (sanitize_needed(opcode)) {
13333 ret = sanitize_val_alu(env, insn);
13334 if (ret < 0)
13335 return sanitize_err(env, insn, ret, NULL, NULL);
13336 }
13337
13338 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
13339 * There are two classes of instructions: The first class we track both
13340 * alu32 and alu64 sign/unsigned bounds independently this provides the
13341 * greatest amount of precision when alu operations are mixed with jmp32
13342 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
13343 * and BPF_OR. This is possible because these ops have fairly easy to
13344 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
13345 * See alu32 verifier tests for examples. The second class of
13346 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
13347 * with regards to tracking sign/unsigned bounds because the bits may
13348 * cross subreg boundaries in the alu64 case. When this happens we mark
13349 * the reg unbounded in the subreg bound space and use the resulting
13350 * tnum to calculate an approximation of the sign/unsigned bounds.
13351 */
13352 switch (opcode) {
13353 case BPF_ADD:
13354 scalar32_min_max_add(dst_reg, &src_reg);
13355 scalar_min_max_add(dst_reg, &src_reg);
13356 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
13357 break;
13358 case BPF_SUB:
13359 scalar32_min_max_sub(dst_reg, &src_reg);
13360 scalar_min_max_sub(dst_reg, &src_reg);
13361 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
13362 break;
13363 case BPF_MUL:
13364 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
13365 scalar32_min_max_mul(dst_reg, &src_reg);
13366 scalar_min_max_mul(dst_reg, &src_reg);
13367 break;
13368 case BPF_AND:
13369 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
13370 scalar32_min_max_and(dst_reg, &src_reg);
13371 scalar_min_max_and(dst_reg, &src_reg);
13372 break;
13373 case BPF_OR:
13374 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
13375 scalar32_min_max_or(dst_reg, &src_reg);
13376 scalar_min_max_or(dst_reg, &src_reg);
13377 break;
13378 case BPF_XOR:
13379 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
13380 scalar32_min_max_xor(dst_reg, &src_reg);
13381 scalar_min_max_xor(dst_reg, &src_reg);
13382 break;
13383 case BPF_LSH:
13384 if (umax_val >= insn_bitness) {
13385 /* Shifts greater than 31 or 63 are undefined.
13386 * This includes shifts by a negative number.
13387 */
13388 mark_reg_unknown(env, regs, insn->dst_reg);
13389 break;
13390 }
13391 if (alu32)
13392 scalar32_min_max_lsh(dst_reg, &src_reg);
13393 else
13394 scalar_min_max_lsh(dst_reg, &src_reg);
13395 break;
13396 case BPF_RSH:
13397 if (umax_val >= insn_bitness) {
13398 /* Shifts greater than 31 or 63 are undefined.
13399 * This includes shifts by a negative number.
13400 */
13401 mark_reg_unknown(env, regs, insn->dst_reg);
13402 break;
13403 }
13404 if (alu32)
13405 scalar32_min_max_rsh(dst_reg, &src_reg);
13406 else
13407 scalar_min_max_rsh(dst_reg, &src_reg);
13408 break;
13409 case BPF_ARSH:
13410 if (umax_val >= insn_bitness) {
13411 /* Shifts greater than 31 or 63 are undefined.
13412 * This includes shifts by a negative number.
13413 */
13414 mark_reg_unknown(env, regs, insn->dst_reg);
13415 break;
13416 }
13417 if (alu32)
13418 scalar32_min_max_arsh(dst_reg, &src_reg);
13419 else
13420 scalar_min_max_arsh(dst_reg, &src_reg);
13421 break;
13422 default:
13423 mark_reg_unknown(env, regs, insn->dst_reg);
13424 break;
13425 }
13426
13427 /* ALU32 ops are zero extended into 64bit register */
13428 if (alu32)
13429 zext_32_to_64(dst_reg);
13430 reg_bounds_sync(dst_reg);
13431 return 0;
13432 }
13433
13434 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
13435 * and var_off.
13436 */
adjust_reg_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn)13437 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
13438 struct bpf_insn *insn)
13439 {
13440 struct bpf_verifier_state *vstate = env->cur_state;
13441 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13442 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
13443 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
13444 u8 opcode = BPF_OP(insn->code);
13445 int err;
13446
13447 dst_reg = ®s[insn->dst_reg];
13448 src_reg = NULL;
13449 if (dst_reg->type != SCALAR_VALUE)
13450 ptr_reg = dst_reg;
13451 else
13452 /* Make sure ID is cleared otherwise dst_reg min/max could be
13453 * incorrectly propagated into other registers by find_equal_scalars()
13454 */
13455 dst_reg->id = 0;
13456 if (BPF_SRC(insn->code) == BPF_X) {
13457 src_reg = ®s[insn->src_reg];
13458 if (src_reg->type != SCALAR_VALUE) {
13459 if (dst_reg->type != SCALAR_VALUE) {
13460 /* Combining two pointers by any ALU op yields
13461 * an arbitrary scalar. Disallow all math except
13462 * pointer subtraction
13463 */
13464 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
13465 mark_reg_unknown(env, regs, insn->dst_reg);
13466 return 0;
13467 }
13468 verbose(env, "R%d pointer %s pointer prohibited\n",
13469 insn->dst_reg,
13470 bpf_alu_string[opcode >> 4]);
13471 return -EACCES;
13472 } else {
13473 /* scalar += pointer
13474 * This is legal, but we have to reverse our
13475 * src/dest handling in computing the range
13476 */
13477 err = mark_chain_precision(env, insn->dst_reg);
13478 if (err)
13479 return err;
13480 return adjust_ptr_min_max_vals(env, insn,
13481 src_reg, dst_reg);
13482 }
13483 } else if (ptr_reg) {
13484 /* pointer += scalar */
13485 err = mark_chain_precision(env, insn->src_reg);
13486 if (err)
13487 return err;
13488 return adjust_ptr_min_max_vals(env, insn,
13489 dst_reg, src_reg);
13490 } else if (dst_reg->precise) {
13491 /* if dst_reg is precise, src_reg should be precise as well */
13492 err = mark_chain_precision(env, insn->src_reg);
13493 if (err)
13494 return err;
13495 }
13496 } else {
13497 /* Pretend the src is a reg with a known value, since we only
13498 * need to be able to read from this state.
13499 */
13500 off_reg.type = SCALAR_VALUE;
13501 __mark_reg_known(&off_reg, insn->imm);
13502 src_reg = &off_reg;
13503 if (ptr_reg) /* pointer += K */
13504 return adjust_ptr_min_max_vals(env, insn,
13505 ptr_reg, src_reg);
13506 }
13507
13508 /* Got here implies adding two SCALAR_VALUEs */
13509 if (WARN_ON_ONCE(ptr_reg)) {
13510 print_verifier_state(env, state, true);
13511 verbose(env, "verifier internal error: unexpected ptr_reg\n");
13512 return -EINVAL;
13513 }
13514 if (WARN_ON(!src_reg)) {
13515 print_verifier_state(env, state, true);
13516 verbose(env, "verifier internal error: no src_reg\n");
13517 return -EINVAL;
13518 }
13519 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
13520 }
13521
13522 /* check validity of 32-bit and 64-bit arithmetic operations */
check_alu_op(struct bpf_verifier_env * env,struct bpf_insn * insn)13523 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
13524 {
13525 struct bpf_reg_state *regs = cur_regs(env);
13526 u8 opcode = BPF_OP(insn->code);
13527 int err;
13528
13529 if (opcode == BPF_END || opcode == BPF_NEG) {
13530 if (opcode == BPF_NEG) {
13531 if (BPF_SRC(insn->code) != BPF_K ||
13532 insn->src_reg != BPF_REG_0 ||
13533 insn->off != 0 || insn->imm != 0) {
13534 verbose(env, "BPF_NEG uses reserved fields\n");
13535 return -EINVAL;
13536 }
13537 } else {
13538 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
13539 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
13540 (BPF_CLASS(insn->code) == BPF_ALU64 &&
13541 BPF_SRC(insn->code) != BPF_TO_LE)) {
13542 verbose(env, "BPF_END uses reserved fields\n");
13543 return -EINVAL;
13544 }
13545 }
13546
13547 /* check src operand */
13548 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13549 if (err)
13550 return err;
13551
13552 if (is_pointer_value(env, insn->dst_reg)) {
13553 verbose(env, "R%d pointer arithmetic prohibited\n",
13554 insn->dst_reg);
13555 return -EACCES;
13556 }
13557
13558 /* check dest operand */
13559 err = check_reg_arg(env, insn->dst_reg, DST_OP);
13560 if (err)
13561 return err;
13562
13563 } else if (opcode == BPF_MOV) {
13564
13565 if (BPF_SRC(insn->code) == BPF_X) {
13566 if (insn->imm != 0) {
13567 verbose(env, "BPF_MOV uses reserved fields\n");
13568 return -EINVAL;
13569 }
13570
13571 if (BPF_CLASS(insn->code) == BPF_ALU) {
13572 if (insn->off != 0 && insn->off != 8 && insn->off != 16) {
13573 verbose(env, "BPF_MOV uses reserved fields\n");
13574 return -EINVAL;
13575 }
13576 } else {
13577 if (insn->off != 0 && insn->off != 8 && insn->off != 16 &&
13578 insn->off != 32) {
13579 verbose(env, "BPF_MOV uses reserved fields\n");
13580 return -EINVAL;
13581 }
13582 }
13583
13584 /* check src operand */
13585 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13586 if (err)
13587 return err;
13588 } else {
13589 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13590 verbose(env, "BPF_MOV uses reserved fields\n");
13591 return -EINVAL;
13592 }
13593 }
13594
13595 /* check dest operand, mark as required later */
13596 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13597 if (err)
13598 return err;
13599
13600 if (BPF_SRC(insn->code) == BPF_X) {
13601 struct bpf_reg_state *src_reg = regs + insn->src_reg;
13602 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
13603 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
13604 !tnum_is_const(src_reg->var_off);
13605
13606 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13607 if (insn->off == 0) {
13608 /* case: R1 = R2
13609 * copy register state to dest reg
13610 */
13611 if (need_id)
13612 /* Assign src and dst registers the same ID
13613 * that will be used by find_equal_scalars()
13614 * to propagate min/max range.
13615 */
13616 src_reg->id = ++env->id_gen;
13617 copy_register_state(dst_reg, src_reg);
13618 dst_reg->live |= REG_LIVE_WRITTEN;
13619 dst_reg->subreg_def = DEF_NOT_SUBREG;
13620 } else {
13621 /* case: R1 = (s8, s16 s32)R2 */
13622 if (is_pointer_value(env, insn->src_reg)) {
13623 verbose(env,
13624 "R%d sign-extension part of pointer\n",
13625 insn->src_reg);
13626 return -EACCES;
13627 } else if (src_reg->type == SCALAR_VALUE) {
13628 bool no_sext;
13629
13630 no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13631 if (no_sext && need_id)
13632 src_reg->id = ++env->id_gen;
13633 copy_register_state(dst_reg, src_reg);
13634 if (!no_sext)
13635 dst_reg->id = 0;
13636 coerce_reg_to_size_sx(dst_reg, insn->off >> 3);
13637 dst_reg->live |= REG_LIVE_WRITTEN;
13638 dst_reg->subreg_def = DEF_NOT_SUBREG;
13639 } else {
13640 mark_reg_unknown(env, regs, insn->dst_reg);
13641 }
13642 }
13643 } else {
13644 /* R1 = (u32) R2 */
13645 if (is_pointer_value(env, insn->src_reg)) {
13646 verbose(env,
13647 "R%d partial copy of pointer\n",
13648 insn->src_reg);
13649 return -EACCES;
13650 } else if (src_reg->type == SCALAR_VALUE) {
13651 if (insn->off == 0) {
13652 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
13653
13654 if (is_src_reg_u32 && need_id)
13655 src_reg->id = ++env->id_gen;
13656 copy_register_state(dst_reg, src_reg);
13657 /* Make sure ID is cleared if src_reg is not in u32
13658 * range otherwise dst_reg min/max could be incorrectly
13659 * propagated into src_reg by find_equal_scalars()
13660 */
13661 if (!is_src_reg_u32)
13662 dst_reg->id = 0;
13663 dst_reg->live |= REG_LIVE_WRITTEN;
13664 dst_reg->subreg_def = env->insn_idx + 1;
13665 } else {
13666 /* case: W1 = (s8, s16)W2 */
13667 bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13668
13669 if (no_sext && need_id)
13670 src_reg->id = ++env->id_gen;
13671 copy_register_state(dst_reg, src_reg);
13672 if (!no_sext)
13673 dst_reg->id = 0;
13674 dst_reg->live |= REG_LIVE_WRITTEN;
13675 dst_reg->subreg_def = env->insn_idx + 1;
13676 coerce_subreg_to_size_sx(dst_reg, insn->off >> 3);
13677 }
13678 } else {
13679 mark_reg_unknown(env, regs,
13680 insn->dst_reg);
13681 }
13682 zext_32_to_64(dst_reg);
13683 reg_bounds_sync(dst_reg);
13684 }
13685 } else {
13686 /* case: R = imm
13687 * remember the value we stored into this reg
13688 */
13689 /* clear any state __mark_reg_known doesn't set */
13690 mark_reg_unknown(env, regs, insn->dst_reg);
13691 regs[insn->dst_reg].type = SCALAR_VALUE;
13692 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13693 __mark_reg_known(regs + insn->dst_reg,
13694 insn->imm);
13695 } else {
13696 __mark_reg_known(regs + insn->dst_reg,
13697 (u32)insn->imm);
13698 }
13699 }
13700
13701 } else if (opcode > BPF_END) {
13702 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13703 return -EINVAL;
13704
13705 } else { /* all other ALU ops: and, sub, xor, add, ... */
13706
13707 if (BPF_SRC(insn->code) == BPF_X) {
13708 if (insn->imm != 0 || insn->off > 1 ||
13709 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13710 verbose(env, "BPF_ALU uses reserved fields\n");
13711 return -EINVAL;
13712 }
13713 /* check src1 operand */
13714 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13715 if (err)
13716 return err;
13717 } else {
13718 if (insn->src_reg != BPF_REG_0 || insn->off > 1 ||
13719 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13720 verbose(env, "BPF_ALU uses reserved fields\n");
13721 return -EINVAL;
13722 }
13723 }
13724
13725 /* check src2 operand */
13726 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13727 if (err)
13728 return err;
13729
13730 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13731 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13732 verbose(env, "div by zero\n");
13733 return -EINVAL;
13734 }
13735
13736 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13737 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13738 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13739
13740 if (insn->imm < 0 || insn->imm >= size) {
13741 verbose(env, "invalid shift %d\n", insn->imm);
13742 return -EINVAL;
13743 }
13744 }
13745
13746 /* check dest operand */
13747 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13748 if (err)
13749 return err;
13750
13751 return adjust_reg_min_max_vals(env, insn);
13752 }
13753
13754 return 0;
13755 }
13756
find_good_pkt_pointers(struct bpf_verifier_state * vstate,struct bpf_reg_state * dst_reg,enum bpf_reg_type type,bool range_right_open)13757 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13758 struct bpf_reg_state *dst_reg,
13759 enum bpf_reg_type type,
13760 bool range_right_open)
13761 {
13762 struct bpf_func_state *state;
13763 struct bpf_reg_state *reg;
13764 int new_range;
13765
13766 if (dst_reg->off < 0 ||
13767 (dst_reg->off == 0 && range_right_open))
13768 /* This doesn't give us any range */
13769 return;
13770
13771 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13772 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13773 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13774 * than pkt_end, but that's because it's also less than pkt.
13775 */
13776 return;
13777
13778 new_range = dst_reg->off;
13779 if (range_right_open)
13780 new_range++;
13781
13782 /* Examples for register markings:
13783 *
13784 * pkt_data in dst register:
13785 *
13786 * r2 = r3;
13787 * r2 += 8;
13788 * if (r2 > pkt_end) goto <handle exception>
13789 * <access okay>
13790 *
13791 * r2 = r3;
13792 * r2 += 8;
13793 * if (r2 < pkt_end) goto <access okay>
13794 * <handle exception>
13795 *
13796 * Where:
13797 * r2 == dst_reg, pkt_end == src_reg
13798 * r2=pkt(id=n,off=8,r=0)
13799 * r3=pkt(id=n,off=0,r=0)
13800 *
13801 * pkt_data in src register:
13802 *
13803 * r2 = r3;
13804 * r2 += 8;
13805 * if (pkt_end >= r2) goto <access okay>
13806 * <handle exception>
13807 *
13808 * r2 = r3;
13809 * r2 += 8;
13810 * if (pkt_end <= r2) goto <handle exception>
13811 * <access okay>
13812 *
13813 * Where:
13814 * pkt_end == dst_reg, r2 == src_reg
13815 * r2=pkt(id=n,off=8,r=0)
13816 * r3=pkt(id=n,off=0,r=0)
13817 *
13818 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13819 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13820 * and [r3, r3 + 8-1) respectively is safe to access depending on
13821 * the check.
13822 */
13823
13824 /* If our ids match, then we must have the same max_value. And we
13825 * don't care about the other reg's fixed offset, since if it's too big
13826 * the range won't allow anything.
13827 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13828 */
13829 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13830 if (reg->type == type && reg->id == dst_reg->id)
13831 /* keep the maximum range already checked */
13832 reg->range = max(reg->range, new_range);
13833 }));
13834 }
13835
is_branch32_taken(struct bpf_reg_state * reg,u32 val,u8 opcode)13836 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13837 {
13838 struct tnum subreg = tnum_subreg(reg->var_off);
13839 s32 sval = (s32)val;
13840
13841 switch (opcode) {
13842 case BPF_JEQ:
13843 if (tnum_is_const(subreg))
13844 return !!tnum_equals_const(subreg, val);
13845 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13846 return 0;
13847 break;
13848 case BPF_JNE:
13849 if (tnum_is_const(subreg))
13850 return !tnum_equals_const(subreg, val);
13851 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13852 return 1;
13853 break;
13854 case BPF_JSET:
13855 if ((~subreg.mask & subreg.value) & val)
13856 return 1;
13857 if (!((subreg.mask | subreg.value) & val))
13858 return 0;
13859 break;
13860 case BPF_JGT:
13861 if (reg->u32_min_value > val)
13862 return 1;
13863 else if (reg->u32_max_value <= val)
13864 return 0;
13865 break;
13866 case BPF_JSGT:
13867 if (reg->s32_min_value > sval)
13868 return 1;
13869 else if (reg->s32_max_value <= sval)
13870 return 0;
13871 break;
13872 case BPF_JLT:
13873 if (reg->u32_max_value < val)
13874 return 1;
13875 else if (reg->u32_min_value >= val)
13876 return 0;
13877 break;
13878 case BPF_JSLT:
13879 if (reg->s32_max_value < sval)
13880 return 1;
13881 else if (reg->s32_min_value >= sval)
13882 return 0;
13883 break;
13884 case BPF_JGE:
13885 if (reg->u32_min_value >= val)
13886 return 1;
13887 else if (reg->u32_max_value < val)
13888 return 0;
13889 break;
13890 case BPF_JSGE:
13891 if (reg->s32_min_value >= sval)
13892 return 1;
13893 else if (reg->s32_max_value < sval)
13894 return 0;
13895 break;
13896 case BPF_JLE:
13897 if (reg->u32_max_value <= val)
13898 return 1;
13899 else if (reg->u32_min_value > val)
13900 return 0;
13901 break;
13902 case BPF_JSLE:
13903 if (reg->s32_max_value <= sval)
13904 return 1;
13905 else if (reg->s32_min_value > sval)
13906 return 0;
13907 break;
13908 }
13909
13910 return -1;
13911 }
13912
13913
is_branch64_taken(struct bpf_reg_state * reg,u64 val,u8 opcode)13914 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13915 {
13916 s64 sval = (s64)val;
13917
13918 switch (opcode) {
13919 case BPF_JEQ:
13920 if (tnum_is_const(reg->var_off))
13921 return !!tnum_equals_const(reg->var_off, val);
13922 else if (val < reg->umin_value || val > reg->umax_value)
13923 return 0;
13924 break;
13925 case BPF_JNE:
13926 if (tnum_is_const(reg->var_off))
13927 return !tnum_equals_const(reg->var_off, val);
13928 else if (val < reg->umin_value || val > reg->umax_value)
13929 return 1;
13930 break;
13931 case BPF_JSET:
13932 if ((~reg->var_off.mask & reg->var_off.value) & val)
13933 return 1;
13934 if (!((reg->var_off.mask | reg->var_off.value) & val))
13935 return 0;
13936 break;
13937 case BPF_JGT:
13938 if (reg->umin_value > val)
13939 return 1;
13940 else if (reg->umax_value <= val)
13941 return 0;
13942 break;
13943 case BPF_JSGT:
13944 if (reg->smin_value > sval)
13945 return 1;
13946 else if (reg->smax_value <= sval)
13947 return 0;
13948 break;
13949 case BPF_JLT:
13950 if (reg->umax_value < val)
13951 return 1;
13952 else if (reg->umin_value >= val)
13953 return 0;
13954 break;
13955 case BPF_JSLT:
13956 if (reg->smax_value < sval)
13957 return 1;
13958 else if (reg->smin_value >= sval)
13959 return 0;
13960 break;
13961 case BPF_JGE:
13962 if (reg->umin_value >= val)
13963 return 1;
13964 else if (reg->umax_value < val)
13965 return 0;
13966 break;
13967 case BPF_JSGE:
13968 if (reg->smin_value >= sval)
13969 return 1;
13970 else if (reg->smax_value < sval)
13971 return 0;
13972 break;
13973 case BPF_JLE:
13974 if (reg->umax_value <= val)
13975 return 1;
13976 else if (reg->umin_value > val)
13977 return 0;
13978 break;
13979 case BPF_JSLE:
13980 if (reg->smax_value <= sval)
13981 return 1;
13982 else if (reg->smin_value > sval)
13983 return 0;
13984 break;
13985 }
13986
13987 return -1;
13988 }
13989
13990 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13991 * and return:
13992 * 1 - branch will be taken and "goto target" will be executed
13993 * 0 - branch will not be taken and fall-through to next insn
13994 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13995 * range [0,10]
13996 */
is_branch_taken(struct bpf_reg_state * reg,u64 val,u8 opcode,bool is_jmp32)13997 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13998 bool is_jmp32)
13999 {
14000 if (__is_pointer_value(false, reg)) {
14001 if (!reg_not_null(reg))
14002 return -1;
14003
14004 /* If pointer is valid tests against zero will fail so we can
14005 * use this to direct branch taken.
14006 */
14007 if (val != 0)
14008 return -1;
14009
14010 switch (opcode) {
14011 case BPF_JEQ:
14012 return 0;
14013 case BPF_JNE:
14014 return 1;
14015 default:
14016 return -1;
14017 }
14018 }
14019
14020 if (is_jmp32)
14021 return is_branch32_taken(reg, val, opcode);
14022 return is_branch64_taken(reg, val, opcode);
14023 }
14024
flip_opcode(u32 opcode)14025 static int flip_opcode(u32 opcode)
14026 {
14027 /* How can we transform "a <op> b" into "b <op> a"? */
14028 static const u8 opcode_flip[16] = {
14029 /* these stay the same */
14030 [BPF_JEQ >> 4] = BPF_JEQ,
14031 [BPF_JNE >> 4] = BPF_JNE,
14032 [BPF_JSET >> 4] = BPF_JSET,
14033 /* these swap "lesser" and "greater" (L and G in the opcodes) */
14034 [BPF_JGE >> 4] = BPF_JLE,
14035 [BPF_JGT >> 4] = BPF_JLT,
14036 [BPF_JLE >> 4] = BPF_JGE,
14037 [BPF_JLT >> 4] = BPF_JGT,
14038 [BPF_JSGE >> 4] = BPF_JSLE,
14039 [BPF_JSGT >> 4] = BPF_JSLT,
14040 [BPF_JSLE >> 4] = BPF_JSGE,
14041 [BPF_JSLT >> 4] = BPF_JSGT
14042 };
14043 return opcode_flip[opcode >> 4];
14044 }
14045
is_pkt_ptr_branch_taken(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg,u8 opcode)14046 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
14047 struct bpf_reg_state *src_reg,
14048 u8 opcode)
14049 {
14050 struct bpf_reg_state *pkt;
14051
14052 if (src_reg->type == PTR_TO_PACKET_END) {
14053 pkt = dst_reg;
14054 } else if (dst_reg->type == PTR_TO_PACKET_END) {
14055 pkt = src_reg;
14056 opcode = flip_opcode(opcode);
14057 } else {
14058 return -1;
14059 }
14060
14061 if (pkt->range >= 0)
14062 return -1;
14063
14064 switch (opcode) {
14065 case BPF_JLE:
14066 /* pkt <= pkt_end */
14067 fallthrough;
14068 case BPF_JGT:
14069 /* pkt > pkt_end */
14070 if (pkt->range == BEYOND_PKT_END)
14071 /* pkt has at last one extra byte beyond pkt_end */
14072 return opcode == BPF_JGT;
14073 break;
14074 case BPF_JLT:
14075 /* pkt < pkt_end */
14076 fallthrough;
14077 case BPF_JGE:
14078 /* pkt >= pkt_end */
14079 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
14080 return opcode == BPF_JGE;
14081 break;
14082 }
14083 return -1;
14084 }
14085
14086 /* Adjusts the register min/max values in the case that the dst_reg is the
14087 * variable register that we are working on, and src_reg is a constant or we're
14088 * simply doing a BPF_K check.
14089 * In JEQ/JNE cases we also adjust the var_off values.
14090 */
reg_set_min_max(struct bpf_reg_state * true_reg,struct bpf_reg_state * false_reg,u64 val,u32 val32,u8 opcode,bool is_jmp32)14091 static void reg_set_min_max(struct bpf_reg_state *true_reg,
14092 struct bpf_reg_state *false_reg,
14093 u64 val, u32 val32,
14094 u8 opcode, bool is_jmp32)
14095 {
14096 struct tnum false_32off = tnum_subreg(false_reg->var_off);
14097 struct tnum false_64off = false_reg->var_off;
14098 struct tnum true_32off = tnum_subreg(true_reg->var_off);
14099 struct tnum true_64off = true_reg->var_off;
14100 s64 sval = (s64)val;
14101 s32 sval32 = (s32)val32;
14102
14103 /* If the dst_reg is a pointer, we can't learn anything about its
14104 * variable offset from the compare (unless src_reg were a pointer into
14105 * the same object, but we don't bother with that.
14106 * Since false_reg and true_reg have the same type by construction, we
14107 * only need to check one of them for pointerness.
14108 */
14109 if (__is_pointer_value(false, false_reg))
14110 return;
14111
14112 switch (opcode) {
14113 /* JEQ/JNE comparison doesn't change the register equivalence.
14114 *
14115 * r1 = r2;
14116 * if (r1 == 42) goto label;
14117 * ...
14118 * label: // here both r1 and r2 are known to be 42.
14119 *
14120 * Hence when marking register as known preserve it's ID.
14121 */
14122 case BPF_JEQ:
14123 if (is_jmp32) {
14124 __mark_reg32_known(true_reg, val32);
14125 true_32off = tnum_subreg(true_reg->var_off);
14126 } else {
14127 ___mark_reg_known(true_reg, val);
14128 true_64off = true_reg->var_off;
14129 }
14130 break;
14131 case BPF_JNE:
14132 if (is_jmp32) {
14133 __mark_reg32_known(false_reg, val32);
14134 false_32off = tnum_subreg(false_reg->var_off);
14135 } else {
14136 ___mark_reg_known(false_reg, val);
14137 false_64off = false_reg->var_off;
14138 }
14139 break;
14140 case BPF_JSET:
14141 if (is_jmp32) {
14142 false_32off = tnum_and(false_32off, tnum_const(~val32));
14143 if (is_power_of_2(val32))
14144 true_32off = tnum_or(true_32off,
14145 tnum_const(val32));
14146 } else {
14147 false_64off = tnum_and(false_64off, tnum_const(~val));
14148 if (is_power_of_2(val))
14149 true_64off = tnum_or(true_64off,
14150 tnum_const(val));
14151 }
14152 break;
14153 case BPF_JGE:
14154 case BPF_JGT:
14155 {
14156 if (is_jmp32) {
14157 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
14158 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
14159
14160 false_reg->u32_max_value = min(false_reg->u32_max_value,
14161 false_umax);
14162 true_reg->u32_min_value = max(true_reg->u32_min_value,
14163 true_umin);
14164 } else {
14165 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
14166 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
14167
14168 false_reg->umax_value = min(false_reg->umax_value, false_umax);
14169 true_reg->umin_value = max(true_reg->umin_value, true_umin);
14170 }
14171 break;
14172 }
14173 case BPF_JSGE:
14174 case BPF_JSGT:
14175 {
14176 if (is_jmp32) {
14177 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
14178 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
14179
14180 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
14181 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
14182 } else {
14183 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
14184 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
14185
14186 false_reg->smax_value = min(false_reg->smax_value, false_smax);
14187 true_reg->smin_value = max(true_reg->smin_value, true_smin);
14188 }
14189 break;
14190 }
14191 case BPF_JLE:
14192 case BPF_JLT:
14193 {
14194 if (is_jmp32) {
14195 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
14196 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
14197
14198 false_reg->u32_min_value = max(false_reg->u32_min_value,
14199 false_umin);
14200 true_reg->u32_max_value = min(true_reg->u32_max_value,
14201 true_umax);
14202 } else {
14203 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
14204 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
14205
14206 false_reg->umin_value = max(false_reg->umin_value, false_umin);
14207 true_reg->umax_value = min(true_reg->umax_value, true_umax);
14208 }
14209 break;
14210 }
14211 case BPF_JSLE:
14212 case BPF_JSLT:
14213 {
14214 if (is_jmp32) {
14215 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
14216 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
14217
14218 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
14219 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
14220 } else {
14221 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
14222 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
14223
14224 false_reg->smin_value = max(false_reg->smin_value, false_smin);
14225 true_reg->smax_value = min(true_reg->smax_value, true_smax);
14226 }
14227 break;
14228 }
14229 default:
14230 return;
14231 }
14232
14233 if (is_jmp32) {
14234 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
14235 tnum_subreg(false_32off));
14236 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
14237 tnum_subreg(true_32off));
14238 __reg_combine_32_into_64(false_reg);
14239 __reg_combine_32_into_64(true_reg);
14240 } else {
14241 false_reg->var_off = false_64off;
14242 true_reg->var_off = true_64off;
14243 __reg_combine_64_into_32(false_reg);
14244 __reg_combine_64_into_32(true_reg);
14245 }
14246 }
14247
14248 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
14249 * the variable reg.
14250 */
reg_set_min_max_inv(struct bpf_reg_state * true_reg,struct bpf_reg_state * false_reg,u64 val,u32 val32,u8 opcode,bool is_jmp32)14251 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
14252 struct bpf_reg_state *false_reg,
14253 u64 val, u32 val32,
14254 u8 opcode, bool is_jmp32)
14255 {
14256 opcode = flip_opcode(opcode);
14257 /* This uses zero as "not present in table"; luckily the zero opcode,
14258 * BPF_JA, can't get here.
14259 */
14260 if (opcode)
14261 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
14262 }
14263
14264 /* Regs are known to be equal, so intersect their min/max/var_off */
__reg_combine_min_max(struct bpf_reg_state * src_reg,struct bpf_reg_state * dst_reg)14265 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
14266 struct bpf_reg_state *dst_reg)
14267 {
14268 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
14269 dst_reg->umin_value);
14270 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
14271 dst_reg->umax_value);
14272 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
14273 dst_reg->smin_value);
14274 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
14275 dst_reg->smax_value);
14276 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
14277 dst_reg->var_off);
14278 reg_bounds_sync(src_reg);
14279 reg_bounds_sync(dst_reg);
14280 }
14281
reg_combine_min_max(struct bpf_reg_state * true_src,struct bpf_reg_state * true_dst,struct bpf_reg_state * false_src,struct bpf_reg_state * false_dst,u8 opcode)14282 static void reg_combine_min_max(struct bpf_reg_state *true_src,
14283 struct bpf_reg_state *true_dst,
14284 struct bpf_reg_state *false_src,
14285 struct bpf_reg_state *false_dst,
14286 u8 opcode)
14287 {
14288 switch (opcode) {
14289 case BPF_JEQ:
14290 __reg_combine_min_max(true_src, true_dst);
14291 break;
14292 case BPF_JNE:
14293 __reg_combine_min_max(false_src, false_dst);
14294 break;
14295 }
14296 }
14297
mark_ptr_or_null_reg(struct bpf_func_state * state,struct bpf_reg_state * reg,u32 id,bool is_null)14298 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
14299 struct bpf_reg_state *reg, u32 id,
14300 bool is_null)
14301 {
14302 if (type_may_be_null(reg->type) && reg->id == id &&
14303 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
14304 /* Old offset (both fixed and variable parts) should have been
14305 * known-zero, because we don't allow pointer arithmetic on
14306 * pointers that might be NULL. If we see this happening, don't
14307 * convert the register.
14308 *
14309 * But in some cases, some helpers that return local kptrs
14310 * advance offset for the returned pointer. In those cases, it
14311 * is fine to expect to see reg->off.
14312 */
14313 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
14314 return;
14315 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
14316 WARN_ON_ONCE(reg->off))
14317 return;
14318
14319 if (is_null) {
14320 reg->type = SCALAR_VALUE;
14321 /* We don't need id and ref_obj_id from this point
14322 * onwards anymore, thus we should better reset it,
14323 * so that state pruning has chances to take effect.
14324 */
14325 reg->id = 0;
14326 reg->ref_obj_id = 0;
14327
14328 return;
14329 }
14330
14331 mark_ptr_not_null_reg(reg);
14332
14333 if (!reg_may_point_to_spin_lock(reg)) {
14334 /* For not-NULL ptr, reg->ref_obj_id will be reset
14335 * in release_reference().
14336 *
14337 * reg->id is still used by spin_lock ptr. Other
14338 * than spin_lock ptr type, reg->id can be reset.
14339 */
14340 reg->id = 0;
14341 }
14342 }
14343 }
14344
14345 /* The logic is similar to find_good_pkt_pointers(), both could eventually
14346 * be folded together at some point.
14347 */
mark_ptr_or_null_regs(struct bpf_verifier_state * vstate,u32 regno,bool is_null)14348 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
14349 bool is_null)
14350 {
14351 struct bpf_func_state *state = vstate->frame[vstate->curframe];
14352 struct bpf_reg_state *regs = state->regs, *reg;
14353 u32 ref_obj_id = regs[regno].ref_obj_id;
14354 u32 id = regs[regno].id;
14355
14356 if (ref_obj_id && ref_obj_id == id && is_null)
14357 /* regs[regno] is in the " == NULL" branch.
14358 * No one could have freed the reference state before
14359 * doing the NULL check.
14360 */
14361 WARN_ON_ONCE(release_reference_state(state, id));
14362
14363 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14364 mark_ptr_or_null_reg(state, reg, id, is_null);
14365 }));
14366 }
14367
try_match_pkt_pointers(const struct bpf_insn * insn,struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg,struct bpf_verifier_state * this_branch,struct bpf_verifier_state * other_branch)14368 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
14369 struct bpf_reg_state *dst_reg,
14370 struct bpf_reg_state *src_reg,
14371 struct bpf_verifier_state *this_branch,
14372 struct bpf_verifier_state *other_branch)
14373 {
14374 if (BPF_SRC(insn->code) != BPF_X)
14375 return false;
14376
14377 /* Pointers are always 64-bit. */
14378 if (BPF_CLASS(insn->code) == BPF_JMP32)
14379 return false;
14380
14381 switch (BPF_OP(insn->code)) {
14382 case BPF_JGT:
14383 if ((dst_reg->type == PTR_TO_PACKET &&
14384 src_reg->type == PTR_TO_PACKET_END) ||
14385 (dst_reg->type == PTR_TO_PACKET_META &&
14386 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14387 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
14388 find_good_pkt_pointers(this_branch, dst_reg,
14389 dst_reg->type, false);
14390 mark_pkt_end(other_branch, insn->dst_reg, true);
14391 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14392 src_reg->type == PTR_TO_PACKET) ||
14393 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14394 src_reg->type == PTR_TO_PACKET_META)) {
14395 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
14396 find_good_pkt_pointers(other_branch, src_reg,
14397 src_reg->type, true);
14398 mark_pkt_end(this_branch, insn->src_reg, false);
14399 } else {
14400 return false;
14401 }
14402 break;
14403 case BPF_JLT:
14404 if ((dst_reg->type == PTR_TO_PACKET &&
14405 src_reg->type == PTR_TO_PACKET_END) ||
14406 (dst_reg->type == PTR_TO_PACKET_META &&
14407 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14408 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
14409 find_good_pkt_pointers(other_branch, dst_reg,
14410 dst_reg->type, true);
14411 mark_pkt_end(this_branch, insn->dst_reg, false);
14412 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14413 src_reg->type == PTR_TO_PACKET) ||
14414 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14415 src_reg->type == PTR_TO_PACKET_META)) {
14416 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
14417 find_good_pkt_pointers(this_branch, src_reg,
14418 src_reg->type, false);
14419 mark_pkt_end(other_branch, insn->src_reg, true);
14420 } else {
14421 return false;
14422 }
14423 break;
14424 case BPF_JGE:
14425 if ((dst_reg->type == PTR_TO_PACKET &&
14426 src_reg->type == PTR_TO_PACKET_END) ||
14427 (dst_reg->type == PTR_TO_PACKET_META &&
14428 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14429 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
14430 find_good_pkt_pointers(this_branch, dst_reg,
14431 dst_reg->type, true);
14432 mark_pkt_end(other_branch, insn->dst_reg, false);
14433 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14434 src_reg->type == PTR_TO_PACKET) ||
14435 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14436 src_reg->type == PTR_TO_PACKET_META)) {
14437 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
14438 find_good_pkt_pointers(other_branch, src_reg,
14439 src_reg->type, false);
14440 mark_pkt_end(this_branch, insn->src_reg, true);
14441 } else {
14442 return false;
14443 }
14444 break;
14445 case BPF_JLE:
14446 if ((dst_reg->type == PTR_TO_PACKET &&
14447 src_reg->type == PTR_TO_PACKET_END) ||
14448 (dst_reg->type == PTR_TO_PACKET_META &&
14449 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14450 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
14451 find_good_pkt_pointers(other_branch, dst_reg,
14452 dst_reg->type, false);
14453 mark_pkt_end(this_branch, insn->dst_reg, true);
14454 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14455 src_reg->type == PTR_TO_PACKET) ||
14456 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14457 src_reg->type == PTR_TO_PACKET_META)) {
14458 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
14459 find_good_pkt_pointers(this_branch, src_reg,
14460 src_reg->type, true);
14461 mark_pkt_end(other_branch, insn->src_reg, false);
14462 } else {
14463 return false;
14464 }
14465 break;
14466 default:
14467 return false;
14468 }
14469
14470 return true;
14471 }
14472
find_equal_scalars(struct bpf_verifier_state * vstate,struct bpf_reg_state * known_reg)14473 static void find_equal_scalars(struct bpf_verifier_state *vstate,
14474 struct bpf_reg_state *known_reg)
14475 {
14476 struct bpf_func_state *state;
14477 struct bpf_reg_state *reg;
14478
14479 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14480 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
14481 copy_register_state(reg, known_reg);
14482 }));
14483 }
14484
check_cond_jmp_op(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)14485 static int check_cond_jmp_op(struct bpf_verifier_env *env,
14486 struct bpf_insn *insn, int *insn_idx)
14487 {
14488 struct bpf_verifier_state *this_branch = env->cur_state;
14489 struct bpf_verifier_state *other_branch;
14490 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
14491 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
14492 struct bpf_reg_state *eq_branch_regs;
14493 u8 opcode = BPF_OP(insn->code);
14494 bool is_jmp32;
14495 int pred = -1;
14496 int err;
14497
14498 /* Only conditional jumps are expected to reach here. */
14499 if (opcode == BPF_JA || opcode > BPF_JSLE) {
14500 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
14501 return -EINVAL;
14502 }
14503
14504 /* check src2 operand */
14505 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
14506 if (err)
14507 return err;
14508
14509 dst_reg = ®s[insn->dst_reg];
14510 if (BPF_SRC(insn->code) == BPF_X) {
14511 if (insn->imm != 0) {
14512 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14513 return -EINVAL;
14514 }
14515
14516 /* check src1 operand */
14517 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14518 if (err)
14519 return err;
14520
14521 src_reg = ®s[insn->src_reg];
14522 if (!(reg_is_pkt_pointer_any(dst_reg) && reg_is_pkt_pointer_any(src_reg)) &&
14523 is_pointer_value(env, insn->src_reg)) {
14524 verbose(env, "R%d pointer comparison prohibited\n",
14525 insn->src_reg);
14526 return -EACCES;
14527 }
14528 } else {
14529 if (insn->src_reg != BPF_REG_0) {
14530 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14531 return -EINVAL;
14532 }
14533 }
14534
14535 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
14536
14537 if (BPF_SRC(insn->code) == BPF_K) {
14538 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
14539 } else if (src_reg->type == SCALAR_VALUE &&
14540 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
14541 pred = is_branch_taken(dst_reg,
14542 tnum_subreg(src_reg->var_off).value,
14543 opcode,
14544 is_jmp32);
14545 } else if (src_reg->type == SCALAR_VALUE &&
14546 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
14547 pred = is_branch_taken(dst_reg,
14548 src_reg->var_off.value,
14549 opcode,
14550 is_jmp32);
14551 } else if (dst_reg->type == SCALAR_VALUE &&
14552 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
14553 pred = is_branch_taken(src_reg,
14554 tnum_subreg(dst_reg->var_off).value,
14555 flip_opcode(opcode),
14556 is_jmp32);
14557 } else if (dst_reg->type == SCALAR_VALUE &&
14558 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
14559 pred = is_branch_taken(src_reg,
14560 dst_reg->var_off.value,
14561 flip_opcode(opcode),
14562 is_jmp32);
14563 } else if (reg_is_pkt_pointer_any(dst_reg) &&
14564 reg_is_pkt_pointer_any(src_reg) &&
14565 !is_jmp32) {
14566 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
14567 }
14568
14569 if (pred >= 0) {
14570 /* If we get here with a dst_reg pointer type it is because
14571 * above is_branch_taken() special cased the 0 comparison.
14572 */
14573 if (!__is_pointer_value(false, dst_reg))
14574 err = mark_chain_precision(env, insn->dst_reg);
14575 if (BPF_SRC(insn->code) == BPF_X && !err &&
14576 !__is_pointer_value(false, src_reg))
14577 err = mark_chain_precision(env, insn->src_reg);
14578 if (err)
14579 return err;
14580 }
14581
14582 if (pred == 1) {
14583 /* Only follow the goto, ignore fall-through. If needed, push
14584 * the fall-through branch for simulation under speculative
14585 * execution.
14586 */
14587 if (!env->bypass_spec_v1 &&
14588 !sanitize_speculative_path(env, insn, *insn_idx + 1,
14589 *insn_idx))
14590 return -EFAULT;
14591 if (env->log.level & BPF_LOG_LEVEL)
14592 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14593 *insn_idx += insn->off;
14594 return 0;
14595 } else if (pred == 0) {
14596 /* Only follow the fall-through branch, since that's where the
14597 * program will go. If needed, push the goto branch for
14598 * simulation under speculative execution.
14599 */
14600 if (!env->bypass_spec_v1 &&
14601 !sanitize_speculative_path(env, insn,
14602 *insn_idx + insn->off + 1,
14603 *insn_idx))
14604 return -EFAULT;
14605 if (env->log.level & BPF_LOG_LEVEL)
14606 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14607 return 0;
14608 }
14609
14610 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
14611 false);
14612 if (!other_branch)
14613 return -EFAULT;
14614 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
14615
14616 /* detect if we are comparing against a constant value so we can adjust
14617 * our min/max values for our dst register.
14618 * this is only legit if both are scalars (or pointers to the same
14619 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
14620 * because otherwise the different base pointers mean the offsets aren't
14621 * comparable.
14622 */
14623 if (BPF_SRC(insn->code) == BPF_X) {
14624 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
14625
14626 if (dst_reg->type == SCALAR_VALUE &&
14627 src_reg->type == SCALAR_VALUE) {
14628 if (tnum_is_const(src_reg->var_off) ||
14629 (is_jmp32 &&
14630 tnum_is_const(tnum_subreg(src_reg->var_off))))
14631 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14632 dst_reg,
14633 src_reg->var_off.value,
14634 tnum_subreg(src_reg->var_off).value,
14635 opcode, is_jmp32);
14636 else if (tnum_is_const(dst_reg->var_off) ||
14637 (is_jmp32 &&
14638 tnum_is_const(tnum_subreg(dst_reg->var_off))))
14639 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
14640 src_reg,
14641 dst_reg->var_off.value,
14642 tnum_subreg(dst_reg->var_off).value,
14643 opcode, is_jmp32);
14644 else if (!is_jmp32 &&
14645 (opcode == BPF_JEQ || opcode == BPF_JNE))
14646 /* Comparing for equality, we can combine knowledge */
14647 reg_combine_min_max(&other_branch_regs[insn->src_reg],
14648 &other_branch_regs[insn->dst_reg],
14649 src_reg, dst_reg, opcode);
14650 if (src_reg->id &&
14651 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
14652 find_equal_scalars(this_branch, src_reg);
14653 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
14654 }
14655
14656 }
14657 } else if (dst_reg->type == SCALAR_VALUE) {
14658 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14659 dst_reg, insn->imm, (u32)insn->imm,
14660 opcode, is_jmp32);
14661 }
14662
14663 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
14664 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
14665 find_equal_scalars(this_branch, dst_reg);
14666 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
14667 }
14668
14669 /* if one pointer register is compared to another pointer
14670 * register check if PTR_MAYBE_NULL could be lifted.
14671 * E.g. register A - maybe null
14672 * register B - not null
14673 * for JNE A, B, ... - A is not null in the false branch;
14674 * for JEQ A, B, ... - A is not null in the true branch.
14675 *
14676 * Since PTR_TO_BTF_ID points to a kernel struct that does
14677 * not need to be null checked by the BPF program, i.e.,
14678 * could be null even without PTR_MAYBE_NULL marking, so
14679 * only propagate nullness when neither reg is that type.
14680 */
14681 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
14682 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
14683 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
14684 base_type(src_reg->type) != PTR_TO_BTF_ID &&
14685 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
14686 eq_branch_regs = NULL;
14687 switch (opcode) {
14688 case BPF_JEQ:
14689 eq_branch_regs = other_branch_regs;
14690 break;
14691 case BPF_JNE:
14692 eq_branch_regs = regs;
14693 break;
14694 default:
14695 /* do nothing */
14696 break;
14697 }
14698 if (eq_branch_regs) {
14699 if (type_may_be_null(src_reg->type))
14700 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
14701 else
14702 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14703 }
14704 }
14705
14706 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14707 * NOTE: these optimizations below are related with pointer comparison
14708 * which will never be JMP32.
14709 */
14710 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14711 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14712 type_may_be_null(dst_reg->type)) {
14713 /* Mark all identical registers in each branch as either
14714 * safe or unknown depending R == 0 or R != 0 conditional.
14715 */
14716 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14717 opcode == BPF_JNE);
14718 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14719 opcode == BPF_JEQ);
14720 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14721 this_branch, other_branch) &&
14722 is_pointer_value(env, insn->dst_reg)) {
14723 verbose(env, "R%d pointer comparison prohibited\n",
14724 insn->dst_reg);
14725 return -EACCES;
14726 }
14727 if (env->log.level & BPF_LOG_LEVEL)
14728 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14729 return 0;
14730 }
14731
14732 /* verify BPF_LD_IMM64 instruction */
check_ld_imm(struct bpf_verifier_env * env,struct bpf_insn * insn)14733 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14734 {
14735 struct bpf_insn_aux_data *aux = cur_aux(env);
14736 struct bpf_reg_state *regs = cur_regs(env);
14737 struct bpf_reg_state *dst_reg;
14738 struct bpf_map *map;
14739 int err;
14740
14741 if (BPF_SIZE(insn->code) != BPF_DW) {
14742 verbose(env, "invalid BPF_LD_IMM insn\n");
14743 return -EINVAL;
14744 }
14745 if (insn->off != 0) {
14746 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14747 return -EINVAL;
14748 }
14749
14750 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14751 if (err)
14752 return err;
14753
14754 dst_reg = ®s[insn->dst_reg];
14755 if (insn->src_reg == 0) {
14756 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14757
14758 dst_reg->type = SCALAR_VALUE;
14759 __mark_reg_known(®s[insn->dst_reg], imm);
14760 return 0;
14761 }
14762
14763 /* All special src_reg cases are listed below. From this point onwards
14764 * we either succeed and assign a corresponding dst_reg->type after
14765 * zeroing the offset, or fail and reject the program.
14766 */
14767 mark_reg_known_zero(env, regs, insn->dst_reg);
14768
14769 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14770 dst_reg->type = aux->btf_var.reg_type;
14771 switch (base_type(dst_reg->type)) {
14772 case PTR_TO_MEM:
14773 dst_reg->mem_size = aux->btf_var.mem_size;
14774 break;
14775 case PTR_TO_BTF_ID:
14776 dst_reg->btf = aux->btf_var.btf;
14777 dst_reg->btf_id = aux->btf_var.btf_id;
14778 break;
14779 default:
14780 verbose(env, "bpf verifier is misconfigured\n");
14781 return -EFAULT;
14782 }
14783 return 0;
14784 }
14785
14786 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14787 struct bpf_prog_aux *aux = env->prog->aux;
14788 u32 subprogno = find_subprog(env,
14789 env->insn_idx + insn->imm + 1);
14790
14791 if (!aux->func_info) {
14792 verbose(env, "missing btf func_info\n");
14793 return -EINVAL;
14794 }
14795 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14796 verbose(env, "callback function not static\n");
14797 return -EINVAL;
14798 }
14799
14800 dst_reg->type = PTR_TO_FUNC;
14801 dst_reg->subprogno = subprogno;
14802 return 0;
14803 }
14804
14805 map = env->used_maps[aux->map_index];
14806 dst_reg->map_ptr = map;
14807
14808 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14809 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14810 dst_reg->type = PTR_TO_MAP_VALUE;
14811 dst_reg->off = aux->map_off;
14812 WARN_ON_ONCE(map->max_entries != 1);
14813 /* We want reg->id to be same (0) as map_value is not distinct */
14814 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14815 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14816 dst_reg->type = CONST_PTR_TO_MAP;
14817 } else {
14818 verbose(env, "bpf verifier is misconfigured\n");
14819 return -EINVAL;
14820 }
14821
14822 return 0;
14823 }
14824
may_access_skb(enum bpf_prog_type type)14825 static bool may_access_skb(enum bpf_prog_type type)
14826 {
14827 switch (type) {
14828 case BPF_PROG_TYPE_SOCKET_FILTER:
14829 case BPF_PROG_TYPE_SCHED_CLS:
14830 case BPF_PROG_TYPE_SCHED_ACT:
14831 return true;
14832 default:
14833 return false;
14834 }
14835 }
14836
14837 /* verify safety of LD_ABS|LD_IND instructions:
14838 * - they can only appear in the programs where ctx == skb
14839 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14840 * preserve R6-R9, and store return value into R0
14841 *
14842 * Implicit input:
14843 * ctx == skb == R6 == CTX
14844 *
14845 * Explicit input:
14846 * SRC == any register
14847 * IMM == 32-bit immediate
14848 *
14849 * Output:
14850 * R0 - 8/16/32-bit skb data converted to cpu endianness
14851 */
check_ld_abs(struct bpf_verifier_env * env,struct bpf_insn * insn)14852 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14853 {
14854 struct bpf_reg_state *regs = cur_regs(env);
14855 static const int ctx_reg = BPF_REG_6;
14856 u8 mode = BPF_MODE(insn->code);
14857 int i, err;
14858
14859 if (!may_access_skb(resolve_prog_type(env->prog))) {
14860 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14861 return -EINVAL;
14862 }
14863
14864 if (!env->ops->gen_ld_abs) {
14865 verbose(env, "bpf verifier is misconfigured\n");
14866 return -EINVAL;
14867 }
14868
14869 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14870 BPF_SIZE(insn->code) == BPF_DW ||
14871 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14872 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14873 return -EINVAL;
14874 }
14875
14876 /* check whether implicit source operand (register R6) is readable */
14877 err = check_reg_arg(env, ctx_reg, SRC_OP);
14878 if (err)
14879 return err;
14880
14881 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14882 * gen_ld_abs() may terminate the program at runtime, leading to
14883 * reference leak.
14884 */
14885 err = check_reference_leak(env);
14886 if (err) {
14887 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14888 return err;
14889 }
14890
14891 if (env->cur_state->active_lock.ptr) {
14892 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14893 return -EINVAL;
14894 }
14895
14896 if (env->cur_state->active_rcu_lock) {
14897 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14898 return -EINVAL;
14899 }
14900
14901 if (regs[ctx_reg].type != PTR_TO_CTX) {
14902 verbose(env,
14903 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14904 return -EINVAL;
14905 }
14906
14907 if (mode == BPF_IND) {
14908 /* check explicit source operand */
14909 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14910 if (err)
14911 return err;
14912 }
14913
14914 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14915 if (err < 0)
14916 return err;
14917
14918 /* reset caller saved regs to unreadable */
14919 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14920 mark_reg_not_init(env, regs, caller_saved[i]);
14921 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14922 }
14923
14924 /* mark destination R0 register as readable, since it contains
14925 * the value fetched from the packet.
14926 * Already marked as written above.
14927 */
14928 mark_reg_unknown(env, regs, BPF_REG_0);
14929 /* ld_abs load up to 32-bit skb data. */
14930 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14931 return 0;
14932 }
14933
check_return_code(struct bpf_verifier_env * env)14934 static int check_return_code(struct bpf_verifier_env *env)
14935 {
14936 struct tnum enforce_attach_type_range = tnum_unknown;
14937 const struct bpf_prog *prog = env->prog;
14938 struct bpf_reg_state *reg;
14939 struct tnum range = tnum_range(0, 1), const_0 = tnum_const(0);
14940 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14941 int err;
14942 struct bpf_func_state *frame = env->cur_state->frame[0];
14943 const bool is_subprog = frame->subprogno;
14944
14945 /* LSM and struct_ops func-ptr's return type could be "void" */
14946 if (!is_subprog) {
14947 switch (prog_type) {
14948 case BPF_PROG_TYPE_LSM:
14949 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14950 /* See below, can be 0 or 0-1 depending on hook. */
14951 break;
14952 fallthrough;
14953 case BPF_PROG_TYPE_STRUCT_OPS:
14954 if (!prog->aux->attach_func_proto->type)
14955 return 0;
14956 break;
14957 default:
14958 break;
14959 }
14960 }
14961
14962 /* eBPF calling convention is such that R0 is used
14963 * to return the value from eBPF program.
14964 * Make sure that it's readable at this time
14965 * of bpf_exit, which means that program wrote
14966 * something into it earlier
14967 */
14968 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14969 if (err)
14970 return err;
14971
14972 if (is_pointer_value(env, BPF_REG_0)) {
14973 verbose(env, "R0 leaks addr as return value\n");
14974 return -EACCES;
14975 }
14976
14977 reg = cur_regs(env) + BPF_REG_0;
14978
14979 if (frame->in_async_callback_fn) {
14980 /* enforce return zero from async callbacks like timer */
14981 if (reg->type != SCALAR_VALUE) {
14982 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14983 reg_type_str(env, reg->type));
14984 return -EINVAL;
14985 }
14986
14987 if (!tnum_in(const_0, reg->var_off)) {
14988 verbose_invalid_scalar(env, reg, &const_0, "async callback", "R0");
14989 return -EINVAL;
14990 }
14991 return 0;
14992 }
14993
14994 if (is_subprog) {
14995 if (reg->type != SCALAR_VALUE) {
14996 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14997 reg_type_str(env, reg->type));
14998 return -EINVAL;
14999 }
15000 return 0;
15001 }
15002
15003 switch (prog_type) {
15004 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
15005 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
15006 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
15007 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
15008 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
15009 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
15010 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
15011 range = tnum_range(1, 1);
15012 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
15013 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
15014 range = tnum_range(0, 3);
15015 break;
15016 case BPF_PROG_TYPE_CGROUP_SKB:
15017 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
15018 range = tnum_range(0, 3);
15019 enforce_attach_type_range = tnum_range(2, 3);
15020 }
15021 break;
15022 case BPF_PROG_TYPE_CGROUP_SOCK:
15023 case BPF_PROG_TYPE_SOCK_OPS:
15024 case BPF_PROG_TYPE_CGROUP_DEVICE:
15025 case BPF_PROG_TYPE_CGROUP_SYSCTL:
15026 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
15027 break;
15028 case BPF_PROG_TYPE_RAW_TRACEPOINT:
15029 if (!env->prog->aux->attach_btf_id)
15030 return 0;
15031 range = tnum_const(0);
15032 break;
15033 case BPF_PROG_TYPE_TRACING:
15034 switch (env->prog->expected_attach_type) {
15035 case BPF_TRACE_FENTRY:
15036 case BPF_TRACE_FEXIT:
15037 range = tnum_const(0);
15038 break;
15039 case BPF_TRACE_RAW_TP:
15040 case BPF_MODIFY_RETURN:
15041 return 0;
15042 case BPF_TRACE_ITER:
15043 break;
15044 default:
15045 return -ENOTSUPP;
15046 }
15047 break;
15048 case BPF_PROG_TYPE_SK_LOOKUP:
15049 range = tnum_range(SK_DROP, SK_PASS);
15050 break;
15051
15052 case BPF_PROG_TYPE_LSM:
15053 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
15054 /* Regular BPF_PROG_TYPE_LSM programs can return
15055 * any value.
15056 */
15057 return 0;
15058 }
15059 if (!env->prog->aux->attach_func_proto->type) {
15060 /* Make sure programs that attach to void
15061 * hooks don't try to modify return value.
15062 */
15063 range = tnum_range(1, 1);
15064 }
15065 break;
15066
15067 case BPF_PROG_TYPE_NETFILTER:
15068 range = tnum_range(NF_DROP, NF_ACCEPT);
15069 break;
15070 case BPF_PROG_TYPE_EXT:
15071 /* freplace program can return anything as its return value
15072 * depends on the to-be-replaced kernel func or bpf program.
15073 */
15074 default:
15075 return 0;
15076 }
15077
15078 if (reg->type != SCALAR_VALUE) {
15079 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
15080 reg_type_str(env, reg->type));
15081 return -EINVAL;
15082 }
15083
15084 if (!tnum_in(range, reg->var_off)) {
15085 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
15086 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
15087 prog_type == BPF_PROG_TYPE_LSM &&
15088 !prog->aux->attach_func_proto->type)
15089 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
15090 return -EINVAL;
15091 }
15092
15093 if (!tnum_is_unknown(enforce_attach_type_range) &&
15094 tnum_in(enforce_attach_type_range, reg->var_off))
15095 env->prog->enforce_expected_attach_type = 1;
15096 return 0;
15097 }
15098
15099 /* non-recursive DFS pseudo code
15100 * 1 procedure DFS-iterative(G,v):
15101 * 2 label v as discovered
15102 * 3 let S be a stack
15103 * 4 S.push(v)
15104 * 5 while S is not empty
15105 * 6 t <- S.peek()
15106 * 7 if t is what we're looking for:
15107 * 8 return t
15108 * 9 for all edges e in G.adjacentEdges(t) do
15109 * 10 if edge e is already labelled
15110 * 11 continue with the next edge
15111 * 12 w <- G.adjacentVertex(t,e)
15112 * 13 if vertex w is not discovered and not explored
15113 * 14 label e as tree-edge
15114 * 15 label w as discovered
15115 * 16 S.push(w)
15116 * 17 continue at 5
15117 * 18 else if vertex w is discovered
15118 * 19 label e as back-edge
15119 * 20 else
15120 * 21 // vertex w is explored
15121 * 22 label e as forward- or cross-edge
15122 * 23 label t as explored
15123 * 24 S.pop()
15124 *
15125 * convention:
15126 * 0x10 - discovered
15127 * 0x11 - discovered and fall-through edge labelled
15128 * 0x12 - discovered and fall-through and branch edges labelled
15129 * 0x20 - explored
15130 */
15131
15132 enum {
15133 DISCOVERED = 0x10,
15134 EXPLORED = 0x20,
15135 FALLTHROUGH = 1,
15136 BRANCH = 2,
15137 };
15138
mark_prune_point(struct bpf_verifier_env * env,int idx)15139 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
15140 {
15141 env->insn_aux_data[idx].prune_point = true;
15142 }
15143
is_prune_point(struct bpf_verifier_env * env,int insn_idx)15144 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
15145 {
15146 return env->insn_aux_data[insn_idx].prune_point;
15147 }
15148
mark_force_checkpoint(struct bpf_verifier_env * env,int idx)15149 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
15150 {
15151 env->insn_aux_data[idx].force_checkpoint = true;
15152 }
15153
is_force_checkpoint(struct bpf_verifier_env * env,int insn_idx)15154 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
15155 {
15156 return env->insn_aux_data[insn_idx].force_checkpoint;
15157 }
15158
mark_calls_callback(struct bpf_verifier_env * env,int idx)15159 static void mark_calls_callback(struct bpf_verifier_env *env, int idx)
15160 {
15161 env->insn_aux_data[idx].calls_callback = true;
15162 }
15163
calls_callback(struct bpf_verifier_env * env,int insn_idx)15164 static bool calls_callback(struct bpf_verifier_env *env, int insn_idx)
15165 {
15166 return env->insn_aux_data[insn_idx].calls_callback;
15167 }
15168
15169 enum {
15170 DONE_EXPLORING = 0,
15171 KEEP_EXPLORING = 1,
15172 };
15173
15174 /* t, w, e - match pseudo-code above:
15175 * t - index of current instruction
15176 * w - next instruction
15177 * e - edge
15178 */
push_insn(int t,int w,int e,struct bpf_verifier_env * env)15179 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
15180 {
15181 int *insn_stack = env->cfg.insn_stack;
15182 int *insn_state = env->cfg.insn_state;
15183
15184 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
15185 return DONE_EXPLORING;
15186
15187 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
15188 return DONE_EXPLORING;
15189
15190 if (w < 0 || w >= env->prog->len) {
15191 verbose_linfo(env, t, "%d: ", t);
15192 verbose(env, "jump out of range from insn %d to %d\n", t, w);
15193 return -EINVAL;
15194 }
15195
15196 if (e == BRANCH) {
15197 /* mark branch target for state pruning */
15198 mark_prune_point(env, w);
15199 mark_jmp_point(env, w);
15200 }
15201
15202 if (insn_state[w] == 0) {
15203 /* tree-edge */
15204 insn_state[t] = DISCOVERED | e;
15205 insn_state[w] = DISCOVERED;
15206 if (env->cfg.cur_stack >= env->prog->len)
15207 return -E2BIG;
15208 insn_stack[env->cfg.cur_stack++] = w;
15209 return KEEP_EXPLORING;
15210 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
15211 if (env->bpf_capable)
15212 return DONE_EXPLORING;
15213 verbose_linfo(env, t, "%d: ", t);
15214 verbose_linfo(env, w, "%d: ", w);
15215 verbose(env, "back-edge from insn %d to %d\n", t, w);
15216 return -EINVAL;
15217 } else if (insn_state[w] == EXPLORED) {
15218 /* forward- or cross-edge */
15219 insn_state[t] = DISCOVERED | e;
15220 } else {
15221 verbose(env, "insn state internal bug\n");
15222 return -EFAULT;
15223 }
15224 return DONE_EXPLORING;
15225 }
15226
visit_func_call_insn(int t,struct bpf_insn * insns,struct bpf_verifier_env * env,bool visit_callee)15227 static int visit_func_call_insn(int t, struct bpf_insn *insns,
15228 struct bpf_verifier_env *env,
15229 bool visit_callee)
15230 {
15231 int ret, insn_sz;
15232
15233 insn_sz = bpf_is_ldimm64(&insns[t]) ? 2 : 1;
15234 ret = push_insn(t, t + insn_sz, FALLTHROUGH, env);
15235 if (ret)
15236 return ret;
15237
15238 mark_prune_point(env, t + insn_sz);
15239 /* when we exit from subprog, we need to record non-linear history */
15240 mark_jmp_point(env, t + insn_sz);
15241
15242 if (visit_callee) {
15243 mark_prune_point(env, t);
15244 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
15245 }
15246 return ret;
15247 }
15248
15249 /* Visits the instruction at index t and returns one of the following:
15250 * < 0 - an error occurred
15251 * DONE_EXPLORING - the instruction was fully explored
15252 * KEEP_EXPLORING - there is still work to be done before it is fully explored
15253 */
visit_insn(int t,struct bpf_verifier_env * env)15254 static int visit_insn(int t, struct bpf_verifier_env *env)
15255 {
15256 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
15257 int ret, off, insn_sz;
15258
15259 if (bpf_pseudo_func(insn))
15260 return visit_func_call_insn(t, insns, env, true);
15261
15262 /* All non-branch instructions have a single fall-through edge. */
15263 if (BPF_CLASS(insn->code) != BPF_JMP &&
15264 BPF_CLASS(insn->code) != BPF_JMP32) {
15265 insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
15266 return push_insn(t, t + insn_sz, FALLTHROUGH, env);
15267 }
15268
15269 switch (BPF_OP(insn->code)) {
15270 case BPF_EXIT:
15271 return DONE_EXPLORING;
15272
15273 case BPF_CALL:
15274 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
15275 /* Mark this call insn as a prune point to trigger
15276 * is_state_visited() check before call itself is
15277 * processed by __check_func_call(). Otherwise new
15278 * async state will be pushed for further exploration.
15279 */
15280 mark_prune_point(env, t);
15281 /* For functions that invoke callbacks it is not known how many times
15282 * callback would be called. Verifier models callback calling functions
15283 * by repeatedly visiting callback bodies and returning to origin call
15284 * instruction.
15285 * In order to stop such iteration verifier needs to identify when a
15286 * state identical some state from a previous iteration is reached.
15287 * Check below forces creation of checkpoint before callback calling
15288 * instruction to allow search for such identical states.
15289 */
15290 if (is_sync_callback_calling_insn(insn)) {
15291 mark_calls_callback(env, t);
15292 mark_force_checkpoint(env, t);
15293 mark_prune_point(env, t);
15294 mark_jmp_point(env, t);
15295 }
15296 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
15297 struct bpf_kfunc_call_arg_meta meta;
15298
15299 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
15300 if (ret == 0 && is_iter_next_kfunc(&meta)) {
15301 mark_prune_point(env, t);
15302 /* Checking and saving state checkpoints at iter_next() call
15303 * is crucial for fast convergence of open-coded iterator loop
15304 * logic, so we need to force it. If we don't do that,
15305 * is_state_visited() might skip saving a checkpoint, causing
15306 * unnecessarily long sequence of not checkpointed
15307 * instructions and jumps, leading to exhaustion of jump
15308 * history buffer, and potentially other undesired outcomes.
15309 * It is expected that with correct open-coded iterators
15310 * convergence will happen quickly, so we don't run a risk of
15311 * exhausting memory.
15312 */
15313 mark_force_checkpoint(env, t);
15314 }
15315 }
15316 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
15317
15318 case BPF_JA:
15319 if (BPF_SRC(insn->code) != BPF_K)
15320 return -EINVAL;
15321
15322 if (BPF_CLASS(insn->code) == BPF_JMP)
15323 off = insn->off;
15324 else
15325 off = insn->imm;
15326
15327 /* unconditional jump with single edge */
15328 ret = push_insn(t, t + off + 1, FALLTHROUGH, env);
15329 if (ret)
15330 return ret;
15331
15332 mark_prune_point(env, t + off + 1);
15333 mark_jmp_point(env, t + off + 1);
15334
15335 return ret;
15336
15337 default:
15338 /* conditional jump with two edges */
15339 mark_prune_point(env, t);
15340
15341 ret = push_insn(t, t + 1, FALLTHROUGH, env);
15342 if (ret)
15343 return ret;
15344
15345 return push_insn(t, t + insn->off + 1, BRANCH, env);
15346 }
15347 }
15348
15349 /* non-recursive depth-first-search to detect loops in BPF program
15350 * loop == back-edge in directed graph
15351 */
check_cfg(struct bpf_verifier_env * env)15352 static int check_cfg(struct bpf_verifier_env *env)
15353 {
15354 int insn_cnt = env->prog->len;
15355 int *insn_stack, *insn_state;
15356 int ret = 0;
15357 int i;
15358
15359 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
15360 if (!insn_state)
15361 return -ENOMEM;
15362
15363 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
15364 if (!insn_stack) {
15365 kvfree(insn_state);
15366 return -ENOMEM;
15367 }
15368
15369 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
15370 insn_stack[0] = 0; /* 0 is the first instruction */
15371 env->cfg.cur_stack = 1;
15372
15373 while (env->cfg.cur_stack > 0) {
15374 int t = insn_stack[env->cfg.cur_stack - 1];
15375
15376 ret = visit_insn(t, env);
15377 switch (ret) {
15378 case DONE_EXPLORING:
15379 insn_state[t] = EXPLORED;
15380 env->cfg.cur_stack--;
15381 break;
15382 case KEEP_EXPLORING:
15383 break;
15384 default:
15385 if (ret > 0) {
15386 verbose(env, "visit_insn internal bug\n");
15387 ret = -EFAULT;
15388 }
15389 goto err_free;
15390 }
15391 }
15392
15393 if (env->cfg.cur_stack < 0) {
15394 verbose(env, "pop stack internal bug\n");
15395 ret = -EFAULT;
15396 goto err_free;
15397 }
15398
15399 for (i = 0; i < insn_cnt; i++) {
15400 struct bpf_insn *insn = &env->prog->insnsi[i];
15401
15402 if (insn_state[i] != EXPLORED) {
15403 verbose(env, "unreachable insn %d\n", i);
15404 ret = -EINVAL;
15405 goto err_free;
15406 }
15407 if (bpf_is_ldimm64(insn)) {
15408 if (insn_state[i + 1] != 0) {
15409 verbose(env, "jump into the middle of ldimm64 insn %d\n", i);
15410 ret = -EINVAL;
15411 goto err_free;
15412 }
15413 i++; /* skip second half of ldimm64 */
15414 }
15415 }
15416 ret = 0; /* cfg looks good */
15417
15418 err_free:
15419 kvfree(insn_state);
15420 kvfree(insn_stack);
15421 env->cfg.insn_state = env->cfg.insn_stack = NULL;
15422 return ret;
15423 }
15424
check_abnormal_return(struct bpf_verifier_env * env)15425 static int check_abnormal_return(struct bpf_verifier_env *env)
15426 {
15427 int i;
15428
15429 for (i = 1; i < env->subprog_cnt; i++) {
15430 if (env->subprog_info[i].has_ld_abs) {
15431 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
15432 return -EINVAL;
15433 }
15434 if (env->subprog_info[i].has_tail_call) {
15435 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
15436 return -EINVAL;
15437 }
15438 }
15439 return 0;
15440 }
15441
15442 /* The minimum supported BTF func info size */
15443 #define MIN_BPF_FUNCINFO_SIZE 8
15444 #define MAX_FUNCINFO_REC_SIZE 252
15445
check_btf_func(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15446 static int check_btf_func(struct bpf_verifier_env *env,
15447 const union bpf_attr *attr,
15448 bpfptr_t uattr)
15449 {
15450 const struct btf_type *type, *func_proto, *ret_type;
15451 u32 i, nfuncs, urec_size, min_size;
15452 u32 krec_size = sizeof(struct bpf_func_info);
15453 struct bpf_func_info *krecord;
15454 struct bpf_func_info_aux *info_aux = NULL;
15455 struct bpf_prog *prog;
15456 const struct btf *btf;
15457 bpfptr_t urecord;
15458 u32 prev_offset = 0;
15459 bool scalar_return;
15460 int ret = -ENOMEM;
15461
15462 nfuncs = attr->func_info_cnt;
15463 if (!nfuncs) {
15464 if (check_abnormal_return(env))
15465 return -EINVAL;
15466 return 0;
15467 }
15468
15469 if (nfuncs != env->subprog_cnt) {
15470 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
15471 return -EINVAL;
15472 }
15473
15474 urec_size = attr->func_info_rec_size;
15475 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
15476 urec_size > MAX_FUNCINFO_REC_SIZE ||
15477 urec_size % sizeof(u32)) {
15478 verbose(env, "invalid func info rec size %u\n", urec_size);
15479 return -EINVAL;
15480 }
15481
15482 prog = env->prog;
15483 btf = prog->aux->btf;
15484
15485 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
15486 min_size = min_t(u32, krec_size, urec_size);
15487
15488 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
15489 if (!krecord)
15490 return -ENOMEM;
15491 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
15492 if (!info_aux)
15493 goto err_free;
15494
15495 for (i = 0; i < nfuncs; i++) {
15496 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
15497 if (ret) {
15498 if (ret == -E2BIG) {
15499 verbose(env, "nonzero tailing record in func info");
15500 /* set the size kernel expects so loader can zero
15501 * out the rest of the record.
15502 */
15503 if (copy_to_bpfptr_offset(uattr,
15504 offsetof(union bpf_attr, func_info_rec_size),
15505 &min_size, sizeof(min_size)))
15506 ret = -EFAULT;
15507 }
15508 goto err_free;
15509 }
15510
15511 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
15512 ret = -EFAULT;
15513 goto err_free;
15514 }
15515
15516 /* check insn_off */
15517 ret = -EINVAL;
15518 if (i == 0) {
15519 if (krecord[i].insn_off) {
15520 verbose(env,
15521 "nonzero insn_off %u for the first func info record",
15522 krecord[i].insn_off);
15523 goto err_free;
15524 }
15525 } else if (krecord[i].insn_off <= prev_offset) {
15526 verbose(env,
15527 "same or smaller insn offset (%u) than previous func info record (%u)",
15528 krecord[i].insn_off, prev_offset);
15529 goto err_free;
15530 }
15531
15532 if (env->subprog_info[i].start != krecord[i].insn_off) {
15533 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
15534 goto err_free;
15535 }
15536
15537 /* check type_id */
15538 type = btf_type_by_id(btf, krecord[i].type_id);
15539 if (!type || !btf_type_is_func(type)) {
15540 verbose(env, "invalid type id %d in func info",
15541 krecord[i].type_id);
15542 goto err_free;
15543 }
15544 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
15545
15546 func_proto = btf_type_by_id(btf, type->type);
15547 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
15548 /* btf_func_check() already verified it during BTF load */
15549 goto err_free;
15550 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
15551 scalar_return =
15552 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
15553 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
15554 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
15555 goto err_free;
15556 }
15557 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
15558 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
15559 goto err_free;
15560 }
15561
15562 prev_offset = krecord[i].insn_off;
15563 bpfptr_add(&urecord, urec_size);
15564 }
15565
15566 prog->aux->func_info = krecord;
15567 prog->aux->func_info_cnt = nfuncs;
15568 prog->aux->func_info_aux = info_aux;
15569 return 0;
15570
15571 err_free:
15572 kvfree(krecord);
15573 kfree(info_aux);
15574 return ret;
15575 }
15576
adjust_btf_func(struct bpf_verifier_env * env)15577 static void adjust_btf_func(struct bpf_verifier_env *env)
15578 {
15579 struct bpf_prog_aux *aux = env->prog->aux;
15580 int i;
15581
15582 if (!aux->func_info)
15583 return;
15584
15585 for (i = 0; i < env->subprog_cnt; i++)
15586 aux->func_info[i].insn_off = env->subprog_info[i].start;
15587 }
15588
15589 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
15590 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
15591
check_btf_line(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15592 static int check_btf_line(struct bpf_verifier_env *env,
15593 const union bpf_attr *attr,
15594 bpfptr_t uattr)
15595 {
15596 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
15597 struct bpf_subprog_info *sub;
15598 struct bpf_line_info *linfo;
15599 struct bpf_prog *prog;
15600 const struct btf *btf;
15601 bpfptr_t ulinfo;
15602 int err;
15603
15604 nr_linfo = attr->line_info_cnt;
15605 if (!nr_linfo)
15606 return 0;
15607 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
15608 return -EINVAL;
15609
15610 rec_size = attr->line_info_rec_size;
15611 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
15612 rec_size > MAX_LINEINFO_REC_SIZE ||
15613 rec_size & (sizeof(u32) - 1))
15614 return -EINVAL;
15615
15616 /* Need to zero it in case the userspace may
15617 * pass in a smaller bpf_line_info object.
15618 */
15619 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
15620 GFP_KERNEL | __GFP_NOWARN);
15621 if (!linfo)
15622 return -ENOMEM;
15623
15624 prog = env->prog;
15625 btf = prog->aux->btf;
15626
15627 s = 0;
15628 sub = env->subprog_info;
15629 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
15630 expected_size = sizeof(struct bpf_line_info);
15631 ncopy = min_t(u32, expected_size, rec_size);
15632 for (i = 0; i < nr_linfo; i++) {
15633 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
15634 if (err) {
15635 if (err == -E2BIG) {
15636 verbose(env, "nonzero tailing record in line_info");
15637 if (copy_to_bpfptr_offset(uattr,
15638 offsetof(union bpf_attr, line_info_rec_size),
15639 &expected_size, sizeof(expected_size)))
15640 err = -EFAULT;
15641 }
15642 goto err_free;
15643 }
15644
15645 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
15646 err = -EFAULT;
15647 goto err_free;
15648 }
15649
15650 /*
15651 * Check insn_off to ensure
15652 * 1) strictly increasing AND
15653 * 2) bounded by prog->len
15654 *
15655 * The linfo[0].insn_off == 0 check logically falls into
15656 * the later "missing bpf_line_info for func..." case
15657 * because the first linfo[0].insn_off must be the
15658 * first sub also and the first sub must have
15659 * subprog_info[0].start == 0.
15660 */
15661 if ((i && linfo[i].insn_off <= prev_offset) ||
15662 linfo[i].insn_off >= prog->len) {
15663 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
15664 i, linfo[i].insn_off, prev_offset,
15665 prog->len);
15666 err = -EINVAL;
15667 goto err_free;
15668 }
15669
15670 if (!prog->insnsi[linfo[i].insn_off].code) {
15671 verbose(env,
15672 "Invalid insn code at line_info[%u].insn_off\n",
15673 i);
15674 err = -EINVAL;
15675 goto err_free;
15676 }
15677
15678 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
15679 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
15680 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
15681 err = -EINVAL;
15682 goto err_free;
15683 }
15684
15685 if (s != env->subprog_cnt) {
15686 if (linfo[i].insn_off == sub[s].start) {
15687 sub[s].linfo_idx = i;
15688 s++;
15689 } else if (sub[s].start < linfo[i].insn_off) {
15690 verbose(env, "missing bpf_line_info for func#%u\n", s);
15691 err = -EINVAL;
15692 goto err_free;
15693 }
15694 }
15695
15696 prev_offset = linfo[i].insn_off;
15697 bpfptr_add(&ulinfo, rec_size);
15698 }
15699
15700 if (s != env->subprog_cnt) {
15701 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
15702 env->subprog_cnt - s, s);
15703 err = -EINVAL;
15704 goto err_free;
15705 }
15706
15707 prog->aux->linfo = linfo;
15708 prog->aux->nr_linfo = nr_linfo;
15709
15710 return 0;
15711
15712 err_free:
15713 kvfree(linfo);
15714 return err;
15715 }
15716
15717 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
15718 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
15719
check_core_relo(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15720 static int check_core_relo(struct bpf_verifier_env *env,
15721 const union bpf_attr *attr,
15722 bpfptr_t uattr)
15723 {
15724 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
15725 struct bpf_core_relo core_relo = {};
15726 struct bpf_prog *prog = env->prog;
15727 const struct btf *btf = prog->aux->btf;
15728 struct bpf_core_ctx ctx = {
15729 .log = &env->log,
15730 .btf = btf,
15731 };
15732 bpfptr_t u_core_relo;
15733 int err;
15734
15735 nr_core_relo = attr->core_relo_cnt;
15736 if (!nr_core_relo)
15737 return 0;
15738 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15739 return -EINVAL;
15740
15741 rec_size = attr->core_relo_rec_size;
15742 if (rec_size < MIN_CORE_RELO_SIZE ||
15743 rec_size > MAX_CORE_RELO_SIZE ||
15744 rec_size % sizeof(u32))
15745 return -EINVAL;
15746
15747 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15748 expected_size = sizeof(struct bpf_core_relo);
15749 ncopy = min_t(u32, expected_size, rec_size);
15750
15751 /* Unlike func_info and line_info, copy and apply each CO-RE
15752 * relocation record one at a time.
15753 */
15754 for (i = 0; i < nr_core_relo; i++) {
15755 /* future proofing when sizeof(bpf_core_relo) changes */
15756 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15757 if (err) {
15758 if (err == -E2BIG) {
15759 verbose(env, "nonzero tailing record in core_relo");
15760 if (copy_to_bpfptr_offset(uattr,
15761 offsetof(union bpf_attr, core_relo_rec_size),
15762 &expected_size, sizeof(expected_size)))
15763 err = -EFAULT;
15764 }
15765 break;
15766 }
15767
15768 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15769 err = -EFAULT;
15770 break;
15771 }
15772
15773 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15774 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15775 i, core_relo.insn_off, prog->len);
15776 err = -EINVAL;
15777 break;
15778 }
15779
15780 err = bpf_core_apply(&ctx, &core_relo, i,
15781 &prog->insnsi[core_relo.insn_off / 8]);
15782 if (err)
15783 break;
15784 bpfptr_add(&u_core_relo, rec_size);
15785 }
15786 return err;
15787 }
15788
check_btf_info(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15789 static int check_btf_info(struct bpf_verifier_env *env,
15790 const union bpf_attr *attr,
15791 bpfptr_t uattr)
15792 {
15793 struct btf *btf;
15794 int err;
15795
15796 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15797 if (check_abnormal_return(env))
15798 return -EINVAL;
15799 return 0;
15800 }
15801
15802 btf = btf_get_by_fd(attr->prog_btf_fd);
15803 if (IS_ERR(btf))
15804 return PTR_ERR(btf);
15805 if (btf_is_kernel(btf)) {
15806 btf_put(btf);
15807 return -EACCES;
15808 }
15809 env->prog->aux->btf = btf;
15810
15811 err = check_btf_func(env, attr, uattr);
15812 if (err)
15813 return err;
15814
15815 err = check_btf_line(env, attr, uattr);
15816 if (err)
15817 return err;
15818
15819 err = check_core_relo(env, attr, uattr);
15820 if (err)
15821 return err;
15822
15823 return 0;
15824 }
15825
15826 /* check %cur's range satisfies %old's */
range_within(struct bpf_reg_state * old,struct bpf_reg_state * cur)15827 static bool range_within(struct bpf_reg_state *old,
15828 struct bpf_reg_state *cur)
15829 {
15830 return old->umin_value <= cur->umin_value &&
15831 old->umax_value >= cur->umax_value &&
15832 old->smin_value <= cur->smin_value &&
15833 old->smax_value >= cur->smax_value &&
15834 old->u32_min_value <= cur->u32_min_value &&
15835 old->u32_max_value >= cur->u32_max_value &&
15836 old->s32_min_value <= cur->s32_min_value &&
15837 old->s32_max_value >= cur->s32_max_value;
15838 }
15839
15840 /* If in the old state two registers had the same id, then they need to have
15841 * the same id in the new state as well. But that id could be different from
15842 * the old state, so we need to track the mapping from old to new ids.
15843 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15844 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15845 * regs with a different old id could still have new id 9, we don't care about
15846 * that.
15847 * So we look through our idmap to see if this old id has been seen before. If
15848 * so, we require the new id to match; otherwise, we add the id pair to the map.
15849 */
check_ids(u32 old_id,u32 cur_id,struct bpf_idmap * idmap)15850 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15851 {
15852 struct bpf_id_pair *map = idmap->map;
15853 unsigned int i;
15854
15855 /* either both IDs should be set or both should be zero */
15856 if (!!old_id != !!cur_id)
15857 return false;
15858
15859 if (old_id == 0) /* cur_id == 0 as well */
15860 return true;
15861
15862 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15863 if (!map[i].old) {
15864 /* Reached an empty slot; haven't seen this id before */
15865 map[i].old = old_id;
15866 map[i].cur = cur_id;
15867 return true;
15868 }
15869 if (map[i].old == old_id)
15870 return map[i].cur == cur_id;
15871 if (map[i].cur == cur_id)
15872 return false;
15873 }
15874 /* We ran out of idmap slots, which should be impossible */
15875 WARN_ON_ONCE(1);
15876 return false;
15877 }
15878
15879 /* Similar to check_ids(), but allocate a unique temporary ID
15880 * for 'old_id' or 'cur_id' of zero.
15881 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15882 */
check_scalar_ids(u32 old_id,u32 cur_id,struct bpf_idmap * idmap)15883 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15884 {
15885 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15886 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15887
15888 return check_ids(old_id, cur_id, idmap);
15889 }
15890
clean_func_state(struct bpf_verifier_env * env,struct bpf_func_state * st)15891 static void clean_func_state(struct bpf_verifier_env *env,
15892 struct bpf_func_state *st)
15893 {
15894 enum bpf_reg_liveness live;
15895 int i, j;
15896
15897 for (i = 0; i < BPF_REG_FP; i++) {
15898 live = st->regs[i].live;
15899 /* liveness must not touch this register anymore */
15900 st->regs[i].live |= REG_LIVE_DONE;
15901 if (!(live & REG_LIVE_READ))
15902 /* since the register is unused, clear its state
15903 * to make further comparison simpler
15904 */
15905 __mark_reg_not_init(env, &st->regs[i]);
15906 }
15907
15908 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15909 live = st->stack[i].spilled_ptr.live;
15910 /* liveness must not touch this stack slot anymore */
15911 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15912 if (!(live & REG_LIVE_READ)) {
15913 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15914 for (j = 0; j < BPF_REG_SIZE; j++)
15915 st->stack[i].slot_type[j] = STACK_INVALID;
15916 }
15917 }
15918 }
15919
clean_verifier_state(struct bpf_verifier_env * env,struct bpf_verifier_state * st)15920 static void clean_verifier_state(struct bpf_verifier_env *env,
15921 struct bpf_verifier_state *st)
15922 {
15923 int i;
15924
15925 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15926 /* all regs in this state in all frames were already marked */
15927 return;
15928
15929 for (i = 0; i <= st->curframe; i++)
15930 clean_func_state(env, st->frame[i]);
15931 }
15932
15933 /* the parentage chains form a tree.
15934 * the verifier states are added to state lists at given insn and
15935 * pushed into state stack for future exploration.
15936 * when the verifier reaches bpf_exit insn some of the verifer states
15937 * stored in the state lists have their final liveness state already,
15938 * but a lot of states will get revised from liveness point of view when
15939 * the verifier explores other branches.
15940 * Example:
15941 * 1: r0 = 1
15942 * 2: if r1 == 100 goto pc+1
15943 * 3: r0 = 2
15944 * 4: exit
15945 * when the verifier reaches exit insn the register r0 in the state list of
15946 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15947 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15948 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15949 *
15950 * Since the verifier pushes the branch states as it sees them while exploring
15951 * the program the condition of walking the branch instruction for the second
15952 * time means that all states below this branch were already explored and
15953 * their final liveness marks are already propagated.
15954 * Hence when the verifier completes the search of state list in is_state_visited()
15955 * we can call this clean_live_states() function to mark all liveness states
15956 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15957 * will not be used.
15958 * This function also clears the registers and stack for states that !READ
15959 * to simplify state merging.
15960 *
15961 * Important note here that walking the same branch instruction in the callee
15962 * doesn't meant that the states are DONE. The verifier has to compare
15963 * the callsites
15964 */
clean_live_states(struct bpf_verifier_env * env,int insn,struct bpf_verifier_state * cur)15965 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15966 struct bpf_verifier_state *cur)
15967 {
15968 struct bpf_verifier_state_list *sl;
15969
15970 sl = *explored_state(env, insn);
15971 while (sl) {
15972 if (sl->state.branches)
15973 goto next;
15974 if (sl->state.insn_idx != insn ||
15975 !same_callsites(&sl->state, cur))
15976 goto next;
15977 clean_verifier_state(env, &sl->state);
15978 next:
15979 sl = sl->next;
15980 }
15981 }
15982
regs_exact(const struct bpf_reg_state * rold,const struct bpf_reg_state * rcur,struct bpf_idmap * idmap)15983 static bool regs_exact(const struct bpf_reg_state *rold,
15984 const struct bpf_reg_state *rcur,
15985 struct bpf_idmap *idmap)
15986 {
15987 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15988 check_ids(rold->id, rcur->id, idmap) &&
15989 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15990 }
15991
15992 /* Returns true if (rold safe implies rcur safe) */
regsafe(struct bpf_verifier_env * env,struct bpf_reg_state * rold,struct bpf_reg_state * rcur,struct bpf_idmap * idmap,bool exact)15993 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15994 struct bpf_reg_state *rcur, struct bpf_idmap *idmap, bool exact)
15995 {
15996 if (exact)
15997 return regs_exact(rold, rcur, idmap);
15998
15999 if (!(rold->live & REG_LIVE_READ))
16000 /* explored state didn't use this */
16001 return true;
16002 if (rold->type == NOT_INIT)
16003 /* explored state can't have used this */
16004 return true;
16005 if (rcur->type == NOT_INIT)
16006 return false;
16007
16008 /* Enforce that register types have to match exactly, including their
16009 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
16010 * rule.
16011 *
16012 * One can make a point that using a pointer register as unbounded
16013 * SCALAR would be technically acceptable, but this could lead to
16014 * pointer leaks because scalars are allowed to leak while pointers
16015 * are not. We could make this safe in special cases if root is
16016 * calling us, but it's probably not worth the hassle.
16017 *
16018 * Also, register types that are *not* MAYBE_NULL could technically be
16019 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
16020 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
16021 * to the same map).
16022 * However, if the old MAYBE_NULL register then got NULL checked,
16023 * doing so could have affected others with the same id, and we can't
16024 * check for that because we lost the id when we converted to
16025 * a non-MAYBE_NULL variant.
16026 * So, as a general rule we don't allow mixing MAYBE_NULL and
16027 * non-MAYBE_NULL registers as well.
16028 */
16029 if (rold->type != rcur->type)
16030 return false;
16031
16032 switch (base_type(rold->type)) {
16033 case SCALAR_VALUE:
16034 if (env->explore_alu_limits) {
16035 /* explore_alu_limits disables tnum_in() and range_within()
16036 * logic and requires everything to be strict
16037 */
16038 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
16039 check_scalar_ids(rold->id, rcur->id, idmap);
16040 }
16041 if (!rold->precise)
16042 return true;
16043 /* Why check_ids() for scalar registers?
16044 *
16045 * Consider the following BPF code:
16046 * 1: r6 = ... unbound scalar, ID=a ...
16047 * 2: r7 = ... unbound scalar, ID=b ...
16048 * 3: if (r6 > r7) goto +1
16049 * 4: r6 = r7
16050 * 5: if (r6 > X) goto ...
16051 * 6: ... memory operation using r7 ...
16052 *
16053 * First verification path is [1-6]:
16054 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
16055 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
16056 * r7 <= X, because r6 and r7 share same id.
16057 * Next verification path is [1-4, 6].
16058 *
16059 * Instruction (6) would be reached in two states:
16060 * I. r6{.id=b}, r7{.id=b} via path 1-6;
16061 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
16062 *
16063 * Use check_ids() to distinguish these states.
16064 * ---
16065 * Also verify that new value satisfies old value range knowledge.
16066 */
16067 return range_within(rold, rcur) &&
16068 tnum_in(rold->var_off, rcur->var_off) &&
16069 check_scalar_ids(rold->id, rcur->id, idmap);
16070 case PTR_TO_MAP_KEY:
16071 case PTR_TO_MAP_VALUE:
16072 case PTR_TO_MEM:
16073 case PTR_TO_BUF:
16074 case PTR_TO_TP_BUFFER:
16075 /* If the new min/max/var_off satisfy the old ones and
16076 * everything else matches, we are OK.
16077 */
16078 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
16079 range_within(rold, rcur) &&
16080 tnum_in(rold->var_off, rcur->var_off) &&
16081 check_ids(rold->id, rcur->id, idmap) &&
16082 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
16083 case PTR_TO_PACKET_META:
16084 case PTR_TO_PACKET:
16085 /* We must have at least as much range as the old ptr
16086 * did, so that any accesses which were safe before are
16087 * still safe. This is true even if old range < old off,
16088 * since someone could have accessed through (ptr - k), or
16089 * even done ptr -= k in a register, to get a safe access.
16090 */
16091 if (rold->range > rcur->range)
16092 return false;
16093 /* If the offsets don't match, we can't trust our alignment;
16094 * nor can we be sure that we won't fall out of range.
16095 */
16096 if (rold->off != rcur->off)
16097 return false;
16098 /* id relations must be preserved */
16099 if (!check_ids(rold->id, rcur->id, idmap))
16100 return false;
16101 /* new val must satisfy old val knowledge */
16102 return range_within(rold, rcur) &&
16103 tnum_in(rold->var_off, rcur->var_off);
16104 case PTR_TO_STACK:
16105 /* two stack pointers are equal only if they're pointing to
16106 * the same stack frame, since fp-8 in foo != fp-8 in bar
16107 */
16108 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
16109 default:
16110 return regs_exact(rold, rcur, idmap);
16111 }
16112 }
16113
stacksafe(struct bpf_verifier_env * env,struct bpf_func_state * old,struct bpf_func_state * cur,struct bpf_idmap * idmap,bool exact)16114 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
16115 struct bpf_func_state *cur, struct bpf_idmap *idmap, bool exact)
16116 {
16117 int i, spi;
16118
16119 /* walk slots of the explored stack and ignore any additional
16120 * slots in the current stack, since explored(safe) state
16121 * didn't use them
16122 */
16123 for (i = 0; i < old->allocated_stack; i++) {
16124 struct bpf_reg_state *old_reg, *cur_reg;
16125
16126 spi = i / BPF_REG_SIZE;
16127
16128 if (exact &&
16129 (i >= cur->allocated_stack ||
16130 old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
16131 cur->stack[spi].slot_type[i % BPF_REG_SIZE]))
16132 return false;
16133
16134 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ) && !exact) {
16135 i += BPF_REG_SIZE - 1;
16136 /* explored state didn't use this */
16137 continue;
16138 }
16139
16140 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
16141 continue;
16142
16143 if (env->allow_uninit_stack &&
16144 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
16145 continue;
16146
16147 /* explored stack has more populated slots than current stack
16148 * and these slots were used
16149 */
16150 if (i >= cur->allocated_stack)
16151 return false;
16152
16153 /* if old state was safe with misc data in the stack
16154 * it will be safe with zero-initialized stack.
16155 * The opposite is not true
16156 */
16157 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
16158 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
16159 continue;
16160 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
16161 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
16162 /* Ex: old explored (safe) state has STACK_SPILL in
16163 * this stack slot, but current has STACK_MISC ->
16164 * this verifier states are not equivalent,
16165 * return false to continue verification of this path
16166 */
16167 return false;
16168 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
16169 continue;
16170 /* Both old and cur are having same slot_type */
16171 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
16172 case STACK_SPILL:
16173 /* when explored and current stack slot are both storing
16174 * spilled registers, check that stored pointers types
16175 * are the same as well.
16176 * Ex: explored safe path could have stored
16177 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
16178 * but current path has stored:
16179 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
16180 * such verifier states are not equivalent.
16181 * return false to continue verification of this path
16182 */
16183 if (!regsafe(env, &old->stack[spi].spilled_ptr,
16184 &cur->stack[spi].spilled_ptr, idmap, exact))
16185 return false;
16186 break;
16187 case STACK_DYNPTR:
16188 old_reg = &old->stack[spi].spilled_ptr;
16189 cur_reg = &cur->stack[spi].spilled_ptr;
16190 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
16191 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
16192 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
16193 return false;
16194 break;
16195 case STACK_ITER:
16196 old_reg = &old->stack[spi].spilled_ptr;
16197 cur_reg = &cur->stack[spi].spilled_ptr;
16198 /* iter.depth is not compared between states as it
16199 * doesn't matter for correctness and would otherwise
16200 * prevent convergence; we maintain it only to prevent
16201 * infinite loop check triggering, see
16202 * iter_active_depths_differ()
16203 */
16204 if (old_reg->iter.btf != cur_reg->iter.btf ||
16205 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
16206 old_reg->iter.state != cur_reg->iter.state ||
16207 /* ignore {old_reg,cur_reg}->iter.depth, see above */
16208 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
16209 return false;
16210 break;
16211 case STACK_MISC:
16212 case STACK_ZERO:
16213 case STACK_INVALID:
16214 continue;
16215 /* Ensure that new unhandled slot types return false by default */
16216 default:
16217 return false;
16218 }
16219 }
16220 return true;
16221 }
16222
refsafe(struct bpf_func_state * old,struct bpf_func_state * cur,struct bpf_idmap * idmap)16223 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
16224 struct bpf_idmap *idmap)
16225 {
16226 int i;
16227
16228 if (old->acquired_refs != cur->acquired_refs)
16229 return false;
16230
16231 for (i = 0; i < old->acquired_refs; i++) {
16232 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
16233 return false;
16234 }
16235
16236 return true;
16237 }
16238
16239 /* compare two verifier states
16240 *
16241 * all states stored in state_list are known to be valid, since
16242 * verifier reached 'bpf_exit' instruction through them
16243 *
16244 * this function is called when verifier exploring different branches of
16245 * execution popped from the state stack. If it sees an old state that has
16246 * more strict register state and more strict stack state then this execution
16247 * branch doesn't need to be explored further, since verifier already
16248 * concluded that more strict state leads to valid finish.
16249 *
16250 * Therefore two states are equivalent if register state is more conservative
16251 * and explored stack state is more conservative than the current one.
16252 * Example:
16253 * explored current
16254 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
16255 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
16256 *
16257 * In other words if current stack state (one being explored) has more
16258 * valid slots than old one that already passed validation, it means
16259 * the verifier can stop exploring and conclude that current state is valid too
16260 *
16261 * Similarly with registers. If explored state has register type as invalid
16262 * whereas register type in current state is meaningful, it means that
16263 * the current state will reach 'bpf_exit' instruction safely
16264 */
func_states_equal(struct bpf_verifier_env * env,struct bpf_func_state * old,struct bpf_func_state * cur,bool exact)16265 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
16266 struct bpf_func_state *cur, bool exact)
16267 {
16268 int i;
16269
16270 if (old->callback_depth > cur->callback_depth)
16271 return false;
16272
16273 for (i = 0; i < MAX_BPF_REG; i++)
16274 if (!regsafe(env, &old->regs[i], &cur->regs[i],
16275 &env->idmap_scratch, exact))
16276 return false;
16277
16278 if (!stacksafe(env, old, cur, &env->idmap_scratch, exact))
16279 return false;
16280
16281 if (!refsafe(old, cur, &env->idmap_scratch))
16282 return false;
16283
16284 return true;
16285 }
16286
reset_idmap_scratch(struct bpf_verifier_env * env)16287 static void reset_idmap_scratch(struct bpf_verifier_env *env)
16288 {
16289 env->idmap_scratch.tmp_id_gen = env->id_gen;
16290 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
16291 }
16292
states_equal(struct bpf_verifier_env * env,struct bpf_verifier_state * old,struct bpf_verifier_state * cur,bool exact)16293 static bool states_equal(struct bpf_verifier_env *env,
16294 struct bpf_verifier_state *old,
16295 struct bpf_verifier_state *cur,
16296 bool exact)
16297 {
16298 int i;
16299
16300 if (old->curframe != cur->curframe)
16301 return false;
16302
16303 reset_idmap_scratch(env);
16304
16305 /* Verification state from speculative execution simulation
16306 * must never prune a non-speculative execution one.
16307 */
16308 if (old->speculative && !cur->speculative)
16309 return false;
16310
16311 if (old->active_lock.ptr != cur->active_lock.ptr)
16312 return false;
16313
16314 /* Old and cur active_lock's have to be either both present
16315 * or both absent.
16316 */
16317 if (!!old->active_lock.id != !!cur->active_lock.id)
16318 return false;
16319
16320 if (old->active_lock.id &&
16321 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
16322 return false;
16323
16324 if (old->active_rcu_lock != cur->active_rcu_lock)
16325 return false;
16326
16327 /* for states to be equal callsites have to be the same
16328 * and all frame states need to be equivalent
16329 */
16330 for (i = 0; i <= old->curframe; i++) {
16331 if (old->frame[i]->callsite != cur->frame[i]->callsite)
16332 return false;
16333 if (!func_states_equal(env, old->frame[i], cur->frame[i], exact))
16334 return false;
16335 }
16336 return true;
16337 }
16338
16339 /* Return 0 if no propagation happened. Return negative error code if error
16340 * happened. Otherwise, return the propagated bit.
16341 */
propagate_liveness_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct bpf_reg_state * parent_reg)16342 static int propagate_liveness_reg(struct bpf_verifier_env *env,
16343 struct bpf_reg_state *reg,
16344 struct bpf_reg_state *parent_reg)
16345 {
16346 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
16347 u8 flag = reg->live & REG_LIVE_READ;
16348 int err;
16349
16350 /* When comes here, read flags of PARENT_REG or REG could be any of
16351 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
16352 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
16353 */
16354 if (parent_flag == REG_LIVE_READ64 ||
16355 /* Or if there is no read flag from REG. */
16356 !flag ||
16357 /* Or if the read flag from REG is the same as PARENT_REG. */
16358 parent_flag == flag)
16359 return 0;
16360
16361 err = mark_reg_read(env, reg, parent_reg, flag);
16362 if (err)
16363 return err;
16364
16365 return flag;
16366 }
16367
16368 /* A write screens off any subsequent reads; but write marks come from the
16369 * straight-line code between a state and its parent. When we arrive at an
16370 * equivalent state (jump target or such) we didn't arrive by the straight-line
16371 * code, so read marks in the state must propagate to the parent regardless
16372 * of the state's write marks. That's what 'parent == state->parent' comparison
16373 * in mark_reg_read() is for.
16374 */
propagate_liveness(struct bpf_verifier_env * env,const struct bpf_verifier_state * vstate,struct bpf_verifier_state * vparent)16375 static int propagate_liveness(struct bpf_verifier_env *env,
16376 const struct bpf_verifier_state *vstate,
16377 struct bpf_verifier_state *vparent)
16378 {
16379 struct bpf_reg_state *state_reg, *parent_reg;
16380 struct bpf_func_state *state, *parent;
16381 int i, frame, err = 0;
16382
16383 if (vparent->curframe != vstate->curframe) {
16384 WARN(1, "propagate_live: parent frame %d current frame %d\n",
16385 vparent->curframe, vstate->curframe);
16386 return -EFAULT;
16387 }
16388 /* Propagate read liveness of registers... */
16389 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
16390 for (frame = 0; frame <= vstate->curframe; frame++) {
16391 parent = vparent->frame[frame];
16392 state = vstate->frame[frame];
16393 parent_reg = parent->regs;
16394 state_reg = state->regs;
16395 /* We don't need to worry about FP liveness, it's read-only */
16396 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
16397 err = propagate_liveness_reg(env, &state_reg[i],
16398 &parent_reg[i]);
16399 if (err < 0)
16400 return err;
16401 if (err == REG_LIVE_READ64)
16402 mark_insn_zext(env, &parent_reg[i]);
16403 }
16404
16405 /* Propagate stack slots. */
16406 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
16407 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
16408 parent_reg = &parent->stack[i].spilled_ptr;
16409 state_reg = &state->stack[i].spilled_ptr;
16410 err = propagate_liveness_reg(env, state_reg,
16411 parent_reg);
16412 if (err < 0)
16413 return err;
16414 }
16415 }
16416 return 0;
16417 }
16418
16419 /* find precise scalars in the previous equivalent state and
16420 * propagate them into the current state
16421 */
propagate_precision(struct bpf_verifier_env * env,const struct bpf_verifier_state * old)16422 static int propagate_precision(struct bpf_verifier_env *env,
16423 const struct bpf_verifier_state *old)
16424 {
16425 struct bpf_reg_state *state_reg;
16426 struct bpf_func_state *state;
16427 int i, err = 0, fr;
16428 bool first;
16429
16430 for (fr = old->curframe; fr >= 0; fr--) {
16431 state = old->frame[fr];
16432 state_reg = state->regs;
16433 first = true;
16434 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
16435 if (state_reg->type != SCALAR_VALUE ||
16436 !state_reg->precise ||
16437 !(state_reg->live & REG_LIVE_READ))
16438 continue;
16439 if (env->log.level & BPF_LOG_LEVEL2) {
16440 if (first)
16441 verbose(env, "frame %d: propagating r%d", fr, i);
16442 else
16443 verbose(env, ",r%d", i);
16444 }
16445 bt_set_frame_reg(&env->bt, fr, i);
16446 first = false;
16447 }
16448
16449 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16450 if (!is_spilled_reg(&state->stack[i]))
16451 continue;
16452 state_reg = &state->stack[i].spilled_ptr;
16453 if (state_reg->type != SCALAR_VALUE ||
16454 !state_reg->precise ||
16455 !(state_reg->live & REG_LIVE_READ))
16456 continue;
16457 if (env->log.level & BPF_LOG_LEVEL2) {
16458 if (first)
16459 verbose(env, "frame %d: propagating fp%d",
16460 fr, (-i - 1) * BPF_REG_SIZE);
16461 else
16462 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
16463 }
16464 bt_set_frame_slot(&env->bt, fr, i);
16465 first = false;
16466 }
16467 if (!first)
16468 verbose(env, "\n");
16469 }
16470
16471 err = mark_chain_precision_batch(env);
16472 if (err < 0)
16473 return err;
16474
16475 return 0;
16476 }
16477
states_maybe_looping(struct bpf_verifier_state * old,struct bpf_verifier_state * cur)16478 static bool states_maybe_looping(struct bpf_verifier_state *old,
16479 struct bpf_verifier_state *cur)
16480 {
16481 struct bpf_func_state *fold, *fcur;
16482 int i, fr = cur->curframe;
16483
16484 if (old->curframe != fr)
16485 return false;
16486
16487 fold = old->frame[fr];
16488 fcur = cur->frame[fr];
16489 for (i = 0; i < MAX_BPF_REG; i++)
16490 if (memcmp(&fold->regs[i], &fcur->regs[i],
16491 offsetof(struct bpf_reg_state, parent)))
16492 return false;
16493 return true;
16494 }
16495
is_iter_next_insn(struct bpf_verifier_env * env,int insn_idx)16496 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
16497 {
16498 return env->insn_aux_data[insn_idx].is_iter_next;
16499 }
16500
16501 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
16502 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
16503 * states to match, which otherwise would look like an infinite loop. So while
16504 * iter_next() calls are taken care of, we still need to be careful and
16505 * prevent erroneous and too eager declaration of "ininite loop", when
16506 * iterators are involved.
16507 *
16508 * Here's a situation in pseudo-BPF assembly form:
16509 *
16510 * 0: again: ; set up iter_next() call args
16511 * 1: r1 = &it ; <CHECKPOINT HERE>
16512 * 2: call bpf_iter_num_next ; this is iter_next() call
16513 * 3: if r0 == 0 goto done
16514 * 4: ... something useful here ...
16515 * 5: goto again ; another iteration
16516 * 6: done:
16517 * 7: r1 = &it
16518 * 8: call bpf_iter_num_destroy ; clean up iter state
16519 * 9: exit
16520 *
16521 * This is a typical loop. Let's assume that we have a prune point at 1:,
16522 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
16523 * again`, assuming other heuristics don't get in a way).
16524 *
16525 * When we first time come to 1:, let's say we have some state X. We proceed
16526 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
16527 * Now we come back to validate that forked ACTIVE state. We proceed through
16528 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
16529 * are converging. But the problem is that we don't know that yet, as this
16530 * convergence has to happen at iter_next() call site only. So if nothing is
16531 * done, at 1: verifier will use bounded loop logic and declare infinite
16532 * looping (and would be *technically* correct, if not for iterator's
16533 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
16534 * don't want that. So what we do in process_iter_next_call() when we go on
16535 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
16536 * a different iteration. So when we suspect an infinite loop, we additionally
16537 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
16538 * pretend we are not looping and wait for next iter_next() call.
16539 *
16540 * This only applies to ACTIVE state. In DRAINED state we don't expect to
16541 * loop, because that would actually mean infinite loop, as DRAINED state is
16542 * "sticky", and so we'll keep returning into the same instruction with the
16543 * same state (at least in one of possible code paths).
16544 *
16545 * This approach allows to keep infinite loop heuristic even in the face of
16546 * active iterator. E.g., C snippet below is and will be detected as
16547 * inifintely looping:
16548 *
16549 * struct bpf_iter_num it;
16550 * int *p, x;
16551 *
16552 * bpf_iter_num_new(&it, 0, 10);
16553 * while ((p = bpf_iter_num_next(&t))) {
16554 * x = p;
16555 * while (x--) {} // <<-- infinite loop here
16556 * }
16557 *
16558 */
iter_active_depths_differ(struct bpf_verifier_state * old,struct bpf_verifier_state * cur)16559 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
16560 {
16561 struct bpf_reg_state *slot, *cur_slot;
16562 struct bpf_func_state *state;
16563 int i, fr;
16564
16565 for (fr = old->curframe; fr >= 0; fr--) {
16566 state = old->frame[fr];
16567 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16568 if (state->stack[i].slot_type[0] != STACK_ITER)
16569 continue;
16570
16571 slot = &state->stack[i].spilled_ptr;
16572 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
16573 continue;
16574
16575 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
16576 if (cur_slot->iter.depth != slot->iter.depth)
16577 return true;
16578 }
16579 }
16580 return false;
16581 }
16582
is_state_visited(struct bpf_verifier_env * env,int insn_idx)16583 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
16584 {
16585 struct bpf_verifier_state_list *new_sl;
16586 struct bpf_verifier_state_list *sl, **pprev;
16587 struct bpf_verifier_state *cur = env->cur_state, *new, *loop_entry;
16588 int i, j, n, err, states_cnt = 0;
16589 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
16590 bool add_new_state = force_new_state;
16591 bool force_exact;
16592
16593 /* bpf progs typically have pruning point every 4 instructions
16594 * http://vger.kernel.org/bpfconf2019.html#session-1
16595 * Do not add new state for future pruning if the verifier hasn't seen
16596 * at least 2 jumps and at least 8 instructions.
16597 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
16598 * In tests that amounts to up to 50% reduction into total verifier
16599 * memory consumption and 20% verifier time speedup.
16600 */
16601 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
16602 env->insn_processed - env->prev_insn_processed >= 8)
16603 add_new_state = true;
16604
16605 pprev = explored_state(env, insn_idx);
16606 sl = *pprev;
16607
16608 clean_live_states(env, insn_idx, cur);
16609
16610 while (sl) {
16611 states_cnt++;
16612 if (sl->state.insn_idx != insn_idx)
16613 goto next;
16614
16615 if (sl->state.branches) {
16616 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
16617
16618 if (frame->in_async_callback_fn &&
16619 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
16620 /* Different async_entry_cnt means that the verifier is
16621 * processing another entry into async callback.
16622 * Seeing the same state is not an indication of infinite
16623 * loop or infinite recursion.
16624 * But finding the same state doesn't mean that it's safe
16625 * to stop processing the current state. The previous state
16626 * hasn't yet reached bpf_exit, since state.branches > 0.
16627 * Checking in_async_callback_fn alone is not enough either.
16628 * Since the verifier still needs to catch infinite loops
16629 * inside async callbacks.
16630 */
16631 goto skip_inf_loop_check;
16632 }
16633 /* BPF open-coded iterators loop detection is special.
16634 * states_maybe_looping() logic is too simplistic in detecting
16635 * states that *might* be equivalent, because it doesn't know
16636 * about ID remapping, so don't even perform it.
16637 * See process_iter_next_call() and iter_active_depths_differ()
16638 * for overview of the logic. When current and one of parent
16639 * states are detected as equivalent, it's a good thing: we prove
16640 * convergence and can stop simulating further iterations.
16641 * It's safe to assume that iterator loop will finish, taking into
16642 * account iter_next() contract of eventually returning
16643 * sticky NULL result.
16644 *
16645 * Note, that states have to be compared exactly in this case because
16646 * read and precision marks might not be finalized inside the loop.
16647 * E.g. as in the program below:
16648 *
16649 * 1. r7 = -16
16650 * 2. r6 = bpf_get_prandom_u32()
16651 * 3. while (bpf_iter_num_next(&fp[-8])) {
16652 * 4. if (r6 != 42) {
16653 * 5. r7 = -32
16654 * 6. r6 = bpf_get_prandom_u32()
16655 * 7. continue
16656 * 8. }
16657 * 9. r0 = r10
16658 * 10. r0 += r7
16659 * 11. r8 = *(u64 *)(r0 + 0)
16660 * 12. r6 = bpf_get_prandom_u32()
16661 * 13. }
16662 *
16663 * Here verifier would first visit path 1-3, create a checkpoint at 3
16664 * with r7=-16, continue to 4-7,3. Existing checkpoint at 3 does
16665 * not have read or precision mark for r7 yet, thus inexact states
16666 * comparison would discard current state with r7=-32
16667 * => unsafe memory access at 11 would not be caught.
16668 */
16669 if (is_iter_next_insn(env, insn_idx)) {
16670 if (states_equal(env, &sl->state, cur, true)) {
16671 struct bpf_func_state *cur_frame;
16672 struct bpf_reg_state *iter_state, *iter_reg;
16673 int spi;
16674
16675 cur_frame = cur->frame[cur->curframe];
16676 /* btf_check_iter_kfuncs() enforces that
16677 * iter state pointer is always the first arg
16678 */
16679 iter_reg = &cur_frame->regs[BPF_REG_1];
16680 /* current state is valid due to states_equal(),
16681 * so we can assume valid iter and reg state,
16682 * no need for extra (re-)validations
16683 */
16684 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
16685 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
16686 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE) {
16687 update_loop_entry(cur, &sl->state);
16688 goto hit;
16689 }
16690 }
16691 goto skip_inf_loop_check;
16692 }
16693 if (calls_callback(env, insn_idx)) {
16694 if (states_equal(env, &sl->state, cur, true))
16695 goto hit;
16696 goto skip_inf_loop_check;
16697 }
16698 /* attempt to detect infinite loop to avoid unnecessary doomed work */
16699 if (states_maybe_looping(&sl->state, cur) &&
16700 states_equal(env, &sl->state, cur, false) &&
16701 !iter_active_depths_differ(&sl->state, cur) &&
16702 sl->state.callback_unroll_depth == cur->callback_unroll_depth) {
16703 verbose_linfo(env, insn_idx, "; ");
16704 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
16705 verbose(env, "cur state:");
16706 print_verifier_state(env, cur->frame[cur->curframe], true);
16707 verbose(env, "old state:");
16708 print_verifier_state(env, sl->state.frame[cur->curframe], true);
16709 return -EINVAL;
16710 }
16711 /* if the verifier is processing a loop, avoid adding new state
16712 * too often, since different loop iterations have distinct
16713 * states and may not help future pruning.
16714 * This threshold shouldn't be too low to make sure that
16715 * a loop with large bound will be rejected quickly.
16716 * The most abusive loop will be:
16717 * r1 += 1
16718 * if r1 < 1000000 goto pc-2
16719 * 1M insn_procssed limit / 100 == 10k peak states.
16720 * This threshold shouldn't be too high either, since states
16721 * at the end of the loop are likely to be useful in pruning.
16722 */
16723 skip_inf_loop_check:
16724 if (!force_new_state &&
16725 env->jmps_processed - env->prev_jmps_processed < 20 &&
16726 env->insn_processed - env->prev_insn_processed < 100)
16727 add_new_state = false;
16728 goto miss;
16729 }
16730 /* If sl->state is a part of a loop and this loop's entry is a part of
16731 * current verification path then states have to be compared exactly.
16732 * 'force_exact' is needed to catch the following case:
16733 *
16734 * initial Here state 'succ' was processed first,
16735 * | it was eventually tracked to produce a
16736 * V state identical to 'hdr'.
16737 * .---------> hdr All branches from 'succ' had been explored
16738 * | | and thus 'succ' has its .branches == 0.
16739 * | V
16740 * | .------... Suppose states 'cur' and 'succ' correspond
16741 * | | | to the same instruction + callsites.
16742 * | V V In such case it is necessary to check
16743 * | ... ... if 'succ' and 'cur' are states_equal().
16744 * | | | If 'succ' and 'cur' are a part of the
16745 * | V V same loop exact flag has to be set.
16746 * | succ <- cur To check if that is the case, verify
16747 * | | if loop entry of 'succ' is in current
16748 * | V DFS path.
16749 * | ...
16750 * | |
16751 * '----'
16752 *
16753 * Additional details are in the comment before get_loop_entry().
16754 */
16755 loop_entry = get_loop_entry(&sl->state);
16756 force_exact = loop_entry && loop_entry->branches > 0;
16757 if (states_equal(env, &sl->state, cur, force_exact)) {
16758 if (force_exact)
16759 update_loop_entry(cur, loop_entry);
16760 hit:
16761 sl->hit_cnt++;
16762 /* reached equivalent register/stack state,
16763 * prune the search.
16764 * Registers read by the continuation are read by us.
16765 * If we have any write marks in env->cur_state, they
16766 * will prevent corresponding reads in the continuation
16767 * from reaching our parent (an explored_state). Our
16768 * own state will get the read marks recorded, but
16769 * they'll be immediately forgotten as we're pruning
16770 * this state and will pop a new one.
16771 */
16772 err = propagate_liveness(env, &sl->state, cur);
16773
16774 /* if previous state reached the exit with precision and
16775 * current state is equivalent to it (except precsion marks)
16776 * the precision needs to be propagated back in
16777 * the current state.
16778 */
16779 err = err ? : push_jmp_history(env, cur);
16780 err = err ? : propagate_precision(env, &sl->state);
16781 if (err)
16782 return err;
16783 return 1;
16784 }
16785 miss:
16786 /* when new state is not going to be added do not increase miss count.
16787 * Otherwise several loop iterations will remove the state
16788 * recorded earlier. The goal of these heuristics is to have
16789 * states from some iterations of the loop (some in the beginning
16790 * and some at the end) to help pruning.
16791 */
16792 if (add_new_state)
16793 sl->miss_cnt++;
16794 /* heuristic to determine whether this state is beneficial
16795 * to keep checking from state equivalence point of view.
16796 * Higher numbers increase max_states_per_insn and verification time,
16797 * but do not meaningfully decrease insn_processed.
16798 * 'n' controls how many times state could miss before eviction.
16799 * Use bigger 'n' for checkpoints because evicting checkpoint states
16800 * too early would hinder iterator convergence.
16801 */
16802 n = is_force_checkpoint(env, insn_idx) && sl->state.branches > 0 ? 64 : 3;
16803 if (sl->miss_cnt > sl->hit_cnt * n + n) {
16804 /* the state is unlikely to be useful. Remove it to
16805 * speed up verification
16806 */
16807 *pprev = sl->next;
16808 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE &&
16809 !sl->state.used_as_loop_entry) {
16810 u32 br = sl->state.branches;
16811
16812 WARN_ONCE(br,
16813 "BUG live_done but branches_to_explore %d\n",
16814 br);
16815 free_verifier_state(&sl->state, false);
16816 kfree(sl);
16817 env->peak_states--;
16818 } else {
16819 /* cannot free this state, since parentage chain may
16820 * walk it later. Add it for free_list instead to
16821 * be freed at the end of verification
16822 */
16823 sl->next = env->free_list;
16824 env->free_list = sl;
16825 }
16826 sl = *pprev;
16827 continue;
16828 }
16829 next:
16830 pprev = &sl->next;
16831 sl = *pprev;
16832 }
16833
16834 if (env->max_states_per_insn < states_cnt)
16835 env->max_states_per_insn = states_cnt;
16836
16837 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16838 return 0;
16839
16840 if (!add_new_state)
16841 return 0;
16842
16843 /* There were no equivalent states, remember the current one.
16844 * Technically the current state is not proven to be safe yet,
16845 * but it will either reach outer most bpf_exit (which means it's safe)
16846 * or it will be rejected. When there are no loops the verifier won't be
16847 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16848 * again on the way to bpf_exit.
16849 * When looping the sl->state.branches will be > 0 and this state
16850 * will not be considered for equivalence until branches == 0.
16851 */
16852 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16853 if (!new_sl)
16854 return -ENOMEM;
16855 env->total_states++;
16856 env->peak_states++;
16857 env->prev_jmps_processed = env->jmps_processed;
16858 env->prev_insn_processed = env->insn_processed;
16859
16860 /* forget precise markings we inherited, see __mark_chain_precision */
16861 if (env->bpf_capable)
16862 mark_all_scalars_imprecise(env, cur);
16863
16864 /* add new state to the head of linked list */
16865 new = &new_sl->state;
16866 err = copy_verifier_state(new, cur);
16867 if (err) {
16868 free_verifier_state(new, false);
16869 kfree(new_sl);
16870 return err;
16871 }
16872 new->insn_idx = insn_idx;
16873 WARN_ONCE(new->branches != 1,
16874 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16875
16876 cur->parent = new;
16877 cur->first_insn_idx = insn_idx;
16878 cur->dfs_depth = new->dfs_depth + 1;
16879 clear_jmp_history(cur);
16880 new_sl->next = *explored_state(env, insn_idx);
16881 *explored_state(env, insn_idx) = new_sl;
16882 /* connect new state to parentage chain. Current frame needs all
16883 * registers connected. Only r6 - r9 of the callers are alive (pushed
16884 * to the stack implicitly by JITs) so in callers' frames connect just
16885 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16886 * the state of the call instruction (with WRITTEN set), and r0 comes
16887 * from callee with its full parentage chain, anyway.
16888 */
16889 /* clear write marks in current state: the writes we did are not writes
16890 * our child did, so they don't screen off its reads from us.
16891 * (There are no read marks in current state, because reads always mark
16892 * their parent and current state never has children yet. Only
16893 * explored_states can get read marks.)
16894 */
16895 for (j = 0; j <= cur->curframe; j++) {
16896 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16897 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16898 for (i = 0; i < BPF_REG_FP; i++)
16899 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16900 }
16901
16902 /* all stack frames are accessible from callee, clear them all */
16903 for (j = 0; j <= cur->curframe; j++) {
16904 struct bpf_func_state *frame = cur->frame[j];
16905 struct bpf_func_state *newframe = new->frame[j];
16906
16907 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16908 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16909 frame->stack[i].spilled_ptr.parent =
16910 &newframe->stack[i].spilled_ptr;
16911 }
16912 }
16913 return 0;
16914 }
16915
16916 /* Return true if it's OK to have the same insn return a different type. */
reg_type_mismatch_ok(enum bpf_reg_type type)16917 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16918 {
16919 switch (base_type(type)) {
16920 case PTR_TO_CTX:
16921 case PTR_TO_SOCKET:
16922 case PTR_TO_SOCK_COMMON:
16923 case PTR_TO_TCP_SOCK:
16924 case PTR_TO_XDP_SOCK:
16925 case PTR_TO_BTF_ID:
16926 return false;
16927 default:
16928 return true;
16929 }
16930 }
16931
16932 /* If an instruction was previously used with particular pointer types, then we
16933 * need to be careful to avoid cases such as the below, where it may be ok
16934 * for one branch accessing the pointer, but not ok for the other branch:
16935 *
16936 * R1 = sock_ptr
16937 * goto X;
16938 * ...
16939 * R1 = some_other_valid_ptr;
16940 * goto X;
16941 * ...
16942 * R2 = *(u32 *)(R1 + 0);
16943 */
reg_type_mismatch(enum bpf_reg_type src,enum bpf_reg_type prev)16944 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16945 {
16946 return src != prev && (!reg_type_mismatch_ok(src) ||
16947 !reg_type_mismatch_ok(prev));
16948 }
16949
save_aux_ptr_type(struct bpf_verifier_env * env,enum bpf_reg_type type,bool allow_trust_missmatch)16950 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16951 bool allow_trust_missmatch)
16952 {
16953 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16954
16955 if (*prev_type == NOT_INIT) {
16956 /* Saw a valid insn
16957 * dst_reg = *(u32 *)(src_reg + off)
16958 * save type to validate intersecting paths
16959 */
16960 *prev_type = type;
16961 } else if (reg_type_mismatch(type, *prev_type)) {
16962 /* Abuser program is trying to use the same insn
16963 * dst_reg = *(u32*) (src_reg + off)
16964 * with different pointer types:
16965 * src_reg == ctx in one branch and
16966 * src_reg == stack|map in some other branch.
16967 * Reject it.
16968 */
16969 if (allow_trust_missmatch &&
16970 base_type(type) == PTR_TO_BTF_ID &&
16971 base_type(*prev_type) == PTR_TO_BTF_ID) {
16972 /*
16973 * Have to support a use case when one path through
16974 * the program yields TRUSTED pointer while another
16975 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16976 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
16977 */
16978 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16979 } else {
16980 verbose(env, "same insn cannot be used with different pointers\n");
16981 return -EINVAL;
16982 }
16983 }
16984
16985 return 0;
16986 }
16987
do_check(struct bpf_verifier_env * env)16988 static int do_check(struct bpf_verifier_env *env)
16989 {
16990 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16991 struct bpf_verifier_state *state = env->cur_state;
16992 struct bpf_insn *insns = env->prog->insnsi;
16993 struct bpf_reg_state *regs;
16994 int insn_cnt = env->prog->len;
16995 bool do_print_state = false;
16996 int prev_insn_idx = -1;
16997
16998 for (;;) {
16999 struct bpf_insn *insn;
17000 u8 class;
17001 int err;
17002
17003 env->prev_insn_idx = prev_insn_idx;
17004 if (env->insn_idx >= insn_cnt) {
17005 verbose(env, "invalid insn idx %d insn_cnt %d\n",
17006 env->insn_idx, insn_cnt);
17007 return -EFAULT;
17008 }
17009
17010 insn = &insns[env->insn_idx];
17011 class = BPF_CLASS(insn->code);
17012
17013 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
17014 verbose(env,
17015 "BPF program is too large. Processed %d insn\n",
17016 env->insn_processed);
17017 return -E2BIG;
17018 }
17019
17020 state->last_insn_idx = env->prev_insn_idx;
17021
17022 if (is_prune_point(env, env->insn_idx)) {
17023 err = is_state_visited(env, env->insn_idx);
17024 if (err < 0)
17025 return err;
17026 if (err == 1) {
17027 /* found equivalent state, can prune the search */
17028 if (env->log.level & BPF_LOG_LEVEL) {
17029 if (do_print_state)
17030 verbose(env, "\nfrom %d to %d%s: safe\n",
17031 env->prev_insn_idx, env->insn_idx,
17032 env->cur_state->speculative ?
17033 " (speculative execution)" : "");
17034 else
17035 verbose(env, "%d: safe\n", env->insn_idx);
17036 }
17037 goto process_bpf_exit;
17038 }
17039 }
17040
17041 if (is_jmp_point(env, env->insn_idx)) {
17042 err = push_jmp_history(env, state);
17043 if (err)
17044 return err;
17045 }
17046
17047 if (signal_pending(current))
17048 return -EAGAIN;
17049
17050 if (need_resched())
17051 cond_resched();
17052
17053 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
17054 verbose(env, "\nfrom %d to %d%s:",
17055 env->prev_insn_idx, env->insn_idx,
17056 env->cur_state->speculative ?
17057 " (speculative execution)" : "");
17058 print_verifier_state(env, state->frame[state->curframe], true);
17059 do_print_state = false;
17060 }
17061
17062 if (env->log.level & BPF_LOG_LEVEL) {
17063 const struct bpf_insn_cbs cbs = {
17064 .cb_call = disasm_kfunc_name,
17065 .cb_print = verbose,
17066 .private_data = env,
17067 };
17068
17069 if (verifier_state_scratched(env))
17070 print_insn_state(env, state->frame[state->curframe]);
17071
17072 verbose_linfo(env, env->insn_idx, "; ");
17073 env->prev_log_pos = env->log.end_pos;
17074 verbose(env, "%d: ", env->insn_idx);
17075 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
17076 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
17077 env->prev_log_pos = env->log.end_pos;
17078 }
17079
17080 if (bpf_prog_is_offloaded(env->prog->aux)) {
17081 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
17082 env->prev_insn_idx);
17083 if (err)
17084 return err;
17085 }
17086
17087 regs = cur_regs(env);
17088 sanitize_mark_insn_seen(env);
17089 prev_insn_idx = env->insn_idx;
17090
17091 if (class == BPF_ALU || class == BPF_ALU64) {
17092 err = check_alu_op(env, insn);
17093 if (err)
17094 return err;
17095
17096 } else if (class == BPF_LDX) {
17097 enum bpf_reg_type src_reg_type;
17098
17099 /* check for reserved fields is already done */
17100
17101 /* check src operand */
17102 err = check_reg_arg(env, insn->src_reg, SRC_OP);
17103 if (err)
17104 return err;
17105
17106 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
17107 if (err)
17108 return err;
17109
17110 src_reg_type = regs[insn->src_reg].type;
17111
17112 /* check that memory (src_reg + off) is readable,
17113 * the state of dst_reg will be updated by this func
17114 */
17115 err = check_mem_access(env, env->insn_idx, insn->src_reg,
17116 insn->off, BPF_SIZE(insn->code),
17117 BPF_READ, insn->dst_reg, false,
17118 BPF_MODE(insn->code) == BPF_MEMSX);
17119 if (err)
17120 return err;
17121
17122 err = save_aux_ptr_type(env, src_reg_type, true);
17123 if (err)
17124 return err;
17125 } else if (class == BPF_STX) {
17126 enum bpf_reg_type dst_reg_type;
17127
17128 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
17129 err = check_atomic(env, env->insn_idx, insn);
17130 if (err)
17131 return err;
17132 env->insn_idx++;
17133 continue;
17134 }
17135
17136 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
17137 verbose(env, "BPF_STX uses reserved fields\n");
17138 return -EINVAL;
17139 }
17140
17141 /* check src1 operand */
17142 err = check_reg_arg(env, insn->src_reg, SRC_OP);
17143 if (err)
17144 return err;
17145 /* check src2 operand */
17146 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
17147 if (err)
17148 return err;
17149
17150 dst_reg_type = regs[insn->dst_reg].type;
17151
17152 /* check that memory (dst_reg + off) is writeable */
17153 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
17154 insn->off, BPF_SIZE(insn->code),
17155 BPF_WRITE, insn->src_reg, false, false);
17156 if (err)
17157 return err;
17158
17159 err = save_aux_ptr_type(env, dst_reg_type, false);
17160 if (err)
17161 return err;
17162 } else if (class == BPF_ST) {
17163 enum bpf_reg_type dst_reg_type;
17164
17165 if (BPF_MODE(insn->code) != BPF_MEM ||
17166 insn->src_reg != BPF_REG_0) {
17167 verbose(env, "BPF_ST uses reserved fields\n");
17168 return -EINVAL;
17169 }
17170 /* check src operand */
17171 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
17172 if (err)
17173 return err;
17174
17175 dst_reg_type = regs[insn->dst_reg].type;
17176
17177 /* check that memory (dst_reg + off) is writeable */
17178 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
17179 insn->off, BPF_SIZE(insn->code),
17180 BPF_WRITE, -1, false, false);
17181 if (err)
17182 return err;
17183
17184 err = save_aux_ptr_type(env, dst_reg_type, false);
17185 if (err)
17186 return err;
17187 } else if (class == BPF_JMP || class == BPF_JMP32) {
17188 u8 opcode = BPF_OP(insn->code);
17189
17190 env->jmps_processed++;
17191 if (opcode == BPF_CALL) {
17192 if (BPF_SRC(insn->code) != BPF_K ||
17193 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
17194 && insn->off != 0) ||
17195 (insn->src_reg != BPF_REG_0 &&
17196 insn->src_reg != BPF_PSEUDO_CALL &&
17197 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
17198 insn->dst_reg != BPF_REG_0 ||
17199 class == BPF_JMP32) {
17200 verbose(env, "BPF_CALL uses reserved fields\n");
17201 return -EINVAL;
17202 }
17203
17204 if (env->cur_state->active_lock.ptr) {
17205 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
17206 (insn->src_reg == BPF_PSEUDO_CALL) ||
17207 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
17208 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
17209 verbose(env, "function calls are not allowed while holding a lock\n");
17210 return -EINVAL;
17211 }
17212 }
17213 if (insn->src_reg == BPF_PSEUDO_CALL)
17214 err = check_func_call(env, insn, &env->insn_idx);
17215 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
17216 err = check_kfunc_call(env, insn, &env->insn_idx);
17217 else
17218 err = check_helper_call(env, insn, &env->insn_idx);
17219 if (err)
17220 return err;
17221
17222 mark_reg_scratched(env, BPF_REG_0);
17223 } else if (opcode == BPF_JA) {
17224 if (BPF_SRC(insn->code) != BPF_K ||
17225 insn->src_reg != BPF_REG_0 ||
17226 insn->dst_reg != BPF_REG_0 ||
17227 (class == BPF_JMP && insn->imm != 0) ||
17228 (class == BPF_JMP32 && insn->off != 0)) {
17229 verbose(env, "BPF_JA uses reserved fields\n");
17230 return -EINVAL;
17231 }
17232
17233 if (class == BPF_JMP)
17234 env->insn_idx += insn->off + 1;
17235 else
17236 env->insn_idx += insn->imm + 1;
17237 continue;
17238
17239 } else if (opcode == BPF_EXIT) {
17240 if (BPF_SRC(insn->code) != BPF_K ||
17241 insn->imm != 0 ||
17242 insn->src_reg != BPF_REG_0 ||
17243 insn->dst_reg != BPF_REG_0 ||
17244 class == BPF_JMP32) {
17245 verbose(env, "BPF_EXIT uses reserved fields\n");
17246 return -EINVAL;
17247 }
17248
17249 if (env->cur_state->active_lock.ptr &&
17250 !in_rbtree_lock_required_cb(env)) {
17251 verbose(env, "bpf_spin_unlock is missing\n");
17252 return -EINVAL;
17253 }
17254
17255 if (env->cur_state->active_rcu_lock &&
17256 !in_rbtree_lock_required_cb(env)) {
17257 verbose(env, "bpf_rcu_read_unlock is missing\n");
17258 return -EINVAL;
17259 }
17260
17261 /* We must do check_reference_leak here before
17262 * prepare_func_exit to handle the case when
17263 * state->curframe > 0, it may be a callback
17264 * function, for which reference_state must
17265 * match caller reference state when it exits.
17266 */
17267 err = check_reference_leak(env);
17268 if (err)
17269 return err;
17270
17271 if (state->curframe) {
17272 /* exit from nested function */
17273 err = prepare_func_exit(env, &env->insn_idx);
17274 if (err)
17275 return err;
17276 do_print_state = true;
17277 continue;
17278 }
17279
17280 err = check_return_code(env);
17281 if (err)
17282 return err;
17283 process_bpf_exit:
17284 mark_verifier_state_scratched(env);
17285 update_branch_counts(env, env->cur_state);
17286 err = pop_stack(env, &prev_insn_idx,
17287 &env->insn_idx, pop_log);
17288 if (err < 0) {
17289 if (err != -ENOENT)
17290 return err;
17291 break;
17292 } else {
17293 do_print_state = true;
17294 continue;
17295 }
17296 } else {
17297 err = check_cond_jmp_op(env, insn, &env->insn_idx);
17298 if (err)
17299 return err;
17300 }
17301 } else if (class == BPF_LD) {
17302 u8 mode = BPF_MODE(insn->code);
17303
17304 if (mode == BPF_ABS || mode == BPF_IND) {
17305 err = check_ld_abs(env, insn);
17306 if (err)
17307 return err;
17308
17309 } else if (mode == BPF_IMM) {
17310 err = check_ld_imm(env, insn);
17311 if (err)
17312 return err;
17313
17314 env->insn_idx++;
17315 sanitize_mark_insn_seen(env);
17316 } else {
17317 verbose(env, "invalid BPF_LD mode\n");
17318 return -EINVAL;
17319 }
17320 } else {
17321 verbose(env, "unknown insn class %d\n", class);
17322 return -EINVAL;
17323 }
17324
17325 env->insn_idx++;
17326 }
17327
17328 return 0;
17329 }
17330
find_btf_percpu_datasec(struct btf * btf)17331 static int find_btf_percpu_datasec(struct btf *btf)
17332 {
17333 const struct btf_type *t;
17334 const char *tname;
17335 int i, n;
17336
17337 /*
17338 * Both vmlinux and module each have their own ".data..percpu"
17339 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
17340 * types to look at only module's own BTF types.
17341 */
17342 n = btf_nr_types(btf);
17343 if (btf_is_module(btf))
17344 i = btf_nr_types(btf_vmlinux);
17345 else
17346 i = 1;
17347
17348 for(; i < n; i++) {
17349 t = btf_type_by_id(btf, i);
17350 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
17351 continue;
17352
17353 tname = btf_name_by_offset(btf, t->name_off);
17354 if (!strcmp(tname, ".data..percpu"))
17355 return i;
17356 }
17357
17358 return -ENOENT;
17359 }
17360
17361 /* replace pseudo btf_id with kernel symbol address */
check_pseudo_btf_id(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_insn_aux_data * aux)17362 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
17363 struct bpf_insn *insn,
17364 struct bpf_insn_aux_data *aux)
17365 {
17366 const struct btf_var_secinfo *vsi;
17367 const struct btf_type *datasec;
17368 struct btf_mod_pair *btf_mod;
17369 const struct btf_type *t;
17370 const char *sym_name;
17371 bool percpu = false;
17372 u32 type, id = insn->imm;
17373 struct btf *btf;
17374 s32 datasec_id;
17375 u64 addr;
17376 int i, btf_fd, err;
17377
17378 btf_fd = insn[1].imm;
17379 if (btf_fd) {
17380 btf = btf_get_by_fd(btf_fd);
17381 if (IS_ERR(btf)) {
17382 verbose(env, "invalid module BTF object FD specified.\n");
17383 return -EINVAL;
17384 }
17385 } else {
17386 if (!btf_vmlinux) {
17387 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
17388 return -EINVAL;
17389 }
17390 btf = btf_vmlinux;
17391 btf_get(btf);
17392 }
17393
17394 t = btf_type_by_id(btf, id);
17395 if (!t) {
17396 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
17397 err = -ENOENT;
17398 goto err_put;
17399 }
17400
17401 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
17402 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
17403 err = -EINVAL;
17404 goto err_put;
17405 }
17406
17407 sym_name = btf_name_by_offset(btf, t->name_off);
17408 addr = kallsyms_lookup_name(sym_name);
17409 if (!addr) {
17410 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
17411 sym_name);
17412 err = -ENOENT;
17413 goto err_put;
17414 }
17415 insn[0].imm = (u32)addr;
17416 insn[1].imm = addr >> 32;
17417
17418 if (btf_type_is_func(t)) {
17419 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
17420 aux->btf_var.mem_size = 0;
17421 goto check_btf;
17422 }
17423
17424 datasec_id = find_btf_percpu_datasec(btf);
17425 if (datasec_id > 0) {
17426 datasec = btf_type_by_id(btf, datasec_id);
17427 for_each_vsi(i, datasec, vsi) {
17428 if (vsi->type == id) {
17429 percpu = true;
17430 break;
17431 }
17432 }
17433 }
17434
17435 type = t->type;
17436 t = btf_type_skip_modifiers(btf, type, NULL);
17437 if (percpu) {
17438 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
17439 aux->btf_var.btf = btf;
17440 aux->btf_var.btf_id = type;
17441 } else if (!btf_type_is_struct(t)) {
17442 const struct btf_type *ret;
17443 const char *tname;
17444 u32 tsize;
17445
17446 /* resolve the type size of ksym. */
17447 ret = btf_resolve_size(btf, t, &tsize);
17448 if (IS_ERR(ret)) {
17449 tname = btf_name_by_offset(btf, t->name_off);
17450 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
17451 tname, PTR_ERR(ret));
17452 err = -EINVAL;
17453 goto err_put;
17454 }
17455 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
17456 aux->btf_var.mem_size = tsize;
17457 } else {
17458 aux->btf_var.reg_type = PTR_TO_BTF_ID;
17459 aux->btf_var.btf = btf;
17460 aux->btf_var.btf_id = type;
17461 }
17462 check_btf:
17463 /* check whether we recorded this BTF (and maybe module) already */
17464 for (i = 0; i < env->used_btf_cnt; i++) {
17465 if (env->used_btfs[i].btf == btf) {
17466 btf_put(btf);
17467 return 0;
17468 }
17469 }
17470
17471 if (env->used_btf_cnt >= MAX_USED_BTFS) {
17472 err = -E2BIG;
17473 goto err_put;
17474 }
17475
17476 btf_mod = &env->used_btfs[env->used_btf_cnt];
17477 btf_mod->btf = btf;
17478 btf_mod->module = NULL;
17479
17480 /* if we reference variables from kernel module, bump its refcount */
17481 if (btf_is_module(btf)) {
17482 btf_mod->module = btf_try_get_module(btf);
17483 if (!btf_mod->module) {
17484 err = -ENXIO;
17485 goto err_put;
17486 }
17487 }
17488
17489 env->used_btf_cnt++;
17490
17491 return 0;
17492 err_put:
17493 btf_put(btf);
17494 return err;
17495 }
17496
is_tracing_prog_type(enum bpf_prog_type type)17497 static bool is_tracing_prog_type(enum bpf_prog_type type)
17498 {
17499 switch (type) {
17500 case BPF_PROG_TYPE_KPROBE:
17501 case BPF_PROG_TYPE_TRACEPOINT:
17502 case BPF_PROG_TYPE_PERF_EVENT:
17503 case BPF_PROG_TYPE_RAW_TRACEPOINT:
17504 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
17505 return true;
17506 default:
17507 return false;
17508 }
17509 }
17510
check_map_prog_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,struct bpf_prog * prog)17511 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
17512 struct bpf_map *map,
17513 struct bpf_prog *prog)
17514
17515 {
17516 enum bpf_prog_type prog_type = resolve_prog_type(prog);
17517
17518 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
17519 btf_record_has_field(map->record, BPF_RB_ROOT)) {
17520 if (is_tracing_prog_type(prog_type)) {
17521 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
17522 return -EINVAL;
17523 }
17524 }
17525
17526 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
17527 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
17528 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
17529 return -EINVAL;
17530 }
17531
17532 if (is_tracing_prog_type(prog_type)) {
17533 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
17534 return -EINVAL;
17535 }
17536 }
17537
17538 if (btf_record_has_field(map->record, BPF_TIMER)) {
17539 if (is_tracing_prog_type(prog_type)) {
17540 verbose(env, "tracing progs cannot use bpf_timer yet\n");
17541 return -EINVAL;
17542 }
17543 }
17544
17545 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
17546 !bpf_offload_prog_map_match(prog, map)) {
17547 verbose(env, "offload device mismatch between prog and map\n");
17548 return -EINVAL;
17549 }
17550
17551 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
17552 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
17553 return -EINVAL;
17554 }
17555
17556 if (prog->aux->sleepable)
17557 switch (map->map_type) {
17558 case BPF_MAP_TYPE_HASH:
17559 case BPF_MAP_TYPE_LRU_HASH:
17560 case BPF_MAP_TYPE_ARRAY:
17561 case BPF_MAP_TYPE_PERCPU_HASH:
17562 case BPF_MAP_TYPE_PERCPU_ARRAY:
17563 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
17564 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
17565 case BPF_MAP_TYPE_HASH_OF_MAPS:
17566 case BPF_MAP_TYPE_RINGBUF:
17567 case BPF_MAP_TYPE_USER_RINGBUF:
17568 case BPF_MAP_TYPE_INODE_STORAGE:
17569 case BPF_MAP_TYPE_SK_STORAGE:
17570 case BPF_MAP_TYPE_TASK_STORAGE:
17571 case BPF_MAP_TYPE_CGRP_STORAGE:
17572 break;
17573 default:
17574 verbose(env,
17575 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
17576 return -EINVAL;
17577 }
17578
17579 return 0;
17580 }
17581
bpf_map_is_cgroup_storage(struct bpf_map * map)17582 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
17583 {
17584 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
17585 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
17586 }
17587
17588 /* find and rewrite pseudo imm in ld_imm64 instructions:
17589 *
17590 * 1. if it accesses map FD, replace it with actual map pointer.
17591 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
17592 *
17593 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
17594 */
resolve_pseudo_ldimm64(struct bpf_verifier_env * env)17595 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
17596 {
17597 struct bpf_insn *insn = env->prog->insnsi;
17598 int insn_cnt = env->prog->len;
17599 int i, j, err;
17600
17601 err = bpf_prog_calc_tag(env->prog);
17602 if (err)
17603 return err;
17604
17605 for (i = 0; i < insn_cnt; i++, insn++) {
17606 if (BPF_CLASS(insn->code) == BPF_LDX &&
17607 ((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
17608 insn->imm != 0)) {
17609 verbose(env, "BPF_LDX uses reserved fields\n");
17610 return -EINVAL;
17611 }
17612
17613 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
17614 struct bpf_insn_aux_data *aux;
17615 struct bpf_map *map;
17616 struct fd f;
17617 u64 addr;
17618 u32 fd;
17619
17620 if (i == insn_cnt - 1 || insn[1].code != 0 ||
17621 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
17622 insn[1].off != 0) {
17623 verbose(env, "invalid bpf_ld_imm64 insn\n");
17624 return -EINVAL;
17625 }
17626
17627 if (insn[0].src_reg == 0)
17628 /* valid generic load 64-bit imm */
17629 goto next_insn;
17630
17631 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
17632 aux = &env->insn_aux_data[i];
17633 err = check_pseudo_btf_id(env, insn, aux);
17634 if (err)
17635 return err;
17636 goto next_insn;
17637 }
17638
17639 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
17640 aux = &env->insn_aux_data[i];
17641 aux->ptr_type = PTR_TO_FUNC;
17642 goto next_insn;
17643 }
17644
17645 /* In final convert_pseudo_ld_imm64() step, this is
17646 * converted into regular 64-bit imm load insn.
17647 */
17648 switch (insn[0].src_reg) {
17649 case BPF_PSEUDO_MAP_VALUE:
17650 case BPF_PSEUDO_MAP_IDX_VALUE:
17651 break;
17652 case BPF_PSEUDO_MAP_FD:
17653 case BPF_PSEUDO_MAP_IDX:
17654 if (insn[1].imm == 0)
17655 break;
17656 fallthrough;
17657 default:
17658 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
17659 return -EINVAL;
17660 }
17661
17662 switch (insn[0].src_reg) {
17663 case BPF_PSEUDO_MAP_IDX_VALUE:
17664 case BPF_PSEUDO_MAP_IDX:
17665 if (bpfptr_is_null(env->fd_array)) {
17666 verbose(env, "fd_idx without fd_array is invalid\n");
17667 return -EPROTO;
17668 }
17669 if (copy_from_bpfptr_offset(&fd, env->fd_array,
17670 insn[0].imm * sizeof(fd),
17671 sizeof(fd)))
17672 return -EFAULT;
17673 break;
17674 default:
17675 fd = insn[0].imm;
17676 break;
17677 }
17678
17679 f = fdget(fd);
17680 map = __bpf_map_get(f);
17681 if (IS_ERR(map)) {
17682 verbose(env, "fd %d is not pointing to valid bpf_map\n", fd);
17683 return PTR_ERR(map);
17684 }
17685
17686 err = check_map_prog_compatibility(env, map, env->prog);
17687 if (err) {
17688 fdput(f);
17689 return err;
17690 }
17691
17692 aux = &env->insn_aux_data[i];
17693 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
17694 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
17695 addr = (unsigned long)map;
17696 } else {
17697 u32 off = insn[1].imm;
17698
17699 if (off >= BPF_MAX_VAR_OFF) {
17700 verbose(env, "direct value offset of %u is not allowed\n", off);
17701 fdput(f);
17702 return -EINVAL;
17703 }
17704
17705 if (!map->ops->map_direct_value_addr) {
17706 verbose(env, "no direct value access support for this map type\n");
17707 fdput(f);
17708 return -EINVAL;
17709 }
17710
17711 err = map->ops->map_direct_value_addr(map, &addr, off);
17712 if (err) {
17713 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
17714 map->value_size, off);
17715 fdput(f);
17716 return err;
17717 }
17718
17719 aux->map_off = off;
17720 addr += off;
17721 }
17722
17723 insn[0].imm = (u32)addr;
17724 insn[1].imm = addr >> 32;
17725
17726 /* check whether we recorded this map already */
17727 for (j = 0; j < env->used_map_cnt; j++) {
17728 if (env->used_maps[j] == map) {
17729 aux->map_index = j;
17730 fdput(f);
17731 goto next_insn;
17732 }
17733 }
17734
17735 if (env->used_map_cnt >= MAX_USED_MAPS) {
17736 fdput(f);
17737 return -E2BIG;
17738 }
17739
17740 if (env->prog->aux->sleepable)
17741 atomic64_inc(&map->sleepable_refcnt);
17742 /* hold the map. If the program is rejected by verifier,
17743 * the map will be released by release_maps() or it
17744 * will be used by the valid program until it's unloaded
17745 * and all maps are released in bpf_free_used_maps()
17746 */
17747 bpf_map_inc(map);
17748
17749 aux->map_index = env->used_map_cnt;
17750 env->used_maps[env->used_map_cnt++] = map;
17751
17752 if (bpf_map_is_cgroup_storage(map) &&
17753 bpf_cgroup_storage_assign(env->prog->aux, map)) {
17754 verbose(env, "only one cgroup storage of each type is allowed\n");
17755 fdput(f);
17756 return -EBUSY;
17757 }
17758
17759 fdput(f);
17760 next_insn:
17761 insn++;
17762 i++;
17763 continue;
17764 }
17765
17766 /* Basic sanity check before we invest more work here. */
17767 if (!bpf_opcode_in_insntable(insn->code)) {
17768 verbose(env, "unknown opcode %02x\n", insn->code);
17769 return -EINVAL;
17770 }
17771 }
17772
17773 /* now all pseudo BPF_LD_IMM64 instructions load valid
17774 * 'struct bpf_map *' into a register instead of user map_fd.
17775 * These pointers will be used later by verifier to validate map access.
17776 */
17777 return 0;
17778 }
17779
17780 /* drop refcnt of maps used by the rejected program */
release_maps(struct bpf_verifier_env * env)17781 static void release_maps(struct bpf_verifier_env *env)
17782 {
17783 __bpf_free_used_maps(env->prog->aux, env->used_maps,
17784 env->used_map_cnt);
17785 }
17786
17787 /* drop refcnt of maps used by the rejected program */
release_btfs(struct bpf_verifier_env * env)17788 static void release_btfs(struct bpf_verifier_env *env)
17789 {
17790 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
17791 env->used_btf_cnt);
17792 }
17793
17794 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
convert_pseudo_ld_imm64(struct bpf_verifier_env * env)17795 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
17796 {
17797 struct bpf_insn *insn = env->prog->insnsi;
17798 int insn_cnt = env->prog->len;
17799 int i;
17800
17801 for (i = 0; i < insn_cnt; i++, insn++) {
17802 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
17803 continue;
17804 if (insn->src_reg == BPF_PSEUDO_FUNC)
17805 continue;
17806 insn->src_reg = 0;
17807 }
17808 }
17809
17810 /* single env->prog->insni[off] instruction was replaced with the range
17811 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
17812 * [0, off) and [off, end) to new locations, so the patched range stays zero
17813 */
adjust_insn_aux_data(struct bpf_verifier_env * env,struct bpf_insn_aux_data * new_data,struct bpf_prog * new_prog,u32 off,u32 cnt)17814 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17815 struct bpf_insn_aux_data *new_data,
17816 struct bpf_prog *new_prog, u32 off, u32 cnt)
17817 {
17818 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17819 struct bpf_insn *insn = new_prog->insnsi;
17820 u32 old_seen = old_data[off].seen;
17821 u32 prog_len;
17822 int i;
17823
17824 /* aux info at OFF always needs adjustment, no matter fast path
17825 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17826 * original insn at old prog.
17827 */
17828 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17829
17830 if (cnt == 1)
17831 return;
17832 prog_len = new_prog->len;
17833
17834 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17835 memcpy(new_data + off + cnt - 1, old_data + off,
17836 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17837 for (i = off; i < off + cnt - 1; i++) {
17838 /* Expand insni[off]'s seen count to the patched range. */
17839 new_data[i].seen = old_seen;
17840 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17841 }
17842 env->insn_aux_data = new_data;
17843 vfree(old_data);
17844 }
17845
adjust_subprog_starts(struct bpf_verifier_env * env,u32 off,u32 len)17846 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17847 {
17848 int i;
17849
17850 if (len == 1)
17851 return;
17852 /* NOTE: fake 'exit' subprog should be updated as well. */
17853 for (i = 0; i <= env->subprog_cnt; i++) {
17854 if (env->subprog_info[i].start <= off)
17855 continue;
17856 env->subprog_info[i].start += len - 1;
17857 }
17858 }
17859
adjust_poke_descs(struct bpf_prog * prog,u32 off,u32 len)17860 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17861 {
17862 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17863 int i, sz = prog->aux->size_poke_tab;
17864 struct bpf_jit_poke_descriptor *desc;
17865
17866 for (i = 0; i < sz; i++) {
17867 desc = &tab[i];
17868 if (desc->insn_idx <= off)
17869 continue;
17870 desc->insn_idx += len - 1;
17871 }
17872 }
17873
bpf_patch_insn_data(struct bpf_verifier_env * env,u32 off,const struct bpf_insn * patch,u32 len)17874 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17875 const struct bpf_insn *patch, u32 len)
17876 {
17877 struct bpf_prog *new_prog;
17878 struct bpf_insn_aux_data *new_data = NULL;
17879
17880 if (len > 1) {
17881 new_data = vzalloc(array_size(env->prog->len + len - 1,
17882 sizeof(struct bpf_insn_aux_data)));
17883 if (!new_data)
17884 return NULL;
17885 }
17886
17887 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17888 if (IS_ERR(new_prog)) {
17889 if (PTR_ERR(new_prog) == -ERANGE)
17890 verbose(env,
17891 "insn %d cannot be patched due to 16-bit range\n",
17892 env->insn_aux_data[off].orig_idx);
17893 vfree(new_data);
17894 return NULL;
17895 }
17896 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17897 adjust_subprog_starts(env, off, len);
17898 adjust_poke_descs(new_prog, off, len);
17899 return new_prog;
17900 }
17901
adjust_subprog_starts_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)17902 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17903 u32 off, u32 cnt)
17904 {
17905 int i, j;
17906
17907 /* find first prog starting at or after off (first to remove) */
17908 for (i = 0; i < env->subprog_cnt; i++)
17909 if (env->subprog_info[i].start >= off)
17910 break;
17911 /* find first prog starting at or after off + cnt (first to stay) */
17912 for (j = i; j < env->subprog_cnt; j++)
17913 if (env->subprog_info[j].start >= off + cnt)
17914 break;
17915 /* if j doesn't start exactly at off + cnt, we are just removing
17916 * the front of previous prog
17917 */
17918 if (env->subprog_info[j].start != off + cnt)
17919 j--;
17920
17921 if (j > i) {
17922 struct bpf_prog_aux *aux = env->prog->aux;
17923 int move;
17924
17925 /* move fake 'exit' subprog as well */
17926 move = env->subprog_cnt + 1 - j;
17927
17928 memmove(env->subprog_info + i,
17929 env->subprog_info + j,
17930 sizeof(*env->subprog_info) * move);
17931 env->subprog_cnt -= j - i;
17932
17933 /* remove func_info */
17934 if (aux->func_info) {
17935 move = aux->func_info_cnt - j;
17936
17937 memmove(aux->func_info + i,
17938 aux->func_info + j,
17939 sizeof(*aux->func_info) * move);
17940 aux->func_info_cnt -= j - i;
17941 /* func_info->insn_off is set after all code rewrites,
17942 * in adjust_btf_func() - no need to adjust
17943 */
17944 }
17945 } else {
17946 /* convert i from "first prog to remove" to "first to adjust" */
17947 if (env->subprog_info[i].start == off)
17948 i++;
17949 }
17950
17951 /* update fake 'exit' subprog as well */
17952 for (; i <= env->subprog_cnt; i++)
17953 env->subprog_info[i].start -= cnt;
17954
17955 return 0;
17956 }
17957
bpf_adj_linfo_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)17958 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17959 u32 cnt)
17960 {
17961 struct bpf_prog *prog = env->prog;
17962 u32 i, l_off, l_cnt, nr_linfo;
17963 struct bpf_line_info *linfo;
17964
17965 nr_linfo = prog->aux->nr_linfo;
17966 if (!nr_linfo)
17967 return 0;
17968
17969 linfo = prog->aux->linfo;
17970
17971 /* find first line info to remove, count lines to be removed */
17972 for (i = 0; i < nr_linfo; i++)
17973 if (linfo[i].insn_off >= off)
17974 break;
17975
17976 l_off = i;
17977 l_cnt = 0;
17978 for (; i < nr_linfo; i++)
17979 if (linfo[i].insn_off < off + cnt)
17980 l_cnt++;
17981 else
17982 break;
17983
17984 /* First live insn doesn't match first live linfo, it needs to "inherit"
17985 * last removed linfo. prog is already modified, so prog->len == off
17986 * means no live instructions after (tail of the program was removed).
17987 */
17988 if (prog->len != off && l_cnt &&
17989 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17990 l_cnt--;
17991 linfo[--i].insn_off = off + cnt;
17992 }
17993
17994 /* remove the line info which refer to the removed instructions */
17995 if (l_cnt) {
17996 memmove(linfo + l_off, linfo + i,
17997 sizeof(*linfo) * (nr_linfo - i));
17998
17999 prog->aux->nr_linfo -= l_cnt;
18000 nr_linfo = prog->aux->nr_linfo;
18001 }
18002
18003 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
18004 for (i = l_off; i < nr_linfo; i++)
18005 linfo[i].insn_off -= cnt;
18006
18007 /* fix up all subprogs (incl. 'exit') which start >= off */
18008 for (i = 0; i <= env->subprog_cnt; i++)
18009 if (env->subprog_info[i].linfo_idx > l_off) {
18010 /* program may have started in the removed region but
18011 * may not be fully removed
18012 */
18013 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
18014 env->subprog_info[i].linfo_idx -= l_cnt;
18015 else
18016 env->subprog_info[i].linfo_idx = l_off;
18017 }
18018
18019 return 0;
18020 }
18021
verifier_remove_insns(struct bpf_verifier_env * env,u32 off,u32 cnt)18022 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
18023 {
18024 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18025 unsigned int orig_prog_len = env->prog->len;
18026 int err;
18027
18028 if (bpf_prog_is_offloaded(env->prog->aux))
18029 bpf_prog_offload_remove_insns(env, off, cnt);
18030
18031 err = bpf_remove_insns(env->prog, off, cnt);
18032 if (err)
18033 return err;
18034
18035 err = adjust_subprog_starts_after_remove(env, off, cnt);
18036 if (err)
18037 return err;
18038
18039 err = bpf_adj_linfo_after_remove(env, off, cnt);
18040 if (err)
18041 return err;
18042
18043 memmove(aux_data + off, aux_data + off + cnt,
18044 sizeof(*aux_data) * (orig_prog_len - off - cnt));
18045
18046 return 0;
18047 }
18048
18049 /* The verifier does more data flow analysis than llvm and will not
18050 * explore branches that are dead at run time. Malicious programs can
18051 * have dead code too. Therefore replace all dead at-run-time code
18052 * with 'ja -1'.
18053 *
18054 * Just nops are not optimal, e.g. if they would sit at the end of the
18055 * program and through another bug we would manage to jump there, then
18056 * we'd execute beyond program memory otherwise. Returning exception
18057 * code also wouldn't work since we can have subprogs where the dead
18058 * code could be located.
18059 */
sanitize_dead_code(struct bpf_verifier_env * env)18060 static void sanitize_dead_code(struct bpf_verifier_env *env)
18061 {
18062 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18063 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
18064 struct bpf_insn *insn = env->prog->insnsi;
18065 const int insn_cnt = env->prog->len;
18066 int i;
18067
18068 for (i = 0; i < insn_cnt; i++) {
18069 if (aux_data[i].seen)
18070 continue;
18071 memcpy(insn + i, &trap, sizeof(trap));
18072 aux_data[i].zext_dst = false;
18073 }
18074 }
18075
insn_is_cond_jump(u8 code)18076 static bool insn_is_cond_jump(u8 code)
18077 {
18078 u8 op;
18079
18080 op = BPF_OP(code);
18081 if (BPF_CLASS(code) == BPF_JMP32)
18082 return op != BPF_JA;
18083
18084 if (BPF_CLASS(code) != BPF_JMP)
18085 return false;
18086
18087 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
18088 }
18089
opt_hard_wire_dead_code_branches(struct bpf_verifier_env * env)18090 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
18091 {
18092 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18093 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
18094 struct bpf_insn *insn = env->prog->insnsi;
18095 const int insn_cnt = env->prog->len;
18096 int i;
18097
18098 for (i = 0; i < insn_cnt; i++, insn++) {
18099 if (!insn_is_cond_jump(insn->code))
18100 continue;
18101
18102 if (!aux_data[i + 1].seen)
18103 ja.off = insn->off;
18104 else if (!aux_data[i + 1 + insn->off].seen)
18105 ja.off = 0;
18106 else
18107 continue;
18108
18109 if (bpf_prog_is_offloaded(env->prog->aux))
18110 bpf_prog_offload_replace_insn(env, i, &ja);
18111
18112 memcpy(insn, &ja, sizeof(ja));
18113 }
18114 }
18115
opt_remove_dead_code(struct bpf_verifier_env * env)18116 static int opt_remove_dead_code(struct bpf_verifier_env *env)
18117 {
18118 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18119 int insn_cnt = env->prog->len;
18120 int i, err;
18121
18122 for (i = 0; i < insn_cnt; i++) {
18123 int j;
18124
18125 j = 0;
18126 while (i + j < insn_cnt && !aux_data[i + j].seen)
18127 j++;
18128 if (!j)
18129 continue;
18130
18131 err = verifier_remove_insns(env, i, j);
18132 if (err)
18133 return err;
18134 insn_cnt = env->prog->len;
18135 }
18136
18137 return 0;
18138 }
18139
opt_remove_nops(struct bpf_verifier_env * env)18140 static int opt_remove_nops(struct bpf_verifier_env *env)
18141 {
18142 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
18143 struct bpf_insn *insn = env->prog->insnsi;
18144 int insn_cnt = env->prog->len;
18145 int i, err;
18146
18147 for (i = 0; i < insn_cnt; i++) {
18148 if (memcmp(&insn[i], &ja, sizeof(ja)))
18149 continue;
18150
18151 err = verifier_remove_insns(env, i, 1);
18152 if (err)
18153 return err;
18154 insn_cnt--;
18155 i--;
18156 }
18157
18158 return 0;
18159 }
18160
opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env * env,const union bpf_attr * attr)18161 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
18162 const union bpf_attr *attr)
18163 {
18164 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
18165 struct bpf_insn_aux_data *aux = env->insn_aux_data;
18166 int i, patch_len, delta = 0, len = env->prog->len;
18167 struct bpf_insn *insns = env->prog->insnsi;
18168 struct bpf_prog *new_prog;
18169 bool rnd_hi32;
18170
18171 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
18172 zext_patch[1] = BPF_ZEXT_REG(0);
18173 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
18174 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
18175 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
18176 for (i = 0; i < len; i++) {
18177 int adj_idx = i + delta;
18178 struct bpf_insn insn;
18179 int load_reg;
18180
18181 insn = insns[adj_idx];
18182 load_reg = insn_def_regno(&insn);
18183 if (!aux[adj_idx].zext_dst) {
18184 u8 code, class;
18185 u32 imm_rnd;
18186
18187 if (!rnd_hi32)
18188 continue;
18189
18190 code = insn.code;
18191 class = BPF_CLASS(code);
18192 if (load_reg == -1)
18193 continue;
18194
18195 /* NOTE: arg "reg" (the fourth one) is only used for
18196 * BPF_STX + SRC_OP, so it is safe to pass NULL
18197 * here.
18198 */
18199 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
18200 if (class == BPF_LD &&
18201 BPF_MODE(code) == BPF_IMM)
18202 i++;
18203 continue;
18204 }
18205
18206 /* ctx load could be transformed into wider load. */
18207 if (class == BPF_LDX &&
18208 aux[adj_idx].ptr_type == PTR_TO_CTX)
18209 continue;
18210
18211 imm_rnd = get_random_u32();
18212 rnd_hi32_patch[0] = insn;
18213 rnd_hi32_patch[1].imm = imm_rnd;
18214 rnd_hi32_patch[3].dst_reg = load_reg;
18215 patch = rnd_hi32_patch;
18216 patch_len = 4;
18217 goto apply_patch_buffer;
18218 }
18219
18220 /* Add in an zero-extend instruction if a) the JIT has requested
18221 * it or b) it's a CMPXCHG.
18222 *
18223 * The latter is because: BPF_CMPXCHG always loads a value into
18224 * R0, therefore always zero-extends. However some archs'
18225 * equivalent instruction only does this load when the
18226 * comparison is successful. This detail of CMPXCHG is
18227 * orthogonal to the general zero-extension behaviour of the
18228 * CPU, so it's treated independently of bpf_jit_needs_zext.
18229 */
18230 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
18231 continue;
18232
18233 /* Zero-extension is done by the caller. */
18234 if (bpf_pseudo_kfunc_call(&insn))
18235 continue;
18236
18237 if (WARN_ON(load_reg == -1)) {
18238 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
18239 return -EFAULT;
18240 }
18241
18242 zext_patch[0] = insn;
18243 zext_patch[1].dst_reg = load_reg;
18244 zext_patch[1].src_reg = load_reg;
18245 patch = zext_patch;
18246 patch_len = 2;
18247 apply_patch_buffer:
18248 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
18249 if (!new_prog)
18250 return -ENOMEM;
18251 env->prog = new_prog;
18252 insns = new_prog->insnsi;
18253 aux = env->insn_aux_data;
18254 delta += patch_len - 1;
18255 }
18256
18257 return 0;
18258 }
18259
18260 /* convert load instructions that access fields of a context type into a
18261 * sequence of instructions that access fields of the underlying structure:
18262 * struct __sk_buff -> struct sk_buff
18263 * struct bpf_sock_ops -> struct sock
18264 */
convert_ctx_accesses(struct bpf_verifier_env * env)18265 static int convert_ctx_accesses(struct bpf_verifier_env *env)
18266 {
18267 const struct bpf_verifier_ops *ops = env->ops;
18268 int i, cnt, size, ctx_field_size, delta = 0;
18269 const int insn_cnt = env->prog->len;
18270 struct bpf_insn insn_buf[16], *insn;
18271 u32 target_size, size_default, off;
18272 struct bpf_prog *new_prog;
18273 enum bpf_access_type type;
18274 bool is_narrower_load;
18275
18276 if (ops->gen_prologue || env->seen_direct_write) {
18277 if (!ops->gen_prologue) {
18278 verbose(env, "bpf verifier is misconfigured\n");
18279 return -EINVAL;
18280 }
18281 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
18282 env->prog);
18283 if (cnt >= ARRAY_SIZE(insn_buf)) {
18284 verbose(env, "bpf verifier is misconfigured\n");
18285 return -EINVAL;
18286 } else if (cnt) {
18287 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
18288 if (!new_prog)
18289 return -ENOMEM;
18290
18291 env->prog = new_prog;
18292 delta += cnt - 1;
18293 }
18294 }
18295
18296 if (bpf_prog_is_offloaded(env->prog->aux))
18297 return 0;
18298
18299 insn = env->prog->insnsi + delta;
18300
18301 for (i = 0; i < insn_cnt; i++, insn++) {
18302 bpf_convert_ctx_access_t convert_ctx_access;
18303 u8 mode;
18304
18305 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
18306 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
18307 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
18308 insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
18309 insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
18310 insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
18311 insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
18312 type = BPF_READ;
18313 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
18314 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
18315 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
18316 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
18317 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
18318 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
18319 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
18320 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
18321 type = BPF_WRITE;
18322 } else {
18323 continue;
18324 }
18325
18326 if (type == BPF_WRITE &&
18327 env->insn_aux_data[i + delta].sanitize_stack_spill) {
18328 struct bpf_insn patch[] = {
18329 *insn,
18330 BPF_ST_NOSPEC(),
18331 };
18332
18333 cnt = ARRAY_SIZE(patch);
18334 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
18335 if (!new_prog)
18336 return -ENOMEM;
18337
18338 delta += cnt - 1;
18339 env->prog = new_prog;
18340 insn = new_prog->insnsi + i + delta;
18341 continue;
18342 }
18343
18344 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
18345 case PTR_TO_CTX:
18346 if (!ops->convert_ctx_access)
18347 continue;
18348 convert_ctx_access = ops->convert_ctx_access;
18349 break;
18350 case PTR_TO_SOCKET:
18351 case PTR_TO_SOCK_COMMON:
18352 convert_ctx_access = bpf_sock_convert_ctx_access;
18353 break;
18354 case PTR_TO_TCP_SOCK:
18355 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
18356 break;
18357 case PTR_TO_XDP_SOCK:
18358 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
18359 break;
18360 case PTR_TO_BTF_ID:
18361 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
18362 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
18363 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
18364 * be said once it is marked PTR_UNTRUSTED, hence we must handle
18365 * any faults for loads into such types. BPF_WRITE is disallowed
18366 * for this case.
18367 */
18368 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
18369 if (type == BPF_READ) {
18370 if (BPF_MODE(insn->code) == BPF_MEM)
18371 insn->code = BPF_LDX | BPF_PROBE_MEM |
18372 BPF_SIZE((insn)->code);
18373 else
18374 insn->code = BPF_LDX | BPF_PROBE_MEMSX |
18375 BPF_SIZE((insn)->code);
18376 env->prog->aux->num_exentries++;
18377 }
18378 continue;
18379 default:
18380 continue;
18381 }
18382
18383 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
18384 size = BPF_LDST_BYTES(insn);
18385 mode = BPF_MODE(insn->code);
18386
18387 /* If the read access is a narrower load of the field,
18388 * convert to a 4/8-byte load, to minimum program type specific
18389 * convert_ctx_access changes. If conversion is successful,
18390 * we will apply proper mask to the result.
18391 */
18392 is_narrower_load = size < ctx_field_size;
18393 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
18394 off = insn->off;
18395 if (is_narrower_load) {
18396 u8 size_code;
18397
18398 if (type == BPF_WRITE) {
18399 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
18400 return -EINVAL;
18401 }
18402
18403 size_code = BPF_H;
18404 if (ctx_field_size == 4)
18405 size_code = BPF_W;
18406 else if (ctx_field_size == 8)
18407 size_code = BPF_DW;
18408
18409 insn->off = off & ~(size_default - 1);
18410 insn->code = BPF_LDX | BPF_MEM | size_code;
18411 }
18412
18413 target_size = 0;
18414 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
18415 &target_size);
18416 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
18417 (ctx_field_size && !target_size)) {
18418 verbose(env, "bpf verifier is misconfigured\n");
18419 return -EINVAL;
18420 }
18421
18422 if (is_narrower_load && size < target_size) {
18423 u8 shift = bpf_ctx_narrow_access_offset(
18424 off, size, size_default) * 8;
18425 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
18426 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
18427 return -EINVAL;
18428 }
18429 if (ctx_field_size <= 4) {
18430 if (shift)
18431 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
18432 insn->dst_reg,
18433 shift);
18434 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
18435 (1 << size * 8) - 1);
18436 } else {
18437 if (shift)
18438 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
18439 insn->dst_reg,
18440 shift);
18441 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
18442 (1ULL << size * 8) - 1);
18443 }
18444 }
18445 if (mode == BPF_MEMSX)
18446 insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
18447 insn->dst_reg, insn->dst_reg,
18448 size * 8, 0);
18449
18450 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18451 if (!new_prog)
18452 return -ENOMEM;
18453
18454 delta += cnt - 1;
18455
18456 /* keep walking new program and skip insns we just inserted */
18457 env->prog = new_prog;
18458 insn = new_prog->insnsi + i + delta;
18459 }
18460
18461 return 0;
18462 }
18463
jit_subprogs(struct bpf_verifier_env * env)18464 static int jit_subprogs(struct bpf_verifier_env *env)
18465 {
18466 struct bpf_prog *prog = env->prog, **func, *tmp;
18467 int i, j, subprog_start, subprog_end = 0, len, subprog;
18468 struct bpf_map *map_ptr;
18469 struct bpf_insn *insn;
18470 void *old_bpf_func;
18471 int err, num_exentries;
18472
18473 if (env->subprog_cnt <= 1)
18474 return 0;
18475
18476 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18477 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
18478 continue;
18479
18480 /* Upon error here we cannot fall back to interpreter but
18481 * need a hard reject of the program. Thus -EFAULT is
18482 * propagated in any case.
18483 */
18484 subprog = find_subprog(env, i + insn->imm + 1);
18485 if (subprog < 0) {
18486 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
18487 i + insn->imm + 1);
18488 return -EFAULT;
18489 }
18490 /* temporarily remember subprog id inside insn instead of
18491 * aux_data, since next loop will split up all insns into funcs
18492 */
18493 insn->off = subprog;
18494 /* remember original imm in case JIT fails and fallback
18495 * to interpreter will be needed
18496 */
18497 env->insn_aux_data[i].call_imm = insn->imm;
18498 /* point imm to __bpf_call_base+1 from JITs point of view */
18499 insn->imm = 1;
18500 if (bpf_pseudo_func(insn))
18501 /* jit (e.g. x86_64) may emit fewer instructions
18502 * if it learns a u32 imm is the same as a u64 imm.
18503 * Force a non zero here.
18504 */
18505 insn[1].imm = 1;
18506 }
18507
18508 err = bpf_prog_alloc_jited_linfo(prog);
18509 if (err)
18510 goto out_undo_insn;
18511
18512 err = -ENOMEM;
18513 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
18514 if (!func)
18515 goto out_undo_insn;
18516
18517 for (i = 0; i < env->subprog_cnt; i++) {
18518 subprog_start = subprog_end;
18519 subprog_end = env->subprog_info[i + 1].start;
18520
18521 len = subprog_end - subprog_start;
18522 /* bpf_prog_run() doesn't call subprogs directly,
18523 * hence main prog stats include the runtime of subprogs.
18524 * subprogs don't have IDs and not reachable via prog_get_next_id
18525 * func[i]->stats will never be accessed and stays NULL
18526 */
18527 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
18528 if (!func[i])
18529 goto out_free;
18530 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
18531 len * sizeof(struct bpf_insn));
18532 func[i]->type = prog->type;
18533 func[i]->len = len;
18534 if (bpf_prog_calc_tag(func[i]))
18535 goto out_free;
18536 func[i]->is_func = 1;
18537 func[i]->aux->func_idx = i;
18538 /* Below members will be freed only at prog->aux */
18539 func[i]->aux->btf = prog->aux->btf;
18540 func[i]->aux->func_info = prog->aux->func_info;
18541 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
18542 func[i]->aux->poke_tab = prog->aux->poke_tab;
18543 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
18544
18545 for (j = 0; j < prog->aux->size_poke_tab; j++) {
18546 struct bpf_jit_poke_descriptor *poke;
18547
18548 poke = &prog->aux->poke_tab[j];
18549 if (poke->insn_idx < subprog_end &&
18550 poke->insn_idx >= subprog_start)
18551 poke->aux = func[i]->aux;
18552 }
18553
18554 func[i]->aux->name[0] = 'F';
18555 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
18556 func[i]->jit_requested = 1;
18557 func[i]->blinding_requested = prog->blinding_requested;
18558 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
18559 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
18560 func[i]->aux->linfo = prog->aux->linfo;
18561 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
18562 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
18563 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
18564 num_exentries = 0;
18565 insn = func[i]->insnsi;
18566 for (j = 0; j < func[i]->len; j++, insn++) {
18567 if (BPF_CLASS(insn->code) == BPF_LDX &&
18568 (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
18569 BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
18570 num_exentries++;
18571 }
18572 func[i]->aux->num_exentries = num_exentries;
18573 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
18574 func[i] = bpf_int_jit_compile(func[i]);
18575 if (!func[i]->jited) {
18576 err = -ENOTSUPP;
18577 goto out_free;
18578 }
18579 cond_resched();
18580 }
18581
18582 /* at this point all bpf functions were successfully JITed
18583 * now populate all bpf_calls with correct addresses and
18584 * run last pass of JIT
18585 */
18586 for (i = 0; i < env->subprog_cnt; i++) {
18587 insn = func[i]->insnsi;
18588 for (j = 0; j < func[i]->len; j++, insn++) {
18589 if (bpf_pseudo_func(insn)) {
18590 subprog = insn->off;
18591 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
18592 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
18593 continue;
18594 }
18595 if (!bpf_pseudo_call(insn))
18596 continue;
18597 subprog = insn->off;
18598 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
18599 }
18600
18601 /* we use the aux data to keep a list of the start addresses
18602 * of the JITed images for each function in the program
18603 *
18604 * for some architectures, such as powerpc64, the imm field
18605 * might not be large enough to hold the offset of the start
18606 * address of the callee's JITed image from __bpf_call_base
18607 *
18608 * in such cases, we can lookup the start address of a callee
18609 * by using its subprog id, available from the off field of
18610 * the call instruction, as an index for this list
18611 */
18612 func[i]->aux->func = func;
18613 func[i]->aux->func_cnt = env->subprog_cnt;
18614 }
18615 for (i = 0; i < env->subprog_cnt; i++) {
18616 old_bpf_func = func[i]->bpf_func;
18617 tmp = bpf_int_jit_compile(func[i]);
18618 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
18619 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
18620 err = -ENOTSUPP;
18621 goto out_free;
18622 }
18623 cond_resched();
18624 }
18625
18626 /* finally lock prog and jit images for all functions and
18627 * populate kallsysm. Begin at the first subprogram, since
18628 * bpf_prog_load will add the kallsyms for the main program.
18629 */
18630 for (i = 1; i < env->subprog_cnt; i++) {
18631 bpf_prog_lock_ro(func[i]);
18632 bpf_prog_kallsyms_add(func[i]);
18633 }
18634
18635 /* Last step: make now unused interpreter insns from main
18636 * prog consistent for later dump requests, so they can
18637 * later look the same as if they were interpreted only.
18638 */
18639 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18640 if (bpf_pseudo_func(insn)) {
18641 insn[0].imm = env->insn_aux_data[i].call_imm;
18642 insn[1].imm = insn->off;
18643 insn->off = 0;
18644 continue;
18645 }
18646 if (!bpf_pseudo_call(insn))
18647 continue;
18648 insn->off = env->insn_aux_data[i].call_imm;
18649 subprog = find_subprog(env, i + insn->off + 1);
18650 insn->imm = subprog;
18651 }
18652
18653 prog->jited = 1;
18654 prog->bpf_func = func[0]->bpf_func;
18655 prog->jited_len = func[0]->jited_len;
18656 prog->aux->extable = func[0]->aux->extable;
18657 prog->aux->num_exentries = func[0]->aux->num_exentries;
18658 prog->aux->func = func;
18659 prog->aux->func_cnt = env->subprog_cnt;
18660 bpf_prog_jit_attempt_done(prog);
18661 return 0;
18662 out_free:
18663 /* We failed JIT'ing, so at this point we need to unregister poke
18664 * descriptors from subprogs, so that kernel is not attempting to
18665 * patch it anymore as we're freeing the subprog JIT memory.
18666 */
18667 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18668 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18669 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
18670 }
18671 /* At this point we're guaranteed that poke descriptors are not
18672 * live anymore. We can just unlink its descriptor table as it's
18673 * released with the main prog.
18674 */
18675 for (i = 0; i < env->subprog_cnt; i++) {
18676 if (!func[i])
18677 continue;
18678 func[i]->aux->poke_tab = NULL;
18679 bpf_jit_free(func[i]);
18680 }
18681 kfree(func);
18682 out_undo_insn:
18683 /* cleanup main prog to be interpreted */
18684 prog->jit_requested = 0;
18685 prog->blinding_requested = 0;
18686 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18687 if (!bpf_pseudo_call(insn))
18688 continue;
18689 insn->off = 0;
18690 insn->imm = env->insn_aux_data[i].call_imm;
18691 }
18692 bpf_prog_jit_attempt_done(prog);
18693 return err;
18694 }
18695
fixup_call_args(struct bpf_verifier_env * env)18696 static int fixup_call_args(struct bpf_verifier_env *env)
18697 {
18698 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18699 struct bpf_prog *prog = env->prog;
18700 struct bpf_insn *insn = prog->insnsi;
18701 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
18702 int i, depth;
18703 #endif
18704 int err = 0;
18705
18706 if (env->prog->jit_requested &&
18707 !bpf_prog_is_offloaded(env->prog->aux)) {
18708 err = jit_subprogs(env);
18709 if (err == 0)
18710 return 0;
18711 if (err == -EFAULT)
18712 return err;
18713 }
18714 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18715 if (has_kfunc_call) {
18716 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
18717 return -EINVAL;
18718 }
18719 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
18720 /* When JIT fails the progs with bpf2bpf calls and tail_calls
18721 * have to be rejected, since interpreter doesn't support them yet.
18722 */
18723 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
18724 return -EINVAL;
18725 }
18726 for (i = 0; i < prog->len; i++, insn++) {
18727 if (bpf_pseudo_func(insn)) {
18728 /* When JIT fails the progs with callback calls
18729 * have to be rejected, since interpreter doesn't support them yet.
18730 */
18731 verbose(env, "callbacks are not allowed in non-JITed programs\n");
18732 return -EINVAL;
18733 }
18734
18735 if (!bpf_pseudo_call(insn))
18736 continue;
18737 depth = get_callee_stack_depth(env, insn, i);
18738 if (depth < 0)
18739 return depth;
18740 bpf_patch_call_args(insn, depth);
18741 }
18742 err = 0;
18743 #endif
18744 return err;
18745 }
18746
18747 /* replace a generic kfunc with a specialized version if necessary */
specialize_kfunc(struct bpf_verifier_env * env,u32 func_id,u16 offset,unsigned long * addr)18748 static void specialize_kfunc(struct bpf_verifier_env *env,
18749 u32 func_id, u16 offset, unsigned long *addr)
18750 {
18751 struct bpf_prog *prog = env->prog;
18752 bool seen_direct_write;
18753 void *xdp_kfunc;
18754 bool is_rdonly;
18755
18756 if (bpf_dev_bound_kfunc_id(func_id)) {
18757 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
18758 if (xdp_kfunc) {
18759 *addr = (unsigned long)xdp_kfunc;
18760 return;
18761 }
18762 /* fallback to default kfunc when not supported by netdev */
18763 }
18764
18765 if (offset)
18766 return;
18767
18768 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
18769 seen_direct_write = env->seen_direct_write;
18770 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
18771
18772 if (is_rdonly)
18773 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
18774
18775 /* restore env->seen_direct_write to its original value, since
18776 * may_access_direct_pkt_data mutates it
18777 */
18778 env->seen_direct_write = seen_direct_write;
18779 }
18780 }
18781
__fixup_collection_insert_kfunc(struct bpf_insn_aux_data * insn_aux,u16 struct_meta_reg,u16 node_offset_reg,struct bpf_insn * insn,struct bpf_insn * insn_buf,int * cnt)18782 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
18783 u16 struct_meta_reg,
18784 u16 node_offset_reg,
18785 struct bpf_insn *insn,
18786 struct bpf_insn *insn_buf,
18787 int *cnt)
18788 {
18789 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
18790 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
18791
18792 insn_buf[0] = addr[0];
18793 insn_buf[1] = addr[1];
18794 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
18795 insn_buf[3] = *insn;
18796 *cnt = 4;
18797 }
18798
fixup_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_insn * insn_buf,int insn_idx,int * cnt)18799 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
18800 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
18801 {
18802 const struct bpf_kfunc_desc *desc;
18803
18804 if (!insn->imm) {
18805 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
18806 return -EINVAL;
18807 }
18808
18809 *cnt = 0;
18810
18811 /* insn->imm has the btf func_id. Replace it with an offset relative to
18812 * __bpf_call_base, unless the JIT needs to call functions that are
18813 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
18814 */
18815 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
18816 if (!desc) {
18817 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
18818 insn->imm);
18819 return -EFAULT;
18820 }
18821
18822 if (!bpf_jit_supports_far_kfunc_call())
18823 insn->imm = BPF_CALL_IMM(desc->addr);
18824 if (insn->off)
18825 return 0;
18826 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
18827 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18828 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18829 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18830
18831 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18832 insn_buf[1] = addr[0];
18833 insn_buf[2] = addr[1];
18834 insn_buf[3] = *insn;
18835 *cnt = 4;
18836 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18837 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18838 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18839 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18840
18841 if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
18842 !kptr_struct_meta) {
18843 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18844 insn_idx);
18845 return -EFAULT;
18846 }
18847
18848 insn_buf[0] = addr[0];
18849 insn_buf[1] = addr[1];
18850 insn_buf[2] = *insn;
18851 *cnt = 3;
18852 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18853 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18854 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18855 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18856 int struct_meta_reg = BPF_REG_3;
18857 int node_offset_reg = BPF_REG_4;
18858
18859 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18860 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18861 struct_meta_reg = BPF_REG_4;
18862 node_offset_reg = BPF_REG_5;
18863 }
18864
18865 if (!kptr_struct_meta) {
18866 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18867 insn_idx);
18868 return -EFAULT;
18869 }
18870
18871 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18872 node_offset_reg, insn, insn_buf, cnt);
18873 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18874 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18875 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18876 *cnt = 1;
18877 }
18878 return 0;
18879 }
18880
18881 /* Do various post-verification rewrites in a single program pass.
18882 * These rewrites simplify JIT and interpreter implementations.
18883 */
do_misc_fixups(struct bpf_verifier_env * env)18884 static int do_misc_fixups(struct bpf_verifier_env *env)
18885 {
18886 struct bpf_prog *prog = env->prog;
18887 enum bpf_attach_type eatype = prog->expected_attach_type;
18888 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18889 struct bpf_insn *insn = prog->insnsi;
18890 const struct bpf_func_proto *fn;
18891 const int insn_cnt = prog->len;
18892 const struct bpf_map_ops *ops;
18893 struct bpf_insn_aux_data *aux;
18894 struct bpf_insn insn_buf[16];
18895 struct bpf_prog *new_prog;
18896 struct bpf_map *map_ptr;
18897 int i, ret, cnt, delta = 0;
18898
18899 for (i = 0; i < insn_cnt; i++, insn++) {
18900 /* Make divide-by-zero exceptions impossible. */
18901 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18902 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18903 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18904 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18905 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18906 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18907 struct bpf_insn *patchlet;
18908 struct bpf_insn chk_and_div[] = {
18909 /* [R,W]x div 0 -> 0 */
18910 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18911 BPF_JNE | BPF_K, insn->src_reg,
18912 0, 2, 0),
18913 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18914 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18915 *insn,
18916 };
18917 struct bpf_insn chk_and_mod[] = {
18918 /* [R,W]x mod 0 -> [R,W]x */
18919 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18920 BPF_JEQ | BPF_K, insn->src_reg,
18921 0, 1 + (is64 ? 0 : 1), 0),
18922 *insn,
18923 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18924 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18925 };
18926
18927 patchlet = isdiv ? chk_and_div : chk_and_mod;
18928 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18929 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18930
18931 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18932 if (!new_prog)
18933 return -ENOMEM;
18934
18935 delta += cnt - 1;
18936 env->prog = prog = new_prog;
18937 insn = new_prog->insnsi + i + delta;
18938 continue;
18939 }
18940
18941 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18942 if (BPF_CLASS(insn->code) == BPF_LD &&
18943 (BPF_MODE(insn->code) == BPF_ABS ||
18944 BPF_MODE(insn->code) == BPF_IND)) {
18945 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18946 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18947 verbose(env, "bpf verifier is misconfigured\n");
18948 return -EINVAL;
18949 }
18950
18951 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18952 if (!new_prog)
18953 return -ENOMEM;
18954
18955 delta += cnt - 1;
18956 env->prog = prog = new_prog;
18957 insn = new_prog->insnsi + i + delta;
18958 continue;
18959 }
18960
18961 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18962 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18963 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18964 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18965 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18966 struct bpf_insn *patch = &insn_buf[0];
18967 bool issrc, isneg, isimm;
18968 u32 off_reg;
18969
18970 aux = &env->insn_aux_data[i + delta];
18971 if (!aux->alu_state ||
18972 aux->alu_state == BPF_ALU_NON_POINTER)
18973 continue;
18974
18975 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18976 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18977 BPF_ALU_SANITIZE_SRC;
18978 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18979
18980 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18981 if (isimm) {
18982 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18983 } else {
18984 if (isneg)
18985 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18986 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18987 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18988 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18989 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18990 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18991 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18992 }
18993 if (!issrc)
18994 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18995 insn->src_reg = BPF_REG_AX;
18996 if (isneg)
18997 insn->code = insn->code == code_add ?
18998 code_sub : code_add;
18999 *patch++ = *insn;
19000 if (issrc && isneg && !isimm)
19001 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
19002 cnt = patch - insn_buf;
19003
19004 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19005 if (!new_prog)
19006 return -ENOMEM;
19007
19008 delta += cnt - 1;
19009 env->prog = prog = new_prog;
19010 insn = new_prog->insnsi + i + delta;
19011 continue;
19012 }
19013
19014 if (insn->code != (BPF_JMP | BPF_CALL))
19015 continue;
19016 if (insn->src_reg == BPF_PSEUDO_CALL)
19017 continue;
19018 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
19019 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
19020 if (ret)
19021 return ret;
19022 if (cnt == 0)
19023 continue;
19024
19025 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19026 if (!new_prog)
19027 return -ENOMEM;
19028
19029 delta += cnt - 1;
19030 env->prog = prog = new_prog;
19031 insn = new_prog->insnsi + i + delta;
19032 continue;
19033 }
19034
19035 if (insn->imm == BPF_FUNC_get_route_realm)
19036 prog->dst_needed = 1;
19037 if (insn->imm == BPF_FUNC_get_prandom_u32)
19038 bpf_user_rnd_init_once();
19039 if (insn->imm == BPF_FUNC_override_return)
19040 prog->kprobe_override = 1;
19041 if (insn->imm == BPF_FUNC_tail_call) {
19042 /* If we tail call into other programs, we
19043 * cannot make any assumptions since they can
19044 * be replaced dynamically during runtime in
19045 * the program array.
19046 */
19047 prog->cb_access = 1;
19048 if (!allow_tail_call_in_subprogs(env))
19049 prog->aux->stack_depth = MAX_BPF_STACK;
19050 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
19051
19052 /* mark bpf_tail_call as different opcode to avoid
19053 * conditional branch in the interpreter for every normal
19054 * call and to prevent accidental JITing by JIT compiler
19055 * that doesn't support bpf_tail_call yet
19056 */
19057 insn->imm = 0;
19058 insn->code = BPF_JMP | BPF_TAIL_CALL;
19059
19060 aux = &env->insn_aux_data[i + delta];
19061 if (env->bpf_capable && !prog->blinding_requested &&
19062 prog->jit_requested &&
19063 !bpf_map_key_poisoned(aux) &&
19064 !bpf_map_ptr_poisoned(aux) &&
19065 !bpf_map_ptr_unpriv(aux)) {
19066 struct bpf_jit_poke_descriptor desc = {
19067 .reason = BPF_POKE_REASON_TAIL_CALL,
19068 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
19069 .tail_call.key = bpf_map_key_immediate(aux),
19070 .insn_idx = i + delta,
19071 };
19072
19073 ret = bpf_jit_add_poke_descriptor(prog, &desc);
19074 if (ret < 0) {
19075 verbose(env, "adding tail call poke descriptor failed\n");
19076 return ret;
19077 }
19078
19079 insn->imm = ret + 1;
19080 continue;
19081 }
19082
19083 if (!bpf_map_ptr_unpriv(aux))
19084 continue;
19085
19086 /* instead of changing every JIT dealing with tail_call
19087 * emit two extra insns:
19088 * if (index >= max_entries) goto out;
19089 * index &= array->index_mask;
19090 * to avoid out-of-bounds cpu speculation
19091 */
19092 if (bpf_map_ptr_poisoned(aux)) {
19093 verbose(env, "tail_call abusing map_ptr\n");
19094 return -EINVAL;
19095 }
19096
19097 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
19098 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
19099 map_ptr->max_entries, 2);
19100 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
19101 container_of(map_ptr,
19102 struct bpf_array,
19103 map)->index_mask);
19104 insn_buf[2] = *insn;
19105 cnt = 3;
19106 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19107 if (!new_prog)
19108 return -ENOMEM;
19109
19110 delta += cnt - 1;
19111 env->prog = prog = new_prog;
19112 insn = new_prog->insnsi + i + delta;
19113 continue;
19114 }
19115
19116 if (insn->imm == BPF_FUNC_timer_set_callback) {
19117 /* The verifier will process callback_fn as many times as necessary
19118 * with different maps and the register states prepared by
19119 * set_timer_callback_state will be accurate.
19120 *
19121 * The following use case is valid:
19122 * map1 is shared by prog1, prog2, prog3.
19123 * prog1 calls bpf_timer_init for some map1 elements
19124 * prog2 calls bpf_timer_set_callback for some map1 elements.
19125 * Those that were not bpf_timer_init-ed will return -EINVAL.
19126 * prog3 calls bpf_timer_start for some map1 elements.
19127 * Those that were not both bpf_timer_init-ed and
19128 * bpf_timer_set_callback-ed will return -EINVAL.
19129 */
19130 struct bpf_insn ld_addrs[2] = {
19131 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
19132 };
19133
19134 insn_buf[0] = ld_addrs[0];
19135 insn_buf[1] = ld_addrs[1];
19136 insn_buf[2] = *insn;
19137 cnt = 3;
19138
19139 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19140 if (!new_prog)
19141 return -ENOMEM;
19142
19143 delta += cnt - 1;
19144 env->prog = prog = new_prog;
19145 insn = new_prog->insnsi + i + delta;
19146 goto patch_call_imm;
19147 }
19148
19149 if (is_storage_get_function(insn->imm)) {
19150 if (!env->prog->aux->sleepable ||
19151 env->insn_aux_data[i + delta].storage_get_func_atomic)
19152 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
19153 else
19154 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
19155 insn_buf[1] = *insn;
19156 cnt = 2;
19157
19158 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19159 if (!new_prog)
19160 return -ENOMEM;
19161
19162 delta += cnt - 1;
19163 env->prog = prog = new_prog;
19164 insn = new_prog->insnsi + i + delta;
19165 goto patch_call_imm;
19166 }
19167
19168 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
19169 * and other inlining handlers are currently limited to 64 bit
19170 * only.
19171 */
19172 if (prog->jit_requested && BITS_PER_LONG == 64 &&
19173 (insn->imm == BPF_FUNC_map_lookup_elem ||
19174 insn->imm == BPF_FUNC_map_update_elem ||
19175 insn->imm == BPF_FUNC_map_delete_elem ||
19176 insn->imm == BPF_FUNC_map_push_elem ||
19177 insn->imm == BPF_FUNC_map_pop_elem ||
19178 insn->imm == BPF_FUNC_map_peek_elem ||
19179 insn->imm == BPF_FUNC_redirect_map ||
19180 insn->imm == BPF_FUNC_for_each_map_elem ||
19181 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
19182 aux = &env->insn_aux_data[i + delta];
19183 if (bpf_map_ptr_poisoned(aux))
19184 goto patch_call_imm;
19185
19186 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
19187 ops = map_ptr->ops;
19188 if (insn->imm == BPF_FUNC_map_lookup_elem &&
19189 ops->map_gen_lookup) {
19190 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
19191 if (cnt == -EOPNOTSUPP)
19192 goto patch_map_ops_generic;
19193 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
19194 verbose(env, "bpf verifier is misconfigured\n");
19195 return -EINVAL;
19196 }
19197
19198 new_prog = bpf_patch_insn_data(env, i + delta,
19199 insn_buf, cnt);
19200 if (!new_prog)
19201 return -ENOMEM;
19202
19203 delta += cnt - 1;
19204 env->prog = prog = new_prog;
19205 insn = new_prog->insnsi + i + delta;
19206 continue;
19207 }
19208
19209 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
19210 (void *(*)(struct bpf_map *map, void *key))NULL));
19211 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
19212 (long (*)(struct bpf_map *map, void *key))NULL));
19213 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
19214 (long (*)(struct bpf_map *map, void *key, void *value,
19215 u64 flags))NULL));
19216 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
19217 (long (*)(struct bpf_map *map, void *value,
19218 u64 flags))NULL));
19219 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
19220 (long (*)(struct bpf_map *map, void *value))NULL));
19221 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
19222 (long (*)(struct bpf_map *map, void *value))NULL));
19223 BUILD_BUG_ON(!__same_type(ops->map_redirect,
19224 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
19225 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
19226 (long (*)(struct bpf_map *map,
19227 bpf_callback_t callback_fn,
19228 void *callback_ctx,
19229 u64 flags))NULL));
19230 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
19231 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
19232
19233 patch_map_ops_generic:
19234 switch (insn->imm) {
19235 case BPF_FUNC_map_lookup_elem:
19236 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
19237 continue;
19238 case BPF_FUNC_map_update_elem:
19239 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
19240 continue;
19241 case BPF_FUNC_map_delete_elem:
19242 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
19243 continue;
19244 case BPF_FUNC_map_push_elem:
19245 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
19246 continue;
19247 case BPF_FUNC_map_pop_elem:
19248 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
19249 continue;
19250 case BPF_FUNC_map_peek_elem:
19251 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
19252 continue;
19253 case BPF_FUNC_redirect_map:
19254 insn->imm = BPF_CALL_IMM(ops->map_redirect);
19255 continue;
19256 case BPF_FUNC_for_each_map_elem:
19257 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
19258 continue;
19259 case BPF_FUNC_map_lookup_percpu_elem:
19260 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
19261 continue;
19262 }
19263
19264 goto patch_call_imm;
19265 }
19266
19267 /* Implement bpf_jiffies64 inline. */
19268 if (prog->jit_requested && BITS_PER_LONG == 64 &&
19269 insn->imm == BPF_FUNC_jiffies64) {
19270 struct bpf_insn ld_jiffies_addr[2] = {
19271 BPF_LD_IMM64(BPF_REG_0,
19272 (unsigned long)&jiffies),
19273 };
19274
19275 insn_buf[0] = ld_jiffies_addr[0];
19276 insn_buf[1] = ld_jiffies_addr[1];
19277 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
19278 BPF_REG_0, 0);
19279 cnt = 3;
19280
19281 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
19282 cnt);
19283 if (!new_prog)
19284 return -ENOMEM;
19285
19286 delta += cnt - 1;
19287 env->prog = prog = new_prog;
19288 insn = new_prog->insnsi + i + delta;
19289 continue;
19290 }
19291
19292 /* Implement bpf_get_func_arg inline. */
19293 if (prog_type == BPF_PROG_TYPE_TRACING &&
19294 insn->imm == BPF_FUNC_get_func_arg) {
19295 /* Load nr_args from ctx - 8 */
19296 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
19297 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
19298 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
19299 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
19300 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
19301 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
19302 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
19303 insn_buf[7] = BPF_JMP_A(1);
19304 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
19305 cnt = 9;
19306
19307 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19308 if (!new_prog)
19309 return -ENOMEM;
19310
19311 delta += cnt - 1;
19312 env->prog = prog = new_prog;
19313 insn = new_prog->insnsi + i + delta;
19314 continue;
19315 }
19316
19317 /* Implement bpf_get_func_ret inline. */
19318 if (prog_type == BPF_PROG_TYPE_TRACING &&
19319 insn->imm == BPF_FUNC_get_func_ret) {
19320 if (eatype == BPF_TRACE_FEXIT ||
19321 eatype == BPF_MODIFY_RETURN) {
19322 /* Load nr_args from ctx - 8 */
19323 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
19324 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
19325 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
19326 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
19327 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
19328 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
19329 cnt = 6;
19330 } else {
19331 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
19332 cnt = 1;
19333 }
19334
19335 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19336 if (!new_prog)
19337 return -ENOMEM;
19338
19339 delta += cnt - 1;
19340 env->prog = prog = new_prog;
19341 insn = new_prog->insnsi + i + delta;
19342 continue;
19343 }
19344
19345 /* Implement get_func_arg_cnt inline. */
19346 if (prog_type == BPF_PROG_TYPE_TRACING &&
19347 insn->imm == BPF_FUNC_get_func_arg_cnt) {
19348 /* Load nr_args from ctx - 8 */
19349 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
19350
19351 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
19352 if (!new_prog)
19353 return -ENOMEM;
19354
19355 env->prog = prog = new_prog;
19356 insn = new_prog->insnsi + i + delta;
19357 continue;
19358 }
19359
19360 /* Implement bpf_get_func_ip inline. */
19361 if (prog_type == BPF_PROG_TYPE_TRACING &&
19362 insn->imm == BPF_FUNC_get_func_ip) {
19363 /* Load IP address from ctx - 16 */
19364 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
19365
19366 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
19367 if (!new_prog)
19368 return -ENOMEM;
19369
19370 env->prog = prog = new_prog;
19371 insn = new_prog->insnsi + i + delta;
19372 continue;
19373 }
19374
19375 patch_call_imm:
19376 fn = env->ops->get_func_proto(insn->imm, env->prog);
19377 /* all functions that have prototype and verifier allowed
19378 * programs to call them, must be real in-kernel functions
19379 */
19380 if (!fn->func) {
19381 verbose(env,
19382 "kernel subsystem misconfigured func %s#%d\n",
19383 func_id_name(insn->imm), insn->imm);
19384 return -EFAULT;
19385 }
19386 insn->imm = fn->func - __bpf_call_base;
19387 }
19388
19389 /* Since poke tab is now finalized, publish aux to tracker. */
19390 for (i = 0; i < prog->aux->size_poke_tab; i++) {
19391 map_ptr = prog->aux->poke_tab[i].tail_call.map;
19392 if (!map_ptr->ops->map_poke_track ||
19393 !map_ptr->ops->map_poke_untrack ||
19394 !map_ptr->ops->map_poke_run) {
19395 verbose(env, "bpf verifier is misconfigured\n");
19396 return -EINVAL;
19397 }
19398
19399 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
19400 if (ret < 0) {
19401 verbose(env, "tracking tail call prog failed\n");
19402 return ret;
19403 }
19404 }
19405
19406 sort_kfunc_descs_by_imm_off(env->prog);
19407
19408 return 0;
19409 }
19410
inline_bpf_loop(struct bpf_verifier_env * env,int position,s32 stack_base,u32 callback_subprogno,u32 * cnt)19411 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
19412 int position,
19413 s32 stack_base,
19414 u32 callback_subprogno,
19415 u32 *cnt)
19416 {
19417 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
19418 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
19419 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
19420 int reg_loop_max = BPF_REG_6;
19421 int reg_loop_cnt = BPF_REG_7;
19422 int reg_loop_ctx = BPF_REG_8;
19423
19424 struct bpf_prog *new_prog;
19425 u32 callback_start;
19426 u32 call_insn_offset;
19427 s32 callback_offset;
19428
19429 /* This represents an inlined version of bpf_iter.c:bpf_loop,
19430 * be careful to modify this code in sync.
19431 */
19432 struct bpf_insn insn_buf[] = {
19433 /* Return error and jump to the end of the patch if
19434 * expected number of iterations is too big.
19435 */
19436 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
19437 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
19438 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
19439 /* spill R6, R7, R8 to use these as loop vars */
19440 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
19441 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
19442 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
19443 /* initialize loop vars */
19444 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
19445 BPF_MOV32_IMM(reg_loop_cnt, 0),
19446 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
19447 /* loop header,
19448 * if reg_loop_cnt >= reg_loop_max skip the loop body
19449 */
19450 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
19451 /* callback call,
19452 * correct callback offset would be set after patching
19453 */
19454 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
19455 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
19456 BPF_CALL_REL(0),
19457 /* increment loop counter */
19458 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
19459 /* jump to loop header if callback returned 0 */
19460 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
19461 /* return value of bpf_loop,
19462 * set R0 to the number of iterations
19463 */
19464 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
19465 /* restore original values of R6, R7, R8 */
19466 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
19467 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
19468 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
19469 };
19470
19471 *cnt = ARRAY_SIZE(insn_buf);
19472 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
19473 if (!new_prog)
19474 return new_prog;
19475
19476 /* callback start is known only after patching */
19477 callback_start = env->subprog_info[callback_subprogno].start;
19478 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
19479 call_insn_offset = position + 12;
19480 callback_offset = callback_start - call_insn_offset - 1;
19481 new_prog->insnsi[call_insn_offset].imm = callback_offset;
19482
19483 return new_prog;
19484 }
19485
is_bpf_loop_call(struct bpf_insn * insn)19486 static bool is_bpf_loop_call(struct bpf_insn *insn)
19487 {
19488 return insn->code == (BPF_JMP | BPF_CALL) &&
19489 insn->src_reg == 0 &&
19490 insn->imm == BPF_FUNC_loop;
19491 }
19492
19493 /* For all sub-programs in the program (including main) check
19494 * insn_aux_data to see if there are bpf_loop calls that require
19495 * inlining. If such calls are found the calls are replaced with a
19496 * sequence of instructions produced by `inline_bpf_loop` function and
19497 * subprog stack_depth is increased by the size of 3 registers.
19498 * This stack space is used to spill values of the R6, R7, R8. These
19499 * registers are used to store the loop bound, counter and context
19500 * variables.
19501 */
optimize_bpf_loop(struct bpf_verifier_env * env)19502 static int optimize_bpf_loop(struct bpf_verifier_env *env)
19503 {
19504 struct bpf_subprog_info *subprogs = env->subprog_info;
19505 int i, cur_subprog = 0, cnt, delta = 0;
19506 struct bpf_insn *insn = env->prog->insnsi;
19507 int insn_cnt = env->prog->len;
19508 u16 stack_depth = subprogs[cur_subprog].stack_depth;
19509 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
19510 u16 stack_depth_extra = 0;
19511
19512 for (i = 0; i < insn_cnt; i++, insn++) {
19513 struct bpf_loop_inline_state *inline_state =
19514 &env->insn_aux_data[i + delta].loop_inline_state;
19515
19516 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
19517 struct bpf_prog *new_prog;
19518
19519 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
19520 new_prog = inline_bpf_loop(env,
19521 i + delta,
19522 -(stack_depth + stack_depth_extra),
19523 inline_state->callback_subprogno,
19524 &cnt);
19525 if (!new_prog)
19526 return -ENOMEM;
19527
19528 delta += cnt - 1;
19529 env->prog = new_prog;
19530 insn = new_prog->insnsi + i + delta;
19531 }
19532
19533 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
19534 subprogs[cur_subprog].stack_depth += stack_depth_extra;
19535 cur_subprog++;
19536 stack_depth = subprogs[cur_subprog].stack_depth;
19537 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
19538 stack_depth_extra = 0;
19539 }
19540 }
19541
19542 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19543
19544 return 0;
19545 }
19546
free_states(struct bpf_verifier_env * env)19547 static void free_states(struct bpf_verifier_env *env)
19548 {
19549 struct bpf_verifier_state_list *sl, *sln;
19550 int i;
19551
19552 sl = env->free_list;
19553 while (sl) {
19554 sln = sl->next;
19555 free_verifier_state(&sl->state, false);
19556 kfree(sl);
19557 sl = sln;
19558 }
19559 env->free_list = NULL;
19560
19561 if (!env->explored_states)
19562 return;
19563
19564 for (i = 0; i < state_htab_size(env); i++) {
19565 sl = env->explored_states[i];
19566
19567 while (sl) {
19568 sln = sl->next;
19569 free_verifier_state(&sl->state, false);
19570 kfree(sl);
19571 sl = sln;
19572 }
19573 env->explored_states[i] = NULL;
19574 }
19575 }
19576
do_check_common(struct bpf_verifier_env * env,int subprog)19577 static int do_check_common(struct bpf_verifier_env *env, int subprog)
19578 {
19579 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
19580 struct bpf_verifier_state *state;
19581 struct bpf_reg_state *regs;
19582 int ret, i;
19583
19584 env->prev_linfo = NULL;
19585 env->pass_cnt++;
19586
19587 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
19588 if (!state)
19589 return -ENOMEM;
19590 state->curframe = 0;
19591 state->speculative = false;
19592 state->branches = 1;
19593 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
19594 if (!state->frame[0]) {
19595 kfree(state);
19596 return -ENOMEM;
19597 }
19598 env->cur_state = state;
19599 init_func_state(env, state->frame[0],
19600 BPF_MAIN_FUNC /* callsite */,
19601 0 /* frameno */,
19602 subprog);
19603 state->first_insn_idx = env->subprog_info[subprog].start;
19604 state->last_insn_idx = -1;
19605
19606 regs = state->frame[state->curframe]->regs;
19607 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
19608 ret = btf_prepare_func_args(env, subprog, regs);
19609 if (ret)
19610 goto out;
19611 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
19612 if (regs[i].type == PTR_TO_CTX)
19613 mark_reg_known_zero(env, regs, i);
19614 else if (regs[i].type == SCALAR_VALUE)
19615 mark_reg_unknown(env, regs, i);
19616 else if (base_type(regs[i].type) == PTR_TO_MEM) {
19617 const u32 mem_size = regs[i].mem_size;
19618
19619 mark_reg_known_zero(env, regs, i);
19620 regs[i].mem_size = mem_size;
19621 regs[i].id = ++env->id_gen;
19622 }
19623 }
19624 } else {
19625 /* 1st arg to a function */
19626 regs[BPF_REG_1].type = PTR_TO_CTX;
19627 mark_reg_known_zero(env, regs, BPF_REG_1);
19628 ret = btf_check_subprog_arg_match(env, subprog, regs);
19629 if (ret == -EFAULT)
19630 /* unlikely verifier bug. abort.
19631 * ret == 0 and ret < 0 are sadly acceptable for
19632 * main() function due to backward compatibility.
19633 * Like socket filter program may be written as:
19634 * int bpf_prog(struct pt_regs *ctx)
19635 * and never dereference that ctx in the program.
19636 * 'struct pt_regs' is a type mismatch for socket
19637 * filter that should be using 'struct __sk_buff'.
19638 */
19639 goto out;
19640 }
19641
19642 ret = do_check(env);
19643 out:
19644 /* check for NULL is necessary, since cur_state can be freed inside
19645 * do_check() under memory pressure.
19646 */
19647 if (env->cur_state) {
19648 free_verifier_state(env->cur_state, true);
19649 env->cur_state = NULL;
19650 }
19651 while (!pop_stack(env, NULL, NULL, false));
19652 if (!ret && pop_log)
19653 bpf_vlog_reset(&env->log, 0);
19654 free_states(env);
19655 return ret;
19656 }
19657
19658 /* Verify all global functions in a BPF program one by one based on their BTF.
19659 * All global functions must pass verification. Otherwise the whole program is rejected.
19660 * Consider:
19661 * int bar(int);
19662 * int foo(int f)
19663 * {
19664 * return bar(f);
19665 * }
19666 * int bar(int b)
19667 * {
19668 * ...
19669 * }
19670 * foo() will be verified first for R1=any_scalar_value. During verification it
19671 * will be assumed that bar() already verified successfully and call to bar()
19672 * from foo() will be checked for type match only. Later bar() will be verified
19673 * independently to check that it's safe for R1=any_scalar_value.
19674 */
do_check_subprogs(struct bpf_verifier_env * env)19675 static int do_check_subprogs(struct bpf_verifier_env *env)
19676 {
19677 struct bpf_prog_aux *aux = env->prog->aux;
19678 int i, ret;
19679
19680 if (!aux->func_info)
19681 return 0;
19682
19683 for (i = 1; i < env->subprog_cnt; i++) {
19684 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
19685 continue;
19686 env->insn_idx = env->subprog_info[i].start;
19687 WARN_ON_ONCE(env->insn_idx == 0);
19688 ret = do_check_common(env, i);
19689 if (ret) {
19690 return ret;
19691 } else if (env->log.level & BPF_LOG_LEVEL) {
19692 verbose(env,
19693 "Func#%d is safe for any args that match its prototype\n",
19694 i);
19695 }
19696 }
19697 return 0;
19698 }
19699
do_check_main(struct bpf_verifier_env * env)19700 static int do_check_main(struct bpf_verifier_env *env)
19701 {
19702 int ret;
19703
19704 env->insn_idx = 0;
19705 ret = do_check_common(env, 0);
19706 if (!ret)
19707 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19708 return ret;
19709 }
19710
19711
print_verification_stats(struct bpf_verifier_env * env)19712 static void print_verification_stats(struct bpf_verifier_env *env)
19713 {
19714 int i;
19715
19716 if (env->log.level & BPF_LOG_STATS) {
19717 verbose(env, "verification time %lld usec\n",
19718 div_u64(env->verification_time, 1000));
19719 verbose(env, "stack depth ");
19720 for (i = 0; i < env->subprog_cnt; i++) {
19721 u32 depth = env->subprog_info[i].stack_depth;
19722
19723 verbose(env, "%d", depth);
19724 if (i + 1 < env->subprog_cnt)
19725 verbose(env, "+");
19726 }
19727 verbose(env, "\n");
19728 }
19729 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
19730 "total_states %d peak_states %d mark_read %d\n",
19731 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
19732 env->max_states_per_insn, env->total_states,
19733 env->peak_states, env->longest_mark_read_walk);
19734 }
19735
check_struct_ops_btf_id(struct bpf_verifier_env * env)19736 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
19737 {
19738 const struct btf_type *t, *func_proto;
19739 const struct bpf_struct_ops *st_ops;
19740 const struct btf_member *member;
19741 struct bpf_prog *prog = env->prog;
19742 u32 btf_id, member_idx;
19743 const char *mname;
19744
19745 if (!prog->gpl_compatible) {
19746 verbose(env, "struct ops programs must have a GPL compatible license\n");
19747 return -EINVAL;
19748 }
19749
19750 btf_id = prog->aux->attach_btf_id;
19751 st_ops = bpf_struct_ops_find(btf_id);
19752 if (!st_ops) {
19753 verbose(env, "attach_btf_id %u is not a supported struct\n",
19754 btf_id);
19755 return -ENOTSUPP;
19756 }
19757
19758 t = st_ops->type;
19759 member_idx = prog->expected_attach_type;
19760 if (member_idx >= btf_type_vlen(t)) {
19761 verbose(env, "attach to invalid member idx %u of struct %s\n",
19762 member_idx, st_ops->name);
19763 return -EINVAL;
19764 }
19765
19766 member = &btf_type_member(t)[member_idx];
19767 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
19768 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
19769 NULL);
19770 if (!func_proto) {
19771 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
19772 mname, member_idx, st_ops->name);
19773 return -EINVAL;
19774 }
19775
19776 if (st_ops->check_member) {
19777 int err = st_ops->check_member(t, member, prog);
19778
19779 if (err) {
19780 verbose(env, "attach to unsupported member %s of struct %s\n",
19781 mname, st_ops->name);
19782 return err;
19783 }
19784 }
19785
19786 prog->aux->attach_func_proto = func_proto;
19787 prog->aux->attach_func_name = mname;
19788 env->ops = st_ops->verifier_ops;
19789
19790 return 0;
19791 }
19792 #define SECURITY_PREFIX "security_"
19793
check_attach_modify_return(unsigned long addr,const char * func_name)19794 static int check_attach_modify_return(unsigned long addr, const char *func_name)
19795 {
19796 if (within_error_injection_list(addr) ||
19797 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
19798 return 0;
19799
19800 return -EINVAL;
19801 }
19802
19803 /* list of non-sleepable functions that are otherwise on
19804 * ALLOW_ERROR_INJECTION list
19805 */
19806 BTF_SET_START(btf_non_sleepable_error_inject)
19807 /* Three functions below can be called from sleepable and non-sleepable context.
19808 * Assume non-sleepable from bpf safety point of view.
19809 */
BTF_ID(func,__filemap_add_folio)19810 BTF_ID(func, __filemap_add_folio)
19811 BTF_ID(func, should_fail_alloc_page)
19812 BTF_ID(func, should_failslab)
19813 BTF_SET_END(btf_non_sleepable_error_inject)
19814
19815 static int check_non_sleepable_error_inject(u32 btf_id)
19816 {
19817 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
19818 }
19819
bpf_check_attach_target(struct bpf_verifier_log * log,const struct bpf_prog * prog,const struct bpf_prog * tgt_prog,u32 btf_id,struct bpf_attach_target_info * tgt_info)19820 int bpf_check_attach_target(struct bpf_verifier_log *log,
19821 const struct bpf_prog *prog,
19822 const struct bpf_prog *tgt_prog,
19823 u32 btf_id,
19824 struct bpf_attach_target_info *tgt_info)
19825 {
19826 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
19827 const char prefix[] = "btf_trace_";
19828 int ret = 0, subprog = -1, i;
19829 const struct btf_type *t;
19830 bool conservative = true;
19831 const char *tname;
19832 struct btf *btf;
19833 long addr = 0;
19834 struct module *mod = NULL;
19835
19836 if (!btf_id) {
19837 bpf_log(log, "Tracing programs must provide btf_id\n");
19838 return -EINVAL;
19839 }
19840 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19841 if (!btf) {
19842 bpf_log(log,
19843 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19844 return -EINVAL;
19845 }
19846 t = btf_type_by_id(btf, btf_id);
19847 if (!t) {
19848 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19849 return -EINVAL;
19850 }
19851 tname = btf_name_by_offset(btf, t->name_off);
19852 if (!tname) {
19853 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19854 return -EINVAL;
19855 }
19856 if (tgt_prog) {
19857 struct bpf_prog_aux *aux = tgt_prog->aux;
19858
19859 if (bpf_prog_is_dev_bound(prog->aux) &&
19860 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19861 bpf_log(log, "Target program bound device mismatch");
19862 return -EINVAL;
19863 }
19864
19865 for (i = 0; i < aux->func_info_cnt; i++)
19866 if (aux->func_info[i].type_id == btf_id) {
19867 subprog = i;
19868 break;
19869 }
19870 if (subprog == -1) {
19871 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19872 return -EINVAL;
19873 }
19874 conservative = aux->func_info_aux[subprog].unreliable;
19875 if (prog_extension) {
19876 if (conservative) {
19877 bpf_log(log,
19878 "Cannot replace static functions\n");
19879 return -EINVAL;
19880 }
19881 if (!prog->jit_requested) {
19882 bpf_log(log,
19883 "Extension programs should be JITed\n");
19884 return -EINVAL;
19885 }
19886 }
19887 if (!tgt_prog->jited) {
19888 bpf_log(log, "Can attach to only JITed progs\n");
19889 return -EINVAL;
19890 }
19891 if (tgt_prog->type == prog->type) {
19892 /* Cannot fentry/fexit another fentry/fexit program.
19893 * Cannot attach program extension to another extension.
19894 * It's ok to attach fentry/fexit to extension program.
19895 */
19896 bpf_log(log, "Cannot recursively attach\n");
19897 return -EINVAL;
19898 }
19899 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19900 prog_extension &&
19901 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19902 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19903 /* Program extensions can extend all program types
19904 * except fentry/fexit. The reason is the following.
19905 * The fentry/fexit programs are used for performance
19906 * analysis, stats and can be attached to any program
19907 * type except themselves. When extension program is
19908 * replacing XDP function it is necessary to allow
19909 * performance analysis of all functions. Both original
19910 * XDP program and its program extension. Hence
19911 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19912 * allowed. If extending of fentry/fexit was allowed it
19913 * would be possible to create long call chain
19914 * fentry->extension->fentry->extension beyond
19915 * reasonable stack size. Hence extending fentry is not
19916 * allowed.
19917 */
19918 bpf_log(log, "Cannot extend fentry/fexit\n");
19919 return -EINVAL;
19920 }
19921 } else {
19922 if (prog_extension) {
19923 bpf_log(log, "Cannot replace kernel functions\n");
19924 return -EINVAL;
19925 }
19926 }
19927
19928 switch (prog->expected_attach_type) {
19929 case BPF_TRACE_RAW_TP:
19930 if (tgt_prog) {
19931 bpf_log(log,
19932 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19933 return -EINVAL;
19934 }
19935 if (!btf_type_is_typedef(t)) {
19936 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19937 btf_id);
19938 return -EINVAL;
19939 }
19940 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19941 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19942 btf_id, tname);
19943 return -EINVAL;
19944 }
19945 tname += sizeof(prefix) - 1;
19946 t = btf_type_by_id(btf, t->type);
19947 if (!btf_type_is_ptr(t))
19948 /* should never happen in valid vmlinux build */
19949 return -EINVAL;
19950 t = btf_type_by_id(btf, t->type);
19951 if (!btf_type_is_func_proto(t))
19952 /* should never happen in valid vmlinux build */
19953 return -EINVAL;
19954
19955 break;
19956 case BPF_TRACE_ITER:
19957 if (!btf_type_is_func(t)) {
19958 bpf_log(log, "attach_btf_id %u is not a function\n",
19959 btf_id);
19960 return -EINVAL;
19961 }
19962 t = btf_type_by_id(btf, t->type);
19963 if (!btf_type_is_func_proto(t))
19964 return -EINVAL;
19965 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19966 if (ret)
19967 return ret;
19968 break;
19969 default:
19970 if (!prog_extension)
19971 return -EINVAL;
19972 fallthrough;
19973 case BPF_MODIFY_RETURN:
19974 case BPF_LSM_MAC:
19975 case BPF_LSM_CGROUP:
19976 case BPF_TRACE_FENTRY:
19977 case BPF_TRACE_FEXIT:
19978 if (!btf_type_is_func(t)) {
19979 bpf_log(log, "attach_btf_id %u is not a function\n",
19980 btf_id);
19981 return -EINVAL;
19982 }
19983 if (prog_extension &&
19984 btf_check_type_match(log, prog, btf, t))
19985 return -EINVAL;
19986 t = btf_type_by_id(btf, t->type);
19987 if (!btf_type_is_func_proto(t))
19988 return -EINVAL;
19989
19990 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19991 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19992 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19993 return -EINVAL;
19994
19995 if (tgt_prog && conservative)
19996 t = NULL;
19997
19998 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19999 if (ret < 0)
20000 return ret;
20001
20002 if (tgt_prog) {
20003 if (subprog == 0)
20004 addr = (long) tgt_prog->bpf_func;
20005 else
20006 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
20007 } else {
20008 if (btf_is_module(btf)) {
20009 mod = btf_try_get_module(btf);
20010 if (mod)
20011 addr = find_kallsyms_symbol_value(mod, tname);
20012 else
20013 addr = 0;
20014 } else {
20015 addr = kallsyms_lookup_name(tname);
20016 }
20017 if (!addr) {
20018 module_put(mod);
20019 bpf_log(log,
20020 "The address of function %s cannot be found\n",
20021 tname);
20022 return -ENOENT;
20023 }
20024 }
20025
20026 if (prog->aux->sleepable) {
20027 ret = -EINVAL;
20028 switch (prog->type) {
20029 case BPF_PROG_TYPE_TRACING:
20030
20031 /* fentry/fexit/fmod_ret progs can be sleepable if they are
20032 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
20033 */
20034 if (!check_non_sleepable_error_inject(btf_id) &&
20035 within_error_injection_list(addr))
20036 ret = 0;
20037 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
20038 * in the fmodret id set with the KF_SLEEPABLE flag.
20039 */
20040 else {
20041 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
20042 prog);
20043
20044 if (flags && (*flags & KF_SLEEPABLE))
20045 ret = 0;
20046 }
20047 break;
20048 case BPF_PROG_TYPE_LSM:
20049 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
20050 * Only some of them are sleepable.
20051 */
20052 if (bpf_lsm_is_sleepable_hook(btf_id))
20053 ret = 0;
20054 break;
20055 default:
20056 break;
20057 }
20058 if (ret) {
20059 module_put(mod);
20060 bpf_log(log, "%s is not sleepable\n", tname);
20061 return ret;
20062 }
20063 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
20064 if (tgt_prog) {
20065 module_put(mod);
20066 bpf_log(log, "can't modify return codes of BPF programs\n");
20067 return -EINVAL;
20068 }
20069 ret = -EINVAL;
20070 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
20071 !check_attach_modify_return(addr, tname))
20072 ret = 0;
20073 if (ret) {
20074 module_put(mod);
20075 bpf_log(log, "%s() is not modifiable\n", tname);
20076 return ret;
20077 }
20078 }
20079
20080 break;
20081 }
20082 tgt_info->tgt_addr = addr;
20083 tgt_info->tgt_name = tname;
20084 tgt_info->tgt_type = t;
20085 tgt_info->tgt_mod = mod;
20086 return 0;
20087 }
20088
BTF_SET_START(btf_id_deny)20089 BTF_SET_START(btf_id_deny)
20090 BTF_ID_UNUSED
20091 #ifdef CONFIG_SMP
20092 BTF_ID(func, migrate_disable)
20093 BTF_ID(func, migrate_enable)
20094 #endif
20095 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
20096 BTF_ID(func, rcu_read_unlock_strict)
20097 #endif
20098 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
20099 BTF_ID(func, preempt_count_add)
20100 BTF_ID(func, preempt_count_sub)
20101 #endif
20102 #ifdef CONFIG_PREEMPT_RCU
20103 BTF_ID(func, __rcu_read_lock)
20104 BTF_ID(func, __rcu_read_unlock)
20105 #endif
20106 BTF_SET_END(btf_id_deny)
20107
20108 static bool can_be_sleepable(struct bpf_prog *prog)
20109 {
20110 if (prog->type == BPF_PROG_TYPE_TRACING) {
20111 switch (prog->expected_attach_type) {
20112 case BPF_TRACE_FENTRY:
20113 case BPF_TRACE_FEXIT:
20114 case BPF_MODIFY_RETURN:
20115 case BPF_TRACE_ITER:
20116 return true;
20117 default:
20118 return false;
20119 }
20120 }
20121 return prog->type == BPF_PROG_TYPE_LSM ||
20122 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
20123 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
20124 }
20125
check_attach_btf_id(struct bpf_verifier_env * env)20126 static int check_attach_btf_id(struct bpf_verifier_env *env)
20127 {
20128 struct bpf_prog *prog = env->prog;
20129 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
20130 struct bpf_attach_target_info tgt_info = {};
20131 u32 btf_id = prog->aux->attach_btf_id;
20132 struct bpf_trampoline *tr;
20133 int ret;
20134 u64 key;
20135
20136 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
20137 if (prog->aux->sleepable)
20138 /* attach_btf_id checked to be zero already */
20139 return 0;
20140 verbose(env, "Syscall programs can only be sleepable\n");
20141 return -EINVAL;
20142 }
20143
20144 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
20145 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
20146 return -EINVAL;
20147 }
20148
20149 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
20150 return check_struct_ops_btf_id(env);
20151
20152 if (prog->type != BPF_PROG_TYPE_TRACING &&
20153 prog->type != BPF_PROG_TYPE_LSM &&
20154 prog->type != BPF_PROG_TYPE_EXT)
20155 return 0;
20156
20157 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
20158 if (ret)
20159 return ret;
20160
20161 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
20162 /* to make freplace equivalent to their targets, they need to
20163 * inherit env->ops and expected_attach_type for the rest of the
20164 * verification
20165 */
20166 env->ops = bpf_verifier_ops[tgt_prog->type];
20167 prog->expected_attach_type = tgt_prog->expected_attach_type;
20168 }
20169
20170 /* store info about the attachment target that will be used later */
20171 prog->aux->attach_func_proto = tgt_info.tgt_type;
20172 prog->aux->attach_func_name = tgt_info.tgt_name;
20173 prog->aux->mod = tgt_info.tgt_mod;
20174
20175 if (tgt_prog) {
20176 prog->aux->saved_dst_prog_type = tgt_prog->type;
20177 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
20178 }
20179
20180 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
20181 prog->aux->attach_btf_trace = true;
20182 return 0;
20183 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
20184 if (!bpf_iter_prog_supported(prog))
20185 return -EINVAL;
20186 return 0;
20187 }
20188
20189 if (prog->type == BPF_PROG_TYPE_LSM) {
20190 ret = bpf_lsm_verify_prog(&env->log, prog);
20191 if (ret < 0)
20192 return ret;
20193 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
20194 btf_id_set_contains(&btf_id_deny, btf_id)) {
20195 return -EINVAL;
20196 }
20197
20198 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
20199 tr = bpf_trampoline_get(key, &tgt_info);
20200 if (!tr)
20201 return -ENOMEM;
20202
20203 if (tgt_prog && tgt_prog->aux->tail_call_reachable)
20204 tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
20205
20206 prog->aux->dst_trampoline = tr;
20207 return 0;
20208 }
20209
bpf_get_btf_vmlinux(void)20210 struct btf *bpf_get_btf_vmlinux(void)
20211 {
20212 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
20213 mutex_lock(&bpf_verifier_lock);
20214 if (!btf_vmlinux)
20215 btf_vmlinux = btf_parse_vmlinux();
20216 mutex_unlock(&bpf_verifier_lock);
20217 }
20218 return btf_vmlinux;
20219 }
20220
bpf_check(struct bpf_prog ** prog,union bpf_attr * attr,bpfptr_t uattr,__u32 uattr_size)20221 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
20222 {
20223 u64 start_time = ktime_get_ns();
20224 struct bpf_verifier_env *env;
20225 int i, len, ret = -EINVAL, err;
20226 u32 log_true_size;
20227 bool is_priv;
20228
20229 /* no program is valid */
20230 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
20231 return -EINVAL;
20232
20233 /* 'struct bpf_verifier_env' can be global, but since it's not small,
20234 * allocate/free it every time bpf_check() is called
20235 */
20236 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
20237 if (!env)
20238 return -ENOMEM;
20239
20240 env->bt.env = env;
20241
20242 len = (*prog)->len;
20243 env->insn_aux_data =
20244 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
20245 ret = -ENOMEM;
20246 if (!env->insn_aux_data)
20247 goto err_free_env;
20248 for (i = 0; i < len; i++)
20249 env->insn_aux_data[i].orig_idx = i;
20250 env->prog = *prog;
20251 env->ops = bpf_verifier_ops[env->prog->type];
20252 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
20253 is_priv = bpf_capable();
20254
20255 bpf_get_btf_vmlinux();
20256
20257 /* grab the mutex to protect few globals used by verifier */
20258 if (!is_priv)
20259 mutex_lock(&bpf_verifier_lock);
20260
20261 /* user could have requested verbose verifier output
20262 * and supplied buffer to store the verification trace
20263 */
20264 ret = bpf_vlog_init(&env->log, attr->log_level,
20265 (char __user *) (unsigned long) attr->log_buf,
20266 attr->log_size);
20267 if (ret)
20268 goto err_unlock;
20269
20270 mark_verifier_state_clean(env);
20271
20272 if (IS_ERR(btf_vmlinux)) {
20273 /* Either gcc or pahole or kernel are broken. */
20274 verbose(env, "in-kernel BTF is malformed\n");
20275 ret = PTR_ERR(btf_vmlinux);
20276 goto skip_full_check;
20277 }
20278
20279 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
20280 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
20281 env->strict_alignment = true;
20282 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
20283 env->strict_alignment = false;
20284
20285 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
20286 env->allow_uninit_stack = bpf_allow_uninit_stack();
20287 env->bypass_spec_v1 = bpf_bypass_spec_v1();
20288 env->bypass_spec_v4 = bpf_bypass_spec_v4();
20289 env->bpf_capable = bpf_capable();
20290
20291 if (is_priv)
20292 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
20293
20294 env->explored_states = kvcalloc(state_htab_size(env),
20295 sizeof(struct bpf_verifier_state_list *),
20296 GFP_USER);
20297 ret = -ENOMEM;
20298 if (!env->explored_states)
20299 goto skip_full_check;
20300
20301 ret = add_subprog_and_kfunc(env);
20302 if (ret < 0)
20303 goto skip_full_check;
20304
20305 ret = check_subprogs(env);
20306 if (ret < 0)
20307 goto skip_full_check;
20308
20309 ret = check_btf_info(env, attr, uattr);
20310 if (ret < 0)
20311 goto skip_full_check;
20312
20313 ret = check_attach_btf_id(env);
20314 if (ret)
20315 goto skip_full_check;
20316
20317 ret = resolve_pseudo_ldimm64(env);
20318 if (ret < 0)
20319 goto skip_full_check;
20320
20321 if (bpf_prog_is_offloaded(env->prog->aux)) {
20322 ret = bpf_prog_offload_verifier_prep(env->prog);
20323 if (ret)
20324 goto skip_full_check;
20325 }
20326
20327 ret = check_cfg(env);
20328 if (ret < 0)
20329 goto skip_full_check;
20330
20331 ret = do_check_subprogs(env);
20332 ret = ret ?: do_check_main(env);
20333
20334 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
20335 ret = bpf_prog_offload_finalize(env);
20336
20337 skip_full_check:
20338 kvfree(env->explored_states);
20339
20340 if (ret == 0)
20341 ret = check_max_stack_depth(env);
20342
20343 /* instruction rewrites happen after this point */
20344 if (ret == 0)
20345 ret = optimize_bpf_loop(env);
20346
20347 if (is_priv) {
20348 if (ret == 0)
20349 opt_hard_wire_dead_code_branches(env);
20350 if (ret == 0)
20351 ret = opt_remove_dead_code(env);
20352 if (ret == 0)
20353 ret = opt_remove_nops(env);
20354 } else {
20355 if (ret == 0)
20356 sanitize_dead_code(env);
20357 }
20358
20359 if (ret == 0)
20360 /* program is valid, convert *(u32*)(ctx + off) accesses */
20361 ret = convert_ctx_accesses(env);
20362
20363 if (ret == 0)
20364 ret = do_misc_fixups(env);
20365
20366 /* do 32-bit optimization after insn patching has done so those patched
20367 * insns could be handled correctly.
20368 */
20369 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
20370 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
20371 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
20372 : false;
20373 }
20374
20375 if (ret == 0)
20376 ret = fixup_call_args(env);
20377
20378 env->verification_time = ktime_get_ns() - start_time;
20379 print_verification_stats(env);
20380 env->prog->aux->verified_insns = env->insn_processed;
20381
20382 /* preserve original error even if log finalization is successful */
20383 err = bpf_vlog_finalize(&env->log, &log_true_size);
20384 if (err)
20385 ret = err;
20386
20387 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
20388 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
20389 &log_true_size, sizeof(log_true_size))) {
20390 ret = -EFAULT;
20391 goto err_release_maps;
20392 }
20393
20394 if (ret)
20395 goto err_release_maps;
20396
20397 if (env->used_map_cnt) {
20398 /* if program passed verifier, update used_maps in bpf_prog_info */
20399 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
20400 sizeof(env->used_maps[0]),
20401 GFP_KERNEL);
20402
20403 if (!env->prog->aux->used_maps) {
20404 ret = -ENOMEM;
20405 goto err_release_maps;
20406 }
20407
20408 memcpy(env->prog->aux->used_maps, env->used_maps,
20409 sizeof(env->used_maps[0]) * env->used_map_cnt);
20410 env->prog->aux->used_map_cnt = env->used_map_cnt;
20411 }
20412 if (env->used_btf_cnt) {
20413 /* if program passed verifier, update used_btfs in bpf_prog_aux */
20414 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
20415 sizeof(env->used_btfs[0]),
20416 GFP_KERNEL);
20417 if (!env->prog->aux->used_btfs) {
20418 ret = -ENOMEM;
20419 goto err_release_maps;
20420 }
20421
20422 memcpy(env->prog->aux->used_btfs, env->used_btfs,
20423 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
20424 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
20425 }
20426 if (env->used_map_cnt || env->used_btf_cnt) {
20427 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
20428 * bpf_ld_imm64 instructions
20429 */
20430 convert_pseudo_ld_imm64(env);
20431 }
20432
20433 adjust_btf_func(env);
20434
20435 err_release_maps:
20436 if (!env->prog->aux->used_maps)
20437 /* if we didn't copy map pointers into bpf_prog_info, release
20438 * them now. Otherwise free_used_maps() will release them.
20439 */
20440 release_maps(env);
20441 if (!env->prog->aux->used_btfs)
20442 release_btfs(env);
20443
20444 /* extension progs temporarily inherit the attach_type of their targets
20445 for verification purposes, so set it back to zero before returning
20446 */
20447 if (env->prog->type == BPF_PROG_TYPE_EXT)
20448 env->prog->expected_attach_type = 0;
20449
20450 *prog = env->prog;
20451 err_unlock:
20452 if (!is_priv)
20453 mutex_unlock(&bpf_verifier_lock);
20454 vfree(env->insn_aux_data);
20455 err_free_env:
20456 kfree(env);
20457 return ret;
20458 }
20459