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 if (BPF_CLASS(code) == BPF_LD &&
3070 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
3071 subprog[cur_subprog].has_ld_abs = true;
3072 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
3073 goto next;
3074 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
3075 goto next;
3076 if (code == (BPF_JMP32 | BPF_JA))
3077 off = i + insn[i].imm + 1;
3078 else
3079 off = i + insn[i].off + 1;
3080 if (off < subprog_start || off >= subprog_end) {
3081 verbose(env, "jump out of range from insn %d to %d\n", i, off);
3082 return -EINVAL;
3083 }
3084 next:
3085 if (i == subprog_end - 1) {
3086 /* to avoid fall-through from one subprog into another
3087 * the last insn of the subprog should be either exit
3088 * or unconditional jump back
3089 */
3090 if (code != (BPF_JMP | BPF_EXIT) &&
3091 code != (BPF_JMP32 | BPF_JA) &&
3092 code != (BPF_JMP | BPF_JA)) {
3093 verbose(env, "last insn is not an exit or jmp\n");
3094 return -EINVAL;
3095 }
3096 subprog_start = subprog_end;
3097 cur_subprog++;
3098 if (cur_subprog < env->subprog_cnt)
3099 subprog_end = subprog[cur_subprog + 1].start;
3100 }
3101 }
3102 return 0;
3103 }
3104
3105 /* Parentage chain of this register (or stack slot) should take care of all
3106 * issues like callee-saved registers, stack slot allocation time, etc.
3107 */
mark_reg_read(struct bpf_verifier_env * env,const struct bpf_reg_state * state,struct bpf_reg_state * parent,u8 flag)3108 static int mark_reg_read(struct bpf_verifier_env *env,
3109 const struct bpf_reg_state *state,
3110 struct bpf_reg_state *parent, u8 flag)
3111 {
3112 bool writes = parent == state->parent; /* Observe write marks */
3113 int cnt = 0;
3114
3115 while (parent) {
3116 /* if read wasn't screened by an earlier write ... */
3117 if (writes && state->live & REG_LIVE_WRITTEN)
3118 break;
3119 if (parent->live & REG_LIVE_DONE) {
3120 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
3121 reg_type_str(env, parent->type),
3122 parent->var_off.value, parent->off);
3123 return -EFAULT;
3124 }
3125 /* The first condition is more likely to be true than the
3126 * second, checked it first.
3127 */
3128 if ((parent->live & REG_LIVE_READ) == flag ||
3129 parent->live & REG_LIVE_READ64)
3130 /* The parentage chain never changes and
3131 * this parent was already marked as LIVE_READ.
3132 * There is no need to keep walking the chain again and
3133 * keep re-marking all parents as LIVE_READ.
3134 * This case happens when the same register is read
3135 * multiple times without writes into it in-between.
3136 * Also, if parent has the stronger REG_LIVE_READ64 set,
3137 * then no need to set the weak REG_LIVE_READ32.
3138 */
3139 break;
3140 /* ... then we depend on parent's value */
3141 parent->live |= flag;
3142 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
3143 if (flag == REG_LIVE_READ64)
3144 parent->live &= ~REG_LIVE_READ32;
3145 state = parent;
3146 parent = state->parent;
3147 writes = true;
3148 cnt++;
3149 }
3150
3151 if (env->longest_mark_read_walk < cnt)
3152 env->longest_mark_read_walk = cnt;
3153 return 0;
3154 }
3155
mark_dynptr_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg)3156 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
3157 {
3158 struct bpf_func_state *state = func(env, reg);
3159 int spi, ret;
3160
3161 /* For CONST_PTR_TO_DYNPTR, it must have already been done by
3162 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
3163 * check_kfunc_call.
3164 */
3165 if (reg->type == CONST_PTR_TO_DYNPTR)
3166 return 0;
3167 spi = dynptr_get_spi(env, reg);
3168 if (spi < 0)
3169 return spi;
3170 /* Caller ensures dynptr is valid and initialized, which means spi is in
3171 * bounds and spi is the first dynptr slot. Simply mark stack slot as
3172 * read.
3173 */
3174 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr,
3175 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64);
3176 if (ret)
3177 return ret;
3178 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr,
3179 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64);
3180 }
3181
mark_iter_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int spi,int nr_slots)3182 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
3183 int spi, int nr_slots)
3184 {
3185 struct bpf_func_state *state = func(env, reg);
3186 int err, i;
3187
3188 for (i = 0; i < nr_slots; i++) {
3189 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
3190
3191 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
3192 if (err)
3193 return err;
3194
3195 mark_stack_slot_scratched(env, spi - i);
3196 }
3197
3198 return 0;
3199 }
3200
3201 /* This function is supposed to be used by the following 32-bit optimization
3202 * code only. It returns TRUE if the source or destination register operates
3203 * on 64-bit, otherwise return FALSE.
3204 */
is_reg64(struct bpf_verifier_env * env,struct bpf_insn * insn,u32 regno,struct bpf_reg_state * reg,enum reg_arg_type t)3205 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
3206 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
3207 {
3208 u8 code, class, op;
3209
3210 code = insn->code;
3211 class = BPF_CLASS(code);
3212 op = BPF_OP(code);
3213 if (class == BPF_JMP) {
3214 /* BPF_EXIT for "main" will reach here. Return TRUE
3215 * conservatively.
3216 */
3217 if (op == BPF_EXIT)
3218 return true;
3219 if (op == BPF_CALL) {
3220 /* BPF to BPF call will reach here because of marking
3221 * caller saved clobber with DST_OP_NO_MARK for which we
3222 * don't care the register def because they are anyway
3223 * marked as NOT_INIT already.
3224 */
3225 if (insn->src_reg == BPF_PSEUDO_CALL)
3226 return false;
3227 /* Helper call will reach here because of arg type
3228 * check, conservatively return TRUE.
3229 */
3230 if (t == SRC_OP)
3231 return true;
3232
3233 return false;
3234 }
3235 }
3236
3237 if (class == BPF_ALU64 && op == BPF_END && (insn->imm == 16 || insn->imm == 32))
3238 return false;
3239
3240 if (class == BPF_ALU64 || class == BPF_JMP ||
3241 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3242 return true;
3243
3244 if (class == BPF_ALU || class == BPF_JMP32)
3245 return false;
3246
3247 if (class == BPF_LDX) {
3248 if (t != SRC_OP)
3249 return BPF_SIZE(code) == BPF_DW;
3250 /* LDX source must be ptr. */
3251 return true;
3252 }
3253
3254 if (class == BPF_STX) {
3255 /* BPF_STX (including atomic variants) has multiple source
3256 * operands, one of which is a ptr. Check whether the caller is
3257 * asking about it.
3258 */
3259 if (t == SRC_OP && reg->type != SCALAR_VALUE)
3260 return true;
3261 return BPF_SIZE(code) == BPF_DW;
3262 }
3263
3264 if (class == BPF_LD) {
3265 u8 mode = BPF_MODE(code);
3266
3267 /* LD_IMM64 */
3268 if (mode == BPF_IMM)
3269 return true;
3270
3271 /* Both LD_IND and LD_ABS return 32-bit data. */
3272 if (t != SRC_OP)
3273 return false;
3274
3275 /* Implicit ctx ptr. */
3276 if (regno == BPF_REG_6)
3277 return true;
3278
3279 /* Explicit source could be any width. */
3280 return true;
3281 }
3282
3283 if (class == BPF_ST)
3284 /* The only source register for BPF_ST is a ptr. */
3285 return true;
3286
3287 /* Conservatively return true at default. */
3288 return true;
3289 }
3290
3291 /* Return the regno defined by the insn, or -1. */
insn_def_regno(const struct bpf_insn * insn)3292 static int insn_def_regno(const struct bpf_insn *insn)
3293 {
3294 switch (BPF_CLASS(insn->code)) {
3295 case BPF_JMP:
3296 case BPF_JMP32:
3297 case BPF_ST:
3298 return -1;
3299 case BPF_STX:
3300 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
3301 (insn->imm & BPF_FETCH)) {
3302 if (insn->imm == BPF_CMPXCHG)
3303 return BPF_REG_0;
3304 else
3305 return insn->src_reg;
3306 } else {
3307 return -1;
3308 }
3309 default:
3310 return insn->dst_reg;
3311 }
3312 }
3313
3314 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
insn_has_def32(struct bpf_verifier_env * env,struct bpf_insn * insn)3315 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3316 {
3317 int dst_reg = insn_def_regno(insn);
3318
3319 if (dst_reg == -1)
3320 return false;
3321
3322 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3323 }
3324
mark_insn_zext(struct bpf_verifier_env * env,struct bpf_reg_state * reg)3325 static void mark_insn_zext(struct bpf_verifier_env *env,
3326 struct bpf_reg_state *reg)
3327 {
3328 s32 def_idx = reg->subreg_def;
3329
3330 if (def_idx == DEF_NOT_SUBREG)
3331 return;
3332
3333 env->insn_aux_data[def_idx - 1].zext_dst = true;
3334 /* The dst will be zero extended, so won't be sub-register anymore. */
3335 reg->subreg_def = DEF_NOT_SUBREG;
3336 }
3337
__check_reg_arg(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,enum reg_arg_type t)3338 static int __check_reg_arg(struct bpf_verifier_env *env, struct bpf_reg_state *regs, u32 regno,
3339 enum reg_arg_type t)
3340 {
3341 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3342 struct bpf_reg_state *reg;
3343 bool rw64;
3344
3345 if (regno >= MAX_BPF_REG) {
3346 verbose(env, "R%d is invalid\n", regno);
3347 return -EINVAL;
3348 }
3349
3350 mark_reg_scratched(env, regno);
3351
3352 reg = ®s[regno];
3353 rw64 = is_reg64(env, insn, regno, reg, t);
3354 if (t == SRC_OP) {
3355 /* check whether register used as source operand can be read */
3356 if (reg->type == NOT_INIT) {
3357 verbose(env, "R%d !read_ok\n", regno);
3358 return -EACCES;
3359 }
3360 /* We don't need to worry about FP liveness because it's read-only */
3361 if (regno == BPF_REG_FP)
3362 return 0;
3363
3364 if (rw64)
3365 mark_insn_zext(env, reg);
3366
3367 return mark_reg_read(env, reg, reg->parent,
3368 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3369 } else {
3370 /* check whether register used as dest operand can be written to */
3371 if (regno == BPF_REG_FP) {
3372 verbose(env, "frame pointer is read only\n");
3373 return -EACCES;
3374 }
3375 reg->live |= REG_LIVE_WRITTEN;
3376 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3377 if (t == DST_OP)
3378 mark_reg_unknown(env, regs, regno);
3379 }
3380 return 0;
3381 }
3382
check_reg_arg(struct bpf_verifier_env * env,u32 regno,enum reg_arg_type t)3383 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3384 enum reg_arg_type t)
3385 {
3386 struct bpf_verifier_state *vstate = env->cur_state;
3387 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3388
3389 return __check_reg_arg(env, state->regs, regno, t);
3390 }
3391
mark_jmp_point(struct bpf_verifier_env * env,int idx)3392 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3393 {
3394 env->insn_aux_data[idx].jmp_point = true;
3395 }
3396
is_jmp_point(struct bpf_verifier_env * env,int insn_idx)3397 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3398 {
3399 return env->insn_aux_data[insn_idx].jmp_point;
3400 }
3401
3402 /* 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)3403 static int push_jmp_history(struct bpf_verifier_env *env,
3404 struct bpf_verifier_state *cur)
3405 {
3406 u32 cnt = cur->jmp_history_cnt;
3407 struct bpf_idx_pair *p;
3408 size_t alloc_size;
3409
3410 if (!is_jmp_point(env, env->insn_idx))
3411 return 0;
3412
3413 cnt++;
3414 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3415 p = krealloc(cur->jmp_history, alloc_size, GFP_USER);
3416 if (!p)
3417 return -ENOMEM;
3418 p[cnt - 1].idx = env->insn_idx;
3419 p[cnt - 1].prev_idx = env->prev_insn_idx;
3420 cur->jmp_history = p;
3421 cur->jmp_history_cnt = cnt;
3422 return 0;
3423 }
3424
3425 /* Backtrack one insn at a time. If idx is not at the top of recorded
3426 * history then previous instruction came from straight line execution.
3427 * Return -ENOENT if we exhausted all instructions within given state.
3428 *
3429 * It's legal to have a bit of a looping with the same starting and ending
3430 * insn index within the same state, e.g.: 3->4->5->3, so just because current
3431 * instruction index is the same as state's first_idx doesn't mean we are
3432 * done. If there is still some jump history left, we should keep going. We
3433 * need to take into account that we might have a jump history between given
3434 * state's parent and itself, due to checkpointing. In this case, we'll have
3435 * history entry recording a jump from last instruction of parent state and
3436 * first instruction of given state.
3437 */
get_prev_insn_idx(struct bpf_verifier_state * st,int i,u32 * history)3438 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3439 u32 *history)
3440 {
3441 u32 cnt = *history;
3442
3443 if (i == st->first_insn_idx) {
3444 if (cnt == 0)
3445 return -ENOENT;
3446 if (cnt == 1 && st->jmp_history[0].idx == i)
3447 return -ENOENT;
3448 }
3449
3450 if (cnt && st->jmp_history[cnt - 1].idx == i) {
3451 i = st->jmp_history[cnt - 1].prev_idx;
3452 (*history)--;
3453 } else {
3454 i--;
3455 }
3456 return i;
3457 }
3458
disasm_kfunc_name(void * data,const struct bpf_insn * insn)3459 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
3460 {
3461 const struct btf_type *func;
3462 struct btf *desc_btf;
3463
3464 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
3465 return NULL;
3466
3467 desc_btf = find_kfunc_desc_btf(data, insn->off);
3468 if (IS_ERR(desc_btf))
3469 return "<error>";
3470
3471 func = btf_type_by_id(desc_btf, insn->imm);
3472 return btf_name_by_offset(desc_btf, func->name_off);
3473 }
3474
bt_init(struct backtrack_state * bt,u32 frame)3475 static inline void bt_init(struct backtrack_state *bt, u32 frame)
3476 {
3477 bt->frame = frame;
3478 }
3479
bt_reset(struct backtrack_state * bt)3480 static inline void bt_reset(struct backtrack_state *bt)
3481 {
3482 struct bpf_verifier_env *env = bt->env;
3483
3484 memset(bt, 0, sizeof(*bt));
3485 bt->env = env;
3486 }
3487
bt_empty(struct backtrack_state * bt)3488 static inline u32 bt_empty(struct backtrack_state *bt)
3489 {
3490 u64 mask = 0;
3491 int i;
3492
3493 for (i = 0; i <= bt->frame; i++)
3494 mask |= bt->reg_masks[i] | bt->stack_masks[i];
3495
3496 return mask == 0;
3497 }
3498
bt_subprog_enter(struct backtrack_state * bt)3499 static inline int bt_subprog_enter(struct backtrack_state *bt)
3500 {
3501 if (bt->frame == MAX_CALL_FRAMES - 1) {
3502 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame);
3503 WARN_ONCE(1, "verifier backtracking bug");
3504 return -EFAULT;
3505 }
3506 bt->frame++;
3507 return 0;
3508 }
3509
bt_subprog_exit(struct backtrack_state * bt)3510 static inline int bt_subprog_exit(struct backtrack_state *bt)
3511 {
3512 if (bt->frame == 0) {
3513 verbose(bt->env, "BUG subprog exit from frame 0\n");
3514 WARN_ONCE(1, "verifier backtracking bug");
3515 return -EFAULT;
3516 }
3517 bt->frame--;
3518 return 0;
3519 }
3520
bt_set_frame_reg(struct backtrack_state * bt,u32 frame,u32 reg)3521 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3522 {
3523 bt->reg_masks[frame] |= 1 << reg;
3524 }
3525
bt_clear_frame_reg(struct backtrack_state * bt,u32 frame,u32 reg)3526 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
3527 {
3528 bt->reg_masks[frame] &= ~(1 << reg);
3529 }
3530
bt_set_reg(struct backtrack_state * bt,u32 reg)3531 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
3532 {
3533 bt_set_frame_reg(bt, bt->frame, reg);
3534 }
3535
bt_clear_reg(struct backtrack_state * bt,u32 reg)3536 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
3537 {
3538 bt_clear_frame_reg(bt, bt->frame, reg);
3539 }
3540
bt_set_frame_slot(struct backtrack_state * bt,u32 frame,u32 slot)3541 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3542 {
3543 bt->stack_masks[frame] |= 1ull << slot;
3544 }
3545
bt_clear_frame_slot(struct backtrack_state * bt,u32 frame,u32 slot)3546 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
3547 {
3548 bt->stack_masks[frame] &= ~(1ull << slot);
3549 }
3550
bt_set_slot(struct backtrack_state * bt,u32 slot)3551 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot)
3552 {
3553 bt_set_frame_slot(bt, bt->frame, slot);
3554 }
3555
bt_clear_slot(struct backtrack_state * bt,u32 slot)3556 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot)
3557 {
3558 bt_clear_frame_slot(bt, bt->frame, slot);
3559 }
3560
bt_frame_reg_mask(struct backtrack_state * bt,u32 frame)3561 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
3562 {
3563 return bt->reg_masks[frame];
3564 }
3565
bt_reg_mask(struct backtrack_state * bt)3566 static inline u32 bt_reg_mask(struct backtrack_state *bt)
3567 {
3568 return bt->reg_masks[bt->frame];
3569 }
3570
bt_frame_stack_mask(struct backtrack_state * bt,u32 frame)3571 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
3572 {
3573 return bt->stack_masks[frame];
3574 }
3575
bt_stack_mask(struct backtrack_state * bt)3576 static inline u64 bt_stack_mask(struct backtrack_state *bt)
3577 {
3578 return bt->stack_masks[bt->frame];
3579 }
3580
bt_is_reg_set(struct backtrack_state * bt,u32 reg)3581 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
3582 {
3583 return bt->reg_masks[bt->frame] & (1 << reg);
3584 }
3585
bt_is_slot_set(struct backtrack_state * bt,u32 slot)3586 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot)
3587 {
3588 return bt->stack_masks[bt->frame] & (1ull << slot);
3589 }
3590
3591 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
fmt_reg_mask(char * buf,ssize_t buf_sz,u32 reg_mask)3592 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
3593 {
3594 DECLARE_BITMAP(mask, 64);
3595 bool first = true;
3596 int i, n;
3597
3598 buf[0] = '\0';
3599
3600 bitmap_from_u64(mask, reg_mask);
3601 for_each_set_bit(i, mask, 32) {
3602 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
3603 first = false;
3604 buf += n;
3605 buf_sz -= n;
3606 if (buf_sz < 0)
3607 break;
3608 }
3609 }
3610 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
fmt_stack_mask(char * buf,ssize_t buf_sz,u64 stack_mask)3611 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
3612 {
3613 DECLARE_BITMAP(mask, 64);
3614 bool first = true;
3615 int i, n;
3616
3617 buf[0] = '\0';
3618
3619 bitmap_from_u64(mask, stack_mask);
3620 for_each_set_bit(i, mask, 64) {
3621 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
3622 first = false;
3623 buf += n;
3624 buf_sz -= n;
3625 if (buf_sz < 0)
3626 break;
3627 }
3628 }
3629
3630 static bool calls_callback(struct bpf_verifier_env *env, int insn_idx);
3631
3632 /* For given verifier state backtrack_insn() is called from the last insn to
3633 * the first insn. Its purpose is to compute a bitmask of registers and
3634 * stack slots that needs precision in the parent verifier state.
3635 *
3636 * @idx is an index of the instruction we are currently processing;
3637 * @subseq_idx is an index of the subsequent instruction that:
3638 * - *would be* executed next, if jump history is viewed in forward order;
3639 * - *was* processed previously during backtracking.
3640 */
backtrack_insn(struct bpf_verifier_env * env,int idx,int subseq_idx,struct backtrack_state * bt)3641 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
3642 struct backtrack_state *bt)
3643 {
3644 const struct bpf_insn_cbs cbs = {
3645 .cb_call = disasm_kfunc_name,
3646 .cb_print = verbose,
3647 .private_data = env,
3648 };
3649 struct bpf_insn *insn = env->prog->insnsi + idx;
3650 u8 class = BPF_CLASS(insn->code);
3651 u8 opcode = BPF_OP(insn->code);
3652 u8 mode = BPF_MODE(insn->code);
3653 u32 dreg = insn->dst_reg;
3654 u32 sreg = insn->src_reg;
3655 u32 spi, i;
3656
3657 if (insn->code == 0)
3658 return 0;
3659 if (env->log.level & BPF_LOG_LEVEL2) {
3660 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
3661 verbose(env, "mark_precise: frame%d: regs=%s ",
3662 bt->frame, env->tmp_str_buf);
3663 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
3664 verbose(env, "stack=%s before ", env->tmp_str_buf);
3665 verbose(env, "%d: ", idx);
3666 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
3667 }
3668
3669 if (class == BPF_ALU || class == BPF_ALU64) {
3670 if (!bt_is_reg_set(bt, dreg))
3671 return 0;
3672 if (opcode == BPF_END || opcode == BPF_NEG) {
3673 /* sreg is reserved and unused
3674 * dreg still need precision before this insn
3675 */
3676 return 0;
3677 } else if (opcode == BPF_MOV) {
3678 if (BPF_SRC(insn->code) == BPF_X) {
3679 /* dreg = sreg or dreg = (s8, s16, s32)sreg
3680 * dreg needs precision after this insn
3681 * sreg needs precision before this insn
3682 */
3683 bt_clear_reg(bt, dreg);
3684 if (sreg != BPF_REG_FP)
3685 bt_set_reg(bt, sreg);
3686 } else {
3687 /* dreg = K
3688 * dreg needs precision after this insn.
3689 * Corresponding register is already marked
3690 * as precise=true in this verifier state.
3691 * No further markings in parent are necessary
3692 */
3693 bt_clear_reg(bt, dreg);
3694 }
3695 } else {
3696 if (BPF_SRC(insn->code) == BPF_X) {
3697 /* dreg += sreg
3698 * both dreg and sreg need precision
3699 * before this insn
3700 */
3701 if (sreg != BPF_REG_FP)
3702 bt_set_reg(bt, sreg);
3703 } /* else dreg += K
3704 * dreg still needs precision before this insn
3705 */
3706 }
3707 } else if (class == BPF_LDX) {
3708 if (!bt_is_reg_set(bt, dreg))
3709 return 0;
3710 bt_clear_reg(bt, dreg);
3711
3712 /* scalars can only be spilled into stack w/o losing precision.
3713 * Load from any other memory can be zero extended.
3714 * The desire to keep that precision is already indicated
3715 * by 'precise' mark in corresponding register of this state.
3716 * No further tracking necessary.
3717 */
3718 if (insn->src_reg != BPF_REG_FP)
3719 return 0;
3720
3721 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
3722 * that [fp - off] slot contains scalar that needs to be
3723 * tracked with precision
3724 */
3725 spi = (-insn->off - 1) / BPF_REG_SIZE;
3726 if (spi >= 64) {
3727 verbose(env, "BUG spi %d\n", spi);
3728 WARN_ONCE(1, "verifier backtracking bug");
3729 return -EFAULT;
3730 }
3731 bt_set_slot(bt, spi);
3732 } else if (class == BPF_STX || class == BPF_ST) {
3733 if (bt_is_reg_set(bt, dreg))
3734 /* stx & st shouldn't be using _scalar_ dst_reg
3735 * to access memory. It means backtracking
3736 * encountered a case of pointer subtraction.
3737 */
3738 return -ENOTSUPP;
3739 /* scalars can only be spilled into stack */
3740 if (insn->dst_reg != BPF_REG_FP)
3741 return 0;
3742 spi = (-insn->off - 1) / BPF_REG_SIZE;
3743 if (spi >= 64) {
3744 verbose(env, "BUG spi %d\n", spi);
3745 WARN_ONCE(1, "verifier backtracking bug");
3746 return -EFAULT;
3747 }
3748 if (!bt_is_slot_set(bt, spi))
3749 return 0;
3750 bt_clear_slot(bt, spi);
3751 if (class == BPF_STX)
3752 bt_set_reg(bt, sreg);
3753 } else if (class == BPF_JMP || class == BPF_JMP32) {
3754 if (bpf_pseudo_call(insn)) {
3755 int subprog_insn_idx, subprog;
3756
3757 subprog_insn_idx = idx + insn->imm + 1;
3758 subprog = find_subprog(env, subprog_insn_idx);
3759 if (subprog < 0)
3760 return -EFAULT;
3761
3762 if (subprog_is_global(env, subprog)) {
3763 /* check that jump history doesn't have any
3764 * extra instructions from subprog; the next
3765 * instruction after call to global subprog
3766 * should be literally next instruction in
3767 * caller program
3768 */
3769 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug");
3770 /* r1-r5 are invalidated after subprog call,
3771 * so for global func call it shouldn't be set
3772 * anymore
3773 */
3774 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3775 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3776 WARN_ONCE(1, "verifier backtracking bug");
3777 return -EFAULT;
3778 }
3779 /* global subprog always sets R0 */
3780 bt_clear_reg(bt, BPF_REG_0);
3781 return 0;
3782 } else {
3783 /* static subprog call instruction, which
3784 * means that we are exiting current subprog,
3785 * so only r1-r5 could be still requested as
3786 * precise, r0 and r6-r10 or any stack slot in
3787 * the current frame should be zero by now
3788 */
3789 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3790 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3791 WARN_ONCE(1, "verifier backtracking bug");
3792 return -EFAULT;
3793 }
3794 /* we don't track register spills perfectly,
3795 * so fallback to force-precise instead of failing */
3796 if (bt_stack_mask(bt) != 0)
3797 return -ENOTSUPP;
3798 /* propagate r1-r5 to the caller */
3799 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
3800 if (bt_is_reg_set(bt, i)) {
3801 bt_clear_reg(bt, i);
3802 bt_set_frame_reg(bt, bt->frame - 1, i);
3803 }
3804 }
3805 if (bt_subprog_exit(bt))
3806 return -EFAULT;
3807 return 0;
3808 }
3809 } else if (is_sync_callback_calling_insn(insn) && idx != subseq_idx - 1) {
3810 /* exit from callback subprog to callback-calling helper or
3811 * kfunc call. Use idx/subseq_idx check to discern it from
3812 * straight line code backtracking.
3813 * Unlike the subprog call handling above, we shouldn't
3814 * propagate precision of r1-r5 (if any requested), as they are
3815 * not actually arguments passed directly to callback subprogs
3816 */
3817 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
3818 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3819 WARN_ONCE(1, "verifier backtracking bug");
3820 return -EFAULT;
3821 }
3822 if (bt_stack_mask(bt) != 0)
3823 return -ENOTSUPP;
3824 /* clear r1-r5 in callback subprog's mask */
3825 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3826 bt_clear_reg(bt, i);
3827 if (bt_subprog_exit(bt))
3828 return -EFAULT;
3829 return 0;
3830 } else if (opcode == BPF_CALL) {
3831 /* kfunc with imm==0 is invalid and fixup_kfunc_call will
3832 * catch this error later. Make backtracking conservative
3833 * with ENOTSUPP.
3834 */
3835 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
3836 return -ENOTSUPP;
3837 /* regular helper call sets R0 */
3838 bt_clear_reg(bt, BPF_REG_0);
3839 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3840 /* if backtracing was looking for registers R1-R5
3841 * they should have been found already.
3842 */
3843 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3844 WARN_ONCE(1, "verifier backtracking bug");
3845 return -EFAULT;
3846 }
3847 } else if (opcode == BPF_EXIT) {
3848 bool r0_precise;
3849
3850 /* Backtracking to a nested function call, 'idx' is a part of
3851 * the inner frame 'subseq_idx' is a part of the outer frame.
3852 * In case of a regular function call, instructions giving
3853 * precision to registers R1-R5 should have been found already.
3854 * In case of a callback, it is ok to have R1-R5 marked for
3855 * backtracking, as these registers are set by the function
3856 * invoking callback.
3857 */
3858 if (subseq_idx >= 0 && calls_callback(env, subseq_idx))
3859 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3860 bt_clear_reg(bt, i);
3861 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
3862 verbose(env, "BUG regs %x\n", bt_reg_mask(bt));
3863 WARN_ONCE(1, "verifier backtracking bug");
3864 return -EFAULT;
3865 }
3866
3867 /* BPF_EXIT in subprog or callback always returns
3868 * right after the call instruction, so by checking
3869 * whether the instruction at subseq_idx-1 is subprog
3870 * call or not we can distinguish actual exit from
3871 * *subprog* from exit from *callback*. In the former
3872 * case, we need to propagate r0 precision, if
3873 * necessary. In the former we never do that.
3874 */
3875 r0_precise = subseq_idx - 1 >= 0 &&
3876 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
3877 bt_is_reg_set(bt, BPF_REG_0);
3878
3879 bt_clear_reg(bt, BPF_REG_0);
3880 if (bt_subprog_enter(bt))
3881 return -EFAULT;
3882
3883 if (r0_precise)
3884 bt_set_reg(bt, BPF_REG_0);
3885 /* r6-r9 and stack slots will stay set in caller frame
3886 * bitmasks until we return back from callee(s)
3887 */
3888 return 0;
3889 } else if (BPF_SRC(insn->code) == BPF_X) {
3890 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
3891 return 0;
3892 /* dreg <cond> sreg
3893 * Both dreg and sreg need precision before
3894 * this insn. If only sreg was marked precise
3895 * before it would be equally necessary to
3896 * propagate it to dreg.
3897 */
3898 bt_set_reg(bt, dreg);
3899 bt_set_reg(bt, sreg);
3900 /* else dreg <cond> K
3901 * Only dreg still needs precision before
3902 * this insn, so for the K-based conditional
3903 * there is nothing new to be marked.
3904 */
3905 }
3906 } else if (class == BPF_LD) {
3907 if (!bt_is_reg_set(bt, dreg))
3908 return 0;
3909 bt_clear_reg(bt, dreg);
3910 /* It's ld_imm64 or ld_abs or ld_ind.
3911 * For ld_imm64 no further tracking of precision
3912 * into parent is necessary
3913 */
3914 if (mode == BPF_IND || mode == BPF_ABS)
3915 /* to be analyzed */
3916 return -ENOTSUPP;
3917 }
3918 return 0;
3919 }
3920
3921 /* the scalar precision tracking algorithm:
3922 * . at the start all registers have precise=false.
3923 * . scalar ranges are tracked as normal through alu and jmp insns.
3924 * . once precise value of the scalar register is used in:
3925 * . ptr + scalar alu
3926 * . if (scalar cond K|scalar)
3927 * . helper_call(.., scalar, ...) where ARG_CONST is expected
3928 * backtrack through the verifier states and mark all registers and
3929 * stack slots with spilled constants that these scalar regisers
3930 * should be precise.
3931 * . during state pruning two registers (or spilled stack slots)
3932 * are equivalent if both are not precise.
3933 *
3934 * Note the verifier cannot simply walk register parentage chain,
3935 * since many different registers and stack slots could have been
3936 * used to compute single precise scalar.
3937 *
3938 * The approach of starting with precise=true for all registers and then
3939 * backtrack to mark a register as not precise when the verifier detects
3940 * that program doesn't care about specific value (e.g., when helper
3941 * takes register as ARG_ANYTHING parameter) is not safe.
3942 *
3943 * It's ok to walk single parentage chain of the verifier states.
3944 * It's possible that this backtracking will go all the way till 1st insn.
3945 * All other branches will be explored for needing precision later.
3946 *
3947 * The backtracking needs to deal with cases like:
3948 * 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)
3949 * r9 -= r8
3950 * r5 = r9
3951 * if r5 > 0x79f goto pc+7
3952 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
3953 * r5 += 1
3954 * ...
3955 * call bpf_perf_event_output#25
3956 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
3957 *
3958 * and this case:
3959 * r6 = 1
3960 * call foo // uses callee's r6 inside to compute r0
3961 * r0 += r6
3962 * if r0 == 0 goto
3963 *
3964 * to track above reg_mask/stack_mask needs to be independent for each frame.
3965 *
3966 * Also if parent's curframe > frame where backtracking started,
3967 * the verifier need to mark registers in both frames, otherwise callees
3968 * may incorrectly prune callers. This is similar to
3969 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
3970 *
3971 * For now backtracking falls back into conservative marking.
3972 */
mark_all_scalars_precise(struct bpf_verifier_env * env,struct bpf_verifier_state * st)3973 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
3974 struct bpf_verifier_state *st)
3975 {
3976 struct bpf_func_state *func;
3977 struct bpf_reg_state *reg;
3978 int i, j;
3979
3980 if (env->log.level & BPF_LOG_LEVEL2) {
3981 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
3982 st->curframe);
3983 }
3984
3985 /* big hammer: mark all scalars precise in this path.
3986 * pop_stack may still get !precise scalars.
3987 * We also skip current state and go straight to first parent state,
3988 * because precision markings in current non-checkpointed state are
3989 * not needed. See why in the comment in __mark_chain_precision below.
3990 */
3991 for (st = st->parent; st; st = st->parent) {
3992 for (i = 0; i <= st->curframe; i++) {
3993 func = st->frame[i];
3994 for (j = 0; j < BPF_REG_FP; j++) {
3995 reg = &func->regs[j];
3996 if (reg->type != SCALAR_VALUE || reg->precise)
3997 continue;
3998 reg->precise = true;
3999 if (env->log.level & BPF_LOG_LEVEL2) {
4000 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
4001 i, j);
4002 }
4003 }
4004 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4005 if (!is_spilled_reg(&func->stack[j]))
4006 continue;
4007 reg = &func->stack[j].spilled_ptr;
4008 if (reg->type != SCALAR_VALUE || reg->precise)
4009 continue;
4010 reg->precise = true;
4011 if (env->log.level & BPF_LOG_LEVEL2) {
4012 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
4013 i, -(j + 1) * 8);
4014 }
4015 }
4016 }
4017 }
4018 }
4019
mark_all_scalars_imprecise(struct bpf_verifier_env * env,struct bpf_verifier_state * st)4020 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
4021 {
4022 struct bpf_func_state *func;
4023 struct bpf_reg_state *reg;
4024 int i, j;
4025
4026 for (i = 0; i <= st->curframe; i++) {
4027 func = st->frame[i];
4028 for (j = 0; j < BPF_REG_FP; j++) {
4029 reg = &func->regs[j];
4030 if (reg->type != SCALAR_VALUE)
4031 continue;
4032 reg->precise = false;
4033 }
4034 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4035 if (!is_spilled_reg(&func->stack[j]))
4036 continue;
4037 reg = &func->stack[j].spilled_ptr;
4038 if (reg->type != SCALAR_VALUE)
4039 continue;
4040 reg->precise = false;
4041 }
4042 }
4043 }
4044
idset_contains(struct bpf_idset * s,u32 id)4045 static bool idset_contains(struct bpf_idset *s, u32 id)
4046 {
4047 u32 i;
4048
4049 for (i = 0; i < s->count; ++i)
4050 if (s->ids[i] == id)
4051 return true;
4052
4053 return false;
4054 }
4055
idset_push(struct bpf_idset * s,u32 id)4056 static int idset_push(struct bpf_idset *s, u32 id)
4057 {
4058 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids)))
4059 return -EFAULT;
4060 s->ids[s->count++] = id;
4061 return 0;
4062 }
4063
idset_reset(struct bpf_idset * s)4064 static void idset_reset(struct bpf_idset *s)
4065 {
4066 s->count = 0;
4067 }
4068
4069 /* Collect a set of IDs for all registers currently marked as precise in env->bt.
4070 * Mark all registers with these IDs as precise.
4071 */
mark_precise_scalar_ids(struct bpf_verifier_env * env,struct bpf_verifier_state * st)4072 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
4073 {
4074 struct bpf_idset *precise_ids = &env->idset_scratch;
4075 struct backtrack_state *bt = &env->bt;
4076 struct bpf_func_state *func;
4077 struct bpf_reg_state *reg;
4078 DECLARE_BITMAP(mask, 64);
4079 int i, fr;
4080
4081 idset_reset(precise_ids);
4082
4083 for (fr = bt->frame; fr >= 0; fr--) {
4084 func = st->frame[fr];
4085
4086 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4087 for_each_set_bit(i, mask, 32) {
4088 reg = &func->regs[i];
4089 if (!reg->id || reg->type != SCALAR_VALUE)
4090 continue;
4091 if (idset_push(precise_ids, reg->id))
4092 return -EFAULT;
4093 }
4094
4095 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4096 for_each_set_bit(i, mask, 64) {
4097 if (i >= func->allocated_stack / BPF_REG_SIZE)
4098 break;
4099 if (!is_spilled_scalar_reg(&func->stack[i]))
4100 continue;
4101 reg = &func->stack[i].spilled_ptr;
4102 if (!reg->id)
4103 continue;
4104 if (idset_push(precise_ids, reg->id))
4105 return -EFAULT;
4106 }
4107 }
4108
4109 for (fr = 0; fr <= st->curframe; ++fr) {
4110 func = st->frame[fr];
4111
4112 for (i = BPF_REG_0; i < BPF_REG_10; ++i) {
4113 reg = &func->regs[i];
4114 if (!reg->id)
4115 continue;
4116 if (!idset_contains(precise_ids, reg->id))
4117 continue;
4118 bt_set_frame_reg(bt, fr, i);
4119 }
4120 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) {
4121 if (!is_spilled_scalar_reg(&func->stack[i]))
4122 continue;
4123 reg = &func->stack[i].spilled_ptr;
4124 if (!reg->id)
4125 continue;
4126 if (!idset_contains(precise_ids, reg->id))
4127 continue;
4128 bt_set_frame_slot(bt, fr, i);
4129 }
4130 }
4131
4132 return 0;
4133 }
4134
4135 /*
4136 * __mark_chain_precision() backtracks BPF program instruction sequence and
4137 * chain of verifier states making sure that register *regno* (if regno >= 0)
4138 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
4139 * SCALARS, as well as any other registers and slots that contribute to
4140 * a tracked state of given registers/stack slots, depending on specific BPF
4141 * assembly instructions (see backtrack_insns() for exact instruction handling
4142 * logic). This backtracking relies on recorded jmp_history and is able to
4143 * traverse entire chain of parent states. This process ends only when all the
4144 * necessary registers/slots and their transitive dependencies are marked as
4145 * precise.
4146 *
4147 * One important and subtle aspect is that precise marks *do not matter* in
4148 * the currently verified state (current state). It is important to understand
4149 * why this is the case.
4150 *
4151 * First, note that current state is the state that is not yet "checkpointed",
4152 * i.e., it is not yet put into env->explored_states, and it has no children
4153 * states as well. It's ephemeral, and can end up either a) being discarded if
4154 * compatible explored state is found at some point or BPF_EXIT instruction is
4155 * reached or b) checkpointed and put into env->explored_states, branching out
4156 * into one or more children states.
4157 *
4158 * In the former case, precise markings in current state are completely
4159 * ignored by state comparison code (see regsafe() for details). Only
4160 * checkpointed ("old") state precise markings are important, and if old
4161 * state's register/slot is precise, regsafe() assumes current state's
4162 * register/slot as precise and checks value ranges exactly and precisely. If
4163 * states turn out to be compatible, current state's necessary precise
4164 * markings and any required parent states' precise markings are enforced
4165 * after the fact with propagate_precision() logic, after the fact. But it's
4166 * important to realize that in this case, even after marking current state
4167 * registers/slots as precise, we immediately discard current state. So what
4168 * actually matters is any of the precise markings propagated into current
4169 * state's parent states, which are always checkpointed (due to b) case above).
4170 * As such, for scenario a) it doesn't matter if current state has precise
4171 * markings set or not.
4172 *
4173 * Now, for the scenario b), checkpointing and forking into child(ren)
4174 * state(s). Note that before current state gets to checkpointing step, any
4175 * processed instruction always assumes precise SCALAR register/slot
4176 * knowledge: if precise value or range is useful to prune jump branch, BPF
4177 * verifier takes this opportunity enthusiastically. Similarly, when
4178 * register's value is used to calculate offset or memory address, exact
4179 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
4180 * what we mentioned above about state comparison ignoring precise markings
4181 * during state comparison, BPF verifier ignores and also assumes precise
4182 * markings *at will* during instruction verification process. But as verifier
4183 * assumes precision, it also propagates any precision dependencies across
4184 * parent states, which are not yet finalized, so can be further restricted
4185 * based on new knowledge gained from restrictions enforced by their children
4186 * states. This is so that once those parent states are finalized, i.e., when
4187 * they have no more active children state, state comparison logic in
4188 * is_state_visited() would enforce strict and precise SCALAR ranges, if
4189 * required for correctness.
4190 *
4191 * To build a bit more intuition, note also that once a state is checkpointed,
4192 * the path we took to get to that state is not important. This is crucial
4193 * property for state pruning. When state is checkpointed and finalized at
4194 * some instruction index, it can be correctly and safely used to "short
4195 * circuit" any *compatible* state that reaches exactly the same instruction
4196 * index. I.e., if we jumped to that instruction from a completely different
4197 * code path than original finalized state was derived from, it doesn't
4198 * matter, current state can be discarded because from that instruction
4199 * forward having a compatible state will ensure we will safely reach the
4200 * exit. States describe preconditions for further exploration, but completely
4201 * forget the history of how we got here.
4202 *
4203 * This also means that even if we needed precise SCALAR range to get to
4204 * finalized state, but from that point forward *that same* SCALAR register is
4205 * never used in a precise context (i.e., it's precise value is not needed for
4206 * correctness), it's correct and safe to mark such register as "imprecise"
4207 * (i.e., precise marking set to false). This is what we rely on when we do
4208 * not set precise marking in current state. If no child state requires
4209 * precision for any given SCALAR register, it's safe to dictate that it can
4210 * be imprecise. If any child state does require this register to be precise,
4211 * we'll mark it precise later retroactively during precise markings
4212 * propagation from child state to parent states.
4213 *
4214 * Skipping precise marking setting in current state is a mild version of
4215 * relying on the above observation. But we can utilize this property even
4216 * more aggressively by proactively forgetting any precise marking in the
4217 * current state (which we inherited from the parent state), right before we
4218 * checkpoint it and branch off into new child state. This is done by
4219 * mark_all_scalars_imprecise() to hopefully get more permissive and generic
4220 * finalized states which help in short circuiting more future states.
4221 */
__mark_chain_precision(struct bpf_verifier_env * env,int regno)4222 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno)
4223 {
4224 struct backtrack_state *bt = &env->bt;
4225 struct bpf_verifier_state *st = env->cur_state;
4226 int first_idx = st->first_insn_idx;
4227 int last_idx = env->insn_idx;
4228 int subseq_idx = -1;
4229 struct bpf_func_state *func;
4230 struct bpf_reg_state *reg;
4231 bool skip_first = true;
4232 int i, fr, err;
4233
4234 if (!env->bpf_capable)
4235 return 0;
4236
4237 /* set frame number from which we are starting to backtrack */
4238 bt_init(bt, env->cur_state->curframe);
4239
4240 /* Do sanity checks against current state of register and/or stack
4241 * slot, but don't set precise flag in current state, as precision
4242 * tracking in the current state is unnecessary.
4243 */
4244 func = st->frame[bt->frame];
4245 if (regno >= 0) {
4246 reg = &func->regs[regno];
4247 if (reg->type != SCALAR_VALUE) {
4248 WARN_ONCE(1, "backtracing misuse");
4249 return -EFAULT;
4250 }
4251 bt_set_reg(bt, regno);
4252 }
4253
4254 if (bt_empty(bt))
4255 return 0;
4256
4257 for (;;) {
4258 DECLARE_BITMAP(mask, 64);
4259 u32 history = st->jmp_history_cnt;
4260
4261 if (env->log.level & BPF_LOG_LEVEL2) {
4262 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4263 bt->frame, last_idx, first_idx, subseq_idx);
4264 }
4265
4266 /* If some register with scalar ID is marked as precise,
4267 * make sure that all registers sharing this ID are also precise.
4268 * This is needed to estimate effect of find_equal_scalars().
4269 * Do this at the last instruction of each state,
4270 * bpf_reg_state::id fields are valid for these instructions.
4271 *
4272 * Allows to track precision in situation like below:
4273 *
4274 * r2 = unknown value
4275 * ...
4276 * --- state #0 ---
4277 * ...
4278 * r1 = r2 // r1 and r2 now share the same ID
4279 * ...
4280 * --- state #1 {r1.id = A, r2.id = A} ---
4281 * ...
4282 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1
4283 * ...
4284 * --- state #2 {r1.id = A, r2.id = A} ---
4285 * r3 = r10
4286 * r3 += r1 // need to mark both r1 and r2
4287 */
4288 if (mark_precise_scalar_ids(env, st))
4289 return -EFAULT;
4290
4291 if (last_idx < 0) {
4292 /* we are at the entry into subprog, which
4293 * is expected for global funcs, but only if
4294 * requested precise registers are R1-R5
4295 * (which are global func's input arguments)
4296 */
4297 if (st->curframe == 0 &&
4298 st->frame[0]->subprogno > 0 &&
4299 st->frame[0]->callsite == BPF_MAIN_FUNC &&
4300 bt_stack_mask(bt) == 0 &&
4301 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4302 bitmap_from_u64(mask, bt_reg_mask(bt));
4303 for_each_set_bit(i, mask, 32) {
4304 reg = &st->frame[0]->regs[i];
4305 bt_clear_reg(bt, i);
4306 if (reg->type == SCALAR_VALUE)
4307 reg->precise = true;
4308 }
4309 return 0;
4310 }
4311
4312 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n",
4313 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4314 WARN_ONCE(1, "verifier backtracking bug");
4315 return -EFAULT;
4316 }
4317
4318 for (i = last_idx;;) {
4319 if (skip_first) {
4320 err = 0;
4321 skip_first = false;
4322 } else {
4323 err = backtrack_insn(env, i, subseq_idx, bt);
4324 }
4325 if (err == -ENOTSUPP) {
4326 mark_all_scalars_precise(env, env->cur_state);
4327 bt_reset(bt);
4328 return 0;
4329 } else if (err) {
4330 return err;
4331 }
4332 if (bt_empty(bt))
4333 /* Found assignment(s) into tracked register in this state.
4334 * Since this state is already marked, just return.
4335 * Nothing to be tracked further in the parent state.
4336 */
4337 return 0;
4338 subseq_idx = i;
4339 i = get_prev_insn_idx(st, i, &history);
4340 if (i == -ENOENT)
4341 break;
4342 if (i >= env->prog->len) {
4343 /* This can happen if backtracking reached insn 0
4344 * and there are still reg_mask or stack_mask
4345 * to backtrack.
4346 * It means the backtracking missed the spot where
4347 * particular register was initialized with a constant.
4348 */
4349 verbose(env, "BUG backtracking idx %d\n", i);
4350 WARN_ONCE(1, "verifier backtracking bug");
4351 return -EFAULT;
4352 }
4353 }
4354 st = st->parent;
4355 if (!st)
4356 break;
4357
4358 for (fr = bt->frame; fr >= 0; fr--) {
4359 func = st->frame[fr];
4360 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4361 for_each_set_bit(i, mask, 32) {
4362 reg = &func->regs[i];
4363 if (reg->type != SCALAR_VALUE) {
4364 bt_clear_frame_reg(bt, fr, i);
4365 continue;
4366 }
4367 if (reg->precise)
4368 bt_clear_frame_reg(bt, fr, i);
4369 else
4370 reg->precise = true;
4371 }
4372
4373 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4374 for_each_set_bit(i, mask, 64) {
4375 if (i >= func->allocated_stack / BPF_REG_SIZE) {
4376 /* the sequence of instructions:
4377 * 2: (bf) r3 = r10
4378 * 3: (7b) *(u64 *)(r3 -8) = r0
4379 * 4: (79) r4 = *(u64 *)(r10 -8)
4380 * doesn't contain jmps. It's backtracked
4381 * as a single block.
4382 * During backtracking insn 3 is not recognized as
4383 * stack access, so at the end of backtracking
4384 * stack slot fp-8 is still marked in stack_mask.
4385 * However the parent state may not have accessed
4386 * fp-8 and it's "unallocated" stack space.
4387 * In such case fallback to conservative.
4388 */
4389 mark_all_scalars_precise(env, env->cur_state);
4390 bt_reset(bt);
4391 return 0;
4392 }
4393
4394 if (!is_spilled_scalar_reg(&func->stack[i])) {
4395 bt_clear_frame_slot(bt, fr, i);
4396 continue;
4397 }
4398 reg = &func->stack[i].spilled_ptr;
4399 if (reg->precise)
4400 bt_clear_frame_slot(bt, fr, i);
4401 else
4402 reg->precise = true;
4403 }
4404 if (env->log.level & BPF_LOG_LEVEL2) {
4405 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4406 bt_frame_reg_mask(bt, fr));
4407 verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4408 fr, env->tmp_str_buf);
4409 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4410 bt_frame_stack_mask(bt, fr));
4411 verbose(env, "stack=%s: ", env->tmp_str_buf);
4412 print_verifier_state(env, func, true);
4413 }
4414 }
4415
4416 if (bt_empty(bt))
4417 return 0;
4418
4419 subseq_idx = first_idx;
4420 last_idx = st->last_insn_idx;
4421 first_idx = st->first_insn_idx;
4422 }
4423
4424 /* if we still have requested precise regs or slots, we missed
4425 * something (e.g., stack access through non-r10 register), so
4426 * fallback to marking all precise
4427 */
4428 if (!bt_empty(bt)) {
4429 mark_all_scalars_precise(env, env->cur_state);
4430 bt_reset(bt);
4431 }
4432
4433 return 0;
4434 }
4435
mark_chain_precision(struct bpf_verifier_env * env,int regno)4436 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4437 {
4438 return __mark_chain_precision(env, regno);
4439 }
4440
4441 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4442 * desired reg and stack masks across all relevant frames
4443 */
mark_chain_precision_batch(struct bpf_verifier_env * env)4444 static int mark_chain_precision_batch(struct bpf_verifier_env *env)
4445 {
4446 return __mark_chain_precision(env, -1);
4447 }
4448
is_spillable_regtype(enum bpf_reg_type type)4449 static bool is_spillable_regtype(enum bpf_reg_type type)
4450 {
4451 switch (base_type(type)) {
4452 case PTR_TO_MAP_VALUE:
4453 case PTR_TO_STACK:
4454 case PTR_TO_CTX:
4455 case PTR_TO_PACKET:
4456 case PTR_TO_PACKET_META:
4457 case PTR_TO_PACKET_END:
4458 case PTR_TO_FLOW_KEYS:
4459 case CONST_PTR_TO_MAP:
4460 case PTR_TO_SOCKET:
4461 case PTR_TO_SOCK_COMMON:
4462 case PTR_TO_TCP_SOCK:
4463 case PTR_TO_XDP_SOCK:
4464 case PTR_TO_BTF_ID:
4465 case PTR_TO_BUF:
4466 case PTR_TO_MEM:
4467 case PTR_TO_FUNC:
4468 case PTR_TO_MAP_KEY:
4469 return true;
4470 default:
4471 return false;
4472 }
4473 }
4474
4475 /* Does this register contain a constant zero? */
register_is_null(struct bpf_reg_state * reg)4476 static bool register_is_null(struct bpf_reg_state *reg)
4477 {
4478 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4479 }
4480
register_is_const(struct bpf_reg_state * reg)4481 static bool register_is_const(struct bpf_reg_state *reg)
4482 {
4483 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
4484 }
4485
__is_scalar_unbounded(struct bpf_reg_state * reg)4486 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
4487 {
4488 return tnum_is_unknown(reg->var_off) &&
4489 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
4490 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
4491 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
4492 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
4493 }
4494
register_is_bounded(struct bpf_reg_state * reg)4495 static bool register_is_bounded(struct bpf_reg_state *reg)
4496 {
4497 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
4498 }
4499
__is_pointer_value(bool allow_ptr_leaks,const struct bpf_reg_state * reg)4500 static bool __is_pointer_value(bool allow_ptr_leaks,
4501 const struct bpf_reg_state *reg)
4502 {
4503 if (allow_ptr_leaks)
4504 return false;
4505
4506 return reg->type != SCALAR_VALUE;
4507 }
4508
4509 /* Copy src state preserving dst->parent and dst->live fields */
copy_register_state(struct bpf_reg_state * dst,const struct bpf_reg_state * src)4510 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
4511 {
4512 struct bpf_reg_state *parent = dst->parent;
4513 enum bpf_reg_liveness live = dst->live;
4514
4515 *dst = *src;
4516 dst->parent = parent;
4517 dst->live = live;
4518 }
4519
save_register_state(struct bpf_func_state * state,int spi,struct bpf_reg_state * reg,int size)4520 static void save_register_state(struct bpf_func_state *state,
4521 int spi, struct bpf_reg_state *reg,
4522 int size)
4523 {
4524 int i;
4525
4526 copy_register_state(&state->stack[spi].spilled_ptr, reg);
4527 if (size == BPF_REG_SIZE)
4528 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4529
4530 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
4531 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
4532
4533 /* size < 8 bytes spill */
4534 for (; i; i--)
4535 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
4536 }
4537
is_bpf_st_mem(struct bpf_insn * insn)4538 static bool is_bpf_st_mem(struct bpf_insn *insn)
4539 {
4540 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
4541 }
4542
4543 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
4544 * stack boundary and alignment are checked in check_mem_access()
4545 */
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)4546 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
4547 /* stack frame we're writing to */
4548 struct bpf_func_state *state,
4549 int off, int size, int value_regno,
4550 int insn_idx)
4551 {
4552 struct bpf_func_state *cur; /* state of the current function */
4553 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
4554 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4555 struct bpf_reg_state *reg = NULL;
4556 u32 dst_reg = insn->dst_reg;
4557
4558 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
4559 * so it's aligned access and [off, off + size) are within stack limits
4560 */
4561 if (!env->allow_ptr_leaks &&
4562 is_spilled_reg(&state->stack[spi]) &&
4563 size != BPF_REG_SIZE) {
4564 verbose(env, "attempt to corrupt spilled pointer on stack\n");
4565 return -EACCES;
4566 }
4567
4568 cur = env->cur_state->frame[env->cur_state->curframe];
4569 if (value_regno >= 0)
4570 reg = &cur->regs[value_regno];
4571 if (!env->bypass_spec_v4) {
4572 bool sanitize = reg && is_spillable_regtype(reg->type);
4573
4574 for (i = 0; i < size; i++) {
4575 u8 type = state->stack[spi].slot_type[i];
4576
4577 if (type != STACK_MISC && type != STACK_ZERO) {
4578 sanitize = true;
4579 break;
4580 }
4581 }
4582
4583 if (sanitize)
4584 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
4585 }
4586
4587 err = destroy_if_dynptr_stack_slot(env, state, spi);
4588 if (err)
4589 return err;
4590
4591 mark_stack_slot_scratched(env, spi);
4592 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
4593 !register_is_null(reg) && env->bpf_capable) {
4594 if (dst_reg != BPF_REG_FP) {
4595 /* The backtracking logic can only recognize explicit
4596 * stack slot address like [fp - 8]. Other spill of
4597 * scalar via different register has to be conservative.
4598 * Backtrack from here and mark all registers as precise
4599 * that contributed into 'reg' being a constant.
4600 */
4601 err = mark_chain_precision(env, value_regno);
4602 if (err)
4603 return err;
4604 }
4605 save_register_state(state, spi, reg, size);
4606 /* Break the relation on a narrowing spill. */
4607 if (fls64(reg->umax_value) > BITS_PER_BYTE * size)
4608 state->stack[spi].spilled_ptr.id = 0;
4609 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
4610 insn->imm != 0 && env->bpf_capable) {
4611 struct bpf_reg_state fake_reg = {};
4612
4613 __mark_reg_known(&fake_reg, insn->imm);
4614 fake_reg.type = SCALAR_VALUE;
4615 save_register_state(state, spi, &fake_reg, size);
4616 } else if (reg && is_spillable_regtype(reg->type)) {
4617 /* register containing pointer is being spilled into stack */
4618 if (size != BPF_REG_SIZE) {
4619 verbose_linfo(env, insn_idx, "; ");
4620 verbose(env, "invalid size of register spill\n");
4621 return -EACCES;
4622 }
4623 if (state != cur && reg->type == PTR_TO_STACK) {
4624 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
4625 return -EINVAL;
4626 }
4627 save_register_state(state, spi, reg, size);
4628 } else {
4629 u8 type = STACK_MISC;
4630
4631 /* regular write of data into stack destroys any spilled ptr */
4632 state->stack[spi].spilled_ptr.type = NOT_INIT;
4633 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
4634 if (is_stack_slot_special(&state->stack[spi]))
4635 for (i = 0; i < BPF_REG_SIZE; i++)
4636 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
4637
4638 /* only mark the slot as written if all 8 bytes were written
4639 * otherwise read propagation may incorrectly stop too soon
4640 * when stack slots are partially written.
4641 * This heuristic means that read propagation will be
4642 * conservative, since it will add reg_live_read marks
4643 * to stack slots all the way to first state when programs
4644 * writes+reads less than 8 bytes
4645 */
4646 if (size == BPF_REG_SIZE)
4647 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
4648
4649 /* when we zero initialize stack slots mark them as such */
4650 if ((reg && register_is_null(reg)) ||
4651 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
4652 /* backtracking doesn't work for STACK_ZERO yet. */
4653 err = mark_chain_precision(env, value_regno);
4654 if (err)
4655 return err;
4656 type = STACK_ZERO;
4657 }
4658
4659 /* Mark slots affected by this stack write. */
4660 for (i = 0; i < size; i++)
4661 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
4662 type;
4663 }
4664 return 0;
4665 }
4666
4667 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
4668 * known to contain a variable offset.
4669 * This function checks whether the write is permitted and conservatively
4670 * tracks the effects of the write, considering that each stack slot in the
4671 * dynamic range is potentially written to.
4672 *
4673 * 'off' includes 'regno->off'.
4674 * 'value_regno' can be -1, meaning that an unknown value is being written to
4675 * the stack.
4676 *
4677 * Spilled pointers in range are not marked as written because we don't know
4678 * what's going to be actually written. This means that read propagation for
4679 * future reads cannot be terminated by this write.
4680 *
4681 * For privileged programs, uninitialized stack slots are considered
4682 * initialized by this write (even though we don't know exactly what offsets
4683 * are going to be written to). The idea is that we don't want the verifier to
4684 * reject future reads that access slots written to through variable offsets.
4685 */
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)4686 static int check_stack_write_var_off(struct bpf_verifier_env *env,
4687 /* func where register points to */
4688 struct bpf_func_state *state,
4689 int ptr_regno, int off, int size,
4690 int value_regno, int insn_idx)
4691 {
4692 struct bpf_func_state *cur; /* state of the current function */
4693 int min_off, max_off;
4694 int i, err;
4695 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
4696 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
4697 bool writing_zero = false;
4698 /* set if the fact that we're writing a zero is used to let any
4699 * stack slots remain STACK_ZERO
4700 */
4701 bool zero_used = false;
4702
4703 cur = env->cur_state->frame[env->cur_state->curframe];
4704 ptr_reg = &cur->regs[ptr_regno];
4705 min_off = ptr_reg->smin_value + off;
4706 max_off = ptr_reg->smax_value + off + size;
4707 if (value_regno >= 0)
4708 value_reg = &cur->regs[value_regno];
4709 if ((value_reg && register_is_null(value_reg)) ||
4710 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
4711 writing_zero = true;
4712
4713 for (i = min_off; i < max_off; i++) {
4714 int spi;
4715
4716 spi = __get_spi(i);
4717 err = destroy_if_dynptr_stack_slot(env, state, spi);
4718 if (err)
4719 return err;
4720 }
4721
4722 /* Variable offset writes destroy any spilled pointers in range. */
4723 for (i = min_off; i < max_off; i++) {
4724 u8 new_type, *stype;
4725 int slot, spi;
4726
4727 slot = -i - 1;
4728 spi = slot / BPF_REG_SIZE;
4729 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4730 mark_stack_slot_scratched(env, spi);
4731
4732 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
4733 /* Reject the write if range we may write to has not
4734 * been initialized beforehand. If we didn't reject
4735 * here, the ptr status would be erased below (even
4736 * though not all slots are actually overwritten),
4737 * possibly opening the door to leaks.
4738 *
4739 * We do however catch STACK_INVALID case below, and
4740 * only allow reading possibly uninitialized memory
4741 * later for CAP_PERFMON, as the write may not happen to
4742 * that slot.
4743 */
4744 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
4745 insn_idx, i);
4746 return -EINVAL;
4747 }
4748
4749 /* Erase all spilled pointers. */
4750 state->stack[spi].spilled_ptr.type = NOT_INIT;
4751
4752 /* Update the slot type. */
4753 new_type = STACK_MISC;
4754 if (writing_zero && *stype == STACK_ZERO) {
4755 new_type = STACK_ZERO;
4756 zero_used = true;
4757 }
4758 /* If the slot is STACK_INVALID, we check whether it's OK to
4759 * pretend that it will be initialized by this write. The slot
4760 * might not actually be written to, and so if we mark it as
4761 * initialized future reads might leak uninitialized memory.
4762 * For privileged programs, we will accept such reads to slots
4763 * that may or may not be written because, if we're reject
4764 * them, the error would be too confusing.
4765 */
4766 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
4767 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
4768 insn_idx, i);
4769 return -EINVAL;
4770 }
4771 *stype = new_type;
4772 }
4773 if (zero_used) {
4774 /* backtracking doesn't work for STACK_ZERO yet. */
4775 err = mark_chain_precision(env, value_regno);
4776 if (err)
4777 return err;
4778 }
4779 return 0;
4780 }
4781
4782 /* When register 'dst_regno' is assigned some values from stack[min_off,
4783 * max_off), we set the register's type according to the types of the
4784 * respective stack slots. If all the stack values are known to be zeros, then
4785 * so is the destination reg. Otherwise, the register is considered to be
4786 * SCALAR. This function does not deal with register filling; the caller must
4787 * ensure that all spilled registers in the stack range have been marked as
4788 * read.
4789 */
mark_reg_stack_read(struct bpf_verifier_env * env,struct bpf_func_state * ptr_state,int min_off,int max_off,int dst_regno)4790 static void mark_reg_stack_read(struct bpf_verifier_env *env,
4791 /* func where src register points to */
4792 struct bpf_func_state *ptr_state,
4793 int min_off, int max_off, int dst_regno)
4794 {
4795 struct bpf_verifier_state *vstate = env->cur_state;
4796 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4797 int i, slot, spi;
4798 u8 *stype;
4799 int zeros = 0;
4800
4801 for (i = min_off; i < max_off; i++) {
4802 slot = -i - 1;
4803 spi = slot / BPF_REG_SIZE;
4804 mark_stack_slot_scratched(env, spi);
4805 stype = ptr_state->stack[spi].slot_type;
4806 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
4807 break;
4808 zeros++;
4809 }
4810 if (zeros == max_off - min_off) {
4811 /* any access_size read into register is zero extended,
4812 * so the whole register == const_zero
4813 */
4814 __mark_reg_const_zero(&state->regs[dst_regno]);
4815 /* backtracking doesn't support STACK_ZERO yet,
4816 * so mark it precise here, so that later
4817 * backtracking can stop here.
4818 * Backtracking may not need this if this register
4819 * doesn't participate in pointer adjustment.
4820 * Forward propagation of precise flag is not
4821 * necessary either. This mark is only to stop
4822 * backtracking. Any register that contributed
4823 * to const 0 was marked precise before spill.
4824 */
4825 state->regs[dst_regno].precise = true;
4826 } else {
4827 /* have read misc data from the stack */
4828 mark_reg_unknown(env, state->regs, dst_regno);
4829 }
4830 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4831 }
4832
4833 /* Read the stack at 'off' and put the results into the register indicated by
4834 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
4835 * spilled reg.
4836 *
4837 * 'dst_regno' can be -1, meaning that the read value is not going to a
4838 * register.
4839 *
4840 * The access is assumed to be within the current stack bounds.
4841 */
check_stack_read_fixed_off(struct bpf_verifier_env * env,struct bpf_func_state * reg_state,int off,int size,int dst_regno)4842 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
4843 /* func where src register points to */
4844 struct bpf_func_state *reg_state,
4845 int off, int size, int dst_regno)
4846 {
4847 struct bpf_verifier_state *vstate = env->cur_state;
4848 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4849 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
4850 struct bpf_reg_state *reg;
4851 u8 *stype, type;
4852
4853 stype = reg_state->stack[spi].slot_type;
4854 reg = ®_state->stack[spi].spilled_ptr;
4855
4856 mark_stack_slot_scratched(env, spi);
4857
4858 if (is_spilled_reg(®_state->stack[spi])) {
4859 u8 spill_size = 1;
4860
4861 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
4862 spill_size++;
4863
4864 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
4865 if (reg->type != SCALAR_VALUE) {
4866 verbose_linfo(env, env->insn_idx, "; ");
4867 verbose(env, "invalid size of register fill\n");
4868 return -EACCES;
4869 }
4870
4871 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4872 if (dst_regno < 0)
4873 return 0;
4874
4875 if (!(off % BPF_REG_SIZE) && size == spill_size) {
4876 /* The earlier check_reg_arg() has decided the
4877 * subreg_def for this insn. Save it first.
4878 */
4879 s32 subreg_def = state->regs[dst_regno].subreg_def;
4880
4881 copy_register_state(&state->regs[dst_regno], reg);
4882 state->regs[dst_regno].subreg_def = subreg_def;
4883 } else {
4884 for (i = 0; i < size; i++) {
4885 type = stype[(slot - i) % BPF_REG_SIZE];
4886 if (type == STACK_SPILL)
4887 continue;
4888 if (type == STACK_MISC)
4889 continue;
4890 if (type == STACK_INVALID && env->allow_uninit_stack)
4891 continue;
4892 verbose(env, "invalid read from stack off %d+%d size %d\n",
4893 off, i, size);
4894 return -EACCES;
4895 }
4896 mark_reg_unknown(env, state->regs, dst_regno);
4897 }
4898 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4899 return 0;
4900 }
4901
4902 if (dst_regno >= 0) {
4903 /* restore register state from stack */
4904 copy_register_state(&state->regs[dst_regno], reg);
4905 /* mark reg as written since spilled pointer state likely
4906 * has its liveness marks cleared by is_state_visited()
4907 * which resets stack/reg liveness for state transitions
4908 */
4909 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
4910 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
4911 /* If dst_regno==-1, the caller is asking us whether
4912 * it is acceptable to use this value as a SCALAR_VALUE
4913 * (e.g. for XADD).
4914 * We must not allow unprivileged callers to do that
4915 * with spilled pointers.
4916 */
4917 verbose(env, "leaking pointer from stack off %d\n",
4918 off);
4919 return -EACCES;
4920 }
4921 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4922 } else {
4923 for (i = 0; i < size; i++) {
4924 type = stype[(slot - i) % BPF_REG_SIZE];
4925 if (type == STACK_MISC)
4926 continue;
4927 if (type == STACK_ZERO)
4928 continue;
4929 if (type == STACK_INVALID && env->allow_uninit_stack)
4930 continue;
4931 verbose(env, "invalid read from stack off %d+%d size %d\n",
4932 off, i, size);
4933 return -EACCES;
4934 }
4935 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
4936 if (dst_regno >= 0)
4937 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
4938 }
4939 return 0;
4940 }
4941
4942 enum bpf_access_src {
4943 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
4944 ACCESS_HELPER = 2, /* the access is performed by a helper */
4945 };
4946
4947 static int check_stack_range_initialized(struct bpf_verifier_env *env,
4948 int regno, int off, int access_size,
4949 bool zero_size_allowed,
4950 enum bpf_access_src type,
4951 struct bpf_call_arg_meta *meta);
4952
reg_state(struct bpf_verifier_env * env,int regno)4953 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
4954 {
4955 return cur_regs(env) + regno;
4956 }
4957
4958 /* Read the stack at 'ptr_regno + off' and put the result into the register
4959 * 'dst_regno'.
4960 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
4961 * but not its variable offset.
4962 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
4963 *
4964 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
4965 * filling registers (i.e. reads of spilled register cannot be detected when
4966 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
4967 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
4968 * offset; for a fixed offset check_stack_read_fixed_off should be used
4969 * instead.
4970 */
check_stack_read_var_off(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)4971 static int check_stack_read_var_off(struct bpf_verifier_env *env,
4972 int ptr_regno, int off, int size, int dst_regno)
4973 {
4974 /* The state of the source register. */
4975 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
4976 struct bpf_func_state *ptr_state = func(env, reg);
4977 int err;
4978 int min_off, max_off;
4979
4980 /* Note that we pass a NULL meta, so raw access will not be permitted.
4981 */
4982 err = check_stack_range_initialized(env, ptr_regno, off, size,
4983 false, ACCESS_DIRECT, NULL);
4984 if (err)
4985 return err;
4986
4987 min_off = reg->smin_value + off;
4988 max_off = reg->smax_value + off;
4989 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
4990 return 0;
4991 }
4992
4993 /* check_stack_read dispatches to check_stack_read_fixed_off or
4994 * check_stack_read_var_off.
4995 *
4996 * The caller must ensure that the offset falls within the allocated stack
4997 * bounds.
4998 *
4999 * 'dst_regno' is a register which will receive the value from the stack. It
5000 * can be -1, meaning that the read value is not going to a register.
5001 */
check_stack_read(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)5002 static int check_stack_read(struct bpf_verifier_env *env,
5003 int ptr_regno, int off, int size,
5004 int dst_regno)
5005 {
5006 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
5007 struct bpf_func_state *state = func(env, reg);
5008 int err;
5009 /* Some accesses are only permitted with a static offset. */
5010 bool var_off = !tnum_is_const(reg->var_off);
5011
5012 /* The offset is required to be static when reads don't go to a
5013 * register, in order to not leak pointers (see
5014 * check_stack_read_fixed_off).
5015 */
5016 if (dst_regno < 0 && var_off) {
5017 char tn_buf[48];
5018
5019 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5020 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
5021 tn_buf, off, size);
5022 return -EACCES;
5023 }
5024 /* Variable offset is prohibited for unprivileged mode for simplicity
5025 * since it requires corresponding support in Spectre masking for stack
5026 * ALU. See also retrieve_ptr_limit(). The check in
5027 * check_stack_access_for_ptr_arithmetic() called by
5028 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
5029 * with variable offsets, therefore no check is required here. Further,
5030 * just checking it here would be insufficient as speculative stack
5031 * writes could still lead to unsafe speculative behaviour.
5032 */
5033 if (!var_off) {
5034 off += reg->var_off.value;
5035 err = check_stack_read_fixed_off(env, state, off, size,
5036 dst_regno);
5037 } else {
5038 /* Variable offset stack reads need more conservative handling
5039 * than fixed offset ones. Note that dst_regno >= 0 on this
5040 * branch.
5041 */
5042 err = check_stack_read_var_off(env, ptr_regno, off, size,
5043 dst_regno);
5044 }
5045 return err;
5046 }
5047
5048
5049 /* check_stack_write dispatches to check_stack_write_fixed_off or
5050 * check_stack_write_var_off.
5051 *
5052 * 'ptr_regno' is the register used as a pointer into the stack.
5053 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
5054 * 'value_regno' is the register whose value we're writing to the stack. It can
5055 * be -1, meaning that we're not writing from a register.
5056 *
5057 * The caller must ensure that the offset falls within the maximum stack size.
5058 */
check_stack_write(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int value_regno,int insn_idx)5059 static int check_stack_write(struct bpf_verifier_env *env,
5060 int ptr_regno, int off, int size,
5061 int value_regno, int insn_idx)
5062 {
5063 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
5064 struct bpf_func_state *state = func(env, reg);
5065 int err;
5066
5067 if (tnum_is_const(reg->var_off)) {
5068 off += reg->var_off.value;
5069 err = check_stack_write_fixed_off(env, state, off, size,
5070 value_regno, insn_idx);
5071 } else {
5072 /* Variable offset stack reads need more conservative handling
5073 * than fixed offset ones.
5074 */
5075 err = check_stack_write_var_off(env, state,
5076 ptr_regno, off, size,
5077 value_regno, insn_idx);
5078 }
5079 return err;
5080 }
5081
check_map_access_type(struct bpf_verifier_env * env,u32 regno,int off,int size,enum bpf_access_type type)5082 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
5083 int off, int size, enum bpf_access_type type)
5084 {
5085 struct bpf_reg_state *regs = cur_regs(env);
5086 struct bpf_map *map = regs[regno].map_ptr;
5087 u32 cap = bpf_map_flags_to_cap(map);
5088
5089 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
5090 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
5091 map->value_size, off, size);
5092 return -EACCES;
5093 }
5094
5095 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
5096 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
5097 map->value_size, off, size);
5098 return -EACCES;
5099 }
5100
5101 return 0;
5102 }
5103
5104 /* 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)5105 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
5106 int off, int size, u32 mem_size,
5107 bool zero_size_allowed)
5108 {
5109 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
5110 struct bpf_reg_state *reg;
5111
5112 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
5113 return 0;
5114
5115 reg = &cur_regs(env)[regno];
5116 switch (reg->type) {
5117 case PTR_TO_MAP_KEY:
5118 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
5119 mem_size, off, size);
5120 break;
5121 case PTR_TO_MAP_VALUE:
5122 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
5123 mem_size, off, size);
5124 break;
5125 case PTR_TO_PACKET:
5126 case PTR_TO_PACKET_META:
5127 case PTR_TO_PACKET_END:
5128 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
5129 off, size, regno, reg->id, off, mem_size);
5130 break;
5131 case PTR_TO_MEM:
5132 default:
5133 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
5134 mem_size, off, size);
5135 }
5136
5137 return -EACCES;
5138 }
5139
5140 /* 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)5141 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
5142 int off, int size, u32 mem_size,
5143 bool zero_size_allowed)
5144 {
5145 struct bpf_verifier_state *vstate = env->cur_state;
5146 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5147 struct bpf_reg_state *reg = &state->regs[regno];
5148 int err;
5149
5150 /* We may have adjusted the register pointing to memory region, so we
5151 * need to try adding each of min_value and max_value to off
5152 * to make sure our theoretical access will be safe.
5153 *
5154 * The minimum value is only important with signed
5155 * comparisons where we can't assume the floor of a
5156 * value is 0. If we are using signed variables for our
5157 * index'es we need to make sure that whatever we use
5158 * will have a set floor within our range.
5159 */
5160 if (reg->smin_value < 0 &&
5161 (reg->smin_value == S64_MIN ||
5162 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
5163 reg->smin_value + off < 0)) {
5164 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5165 regno);
5166 return -EACCES;
5167 }
5168 err = __check_mem_access(env, regno, reg->smin_value + off, size,
5169 mem_size, zero_size_allowed);
5170 if (err) {
5171 verbose(env, "R%d min value is outside of the allowed memory range\n",
5172 regno);
5173 return err;
5174 }
5175
5176 /* If we haven't set a max value then we need to bail since we can't be
5177 * sure we won't do bad things.
5178 * If reg->umax_value + off could overflow, treat that as unbounded too.
5179 */
5180 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
5181 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
5182 regno);
5183 return -EACCES;
5184 }
5185 err = __check_mem_access(env, regno, reg->umax_value + off, size,
5186 mem_size, zero_size_allowed);
5187 if (err) {
5188 verbose(env, "R%d max value is outside of the allowed memory range\n",
5189 regno);
5190 return err;
5191 }
5192
5193 return 0;
5194 }
5195
__check_ptr_off_reg(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,bool fixed_off_ok)5196 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
5197 const struct bpf_reg_state *reg, int regno,
5198 bool fixed_off_ok)
5199 {
5200 /* Access to this pointer-typed register or passing it to a helper
5201 * is only allowed in its original, unmodified form.
5202 */
5203
5204 if (reg->off < 0) {
5205 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
5206 reg_type_str(env, reg->type), regno, reg->off);
5207 return -EACCES;
5208 }
5209
5210 if (!fixed_off_ok && reg->off) {
5211 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
5212 reg_type_str(env, reg->type), regno, reg->off);
5213 return -EACCES;
5214 }
5215
5216 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5217 char tn_buf[48];
5218
5219 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5220 verbose(env, "variable %s access var_off=%s disallowed\n",
5221 reg_type_str(env, reg->type), tn_buf);
5222 return -EACCES;
5223 }
5224
5225 return 0;
5226 }
5227
check_ptr_off_reg(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno)5228 int check_ptr_off_reg(struct bpf_verifier_env *env,
5229 const struct bpf_reg_state *reg, int regno)
5230 {
5231 return __check_ptr_off_reg(env, reg, regno, false);
5232 }
5233
map_kptr_match_type(struct bpf_verifier_env * env,struct btf_field * kptr_field,struct bpf_reg_state * reg,u32 regno)5234 static int map_kptr_match_type(struct bpf_verifier_env *env,
5235 struct btf_field *kptr_field,
5236 struct bpf_reg_state *reg, u32 regno)
5237 {
5238 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
5239 int perm_flags;
5240 const char *reg_name = "";
5241
5242 if (btf_is_kernel(reg->btf)) {
5243 perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
5244
5245 /* Only unreferenced case accepts untrusted pointers */
5246 if (kptr_field->type == BPF_KPTR_UNREF)
5247 perm_flags |= PTR_UNTRUSTED;
5248 } else {
5249 perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5250 }
5251
5252 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
5253 goto bad_type;
5254
5255 /* We need to verify reg->type and reg->btf, before accessing reg->btf */
5256 reg_name = btf_type_name(reg->btf, reg->btf_id);
5257
5258 /* For ref_ptr case, release function check should ensure we get one
5259 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5260 * normal store of unreferenced kptr, we must ensure var_off is zero.
5261 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5262 * reg->off and reg->ref_obj_id are not needed here.
5263 */
5264 if (__check_ptr_off_reg(env, reg, regno, true))
5265 return -EACCES;
5266
5267 /* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5268 * we also need to take into account the reg->off.
5269 *
5270 * We want to support cases like:
5271 *
5272 * struct foo {
5273 * struct bar br;
5274 * struct baz bz;
5275 * };
5276 *
5277 * struct foo *v;
5278 * v = func(); // PTR_TO_BTF_ID
5279 * val->foo = v; // reg->off is zero, btf and btf_id match type
5280 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5281 * // first member type of struct after comparison fails
5282 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5283 * // to match type
5284 *
5285 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5286 * is zero. We must also ensure that btf_struct_ids_match does not walk
5287 * the struct to match type against first member of struct, i.e. reject
5288 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5289 * strict mode to true for type match.
5290 */
5291 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5292 kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5293 kptr_field->type == BPF_KPTR_REF))
5294 goto bad_type;
5295 return 0;
5296 bad_type:
5297 verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5298 reg_type_str(env, reg->type), reg_name);
5299 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5300 if (kptr_field->type == BPF_KPTR_UNREF)
5301 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5302 targ_name);
5303 else
5304 verbose(env, "\n");
5305 return -EINVAL;
5306 }
5307
5308 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5309 * can dereference RCU protected pointers and result is PTR_TRUSTED.
5310 */
in_rcu_cs(struct bpf_verifier_env * env)5311 static bool in_rcu_cs(struct bpf_verifier_env *env)
5312 {
5313 return env->cur_state->active_rcu_lock ||
5314 env->cur_state->active_lock.ptr ||
5315 !env->prog->aux->sleepable;
5316 }
5317
5318 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5319 BTF_SET_START(rcu_protected_types)
BTF_ID(struct,prog_test_ref_kfunc)5320 BTF_ID(struct, prog_test_ref_kfunc)
5321 BTF_ID(struct, cgroup)
5322 BTF_ID(struct, bpf_cpumask)
5323 BTF_ID(struct, task_struct)
5324 BTF_SET_END(rcu_protected_types)
5325
5326 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5327 {
5328 if (!btf_is_kernel(btf))
5329 return false;
5330 return btf_id_set_contains(&rcu_protected_types, btf_id);
5331 }
5332
rcu_safe_kptr(const struct btf_field * field)5333 static bool rcu_safe_kptr(const struct btf_field *field)
5334 {
5335 const struct btf_field_kptr *kptr = &field->kptr;
5336
5337 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id);
5338 }
5339
check_map_kptr_access(struct bpf_verifier_env * env,u32 regno,int value_regno,int insn_idx,struct btf_field * kptr_field)5340 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5341 int value_regno, int insn_idx,
5342 struct btf_field *kptr_field)
5343 {
5344 struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5345 int class = BPF_CLASS(insn->code);
5346 struct bpf_reg_state *val_reg;
5347
5348 /* Things we already checked for in check_map_access and caller:
5349 * - Reject cases where variable offset may touch kptr
5350 * - size of access (must be BPF_DW)
5351 * - tnum_is_const(reg->var_off)
5352 * - kptr_field->offset == off + reg->var_off.value
5353 */
5354 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5355 if (BPF_MODE(insn->code) != BPF_MEM) {
5356 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5357 return -EACCES;
5358 }
5359
5360 /* We only allow loading referenced kptr, since it will be marked as
5361 * untrusted, similar to unreferenced kptr.
5362 */
5363 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) {
5364 verbose(env, "store to referenced kptr disallowed\n");
5365 return -EACCES;
5366 }
5367
5368 if (class == BPF_LDX) {
5369 val_reg = reg_state(env, value_regno);
5370 /* We can simply mark the value_regno receiving the pointer
5371 * value from map as PTR_TO_BTF_ID, with the correct type.
5372 */
5373 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf,
5374 kptr_field->kptr.btf_id,
5375 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ?
5376 PTR_MAYBE_NULL | MEM_RCU :
5377 PTR_MAYBE_NULL | PTR_UNTRUSTED);
5378 } else if (class == BPF_STX) {
5379 val_reg = reg_state(env, value_regno);
5380 if (!register_is_null(val_reg) &&
5381 map_kptr_match_type(env, kptr_field, val_reg, value_regno))
5382 return -EACCES;
5383 } else if (class == BPF_ST) {
5384 if (insn->imm) {
5385 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
5386 kptr_field->offset);
5387 return -EACCES;
5388 }
5389 } else {
5390 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
5391 return -EACCES;
5392 }
5393 return 0;
5394 }
5395
5396 /* 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)5397 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
5398 int off, int size, bool zero_size_allowed,
5399 enum bpf_access_src src)
5400 {
5401 struct bpf_verifier_state *vstate = env->cur_state;
5402 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5403 struct bpf_reg_state *reg = &state->regs[regno];
5404 struct bpf_map *map = reg->map_ptr;
5405 struct btf_record *rec;
5406 int err, i;
5407
5408 err = check_mem_region_access(env, regno, off, size, map->value_size,
5409 zero_size_allowed);
5410 if (err)
5411 return err;
5412
5413 if (IS_ERR_OR_NULL(map->record))
5414 return 0;
5415 rec = map->record;
5416 for (i = 0; i < rec->cnt; i++) {
5417 struct btf_field *field = &rec->fields[i];
5418 u32 p = field->offset;
5419
5420 /* If any part of a field can be touched by load/store, reject
5421 * this program. To check that [x1, x2) overlaps with [y1, y2),
5422 * it is sufficient to check x1 < y2 && y1 < x2.
5423 */
5424 if (reg->smin_value + off < p + btf_field_type_size(field->type) &&
5425 p < reg->umax_value + off + size) {
5426 switch (field->type) {
5427 case BPF_KPTR_UNREF:
5428 case BPF_KPTR_REF:
5429 if (src != ACCESS_DIRECT) {
5430 verbose(env, "kptr cannot be accessed indirectly by helper\n");
5431 return -EACCES;
5432 }
5433 if (!tnum_is_const(reg->var_off)) {
5434 verbose(env, "kptr access cannot have variable offset\n");
5435 return -EACCES;
5436 }
5437 if (p != off + reg->var_off.value) {
5438 verbose(env, "kptr access misaligned expected=%u off=%llu\n",
5439 p, off + reg->var_off.value);
5440 return -EACCES;
5441 }
5442 if (size != bpf_size_to_bytes(BPF_DW)) {
5443 verbose(env, "kptr access size must be BPF_DW\n");
5444 return -EACCES;
5445 }
5446 break;
5447 default:
5448 verbose(env, "%s cannot be accessed directly by load/store\n",
5449 btf_field_type_name(field->type));
5450 return -EACCES;
5451 }
5452 }
5453 }
5454 return 0;
5455 }
5456
5457 #define MAX_PACKET_OFF 0xffff
5458
may_access_direct_pkt_data(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_access_type t)5459 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
5460 const struct bpf_call_arg_meta *meta,
5461 enum bpf_access_type t)
5462 {
5463 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
5464
5465 switch (prog_type) {
5466 /* Program types only with direct read access go here! */
5467 case BPF_PROG_TYPE_LWT_IN:
5468 case BPF_PROG_TYPE_LWT_OUT:
5469 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
5470 case BPF_PROG_TYPE_SK_REUSEPORT:
5471 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5472 case BPF_PROG_TYPE_CGROUP_SKB:
5473 if (t == BPF_WRITE)
5474 return false;
5475 fallthrough;
5476
5477 /* Program types with direct read + write access go here! */
5478 case BPF_PROG_TYPE_SCHED_CLS:
5479 case BPF_PROG_TYPE_SCHED_ACT:
5480 case BPF_PROG_TYPE_XDP:
5481 case BPF_PROG_TYPE_LWT_XMIT:
5482 case BPF_PROG_TYPE_SK_SKB:
5483 case BPF_PROG_TYPE_SK_MSG:
5484 if (meta)
5485 return meta->pkt_access;
5486
5487 env->seen_direct_write = true;
5488 return true;
5489
5490 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
5491 if (t == BPF_WRITE)
5492 env->seen_direct_write = true;
5493
5494 return true;
5495
5496 default:
5497 return false;
5498 }
5499 }
5500
check_packet_access(struct bpf_verifier_env * env,u32 regno,int off,int size,bool zero_size_allowed)5501 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
5502 int size, bool zero_size_allowed)
5503 {
5504 struct bpf_reg_state *regs = cur_regs(env);
5505 struct bpf_reg_state *reg = ®s[regno];
5506 int err;
5507
5508 /* We may have added a variable offset to the packet pointer; but any
5509 * reg->range we have comes after that. We are only checking the fixed
5510 * offset.
5511 */
5512
5513 /* We don't allow negative numbers, because we aren't tracking enough
5514 * detail to prove they're safe.
5515 */
5516 if (reg->smin_value < 0) {
5517 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5518 regno);
5519 return -EACCES;
5520 }
5521
5522 err = reg->range < 0 ? -EINVAL :
5523 __check_mem_access(env, regno, off, size, reg->range,
5524 zero_size_allowed);
5525 if (err) {
5526 verbose(env, "R%d offset is outside of the packet\n", regno);
5527 return err;
5528 }
5529
5530 /* __check_mem_access has made sure "off + size - 1" is within u16.
5531 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
5532 * otherwise find_good_pkt_pointers would have refused to set range info
5533 * that __check_mem_access would have rejected this pkt access.
5534 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
5535 */
5536 env->prog->aux->max_pkt_offset =
5537 max_t(u32, env->prog->aux->max_pkt_offset,
5538 off + reg->umax_value + size - 1);
5539
5540 return err;
5541 }
5542
5543 /* 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)5544 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
5545 enum bpf_access_type t, enum bpf_reg_type *reg_type,
5546 struct btf **btf, u32 *btf_id)
5547 {
5548 struct bpf_insn_access_aux info = {
5549 .reg_type = *reg_type,
5550 .log = &env->log,
5551 };
5552
5553 if (env->ops->is_valid_access &&
5554 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
5555 /* A non zero info.ctx_field_size indicates that this field is a
5556 * candidate for later verifier transformation to load the whole
5557 * field and then apply a mask when accessed with a narrower
5558 * access than actual ctx access size. A zero info.ctx_field_size
5559 * will only allow for whole field access and rejects any other
5560 * type of narrower access.
5561 */
5562 *reg_type = info.reg_type;
5563
5564 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
5565 *btf = info.btf;
5566 *btf_id = info.btf_id;
5567 } else {
5568 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
5569 }
5570 /* remember the offset of last byte accessed in ctx */
5571 if (env->prog->aux->max_ctx_offset < off + size)
5572 env->prog->aux->max_ctx_offset = off + size;
5573 return 0;
5574 }
5575
5576 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
5577 return -EACCES;
5578 }
5579
check_flow_keys_access(struct bpf_verifier_env * env,int off,int size)5580 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
5581 int size)
5582 {
5583 if (size < 0 || off < 0 ||
5584 (u64)off + size > sizeof(struct bpf_flow_keys)) {
5585 verbose(env, "invalid access to flow keys off=%d size=%d\n",
5586 off, size);
5587 return -EACCES;
5588 }
5589 return 0;
5590 }
5591
check_sock_access(struct bpf_verifier_env * env,int insn_idx,u32 regno,int off,int size,enum bpf_access_type t)5592 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
5593 u32 regno, int off, int size,
5594 enum bpf_access_type t)
5595 {
5596 struct bpf_reg_state *regs = cur_regs(env);
5597 struct bpf_reg_state *reg = ®s[regno];
5598 struct bpf_insn_access_aux info = {};
5599 bool valid;
5600
5601 if (reg->smin_value < 0) {
5602 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5603 regno);
5604 return -EACCES;
5605 }
5606
5607 switch (reg->type) {
5608 case PTR_TO_SOCK_COMMON:
5609 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
5610 break;
5611 case PTR_TO_SOCKET:
5612 valid = bpf_sock_is_valid_access(off, size, t, &info);
5613 break;
5614 case PTR_TO_TCP_SOCK:
5615 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
5616 break;
5617 case PTR_TO_XDP_SOCK:
5618 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
5619 break;
5620 default:
5621 valid = false;
5622 }
5623
5624
5625 if (valid) {
5626 env->insn_aux_data[insn_idx].ctx_field_size =
5627 info.ctx_field_size;
5628 return 0;
5629 }
5630
5631 verbose(env, "R%d invalid %s access off=%d size=%d\n",
5632 regno, reg_type_str(env, reg->type), off, size);
5633
5634 return -EACCES;
5635 }
5636
is_pointer_value(struct bpf_verifier_env * env,int regno)5637 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
5638 {
5639 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
5640 }
5641
is_ctx_reg(struct bpf_verifier_env * env,int regno)5642 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
5643 {
5644 const struct bpf_reg_state *reg = reg_state(env, regno);
5645
5646 return reg->type == PTR_TO_CTX;
5647 }
5648
is_sk_reg(struct bpf_verifier_env * env,int regno)5649 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
5650 {
5651 const struct bpf_reg_state *reg = reg_state(env, regno);
5652
5653 return type_is_sk_pointer(reg->type);
5654 }
5655
is_pkt_reg(struct bpf_verifier_env * env,int regno)5656 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
5657 {
5658 const struct bpf_reg_state *reg = reg_state(env, regno);
5659
5660 return type_is_pkt_pointer(reg->type);
5661 }
5662
is_flow_key_reg(struct bpf_verifier_env * env,int regno)5663 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
5664 {
5665 const struct bpf_reg_state *reg = reg_state(env, regno);
5666
5667 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
5668 return reg->type == PTR_TO_FLOW_KEYS;
5669 }
5670
5671 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
5672 #ifdef CONFIG_NET
5673 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
5674 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5675 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
5676 #endif
5677 [CONST_PTR_TO_MAP] = btf_bpf_map_id,
5678 };
5679
is_trusted_reg(const struct bpf_reg_state * reg)5680 static bool is_trusted_reg(const struct bpf_reg_state *reg)
5681 {
5682 /* A referenced register is always trusted. */
5683 if (reg->ref_obj_id)
5684 return true;
5685
5686 /* Types listed in the reg2btf_ids are always trusted */
5687 if (reg2btf_ids[base_type(reg->type)] &&
5688 !bpf_type_has_unsafe_modifiers(reg->type))
5689 return true;
5690
5691 /* If a register is not referenced, it is trusted if it has the
5692 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
5693 * other type modifiers may be safe, but we elect to take an opt-in
5694 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
5695 * not.
5696 *
5697 * Eventually, we should make PTR_TRUSTED the single source of truth
5698 * for whether a register is trusted.
5699 */
5700 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
5701 !bpf_type_has_unsafe_modifiers(reg->type);
5702 }
5703
is_rcu_reg(const struct bpf_reg_state * reg)5704 static bool is_rcu_reg(const struct bpf_reg_state *reg)
5705 {
5706 return reg->type & MEM_RCU;
5707 }
5708
clear_trusted_flags(enum bpf_type_flag * flag)5709 static void clear_trusted_flags(enum bpf_type_flag *flag)
5710 {
5711 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
5712 }
5713
check_pkt_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict)5714 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
5715 const struct bpf_reg_state *reg,
5716 int off, int size, bool strict)
5717 {
5718 struct tnum reg_off;
5719 int ip_align;
5720
5721 /* Byte size accesses are always allowed. */
5722 if (!strict || size == 1)
5723 return 0;
5724
5725 /* For platforms that do not have a Kconfig enabling
5726 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
5727 * NET_IP_ALIGN is universally set to '2'. And on platforms
5728 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
5729 * to this code only in strict mode where we want to emulate
5730 * the NET_IP_ALIGN==2 checking. Therefore use an
5731 * unconditional IP align value of '2'.
5732 */
5733 ip_align = 2;
5734
5735 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
5736 if (!tnum_is_aligned(reg_off, size)) {
5737 char tn_buf[48];
5738
5739 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5740 verbose(env,
5741 "misaligned packet access off %d+%s+%d+%d size %d\n",
5742 ip_align, tn_buf, reg->off, off, size);
5743 return -EACCES;
5744 }
5745
5746 return 0;
5747 }
5748
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)5749 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
5750 const struct bpf_reg_state *reg,
5751 const char *pointer_desc,
5752 int off, int size, bool strict)
5753 {
5754 struct tnum reg_off;
5755
5756 /* Byte size accesses are always allowed. */
5757 if (!strict || size == 1)
5758 return 0;
5759
5760 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
5761 if (!tnum_is_aligned(reg_off, size)) {
5762 char tn_buf[48];
5763
5764 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5765 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
5766 pointer_desc, tn_buf, reg->off, off, size);
5767 return -EACCES;
5768 }
5769
5770 return 0;
5771 }
5772
check_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict_alignment_once)5773 static int check_ptr_alignment(struct bpf_verifier_env *env,
5774 const struct bpf_reg_state *reg, int off,
5775 int size, bool strict_alignment_once)
5776 {
5777 bool strict = env->strict_alignment || strict_alignment_once;
5778 const char *pointer_desc = "";
5779
5780 switch (reg->type) {
5781 case PTR_TO_PACKET:
5782 case PTR_TO_PACKET_META:
5783 /* Special case, because of NET_IP_ALIGN. Given metadata sits
5784 * right in front, treat it the very same way.
5785 */
5786 return check_pkt_ptr_alignment(env, reg, off, size, strict);
5787 case PTR_TO_FLOW_KEYS:
5788 pointer_desc = "flow keys ";
5789 break;
5790 case PTR_TO_MAP_KEY:
5791 pointer_desc = "key ";
5792 break;
5793 case PTR_TO_MAP_VALUE:
5794 pointer_desc = "value ";
5795 break;
5796 case PTR_TO_CTX:
5797 pointer_desc = "context ";
5798 break;
5799 case PTR_TO_STACK:
5800 pointer_desc = "stack ";
5801 /* The stack spill tracking logic in check_stack_write_fixed_off()
5802 * and check_stack_read_fixed_off() relies on stack accesses being
5803 * aligned.
5804 */
5805 strict = true;
5806 break;
5807 case PTR_TO_SOCKET:
5808 pointer_desc = "sock ";
5809 break;
5810 case PTR_TO_SOCK_COMMON:
5811 pointer_desc = "sock_common ";
5812 break;
5813 case PTR_TO_TCP_SOCK:
5814 pointer_desc = "tcp_sock ";
5815 break;
5816 case PTR_TO_XDP_SOCK:
5817 pointer_desc = "xdp_sock ";
5818 break;
5819 default:
5820 break;
5821 }
5822 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
5823 strict);
5824 }
5825
5826 /* starting from main bpf function walk all instructions of the function
5827 * and recursively walk all callees that given function can call.
5828 * Ignore jump and exit insns.
5829 * Since recursion is prevented by check_cfg() this algorithm
5830 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
5831 */
check_max_stack_depth_subprog(struct bpf_verifier_env * env,int idx)5832 static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx)
5833 {
5834 struct bpf_subprog_info *subprog = env->subprog_info;
5835 struct bpf_insn *insn = env->prog->insnsi;
5836 int depth = 0, frame = 0, i, subprog_end;
5837 bool tail_call_reachable = false;
5838 int ret_insn[MAX_CALL_FRAMES];
5839 int ret_prog[MAX_CALL_FRAMES];
5840 int j;
5841
5842 i = subprog[idx].start;
5843 process_func:
5844 /* protect against potential stack overflow that might happen when
5845 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
5846 * depth for such case down to 256 so that the worst case scenario
5847 * would result in 8k stack size (32 which is tailcall limit * 256 =
5848 * 8k).
5849 *
5850 * To get the idea what might happen, see an example:
5851 * func1 -> sub rsp, 128
5852 * subfunc1 -> sub rsp, 256
5853 * tailcall1 -> add rsp, 256
5854 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
5855 * subfunc2 -> sub rsp, 64
5856 * subfunc22 -> sub rsp, 128
5857 * tailcall2 -> add rsp, 128
5858 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
5859 *
5860 * tailcall will unwind the current stack frame but it will not get rid
5861 * of caller's stack as shown on the example above.
5862 */
5863 if (idx && subprog[idx].has_tail_call && depth >= 256) {
5864 verbose(env,
5865 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
5866 depth);
5867 return -EACCES;
5868 }
5869 /* round up to 32-bytes, since this is granularity
5870 * of interpreter stack size
5871 */
5872 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5873 if (depth > MAX_BPF_STACK) {
5874 verbose(env, "combined stack size of %d calls is %d. Too large\n",
5875 frame + 1, depth);
5876 return -EACCES;
5877 }
5878 continue_func:
5879 subprog_end = subprog[idx + 1].start;
5880 for (; i < subprog_end; i++) {
5881 int next_insn, sidx;
5882
5883 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
5884 continue;
5885 /* remember insn and function to return to */
5886 ret_insn[frame] = i + 1;
5887 ret_prog[frame] = idx;
5888
5889 /* find the callee */
5890 next_insn = i + insn[i].imm + 1;
5891 sidx = find_subprog(env, next_insn);
5892 if (sidx < 0) {
5893 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5894 next_insn);
5895 return -EFAULT;
5896 }
5897 if (subprog[sidx].is_async_cb) {
5898 if (subprog[sidx].has_tail_call) {
5899 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
5900 return -EFAULT;
5901 }
5902 /* async callbacks don't increase bpf prog stack size unless called directly */
5903 if (!bpf_pseudo_call(insn + i))
5904 continue;
5905 }
5906 i = next_insn;
5907 idx = sidx;
5908
5909 if (subprog[idx].has_tail_call)
5910 tail_call_reachable = true;
5911
5912 frame++;
5913 if (frame >= MAX_CALL_FRAMES) {
5914 verbose(env, "the call stack of %d frames is too deep !\n",
5915 frame);
5916 return -E2BIG;
5917 }
5918 goto process_func;
5919 }
5920 /* if tail call got detected across bpf2bpf calls then mark each of the
5921 * currently present subprog frames as tail call reachable subprogs;
5922 * this info will be utilized by JIT so that we will be preserving the
5923 * tail call counter throughout bpf2bpf calls combined with tailcalls
5924 */
5925 if (tail_call_reachable)
5926 for (j = 0; j < frame; j++)
5927 subprog[ret_prog[j]].tail_call_reachable = true;
5928 if (subprog[0].tail_call_reachable)
5929 env->prog->aux->tail_call_reachable = true;
5930
5931 /* end of for() loop means the last insn of the 'subprog'
5932 * was reached. Doesn't matter whether it was JA or EXIT
5933 */
5934 if (frame == 0)
5935 return 0;
5936 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
5937 frame--;
5938 i = ret_insn[frame];
5939 idx = ret_prog[frame];
5940 goto continue_func;
5941 }
5942
check_max_stack_depth(struct bpf_verifier_env * env)5943 static int check_max_stack_depth(struct bpf_verifier_env *env)
5944 {
5945 struct bpf_subprog_info *si = env->subprog_info;
5946 int ret;
5947
5948 for (int i = 0; i < env->subprog_cnt; i++) {
5949 if (!i || si[i].is_async_cb) {
5950 ret = check_max_stack_depth_subprog(env, i);
5951 if (ret < 0)
5952 return ret;
5953 }
5954 continue;
5955 }
5956 return 0;
5957 }
5958
5959 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
get_callee_stack_depth(struct bpf_verifier_env * env,const struct bpf_insn * insn,int idx)5960 static int get_callee_stack_depth(struct bpf_verifier_env *env,
5961 const struct bpf_insn *insn, int idx)
5962 {
5963 int start = idx + insn->imm + 1, subprog;
5964
5965 subprog = find_subprog(env, start);
5966 if (subprog < 0) {
5967 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5968 start);
5969 return -EFAULT;
5970 }
5971 return env->subprog_info[subprog].stack_depth;
5972 }
5973 #endif
5974
__check_buffer_access(struct bpf_verifier_env * env,const char * buf_info,const struct bpf_reg_state * reg,int regno,int off,int size)5975 static int __check_buffer_access(struct bpf_verifier_env *env,
5976 const char *buf_info,
5977 const struct bpf_reg_state *reg,
5978 int regno, int off, int size)
5979 {
5980 if (off < 0) {
5981 verbose(env,
5982 "R%d invalid %s buffer access: off=%d, size=%d\n",
5983 regno, buf_info, off, size);
5984 return -EACCES;
5985 }
5986 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5987 char tn_buf[48];
5988
5989 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5990 verbose(env,
5991 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
5992 regno, off, tn_buf);
5993 return -EACCES;
5994 }
5995
5996 return 0;
5997 }
5998
check_tp_buffer_access(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,int off,int size)5999 static int check_tp_buffer_access(struct bpf_verifier_env *env,
6000 const struct bpf_reg_state *reg,
6001 int regno, int off, int size)
6002 {
6003 int err;
6004
6005 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
6006 if (err)
6007 return err;
6008
6009 if (off + size > env->prog->aux->max_tp_access)
6010 env->prog->aux->max_tp_access = off + size;
6011
6012 return 0;
6013 }
6014
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)6015 static int check_buffer_access(struct bpf_verifier_env *env,
6016 const struct bpf_reg_state *reg,
6017 int regno, int off, int size,
6018 bool zero_size_allowed,
6019 u32 *max_access)
6020 {
6021 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
6022 int err;
6023
6024 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
6025 if (err)
6026 return err;
6027
6028 if (off + size > *max_access)
6029 *max_access = off + size;
6030
6031 return 0;
6032 }
6033
6034 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
zext_32_to_64(struct bpf_reg_state * reg)6035 static void zext_32_to_64(struct bpf_reg_state *reg)
6036 {
6037 reg->var_off = tnum_subreg(reg->var_off);
6038 __reg_assign_32_into_64(reg);
6039 }
6040
6041 /* truncate register to smaller size (in bytes)
6042 * must be called with size < BPF_REG_SIZE
6043 */
coerce_reg_to_size(struct bpf_reg_state * reg,int size)6044 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
6045 {
6046 u64 mask;
6047
6048 /* clear high bits in bit representation */
6049 reg->var_off = tnum_cast(reg->var_off, size);
6050
6051 /* fix arithmetic bounds */
6052 mask = ((u64)1 << (size * 8)) - 1;
6053 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
6054 reg->umin_value &= mask;
6055 reg->umax_value &= mask;
6056 } else {
6057 reg->umin_value = 0;
6058 reg->umax_value = mask;
6059 }
6060 reg->smin_value = reg->umin_value;
6061 reg->smax_value = reg->umax_value;
6062
6063 /* If size is smaller than 32bit register the 32bit register
6064 * values are also truncated so we push 64-bit bounds into
6065 * 32-bit bounds. Above were truncated < 32-bits already.
6066 */
6067 if (size >= 4)
6068 return;
6069 __reg_combine_64_into_32(reg);
6070 }
6071
set_sext64_default_val(struct bpf_reg_state * reg,int size)6072 static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
6073 {
6074 if (size == 1) {
6075 reg->smin_value = reg->s32_min_value = S8_MIN;
6076 reg->smax_value = reg->s32_max_value = S8_MAX;
6077 } else if (size == 2) {
6078 reg->smin_value = reg->s32_min_value = S16_MIN;
6079 reg->smax_value = reg->s32_max_value = S16_MAX;
6080 } else {
6081 /* size == 4 */
6082 reg->smin_value = reg->s32_min_value = S32_MIN;
6083 reg->smax_value = reg->s32_max_value = S32_MAX;
6084 }
6085 reg->umin_value = reg->u32_min_value = 0;
6086 reg->umax_value = U64_MAX;
6087 reg->u32_max_value = U32_MAX;
6088 reg->var_off = tnum_unknown;
6089 }
6090
coerce_reg_to_size_sx(struct bpf_reg_state * reg,int size)6091 static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
6092 {
6093 s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
6094 u64 top_smax_value, top_smin_value;
6095 u64 num_bits = size * 8;
6096
6097 if (tnum_is_const(reg->var_off)) {
6098 u64_cval = reg->var_off.value;
6099 if (size == 1)
6100 reg->var_off = tnum_const((s8)u64_cval);
6101 else if (size == 2)
6102 reg->var_off = tnum_const((s16)u64_cval);
6103 else
6104 /* size == 4 */
6105 reg->var_off = tnum_const((s32)u64_cval);
6106
6107 u64_cval = reg->var_off.value;
6108 reg->smax_value = reg->smin_value = u64_cval;
6109 reg->umax_value = reg->umin_value = u64_cval;
6110 reg->s32_max_value = reg->s32_min_value = u64_cval;
6111 reg->u32_max_value = reg->u32_min_value = u64_cval;
6112 return;
6113 }
6114
6115 top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
6116 top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
6117
6118 if (top_smax_value != top_smin_value)
6119 goto out;
6120
6121 /* find the s64_min and s64_min after sign extension */
6122 if (size == 1) {
6123 init_s64_max = (s8)reg->smax_value;
6124 init_s64_min = (s8)reg->smin_value;
6125 } else if (size == 2) {
6126 init_s64_max = (s16)reg->smax_value;
6127 init_s64_min = (s16)reg->smin_value;
6128 } else {
6129 init_s64_max = (s32)reg->smax_value;
6130 init_s64_min = (s32)reg->smin_value;
6131 }
6132
6133 s64_max = max(init_s64_max, init_s64_min);
6134 s64_min = min(init_s64_max, init_s64_min);
6135
6136 /* both of s64_max/s64_min positive or negative */
6137 if ((s64_max >= 0) == (s64_min >= 0)) {
6138 reg->smin_value = reg->s32_min_value = s64_min;
6139 reg->smax_value = reg->s32_max_value = s64_max;
6140 reg->umin_value = reg->u32_min_value = s64_min;
6141 reg->umax_value = reg->u32_max_value = s64_max;
6142 reg->var_off = tnum_range(s64_min, s64_max);
6143 return;
6144 }
6145
6146 out:
6147 set_sext64_default_val(reg, size);
6148 }
6149
set_sext32_default_val(struct bpf_reg_state * reg,int size)6150 static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
6151 {
6152 if (size == 1) {
6153 reg->s32_min_value = S8_MIN;
6154 reg->s32_max_value = S8_MAX;
6155 } else {
6156 /* size == 2 */
6157 reg->s32_min_value = S16_MIN;
6158 reg->s32_max_value = S16_MAX;
6159 }
6160 reg->u32_min_value = 0;
6161 reg->u32_max_value = U32_MAX;
6162 }
6163
coerce_subreg_to_size_sx(struct bpf_reg_state * reg,int size)6164 static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
6165 {
6166 s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
6167 u32 top_smax_value, top_smin_value;
6168 u32 num_bits = size * 8;
6169
6170 if (tnum_is_const(reg->var_off)) {
6171 u32_val = reg->var_off.value;
6172 if (size == 1)
6173 reg->var_off = tnum_const((s8)u32_val);
6174 else
6175 reg->var_off = tnum_const((s16)u32_val);
6176
6177 u32_val = reg->var_off.value;
6178 reg->s32_min_value = reg->s32_max_value = u32_val;
6179 reg->u32_min_value = reg->u32_max_value = u32_val;
6180 return;
6181 }
6182
6183 top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
6184 top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
6185
6186 if (top_smax_value != top_smin_value)
6187 goto out;
6188
6189 /* find the s32_min and s32_min after sign extension */
6190 if (size == 1) {
6191 init_s32_max = (s8)reg->s32_max_value;
6192 init_s32_min = (s8)reg->s32_min_value;
6193 } else {
6194 /* size == 2 */
6195 init_s32_max = (s16)reg->s32_max_value;
6196 init_s32_min = (s16)reg->s32_min_value;
6197 }
6198 s32_max = max(init_s32_max, init_s32_min);
6199 s32_min = min(init_s32_max, init_s32_min);
6200
6201 if ((s32_min >= 0) == (s32_max >= 0)) {
6202 reg->s32_min_value = s32_min;
6203 reg->s32_max_value = s32_max;
6204 reg->u32_min_value = (u32)s32_min;
6205 reg->u32_max_value = (u32)s32_max;
6206 return;
6207 }
6208
6209 out:
6210 set_sext32_default_val(reg, size);
6211 }
6212
bpf_map_is_rdonly(const struct bpf_map * map)6213 static bool bpf_map_is_rdonly(const struct bpf_map *map)
6214 {
6215 /* A map is considered read-only if the following condition are true:
6216 *
6217 * 1) BPF program side cannot change any of the map content. The
6218 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
6219 * and was set at map creation time.
6220 * 2) The map value(s) have been initialized from user space by a
6221 * loader and then "frozen", such that no new map update/delete
6222 * operations from syscall side are possible for the rest of
6223 * the map's lifetime from that point onwards.
6224 * 3) Any parallel/pending map update/delete operations from syscall
6225 * side have been completed. Only after that point, it's safe to
6226 * assume that map value(s) are immutable.
6227 */
6228 return (map->map_flags & BPF_F_RDONLY_PROG) &&
6229 READ_ONCE(map->frozen) &&
6230 !bpf_map_write_active(map);
6231 }
6232
bpf_map_direct_read(struct bpf_map * map,int off,int size,u64 * val,bool is_ldsx)6233 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
6234 bool is_ldsx)
6235 {
6236 void *ptr;
6237 u64 addr;
6238 int err;
6239
6240 err = map->ops->map_direct_value_addr(map, &addr, off);
6241 if (err)
6242 return err;
6243 ptr = (void *)(long)addr + off;
6244
6245 switch (size) {
6246 case sizeof(u8):
6247 *val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
6248 break;
6249 case sizeof(u16):
6250 *val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
6251 break;
6252 case sizeof(u32):
6253 *val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
6254 break;
6255 case sizeof(u64):
6256 *val = *(u64 *)ptr;
6257 break;
6258 default:
6259 return -EINVAL;
6260 }
6261 return 0;
6262 }
6263
6264 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
6265 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
6266 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
6267 #define BTF_TYPE_SAFE_TRUSTED_OR_NULL(__type) __PASTE(__type, __safe_trusted_or_null)
6268
6269 /*
6270 * Allow list few fields as RCU trusted or full trusted.
6271 * This logic doesn't allow mix tagging and will be removed once GCC supports
6272 * btf_type_tag.
6273 */
6274
6275 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
BTF_TYPE_SAFE_RCU(struct task_struct)6276 BTF_TYPE_SAFE_RCU(struct task_struct) {
6277 const cpumask_t *cpus_ptr;
6278 struct css_set __rcu *cgroups;
6279 struct task_struct __rcu *real_parent;
6280 struct task_struct *group_leader;
6281 };
6282
BTF_TYPE_SAFE_RCU(struct cgroup)6283 BTF_TYPE_SAFE_RCU(struct cgroup) {
6284 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
6285 struct kernfs_node *kn;
6286 };
6287
BTF_TYPE_SAFE_RCU(struct css_set)6288 BTF_TYPE_SAFE_RCU(struct css_set) {
6289 struct cgroup *dfl_cgrp;
6290 };
6291
6292 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct)6293 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
6294 struct file __rcu *exe_file;
6295 };
6296
6297 /* skb->sk, req->sk are not RCU protected, but we mark them as such
6298 * because bpf prog accessible sockets are SOCK_RCU_FREE.
6299 */
BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff)6300 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
6301 struct sock *sk;
6302 };
6303
BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock)6304 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
6305 struct sock *sk;
6306 };
6307
6308 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta)6309 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
6310 struct seq_file *seq;
6311 };
6312
BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task)6313 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
6314 struct bpf_iter_meta *meta;
6315 struct task_struct *task;
6316 };
6317
BTF_TYPE_SAFE_TRUSTED(struct linux_binprm)6318 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
6319 struct file *file;
6320 };
6321
BTF_TYPE_SAFE_TRUSTED(struct file)6322 BTF_TYPE_SAFE_TRUSTED(struct file) {
6323 struct inode *f_inode;
6324 };
6325
BTF_TYPE_SAFE_TRUSTED(struct dentry)6326 BTF_TYPE_SAFE_TRUSTED(struct dentry) {
6327 /* no negative dentry-s in places where bpf can see it */
6328 struct inode *d_inode;
6329 };
6330
BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket)6331 BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket) {
6332 struct sock *sk;
6333 };
6334
type_is_rcu(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6335 static bool type_is_rcu(struct bpf_verifier_env *env,
6336 struct bpf_reg_state *reg,
6337 const char *field_name, u32 btf_id)
6338 {
6339 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
6340 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
6341 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
6342
6343 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
6344 }
6345
type_is_rcu_or_null(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6346 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
6347 struct bpf_reg_state *reg,
6348 const char *field_name, u32 btf_id)
6349 {
6350 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
6351 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
6352 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
6353
6354 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
6355 }
6356
type_is_trusted(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6357 static bool type_is_trusted(struct bpf_verifier_env *env,
6358 struct bpf_reg_state *reg,
6359 const char *field_name, u32 btf_id)
6360 {
6361 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
6362 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
6363 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
6364 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
6365 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry));
6366
6367 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
6368 }
6369
type_is_trusted_or_null(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)6370 static bool type_is_trusted_or_null(struct bpf_verifier_env *env,
6371 struct bpf_reg_state *reg,
6372 const char *field_name, u32 btf_id)
6373 {
6374 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket));
6375
6376 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id,
6377 "__safe_trusted_or_null");
6378 }
6379
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)6380 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
6381 struct bpf_reg_state *regs,
6382 int regno, int off, int size,
6383 enum bpf_access_type atype,
6384 int value_regno)
6385 {
6386 struct bpf_reg_state *reg = regs + regno;
6387 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
6388 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
6389 const char *field_name = NULL;
6390 enum bpf_type_flag flag = 0;
6391 u32 btf_id = 0;
6392 int ret;
6393
6394 if (!env->allow_ptr_leaks) {
6395 verbose(env,
6396 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6397 tname);
6398 return -EPERM;
6399 }
6400 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
6401 verbose(env,
6402 "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
6403 tname);
6404 return -EINVAL;
6405 }
6406 if (off < 0) {
6407 verbose(env,
6408 "R%d is ptr_%s invalid negative access: off=%d\n",
6409 regno, tname, off);
6410 return -EACCES;
6411 }
6412 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6413 char tn_buf[48];
6414
6415 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6416 verbose(env,
6417 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
6418 regno, tname, off, tn_buf);
6419 return -EACCES;
6420 }
6421
6422 if (reg->type & MEM_USER) {
6423 verbose(env,
6424 "R%d is ptr_%s access user memory: off=%d\n",
6425 regno, tname, off);
6426 return -EACCES;
6427 }
6428
6429 if (reg->type & MEM_PERCPU) {
6430 verbose(env,
6431 "R%d is ptr_%s access percpu memory: off=%d\n",
6432 regno, tname, off);
6433 return -EACCES;
6434 }
6435
6436 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
6437 if (!btf_is_kernel(reg->btf)) {
6438 verbose(env, "verifier internal error: reg->btf must be kernel btf\n");
6439 return -EFAULT;
6440 }
6441 ret = env->ops->btf_struct_access(&env->log, reg, off, size);
6442 } else {
6443 /* Writes are permitted with default btf_struct_access for
6444 * program allocated objects (which always have ref_obj_id > 0),
6445 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
6446 */
6447 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
6448 verbose(env, "only read is supported\n");
6449 return -EACCES;
6450 }
6451
6452 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
6453 !reg->ref_obj_id) {
6454 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n");
6455 return -EFAULT;
6456 }
6457
6458 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
6459 }
6460
6461 if (ret < 0)
6462 return ret;
6463
6464 if (ret != PTR_TO_BTF_ID) {
6465 /* just mark; */
6466
6467 } else if (type_flag(reg->type) & PTR_UNTRUSTED) {
6468 /* If this is an untrusted pointer, all pointers formed by walking it
6469 * also inherit the untrusted flag.
6470 */
6471 flag = PTR_UNTRUSTED;
6472
6473 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
6474 /* By default any pointer obtained from walking a trusted pointer is no
6475 * longer trusted, unless the field being accessed has explicitly been
6476 * marked as inheriting its parent's state of trust (either full or RCU).
6477 * For example:
6478 * 'cgroups' pointer is untrusted if task->cgroups dereference
6479 * happened in a sleepable program outside of bpf_rcu_read_lock()
6480 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
6481 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
6482 *
6483 * A regular RCU-protected pointer with __rcu tag can also be deemed
6484 * trusted if we are in an RCU CS. Such pointer can be NULL.
6485 */
6486 if (type_is_trusted(env, reg, field_name, btf_id)) {
6487 flag |= PTR_TRUSTED;
6488 } else if (type_is_trusted_or_null(env, reg, field_name, btf_id)) {
6489 flag |= PTR_TRUSTED | PTR_MAYBE_NULL;
6490 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
6491 if (type_is_rcu(env, reg, field_name, btf_id)) {
6492 /* ignore __rcu tag and mark it MEM_RCU */
6493 flag |= MEM_RCU;
6494 } else if (flag & MEM_RCU ||
6495 type_is_rcu_or_null(env, reg, field_name, btf_id)) {
6496 /* __rcu tagged pointers can be NULL */
6497 flag |= MEM_RCU | PTR_MAYBE_NULL;
6498
6499 /* We always trust them */
6500 if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
6501 flag & PTR_UNTRUSTED)
6502 flag &= ~PTR_UNTRUSTED;
6503 } else if (flag & (MEM_PERCPU | MEM_USER)) {
6504 /* keep as-is */
6505 } else {
6506 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
6507 clear_trusted_flags(&flag);
6508 }
6509 } else {
6510 /*
6511 * If not in RCU CS or MEM_RCU pointer can be NULL then
6512 * aggressively mark as untrusted otherwise such
6513 * pointers will be plain PTR_TO_BTF_ID without flags
6514 * and will be allowed to be passed into helpers for
6515 * compat reasons.
6516 */
6517 flag = PTR_UNTRUSTED;
6518 }
6519 } else {
6520 /* Old compat. Deprecated */
6521 clear_trusted_flags(&flag);
6522 }
6523
6524 if (atype == BPF_READ && value_regno >= 0)
6525 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
6526
6527 return 0;
6528 }
6529
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)6530 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
6531 struct bpf_reg_state *regs,
6532 int regno, int off, int size,
6533 enum bpf_access_type atype,
6534 int value_regno)
6535 {
6536 struct bpf_reg_state *reg = regs + regno;
6537 struct bpf_map *map = reg->map_ptr;
6538 struct bpf_reg_state map_reg;
6539 enum bpf_type_flag flag = 0;
6540 const struct btf_type *t;
6541 const char *tname;
6542 u32 btf_id;
6543 int ret;
6544
6545 if (!btf_vmlinux) {
6546 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
6547 return -ENOTSUPP;
6548 }
6549
6550 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
6551 verbose(env, "map_ptr access not supported for map type %d\n",
6552 map->map_type);
6553 return -ENOTSUPP;
6554 }
6555
6556 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
6557 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
6558
6559 if (!env->allow_ptr_leaks) {
6560 verbose(env,
6561 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
6562 tname);
6563 return -EPERM;
6564 }
6565
6566 if (off < 0) {
6567 verbose(env, "R%d is %s invalid negative access: off=%d\n",
6568 regno, tname, off);
6569 return -EACCES;
6570 }
6571
6572 if (atype != BPF_READ) {
6573 verbose(env, "only read from %s is supported\n", tname);
6574 return -EACCES;
6575 }
6576
6577 /* Simulate access to a PTR_TO_BTF_ID */
6578 memset(&map_reg, 0, sizeof(map_reg));
6579 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0);
6580 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
6581 if (ret < 0)
6582 return ret;
6583
6584 if (value_regno >= 0)
6585 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
6586
6587 return 0;
6588 }
6589
6590 /* Check that the stack access at the given offset is within bounds. The
6591 * maximum valid offset is -1.
6592 *
6593 * The minimum valid offset is -MAX_BPF_STACK for writes, and
6594 * -state->allocated_stack for reads.
6595 */
check_stack_slot_within_bounds(struct bpf_verifier_env * env,s64 off,struct bpf_func_state * state,enum bpf_access_type t)6596 static int check_stack_slot_within_bounds(struct bpf_verifier_env *env,
6597 s64 off,
6598 struct bpf_func_state *state,
6599 enum bpf_access_type t)
6600 {
6601 int min_valid_off;
6602
6603 if (t == BPF_WRITE || env->allow_uninit_stack)
6604 min_valid_off = -MAX_BPF_STACK;
6605 else
6606 min_valid_off = -state->allocated_stack;
6607
6608 if (off < min_valid_off || off > -1)
6609 return -EACCES;
6610 return 0;
6611 }
6612
6613 /* Check that the stack access at 'regno + off' falls within the maximum stack
6614 * bounds.
6615 *
6616 * 'off' includes `regno->offset`, but not its dynamic part (if any).
6617 */
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)6618 static int check_stack_access_within_bounds(
6619 struct bpf_verifier_env *env,
6620 int regno, int off, int access_size,
6621 enum bpf_access_src src, enum bpf_access_type type)
6622 {
6623 struct bpf_reg_state *regs = cur_regs(env);
6624 struct bpf_reg_state *reg = regs + regno;
6625 struct bpf_func_state *state = func(env, reg);
6626 s64 min_off, max_off;
6627 int err;
6628 char *err_extra;
6629
6630 if (src == ACCESS_HELPER)
6631 /* We don't know if helpers are reading or writing (or both). */
6632 err_extra = " indirect access to";
6633 else if (type == BPF_READ)
6634 err_extra = " read from";
6635 else
6636 err_extra = " write to";
6637
6638 if (tnum_is_const(reg->var_off)) {
6639 min_off = (s64)reg->var_off.value + off;
6640 max_off = min_off + access_size;
6641 } else {
6642 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
6643 reg->smin_value <= -BPF_MAX_VAR_OFF) {
6644 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
6645 err_extra, regno);
6646 return -EACCES;
6647 }
6648 min_off = reg->smin_value + off;
6649 max_off = reg->smax_value + off + access_size;
6650 }
6651
6652 err = check_stack_slot_within_bounds(env, min_off, state, type);
6653 if (!err && max_off > 0)
6654 err = -EINVAL; /* out of stack access into non-negative offsets */
6655 if (!err && access_size < 0)
6656 /* access_size should not be negative (or overflow an int); others checks
6657 * along the way should have prevented such an access.
6658 */
6659 err = -EFAULT; /* invalid negative access size; integer overflow? */
6660
6661 if (err) {
6662 if (tnum_is_const(reg->var_off)) {
6663 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
6664 err_extra, regno, off, access_size);
6665 } else {
6666 char tn_buf[48];
6667
6668 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6669 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
6670 err_extra, regno, tn_buf, access_size);
6671 }
6672 return err;
6673 }
6674
6675 return grow_stack_state(env, state, round_up(-min_off, BPF_REG_SIZE));
6676 }
6677
6678 /* check whether memory at (regno + off) is accessible for t = (read | write)
6679 * if t==write, value_regno is a register which value is stored into memory
6680 * if t==read, value_regno is a register which will receive the value from memory
6681 * if t==write && value_regno==-1, some unknown value is stored into memory
6682 * if t==read && value_regno==-1, don't care what we read from memory
6683 */
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)6684 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
6685 int off, int bpf_size, enum bpf_access_type t,
6686 int value_regno, bool strict_alignment_once, bool is_ldsx)
6687 {
6688 struct bpf_reg_state *regs = cur_regs(env);
6689 struct bpf_reg_state *reg = regs + regno;
6690 int size, err = 0;
6691
6692 size = bpf_size_to_bytes(bpf_size);
6693 if (size < 0)
6694 return size;
6695
6696 /* alignment checks will add in reg->off themselves */
6697 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
6698 if (err)
6699 return err;
6700
6701 /* for access checks, reg->off is just part of off */
6702 off += reg->off;
6703
6704 if (reg->type == PTR_TO_MAP_KEY) {
6705 if (t == BPF_WRITE) {
6706 verbose(env, "write to change key R%d not allowed\n", regno);
6707 return -EACCES;
6708 }
6709
6710 err = check_mem_region_access(env, regno, off, size,
6711 reg->map_ptr->key_size, false);
6712 if (err)
6713 return err;
6714 if (value_regno >= 0)
6715 mark_reg_unknown(env, regs, value_regno);
6716 } else if (reg->type == PTR_TO_MAP_VALUE) {
6717 struct btf_field *kptr_field = NULL;
6718
6719 if (t == BPF_WRITE && value_regno >= 0 &&
6720 is_pointer_value(env, value_regno)) {
6721 verbose(env, "R%d leaks addr into map\n", value_regno);
6722 return -EACCES;
6723 }
6724 err = check_map_access_type(env, regno, off, size, t);
6725 if (err)
6726 return err;
6727 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
6728 if (err)
6729 return err;
6730 if (tnum_is_const(reg->var_off))
6731 kptr_field = btf_record_find(reg->map_ptr->record,
6732 off + reg->var_off.value, BPF_KPTR);
6733 if (kptr_field) {
6734 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
6735 } else if (t == BPF_READ && value_regno >= 0) {
6736 struct bpf_map *map = reg->map_ptr;
6737
6738 /* if map is read-only, track its contents as scalars */
6739 if (tnum_is_const(reg->var_off) &&
6740 bpf_map_is_rdonly(map) &&
6741 map->ops->map_direct_value_addr) {
6742 int map_off = off + reg->var_off.value;
6743 u64 val = 0;
6744
6745 err = bpf_map_direct_read(map, map_off, size,
6746 &val, is_ldsx);
6747 if (err)
6748 return err;
6749
6750 regs[value_regno].type = SCALAR_VALUE;
6751 __mark_reg_known(®s[value_regno], val);
6752 } else {
6753 mark_reg_unknown(env, regs, value_regno);
6754 }
6755 }
6756 } else if (base_type(reg->type) == PTR_TO_MEM) {
6757 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6758
6759 if (type_may_be_null(reg->type)) {
6760 verbose(env, "R%d invalid mem access '%s'\n", regno,
6761 reg_type_str(env, reg->type));
6762 return -EACCES;
6763 }
6764
6765 if (t == BPF_WRITE && rdonly_mem) {
6766 verbose(env, "R%d cannot write into %s\n",
6767 regno, reg_type_str(env, reg->type));
6768 return -EACCES;
6769 }
6770
6771 if (t == BPF_WRITE && value_regno >= 0 &&
6772 is_pointer_value(env, value_regno)) {
6773 verbose(env, "R%d leaks addr into mem\n", value_regno);
6774 return -EACCES;
6775 }
6776
6777 err = check_mem_region_access(env, regno, off, size,
6778 reg->mem_size, false);
6779 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
6780 mark_reg_unknown(env, regs, value_regno);
6781 } else if (reg->type == PTR_TO_CTX) {
6782 enum bpf_reg_type reg_type = SCALAR_VALUE;
6783 struct btf *btf = NULL;
6784 u32 btf_id = 0;
6785
6786 if (t == BPF_WRITE && value_regno >= 0 &&
6787 is_pointer_value(env, value_regno)) {
6788 verbose(env, "R%d leaks addr into ctx\n", value_regno);
6789 return -EACCES;
6790 }
6791
6792 err = check_ptr_off_reg(env, reg, regno);
6793 if (err < 0)
6794 return err;
6795
6796 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf,
6797 &btf_id);
6798 if (err)
6799 verbose_linfo(env, insn_idx, "; ");
6800 if (!err && t == BPF_READ && value_regno >= 0) {
6801 /* ctx access returns either a scalar, or a
6802 * PTR_TO_PACKET[_META,_END]. In the latter
6803 * case, we know the offset is zero.
6804 */
6805 if (reg_type == SCALAR_VALUE) {
6806 mark_reg_unknown(env, regs, value_regno);
6807 } else {
6808 mark_reg_known_zero(env, regs,
6809 value_regno);
6810 if (type_may_be_null(reg_type))
6811 regs[value_regno].id = ++env->id_gen;
6812 /* A load of ctx field could have different
6813 * actual load size with the one encoded in the
6814 * insn. When the dst is PTR, it is for sure not
6815 * a sub-register.
6816 */
6817 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
6818 if (base_type(reg_type) == PTR_TO_BTF_ID) {
6819 regs[value_regno].btf = btf;
6820 regs[value_regno].btf_id = btf_id;
6821 }
6822 }
6823 regs[value_regno].type = reg_type;
6824 }
6825
6826 } else if (reg->type == PTR_TO_STACK) {
6827 /* Basic bounds checks. */
6828 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
6829 if (err)
6830 return err;
6831
6832 if (t == BPF_READ)
6833 err = check_stack_read(env, regno, off, size,
6834 value_regno);
6835 else
6836 err = check_stack_write(env, regno, off, size,
6837 value_regno, insn_idx);
6838 } else if (reg_is_pkt_pointer(reg)) {
6839 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
6840 verbose(env, "cannot write into packet\n");
6841 return -EACCES;
6842 }
6843 if (t == BPF_WRITE && value_regno >= 0 &&
6844 is_pointer_value(env, value_regno)) {
6845 verbose(env, "R%d leaks addr into packet\n",
6846 value_regno);
6847 return -EACCES;
6848 }
6849 err = check_packet_access(env, regno, off, size, false);
6850 if (!err && t == BPF_READ && value_regno >= 0)
6851 mark_reg_unknown(env, regs, value_regno);
6852 } else if (reg->type == PTR_TO_FLOW_KEYS) {
6853 if (t == BPF_WRITE && value_regno >= 0 &&
6854 is_pointer_value(env, value_regno)) {
6855 verbose(env, "R%d leaks addr into flow keys\n",
6856 value_regno);
6857 return -EACCES;
6858 }
6859
6860 err = check_flow_keys_access(env, off, size);
6861 if (!err && t == BPF_READ && value_regno >= 0)
6862 mark_reg_unknown(env, regs, value_regno);
6863 } else if (type_is_sk_pointer(reg->type)) {
6864 if (t == BPF_WRITE) {
6865 verbose(env, "R%d cannot write into %s\n",
6866 regno, reg_type_str(env, reg->type));
6867 return -EACCES;
6868 }
6869 err = check_sock_access(env, insn_idx, regno, off, size, t);
6870 if (!err && value_regno >= 0)
6871 mark_reg_unknown(env, regs, value_regno);
6872 } else if (reg->type == PTR_TO_TP_BUFFER) {
6873 err = check_tp_buffer_access(env, reg, regno, off, size);
6874 if (!err && t == BPF_READ && value_regno >= 0)
6875 mark_reg_unknown(env, regs, value_regno);
6876 } else if (base_type(reg->type) == PTR_TO_BTF_ID &&
6877 !type_may_be_null(reg->type)) {
6878 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
6879 value_regno);
6880 } else if (reg->type == CONST_PTR_TO_MAP) {
6881 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
6882 value_regno);
6883 } else if (base_type(reg->type) == PTR_TO_BUF) {
6884 bool rdonly_mem = type_is_rdonly_mem(reg->type);
6885 u32 *max_access;
6886
6887 if (rdonly_mem) {
6888 if (t == BPF_WRITE) {
6889 verbose(env, "R%d cannot write into %s\n",
6890 regno, reg_type_str(env, reg->type));
6891 return -EACCES;
6892 }
6893 max_access = &env->prog->aux->max_rdonly_access;
6894 } else {
6895 max_access = &env->prog->aux->max_rdwr_access;
6896 }
6897
6898 err = check_buffer_access(env, reg, regno, off, size, false,
6899 max_access);
6900
6901 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
6902 mark_reg_unknown(env, regs, value_regno);
6903 } else {
6904 verbose(env, "R%d invalid mem access '%s'\n", regno,
6905 reg_type_str(env, reg->type));
6906 return -EACCES;
6907 }
6908
6909 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
6910 regs[value_regno].type == SCALAR_VALUE) {
6911 if (!is_ldsx)
6912 /* b/h/w load zero-extends, mark upper bits as known 0 */
6913 coerce_reg_to_size(®s[value_regno], size);
6914 else
6915 coerce_reg_to_size_sx(®s[value_regno], size);
6916 }
6917 return err;
6918 }
6919
check_atomic(struct bpf_verifier_env * env,int insn_idx,struct bpf_insn * insn)6920 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
6921 {
6922 int load_reg;
6923 int err;
6924
6925 switch (insn->imm) {
6926 case BPF_ADD:
6927 case BPF_ADD | BPF_FETCH:
6928 case BPF_AND:
6929 case BPF_AND | BPF_FETCH:
6930 case BPF_OR:
6931 case BPF_OR | BPF_FETCH:
6932 case BPF_XOR:
6933 case BPF_XOR | BPF_FETCH:
6934 case BPF_XCHG:
6935 case BPF_CMPXCHG:
6936 break;
6937 default:
6938 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
6939 return -EINVAL;
6940 }
6941
6942 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
6943 verbose(env, "invalid atomic operand size\n");
6944 return -EINVAL;
6945 }
6946
6947 /* check src1 operand */
6948 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6949 if (err)
6950 return err;
6951
6952 /* check src2 operand */
6953 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6954 if (err)
6955 return err;
6956
6957 if (insn->imm == BPF_CMPXCHG) {
6958 /* Check comparison of R0 with memory location */
6959 const u32 aux_reg = BPF_REG_0;
6960
6961 err = check_reg_arg(env, aux_reg, SRC_OP);
6962 if (err)
6963 return err;
6964
6965 if (is_pointer_value(env, aux_reg)) {
6966 verbose(env, "R%d leaks addr into mem\n", aux_reg);
6967 return -EACCES;
6968 }
6969 }
6970
6971 if (is_pointer_value(env, insn->src_reg)) {
6972 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
6973 return -EACCES;
6974 }
6975
6976 if (is_ctx_reg(env, insn->dst_reg) ||
6977 is_pkt_reg(env, insn->dst_reg) ||
6978 is_flow_key_reg(env, insn->dst_reg) ||
6979 is_sk_reg(env, insn->dst_reg)) {
6980 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
6981 insn->dst_reg,
6982 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
6983 return -EACCES;
6984 }
6985
6986 if (insn->imm & BPF_FETCH) {
6987 if (insn->imm == BPF_CMPXCHG)
6988 load_reg = BPF_REG_0;
6989 else
6990 load_reg = insn->src_reg;
6991
6992 /* check and record load of old value */
6993 err = check_reg_arg(env, load_reg, DST_OP);
6994 if (err)
6995 return err;
6996 } else {
6997 /* This instruction accesses a memory location but doesn't
6998 * actually load it into a register.
6999 */
7000 load_reg = -1;
7001 }
7002
7003 /* Check whether we can read the memory, with second call for fetch
7004 * case to simulate the register fill.
7005 */
7006 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
7007 BPF_SIZE(insn->code), BPF_READ, -1, true, false);
7008 if (!err && load_reg >= 0)
7009 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
7010 BPF_SIZE(insn->code), BPF_READ, load_reg,
7011 true, false);
7012 if (err)
7013 return err;
7014
7015 /* Check whether we can write into the same memory. */
7016 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
7017 BPF_SIZE(insn->code), BPF_WRITE, -1, true, false);
7018 if (err)
7019 return err;
7020
7021 return 0;
7022 }
7023
7024 /* When register 'regno' is used to read the stack (either directly or through
7025 * a helper function) make sure that it's within stack boundary and, depending
7026 * on the access type and privileges, that all elements of the stack are
7027 * initialized.
7028 *
7029 * 'off' includes 'regno->off', but not its dynamic part (if any).
7030 *
7031 * All registers that have been spilled on the stack in the slots within the
7032 * read offsets are marked as read.
7033 */
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)7034 static int check_stack_range_initialized(
7035 struct bpf_verifier_env *env, int regno, int off,
7036 int access_size, bool zero_size_allowed,
7037 enum bpf_access_src type, struct bpf_call_arg_meta *meta)
7038 {
7039 struct bpf_reg_state *reg = reg_state(env, regno);
7040 struct bpf_func_state *state = func(env, reg);
7041 int err, min_off, max_off, i, j, slot, spi;
7042 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
7043 enum bpf_access_type bounds_check_type;
7044 /* Some accesses can write anything into the stack, others are
7045 * read-only.
7046 */
7047 bool clobber = false;
7048
7049 if (access_size == 0 && !zero_size_allowed) {
7050 verbose(env, "invalid zero-sized read\n");
7051 return -EACCES;
7052 }
7053
7054 if (type == ACCESS_HELPER) {
7055 /* The bounds checks for writes are more permissive than for
7056 * reads. However, if raw_mode is not set, we'll do extra
7057 * checks below.
7058 */
7059 bounds_check_type = BPF_WRITE;
7060 clobber = true;
7061 } else {
7062 bounds_check_type = BPF_READ;
7063 }
7064 err = check_stack_access_within_bounds(env, regno, off, access_size,
7065 type, bounds_check_type);
7066 if (err)
7067 return err;
7068
7069
7070 if (tnum_is_const(reg->var_off)) {
7071 min_off = max_off = reg->var_off.value + off;
7072 } else {
7073 /* Variable offset is prohibited for unprivileged mode for
7074 * simplicity since it requires corresponding support in
7075 * Spectre masking for stack ALU.
7076 * See also retrieve_ptr_limit().
7077 */
7078 if (!env->bypass_spec_v1) {
7079 char tn_buf[48];
7080
7081 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7082 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
7083 regno, err_extra, tn_buf);
7084 return -EACCES;
7085 }
7086 /* Only initialized buffer on stack is allowed to be accessed
7087 * with variable offset. With uninitialized buffer it's hard to
7088 * guarantee that whole memory is marked as initialized on
7089 * helper return since specific bounds are unknown what may
7090 * cause uninitialized stack leaking.
7091 */
7092 if (meta && meta->raw_mode)
7093 meta = NULL;
7094
7095 min_off = reg->smin_value + off;
7096 max_off = reg->smax_value + off;
7097 }
7098
7099 if (meta && meta->raw_mode) {
7100 /* Ensure we won't be overwriting dynptrs when simulating byte
7101 * by byte access in check_helper_call using meta.access_size.
7102 * This would be a problem if we have a helper in the future
7103 * which takes:
7104 *
7105 * helper(uninit_mem, len, dynptr)
7106 *
7107 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
7108 * may end up writing to dynptr itself when touching memory from
7109 * arg 1. This can be relaxed on a case by case basis for known
7110 * safe cases, but reject due to the possibilitiy of aliasing by
7111 * default.
7112 */
7113 for (i = min_off; i < max_off + access_size; i++) {
7114 int stack_off = -i - 1;
7115
7116 spi = __get_spi(i);
7117 /* raw_mode may write past allocated_stack */
7118 if (state->allocated_stack <= stack_off)
7119 continue;
7120 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
7121 verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
7122 return -EACCES;
7123 }
7124 }
7125 meta->access_size = access_size;
7126 meta->regno = regno;
7127 return 0;
7128 }
7129
7130 for (i = min_off; i < max_off + access_size; i++) {
7131 u8 *stype;
7132
7133 slot = -i - 1;
7134 spi = slot / BPF_REG_SIZE;
7135 if (state->allocated_stack <= slot) {
7136 verbose(env, "verifier bug: allocated_stack too small");
7137 return -EFAULT;
7138 }
7139
7140 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
7141 if (*stype == STACK_MISC)
7142 goto mark;
7143 if ((*stype == STACK_ZERO) ||
7144 (*stype == STACK_INVALID && env->allow_uninit_stack)) {
7145 if (clobber) {
7146 /* helper can write anything into the stack */
7147 *stype = STACK_MISC;
7148 }
7149 goto mark;
7150 }
7151
7152 if (is_spilled_reg(&state->stack[spi]) &&
7153 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
7154 env->allow_ptr_leaks)) {
7155 if (clobber) {
7156 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
7157 for (j = 0; j < BPF_REG_SIZE; j++)
7158 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
7159 }
7160 goto mark;
7161 }
7162
7163 if (tnum_is_const(reg->var_off)) {
7164 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
7165 err_extra, regno, min_off, i - min_off, access_size);
7166 } else {
7167 char tn_buf[48];
7168
7169 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7170 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
7171 err_extra, regno, tn_buf, i - min_off, access_size);
7172 }
7173 return -EACCES;
7174 mark:
7175 /* reading any byte out of 8-byte 'spill_slot' will cause
7176 * the whole slot to be marked as 'read'
7177 */
7178 mark_reg_read(env, &state->stack[spi].spilled_ptr,
7179 state->stack[spi].spilled_ptr.parent,
7180 REG_LIVE_READ64);
7181 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
7182 * be sure that whether stack slot is written to or not. Hence,
7183 * we must still conservatively propagate reads upwards even if
7184 * helper may write to the entire memory range.
7185 */
7186 }
7187 return 0;
7188 }
7189
check_helper_mem_access(struct bpf_verifier_env * env,int regno,int access_size,bool zero_size_allowed,struct bpf_call_arg_meta * meta)7190 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
7191 int access_size, bool zero_size_allowed,
7192 struct bpf_call_arg_meta *meta)
7193 {
7194 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7195 u32 *max_access;
7196
7197 switch (base_type(reg->type)) {
7198 case PTR_TO_PACKET:
7199 case PTR_TO_PACKET_META:
7200 return check_packet_access(env, regno, reg->off, access_size,
7201 zero_size_allowed);
7202 case PTR_TO_MAP_KEY:
7203 if (meta && meta->raw_mode) {
7204 verbose(env, "R%d cannot write into %s\n", regno,
7205 reg_type_str(env, reg->type));
7206 return -EACCES;
7207 }
7208 return check_mem_region_access(env, regno, reg->off, access_size,
7209 reg->map_ptr->key_size, false);
7210 case PTR_TO_MAP_VALUE:
7211 if (check_map_access_type(env, regno, reg->off, access_size,
7212 meta && meta->raw_mode ? BPF_WRITE :
7213 BPF_READ))
7214 return -EACCES;
7215 return check_map_access(env, regno, reg->off, access_size,
7216 zero_size_allowed, ACCESS_HELPER);
7217 case PTR_TO_MEM:
7218 if (type_is_rdonly_mem(reg->type)) {
7219 if (meta && meta->raw_mode) {
7220 verbose(env, "R%d cannot write into %s\n", regno,
7221 reg_type_str(env, reg->type));
7222 return -EACCES;
7223 }
7224 }
7225 return check_mem_region_access(env, regno, reg->off,
7226 access_size, reg->mem_size,
7227 zero_size_allowed);
7228 case PTR_TO_BUF:
7229 if (type_is_rdonly_mem(reg->type)) {
7230 if (meta && meta->raw_mode) {
7231 verbose(env, "R%d cannot write into %s\n", regno,
7232 reg_type_str(env, reg->type));
7233 return -EACCES;
7234 }
7235
7236 max_access = &env->prog->aux->max_rdonly_access;
7237 } else {
7238 max_access = &env->prog->aux->max_rdwr_access;
7239 }
7240 return check_buffer_access(env, reg, regno, reg->off,
7241 access_size, zero_size_allowed,
7242 max_access);
7243 case PTR_TO_STACK:
7244 return check_stack_range_initialized(
7245 env,
7246 regno, reg->off, access_size,
7247 zero_size_allowed, ACCESS_HELPER, meta);
7248 case PTR_TO_BTF_ID:
7249 return check_ptr_to_btf_access(env, regs, regno, reg->off,
7250 access_size, BPF_READ, -1);
7251 case PTR_TO_CTX:
7252 /* in case the function doesn't know how to access the context,
7253 * (because we are in a program of type SYSCALL for example), we
7254 * can not statically check its size.
7255 * Dynamically check it now.
7256 */
7257 if (!env->ops->convert_ctx_access) {
7258 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ;
7259 int offset = access_size - 1;
7260
7261 /* Allow zero-byte read from PTR_TO_CTX */
7262 if (access_size == 0)
7263 return zero_size_allowed ? 0 : -EACCES;
7264
7265 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
7266 atype, -1, false, false);
7267 }
7268
7269 fallthrough;
7270 default: /* scalar_value or invalid ptr */
7271 /* Allow zero-byte read from NULL, regardless of pointer type */
7272 if (zero_size_allowed && access_size == 0 &&
7273 register_is_null(reg))
7274 return 0;
7275
7276 verbose(env, "R%d type=%s ", regno,
7277 reg_type_str(env, reg->type));
7278 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
7279 return -EACCES;
7280 }
7281 }
7282
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)7283 static int check_mem_size_reg(struct bpf_verifier_env *env,
7284 struct bpf_reg_state *reg, u32 regno,
7285 bool zero_size_allowed,
7286 struct bpf_call_arg_meta *meta)
7287 {
7288 int err;
7289
7290 /* This is used to refine r0 return value bounds for helpers
7291 * that enforce this value as an upper bound on return values.
7292 * See do_refine_retval_range() for helpers that can refine
7293 * the return value. C type of helper is u32 so we pull register
7294 * bound from umax_value however, if negative verifier errors
7295 * out. Only upper bounds can be learned because retval is an
7296 * int type and negative retvals are allowed.
7297 */
7298 meta->msize_max_value = reg->umax_value;
7299
7300 /* The register is SCALAR_VALUE; the access check
7301 * happens using its boundaries.
7302 */
7303 if (!tnum_is_const(reg->var_off))
7304 /* For unprivileged variable accesses, disable raw
7305 * mode so that the program is required to
7306 * initialize all the memory that the helper could
7307 * just partially fill up.
7308 */
7309 meta = NULL;
7310
7311 if (reg->smin_value < 0) {
7312 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
7313 regno);
7314 return -EACCES;
7315 }
7316
7317 if (reg->umin_value == 0) {
7318 err = check_helper_mem_access(env, regno - 1, 0,
7319 zero_size_allowed,
7320 meta);
7321 if (err)
7322 return err;
7323 }
7324
7325 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
7326 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
7327 regno);
7328 return -EACCES;
7329 }
7330 err = check_helper_mem_access(env, regno - 1,
7331 reg->umax_value,
7332 zero_size_allowed, meta);
7333 if (!err)
7334 err = mark_chain_precision(env, regno);
7335 return err;
7336 }
7337
check_mem_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,u32 mem_size)7338 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7339 u32 regno, u32 mem_size)
7340 {
7341 bool may_be_null = type_may_be_null(reg->type);
7342 struct bpf_reg_state saved_reg;
7343 struct bpf_call_arg_meta meta;
7344 int err;
7345
7346 if (register_is_null(reg))
7347 return 0;
7348
7349 memset(&meta, 0, sizeof(meta));
7350 /* Assuming that the register contains a value check if the memory
7351 * access is safe. Temporarily save and restore the register's state as
7352 * the conversion shouldn't be visible to a caller.
7353 */
7354 if (may_be_null) {
7355 saved_reg = *reg;
7356 mark_ptr_not_null_reg(reg);
7357 }
7358
7359 err = check_helper_mem_access(env, regno, mem_size, true, &meta);
7360 /* Check access for BPF_WRITE */
7361 meta.raw_mode = true;
7362 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta);
7363
7364 if (may_be_null)
7365 *reg = saved_reg;
7366
7367 return err;
7368 }
7369
check_kfunc_mem_size_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno)7370 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
7371 u32 regno)
7372 {
7373 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
7374 bool may_be_null = type_may_be_null(mem_reg->type);
7375 struct bpf_reg_state saved_reg;
7376 struct bpf_call_arg_meta meta;
7377 int err;
7378
7379 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
7380
7381 memset(&meta, 0, sizeof(meta));
7382
7383 if (may_be_null) {
7384 saved_reg = *mem_reg;
7385 mark_ptr_not_null_reg(mem_reg);
7386 }
7387
7388 err = check_mem_size_reg(env, reg, regno, true, &meta);
7389 /* Check access for BPF_WRITE */
7390 meta.raw_mode = true;
7391 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta);
7392
7393 if (may_be_null)
7394 *mem_reg = saved_reg;
7395 return err;
7396 }
7397
7398 /* Implementation details:
7399 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
7400 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
7401 * Two bpf_map_lookups (even with the same key) will have different reg->id.
7402 * Two separate bpf_obj_new will also have different reg->id.
7403 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
7404 * clears reg->id after value_or_null->value transition, since the verifier only
7405 * cares about the range of access to valid map value pointer and doesn't care
7406 * about actual address of the map element.
7407 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
7408 * reg->id > 0 after value_or_null->value transition. By doing so
7409 * two bpf_map_lookups will be considered two different pointers that
7410 * point to different bpf_spin_locks. Likewise for pointers to allocated objects
7411 * returned from bpf_obj_new.
7412 * The verifier allows taking only one bpf_spin_lock at a time to avoid
7413 * dead-locks.
7414 * Since only one bpf_spin_lock is allowed the checks are simpler than
7415 * reg_is_refcounted() logic. The verifier needs to remember only
7416 * one spin_lock instead of array of acquired_refs.
7417 * cur_state->active_lock remembers which map value element or allocated
7418 * object got locked and clears it after bpf_spin_unlock.
7419 */
process_spin_lock(struct bpf_verifier_env * env,int regno,bool is_lock)7420 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
7421 bool is_lock)
7422 {
7423 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7424 struct bpf_verifier_state *cur = env->cur_state;
7425 bool is_const = tnum_is_const(reg->var_off);
7426 u64 val = reg->var_off.value;
7427 struct bpf_map *map = NULL;
7428 struct btf *btf = NULL;
7429 struct btf_record *rec;
7430
7431 if (!is_const) {
7432 verbose(env,
7433 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
7434 regno);
7435 return -EINVAL;
7436 }
7437 if (reg->type == PTR_TO_MAP_VALUE) {
7438 map = reg->map_ptr;
7439 if (!map->btf) {
7440 verbose(env,
7441 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
7442 map->name);
7443 return -EINVAL;
7444 }
7445 } else {
7446 btf = reg->btf;
7447 }
7448
7449 rec = reg_btf_record(reg);
7450 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) {
7451 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local",
7452 map ? map->name : "kptr");
7453 return -EINVAL;
7454 }
7455 if (rec->spin_lock_off != val + reg->off) {
7456 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n",
7457 val + reg->off, rec->spin_lock_off);
7458 return -EINVAL;
7459 }
7460 if (is_lock) {
7461 if (cur->active_lock.ptr) {
7462 verbose(env,
7463 "Locking two bpf_spin_locks are not allowed\n");
7464 return -EINVAL;
7465 }
7466 if (map)
7467 cur->active_lock.ptr = map;
7468 else
7469 cur->active_lock.ptr = btf;
7470 cur->active_lock.id = reg->id;
7471 } else {
7472 void *ptr;
7473
7474 if (map)
7475 ptr = map;
7476 else
7477 ptr = btf;
7478
7479 if (!cur->active_lock.ptr) {
7480 verbose(env, "bpf_spin_unlock without taking a lock\n");
7481 return -EINVAL;
7482 }
7483 if (cur->active_lock.ptr != ptr ||
7484 cur->active_lock.id != reg->id) {
7485 verbose(env, "bpf_spin_unlock of different lock\n");
7486 return -EINVAL;
7487 }
7488
7489 invalidate_non_owning_refs(env);
7490
7491 cur->active_lock.ptr = NULL;
7492 cur->active_lock.id = 0;
7493 }
7494 return 0;
7495 }
7496
process_timer_func(struct bpf_verifier_env * env,int regno,struct bpf_call_arg_meta * meta)7497 static int process_timer_func(struct bpf_verifier_env *env, int regno,
7498 struct bpf_call_arg_meta *meta)
7499 {
7500 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7501 bool is_const = tnum_is_const(reg->var_off);
7502 struct bpf_map *map = reg->map_ptr;
7503 u64 val = reg->var_off.value;
7504
7505 if (!is_const) {
7506 verbose(env,
7507 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
7508 regno);
7509 return -EINVAL;
7510 }
7511 if (!map->btf) {
7512 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
7513 map->name);
7514 return -EINVAL;
7515 }
7516 if (!btf_record_has_field(map->record, BPF_TIMER)) {
7517 verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
7518 return -EINVAL;
7519 }
7520 if (map->record->timer_off != val + reg->off) {
7521 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
7522 val + reg->off, map->record->timer_off);
7523 return -EINVAL;
7524 }
7525 if (meta->map_ptr) {
7526 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
7527 return -EFAULT;
7528 }
7529 meta->map_uid = reg->map_uid;
7530 meta->map_ptr = map;
7531 return 0;
7532 }
7533
process_kptr_func(struct bpf_verifier_env * env,int regno,struct bpf_call_arg_meta * meta)7534 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
7535 struct bpf_call_arg_meta *meta)
7536 {
7537 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7538 struct bpf_map *map_ptr = reg->map_ptr;
7539 struct btf_field *kptr_field;
7540 u32 kptr_off;
7541
7542 if (!tnum_is_const(reg->var_off)) {
7543 verbose(env,
7544 "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
7545 regno);
7546 return -EINVAL;
7547 }
7548 if (!map_ptr->btf) {
7549 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
7550 map_ptr->name);
7551 return -EINVAL;
7552 }
7553 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) {
7554 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name);
7555 return -EINVAL;
7556 }
7557
7558 meta->map_ptr = map_ptr;
7559 kptr_off = reg->off + reg->var_off.value;
7560 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR);
7561 if (!kptr_field) {
7562 verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
7563 return -EACCES;
7564 }
7565 if (kptr_field->type != BPF_KPTR_REF) {
7566 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
7567 return -EACCES;
7568 }
7569 meta->kptr_field = kptr_field;
7570 return 0;
7571 }
7572
7573 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
7574 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
7575 *
7576 * In both cases we deal with the first 8 bytes, but need to mark the next 8
7577 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
7578 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
7579 *
7580 * Mutability of bpf_dynptr is at two levels, one is at the level of struct
7581 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
7582 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
7583 * mutate the view of the dynptr and also possibly destroy it. In the latter
7584 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
7585 * memory that dynptr points to.
7586 *
7587 * The verifier will keep track both levels of mutation (bpf_dynptr's in
7588 * reg->type and the memory's in reg->dynptr.type), but there is no support for
7589 * readonly dynptr view yet, hence only the first case is tracked and checked.
7590 *
7591 * This is consistent with how C applies the const modifier to a struct object,
7592 * where the pointer itself inside bpf_dynptr becomes const but not what it
7593 * points to.
7594 *
7595 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
7596 * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
7597 */
process_dynptr_func(struct bpf_verifier_env * env,int regno,int insn_idx,enum bpf_arg_type arg_type,int clone_ref_obj_id)7598 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
7599 enum bpf_arg_type arg_type, int clone_ref_obj_id)
7600 {
7601 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7602 int err;
7603
7604 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
7605 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
7606 */
7607 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
7608 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n");
7609 return -EFAULT;
7610 }
7611
7612 /* MEM_UNINIT - Points to memory that is an appropriate candidate for
7613 * constructing a mutable bpf_dynptr object.
7614 *
7615 * Currently, this is only possible with PTR_TO_STACK
7616 * pointing to a region of at least 16 bytes which doesn't
7617 * contain an existing bpf_dynptr.
7618 *
7619 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
7620 * mutated or destroyed. However, the memory it points to
7621 * may be mutated.
7622 *
7623 * None - Points to a initialized dynptr that can be mutated and
7624 * destroyed, including mutation of the memory it points
7625 * to.
7626 */
7627 if (arg_type & MEM_UNINIT) {
7628 int i;
7629
7630 if (!is_dynptr_reg_valid_uninit(env, reg)) {
7631 verbose(env, "Dynptr has to be an uninitialized dynptr\n");
7632 return -EINVAL;
7633 }
7634
7635 /* we write BPF_DW bits (8 bytes) at a time */
7636 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
7637 err = check_mem_access(env, insn_idx, regno,
7638 i, BPF_DW, BPF_WRITE, -1, false, false);
7639 if (err)
7640 return err;
7641 }
7642
7643 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
7644 } else /* MEM_RDONLY and None case from above */ {
7645 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
7646 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
7647 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
7648 return -EINVAL;
7649 }
7650
7651 if (!is_dynptr_reg_valid_init(env, reg)) {
7652 verbose(env,
7653 "Expected an initialized dynptr as arg #%d\n",
7654 regno);
7655 return -EINVAL;
7656 }
7657
7658 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
7659 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
7660 verbose(env,
7661 "Expected a dynptr of type %s as arg #%d\n",
7662 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno);
7663 return -EINVAL;
7664 }
7665
7666 err = mark_dynptr_read(env, reg);
7667 }
7668 return err;
7669 }
7670
iter_ref_obj_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int spi)7671 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
7672 {
7673 struct bpf_func_state *state = func(env, reg);
7674
7675 return state->stack[spi].spilled_ptr.ref_obj_id;
7676 }
7677
is_iter_kfunc(struct bpf_kfunc_call_arg_meta * meta)7678 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7679 {
7680 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
7681 }
7682
is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta * meta)7683 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7684 {
7685 return meta->kfunc_flags & KF_ITER_NEW;
7686 }
7687
is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta * meta)7688 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7689 {
7690 return meta->kfunc_flags & KF_ITER_NEXT;
7691 }
7692
is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta * meta)7693 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
7694 {
7695 return meta->kfunc_flags & KF_ITER_DESTROY;
7696 }
7697
is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta * meta,int arg)7698 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg)
7699 {
7700 /* btf_check_iter_kfuncs() guarantees that first argument of any iter
7701 * kfunc is iter state pointer
7702 */
7703 return arg == 0 && is_iter_kfunc(meta);
7704 }
7705
process_iter_arg(struct bpf_verifier_env * env,int regno,int insn_idx,struct bpf_kfunc_call_arg_meta * meta)7706 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
7707 struct bpf_kfunc_call_arg_meta *meta)
7708 {
7709 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
7710 const struct btf_type *t;
7711 const struct btf_param *arg;
7712 int spi, err, i, nr_slots;
7713 u32 btf_id;
7714
7715 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */
7716 arg = &btf_params(meta->func_proto)[0];
7717 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */
7718 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */
7719 nr_slots = t->size / BPF_REG_SIZE;
7720
7721 if (is_iter_new_kfunc(meta)) {
7722 /* bpf_iter_<type>_new() expects pointer to uninit iter state */
7723 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
7724 verbose(env, "expected uninitialized iter_%s as arg #%d\n",
7725 iter_type_str(meta->btf, btf_id), regno);
7726 return -EINVAL;
7727 }
7728
7729 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
7730 err = check_mem_access(env, insn_idx, regno,
7731 i, BPF_DW, BPF_WRITE, -1, false, false);
7732 if (err)
7733 return err;
7734 }
7735
7736 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots);
7737 if (err)
7738 return err;
7739 } else {
7740 /* iter_next() or iter_destroy() expect initialized iter state*/
7741 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) {
7742 verbose(env, "expected an initialized iter_%s as arg #%d\n",
7743 iter_type_str(meta->btf, btf_id), regno);
7744 return -EINVAL;
7745 }
7746
7747 spi = iter_get_spi(env, reg, nr_slots);
7748 if (spi < 0)
7749 return spi;
7750
7751 err = mark_iter_read(env, reg, spi, nr_slots);
7752 if (err)
7753 return err;
7754
7755 /* remember meta->iter info for process_iter_next_call() */
7756 meta->iter.spi = spi;
7757 meta->iter.frameno = reg->frameno;
7758 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
7759
7760 if (is_iter_destroy_kfunc(meta)) {
7761 err = unmark_stack_slots_iter(env, reg, nr_slots);
7762 if (err)
7763 return err;
7764 }
7765 }
7766
7767 return 0;
7768 }
7769
7770 /* Look for a previous loop entry at insn_idx: nearest parent state
7771 * stopped at insn_idx with callsites matching those in cur->frame.
7772 */
find_prev_entry(struct bpf_verifier_env * env,struct bpf_verifier_state * cur,int insn_idx)7773 static struct bpf_verifier_state *find_prev_entry(struct bpf_verifier_env *env,
7774 struct bpf_verifier_state *cur,
7775 int insn_idx)
7776 {
7777 struct bpf_verifier_state_list *sl;
7778 struct bpf_verifier_state *st;
7779
7780 /* Explored states are pushed in stack order, most recent states come first */
7781 sl = *explored_state(env, insn_idx);
7782 for (; sl; sl = sl->next) {
7783 /* If st->branches != 0 state is a part of current DFS verification path,
7784 * hence cur & st for a loop.
7785 */
7786 st = &sl->state;
7787 if (st->insn_idx == insn_idx && st->branches && same_callsites(st, cur) &&
7788 st->dfs_depth < cur->dfs_depth)
7789 return st;
7790 }
7791
7792 return NULL;
7793 }
7794
7795 static void reset_idmap_scratch(struct bpf_verifier_env *env);
7796 static bool regs_exact(const struct bpf_reg_state *rold,
7797 const struct bpf_reg_state *rcur,
7798 struct bpf_idmap *idmap);
7799
maybe_widen_reg(struct bpf_verifier_env * env,struct bpf_reg_state * rold,struct bpf_reg_state * rcur,struct bpf_idmap * idmap)7800 static void maybe_widen_reg(struct bpf_verifier_env *env,
7801 struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
7802 struct bpf_idmap *idmap)
7803 {
7804 if (rold->type != SCALAR_VALUE)
7805 return;
7806 if (rold->type != rcur->type)
7807 return;
7808 if (rold->precise || rcur->precise || regs_exact(rold, rcur, idmap))
7809 return;
7810 __mark_reg_unknown(env, rcur);
7811 }
7812
widen_imprecise_scalars(struct bpf_verifier_env * env,struct bpf_verifier_state * old,struct bpf_verifier_state * cur)7813 static int widen_imprecise_scalars(struct bpf_verifier_env *env,
7814 struct bpf_verifier_state *old,
7815 struct bpf_verifier_state *cur)
7816 {
7817 struct bpf_func_state *fold, *fcur;
7818 int i, fr;
7819
7820 reset_idmap_scratch(env);
7821 for (fr = old->curframe; fr >= 0; fr--) {
7822 fold = old->frame[fr];
7823 fcur = cur->frame[fr];
7824
7825 for (i = 0; i < MAX_BPF_REG; i++)
7826 maybe_widen_reg(env,
7827 &fold->regs[i],
7828 &fcur->regs[i],
7829 &env->idmap_scratch);
7830
7831 for (i = 0; i < fold->allocated_stack / BPF_REG_SIZE; i++) {
7832 if (!is_spilled_reg(&fold->stack[i]) ||
7833 !is_spilled_reg(&fcur->stack[i]))
7834 continue;
7835
7836 maybe_widen_reg(env,
7837 &fold->stack[i].spilled_ptr,
7838 &fcur->stack[i].spilled_ptr,
7839 &env->idmap_scratch);
7840 }
7841 }
7842 return 0;
7843 }
7844
7845 /* process_iter_next_call() is called when verifier gets to iterator's next
7846 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
7847 * to it as just "iter_next()" in comments below.
7848 *
7849 * BPF verifier relies on a crucial contract for any iter_next()
7850 * implementation: it should *eventually* return NULL, and once that happens
7851 * it should keep returning NULL. That is, once iterator exhausts elements to
7852 * iterate, it should never reset or spuriously return new elements.
7853 *
7854 * With the assumption of such contract, process_iter_next_call() simulates
7855 * a fork in the verifier state to validate loop logic correctness and safety
7856 * without having to simulate infinite amount of iterations.
7857 *
7858 * In current state, we first assume that iter_next() returned NULL and
7859 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
7860 * conditions we should not form an infinite loop and should eventually reach
7861 * exit.
7862 *
7863 * Besides that, we also fork current state and enqueue it for later
7864 * verification. In a forked state we keep iterator state as ACTIVE
7865 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
7866 * also bump iteration depth to prevent erroneous infinite loop detection
7867 * later on (see iter_active_depths_differ() comment for details). In this
7868 * state we assume that we'll eventually loop back to another iter_next()
7869 * calls (it could be in exactly same location or in some other instruction,
7870 * it doesn't matter, we don't make any unnecessary assumptions about this,
7871 * everything revolves around iterator state in a stack slot, not which
7872 * instruction is calling iter_next()). When that happens, we either will come
7873 * to iter_next() with equivalent state and can conclude that next iteration
7874 * will proceed in exactly the same way as we just verified, so it's safe to
7875 * assume that loop converges. If not, we'll go on another iteration
7876 * simulation with a different input state, until all possible starting states
7877 * are validated or we reach maximum number of instructions limit.
7878 *
7879 * This way, we will either exhaustively discover all possible input states
7880 * that iterator loop can start with and eventually will converge, or we'll
7881 * effectively regress into bounded loop simulation logic and either reach
7882 * maximum number of instructions if loop is not provably convergent, or there
7883 * is some statically known limit on number of iterations (e.g., if there is
7884 * an explicit `if n > 100 then break;` statement somewhere in the loop).
7885 *
7886 * Iteration convergence logic in is_state_visited() relies on exact
7887 * states comparison, which ignores read and precision marks.
7888 * This is necessary because read and precision marks are not finalized
7889 * while in the loop. Exact comparison might preclude convergence for
7890 * simple programs like below:
7891 *
7892 * i = 0;
7893 * while(iter_next(&it))
7894 * i++;
7895 *
7896 * At each iteration step i++ would produce a new distinct state and
7897 * eventually instruction processing limit would be reached.
7898 *
7899 * To avoid such behavior speculatively forget (widen) range for
7900 * imprecise scalar registers, if those registers were not precise at the
7901 * end of the previous iteration and do not match exactly.
7902 *
7903 * This is a conservative heuristic that allows to verify wide range of programs,
7904 * however it precludes verification of programs that conjure an
7905 * imprecise value on the first loop iteration and use it as precise on a second.
7906 * For example, the following safe program would fail to verify:
7907 *
7908 * struct bpf_num_iter it;
7909 * int arr[10];
7910 * int i = 0, a = 0;
7911 * bpf_iter_num_new(&it, 0, 10);
7912 * while (bpf_iter_num_next(&it)) {
7913 * if (a == 0) {
7914 * a = 1;
7915 * i = 7; // Because i changed verifier would forget
7916 * // it's range on second loop entry.
7917 * } else {
7918 * arr[i] = 42; // This would fail to verify.
7919 * }
7920 * }
7921 * bpf_iter_num_destroy(&it);
7922 */
process_iter_next_call(struct bpf_verifier_env * env,int insn_idx,struct bpf_kfunc_call_arg_meta * meta)7923 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
7924 struct bpf_kfunc_call_arg_meta *meta)
7925 {
7926 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st, *prev_st;
7927 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
7928 struct bpf_reg_state *cur_iter, *queued_iter;
7929 int iter_frameno = meta->iter.frameno;
7930 int iter_spi = meta->iter.spi;
7931
7932 BTF_TYPE_EMIT(struct bpf_iter);
7933
7934 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7935
7936 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
7937 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
7938 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n",
7939 cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
7940 return -EFAULT;
7941 }
7942
7943 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
7944 /* Because iter_next() call is a checkpoint is_state_visitied()
7945 * should guarantee parent state with same call sites and insn_idx.
7946 */
7947 if (!cur_st->parent || cur_st->parent->insn_idx != insn_idx ||
7948 !same_callsites(cur_st->parent, cur_st)) {
7949 verbose(env, "bug: bad parent state for iter next call");
7950 return -EFAULT;
7951 }
7952 /* Note cur_st->parent in the call below, it is necessary to skip
7953 * checkpoint created for cur_st by is_state_visited()
7954 * right at this instruction.
7955 */
7956 prev_st = find_prev_entry(env, cur_st->parent, insn_idx);
7957 /* branch out active iter state */
7958 queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
7959 if (!queued_st)
7960 return -ENOMEM;
7961
7962 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
7963 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
7964 queued_iter->iter.depth++;
7965 if (prev_st)
7966 widen_imprecise_scalars(env, prev_st, queued_st);
7967
7968 queued_fr = queued_st->frame[queued_st->curframe];
7969 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
7970 }
7971
7972 /* switch to DRAINED state, but keep the depth unchanged */
7973 /* mark current iter state as drained and assume returned NULL */
7974 cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
7975 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]);
7976
7977 return 0;
7978 }
7979
arg_type_is_mem_size(enum bpf_arg_type type)7980 static bool arg_type_is_mem_size(enum bpf_arg_type type)
7981 {
7982 return type == ARG_CONST_SIZE ||
7983 type == ARG_CONST_SIZE_OR_ZERO;
7984 }
7985
arg_type_is_release(enum bpf_arg_type type)7986 static bool arg_type_is_release(enum bpf_arg_type type)
7987 {
7988 return type & OBJ_RELEASE;
7989 }
7990
arg_type_is_dynptr(enum bpf_arg_type type)7991 static bool arg_type_is_dynptr(enum bpf_arg_type type)
7992 {
7993 return base_type(type) == ARG_PTR_TO_DYNPTR;
7994 }
7995
int_ptr_type_to_size(enum bpf_arg_type type)7996 static int int_ptr_type_to_size(enum bpf_arg_type type)
7997 {
7998 if (type == ARG_PTR_TO_INT)
7999 return sizeof(u32);
8000 else if (type == ARG_PTR_TO_LONG)
8001 return sizeof(u64);
8002
8003 return -EINVAL;
8004 }
8005
resolve_map_arg_type(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_arg_type * arg_type)8006 static int resolve_map_arg_type(struct bpf_verifier_env *env,
8007 const struct bpf_call_arg_meta *meta,
8008 enum bpf_arg_type *arg_type)
8009 {
8010 if (!meta->map_ptr) {
8011 /* kernel subsystem misconfigured verifier */
8012 verbose(env, "invalid map_ptr to access map->type\n");
8013 return -EACCES;
8014 }
8015
8016 switch (meta->map_ptr->map_type) {
8017 case BPF_MAP_TYPE_SOCKMAP:
8018 case BPF_MAP_TYPE_SOCKHASH:
8019 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
8020 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
8021 } else {
8022 verbose(env, "invalid arg_type for sockmap/sockhash\n");
8023 return -EINVAL;
8024 }
8025 break;
8026 case BPF_MAP_TYPE_BLOOM_FILTER:
8027 if (meta->func_id == BPF_FUNC_map_peek_elem)
8028 *arg_type = ARG_PTR_TO_MAP_VALUE;
8029 break;
8030 default:
8031 break;
8032 }
8033 return 0;
8034 }
8035
8036 struct bpf_reg_types {
8037 const enum bpf_reg_type types[10];
8038 u32 *btf_id;
8039 };
8040
8041 static const struct bpf_reg_types sock_types = {
8042 .types = {
8043 PTR_TO_SOCK_COMMON,
8044 PTR_TO_SOCKET,
8045 PTR_TO_TCP_SOCK,
8046 PTR_TO_XDP_SOCK,
8047 },
8048 };
8049
8050 #ifdef CONFIG_NET
8051 static const struct bpf_reg_types btf_id_sock_common_types = {
8052 .types = {
8053 PTR_TO_SOCK_COMMON,
8054 PTR_TO_SOCKET,
8055 PTR_TO_TCP_SOCK,
8056 PTR_TO_XDP_SOCK,
8057 PTR_TO_BTF_ID,
8058 PTR_TO_BTF_ID | PTR_TRUSTED,
8059 },
8060 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
8061 };
8062 #endif
8063
8064 static const struct bpf_reg_types mem_types = {
8065 .types = {
8066 PTR_TO_STACK,
8067 PTR_TO_PACKET,
8068 PTR_TO_PACKET_META,
8069 PTR_TO_MAP_KEY,
8070 PTR_TO_MAP_VALUE,
8071 PTR_TO_MEM,
8072 PTR_TO_MEM | MEM_RINGBUF,
8073 PTR_TO_BUF,
8074 PTR_TO_BTF_ID | PTR_TRUSTED,
8075 },
8076 };
8077
8078 static const struct bpf_reg_types int_ptr_types = {
8079 .types = {
8080 PTR_TO_STACK,
8081 PTR_TO_PACKET,
8082 PTR_TO_PACKET_META,
8083 PTR_TO_MAP_KEY,
8084 PTR_TO_MAP_VALUE,
8085 },
8086 };
8087
8088 static const struct bpf_reg_types spin_lock_types = {
8089 .types = {
8090 PTR_TO_MAP_VALUE,
8091 PTR_TO_BTF_ID | MEM_ALLOC,
8092 }
8093 };
8094
8095 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
8096 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
8097 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
8098 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
8099 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
8100 static const struct bpf_reg_types btf_ptr_types = {
8101 .types = {
8102 PTR_TO_BTF_ID,
8103 PTR_TO_BTF_ID | PTR_TRUSTED,
8104 PTR_TO_BTF_ID | MEM_RCU,
8105 },
8106 };
8107 static const struct bpf_reg_types percpu_btf_ptr_types = {
8108 .types = {
8109 PTR_TO_BTF_ID | MEM_PERCPU,
8110 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
8111 }
8112 };
8113 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
8114 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
8115 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
8116 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
8117 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } };
8118 static const struct bpf_reg_types dynptr_types = {
8119 .types = {
8120 PTR_TO_STACK,
8121 CONST_PTR_TO_DYNPTR,
8122 }
8123 };
8124
8125 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
8126 [ARG_PTR_TO_MAP_KEY] = &mem_types,
8127 [ARG_PTR_TO_MAP_VALUE] = &mem_types,
8128 [ARG_CONST_SIZE] = &scalar_types,
8129 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
8130 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
8131 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
8132 [ARG_PTR_TO_CTX] = &context_types,
8133 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
8134 #ifdef CONFIG_NET
8135 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
8136 #endif
8137 [ARG_PTR_TO_SOCKET] = &fullsock_types,
8138 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
8139 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
8140 [ARG_PTR_TO_MEM] = &mem_types,
8141 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
8142 [ARG_PTR_TO_INT] = &int_ptr_types,
8143 [ARG_PTR_TO_LONG] = &int_ptr_types,
8144 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
8145 [ARG_PTR_TO_FUNC] = &func_ptr_types,
8146 [ARG_PTR_TO_STACK] = &stack_ptr_types,
8147 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
8148 [ARG_PTR_TO_TIMER] = &timer_types,
8149 [ARG_PTR_TO_KPTR] = &kptr_types,
8150 [ARG_PTR_TO_DYNPTR] = &dynptr_types,
8151 };
8152
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)8153 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
8154 enum bpf_arg_type arg_type,
8155 const u32 *arg_btf_id,
8156 struct bpf_call_arg_meta *meta)
8157 {
8158 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8159 enum bpf_reg_type expected, type = reg->type;
8160 const struct bpf_reg_types *compatible;
8161 int i, j;
8162
8163 compatible = compatible_reg_types[base_type(arg_type)];
8164 if (!compatible) {
8165 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
8166 return -EFAULT;
8167 }
8168
8169 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
8170 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
8171 *
8172 * Same for MAYBE_NULL:
8173 *
8174 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
8175 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
8176 *
8177 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
8178 *
8179 * Therefore we fold these flags depending on the arg_type before comparison.
8180 */
8181 if (arg_type & MEM_RDONLY)
8182 type &= ~MEM_RDONLY;
8183 if (arg_type & PTR_MAYBE_NULL)
8184 type &= ~PTR_MAYBE_NULL;
8185 if (base_type(arg_type) == ARG_PTR_TO_MEM)
8186 type &= ~DYNPTR_TYPE_FLAG_MASK;
8187
8188 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type))
8189 type &= ~MEM_ALLOC;
8190
8191 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
8192 expected = compatible->types[i];
8193 if (expected == NOT_INIT)
8194 break;
8195
8196 if (type == expected)
8197 goto found;
8198 }
8199
8200 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
8201 for (j = 0; j + 1 < i; j++)
8202 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
8203 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
8204 return -EACCES;
8205
8206 found:
8207 if (base_type(reg->type) != PTR_TO_BTF_ID)
8208 return 0;
8209
8210 if (compatible == &mem_types) {
8211 if (!(arg_type & MEM_RDONLY)) {
8212 verbose(env,
8213 "%s() may write into memory pointed by R%d type=%s\n",
8214 func_id_name(meta->func_id),
8215 regno, reg_type_str(env, reg->type));
8216 return -EACCES;
8217 }
8218 return 0;
8219 }
8220
8221 switch ((int)reg->type) {
8222 case PTR_TO_BTF_ID:
8223 case PTR_TO_BTF_ID | PTR_TRUSTED:
8224 case PTR_TO_BTF_ID | MEM_RCU:
8225 case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
8226 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
8227 {
8228 /* For bpf_sk_release, it needs to match against first member
8229 * 'struct sock_common', hence make an exception for it. This
8230 * allows bpf_sk_release to work for multiple socket types.
8231 */
8232 bool strict_type_match = arg_type_is_release(arg_type) &&
8233 meta->func_id != BPF_FUNC_sk_release;
8234
8235 if (type_may_be_null(reg->type) &&
8236 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
8237 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
8238 return -EACCES;
8239 }
8240
8241 if (!arg_btf_id) {
8242 if (!compatible->btf_id) {
8243 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
8244 return -EFAULT;
8245 }
8246 arg_btf_id = compatible->btf_id;
8247 }
8248
8249 if (meta->func_id == BPF_FUNC_kptr_xchg) {
8250 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
8251 return -EACCES;
8252 } else {
8253 if (arg_btf_id == BPF_PTR_POISON) {
8254 verbose(env, "verifier internal error:");
8255 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
8256 regno);
8257 return -EACCES;
8258 }
8259
8260 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
8261 btf_vmlinux, *arg_btf_id,
8262 strict_type_match)) {
8263 verbose(env, "R%d is of type %s but %s is expected\n",
8264 regno, btf_type_name(reg->btf, reg->btf_id),
8265 btf_type_name(btf_vmlinux, *arg_btf_id));
8266 return -EACCES;
8267 }
8268 }
8269 break;
8270 }
8271 case PTR_TO_BTF_ID | MEM_ALLOC:
8272 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
8273 meta->func_id != BPF_FUNC_kptr_xchg) {
8274 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n");
8275 return -EFAULT;
8276 }
8277 if (meta->func_id == BPF_FUNC_kptr_xchg) {
8278 if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
8279 return -EACCES;
8280 }
8281 break;
8282 case PTR_TO_BTF_ID | MEM_PERCPU:
8283 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
8284 /* Handled by helper specific checks */
8285 break;
8286 default:
8287 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n");
8288 return -EFAULT;
8289 }
8290 return 0;
8291 }
8292
8293 static struct btf_field *
reg_find_field_offset(const struct bpf_reg_state * reg,s32 off,u32 fields)8294 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
8295 {
8296 struct btf_field *field;
8297 struct btf_record *rec;
8298
8299 rec = reg_btf_record(reg);
8300 if (!rec)
8301 return NULL;
8302
8303 field = btf_record_find(rec, off, fields);
8304 if (!field)
8305 return NULL;
8306
8307 return field;
8308 }
8309
check_func_arg_reg_off(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,enum bpf_arg_type arg_type)8310 int check_func_arg_reg_off(struct bpf_verifier_env *env,
8311 const struct bpf_reg_state *reg, int regno,
8312 enum bpf_arg_type arg_type)
8313 {
8314 u32 type = reg->type;
8315
8316 /* When referenced register is passed to release function, its fixed
8317 * offset must be 0.
8318 *
8319 * We will check arg_type_is_release reg has ref_obj_id when storing
8320 * meta->release_regno.
8321 */
8322 if (arg_type_is_release(arg_type)) {
8323 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
8324 * may not directly point to the object being released, but to
8325 * dynptr pointing to such object, which might be at some offset
8326 * on the stack. In that case, we simply to fallback to the
8327 * default handling.
8328 */
8329 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
8330 return 0;
8331
8332 /* Doing check_ptr_off_reg check for the offset will catch this
8333 * because fixed_off_ok is false, but checking here allows us
8334 * to give the user a better error message.
8335 */
8336 if (reg->off) {
8337 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
8338 regno);
8339 return -EINVAL;
8340 }
8341 return __check_ptr_off_reg(env, reg, regno, false);
8342 }
8343
8344 switch (type) {
8345 /* Pointer types where both fixed and variable offset is explicitly allowed: */
8346 case PTR_TO_STACK:
8347 case PTR_TO_PACKET:
8348 case PTR_TO_PACKET_META:
8349 case PTR_TO_MAP_KEY:
8350 case PTR_TO_MAP_VALUE:
8351 case PTR_TO_MEM:
8352 case PTR_TO_MEM | MEM_RDONLY:
8353 case PTR_TO_MEM | MEM_RINGBUF:
8354 case PTR_TO_BUF:
8355 case PTR_TO_BUF | MEM_RDONLY:
8356 case SCALAR_VALUE:
8357 return 0;
8358 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows
8359 * fixed offset.
8360 */
8361 case PTR_TO_BTF_ID:
8362 case PTR_TO_BTF_ID | MEM_ALLOC:
8363 case PTR_TO_BTF_ID | PTR_TRUSTED:
8364 case PTR_TO_BTF_ID | MEM_RCU:
8365 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
8366 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
8367 /* When referenced PTR_TO_BTF_ID is passed to release function,
8368 * its fixed offset must be 0. In the other cases, fixed offset
8369 * can be non-zero. This was already checked above. So pass
8370 * fixed_off_ok as true to allow fixed offset for all other
8371 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
8372 * still need to do checks instead of returning.
8373 */
8374 return __check_ptr_off_reg(env, reg, regno, true);
8375 default:
8376 return __check_ptr_off_reg(env, reg, regno, false);
8377 }
8378 }
8379
get_dynptr_arg_reg(struct bpf_verifier_env * env,const struct bpf_func_proto * fn,struct bpf_reg_state * regs)8380 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
8381 const struct bpf_func_proto *fn,
8382 struct bpf_reg_state *regs)
8383 {
8384 struct bpf_reg_state *state = NULL;
8385 int i;
8386
8387 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
8388 if (arg_type_is_dynptr(fn->arg_type[i])) {
8389 if (state) {
8390 verbose(env, "verifier internal error: multiple dynptr args\n");
8391 return NULL;
8392 }
8393 state = ®s[BPF_REG_1 + i];
8394 }
8395
8396 if (!state)
8397 verbose(env, "verifier internal error: no dynptr arg found\n");
8398
8399 return state;
8400 }
8401
dynptr_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg)8402 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8403 {
8404 struct bpf_func_state *state = func(env, reg);
8405 int spi;
8406
8407 if (reg->type == CONST_PTR_TO_DYNPTR)
8408 return reg->id;
8409 spi = dynptr_get_spi(env, reg);
8410 if (spi < 0)
8411 return spi;
8412 return state->stack[spi].spilled_ptr.id;
8413 }
8414
dynptr_ref_obj_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg)8415 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
8416 {
8417 struct bpf_func_state *state = func(env, reg);
8418 int spi;
8419
8420 if (reg->type == CONST_PTR_TO_DYNPTR)
8421 return reg->ref_obj_id;
8422 spi = dynptr_get_spi(env, reg);
8423 if (spi < 0)
8424 return spi;
8425 return state->stack[spi].spilled_ptr.ref_obj_id;
8426 }
8427
dynptr_get_type(struct bpf_verifier_env * env,struct bpf_reg_state * reg)8428 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
8429 struct bpf_reg_state *reg)
8430 {
8431 struct bpf_func_state *state = func(env, reg);
8432 int spi;
8433
8434 if (reg->type == CONST_PTR_TO_DYNPTR)
8435 return reg->dynptr.type;
8436
8437 spi = __get_spi(reg->off);
8438 if (spi < 0) {
8439 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
8440 return BPF_DYNPTR_TYPE_INVALID;
8441 }
8442
8443 return state->stack[spi].spilled_ptr.dynptr.type;
8444 }
8445
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)8446 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
8447 struct bpf_call_arg_meta *meta,
8448 const struct bpf_func_proto *fn,
8449 int insn_idx)
8450 {
8451 u32 regno = BPF_REG_1 + arg;
8452 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
8453 enum bpf_arg_type arg_type = fn->arg_type[arg];
8454 enum bpf_reg_type type = reg->type;
8455 u32 *arg_btf_id = NULL;
8456 int err = 0;
8457
8458 if (arg_type == ARG_DONTCARE)
8459 return 0;
8460
8461 err = check_reg_arg(env, regno, SRC_OP);
8462 if (err)
8463 return err;
8464
8465 if (arg_type == ARG_ANYTHING) {
8466 if (is_pointer_value(env, regno)) {
8467 verbose(env, "R%d leaks addr into helper function\n",
8468 regno);
8469 return -EACCES;
8470 }
8471 return 0;
8472 }
8473
8474 if (type_is_pkt_pointer(type) &&
8475 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
8476 verbose(env, "helper access to the packet is not allowed\n");
8477 return -EACCES;
8478 }
8479
8480 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
8481 err = resolve_map_arg_type(env, meta, &arg_type);
8482 if (err)
8483 return err;
8484 }
8485
8486 if (register_is_null(reg) && type_may_be_null(arg_type))
8487 /* A NULL register has a SCALAR_VALUE type, so skip
8488 * type checking.
8489 */
8490 goto skip_type_check;
8491
8492 /* arg_btf_id and arg_size are in a union. */
8493 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
8494 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
8495 arg_btf_id = fn->arg_btf_id[arg];
8496
8497 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
8498 if (err)
8499 return err;
8500
8501 err = check_func_arg_reg_off(env, reg, regno, arg_type);
8502 if (err)
8503 return err;
8504
8505 skip_type_check:
8506 if (arg_type_is_release(arg_type)) {
8507 if (arg_type_is_dynptr(arg_type)) {
8508 struct bpf_func_state *state = func(env, reg);
8509 int spi;
8510
8511 /* Only dynptr created on stack can be released, thus
8512 * the get_spi and stack state checks for spilled_ptr
8513 * should only be done before process_dynptr_func for
8514 * PTR_TO_STACK.
8515 */
8516 if (reg->type == PTR_TO_STACK) {
8517 spi = dynptr_get_spi(env, reg);
8518 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
8519 verbose(env, "arg %d is an unacquired reference\n", regno);
8520 return -EINVAL;
8521 }
8522 } else {
8523 verbose(env, "cannot release unowned const bpf_dynptr\n");
8524 return -EINVAL;
8525 }
8526 } else if (!reg->ref_obj_id && !register_is_null(reg)) {
8527 verbose(env, "R%d must be referenced when passed to release function\n",
8528 regno);
8529 return -EINVAL;
8530 }
8531 if (meta->release_regno) {
8532 verbose(env, "verifier internal error: more than one release argument\n");
8533 return -EFAULT;
8534 }
8535 meta->release_regno = regno;
8536 }
8537
8538 if (reg->ref_obj_id) {
8539 if (meta->ref_obj_id) {
8540 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
8541 regno, reg->ref_obj_id,
8542 meta->ref_obj_id);
8543 return -EFAULT;
8544 }
8545 meta->ref_obj_id = reg->ref_obj_id;
8546 }
8547
8548 switch (base_type(arg_type)) {
8549 case ARG_CONST_MAP_PTR:
8550 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
8551 if (meta->map_ptr) {
8552 /* Use map_uid (which is unique id of inner map) to reject:
8553 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
8554 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
8555 * if (inner_map1 && inner_map2) {
8556 * timer = bpf_map_lookup_elem(inner_map1);
8557 * if (timer)
8558 * // mismatch would have been allowed
8559 * bpf_timer_init(timer, inner_map2);
8560 * }
8561 *
8562 * Comparing map_ptr is enough to distinguish normal and outer maps.
8563 */
8564 if (meta->map_ptr != reg->map_ptr ||
8565 meta->map_uid != reg->map_uid) {
8566 verbose(env,
8567 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
8568 meta->map_uid, reg->map_uid);
8569 return -EINVAL;
8570 }
8571 }
8572 meta->map_ptr = reg->map_ptr;
8573 meta->map_uid = reg->map_uid;
8574 break;
8575 case ARG_PTR_TO_MAP_KEY:
8576 /* bpf_map_xxx(..., map_ptr, ..., key) call:
8577 * check that [key, key + map->key_size) are within
8578 * stack limits and initialized
8579 */
8580 if (!meta->map_ptr) {
8581 /* in function declaration map_ptr must come before
8582 * map_key, so that it's verified and known before
8583 * we have to check map_key here. Otherwise it means
8584 * that kernel subsystem misconfigured verifier
8585 */
8586 verbose(env, "invalid map_ptr to access map->key\n");
8587 return -EACCES;
8588 }
8589 err = check_helper_mem_access(env, regno,
8590 meta->map_ptr->key_size, false,
8591 NULL);
8592 break;
8593 case ARG_PTR_TO_MAP_VALUE:
8594 if (type_may_be_null(arg_type) && register_is_null(reg))
8595 return 0;
8596
8597 /* bpf_map_xxx(..., map_ptr, ..., value) call:
8598 * check [value, value + map->value_size) validity
8599 */
8600 if (!meta->map_ptr) {
8601 /* kernel subsystem misconfigured verifier */
8602 verbose(env, "invalid map_ptr to access map->value\n");
8603 return -EACCES;
8604 }
8605 meta->raw_mode = arg_type & MEM_UNINIT;
8606 err = check_helper_mem_access(env, regno,
8607 meta->map_ptr->value_size, false,
8608 meta);
8609 break;
8610 case ARG_PTR_TO_PERCPU_BTF_ID:
8611 if (!reg->btf_id) {
8612 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
8613 return -EACCES;
8614 }
8615 meta->ret_btf = reg->btf;
8616 meta->ret_btf_id = reg->btf_id;
8617 break;
8618 case ARG_PTR_TO_SPIN_LOCK:
8619 if (in_rbtree_lock_required_cb(env)) {
8620 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
8621 return -EACCES;
8622 }
8623 if (meta->func_id == BPF_FUNC_spin_lock) {
8624 err = process_spin_lock(env, regno, true);
8625 if (err)
8626 return err;
8627 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
8628 err = process_spin_lock(env, regno, false);
8629 if (err)
8630 return err;
8631 } else {
8632 verbose(env, "verifier internal error\n");
8633 return -EFAULT;
8634 }
8635 break;
8636 case ARG_PTR_TO_TIMER:
8637 err = process_timer_func(env, regno, meta);
8638 if (err)
8639 return err;
8640 break;
8641 case ARG_PTR_TO_FUNC:
8642 meta->subprogno = reg->subprogno;
8643 break;
8644 case ARG_PTR_TO_MEM:
8645 /* The access to this pointer is only checked when we hit the
8646 * next is_mem_size argument below.
8647 */
8648 meta->raw_mode = arg_type & MEM_UNINIT;
8649 if (arg_type & MEM_FIXED_SIZE) {
8650 err = check_helper_mem_access(env, regno,
8651 fn->arg_size[arg], false,
8652 meta);
8653 }
8654 break;
8655 case ARG_CONST_SIZE:
8656 err = check_mem_size_reg(env, reg, regno, false, meta);
8657 break;
8658 case ARG_CONST_SIZE_OR_ZERO:
8659 err = check_mem_size_reg(env, reg, regno, true, meta);
8660 break;
8661 case ARG_PTR_TO_DYNPTR:
8662 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
8663 if (err)
8664 return err;
8665 break;
8666 case ARG_CONST_ALLOC_SIZE_OR_ZERO:
8667 if (!tnum_is_const(reg->var_off)) {
8668 verbose(env, "R%d is not a known constant'\n",
8669 regno);
8670 return -EACCES;
8671 }
8672 meta->mem_size = reg->var_off.value;
8673 err = mark_chain_precision(env, regno);
8674 if (err)
8675 return err;
8676 break;
8677 case ARG_PTR_TO_INT:
8678 case ARG_PTR_TO_LONG:
8679 {
8680 int size = int_ptr_type_to_size(arg_type);
8681
8682 err = check_helper_mem_access(env, regno, size, false, meta);
8683 if (err)
8684 return err;
8685 err = check_ptr_alignment(env, reg, 0, size, true);
8686 break;
8687 }
8688 case ARG_PTR_TO_CONST_STR:
8689 {
8690 struct bpf_map *map = reg->map_ptr;
8691 int map_off;
8692 u64 map_addr;
8693 char *str_ptr;
8694
8695 if (!bpf_map_is_rdonly(map)) {
8696 verbose(env, "R%d does not point to a readonly map'\n", regno);
8697 return -EACCES;
8698 }
8699
8700 if (!tnum_is_const(reg->var_off)) {
8701 verbose(env, "R%d is not a constant address'\n", regno);
8702 return -EACCES;
8703 }
8704
8705 if (!map->ops->map_direct_value_addr) {
8706 verbose(env, "no direct value access support for this map type\n");
8707 return -EACCES;
8708 }
8709
8710 err = check_map_access(env, regno, reg->off,
8711 map->value_size - reg->off, false,
8712 ACCESS_HELPER);
8713 if (err)
8714 return err;
8715
8716 map_off = reg->off + reg->var_off.value;
8717 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
8718 if (err) {
8719 verbose(env, "direct value access on string failed\n");
8720 return err;
8721 }
8722
8723 str_ptr = (char *)(long)(map_addr);
8724 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
8725 verbose(env, "string is not zero-terminated\n");
8726 return -EINVAL;
8727 }
8728 break;
8729 }
8730 case ARG_PTR_TO_KPTR:
8731 err = process_kptr_func(env, regno, meta);
8732 if (err)
8733 return err;
8734 break;
8735 }
8736
8737 return err;
8738 }
8739
may_update_sockmap(struct bpf_verifier_env * env,int func_id)8740 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
8741 {
8742 enum bpf_attach_type eatype = env->prog->expected_attach_type;
8743 enum bpf_prog_type type = resolve_prog_type(env->prog);
8744
8745 if (func_id != BPF_FUNC_map_update_elem &&
8746 func_id != BPF_FUNC_map_delete_elem)
8747 return false;
8748
8749 /* It's not possible to get access to a locked struct sock in these
8750 * contexts, so updating is safe.
8751 */
8752 switch (type) {
8753 case BPF_PROG_TYPE_TRACING:
8754 if (eatype == BPF_TRACE_ITER)
8755 return true;
8756 break;
8757 case BPF_PROG_TYPE_SOCK_OPS:
8758 /* map_update allowed only via dedicated helpers with event type checks */
8759 if (func_id == BPF_FUNC_map_delete_elem)
8760 return true;
8761 break;
8762 case BPF_PROG_TYPE_SOCKET_FILTER:
8763 case BPF_PROG_TYPE_SCHED_CLS:
8764 case BPF_PROG_TYPE_SCHED_ACT:
8765 case BPF_PROG_TYPE_XDP:
8766 case BPF_PROG_TYPE_SK_REUSEPORT:
8767 case BPF_PROG_TYPE_FLOW_DISSECTOR:
8768 case BPF_PROG_TYPE_SK_LOOKUP:
8769 return true;
8770 default:
8771 break;
8772 }
8773
8774 verbose(env, "cannot update sockmap in this context\n");
8775 return false;
8776 }
8777
allow_tail_call_in_subprogs(struct bpf_verifier_env * env)8778 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
8779 {
8780 return env->prog->jit_requested &&
8781 bpf_jit_supports_subprog_tailcalls();
8782 }
8783
check_map_func_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,int func_id)8784 static int check_map_func_compatibility(struct bpf_verifier_env *env,
8785 struct bpf_map *map, int func_id)
8786 {
8787 if (!map)
8788 return 0;
8789
8790 /* We need a two way check, first is from map perspective ... */
8791 switch (map->map_type) {
8792 case BPF_MAP_TYPE_PROG_ARRAY:
8793 if (func_id != BPF_FUNC_tail_call)
8794 goto error;
8795 break;
8796 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
8797 if (func_id != BPF_FUNC_perf_event_read &&
8798 func_id != BPF_FUNC_perf_event_output &&
8799 func_id != BPF_FUNC_skb_output &&
8800 func_id != BPF_FUNC_perf_event_read_value &&
8801 func_id != BPF_FUNC_xdp_output)
8802 goto error;
8803 break;
8804 case BPF_MAP_TYPE_RINGBUF:
8805 if (func_id != BPF_FUNC_ringbuf_output &&
8806 func_id != BPF_FUNC_ringbuf_reserve &&
8807 func_id != BPF_FUNC_ringbuf_query &&
8808 func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
8809 func_id != BPF_FUNC_ringbuf_submit_dynptr &&
8810 func_id != BPF_FUNC_ringbuf_discard_dynptr)
8811 goto error;
8812 break;
8813 case BPF_MAP_TYPE_USER_RINGBUF:
8814 if (func_id != BPF_FUNC_user_ringbuf_drain)
8815 goto error;
8816 break;
8817 case BPF_MAP_TYPE_STACK_TRACE:
8818 if (func_id != BPF_FUNC_get_stackid)
8819 goto error;
8820 break;
8821 case BPF_MAP_TYPE_CGROUP_ARRAY:
8822 if (func_id != BPF_FUNC_skb_under_cgroup &&
8823 func_id != BPF_FUNC_current_task_under_cgroup)
8824 goto error;
8825 break;
8826 case BPF_MAP_TYPE_CGROUP_STORAGE:
8827 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
8828 if (func_id != BPF_FUNC_get_local_storage)
8829 goto error;
8830 break;
8831 case BPF_MAP_TYPE_DEVMAP:
8832 case BPF_MAP_TYPE_DEVMAP_HASH:
8833 if (func_id != BPF_FUNC_redirect_map &&
8834 func_id != BPF_FUNC_map_lookup_elem)
8835 goto error;
8836 break;
8837 /* Restrict bpf side of cpumap and xskmap, open when use-cases
8838 * appear.
8839 */
8840 case BPF_MAP_TYPE_CPUMAP:
8841 if (func_id != BPF_FUNC_redirect_map)
8842 goto error;
8843 break;
8844 case BPF_MAP_TYPE_XSKMAP:
8845 if (func_id != BPF_FUNC_redirect_map &&
8846 func_id != BPF_FUNC_map_lookup_elem)
8847 goto error;
8848 break;
8849 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
8850 case BPF_MAP_TYPE_HASH_OF_MAPS:
8851 if (func_id != BPF_FUNC_map_lookup_elem)
8852 goto error;
8853 break;
8854 case BPF_MAP_TYPE_SOCKMAP:
8855 if (func_id != BPF_FUNC_sk_redirect_map &&
8856 func_id != BPF_FUNC_sock_map_update &&
8857 func_id != BPF_FUNC_msg_redirect_map &&
8858 func_id != BPF_FUNC_sk_select_reuseport &&
8859 func_id != BPF_FUNC_map_lookup_elem &&
8860 !may_update_sockmap(env, func_id))
8861 goto error;
8862 break;
8863 case BPF_MAP_TYPE_SOCKHASH:
8864 if (func_id != BPF_FUNC_sk_redirect_hash &&
8865 func_id != BPF_FUNC_sock_hash_update &&
8866 func_id != BPF_FUNC_msg_redirect_hash &&
8867 func_id != BPF_FUNC_sk_select_reuseport &&
8868 func_id != BPF_FUNC_map_lookup_elem &&
8869 !may_update_sockmap(env, func_id))
8870 goto error;
8871 break;
8872 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
8873 if (func_id != BPF_FUNC_sk_select_reuseport)
8874 goto error;
8875 break;
8876 case BPF_MAP_TYPE_QUEUE:
8877 case BPF_MAP_TYPE_STACK:
8878 if (func_id != BPF_FUNC_map_peek_elem &&
8879 func_id != BPF_FUNC_map_pop_elem &&
8880 func_id != BPF_FUNC_map_push_elem)
8881 goto error;
8882 break;
8883 case BPF_MAP_TYPE_SK_STORAGE:
8884 if (func_id != BPF_FUNC_sk_storage_get &&
8885 func_id != BPF_FUNC_sk_storage_delete &&
8886 func_id != BPF_FUNC_kptr_xchg)
8887 goto error;
8888 break;
8889 case BPF_MAP_TYPE_INODE_STORAGE:
8890 if (func_id != BPF_FUNC_inode_storage_get &&
8891 func_id != BPF_FUNC_inode_storage_delete &&
8892 func_id != BPF_FUNC_kptr_xchg)
8893 goto error;
8894 break;
8895 case BPF_MAP_TYPE_TASK_STORAGE:
8896 if (func_id != BPF_FUNC_task_storage_get &&
8897 func_id != BPF_FUNC_task_storage_delete &&
8898 func_id != BPF_FUNC_kptr_xchg)
8899 goto error;
8900 break;
8901 case BPF_MAP_TYPE_CGRP_STORAGE:
8902 if (func_id != BPF_FUNC_cgrp_storage_get &&
8903 func_id != BPF_FUNC_cgrp_storage_delete &&
8904 func_id != BPF_FUNC_kptr_xchg)
8905 goto error;
8906 break;
8907 case BPF_MAP_TYPE_BLOOM_FILTER:
8908 if (func_id != BPF_FUNC_map_peek_elem &&
8909 func_id != BPF_FUNC_map_push_elem)
8910 goto error;
8911 break;
8912 default:
8913 break;
8914 }
8915
8916 /* ... and second from the function itself. */
8917 switch (func_id) {
8918 case BPF_FUNC_tail_call:
8919 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
8920 goto error;
8921 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
8922 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
8923 return -EINVAL;
8924 }
8925 break;
8926 case BPF_FUNC_perf_event_read:
8927 case BPF_FUNC_perf_event_output:
8928 case BPF_FUNC_perf_event_read_value:
8929 case BPF_FUNC_skb_output:
8930 case BPF_FUNC_xdp_output:
8931 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
8932 goto error;
8933 break;
8934 case BPF_FUNC_ringbuf_output:
8935 case BPF_FUNC_ringbuf_reserve:
8936 case BPF_FUNC_ringbuf_query:
8937 case BPF_FUNC_ringbuf_reserve_dynptr:
8938 case BPF_FUNC_ringbuf_submit_dynptr:
8939 case BPF_FUNC_ringbuf_discard_dynptr:
8940 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
8941 goto error;
8942 break;
8943 case BPF_FUNC_user_ringbuf_drain:
8944 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
8945 goto error;
8946 break;
8947 case BPF_FUNC_get_stackid:
8948 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
8949 goto error;
8950 break;
8951 case BPF_FUNC_current_task_under_cgroup:
8952 case BPF_FUNC_skb_under_cgroup:
8953 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
8954 goto error;
8955 break;
8956 case BPF_FUNC_redirect_map:
8957 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
8958 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
8959 map->map_type != BPF_MAP_TYPE_CPUMAP &&
8960 map->map_type != BPF_MAP_TYPE_XSKMAP)
8961 goto error;
8962 break;
8963 case BPF_FUNC_sk_redirect_map:
8964 case BPF_FUNC_msg_redirect_map:
8965 case BPF_FUNC_sock_map_update:
8966 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
8967 goto error;
8968 break;
8969 case BPF_FUNC_sk_redirect_hash:
8970 case BPF_FUNC_msg_redirect_hash:
8971 case BPF_FUNC_sock_hash_update:
8972 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
8973 goto error;
8974 break;
8975 case BPF_FUNC_get_local_storage:
8976 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
8977 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
8978 goto error;
8979 break;
8980 case BPF_FUNC_sk_select_reuseport:
8981 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
8982 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
8983 map->map_type != BPF_MAP_TYPE_SOCKHASH)
8984 goto error;
8985 break;
8986 case BPF_FUNC_map_pop_elem:
8987 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8988 map->map_type != BPF_MAP_TYPE_STACK)
8989 goto error;
8990 break;
8991 case BPF_FUNC_map_peek_elem:
8992 case BPF_FUNC_map_push_elem:
8993 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
8994 map->map_type != BPF_MAP_TYPE_STACK &&
8995 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
8996 goto error;
8997 break;
8998 case BPF_FUNC_map_lookup_percpu_elem:
8999 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
9000 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
9001 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
9002 goto error;
9003 break;
9004 case BPF_FUNC_sk_storage_get:
9005 case BPF_FUNC_sk_storage_delete:
9006 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
9007 goto error;
9008 break;
9009 case BPF_FUNC_inode_storage_get:
9010 case BPF_FUNC_inode_storage_delete:
9011 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
9012 goto error;
9013 break;
9014 case BPF_FUNC_task_storage_get:
9015 case BPF_FUNC_task_storage_delete:
9016 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
9017 goto error;
9018 break;
9019 case BPF_FUNC_cgrp_storage_get:
9020 case BPF_FUNC_cgrp_storage_delete:
9021 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
9022 goto error;
9023 break;
9024 default:
9025 break;
9026 }
9027
9028 return 0;
9029 error:
9030 verbose(env, "cannot pass map_type %d into func %s#%d\n",
9031 map->map_type, func_id_name(func_id), func_id);
9032 return -EINVAL;
9033 }
9034
check_raw_mode_ok(const struct bpf_func_proto * fn)9035 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
9036 {
9037 int count = 0;
9038
9039 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
9040 count++;
9041 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
9042 count++;
9043 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
9044 count++;
9045 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
9046 count++;
9047 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
9048 count++;
9049
9050 /* We only support one arg being in raw mode at the moment,
9051 * which is sufficient for the helper functions we have
9052 * right now.
9053 */
9054 return count <= 1;
9055 }
9056
check_args_pair_invalid(const struct bpf_func_proto * fn,int arg)9057 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
9058 {
9059 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
9060 bool has_size = fn->arg_size[arg] != 0;
9061 bool is_next_size = false;
9062
9063 if (arg + 1 < ARRAY_SIZE(fn->arg_type))
9064 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
9065
9066 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
9067 return is_next_size;
9068
9069 return has_size == is_next_size || is_next_size == is_fixed;
9070 }
9071
check_arg_pair_ok(const struct bpf_func_proto * fn)9072 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
9073 {
9074 /* bpf_xxx(..., buf, len) call will access 'len'
9075 * bytes from memory 'buf'. Both arg types need
9076 * to be paired, so make sure there's no buggy
9077 * helper function specification.
9078 */
9079 if (arg_type_is_mem_size(fn->arg1_type) ||
9080 check_args_pair_invalid(fn, 0) ||
9081 check_args_pair_invalid(fn, 1) ||
9082 check_args_pair_invalid(fn, 2) ||
9083 check_args_pair_invalid(fn, 3) ||
9084 check_args_pair_invalid(fn, 4))
9085 return false;
9086
9087 return true;
9088 }
9089
check_btf_id_ok(const struct bpf_func_proto * fn)9090 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
9091 {
9092 int i;
9093
9094 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
9095 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
9096 return !!fn->arg_btf_id[i];
9097 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
9098 return fn->arg_btf_id[i] == BPF_PTR_POISON;
9099 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
9100 /* arg_btf_id and arg_size are in a union. */
9101 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
9102 !(fn->arg_type[i] & MEM_FIXED_SIZE)))
9103 return false;
9104 }
9105
9106 return true;
9107 }
9108
check_func_proto(const struct bpf_func_proto * fn,int func_id)9109 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
9110 {
9111 return check_raw_mode_ok(fn) &&
9112 check_arg_pair_ok(fn) &&
9113 check_btf_id_ok(fn) ? 0 : -EINVAL;
9114 }
9115
9116 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
9117 * are now invalid, so turn them into unknown SCALAR_VALUE.
9118 *
9119 * This also applies to dynptr slices belonging to skb and xdp dynptrs,
9120 * since these slices point to packet data.
9121 */
clear_all_pkt_pointers(struct bpf_verifier_env * env)9122 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
9123 {
9124 struct bpf_func_state *state;
9125 struct bpf_reg_state *reg;
9126
9127 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
9128 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
9129 mark_reg_invalid(env, reg);
9130 }));
9131 }
9132
9133 enum {
9134 AT_PKT_END = -1,
9135 BEYOND_PKT_END = -2,
9136 };
9137
mark_pkt_end(struct bpf_verifier_state * vstate,int regn,bool range_open)9138 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
9139 {
9140 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9141 struct bpf_reg_state *reg = &state->regs[regn];
9142
9143 if (reg->type != PTR_TO_PACKET)
9144 /* PTR_TO_PACKET_META is not supported yet */
9145 return;
9146
9147 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
9148 * How far beyond pkt_end it goes is unknown.
9149 * if (!range_open) it's the case of pkt >= pkt_end
9150 * if (range_open) it's the case of pkt > pkt_end
9151 * hence this pointer is at least 1 byte bigger than pkt_end
9152 */
9153 if (range_open)
9154 reg->range = BEYOND_PKT_END;
9155 else
9156 reg->range = AT_PKT_END;
9157 }
9158
9159 /* The pointer with the specified id has released its reference to kernel
9160 * resources. Identify all copies of the same pointer and clear the reference.
9161 */
release_reference(struct bpf_verifier_env * env,int ref_obj_id)9162 static int release_reference(struct bpf_verifier_env *env,
9163 int ref_obj_id)
9164 {
9165 struct bpf_func_state *state;
9166 struct bpf_reg_state *reg;
9167 int err;
9168
9169 err = release_reference_state(cur_func(env), ref_obj_id);
9170 if (err)
9171 return err;
9172
9173 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
9174 if (reg->ref_obj_id == ref_obj_id)
9175 mark_reg_invalid(env, reg);
9176 }));
9177
9178 return 0;
9179 }
9180
invalidate_non_owning_refs(struct bpf_verifier_env * env)9181 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
9182 {
9183 struct bpf_func_state *unused;
9184 struct bpf_reg_state *reg;
9185
9186 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
9187 if (type_is_non_owning_ref(reg->type))
9188 mark_reg_invalid(env, reg);
9189 }));
9190 }
9191
clear_caller_saved_regs(struct bpf_verifier_env * env,struct bpf_reg_state * regs)9192 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
9193 struct bpf_reg_state *regs)
9194 {
9195 int i;
9196
9197 /* after the call registers r0 - r5 were scratched */
9198 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9199 mark_reg_not_init(env, regs, caller_saved[i]);
9200 __check_reg_arg(env, regs, caller_saved[i], DST_OP_NO_MARK);
9201 }
9202 }
9203
9204 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
9205 struct bpf_func_state *caller,
9206 struct bpf_func_state *callee,
9207 int insn_idx);
9208
9209 static int set_callee_state(struct bpf_verifier_env *env,
9210 struct bpf_func_state *caller,
9211 struct bpf_func_state *callee, int insn_idx);
9212
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)9213 static int setup_func_entry(struct bpf_verifier_env *env, int subprog, int callsite,
9214 set_callee_state_fn set_callee_state_cb,
9215 struct bpf_verifier_state *state)
9216 {
9217 struct bpf_func_state *caller, *callee;
9218 int err;
9219
9220 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
9221 verbose(env, "the call stack of %d frames is too deep\n",
9222 state->curframe + 2);
9223 return -E2BIG;
9224 }
9225
9226 if (state->frame[state->curframe + 1]) {
9227 verbose(env, "verifier bug. Frame %d already allocated\n",
9228 state->curframe + 1);
9229 return -EFAULT;
9230 }
9231
9232 caller = state->frame[state->curframe];
9233 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
9234 if (!callee)
9235 return -ENOMEM;
9236 state->frame[state->curframe + 1] = callee;
9237
9238 /* callee cannot access r0, r6 - r9 for reading and has to write
9239 * into its own stack before reading from it.
9240 * callee can read/write into caller's stack
9241 */
9242 init_func_state(env, callee,
9243 /* remember the callsite, it will be used by bpf_exit */
9244 callsite,
9245 state->curframe + 1 /* frameno within this callchain */,
9246 subprog /* subprog number within this prog */);
9247 /* Transfer references to the callee */
9248 err = copy_reference_state(callee, caller);
9249 err = err ?: set_callee_state_cb(env, caller, callee, callsite);
9250 if (err)
9251 goto err_out;
9252
9253 /* only increment it after check_reg_arg() finished */
9254 state->curframe++;
9255
9256 return 0;
9257
9258 err_out:
9259 free_func_state(callee);
9260 state->frame[state->curframe + 1] = NULL;
9261 return err;
9262 }
9263
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)9264 static int push_callback_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9265 int insn_idx, int subprog,
9266 set_callee_state_fn set_callee_state_cb)
9267 {
9268 struct bpf_verifier_state *state = env->cur_state, *callback_state;
9269 struct bpf_func_state *caller, *callee;
9270 int err;
9271
9272 caller = state->frame[state->curframe];
9273 err = btf_check_subprog_call(env, subprog, caller->regs);
9274 if (err == -EFAULT)
9275 return err;
9276
9277 /* set_callee_state is used for direct subprog calls, but we are
9278 * interested in validating only BPF helpers that can call subprogs as
9279 * callbacks
9280 */
9281 if (bpf_pseudo_kfunc_call(insn) &&
9282 !is_sync_callback_calling_kfunc(insn->imm)) {
9283 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n",
9284 func_id_name(insn->imm), insn->imm);
9285 return -EFAULT;
9286 } else if (!bpf_pseudo_kfunc_call(insn) &&
9287 !is_callback_calling_function(insn->imm)) { /* helper */
9288 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n",
9289 func_id_name(insn->imm), insn->imm);
9290 return -EFAULT;
9291 }
9292
9293 if (insn->code == (BPF_JMP | BPF_CALL) &&
9294 insn->src_reg == 0 &&
9295 insn->imm == BPF_FUNC_timer_set_callback) {
9296 struct bpf_verifier_state *async_cb;
9297
9298 /* there is no real recursion here. timer callbacks are async */
9299 env->subprog_info[subprog].is_async_cb = true;
9300 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
9301 insn_idx, subprog);
9302 if (!async_cb)
9303 return -EFAULT;
9304 callee = async_cb->frame[0];
9305 callee->async_entry_cnt = caller->async_entry_cnt + 1;
9306
9307 /* Convert bpf_timer_set_callback() args into timer callback args */
9308 err = set_callee_state_cb(env, caller, callee, insn_idx);
9309 if (err)
9310 return err;
9311
9312 return 0;
9313 }
9314
9315 /* for callback functions enqueue entry to callback and
9316 * proceed with next instruction within current frame.
9317 */
9318 callback_state = push_stack(env, env->subprog_info[subprog].start, insn_idx, false);
9319 if (!callback_state)
9320 return -ENOMEM;
9321
9322 err = setup_func_entry(env, subprog, insn_idx, set_callee_state_cb,
9323 callback_state);
9324 if (err)
9325 return err;
9326
9327 callback_state->callback_unroll_depth++;
9328 callback_state->frame[callback_state->curframe - 1]->callback_depth++;
9329 caller->callback_depth = 0;
9330 return 0;
9331 }
9332
check_func_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)9333 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9334 int *insn_idx)
9335 {
9336 struct bpf_verifier_state *state = env->cur_state;
9337 struct bpf_func_state *caller;
9338 int err, subprog, target_insn;
9339
9340 target_insn = *insn_idx + insn->imm + 1;
9341 subprog = find_subprog(env, target_insn);
9342 if (subprog < 0) {
9343 verbose(env, "verifier bug. No program starts at insn %d\n", target_insn);
9344 return -EFAULT;
9345 }
9346
9347 caller = state->frame[state->curframe];
9348 err = btf_check_subprog_call(env, subprog, caller->regs);
9349 if (err == -EFAULT)
9350 return err;
9351 if (subprog_is_global(env, subprog)) {
9352 if (err) {
9353 verbose(env, "Caller passes invalid args into func#%d\n", subprog);
9354 return err;
9355 }
9356
9357 if (env->log.level & BPF_LOG_LEVEL)
9358 verbose(env, "Func#%d is global and valid. Skipping.\n", subprog);
9359 clear_caller_saved_regs(env, caller->regs);
9360
9361 /* All global functions return a 64-bit SCALAR_VALUE */
9362 mark_reg_unknown(env, caller->regs, BPF_REG_0);
9363 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
9364
9365 /* continue with next insn after call */
9366 return 0;
9367 }
9368
9369 /* for regular function entry setup new frame and continue
9370 * from that frame.
9371 */
9372 err = setup_func_entry(env, subprog, *insn_idx, set_callee_state, state);
9373 if (err)
9374 return err;
9375
9376 clear_caller_saved_regs(env, caller->regs);
9377
9378 /* and go analyze first insn of the callee */
9379 *insn_idx = env->subprog_info[subprog].start - 1;
9380
9381 if (env->log.level & BPF_LOG_LEVEL) {
9382 verbose(env, "caller:\n");
9383 print_verifier_state(env, caller, true);
9384 verbose(env, "callee:\n");
9385 print_verifier_state(env, state->frame[state->curframe], true);
9386 }
9387
9388 return 0;
9389 }
9390
map_set_for_each_callback_args(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee)9391 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
9392 struct bpf_func_state *caller,
9393 struct bpf_func_state *callee)
9394 {
9395 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
9396 * void *callback_ctx, u64 flags);
9397 * callback_fn(struct bpf_map *map, void *key, void *value,
9398 * void *callback_ctx);
9399 */
9400 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9401
9402 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9403 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9404 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9405
9406 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9407 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9408 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
9409
9410 /* pointer to stack or null */
9411 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
9412
9413 /* unused */
9414 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9415 return 0;
9416 }
9417
set_callee_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9418 static int set_callee_state(struct bpf_verifier_env *env,
9419 struct bpf_func_state *caller,
9420 struct bpf_func_state *callee, int insn_idx)
9421 {
9422 int i;
9423
9424 /* copy r1 - r5 args that callee can access. The copy includes parent
9425 * pointers, which connects us up to the liveness chain
9426 */
9427 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
9428 callee->regs[i] = caller->regs[i];
9429 return 0;
9430 }
9431
set_map_elem_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9432 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
9433 struct bpf_func_state *caller,
9434 struct bpf_func_state *callee,
9435 int insn_idx)
9436 {
9437 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
9438 struct bpf_map *map;
9439 int err;
9440
9441 if (bpf_map_ptr_poisoned(insn_aux)) {
9442 verbose(env, "tail_call abusing map_ptr\n");
9443 return -EINVAL;
9444 }
9445
9446 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
9447 if (!map->ops->map_set_for_each_callback_args ||
9448 !map->ops->map_for_each_callback) {
9449 verbose(env, "callback function not allowed for map\n");
9450 return -ENOTSUPP;
9451 }
9452
9453 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
9454 if (err)
9455 return err;
9456
9457 callee->in_callback_fn = true;
9458 callee->callback_ret_range = tnum_range(0, 1);
9459 return 0;
9460 }
9461
set_loop_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9462 static int set_loop_callback_state(struct bpf_verifier_env *env,
9463 struct bpf_func_state *caller,
9464 struct bpf_func_state *callee,
9465 int insn_idx)
9466 {
9467 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
9468 * u64 flags);
9469 * callback_fn(u32 index, void *callback_ctx);
9470 */
9471 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
9472 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9473
9474 /* unused */
9475 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9476 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9477 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9478
9479 callee->in_callback_fn = true;
9480 callee->callback_ret_range = tnum_range(0, 1);
9481 return 0;
9482 }
9483
set_timer_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9484 static int set_timer_callback_state(struct bpf_verifier_env *env,
9485 struct bpf_func_state *caller,
9486 struct bpf_func_state *callee,
9487 int insn_idx)
9488 {
9489 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
9490
9491 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
9492 * callback_fn(struct bpf_map *map, void *key, void *value);
9493 */
9494 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
9495 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
9496 callee->regs[BPF_REG_1].map_ptr = map_ptr;
9497
9498 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
9499 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9500 callee->regs[BPF_REG_2].map_ptr = map_ptr;
9501
9502 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
9503 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
9504 callee->regs[BPF_REG_3].map_ptr = map_ptr;
9505
9506 /* unused */
9507 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9508 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9509 callee->in_async_callback_fn = true;
9510 callee->callback_ret_range = tnum_range(0, 1);
9511 return 0;
9512 }
9513
set_find_vma_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9514 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
9515 struct bpf_func_state *caller,
9516 struct bpf_func_state *callee,
9517 int insn_idx)
9518 {
9519 /* bpf_find_vma(struct task_struct *task, u64 addr,
9520 * void *callback_fn, void *callback_ctx, u64 flags)
9521 * (callback_fn)(struct task_struct *task,
9522 * struct vm_area_struct *vma, void *callback_ctx);
9523 */
9524 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
9525
9526 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
9527 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
9528 callee->regs[BPF_REG_2].btf = btf_vmlinux;
9529 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
9530
9531 /* pointer to stack or null */
9532 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
9533
9534 /* unused */
9535 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9536 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9537 callee->in_callback_fn = true;
9538 callee->callback_ret_range = tnum_range(0, 1);
9539 return 0;
9540 }
9541
set_user_ringbuf_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9542 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
9543 struct bpf_func_state *caller,
9544 struct bpf_func_state *callee,
9545 int insn_idx)
9546 {
9547 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
9548 * callback_ctx, u64 flags);
9549 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
9550 */
9551 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
9552 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
9553 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
9554
9555 /* unused */
9556 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9557 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9558 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9559
9560 callee->in_callback_fn = true;
9561 callee->callback_ret_range = tnum_range(0, 1);
9562 return 0;
9563 }
9564
set_rbtree_add_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)9565 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
9566 struct bpf_func_state *caller,
9567 struct bpf_func_state *callee,
9568 int insn_idx)
9569 {
9570 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
9571 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
9572 *
9573 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
9574 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
9575 * by this point, so look at 'root'
9576 */
9577 struct btf_field *field;
9578
9579 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
9580 BPF_RB_ROOT);
9581 if (!field || !field->graph_root.value_btf_id)
9582 return -EFAULT;
9583
9584 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
9585 ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
9586 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
9587 ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
9588
9589 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
9590 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
9591 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
9592 callee->in_callback_fn = true;
9593 callee->callback_ret_range = tnum_range(0, 1);
9594 return 0;
9595 }
9596
9597 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
9598
9599 /* Are we currently verifying the callback for a rbtree helper that must
9600 * be called with lock held? If so, no need to complain about unreleased
9601 * lock
9602 */
in_rbtree_lock_required_cb(struct bpf_verifier_env * env)9603 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
9604 {
9605 struct bpf_verifier_state *state = env->cur_state;
9606 struct bpf_insn *insn = env->prog->insnsi;
9607 struct bpf_func_state *callee;
9608 int kfunc_btf_id;
9609
9610 if (!state->curframe)
9611 return false;
9612
9613 callee = state->frame[state->curframe];
9614
9615 if (!callee->in_callback_fn)
9616 return false;
9617
9618 kfunc_btf_id = insn[callee->callsite].imm;
9619 return is_rbtree_lock_required_kfunc(kfunc_btf_id);
9620 }
9621
prepare_func_exit(struct bpf_verifier_env * env,int * insn_idx)9622 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
9623 {
9624 struct bpf_verifier_state *state = env->cur_state, *prev_st;
9625 struct bpf_func_state *caller, *callee;
9626 struct bpf_reg_state *r0;
9627 bool in_callback_fn;
9628 int err;
9629
9630 callee = state->frame[state->curframe];
9631 r0 = &callee->regs[BPF_REG_0];
9632 if (r0->type == PTR_TO_STACK) {
9633 /* technically it's ok to return caller's stack pointer
9634 * (or caller's caller's pointer) back to the caller,
9635 * since these pointers are valid. Only current stack
9636 * pointer will be invalid as soon as function exits,
9637 * but let's be conservative
9638 */
9639 verbose(env, "cannot return stack pointer to the caller\n");
9640 return -EINVAL;
9641 }
9642
9643 caller = state->frame[state->curframe - 1];
9644 if (callee->in_callback_fn) {
9645 /* enforce R0 return value range [0, 1]. */
9646 struct tnum range = callee->callback_ret_range;
9647
9648 if (r0->type != SCALAR_VALUE) {
9649 verbose(env, "R0 not a scalar value\n");
9650 return -EACCES;
9651 }
9652
9653 /* we are going to rely on register's precise value */
9654 err = mark_reg_read(env, r0, r0->parent, REG_LIVE_READ64);
9655 err = err ?: mark_chain_precision(env, BPF_REG_0);
9656 if (err)
9657 return err;
9658
9659 if (!tnum_in(range, r0->var_off)) {
9660 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
9661 return -EINVAL;
9662 }
9663 if (!calls_callback(env, callee->callsite)) {
9664 verbose(env, "BUG: in callback at %d, callsite %d !calls_callback\n",
9665 *insn_idx, callee->callsite);
9666 return -EFAULT;
9667 }
9668 } else {
9669 /* return to the caller whatever r0 had in the callee */
9670 caller->regs[BPF_REG_0] = *r0;
9671 }
9672
9673 /* callback_fn frame should have released its own additions to parent's
9674 * reference state at this point, or check_reference_leak would
9675 * complain, hence it must be the same as the caller. There is no need
9676 * to copy it back.
9677 */
9678 if (!callee->in_callback_fn) {
9679 /* Transfer references to the caller */
9680 err = copy_reference_state(caller, callee);
9681 if (err)
9682 return err;
9683 }
9684
9685 /* for callbacks like bpf_loop or bpf_for_each_map_elem go back to callsite,
9686 * there function call logic would reschedule callback visit. If iteration
9687 * converges is_state_visited() would prune that visit eventually.
9688 */
9689 in_callback_fn = callee->in_callback_fn;
9690 if (in_callback_fn)
9691 *insn_idx = callee->callsite;
9692 else
9693 *insn_idx = callee->callsite + 1;
9694
9695 if (env->log.level & BPF_LOG_LEVEL) {
9696 verbose(env, "returning from callee:\n");
9697 print_verifier_state(env, callee, true);
9698 verbose(env, "to caller at %d:\n", *insn_idx);
9699 print_verifier_state(env, caller, true);
9700 }
9701 /* clear everything in the callee */
9702 free_func_state(callee);
9703 state->frame[state->curframe--] = NULL;
9704
9705 /* for callbacks widen imprecise scalars to make programs like below verify:
9706 *
9707 * struct ctx { int i; }
9708 * void cb(int idx, struct ctx *ctx) { ctx->i++; ... }
9709 * ...
9710 * struct ctx = { .i = 0; }
9711 * bpf_loop(100, cb, &ctx, 0);
9712 *
9713 * This is similar to what is done in process_iter_next_call() for open
9714 * coded iterators.
9715 */
9716 prev_st = in_callback_fn ? find_prev_entry(env, state, *insn_idx) : NULL;
9717 if (prev_st) {
9718 err = widen_imprecise_scalars(env, prev_st, state);
9719 if (err)
9720 return err;
9721 }
9722 return 0;
9723 }
9724
do_refine_retval_range(struct bpf_reg_state * regs,int ret_type,int func_id,struct bpf_call_arg_meta * meta)9725 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
9726 int func_id,
9727 struct bpf_call_arg_meta *meta)
9728 {
9729 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
9730
9731 if (ret_type != RET_INTEGER)
9732 return;
9733
9734 switch (func_id) {
9735 case BPF_FUNC_get_stack:
9736 case BPF_FUNC_get_task_stack:
9737 case BPF_FUNC_probe_read_str:
9738 case BPF_FUNC_probe_read_kernel_str:
9739 case BPF_FUNC_probe_read_user_str:
9740 ret_reg->smax_value = meta->msize_max_value;
9741 ret_reg->s32_max_value = meta->msize_max_value;
9742 ret_reg->smin_value = -MAX_ERRNO;
9743 ret_reg->s32_min_value = -MAX_ERRNO;
9744 reg_bounds_sync(ret_reg);
9745 break;
9746 case BPF_FUNC_get_smp_processor_id:
9747 ret_reg->umax_value = nr_cpu_ids - 1;
9748 ret_reg->u32_max_value = nr_cpu_ids - 1;
9749 ret_reg->smax_value = nr_cpu_ids - 1;
9750 ret_reg->s32_max_value = nr_cpu_ids - 1;
9751 ret_reg->umin_value = 0;
9752 ret_reg->u32_min_value = 0;
9753 ret_reg->smin_value = 0;
9754 ret_reg->s32_min_value = 0;
9755 reg_bounds_sync(ret_reg);
9756 break;
9757 }
9758 }
9759
9760 static int
record_func_map(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)9761 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9762 int func_id, int insn_idx)
9763 {
9764 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9765 struct bpf_map *map = meta->map_ptr;
9766
9767 if (func_id != BPF_FUNC_tail_call &&
9768 func_id != BPF_FUNC_map_lookup_elem &&
9769 func_id != BPF_FUNC_map_update_elem &&
9770 func_id != BPF_FUNC_map_delete_elem &&
9771 func_id != BPF_FUNC_map_push_elem &&
9772 func_id != BPF_FUNC_map_pop_elem &&
9773 func_id != BPF_FUNC_map_peek_elem &&
9774 func_id != BPF_FUNC_for_each_map_elem &&
9775 func_id != BPF_FUNC_redirect_map &&
9776 func_id != BPF_FUNC_map_lookup_percpu_elem)
9777 return 0;
9778
9779 if (map == NULL) {
9780 verbose(env, "kernel subsystem misconfigured verifier\n");
9781 return -EINVAL;
9782 }
9783
9784 /* In case of read-only, some additional restrictions
9785 * need to be applied in order to prevent altering the
9786 * state of the map from program side.
9787 */
9788 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
9789 (func_id == BPF_FUNC_map_delete_elem ||
9790 func_id == BPF_FUNC_map_update_elem ||
9791 func_id == BPF_FUNC_map_push_elem ||
9792 func_id == BPF_FUNC_map_pop_elem)) {
9793 verbose(env, "write into map forbidden\n");
9794 return -EACCES;
9795 }
9796
9797 if (!BPF_MAP_PTR(aux->map_ptr_state))
9798 bpf_map_ptr_store(aux, meta->map_ptr,
9799 !meta->map_ptr->bypass_spec_v1);
9800 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
9801 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
9802 !meta->map_ptr->bypass_spec_v1);
9803 return 0;
9804 }
9805
9806 static int
record_func_key(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)9807 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
9808 int func_id, int insn_idx)
9809 {
9810 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
9811 struct bpf_reg_state *regs = cur_regs(env), *reg;
9812 struct bpf_map *map = meta->map_ptr;
9813 u64 val, max;
9814 int err;
9815
9816 if (func_id != BPF_FUNC_tail_call)
9817 return 0;
9818 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
9819 verbose(env, "kernel subsystem misconfigured verifier\n");
9820 return -EINVAL;
9821 }
9822
9823 reg = ®s[BPF_REG_3];
9824 val = reg->var_off.value;
9825 max = map->max_entries;
9826
9827 if (!(register_is_const(reg) && val < max)) {
9828 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9829 return 0;
9830 }
9831
9832 err = mark_chain_precision(env, BPF_REG_3);
9833 if (err)
9834 return err;
9835 if (bpf_map_key_unseen(aux))
9836 bpf_map_key_store(aux, val);
9837 else if (!bpf_map_key_poisoned(aux) &&
9838 bpf_map_key_immediate(aux) != val)
9839 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
9840 return 0;
9841 }
9842
check_reference_leak(struct bpf_verifier_env * env)9843 static int check_reference_leak(struct bpf_verifier_env *env)
9844 {
9845 struct bpf_func_state *state = cur_func(env);
9846 bool refs_lingering = false;
9847 int i;
9848
9849 if (state->frameno && !state->in_callback_fn)
9850 return 0;
9851
9852 for (i = 0; i < state->acquired_refs; i++) {
9853 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
9854 continue;
9855 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
9856 state->refs[i].id, state->refs[i].insn_idx);
9857 refs_lingering = true;
9858 }
9859 return refs_lingering ? -EINVAL : 0;
9860 }
9861
check_bpf_snprintf_call(struct bpf_verifier_env * env,struct bpf_reg_state * regs)9862 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
9863 struct bpf_reg_state *regs)
9864 {
9865 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
9866 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
9867 struct bpf_map *fmt_map = fmt_reg->map_ptr;
9868 struct bpf_bprintf_data data = {};
9869 int err, fmt_map_off, num_args;
9870 u64 fmt_addr;
9871 char *fmt;
9872
9873 /* data must be an array of u64 */
9874 if (data_len_reg->var_off.value % 8)
9875 return -EINVAL;
9876 num_args = data_len_reg->var_off.value / 8;
9877
9878 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
9879 * and map_direct_value_addr is set.
9880 */
9881 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
9882 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
9883 fmt_map_off);
9884 if (err) {
9885 verbose(env, "verifier bug\n");
9886 return -EFAULT;
9887 }
9888 fmt = (char *)(long)fmt_addr + fmt_map_off;
9889
9890 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
9891 * can focus on validating the format specifiers.
9892 */
9893 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
9894 if (err < 0)
9895 verbose(env, "Invalid format string\n");
9896
9897 return err;
9898 }
9899
check_get_func_ip(struct bpf_verifier_env * env)9900 static int check_get_func_ip(struct bpf_verifier_env *env)
9901 {
9902 enum bpf_prog_type type = resolve_prog_type(env->prog);
9903 int func_id = BPF_FUNC_get_func_ip;
9904
9905 if (type == BPF_PROG_TYPE_TRACING) {
9906 if (!bpf_prog_has_trampoline(env->prog)) {
9907 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
9908 func_id_name(func_id), func_id);
9909 return -ENOTSUPP;
9910 }
9911 return 0;
9912 } else if (type == BPF_PROG_TYPE_KPROBE) {
9913 return 0;
9914 }
9915
9916 verbose(env, "func %s#%d not supported for program type %d\n",
9917 func_id_name(func_id), func_id, type);
9918 return -ENOTSUPP;
9919 }
9920
cur_aux(struct bpf_verifier_env * env)9921 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
9922 {
9923 return &env->insn_aux_data[env->insn_idx];
9924 }
9925
loop_flag_is_zero(struct bpf_verifier_env * env)9926 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
9927 {
9928 struct bpf_reg_state *regs = cur_regs(env);
9929 struct bpf_reg_state *reg = ®s[BPF_REG_4];
9930 bool reg_is_null = register_is_null(reg);
9931
9932 if (reg_is_null)
9933 mark_chain_precision(env, BPF_REG_4);
9934
9935 return reg_is_null;
9936 }
9937
update_loop_inline_state(struct bpf_verifier_env * env,u32 subprogno)9938 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
9939 {
9940 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
9941
9942 if (!state->initialized) {
9943 state->initialized = 1;
9944 state->fit_for_inline = loop_flag_is_zero(env);
9945 state->callback_subprogno = subprogno;
9946 return;
9947 }
9948
9949 if (!state->fit_for_inline)
9950 return;
9951
9952 state->fit_for_inline = (loop_flag_is_zero(env) &&
9953 state->callback_subprogno == subprogno);
9954 }
9955
check_helper_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx_p)9956 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
9957 int *insn_idx_p)
9958 {
9959 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9960 const struct bpf_func_proto *fn = NULL;
9961 enum bpf_return_type ret_type;
9962 enum bpf_type_flag ret_flag;
9963 struct bpf_reg_state *regs;
9964 struct bpf_call_arg_meta meta;
9965 int insn_idx = *insn_idx_p;
9966 bool changes_data;
9967 int i, err, func_id;
9968
9969 /* find function prototype */
9970 func_id = insn->imm;
9971 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
9972 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
9973 func_id);
9974 return -EINVAL;
9975 }
9976
9977 if (env->ops->get_func_proto)
9978 fn = env->ops->get_func_proto(func_id, env->prog);
9979 if (!fn) {
9980 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
9981 func_id);
9982 return -EINVAL;
9983 }
9984
9985 /* eBPF programs must be GPL compatible to use GPL-ed functions */
9986 if (!env->prog->gpl_compatible && fn->gpl_only) {
9987 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
9988 return -EINVAL;
9989 }
9990
9991 if (fn->allowed && !fn->allowed(env->prog)) {
9992 verbose(env, "helper call is not allowed in probe\n");
9993 return -EINVAL;
9994 }
9995
9996 if (!env->prog->aux->sleepable && fn->might_sleep) {
9997 verbose(env, "helper call might sleep in a non-sleepable prog\n");
9998 return -EINVAL;
9999 }
10000
10001 /* With LD_ABS/IND some JITs save/restore skb from r1. */
10002 changes_data = bpf_helper_changes_pkt_data(fn->func);
10003 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
10004 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
10005 func_id_name(func_id), func_id);
10006 return -EINVAL;
10007 }
10008
10009 memset(&meta, 0, sizeof(meta));
10010 meta.pkt_access = fn->pkt_access;
10011
10012 err = check_func_proto(fn, func_id);
10013 if (err) {
10014 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
10015 func_id_name(func_id), func_id);
10016 return err;
10017 }
10018
10019 if (env->cur_state->active_rcu_lock) {
10020 if (fn->might_sleep) {
10021 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
10022 func_id_name(func_id), func_id);
10023 return -EINVAL;
10024 }
10025
10026 if (env->prog->aux->sleepable && is_storage_get_function(func_id))
10027 env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
10028 }
10029
10030 meta.func_id = func_id;
10031 /* check args */
10032 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
10033 err = check_func_arg(env, i, &meta, fn, insn_idx);
10034 if (err)
10035 return err;
10036 }
10037
10038 err = record_func_map(env, &meta, func_id, insn_idx);
10039 if (err)
10040 return err;
10041
10042 err = record_func_key(env, &meta, func_id, insn_idx);
10043 if (err)
10044 return err;
10045
10046 /* Mark slots with STACK_MISC in case of raw mode, stack offset
10047 * is inferred from register state.
10048 */
10049 for (i = 0; i < meta.access_size; i++) {
10050 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
10051 BPF_WRITE, -1, false, false);
10052 if (err)
10053 return err;
10054 }
10055
10056 regs = cur_regs(env);
10057
10058 if (meta.release_regno) {
10059 err = -EINVAL;
10060 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
10061 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
10062 * is safe to do directly.
10063 */
10064 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
10065 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
10066 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n");
10067 return -EFAULT;
10068 }
10069 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]);
10070 } else if (meta.ref_obj_id) {
10071 err = release_reference(env, meta.ref_obj_id);
10072 } else if (register_is_null(®s[meta.release_regno])) {
10073 /* meta.ref_obj_id can only be 0 if register that is meant to be
10074 * released is NULL, which must be > R0.
10075 */
10076 err = 0;
10077 }
10078 if (err) {
10079 verbose(env, "func %s#%d reference has not been acquired before\n",
10080 func_id_name(func_id), func_id);
10081 return err;
10082 }
10083 }
10084
10085 switch (func_id) {
10086 case BPF_FUNC_tail_call:
10087 err = check_reference_leak(env);
10088 if (err) {
10089 verbose(env, "tail_call would lead to reference leak\n");
10090 return err;
10091 }
10092 break;
10093 case BPF_FUNC_get_local_storage:
10094 /* check that flags argument in get_local_storage(map, flags) is 0,
10095 * this is required because get_local_storage() can't return an error.
10096 */
10097 if (!register_is_null(®s[BPF_REG_2])) {
10098 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
10099 return -EINVAL;
10100 }
10101 break;
10102 case BPF_FUNC_for_each_map_elem:
10103 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10104 set_map_elem_callback_state);
10105 break;
10106 case BPF_FUNC_timer_set_callback:
10107 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10108 set_timer_callback_state);
10109 break;
10110 case BPF_FUNC_find_vma:
10111 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10112 set_find_vma_callback_state);
10113 break;
10114 case BPF_FUNC_snprintf:
10115 err = check_bpf_snprintf_call(env, regs);
10116 break;
10117 case BPF_FUNC_loop:
10118 update_loop_inline_state(env, meta.subprogno);
10119 /* Verifier relies on R1 value to determine if bpf_loop() iteration
10120 * is finished, thus mark it precise.
10121 */
10122 err = mark_chain_precision(env, BPF_REG_1);
10123 if (err)
10124 return err;
10125 if (cur_func(env)->callback_depth < regs[BPF_REG_1].umax_value) {
10126 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10127 set_loop_callback_state);
10128 } else {
10129 cur_func(env)->callback_depth = 0;
10130 if (env->log.level & BPF_LOG_LEVEL2)
10131 verbose(env, "frame%d bpf_loop iteration limit reached\n",
10132 env->cur_state->curframe);
10133 }
10134 break;
10135 case BPF_FUNC_dynptr_from_mem:
10136 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
10137 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
10138 reg_type_str(env, regs[BPF_REG_1].type));
10139 return -EACCES;
10140 }
10141 break;
10142 case BPF_FUNC_set_retval:
10143 if (prog_type == BPF_PROG_TYPE_LSM &&
10144 env->prog->expected_attach_type == BPF_LSM_CGROUP) {
10145 if (!env->prog->aux->attach_func_proto->type) {
10146 /* Make sure programs that attach to void
10147 * hooks don't try to modify return value.
10148 */
10149 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
10150 return -EINVAL;
10151 }
10152 }
10153 break;
10154 case BPF_FUNC_dynptr_data:
10155 {
10156 struct bpf_reg_state *reg;
10157 int id, ref_obj_id;
10158
10159 reg = get_dynptr_arg_reg(env, fn, regs);
10160 if (!reg)
10161 return -EFAULT;
10162
10163
10164 if (meta.dynptr_id) {
10165 verbose(env, "verifier internal error: meta.dynptr_id already set\n");
10166 return -EFAULT;
10167 }
10168 if (meta.ref_obj_id) {
10169 verbose(env, "verifier internal error: meta.ref_obj_id already set\n");
10170 return -EFAULT;
10171 }
10172
10173 id = dynptr_id(env, reg);
10174 if (id < 0) {
10175 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
10176 return id;
10177 }
10178
10179 ref_obj_id = dynptr_ref_obj_id(env, reg);
10180 if (ref_obj_id < 0) {
10181 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n");
10182 return ref_obj_id;
10183 }
10184
10185 meta.dynptr_id = id;
10186 meta.ref_obj_id = ref_obj_id;
10187
10188 break;
10189 }
10190 case BPF_FUNC_dynptr_write:
10191 {
10192 enum bpf_dynptr_type dynptr_type;
10193 struct bpf_reg_state *reg;
10194
10195 reg = get_dynptr_arg_reg(env, fn, regs);
10196 if (!reg)
10197 return -EFAULT;
10198
10199 dynptr_type = dynptr_get_type(env, reg);
10200 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
10201 return -EFAULT;
10202
10203 if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
10204 /* this will trigger clear_all_pkt_pointers(), which will
10205 * invalidate all dynptr slices associated with the skb
10206 */
10207 changes_data = true;
10208
10209 break;
10210 }
10211 case BPF_FUNC_user_ringbuf_drain:
10212 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
10213 set_user_ringbuf_callback_state);
10214 break;
10215 }
10216
10217 if (err)
10218 return err;
10219
10220 /* reset caller saved regs */
10221 for (i = 0; i < CALLER_SAVED_REGS; i++) {
10222 mark_reg_not_init(env, regs, caller_saved[i]);
10223 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
10224 }
10225
10226 /* helper call returns 64-bit value. */
10227 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
10228
10229 /* update return register (already marked as written above) */
10230 ret_type = fn->ret_type;
10231 ret_flag = type_flag(ret_type);
10232
10233 switch (base_type(ret_type)) {
10234 case RET_INTEGER:
10235 /* sets type to SCALAR_VALUE */
10236 mark_reg_unknown(env, regs, BPF_REG_0);
10237 break;
10238 case RET_VOID:
10239 regs[BPF_REG_0].type = NOT_INIT;
10240 break;
10241 case RET_PTR_TO_MAP_VALUE:
10242 /* There is no offset yet applied, variable or fixed */
10243 mark_reg_known_zero(env, regs, BPF_REG_0);
10244 /* remember map_ptr, so that check_map_access()
10245 * can check 'value_size' boundary of memory access
10246 * to map element returned from bpf_map_lookup_elem()
10247 */
10248 if (meta.map_ptr == NULL) {
10249 verbose(env,
10250 "kernel subsystem misconfigured verifier\n");
10251 return -EINVAL;
10252 }
10253 regs[BPF_REG_0].map_ptr = meta.map_ptr;
10254 regs[BPF_REG_0].map_uid = meta.map_uid;
10255 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
10256 if (!type_may_be_null(ret_type) &&
10257 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) {
10258 regs[BPF_REG_0].id = ++env->id_gen;
10259 }
10260 break;
10261 case RET_PTR_TO_SOCKET:
10262 mark_reg_known_zero(env, regs, BPF_REG_0);
10263 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
10264 break;
10265 case RET_PTR_TO_SOCK_COMMON:
10266 mark_reg_known_zero(env, regs, BPF_REG_0);
10267 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
10268 break;
10269 case RET_PTR_TO_TCP_SOCK:
10270 mark_reg_known_zero(env, regs, BPF_REG_0);
10271 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
10272 break;
10273 case RET_PTR_TO_MEM:
10274 mark_reg_known_zero(env, regs, BPF_REG_0);
10275 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
10276 regs[BPF_REG_0].mem_size = meta.mem_size;
10277 break;
10278 case RET_PTR_TO_MEM_OR_BTF_ID:
10279 {
10280 const struct btf_type *t;
10281
10282 mark_reg_known_zero(env, regs, BPF_REG_0);
10283 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
10284 if (!btf_type_is_struct(t)) {
10285 u32 tsize;
10286 const struct btf_type *ret;
10287 const char *tname;
10288
10289 /* resolve the type size of ksym. */
10290 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
10291 if (IS_ERR(ret)) {
10292 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
10293 verbose(env, "unable to resolve the size of type '%s': %ld\n",
10294 tname, PTR_ERR(ret));
10295 return -EINVAL;
10296 }
10297 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
10298 regs[BPF_REG_0].mem_size = tsize;
10299 } else {
10300 /* MEM_RDONLY may be carried from ret_flag, but it
10301 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
10302 * it will confuse the check of PTR_TO_BTF_ID in
10303 * check_mem_access().
10304 */
10305 ret_flag &= ~MEM_RDONLY;
10306
10307 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
10308 regs[BPF_REG_0].btf = meta.ret_btf;
10309 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
10310 }
10311 break;
10312 }
10313 case RET_PTR_TO_BTF_ID:
10314 {
10315 struct btf *ret_btf;
10316 int ret_btf_id;
10317
10318 mark_reg_known_zero(env, regs, BPF_REG_0);
10319 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
10320 if (func_id == BPF_FUNC_kptr_xchg) {
10321 ret_btf = meta.kptr_field->kptr.btf;
10322 ret_btf_id = meta.kptr_field->kptr.btf_id;
10323 if (!btf_is_kernel(ret_btf))
10324 regs[BPF_REG_0].type |= MEM_ALLOC;
10325 } else {
10326 if (fn->ret_btf_id == BPF_PTR_POISON) {
10327 verbose(env, "verifier internal error:");
10328 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n",
10329 func_id_name(func_id));
10330 return -EINVAL;
10331 }
10332 ret_btf = btf_vmlinux;
10333 ret_btf_id = *fn->ret_btf_id;
10334 }
10335 if (ret_btf_id == 0) {
10336 verbose(env, "invalid return type %u of func %s#%d\n",
10337 base_type(ret_type), func_id_name(func_id),
10338 func_id);
10339 return -EINVAL;
10340 }
10341 regs[BPF_REG_0].btf = ret_btf;
10342 regs[BPF_REG_0].btf_id = ret_btf_id;
10343 break;
10344 }
10345 default:
10346 verbose(env, "unknown return type %u of func %s#%d\n",
10347 base_type(ret_type), func_id_name(func_id), func_id);
10348 return -EINVAL;
10349 }
10350
10351 if (type_may_be_null(regs[BPF_REG_0].type))
10352 regs[BPF_REG_0].id = ++env->id_gen;
10353
10354 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
10355 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n",
10356 func_id_name(func_id), func_id);
10357 return -EFAULT;
10358 }
10359
10360 if (is_dynptr_ref_function(func_id))
10361 regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
10362
10363 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
10364 /* For release_reference() */
10365 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
10366 } else if (is_acquire_function(func_id, meta.map_ptr)) {
10367 int id = acquire_reference_state(env, insn_idx);
10368
10369 if (id < 0)
10370 return id;
10371 /* For mark_ptr_or_null_reg() */
10372 regs[BPF_REG_0].id = id;
10373 /* For release_reference() */
10374 regs[BPF_REG_0].ref_obj_id = id;
10375 }
10376
10377 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
10378
10379 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
10380 if (err)
10381 return err;
10382
10383 if ((func_id == BPF_FUNC_get_stack ||
10384 func_id == BPF_FUNC_get_task_stack) &&
10385 !env->prog->has_callchain_buf) {
10386 const char *err_str;
10387
10388 #ifdef CONFIG_PERF_EVENTS
10389 err = get_callchain_buffers(sysctl_perf_event_max_stack);
10390 err_str = "cannot get callchain buffer for func %s#%d\n";
10391 #else
10392 err = -ENOTSUPP;
10393 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
10394 #endif
10395 if (err) {
10396 verbose(env, err_str, func_id_name(func_id), func_id);
10397 return err;
10398 }
10399
10400 env->prog->has_callchain_buf = true;
10401 }
10402
10403 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
10404 env->prog->call_get_stack = true;
10405
10406 if (func_id == BPF_FUNC_get_func_ip) {
10407 if (check_get_func_ip(env))
10408 return -ENOTSUPP;
10409 env->prog->call_get_func_ip = true;
10410 }
10411
10412 if (changes_data)
10413 clear_all_pkt_pointers(env);
10414 return 0;
10415 }
10416
10417 /* mark_btf_func_reg_size() is used when the reg size is determined by
10418 * the BTF func_proto's return value size and argument.
10419 */
mark_btf_func_reg_size(struct bpf_verifier_env * env,u32 regno,size_t reg_size)10420 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
10421 size_t reg_size)
10422 {
10423 struct bpf_reg_state *reg = &cur_regs(env)[regno];
10424
10425 if (regno == BPF_REG_0) {
10426 /* Function return value */
10427 reg->live |= REG_LIVE_WRITTEN;
10428 reg->subreg_def = reg_size == sizeof(u64) ?
10429 DEF_NOT_SUBREG : env->insn_idx + 1;
10430 } else {
10431 /* Function argument */
10432 if (reg_size == sizeof(u64)) {
10433 mark_insn_zext(env, reg);
10434 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
10435 } else {
10436 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
10437 }
10438 }
10439 }
10440
is_kfunc_acquire(struct bpf_kfunc_call_arg_meta * meta)10441 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
10442 {
10443 return meta->kfunc_flags & KF_ACQUIRE;
10444 }
10445
is_kfunc_release(struct bpf_kfunc_call_arg_meta * meta)10446 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
10447 {
10448 return meta->kfunc_flags & KF_RELEASE;
10449 }
10450
is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta * meta)10451 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
10452 {
10453 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
10454 }
10455
is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta * meta)10456 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
10457 {
10458 return meta->kfunc_flags & KF_SLEEPABLE;
10459 }
10460
is_kfunc_destructive(struct bpf_kfunc_call_arg_meta * meta)10461 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
10462 {
10463 return meta->kfunc_flags & KF_DESTRUCTIVE;
10464 }
10465
is_kfunc_rcu(struct bpf_kfunc_call_arg_meta * meta)10466 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
10467 {
10468 return meta->kfunc_flags & KF_RCU;
10469 }
10470
__kfunc_param_match_suffix(const struct btf * btf,const struct btf_param * arg,const char * suffix)10471 static bool __kfunc_param_match_suffix(const struct btf *btf,
10472 const struct btf_param *arg,
10473 const char *suffix)
10474 {
10475 int suffix_len = strlen(suffix), len;
10476 const char *param_name;
10477
10478 /* In the future, this can be ported to use BTF tagging */
10479 param_name = btf_name_by_offset(btf, arg->name_off);
10480 if (str_is_empty(param_name))
10481 return false;
10482 len = strlen(param_name);
10483 if (len < suffix_len)
10484 return false;
10485 param_name += len - suffix_len;
10486 return !strncmp(param_name, suffix, suffix_len);
10487 }
10488
is_kfunc_arg_mem_size(const struct btf * btf,const struct btf_param * arg,const struct bpf_reg_state * reg)10489 static bool is_kfunc_arg_mem_size(const struct btf *btf,
10490 const struct btf_param *arg,
10491 const struct bpf_reg_state *reg)
10492 {
10493 const struct btf_type *t;
10494
10495 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10496 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10497 return false;
10498
10499 return __kfunc_param_match_suffix(btf, arg, "__sz");
10500 }
10501
is_kfunc_arg_const_mem_size(const struct btf * btf,const struct btf_param * arg,const struct bpf_reg_state * reg)10502 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
10503 const struct btf_param *arg,
10504 const struct bpf_reg_state *reg)
10505 {
10506 const struct btf_type *t;
10507
10508 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10509 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
10510 return false;
10511
10512 return __kfunc_param_match_suffix(btf, arg, "__szk");
10513 }
10514
is_kfunc_arg_optional(const struct btf * btf,const struct btf_param * arg)10515 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
10516 {
10517 return __kfunc_param_match_suffix(btf, arg, "__opt");
10518 }
10519
is_kfunc_arg_constant(const struct btf * btf,const struct btf_param * arg)10520 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
10521 {
10522 return __kfunc_param_match_suffix(btf, arg, "__k");
10523 }
10524
is_kfunc_arg_ignore(const struct btf * btf,const struct btf_param * arg)10525 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
10526 {
10527 return __kfunc_param_match_suffix(btf, arg, "__ign");
10528 }
10529
is_kfunc_arg_alloc_obj(const struct btf * btf,const struct btf_param * arg)10530 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
10531 {
10532 return __kfunc_param_match_suffix(btf, arg, "__alloc");
10533 }
10534
is_kfunc_arg_uninit(const struct btf * btf,const struct btf_param * arg)10535 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
10536 {
10537 return __kfunc_param_match_suffix(btf, arg, "__uninit");
10538 }
10539
is_kfunc_arg_refcounted_kptr(const struct btf * btf,const struct btf_param * arg)10540 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
10541 {
10542 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr");
10543 }
10544
is_kfunc_arg_scalar_with_name(const struct btf * btf,const struct btf_param * arg,const char * name)10545 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
10546 const struct btf_param *arg,
10547 const char *name)
10548 {
10549 int len, target_len = strlen(name);
10550 const char *param_name;
10551
10552 param_name = btf_name_by_offset(btf, arg->name_off);
10553 if (str_is_empty(param_name))
10554 return false;
10555 len = strlen(param_name);
10556 if (len != target_len)
10557 return false;
10558 if (strcmp(param_name, name))
10559 return false;
10560
10561 return true;
10562 }
10563
10564 enum {
10565 KF_ARG_DYNPTR_ID,
10566 KF_ARG_LIST_HEAD_ID,
10567 KF_ARG_LIST_NODE_ID,
10568 KF_ARG_RB_ROOT_ID,
10569 KF_ARG_RB_NODE_ID,
10570 };
10571
10572 BTF_ID_LIST(kf_arg_btf_ids)
BTF_ID(struct,bpf_dynptr_kern)10573 BTF_ID(struct, bpf_dynptr_kern)
10574 BTF_ID(struct, bpf_list_head)
10575 BTF_ID(struct, bpf_list_node)
10576 BTF_ID(struct, bpf_rb_root)
10577 BTF_ID(struct, bpf_rb_node)
10578
10579 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
10580 const struct btf_param *arg, int type)
10581 {
10582 const struct btf_type *t;
10583 u32 res_id;
10584
10585 t = btf_type_skip_modifiers(btf, arg->type, NULL);
10586 if (!t)
10587 return false;
10588 if (!btf_type_is_ptr(t))
10589 return false;
10590 t = btf_type_skip_modifiers(btf, t->type, &res_id);
10591 if (!t)
10592 return false;
10593 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
10594 }
10595
is_kfunc_arg_dynptr(const struct btf * btf,const struct btf_param * arg)10596 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
10597 {
10598 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
10599 }
10600
is_kfunc_arg_list_head(const struct btf * btf,const struct btf_param * arg)10601 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
10602 {
10603 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
10604 }
10605
is_kfunc_arg_list_node(const struct btf * btf,const struct btf_param * arg)10606 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
10607 {
10608 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
10609 }
10610
is_kfunc_arg_rbtree_root(const struct btf * btf,const struct btf_param * arg)10611 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
10612 {
10613 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
10614 }
10615
is_kfunc_arg_rbtree_node(const struct btf * btf,const struct btf_param * arg)10616 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
10617 {
10618 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
10619 }
10620
is_kfunc_arg_callback(struct bpf_verifier_env * env,const struct btf * btf,const struct btf_param * arg)10621 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
10622 const struct btf_param *arg)
10623 {
10624 const struct btf_type *t;
10625
10626 t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
10627 if (!t)
10628 return false;
10629
10630 return true;
10631 }
10632
10633 /* 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)10634 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
10635 const struct btf *btf,
10636 const struct btf_type *t, int rec)
10637 {
10638 const struct btf_type *member_type;
10639 const struct btf_member *member;
10640 u32 i;
10641
10642 if (!btf_type_is_struct(t))
10643 return false;
10644
10645 for_each_member(i, t, member) {
10646 const struct btf_array *array;
10647
10648 member_type = btf_type_skip_modifiers(btf, member->type, NULL);
10649 if (btf_type_is_struct(member_type)) {
10650 if (rec >= 3) {
10651 verbose(env, "max struct nesting depth exceeded\n");
10652 return false;
10653 }
10654 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
10655 return false;
10656 continue;
10657 }
10658 if (btf_type_is_array(member_type)) {
10659 array = btf_array(member_type);
10660 if (!array->nelems)
10661 return false;
10662 member_type = btf_type_skip_modifiers(btf, array->type, NULL);
10663 if (!btf_type_is_scalar(member_type))
10664 return false;
10665 continue;
10666 }
10667 if (!btf_type_is_scalar(member_type))
10668 return false;
10669 }
10670 return true;
10671 }
10672
10673 enum kfunc_ptr_arg_type {
10674 KF_ARG_PTR_TO_CTX,
10675 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
10676 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
10677 KF_ARG_PTR_TO_DYNPTR,
10678 KF_ARG_PTR_TO_ITER,
10679 KF_ARG_PTR_TO_LIST_HEAD,
10680 KF_ARG_PTR_TO_LIST_NODE,
10681 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
10682 KF_ARG_PTR_TO_MEM,
10683 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
10684 KF_ARG_PTR_TO_CALLBACK,
10685 KF_ARG_PTR_TO_RB_ROOT,
10686 KF_ARG_PTR_TO_RB_NODE,
10687 };
10688
10689 enum special_kfunc_type {
10690 KF_bpf_obj_new_impl,
10691 KF_bpf_obj_drop_impl,
10692 KF_bpf_refcount_acquire_impl,
10693 KF_bpf_list_push_front_impl,
10694 KF_bpf_list_push_back_impl,
10695 KF_bpf_list_pop_front,
10696 KF_bpf_list_pop_back,
10697 KF_bpf_cast_to_kern_ctx,
10698 KF_bpf_rdonly_cast,
10699 KF_bpf_rcu_read_lock,
10700 KF_bpf_rcu_read_unlock,
10701 KF_bpf_rbtree_remove,
10702 KF_bpf_rbtree_add_impl,
10703 KF_bpf_rbtree_first,
10704 KF_bpf_dynptr_from_skb,
10705 KF_bpf_dynptr_from_xdp,
10706 KF_bpf_dynptr_slice,
10707 KF_bpf_dynptr_slice_rdwr,
10708 KF_bpf_dynptr_clone,
10709 };
10710
10711 BTF_SET_START(special_kfunc_set)
BTF_ID(func,bpf_obj_new_impl)10712 BTF_ID(func, bpf_obj_new_impl)
10713 BTF_ID(func, bpf_obj_drop_impl)
10714 BTF_ID(func, bpf_refcount_acquire_impl)
10715 BTF_ID(func, bpf_list_push_front_impl)
10716 BTF_ID(func, bpf_list_push_back_impl)
10717 BTF_ID(func, bpf_list_pop_front)
10718 BTF_ID(func, bpf_list_pop_back)
10719 BTF_ID(func, bpf_cast_to_kern_ctx)
10720 BTF_ID(func, bpf_rdonly_cast)
10721 BTF_ID(func, bpf_rbtree_remove)
10722 BTF_ID(func, bpf_rbtree_add_impl)
10723 BTF_ID(func, bpf_rbtree_first)
10724 BTF_ID(func, bpf_dynptr_from_skb)
10725 BTF_ID(func, bpf_dynptr_from_xdp)
10726 BTF_ID(func, bpf_dynptr_slice)
10727 BTF_ID(func, bpf_dynptr_slice_rdwr)
10728 BTF_ID(func, bpf_dynptr_clone)
10729 BTF_SET_END(special_kfunc_set)
10730
10731 BTF_ID_LIST(special_kfunc_list)
10732 BTF_ID(func, bpf_obj_new_impl)
10733 BTF_ID(func, bpf_obj_drop_impl)
10734 BTF_ID(func, bpf_refcount_acquire_impl)
10735 BTF_ID(func, bpf_list_push_front_impl)
10736 BTF_ID(func, bpf_list_push_back_impl)
10737 BTF_ID(func, bpf_list_pop_front)
10738 BTF_ID(func, bpf_list_pop_back)
10739 BTF_ID(func, bpf_cast_to_kern_ctx)
10740 BTF_ID(func, bpf_rdonly_cast)
10741 BTF_ID(func, bpf_rcu_read_lock)
10742 BTF_ID(func, bpf_rcu_read_unlock)
10743 BTF_ID(func, bpf_rbtree_remove)
10744 BTF_ID(func, bpf_rbtree_add_impl)
10745 BTF_ID(func, bpf_rbtree_first)
10746 BTF_ID(func, bpf_dynptr_from_skb)
10747 BTF_ID(func, bpf_dynptr_from_xdp)
10748 BTF_ID(func, bpf_dynptr_slice)
10749 BTF_ID(func, bpf_dynptr_slice_rdwr)
10750 BTF_ID(func, bpf_dynptr_clone)
10751
10752 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
10753 {
10754 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
10755 meta->arg_owning_ref) {
10756 return false;
10757 }
10758
10759 return meta->kfunc_flags & KF_RET_NULL;
10760 }
10761
is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta * meta)10762 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
10763 {
10764 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
10765 }
10766
is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta * meta)10767 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
10768 {
10769 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
10770 }
10771
10772 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)10773 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
10774 struct bpf_kfunc_call_arg_meta *meta,
10775 const struct btf_type *t, const struct btf_type *ref_t,
10776 const char *ref_tname, const struct btf_param *args,
10777 int argno, int nargs)
10778 {
10779 u32 regno = argno + 1;
10780 struct bpf_reg_state *regs = cur_regs(env);
10781 struct bpf_reg_state *reg = ®s[regno];
10782 bool arg_mem_size = false;
10783
10784 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
10785 return KF_ARG_PTR_TO_CTX;
10786
10787 /* In this function, we verify the kfunc's BTF as per the argument type,
10788 * leaving the rest of the verification with respect to the register
10789 * type to our caller. When a set of conditions hold in the BTF type of
10790 * arguments, we resolve it to a known kfunc_ptr_arg_type.
10791 */
10792 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
10793 return KF_ARG_PTR_TO_CTX;
10794
10795 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
10796 return KF_ARG_PTR_TO_ALLOC_BTF_ID;
10797
10798 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
10799 return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
10800
10801 if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
10802 return KF_ARG_PTR_TO_DYNPTR;
10803
10804 if (is_kfunc_arg_iter(meta, argno))
10805 return KF_ARG_PTR_TO_ITER;
10806
10807 if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
10808 return KF_ARG_PTR_TO_LIST_HEAD;
10809
10810 if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
10811 return KF_ARG_PTR_TO_LIST_NODE;
10812
10813 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
10814 return KF_ARG_PTR_TO_RB_ROOT;
10815
10816 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
10817 return KF_ARG_PTR_TO_RB_NODE;
10818
10819 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
10820 if (!btf_type_is_struct(ref_t)) {
10821 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
10822 meta->func_name, argno, btf_type_str(ref_t), ref_tname);
10823 return -EINVAL;
10824 }
10825 return KF_ARG_PTR_TO_BTF_ID;
10826 }
10827
10828 if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
10829 return KF_ARG_PTR_TO_CALLBACK;
10830
10831
10832 if (argno + 1 < nargs &&
10833 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) ||
10834 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1])))
10835 arg_mem_size = true;
10836
10837 /* This is the catch all argument type of register types supported by
10838 * check_helper_mem_access. However, we only allow when argument type is
10839 * pointer to scalar, or struct composed (recursively) of scalars. When
10840 * arg_mem_size is true, the pointer can be void *.
10841 */
10842 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
10843 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
10844 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
10845 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
10846 return -EINVAL;
10847 }
10848 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
10849 }
10850
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)10851 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
10852 struct bpf_reg_state *reg,
10853 const struct btf_type *ref_t,
10854 const char *ref_tname, u32 ref_id,
10855 struct bpf_kfunc_call_arg_meta *meta,
10856 int argno)
10857 {
10858 const struct btf_type *reg_ref_t;
10859 bool strict_type_match = false;
10860 const struct btf *reg_btf;
10861 const char *reg_ref_tname;
10862 u32 reg_ref_id;
10863
10864 if (base_type(reg->type) == PTR_TO_BTF_ID) {
10865 reg_btf = reg->btf;
10866 reg_ref_id = reg->btf_id;
10867 } else {
10868 reg_btf = btf_vmlinux;
10869 reg_ref_id = *reg2btf_ids[base_type(reg->type)];
10870 }
10871
10872 /* Enforce strict type matching for calls to kfuncs that are acquiring
10873 * or releasing a reference, or are no-cast aliases. We do _not_
10874 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
10875 * as we want to enable BPF programs to pass types that are bitwise
10876 * equivalent without forcing them to explicitly cast with something
10877 * like bpf_cast_to_kern_ctx().
10878 *
10879 * For example, say we had a type like the following:
10880 *
10881 * struct bpf_cpumask {
10882 * cpumask_t cpumask;
10883 * refcount_t usage;
10884 * };
10885 *
10886 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
10887 * to a struct cpumask, so it would be safe to pass a struct
10888 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
10889 *
10890 * The philosophy here is similar to how we allow scalars of different
10891 * types to be passed to kfuncs as long as the size is the same. The
10892 * only difference here is that we're simply allowing
10893 * btf_struct_ids_match() to walk the struct at the 0th offset, and
10894 * resolve types.
10895 */
10896 if (is_kfunc_acquire(meta) ||
10897 (is_kfunc_release(meta) && reg->ref_obj_id) ||
10898 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
10899 strict_type_match = true;
10900
10901 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off);
10902
10903 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id);
10904 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
10905 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) {
10906 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
10907 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
10908 btf_type_str(reg_ref_t), reg_ref_tname);
10909 return -EINVAL;
10910 }
10911 return 0;
10912 }
10913
ref_set_non_owning(struct bpf_verifier_env * env,struct bpf_reg_state * reg)10914 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
10915 {
10916 struct bpf_verifier_state *state = env->cur_state;
10917 struct btf_record *rec = reg_btf_record(reg);
10918
10919 if (!state->active_lock.ptr) {
10920 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n");
10921 return -EFAULT;
10922 }
10923
10924 if (type_flag(reg->type) & NON_OWN_REF) {
10925 verbose(env, "verifier internal error: NON_OWN_REF already set\n");
10926 return -EFAULT;
10927 }
10928
10929 reg->type |= NON_OWN_REF;
10930 if (rec->refcount_off >= 0)
10931 reg->type |= MEM_RCU;
10932
10933 return 0;
10934 }
10935
ref_convert_owning_non_owning(struct bpf_verifier_env * env,u32 ref_obj_id)10936 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
10937 {
10938 struct bpf_func_state *state, *unused;
10939 struct bpf_reg_state *reg;
10940 int i;
10941
10942 state = cur_func(env);
10943
10944 if (!ref_obj_id) {
10945 verbose(env, "verifier internal error: ref_obj_id is zero for "
10946 "owning -> non-owning conversion\n");
10947 return -EFAULT;
10948 }
10949
10950 for (i = 0; i < state->acquired_refs; i++) {
10951 if (state->refs[i].id != ref_obj_id)
10952 continue;
10953
10954 /* Clear ref_obj_id here so release_reference doesn't clobber
10955 * the whole reg
10956 */
10957 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10958 if (reg->ref_obj_id == ref_obj_id) {
10959 reg->ref_obj_id = 0;
10960 ref_set_non_owning(env, reg);
10961 }
10962 }));
10963 return 0;
10964 }
10965
10966 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n");
10967 return -EFAULT;
10968 }
10969
10970 /* Implementation details:
10971 *
10972 * Each register points to some region of memory, which we define as an
10973 * allocation. Each allocation may embed a bpf_spin_lock which protects any
10974 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
10975 * allocation. The lock and the data it protects are colocated in the same
10976 * memory region.
10977 *
10978 * Hence, everytime a register holds a pointer value pointing to such
10979 * allocation, the verifier preserves a unique reg->id for it.
10980 *
10981 * The verifier remembers the lock 'ptr' and the lock 'id' whenever
10982 * bpf_spin_lock is called.
10983 *
10984 * To enable this, lock state in the verifier captures two values:
10985 * active_lock.ptr = Register's type specific pointer
10986 * active_lock.id = A unique ID for each register pointer value
10987 *
10988 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
10989 * supported register types.
10990 *
10991 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
10992 * allocated objects is the reg->btf pointer.
10993 *
10994 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
10995 * can establish the provenance of the map value statically for each distinct
10996 * lookup into such maps. They always contain a single map value hence unique
10997 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
10998 *
10999 * So, in case of global variables, they use array maps with max_entries = 1,
11000 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
11001 * into the same map value as max_entries is 1, as described above).
11002 *
11003 * In case of inner map lookups, the inner map pointer has same map_ptr as the
11004 * outer map pointer (in verifier context), but each lookup into an inner map
11005 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
11006 * maps from the same outer map share the same map_ptr as active_lock.ptr, they
11007 * will get different reg->id assigned to each lookup, hence different
11008 * active_lock.id.
11009 *
11010 * In case of allocated objects, active_lock.ptr is the reg->btf, and the
11011 * reg->id is a unique ID preserved after the NULL pointer check on the pointer
11012 * returned from bpf_obj_new. Each allocation receives a new reg->id.
11013 */
check_reg_allocation_locked(struct bpf_verifier_env * env,struct bpf_reg_state * reg)11014 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
11015 {
11016 void *ptr;
11017 u32 id;
11018
11019 switch ((int)reg->type) {
11020 case PTR_TO_MAP_VALUE:
11021 ptr = reg->map_ptr;
11022 break;
11023 case PTR_TO_BTF_ID | MEM_ALLOC:
11024 ptr = reg->btf;
11025 break;
11026 default:
11027 verbose(env, "verifier internal error: unknown reg type for lock check\n");
11028 return -EFAULT;
11029 }
11030 id = reg->id;
11031
11032 if (!env->cur_state->active_lock.ptr)
11033 return -EINVAL;
11034 if (env->cur_state->active_lock.ptr != ptr ||
11035 env->cur_state->active_lock.id != id) {
11036 verbose(env, "held lock and object are not in the same allocation\n");
11037 return -EINVAL;
11038 }
11039 return 0;
11040 }
11041
is_bpf_list_api_kfunc(u32 btf_id)11042 static bool is_bpf_list_api_kfunc(u32 btf_id)
11043 {
11044 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11045 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11046 btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11047 btf_id == special_kfunc_list[KF_bpf_list_pop_back];
11048 }
11049
is_bpf_rbtree_api_kfunc(u32 btf_id)11050 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
11051 {
11052 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
11053 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11054 btf_id == special_kfunc_list[KF_bpf_rbtree_first];
11055 }
11056
is_bpf_graph_api_kfunc(u32 btf_id)11057 static bool is_bpf_graph_api_kfunc(u32 btf_id)
11058 {
11059 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
11060 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
11061 }
11062
is_sync_callback_calling_kfunc(u32 btf_id)11063 static bool is_sync_callback_calling_kfunc(u32 btf_id)
11064 {
11065 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
11066 }
11067
is_rbtree_lock_required_kfunc(u32 btf_id)11068 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
11069 {
11070 return is_bpf_rbtree_api_kfunc(btf_id);
11071 }
11072
check_kfunc_is_graph_root_api(struct bpf_verifier_env * env,enum btf_field_type head_field_type,u32 kfunc_btf_id)11073 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
11074 enum btf_field_type head_field_type,
11075 u32 kfunc_btf_id)
11076 {
11077 bool ret;
11078
11079 switch (head_field_type) {
11080 case BPF_LIST_HEAD:
11081 ret = is_bpf_list_api_kfunc(kfunc_btf_id);
11082 break;
11083 case BPF_RB_ROOT:
11084 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
11085 break;
11086 default:
11087 verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
11088 btf_field_type_name(head_field_type));
11089 return false;
11090 }
11091
11092 if (!ret)
11093 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
11094 btf_field_type_name(head_field_type));
11095 return ret;
11096 }
11097
check_kfunc_is_graph_node_api(struct bpf_verifier_env * env,enum btf_field_type node_field_type,u32 kfunc_btf_id)11098 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
11099 enum btf_field_type node_field_type,
11100 u32 kfunc_btf_id)
11101 {
11102 bool ret;
11103
11104 switch (node_field_type) {
11105 case BPF_LIST_NODE:
11106 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11107 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
11108 break;
11109 case BPF_RB_NODE:
11110 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11111 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]);
11112 break;
11113 default:
11114 verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
11115 btf_field_type_name(node_field_type));
11116 return false;
11117 }
11118
11119 if (!ret)
11120 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
11121 btf_field_type_name(node_field_type));
11122 return ret;
11123 }
11124
11125 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)11126 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
11127 struct bpf_reg_state *reg, u32 regno,
11128 struct bpf_kfunc_call_arg_meta *meta,
11129 enum btf_field_type head_field_type,
11130 struct btf_field **head_field)
11131 {
11132 const char *head_type_name;
11133 struct btf_field *field;
11134 struct btf_record *rec;
11135 u32 head_off;
11136
11137 if (meta->btf != btf_vmlinux) {
11138 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
11139 return -EFAULT;
11140 }
11141
11142 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
11143 return -EFAULT;
11144
11145 head_type_name = btf_field_type_name(head_field_type);
11146 if (!tnum_is_const(reg->var_off)) {
11147 verbose(env,
11148 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
11149 regno, head_type_name);
11150 return -EINVAL;
11151 }
11152
11153 rec = reg_btf_record(reg);
11154 head_off = reg->off + reg->var_off.value;
11155 field = btf_record_find(rec, head_off, head_field_type);
11156 if (!field) {
11157 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
11158 return -EINVAL;
11159 }
11160
11161 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */
11162 if (check_reg_allocation_locked(env, reg)) {
11163 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
11164 rec->spin_lock_off, head_type_name);
11165 return -EINVAL;
11166 }
11167
11168 if (*head_field) {
11169 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name);
11170 return -EFAULT;
11171 }
11172 *head_field = field;
11173 return 0;
11174 }
11175
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)11176 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
11177 struct bpf_reg_state *reg, u32 regno,
11178 struct bpf_kfunc_call_arg_meta *meta)
11179 {
11180 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
11181 &meta->arg_list_head.field);
11182 }
11183
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)11184 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
11185 struct bpf_reg_state *reg, u32 regno,
11186 struct bpf_kfunc_call_arg_meta *meta)
11187 {
11188 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
11189 &meta->arg_rbtree_root.field);
11190 }
11191
11192 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)11193 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
11194 struct bpf_reg_state *reg, u32 regno,
11195 struct bpf_kfunc_call_arg_meta *meta,
11196 enum btf_field_type head_field_type,
11197 enum btf_field_type node_field_type,
11198 struct btf_field **node_field)
11199 {
11200 const char *node_type_name;
11201 const struct btf_type *et, *t;
11202 struct btf_field *field;
11203 u32 node_off;
11204
11205 if (meta->btf != btf_vmlinux) {
11206 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n");
11207 return -EFAULT;
11208 }
11209
11210 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
11211 return -EFAULT;
11212
11213 node_type_name = btf_field_type_name(node_field_type);
11214 if (!tnum_is_const(reg->var_off)) {
11215 verbose(env,
11216 "R%d doesn't have constant offset. %s has to be at the constant offset\n",
11217 regno, node_type_name);
11218 return -EINVAL;
11219 }
11220
11221 node_off = reg->off + reg->var_off.value;
11222 field = reg_find_field_offset(reg, node_off, node_field_type);
11223 if (!field || field->offset != node_off) {
11224 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
11225 return -EINVAL;
11226 }
11227
11228 field = *node_field;
11229
11230 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
11231 t = btf_type_by_id(reg->btf, reg->btf_id);
11232 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
11233 field->graph_root.value_btf_id, true)) {
11234 verbose(env, "operation on %s expects arg#1 %s at offset=%d "
11235 "in struct %s, but arg is at offset=%d in struct %s\n",
11236 btf_field_type_name(head_field_type),
11237 btf_field_type_name(node_field_type),
11238 field->graph_root.node_offset,
11239 btf_name_by_offset(field->graph_root.btf, et->name_off),
11240 node_off, btf_name_by_offset(reg->btf, t->name_off));
11241 return -EINVAL;
11242 }
11243 meta->arg_btf = reg->btf;
11244 meta->arg_btf_id = reg->btf_id;
11245
11246 if (node_off != field->graph_root.node_offset) {
11247 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
11248 node_off, btf_field_type_name(node_field_type),
11249 field->graph_root.node_offset,
11250 btf_name_by_offset(field->graph_root.btf, et->name_off));
11251 return -EINVAL;
11252 }
11253
11254 return 0;
11255 }
11256
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)11257 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
11258 struct bpf_reg_state *reg, u32 regno,
11259 struct bpf_kfunc_call_arg_meta *meta)
11260 {
11261 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
11262 BPF_LIST_HEAD, BPF_LIST_NODE,
11263 &meta->arg_list_head.field);
11264 }
11265
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)11266 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
11267 struct bpf_reg_state *reg, u32 regno,
11268 struct bpf_kfunc_call_arg_meta *meta)
11269 {
11270 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
11271 BPF_RB_ROOT, BPF_RB_NODE,
11272 &meta->arg_rbtree_root.field);
11273 }
11274
check_kfunc_args(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,int insn_idx)11275 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
11276 int insn_idx)
11277 {
11278 const char *func_name = meta->func_name, *ref_tname;
11279 const struct btf *btf = meta->btf;
11280 const struct btf_param *args;
11281 struct btf_record *rec;
11282 u32 i, nargs;
11283 int ret;
11284
11285 args = (const struct btf_param *)(meta->func_proto + 1);
11286 nargs = btf_type_vlen(meta->func_proto);
11287 if (nargs > MAX_BPF_FUNC_REG_ARGS) {
11288 verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
11289 MAX_BPF_FUNC_REG_ARGS);
11290 return -EINVAL;
11291 }
11292
11293 /* Check that BTF function arguments match actual types that the
11294 * verifier sees.
11295 */
11296 for (i = 0; i < nargs; i++) {
11297 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1];
11298 const struct btf_type *t, *ref_t, *resolve_ret;
11299 enum bpf_arg_type arg_type = ARG_DONTCARE;
11300 u32 regno = i + 1, ref_id, type_size;
11301 bool is_ret_buf_sz = false;
11302 int kf_arg_type;
11303
11304 t = btf_type_skip_modifiers(btf, args[i].type, NULL);
11305
11306 if (is_kfunc_arg_ignore(btf, &args[i]))
11307 continue;
11308
11309 if (btf_type_is_scalar(t)) {
11310 if (reg->type != SCALAR_VALUE) {
11311 verbose(env, "R%d is not a scalar\n", regno);
11312 return -EINVAL;
11313 }
11314
11315 if (is_kfunc_arg_constant(meta->btf, &args[i])) {
11316 if (meta->arg_constant.found) {
11317 verbose(env, "verifier internal error: only one constant argument permitted\n");
11318 return -EFAULT;
11319 }
11320 if (!tnum_is_const(reg->var_off)) {
11321 verbose(env, "R%d must be a known constant\n", regno);
11322 return -EINVAL;
11323 }
11324 ret = mark_chain_precision(env, regno);
11325 if (ret < 0)
11326 return ret;
11327 meta->arg_constant.found = true;
11328 meta->arg_constant.value = reg->var_off.value;
11329 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
11330 meta->r0_rdonly = true;
11331 is_ret_buf_sz = true;
11332 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
11333 is_ret_buf_sz = true;
11334 }
11335
11336 if (is_ret_buf_sz) {
11337 if (meta->r0_size) {
11338 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
11339 return -EINVAL;
11340 }
11341
11342 if (!tnum_is_const(reg->var_off)) {
11343 verbose(env, "R%d is not a const\n", regno);
11344 return -EINVAL;
11345 }
11346
11347 meta->r0_size = reg->var_off.value;
11348 ret = mark_chain_precision(env, regno);
11349 if (ret)
11350 return ret;
11351 }
11352 continue;
11353 }
11354
11355 if (!btf_type_is_ptr(t)) {
11356 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
11357 return -EINVAL;
11358 }
11359
11360 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
11361 (register_is_null(reg) || type_may_be_null(reg->type))) {
11362 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
11363 return -EACCES;
11364 }
11365
11366 if (reg->ref_obj_id) {
11367 if (is_kfunc_release(meta) && meta->ref_obj_id) {
11368 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
11369 regno, reg->ref_obj_id,
11370 meta->ref_obj_id);
11371 return -EFAULT;
11372 }
11373 meta->ref_obj_id = reg->ref_obj_id;
11374 if (is_kfunc_release(meta))
11375 meta->release_regno = regno;
11376 }
11377
11378 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
11379 ref_tname = btf_name_by_offset(btf, ref_t->name_off);
11380
11381 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
11382 if (kf_arg_type < 0)
11383 return kf_arg_type;
11384
11385 switch (kf_arg_type) {
11386 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11387 case KF_ARG_PTR_TO_BTF_ID:
11388 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
11389 break;
11390
11391 if (!is_trusted_reg(reg)) {
11392 if (!is_kfunc_rcu(meta)) {
11393 verbose(env, "R%d must be referenced or trusted\n", regno);
11394 return -EINVAL;
11395 }
11396 if (!is_rcu_reg(reg)) {
11397 verbose(env, "R%d must be a rcu pointer\n", regno);
11398 return -EINVAL;
11399 }
11400 }
11401
11402 fallthrough;
11403 case KF_ARG_PTR_TO_CTX:
11404 /* Trusted arguments have the same offset checks as release arguments */
11405 arg_type |= OBJ_RELEASE;
11406 break;
11407 case KF_ARG_PTR_TO_DYNPTR:
11408 case KF_ARG_PTR_TO_ITER:
11409 case KF_ARG_PTR_TO_LIST_HEAD:
11410 case KF_ARG_PTR_TO_LIST_NODE:
11411 case KF_ARG_PTR_TO_RB_ROOT:
11412 case KF_ARG_PTR_TO_RB_NODE:
11413 case KF_ARG_PTR_TO_MEM:
11414 case KF_ARG_PTR_TO_MEM_SIZE:
11415 case KF_ARG_PTR_TO_CALLBACK:
11416 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11417 /* Trusted by default */
11418 break;
11419 default:
11420 WARN_ON_ONCE(1);
11421 return -EFAULT;
11422 }
11423
11424 if (is_kfunc_release(meta) && reg->ref_obj_id)
11425 arg_type |= OBJ_RELEASE;
11426 ret = check_func_arg_reg_off(env, reg, regno, arg_type);
11427 if (ret < 0)
11428 return ret;
11429
11430 switch (kf_arg_type) {
11431 case KF_ARG_PTR_TO_CTX:
11432 if (reg->type != PTR_TO_CTX) {
11433 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t));
11434 return -EINVAL;
11435 }
11436
11437 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11438 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
11439 if (ret < 0)
11440 return -EINVAL;
11441 meta->ret_btf_id = ret;
11442 }
11443 break;
11444 case KF_ARG_PTR_TO_ALLOC_BTF_ID:
11445 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11446 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11447 return -EINVAL;
11448 }
11449 if (!reg->ref_obj_id) {
11450 verbose(env, "allocated object must be referenced\n");
11451 return -EINVAL;
11452 }
11453 if (meta->btf == btf_vmlinux &&
11454 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
11455 meta->arg_btf = reg->btf;
11456 meta->arg_btf_id = reg->btf_id;
11457 }
11458 break;
11459 case KF_ARG_PTR_TO_DYNPTR:
11460 {
11461 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
11462 int clone_ref_obj_id = 0;
11463
11464 if (reg->type != PTR_TO_STACK &&
11465 reg->type != CONST_PTR_TO_DYNPTR) {
11466 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i);
11467 return -EINVAL;
11468 }
11469
11470 if (reg->type == CONST_PTR_TO_DYNPTR)
11471 dynptr_arg_type |= MEM_RDONLY;
11472
11473 if (is_kfunc_arg_uninit(btf, &args[i]))
11474 dynptr_arg_type |= MEM_UNINIT;
11475
11476 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
11477 dynptr_arg_type |= DYNPTR_TYPE_SKB;
11478 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
11479 dynptr_arg_type |= DYNPTR_TYPE_XDP;
11480 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
11481 (dynptr_arg_type & MEM_UNINIT)) {
11482 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
11483
11484 if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
11485 verbose(env, "verifier internal error: no dynptr type for parent of clone\n");
11486 return -EFAULT;
11487 }
11488
11489 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
11490 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
11491 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
11492 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n");
11493 return -EFAULT;
11494 }
11495 }
11496
11497 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
11498 if (ret < 0)
11499 return ret;
11500
11501 if (!(dynptr_arg_type & MEM_UNINIT)) {
11502 int id = dynptr_id(env, reg);
11503
11504 if (id < 0) {
11505 verbose(env, "verifier internal error: failed to obtain dynptr id\n");
11506 return id;
11507 }
11508 meta->initialized_dynptr.id = id;
11509 meta->initialized_dynptr.type = dynptr_get_type(env, reg);
11510 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
11511 }
11512
11513 break;
11514 }
11515 case KF_ARG_PTR_TO_ITER:
11516 ret = process_iter_arg(env, regno, insn_idx, meta);
11517 if (ret < 0)
11518 return ret;
11519 break;
11520 case KF_ARG_PTR_TO_LIST_HEAD:
11521 if (reg->type != PTR_TO_MAP_VALUE &&
11522 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11523 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11524 return -EINVAL;
11525 }
11526 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11527 verbose(env, "allocated object must be referenced\n");
11528 return -EINVAL;
11529 }
11530 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
11531 if (ret < 0)
11532 return ret;
11533 break;
11534 case KF_ARG_PTR_TO_RB_ROOT:
11535 if (reg->type != PTR_TO_MAP_VALUE &&
11536 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11537 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
11538 return -EINVAL;
11539 }
11540 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
11541 verbose(env, "allocated object must be referenced\n");
11542 return -EINVAL;
11543 }
11544 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
11545 if (ret < 0)
11546 return ret;
11547 break;
11548 case KF_ARG_PTR_TO_LIST_NODE:
11549 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11550 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11551 return -EINVAL;
11552 }
11553 if (!reg->ref_obj_id) {
11554 verbose(env, "allocated object must be referenced\n");
11555 return -EINVAL;
11556 }
11557 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
11558 if (ret < 0)
11559 return ret;
11560 break;
11561 case KF_ARG_PTR_TO_RB_NODE:
11562 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) {
11563 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) {
11564 verbose(env, "rbtree_remove node input must be non-owning ref\n");
11565 return -EINVAL;
11566 }
11567 if (in_rbtree_lock_required_cb(env)) {
11568 verbose(env, "rbtree_remove not allowed in rbtree cb\n");
11569 return -EINVAL;
11570 }
11571 } else {
11572 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
11573 verbose(env, "arg#%d expected pointer to allocated object\n", i);
11574 return -EINVAL;
11575 }
11576 if (!reg->ref_obj_id) {
11577 verbose(env, "allocated object must be referenced\n");
11578 return -EINVAL;
11579 }
11580 }
11581
11582 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
11583 if (ret < 0)
11584 return ret;
11585 break;
11586 case KF_ARG_PTR_TO_BTF_ID:
11587 /* Only base_type is checked, further checks are done here */
11588 if ((base_type(reg->type) != PTR_TO_BTF_ID ||
11589 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
11590 !reg2btf_ids[base_type(reg->type)]) {
11591 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
11592 verbose(env, "expected %s or socket\n",
11593 reg_type_str(env, base_type(reg->type) |
11594 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
11595 return -EINVAL;
11596 }
11597 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
11598 if (ret < 0)
11599 return ret;
11600 break;
11601 case KF_ARG_PTR_TO_MEM:
11602 resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
11603 if (IS_ERR(resolve_ret)) {
11604 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
11605 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
11606 return -EINVAL;
11607 }
11608 ret = check_mem_reg(env, reg, regno, type_size);
11609 if (ret < 0)
11610 return ret;
11611 break;
11612 case KF_ARG_PTR_TO_MEM_SIZE:
11613 {
11614 struct bpf_reg_state *buff_reg = ®s[regno];
11615 const struct btf_param *buff_arg = &args[i];
11616 struct bpf_reg_state *size_reg = ®s[regno + 1];
11617 const struct btf_param *size_arg = &args[i + 1];
11618
11619 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
11620 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
11621 if (ret < 0) {
11622 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
11623 return ret;
11624 }
11625 }
11626
11627 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
11628 if (meta->arg_constant.found) {
11629 verbose(env, "verifier internal error: only one constant argument permitted\n");
11630 return -EFAULT;
11631 }
11632 if (!tnum_is_const(size_reg->var_off)) {
11633 verbose(env, "R%d must be a known constant\n", regno + 1);
11634 return -EINVAL;
11635 }
11636 meta->arg_constant.found = true;
11637 meta->arg_constant.value = size_reg->var_off.value;
11638 }
11639
11640 /* Skip next '__sz' or '__szk' argument */
11641 i++;
11642 break;
11643 }
11644 case KF_ARG_PTR_TO_CALLBACK:
11645 if (reg->type != PTR_TO_FUNC) {
11646 verbose(env, "arg%d expected pointer to func\n", i);
11647 return -EINVAL;
11648 }
11649 meta->subprogno = reg->subprogno;
11650 break;
11651 case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
11652 if (!type_is_ptr_alloc_obj(reg->type)) {
11653 verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
11654 return -EINVAL;
11655 }
11656 if (!type_is_non_owning_ref(reg->type))
11657 meta->arg_owning_ref = true;
11658
11659 rec = reg_btf_record(reg);
11660 if (!rec) {
11661 verbose(env, "verifier internal error: Couldn't find btf_record\n");
11662 return -EFAULT;
11663 }
11664
11665 if (rec->refcount_off < 0) {
11666 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
11667 return -EINVAL;
11668 }
11669
11670 meta->arg_btf = reg->btf;
11671 meta->arg_btf_id = reg->btf_id;
11672 break;
11673 }
11674 }
11675
11676 if (is_kfunc_release(meta) && !meta->release_regno) {
11677 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
11678 func_name);
11679 return -EINVAL;
11680 }
11681
11682 return 0;
11683 }
11684
fetch_kfunc_meta(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_kfunc_call_arg_meta * meta,const char ** kfunc_name)11685 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
11686 struct bpf_insn *insn,
11687 struct bpf_kfunc_call_arg_meta *meta,
11688 const char **kfunc_name)
11689 {
11690 const struct btf_type *func, *func_proto;
11691 u32 func_id, *kfunc_flags;
11692 const char *func_name;
11693 struct btf *desc_btf;
11694
11695 if (kfunc_name)
11696 *kfunc_name = NULL;
11697
11698 if (!insn->imm)
11699 return -EINVAL;
11700
11701 desc_btf = find_kfunc_desc_btf(env, insn->off);
11702 if (IS_ERR(desc_btf))
11703 return PTR_ERR(desc_btf);
11704
11705 func_id = insn->imm;
11706 func = btf_type_by_id(desc_btf, func_id);
11707 func_name = btf_name_by_offset(desc_btf, func->name_off);
11708 if (kfunc_name)
11709 *kfunc_name = func_name;
11710 func_proto = btf_type_by_id(desc_btf, func->type);
11711
11712 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
11713 if (!kfunc_flags) {
11714 return -EACCES;
11715 }
11716
11717 memset(meta, 0, sizeof(*meta));
11718 meta->btf = desc_btf;
11719 meta->func_id = func_id;
11720 meta->kfunc_flags = *kfunc_flags;
11721 meta->func_proto = func_proto;
11722 meta->func_name = func_name;
11723
11724 return 0;
11725 }
11726
check_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx_p)11727 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11728 int *insn_idx_p)
11729 {
11730 const struct btf_type *t, *ptr_type;
11731 u32 i, nargs, ptr_type_id, release_ref_obj_id;
11732 struct bpf_reg_state *regs = cur_regs(env);
11733 const char *func_name, *ptr_type_name;
11734 bool sleepable, rcu_lock, rcu_unlock;
11735 struct bpf_kfunc_call_arg_meta meta;
11736 struct bpf_insn_aux_data *insn_aux;
11737 int err, insn_idx = *insn_idx_p;
11738 const struct btf_param *args;
11739 const struct btf_type *ret_t;
11740 struct btf *desc_btf;
11741
11742 /* skip for now, but return error when we find this in fixup_kfunc_call */
11743 if (!insn->imm)
11744 return 0;
11745
11746 err = fetch_kfunc_meta(env, insn, &meta, &func_name);
11747 if (err == -EACCES && func_name)
11748 verbose(env, "calling kernel function %s is not allowed\n", func_name);
11749 if (err)
11750 return err;
11751 desc_btf = meta.btf;
11752 insn_aux = &env->insn_aux_data[insn_idx];
11753
11754 insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
11755
11756 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
11757 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
11758 return -EACCES;
11759 }
11760
11761 sleepable = is_kfunc_sleepable(&meta);
11762 if (sleepable && !env->prog->aux->sleepable) {
11763 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
11764 return -EACCES;
11765 }
11766
11767 /* Check the arguments */
11768 err = check_kfunc_args(env, &meta, insn_idx);
11769 if (err < 0)
11770 return err;
11771
11772 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11773 err = push_callback_call(env, insn, insn_idx, meta.subprogno,
11774 set_rbtree_add_callback_state);
11775 if (err) {
11776 verbose(env, "kfunc %s#%d failed callback verification\n",
11777 func_name, meta.func_id);
11778 return err;
11779 }
11780 }
11781
11782 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
11783 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
11784
11785 if (env->cur_state->active_rcu_lock) {
11786 struct bpf_func_state *state;
11787 struct bpf_reg_state *reg;
11788
11789 if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
11790 verbose(env, "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
11791 return -EACCES;
11792 }
11793
11794 if (rcu_lock) {
11795 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
11796 return -EINVAL;
11797 } else if (rcu_unlock) {
11798 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11799 if (reg->type & MEM_RCU) {
11800 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
11801 reg->type |= PTR_UNTRUSTED;
11802 }
11803 }));
11804 env->cur_state->active_rcu_lock = false;
11805 } else if (sleepable) {
11806 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
11807 return -EACCES;
11808 }
11809 } else if (rcu_lock) {
11810 env->cur_state->active_rcu_lock = true;
11811 } else if (rcu_unlock) {
11812 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
11813 return -EINVAL;
11814 }
11815
11816 /* In case of release function, we get register number of refcounted
11817 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
11818 */
11819 if (meta.release_regno) {
11820 err = release_reference(env, regs[meta.release_regno].ref_obj_id);
11821 if (err) {
11822 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11823 func_name, meta.func_id);
11824 return err;
11825 }
11826 }
11827
11828 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
11829 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
11830 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
11831 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
11832 insn_aux->insert_off = regs[BPF_REG_2].off;
11833 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
11834 err = ref_convert_owning_non_owning(env, release_ref_obj_id);
11835 if (err) {
11836 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
11837 func_name, meta.func_id);
11838 return err;
11839 }
11840
11841 err = release_reference(env, release_ref_obj_id);
11842 if (err) {
11843 verbose(env, "kfunc %s#%d reference has not been acquired before\n",
11844 func_name, meta.func_id);
11845 return err;
11846 }
11847 }
11848
11849 for (i = 0; i < CALLER_SAVED_REGS; i++)
11850 mark_reg_not_init(env, regs, caller_saved[i]);
11851
11852 /* Check return type */
11853 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
11854
11855 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
11856 /* Only exception is bpf_obj_new_impl */
11857 if (meta.btf != btf_vmlinux ||
11858 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
11859 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
11860 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
11861 return -EINVAL;
11862 }
11863 }
11864
11865 if (btf_type_is_scalar(t)) {
11866 mark_reg_unknown(env, regs, BPF_REG_0);
11867 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
11868 } else if (btf_type_is_ptr(t)) {
11869 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
11870
11871 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
11872 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
11873 struct btf *ret_btf;
11874 u32 ret_btf_id;
11875
11876 if (unlikely(!bpf_global_ma_set))
11877 return -ENOMEM;
11878
11879 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) {
11880 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
11881 return -EINVAL;
11882 }
11883
11884 ret_btf = env->prog->aux->btf;
11885 ret_btf_id = meta.arg_constant.value;
11886
11887 /* This may be NULL due to user not supplying a BTF */
11888 if (!ret_btf) {
11889 verbose(env, "bpf_obj_new requires prog BTF\n");
11890 return -EINVAL;
11891 }
11892
11893 ret_t = btf_type_by_id(ret_btf, ret_btf_id);
11894 if (!ret_t || !__btf_type_is_struct(ret_t)) {
11895 verbose(env, "bpf_obj_new type ID argument must be of a struct\n");
11896 return -EINVAL;
11897 }
11898
11899 mark_reg_known_zero(env, regs, BPF_REG_0);
11900 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11901 regs[BPF_REG_0].btf = ret_btf;
11902 regs[BPF_REG_0].btf_id = ret_btf_id;
11903
11904 insn_aux->obj_new_size = ret_t->size;
11905 insn_aux->kptr_struct_meta =
11906 btf_find_struct_meta(ret_btf, ret_btf_id);
11907 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
11908 mark_reg_known_zero(env, regs, BPF_REG_0);
11909 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
11910 regs[BPF_REG_0].btf = meta.arg_btf;
11911 regs[BPF_REG_0].btf_id = meta.arg_btf_id;
11912
11913 insn_aux->kptr_struct_meta =
11914 btf_find_struct_meta(meta.arg_btf,
11915 meta.arg_btf_id);
11916 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] ||
11917 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) {
11918 struct btf_field *field = meta.arg_list_head.field;
11919
11920 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11921 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
11922 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
11923 struct btf_field *field = meta.arg_rbtree_root.field;
11924
11925 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
11926 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
11927 mark_reg_known_zero(env, regs, BPF_REG_0);
11928 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
11929 regs[BPF_REG_0].btf = desc_btf;
11930 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11931 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
11932 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value);
11933 if (!ret_t || !btf_type_is_struct(ret_t)) {
11934 verbose(env,
11935 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n");
11936 return -EINVAL;
11937 }
11938
11939 mark_reg_known_zero(env, regs, BPF_REG_0);
11940 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
11941 regs[BPF_REG_0].btf = desc_btf;
11942 regs[BPF_REG_0].btf_id = meta.arg_constant.value;
11943 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
11944 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
11945 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type);
11946
11947 mark_reg_known_zero(env, regs, BPF_REG_0);
11948
11949 if (!meta.arg_constant.found) {
11950 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n");
11951 return -EFAULT;
11952 }
11953
11954 regs[BPF_REG_0].mem_size = meta.arg_constant.value;
11955
11956 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
11957 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
11958
11959 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
11960 regs[BPF_REG_0].type |= MEM_RDONLY;
11961 } else {
11962 /* this will set env->seen_direct_write to true */
11963 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
11964 verbose(env, "the prog does not allow writes to packet data\n");
11965 return -EINVAL;
11966 }
11967 }
11968
11969 if (!meta.initialized_dynptr.id) {
11970 verbose(env, "verifier internal error: no dynptr id\n");
11971 return -EFAULT;
11972 }
11973 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id;
11974
11975 /* we don't need to set BPF_REG_0's ref obj id
11976 * because packet slices are not refcounted (see
11977 * dynptr_type_refcounted)
11978 */
11979 } else {
11980 verbose(env, "kernel function %s unhandled dynamic return type\n",
11981 meta.func_name);
11982 return -EFAULT;
11983 }
11984 } else if (!__btf_type_is_struct(ptr_type)) {
11985 if (!meta.r0_size) {
11986 __u32 sz;
11987
11988 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
11989 meta.r0_size = sz;
11990 meta.r0_rdonly = true;
11991 }
11992 }
11993 if (!meta.r0_size) {
11994 ptr_type_name = btf_name_by_offset(desc_btf,
11995 ptr_type->name_off);
11996 verbose(env,
11997 "kernel function %s returns pointer type %s %s is not supported\n",
11998 func_name,
11999 btf_type_str(ptr_type),
12000 ptr_type_name);
12001 return -EINVAL;
12002 }
12003
12004 mark_reg_known_zero(env, regs, BPF_REG_0);
12005 regs[BPF_REG_0].type = PTR_TO_MEM;
12006 regs[BPF_REG_0].mem_size = meta.r0_size;
12007
12008 if (meta.r0_rdonly)
12009 regs[BPF_REG_0].type |= MEM_RDONLY;
12010
12011 /* Ensures we don't access the memory after a release_reference() */
12012 if (meta.ref_obj_id)
12013 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
12014 } else {
12015 mark_reg_known_zero(env, regs, BPF_REG_0);
12016 regs[BPF_REG_0].btf = desc_btf;
12017 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
12018 regs[BPF_REG_0].btf_id = ptr_type_id;
12019 }
12020
12021 if (is_kfunc_ret_null(&meta)) {
12022 regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
12023 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
12024 regs[BPF_REG_0].id = ++env->id_gen;
12025 }
12026 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
12027 if (is_kfunc_acquire(&meta)) {
12028 int id = acquire_reference_state(env, insn_idx);
12029
12030 if (id < 0)
12031 return id;
12032 if (is_kfunc_ret_null(&meta))
12033 regs[BPF_REG_0].id = id;
12034 regs[BPF_REG_0].ref_obj_id = id;
12035 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) {
12036 ref_set_non_owning(env, ®s[BPF_REG_0]);
12037 }
12038
12039 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id)
12040 regs[BPF_REG_0].id = ++env->id_gen;
12041 } else if (btf_type_is_void(t)) {
12042 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) {
12043 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) {
12044 insn_aux->kptr_struct_meta =
12045 btf_find_struct_meta(meta.arg_btf,
12046 meta.arg_btf_id);
12047 }
12048 }
12049 }
12050
12051 nargs = btf_type_vlen(meta.func_proto);
12052 args = (const struct btf_param *)(meta.func_proto + 1);
12053 for (i = 0; i < nargs; i++) {
12054 u32 regno = i + 1;
12055
12056 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
12057 if (btf_type_is_ptr(t))
12058 mark_btf_func_reg_size(env, regno, sizeof(void *));
12059 else
12060 /* scalar. ensured by btf_check_kfunc_arg_match() */
12061 mark_btf_func_reg_size(env, regno, t->size);
12062 }
12063
12064 if (is_iter_next_kfunc(&meta)) {
12065 err = process_iter_next_call(env, insn_idx, &meta);
12066 if (err)
12067 return err;
12068 }
12069
12070 return 0;
12071 }
12072
signed_add_overflows(s64 a,s64 b)12073 static bool signed_add_overflows(s64 a, s64 b)
12074 {
12075 /* Do the add in u64, where overflow is well-defined */
12076 s64 res = (s64)((u64)a + (u64)b);
12077
12078 if (b < 0)
12079 return res > a;
12080 return res < a;
12081 }
12082
signed_add32_overflows(s32 a,s32 b)12083 static bool signed_add32_overflows(s32 a, s32 b)
12084 {
12085 /* Do the add in u32, where overflow is well-defined */
12086 s32 res = (s32)((u32)a + (u32)b);
12087
12088 if (b < 0)
12089 return res > a;
12090 return res < a;
12091 }
12092
signed_sub_overflows(s64 a,s64 b)12093 static bool signed_sub_overflows(s64 a, s64 b)
12094 {
12095 /* Do the sub in u64, where overflow is well-defined */
12096 s64 res = (s64)((u64)a - (u64)b);
12097
12098 if (b < 0)
12099 return res < a;
12100 return res > a;
12101 }
12102
signed_sub32_overflows(s32 a,s32 b)12103 static bool signed_sub32_overflows(s32 a, s32 b)
12104 {
12105 /* Do the sub in u32, where overflow is well-defined */
12106 s32 res = (s32)((u32)a - (u32)b);
12107
12108 if (b < 0)
12109 return res < a;
12110 return res > a;
12111 }
12112
check_reg_sane_offset(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,enum bpf_reg_type type)12113 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
12114 const struct bpf_reg_state *reg,
12115 enum bpf_reg_type type)
12116 {
12117 bool known = tnum_is_const(reg->var_off);
12118 s64 val = reg->var_off.value;
12119 s64 smin = reg->smin_value;
12120
12121 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
12122 verbose(env, "math between %s pointer and %lld is not allowed\n",
12123 reg_type_str(env, type), val);
12124 return false;
12125 }
12126
12127 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
12128 verbose(env, "%s pointer offset %d is not allowed\n",
12129 reg_type_str(env, type), reg->off);
12130 return false;
12131 }
12132
12133 if (smin == S64_MIN) {
12134 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
12135 reg_type_str(env, type));
12136 return false;
12137 }
12138
12139 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
12140 verbose(env, "value %lld makes %s pointer be out of bounds\n",
12141 smin, reg_type_str(env, type));
12142 return false;
12143 }
12144
12145 return true;
12146 }
12147
12148 enum {
12149 REASON_BOUNDS = -1,
12150 REASON_TYPE = -2,
12151 REASON_PATHS = -3,
12152 REASON_LIMIT = -4,
12153 REASON_STACK = -5,
12154 };
12155
retrieve_ptr_limit(const struct bpf_reg_state * ptr_reg,u32 * alu_limit,bool mask_to_left)12156 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
12157 u32 *alu_limit, bool mask_to_left)
12158 {
12159 u32 max = 0, ptr_limit = 0;
12160
12161 switch (ptr_reg->type) {
12162 case PTR_TO_STACK:
12163 /* Offset 0 is out-of-bounds, but acceptable start for the
12164 * left direction, see BPF_REG_FP. Also, unknown scalar
12165 * offset where we would need to deal with min/max bounds is
12166 * currently prohibited for unprivileged.
12167 */
12168 max = MAX_BPF_STACK + mask_to_left;
12169 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
12170 break;
12171 case PTR_TO_MAP_VALUE:
12172 max = ptr_reg->map_ptr->value_size;
12173 ptr_limit = (mask_to_left ?
12174 ptr_reg->smin_value :
12175 ptr_reg->umax_value) + ptr_reg->off;
12176 break;
12177 default:
12178 return REASON_TYPE;
12179 }
12180
12181 if (ptr_limit >= max)
12182 return REASON_LIMIT;
12183 *alu_limit = ptr_limit;
12184 return 0;
12185 }
12186
can_skip_alu_sanitation(const struct bpf_verifier_env * env,const struct bpf_insn * insn)12187 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
12188 const struct bpf_insn *insn)
12189 {
12190 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
12191 }
12192
update_alu_sanitation_state(struct bpf_insn_aux_data * aux,u32 alu_state,u32 alu_limit)12193 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
12194 u32 alu_state, u32 alu_limit)
12195 {
12196 /* If we arrived here from different branches with different
12197 * state or limits to sanitize, then this won't work.
12198 */
12199 if (aux->alu_state &&
12200 (aux->alu_state != alu_state ||
12201 aux->alu_limit != alu_limit))
12202 return REASON_PATHS;
12203
12204 /* Corresponding fixup done in do_misc_fixups(). */
12205 aux->alu_state = alu_state;
12206 aux->alu_limit = alu_limit;
12207 return 0;
12208 }
12209
sanitize_val_alu(struct bpf_verifier_env * env,struct bpf_insn * insn)12210 static int sanitize_val_alu(struct bpf_verifier_env *env,
12211 struct bpf_insn *insn)
12212 {
12213 struct bpf_insn_aux_data *aux = cur_aux(env);
12214
12215 if (can_skip_alu_sanitation(env, insn))
12216 return 0;
12217
12218 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
12219 }
12220
sanitize_needed(u8 opcode)12221 static bool sanitize_needed(u8 opcode)
12222 {
12223 return opcode == BPF_ADD || opcode == BPF_SUB;
12224 }
12225
12226 struct bpf_sanitize_info {
12227 struct bpf_insn_aux_data aux;
12228 bool mask_to_left;
12229 };
12230
12231 static struct bpf_verifier_state *
sanitize_speculative_path(struct bpf_verifier_env * env,const struct bpf_insn * insn,u32 next_idx,u32 curr_idx)12232 sanitize_speculative_path(struct bpf_verifier_env *env,
12233 const struct bpf_insn *insn,
12234 u32 next_idx, u32 curr_idx)
12235 {
12236 struct bpf_verifier_state *branch;
12237 struct bpf_reg_state *regs;
12238
12239 branch = push_stack(env, next_idx, curr_idx, true);
12240 if (branch && insn) {
12241 regs = branch->frame[branch->curframe]->regs;
12242 if (BPF_SRC(insn->code) == BPF_K) {
12243 mark_reg_unknown(env, regs, insn->dst_reg);
12244 } else if (BPF_SRC(insn->code) == BPF_X) {
12245 mark_reg_unknown(env, regs, insn->dst_reg);
12246 mark_reg_unknown(env, regs, insn->src_reg);
12247 }
12248 }
12249 return branch;
12250 }
12251
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)12252 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
12253 struct bpf_insn *insn,
12254 const struct bpf_reg_state *ptr_reg,
12255 const struct bpf_reg_state *off_reg,
12256 struct bpf_reg_state *dst_reg,
12257 struct bpf_sanitize_info *info,
12258 const bool commit_window)
12259 {
12260 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
12261 struct bpf_verifier_state *vstate = env->cur_state;
12262 bool off_is_imm = tnum_is_const(off_reg->var_off);
12263 bool off_is_neg = off_reg->smin_value < 0;
12264 bool ptr_is_dst_reg = ptr_reg == dst_reg;
12265 u8 opcode = BPF_OP(insn->code);
12266 u32 alu_state, alu_limit;
12267 struct bpf_reg_state tmp;
12268 bool ret;
12269 int err;
12270
12271 if (can_skip_alu_sanitation(env, insn))
12272 return 0;
12273
12274 /* We already marked aux for masking from non-speculative
12275 * paths, thus we got here in the first place. We only care
12276 * to explore bad access from here.
12277 */
12278 if (vstate->speculative)
12279 goto do_sim;
12280
12281 if (!commit_window) {
12282 if (!tnum_is_const(off_reg->var_off) &&
12283 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
12284 return REASON_BOUNDS;
12285
12286 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
12287 (opcode == BPF_SUB && !off_is_neg);
12288 }
12289
12290 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
12291 if (err < 0)
12292 return err;
12293
12294 if (commit_window) {
12295 /* In commit phase we narrow the masking window based on
12296 * the observed pointer move after the simulated operation.
12297 */
12298 alu_state = info->aux.alu_state;
12299 alu_limit = abs(info->aux.alu_limit - alu_limit);
12300 } else {
12301 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
12302 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
12303 alu_state |= ptr_is_dst_reg ?
12304 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
12305
12306 /* Limit pruning on unknown scalars to enable deep search for
12307 * potential masking differences from other program paths.
12308 */
12309 if (!off_is_imm)
12310 env->explore_alu_limits = true;
12311 }
12312
12313 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
12314 if (err < 0)
12315 return err;
12316 do_sim:
12317 /* If we're in commit phase, we're done here given we already
12318 * pushed the truncated dst_reg into the speculative verification
12319 * stack.
12320 *
12321 * Also, when register is a known constant, we rewrite register-based
12322 * operation to immediate-based, and thus do not need masking (and as
12323 * a consequence, do not need to simulate the zero-truncation either).
12324 */
12325 if (commit_window || off_is_imm)
12326 return 0;
12327
12328 /* Simulate and find potential out-of-bounds access under
12329 * speculative execution from truncation as a result of
12330 * masking when off was not within expected range. If off
12331 * sits in dst, then we temporarily need to move ptr there
12332 * to simulate dst (== 0) +/-= ptr. Needed, for example,
12333 * for cases where we use K-based arithmetic in one direction
12334 * and truncated reg-based in the other in order to explore
12335 * bad access.
12336 */
12337 if (!ptr_is_dst_reg) {
12338 tmp = *dst_reg;
12339 copy_register_state(dst_reg, ptr_reg);
12340 }
12341 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
12342 env->insn_idx);
12343 if (!ptr_is_dst_reg && ret)
12344 *dst_reg = tmp;
12345 return !ret ? REASON_STACK : 0;
12346 }
12347
sanitize_mark_insn_seen(struct bpf_verifier_env * env)12348 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
12349 {
12350 struct bpf_verifier_state *vstate = env->cur_state;
12351
12352 /* If we simulate paths under speculation, we don't update the
12353 * insn as 'seen' such that when we verify unreachable paths in
12354 * the non-speculative domain, sanitize_dead_code() can still
12355 * rewrite/sanitize them.
12356 */
12357 if (!vstate->speculative)
12358 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
12359 }
12360
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)12361 static int sanitize_err(struct bpf_verifier_env *env,
12362 const struct bpf_insn *insn, int reason,
12363 const struct bpf_reg_state *off_reg,
12364 const struct bpf_reg_state *dst_reg)
12365 {
12366 static const char *err = "pointer arithmetic with it prohibited for !root";
12367 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
12368 u32 dst = insn->dst_reg, src = insn->src_reg;
12369
12370 switch (reason) {
12371 case REASON_BOUNDS:
12372 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
12373 off_reg == dst_reg ? dst : src, err);
12374 break;
12375 case REASON_TYPE:
12376 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
12377 off_reg == dst_reg ? src : dst, err);
12378 break;
12379 case REASON_PATHS:
12380 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
12381 dst, op, err);
12382 break;
12383 case REASON_LIMIT:
12384 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
12385 dst, op, err);
12386 break;
12387 case REASON_STACK:
12388 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
12389 dst, err);
12390 break;
12391 default:
12392 verbose(env, "verifier internal error: unknown reason (%d)\n",
12393 reason);
12394 break;
12395 }
12396
12397 return -EACCES;
12398 }
12399
12400 /* check that stack access falls within stack limits and that 'reg' doesn't
12401 * have a variable offset.
12402 *
12403 * Variable offset is prohibited for unprivileged mode for simplicity since it
12404 * requires corresponding support in Spectre masking for stack ALU. See also
12405 * retrieve_ptr_limit().
12406 *
12407 *
12408 * 'off' includes 'reg->off'.
12409 */
check_stack_access_for_ptr_arithmetic(struct bpf_verifier_env * env,int regno,const struct bpf_reg_state * reg,int off)12410 static int check_stack_access_for_ptr_arithmetic(
12411 struct bpf_verifier_env *env,
12412 int regno,
12413 const struct bpf_reg_state *reg,
12414 int off)
12415 {
12416 if (!tnum_is_const(reg->var_off)) {
12417 char tn_buf[48];
12418
12419 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
12420 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
12421 regno, tn_buf, off);
12422 return -EACCES;
12423 }
12424
12425 if (off >= 0 || off < -MAX_BPF_STACK) {
12426 verbose(env, "R%d stack pointer arithmetic goes out of range, "
12427 "prohibited for !root; off=%d\n", regno, off);
12428 return -EACCES;
12429 }
12430
12431 return 0;
12432 }
12433
sanitize_check_bounds(struct bpf_verifier_env * env,const struct bpf_insn * insn,const struct bpf_reg_state * dst_reg)12434 static int sanitize_check_bounds(struct bpf_verifier_env *env,
12435 const struct bpf_insn *insn,
12436 const struct bpf_reg_state *dst_reg)
12437 {
12438 u32 dst = insn->dst_reg;
12439
12440 /* For unprivileged we require that resulting offset must be in bounds
12441 * in order to be able to sanitize access later on.
12442 */
12443 if (env->bypass_spec_v1)
12444 return 0;
12445
12446 switch (dst_reg->type) {
12447 case PTR_TO_STACK:
12448 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
12449 dst_reg->off + dst_reg->var_off.value))
12450 return -EACCES;
12451 break;
12452 case PTR_TO_MAP_VALUE:
12453 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
12454 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
12455 "prohibited for !root\n", dst);
12456 return -EACCES;
12457 }
12458 break;
12459 default:
12460 break;
12461 }
12462
12463 return 0;
12464 }
12465
12466 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
12467 * Caller should also handle BPF_MOV case separately.
12468 * If we return -EACCES, caller may want to try again treating pointer as a
12469 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
12470 */
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)12471 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
12472 struct bpf_insn *insn,
12473 const struct bpf_reg_state *ptr_reg,
12474 const struct bpf_reg_state *off_reg)
12475 {
12476 struct bpf_verifier_state *vstate = env->cur_state;
12477 struct bpf_func_state *state = vstate->frame[vstate->curframe];
12478 struct bpf_reg_state *regs = state->regs, *dst_reg;
12479 bool known = tnum_is_const(off_reg->var_off);
12480 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
12481 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
12482 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
12483 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
12484 struct bpf_sanitize_info info = {};
12485 u8 opcode = BPF_OP(insn->code);
12486 u32 dst = insn->dst_reg;
12487 int ret;
12488
12489 dst_reg = ®s[dst];
12490
12491 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
12492 smin_val > smax_val || umin_val > umax_val) {
12493 /* Taint dst register if offset had invalid bounds derived from
12494 * e.g. dead branches.
12495 */
12496 __mark_reg_unknown(env, dst_reg);
12497 return 0;
12498 }
12499
12500 if (BPF_CLASS(insn->code) != BPF_ALU64) {
12501 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
12502 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
12503 __mark_reg_unknown(env, dst_reg);
12504 return 0;
12505 }
12506
12507 verbose(env,
12508 "R%d 32-bit pointer arithmetic prohibited\n",
12509 dst);
12510 return -EACCES;
12511 }
12512
12513 if (ptr_reg->type & PTR_MAYBE_NULL) {
12514 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
12515 dst, reg_type_str(env, ptr_reg->type));
12516 return -EACCES;
12517 }
12518
12519 switch (base_type(ptr_reg->type)) {
12520 case PTR_TO_FLOW_KEYS:
12521 if (known)
12522 break;
12523 fallthrough;
12524 case CONST_PTR_TO_MAP:
12525 /* smin_val represents the known value */
12526 if (known && smin_val == 0 && opcode == BPF_ADD)
12527 break;
12528 fallthrough;
12529 case PTR_TO_PACKET_END:
12530 case PTR_TO_SOCKET:
12531 case PTR_TO_SOCK_COMMON:
12532 case PTR_TO_TCP_SOCK:
12533 case PTR_TO_XDP_SOCK:
12534 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
12535 dst, reg_type_str(env, ptr_reg->type));
12536 return -EACCES;
12537 default:
12538 break;
12539 }
12540
12541 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
12542 * The id may be overwritten later if we create a new variable offset.
12543 */
12544 dst_reg->type = ptr_reg->type;
12545 dst_reg->id = ptr_reg->id;
12546
12547 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
12548 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
12549 return -EINVAL;
12550
12551 /* pointer types do not carry 32-bit bounds at the moment. */
12552 __mark_reg32_unbounded(dst_reg);
12553
12554 if (sanitize_needed(opcode)) {
12555 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
12556 &info, false);
12557 if (ret < 0)
12558 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12559 }
12560
12561 switch (opcode) {
12562 case BPF_ADD:
12563 /* We can take a fixed offset as long as it doesn't overflow
12564 * the s32 'off' field
12565 */
12566 if (known && (ptr_reg->off + smin_val ==
12567 (s64)(s32)(ptr_reg->off + smin_val))) {
12568 /* pointer += K. Accumulate it into fixed offset */
12569 dst_reg->smin_value = smin_ptr;
12570 dst_reg->smax_value = smax_ptr;
12571 dst_reg->umin_value = umin_ptr;
12572 dst_reg->umax_value = umax_ptr;
12573 dst_reg->var_off = ptr_reg->var_off;
12574 dst_reg->off = ptr_reg->off + smin_val;
12575 dst_reg->raw = ptr_reg->raw;
12576 break;
12577 }
12578 /* A new variable offset is created. Note that off_reg->off
12579 * == 0, since it's a scalar.
12580 * dst_reg gets the pointer type and since some positive
12581 * integer value was added to the pointer, give it a new 'id'
12582 * if it's a PTR_TO_PACKET.
12583 * this creates a new 'base' pointer, off_reg (variable) gets
12584 * added into the variable offset, and we copy the fixed offset
12585 * from ptr_reg.
12586 */
12587 if (signed_add_overflows(smin_ptr, smin_val) ||
12588 signed_add_overflows(smax_ptr, smax_val)) {
12589 dst_reg->smin_value = S64_MIN;
12590 dst_reg->smax_value = S64_MAX;
12591 } else {
12592 dst_reg->smin_value = smin_ptr + smin_val;
12593 dst_reg->smax_value = smax_ptr + smax_val;
12594 }
12595 if (umin_ptr + umin_val < umin_ptr ||
12596 umax_ptr + umax_val < umax_ptr) {
12597 dst_reg->umin_value = 0;
12598 dst_reg->umax_value = U64_MAX;
12599 } else {
12600 dst_reg->umin_value = umin_ptr + umin_val;
12601 dst_reg->umax_value = umax_ptr + umax_val;
12602 }
12603 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
12604 dst_reg->off = ptr_reg->off;
12605 dst_reg->raw = ptr_reg->raw;
12606 if (reg_is_pkt_pointer(ptr_reg)) {
12607 dst_reg->id = ++env->id_gen;
12608 /* something was added to pkt_ptr, set range to zero */
12609 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12610 }
12611 break;
12612 case BPF_SUB:
12613 if (dst_reg == off_reg) {
12614 /* scalar -= pointer. Creates an unknown scalar */
12615 verbose(env, "R%d tried to subtract pointer from scalar\n",
12616 dst);
12617 return -EACCES;
12618 }
12619 /* We don't allow subtraction from FP, because (according to
12620 * test_verifier.c test "invalid fp arithmetic", JITs might not
12621 * be able to deal with it.
12622 */
12623 if (ptr_reg->type == PTR_TO_STACK) {
12624 verbose(env, "R%d subtraction from stack pointer prohibited\n",
12625 dst);
12626 return -EACCES;
12627 }
12628 if (known && (ptr_reg->off - smin_val ==
12629 (s64)(s32)(ptr_reg->off - smin_val))) {
12630 /* pointer -= K. Subtract it from fixed offset */
12631 dst_reg->smin_value = smin_ptr;
12632 dst_reg->smax_value = smax_ptr;
12633 dst_reg->umin_value = umin_ptr;
12634 dst_reg->umax_value = umax_ptr;
12635 dst_reg->var_off = ptr_reg->var_off;
12636 dst_reg->id = ptr_reg->id;
12637 dst_reg->off = ptr_reg->off - smin_val;
12638 dst_reg->raw = ptr_reg->raw;
12639 break;
12640 }
12641 /* A new variable offset is created. If the subtrahend is known
12642 * nonnegative, then any reg->range we had before is still good.
12643 */
12644 if (signed_sub_overflows(smin_ptr, smax_val) ||
12645 signed_sub_overflows(smax_ptr, smin_val)) {
12646 /* Overflow possible, we know nothing */
12647 dst_reg->smin_value = S64_MIN;
12648 dst_reg->smax_value = S64_MAX;
12649 } else {
12650 dst_reg->smin_value = smin_ptr - smax_val;
12651 dst_reg->smax_value = smax_ptr - smin_val;
12652 }
12653 if (umin_ptr < umax_val) {
12654 /* Overflow possible, we know nothing */
12655 dst_reg->umin_value = 0;
12656 dst_reg->umax_value = U64_MAX;
12657 } else {
12658 /* Cannot overflow (as long as bounds are consistent) */
12659 dst_reg->umin_value = umin_ptr - umax_val;
12660 dst_reg->umax_value = umax_ptr - umin_val;
12661 }
12662 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
12663 dst_reg->off = ptr_reg->off;
12664 dst_reg->raw = ptr_reg->raw;
12665 if (reg_is_pkt_pointer(ptr_reg)) {
12666 dst_reg->id = ++env->id_gen;
12667 /* something was added to pkt_ptr, set range to zero */
12668 if (smin_val < 0)
12669 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
12670 }
12671 break;
12672 case BPF_AND:
12673 case BPF_OR:
12674 case BPF_XOR:
12675 /* bitwise ops on pointers are troublesome, prohibit. */
12676 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
12677 dst, bpf_alu_string[opcode >> 4]);
12678 return -EACCES;
12679 default:
12680 /* other operators (e.g. MUL,LSH) produce non-pointer results */
12681 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
12682 dst, bpf_alu_string[opcode >> 4]);
12683 return -EACCES;
12684 }
12685
12686 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
12687 return -EINVAL;
12688 reg_bounds_sync(dst_reg);
12689 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
12690 return -EACCES;
12691 if (sanitize_needed(opcode)) {
12692 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
12693 &info, true);
12694 if (ret < 0)
12695 return sanitize_err(env, insn, ret, off_reg, dst_reg);
12696 }
12697
12698 return 0;
12699 }
12700
scalar32_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12701 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
12702 struct bpf_reg_state *src_reg)
12703 {
12704 s32 smin_val = src_reg->s32_min_value;
12705 s32 smax_val = src_reg->s32_max_value;
12706 u32 umin_val = src_reg->u32_min_value;
12707 u32 umax_val = src_reg->u32_max_value;
12708
12709 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
12710 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
12711 dst_reg->s32_min_value = S32_MIN;
12712 dst_reg->s32_max_value = S32_MAX;
12713 } else {
12714 dst_reg->s32_min_value += smin_val;
12715 dst_reg->s32_max_value += smax_val;
12716 }
12717 if (dst_reg->u32_min_value + umin_val < umin_val ||
12718 dst_reg->u32_max_value + umax_val < umax_val) {
12719 dst_reg->u32_min_value = 0;
12720 dst_reg->u32_max_value = U32_MAX;
12721 } else {
12722 dst_reg->u32_min_value += umin_val;
12723 dst_reg->u32_max_value += umax_val;
12724 }
12725 }
12726
scalar_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12727 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
12728 struct bpf_reg_state *src_reg)
12729 {
12730 s64 smin_val = src_reg->smin_value;
12731 s64 smax_val = src_reg->smax_value;
12732 u64 umin_val = src_reg->umin_value;
12733 u64 umax_val = src_reg->umax_value;
12734
12735 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
12736 signed_add_overflows(dst_reg->smax_value, smax_val)) {
12737 dst_reg->smin_value = S64_MIN;
12738 dst_reg->smax_value = S64_MAX;
12739 } else {
12740 dst_reg->smin_value += smin_val;
12741 dst_reg->smax_value += smax_val;
12742 }
12743 if (dst_reg->umin_value + umin_val < umin_val ||
12744 dst_reg->umax_value + umax_val < umax_val) {
12745 dst_reg->umin_value = 0;
12746 dst_reg->umax_value = U64_MAX;
12747 } else {
12748 dst_reg->umin_value += umin_val;
12749 dst_reg->umax_value += umax_val;
12750 }
12751 }
12752
scalar32_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12753 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
12754 struct bpf_reg_state *src_reg)
12755 {
12756 s32 smin_val = src_reg->s32_min_value;
12757 s32 smax_val = src_reg->s32_max_value;
12758 u32 umin_val = src_reg->u32_min_value;
12759 u32 umax_val = src_reg->u32_max_value;
12760
12761 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
12762 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
12763 /* Overflow possible, we know nothing */
12764 dst_reg->s32_min_value = S32_MIN;
12765 dst_reg->s32_max_value = S32_MAX;
12766 } else {
12767 dst_reg->s32_min_value -= smax_val;
12768 dst_reg->s32_max_value -= smin_val;
12769 }
12770 if (dst_reg->u32_min_value < umax_val) {
12771 /* Overflow possible, we know nothing */
12772 dst_reg->u32_min_value = 0;
12773 dst_reg->u32_max_value = U32_MAX;
12774 } else {
12775 /* Cannot overflow (as long as bounds are consistent) */
12776 dst_reg->u32_min_value -= umax_val;
12777 dst_reg->u32_max_value -= umin_val;
12778 }
12779 }
12780
scalar_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12781 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
12782 struct bpf_reg_state *src_reg)
12783 {
12784 s64 smin_val = src_reg->smin_value;
12785 s64 smax_val = src_reg->smax_value;
12786 u64 umin_val = src_reg->umin_value;
12787 u64 umax_val = src_reg->umax_value;
12788
12789 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
12790 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
12791 /* Overflow possible, we know nothing */
12792 dst_reg->smin_value = S64_MIN;
12793 dst_reg->smax_value = S64_MAX;
12794 } else {
12795 dst_reg->smin_value -= smax_val;
12796 dst_reg->smax_value -= smin_val;
12797 }
12798 if (dst_reg->umin_value < umax_val) {
12799 /* Overflow possible, we know nothing */
12800 dst_reg->umin_value = 0;
12801 dst_reg->umax_value = U64_MAX;
12802 } else {
12803 /* Cannot overflow (as long as bounds are consistent) */
12804 dst_reg->umin_value -= umax_val;
12805 dst_reg->umax_value -= umin_val;
12806 }
12807 }
12808
scalar32_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12809 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
12810 struct bpf_reg_state *src_reg)
12811 {
12812 s32 smin_val = src_reg->s32_min_value;
12813 u32 umin_val = src_reg->u32_min_value;
12814 u32 umax_val = src_reg->u32_max_value;
12815
12816 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
12817 /* Ain't nobody got time to multiply that sign */
12818 __mark_reg32_unbounded(dst_reg);
12819 return;
12820 }
12821 /* Both values are positive, so we can work with unsigned and
12822 * copy the result to signed (unless it exceeds S32_MAX).
12823 */
12824 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
12825 /* Potential overflow, we know nothing */
12826 __mark_reg32_unbounded(dst_reg);
12827 return;
12828 }
12829 dst_reg->u32_min_value *= umin_val;
12830 dst_reg->u32_max_value *= umax_val;
12831 if (dst_reg->u32_max_value > S32_MAX) {
12832 /* Overflow possible, we know nothing */
12833 dst_reg->s32_min_value = S32_MIN;
12834 dst_reg->s32_max_value = S32_MAX;
12835 } else {
12836 dst_reg->s32_min_value = dst_reg->u32_min_value;
12837 dst_reg->s32_max_value = dst_reg->u32_max_value;
12838 }
12839 }
12840
scalar_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12841 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
12842 struct bpf_reg_state *src_reg)
12843 {
12844 s64 smin_val = src_reg->smin_value;
12845 u64 umin_val = src_reg->umin_value;
12846 u64 umax_val = src_reg->umax_value;
12847
12848 if (smin_val < 0 || dst_reg->smin_value < 0) {
12849 /* Ain't nobody got time to multiply that sign */
12850 __mark_reg64_unbounded(dst_reg);
12851 return;
12852 }
12853 /* Both values are positive, so we can work with unsigned and
12854 * copy the result to signed (unless it exceeds S64_MAX).
12855 */
12856 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
12857 /* Potential overflow, we know nothing */
12858 __mark_reg64_unbounded(dst_reg);
12859 return;
12860 }
12861 dst_reg->umin_value *= umin_val;
12862 dst_reg->umax_value *= umax_val;
12863 if (dst_reg->umax_value > S64_MAX) {
12864 /* Overflow possible, we know nothing */
12865 dst_reg->smin_value = S64_MIN;
12866 dst_reg->smax_value = S64_MAX;
12867 } else {
12868 dst_reg->smin_value = dst_reg->umin_value;
12869 dst_reg->smax_value = dst_reg->umax_value;
12870 }
12871 }
12872
scalar32_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12873 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
12874 struct bpf_reg_state *src_reg)
12875 {
12876 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12877 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12878 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12879 s32 smin_val = src_reg->s32_min_value;
12880 u32 umax_val = src_reg->u32_max_value;
12881
12882 if (src_known && dst_known) {
12883 __mark_reg32_known(dst_reg, var32_off.value);
12884 return;
12885 }
12886
12887 /* We get our minimum from the var_off, since that's inherently
12888 * bitwise. Our maximum is the minimum of the operands' maxima.
12889 */
12890 dst_reg->u32_min_value = var32_off.value;
12891 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
12892 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12893 /* Lose signed bounds when ANDing negative numbers,
12894 * ain't nobody got time for that.
12895 */
12896 dst_reg->s32_min_value = S32_MIN;
12897 dst_reg->s32_max_value = S32_MAX;
12898 } else {
12899 /* ANDing two positives gives a positive, so safe to
12900 * cast result into s64.
12901 */
12902 dst_reg->s32_min_value = dst_reg->u32_min_value;
12903 dst_reg->s32_max_value = dst_reg->u32_max_value;
12904 }
12905 }
12906
scalar_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12907 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
12908 struct bpf_reg_state *src_reg)
12909 {
12910 bool src_known = tnum_is_const(src_reg->var_off);
12911 bool dst_known = tnum_is_const(dst_reg->var_off);
12912 s64 smin_val = src_reg->smin_value;
12913 u64 umax_val = src_reg->umax_value;
12914
12915 if (src_known && dst_known) {
12916 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12917 return;
12918 }
12919
12920 /* We get our minimum from the var_off, since that's inherently
12921 * bitwise. Our maximum is the minimum of the operands' maxima.
12922 */
12923 dst_reg->umin_value = dst_reg->var_off.value;
12924 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
12925 if (dst_reg->smin_value < 0 || smin_val < 0) {
12926 /* Lose signed bounds when ANDing negative numbers,
12927 * ain't nobody got time for that.
12928 */
12929 dst_reg->smin_value = S64_MIN;
12930 dst_reg->smax_value = S64_MAX;
12931 } else {
12932 /* ANDing two positives gives a positive, so safe to
12933 * cast result into s64.
12934 */
12935 dst_reg->smin_value = dst_reg->umin_value;
12936 dst_reg->smax_value = dst_reg->umax_value;
12937 }
12938 /* We may learn something more from the var_off */
12939 __update_reg_bounds(dst_reg);
12940 }
12941
scalar32_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12942 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
12943 struct bpf_reg_state *src_reg)
12944 {
12945 bool src_known = tnum_subreg_is_const(src_reg->var_off);
12946 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
12947 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
12948 s32 smin_val = src_reg->s32_min_value;
12949 u32 umin_val = src_reg->u32_min_value;
12950
12951 if (src_known && dst_known) {
12952 __mark_reg32_known(dst_reg, var32_off.value);
12953 return;
12954 }
12955
12956 /* We get our maximum from the var_off, and our minimum is the
12957 * maximum of the operands' minima
12958 */
12959 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
12960 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
12961 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
12962 /* Lose signed bounds when ORing negative numbers,
12963 * ain't nobody got time for that.
12964 */
12965 dst_reg->s32_min_value = S32_MIN;
12966 dst_reg->s32_max_value = S32_MAX;
12967 } else {
12968 /* ORing two positives gives a positive, so safe to
12969 * cast result into s64.
12970 */
12971 dst_reg->s32_min_value = dst_reg->u32_min_value;
12972 dst_reg->s32_max_value = dst_reg->u32_max_value;
12973 }
12974 }
12975
scalar_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)12976 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
12977 struct bpf_reg_state *src_reg)
12978 {
12979 bool src_known = tnum_is_const(src_reg->var_off);
12980 bool dst_known = tnum_is_const(dst_reg->var_off);
12981 s64 smin_val = src_reg->smin_value;
12982 u64 umin_val = src_reg->umin_value;
12983
12984 if (src_known && dst_known) {
12985 __mark_reg_known(dst_reg, dst_reg->var_off.value);
12986 return;
12987 }
12988
12989 /* We get our maximum from the var_off, and our minimum is the
12990 * maximum of the operands' minima
12991 */
12992 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
12993 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
12994 if (dst_reg->smin_value < 0 || smin_val < 0) {
12995 /* Lose signed bounds when ORing negative numbers,
12996 * ain't nobody got time for that.
12997 */
12998 dst_reg->smin_value = S64_MIN;
12999 dst_reg->smax_value = S64_MAX;
13000 } else {
13001 /* ORing two positives gives a positive, so safe to
13002 * cast result into s64.
13003 */
13004 dst_reg->smin_value = dst_reg->umin_value;
13005 dst_reg->smax_value = dst_reg->umax_value;
13006 }
13007 /* We may learn something more from the var_off */
13008 __update_reg_bounds(dst_reg);
13009 }
13010
scalar32_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13011 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
13012 struct bpf_reg_state *src_reg)
13013 {
13014 bool src_known = tnum_subreg_is_const(src_reg->var_off);
13015 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
13016 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
13017 s32 smin_val = src_reg->s32_min_value;
13018
13019 if (src_known && dst_known) {
13020 __mark_reg32_known(dst_reg, var32_off.value);
13021 return;
13022 }
13023
13024 /* We get both minimum and maximum from the var32_off. */
13025 dst_reg->u32_min_value = var32_off.value;
13026 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
13027
13028 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
13029 /* XORing two positive sign numbers gives a positive,
13030 * so safe to cast u32 result into s32.
13031 */
13032 dst_reg->s32_min_value = dst_reg->u32_min_value;
13033 dst_reg->s32_max_value = dst_reg->u32_max_value;
13034 } else {
13035 dst_reg->s32_min_value = S32_MIN;
13036 dst_reg->s32_max_value = S32_MAX;
13037 }
13038 }
13039
scalar_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13040 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
13041 struct bpf_reg_state *src_reg)
13042 {
13043 bool src_known = tnum_is_const(src_reg->var_off);
13044 bool dst_known = tnum_is_const(dst_reg->var_off);
13045 s64 smin_val = src_reg->smin_value;
13046
13047 if (src_known && dst_known) {
13048 /* dst_reg->var_off.value has been updated earlier */
13049 __mark_reg_known(dst_reg, dst_reg->var_off.value);
13050 return;
13051 }
13052
13053 /* We get both minimum and maximum from the var_off. */
13054 dst_reg->umin_value = dst_reg->var_off.value;
13055 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
13056
13057 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
13058 /* XORing two positive sign numbers gives a positive,
13059 * so safe to cast u64 result into s64.
13060 */
13061 dst_reg->smin_value = dst_reg->umin_value;
13062 dst_reg->smax_value = dst_reg->umax_value;
13063 } else {
13064 dst_reg->smin_value = S64_MIN;
13065 dst_reg->smax_value = S64_MAX;
13066 }
13067
13068 __update_reg_bounds(dst_reg);
13069 }
13070
__scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)13071 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
13072 u64 umin_val, u64 umax_val)
13073 {
13074 /* We lose all sign bit information (except what we can pick
13075 * up from var_off)
13076 */
13077 dst_reg->s32_min_value = S32_MIN;
13078 dst_reg->s32_max_value = S32_MAX;
13079 /* If we might shift our top bit out, then we know nothing */
13080 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
13081 dst_reg->u32_min_value = 0;
13082 dst_reg->u32_max_value = U32_MAX;
13083 } else {
13084 dst_reg->u32_min_value <<= umin_val;
13085 dst_reg->u32_max_value <<= umax_val;
13086 }
13087 }
13088
scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13089 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
13090 struct bpf_reg_state *src_reg)
13091 {
13092 u32 umax_val = src_reg->u32_max_value;
13093 u32 umin_val = src_reg->u32_min_value;
13094 /* u32 alu operation will zext upper bits */
13095 struct tnum subreg = tnum_subreg(dst_reg->var_off);
13096
13097 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
13098 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
13099 /* Not required but being careful mark reg64 bounds as unknown so
13100 * that we are forced to pick them up from tnum and zext later and
13101 * if some path skips this step we are still safe.
13102 */
13103 __mark_reg64_unbounded(dst_reg);
13104 __update_reg32_bounds(dst_reg);
13105 }
13106
__scalar64_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)13107 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
13108 u64 umin_val, u64 umax_val)
13109 {
13110 /* Special case <<32 because it is a common compiler pattern to sign
13111 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
13112 * positive we know this shift will also be positive so we can track
13113 * bounds correctly. Otherwise we lose all sign bit information except
13114 * what we can pick up from var_off. Perhaps we can generalize this
13115 * later to shifts of any length.
13116 */
13117 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
13118 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
13119 else
13120 dst_reg->smax_value = S64_MAX;
13121
13122 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
13123 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
13124 else
13125 dst_reg->smin_value = S64_MIN;
13126
13127 /* If we might shift our top bit out, then we know nothing */
13128 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
13129 dst_reg->umin_value = 0;
13130 dst_reg->umax_value = U64_MAX;
13131 } else {
13132 dst_reg->umin_value <<= umin_val;
13133 dst_reg->umax_value <<= umax_val;
13134 }
13135 }
13136
scalar_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13137 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
13138 struct bpf_reg_state *src_reg)
13139 {
13140 u64 umax_val = src_reg->umax_value;
13141 u64 umin_val = src_reg->umin_value;
13142
13143 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
13144 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
13145 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
13146
13147 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
13148 /* We may learn something more from the var_off */
13149 __update_reg_bounds(dst_reg);
13150 }
13151
scalar32_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13152 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
13153 struct bpf_reg_state *src_reg)
13154 {
13155 struct tnum subreg = tnum_subreg(dst_reg->var_off);
13156 u32 umax_val = src_reg->u32_max_value;
13157 u32 umin_val = src_reg->u32_min_value;
13158
13159 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
13160 * be negative, then either:
13161 * 1) src_reg might be zero, so the sign bit of the result is
13162 * unknown, so we lose our signed bounds
13163 * 2) it's known negative, thus the unsigned bounds capture the
13164 * signed bounds
13165 * 3) the signed bounds cross zero, so they tell us nothing
13166 * about the result
13167 * If the value in dst_reg is known nonnegative, then again the
13168 * unsigned bounds capture the signed bounds.
13169 * Thus, in all cases it suffices to blow away our signed bounds
13170 * and rely on inferring new ones from the unsigned bounds and
13171 * var_off of the result.
13172 */
13173 dst_reg->s32_min_value = S32_MIN;
13174 dst_reg->s32_max_value = S32_MAX;
13175
13176 dst_reg->var_off = tnum_rshift(subreg, umin_val);
13177 dst_reg->u32_min_value >>= umax_val;
13178 dst_reg->u32_max_value >>= umin_val;
13179
13180 __mark_reg64_unbounded(dst_reg);
13181 __update_reg32_bounds(dst_reg);
13182 }
13183
scalar_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13184 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
13185 struct bpf_reg_state *src_reg)
13186 {
13187 u64 umax_val = src_reg->umax_value;
13188 u64 umin_val = src_reg->umin_value;
13189
13190 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
13191 * be negative, then either:
13192 * 1) src_reg might be zero, so the sign bit of the result is
13193 * unknown, so we lose our signed bounds
13194 * 2) it's known negative, thus the unsigned bounds capture the
13195 * signed bounds
13196 * 3) the signed bounds cross zero, so they tell us nothing
13197 * about the result
13198 * If the value in dst_reg is known nonnegative, then again the
13199 * unsigned bounds capture the signed bounds.
13200 * Thus, in all cases it suffices to blow away our signed bounds
13201 * and rely on inferring new ones from the unsigned bounds and
13202 * var_off of the result.
13203 */
13204 dst_reg->smin_value = S64_MIN;
13205 dst_reg->smax_value = S64_MAX;
13206 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
13207 dst_reg->umin_value >>= umax_val;
13208 dst_reg->umax_value >>= umin_val;
13209
13210 /* Its not easy to operate on alu32 bounds here because it depends
13211 * on bits being shifted in. Take easy way out and mark unbounded
13212 * so we can recalculate later from tnum.
13213 */
13214 __mark_reg32_unbounded(dst_reg);
13215 __update_reg_bounds(dst_reg);
13216 }
13217
scalar32_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13218 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
13219 struct bpf_reg_state *src_reg)
13220 {
13221 u64 umin_val = src_reg->u32_min_value;
13222
13223 /* Upon reaching here, src_known is true and
13224 * umax_val is equal to umin_val.
13225 */
13226 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
13227 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
13228
13229 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
13230
13231 /* blow away the dst_reg umin_value/umax_value and rely on
13232 * dst_reg var_off to refine the result.
13233 */
13234 dst_reg->u32_min_value = 0;
13235 dst_reg->u32_max_value = U32_MAX;
13236
13237 __mark_reg64_unbounded(dst_reg);
13238 __update_reg32_bounds(dst_reg);
13239 }
13240
scalar_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)13241 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
13242 struct bpf_reg_state *src_reg)
13243 {
13244 u64 umin_val = src_reg->umin_value;
13245
13246 /* Upon reaching here, src_known is true and umax_val is equal
13247 * to umin_val.
13248 */
13249 dst_reg->smin_value >>= umin_val;
13250 dst_reg->smax_value >>= umin_val;
13251
13252 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
13253
13254 /* blow away the dst_reg umin_value/umax_value and rely on
13255 * dst_reg var_off to refine the result.
13256 */
13257 dst_reg->umin_value = 0;
13258 dst_reg->umax_value = U64_MAX;
13259
13260 /* Its not easy to operate on alu32 bounds here because it depends
13261 * on bits being shifted in from upper 32-bits. Take easy way out
13262 * and mark unbounded so we can recalculate later from tnum.
13263 */
13264 __mark_reg32_unbounded(dst_reg);
13265 __update_reg_bounds(dst_reg);
13266 }
13267
13268 /* WARNING: This function does calculations on 64-bit values, but the actual
13269 * execution may occur on 32-bit values. Therefore, things like bitshifts
13270 * need extra checks in the 32-bit case.
13271 */
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)13272 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
13273 struct bpf_insn *insn,
13274 struct bpf_reg_state *dst_reg,
13275 struct bpf_reg_state src_reg)
13276 {
13277 struct bpf_reg_state *regs = cur_regs(env);
13278 u8 opcode = BPF_OP(insn->code);
13279 bool src_known;
13280 s64 smin_val, smax_val;
13281 u64 umin_val, umax_val;
13282 s32 s32_min_val, s32_max_val;
13283 u32 u32_min_val, u32_max_val;
13284 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
13285 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
13286 int ret;
13287
13288 smin_val = src_reg.smin_value;
13289 smax_val = src_reg.smax_value;
13290 umin_val = src_reg.umin_value;
13291 umax_val = src_reg.umax_value;
13292
13293 s32_min_val = src_reg.s32_min_value;
13294 s32_max_val = src_reg.s32_max_value;
13295 u32_min_val = src_reg.u32_min_value;
13296 u32_max_val = src_reg.u32_max_value;
13297
13298 if (alu32) {
13299 src_known = tnum_subreg_is_const(src_reg.var_off);
13300 if ((src_known &&
13301 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
13302 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
13303 /* Taint dst register if offset had invalid bounds
13304 * derived from e.g. dead branches.
13305 */
13306 __mark_reg_unknown(env, dst_reg);
13307 return 0;
13308 }
13309 } else {
13310 src_known = tnum_is_const(src_reg.var_off);
13311 if ((src_known &&
13312 (smin_val != smax_val || umin_val != umax_val)) ||
13313 smin_val > smax_val || umin_val > umax_val) {
13314 /* Taint dst register if offset had invalid bounds
13315 * derived from e.g. dead branches.
13316 */
13317 __mark_reg_unknown(env, dst_reg);
13318 return 0;
13319 }
13320 }
13321
13322 if (!src_known &&
13323 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
13324 __mark_reg_unknown(env, dst_reg);
13325 return 0;
13326 }
13327
13328 if (sanitize_needed(opcode)) {
13329 ret = sanitize_val_alu(env, insn);
13330 if (ret < 0)
13331 return sanitize_err(env, insn, ret, NULL, NULL);
13332 }
13333
13334 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
13335 * There are two classes of instructions: The first class we track both
13336 * alu32 and alu64 sign/unsigned bounds independently this provides the
13337 * greatest amount of precision when alu operations are mixed with jmp32
13338 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
13339 * and BPF_OR. This is possible because these ops have fairly easy to
13340 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
13341 * See alu32 verifier tests for examples. The second class of
13342 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
13343 * with regards to tracking sign/unsigned bounds because the bits may
13344 * cross subreg boundaries in the alu64 case. When this happens we mark
13345 * the reg unbounded in the subreg bound space and use the resulting
13346 * tnum to calculate an approximation of the sign/unsigned bounds.
13347 */
13348 switch (opcode) {
13349 case BPF_ADD:
13350 scalar32_min_max_add(dst_reg, &src_reg);
13351 scalar_min_max_add(dst_reg, &src_reg);
13352 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
13353 break;
13354 case BPF_SUB:
13355 scalar32_min_max_sub(dst_reg, &src_reg);
13356 scalar_min_max_sub(dst_reg, &src_reg);
13357 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
13358 break;
13359 case BPF_MUL:
13360 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
13361 scalar32_min_max_mul(dst_reg, &src_reg);
13362 scalar_min_max_mul(dst_reg, &src_reg);
13363 break;
13364 case BPF_AND:
13365 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
13366 scalar32_min_max_and(dst_reg, &src_reg);
13367 scalar_min_max_and(dst_reg, &src_reg);
13368 break;
13369 case BPF_OR:
13370 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
13371 scalar32_min_max_or(dst_reg, &src_reg);
13372 scalar_min_max_or(dst_reg, &src_reg);
13373 break;
13374 case BPF_XOR:
13375 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
13376 scalar32_min_max_xor(dst_reg, &src_reg);
13377 scalar_min_max_xor(dst_reg, &src_reg);
13378 break;
13379 case BPF_LSH:
13380 if (umax_val >= insn_bitness) {
13381 /* Shifts greater than 31 or 63 are undefined.
13382 * This includes shifts by a negative number.
13383 */
13384 mark_reg_unknown(env, regs, insn->dst_reg);
13385 break;
13386 }
13387 if (alu32)
13388 scalar32_min_max_lsh(dst_reg, &src_reg);
13389 else
13390 scalar_min_max_lsh(dst_reg, &src_reg);
13391 break;
13392 case BPF_RSH:
13393 if (umax_val >= insn_bitness) {
13394 /* Shifts greater than 31 or 63 are undefined.
13395 * This includes shifts by a negative number.
13396 */
13397 mark_reg_unknown(env, regs, insn->dst_reg);
13398 break;
13399 }
13400 if (alu32)
13401 scalar32_min_max_rsh(dst_reg, &src_reg);
13402 else
13403 scalar_min_max_rsh(dst_reg, &src_reg);
13404 break;
13405 case BPF_ARSH:
13406 if (umax_val >= insn_bitness) {
13407 /* Shifts greater than 31 or 63 are undefined.
13408 * This includes shifts by a negative number.
13409 */
13410 mark_reg_unknown(env, regs, insn->dst_reg);
13411 break;
13412 }
13413 if (alu32)
13414 scalar32_min_max_arsh(dst_reg, &src_reg);
13415 else
13416 scalar_min_max_arsh(dst_reg, &src_reg);
13417 break;
13418 default:
13419 mark_reg_unknown(env, regs, insn->dst_reg);
13420 break;
13421 }
13422
13423 /* ALU32 ops are zero extended into 64bit register */
13424 if (alu32)
13425 zext_32_to_64(dst_reg);
13426 reg_bounds_sync(dst_reg);
13427 return 0;
13428 }
13429
13430 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
13431 * and var_off.
13432 */
adjust_reg_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn)13433 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
13434 struct bpf_insn *insn)
13435 {
13436 struct bpf_verifier_state *vstate = env->cur_state;
13437 struct bpf_func_state *state = vstate->frame[vstate->curframe];
13438 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
13439 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
13440 u8 opcode = BPF_OP(insn->code);
13441 int err;
13442
13443 dst_reg = ®s[insn->dst_reg];
13444 src_reg = NULL;
13445 if (dst_reg->type != SCALAR_VALUE)
13446 ptr_reg = dst_reg;
13447 else
13448 /* Make sure ID is cleared otherwise dst_reg min/max could be
13449 * incorrectly propagated into other registers by find_equal_scalars()
13450 */
13451 dst_reg->id = 0;
13452 if (BPF_SRC(insn->code) == BPF_X) {
13453 src_reg = ®s[insn->src_reg];
13454 if (src_reg->type != SCALAR_VALUE) {
13455 if (dst_reg->type != SCALAR_VALUE) {
13456 /* Combining two pointers by any ALU op yields
13457 * an arbitrary scalar. Disallow all math except
13458 * pointer subtraction
13459 */
13460 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
13461 mark_reg_unknown(env, regs, insn->dst_reg);
13462 return 0;
13463 }
13464 verbose(env, "R%d pointer %s pointer prohibited\n",
13465 insn->dst_reg,
13466 bpf_alu_string[opcode >> 4]);
13467 return -EACCES;
13468 } else {
13469 /* scalar += pointer
13470 * This is legal, but we have to reverse our
13471 * src/dest handling in computing the range
13472 */
13473 err = mark_chain_precision(env, insn->dst_reg);
13474 if (err)
13475 return err;
13476 return adjust_ptr_min_max_vals(env, insn,
13477 src_reg, dst_reg);
13478 }
13479 } else if (ptr_reg) {
13480 /* pointer += scalar */
13481 err = mark_chain_precision(env, insn->src_reg);
13482 if (err)
13483 return err;
13484 return adjust_ptr_min_max_vals(env, insn,
13485 dst_reg, src_reg);
13486 } else if (dst_reg->precise) {
13487 /* if dst_reg is precise, src_reg should be precise as well */
13488 err = mark_chain_precision(env, insn->src_reg);
13489 if (err)
13490 return err;
13491 }
13492 } else {
13493 /* Pretend the src is a reg with a known value, since we only
13494 * need to be able to read from this state.
13495 */
13496 off_reg.type = SCALAR_VALUE;
13497 __mark_reg_known(&off_reg, insn->imm);
13498 src_reg = &off_reg;
13499 if (ptr_reg) /* pointer += K */
13500 return adjust_ptr_min_max_vals(env, insn,
13501 ptr_reg, src_reg);
13502 }
13503
13504 /* Got here implies adding two SCALAR_VALUEs */
13505 if (WARN_ON_ONCE(ptr_reg)) {
13506 print_verifier_state(env, state, true);
13507 verbose(env, "verifier internal error: unexpected ptr_reg\n");
13508 return -EINVAL;
13509 }
13510 if (WARN_ON(!src_reg)) {
13511 print_verifier_state(env, state, true);
13512 verbose(env, "verifier internal error: no src_reg\n");
13513 return -EINVAL;
13514 }
13515 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
13516 }
13517
13518 /* check validity of 32-bit and 64-bit arithmetic operations */
check_alu_op(struct bpf_verifier_env * env,struct bpf_insn * insn)13519 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
13520 {
13521 struct bpf_reg_state *regs = cur_regs(env);
13522 u8 opcode = BPF_OP(insn->code);
13523 int err;
13524
13525 if (opcode == BPF_END || opcode == BPF_NEG) {
13526 if (opcode == BPF_NEG) {
13527 if (BPF_SRC(insn->code) != BPF_K ||
13528 insn->src_reg != BPF_REG_0 ||
13529 insn->off != 0 || insn->imm != 0) {
13530 verbose(env, "BPF_NEG uses reserved fields\n");
13531 return -EINVAL;
13532 }
13533 } else {
13534 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
13535 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
13536 (BPF_CLASS(insn->code) == BPF_ALU64 &&
13537 BPF_SRC(insn->code) != BPF_TO_LE)) {
13538 verbose(env, "BPF_END uses reserved fields\n");
13539 return -EINVAL;
13540 }
13541 }
13542
13543 /* check src operand */
13544 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13545 if (err)
13546 return err;
13547
13548 if (is_pointer_value(env, insn->dst_reg)) {
13549 verbose(env, "R%d pointer arithmetic prohibited\n",
13550 insn->dst_reg);
13551 return -EACCES;
13552 }
13553
13554 /* check dest operand */
13555 err = check_reg_arg(env, insn->dst_reg, DST_OP);
13556 if (err)
13557 return err;
13558
13559 } else if (opcode == BPF_MOV) {
13560
13561 if (BPF_SRC(insn->code) == BPF_X) {
13562 if (insn->imm != 0) {
13563 verbose(env, "BPF_MOV uses reserved fields\n");
13564 return -EINVAL;
13565 }
13566
13567 if (BPF_CLASS(insn->code) == BPF_ALU) {
13568 if (insn->off != 0 && insn->off != 8 && insn->off != 16) {
13569 verbose(env, "BPF_MOV uses reserved fields\n");
13570 return -EINVAL;
13571 }
13572 } else {
13573 if (insn->off != 0 && insn->off != 8 && insn->off != 16 &&
13574 insn->off != 32) {
13575 verbose(env, "BPF_MOV uses reserved fields\n");
13576 return -EINVAL;
13577 }
13578 }
13579
13580 /* check src operand */
13581 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13582 if (err)
13583 return err;
13584 } else {
13585 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
13586 verbose(env, "BPF_MOV uses reserved fields\n");
13587 return -EINVAL;
13588 }
13589 }
13590
13591 /* check dest operand, mark as required later */
13592 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13593 if (err)
13594 return err;
13595
13596 if (BPF_SRC(insn->code) == BPF_X) {
13597 struct bpf_reg_state *src_reg = regs + insn->src_reg;
13598 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
13599 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id &&
13600 !tnum_is_const(src_reg->var_off);
13601
13602 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13603 if (insn->off == 0) {
13604 /* case: R1 = R2
13605 * copy register state to dest reg
13606 */
13607 if (need_id)
13608 /* Assign src and dst registers the same ID
13609 * that will be used by find_equal_scalars()
13610 * to propagate min/max range.
13611 */
13612 src_reg->id = ++env->id_gen;
13613 copy_register_state(dst_reg, src_reg);
13614 dst_reg->live |= REG_LIVE_WRITTEN;
13615 dst_reg->subreg_def = DEF_NOT_SUBREG;
13616 } else {
13617 /* case: R1 = (s8, s16 s32)R2 */
13618 if (is_pointer_value(env, insn->src_reg)) {
13619 verbose(env,
13620 "R%d sign-extension part of pointer\n",
13621 insn->src_reg);
13622 return -EACCES;
13623 } else if (src_reg->type == SCALAR_VALUE) {
13624 bool no_sext;
13625
13626 no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13627 if (no_sext && need_id)
13628 src_reg->id = ++env->id_gen;
13629 copy_register_state(dst_reg, src_reg);
13630 if (!no_sext)
13631 dst_reg->id = 0;
13632 coerce_reg_to_size_sx(dst_reg, insn->off >> 3);
13633 dst_reg->live |= REG_LIVE_WRITTEN;
13634 dst_reg->subreg_def = DEF_NOT_SUBREG;
13635 } else {
13636 mark_reg_unknown(env, regs, insn->dst_reg);
13637 }
13638 }
13639 } else {
13640 /* R1 = (u32) R2 */
13641 if (is_pointer_value(env, insn->src_reg)) {
13642 verbose(env,
13643 "R%d partial copy of pointer\n",
13644 insn->src_reg);
13645 return -EACCES;
13646 } else if (src_reg->type == SCALAR_VALUE) {
13647 if (insn->off == 0) {
13648 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX;
13649
13650 if (is_src_reg_u32 && need_id)
13651 src_reg->id = ++env->id_gen;
13652 copy_register_state(dst_reg, src_reg);
13653 /* Make sure ID is cleared if src_reg is not in u32
13654 * range otherwise dst_reg min/max could be incorrectly
13655 * propagated into src_reg by find_equal_scalars()
13656 */
13657 if (!is_src_reg_u32)
13658 dst_reg->id = 0;
13659 dst_reg->live |= REG_LIVE_WRITTEN;
13660 dst_reg->subreg_def = env->insn_idx + 1;
13661 } else {
13662 /* case: W1 = (s8, s16)W2 */
13663 bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
13664
13665 if (no_sext && need_id)
13666 src_reg->id = ++env->id_gen;
13667 copy_register_state(dst_reg, src_reg);
13668 if (!no_sext)
13669 dst_reg->id = 0;
13670 dst_reg->live |= REG_LIVE_WRITTEN;
13671 dst_reg->subreg_def = env->insn_idx + 1;
13672 coerce_subreg_to_size_sx(dst_reg, insn->off >> 3);
13673 }
13674 } else {
13675 mark_reg_unknown(env, regs,
13676 insn->dst_reg);
13677 }
13678 zext_32_to_64(dst_reg);
13679 reg_bounds_sync(dst_reg);
13680 }
13681 } else {
13682 /* case: R = imm
13683 * remember the value we stored into this reg
13684 */
13685 /* clear any state __mark_reg_known doesn't set */
13686 mark_reg_unknown(env, regs, insn->dst_reg);
13687 regs[insn->dst_reg].type = SCALAR_VALUE;
13688 if (BPF_CLASS(insn->code) == BPF_ALU64) {
13689 __mark_reg_known(regs + insn->dst_reg,
13690 insn->imm);
13691 } else {
13692 __mark_reg_known(regs + insn->dst_reg,
13693 (u32)insn->imm);
13694 }
13695 }
13696
13697 } else if (opcode > BPF_END) {
13698 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
13699 return -EINVAL;
13700
13701 } else { /* all other ALU ops: and, sub, xor, add, ... */
13702
13703 if (BPF_SRC(insn->code) == BPF_X) {
13704 if (insn->imm != 0 || insn->off > 1 ||
13705 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13706 verbose(env, "BPF_ALU uses reserved fields\n");
13707 return -EINVAL;
13708 }
13709 /* check src1 operand */
13710 err = check_reg_arg(env, insn->src_reg, SRC_OP);
13711 if (err)
13712 return err;
13713 } else {
13714 if (insn->src_reg != BPF_REG_0 || insn->off > 1 ||
13715 (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
13716 verbose(env, "BPF_ALU uses reserved fields\n");
13717 return -EINVAL;
13718 }
13719 }
13720
13721 /* check src2 operand */
13722 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
13723 if (err)
13724 return err;
13725
13726 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
13727 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
13728 verbose(env, "div by zero\n");
13729 return -EINVAL;
13730 }
13731
13732 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
13733 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
13734 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
13735
13736 if (insn->imm < 0 || insn->imm >= size) {
13737 verbose(env, "invalid shift %d\n", insn->imm);
13738 return -EINVAL;
13739 }
13740 }
13741
13742 /* check dest operand */
13743 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
13744 if (err)
13745 return err;
13746
13747 return adjust_reg_min_max_vals(env, insn);
13748 }
13749
13750 return 0;
13751 }
13752
find_good_pkt_pointers(struct bpf_verifier_state * vstate,struct bpf_reg_state * dst_reg,enum bpf_reg_type type,bool range_right_open)13753 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
13754 struct bpf_reg_state *dst_reg,
13755 enum bpf_reg_type type,
13756 bool range_right_open)
13757 {
13758 struct bpf_func_state *state;
13759 struct bpf_reg_state *reg;
13760 int new_range;
13761
13762 if (dst_reg->off < 0 ||
13763 (dst_reg->off == 0 && range_right_open))
13764 /* This doesn't give us any range */
13765 return;
13766
13767 if (dst_reg->umax_value > MAX_PACKET_OFF ||
13768 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
13769 /* Risk of overflow. For instance, ptr + (1<<63) may be less
13770 * than pkt_end, but that's because it's also less than pkt.
13771 */
13772 return;
13773
13774 new_range = dst_reg->off;
13775 if (range_right_open)
13776 new_range++;
13777
13778 /* Examples for register markings:
13779 *
13780 * pkt_data in dst register:
13781 *
13782 * r2 = r3;
13783 * r2 += 8;
13784 * if (r2 > pkt_end) goto <handle exception>
13785 * <access okay>
13786 *
13787 * r2 = r3;
13788 * r2 += 8;
13789 * if (r2 < pkt_end) goto <access okay>
13790 * <handle exception>
13791 *
13792 * Where:
13793 * r2 == dst_reg, pkt_end == src_reg
13794 * r2=pkt(id=n,off=8,r=0)
13795 * r3=pkt(id=n,off=0,r=0)
13796 *
13797 * pkt_data in src register:
13798 *
13799 * r2 = r3;
13800 * r2 += 8;
13801 * if (pkt_end >= r2) goto <access okay>
13802 * <handle exception>
13803 *
13804 * r2 = r3;
13805 * r2 += 8;
13806 * if (pkt_end <= r2) goto <handle exception>
13807 * <access okay>
13808 *
13809 * Where:
13810 * pkt_end == dst_reg, r2 == src_reg
13811 * r2=pkt(id=n,off=8,r=0)
13812 * r3=pkt(id=n,off=0,r=0)
13813 *
13814 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
13815 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
13816 * and [r3, r3 + 8-1) respectively is safe to access depending on
13817 * the check.
13818 */
13819
13820 /* If our ids match, then we must have the same max_value. And we
13821 * don't care about the other reg's fixed offset, since if it's too big
13822 * the range won't allow anything.
13823 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
13824 */
13825 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
13826 if (reg->type == type && reg->id == dst_reg->id)
13827 /* keep the maximum range already checked */
13828 reg->range = max(reg->range, new_range);
13829 }));
13830 }
13831
is_branch32_taken(struct bpf_reg_state * reg,u32 val,u8 opcode)13832 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
13833 {
13834 struct tnum subreg = tnum_subreg(reg->var_off);
13835 s32 sval = (s32)val;
13836
13837 switch (opcode) {
13838 case BPF_JEQ:
13839 if (tnum_is_const(subreg))
13840 return !!tnum_equals_const(subreg, val);
13841 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13842 return 0;
13843 break;
13844 case BPF_JNE:
13845 if (tnum_is_const(subreg))
13846 return !tnum_equals_const(subreg, val);
13847 else if (val < reg->u32_min_value || val > reg->u32_max_value)
13848 return 1;
13849 break;
13850 case BPF_JSET:
13851 if ((~subreg.mask & subreg.value) & val)
13852 return 1;
13853 if (!((subreg.mask | subreg.value) & val))
13854 return 0;
13855 break;
13856 case BPF_JGT:
13857 if (reg->u32_min_value > val)
13858 return 1;
13859 else if (reg->u32_max_value <= val)
13860 return 0;
13861 break;
13862 case BPF_JSGT:
13863 if (reg->s32_min_value > sval)
13864 return 1;
13865 else if (reg->s32_max_value <= sval)
13866 return 0;
13867 break;
13868 case BPF_JLT:
13869 if (reg->u32_max_value < val)
13870 return 1;
13871 else if (reg->u32_min_value >= val)
13872 return 0;
13873 break;
13874 case BPF_JSLT:
13875 if (reg->s32_max_value < sval)
13876 return 1;
13877 else if (reg->s32_min_value >= sval)
13878 return 0;
13879 break;
13880 case BPF_JGE:
13881 if (reg->u32_min_value >= val)
13882 return 1;
13883 else if (reg->u32_max_value < val)
13884 return 0;
13885 break;
13886 case BPF_JSGE:
13887 if (reg->s32_min_value >= sval)
13888 return 1;
13889 else if (reg->s32_max_value < sval)
13890 return 0;
13891 break;
13892 case BPF_JLE:
13893 if (reg->u32_max_value <= val)
13894 return 1;
13895 else if (reg->u32_min_value > val)
13896 return 0;
13897 break;
13898 case BPF_JSLE:
13899 if (reg->s32_max_value <= sval)
13900 return 1;
13901 else if (reg->s32_min_value > sval)
13902 return 0;
13903 break;
13904 }
13905
13906 return -1;
13907 }
13908
13909
is_branch64_taken(struct bpf_reg_state * reg,u64 val,u8 opcode)13910 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
13911 {
13912 s64 sval = (s64)val;
13913
13914 switch (opcode) {
13915 case BPF_JEQ:
13916 if (tnum_is_const(reg->var_off))
13917 return !!tnum_equals_const(reg->var_off, val);
13918 else if (val < reg->umin_value || val > reg->umax_value)
13919 return 0;
13920 break;
13921 case BPF_JNE:
13922 if (tnum_is_const(reg->var_off))
13923 return !tnum_equals_const(reg->var_off, val);
13924 else if (val < reg->umin_value || val > reg->umax_value)
13925 return 1;
13926 break;
13927 case BPF_JSET:
13928 if ((~reg->var_off.mask & reg->var_off.value) & val)
13929 return 1;
13930 if (!((reg->var_off.mask | reg->var_off.value) & val))
13931 return 0;
13932 break;
13933 case BPF_JGT:
13934 if (reg->umin_value > val)
13935 return 1;
13936 else if (reg->umax_value <= val)
13937 return 0;
13938 break;
13939 case BPF_JSGT:
13940 if (reg->smin_value > sval)
13941 return 1;
13942 else if (reg->smax_value <= sval)
13943 return 0;
13944 break;
13945 case BPF_JLT:
13946 if (reg->umax_value < val)
13947 return 1;
13948 else if (reg->umin_value >= val)
13949 return 0;
13950 break;
13951 case BPF_JSLT:
13952 if (reg->smax_value < sval)
13953 return 1;
13954 else if (reg->smin_value >= sval)
13955 return 0;
13956 break;
13957 case BPF_JGE:
13958 if (reg->umin_value >= val)
13959 return 1;
13960 else if (reg->umax_value < val)
13961 return 0;
13962 break;
13963 case BPF_JSGE:
13964 if (reg->smin_value >= sval)
13965 return 1;
13966 else if (reg->smax_value < sval)
13967 return 0;
13968 break;
13969 case BPF_JLE:
13970 if (reg->umax_value <= val)
13971 return 1;
13972 else if (reg->umin_value > val)
13973 return 0;
13974 break;
13975 case BPF_JSLE:
13976 if (reg->smax_value <= sval)
13977 return 1;
13978 else if (reg->smin_value > sval)
13979 return 0;
13980 break;
13981 }
13982
13983 return -1;
13984 }
13985
13986 /* compute branch direction of the expression "if (reg opcode val) goto target;"
13987 * and return:
13988 * 1 - branch will be taken and "goto target" will be executed
13989 * 0 - branch will not be taken and fall-through to next insn
13990 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
13991 * range [0,10]
13992 */
is_branch_taken(struct bpf_reg_state * reg,u64 val,u8 opcode,bool is_jmp32)13993 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
13994 bool is_jmp32)
13995 {
13996 if (__is_pointer_value(false, reg)) {
13997 if (!reg_not_null(reg))
13998 return -1;
13999
14000 /* If pointer is valid tests against zero will fail so we can
14001 * use this to direct branch taken.
14002 */
14003 if (val != 0)
14004 return -1;
14005
14006 switch (opcode) {
14007 case BPF_JEQ:
14008 return 0;
14009 case BPF_JNE:
14010 return 1;
14011 default:
14012 return -1;
14013 }
14014 }
14015
14016 if (is_jmp32)
14017 return is_branch32_taken(reg, val, opcode);
14018 return is_branch64_taken(reg, val, opcode);
14019 }
14020
flip_opcode(u32 opcode)14021 static int flip_opcode(u32 opcode)
14022 {
14023 /* How can we transform "a <op> b" into "b <op> a"? */
14024 static const u8 opcode_flip[16] = {
14025 /* these stay the same */
14026 [BPF_JEQ >> 4] = BPF_JEQ,
14027 [BPF_JNE >> 4] = BPF_JNE,
14028 [BPF_JSET >> 4] = BPF_JSET,
14029 /* these swap "lesser" and "greater" (L and G in the opcodes) */
14030 [BPF_JGE >> 4] = BPF_JLE,
14031 [BPF_JGT >> 4] = BPF_JLT,
14032 [BPF_JLE >> 4] = BPF_JGE,
14033 [BPF_JLT >> 4] = BPF_JGT,
14034 [BPF_JSGE >> 4] = BPF_JSLE,
14035 [BPF_JSGT >> 4] = BPF_JSLT,
14036 [BPF_JSLE >> 4] = BPF_JSGE,
14037 [BPF_JSLT >> 4] = BPF_JSGT
14038 };
14039 return opcode_flip[opcode >> 4];
14040 }
14041
is_pkt_ptr_branch_taken(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg,u8 opcode)14042 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
14043 struct bpf_reg_state *src_reg,
14044 u8 opcode)
14045 {
14046 struct bpf_reg_state *pkt;
14047
14048 if (src_reg->type == PTR_TO_PACKET_END) {
14049 pkt = dst_reg;
14050 } else if (dst_reg->type == PTR_TO_PACKET_END) {
14051 pkt = src_reg;
14052 opcode = flip_opcode(opcode);
14053 } else {
14054 return -1;
14055 }
14056
14057 if (pkt->range >= 0)
14058 return -1;
14059
14060 switch (opcode) {
14061 case BPF_JLE:
14062 /* pkt <= pkt_end */
14063 fallthrough;
14064 case BPF_JGT:
14065 /* pkt > pkt_end */
14066 if (pkt->range == BEYOND_PKT_END)
14067 /* pkt has at last one extra byte beyond pkt_end */
14068 return opcode == BPF_JGT;
14069 break;
14070 case BPF_JLT:
14071 /* pkt < pkt_end */
14072 fallthrough;
14073 case BPF_JGE:
14074 /* pkt >= pkt_end */
14075 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
14076 return opcode == BPF_JGE;
14077 break;
14078 }
14079 return -1;
14080 }
14081
14082 /* Adjusts the register min/max values in the case that the dst_reg is the
14083 * variable register that we are working on, and src_reg is a constant or we're
14084 * simply doing a BPF_K check.
14085 * In JEQ/JNE cases we also adjust the var_off values.
14086 */
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)14087 static void reg_set_min_max(struct bpf_reg_state *true_reg,
14088 struct bpf_reg_state *false_reg,
14089 u64 val, u32 val32,
14090 u8 opcode, bool is_jmp32)
14091 {
14092 struct tnum false_32off = tnum_subreg(false_reg->var_off);
14093 struct tnum false_64off = false_reg->var_off;
14094 struct tnum true_32off = tnum_subreg(true_reg->var_off);
14095 struct tnum true_64off = true_reg->var_off;
14096 s64 sval = (s64)val;
14097 s32 sval32 = (s32)val32;
14098
14099 /* If the dst_reg is a pointer, we can't learn anything about its
14100 * variable offset from the compare (unless src_reg were a pointer into
14101 * the same object, but we don't bother with that.
14102 * Since false_reg and true_reg have the same type by construction, we
14103 * only need to check one of them for pointerness.
14104 */
14105 if (__is_pointer_value(false, false_reg))
14106 return;
14107
14108 switch (opcode) {
14109 /* JEQ/JNE comparison doesn't change the register equivalence.
14110 *
14111 * r1 = r2;
14112 * if (r1 == 42) goto label;
14113 * ...
14114 * label: // here both r1 and r2 are known to be 42.
14115 *
14116 * Hence when marking register as known preserve it's ID.
14117 */
14118 case BPF_JEQ:
14119 if (is_jmp32) {
14120 __mark_reg32_known(true_reg, val32);
14121 true_32off = tnum_subreg(true_reg->var_off);
14122 } else {
14123 ___mark_reg_known(true_reg, val);
14124 true_64off = true_reg->var_off;
14125 }
14126 break;
14127 case BPF_JNE:
14128 if (is_jmp32) {
14129 __mark_reg32_known(false_reg, val32);
14130 false_32off = tnum_subreg(false_reg->var_off);
14131 } else {
14132 ___mark_reg_known(false_reg, val);
14133 false_64off = false_reg->var_off;
14134 }
14135 break;
14136 case BPF_JSET:
14137 if (is_jmp32) {
14138 false_32off = tnum_and(false_32off, tnum_const(~val32));
14139 if (is_power_of_2(val32))
14140 true_32off = tnum_or(true_32off,
14141 tnum_const(val32));
14142 } else {
14143 false_64off = tnum_and(false_64off, tnum_const(~val));
14144 if (is_power_of_2(val))
14145 true_64off = tnum_or(true_64off,
14146 tnum_const(val));
14147 }
14148 break;
14149 case BPF_JGE:
14150 case BPF_JGT:
14151 {
14152 if (is_jmp32) {
14153 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
14154 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
14155
14156 false_reg->u32_max_value = min(false_reg->u32_max_value,
14157 false_umax);
14158 true_reg->u32_min_value = max(true_reg->u32_min_value,
14159 true_umin);
14160 } else {
14161 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
14162 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
14163
14164 false_reg->umax_value = min(false_reg->umax_value, false_umax);
14165 true_reg->umin_value = max(true_reg->umin_value, true_umin);
14166 }
14167 break;
14168 }
14169 case BPF_JSGE:
14170 case BPF_JSGT:
14171 {
14172 if (is_jmp32) {
14173 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
14174 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
14175
14176 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
14177 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
14178 } else {
14179 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
14180 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
14181
14182 false_reg->smax_value = min(false_reg->smax_value, false_smax);
14183 true_reg->smin_value = max(true_reg->smin_value, true_smin);
14184 }
14185 break;
14186 }
14187 case BPF_JLE:
14188 case BPF_JLT:
14189 {
14190 if (is_jmp32) {
14191 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
14192 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
14193
14194 false_reg->u32_min_value = max(false_reg->u32_min_value,
14195 false_umin);
14196 true_reg->u32_max_value = min(true_reg->u32_max_value,
14197 true_umax);
14198 } else {
14199 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
14200 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
14201
14202 false_reg->umin_value = max(false_reg->umin_value, false_umin);
14203 true_reg->umax_value = min(true_reg->umax_value, true_umax);
14204 }
14205 break;
14206 }
14207 case BPF_JSLE:
14208 case BPF_JSLT:
14209 {
14210 if (is_jmp32) {
14211 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
14212 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
14213
14214 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
14215 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
14216 } else {
14217 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
14218 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
14219
14220 false_reg->smin_value = max(false_reg->smin_value, false_smin);
14221 true_reg->smax_value = min(true_reg->smax_value, true_smax);
14222 }
14223 break;
14224 }
14225 default:
14226 return;
14227 }
14228
14229 if (is_jmp32) {
14230 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
14231 tnum_subreg(false_32off));
14232 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
14233 tnum_subreg(true_32off));
14234 __reg_combine_32_into_64(false_reg);
14235 __reg_combine_32_into_64(true_reg);
14236 } else {
14237 false_reg->var_off = false_64off;
14238 true_reg->var_off = true_64off;
14239 __reg_combine_64_into_32(false_reg);
14240 __reg_combine_64_into_32(true_reg);
14241 }
14242 }
14243
14244 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
14245 * the variable reg.
14246 */
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)14247 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
14248 struct bpf_reg_state *false_reg,
14249 u64 val, u32 val32,
14250 u8 opcode, bool is_jmp32)
14251 {
14252 opcode = flip_opcode(opcode);
14253 /* This uses zero as "not present in table"; luckily the zero opcode,
14254 * BPF_JA, can't get here.
14255 */
14256 if (opcode)
14257 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
14258 }
14259
14260 /* 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)14261 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
14262 struct bpf_reg_state *dst_reg)
14263 {
14264 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
14265 dst_reg->umin_value);
14266 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
14267 dst_reg->umax_value);
14268 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
14269 dst_reg->smin_value);
14270 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
14271 dst_reg->smax_value);
14272 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
14273 dst_reg->var_off);
14274 reg_bounds_sync(src_reg);
14275 reg_bounds_sync(dst_reg);
14276 }
14277
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)14278 static void reg_combine_min_max(struct bpf_reg_state *true_src,
14279 struct bpf_reg_state *true_dst,
14280 struct bpf_reg_state *false_src,
14281 struct bpf_reg_state *false_dst,
14282 u8 opcode)
14283 {
14284 switch (opcode) {
14285 case BPF_JEQ:
14286 __reg_combine_min_max(true_src, true_dst);
14287 break;
14288 case BPF_JNE:
14289 __reg_combine_min_max(false_src, false_dst);
14290 break;
14291 }
14292 }
14293
mark_ptr_or_null_reg(struct bpf_func_state * state,struct bpf_reg_state * reg,u32 id,bool is_null)14294 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
14295 struct bpf_reg_state *reg, u32 id,
14296 bool is_null)
14297 {
14298 if (type_may_be_null(reg->type) && reg->id == id &&
14299 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
14300 /* Old offset (both fixed and variable parts) should have been
14301 * known-zero, because we don't allow pointer arithmetic on
14302 * pointers that might be NULL. If we see this happening, don't
14303 * convert the register.
14304 *
14305 * But in some cases, some helpers that return local kptrs
14306 * advance offset for the returned pointer. In those cases, it
14307 * is fine to expect to see reg->off.
14308 */
14309 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
14310 return;
14311 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
14312 WARN_ON_ONCE(reg->off))
14313 return;
14314
14315 if (is_null) {
14316 reg->type = SCALAR_VALUE;
14317 /* We don't need id and ref_obj_id from this point
14318 * onwards anymore, thus we should better reset it,
14319 * so that state pruning has chances to take effect.
14320 */
14321 reg->id = 0;
14322 reg->ref_obj_id = 0;
14323
14324 return;
14325 }
14326
14327 mark_ptr_not_null_reg(reg);
14328
14329 if (!reg_may_point_to_spin_lock(reg)) {
14330 /* For not-NULL ptr, reg->ref_obj_id will be reset
14331 * in release_reference().
14332 *
14333 * reg->id is still used by spin_lock ptr. Other
14334 * than spin_lock ptr type, reg->id can be reset.
14335 */
14336 reg->id = 0;
14337 }
14338 }
14339 }
14340
14341 /* The logic is similar to find_good_pkt_pointers(), both could eventually
14342 * be folded together at some point.
14343 */
mark_ptr_or_null_regs(struct bpf_verifier_state * vstate,u32 regno,bool is_null)14344 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
14345 bool is_null)
14346 {
14347 struct bpf_func_state *state = vstate->frame[vstate->curframe];
14348 struct bpf_reg_state *regs = state->regs, *reg;
14349 u32 ref_obj_id = regs[regno].ref_obj_id;
14350 u32 id = regs[regno].id;
14351
14352 if (ref_obj_id && ref_obj_id == id && is_null)
14353 /* regs[regno] is in the " == NULL" branch.
14354 * No one could have freed the reference state before
14355 * doing the NULL check.
14356 */
14357 WARN_ON_ONCE(release_reference_state(state, id));
14358
14359 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14360 mark_ptr_or_null_reg(state, reg, id, is_null);
14361 }));
14362 }
14363
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)14364 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
14365 struct bpf_reg_state *dst_reg,
14366 struct bpf_reg_state *src_reg,
14367 struct bpf_verifier_state *this_branch,
14368 struct bpf_verifier_state *other_branch)
14369 {
14370 if (BPF_SRC(insn->code) != BPF_X)
14371 return false;
14372
14373 /* Pointers are always 64-bit. */
14374 if (BPF_CLASS(insn->code) == BPF_JMP32)
14375 return false;
14376
14377 switch (BPF_OP(insn->code)) {
14378 case BPF_JGT:
14379 if ((dst_reg->type == PTR_TO_PACKET &&
14380 src_reg->type == PTR_TO_PACKET_END) ||
14381 (dst_reg->type == PTR_TO_PACKET_META &&
14382 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14383 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
14384 find_good_pkt_pointers(this_branch, dst_reg,
14385 dst_reg->type, false);
14386 mark_pkt_end(other_branch, insn->dst_reg, true);
14387 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14388 src_reg->type == PTR_TO_PACKET) ||
14389 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14390 src_reg->type == PTR_TO_PACKET_META)) {
14391 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
14392 find_good_pkt_pointers(other_branch, src_reg,
14393 src_reg->type, true);
14394 mark_pkt_end(this_branch, insn->src_reg, false);
14395 } else {
14396 return false;
14397 }
14398 break;
14399 case BPF_JLT:
14400 if ((dst_reg->type == PTR_TO_PACKET &&
14401 src_reg->type == PTR_TO_PACKET_END) ||
14402 (dst_reg->type == PTR_TO_PACKET_META &&
14403 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14404 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
14405 find_good_pkt_pointers(other_branch, dst_reg,
14406 dst_reg->type, true);
14407 mark_pkt_end(this_branch, insn->dst_reg, false);
14408 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14409 src_reg->type == PTR_TO_PACKET) ||
14410 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14411 src_reg->type == PTR_TO_PACKET_META)) {
14412 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
14413 find_good_pkt_pointers(this_branch, src_reg,
14414 src_reg->type, false);
14415 mark_pkt_end(other_branch, insn->src_reg, true);
14416 } else {
14417 return false;
14418 }
14419 break;
14420 case BPF_JGE:
14421 if ((dst_reg->type == PTR_TO_PACKET &&
14422 src_reg->type == PTR_TO_PACKET_END) ||
14423 (dst_reg->type == PTR_TO_PACKET_META &&
14424 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14425 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
14426 find_good_pkt_pointers(this_branch, dst_reg,
14427 dst_reg->type, true);
14428 mark_pkt_end(other_branch, insn->dst_reg, false);
14429 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14430 src_reg->type == PTR_TO_PACKET) ||
14431 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14432 src_reg->type == PTR_TO_PACKET_META)) {
14433 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
14434 find_good_pkt_pointers(other_branch, src_reg,
14435 src_reg->type, false);
14436 mark_pkt_end(this_branch, insn->src_reg, true);
14437 } else {
14438 return false;
14439 }
14440 break;
14441 case BPF_JLE:
14442 if ((dst_reg->type == PTR_TO_PACKET &&
14443 src_reg->type == PTR_TO_PACKET_END) ||
14444 (dst_reg->type == PTR_TO_PACKET_META &&
14445 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
14446 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
14447 find_good_pkt_pointers(other_branch, dst_reg,
14448 dst_reg->type, false);
14449 mark_pkt_end(this_branch, insn->dst_reg, true);
14450 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
14451 src_reg->type == PTR_TO_PACKET) ||
14452 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
14453 src_reg->type == PTR_TO_PACKET_META)) {
14454 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
14455 find_good_pkt_pointers(this_branch, src_reg,
14456 src_reg->type, true);
14457 mark_pkt_end(other_branch, insn->src_reg, false);
14458 } else {
14459 return false;
14460 }
14461 break;
14462 default:
14463 return false;
14464 }
14465
14466 return true;
14467 }
14468
find_equal_scalars(struct bpf_verifier_state * vstate,struct bpf_reg_state * known_reg)14469 static void find_equal_scalars(struct bpf_verifier_state *vstate,
14470 struct bpf_reg_state *known_reg)
14471 {
14472 struct bpf_func_state *state;
14473 struct bpf_reg_state *reg;
14474
14475 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
14476 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
14477 copy_register_state(reg, known_reg);
14478 }));
14479 }
14480
check_cond_jmp_op(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)14481 static int check_cond_jmp_op(struct bpf_verifier_env *env,
14482 struct bpf_insn *insn, int *insn_idx)
14483 {
14484 struct bpf_verifier_state *this_branch = env->cur_state;
14485 struct bpf_verifier_state *other_branch;
14486 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
14487 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
14488 struct bpf_reg_state *eq_branch_regs;
14489 u8 opcode = BPF_OP(insn->code);
14490 bool is_jmp32;
14491 int pred = -1;
14492 int err;
14493
14494 /* Only conditional jumps are expected to reach here. */
14495 if (opcode == BPF_JA || opcode > BPF_JSLE) {
14496 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
14497 return -EINVAL;
14498 }
14499
14500 /* check src2 operand */
14501 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
14502 if (err)
14503 return err;
14504
14505 dst_reg = ®s[insn->dst_reg];
14506 if (BPF_SRC(insn->code) == BPF_X) {
14507 if (insn->imm != 0) {
14508 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14509 return -EINVAL;
14510 }
14511
14512 /* check src1 operand */
14513 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14514 if (err)
14515 return err;
14516
14517 src_reg = ®s[insn->src_reg];
14518 if (!(reg_is_pkt_pointer_any(dst_reg) && reg_is_pkt_pointer_any(src_reg)) &&
14519 is_pointer_value(env, insn->src_reg)) {
14520 verbose(env, "R%d pointer comparison prohibited\n",
14521 insn->src_reg);
14522 return -EACCES;
14523 }
14524 } else {
14525 if (insn->src_reg != BPF_REG_0) {
14526 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
14527 return -EINVAL;
14528 }
14529 }
14530
14531 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
14532
14533 if (BPF_SRC(insn->code) == BPF_K) {
14534 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
14535 } else if (src_reg->type == SCALAR_VALUE &&
14536 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
14537 pred = is_branch_taken(dst_reg,
14538 tnum_subreg(src_reg->var_off).value,
14539 opcode,
14540 is_jmp32);
14541 } else if (src_reg->type == SCALAR_VALUE &&
14542 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
14543 pred = is_branch_taken(dst_reg,
14544 src_reg->var_off.value,
14545 opcode,
14546 is_jmp32);
14547 } else if (dst_reg->type == SCALAR_VALUE &&
14548 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) {
14549 pred = is_branch_taken(src_reg,
14550 tnum_subreg(dst_reg->var_off).value,
14551 flip_opcode(opcode),
14552 is_jmp32);
14553 } else if (dst_reg->type == SCALAR_VALUE &&
14554 !is_jmp32 && tnum_is_const(dst_reg->var_off)) {
14555 pred = is_branch_taken(src_reg,
14556 dst_reg->var_off.value,
14557 flip_opcode(opcode),
14558 is_jmp32);
14559 } else if (reg_is_pkt_pointer_any(dst_reg) &&
14560 reg_is_pkt_pointer_any(src_reg) &&
14561 !is_jmp32) {
14562 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
14563 }
14564
14565 if (pred >= 0) {
14566 /* If we get here with a dst_reg pointer type it is because
14567 * above is_branch_taken() special cased the 0 comparison.
14568 */
14569 if (!__is_pointer_value(false, dst_reg))
14570 err = mark_chain_precision(env, insn->dst_reg);
14571 if (BPF_SRC(insn->code) == BPF_X && !err &&
14572 !__is_pointer_value(false, src_reg))
14573 err = mark_chain_precision(env, insn->src_reg);
14574 if (err)
14575 return err;
14576 }
14577
14578 if (pred == 1) {
14579 /* Only follow the goto, ignore fall-through. If needed, push
14580 * the fall-through branch for simulation under speculative
14581 * execution.
14582 */
14583 if (!env->bypass_spec_v1 &&
14584 !sanitize_speculative_path(env, insn, *insn_idx + 1,
14585 *insn_idx))
14586 return -EFAULT;
14587 if (env->log.level & BPF_LOG_LEVEL)
14588 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14589 *insn_idx += insn->off;
14590 return 0;
14591 } else if (pred == 0) {
14592 /* Only follow the fall-through branch, since that's where the
14593 * program will go. If needed, push the goto branch for
14594 * simulation under speculative execution.
14595 */
14596 if (!env->bypass_spec_v1 &&
14597 !sanitize_speculative_path(env, insn,
14598 *insn_idx + insn->off + 1,
14599 *insn_idx))
14600 return -EFAULT;
14601 if (env->log.level & BPF_LOG_LEVEL)
14602 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14603 return 0;
14604 }
14605
14606 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
14607 false);
14608 if (!other_branch)
14609 return -EFAULT;
14610 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
14611
14612 /* detect if we are comparing against a constant value so we can adjust
14613 * our min/max values for our dst register.
14614 * this is only legit if both are scalars (or pointers to the same
14615 * object, I suppose, see the PTR_MAYBE_NULL related if block below),
14616 * because otherwise the different base pointers mean the offsets aren't
14617 * comparable.
14618 */
14619 if (BPF_SRC(insn->code) == BPF_X) {
14620 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
14621
14622 if (dst_reg->type == SCALAR_VALUE &&
14623 src_reg->type == SCALAR_VALUE) {
14624 if (tnum_is_const(src_reg->var_off) ||
14625 (is_jmp32 &&
14626 tnum_is_const(tnum_subreg(src_reg->var_off))))
14627 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14628 dst_reg,
14629 src_reg->var_off.value,
14630 tnum_subreg(src_reg->var_off).value,
14631 opcode, is_jmp32);
14632 else if (tnum_is_const(dst_reg->var_off) ||
14633 (is_jmp32 &&
14634 tnum_is_const(tnum_subreg(dst_reg->var_off))))
14635 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
14636 src_reg,
14637 dst_reg->var_off.value,
14638 tnum_subreg(dst_reg->var_off).value,
14639 opcode, is_jmp32);
14640 else if (!is_jmp32 &&
14641 (opcode == BPF_JEQ || opcode == BPF_JNE))
14642 /* Comparing for equality, we can combine knowledge */
14643 reg_combine_min_max(&other_branch_regs[insn->src_reg],
14644 &other_branch_regs[insn->dst_reg],
14645 src_reg, dst_reg, opcode);
14646 if (src_reg->id &&
14647 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
14648 find_equal_scalars(this_branch, src_reg);
14649 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
14650 }
14651
14652 }
14653 } else if (dst_reg->type == SCALAR_VALUE) {
14654 reg_set_min_max(&other_branch_regs[insn->dst_reg],
14655 dst_reg, insn->imm, (u32)insn->imm,
14656 opcode, is_jmp32);
14657 }
14658
14659 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
14660 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
14661 find_equal_scalars(this_branch, dst_reg);
14662 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
14663 }
14664
14665 /* if one pointer register is compared to another pointer
14666 * register check if PTR_MAYBE_NULL could be lifted.
14667 * E.g. register A - maybe null
14668 * register B - not null
14669 * for JNE A, B, ... - A is not null in the false branch;
14670 * for JEQ A, B, ... - A is not null in the true branch.
14671 *
14672 * Since PTR_TO_BTF_ID points to a kernel struct that does
14673 * not need to be null checked by the BPF program, i.e.,
14674 * could be null even without PTR_MAYBE_NULL marking, so
14675 * only propagate nullness when neither reg is that type.
14676 */
14677 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
14678 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
14679 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
14680 base_type(src_reg->type) != PTR_TO_BTF_ID &&
14681 base_type(dst_reg->type) != PTR_TO_BTF_ID) {
14682 eq_branch_regs = NULL;
14683 switch (opcode) {
14684 case BPF_JEQ:
14685 eq_branch_regs = other_branch_regs;
14686 break;
14687 case BPF_JNE:
14688 eq_branch_regs = regs;
14689 break;
14690 default:
14691 /* do nothing */
14692 break;
14693 }
14694 if (eq_branch_regs) {
14695 if (type_may_be_null(src_reg->type))
14696 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
14697 else
14698 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
14699 }
14700 }
14701
14702 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
14703 * NOTE: these optimizations below are related with pointer comparison
14704 * which will never be JMP32.
14705 */
14706 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
14707 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
14708 type_may_be_null(dst_reg->type)) {
14709 /* Mark all identical registers in each branch as either
14710 * safe or unknown depending R == 0 or R != 0 conditional.
14711 */
14712 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
14713 opcode == BPF_JNE);
14714 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
14715 opcode == BPF_JEQ);
14716 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
14717 this_branch, other_branch) &&
14718 is_pointer_value(env, insn->dst_reg)) {
14719 verbose(env, "R%d pointer comparison prohibited\n",
14720 insn->dst_reg);
14721 return -EACCES;
14722 }
14723 if (env->log.level & BPF_LOG_LEVEL)
14724 print_insn_state(env, this_branch->frame[this_branch->curframe]);
14725 return 0;
14726 }
14727
14728 /* verify BPF_LD_IMM64 instruction */
check_ld_imm(struct bpf_verifier_env * env,struct bpf_insn * insn)14729 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
14730 {
14731 struct bpf_insn_aux_data *aux = cur_aux(env);
14732 struct bpf_reg_state *regs = cur_regs(env);
14733 struct bpf_reg_state *dst_reg;
14734 struct bpf_map *map;
14735 int err;
14736
14737 if (BPF_SIZE(insn->code) != BPF_DW) {
14738 verbose(env, "invalid BPF_LD_IMM insn\n");
14739 return -EINVAL;
14740 }
14741 if (insn->off != 0) {
14742 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
14743 return -EINVAL;
14744 }
14745
14746 err = check_reg_arg(env, insn->dst_reg, DST_OP);
14747 if (err)
14748 return err;
14749
14750 dst_reg = ®s[insn->dst_reg];
14751 if (insn->src_reg == 0) {
14752 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
14753
14754 dst_reg->type = SCALAR_VALUE;
14755 __mark_reg_known(®s[insn->dst_reg], imm);
14756 return 0;
14757 }
14758
14759 /* All special src_reg cases are listed below. From this point onwards
14760 * we either succeed and assign a corresponding dst_reg->type after
14761 * zeroing the offset, or fail and reject the program.
14762 */
14763 mark_reg_known_zero(env, regs, insn->dst_reg);
14764
14765 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
14766 dst_reg->type = aux->btf_var.reg_type;
14767 switch (base_type(dst_reg->type)) {
14768 case PTR_TO_MEM:
14769 dst_reg->mem_size = aux->btf_var.mem_size;
14770 break;
14771 case PTR_TO_BTF_ID:
14772 dst_reg->btf = aux->btf_var.btf;
14773 dst_reg->btf_id = aux->btf_var.btf_id;
14774 break;
14775 default:
14776 verbose(env, "bpf verifier is misconfigured\n");
14777 return -EFAULT;
14778 }
14779 return 0;
14780 }
14781
14782 if (insn->src_reg == BPF_PSEUDO_FUNC) {
14783 struct bpf_prog_aux *aux = env->prog->aux;
14784 u32 subprogno = find_subprog(env,
14785 env->insn_idx + insn->imm + 1);
14786
14787 if (!aux->func_info) {
14788 verbose(env, "missing btf func_info\n");
14789 return -EINVAL;
14790 }
14791 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
14792 verbose(env, "callback function not static\n");
14793 return -EINVAL;
14794 }
14795
14796 dst_reg->type = PTR_TO_FUNC;
14797 dst_reg->subprogno = subprogno;
14798 return 0;
14799 }
14800
14801 map = env->used_maps[aux->map_index];
14802 dst_reg->map_ptr = map;
14803
14804 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
14805 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
14806 dst_reg->type = PTR_TO_MAP_VALUE;
14807 dst_reg->off = aux->map_off;
14808 WARN_ON_ONCE(map->max_entries != 1);
14809 /* We want reg->id to be same (0) as map_value is not distinct */
14810 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
14811 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
14812 dst_reg->type = CONST_PTR_TO_MAP;
14813 } else {
14814 verbose(env, "bpf verifier is misconfigured\n");
14815 return -EINVAL;
14816 }
14817
14818 return 0;
14819 }
14820
may_access_skb(enum bpf_prog_type type)14821 static bool may_access_skb(enum bpf_prog_type type)
14822 {
14823 switch (type) {
14824 case BPF_PROG_TYPE_SOCKET_FILTER:
14825 case BPF_PROG_TYPE_SCHED_CLS:
14826 case BPF_PROG_TYPE_SCHED_ACT:
14827 return true;
14828 default:
14829 return false;
14830 }
14831 }
14832
14833 /* verify safety of LD_ABS|LD_IND instructions:
14834 * - they can only appear in the programs where ctx == skb
14835 * - since they are wrappers of function calls, they scratch R1-R5 registers,
14836 * preserve R6-R9, and store return value into R0
14837 *
14838 * Implicit input:
14839 * ctx == skb == R6 == CTX
14840 *
14841 * Explicit input:
14842 * SRC == any register
14843 * IMM == 32-bit immediate
14844 *
14845 * Output:
14846 * R0 - 8/16/32-bit skb data converted to cpu endianness
14847 */
check_ld_abs(struct bpf_verifier_env * env,struct bpf_insn * insn)14848 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
14849 {
14850 struct bpf_reg_state *regs = cur_regs(env);
14851 static const int ctx_reg = BPF_REG_6;
14852 u8 mode = BPF_MODE(insn->code);
14853 int i, err;
14854
14855 if (!may_access_skb(resolve_prog_type(env->prog))) {
14856 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
14857 return -EINVAL;
14858 }
14859
14860 if (!env->ops->gen_ld_abs) {
14861 verbose(env, "bpf verifier is misconfigured\n");
14862 return -EINVAL;
14863 }
14864
14865 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
14866 BPF_SIZE(insn->code) == BPF_DW ||
14867 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
14868 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
14869 return -EINVAL;
14870 }
14871
14872 /* check whether implicit source operand (register R6) is readable */
14873 err = check_reg_arg(env, ctx_reg, SRC_OP);
14874 if (err)
14875 return err;
14876
14877 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
14878 * gen_ld_abs() may terminate the program at runtime, leading to
14879 * reference leak.
14880 */
14881 err = check_reference_leak(env);
14882 if (err) {
14883 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
14884 return err;
14885 }
14886
14887 if (env->cur_state->active_lock.ptr) {
14888 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
14889 return -EINVAL;
14890 }
14891
14892 if (env->cur_state->active_rcu_lock) {
14893 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n");
14894 return -EINVAL;
14895 }
14896
14897 if (regs[ctx_reg].type != PTR_TO_CTX) {
14898 verbose(env,
14899 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
14900 return -EINVAL;
14901 }
14902
14903 if (mode == BPF_IND) {
14904 /* check explicit source operand */
14905 err = check_reg_arg(env, insn->src_reg, SRC_OP);
14906 if (err)
14907 return err;
14908 }
14909
14910 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
14911 if (err < 0)
14912 return err;
14913
14914 /* reset caller saved regs to unreadable */
14915 for (i = 0; i < CALLER_SAVED_REGS; i++) {
14916 mark_reg_not_init(env, regs, caller_saved[i]);
14917 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
14918 }
14919
14920 /* mark destination R0 register as readable, since it contains
14921 * the value fetched from the packet.
14922 * Already marked as written above.
14923 */
14924 mark_reg_unknown(env, regs, BPF_REG_0);
14925 /* ld_abs load up to 32-bit skb data. */
14926 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
14927 return 0;
14928 }
14929
check_return_code(struct bpf_verifier_env * env)14930 static int check_return_code(struct bpf_verifier_env *env)
14931 {
14932 struct tnum enforce_attach_type_range = tnum_unknown;
14933 const struct bpf_prog *prog = env->prog;
14934 struct bpf_reg_state *reg;
14935 struct tnum range = tnum_range(0, 1), const_0 = tnum_const(0);
14936 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
14937 int err;
14938 struct bpf_func_state *frame = env->cur_state->frame[0];
14939 const bool is_subprog = frame->subprogno;
14940
14941 /* LSM and struct_ops func-ptr's return type could be "void" */
14942 if (!is_subprog) {
14943 switch (prog_type) {
14944 case BPF_PROG_TYPE_LSM:
14945 if (prog->expected_attach_type == BPF_LSM_CGROUP)
14946 /* See below, can be 0 or 0-1 depending on hook. */
14947 break;
14948 fallthrough;
14949 case BPF_PROG_TYPE_STRUCT_OPS:
14950 if (!prog->aux->attach_func_proto->type)
14951 return 0;
14952 break;
14953 default:
14954 break;
14955 }
14956 }
14957
14958 /* eBPF calling convention is such that R0 is used
14959 * to return the value from eBPF program.
14960 * Make sure that it's readable at this time
14961 * of bpf_exit, which means that program wrote
14962 * something into it earlier
14963 */
14964 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
14965 if (err)
14966 return err;
14967
14968 if (is_pointer_value(env, BPF_REG_0)) {
14969 verbose(env, "R0 leaks addr as return value\n");
14970 return -EACCES;
14971 }
14972
14973 reg = cur_regs(env) + BPF_REG_0;
14974
14975 if (frame->in_async_callback_fn) {
14976 /* enforce return zero from async callbacks like timer */
14977 if (reg->type != SCALAR_VALUE) {
14978 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
14979 reg_type_str(env, reg->type));
14980 return -EINVAL;
14981 }
14982
14983 if (!tnum_in(const_0, reg->var_off)) {
14984 verbose_invalid_scalar(env, reg, &const_0, "async callback", "R0");
14985 return -EINVAL;
14986 }
14987 return 0;
14988 }
14989
14990 if (is_subprog) {
14991 if (reg->type != SCALAR_VALUE) {
14992 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
14993 reg_type_str(env, reg->type));
14994 return -EINVAL;
14995 }
14996 return 0;
14997 }
14998
14999 switch (prog_type) {
15000 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
15001 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
15002 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
15003 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
15004 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
15005 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
15006 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
15007 range = tnum_range(1, 1);
15008 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
15009 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
15010 range = tnum_range(0, 3);
15011 break;
15012 case BPF_PROG_TYPE_CGROUP_SKB:
15013 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
15014 range = tnum_range(0, 3);
15015 enforce_attach_type_range = tnum_range(2, 3);
15016 }
15017 break;
15018 case BPF_PROG_TYPE_CGROUP_SOCK:
15019 case BPF_PROG_TYPE_SOCK_OPS:
15020 case BPF_PROG_TYPE_CGROUP_DEVICE:
15021 case BPF_PROG_TYPE_CGROUP_SYSCTL:
15022 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
15023 break;
15024 case BPF_PROG_TYPE_RAW_TRACEPOINT:
15025 if (!env->prog->aux->attach_btf_id)
15026 return 0;
15027 range = tnum_const(0);
15028 break;
15029 case BPF_PROG_TYPE_TRACING:
15030 switch (env->prog->expected_attach_type) {
15031 case BPF_TRACE_FENTRY:
15032 case BPF_TRACE_FEXIT:
15033 range = tnum_const(0);
15034 break;
15035 case BPF_TRACE_RAW_TP:
15036 case BPF_MODIFY_RETURN:
15037 return 0;
15038 case BPF_TRACE_ITER:
15039 break;
15040 default:
15041 return -ENOTSUPP;
15042 }
15043 break;
15044 case BPF_PROG_TYPE_SK_LOOKUP:
15045 range = tnum_range(SK_DROP, SK_PASS);
15046 break;
15047
15048 case BPF_PROG_TYPE_LSM:
15049 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
15050 /* Regular BPF_PROG_TYPE_LSM programs can return
15051 * any value.
15052 */
15053 return 0;
15054 }
15055 if (!env->prog->aux->attach_func_proto->type) {
15056 /* Make sure programs that attach to void
15057 * hooks don't try to modify return value.
15058 */
15059 range = tnum_range(1, 1);
15060 }
15061 break;
15062
15063 case BPF_PROG_TYPE_NETFILTER:
15064 range = tnum_range(NF_DROP, NF_ACCEPT);
15065 break;
15066 case BPF_PROG_TYPE_EXT:
15067 /* freplace program can return anything as its return value
15068 * depends on the to-be-replaced kernel func or bpf program.
15069 */
15070 default:
15071 return 0;
15072 }
15073
15074 if (reg->type != SCALAR_VALUE) {
15075 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
15076 reg_type_str(env, reg->type));
15077 return -EINVAL;
15078 }
15079
15080 if (!tnum_in(range, reg->var_off)) {
15081 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
15082 if (prog->expected_attach_type == BPF_LSM_CGROUP &&
15083 prog_type == BPF_PROG_TYPE_LSM &&
15084 !prog->aux->attach_func_proto->type)
15085 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
15086 return -EINVAL;
15087 }
15088
15089 if (!tnum_is_unknown(enforce_attach_type_range) &&
15090 tnum_in(enforce_attach_type_range, reg->var_off))
15091 env->prog->enforce_expected_attach_type = 1;
15092 return 0;
15093 }
15094
15095 /* non-recursive DFS pseudo code
15096 * 1 procedure DFS-iterative(G,v):
15097 * 2 label v as discovered
15098 * 3 let S be a stack
15099 * 4 S.push(v)
15100 * 5 while S is not empty
15101 * 6 t <- S.peek()
15102 * 7 if t is what we're looking for:
15103 * 8 return t
15104 * 9 for all edges e in G.adjacentEdges(t) do
15105 * 10 if edge e is already labelled
15106 * 11 continue with the next edge
15107 * 12 w <- G.adjacentVertex(t,e)
15108 * 13 if vertex w is not discovered and not explored
15109 * 14 label e as tree-edge
15110 * 15 label w as discovered
15111 * 16 S.push(w)
15112 * 17 continue at 5
15113 * 18 else if vertex w is discovered
15114 * 19 label e as back-edge
15115 * 20 else
15116 * 21 // vertex w is explored
15117 * 22 label e as forward- or cross-edge
15118 * 23 label t as explored
15119 * 24 S.pop()
15120 *
15121 * convention:
15122 * 0x10 - discovered
15123 * 0x11 - discovered and fall-through edge labelled
15124 * 0x12 - discovered and fall-through and branch edges labelled
15125 * 0x20 - explored
15126 */
15127
15128 enum {
15129 DISCOVERED = 0x10,
15130 EXPLORED = 0x20,
15131 FALLTHROUGH = 1,
15132 BRANCH = 2,
15133 };
15134
mark_prune_point(struct bpf_verifier_env * env,int idx)15135 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
15136 {
15137 env->insn_aux_data[idx].prune_point = true;
15138 }
15139
is_prune_point(struct bpf_verifier_env * env,int insn_idx)15140 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
15141 {
15142 return env->insn_aux_data[insn_idx].prune_point;
15143 }
15144
mark_force_checkpoint(struct bpf_verifier_env * env,int idx)15145 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
15146 {
15147 env->insn_aux_data[idx].force_checkpoint = true;
15148 }
15149
is_force_checkpoint(struct bpf_verifier_env * env,int insn_idx)15150 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
15151 {
15152 return env->insn_aux_data[insn_idx].force_checkpoint;
15153 }
15154
mark_calls_callback(struct bpf_verifier_env * env,int idx)15155 static void mark_calls_callback(struct bpf_verifier_env *env, int idx)
15156 {
15157 env->insn_aux_data[idx].calls_callback = true;
15158 }
15159
calls_callback(struct bpf_verifier_env * env,int insn_idx)15160 static bool calls_callback(struct bpf_verifier_env *env, int insn_idx)
15161 {
15162 return env->insn_aux_data[insn_idx].calls_callback;
15163 }
15164
15165 enum {
15166 DONE_EXPLORING = 0,
15167 KEEP_EXPLORING = 1,
15168 };
15169
15170 /* t, w, e - match pseudo-code above:
15171 * t - index of current instruction
15172 * w - next instruction
15173 * e - edge
15174 */
push_insn(int t,int w,int e,struct bpf_verifier_env * env)15175 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
15176 {
15177 int *insn_stack = env->cfg.insn_stack;
15178 int *insn_state = env->cfg.insn_state;
15179
15180 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
15181 return DONE_EXPLORING;
15182
15183 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
15184 return DONE_EXPLORING;
15185
15186 if (w < 0 || w >= env->prog->len) {
15187 verbose_linfo(env, t, "%d: ", t);
15188 verbose(env, "jump out of range from insn %d to %d\n", t, w);
15189 return -EINVAL;
15190 }
15191
15192 if (e == BRANCH) {
15193 /* mark branch target for state pruning */
15194 mark_prune_point(env, w);
15195 mark_jmp_point(env, w);
15196 }
15197
15198 if (insn_state[w] == 0) {
15199 /* tree-edge */
15200 insn_state[t] = DISCOVERED | e;
15201 insn_state[w] = DISCOVERED;
15202 if (env->cfg.cur_stack >= env->prog->len)
15203 return -E2BIG;
15204 insn_stack[env->cfg.cur_stack++] = w;
15205 return KEEP_EXPLORING;
15206 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
15207 if (env->bpf_capable)
15208 return DONE_EXPLORING;
15209 verbose_linfo(env, t, "%d: ", t);
15210 verbose_linfo(env, w, "%d: ", w);
15211 verbose(env, "back-edge from insn %d to %d\n", t, w);
15212 return -EINVAL;
15213 } else if (insn_state[w] == EXPLORED) {
15214 /* forward- or cross-edge */
15215 insn_state[t] = DISCOVERED | e;
15216 } else {
15217 verbose(env, "insn state internal bug\n");
15218 return -EFAULT;
15219 }
15220 return DONE_EXPLORING;
15221 }
15222
visit_func_call_insn(int t,struct bpf_insn * insns,struct bpf_verifier_env * env,bool visit_callee)15223 static int visit_func_call_insn(int t, struct bpf_insn *insns,
15224 struct bpf_verifier_env *env,
15225 bool visit_callee)
15226 {
15227 int ret, insn_sz;
15228
15229 insn_sz = bpf_is_ldimm64(&insns[t]) ? 2 : 1;
15230 ret = push_insn(t, t + insn_sz, FALLTHROUGH, env);
15231 if (ret)
15232 return ret;
15233
15234 mark_prune_point(env, t + insn_sz);
15235 /* when we exit from subprog, we need to record non-linear history */
15236 mark_jmp_point(env, t + insn_sz);
15237
15238 if (visit_callee) {
15239 mark_prune_point(env, t);
15240 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
15241 }
15242 return ret;
15243 }
15244
15245 /* Visits the instruction at index t and returns one of the following:
15246 * < 0 - an error occurred
15247 * DONE_EXPLORING - the instruction was fully explored
15248 * KEEP_EXPLORING - there is still work to be done before it is fully explored
15249 */
visit_insn(int t,struct bpf_verifier_env * env)15250 static int visit_insn(int t, struct bpf_verifier_env *env)
15251 {
15252 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
15253 int ret, off, insn_sz;
15254
15255 if (bpf_pseudo_func(insn))
15256 return visit_func_call_insn(t, insns, env, true);
15257
15258 /* All non-branch instructions have a single fall-through edge. */
15259 if (BPF_CLASS(insn->code) != BPF_JMP &&
15260 BPF_CLASS(insn->code) != BPF_JMP32) {
15261 insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
15262 return push_insn(t, t + insn_sz, FALLTHROUGH, env);
15263 }
15264
15265 switch (BPF_OP(insn->code)) {
15266 case BPF_EXIT:
15267 return DONE_EXPLORING;
15268
15269 case BPF_CALL:
15270 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback)
15271 /* Mark this call insn as a prune point to trigger
15272 * is_state_visited() check before call itself is
15273 * processed by __check_func_call(). Otherwise new
15274 * async state will be pushed for further exploration.
15275 */
15276 mark_prune_point(env, t);
15277 /* For functions that invoke callbacks it is not known how many times
15278 * callback would be called. Verifier models callback calling functions
15279 * by repeatedly visiting callback bodies and returning to origin call
15280 * instruction.
15281 * In order to stop such iteration verifier needs to identify when a
15282 * state identical some state from a previous iteration is reached.
15283 * Check below forces creation of checkpoint before callback calling
15284 * instruction to allow search for such identical states.
15285 */
15286 if (is_sync_callback_calling_insn(insn)) {
15287 mark_calls_callback(env, t);
15288 mark_force_checkpoint(env, t);
15289 mark_prune_point(env, t);
15290 mark_jmp_point(env, t);
15291 }
15292 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
15293 struct bpf_kfunc_call_arg_meta meta;
15294
15295 ret = fetch_kfunc_meta(env, insn, &meta, NULL);
15296 if (ret == 0 && is_iter_next_kfunc(&meta)) {
15297 mark_prune_point(env, t);
15298 /* Checking and saving state checkpoints at iter_next() call
15299 * is crucial for fast convergence of open-coded iterator loop
15300 * logic, so we need to force it. If we don't do that,
15301 * is_state_visited() might skip saving a checkpoint, causing
15302 * unnecessarily long sequence of not checkpointed
15303 * instructions and jumps, leading to exhaustion of jump
15304 * history buffer, and potentially other undesired outcomes.
15305 * It is expected that with correct open-coded iterators
15306 * convergence will happen quickly, so we don't run a risk of
15307 * exhausting memory.
15308 */
15309 mark_force_checkpoint(env, t);
15310 }
15311 }
15312 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
15313
15314 case BPF_JA:
15315 if (BPF_SRC(insn->code) != BPF_K)
15316 return -EINVAL;
15317
15318 if (BPF_CLASS(insn->code) == BPF_JMP)
15319 off = insn->off;
15320 else
15321 off = insn->imm;
15322
15323 /* unconditional jump with single edge */
15324 ret = push_insn(t, t + off + 1, FALLTHROUGH, env);
15325 if (ret)
15326 return ret;
15327
15328 mark_prune_point(env, t + off + 1);
15329 mark_jmp_point(env, t + off + 1);
15330
15331 return ret;
15332
15333 default:
15334 /* conditional jump with two edges */
15335 mark_prune_point(env, t);
15336
15337 ret = push_insn(t, t + 1, FALLTHROUGH, env);
15338 if (ret)
15339 return ret;
15340
15341 return push_insn(t, t + insn->off + 1, BRANCH, env);
15342 }
15343 }
15344
15345 /* non-recursive depth-first-search to detect loops in BPF program
15346 * loop == back-edge in directed graph
15347 */
check_cfg(struct bpf_verifier_env * env)15348 static int check_cfg(struct bpf_verifier_env *env)
15349 {
15350 int insn_cnt = env->prog->len;
15351 int *insn_stack, *insn_state;
15352 int ret = 0;
15353 int i;
15354
15355 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
15356 if (!insn_state)
15357 return -ENOMEM;
15358
15359 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
15360 if (!insn_stack) {
15361 kvfree(insn_state);
15362 return -ENOMEM;
15363 }
15364
15365 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
15366 insn_stack[0] = 0; /* 0 is the first instruction */
15367 env->cfg.cur_stack = 1;
15368
15369 while (env->cfg.cur_stack > 0) {
15370 int t = insn_stack[env->cfg.cur_stack - 1];
15371
15372 ret = visit_insn(t, env);
15373 switch (ret) {
15374 case DONE_EXPLORING:
15375 insn_state[t] = EXPLORED;
15376 env->cfg.cur_stack--;
15377 break;
15378 case KEEP_EXPLORING:
15379 break;
15380 default:
15381 if (ret > 0) {
15382 verbose(env, "visit_insn internal bug\n");
15383 ret = -EFAULT;
15384 }
15385 goto err_free;
15386 }
15387 }
15388
15389 if (env->cfg.cur_stack < 0) {
15390 verbose(env, "pop stack internal bug\n");
15391 ret = -EFAULT;
15392 goto err_free;
15393 }
15394
15395 for (i = 0; i < insn_cnt; i++) {
15396 struct bpf_insn *insn = &env->prog->insnsi[i];
15397
15398 if (insn_state[i] != EXPLORED) {
15399 verbose(env, "unreachable insn %d\n", i);
15400 ret = -EINVAL;
15401 goto err_free;
15402 }
15403 if (bpf_is_ldimm64(insn)) {
15404 if (insn_state[i + 1] != 0) {
15405 verbose(env, "jump into the middle of ldimm64 insn %d\n", i);
15406 ret = -EINVAL;
15407 goto err_free;
15408 }
15409 i++; /* skip second half of ldimm64 */
15410 }
15411 }
15412 ret = 0; /* cfg looks good */
15413
15414 err_free:
15415 kvfree(insn_state);
15416 kvfree(insn_stack);
15417 env->cfg.insn_state = env->cfg.insn_stack = NULL;
15418 return ret;
15419 }
15420
check_abnormal_return(struct bpf_verifier_env * env)15421 static int check_abnormal_return(struct bpf_verifier_env *env)
15422 {
15423 int i;
15424
15425 for (i = 1; i < env->subprog_cnt; i++) {
15426 if (env->subprog_info[i].has_ld_abs) {
15427 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
15428 return -EINVAL;
15429 }
15430 if (env->subprog_info[i].has_tail_call) {
15431 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
15432 return -EINVAL;
15433 }
15434 }
15435 return 0;
15436 }
15437
15438 /* The minimum supported BTF func info size */
15439 #define MIN_BPF_FUNCINFO_SIZE 8
15440 #define MAX_FUNCINFO_REC_SIZE 252
15441
check_btf_func(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15442 static int check_btf_func(struct bpf_verifier_env *env,
15443 const union bpf_attr *attr,
15444 bpfptr_t uattr)
15445 {
15446 const struct btf_type *type, *func_proto, *ret_type;
15447 u32 i, nfuncs, urec_size, min_size;
15448 u32 krec_size = sizeof(struct bpf_func_info);
15449 struct bpf_func_info *krecord;
15450 struct bpf_func_info_aux *info_aux = NULL;
15451 struct bpf_prog *prog;
15452 const struct btf *btf;
15453 bpfptr_t urecord;
15454 u32 prev_offset = 0;
15455 bool scalar_return;
15456 int ret = -ENOMEM;
15457
15458 nfuncs = attr->func_info_cnt;
15459 if (!nfuncs) {
15460 if (check_abnormal_return(env))
15461 return -EINVAL;
15462 return 0;
15463 }
15464
15465 if (nfuncs != env->subprog_cnt) {
15466 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
15467 return -EINVAL;
15468 }
15469
15470 urec_size = attr->func_info_rec_size;
15471 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
15472 urec_size > MAX_FUNCINFO_REC_SIZE ||
15473 urec_size % sizeof(u32)) {
15474 verbose(env, "invalid func info rec size %u\n", urec_size);
15475 return -EINVAL;
15476 }
15477
15478 prog = env->prog;
15479 btf = prog->aux->btf;
15480
15481 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
15482 min_size = min_t(u32, krec_size, urec_size);
15483
15484 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
15485 if (!krecord)
15486 return -ENOMEM;
15487 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
15488 if (!info_aux)
15489 goto err_free;
15490
15491 for (i = 0; i < nfuncs; i++) {
15492 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
15493 if (ret) {
15494 if (ret == -E2BIG) {
15495 verbose(env, "nonzero tailing record in func info");
15496 /* set the size kernel expects so loader can zero
15497 * out the rest of the record.
15498 */
15499 if (copy_to_bpfptr_offset(uattr,
15500 offsetof(union bpf_attr, func_info_rec_size),
15501 &min_size, sizeof(min_size)))
15502 ret = -EFAULT;
15503 }
15504 goto err_free;
15505 }
15506
15507 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
15508 ret = -EFAULT;
15509 goto err_free;
15510 }
15511
15512 /* check insn_off */
15513 ret = -EINVAL;
15514 if (i == 0) {
15515 if (krecord[i].insn_off) {
15516 verbose(env,
15517 "nonzero insn_off %u for the first func info record",
15518 krecord[i].insn_off);
15519 goto err_free;
15520 }
15521 } else if (krecord[i].insn_off <= prev_offset) {
15522 verbose(env,
15523 "same or smaller insn offset (%u) than previous func info record (%u)",
15524 krecord[i].insn_off, prev_offset);
15525 goto err_free;
15526 }
15527
15528 if (env->subprog_info[i].start != krecord[i].insn_off) {
15529 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
15530 goto err_free;
15531 }
15532
15533 /* check type_id */
15534 type = btf_type_by_id(btf, krecord[i].type_id);
15535 if (!type || !btf_type_is_func(type)) {
15536 verbose(env, "invalid type id %d in func info",
15537 krecord[i].type_id);
15538 goto err_free;
15539 }
15540 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
15541
15542 func_proto = btf_type_by_id(btf, type->type);
15543 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
15544 /* btf_func_check() already verified it during BTF load */
15545 goto err_free;
15546 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
15547 scalar_return =
15548 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
15549 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
15550 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
15551 goto err_free;
15552 }
15553 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
15554 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
15555 goto err_free;
15556 }
15557
15558 prev_offset = krecord[i].insn_off;
15559 bpfptr_add(&urecord, urec_size);
15560 }
15561
15562 prog->aux->func_info = krecord;
15563 prog->aux->func_info_cnt = nfuncs;
15564 prog->aux->func_info_aux = info_aux;
15565 return 0;
15566
15567 err_free:
15568 kvfree(krecord);
15569 kfree(info_aux);
15570 return ret;
15571 }
15572
adjust_btf_func(struct bpf_verifier_env * env)15573 static void adjust_btf_func(struct bpf_verifier_env *env)
15574 {
15575 struct bpf_prog_aux *aux = env->prog->aux;
15576 int i;
15577
15578 if (!aux->func_info)
15579 return;
15580
15581 for (i = 0; i < env->subprog_cnt; i++)
15582 aux->func_info[i].insn_off = env->subprog_info[i].start;
15583 }
15584
15585 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
15586 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
15587
check_btf_line(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15588 static int check_btf_line(struct bpf_verifier_env *env,
15589 const union bpf_attr *attr,
15590 bpfptr_t uattr)
15591 {
15592 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
15593 struct bpf_subprog_info *sub;
15594 struct bpf_line_info *linfo;
15595 struct bpf_prog *prog;
15596 const struct btf *btf;
15597 bpfptr_t ulinfo;
15598 int err;
15599
15600 nr_linfo = attr->line_info_cnt;
15601 if (!nr_linfo)
15602 return 0;
15603 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
15604 return -EINVAL;
15605
15606 rec_size = attr->line_info_rec_size;
15607 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
15608 rec_size > MAX_LINEINFO_REC_SIZE ||
15609 rec_size & (sizeof(u32) - 1))
15610 return -EINVAL;
15611
15612 /* Need to zero it in case the userspace may
15613 * pass in a smaller bpf_line_info object.
15614 */
15615 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
15616 GFP_KERNEL | __GFP_NOWARN);
15617 if (!linfo)
15618 return -ENOMEM;
15619
15620 prog = env->prog;
15621 btf = prog->aux->btf;
15622
15623 s = 0;
15624 sub = env->subprog_info;
15625 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
15626 expected_size = sizeof(struct bpf_line_info);
15627 ncopy = min_t(u32, expected_size, rec_size);
15628 for (i = 0; i < nr_linfo; i++) {
15629 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
15630 if (err) {
15631 if (err == -E2BIG) {
15632 verbose(env, "nonzero tailing record in line_info");
15633 if (copy_to_bpfptr_offset(uattr,
15634 offsetof(union bpf_attr, line_info_rec_size),
15635 &expected_size, sizeof(expected_size)))
15636 err = -EFAULT;
15637 }
15638 goto err_free;
15639 }
15640
15641 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
15642 err = -EFAULT;
15643 goto err_free;
15644 }
15645
15646 /*
15647 * Check insn_off to ensure
15648 * 1) strictly increasing AND
15649 * 2) bounded by prog->len
15650 *
15651 * The linfo[0].insn_off == 0 check logically falls into
15652 * the later "missing bpf_line_info for func..." case
15653 * because the first linfo[0].insn_off must be the
15654 * first sub also and the first sub must have
15655 * subprog_info[0].start == 0.
15656 */
15657 if ((i && linfo[i].insn_off <= prev_offset) ||
15658 linfo[i].insn_off >= prog->len) {
15659 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
15660 i, linfo[i].insn_off, prev_offset,
15661 prog->len);
15662 err = -EINVAL;
15663 goto err_free;
15664 }
15665
15666 if (!prog->insnsi[linfo[i].insn_off].code) {
15667 verbose(env,
15668 "Invalid insn code at line_info[%u].insn_off\n",
15669 i);
15670 err = -EINVAL;
15671 goto err_free;
15672 }
15673
15674 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
15675 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
15676 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
15677 err = -EINVAL;
15678 goto err_free;
15679 }
15680
15681 if (s != env->subprog_cnt) {
15682 if (linfo[i].insn_off == sub[s].start) {
15683 sub[s].linfo_idx = i;
15684 s++;
15685 } else if (sub[s].start < linfo[i].insn_off) {
15686 verbose(env, "missing bpf_line_info for func#%u\n", s);
15687 err = -EINVAL;
15688 goto err_free;
15689 }
15690 }
15691
15692 prev_offset = linfo[i].insn_off;
15693 bpfptr_add(&ulinfo, rec_size);
15694 }
15695
15696 if (s != env->subprog_cnt) {
15697 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
15698 env->subprog_cnt - s, s);
15699 err = -EINVAL;
15700 goto err_free;
15701 }
15702
15703 prog->aux->linfo = linfo;
15704 prog->aux->nr_linfo = nr_linfo;
15705
15706 return 0;
15707
15708 err_free:
15709 kvfree(linfo);
15710 return err;
15711 }
15712
15713 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
15714 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
15715
check_core_relo(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15716 static int check_core_relo(struct bpf_verifier_env *env,
15717 const union bpf_attr *attr,
15718 bpfptr_t uattr)
15719 {
15720 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
15721 struct bpf_core_relo core_relo = {};
15722 struct bpf_prog *prog = env->prog;
15723 const struct btf *btf = prog->aux->btf;
15724 struct bpf_core_ctx ctx = {
15725 .log = &env->log,
15726 .btf = btf,
15727 };
15728 bpfptr_t u_core_relo;
15729 int err;
15730
15731 nr_core_relo = attr->core_relo_cnt;
15732 if (!nr_core_relo)
15733 return 0;
15734 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
15735 return -EINVAL;
15736
15737 rec_size = attr->core_relo_rec_size;
15738 if (rec_size < MIN_CORE_RELO_SIZE ||
15739 rec_size > MAX_CORE_RELO_SIZE ||
15740 rec_size % sizeof(u32))
15741 return -EINVAL;
15742
15743 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
15744 expected_size = sizeof(struct bpf_core_relo);
15745 ncopy = min_t(u32, expected_size, rec_size);
15746
15747 /* Unlike func_info and line_info, copy and apply each CO-RE
15748 * relocation record one at a time.
15749 */
15750 for (i = 0; i < nr_core_relo; i++) {
15751 /* future proofing when sizeof(bpf_core_relo) changes */
15752 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
15753 if (err) {
15754 if (err == -E2BIG) {
15755 verbose(env, "nonzero tailing record in core_relo");
15756 if (copy_to_bpfptr_offset(uattr,
15757 offsetof(union bpf_attr, core_relo_rec_size),
15758 &expected_size, sizeof(expected_size)))
15759 err = -EFAULT;
15760 }
15761 break;
15762 }
15763
15764 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
15765 err = -EFAULT;
15766 break;
15767 }
15768
15769 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
15770 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
15771 i, core_relo.insn_off, prog->len);
15772 err = -EINVAL;
15773 break;
15774 }
15775
15776 err = bpf_core_apply(&ctx, &core_relo, i,
15777 &prog->insnsi[core_relo.insn_off / 8]);
15778 if (err)
15779 break;
15780 bpfptr_add(&u_core_relo, rec_size);
15781 }
15782 return err;
15783 }
15784
check_btf_info(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)15785 static int check_btf_info(struct bpf_verifier_env *env,
15786 const union bpf_attr *attr,
15787 bpfptr_t uattr)
15788 {
15789 struct btf *btf;
15790 int err;
15791
15792 if (!attr->func_info_cnt && !attr->line_info_cnt) {
15793 if (check_abnormal_return(env))
15794 return -EINVAL;
15795 return 0;
15796 }
15797
15798 btf = btf_get_by_fd(attr->prog_btf_fd);
15799 if (IS_ERR(btf))
15800 return PTR_ERR(btf);
15801 if (btf_is_kernel(btf)) {
15802 btf_put(btf);
15803 return -EACCES;
15804 }
15805 env->prog->aux->btf = btf;
15806
15807 err = check_btf_func(env, attr, uattr);
15808 if (err)
15809 return err;
15810
15811 err = check_btf_line(env, attr, uattr);
15812 if (err)
15813 return err;
15814
15815 err = check_core_relo(env, attr, uattr);
15816 if (err)
15817 return err;
15818
15819 return 0;
15820 }
15821
15822 /* check %cur's range satisfies %old's */
range_within(struct bpf_reg_state * old,struct bpf_reg_state * cur)15823 static bool range_within(struct bpf_reg_state *old,
15824 struct bpf_reg_state *cur)
15825 {
15826 return old->umin_value <= cur->umin_value &&
15827 old->umax_value >= cur->umax_value &&
15828 old->smin_value <= cur->smin_value &&
15829 old->smax_value >= cur->smax_value &&
15830 old->u32_min_value <= cur->u32_min_value &&
15831 old->u32_max_value >= cur->u32_max_value &&
15832 old->s32_min_value <= cur->s32_min_value &&
15833 old->s32_max_value >= cur->s32_max_value;
15834 }
15835
15836 /* If in the old state two registers had the same id, then they need to have
15837 * the same id in the new state as well. But that id could be different from
15838 * the old state, so we need to track the mapping from old to new ids.
15839 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
15840 * regs with old id 5 must also have new id 9 for the new state to be safe. But
15841 * regs with a different old id could still have new id 9, we don't care about
15842 * that.
15843 * So we look through our idmap to see if this old id has been seen before. If
15844 * so, we require the new id to match; otherwise, we add the id pair to the map.
15845 */
check_ids(u32 old_id,u32 cur_id,struct bpf_idmap * idmap)15846 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15847 {
15848 struct bpf_id_pair *map = idmap->map;
15849 unsigned int i;
15850
15851 /* either both IDs should be set or both should be zero */
15852 if (!!old_id != !!cur_id)
15853 return false;
15854
15855 if (old_id == 0) /* cur_id == 0 as well */
15856 return true;
15857
15858 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
15859 if (!map[i].old) {
15860 /* Reached an empty slot; haven't seen this id before */
15861 map[i].old = old_id;
15862 map[i].cur = cur_id;
15863 return true;
15864 }
15865 if (map[i].old == old_id)
15866 return map[i].cur == cur_id;
15867 if (map[i].cur == cur_id)
15868 return false;
15869 }
15870 /* We ran out of idmap slots, which should be impossible */
15871 WARN_ON_ONCE(1);
15872 return false;
15873 }
15874
15875 /* Similar to check_ids(), but allocate a unique temporary ID
15876 * for 'old_id' or 'cur_id' of zero.
15877 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
15878 */
check_scalar_ids(u32 old_id,u32 cur_id,struct bpf_idmap * idmap)15879 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
15880 {
15881 old_id = old_id ? old_id : ++idmap->tmp_id_gen;
15882 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
15883
15884 return check_ids(old_id, cur_id, idmap);
15885 }
15886
clean_func_state(struct bpf_verifier_env * env,struct bpf_func_state * st)15887 static void clean_func_state(struct bpf_verifier_env *env,
15888 struct bpf_func_state *st)
15889 {
15890 enum bpf_reg_liveness live;
15891 int i, j;
15892
15893 for (i = 0; i < BPF_REG_FP; i++) {
15894 live = st->regs[i].live;
15895 /* liveness must not touch this register anymore */
15896 st->regs[i].live |= REG_LIVE_DONE;
15897 if (!(live & REG_LIVE_READ))
15898 /* since the register is unused, clear its state
15899 * to make further comparison simpler
15900 */
15901 __mark_reg_not_init(env, &st->regs[i]);
15902 }
15903
15904 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
15905 live = st->stack[i].spilled_ptr.live;
15906 /* liveness must not touch this stack slot anymore */
15907 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
15908 if (!(live & REG_LIVE_READ)) {
15909 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
15910 for (j = 0; j < BPF_REG_SIZE; j++)
15911 st->stack[i].slot_type[j] = STACK_INVALID;
15912 }
15913 }
15914 }
15915
clean_verifier_state(struct bpf_verifier_env * env,struct bpf_verifier_state * st)15916 static void clean_verifier_state(struct bpf_verifier_env *env,
15917 struct bpf_verifier_state *st)
15918 {
15919 int i;
15920
15921 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
15922 /* all regs in this state in all frames were already marked */
15923 return;
15924
15925 for (i = 0; i <= st->curframe; i++)
15926 clean_func_state(env, st->frame[i]);
15927 }
15928
15929 /* the parentage chains form a tree.
15930 * the verifier states are added to state lists at given insn and
15931 * pushed into state stack for future exploration.
15932 * when the verifier reaches bpf_exit insn some of the verifer states
15933 * stored in the state lists have their final liveness state already,
15934 * but a lot of states will get revised from liveness point of view when
15935 * the verifier explores other branches.
15936 * Example:
15937 * 1: r0 = 1
15938 * 2: if r1 == 100 goto pc+1
15939 * 3: r0 = 2
15940 * 4: exit
15941 * when the verifier reaches exit insn the register r0 in the state list of
15942 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
15943 * of insn 2 and goes exploring further. At the insn 4 it will walk the
15944 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
15945 *
15946 * Since the verifier pushes the branch states as it sees them while exploring
15947 * the program the condition of walking the branch instruction for the second
15948 * time means that all states below this branch were already explored and
15949 * their final liveness marks are already propagated.
15950 * Hence when the verifier completes the search of state list in is_state_visited()
15951 * we can call this clean_live_states() function to mark all liveness states
15952 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
15953 * will not be used.
15954 * This function also clears the registers and stack for states that !READ
15955 * to simplify state merging.
15956 *
15957 * Important note here that walking the same branch instruction in the callee
15958 * doesn't meant that the states are DONE. The verifier has to compare
15959 * the callsites
15960 */
clean_live_states(struct bpf_verifier_env * env,int insn,struct bpf_verifier_state * cur)15961 static void clean_live_states(struct bpf_verifier_env *env, int insn,
15962 struct bpf_verifier_state *cur)
15963 {
15964 struct bpf_verifier_state_list *sl;
15965
15966 sl = *explored_state(env, insn);
15967 while (sl) {
15968 if (sl->state.branches)
15969 goto next;
15970 if (sl->state.insn_idx != insn ||
15971 !same_callsites(&sl->state, cur))
15972 goto next;
15973 clean_verifier_state(env, &sl->state);
15974 next:
15975 sl = sl->next;
15976 }
15977 }
15978
regs_exact(const struct bpf_reg_state * rold,const struct bpf_reg_state * rcur,struct bpf_idmap * idmap)15979 static bool regs_exact(const struct bpf_reg_state *rold,
15980 const struct bpf_reg_state *rcur,
15981 struct bpf_idmap *idmap)
15982 {
15983 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
15984 check_ids(rold->id, rcur->id, idmap) &&
15985 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
15986 }
15987
15988 /* 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)15989 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
15990 struct bpf_reg_state *rcur, struct bpf_idmap *idmap, bool exact)
15991 {
15992 if (exact)
15993 return regs_exact(rold, rcur, idmap);
15994
15995 if (!(rold->live & REG_LIVE_READ))
15996 /* explored state didn't use this */
15997 return true;
15998 if (rold->type == NOT_INIT)
15999 /* explored state can't have used this */
16000 return true;
16001 if (rcur->type == NOT_INIT)
16002 return false;
16003
16004 /* Enforce that register types have to match exactly, including their
16005 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
16006 * rule.
16007 *
16008 * One can make a point that using a pointer register as unbounded
16009 * SCALAR would be technically acceptable, but this could lead to
16010 * pointer leaks because scalars are allowed to leak while pointers
16011 * are not. We could make this safe in special cases if root is
16012 * calling us, but it's probably not worth the hassle.
16013 *
16014 * Also, register types that are *not* MAYBE_NULL could technically be
16015 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
16016 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
16017 * to the same map).
16018 * However, if the old MAYBE_NULL register then got NULL checked,
16019 * doing so could have affected others with the same id, and we can't
16020 * check for that because we lost the id when we converted to
16021 * a non-MAYBE_NULL variant.
16022 * So, as a general rule we don't allow mixing MAYBE_NULL and
16023 * non-MAYBE_NULL registers as well.
16024 */
16025 if (rold->type != rcur->type)
16026 return false;
16027
16028 switch (base_type(rold->type)) {
16029 case SCALAR_VALUE:
16030 if (env->explore_alu_limits) {
16031 /* explore_alu_limits disables tnum_in() and range_within()
16032 * logic and requires everything to be strict
16033 */
16034 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
16035 check_scalar_ids(rold->id, rcur->id, idmap);
16036 }
16037 if (!rold->precise)
16038 return true;
16039 /* Why check_ids() for scalar registers?
16040 *
16041 * Consider the following BPF code:
16042 * 1: r6 = ... unbound scalar, ID=a ...
16043 * 2: r7 = ... unbound scalar, ID=b ...
16044 * 3: if (r6 > r7) goto +1
16045 * 4: r6 = r7
16046 * 5: if (r6 > X) goto ...
16047 * 6: ... memory operation using r7 ...
16048 *
16049 * First verification path is [1-6]:
16050 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
16051 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark
16052 * r7 <= X, because r6 and r7 share same id.
16053 * Next verification path is [1-4, 6].
16054 *
16055 * Instruction (6) would be reached in two states:
16056 * I. r6{.id=b}, r7{.id=b} via path 1-6;
16057 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
16058 *
16059 * Use check_ids() to distinguish these states.
16060 * ---
16061 * Also verify that new value satisfies old value range knowledge.
16062 */
16063 return range_within(rold, rcur) &&
16064 tnum_in(rold->var_off, rcur->var_off) &&
16065 check_scalar_ids(rold->id, rcur->id, idmap);
16066 case PTR_TO_MAP_KEY:
16067 case PTR_TO_MAP_VALUE:
16068 case PTR_TO_MEM:
16069 case PTR_TO_BUF:
16070 case PTR_TO_TP_BUFFER:
16071 /* If the new min/max/var_off satisfy the old ones and
16072 * everything else matches, we are OK.
16073 */
16074 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
16075 range_within(rold, rcur) &&
16076 tnum_in(rold->var_off, rcur->var_off) &&
16077 check_ids(rold->id, rcur->id, idmap) &&
16078 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
16079 case PTR_TO_PACKET_META:
16080 case PTR_TO_PACKET:
16081 /* We must have at least as much range as the old ptr
16082 * did, so that any accesses which were safe before are
16083 * still safe. This is true even if old range < old off,
16084 * since someone could have accessed through (ptr - k), or
16085 * even done ptr -= k in a register, to get a safe access.
16086 */
16087 if (rold->range > rcur->range)
16088 return false;
16089 /* If the offsets don't match, we can't trust our alignment;
16090 * nor can we be sure that we won't fall out of range.
16091 */
16092 if (rold->off != rcur->off)
16093 return false;
16094 /* id relations must be preserved */
16095 if (!check_ids(rold->id, rcur->id, idmap))
16096 return false;
16097 /* new val must satisfy old val knowledge */
16098 return range_within(rold, rcur) &&
16099 tnum_in(rold->var_off, rcur->var_off);
16100 case PTR_TO_STACK:
16101 /* two stack pointers are equal only if they're pointing to
16102 * the same stack frame, since fp-8 in foo != fp-8 in bar
16103 */
16104 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
16105 default:
16106 return regs_exact(rold, rcur, idmap);
16107 }
16108 }
16109
stacksafe(struct bpf_verifier_env * env,struct bpf_func_state * old,struct bpf_func_state * cur,struct bpf_idmap * idmap,bool exact)16110 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
16111 struct bpf_func_state *cur, struct bpf_idmap *idmap, bool exact)
16112 {
16113 int i, spi;
16114
16115 /* walk slots of the explored stack and ignore any additional
16116 * slots in the current stack, since explored(safe) state
16117 * didn't use them
16118 */
16119 for (i = 0; i < old->allocated_stack; i++) {
16120 struct bpf_reg_state *old_reg, *cur_reg;
16121
16122 spi = i / BPF_REG_SIZE;
16123
16124 if (exact &&
16125 old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
16126 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
16127 return false;
16128
16129 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ) && !exact) {
16130 i += BPF_REG_SIZE - 1;
16131 /* explored state didn't use this */
16132 continue;
16133 }
16134
16135 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
16136 continue;
16137
16138 if (env->allow_uninit_stack &&
16139 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
16140 continue;
16141
16142 /* explored stack has more populated slots than current stack
16143 * and these slots were used
16144 */
16145 if (i >= cur->allocated_stack)
16146 return false;
16147
16148 /* if old state was safe with misc data in the stack
16149 * it will be safe with zero-initialized stack.
16150 * The opposite is not true
16151 */
16152 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
16153 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
16154 continue;
16155 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
16156 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
16157 /* Ex: old explored (safe) state has STACK_SPILL in
16158 * this stack slot, but current has STACK_MISC ->
16159 * this verifier states are not equivalent,
16160 * return false to continue verification of this path
16161 */
16162 return false;
16163 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
16164 continue;
16165 /* Both old and cur are having same slot_type */
16166 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
16167 case STACK_SPILL:
16168 /* when explored and current stack slot are both storing
16169 * spilled registers, check that stored pointers types
16170 * are the same as well.
16171 * Ex: explored safe path could have stored
16172 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
16173 * but current path has stored:
16174 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
16175 * such verifier states are not equivalent.
16176 * return false to continue verification of this path
16177 */
16178 if (!regsafe(env, &old->stack[spi].spilled_ptr,
16179 &cur->stack[spi].spilled_ptr, idmap, exact))
16180 return false;
16181 break;
16182 case STACK_DYNPTR:
16183 old_reg = &old->stack[spi].spilled_ptr;
16184 cur_reg = &cur->stack[spi].spilled_ptr;
16185 if (old_reg->dynptr.type != cur_reg->dynptr.type ||
16186 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
16187 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
16188 return false;
16189 break;
16190 case STACK_ITER:
16191 old_reg = &old->stack[spi].spilled_ptr;
16192 cur_reg = &cur->stack[spi].spilled_ptr;
16193 /* iter.depth is not compared between states as it
16194 * doesn't matter for correctness and would otherwise
16195 * prevent convergence; we maintain it only to prevent
16196 * infinite loop check triggering, see
16197 * iter_active_depths_differ()
16198 */
16199 if (old_reg->iter.btf != cur_reg->iter.btf ||
16200 old_reg->iter.btf_id != cur_reg->iter.btf_id ||
16201 old_reg->iter.state != cur_reg->iter.state ||
16202 /* ignore {old_reg,cur_reg}->iter.depth, see above */
16203 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
16204 return false;
16205 break;
16206 case STACK_MISC:
16207 case STACK_ZERO:
16208 case STACK_INVALID:
16209 continue;
16210 /* Ensure that new unhandled slot types return false by default */
16211 default:
16212 return false;
16213 }
16214 }
16215 return true;
16216 }
16217
refsafe(struct bpf_func_state * old,struct bpf_func_state * cur,struct bpf_idmap * idmap)16218 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur,
16219 struct bpf_idmap *idmap)
16220 {
16221 int i;
16222
16223 if (old->acquired_refs != cur->acquired_refs)
16224 return false;
16225
16226 for (i = 0; i < old->acquired_refs; i++) {
16227 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap))
16228 return false;
16229 }
16230
16231 return true;
16232 }
16233
16234 /* compare two verifier states
16235 *
16236 * all states stored in state_list are known to be valid, since
16237 * verifier reached 'bpf_exit' instruction through them
16238 *
16239 * this function is called when verifier exploring different branches of
16240 * execution popped from the state stack. If it sees an old state that has
16241 * more strict register state and more strict stack state then this execution
16242 * branch doesn't need to be explored further, since verifier already
16243 * concluded that more strict state leads to valid finish.
16244 *
16245 * Therefore two states are equivalent if register state is more conservative
16246 * and explored stack state is more conservative than the current one.
16247 * Example:
16248 * explored current
16249 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
16250 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
16251 *
16252 * In other words if current stack state (one being explored) has more
16253 * valid slots than old one that already passed validation, it means
16254 * the verifier can stop exploring and conclude that current state is valid too
16255 *
16256 * Similarly with registers. If explored state has register type as invalid
16257 * whereas register type in current state is meaningful, it means that
16258 * the current state will reach 'bpf_exit' instruction safely
16259 */
func_states_equal(struct bpf_verifier_env * env,struct bpf_func_state * old,struct bpf_func_state * cur,bool exact)16260 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
16261 struct bpf_func_state *cur, bool exact)
16262 {
16263 int i;
16264
16265 if (old->callback_depth > cur->callback_depth)
16266 return false;
16267
16268 for (i = 0; i < MAX_BPF_REG; i++)
16269 if (!regsafe(env, &old->regs[i], &cur->regs[i],
16270 &env->idmap_scratch, exact))
16271 return false;
16272
16273 if (!stacksafe(env, old, cur, &env->idmap_scratch, exact))
16274 return false;
16275
16276 if (!refsafe(old, cur, &env->idmap_scratch))
16277 return false;
16278
16279 return true;
16280 }
16281
reset_idmap_scratch(struct bpf_verifier_env * env)16282 static void reset_idmap_scratch(struct bpf_verifier_env *env)
16283 {
16284 env->idmap_scratch.tmp_id_gen = env->id_gen;
16285 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
16286 }
16287
states_equal(struct bpf_verifier_env * env,struct bpf_verifier_state * old,struct bpf_verifier_state * cur,bool exact)16288 static bool states_equal(struct bpf_verifier_env *env,
16289 struct bpf_verifier_state *old,
16290 struct bpf_verifier_state *cur,
16291 bool exact)
16292 {
16293 int i;
16294
16295 if (old->curframe != cur->curframe)
16296 return false;
16297
16298 reset_idmap_scratch(env);
16299
16300 /* Verification state from speculative execution simulation
16301 * must never prune a non-speculative execution one.
16302 */
16303 if (old->speculative && !cur->speculative)
16304 return false;
16305
16306 if (old->active_lock.ptr != cur->active_lock.ptr)
16307 return false;
16308
16309 /* Old and cur active_lock's have to be either both present
16310 * or both absent.
16311 */
16312 if (!!old->active_lock.id != !!cur->active_lock.id)
16313 return false;
16314
16315 if (old->active_lock.id &&
16316 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch))
16317 return false;
16318
16319 if (old->active_rcu_lock != cur->active_rcu_lock)
16320 return false;
16321
16322 /* for states to be equal callsites have to be the same
16323 * and all frame states need to be equivalent
16324 */
16325 for (i = 0; i <= old->curframe; i++) {
16326 if (old->frame[i]->callsite != cur->frame[i]->callsite)
16327 return false;
16328 if (!func_states_equal(env, old->frame[i], cur->frame[i], exact))
16329 return false;
16330 }
16331 return true;
16332 }
16333
16334 /* Return 0 if no propagation happened. Return negative error code if error
16335 * happened. Otherwise, return the propagated bit.
16336 */
propagate_liveness_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct bpf_reg_state * parent_reg)16337 static int propagate_liveness_reg(struct bpf_verifier_env *env,
16338 struct bpf_reg_state *reg,
16339 struct bpf_reg_state *parent_reg)
16340 {
16341 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
16342 u8 flag = reg->live & REG_LIVE_READ;
16343 int err;
16344
16345 /* When comes here, read flags of PARENT_REG or REG could be any of
16346 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
16347 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
16348 */
16349 if (parent_flag == REG_LIVE_READ64 ||
16350 /* Or if there is no read flag from REG. */
16351 !flag ||
16352 /* Or if the read flag from REG is the same as PARENT_REG. */
16353 parent_flag == flag)
16354 return 0;
16355
16356 err = mark_reg_read(env, reg, parent_reg, flag);
16357 if (err)
16358 return err;
16359
16360 return flag;
16361 }
16362
16363 /* A write screens off any subsequent reads; but write marks come from the
16364 * straight-line code between a state and its parent. When we arrive at an
16365 * equivalent state (jump target or such) we didn't arrive by the straight-line
16366 * code, so read marks in the state must propagate to the parent regardless
16367 * of the state's write marks. That's what 'parent == state->parent' comparison
16368 * in mark_reg_read() is for.
16369 */
propagate_liveness(struct bpf_verifier_env * env,const struct bpf_verifier_state * vstate,struct bpf_verifier_state * vparent)16370 static int propagate_liveness(struct bpf_verifier_env *env,
16371 const struct bpf_verifier_state *vstate,
16372 struct bpf_verifier_state *vparent)
16373 {
16374 struct bpf_reg_state *state_reg, *parent_reg;
16375 struct bpf_func_state *state, *parent;
16376 int i, frame, err = 0;
16377
16378 if (vparent->curframe != vstate->curframe) {
16379 WARN(1, "propagate_live: parent frame %d current frame %d\n",
16380 vparent->curframe, vstate->curframe);
16381 return -EFAULT;
16382 }
16383 /* Propagate read liveness of registers... */
16384 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
16385 for (frame = 0; frame <= vstate->curframe; frame++) {
16386 parent = vparent->frame[frame];
16387 state = vstate->frame[frame];
16388 parent_reg = parent->regs;
16389 state_reg = state->regs;
16390 /* We don't need to worry about FP liveness, it's read-only */
16391 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
16392 err = propagate_liveness_reg(env, &state_reg[i],
16393 &parent_reg[i]);
16394 if (err < 0)
16395 return err;
16396 if (err == REG_LIVE_READ64)
16397 mark_insn_zext(env, &parent_reg[i]);
16398 }
16399
16400 /* Propagate stack slots. */
16401 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
16402 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
16403 parent_reg = &parent->stack[i].spilled_ptr;
16404 state_reg = &state->stack[i].spilled_ptr;
16405 err = propagate_liveness_reg(env, state_reg,
16406 parent_reg);
16407 if (err < 0)
16408 return err;
16409 }
16410 }
16411 return 0;
16412 }
16413
16414 /* find precise scalars in the previous equivalent state and
16415 * propagate them into the current state
16416 */
propagate_precision(struct bpf_verifier_env * env,const struct bpf_verifier_state * old)16417 static int propagate_precision(struct bpf_verifier_env *env,
16418 const struct bpf_verifier_state *old)
16419 {
16420 struct bpf_reg_state *state_reg;
16421 struct bpf_func_state *state;
16422 int i, err = 0, fr;
16423 bool first;
16424
16425 for (fr = old->curframe; fr >= 0; fr--) {
16426 state = old->frame[fr];
16427 state_reg = state->regs;
16428 first = true;
16429 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
16430 if (state_reg->type != SCALAR_VALUE ||
16431 !state_reg->precise ||
16432 !(state_reg->live & REG_LIVE_READ))
16433 continue;
16434 if (env->log.level & BPF_LOG_LEVEL2) {
16435 if (first)
16436 verbose(env, "frame %d: propagating r%d", fr, i);
16437 else
16438 verbose(env, ",r%d", i);
16439 }
16440 bt_set_frame_reg(&env->bt, fr, i);
16441 first = false;
16442 }
16443
16444 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16445 if (!is_spilled_reg(&state->stack[i]))
16446 continue;
16447 state_reg = &state->stack[i].spilled_ptr;
16448 if (state_reg->type != SCALAR_VALUE ||
16449 !state_reg->precise ||
16450 !(state_reg->live & REG_LIVE_READ))
16451 continue;
16452 if (env->log.level & BPF_LOG_LEVEL2) {
16453 if (first)
16454 verbose(env, "frame %d: propagating fp%d",
16455 fr, (-i - 1) * BPF_REG_SIZE);
16456 else
16457 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
16458 }
16459 bt_set_frame_slot(&env->bt, fr, i);
16460 first = false;
16461 }
16462 if (!first)
16463 verbose(env, "\n");
16464 }
16465
16466 err = mark_chain_precision_batch(env);
16467 if (err < 0)
16468 return err;
16469
16470 return 0;
16471 }
16472
states_maybe_looping(struct bpf_verifier_state * old,struct bpf_verifier_state * cur)16473 static bool states_maybe_looping(struct bpf_verifier_state *old,
16474 struct bpf_verifier_state *cur)
16475 {
16476 struct bpf_func_state *fold, *fcur;
16477 int i, fr = cur->curframe;
16478
16479 if (old->curframe != fr)
16480 return false;
16481
16482 fold = old->frame[fr];
16483 fcur = cur->frame[fr];
16484 for (i = 0; i < MAX_BPF_REG; i++)
16485 if (memcmp(&fold->regs[i], &fcur->regs[i],
16486 offsetof(struct bpf_reg_state, parent)))
16487 return false;
16488 return true;
16489 }
16490
is_iter_next_insn(struct bpf_verifier_env * env,int insn_idx)16491 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
16492 {
16493 return env->insn_aux_data[insn_idx].is_iter_next;
16494 }
16495
16496 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
16497 * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
16498 * states to match, which otherwise would look like an infinite loop. So while
16499 * iter_next() calls are taken care of, we still need to be careful and
16500 * prevent erroneous and too eager declaration of "ininite loop", when
16501 * iterators are involved.
16502 *
16503 * Here's a situation in pseudo-BPF assembly form:
16504 *
16505 * 0: again: ; set up iter_next() call args
16506 * 1: r1 = &it ; <CHECKPOINT HERE>
16507 * 2: call bpf_iter_num_next ; this is iter_next() call
16508 * 3: if r0 == 0 goto done
16509 * 4: ... something useful here ...
16510 * 5: goto again ; another iteration
16511 * 6: done:
16512 * 7: r1 = &it
16513 * 8: call bpf_iter_num_destroy ; clean up iter state
16514 * 9: exit
16515 *
16516 * This is a typical loop. Let's assume that we have a prune point at 1:,
16517 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
16518 * again`, assuming other heuristics don't get in a way).
16519 *
16520 * When we first time come to 1:, let's say we have some state X. We proceed
16521 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
16522 * Now we come back to validate that forked ACTIVE state. We proceed through
16523 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
16524 * are converging. But the problem is that we don't know that yet, as this
16525 * convergence has to happen at iter_next() call site only. So if nothing is
16526 * done, at 1: verifier will use bounded loop logic and declare infinite
16527 * looping (and would be *technically* correct, if not for iterator's
16528 * "eventual sticky NULL" contract, see process_iter_next_call()). But we
16529 * don't want that. So what we do in process_iter_next_call() when we go on
16530 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
16531 * a different iteration. So when we suspect an infinite loop, we additionally
16532 * check if any of the *ACTIVE* iterator states depths differ. If yes, we
16533 * pretend we are not looping and wait for next iter_next() call.
16534 *
16535 * This only applies to ACTIVE state. In DRAINED state we don't expect to
16536 * loop, because that would actually mean infinite loop, as DRAINED state is
16537 * "sticky", and so we'll keep returning into the same instruction with the
16538 * same state (at least in one of possible code paths).
16539 *
16540 * This approach allows to keep infinite loop heuristic even in the face of
16541 * active iterator. E.g., C snippet below is and will be detected as
16542 * inifintely looping:
16543 *
16544 * struct bpf_iter_num it;
16545 * int *p, x;
16546 *
16547 * bpf_iter_num_new(&it, 0, 10);
16548 * while ((p = bpf_iter_num_next(&t))) {
16549 * x = p;
16550 * while (x--) {} // <<-- infinite loop here
16551 * }
16552 *
16553 */
iter_active_depths_differ(struct bpf_verifier_state * old,struct bpf_verifier_state * cur)16554 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
16555 {
16556 struct bpf_reg_state *slot, *cur_slot;
16557 struct bpf_func_state *state;
16558 int i, fr;
16559
16560 for (fr = old->curframe; fr >= 0; fr--) {
16561 state = old->frame[fr];
16562 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
16563 if (state->stack[i].slot_type[0] != STACK_ITER)
16564 continue;
16565
16566 slot = &state->stack[i].spilled_ptr;
16567 if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
16568 continue;
16569
16570 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
16571 if (cur_slot->iter.depth != slot->iter.depth)
16572 return true;
16573 }
16574 }
16575 return false;
16576 }
16577
is_state_visited(struct bpf_verifier_env * env,int insn_idx)16578 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
16579 {
16580 struct bpf_verifier_state_list *new_sl;
16581 struct bpf_verifier_state_list *sl, **pprev;
16582 struct bpf_verifier_state *cur = env->cur_state, *new, *loop_entry;
16583 int i, j, n, err, states_cnt = 0;
16584 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx);
16585 bool add_new_state = force_new_state;
16586 bool force_exact;
16587
16588 /* bpf progs typically have pruning point every 4 instructions
16589 * http://vger.kernel.org/bpfconf2019.html#session-1
16590 * Do not add new state for future pruning if the verifier hasn't seen
16591 * at least 2 jumps and at least 8 instructions.
16592 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
16593 * In tests that amounts to up to 50% reduction into total verifier
16594 * memory consumption and 20% verifier time speedup.
16595 */
16596 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
16597 env->insn_processed - env->prev_insn_processed >= 8)
16598 add_new_state = true;
16599
16600 pprev = explored_state(env, insn_idx);
16601 sl = *pprev;
16602
16603 clean_live_states(env, insn_idx, cur);
16604
16605 while (sl) {
16606 states_cnt++;
16607 if (sl->state.insn_idx != insn_idx)
16608 goto next;
16609
16610 if (sl->state.branches) {
16611 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
16612
16613 if (frame->in_async_callback_fn &&
16614 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
16615 /* Different async_entry_cnt means that the verifier is
16616 * processing another entry into async callback.
16617 * Seeing the same state is not an indication of infinite
16618 * loop or infinite recursion.
16619 * But finding the same state doesn't mean that it's safe
16620 * to stop processing the current state. The previous state
16621 * hasn't yet reached bpf_exit, since state.branches > 0.
16622 * Checking in_async_callback_fn alone is not enough either.
16623 * Since the verifier still needs to catch infinite loops
16624 * inside async callbacks.
16625 */
16626 goto skip_inf_loop_check;
16627 }
16628 /* BPF open-coded iterators loop detection is special.
16629 * states_maybe_looping() logic is too simplistic in detecting
16630 * states that *might* be equivalent, because it doesn't know
16631 * about ID remapping, so don't even perform it.
16632 * See process_iter_next_call() and iter_active_depths_differ()
16633 * for overview of the logic. When current and one of parent
16634 * states are detected as equivalent, it's a good thing: we prove
16635 * convergence and can stop simulating further iterations.
16636 * It's safe to assume that iterator loop will finish, taking into
16637 * account iter_next() contract of eventually returning
16638 * sticky NULL result.
16639 *
16640 * Note, that states have to be compared exactly in this case because
16641 * read and precision marks might not be finalized inside the loop.
16642 * E.g. as in the program below:
16643 *
16644 * 1. r7 = -16
16645 * 2. r6 = bpf_get_prandom_u32()
16646 * 3. while (bpf_iter_num_next(&fp[-8])) {
16647 * 4. if (r6 != 42) {
16648 * 5. r7 = -32
16649 * 6. r6 = bpf_get_prandom_u32()
16650 * 7. continue
16651 * 8. }
16652 * 9. r0 = r10
16653 * 10. r0 += r7
16654 * 11. r8 = *(u64 *)(r0 + 0)
16655 * 12. r6 = bpf_get_prandom_u32()
16656 * 13. }
16657 *
16658 * Here verifier would first visit path 1-3, create a checkpoint at 3
16659 * with r7=-16, continue to 4-7,3. Existing checkpoint at 3 does
16660 * not have read or precision mark for r7 yet, thus inexact states
16661 * comparison would discard current state with r7=-32
16662 * => unsafe memory access at 11 would not be caught.
16663 */
16664 if (is_iter_next_insn(env, insn_idx)) {
16665 if (states_equal(env, &sl->state, cur, true)) {
16666 struct bpf_func_state *cur_frame;
16667 struct bpf_reg_state *iter_state, *iter_reg;
16668 int spi;
16669
16670 cur_frame = cur->frame[cur->curframe];
16671 /* btf_check_iter_kfuncs() enforces that
16672 * iter state pointer is always the first arg
16673 */
16674 iter_reg = &cur_frame->regs[BPF_REG_1];
16675 /* current state is valid due to states_equal(),
16676 * so we can assume valid iter and reg state,
16677 * no need for extra (re-)validations
16678 */
16679 spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
16680 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
16681 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE) {
16682 update_loop_entry(cur, &sl->state);
16683 goto hit;
16684 }
16685 }
16686 goto skip_inf_loop_check;
16687 }
16688 if (calls_callback(env, insn_idx)) {
16689 if (states_equal(env, &sl->state, cur, true))
16690 goto hit;
16691 goto skip_inf_loop_check;
16692 }
16693 /* attempt to detect infinite loop to avoid unnecessary doomed work */
16694 if (states_maybe_looping(&sl->state, cur) &&
16695 states_equal(env, &sl->state, cur, false) &&
16696 !iter_active_depths_differ(&sl->state, cur) &&
16697 sl->state.callback_unroll_depth == cur->callback_unroll_depth) {
16698 verbose_linfo(env, insn_idx, "; ");
16699 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
16700 verbose(env, "cur state:");
16701 print_verifier_state(env, cur->frame[cur->curframe], true);
16702 verbose(env, "old state:");
16703 print_verifier_state(env, sl->state.frame[cur->curframe], true);
16704 return -EINVAL;
16705 }
16706 /* if the verifier is processing a loop, avoid adding new state
16707 * too often, since different loop iterations have distinct
16708 * states and may not help future pruning.
16709 * This threshold shouldn't be too low to make sure that
16710 * a loop with large bound will be rejected quickly.
16711 * The most abusive loop will be:
16712 * r1 += 1
16713 * if r1 < 1000000 goto pc-2
16714 * 1M insn_procssed limit / 100 == 10k peak states.
16715 * This threshold shouldn't be too high either, since states
16716 * at the end of the loop are likely to be useful in pruning.
16717 */
16718 skip_inf_loop_check:
16719 if (!force_new_state &&
16720 env->jmps_processed - env->prev_jmps_processed < 20 &&
16721 env->insn_processed - env->prev_insn_processed < 100)
16722 add_new_state = false;
16723 goto miss;
16724 }
16725 /* If sl->state is a part of a loop and this loop's entry is a part of
16726 * current verification path then states have to be compared exactly.
16727 * 'force_exact' is needed to catch the following case:
16728 *
16729 * initial Here state 'succ' was processed first,
16730 * | it was eventually tracked to produce a
16731 * V state identical to 'hdr'.
16732 * .---------> hdr All branches from 'succ' had been explored
16733 * | | and thus 'succ' has its .branches == 0.
16734 * | V
16735 * | .------... Suppose states 'cur' and 'succ' correspond
16736 * | | | to the same instruction + callsites.
16737 * | V V In such case it is necessary to check
16738 * | ... ... if 'succ' and 'cur' are states_equal().
16739 * | | | If 'succ' and 'cur' are a part of the
16740 * | V V same loop exact flag has to be set.
16741 * | succ <- cur To check if that is the case, verify
16742 * | | if loop entry of 'succ' is in current
16743 * | V DFS path.
16744 * | ...
16745 * | |
16746 * '----'
16747 *
16748 * Additional details are in the comment before get_loop_entry().
16749 */
16750 loop_entry = get_loop_entry(&sl->state);
16751 force_exact = loop_entry && loop_entry->branches > 0;
16752 if (states_equal(env, &sl->state, cur, force_exact)) {
16753 if (force_exact)
16754 update_loop_entry(cur, loop_entry);
16755 hit:
16756 sl->hit_cnt++;
16757 /* reached equivalent register/stack state,
16758 * prune the search.
16759 * Registers read by the continuation are read by us.
16760 * If we have any write marks in env->cur_state, they
16761 * will prevent corresponding reads in the continuation
16762 * from reaching our parent (an explored_state). Our
16763 * own state will get the read marks recorded, but
16764 * they'll be immediately forgotten as we're pruning
16765 * this state and will pop a new one.
16766 */
16767 err = propagate_liveness(env, &sl->state, cur);
16768
16769 /* if previous state reached the exit with precision and
16770 * current state is equivalent to it (except precsion marks)
16771 * the precision needs to be propagated back in
16772 * the current state.
16773 */
16774 err = err ? : push_jmp_history(env, cur);
16775 err = err ? : propagate_precision(env, &sl->state);
16776 if (err)
16777 return err;
16778 return 1;
16779 }
16780 miss:
16781 /* when new state is not going to be added do not increase miss count.
16782 * Otherwise several loop iterations will remove the state
16783 * recorded earlier. The goal of these heuristics is to have
16784 * states from some iterations of the loop (some in the beginning
16785 * and some at the end) to help pruning.
16786 */
16787 if (add_new_state)
16788 sl->miss_cnt++;
16789 /* heuristic to determine whether this state is beneficial
16790 * to keep checking from state equivalence point of view.
16791 * Higher numbers increase max_states_per_insn and verification time,
16792 * but do not meaningfully decrease insn_processed.
16793 * 'n' controls how many times state could miss before eviction.
16794 * Use bigger 'n' for checkpoints because evicting checkpoint states
16795 * too early would hinder iterator convergence.
16796 */
16797 n = is_force_checkpoint(env, insn_idx) && sl->state.branches > 0 ? 64 : 3;
16798 if (sl->miss_cnt > sl->hit_cnt * n + n) {
16799 /* the state is unlikely to be useful. Remove it to
16800 * speed up verification
16801 */
16802 *pprev = sl->next;
16803 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE &&
16804 !sl->state.used_as_loop_entry) {
16805 u32 br = sl->state.branches;
16806
16807 WARN_ONCE(br,
16808 "BUG live_done but branches_to_explore %d\n",
16809 br);
16810 free_verifier_state(&sl->state, false);
16811 kfree(sl);
16812 env->peak_states--;
16813 } else {
16814 /* cannot free this state, since parentage chain may
16815 * walk it later. Add it for free_list instead to
16816 * be freed at the end of verification
16817 */
16818 sl->next = env->free_list;
16819 env->free_list = sl;
16820 }
16821 sl = *pprev;
16822 continue;
16823 }
16824 next:
16825 pprev = &sl->next;
16826 sl = *pprev;
16827 }
16828
16829 if (env->max_states_per_insn < states_cnt)
16830 env->max_states_per_insn = states_cnt;
16831
16832 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
16833 return 0;
16834
16835 if (!add_new_state)
16836 return 0;
16837
16838 /* There were no equivalent states, remember the current one.
16839 * Technically the current state is not proven to be safe yet,
16840 * but it will either reach outer most bpf_exit (which means it's safe)
16841 * or it will be rejected. When there are no loops the verifier won't be
16842 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
16843 * again on the way to bpf_exit.
16844 * When looping the sl->state.branches will be > 0 and this state
16845 * will not be considered for equivalence until branches == 0.
16846 */
16847 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
16848 if (!new_sl)
16849 return -ENOMEM;
16850 env->total_states++;
16851 env->peak_states++;
16852 env->prev_jmps_processed = env->jmps_processed;
16853 env->prev_insn_processed = env->insn_processed;
16854
16855 /* forget precise markings we inherited, see __mark_chain_precision */
16856 if (env->bpf_capable)
16857 mark_all_scalars_imprecise(env, cur);
16858
16859 /* add new state to the head of linked list */
16860 new = &new_sl->state;
16861 err = copy_verifier_state(new, cur);
16862 if (err) {
16863 free_verifier_state(new, false);
16864 kfree(new_sl);
16865 return err;
16866 }
16867 new->insn_idx = insn_idx;
16868 WARN_ONCE(new->branches != 1,
16869 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
16870
16871 cur->parent = new;
16872 cur->first_insn_idx = insn_idx;
16873 cur->dfs_depth = new->dfs_depth + 1;
16874 clear_jmp_history(cur);
16875 new_sl->next = *explored_state(env, insn_idx);
16876 *explored_state(env, insn_idx) = new_sl;
16877 /* connect new state to parentage chain. Current frame needs all
16878 * registers connected. Only r6 - r9 of the callers are alive (pushed
16879 * to the stack implicitly by JITs) so in callers' frames connect just
16880 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
16881 * the state of the call instruction (with WRITTEN set), and r0 comes
16882 * from callee with its full parentage chain, anyway.
16883 */
16884 /* clear write marks in current state: the writes we did are not writes
16885 * our child did, so they don't screen off its reads from us.
16886 * (There are no read marks in current state, because reads always mark
16887 * their parent and current state never has children yet. Only
16888 * explored_states can get read marks.)
16889 */
16890 for (j = 0; j <= cur->curframe; j++) {
16891 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
16892 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
16893 for (i = 0; i < BPF_REG_FP; i++)
16894 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
16895 }
16896
16897 /* all stack frames are accessible from callee, clear them all */
16898 for (j = 0; j <= cur->curframe; j++) {
16899 struct bpf_func_state *frame = cur->frame[j];
16900 struct bpf_func_state *newframe = new->frame[j];
16901
16902 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
16903 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
16904 frame->stack[i].spilled_ptr.parent =
16905 &newframe->stack[i].spilled_ptr;
16906 }
16907 }
16908 return 0;
16909 }
16910
16911 /* Return true if it's OK to have the same insn return a different type. */
reg_type_mismatch_ok(enum bpf_reg_type type)16912 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
16913 {
16914 switch (base_type(type)) {
16915 case PTR_TO_CTX:
16916 case PTR_TO_SOCKET:
16917 case PTR_TO_SOCK_COMMON:
16918 case PTR_TO_TCP_SOCK:
16919 case PTR_TO_XDP_SOCK:
16920 case PTR_TO_BTF_ID:
16921 return false;
16922 default:
16923 return true;
16924 }
16925 }
16926
16927 /* If an instruction was previously used with particular pointer types, then we
16928 * need to be careful to avoid cases such as the below, where it may be ok
16929 * for one branch accessing the pointer, but not ok for the other branch:
16930 *
16931 * R1 = sock_ptr
16932 * goto X;
16933 * ...
16934 * R1 = some_other_valid_ptr;
16935 * goto X;
16936 * ...
16937 * R2 = *(u32 *)(R1 + 0);
16938 */
reg_type_mismatch(enum bpf_reg_type src,enum bpf_reg_type prev)16939 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
16940 {
16941 return src != prev && (!reg_type_mismatch_ok(src) ||
16942 !reg_type_mismatch_ok(prev));
16943 }
16944
save_aux_ptr_type(struct bpf_verifier_env * env,enum bpf_reg_type type,bool allow_trust_missmatch)16945 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
16946 bool allow_trust_missmatch)
16947 {
16948 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
16949
16950 if (*prev_type == NOT_INIT) {
16951 /* Saw a valid insn
16952 * dst_reg = *(u32 *)(src_reg + off)
16953 * save type to validate intersecting paths
16954 */
16955 *prev_type = type;
16956 } else if (reg_type_mismatch(type, *prev_type)) {
16957 /* Abuser program is trying to use the same insn
16958 * dst_reg = *(u32*) (src_reg + off)
16959 * with different pointer types:
16960 * src_reg == ctx in one branch and
16961 * src_reg == stack|map in some other branch.
16962 * Reject it.
16963 */
16964 if (allow_trust_missmatch &&
16965 base_type(type) == PTR_TO_BTF_ID &&
16966 base_type(*prev_type) == PTR_TO_BTF_ID) {
16967 /*
16968 * Have to support a use case when one path through
16969 * the program yields TRUSTED pointer while another
16970 * is UNTRUSTED. Fallback to UNTRUSTED to generate
16971 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
16972 */
16973 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
16974 } else {
16975 verbose(env, "same insn cannot be used with different pointers\n");
16976 return -EINVAL;
16977 }
16978 }
16979
16980 return 0;
16981 }
16982
do_check(struct bpf_verifier_env * env)16983 static int do_check(struct bpf_verifier_env *env)
16984 {
16985 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
16986 struct bpf_verifier_state *state = env->cur_state;
16987 struct bpf_insn *insns = env->prog->insnsi;
16988 struct bpf_reg_state *regs;
16989 int insn_cnt = env->prog->len;
16990 bool do_print_state = false;
16991 int prev_insn_idx = -1;
16992
16993 for (;;) {
16994 struct bpf_insn *insn;
16995 u8 class;
16996 int err;
16997
16998 env->prev_insn_idx = prev_insn_idx;
16999 if (env->insn_idx >= insn_cnt) {
17000 verbose(env, "invalid insn idx %d insn_cnt %d\n",
17001 env->insn_idx, insn_cnt);
17002 return -EFAULT;
17003 }
17004
17005 insn = &insns[env->insn_idx];
17006 class = BPF_CLASS(insn->code);
17007
17008 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
17009 verbose(env,
17010 "BPF program is too large. Processed %d insn\n",
17011 env->insn_processed);
17012 return -E2BIG;
17013 }
17014
17015 state->last_insn_idx = env->prev_insn_idx;
17016
17017 if (is_prune_point(env, env->insn_idx)) {
17018 err = is_state_visited(env, env->insn_idx);
17019 if (err < 0)
17020 return err;
17021 if (err == 1) {
17022 /* found equivalent state, can prune the search */
17023 if (env->log.level & BPF_LOG_LEVEL) {
17024 if (do_print_state)
17025 verbose(env, "\nfrom %d to %d%s: safe\n",
17026 env->prev_insn_idx, env->insn_idx,
17027 env->cur_state->speculative ?
17028 " (speculative execution)" : "");
17029 else
17030 verbose(env, "%d: safe\n", env->insn_idx);
17031 }
17032 goto process_bpf_exit;
17033 }
17034 }
17035
17036 if (is_jmp_point(env, env->insn_idx)) {
17037 err = push_jmp_history(env, state);
17038 if (err)
17039 return err;
17040 }
17041
17042 if (signal_pending(current))
17043 return -EAGAIN;
17044
17045 if (need_resched())
17046 cond_resched();
17047
17048 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
17049 verbose(env, "\nfrom %d to %d%s:",
17050 env->prev_insn_idx, env->insn_idx,
17051 env->cur_state->speculative ?
17052 " (speculative execution)" : "");
17053 print_verifier_state(env, state->frame[state->curframe], true);
17054 do_print_state = false;
17055 }
17056
17057 if (env->log.level & BPF_LOG_LEVEL) {
17058 const struct bpf_insn_cbs cbs = {
17059 .cb_call = disasm_kfunc_name,
17060 .cb_print = verbose,
17061 .private_data = env,
17062 };
17063
17064 if (verifier_state_scratched(env))
17065 print_insn_state(env, state->frame[state->curframe]);
17066
17067 verbose_linfo(env, env->insn_idx, "; ");
17068 env->prev_log_pos = env->log.end_pos;
17069 verbose(env, "%d: ", env->insn_idx);
17070 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
17071 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
17072 env->prev_log_pos = env->log.end_pos;
17073 }
17074
17075 if (bpf_prog_is_offloaded(env->prog->aux)) {
17076 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
17077 env->prev_insn_idx);
17078 if (err)
17079 return err;
17080 }
17081
17082 regs = cur_regs(env);
17083 sanitize_mark_insn_seen(env);
17084 prev_insn_idx = env->insn_idx;
17085
17086 if (class == BPF_ALU || class == BPF_ALU64) {
17087 err = check_alu_op(env, insn);
17088 if (err)
17089 return err;
17090
17091 } else if (class == BPF_LDX) {
17092 enum bpf_reg_type src_reg_type;
17093
17094 /* check for reserved fields is already done */
17095
17096 /* check src operand */
17097 err = check_reg_arg(env, insn->src_reg, SRC_OP);
17098 if (err)
17099 return err;
17100
17101 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
17102 if (err)
17103 return err;
17104
17105 src_reg_type = regs[insn->src_reg].type;
17106
17107 /* check that memory (src_reg + off) is readable,
17108 * the state of dst_reg will be updated by this func
17109 */
17110 err = check_mem_access(env, env->insn_idx, insn->src_reg,
17111 insn->off, BPF_SIZE(insn->code),
17112 BPF_READ, insn->dst_reg, false,
17113 BPF_MODE(insn->code) == BPF_MEMSX);
17114 if (err)
17115 return err;
17116
17117 err = save_aux_ptr_type(env, src_reg_type, true);
17118 if (err)
17119 return err;
17120 } else if (class == BPF_STX) {
17121 enum bpf_reg_type dst_reg_type;
17122
17123 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
17124 err = check_atomic(env, env->insn_idx, insn);
17125 if (err)
17126 return err;
17127 env->insn_idx++;
17128 continue;
17129 }
17130
17131 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
17132 verbose(env, "BPF_STX uses reserved fields\n");
17133 return -EINVAL;
17134 }
17135
17136 /* check src1 operand */
17137 err = check_reg_arg(env, insn->src_reg, SRC_OP);
17138 if (err)
17139 return err;
17140 /* check src2 operand */
17141 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
17142 if (err)
17143 return err;
17144
17145 dst_reg_type = regs[insn->dst_reg].type;
17146
17147 /* check that memory (dst_reg + off) is writeable */
17148 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
17149 insn->off, BPF_SIZE(insn->code),
17150 BPF_WRITE, insn->src_reg, false, false);
17151 if (err)
17152 return err;
17153
17154 err = save_aux_ptr_type(env, dst_reg_type, false);
17155 if (err)
17156 return err;
17157 } else if (class == BPF_ST) {
17158 enum bpf_reg_type dst_reg_type;
17159
17160 if (BPF_MODE(insn->code) != BPF_MEM ||
17161 insn->src_reg != BPF_REG_0) {
17162 verbose(env, "BPF_ST uses reserved fields\n");
17163 return -EINVAL;
17164 }
17165 /* check src operand */
17166 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
17167 if (err)
17168 return err;
17169
17170 dst_reg_type = regs[insn->dst_reg].type;
17171
17172 /* check that memory (dst_reg + off) is writeable */
17173 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
17174 insn->off, BPF_SIZE(insn->code),
17175 BPF_WRITE, -1, false, false);
17176 if (err)
17177 return err;
17178
17179 err = save_aux_ptr_type(env, dst_reg_type, false);
17180 if (err)
17181 return err;
17182 } else if (class == BPF_JMP || class == BPF_JMP32) {
17183 u8 opcode = BPF_OP(insn->code);
17184
17185 env->jmps_processed++;
17186 if (opcode == BPF_CALL) {
17187 if (BPF_SRC(insn->code) != BPF_K ||
17188 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
17189 && insn->off != 0) ||
17190 (insn->src_reg != BPF_REG_0 &&
17191 insn->src_reg != BPF_PSEUDO_CALL &&
17192 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
17193 insn->dst_reg != BPF_REG_0 ||
17194 class == BPF_JMP32) {
17195 verbose(env, "BPF_CALL uses reserved fields\n");
17196 return -EINVAL;
17197 }
17198
17199 if (env->cur_state->active_lock.ptr) {
17200 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) ||
17201 (insn->src_reg == BPF_PSEUDO_CALL) ||
17202 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
17203 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) {
17204 verbose(env, "function calls are not allowed while holding a lock\n");
17205 return -EINVAL;
17206 }
17207 }
17208 if (insn->src_reg == BPF_PSEUDO_CALL)
17209 err = check_func_call(env, insn, &env->insn_idx);
17210 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
17211 err = check_kfunc_call(env, insn, &env->insn_idx);
17212 else
17213 err = check_helper_call(env, insn, &env->insn_idx);
17214 if (err)
17215 return err;
17216
17217 mark_reg_scratched(env, BPF_REG_0);
17218 } else if (opcode == BPF_JA) {
17219 if (BPF_SRC(insn->code) != BPF_K ||
17220 insn->src_reg != BPF_REG_0 ||
17221 insn->dst_reg != BPF_REG_0 ||
17222 (class == BPF_JMP && insn->imm != 0) ||
17223 (class == BPF_JMP32 && insn->off != 0)) {
17224 verbose(env, "BPF_JA uses reserved fields\n");
17225 return -EINVAL;
17226 }
17227
17228 if (class == BPF_JMP)
17229 env->insn_idx += insn->off + 1;
17230 else
17231 env->insn_idx += insn->imm + 1;
17232 continue;
17233
17234 } else if (opcode == BPF_EXIT) {
17235 if (BPF_SRC(insn->code) != BPF_K ||
17236 insn->imm != 0 ||
17237 insn->src_reg != BPF_REG_0 ||
17238 insn->dst_reg != BPF_REG_0 ||
17239 class == BPF_JMP32) {
17240 verbose(env, "BPF_EXIT uses reserved fields\n");
17241 return -EINVAL;
17242 }
17243
17244 if (env->cur_state->active_lock.ptr &&
17245 !in_rbtree_lock_required_cb(env)) {
17246 verbose(env, "bpf_spin_unlock is missing\n");
17247 return -EINVAL;
17248 }
17249
17250 if (env->cur_state->active_rcu_lock &&
17251 !in_rbtree_lock_required_cb(env)) {
17252 verbose(env, "bpf_rcu_read_unlock is missing\n");
17253 return -EINVAL;
17254 }
17255
17256 /* We must do check_reference_leak here before
17257 * prepare_func_exit to handle the case when
17258 * state->curframe > 0, it may be a callback
17259 * function, for which reference_state must
17260 * match caller reference state when it exits.
17261 */
17262 err = check_reference_leak(env);
17263 if (err)
17264 return err;
17265
17266 if (state->curframe) {
17267 /* exit from nested function */
17268 err = prepare_func_exit(env, &env->insn_idx);
17269 if (err)
17270 return err;
17271 do_print_state = true;
17272 continue;
17273 }
17274
17275 err = check_return_code(env);
17276 if (err)
17277 return err;
17278 process_bpf_exit:
17279 mark_verifier_state_scratched(env);
17280 update_branch_counts(env, env->cur_state);
17281 err = pop_stack(env, &prev_insn_idx,
17282 &env->insn_idx, pop_log);
17283 if (err < 0) {
17284 if (err != -ENOENT)
17285 return err;
17286 break;
17287 } else {
17288 do_print_state = true;
17289 continue;
17290 }
17291 } else {
17292 err = check_cond_jmp_op(env, insn, &env->insn_idx);
17293 if (err)
17294 return err;
17295 }
17296 } else if (class == BPF_LD) {
17297 u8 mode = BPF_MODE(insn->code);
17298
17299 if (mode == BPF_ABS || mode == BPF_IND) {
17300 err = check_ld_abs(env, insn);
17301 if (err)
17302 return err;
17303
17304 } else if (mode == BPF_IMM) {
17305 err = check_ld_imm(env, insn);
17306 if (err)
17307 return err;
17308
17309 env->insn_idx++;
17310 sanitize_mark_insn_seen(env);
17311 } else {
17312 verbose(env, "invalid BPF_LD mode\n");
17313 return -EINVAL;
17314 }
17315 } else {
17316 verbose(env, "unknown insn class %d\n", class);
17317 return -EINVAL;
17318 }
17319
17320 env->insn_idx++;
17321 }
17322
17323 return 0;
17324 }
17325
find_btf_percpu_datasec(struct btf * btf)17326 static int find_btf_percpu_datasec(struct btf *btf)
17327 {
17328 const struct btf_type *t;
17329 const char *tname;
17330 int i, n;
17331
17332 /*
17333 * Both vmlinux and module each have their own ".data..percpu"
17334 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
17335 * types to look at only module's own BTF types.
17336 */
17337 n = btf_nr_types(btf);
17338 if (btf_is_module(btf))
17339 i = btf_nr_types(btf_vmlinux);
17340 else
17341 i = 1;
17342
17343 for(; i < n; i++) {
17344 t = btf_type_by_id(btf, i);
17345 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
17346 continue;
17347
17348 tname = btf_name_by_offset(btf, t->name_off);
17349 if (!strcmp(tname, ".data..percpu"))
17350 return i;
17351 }
17352
17353 return -ENOENT;
17354 }
17355
17356 /* 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)17357 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
17358 struct bpf_insn *insn,
17359 struct bpf_insn_aux_data *aux)
17360 {
17361 const struct btf_var_secinfo *vsi;
17362 const struct btf_type *datasec;
17363 struct btf_mod_pair *btf_mod;
17364 const struct btf_type *t;
17365 const char *sym_name;
17366 bool percpu = false;
17367 u32 type, id = insn->imm;
17368 struct btf *btf;
17369 s32 datasec_id;
17370 u64 addr;
17371 int i, btf_fd, err;
17372
17373 btf_fd = insn[1].imm;
17374 if (btf_fd) {
17375 btf = btf_get_by_fd(btf_fd);
17376 if (IS_ERR(btf)) {
17377 verbose(env, "invalid module BTF object FD specified.\n");
17378 return -EINVAL;
17379 }
17380 } else {
17381 if (!btf_vmlinux) {
17382 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
17383 return -EINVAL;
17384 }
17385 btf = btf_vmlinux;
17386 btf_get(btf);
17387 }
17388
17389 t = btf_type_by_id(btf, id);
17390 if (!t) {
17391 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
17392 err = -ENOENT;
17393 goto err_put;
17394 }
17395
17396 if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
17397 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
17398 err = -EINVAL;
17399 goto err_put;
17400 }
17401
17402 sym_name = btf_name_by_offset(btf, t->name_off);
17403 addr = kallsyms_lookup_name(sym_name);
17404 if (!addr) {
17405 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
17406 sym_name);
17407 err = -ENOENT;
17408 goto err_put;
17409 }
17410 insn[0].imm = (u32)addr;
17411 insn[1].imm = addr >> 32;
17412
17413 if (btf_type_is_func(t)) {
17414 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
17415 aux->btf_var.mem_size = 0;
17416 goto check_btf;
17417 }
17418
17419 datasec_id = find_btf_percpu_datasec(btf);
17420 if (datasec_id > 0) {
17421 datasec = btf_type_by_id(btf, datasec_id);
17422 for_each_vsi(i, datasec, vsi) {
17423 if (vsi->type == id) {
17424 percpu = true;
17425 break;
17426 }
17427 }
17428 }
17429
17430 type = t->type;
17431 t = btf_type_skip_modifiers(btf, type, NULL);
17432 if (percpu) {
17433 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
17434 aux->btf_var.btf = btf;
17435 aux->btf_var.btf_id = type;
17436 } else if (!btf_type_is_struct(t)) {
17437 const struct btf_type *ret;
17438 const char *tname;
17439 u32 tsize;
17440
17441 /* resolve the type size of ksym. */
17442 ret = btf_resolve_size(btf, t, &tsize);
17443 if (IS_ERR(ret)) {
17444 tname = btf_name_by_offset(btf, t->name_off);
17445 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
17446 tname, PTR_ERR(ret));
17447 err = -EINVAL;
17448 goto err_put;
17449 }
17450 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
17451 aux->btf_var.mem_size = tsize;
17452 } else {
17453 aux->btf_var.reg_type = PTR_TO_BTF_ID;
17454 aux->btf_var.btf = btf;
17455 aux->btf_var.btf_id = type;
17456 }
17457 check_btf:
17458 /* check whether we recorded this BTF (and maybe module) already */
17459 for (i = 0; i < env->used_btf_cnt; i++) {
17460 if (env->used_btfs[i].btf == btf) {
17461 btf_put(btf);
17462 return 0;
17463 }
17464 }
17465
17466 if (env->used_btf_cnt >= MAX_USED_BTFS) {
17467 err = -E2BIG;
17468 goto err_put;
17469 }
17470
17471 btf_mod = &env->used_btfs[env->used_btf_cnt];
17472 btf_mod->btf = btf;
17473 btf_mod->module = NULL;
17474
17475 /* if we reference variables from kernel module, bump its refcount */
17476 if (btf_is_module(btf)) {
17477 btf_mod->module = btf_try_get_module(btf);
17478 if (!btf_mod->module) {
17479 err = -ENXIO;
17480 goto err_put;
17481 }
17482 }
17483
17484 env->used_btf_cnt++;
17485
17486 return 0;
17487 err_put:
17488 btf_put(btf);
17489 return err;
17490 }
17491
is_tracing_prog_type(enum bpf_prog_type type)17492 static bool is_tracing_prog_type(enum bpf_prog_type type)
17493 {
17494 switch (type) {
17495 case BPF_PROG_TYPE_KPROBE:
17496 case BPF_PROG_TYPE_TRACEPOINT:
17497 case BPF_PROG_TYPE_PERF_EVENT:
17498 case BPF_PROG_TYPE_RAW_TRACEPOINT:
17499 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
17500 return true;
17501 default:
17502 return false;
17503 }
17504 }
17505
check_map_prog_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,struct bpf_prog * prog)17506 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
17507 struct bpf_map *map,
17508 struct bpf_prog *prog)
17509
17510 {
17511 enum bpf_prog_type prog_type = resolve_prog_type(prog);
17512
17513 if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
17514 btf_record_has_field(map->record, BPF_RB_ROOT)) {
17515 if (is_tracing_prog_type(prog_type)) {
17516 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
17517 return -EINVAL;
17518 }
17519 }
17520
17521 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) {
17522 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
17523 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
17524 return -EINVAL;
17525 }
17526
17527 if (is_tracing_prog_type(prog_type)) {
17528 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
17529 return -EINVAL;
17530 }
17531 }
17532
17533 if (btf_record_has_field(map->record, BPF_TIMER)) {
17534 if (is_tracing_prog_type(prog_type)) {
17535 verbose(env, "tracing progs cannot use bpf_timer yet\n");
17536 return -EINVAL;
17537 }
17538 }
17539
17540 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
17541 !bpf_offload_prog_map_match(prog, map)) {
17542 verbose(env, "offload device mismatch between prog and map\n");
17543 return -EINVAL;
17544 }
17545
17546 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
17547 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
17548 return -EINVAL;
17549 }
17550
17551 if (prog->aux->sleepable)
17552 switch (map->map_type) {
17553 case BPF_MAP_TYPE_HASH:
17554 case BPF_MAP_TYPE_LRU_HASH:
17555 case BPF_MAP_TYPE_ARRAY:
17556 case BPF_MAP_TYPE_PERCPU_HASH:
17557 case BPF_MAP_TYPE_PERCPU_ARRAY:
17558 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
17559 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
17560 case BPF_MAP_TYPE_HASH_OF_MAPS:
17561 case BPF_MAP_TYPE_RINGBUF:
17562 case BPF_MAP_TYPE_USER_RINGBUF:
17563 case BPF_MAP_TYPE_INODE_STORAGE:
17564 case BPF_MAP_TYPE_SK_STORAGE:
17565 case BPF_MAP_TYPE_TASK_STORAGE:
17566 case BPF_MAP_TYPE_CGRP_STORAGE:
17567 break;
17568 default:
17569 verbose(env,
17570 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
17571 return -EINVAL;
17572 }
17573
17574 return 0;
17575 }
17576
bpf_map_is_cgroup_storage(struct bpf_map * map)17577 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
17578 {
17579 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
17580 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
17581 }
17582
17583 /* find and rewrite pseudo imm in ld_imm64 instructions:
17584 *
17585 * 1. if it accesses map FD, replace it with actual map pointer.
17586 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
17587 *
17588 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
17589 */
resolve_pseudo_ldimm64(struct bpf_verifier_env * env)17590 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
17591 {
17592 struct bpf_insn *insn = env->prog->insnsi;
17593 int insn_cnt = env->prog->len;
17594 int i, j, err;
17595
17596 err = bpf_prog_calc_tag(env->prog);
17597 if (err)
17598 return err;
17599
17600 for (i = 0; i < insn_cnt; i++, insn++) {
17601 if (BPF_CLASS(insn->code) == BPF_LDX &&
17602 ((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
17603 insn->imm != 0)) {
17604 verbose(env, "BPF_LDX uses reserved fields\n");
17605 return -EINVAL;
17606 }
17607
17608 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
17609 struct bpf_insn_aux_data *aux;
17610 struct bpf_map *map;
17611 struct fd f;
17612 u64 addr;
17613 u32 fd;
17614
17615 if (i == insn_cnt - 1 || insn[1].code != 0 ||
17616 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
17617 insn[1].off != 0) {
17618 verbose(env, "invalid bpf_ld_imm64 insn\n");
17619 return -EINVAL;
17620 }
17621
17622 if (insn[0].src_reg == 0)
17623 /* valid generic load 64-bit imm */
17624 goto next_insn;
17625
17626 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
17627 aux = &env->insn_aux_data[i];
17628 err = check_pseudo_btf_id(env, insn, aux);
17629 if (err)
17630 return err;
17631 goto next_insn;
17632 }
17633
17634 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
17635 aux = &env->insn_aux_data[i];
17636 aux->ptr_type = PTR_TO_FUNC;
17637 goto next_insn;
17638 }
17639
17640 /* In final convert_pseudo_ld_imm64() step, this is
17641 * converted into regular 64-bit imm load insn.
17642 */
17643 switch (insn[0].src_reg) {
17644 case BPF_PSEUDO_MAP_VALUE:
17645 case BPF_PSEUDO_MAP_IDX_VALUE:
17646 break;
17647 case BPF_PSEUDO_MAP_FD:
17648 case BPF_PSEUDO_MAP_IDX:
17649 if (insn[1].imm == 0)
17650 break;
17651 fallthrough;
17652 default:
17653 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
17654 return -EINVAL;
17655 }
17656
17657 switch (insn[0].src_reg) {
17658 case BPF_PSEUDO_MAP_IDX_VALUE:
17659 case BPF_PSEUDO_MAP_IDX:
17660 if (bpfptr_is_null(env->fd_array)) {
17661 verbose(env, "fd_idx without fd_array is invalid\n");
17662 return -EPROTO;
17663 }
17664 if (copy_from_bpfptr_offset(&fd, env->fd_array,
17665 insn[0].imm * sizeof(fd),
17666 sizeof(fd)))
17667 return -EFAULT;
17668 break;
17669 default:
17670 fd = insn[0].imm;
17671 break;
17672 }
17673
17674 f = fdget(fd);
17675 map = __bpf_map_get(f);
17676 if (IS_ERR(map)) {
17677 verbose(env, "fd %d is not pointing to valid bpf_map\n", fd);
17678 return PTR_ERR(map);
17679 }
17680
17681 err = check_map_prog_compatibility(env, map, env->prog);
17682 if (err) {
17683 fdput(f);
17684 return err;
17685 }
17686
17687 aux = &env->insn_aux_data[i];
17688 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
17689 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
17690 addr = (unsigned long)map;
17691 } else {
17692 u32 off = insn[1].imm;
17693
17694 if (off >= BPF_MAX_VAR_OFF) {
17695 verbose(env, "direct value offset of %u is not allowed\n", off);
17696 fdput(f);
17697 return -EINVAL;
17698 }
17699
17700 if (!map->ops->map_direct_value_addr) {
17701 verbose(env, "no direct value access support for this map type\n");
17702 fdput(f);
17703 return -EINVAL;
17704 }
17705
17706 err = map->ops->map_direct_value_addr(map, &addr, off);
17707 if (err) {
17708 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
17709 map->value_size, off);
17710 fdput(f);
17711 return err;
17712 }
17713
17714 aux->map_off = off;
17715 addr += off;
17716 }
17717
17718 insn[0].imm = (u32)addr;
17719 insn[1].imm = addr >> 32;
17720
17721 /* check whether we recorded this map already */
17722 for (j = 0; j < env->used_map_cnt; j++) {
17723 if (env->used_maps[j] == map) {
17724 aux->map_index = j;
17725 fdput(f);
17726 goto next_insn;
17727 }
17728 }
17729
17730 if (env->used_map_cnt >= MAX_USED_MAPS) {
17731 fdput(f);
17732 return -E2BIG;
17733 }
17734
17735 if (env->prog->aux->sleepable)
17736 atomic64_inc(&map->sleepable_refcnt);
17737 /* hold the map. If the program is rejected by verifier,
17738 * the map will be released by release_maps() or it
17739 * will be used by the valid program until it's unloaded
17740 * and all maps are released in bpf_free_used_maps()
17741 */
17742 bpf_map_inc(map);
17743
17744 aux->map_index = env->used_map_cnt;
17745 env->used_maps[env->used_map_cnt++] = map;
17746
17747 if (bpf_map_is_cgroup_storage(map) &&
17748 bpf_cgroup_storage_assign(env->prog->aux, map)) {
17749 verbose(env, "only one cgroup storage of each type is allowed\n");
17750 fdput(f);
17751 return -EBUSY;
17752 }
17753
17754 fdput(f);
17755 next_insn:
17756 insn++;
17757 i++;
17758 continue;
17759 }
17760
17761 /* Basic sanity check before we invest more work here. */
17762 if (!bpf_opcode_in_insntable(insn->code)) {
17763 verbose(env, "unknown opcode %02x\n", insn->code);
17764 return -EINVAL;
17765 }
17766 }
17767
17768 /* now all pseudo BPF_LD_IMM64 instructions load valid
17769 * 'struct bpf_map *' into a register instead of user map_fd.
17770 * These pointers will be used later by verifier to validate map access.
17771 */
17772 return 0;
17773 }
17774
17775 /* drop refcnt of maps used by the rejected program */
release_maps(struct bpf_verifier_env * env)17776 static void release_maps(struct bpf_verifier_env *env)
17777 {
17778 __bpf_free_used_maps(env->prog->aux, env->used_maps,
17779 env->used_map_cnt);
17780 }
17781
17782 /* drop refcnt of maps used by the rejected program */
release_btfs(struct bpf_verifier_env * env)17783 static void release_btfs(struct bpf_verifier_env *env)
17784 {
17785 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
17786 env->used_btf_cnt);
17787 }
17788
17789 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
convert_pseudo_ld_imm64(struct bpf_verifier_env * env)17790 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
17791 {
17792 struct bpf_insn *insn = env->prog->insnsi;
17793 int insn_cnt = env->prog->len;
17794 int i;
17795
17796 for (i = 0; i < insn_cnt; i++, insn++) {
17797 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
17798 continue;
17799 if (insn->src_reg == BPF_PSEUDO_FUNC)
17800 continue;
17801 insn->src_reg = 0;
17802 }
17803 }
17804
17805 /* single env->prog->insni[off] instruction was replaced with the range
17806 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
17807 * [0, off) and [off, end) to new locations, so the patched range stays zero
17808 */
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)17809 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
17810 struct bpf_insn_aux_data *new_data,
17811 struct bpf_prog *new_prog, u32 off, u32 cnt)
17812 {
17813 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
17814 struct bpf_insn *insn = new_prog->insnsi;
17815 u32 old_seen = old_data[off].seen;
17816 u32 prog_len;
17817 int i;
17818
17819 /* aux info at OFF always needs adjustment, no matter fast path
17820 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
17821 * original insn at old prog.
17822 */
17823 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
17824
17825 if (cnt == 1)
17826 return;
17827 prog_len = new_prog->len;
17828
17829 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
17830 memcpy(new_data + off + cnt - 1, old_data + off,
17831 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
17832 for (i = off; i < off + cnt - 1; i++) {
17833 /* Expand insni[off]'s seen count to the patched range. */
17834 new_data[i].seen = old_seen;
17835 new_data[i].zext_dst = insn_has_def32(env, insn + i);
17836 }
17837 env->insn_aux_data = new_data;
17838 vfree(old_data);
17839 }
17840
adjust_subprog_starts(struct bpf_verifier_env * env,u32 off,u32 len)17841 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
17842 {
17843 int i;
17844
17845 if (len == 1)
17846 return;
17847 /* NOTE: fake 'exit' subprog should be updated as well. */
17848 for (i = 0; i <= env->subprog_cnt; i++) {
17849 if (env->subprog_info[i].start <= off)
17850 continue;
17851 env->subprog_info[i].start += len - 1;
17852 }
17853 }
17854
adjust_poke_descs(struct bpf_prog * prog,u32 off,u32 len)17855 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
17856 {
17857 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
17858 int i, sz = prog->aux->size_poke_tab;
17859 struct bpf_jit_poke_descriptor *desc;
17860
17861 for (i = 0; i < sz; i++) {
17862 desc = &tab[i];
17863 if (desc->insn_idx <= off)
17864 continue;
17865 desc->insn_idx += len - 1;
17866 }
17867 }
17868
bpf_patch_insn_data(struct bpf_verifier_env * env,u32 off,const struct bpf_insn * patch,u32 len)17869 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
17870 const struct bpf_insn *patch, u32 len)
17871 {
17872 struct bpf_prog *new_prog;
17873 struct bpf_insn_aux_data *new_data = NULL;
17874
17875 if (len > 1) {
17876 new_data = vzalloc(array_size(env->prog->len + len - 1,
17877 sizeof(struct bpf_insn_aux_data)));
17878 if (!new_data)
17879 return NULL;
17880 }
17881
17882 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
17883 if (IS_ERR(new_prog)) {
17884 if (PTR_ERR(new_prog) == -ERANGE)
17885 verbose(env,
17886 "insn %d cannot be patched due to 16-bit range\n",
17887 env->insn_aux_data[off].orig_idx);
17888 vfree(new_data);
17889 return NULL;
17890 }
17891 adjust_insn_aux_data(env, new_data, new_prog, off, len);
17892 adjust_subprog_starts(env, off, len);
17893 adjust_poke_descs(new_prog, off, len);
17894 return new_prog;
17895 }
17896
adjust_subprog_starts_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)17897 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
17898 u32 off, u32 cnt)
17899 {
17900 int i, j;
17901
17902 /* find first prog starting at or after off (first to remove) */
17903 for (i = 0; i < env->subprog_cnt; i++)
17904 if (env->subprog_info[i].start >= off)
17905 break;
17906 /* find first prog starting at or after off + cnt (first to stay) */
17907 for (j = i; j < env->subprog_cnt; j++)
17908 if (env->subprog_info[j].start >= off + cnt)
17909 break;
17910 /* if j doesn't start exactly at off + cnt, we are just removing
17911 * the front of previous prog
17912 */
17913 if (env->subprog_info[j].start != off + cnt)
17914 j--;
17915
17916 if (j > i) {
17917 struct bpf_prog_aux *aux = env->prog->aux;
17918 int move;
17919
17920 /* move fake 'exit' subprog as well */
17921 move = env->subprog_cnt + 1 - j;
17922
17923 memmove(env->subprog_info + i,
17924 env->subprog_info + j,
17925 sizeof(*env->subprog_info) * move);
17926 env->subprog_cnt -= j - i;
17927
17928 /* remove func_info */
17929 if (aux->func_info) {
17930 move = aux->func_info_cnt - j;
17931
17932 memmove(aux->func_info + i,
17933 aux->func_info + j,
17934 sizeof(*aux->func_info) * move);
17935 aux->func_info_cnt -= j - i;
17936 /* func_info->insn_off is set after all code rewrites,
17937 * in adjust_btf_func() - no need to adjust
17938 */
17939 }
17940 } else {
17941 /* convert i from "first prog to remove" to "first to adjust" */
17942 if (env->subprog_info[i].start == off)
17943 i++;
17944 }
17945
17946 /* update fake 'exit' subprog as well */
17947 for (; i <= env->subprog_cnt; i++)
17948 env->subprog_info[i].start -= cnt;
17949
17950 return 0;
17951 }
17952
bpf_adj_linfo_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)17953 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
17954 u32 cnt)
17955 {
17956 struct bpf_prog *prog = env->prog;
17957 u32 i, l_off, l_cnt, nr_linfo;
17958 struct bpf_line_info *linfo;
17959
17960 nr_linfo = prog->aux->nr_linfo;
17961 if (!nr_linfo)
17962 return 0;
17963
17964 linfo = prog->aux->linfo;
17965
17966 /* find first line info to remove, count lines to be removed */
17967 for (i = 0; i < nr_linfo; i++)
17968 if (linfo[i].insn_off >= off)
17969 break;
17970
17971 l_off = i;
17972 l_cnt = 0;
17973 for (; i < nr_linfo; i++)
17974 if (linfo[i].insn_off < off + cnt)
17975 l_cnt++;
17976 else
17977 break;
17978
17979 /* First live insn doesn't match first live linfo, it needs to "inherit"
17980 * last removed linfo. prog is already modified, so prog->len == off
17981 * means no live instructions after (tail of the program was removed).
17982 */
17983 if (prog->len != off && l_cnt &&
17984 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
17985 l_cnt--;
17986 linfo[--i].insn_off = off + cnt;
17987 }
17988
17989 /* remove the line info which refer to the removed instructions */
17990 if (l_cnt) {
17991 memmove(linfo + l_off, linfo + i,
17992 sizeof(*linfo) * (nr_linfo - i));
17993
17994 prog->aux->nr_linfo -= l_cnt;
17995 nr_linfo = prog->aux->nr_linfo;
17996 }
17997
17998 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
17999 for (i = l_off; i < nr_linfo; i++)
18000 linfo[i].insn_off -= cnt;
18001
18002 /* fix up all subprogs (incl. 'exit') which start >= off */
18003 for (i = 0; i <= env->subprog_cnt; i++)
18004 if (env->subprog_info[i].linfo_idx > l_off) {
18005 /* program may have started in the removed region but
18006 * may not be fully removed
18007 */
18008 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
18009 env->subprog_info[i].linfo_idx -= l_cnt;
18010 else
18011 env->subprog_info[i].linfo_idx = l_off;
18012 }
18013
18014 return 0;
18015 }
18016
verifier_remove_insns(struct bpf_verifier_env * env,u32 off,u32 cnt)18017 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
18018 {
18019 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18020 unsigned int orig_prog_len = env->prog->len;
18021 int err;
18022
18023 if (bpf_prog_is_offloaded(env->prog->aux))
18024 bpf_prog_offload_remove_insns(env, off, cnt);
18025
18026 err = bpf_remove_insns(env->prog, off, cnt);
18027 if (err)
18028 return err;
18029
18030 err = adjust_subprog_starts_after_remove(env, off, cnt);
18031 if (err)
18032 return err;
18033
18034 err = bpf_adj_linfo_after_remove(env, off, cnt);
18035 if (err)
18036 return err;
18037
18038 memmove(aux_data + off, aux_data + off + cnt,
18039 sizeof(*aux_data) * (orig_prog_len - off - cnt));
18040
18041 return 0;
18042 }
18043
18044 /* The verifier does more data flow analysis than llvm and will not
18045 * explore branches that are dead at run time. Malicious programs can
18046 * have dead code too. Therefore replace all dead at-run-time code
18047 * with 'ja -1'.
18048 *
18049 * Just nops are not optimal, e.g. if they would sit at the end of the
18050 * program and through another bug we would manage to jump there, then
18051 * we'd execute beyond program memory otherwise. Returning exception
18052 * code also wouldn't work since we can have subprogs where the dead
18053 * code could be located.
18054 */
sanitize_dead_code(struct bpf_verifier_env * env)18055 static void sanitize_dead_code(struct bpf_verifier_env *env)
18056 {
18057 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18058 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
18059 struct bpf_insn *insn = env->prog->insnsi;
18060 const int insn_cnt = env->prog->len;
18061 int i;
18062
18063 for (i = 0; i < insn_cnt; i++) {
18064 if (aux_data[i].seen)
18065 continue;
18066 memcpy(insn + i, &trap, sizeof(trap));
18067 aux_data[i].zext_dst = false;
18068 }
18069 }
18070
insn_is_cond_jump(u8 code)18071 static bool insn_is_cond_jump(u8 code)
18072 {
18073 u8 op;
18074
18075 op = BPF_OP(code);
18076 if (BPF_CLASS(code) == BPF_JMP32)
18077 return op != BPF_JA;
18078
18079 if (BPF_CLASS(code) != BPF_JMP)
18080 return false;
18081
18082 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
18083 }
18084
opt_hard_wire_dead_code_branches(struct bpf_verifier_env * env)18085 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
18086 {
18087 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18088 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
18089 struct bpf_insn *insn = env->prog->insnsi;
18090 const int insn_cnt = env->prog->len;
18091 int i;
18092
18093 for (i = 0; i < insn_cnt; i++, insn++) {
18094 if (!insn_is_cond_jump(insn->code))
18095 continue;
18096
18097 if (!aux_data[i + 1].seen)
18098 ja.off = insn->off;
18099 else if (!aux_data[i + 1 + insn->off].seen)
18100 ja.off = 0;
18101 else
18102 continue;
18103
18104 if (bpf_prog_is_offloaded(env->prog->aux))
18105 bpf_prog_offload_replace_insn(env, i, &ja);
18106
18107 memcpy(insn, &ja, sizeof(ja));
18108 }
18109 }
18110
opt_remove_dead_code(struct bpf_verifier_env * env)18111 static int opt_remove_dead_code(struct bpf_verifier_env *env)
18112 {
18113 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
18114 int insn_cnt = env->prog->len;
18115 int i, err;
18116
18117 for (i = 0; i < insn_cnt; i++) {
18118 int j;
18119
18120 j = 0;
18121 while (i + j < insn_cnt && !aux_data[i + j].seen)
18122 j++;
18123 if (!j)
18124 continue;
18125
18126 err = verifier_remove_insns(env, i, j);
18127 if (err)
18128 return err;
18129 insn_cnt = env->prog->len;
18130 }
18131
18132 return 0;
18133 }
18134
opt_remove_nops(struct bpf_verifier_env * env)18135 static int opt_remove_nops(struct bpf_verifier_env *env)
18136 {
18137 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
18138 struct bpf_insn *insn = env->prog->insnsi;
18139 int insn_cnt = env->prog->len;
18140 int i, err;
18141
18142 for (i = 0; i < insn_cnt; i++) {
18143 if (memcmp(&insn[i], &ja, sizeof(ja)))
18144 continue;
18145
18146 err = verifier_remove_insns(env, i, 1);
18147 if (err)
18148 return err;
18149 insn_cnt--;
18150 i--;
18151 }
18152
18153 return 0;
18154 }
18155
opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env * env,const union bpf_attr * attr)18156 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
18157 const union bpf_attr *attr)
18158 {
18159 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
18160 struct bpf_insn_aux_data *aux = env->insn_aux_data;
18161 int i, patch_len, delta = 0, len = env->prog->len;
18162 struct bpf_insn *insns = env->prog->insnsi;
18163 struct bpf_prog *new_prog;
18164 bool rnd_hi32;
18165
18166 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
18167 zext_patch[1] = BPF_ZEXT_REG(0);
18168 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
18169 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
18170 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
18171 for (i = 0; i < len; i++) {
18172 int adj_idx = i + delta;
18173 struct bpf_insn insn;
18174 int load_reg;
18175
18176 insn = insns[adj_idx];
18177 load_reg = insn_def_regno(&insn);
18178 if (!aux[adj_idx].zext_dst) {
18179 u8 code, class;
18180 u32 imm_rnd;
18181
18182 if (!rnd_hi32)
18183 continue;
18184
18185 code = insn.code;
18186 class = BPF_CLASS(code);
18187 if (load_reg == -1)
18188 continue;
18189
18190 /* NOTE: arg "reg" (the fourth one) is only used for
18191 * BPF_STX + SRC_OP, so it is safe to pass NULL
18192 * here.
18193 */
18194 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
18195 if (class == BPF_LD &&
18196 BPF_MODE(code) == BPF_IMM)
18197 i++;
18198 continue;
18199 }
18200
18201 /* ctx load could be transformed into wider load. */
18202 if (class == BPF_LDX &&
18203 aux[adj_idx].ptr_type == PTR_TO_CTX)
18204 continue;
18205
18206 imm_rnd = get_random_u32();
18207 rnd_hi32_patch[0] = insn;
18208 rnd_hi32_patch[1].imm = imm_rnd;
18209 rnd_hi32_patch[3].dst_reg = load_reg;
18210 patch = rnd_hi32_patch;
18211 patch_len = 4;
18212 goto apply_patch_buffer;
18213 }
18214
18215 /* Add in an zero-extend instruction if a) the JIT has requested
18216 * it or b) it's a CMPXCHG.
18217 *
18218 * The latter is because: BPF_CMPXCHG always loads a value into
18219 * R0, therefore always zero-extends. However some archs'
18220 * equivalent instruction only does this load when the
18221 * comparison is successful. This detail of CMPXCHG is
18222 * orthogonal to the general zero-extension behaviour of the
18223 * CPU, so it's treated independently of bpf_jit_needs_zext.
18224 */
18225 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
18226 continue;
18227
18228 /* Zero-extension is done by the caller. */
18229 if (bpf_pseudo_kfunc_call(&insn))
18230 continue;
18231
18232 if (WARN_ON(load_reg == -1)) {
18233 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
18234 return -EFAULT;
18235 }
18236
18237 zext_patch[0] = insn;
18238 zext_patch[1].dst_reg = load_reg;
18239 zext_patch[1].src_reg = load_reg;
18240 patch = zext_patch;
18241 patch_len = 2;
18242 apply_patch_buffer:
18243 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
18244 if (!new_prog)
18245 return -ENOMEM;
18246 env->prog = new_prog;
18247 insns = new_prog->insnsi;
18248 aux = env->insn_aux_data;
18249 delta += patch_len - 1;
18250 }
18251
18252 return 0;
18253 }
18254
18255 /* convert load instructions that access fields of a context type into a
18256 * sequence of instructions that access fields of the underlying structure:
18257 * struct __sk_buff -> struct sk_buff
18258 * struct bpf_sock_ops -> struct sock
18259 */
convert_ctx_accesses(struct bpf_verifier_env * env)18260 static int convert_ctx_accesses(struct bpf_verifier_env *env)
18261 {
18262 const struct bpf_verifier_ops *ops = env->ops;
18263 int i, cnt, size, ctx_field_size, delta = 0;
18264 const int insn_cnt = env->prog->len;
18265 struct bpf_insn insn_buf[16], *insn;
18266 u32 target_size, size_default, off;
18267 struct bpf_prog *new_prog;
18268 enum bpf_access_type type;
18269 bool is_narrower_load;
18270
18271 if (ops->gen_prologue || env->seen_direct_write) {
18272 if (!ops->gen_prologue) {
18273 verbose(env, "bpf verifier is misconfigured\n");
18274 return -EINVAL;
18275 }
18276 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
18277 env->prog);
18278 if (cnt >= ARRAY_SIZE(insn_buf)) {
18279 verbose(env, "bpf verifier is misconfigured\n");
18280 return -EINVAL;
18281 } else if (cnt) {
18282 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
18283 if (!new_prog)
18284 return -ENOMEM;
18285
18286 env->prog = new_prog;
18287 delta += cnt - 1;
18288 }
18289 }
18290
18291 if (bpf_prog_is_offloaded(env->prog->aux))
18292 return 0;
18293
18294 insn = env->prog->insnsi + delta;
18295
18296 for (i = 0; i < insn_cnt; i++, insn++) {
18297 bpf_convert_ctx_access_t convert_ctx_access;
18298 u8 mode;
18299
18300 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
18301 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
18302 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
18303 insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
18304 insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
18305 insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
18306 insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
18307 type = BPF_READ;
18308 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
18309 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
18310 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
18311 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
18312 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
18313 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
18314 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
18315 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
18316 type = BPF_WRITE;
18317 } else {
18318 continue;
18319 }
18320
18321 if (type == BPF_WRITE &&
18322 env->insn_aux_data[i + delta].sanitize_stack_spill) {
18323 struct bpf_insn patch[] = {
18324 *insn,
18325 BPF_ST_NOSPEC(),
18326 };
18327
18328 cnt = ARRAY_SIZE(patch);
18329 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
18330 if (!new_prog)
18331 return -ENOMEM;
18332
18333 delta += cnt - 1;
18334 env->prog = new_prog;
18335 insn = new_prog->insnsi + i + delta;
18336 continue;
18337 }
18338
18339 switch ((int)env->insn_aux_data[i + delta].ptr_type) {
18340 case PTR_TO_CTX:
18341 if (!ops->convert_ctx_access)
18342 continue;
18343 convert_ctx_access = ops->convert_ctx_access;
18344 break;
18345 case PTR_TO_SOCKET:
18346 case PTR_TO_SOCK_COMMON:
18347 convert_ctx_access = bpf_sock_convert_ctx_access;
18348 break;
18349 case PTR_TO_TCP_SOCK:
18350 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
18351 break;
18352 case PTR_TO_XDP_SOCK:
18353 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
18354 break;
18355 case PTR_TO_BTF_ID:
18356 case PTR_TO_BTF_ID | PTR_UNTRUSTED:
18357 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
18358 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
18359 * be said once it is marked PTR_UNTRUSTED, hence we must handle
18360 * any faults for loads into such types. BPF_WRITE is disallowed
18361 * for this case.
18362 */
18363 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
18364 if (type == BPF_READ) {
18365 if (BPF_MODE(insn->code) == BPF_MEM)
18366 insn->code = BPF_LDX | BPF_PROBE_MEM |
18367 BPF_SIZE((insn)->code);
18368 else
18369 insn->code = BPF_LDX | BPF_PROBE_MEMSX |
18370 BPF_SIZE((insn)->code);
18371 env->prog->aux->num_exentries++;
18372 }
18373 continue;
18374 default:
18375 continue;
18376 }
18377
18378 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
18379 size = BPF_LDST_BYTES(insn);
18380 mode = BPF_MODE(insn->code);
18381
18382 /* If the read access is a narrower load of the field,
18383 * convert to a 4/8-byte load, to minimum program type specific
18384 * convert_ctx_access changes. If conversion is successful,
18385 * we will apply proper mask to the result.
18386 */
18387 is_narrower_load = size < ctx_field_size;
18388 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
18389 off = insn->off;
18390 if (is_narrower_load) {
18391 u8 size_code;
18392
18393 if (type == BPF_WRITE) {
18394 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
18395 return -EINVAL;
18396 }
18397
18398 size_code = BPF_H;
18399 if (ctx_field_size == 4)
18400 size_code = BPF_W;
18401 else if (ctx_field_size == 8)
18402 size_code = BPF_DW;
18403
18404 insn->off = off & ~(size_default - 1);
18405 insn->code = BPF_LDX | BPF_MEM | size_code;
18406 }
18407
18408 target_size = 0;
18409 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
18410 &target_size);
18411 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
18412 (ctx_field_size && !target_size)) {
18413 verbose(env, "bpf verifier is misconfigured\n");
18414 return -EINVAL;
18415 }
18416
18417 if (is_narrower_load && size < target_size) {
18418 u8 shift = bpf_ctx_narrow_access_offset(
18419 off, size, size_default) * 8;
18420 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
18421 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
18422 return -EINVAL;
18423 }
18424 if (ctx_field_size <= 4) {
18425 if (shift)
18426 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
18427 insn->dst_reg,
18428 shift);
18429 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
18430 (1 << size * 8) - 1);
18431 } else {
18432 if (shift)
18433 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
18434 insn->dst_reg,
18435 shift);
18436 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
18437 (1ULL << size * 8) - 1);
18438 }
18439 }
18440 if (mode == BPF_MEMSX)
18441 insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
18442 insn->dst_reg, insn->dst_reg,
18443 size * 8, 0);
18444
18445 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18446 if (!new_prog)
18447 return -ENOMEM;
18448
18449 delta += cnt - 1;
18450
18451 /* keep walking new program and skip insns we just inserted */
18452 env->prog = new_prog;
18453 insn = new_prog->insnsi + i + delta;
18454 }
18455
18456 return 0;
18457 }
18458
jit_subprogs(struct bpf_verifier_env * env)18459 static int jit_subprogs(struct bpf_verifier_env *env)
18460 {
18461 struct bpf_prog *prog = env->prog, **func, *tmp;
18462 int i, j, subprog_start, subprog_end = 0, len, subprog;
18463 struct bpf_map *map_ptr;
18464 struct bpf_insn *insn;
18465 void *old_bpf_func;
18466 int err, num_exentries;
18467
18468 if (env->subprog_cnt <= 1)
18469 return 0;
18470
18471 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18472 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
18473 continue;
18474
18475 /* Upon error here we cannot fall back to interpreter but
18476 * need a hard reject of the program. Thus -EFAULT is
18477 * propagated in any case.
18478 */
18479 subprog = find_subprog(env, i + insn->imm + 1);
18480 if (subprog < 0) {
18481 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
18482 i + insn->imm + 1);
18483 return -EFAULT;
18484 }
18485 /* temporarily remember subprog id inside insn instead of
18486 * aux_data, since next loop will split up all insns into funcs
18487 */
18488 insn->off = subprog;
18489 /* remember original imm in case JIT fails and fallback
18490 * to interpreter will be needed
18491 */
18492 env->insn_aux_data[i].call_imm = insn->imm;
18493 /* point imm to __bpf_call_base+1 from JITs point of view */
18494 insn->imm = 1;
18495 if (bpf_pseudo_func(insn))
18496 /* jit (e.g. x86_64) may emit fewer instructions
18497 * if it learns a u32 imm is the same as a u64 imm.
18498 * Force a non zero here.
18499 */
18500 insn[1].imm = 1;
18501 }
18502
18503 err = bpf_prog_alloc_jited_linfo(prog);
18504 if (err)
18505 goto out_undo_insn;
18506
18507 err = -ENOMEM;
18508 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
18509 if (!func)
18510 goto out_undo_insn;
18511
18512 for (i = 0; i < env->subprog_cnt; i++) {
18513 subprog_start = subprog_end;
18514 subprog_end = env->subprog_info[i + 1].start;
18515
18516 len = subprog_end - subprog_start;
18517 /* bpf_prog_run() doesn't call subprogs directly,
18518 * hence main prog stats include the runtime of subprogs.
18519 * subprogs don't have IDs and not reachable via prog_get_next_id
18520 * func[i]->stats will never be accessed and stays NULL
18521 */
18522 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
18523 if (!func[i])
18524 goto out_free;
18525 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
18526 len * sizeof(struct bpf_insn));
18527 func[i]->type = prog->type;
18528 func[i]->len = len;
18529 if (bpf_prog_calc_tag(func[i]))
18530 goto out_free;
18531 func[i]->is_func = 1;
18532 func[i]->aux->func_idx = i;
18533 /* Below members will be freed only at prog->aux */
18534 func[i]->aux->btf = prog->aux->btf;
18535 func[i]->aux->func_info = prog->aux->func_info;
18536 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
18537 func[i]->aux->poke_tab = prog->aux->poke_tab;
18538 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
18539
18540 for (j = 0; j < prog->aux->size_poke_tab; j++) {
18541 struct bpf_jit_poke_descriptor *poke;
18542
18543 poke = &prog->aux->poke_tab[j];
18544 if (poke->insn_idx < subprog_end &&
18545 poke->insn_idx >= subprog_start)
18546 poke->aux = func[i]->aux;
18547 }
18548
18549 func[i]->aux->name[0] = 'F';
18550 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
18551 func[i]->jit_requested = 1;
18552 func[i]->blinding_requested = prog->blinding_requested;
18553 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
18554 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
18555 func[i]->aux->linfo = prog->aux->linfo;
18556 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
18557 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
18558 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
18559 num_exentries = 0;
18560 insn = func[i]->insnsi;
18561 for (j = 0; j < func[i]->len; j++, insn++) {
18562 if (BPF_CLASS(insn->code) == BPF_LDX &&
18563 (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
18564 BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
18565 num_exentries++;
18566 }
18567 func[i]->aux->num_exentries = num_exentries;
18568 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
18569 func[i] = bpf_int_jit_compile(func[i]);
18570 if (!func[i]->jited) {
18571 err = -ENOTSUPP;
18572 goto out_free;
18573 }
18574 cond_resched();
18575 }
18576
18577 /* at this point all bpf functions were successfully JITed
18578 * now populate all bpf_calls with correct addresses and
18579 * run last pass of JIT
18580 */
18581 for (i = 0; i < env->subprog_cnt; i++) {
18582 insn = func[i]->insnsi;
18583 for (j = 0; j < func[i]->len; j++, insn++) {
18584 if (bpf_pseudo_func(insn)) {
18585 subprog = insn->off;
18586 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
18587 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
18588 continue;
18589 }
18590 if (!bpf_pseudo_call(insn))
18591 continue;
18592 subprog = insn->off;
18593 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
18594 }
18595
18596 /* we use the aux data to keep a list of the start addresses
18597 * of the JITed images for each function in the program
18598 *
18599 * for some architectures, such as powerpc64, the imm field
18600 * might not be large enough to hold the offset of the start
18601 * address of the callee's JITed image from __bpf_call_base
18602 *
18603 * in such cases, we can lookup the start address of a callee
18604 * by using its subprog id, available from the off field of
18605 * the call instruction, as an index for this list
18606 */
18607 func[i]->aux->func = func;
18608 func[i]->aux->func_cnt = env->subprog_cnt;
18609 }
18610 for (i = 0; i < env->subprog_cnt; i++) {
18611 old_bpf_func = func[i]->bpf_func;
18612 tmp = bpf_int_jit_compile(func[i]);
18613 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
18614 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
18615 err = -ENOTSUPP;
18616 goto out_free;
18617 }
18618 cond_resched();
18619 }
18620
18621 /* finally lock prog and jit images for all functions and
18622 * populate kallsysm. Begin at the first subprogram, since
18623 * bpf_prog_load will add the kallsyms for the main program.
18624 */
18625 for (i = 1; i < env->subprog_cnt; i++) {
18626 bpf_prog_lock_ro(func[i]);
18627 bpf_prog_kallsyms_add(func[i]);
18628 }
18629
18630 /* Last step: make now unused interpreter insns from main
18631 * prog consistent for later dump requests, so they can
18632 * later look the same as if they were interpreted only.
18633 */
18634 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18635 if (bpf_pseudo_func(insn)) {
18636 insn[0].imm = env->insn_aux_data[i].call_imm;
18637 insn[1].imm = insn->off;
18638 insn->off = 0;
18639 continue;
18640 }
18641 if (!bpf_pseudo_call(insn))
18642 continue;
18643 insn->off = env->insn_aux_data[i].call_imm;
18644 subprog = find_subprog(env, i + insn->off + 1);
18645 insn->imm = subprog;
18646 }
18647
18648 prog->jited = 1;
18649 prog->bpf_func = func[0]->bpf_func;
18650 prog->jited_len = func[0]->jited_len;
18651 prog->aux->extable = func[0]->aux->extable;
18652 prog->aux->num_exentries = func[0]->aux->num_exentries;
18653 prog->aux->func = func;
18654 prog->aux->func_cnt = env->subprog_cnt;
18655 bpf_prog_jit_attempt_done(prog);
18656 return 0;
18657 out_free:
18658 /* We failed JIT'ing, so at this point we need to unregister poke
18659 * descriptors from subprogs, so that kernel is not attempting to
18660 * patch it anymore as we're freeing the subprog JIT memory.
18661 */
18662 for (i = 0; i < prog->aux->size_poke_tab; i++) {
18663 map_ptr = prog->aux->poke_tab[i].tail_call.map;
18664 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
18665 }
18666 /* At this point we're guaranteed that poke descriptors are not
18667 * live anymore. We can just unlink its descriptor table as it's
18668 * released with the main prog.
18669 */
18670 for (i = 0; i < env->subprog_cnt; i++) {
18671 if (!func[i])
18672 continue;
18673 func[i]->aux->poke_tab = NULL;
18674 bpf_jit_free(func[i]);
18675 }
18676 kfree(func);
18677 out_undo_insn:
18678 /* cleanup main prog to be interpreted */
18679 prog->jit_requested = 0;
18680 prog->blinding_requested = 0;
18681 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
18682 if (!bpf_pseudo_call(insn))
18683 continue;
18684 insn->off = 0;
18685 insn->imm = env->insn_aux_data[i].call_imm;
18686 }
18687 bpf_prog_jit_attempt_done(prog);
18688 return err;
18689 }
18690
fixup_call_args(struct bpf_verifier_env * env)18691 static int fixup_call_args(struct bpf_verifier_env *env)
18692 {
18693 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18694 struct bpf_prog *prog = env->prog;
18695 struct bpf_insn *insn = prog->insnsi;
18696 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
18697 int i, depth;
18698 #endif
18699 int err = 0;
18700
18701 if (env->prog->jit_requested &&
18702 !bpf_prog_is_offloaded(env->prog->aux)) {
18703 err = jit_subprogs(env);
18704 if (err == 0)
18705 return 0;
18706 if (err == -EFAULT)
18707 return err;
18708 }
18709 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
18710 if (has_kfunc_call) {
18711 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
18712 return -EINVAL;
18713 }
18714 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
18715 /* When JIT fails the progs with bpf2bpf calls and tail_calls
18716 * have to be rejected, since interpreter doesn't support them yet.
18717 */
18718 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
18719 return -EINVAL;
18720 }
18721 for (i = 0; i < prog->len; i++, insn++) {
18722 if (bpf_pseudo_func(insn)) {
18723 /* When JIT fails the progs with callback calls
18724 * have to be rejected, since interpreter doesn't support them yet.
18725 */
18726 verbose(env, "callbacks are not allowed in non-JITed programs\n");
18727 return -EINVAL;
18728 }
18729
18730 if (!bpf_pseudo_call(insn))
18731 continue;
18732 depth = get_callee_stack_depth(env, insn, i);
18733 if (depth < 0)
18734 return depth;
18735 bpf_patch_call_args(insn, depth);
18736 }
18737 err = 0;
18738 #endif
18739 return err;
18740 }
18741
18742 /* 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)18743 static void specialize_kfunc(struct bpf_verifier_env *env,
18744 u32 func_id, u16 offset, unsigned long *addr)
18745 {
18746 struct bpf_prog *prog = env->prog;
18747 bool seen_direct_write;
18748 void *xdp_kfunc;
18749 bool is_rdonly;
18750
18751 if (bpf_dev_bound_kfunc_id(func_id)) {
18752 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
18753 if (xdp_kfunc) {
18754 *addr = (unsigned long)xdp_kfunc;
18755 return;
18756 }
18757 /* fallback to default kfunc when not supported by netdev */
18758 }
18759
18760 if (offset)
18761 return;
18762
18763 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
18764 seen_direct_write = env->seen_direct_write;
18765 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
18766
18767 if (is_rdonly)
18768 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
18769
18770 /* restore env->seen_direct_write to its original value, since
18771 * may_access_direct_pkt_data mutates it
18772 */
18773 env->seen_direct_write = seen_direct_write;
18774 }
18775 }
18776
__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)18777 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
18778 u16 struct_meta_reg,
18779 u16 node_offset_reg,
18780 struct bpf_insn *insn,
18781 struct bpf_insn *insn_buf,
18782 int *cnt)
18783 {
18784 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
18785 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
18786
18787 insn_buf[0] = addr[0];
18788 insn_buf[1] = addr[1];
18789 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
18790 insn_buf[3] = *insn;
18791 *cnt = 4;
18792 }
18793
fixup_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_insn * insn_buf,int insn_idx,int * cnt)18794 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
18795 struct bpf_insn *insn_buf, int insn_idx, int *cnt)
18796 {
18797 const struct bpf_kfunc_desc *desc;
18798
18799 if (!insn->imm) {
18800 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
18801 return -EINVAL;
18802 }
18803
18804 *cnt = 0;
18805
18806 /* insn->imm has the btf func_id. Replace it with an offset relative to
18807 * __bpf_call_base, unless the JIT needs to call functions that are
18808 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
18809 */
18810 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
18811 if (!desc) {
18812 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
18813 insn->imm);
18814 return -EFAULT;
18815 }
18816
18817 if (!bpf_jit_supports_far_kfunc_call())
18818 insn->imm = BPF_CALL_IMM(desc->addr);
18819 if (insn->off)
18820 return 0;
18821 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) {
18822 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18823 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18824 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
18825
18826 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
18827 insn_buf[1] = addr[0];
18828 insn_buf[2] = addr[1];
18829 insn_buf[3] = *insn;
18830 *cnt = 4;
18831 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
18832 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
18833 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18834 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
18835
18836 if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
18837 !kptr_struct_meta) {
18838 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18839 insn_idx);
18840 return -EFAULT;
18841 }
18842
18843 insn_buf[0] = addr[0];
18844 insn_buf[1] = addr[1];
18845 insn_buf[2] = *insn;
18846 *cnt = 3;
18847 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
18848 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
18849 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18850 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
18851 int struct_meta_reg = BPF_REG_3;
18852 int node_offset_reg = BPF_REG_4;
18853
18854 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
18855 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
18856 struct_meta_reg = BPF_REG_4;
18857 node_offset_reg = BPF_REG_5;
18858 }
18859
18860 if (!kptr_struct_meta) {
18861 verbose(env, "verifier internal error: kptr_struct_meta expected at insn_idx %d\n",
18862 insn_idx);
18863 return -EFAULT;
18864 }
18865
18866 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
18867 node_offset_reg, insn, insn_buf, cnt);
18868 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
18869 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
18870 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
18871 *cnt = 1;
18872 }
18873 return 0;
18874 }
18875
18876 /* Do various post-verification rewrites in a single program pass.
18877 * These rewrites simplify JIT and interpreter implementations.
18878 */
do_misc_fixups(struct bpf_verifier_env * env)18879 static int do_misc_fixups(struct bpf_verifier_env *env)
18880 {
18881 struct bpf_prog *prog = env->prog;
18882 enum bpf_attach_type eatype = prog->expected_attach_type;
18883 enum bpf_prog_type prog_type = resolve_prog_type(prog);
18884 struct bpf_insn *insn = prog->insnsi;
18885 const struct bpf_func_proto *fn;
18886 const int insn_cnt = prog->len;
18887 const struct bpf_map_ops *ops;
18888 struct bpf_insn_aux_data *aux;
18889 struct bpf_insn insn_buf[16];
18890 struct bpf_prog *new_prog;
18891 struct bpf_map *map_ptr;
18892 int i, ret, cnt, delta = 0;
18893
18894 for (i = 0; i < insn_cnt; i++, insn++) {
18895 /* Make divide-by-zero exceptions impossible. */
18896 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
18897 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
18898 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
18899 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
18900 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
18901 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
18902 struct bpf_insn *patchlet;
18903 struct bpf_insn chk_and_div[] = {
18904 /* [R,W]x div 0 -> 0 */
18905 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18906 BPF_JNE | BPF_K, insn->src_reg,
18907 0, 2, 0),
18908 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
18909 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18910 *insn,
18911 };
18912 struct bpf_insn chk_and_mod[] = {
18913 /* [R,W]x mod 0 -> [R,W]x */
18914 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
18915 BPF_JEQ | BPF_K, insn->src_reg,
18916 0, 1 + (is64 ? 0 : 1), 0),
18917 *insn,
18918 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
18919 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
18920 };
18921
18922 patchlet = isdiv ? chk_and_div : chk_and_mod;
18923 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
18924 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
18925
18926 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
18927 if (!new_prog)
18928 return -ENOMEM;
18929
18930 delta += cnt - 1;
18931 env->prog = prog = new_prog;
18932 insn = new_prog->insnsi + i + delta;
18933 continue;
18934 }
18935
18936 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
18937 if (BPF_CLASS(insn->code) == BPF_LD &&
18938 (BPF_MODE(insn->code) == BPF_ABS ||
18939 BPF_MODE(insn->code) == BPF_IND)) {
18940 cnt = env->ops->gen_ld_abs(insn, insn_buf);
18941 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
18942 verbose(env, "bpf verifier is misconfigured\n");
18943 return -EINVAL;
18944 }
18945
18946 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
18947 if (!new_prog)
18948 return -ENOMEM;
18949
18950 delta += cnt - 1;
18951 env->prog = prog = new_prog;
18952 insn = new_prog->insnsi + i + delta;
18953 continue;
18954 }
18955
18956 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
18957 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
18958 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
18959 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
18960 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
18961 struct bpf_insn *patch = &insn_buf[0];
18962 bool issrc, isneg, isimm;
18963 u32 off_reg;
18964
18965 aux = &env->insn_aux_data[i + delta];
18966 if (!aux->alu_state ||
18967 aux->alu_state == BPF_ALU_NON_POINTER)
18968 continue;
18969
18970 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
18971 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
18972 BPF_ALU_SANITIZE_SRC;
18973 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
18974
18975 off_reg = issrc ? insn->src_reg : insn->dst_reg;
18976 if (isimm) {
18977 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18978 } else {
18979 if (isneg)
18980 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18981 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
18982 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
18983 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
18984 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
18985 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
18986 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
18987 }
18988 if (!issrc)
18989 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
18990 insn->src_reg = BPF_REG_AX;
18991 if (isneg)
18992 insn->code = insn->code == code_add ?
18993 code_sub : code_add;
18994 *patch++ = *insn;
18995 if (issrc && isneg && !isimm)
18996 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
18997 cnt = patch - insn_buf;
18998
18999 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19000 if (!new_prog)
19001 return -ENOMEM;
19002
19003 delta += cnt - 1;
19004 env->prog = prog = new_prog;
19005 insn = new_prog->insnsi + i + delta;
19006 continue;
19007 }
19008
19009 if (insn->code != (BPF_JMP | BPF_CALL))
19010 continue;
19011 if (insn->src_reg == BPF_PSEUDO_CALL)
19012 continue;
19013 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
19014 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
19015 if (ret)
19016 return ret;
19017 if (cnt == 0)
19018 continue;
19019
19020 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19021 if (!new_prog)
19022 return -ENOMEM;
19023
19024 delta += cnt - 1;
19025 env->prog = prog = new_prog;
19026 insn = new_prog->insnsi + i + delta;
19027 continue;
19028 }
19029
19030 if (insn->imm == BPF_FUNC_get_route_realm)
19031 prog->dst_needed = 1;
19032 if (insn->imm == BPF_FUNC_get_prandom_u32)
19033 bpf_user_rnd_init_once();
19034 if (insn->imm == BPF_FUNC_override_return)
19035 prog->kprobe_override = 1;
19036 if (insn->imm == BPF_FUNC_tail_call) {
19037 /* If we tail call into other programs, we
19038 * cannot make any assumptions since they can
19039 * be replaced dynamically during runtime in
19040 * the program array.
19041 */
19042 prog->cb_access = 1;
19043 if (!allow_tail_call_in_subprogs(env))
19044 prog->aux->stack_depth = MAX_BPF_STACK;
19045 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
19046
19047 /* mark bpf_tail_call as different opcode to avoid
19048 * conditional branch in the interpreter for every normal
19049 * call and to prevent accidental JITing by JIT compiler
19050 * that doesn't support bpf_tail_call yet
19051 */
19052 insn->imm = 0;
19053 insn->code = BPF_JMP | BPF_TAIL_CALL;
19054
19055 aux = &env->insn_aux_data[i + delta];
19056 if (env->bpf_capable && !prog->blinding_requested &&
19057 prog->jit_requested &&
19058 !bpf_map_key_poisoned(aux) &&
19059 !bpf_map_ptr_poisoned(aux) &&
19060 !bpf_map_ptr_unpriv(aux)) {
19061 struct bpf_jit_poke_descriptor desc = {
19062 .reason = BPF_POKE_REASON_TAIL_CALL,
19063 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
19064 .tail_call.key = bpf_map_key_immediate(aux),
19065 .insn_idx = i + delta,
19066 };
19067
19068 ret = bpf_jit_add_poke_descriptor(prog, &desc);
19069 if (ret < 0) {
19070 verbose(env, "adding tail call poke descriptor failed\n");
19071 return ret;
19072 }
19073
19074 insn->imm = ret + 1;
19075 continue;
19076 }
19077
19078 if (!bpf_map_ptr_unpriv(aux))
19079 continue;
19080
19081 /* instead of changing every JIT dealing with tail_call
19082 * emit two extra insns:
19083 * if (index >= max_entries) goto out;
19084 * index &= array->index_mask;
19085 * to avoid out-of-bounds cpu speculation
19086 */
19087 if (bpf_map_ptr_poisoned(aux)) {
19088 verbose(env, "tail_call abusing map_ptr\n");
19089 return -EINVAL;
19090 }
19091
19092 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
19093 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
19094 map_ptr->max_entries, 2);
19095 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
19096 container_of(map_ptr,
19097 struct bpf_array,
19098 map)->index_mask);
19099 insn_buf[2] = *insn;
19100 cnt = 3;
19101 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19102 if (!new_prog)
19103 return -ENOMEM;
19104
19105 delta += cnt - 1;
19106 env->prog = prog = new_prog;
19107 insn = new_prog->insnsi + i + delta;
19108 continue;
19109 }
19110
19111 if (insn->imm == BPF_FUNC_timer_set_callback) {
19112 /* The verifier will process callback_fn as many times as necessary
19113 * with different maps and the register states prepared by
19114 * set_timer_callback_state will be accurate.
19115 *
19116 * The following use case is valid:
19117 * map1 is shared by prog1, prog2, prog3.
19118 * prog1 calls bpf_timer_init for some map1 elements
19119 * prog2 calls bpf_timer_set_callback for some map1 elements.
19120 * Those that were not bpf_timer_init-ed will return -EINVAL.
19121 * prog3 calls bpf_timer_start for some map1 elements.
19122 * Those that were not both bpf_timer_init-ed and
19123 * bpf_timer_set_callback-ed will return -EINVAL.
19124 */
19125 struct bpf_insn ld_addrs[2] = {
19126 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
19127 };
19128
19129 insn_buf[0] = ld_addrs[0];
19130 insn_buf[1] = ld_addrs[1];
19131 insn_buf[2] = *insn;
19132 cnt = 3;
19133
19134 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19135 if (!new_prog)
19136 return -ENOMEM;
19137
19138 delta += cnt - 1;
19139 env->prog = prog = new_prog;
19140 insn = new_prog->insnsi + i + delta;
19141 goto patch_call_imm;
19142 }
19143
19144 if (is_storage_get_function(insn->imm)) {
19145 if (!env->prog->aux->sleepable ||
19146 env->insn_aux_data[i + delta].storage_get_func_atomic)
19147 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
19148 else
19149 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
19150 insn_buf[1] = *insn;
19151 cnt = 2;
19152
19153 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19154 if (!new_prog)
19155 return -ENOMEM;
19156
19157 delta += cnt - 1;
19158 env->prog = prog = new_prog;
19159 insn = new_prog->insnsi + i + delta;
19160 goto patch_call_imm;
19161 }
19162
19163 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
19164 * and other inlining handlers are currently limited to 64 bit
19165 * only.
19166 */
19167 if (prog->jit_requested && BITS_PER_LONG == 64 &&
19168 (insn->imm == BPF_FUNC_map_lookup_elem ||
19169 insn->imm == BPF_FUNC_map_update_elem ||
19170 insn->imm == BPF_FUNC_map_delete_elem ||
19171 insn->imm == BPF_FUNC_map_push_elem ||
19172 insn->imm == BPF_FUNC_map_pop_elem ||
19173 insn->imm == BPF_FUNC_map_peek_elem ||
19174 insn->imm == BPF_FUNC_redirect_map ||
19175 insn->imm == BPF_FUNC_for_each_map_elem ||
19176 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
19177 aux = &env->insn_aux_data[i + delta];
19178 if (bpf_map_ptr_poisoned(aux))
19179 goto patch_call_imm;
19180
19181 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
19182 ops = map_ptr->ops;
19183 if (insn->imm == BPF_FUNC_map_lookup_elem &&
19184 ops->map_gen_lookup) {
19185 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
19186 if (cnt == -EOPNOTSUPP)
19187 goto patch_map_ops_generic;
19188 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
19189 verbose(env, "bpf verifier is misconfigured\n");
19190 return -EINVAL;
19191 }
19192
19193 new_prog = bpf_patch_insn_data(env, i + delta,
19194 insn_buf, cnt);
19195 if (!new_prog)
19196 return -ENOMEM;
19197
19198 delta += cnt - 1;
19199 env->prog = prog = new_prog;
19200 insn = new_prog->insnsi + i + delta;
19201 continue;
19202 }
19203
19204 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
19205 (void *(*)(struct bpf_map *map, void *key))NULL));
19206 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
19207 (long (*)(struct bpf_map *map, void *key))NULL));
19208 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
19209 (long (*)(struct bpf_map *map, void *key, void *value,
19210 u64 flags))NULL));
19211 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
19212 (long (*)(struct bpf_map *map, void *value,
19213 u64 flags))NULL));
19214 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
19215 (long (*)(struct bpf_map *map, void *value))NULL));
19216 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
19217 (long (*)(struct bpf_map *map, void *value))NULL));
19218 BUILD_BUG_ON(!__same_type(ops->map_redirect,
19219 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
19220 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
19221 (long (*)(struct bpf_map *map,
19222 bpf_callback_t callback_fn,
19223 void *callback_ctx,
19224 u64 flags))NULL));
19225 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
19226 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
19227
19228 patch_map_ops_generic:
19229 switch (insn->imm) {
19230 case BPF_FUNC_map_lookup_elem:
19231 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
19232 continue;
19233 case BPF_FUNC_map_update_elem:
19234 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
19235 continue;
19236 case BPF_FUNC_map_delete_elem:
19237 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
19238 continue;
19239 case BPF_FUNC_map_push_elem:
19240 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
19241 continue;
19242 case BPF_FUNC_map_pop_elem:
19243 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
19244 continue;
19245 case BPF_FUNC_map_peek_elem:
19246 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
19247 continue;
19248 case BPF_FUNC_redirect_map:
19249 insn->imm = BPF_CALL_IMM(ops->map_redirect);
19250 continue;
19251 case BPF_FUNC_for_each_map_elem:
19252 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
19253 continue;
19254 case BPF_FUNC_map_lookup_percpu_elem:
19255 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
19256 continue;
19257 }
19258
19259 goto patch_call_imm;
19260 }
19261
19262 /* Implement bpf_jiffies64 inline. */
19263 if (prog->jit_requested && BITS_PER_LONG == 64 &&
19264 insn->imm == BPF_FUNC_jiffies64) {
19265 struct bpf_insn ld_jiffies_addr[2] = {
19266 BPF_LD_IMM64(BPF_REG_0,
19267 (unsigned long)&jiffies),
19268 };
19269
19270 insn_buf[0] = ld_jiffies_addr[0];
19271 insn_buf[1] = ld_jiffies_addr[1];
19272 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
19273 BPF_REG_0, 0);
19274 cnt = 3;
19275
19276 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
19277 cnt);
19278 if (!new_prog)
19279 return -ENOMEM;
19280
19281 delta += cnt - 1;
19282 env->prog = prog = new_prog;
19283 insn = new_prog->insnsi + i + delta;
19284 continue;
19285 }
19286
19287 /* Implement bpf_get_func_arg inline. */
19288 if (prog_type == BPF_PROG_TYPE_TRACING &&
19289 insn->imm == BPF_FUNC_get_func_arg) {
19290 /* Load nr_args from ctx - 8 */
19291 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
19292 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
19293 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
19294 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
19295 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
19296 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
19297 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
19298 insn_buf[7] = BPF_JMP_A(1);
19299 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
19300 cnt = 9;
19301
19302 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19303 if (!new_prog)
19304 return -ENOMEM;
19305
19306 delta += cnt - 1;
19307 env->prog = prog = new_prog;
19308 insn = new_prog->insnsi + i + delta;
19309 continue;
19310 }
19311
19312 /* Implement bpf_get_func_ret inline. */
19313 if (prog_type == BPF_PROG_TYPE_TRACING &&
19314 insn->imm == BPF_FUNC_get_func_ret) {
19315 if (eatype == BPF_TRACE_FEXIT ||
19316 eatype == BPF_MODIFY_RETURN) {
19317 /* Load nr_args from ctx - 8 */
19318 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
19319 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
19320 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
19321 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
19322 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
19323 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
19324 cnt = 6;
19325 } else {
19326 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
19327 cnt = 1;
19328 }
19329
19330 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
19331 if (!new_prog)
19332 return -ENOMEM;
19333
19334 delta += cnt - 1;
19335 env->prog = prog = new_prog;
19336 insn = new_prog->insnsi + i + delta;
19337 continue;
19338 }
19339
19340 /* Implement get_func_arg_cnt inline. */
19341 if (prog_type == BPF_PROG_TYPE_TRACING &&
19342 insn->imm == BPF_FUNC_get_func_arg_cnt) {
19343 /* Load nr_args from ctx - 8 */
19344 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
19345
19346 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
19347 if (!new_prog)
19348 return -ENOMEM;
19349
19350 env->prog = prog = new_prog;
19351 insn = new_prog->insnsi + i + delta;
19352 continue;
19353 }
19354
19355 /* Implement bpf_get_func_ip inline. */
19356 if (prog_type == BPF_PROG_TYPE_TRACING &&
19357 insn->imm == BPF_FUNC_get_func_ip) {
19358 /* Load IP address from ctx - 16 */
19359 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
19360
19361 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
19362 if (!new_prog)
19363 return -ENOMEM;
19364
19365 env->prog = prog = new_prog;
19366 insn = new_prog->insnsi + i + delta;
19367 continue;
19368 }
19369
19370 patch_call_imm:
19371 fn = env->ops->get_func_proto(insn->imm, env->prog);
19372 /* all functions that have prototype and verifier allowed
19373 * programs to call them, must be real in-kernel functions
19374 */
19375 if (!fn->func) {
19376 verbose(env,
19377 "kernel subsystem misconfigured func %s#%d\n",
19378 func_id_name(insn->imm), insn->imm);
19379 return -EFAULT;
19380 }
19381 insn->imm = fn->func - __bpf_call_base;
19382 }
19383
19384 /* Since poke tab is now finalized, publish aux to tracker. */
19385 for (i = 0; i < prog->aux->size_poke_tab; i++) {
19386 map_ptr = prog->aux->poke_tab[i].tail_call.map;
19387 if (!map_ptr->ops->map_poke_track ||
19388 !map_ptr->ops->map_poke_untrack ||
19389 !map_ptr->ops->map_poke_run) {
19390 verbose(env, "bpf verifier is misconfigured\n");
19391 return -EINVAL;
19392 }
19393
19394 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
19395 if (ret < 0) {
19396 verbose(env, "tracking tail call prog failed\n");
19397 return ret;
19398 }
19399 }
19400
19401 sort_kfunc_descs_by_imm_off(env->prog);
19402
19403 return 0;
19404 }
19405
inline_bpf_loop(struct bpf_verifier_env * env,int position,s32 stack_base,u32 callback_subprogno,u32 * cnt)19406 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
19407 int position,
19408 s32 stack_base,
19409 u32 callback_subprogno,
19410 u32 *cnt)
19411 {
19412 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
19413 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
19414 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
19415 int reg_loop_max = BPF_REG_6;
19416 int reg_loop_cnt = BPF_REG_7;
19417 int reg_loop_ctx = BPF_REG_8;
19418
19419 struct bpf_prog *new_prog;
19420 u32 callback_start;
19421 u32 call_insn_offset;
19422 s32 callback_offset;
19423
19424 /* This represents an inlined version of bpf_iter.c:bpf_loop,
19425 * be careful to modify this code in sync.
19426 */
19427 struct bpf_insn insn_buf[] = {
19428 /* Return error and jump to the end of the patch if
19429 * expected number of iterations is too big.
19430 */
19431 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
19432 BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
19433 BPF_JMP_IMM(BPF_JA, 0, 0, 16),
19434 /* spill R6, R7, R8 to use these as loop vars */
19435 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
19436 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
19437 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
19438 /* initialize loop vars */
19439 BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
19440 BPF_MOV32_IMM(reg_loop_cnt, 0),
19441 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
19442 /* loop header,
19443 * if reg_loop_cnt >= reg_loop_max skip the loop body
19444 */
19445 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
19446 /* callback call,
19447 * correct callback offset would be set after patching
19448 */
19449 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
19450 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
19451 BPF_CALL_REL(0),
19452 /* increment loop counter */
19453 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
19454 /* jump to loop header if callback returned 0 */
19455 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
19456 /* return value of bpf_loop,
19457 * set R0 to the number of iterations
19458 */
19459 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
19460 /* restore original values of R6, R7, R8 */
19461 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
19462 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
19463 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
19464 };
19465
19466 *cnt = ARRAY_SIZE(insn_buf);
19467 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
19468 if (!new_prog)
19469 return new_prog;
19470
19471 /* callback start is known only after patching */
19472 callback_start = env->subprog_info[callback_subprogno].start;
19473 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
19474 call_insn_offset = position + 12;
19475 callback_offset = callback_start - call_insn_offset - 1;
19476 new_prog->insnsi[call_insn_offset].imm = callback_offset;
19477
19478 return new_prog;
19479 }
19480
is_bpf_loop_call(struct bpf_insn * insn)19481 static bool is_bpf_loop_call(struct bpf_insn *insn)
19482 {
19483 return insn->code == (BPF_JMP | BPF_CALL) &&
19484 insn->src_reg == 0 &&
19485 insn->imm == BPF_FUNC_loop;
19486 }
19487
19488 /* For all sub-programs in the program (including main) check
19489 * insn_aux_data to see if there are bpf_loop calls that require
19490 * inlining. If such calls are found the calls are replaced with a
19491 * sequence of instructions produced by `inline_bpf_loop` function and
19492 * subprog stack_depth is increased by the size of 3 registers.
19493 * This stack space is used to spill values of the R6, R7, R8. These
19494 * registers are used to store the loop bound, counter and context
19495 * variables.
19496 */
optimize_bpf_loop(struct bpf_verifier_env * env)19497 static int optimize_bpf_loop(struct bpf_verifier_env *env)
19498 {
19499 struct bpf_subprog_info *subprogs = env->subprog_info;
19500 int i, cur_subprog = 0, cnt, delta = 0;
19501 struct bpf_insn *insn = env->prog->insnsi;
19502 int insn_cnt = env->prog->len;
19503 u16 stack_depth = subprogs[cur_subprog].stack_depth;
19504 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
19505 u16 stack_depth_extra = 0;
19506
19507 for (i = 0; i < insn_cnt; i++, insn++) {
19508 struct bpf_loop_inline_state *inline_state =
19509 &env->insn_aux_data[i + delta].loop_inline_state;
19510
19511 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
19512 struct bpf_prog *new_prog;
19513
19514 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
19515 new_prog = inline_bpf_loop(env,
19516 i + delta,
19517 -(stack_depth + stack_depth_extra),
19518 inline_state->callback_subprogno,
19519 &cnt);
19520 if (!new_prog)
19521 return -ENOMEM;
19522
19523 delta += cnt - 1;
19524 env->prog = new_prog;
19525 insn = new_prog->insnsi + i + delta;
19526 }
19527
19528 if (subprogs[cur_subprog + 1].start == i + delta + 1) {
19529 subprogs[cur_subprog].stack_depth += stack_depth_extra;
19530 cur_subprog++;
19531 stack_depth = subprogs[cur_subprog].stack_depth;
19532 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
19533 stack_depth_extra = 0;
19534 }
19535 }
19536
19537 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19538
19539 return 0;
19540 }
19541
free_states(struct bpf_verifier_env * env)19542 static void free_states(struct bpf_verifier_env *env)
19543 {
19544 struct bpf_verifier_state_list *sl, *sln;
19545 int i;
19546
19547 sl = env->free_list;
19548 while (sl) {
19549 sln = sl->next;
19550 free_verifier_state(&sl->state, false);
19551 kfree(sl);
19552 sl = sln;
19553 }
19554 env->free_list = NULL;
19555
19556 if (!env->explored_states)
19557 return;
19558
19559 for (i = 0; i < state_htab_size(env); i++) {
19560 sl = env->explored_states[i];
19561
19562 while (sl) {
19563 sln = sl->next;
19564 free_verifier_state(&sl->state, false);
19565 kfree(sl);
19566 sl = sln;
19567 }
19568 env->explored_states[i] = NULL;
19569 }
19570 }
19571
do_check_common(struct bpf_verifier_env * env,int subprog)19572 static int do_check_common(struct bpf_verifier_env *env, int subprog)
19573 {
19574 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
19575 struct bpf_verifier_state *state;
19576 struct bpf_reg_state *regs;
19577 int ret, i;
19578
19579 env->prev_linfo = NULL;
19580 env->pass_cnt++;
19581
19582 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
19583 if (!state)
19584 return -ENOMEM;
19585 state->curframe = 0;
19586 state->speculative = false;
19587 state->branches = 1;
19588 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
19589 if (!state->frame[0]) {
19590 kfree(state);
19591 return -ENOMEM;
19592 }
19593 env->cur_state = state;
19594 init_func_state(env, state->frame[0],
19595 BPF_MAIN_FUNC /* callsite */,
19596 0 /* frameno */,
19597 subprog);
19598 state->first_insn_idx = env->subprog_info[subprog].start;
19599 state->last_insn_idx = -1;
19600
19601 regs = state->frame[state->curframe]->regs;
19602 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
19603 ret = btf_prepare_func_args(env, subprog, regs);
19604 if (ret)
19605 goto out;
19606 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
19607 if (regs[i].type == PTR_TO_CTX)
19608 mark_reg_known_zero(env, regs, i);
19609 else if (regs[i].type == SCALAR_VALUE)
19610 mark_reg_unknown(env, regs, i);
19611 else if (base_type(regs[i].type) == PTR_TO_MEM) {
19612 const u32 mem_size = regs[i].mem_size;
19613
19614 mark_reg_known_zero(env, regs, i);
19615 regs[i].mem_size = mem_size;
19616 regs[i].id = ++env->id_gen;
19617 }
19618 }
19619 } else {
19620 /* 1st arg to a function */
19621 regs[BPF_REG_1].type = PTR_TO_CTX;
19622 mark_reg_known_zero(env, regs, BPF_REG_1);
19623 ret = btf_check_subprog_arg_match(env, subprog, regs);
19624 if (ret == -EFAULT)
19625 /* unlikely verifier bug. abort.
19626 * ret == 0 and ret < 0 are sadly acceptable for
19627 * main() function due to backward compatibility.
19628 * Like socket filter program may be written as:
19629 * int bpf_prog(struct pt_regs *ctx)
19630 * and never dereference that ctx in the program.
19631 * 'struct pt_regs' is a type mismatch for socket
19632 * filter that should be using 'struct __sk_buff'.
19633 */
19634 goto out;
19635 }
19636
19637 ret = do_check(env);
19638 out:
19639 /* check for NULL is necessary, since cur_state can be freed inside
19640 * do_check() under memory pressure.
19641 */
19642 if (env->cur_state) {
19643 free_verifier_state(env->cur_state, true);
19644 env->cur_state = NULL;
19645 }
19646 while (!pop_stack(env, NULL, NULL, false));
19647 if (!ret && pop_log)
19648 bpf_vlog_reset(&env->log, 0);
19649 free_states(env);
19650 return ret;
19651 }
19652
19653 /* Verify all global functions in a BPF program one by one based on their BTF.
19654 * All global functions must pass verification. Otherwise the whole program is rejected.
19655 * Consider:
19656 * int bar(int);
19657 * int foo(int f)
19658 * {
19659 * return bar(f);
19660 * }
19661 * int bar(int b)
19662 * {
19663 * ...
19664 * }
19665 * foo() will be verified first for R1=any_scalar_value. During verification it
19666 * will be assumed that bar() already verified successfully and call to bar()
19667 * from foo() will be checked for type match only. Later bar() will be verified
19668 * independently to check that it's safe for R1=any_scalar_value.
19669 */
do_check_subprogs(struct bpf_verifier_env * env)19670 static int do_check_subprogs(struct bpf_verifier_env *env)
19671 {
19672 struct bpf_prog_aux *aux = env->prog->aux;
19673 int i, ret;
19674
19675 if (!aux->func_info)
19676 return 0;
19677
19678 for (i = 1; i < env->subprog_cnt; i++) {
19679 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
19680 continue;
19681 env->insn_idx = env->subprog_info[i].start;
19682 WARN_ON_ONCE(env->insn_idx == 0);
19683 ret = do_check_common(env, i);
19684 if (ret) {
19685 return ret;
19686 } else if (env->log.level & BPF_LOG_LEVEL) {
19687 verbose(env,
19688 "Func#%d is safe for any args that match its prototype\n",
19689 i);
19690 }
19691 }
19692 return 0;
19693 }
19694
do_check_main(struct bpf_verifier_env * env)19695 static int do_check_main(struct bpf_verifier_env *env)
19696 {
19697 int ret;
19698
19699 env->insn_idx = 0;
19700 ret = do_check_common(env, 0);
19701 if (!ret)
19702 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
19703 return ret;
19704 }
19705
19706
print_verification_stats(struct bpf_verifier_env * env)19707 static void print_verification_stats(struct bpf_verifier_env *env)
19708 {
19709 int i;
19710
19711 if (env->log.level & BPF_LOG_STATS) {
19712 verbose(env, "verification time %lld usec\n",
19713 div_u64(env->verification_time, 1000));
19714 verbose(env, "stack depth ");
19715 for (i = 0; i < env->subprog_cnt; i++) {
19716 u32 depth = env->subprog_info[i].stack_depth;
19717
19718 verbose(env, "%d", depth);
19719 if (i + 1 < env->subprog_cnt)
19720 verbose(env, "+");
19721 }
19722 verbose(env, "\n");
19723 }
19724 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
19725 "total_states %d peak_states %d mark_read %d\n",
19726 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
19727 env->max_states_per_insn, env->total_states,
19728 env->peak_states, env->longest_mark_read_walk);
19729 }
19730
check_struct_ops_btf_id(struct bpf_verifier_env * env)19731 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
19732 {
19733 const struct btf_type *t, *func_proto;
19734 const struct bpf_struct_ops *st_ops;
19735 const struct btf_member *member;
19736 struct bpf_prog *prog = env->prog;
19737 u32 btf_id, member_idx;
19738 const char *mname;
19739
19740 if (!prog->gpl_compatible) {
19741 verbose(env, "struct ops programs must have a GPL compatible license\n");
19742 return -EINVAL;
19743 }
19744
19745 btf_id = prog->aux->attach_btf_id;
19746 st_ops = bpf_struct_ops_find(btf_id);
19747 if (!st_ops) {
19748 verbose(env, "attach_btf_id %u is not a supported struct\n",
19749 btf_id);
19750 return -ENOTSUPP;
19751 }
19752
19753 t = st_ops->type;
19754 member_idx = prog->expected_attach_type;
19755 if (member_idx >= btf_type_vlen(t)) {
19756 verbose(env, "attach to invalid member idx %u of struct %s\n",
19757 member_idx, st_ops->name);
19758 return -EINVAL;
19759 }
19760
19761 member = &btf_type_member(t)[member_idx];
19762 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
19763 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
19764 NULL);
19765 if (!func_proto) {
19766 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
19767 mname, member_idx, st_ops->name);
19768 return -EINVAL;
19769 }
19770
19771 if (st_ops->check_member) {
19772 int err = st_ops->check_member(t, member, prog);
19773
19774 if (err) {
19775 verbose(env, "attach to unsupported member %s of struct %s\n",
19776 mname, st_ops->name);
19777 return err;
19778 }
19779 }
19780
19781 prog->aux->attach_func_proto = func_proto;
19782 prog->aux->attach_func_name = mname;
19783 env->ops = st_ops->verifier_ops;
19784
19785 return 0;
19786 }
19787 #define SECURITY_PREFIX "security_"
19788
check_attach_modify_return(unsigned long addr,const char * func_name)19789 static int check_attach_modify_return(unsigned long addr, const char *func_name)
19790 {
19791 if (within_error_injection_list(addr) ||
19792 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
19793 return 0;
19794
19795 return -EINVAL;
19796 }
19797
19798 /* list of non-sleepable functions that are otherwise on
19799 * ALLOW_ERROR_INJECTION list
19800 */
19801 BTF_SET_START(btf_non_sleepable_error_inject)
19802 /* Three functions below can be called from sleepable and non-sleepable context.
19803 * Assume non-sleepable from bpf safety point of view.
19804 */
BTF_ID(func,__filemap_add_folio)19805 BTF_ID(func, __filemap_add_folio)
19806 BTF_ID(func, should_fail_alloc_page)
19807 BTF_ID(func, should_failslab)
19808 BTF_SET_END(btf_non_sleepable_error_inject)
19809
19810 static int check_non_sleepable_error_inject(u32 btf_id)
19811 {
19812 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
19813 }
19814
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)19815 int bpf_check_attach_target(struct bpf_verifier_log *log,
19816 const struct bpf_prog *prog,
19817 const struct bpf_prog *tgt_prog,
19818 u32 btf_id,
19819 struct bpf_attach_target_info *tgt_info)
19820 {
19821 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
19822 const char prefix[] = "btf_trace_";
19823 int ret = 0, subprog = -1, i;
19824 const struct btf_type *t;
19825 bool conservative = true;
19826 const char *tname;
19827 struct btf *btf;
19828 long addr = 0;
19829 struct module *mod = NULL;
19830
19831 if (!btf_id) {
19832 bpf_log(log, "Tracing programs must provide btf_id\n");
19833 return -EINVAL;
19834 }
19835 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
19836 if (!btf) {
19837 bpf_log(log,
19838 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
19839 return -EINVAL;
19840 }
19841 t = btf_type_by_id(btf, btf_id);
19842 if (!t) {
19843 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
19844 return -EINVAL;
19845 }
19846 tname = btf_name_by_offset(btf, t->name_off);
19847 if (!tname) {
19848 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
19849 return -EINVAL;
19850 }
19851 if (tgt_prog) {
19852 struct bpf_prog_aux *aux = tgt_prog->aux;
19853
19854 if (bpf_prog_is_dev_bound(prog->aux) &&
19855 !bpf_prog_dev_bound_match(prog, tgt_prog)) {
19856 bpf_log(log, "Target program bound device mismatch");
19857 return -EINVAL;
19858 }
19859
19860 for (i = 0; i < aux->func_info_cnt; i++)
19861 if (aux->func_info[i].type_id == btf_id) {
19862 subprog = i;
19863 break;
19864 }
19865 if (subprog == -1) {
19866 bpf_log(log, "Subprog %s doesn't exist\n", tname);
19867 return -EINVAL;
19868 }
19869 conservative = aux->func_info_aux[subprog].unreliable;
19870 if (prog_extension) {
19871 if (conservative) {
19872 bpf_log(log,
19873 "Cannot replace static functions\n");
19874 return -EINVAL;
19875 }
19876 if (!prog->jit_requested) {
19877 bpf_log(log,
19878 "Extension programs should be JITed\n");
19879 return -EINVAL;
19880 }
19881 }
19882 if (!tgt_prog->jited) {
19883 bpf_log(log, "Can attach to only JITed progs\n");
19884 return -EINVAL;
19885 }
19886 if (tgt_prog->type == prog->type) {
19887 /* Cannot fentry/fexit another fentry/fexit program.
19888 * Cannot attach program extension to another extension.
19889 * It's ok to attach fentry/fexit to extension program.
19890 */
19891 bpf_log(log, "Cannot recursively attach\n");
19892 return -EINVAL;
19893 }
19894 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
19895 prog_extension &&
19896 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
19897 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
19898 /* Program extensions can extend all program types
19899 * except fentry/fexit. The reason is the following.
19900 * The fentry/fexit programs are used for performance
19901 * analysis, stats and can be attached to any program
19902 * type except themselves. When extension program is
19903 * replacing XDP function it is necessary to allow
19904 * performance analysis of all functions. Both original
19905 * XDP program and its program extension. Hence
19906 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
19907 * allowed. If extending of fentry/fexit was allowed it
19908 * would be possible to create long call chain
19909 * fentry->extension->fentry->extension beyond
19910 * reasonable stack size. Hence extending fentry is not
19911 * allowed.
19912 */
19913 bpf_log(log, "Cannot extend fentry/fexit\n");
19914 return -EINVAL;
19915 }
19916 } else {
19917 if (prog_extension) {
19918 bpf_log(log, "Cannot replace kernel functions\n");
19919 return -EINVAL;
19920 }
19921 }
19922
19923 switch (prog->expected_attach_type) {
19924 case BPF_TRACE_RAW_TP:
19925 if (tgt_prog) {
19926 bpf_log(log,
19927 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
19928 return -EINVAL;
19929 }
19930 if (!btf_type_is_typedef(t)) {
19931 bpf_log(log, "attach_btf_id %u is not a typedef\n",
19932 btf_id);
19933 return -EINVAL;
19934 }
19935 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
19936 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
19937 btf_id, tname);
19938 return -EINVAL;
19939 }
19940 tname += sizeof(prefix) - 1;
19941 t = btf_type_by_id(btf, t->type);
19942 if (!btf_type_is_ptr(t))
19943 /* should never happen in valid vmlinux build */
19944 return -EINVAL;
19945 t = btf_type_by_id(btf, t->type);
19946 if (!btf_type_is_func_proto(t))
19947 /* should never happen in valid vmlinux build */
19948 return -EINVAL;
19949
19950 break;
19951 case BPF_TRACE_ITER:
19952 if (!btf_type_is_func(t)) {
19953 bpf_log(log, "attach_btf_id %u is not a function\n",
19954 btf_id);
19955 return -EINVAL;
19956 }
19957 t = btf_type_by_id(btf, t->type);
19958 if (!btf_type_is_func_proto(t))
19959 return -EINVAL;
19960 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19961 if (ret)
19962 return ret;
19963 break;
19964 default:
19965 if (!prog_extension)
19966 return -EINVAL;
19967 fallthrough;
19968 case BPF_MODIFY_RETURN:
19969 case BPF_LSM_MAC:
19970 case BPF_LSM_CGROUP:
19971 case BPF_TRACE_FENTRY:
19972 case BPF_TRACE_FEXIT:
19973 if (!btf_type_is_func(t)) {
19974 bpf_log(log, "attach_btf_id %u is not a function\n",
19975 btf_id);
19976 return -EINVAL;
19977 }
19978 if (prog_extension &&
19979 btf_check_type_match(log, prog, btf, t))
19980 return -EINVAL;
19981 t = btf_type_by_id(btf, t->type);
19982 if (!btf_type_is_func_proto(t))
19983 return -EINVAL;
19984
19985 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
19986 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
19987 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
19988 return -EINVAL;
19989
19990 if (tgt_prog && conservative)
19991 t = NULL;
19992
19993 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
19994 if (ret < 0)
19995 return ret;
19996
19997 if (tgt_prog) {
19998 if (subprog == 0)
19999 addr = (long) tgt_prog->bpf_func;
20000 else
20001 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
20002 } else {
20003 if (btf_is_module(btf)) {
20004 mod = btf_try_get_module(btf);
20005 if (mod)
20006 addr = find_kallsyms_symbol_value(mod, tname);
20007 else
20008 addr = 0;
20009 } else {
20010 addr = kallsyms_lookup_name(tname);
20011 }
20012 if (!addr) {
20013 module_put(mod);
20014 bpf_log(log,
20015 "The address of function %s cannot be found\n",
20016 tname);
20017 return -ENOENT;
20018 }
20019 }
20020
20021 if (prog->aux->sleepable) {
20022 ret = -EINVAL;
20023 switch (prog->type) {
20024 case BPF_PROG_TYPE_TRACING:
20025
20026 /* fentry/fexit/fmod_ret progs can be sleepable if they are
20027 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
20028 */
20029 if (!check_non_sleepable_error_inject(btf_id) &&
20030 within_error_injection_list(addr))
20031 ret = 0;
20032 /* fentry/fexit/fmod_ret progs can also be sleepable if they are
20033 * in the fmodret id set with the KF_SLEEPABLE flag.
20034 */
20035 else {
20036 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
20037 prog);
20038
20039 if (flags && (*flags & KF_SLEEPABLE))
20040 ret = 0;
20041 }
20042 break;
20043 case BPF_PROG_TYPE_LSM:
20044 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
20045 * Only some of them are sleepable.
20046 */
20047 if (bpf_lsm_is_sleepable_hook(btf_id))
20048 ret = 0;
20049 break;
20050 default:
20051 break;
20052 }
20053 if (ret) {
20054 module_put(mod);
20055 bpf_log(log, "%s is not sleepable\n", tname);
20056 return ret;
20057 }
20058 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
20059 if (tgt_prog) {
20060 module_put(mod);
20061 bpf_log(log, "can't modify return codes of BPF programs\n");
20062 return -EINVAL;
20063 }
20064 ret = -EINVAL;
20065 if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
20066 !check_attach_modify_return(addr, tname))
20067 ret = 0;
20068 if (ret) {
20069 module_put(mod);
20070 bpf_log(log, "%s() is not modifiable\n", tname);
20071 return ret;
20072 }
20073 }
20074
20075 break;
20076 }
20077 tgt_info->tgt_addr = addr;
20078 tgt_info->tgt_name = tname;
20079 tgt_info->tgt_type = t;
20080 tgt_info->tgt_mod = mod;
20081 return 0;
20082 }
20083
BTF_SET_START(btf_id_deny)20084 BTF_SET_START(btf_id_deny)
20085 BTF_ID_UNUSED
20086 #ifdef CONFIG_SMP
20087 BTF_ID(func, migrate_disable)
20088 BTF_ID(func, migrate_enable)
20089 #endif
20090 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
20091 BTF_ID(func, rcu_read_unlock_strict)
20092 #endif
20093 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
20094 BTF_ID(func, preempt_count_add)
20095 BTF_ID(func, preempt_count_sub)
20096 #endif
20097 #ifdef CONFIG_PREEMPT_RCU
20098 BTF_ID(func, __rcu_read_lock)
20099 BTF_ID(func, __rcu_read_unlock)
20100 #endif
20101 BTF_SET_END(btf_id_deny)
20102
20103 static bool can_be_sleepable(struct bpf_prog *prog)
20104 {
20105 if (prog->type == BPF_PROG_TYPE_TRACING) {
20106 switch (prog->expected_attach_type) {
20107 case BPF_TRACE_FENTRY:
20108 case BPF_TRACE_FEXIT:
20109 case BPF_MODIFY_RETURN:
20110 case BPF_TRACE_ITER:
20111 return true;
20112 default:
20113 return false;
20114 }
20115 }
20116 return prog->type == BPF_PROG_TYPE_LSM ||
20117 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
20118 prog->type == BPF_PROG_TYPE_STRUCT_OPS;
20119 }
20120
check_attach_btf_id(struct bpf_verifier_env * env)20121 static int check_attach_btf_id(struct bpf_verifier_env *env)
20122 {
20123 struct bpf_prog *prog = env->prog;
20124 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
20125 struct bpf_attach_target_info tgt_info = {};
20126 u32 btf_id = prog->aux->attach_btf_id;
20127 struct bpf_trampoline *tr;
20128 int ret;
20129 u64 key;
20130
20131 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
20132 if (prog->aux->sleepable)
20133 /* attach_btf_id checked to be zero already */
20134 return 0;
20135 verbose(env, "Syscall programs can only be sleepable\n");
20136 return -EINVAL;
20137 }
20138
20139 if (prog->aux->sleepable && !can_be_sleepable(prog)) {
20140 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
20141 return -EINVAL;
20142 }
20143
20144 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
20145 return check_struct_ops_btf_id(env);
20146
20147 if (prog->type != BPF_PROG_TYPE_TRACING &&
20148 prog->type != BPF_PROG_TYPE_LSM &&
20149 prog->type != BPF_PROG_TYPE_EXT)
20150 return 0;
20151
20152 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
20153 if (ret)
20154 return ret;
20155
20156 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
20157 /* to make freplace equivalent to their targets, they need to
20158 * inherit env->ops and expected_attach_type for the rest of the
20159 * verification
20160 */
20161 env->ops = bpf_verifier_ops[tgt_prog->type];
20162 prog->expected_attach_type = tgt_prog->expected_attach_type;
20163 }
20164
20165 /* store info about the attachment target that will be used later */
20166 prog->aux->attach_func_proto = tgt_info.tgt_type;
20167 prog->aux->attach_func_name = tgt_info.tgt_name;
20168 prog->aux->mod = tgt_info.tgt_mod;
20169
20170 if (tgt_prog) {
20171 prog->aux->saved_dst_prog_type = tgt_prog->type;
20172 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
20173 }
20174
20175 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
20176 prog->aux->attach_btf_trace = true;
20177 return 0;
20178 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
20179 if (!bpf_iter_prog_supported(prog))
20180 return -EINVAL;
20181 return 0;
20182 }
20183
20184 if (prog->type == BPF_PROG_TYPE_LSM) {
20185 ret = bpf_lsm_verify_prog(&env->log, prog);
20186 if (ret < 0)
20187 return ret;
20188 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
20189 btf_id_set_contains(&btf_id_deny, btf_id)) {
20190 return -EINVAL;
20191 }
20192
20193 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
20194 tr = bpf_trampoline_get(key, &tgt_info);
20195 if (!tr)
20196 return -ENOMEM;
20197
20198 if (tgt_prog && tgt_prog->aux->tail_call_reachable)
20199 tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
20200
20201 prog->aux->dst_trampoline = tr;
20202 return 0;
20203 }
20204
bpf_get_btf_vmlinux(void)20205 struct btf *bpf_get_btf_vmlinux(void)
20206 {
20207 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
20208 mutex_lock(&bpf_verifier_lock);
20209 if (!btf_vmlinux)
20210 btf_vmlinux = btf_parse_vmlinux();
20211 mutex_unlock(&bpf_verifier_lock);
20212 }
20213 return btf_vmlinux;
20214 }
20215
bpf_check(struct bpf_prog ** prog,union bpf_attr * attr,bpfptr_t uattr,__u32 uattr_size)20216 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
20217 {
20218 u64 start_time = ktime_get_ns();
20219 struct bpf_verifier_env *env;
20220 int i, len, ret = -EINVAL, err;
20221 u32 log_true_size;
20222 bool is_priv;
20223
20224 /* no program is valid */
20225 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
20226 return -EINVAL;
20227
20228 /* 'struct bpf_verifier_env' can be global, but since it's not small,
20229 * allocate/free it every time bpf_check() is called
20230 */
20231 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
20232 if (!env)
20233 return -ENOMEM;
20234
20235 env->bt.env = env;
20236
20237 len = (*prog)->len;
20238 env->insn_aux_data =
20239 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
20240 ret = -ENOMEM;
20241 if (!env->insn_aux_data)
20242 goto err_free_env;
20243 for (i = 0; i < len; i++)
20244 env->insn_aux_data[i].orig_idx = i;
20245 env->prog = *prog;
20246 env->ops = bpf_verifier_ops[env->prog->type];
20247 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
20248 is_priv = bpf_capable();
20249
20250 bpf_get_btf_vmlinux();
20251
20252 /* grab the mutex to protect few globals used by verifier */
20253 if (!is_priv)
20254 mutex_lock(&bpf_verifier_lock);
20255
20256 /* user could have requested verbose verifier output
20257 * and supplied buffer to store the verification trace
20258 */
20259 ret = bpf_vlog_init(&env->log, attr->log_level,
20260 (char __user *) (unsigned long) attr->log_buf,
20261 attr->log_size);
20262 if (ret)
20263 goto err_unlock;
20264
20265 mark_verifier_state_clean(env);
20266
20267 if (IS_ERR(btf_vmlinux)) {
20268 /* Either gcc or pahole or kernel are broken. */
20269 verbose(env, "in-kernel BTF is malformed\n");
20270 ret = PTR_ERR(btf_vmlinux);
20271 goto skip_full_check;
20272 }
20273
20274 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
20275 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
20276 env->strict_alignment = true;
20277 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
20278 env->strict_alignment = false;
20279
20280 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
20281 env->allow_uninit_stack = bpf_allow_uninit_stack();
20282 env->bypass_spec_v1 = bpf_bypass_spec_v1();
20283 env->bypass_spec_v4 = bpf_bypass_spec_v4();
20284 env->bpf_capable = bpf_capable();
20285
20286 if (is_priv)
20287 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
20288
20289 env->explored_states = kvcalloc(state_htab_size(env),
20290 sizeof(struct bpf_verifier_state_list *),
20291 GFP_USER);
20292 ret = -ENOMEM;
20293 if (!env->explored_states)
20294 goto skip_full_check;
20295
20296 ret = add_subprog_and_kfunc(env);
20297 if (ret < 0)
20298 goto skip_full_check;
20299
20300 ret = check_subprogs(env);
20301 if (ret < 0)
20302 goto skip_full_check;
20303
20304 ret = check_btf_info(env, attr, uattr);
20305 if (ret < 0)
20306 goto skip_full_check;
20307
20308 ret = check_attach_btf_id(env);
20309 if (ret)
20310 goto skip_full_check;
20311
20312 ret = resolve_pseudo_ldimm64(env);
20313 if (ret < 0)
20314 goto skip_full_check;
20315
20316 if (bpf_prog_is_offloaded(env->prog->aux)) {
20317 ret = bpf_prog_offload_verifier_prep(env->prog);
20318 if (ret)
20319 goto skip_full_check;
20320 }
20321
20322 ret = check_cfg(env);
20323 if (ret < 0)
20324 goto skip_full_check;
20325
20326 ret = do_check_subprogs(env);
20327 ret = ret ?: do_check_main(env);
20328
20329 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
20330 ret = bpf_prog_offload_finalize(env);
20331
20332 skip_full_check:
20333 kvfree(env->explored_states);
20334
20335 if (ret == 0)
20336 ret = check_max_stack_depth(env);
20337
20338 /* instruction rewrites happen after this point */
20339 if (ret == 0)
20340 ret = optimize_bpf_loop(env);
20341
20342 if (is_priv) {
20343 if (ret == 0)
20344 opt_hard_wire_dead_code_branches(env);
20345 if (ret == 0)
20346 ret = opt_remove_dead_code(env);
20347 if (ret == 0)
20348 ret = opt_remove_nops(env);
20349 } else {
20350 if (ret == 0)
20351 sanitize_dead_code(env);
20352 }
20353
20354 if (ret == 0)
20355 /* program is valid, convert *(u32*)(ctx + off) accesses */
20356 ret = convert_ctx_accesses(env);
20357
20358 if (ret == 0)
20359 ret = do_misc_fixups(env);
20360
20361 /* do 32-bit optimization after insn patching has done so those patched
20362 * insns could be handled correctly.
20363 */
20364 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
20365 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
20366 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
20367 : false;
20368 }
20369
20370 if (ret == 0)
20371 ret = fixup_call_args(env);
20372
20373 env->verification_time = ktime_get_ns() - start_time;
20374 print_verification_stats(env);
20375 env->prog->aux->verified_insns = env->insn_processed;
20376
20377 /* preserve original error even if log finalization is successful */
20378 err = bpf_vlog_finalize(&env->log, &log_true_size);
20379 if (err)
20380 ret = err;
20381
20382 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
20383 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
20384 &log_true_size, sizeof(log_true_size))) {
20385 ret = -EFAULT;
20386 goto err_release_maps;
20387 }
20388
20389 if (ret)
20390 goto err_release_maps;
20391
20392 if (env->used_map_cnt) {
20393 /* if program passed verifier, update used_maps in bpf_prog_info */
20394 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
20395 sizeof(env->used_maps[0]),
20396 GFP_KERNEL);
20397
20398 if (!env->prog->aux->used_maps) {
20399 ret = -ENOMEM;
20400 goto err_release_maps;
20401 }
20402
20403 memcpy(env->prog->aux->used_maps, env->used_maps,
20404 sizeof(env->used_maps[0]) * env->used_map_cnt);
20405 env->prog->aux->used_map_cnt = env->used_map_cnt;
20406 }
20407 if (env->used_btf_cnt) {
20408 /* if program passed verifier, update used_btfs in bpf_prog_aux */
20409 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
20410 sizeof(env->used_btfs[0]),
20411 GFP_KERNEL);
20412 if (!env->prog->aux->used_btfs) {
20413 ret = -ENOMEM;
20414 goto err_release_maps;
20415 }
20416
20417 memcpy(env->prog->aux->used_btfs, env->used_btfs,
20418 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
20419 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
20420 }
20421 if (env->used_map_cnt || env->used_btf_cnt) {
20422 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
20423 * bpf_ld_imm64 instructions
20424 */
20425 convert_pseudo_ld_imm64(env);
20426 }
20427
20428 adjust_btf_func(env);
20429
20430 err_release_maps:
20431 if (!env->prog->aux->used_maps)
20432 /* if we didn't copy map pointers into bpf_prog_info, release
20433 * them now. Otherwise free_used_maps() will release them.
20434 */
20435 release_maps(env);
20436 if (!env->prog->aux->used_btfs)
20437 release_btfs(env);
20438
20439 /* extension progs temporarily inherit the attach_type of their targets
20440 for verification purposes, so set it back to zero before returning
20441 */
20442 if (env->prog->type == BPF_PROG_TYPE_EXT)
20443 env->prog->expected_attach_type = 0;
20444
20445 *prog = env->prog;
20446 err_unlock:
20447 if (!is_priv)
20448 mutex_unlock(&bpf_verifier_lock);
20449 vfree(env->insn_aux_data);
20450 err_free_env:
20451 kfree(env);
20452 return ret;
20453 }
20454