xref: /openbmc/linux/kernel/bpf/verifier.c (revision 4bce6fce)
1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2  *
3  * This program is free software; you can redistribute it and/or
4  * modify it under the terms of version 2 of the GNU General Public
5  * License as published by the Free Software Foundation.
6  *
7  * This program is distributed in the hope that it will be useful, but
8  * WITHOUT ANY WARRANTY; without even the implied warranty of
9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
10  * General Public License for more details.
11  */
12 #include <linux/kernel.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/bpf.h>
16 #include <linux/filter.h>
17 #include <net/netlink.h>
18 #include <linux/file.h>
19 #include <linux/vmalloc.h>
20 
21 /* bpf_check() is a static code analyzer that walks eBPF program
22  * instruction by instruction and updates register/stack state.
23  * All paths of conditional branches are analyzed until 'bpf_exit' insn.
24  *
25  * The first pass is depth-first-search to check that the program is a DAG.
26  * It rejects the following programs:
27  * - larger than BPF_MAXINSNS insns
28  * - if loop is present (detected via back-edge)
29  * - unreachable insns exist (shouldn't be a forest. program = one function)
30  * - out of bounds or malformed jumps
31  * The second pass is all possible path descent from the 1st insn.
32  * Since it's analyzing all pathes through the program, the length of the
33  * analysis is limited to 32k insn, which may be hit even if total number of
34  * insn is less then 4K, but there are too many branches that change stack/regs.
35  * Number of 'branches to be analyzed' is limited to 1k
36  *
37  * On entry to each instruction, each register has a type, and the instruction
38  * changes the types of the registers depending on instruction semantics.
39  * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
40  * copied to R1.
41  *
42  * All registers are 64-bit.
43  * R0 - return register
44  * R1-R5 argument passing registers
45  * R6-R9 callee saved registers
46  * R10 - frame pointer read-only
47  *
48  * At the start of BPF program the register R1 contains a pointer to bpf_context
49  * and has type PTR_TO_CTX.
50  *
51  * Verifier tracks arithmetic operations on pointers in case:
52  *    BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
53  *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
54  * 1st insn copies R10 (which has FRAME_PTR) type into R1
55  * and 2nd arithmetic instruction is pattern matched to recognize
56  * that it wants to construct a pointer to some element within stack.
57  * So after 2nd insn, the register R1 has type PTR_TO_STACK
58  * (and -20 constant is saved for further stack bounds checking).
59  * Meaning that this reg is a pointer to stack plus known immediate constant.
60  *
61  * Most of the time the registers have UNKNOWN_VALUE type, which
62  * means the register has some value, but it's not a valid pointer.
63  * (like pointer plus pointer becomes UNKNOWN_VALUE type)
64  *
65  * When verifier sees load or store instructions the type of base register
66  * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, FRAME_PTR. These are three pointer
67  * types recognized by check_mem_access() function.
68  *
69  * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
70  * and the range of [ptr, ptr + map's value_size) is accessible.
71  *
72  * registers used to pass values to function calls are checked against
73  * function argument constraints.
74  *
75  * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
76  * It means that the register type passed to this function must be
77  * PTR_TO_STACK and it will be used inside the function as
78  * 'pointer to map element key'
79  *
80  * For example the argument constraints for bpf_map_lookup_elem():
81  *   .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
82  *   .arg1_type = ARG_CONST_MAP_PTR,
83  *   .arg2_type = ARG_PTR_TO_MAP_KEY,
84  *
85  * ret_type says that this function returns 'pointer to map elem value or null'
86  * function expects 1st argument to be a const pointer to 'struct bpf_map' and
87  * 2nd argument should be a pointer to stack, which will be used inside
88  * the helper function as a pointer to map element key.
89  *
90  * On the kernel side the helper function looks like:
91  * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
92  * {
93  *    struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
94  *    void *key = (void *) (unsigned long) r2;
95  *    void *value;
96  *
97  *    here kernel can access 'key' and 'map' pointers safely, knowing that
98  *    [key, key + map->key_size) bytes are valid and were initialized on
99  *    the stack of eBPF program.
100  * }
101  *
102  * Corresponding eBPF program may look like:
103  *    BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),  // after this insn R2 type is FRAME_PTR
104  *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
105  *    BPF_LD_MAP_FD(BPF_REG_1, map_fd),      // after this insn R1 type is CONST_PTR_TO_MAP
106  *    BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
107  * here verifier looks at prototype of map_lookup_elem() and sees:
108  * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
109  * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
110  *
111  * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
112  * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
113  * and were initialized prior to this call.
114  * If it's ok, then verifier allows this BPF_CALL insn and looks at
115  * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
116  * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
117  * returns ether pointer to map value or NULL.
118  *
119  * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
120  * insn, the register holding that pointer in the true branch changes state to
121  * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
122  * branch. See check_cond_jmp_op().
123  *
124  * After the call R0 is set to return type of the function and registers R1-R5
125  * are set to NOT_INIT to indicate that they are no longer readable.
126  */
127 
128 /* types of values stored in eBPF registers */
129 enum bpf_reg_type {
130 	NOT_INIT = 0,		 /* nothing was written into register */
131 	UNKNOWN_VALUE,		 /* reg doesn't contain a valid pointer */
132 	PTR_TO_CTX,		 /* reg points to bpf_context */
133 	CONST_PTR_TO_MAP,	 /* reg points to struct bpf_map */
134 	PTR_TO_MAP_VALUE,	 /* reg points to map element value */
135 	PTR_TO_MAP_VALUE_OR_NULL,/* points to map elem value or NULL */
136 	FRAME_PTR,		 /* reg == frame_pointer */
137 	PTR_TO_STACK,		 /* reg == frame_pointer + imm */
138 	CONST_IMM,		 /* constant integer value */
139 };
140 
141 struct reg_state {
142 	enum bpf_reg_type type;
143 	union {
144 		/* valid when type == CONST_IMM | PTR_TO_STACK */
145 		int imm;
146 
147 		/* valid when type == CONST_PTR_TO_MAP | PTR_TO_MAP_VALUE |
148 		 *   PTR_TO_MAP_VALUE_OR_NULL
149 		 */
150 		struct bpf_map *map_ptr;
151 	};
152 };
153 
154 enum bpf_stack_slot_type {
155 	STACK_INVALID,    /* nothing was stored in this stack slot */
156 	STACK_SPILL,      /* register spilled into stack */
157 	STACK_MISC	  /* BPF program wrote some data into this slot */
158 };
159 
160 #define BPF_REG_SIZE 8	/* size of eBPF register in bytes */
161 
162 /* state of the program:
163  * type of all registers and stack info
164  */
165 struct verifier_state {
166 	struct reg_state regs[MAX_BPF_REG];
167 	u8 stack_slot_type[MAX_BPF_STACK];
168 	struct reg_state spilled_regs[MAX_BPF_STACK / BPF_REG_SIZE];
169 };
170 
171 /* linked list of verifier states used to prune search */
172 struct verifier_state_list {
173 	struct verifier_state state;
174 	struct verifier_state_list *next;
175 };
176 
177 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
178 struct verifier_stack_elem {
179 	/* verifer state is 'st'
180 	 * before processing instruction 'insn_idx'
181 	 * and after processing instruction 'prev_insn_idx'
182 	 */
183 	struct verifier_state st;
184 	int insn_idx;
185 	int prev_insn_idx;
186 	struct verifier_stack_elem *next;
187 };
188 
189 #define MAX_USED_MAPS 64 /* max number of maps accessed by one eBPF program */
190 
191 /* single container for all structs
192  * one verifier_env per bpf_check() call
193  */
194 struct verifier_env {
195 	struct bpf_prog *prog;		/* eBPF program being verified */
196 	struct verifier_stack_elem *head; /* stack of verifier states to be processed */
197 	int stack_size;			/* number of states to be processed */
198 	struct verifier_state cur_state; /* current verifier state */
199 	struct verifier_state_list **explored_states; /* search pruning optimization */
200 	struct bpf_map *used_maps[MAX_USED_MAPS]; /* array of map's used by eBPF program */
201 	u32 used_map_cnt;		/* number of used maps */
202 };
203 
204 /* verbose verifier prints what it's seeing
205  * bpf_check() is called under lock, so no race to access these global vars
206  */
207 static u32 log_level, log_size, log_len;
208 static char *log_buf;
209 
210 static DEFINE_MUTEX(bpf_verifier_lock);
211 
212 /* log_level controls verbosity level of eBPF verifier.
213  * verbose() is used to dump the verification trace to the log, so the user
214  * can figure out what's wrong with the program
215  */
216 static void verbose(const char *fmt, ...)
217 {
218 	va_list args;
219 
220 	if (log_level == 0 || log_len >= log_size - 1)
221 		return;
222 
223 	va_start(args, fmt);
224 	log_len += vscnprintf(log_buf + log_len, log_size - log_len, fmt, args);
225 	va_end(args);
226 }
227 
228 /* string representation of 'enum bpf_reg_type' */
229 static const char * const reg_type_str[] = {
230 	[NOT_INIT]		= "?",
231 	[UNKNOWN_VALUE]		= "inv",
232 	[PTR_TO_CTX]		= "ctx",
233 	[CONST_PTR_TO_MAP]	= "map_ptr",
234 	[PTR_TO_MAP_VALUE]	= "map_value",
235 	[PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
236 	[FRAME_PTR]		= "fp",
237 	[PTR_TO_STACK]		= "fp",
238 	[CONST_IMM]		= "imm",
239 };
240 
241 static void print_verifier_state(struct verifier_env *env)
242 {
243 	enum bpf_reg_type t;
244 	int i;
245 
246 	for (i = 0; i < MAX_BPF_REG; i++) {
247 		t = env->cur_state.regs[i].type;
248 		if (t == NOT_INIT)
249 			continue;
250 		verbose(" R%d=%s", i, reg_type_str[t]);
251 		if (t == CONST_IMM || t == PTR_TO_STACK)
252 			verbose("%d", env->cur_state.regs[i].imm);
253 		else if (t == CONST_PTR_TO_MAP || t == PTR_TO_MAP_VALUE ||
254 			 t == PTR_TO_MAP_VALUE_OR_NULL)
255 			verbose("(ks=%d,vs=%d)",
256 				env->cur_state.regs[i].map_ptr->key_size,
257 				env->cur_state.regs[i].map_ptr->value_size);
258 	}
259 	for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
260 		if (env->cur_state.stack_slot_type[i] == STACK_SPILL)
261 			verbose(" fp%d=%s", -MAX_BPF_STACK + i,
262 				reg_type_str[env->cur_state.spilled_regs[i / BPF_REG_SIZE].type]);
263 	}
264 	verbose("\n");
265 }
266 
267 static const char *const bpf_class_string[] = {
268 	[BPF_LD]    = "ld",
269 	[BPF_LDX]   = "ldx",
270 	[BPF_ST]    = "st",
271 	[BPF_STX]   = "stx",
272 	[BPF_ALU]   = "alu",
273 	[BPF_JMP]   = "jmp",
274 	[BPF_RET]   = "BUG",
275 	[BPF_ALU64] = "alu64",
276 };
277 
278 static const char *const bpf_alu_string[] = {
279 	[BPF_ADD >> 4]  = "+=",
280 	[BPF_SUB >> 4]  = "-=",
281 	[BPF_MUL >> 4]  = "*=",
282 	[BPF_DIV >> 4]  = "/=",
283 	[BPF_OR  >> 4]  = "|=",
284 	[BPF_AND >> 4]  = "&=",
285 	[BPF_LSH >> 4]  = "<<=",
286 	[BPF_RSH >> 4]  = ">>=",
287 	[BPF_NEG >> 4]  = "neg",
288 	[BPF_MOD >> 4]  = "%=",
289 	[BPF_XOR >> 4]  = "^=",
290 	[BPF_MOV >> 4]  = "=",
291 	[BPF_ARSH >> 4] = "s>>=",
292 	[BPF_END >> 4]  = "endian",
293 };
294 
295 static const char *const bpf_ldst_string[] = {
296 	[BPF_W >> 3]  = "u32",
297 	[BPF_H >> 3]  = "u16",
298 	[BPF_B >> 3]  = "u8",
299 	[BPF_DW >> 3] = "u64",
300 };
301 
302 static const char *const bpf_jmp_string[] = {
303 	[BPF_JA >> 4]   = "jmp",
304 	[BPF_JEQ >> 4]  = "==",
305 	[BPF_JGT >> 4]  = ">",
306 	[BPF_JGE >> 4]  = ">=",
307 	[BPF_JSET >> 4] = "&",
308 	[BPF_JNE >> 4]  = "!=",
309 	[BPF_JSGT >> 4] = "s>",
310 	[BPF_JSGE >> 4] = "s>=",
311 	[BPF_CALL >> 4] = "call",
312 	[BPF_EXIT >> 4] = "exit",
313 };
314 
315 static void print_bpf_insn(struct bpf_insn *insn)
316 {
317 	u8 class = BPF_CLASS(insn->code);
318 
319 	if (class == BPF_ALU || class == BPF_ALU64) {
320 		if (BPF_SRC(insn->code) == BPF_X)
321 			verbose("(%02x) %sr%d %s %sr%d\n",
322 				insn->code, class == BPF_ALU ? "(u32) " : "",
323 				insn->dst_reg,
324 				bpf_alu_string[BPF_OP(insn->code) >> 4],
325 				class == BPF_ALU ? "(u32) " : "",
326 				insn->src_reg);
327 		else
328 			verbose("(%02x) %sr%d %s %s%d\n",
329 				insn->code, class == BPF_ALU ? "(u32) " : "",
330 				insn->dst_reg,
331 				bpf_alu_string[BPF_OP(insn->code) >> 4],
332 				class == BPF_ALU ? "(u32) " : "",
333 				insn->imm);
334 	} else if (class == BPF_STX) {
335 		if (BPF_MODE(insn->code) == BPF_MEM)
336 			verbose("(%02x) *(%s *)(r%d %+d) = r%d\n",
337 				insn->code,
338 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
339 				insn->dst_reg,
340 				insn->off, insn->src_reg);
341 		else if (BPF_MODE(insn->code) == BPF_XADD)
342 			verbose("(%02x) lock *(%s *)(r%d %+d) += r%d\n",
343 				insn->code,
344 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
345 				insn->dst_reg, insn->off,
346 				insn->src_reg);
347 		else
348 			verbose("BUG_%02x\n", insn->code);
349 	} else if (class == BPF_ST) {
350 		if (BPF_MODE(insn->code) != BPF_MEM) {
351 			verbose("BUG_st_%02x\n", insn->code);
352 			return;
353 		}
354 		verbose("(%02x) *(%s *)(r%d %+d) = %d\n",
355 			insn->code,
356 			bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
357 			insn->dst_reg,
358 			insn->off, insn->imm);
359 	} else if (class == BPF_LDX) {
360 		if (BPF_MODE(insn->code) != BPF_MEM) {
361 			verbose("BUG_ldx_%02x\n", insn->code);
362 			return;
363 		}
364 		verbose("(%02x) r%d = *(%s *)(r%d %+d)\n",
365 			insn->code, insn->dst_reg,
366 			bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
367 			insn->src_reg, insn->off);
368 	} else if (class == BPF_LD) {
369 		if (BPF_MODE(insn->code) == BPF_ABS) {
370 			verbose("(%02x) r0 = *(%s *)skb[%d]\n",
371 				insn->code,
372 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
373 				insn->imm);
374 		} else if (BPF_MODE(insn->code) == BPF_IND) {
375 			verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n",
376 				insn->code,
377 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
378 				insn->src_reg, insn->imm);
379 		} else if (BPF_MODE(insn->code) == BPF_IMM) {
380 			verbose("(%02x) r%d = 0x%x\n",
381 				insn->code, insn->dst_reg, insn->imm);
382 		} else {
383 			verbose("BUG_ld_%02x\n", insn->code);
384 			return;
385 		}
386 	} else if (class == BPF_JMP) {
387 		u8 opcode = BPF_OP(insn->code);
388 
389 		if (opcode == BPF_CALL) {
390 			verbose("(%02x) call %d\n", insn->code, insn->imm);
391 		} else if (insn->code == (BPF_JMP | BPF_JA)) {
392 			verbose("(%02x) goto pc%+d\n",
393 				insn->code, insn->off);
394 		} else if (insn->code == (BPF_JMP | BPF_EXIT)) {
395 			verbose("(%02x) exit\n", insn->code);
396 		} else if (BPF_SRC(insn->code) == BPF_X) {
397 			verbose("(%02x) if r%d %s r%d goto pc%+d\n",
398 				insn->code, insn->dst_reg,
399 				bpf_jmp_string[BPF_OP(insn->code) >> 4],
400 				insn->src_reg, insn->off);
401 		} else {
402 			verbose("(%02x) if r%d %s 0x%x goto pc%+d\n",
403 				insn->code, insn->dst_reg,
404 				bpf_jmp_string[BPF_OP(insn->code) >> 4],
405 				insn->imm, insn->off);
406 		}
407 	} else {
408 		verbose("(%02x) %s\n", insn->code, bpf_class_string[class]);
409 	}
410 }
411 
412 static int pop_stack(struct verifier_env *env, int *prev_insn_idx)
413 {
414 	struct verifier_stack_elem *elem;
415 	int insn_idx;
416 
417 	if (env->head == NULL)
418 		return -1;
419 
420 	memcpy(&env->cur_state, &env->head->st, sizeof(env->cur_state));
421 	insn_idx = env->head->insn_idx;
422 	if (prev_insn_idx)
423 		*prev_insn_idx = env->head->prev_insn_idx;
424 	elem = env->head->next;
425 	kfree(env->head);
426 	env->head = elem;
427 	env->stack_size--;
428 	return insn_idx;
429 }
430 
431 static struct verifier_state *push_stack(struct verifier_env *env, int insn_idx,
432 					 int prev_insn_idx)
433 {
434 	struct verifier_stack_elem *elem;
435 
436 	elem = kmalloc(sizeof(struct verifier_stack_elem), GFP_KERNEL);
437 	if (!elem)
438 		goto err;
439 
440 	memcpy(&elem->st, &env->cur_state, sizeof(env->cur_state));
441 	elem->insn_idx = insn_idx;
442 	elem->prev_insn_idx = prev_insn_idx;
443 	elem->next = env->head;
444 	env->head = elem;
445 	env->stack_size++;
446 	if (env->stack_size > 1024) {
447 		verbose("BPF program is too complex\n");
448 		goto err;
449 	}
450 	return &elem->st;
451 err:
452 	/* pop all elements and return */
453 	while (pop_stack(env, NULL) >= 0);
454 	return NULL;
455 }
456 
457 #define CALLER_SAVED_REGS 6
458 static const int caller_saved[CALLER_SAVED_REGS] = {
459 	BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
460 };
461 
462 static void init_reg_state(struct reg_state *regs)
463 {
464 	int i;
465 
466 	for (i = 0; i < MAX_BPF_REG; i++) {
467 		regs[i].type = NOT_INIT;
468 		regs[i].imm = 0;
469 		regs[i].map_ptr = NULL;
470 	}
471 
472 	/* frame pointer */
473 	regs[BPF_REG_FP].type = FRAME_PTR;
474 
475 	/* 1st arg to a function */
476 	regs[BPF_REG_1].type = PTR_TO_CTX;
477 }
478 
479 static void mark_reg_unknown_value(struct reg_state *regs, u32 regno)
480 {
481 	BUG_ON(regno >= MAX_BPF_REG);
482 	regs[regno].type = UNKNOWN_VALUE;
483 	regs[regno].imm = 0;
484 	regs[regno].map_ptr = NULL;
485 }
486 
487 enum reg_arg_type {
488 	SRC_OP,		/* register is used as source operand */
489 	DST_OP,		/* register is used as destination operand */
490 	DST_OP_NO_MARK	/* same as above, check only, don't mark */
491 };
492 
493 static int check_reg_arg(struct reg_state *regs, u32 regno,
494 			 enum reg_arg_type t)
495 {
496 	if (regno >= MAX_BPF_REG) {
497 		verbose("R%d is invalid\n", regno);
498 		return -EINVAL;
499 	}
500 
501 	if (t == SRC_OP) {
502 		/* check whether register used as source operand can be read */
503 		if (regs[regno].type == NOT_INIT) {
504 			verbose("R%d !read_ok\n", regno);
505 			return -EACCES;
506 		}
507 	} else {
508 		/* check whether register used as dest operand can be written to */
509 		if (regno == BPF_REG_FP) {
510 			verbose("frame pointer is read only\n");
511 			return -EACCES;
512 		}
513 		if (t == DST_OP)
514 			mark_reg_unknown_value(regs, regno);
515 	}
516 	return 0;
517 }
518 
519 static int bpf_size_to_bytes(int bpf_size)
520 {
521 	if (bpf_size == BPF_W)
522 		return 4;
523 	else if (bpf_size == BPF_H)
524 		return 2;
525 	else if (bpf_size == BPF_B)
526 		return 1;
527 	else if (bpf_size == BPF_DW)
528 		return 8;
529 	else
530 		return -EINVAL;
531 }
532 
533 /* check_stack_read/write functions track spill/fill of registers,
534  * stack boundary and alignment are checked in check_mem_access()
535  */
536 static int check_stack_write(struct verifier_state *state, int off, int size,
537 			     int value_regno)
538 {
539 	int i;
540 	/* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
541 	 * so it's aligned access and [off, off + size) are within stack limits
542 	 */
543 
544 	if (value_regno >= 0 &&
545 	    (state->regs[value_regno].type == PTR_TO_MAP_VALUE ||
546 	     state->regs[value_regno].type == PTR_TO_STACK ||
547 	     state->regs[value_regno].type == PTR_TO_CTX)) {
548 
549 		/* register containing pointer is being spilled into stack */
550 		if (size != BPF_REG_SIZE) {
551 			verbose("invalid size of register spill\n");
552 			return -EACCES;
553 		}
554 
555 		/* save register state */
556 		state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] =
557 			state->regs[value_regno];
558 
559 		for (i = 0; i < BPF_REG_SIZE; i++)
560 			state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL;
561 	} else {
562 		/* regular write of data into stack */
563 		state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] =
564 			(struct reg_state) {};
565 
566 		for (i = 0; i < size; i++)
567 			state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC;
568 	}
569 	return 0;
570 }
571 
572 static int check_stack_read(struct verifier_state *state, int off, int size,
573 			    int value_regno)
574 {
575 	u8 *slot_type;
576 	int i;
577 
578 	slot_type = &state->stack_slot_type[MAX_BPF_STACK + off];
579 
580 	if (slot_type[0] == STACK_SPILL) {
581 		if (size != BPF_REG_SIZE) {
582 			verbose("invalid size of register spill\n");
583 			return -EACCES;
584 		}
585 		for (i = 1; i < BPF_REG_SIZE; i++) {
586 			if (slot_type[i] != STACK_SPILL) {
587 				verbose("corrupted spill memory\n");
588 				return -EACCES;
589 			}
590 		}
591 
592 		if (value_regno >= 0)
593 			/* restore register state from stack */
594 			state->regs[value_regno] =
595 				state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE];
596 		return 0;
597 	} else {
598 		for (i = 0; i < size; i++) {
599 			if (slot_type[i] != STACK_MISC) {
600 				verbose("invalid read from stack off %d+%d size %d\n",
601 					off, i, size);
602 				return -EACCES;
603 			}
604 		}
605 		if (value_regno >= 0)
606 			/* have read misc data from the stack */
607 			mark_reg_unknown_value(state->regs, value_regno);
608 		return 0;
609 	}
610 }
611 
612 /* check read/write into map element returned by bpf_map_lookup_elem() */
613 static int check_map_access(struct verifier_env *env, u32 regno, int off,
614 			    int size)
615 {
616 	struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
617 
618 	if (off < 0 || off + size > map->value_size) {
619 		verbose("invalid access to map value, value_size=%d off=%d size=%d\n",
620 			map->value_size, off, size);
621 		return -EACCES;
622 	}
623 	return 0;
624 }
625 
626 /* check access to 'struct bpf_context' fields */
627 static int check_ctx_access(struct verifier_env *env, int off, int size,
628 			    enum bpf_access_type t)
629 {
630 	if (env->prog->aux->ops->is_valid_access &&
631 	    env->prog->aux->ops->is_valid_access(off, size, t))
632 		return 0;
633 
634 	verbose("invalid bpf_context access off=%d size=%d\n", off, size);
635 	return -EACCES;
636 }
637 
638 /* check whether memory at (regno + off) is accessible for t = (read | write)
639  * if t==write, value_regno is a register which value is stored into memory
640  * if t==read, value_regno is a register which will receive the value from memory
641  * if t==write && value_regno==-1, some unknown value is stored into memory
642  * if t==read && value_regno==-1, don't care what we read from memory
643  */
644 static int check_mem_access(struct verifier_env *env, u32 regno, int off,
645 			    int bpf_size, enum bpf_access_type t,
646 			    int value_regno)
647 {
648 	struct verifier_state *state = &env->cur_state;
649 	int size, err = 0;
650 
651 	size = bpf_size_to_bytes(bpf_size);
652 	if (size < 0)
653 		return size;
654 
655 	if (off % size != 0) {
656 		verbose("misaligned access off %d size %d\n", off, size);
657 		return -EACCES;
658 	}
659 
660 	if (state->regs[regno].type == PTR_TO_MAP_VALUE) {
661 		err = check_map_access(env, regno, off, size);
662 		if (!err && t == BPF_READ && value_regno >= 0)
663 			mark_reg_unknown_value(state->regs, value_regno);
664 
665 	} else if (state->regs[regno].type == PTR_TO_CTX) {
666 		err = check_ctx_access(env, off, size, t);
667 		if (!err && t == BPF_READ && value_regno >= 0)
668 			mark_reg_unknown_value(state->regs, value_regno);
669 
670 	} else if (state->regs[regno].type == FRAME_PTR) {
671 		if (off >= 0 || off < -MAX_BPF_STACK) {
672 			verbose("invalid stack off=%d size=%d\n", off, size);
673 			return -EACCES;
674 		}
675 		if (t == BPF_WRITE)
676 			err = check_stack_write(state, off, size, value_regno);
677 		else
678 			err = check_stack_read(state, off, size, value_regno);
679 	} else {
680 		verbose("R%d invalid mem access '%s'\n",
681 			regno, reg_type_str[state->regs[regno].type]);
682 		return -EACCES;
683 	}
684 	return err;
685 }
686 
687 static int check_xadd(struct verifier_env *env, struct bpf_insn *insn)
688 {
689 	struct reg_state *regs = env->cur_state.regs;
690 	int err;
691 
692 	if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
693 	    insn->imm != 0) {
694 		verbose("BPF_XADD uses reserved fields\n");
695 		return -EINVAL;
696 	}
697 
698 	/* check src1 operand */
699 	err = check_reg_arg(regs, insn->src_reg, SRC_OP);
700 	if (err)
701 		return err;
702 
703 	/* check src2 operand */
704 	err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
705 	if (err)
706 		return err;
707 
708 	/* check whether atomic_add can read the memory */
709 	err = check_mem_access(env, insn->dst_reg, insn->off,
710 			       BPF_SIZE(insn->code), BPF_READ, -1);
711 	if (err)
712 		return err;
713 
714 	/* check whether atomic_add can write into the same memory */
715 	return check_mem_access(env, insn->dst_reg, insn->off,
716 				BPF_SIZE(insn->code), BPF_WRITE, -1);
717 }
718 
719 /* when register 'regno' is passed into function that will read 'access_size'
720  * bytes from that pointer, make sure that it's within stack boundary
721  * and all elements of stack are initialized
722  */
723 static int check_stack_boundary(struct verifier_env *env,
724 				int regno, int access_size)
725 {
726 	struct verifier_state *state = &env->cur_state;
727 	struct reg_state *regs = state->regs;
728 	int off, i;
729 
730 	if (regs[regno].type != PTR_TO_STACK)
731 		return -EACCES;
732 
733 	off = regs[regno].imm;
734 	if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
735 	    access_size <= 0) {
736 		verbose("invalid stack type R%d off=%d access_size=%d\n",
737 			regno, off, access_size);
738 		return -EACCES;
739 	}
740 
741 	for (i = 0; i < access_size; i++) {
742 		if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
743 			verbose("invalid indirect read from stack off %d+%d size %d\n",
744 				off, i, access_size);
745 			return -EACCES;
746 		}
747 	}
748 	return 0;
749 }
750 
751 static int check_func_arg(struct verifier_env *env, u32 regno,
752 			  enum bpf_arg_type arg_type, struct bpf_map **mapp)
753 {
754 	struct reg_state *reg = env->cur_state.regs + regno;
755 	enum bpf_reg_type expected_type;
756 	int err = 0;
757 
758 	if (arg_type == ARG_DONTCARE)
759 		return 0;
760 
761 	if (reg->type == NOT_INIT) {
762 		verbose("R%d !read_ok\n", regno);
763 		return -EACCES;
764 	}
765 
766 	if (arg_type == ARG_ANYTHING)
767 		return 0;
768 
769 	if (arg_type == ARG_PTR_TO_STACK || arg_type == ARG_PTR_TO_MAP_KEY ||
770 	    arg_type == ARG_PTR_TO_MAP_VALUE) {
771 		expected_type = PTR_TO_STACK;
772 	} else if (arg_type == ARG_CONST_STACK_SIZE) {
773 		expected_type = CONST_IMM;
774 	} else if (arg_type == ARG_CONST_MAP_PTR) {
775 		expected_type = CONST_PTR_TO_MAP;
776 	} else if (arg_type == ARG_PTR_TO_CTX) {
777 		expected_type = PTR_TO_CTX;
778 	} else {
779 		verbose("unsupported arg_type %d\n", arg_type);
780 		return -EFAULT;
781 	}
782 
783 	if (reg->type != expected_type) {
784 		verbose("R%d type=%s expected=%s\n", regno,
785 			reg_type_str[reg->type], reg_type_str[expected_type]);
786 		return -EACCES;
787 	}
788 
789 	if (arg_type == ARG_CONST_MAP_PTR) {
790 		/* bpf_map_xxx(map_ptr) call: remember that map_ptr */
791 		*mapp = reg->map_ptr;
792 
793 	} else if (arg_type == ARG_PTR_TO_MAP_KEY) {
794 		/* bpf_map_xxx(..., map_ptr, ..., key) call:
795 		 * check that [key, key + map->key_size) are within
796 		 * stack limits and initialized
797 		 */
798 		if (!*mapp) {
799 			/* in function declaration map_ptr must come before
800 			 * map_key, so that it's verified and known before
801 			 * we have to check map_key here. Otherwise it means
802 			 * that kernel subsystem misconfigured verifier
803 			 */
804 			verbose("invalid map_ptr to access map->key\n");
805 			return -EACCES;
806 		}
807 		err = check_stack_boundary(env, regno, (*mapp)->key_size);
808 
809 	} else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
810 		/* bpf_map_xxx(..., map_ptr, ..., value) call:
811 		 * check [value, value + map->value_size) validity
812 		 */
813 		if (!*mapp) {
814 			/* kernel subsystem misconfigured verifier */
815 			verbose("invalid map_ptr to access map->value\n");
816 			return -EACCES;
817 		}
818 		err = check_stack_boundary(env, regno, (*mapp)->value_size);
819 
820 	} else if (arg_type == ARG_CONST_STACK_SIZE) {
821 		/* bpf_xxx(..., buf, len) call will access 'len' bytes
822 		 * from stack pointer 'buf'. Check it
823 		 * note: regno == len, regno - 1 == buf
824 		 */
825 		if (regno == 0) {
826 			/* kernel subsystem misconfigured verifier */
827 			verbose("ARG_CONST_STACK_SIZE cannot be first argument\n");
828 			return -EACCES;
829 		}
830 		err = check_stack_boundary(env, regno - 1, reg->imm);
831 	}
832 
833 	return err;
834 }
835 
836 static int check_call(struct verifier_env *env, int func_id)
837 {
838 	struct verifier_state *state = &env->cur_state;
839 	const struct bpf_func_proto *fn = NULL;
840 	struct reg_state *regs = state->regs;
841 	struct bpf_map *map = NULL;
842 	struct reg_state *reg;
843 	int i, err;
844 
845 	/* find function prototype */
846 	if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
847 		verbose("invalid func %d\n", func_id);
848 		return -EINVAL;
849 	}
850 
851 	if (env->prog->aux->ops->get_func_proto)
852 		fn = env->prog->aux->ops->get_func_proto(func_id);
853 
854 	if (!fn) {
855 		verbose("unknown func %d\n", func_id);
856 		return -EINVAL;
857 	}
858 
859 	/* eBPF programs must be GPL compatible to use GPL-ed functions */
860 	if (!env->prog->gpl_compatible && fn->gpl_only) {
861 		verbose("cannot call GPL only function from proprietary program\n");
862 		return -EINVAL;
863 	}
864 
865 	/* check args */
866 	err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &map);
867 	if (err)
868 		return err;
869 	err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &map);
870 	if (err)
871 		return err;
872 	err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &map);
873 	if (err)
874 		return err;
875 	err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &map);
876 	if (err)
877 		return err;
878 	err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &map);
879 	if (err)
880 		return err;
881 
882 	/* reset caller saved regs */
883 	for (i = 0; i < CALLER_SAVED_REGS; i++) {
884 		reg = regs + caller_saved[i];
885 		reg->type = NOT_INIT;
886 		reg->imm = 0;
887 	}
888 
889 	/* update return register */
890 	if (fn->ret_type == RET_INTEGER) {
891 		regs[BPF_REG_0].type = UNKNOWN_VALUE;
892 	} else if (fn->ret_type == RET_VOID) {
893 		regs[BPF_REG_0].type = NOT_INIT;
894 	} else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
895 		regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
896 		/* remember map_ptr, so that check_map_access()
897 		 * can check 'value_size' boundary of memory access
898 		 * to map element returned from bpf_map_lookup_elem()
899 		 */
900 		if (map == NULL) {
901 			verbose("kernel subsystem misconfigured verifier\n");
902 			return -EINVAL;
903 		}
904 		regs[BPF_REG_0].map_ptr = map;
905 	} else {
906 		verbose("unknown return type %d of func %d\n",
907 			fn->ret_type, func_id);
908 		return -EINVAL;
909 	}
910 	return 0;
911 }
912 
913 /* check validity of 32-bit and 64-bit arithmetic operations */
914 static int check_alu_op(struct reg_state *regs, struct bpf_insn *insn)
915 {
916 	u8 opcode = BPF_OP(insn->code);
917 	int err;
918 
919 	if (opcode == BPF_END || opcode == BPF_NEG) {
920 		if (opcode == BPF_NEG) {
921 			if (BPF_SRC(insn->code) != 0 ||
922 			    insn->src_reg != BPF_REG_0 ||
923 			    insn->off != 0 || insn->imm != 0) {
924 				verbose("BPF_NEG uses reserved fields\n");
925 				return -EINVAL;
926 			}
927 		} else {
928 			if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
929 			    (insn->imm != 16 && insn->imm != 32 && insn->imm != 64)) {
930 				verbose("BPF_END uses reserved fields\n");
931 				return -EINVAL;
932 			}
933 		}
934 
935 		/* check src operand */
936 		err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
937 		if (err)
938 			return err;
939 
940 		/* check dest operand */
941 		err = check_reg_arg(regs, insn->dst_reg, DST_OP);
942 		if (err)
943 			return err;
944 
945 	} else if (opcode == BPF_MOV) {
946 
947 		if (BPF_SRC(insn->code) == BPF_X) {
948 			if (insn->imm != 0 || insn->off != 0) {
949 				verbose("BPF_MOV uses reserved fields\n");
950 				return -EINVAL;
951 			}
952 
953 			/* check src operand */
954 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
955 			if (err)
956 				return err;
957 		} else {
958 			if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
959 				verbose("BPF_MOV uses reserved fields\n");
960 				return -EINVAL;
961 			}
962 		}
963 
964 		/* check dest operand */
965 		err = check_reg_arg(regs, insn->dst_reg, DST_OP);
966 		if (err)
967 			return err;
968 
969 		if (BPF_SRC(insn->code) == BPF_X) {
970 			if (BPF_CLASS(insn->code) == BPF_ALU64) {
971 				/* case: R1 = R2
972 				 * copy register state to dest reg
973 				 */
974 				regs[insn->dst_reg] = regs[insn->src_reg];
975 			} else {
976 				regs[insn->dst_reg].type = UNKNOWN_VALUE;
977 				regs[insn->dst_reg].map_ptr = NULL;
978 			}
979 		} else {
980 			/* case: R = imm
981 			 * remember the value we stored into this reg
982 			 */
983 			regs[insn->dst_reg].type = CONST_IMM;
984 			regs[insn->dst_reg].imm = insn->imm;
985 		}
986 
987 	} else if (opcode > BPF_END) {
988 		verbose("invalid BPF_ALU opcode %x\n", opcode);
989 		return -EINVAL;
990 
991 	} else {	/* all other ALU ops: and, sub, xor, add, ... */
992 
993 		bool stack_relative = false;
994 
995 		if (BPF_SRC(insn->code) == BPF_X) {
996 			if (insn->imm != 0 || insn->off != 0) {
997 				verbose("BPF_ALU uses reserved fields\n");
998 				return -EINVAL;
999 			}
1000 			/* check src1 operand */
1001 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1002 			if (err)
1003 				return err;
1004 		} else {
1005 			if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
1006 				verbose("BPF_ALU uses reserved fields\n");
1007 				return -EINVAL;
1008 			}
1009 		}
1010 
1011 		/* check src2 operand */
1012 		err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1013 		if (err)
1014 			return err;
1015 
1016 		if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
1017 		    BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
1018 			verbose("div by zero\n");
1019 			return -EINVAL;
1020 		}
1021 
1022 		/* pattern match 'bpf_add Rx, imm' instruction */
1023 		if (opcode == BPF_ADD && BPF_CLASS(insn->code) == BPF_ALU64 &&
1024 		    regs[insn->dst_reg].type == FRAME_PTR &&
1025 		    BPF_SRC(insn->code) == BPF_K)
1026 			stack_relative = true;
1027 
1028 		/* check dest operand */
1029 		err = check_reg_arg(regs, insn->dst_reg, DST_OP);
1030 		if (err)
1031 			return err;
1032 
1033 		if (stack_relative) {
1034 			regs[insn->dst_reg].type = PTR_TO_STACK;
1035 			regs[insn->dst_reg].imm = insn->imm;
1036 		}
1037 	}
1038 
1039 	return 0;
1040 }
1041 
1042 static int check_cond_jmp_op(struct verifier_env *env,
1043 			     struct bpf_insn *insn, int *insn_idx)
1044 {
1045 	struct reg_state *regs = env->cur_state.regs;
1046 	struct verifier_state *other_branch;
1047 	u8 opcode = BPF_OP(insn->code);
1048 	int err;
1049 
1050 	if (opcode > BPF_EXIT) {
1051 		verbose("invalid BPF_JMP opcode %x\n", opcode);
1052 		return -EINVAL;
1053 	}
1054 
1055 	if (BPF_SRC(insn->code) == BPF_X) {
1056 		if (insn->imm != 0) {
1057 			verbose("BPF_JMP uses reserved fields\n");
1058 			return -EINVAL;
1059 		}
1060 
1061 		/* check src1 operand */
1062 		err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1063 		if (err)
1064 			return err;
1065 	} else {
1066 		if (insn->src_reg != BPF_REG_0) {
1067 			verbose("BPF_JMP uses reserved fields\n");
1068 			return -EINVAL;
1069 		}
1070 	}
1071 
1072 	/* check src2 operand */
1073 	err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1074 	if (err)
1075 		return err;
1076 
1077 	/* detect if R == 0 where R was initialized to zero earlier */
1078 	if (BPF_SRC(insn->code) == BPF_K &&
1079 	    (opcode == BPF_JEQ || opcode == BPF_JNE) &&
1080 	    regs[insn->dst_reg].type == CONST_IMM &&
1081 	    regs[insn->dst_reg].imm == insn->imm) {
1082 		if (opcode == BPF_JEQ) {
1083 			/* if (imm == imm) goto pc+off;
1084 			 * only follow the goto, ignore fall-through
1085 			 */
1086 			*insn_idx += insn->off;
1087 			return 0;
1088 		} else {
1089 			/* if (imm != imm) goto pc+off;
1090 			 * only follow fall-through branch, since
1091 			 * that's where the program will go
1092 			 */
1093 			return 0;
1094 		}
1095 	}
1096 
1097 	other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
1098 	if (!other_branch)
1099 		return -EFAULT;
1100 
1101 	/* detect if R == 0 where R is returned value from bpf_map_lookup_elem() */
1102 	if (BPF_SRC(insn->code) == BPF_K &&
1103 	    insn->imm == 0 && (opcode == BPF_JEQ ||
1104 			       opcode == BPF_JNE) &&
1105 	    regs[insn->dst_reg].type == PTR_TO_MAP_VALUE_OR_NULL) {
1106 		if (opcode == BPF_JEQ) {
1107 			/* next fallthrough insn can access memory via
1108 			 * this register
1109 			 */
1110 			regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
1111 			/* branch targer cannot access it, since reg == 0 */
1112 			other_branch->regs[insn->dst_reg].type = CONST_IMM;
1113 			other_branch->regs[insn->dst_reg].imm = 0;
1114 		} else {
1115 			other_branch->regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
1116 			regs[insn->dst_reg].type = CONST_IMM;
1117 			regs[insn->dst_reg].imm = 0;
1118 		}
1119 	} else if (BPF_SRC(insn->code) == BPF_K &&
1120 		   (opcode == BPF_JEQ || opcode == BPF_JNE)) {
1121 
1122 		if (opcode == BPF_JEQ) {
1123 			/* detect if (R == imm) goto
1124 			 * and in the target state recognize that R = imm
1125 			 */
1126 			other_branch->regs[insn->dst_reg].type = CONST_IMM;
1127 			other_branch->regs[insn->dst_reg].imm = insn->imm;
1128 		} else {
1129 			/* detect if (R != imm) goto
1130 			 * and in the fall-through state recognize that R = imm
1131 			 */
1132 			regs[insn->dst_reg].type = CONST_IMM;
1133 			regs[insn->dst_reg].imm = insn->imm;
1134 		}
1135 	}
1136 	if (log_level)
1137 		print_verifier_state(env);
1138 	return 0;
1139 }
1140 
1141 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
1142 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
1143 {
1144 	u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
1145 
1146 	return (struct bpf_map *) (unsigned long) imm64;
1147 }
1148 
1149 /* verify BPF_LD_IMM64 instruction */
1150 static int check_ld_imm(struct verifier_env *env, struct bpf_insn *insn)
1151 {
1152 	struct reg_state *regs = env->cur_state.regs;
1153 	int err;
1154 
1155 	if (BPF_SIZE(insn->code) != BPF_DW) {
1156 		verbose("invalid BPF_LD_IMM insn\n");
1157 		return -EINVAL;
1158 	}
1159 	if (insn->off != 0) {
1160 		verbose("BPF_LD_IMM64 uses reserved fields\n");
1161 		return -EINVAL;
1162 	}
1163 
1164 	err = check_reg_arg(regs, insn->dst_reg, DST_OP);
1165 	if (err)
1166 		return err;
1167 
1168 	if (insn->src_reg == 0)
1169 		/* generic move 64-bit immediate into a register */
1170 		return 0;
1171 
1172 	/* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
1173 	BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
1174 
1175 	regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
1176 	regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
1177 	return 0;
1178 }
1179 
1180 static bool may_access_skb(enum bpf_prog_type type)
1181 {
1182 	switch (type) {
1183 	case BPF_PROG_TYPE_SOCKET_FILTER:
1184 	case BPF_PROG_TYPE_SCHED_CLS:
1185 	case BPF_PROG_TYPE_SCHED_ACT:
1186 		return true;
1187 	default:
1188 		return false;
1189 	}
1190 }
1191 
1192 /* verify safety of LD_ABS|LD_IND instructions:
1193  * - they can only appear in the programs where ctx == skb
1194  * - since they are wrappers of function calls, they scratch R1-R5 registers,
1195  *   preserve R6-R9, and store return value into R0
1196  *
1197  * Implicit input:
1198  *   ctx == skb == R6 == CTX
1199  *
1200  * Explicit input:
1201  *   SRC == any register
1202  *   IMM == 32-bit immediate
1203  *
1204  * Output:
1205  *   R0 - 8/16/32-bit skb data converted to cpu endianness
1206  */
1207 static int check_ld_abs(struct verifier_env *env, struct bpf_insn *insn)
1208 {
1209 	struct reg_state *regs = env->cur_state.regs;
1210 	u8 mode = BPF_MODE(insn->code);
1211 	struct reg_state *reg;
1212 	int i, err;
1213 
1214 	if (!may_access_skb(env->prog->type)) {
1215 		verbose("BPF_LD_ABS|IND instructions not allowed for this program type\n");
1216 		return -EINVAL;
1217 	}
1218 
1219 	if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
1220 	    (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
1221 		verbose("BPF_LD_ABS uses reserved fields\n");
1222 		return -EINVAL;
1223 	}
1224 
1225 	/* check whether implicit source operand (register R6) is readable */
1226 	err = check_reg_arg(regs, BPF_REG_6, SRC_OP);
1227 	if (err)
1228 		return err;
1229 
1230 	if (regs[BPF_REG_6].type != PTR_TO_CTX) {
1231 		verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
1232 		return -EINVAL;
1233 	}
1234 
1235 	if (mode == BPF_IND) {
1236 		/* check explicit source operand */
1237 		err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1238 		if (err)
1239 			return err;
1240 	}
1241 
1242 	/* reset caller saved regs to unreadable */
1243 	for (i = 0; i < CALLER_SAVED_REGS; i++) {
1244 		reg = regs + caller_saved[i];
1245 		reg->type = NOT_INIT;
1246 		reg->imm = 0;
1247 	}
1248 
1249 	/* mark destination R0 register as readable, since it contains
1250 	 * the value fetched from the packet
1251 	 */
1252 	regs[BPF_REG_0].type = UNKNOWN_VALUE;
1253 	return 0;
1254 }
1255 
1256 /* non-recursive DFS pseudo code
1257  * 1  procedure DFS-iterative(G,v):
1258  * 2      label v as discovered
1259  * 3      let S be a stack
1260  * 4      S.push(v)
1261  * 5      while S is not empty
1262  * 6            t <- S.pop()
1263  * 7            if t is what we're looking for:
1264  * 8                return t
1265  * 9            for all edges e in G.adjacentEdges(t) do
1266  * 10               if edge e is already labelled
1267  * 11                   continue with the next edge
1268  * 12               w <- G.adjacentVertex(t,e)
1269  * 13               if vertex w is not discovered and not explored
1270  * 14                   label e as tree-edge
1271  * 15                   label w as discovered
1272  * 16                   S.push(w)
1273  * 17                   continue at 5
1274  * 18               else if vertex w is discovered
1275  * 19                   label e as back-edge
1276  * 20               else
1277  * 21                   // vertex w is explored
1278  * 22                   label e as forward- or cross-edge
1279  * 23           label t as explored
1280  * 24           S.pop()
1281  *
1282  * convention:
1283  * 0x10 - discovered
1284  * 0x11 - discovered and fall-through edge labelled
1285  * 0x12 - discovered and fall-through and branch edges labelled
1286  * 0x20 - explored
1287  */
1288 
1289 enum {
1290 	DISCOVERED = 0x10,
1291 	EXPLORED = 0x20,
1292 	FALLTHROUGH = 1,
1293 	BRANCH = 2,
1294 };
1295 
1296 #define STATE_LIST_MARK ((struct verifier_state_list *) -1L)
1297 
1298 static int *insn_stack;	/* stack of insns to process */
1299 static int cur_stack;	/* current stack index */
1300 static int *insn_state;
1301 
1302 /* t, w, e - match pseudo-code above:
1303  * t - index of current instruction
1304  * w - next instruction
1305  * e - edge
1306  */
1307 static int push_insn(int t, int w, int e, struct verifier_env *env)
1308 {
1309 	if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
1310 		return 0;
1311 
1312 	if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
1313 		return 0;
1314 
1315 	if (w < 0 || w >= env->prog->len) {
1316 		verbose("jump out of range from insn %d to %d\n", t, w);
1317 		return -EINVAL;
1318 	}
1319 
1320 	if (e == BRANCH)
1321 		/* mark branch target for state pruning */
1322 		env->explored_states[w] = STATE_LIST_MARK;
1323 
1324 	if (insn_state[w] == 0) {
1325 		/* tree-edge */
1326 		insn_state[t] = DISCOVERED | e;
1327 		insn_state[w] = DISCOVERED;
1328 		if (cur_stack >= env->prog->len)
1329 			return -E2BIG;
1330 		insn_stack[cur_stack++] = w;
1331 		return 1;
1332 	} else if ((insn_state[w] & 0xF0) == DISCOVERED) {
1333 		verbose("back-edge from insn %d to %d\n", t, w);
1334 		return -EINVAL;
1335 	} else if (insn_state[w] == EXPLORED) {
1336 		/* forward- or cross-edge */
1337 		insn_state[t] = DISCOVERED | e;
1338 	} else {
1339 		verbose("insn state internal bug\n");
1340 		return -EFAULT;
1341 	}
1342 	return 0;
1343 }
1344 
1345 /* non-recursive depth-first-search to detect loops in BPF program
1346  * loop == back-edge in directed graph
1347  */
1348 static int check_cfg(struct verifier_env *env)
1349 {
1350 	struct bpf_insn *insns = env->prog->insnsi;
1351 	int insn_cnt = env->prog->len;
1352 	int ret = 0;
1353 	int i, t;
1354 
1355 	insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
1356 	if (!insn_state)
1357 		return -ENOMEM;
1358 
1359 	insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
1360 	if (!insn_stack) {
1361 		kfree(insn_state);
1362 		return -ENOMEM;
1363 	}
1364 
1365 	insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
1366 	insn_stack[0] = 0; /* 0 is the first instruction */
1367 	cur_stack = 1;
1368 
1369 peek_stack:
1370 	if (cur_stack == 0)
1371 		goto check_state;
1372 	t = insn_stack[cur_stack - 1];
1373 
1374 	if (BPF_CLASS(insns[t].code) == BPF_JMP) {
1375 		u8 opcode = BPF_OP(insns[t].code);
1376 
1377 		if (opcode == BPF_EXIT) {
1378 			goto mark_explored;
1379 		} else if (opcode == BPF_CALL) {
1380 			ret = push_insn(t, t + 1, FALLTHROUGH, env);
1381 			if (ret == 1)
1382 				goto peek_stack;
1383 			else if (ret < 0)
1384 				goto err_free;
1385 		} else if (opcode == BPF_JA) {
1386 			if (BPF_SRC(insns[t].code) != BPF_K) {
1387 				ret = -EINVAL;
1388 				goto err_free;
1389 			}
1390 			/* unconditional jump with single edge */
1391 			ret = push_insn(t, t + insns[t].off + 1,
1392 					FALLTHROUGH, env);
1393 			if (ret == 1)
1394 				goto peek_stack;
1395 			else if (ret < 0)
1396 				goto err_free;
1397 			/* tell verifier to check for equivalent states
1398 			 * after every call and jump
1399 			 */
1400 			if (t + 1 < insn_cnt)
1401 				env->explored_states[t + 1] = STATE_LIST_MARK;
1402 		} else {
1403 			/* conditional jump with two edges */
1404 			ret = push_insn(t, t + 1, FALLTHROUGH, env);
1405 			if (ret == 1)
1406 				goto peek_stack;
1407 			else if (ret < 0)
1408 				goto err_free;
1409 
1410 			ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
1411 			if (ret == 1)
1412 				goto peek_stack;
1413 			else if (ret < 0)
1414 				goto err_free;
1415 		}
1416 	} else {
1417 		/* all other non-branch instructions with single
1418 		 * fall-through edge
1419 		 */
1420 		ret = push_insn(t, t + 1, FALLTHROUGH, env);
1421 		if (ret == 1)
1422 			goto peek_stack;
1423 		else if (ret < 0)
1424 			goto err_free;
1425 	}
1426 
1427 mark_explored:
1428 	insn_state[t] = EXPLORED;
1429 	if (cur_stack-- <= 0) {
1430 		verbose("pop stack internal bug\n");
1431 		ret = -EFAULT;
1432 		goto err_free;
1433 	}
1434 	goto peek_stack;
1435 
1436 check_state:
1437 	for (i = 0; i < insn_cnt; i++) {
1438 		if (insn_state[i] != EXPLORED) {
1439 			verbose("unreachable insn %d\n", i);
1440 			ret = -EINVAL;
1441 			goto err_free;
1442 		}
1443 	}
1444 	ret = 0; /* cfg looks good */
1445 
1446 err_free:
1447 	kfree(insn_state);
1448 	kfree(insn_stack);
1449 	return ret;
1450 }
1451 
1452 /* compare two verifier states
1453  *
1454  * all states stored in state_list are known to be valid, since
1455  * verifier reached 'bpf_exit' instruction through them
1456  *
1457  * this function is called when verifier exploring different branches of
1458  * execution popped from the state stack. If it sees an old state that has
1459  * more strict register state and more strict stack state then this execution
1460  * branch doesn't need to be explored further, since verifier already
1461  * concluded that more strict state leads to valid finish.
1462  *
1463  * Therefore two states are equivalent if register state is more conservative
1464  * and explored stack state is more conservative than the current one.
1465  * Example:
1466  *       explored                   current
1467  * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
1468  * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
1469  *
1470  * In other words if current stack state (one being explored) has more
1471  * valid slots than old one that already passed validation, it means
1472  * the verifier can stop exploring and conclude that current state is valid too
1473  *
1474  * Similarly with registers. If explored state has register type as invalid
1475  * whereas register type in current state is meaningful, it means that
1476  * the current state will reach 'bpf_exit' instruction safely
1477  */
1478 static bool states_equal(struct verifier_state *old, struct verifier_state *cur)
1479 {
1480 	int i;
1481 
1482 	for (i = 0; i < MAX_BPF_REG; i++) {
1483 		if (memcmp(&old->regs[i], &cur->regs[i],
1484 			   sizeof(old->regs[0])) != 0) {
1485 			if (old->regs[i].type == NOT_INIT ||
1486 			    (old->regs[i].type == UNKNOWN_VALUE &&
1487 			     cur->regs[i].type != NOT_INIT))
1488 				continue;
1489 			return false;
1490 		}
1491 	}
1492 
1493 	for (i = 0; i < MAX_BPF_STACK; i++) {
1494 		if (old->stack_slot_type[i] == STACK_INVALID)
1495 			continue;
1496 		if (old->stack_slot_type[i] != cur->stack_slot_type[i])
1497 			/* Ex: old explored (safe) state has STACK_SPILL in
1498 			 * this stack slot, but current has has STACK_MISC ->
1499 			 * this verifier states are not equivalent,
1500 			 * return false to continue verification of this path
1501 			 */
1502 			return false;
1503 		if (i % BPF_REG_SIZE)
1504 			continue;
1505 		if (memcmp(&old->spilled_regs[i / BPF_REG_SIZE],
1506 			   &cur->spilled_regs[i / BPF_REG_SIZE],
1507 			   sizeof(old->spilled_regs[0])))
1508 			/* when explored and current stack slot types are
1509 			 * the same, check that stored pointers types
1510 			 * are the same as well.
1511 			 * Ex: explored safe path could have stored
1512 			 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -8}
1513 			 * but current path has stored:
1514 			 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -16}
1515 			 * such verifier states are not equivalent.
1516 			 * return false to continue verification of this path
1517 			 */
1518 			return false;
1519 		else
1520 			continue;
1521 	}
1522 	return true;
1523 }
1524 
1525 static int is_state_visited(struct verifier_env *env, int insn_idx)
1526 {
1527 	struct verifier_state_list *new_sl;
1528 	struct verifier_state_list *sl;
1529 
1530 	sl = env->explored_states[insn_idx];
1531 	if (!sl)
1532 		/* this 'insn_idx' instruction wasn't marked, so we will not
1533 		 * be doing state search here
1534 		 */
1535 		return 0;
1536 
1537 	while (sl != STATE_LIST_MARK) {
1538 		if (states_equal(&sl->state, &env->cur_state))
1539 			/* reached equivalent register/stack state,
1540 			 * prune the search
1541 			 */
1542 			return 1;
1543 		sl = sl->next;
1544 	}
1545 
1546 	/* there were no equivalent states, remember current one.
1547 	 * technically the current state is not proven to be safe yet,
1548 	 * but it will either reach bpf_exit (which means it's safe) or
1549 	 * it will be rejected. Since there are no loops, we won't be
1550 	 * seeing this 'insn_idx' instruction again on the way to bpf_exit
1551 	 */
1552 	new_sl = kmalloc(sizeof(struct verifier_state_list), GFP_USER);
1553 	if (!new_sl)
1554 		return -ENOMEM;
1555 
1556 	/* add new state to the head of linked list */
1557 	memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
1558 	new_sl->next = env->explored_states[insn_idx];
1559 	env->explored_states[insn_idx] = new_sl;
1560 	return 0;
1561 }
1562 
1563 static int do_check(struct verifier_env *env)
1564 {
1565 	struct verifier_state *state = &env->cur_state;
1566 	struct bpf_insn *insns = env->prog->insnsi;
1567 	struct reg_state *regs = state->regs;
1568 	int insn_cnt = env->prog->len;
1569 	int insn_idx, prev_insn_idx = 0;
1570 	int insn_processed = 0;
1571 	bool do_print_state = false;
1572 
1573 	init_reg_state(regs);
1574 	insn_idx = 0;
1575 	for (;;) {
1576 		struct bpf_insn *insn;
1577 		u8 class;
1578 		int err;
1579 
1580 		if (insn_idx >= insn_cnt) {
1581 			verbose("invalid insn idx %d insn_cnt %d\n",
1582 				insn_idx, insn_cnt);
1583 			return -EFAULT;
1584 		}
1585 
1586 		insn = &insns[insn_idx];
1587 		class = BPF_CLASS(insn->code);
1588 
1589 		if (++insn_processed > 32768) {
1590 			verbose("BPF program is too large. Proccessed %d insn\n",
1591 				insn_processed);
1592 			return -E2BIG;
1593 		}
1594 
1595 		err = is_state_visited(env, insn_idx);
1596 		if (err < 0)
1597 			return err;
1598 		if (err == 1) {
1599 			/* found equivalent state, can prune the search */
1600 			if (log_level) {
1601 				if (do_print_state)
1602 					verbose("\nfrom %d to %d: safe\n",
1603 						prev_insn_idx, insn_idx);
1604 				else
1605 					verbose("%d: safe\n", insn_idx);
1606 			}
1607 			goto process_bpf_exit;
1608 		}
1609 
1610 		if (log_level && do_print_state) {
1611 			verbose("\nfrom %d to %d:", prev_insn_idx, insn_idx);
1612 			print_verifier_state(env);
1613 			do_print_state = false;
1614 		}
1615 
1616 		if (log_level) {
1617 			verbose("%d: ", insn_idx);
1618 			print_bpf_insn(insn);
1619 		}
1620 
1621 		if (class == BPF_ALU || class == BPF_ALU64) {
1622 			err = check_alu_op(regs, insn);
1623 			if (err)
1624 				return err;
1625 
1626 		} else if (class == BPF_LDX) {
1627 			enum bpf_reg_type src_reg_type;
1628 
1629 			/* check for reserved fields is already done */
1630 
1631 			/* check src operand */
1632 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1633 			if (err)
1634 				return err;
1635 
1636 			err = check_reg_arg(regs, insn->dst_reg, DST_OP_NO_MARK);
1637 			if (err)
1638 				return err;
1639 
1640 			src_reg_type = regs[insn->src_reg].type;
1641 
1642 			/* check that memory (src_reg + off) is readable,
1643 			 * the state of dst_reg will be updated by this func
1644 			 */
1645 			err = check_mem_access(env, insn->src_reg, insn->off,
1646 					       BPF_SIZE(insn->code), BPF_READ,
1647 					       insn->dst_reg);
1648 			if (err)
1649 				return err;
1650 
1651 			if (BPF_SIZE(insn->code) != BPF_W) {
1652 				insn_idx++;
1653 				continue;
1654 			}
1655 
1656 			if (insn->imm == 0) {
1657 				/* saw a valid insn
1658 				 * dst_reg = *(u32 *)(src_reg + off)
1659 				 * use reserved 'imm' field to mark this insn
1660 				 */
1661 				insn->imm = src_reg_type;
1662 
1663 			} else if (src_reg_type != insn->imm &&
1664 				   (src_reg_type == PTR_TO_CTX ||
1665 				    insn->imm == PTR_TO_CTX)) {
1666 				/* ABuser program is trying to use the same insn
1667 				 * dst_reg = *(u32*) (src_reg + off)
1668 				 * with different pointer types:
1669 				 * src_reg == ctx in one branch and
1670 				 * src_reg == stack|map in some other branch.
1671 				 * Reject it.
1672 				 */
1673 				verbose("same insn cannot be used with different pointers\n");
1674 				return -EINVAL;
1675 			}
1676 
1677 		} else if (class == BPF_STX) {
1678 			if (BPF_MODE(insn->code) == BPF_XADD) {
1679 				err = check_xadd(env, insn);
1680 				if (err)
1681 					return err;
1682 				insn_idx++;
1683 				continue;
1684 			}
1685 
1686 			if (BPF_MODE(insn->code) != BPF_MEM ||
1687 			    insn->imm != 0) {
1688 				verbose("BPF_STX uses reserved fields\n");
1689 				return -EINVAL;
1690 			}
1691 			/* check src1 operand */
1692 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1693 			if (err)
1694 				return err;
1695 			/* check src2 operand */
1696 			err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1697 			if (err)
1698 				return err;
1699 
1700 			/* check that memory (dst_reg + off) is writeable */
1701 			err = check_mem_access(env, insn->dst_reg, insn->off,
1702 					       BPF_SIZE(insn->code), BPF_WRITE,
1703 					       insn->src_reg);
1704 			if (err)
1705 				return err;
1706 
1707 		} else if (class == BPF_ST) {
1708 			if (BPF_MODE(insn->code) != BPF_MEM ||
1709 			    insn->src_reg != BPF_REG_0) {
1710 				verbose("BPF_ST uses reserved fields\n");
1711 				return -EINVAL;
1712 			}
1713 			/* check src operand */
1714 			err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1715 			if (err)
1716 				return err;
1717 
1718 			/* check that memory (dst_reg + off) is writeable */
1719 			err = check_mem_access(env, insn->dst_reg, insn->off,
1720 					       BPF_SIZE(insn->code), BPF_WRITE,
1721 					       -1);
1722 			if (err)
1723 				return err;
1724 
1725 		} else if (class == BPF_JMP) {
1726 			u8 opcode = BPF_OP(insn->code);
1727 
1728 			if (opcode == BPF_CALL) {
1729 				if (BPF_SRC(insn->code) != BPF_K ||
1730 				    insn->off != 0 ||
1731 				    insn->src_reg != BPF_REG_0 ||
1732 				    insn->dst_reg != BPF_REG_0) {
1733 					verbose("BPF_CALL uses reserved fields\n");
1734 					return -EINVAL;
1735 				}
1736 
1737 				err = check_call(env, insn->imm);
1738 				if (err)
1739 					return err;
1740 
1741 			} else if (opcode == BPF_JA) {
1742 				if (BPF_SRC(insn->code) != BPF_K ||
1743 				    insn->imm != 0 ||
1744 				    insn->src_reg != BPF_REG_0 ||
1745 				    insn->dst_reg != BPF_REG_0) {
1746 					verbose("BPF_JA uses reserved fields\n");
1747 					return -EINVAL;
1748 				}
1749 
1750 				insn_idx += insn->off + 1;
1751 				continue;
1752 
1753 			} else if (opcode == BPF_EXIT) {
1754 				if (BPF_SRC(insn->code) != BPF_K ||
1755 				    insn->imm != 0 ||
1756 				    insn->src_reg != BPF_REG_0 ||
1757 				    insn->dst_reg != BPF_REG_0) {
1758 					verbose("BPF_EXIT uses reserved fields\n");
1759 					return -EINVAL;
1760 				}
1761 
1762 				/* eBPF calling convetion is such that R0 is used
1763 				 * to return the value from eBPF program.
1764 				 * Make sure that it's readable at this time
1765 				 * of bpf_exit, which means that program wrote
1766 				 * something into it earlier
1767 				 */
1768 				err = check_reg_arg(regs, BPF_REG_0, SRC_OP);
1769 				if (err)
1770 					return err;
1771 
1772 process_bpf_exit:
1773 				insn_idx = pop_stack(env, &prev_insn_idx);
1774 				if (insn_idx < 0) {
1775 					break;
1776 				} else {
1777 					do_print_state = true;
1778 					continue;
1779 				}
1780 			} else {
1781 				err = check_cond_jmp_op(env, insn, &insn_idx);
1782 				if (err)
1783 					return err;
1784 			}
1785 		} else if (class == BPF_LD) {
1786 			u8 mode = BPF_MODE(insn->code);
1787 
1788 			if (mode == BPF_ABS || mode == BPF_IND) {
1789 				err = check_ld_abs(env, insn);
1790 				if (err)
1791 					return err;
1792 
1793 			} else if (mode == BPF_IMM) {
1794 				err = check_ld_imm(env, insn);
1795 				if (err)
1796 					return err;
1797 
1798 				insn_idx++;
1799 			} else {
1800 				verbose("invalid BPF_LD mode\n");
1801 				return -EINVAL;
1802 			}
1803 		} else {
1804 			verbose("unknown insn class %d\n", class);
1805 			return -EINVAL;
1806 		}
1807 
1808 		insn_idx++;
1809 	}
1810 
1811 	return 0;
1812 }
1813 
1814 /* look for pseudo eBPF instructions that access map FDs and
1815  * replace them with actual map pointers
1816  */
1817 static int replace_map_fd_with_map_ptr(struct verifier_env *env)
1818 {
1819 	struct bpf_insn *insn = env->prog->insnsi;
1820 	int insn_cnt = env->prog->len;
1821 	int i, j;
1822 
1823 	for (i = 0; i < insn_cnt; i++, insn++) {
1824 		if (BPF_CLASS(insn->code) == BPF_LDX &&
1825 		    (BPF_MODE(insn->code) != BPF_MEM ||
1826 		     insn->imm != 0)) {
1827 			verbose("BPF_LDX uses reserved fields\n");
1828 			return -EINVAL;
1829 		}
1830 
1831 		if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
1832 			struct bpf_map *map;
1833 			struct fd f;
1834 
1835 			if (i == insn_cnt - 1 || insn[1].code != 0 ||
1836 			    insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
1837 			    insn[1].off != 0) {
1838 				verbose("invalid bpf_ld_imm64 insn\n");
1839 				return -EINVAL;
1840 			}
1841 
1842 			if (insn->src_reg == 0)
1843 				/* valid generic load 64-bit imm */
1844 				goto next_insn;
1845 
1846 			if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
1847 				verbose("unrecognized bpf_ld_imm64 insn\n");
1848 				return -EINVAL;
1849 			}
1850 
1851 			f = fdget(insn->imm);
1852 
1853 			map = bpf_map_get(f);
1854 			if (IS_ERR(map)) {
1855 				verbose("fd %d is not pointing to valid bpf_map\n",
1856 					insn->imm);
1857 				fdput(f);
1858 				return PTR_ERR(map);
1859 			}
1860 
1861 			/* store map pointer inside BPF_LD_IMM64 instruction */
1862 			insn[0].imm = (u32) (unsigned long) map;
1863 			insn[1].imm = ((u64) (unsigned long) map) >> 32;
1864 
1865 			/* check whether we recorded this map already */
1866 			for (j = 0; j < env->used_map_cnt; j++)
1867 				if (env->used_maps[j] == map) {
1868 					fdput(f);
1869 					goto next_insn;
1870 				}
1871 
1872 			if (env->used_map_cnt >= MAX_USED_MAPS) {
1873 				fdput(f);
1874 				return -E2BIG;
1875 			}
1876 
1877 			/* remember this map */
1878 			env->used_maps[env->used_map_cnt++] = map;
1879 
1880 			/* hold the map. If the program is rejected by verifier,
1881 			 * the map will be released by release_maps() or it
1882 			 * will be used by the valid program until it's unloaded
1883 			 * and all maps are released in free_bpf_prog_info()
1884 			 */
1885 			atomic_inc(&map->refcnt);
1886 
1887 			fdput(f);
1888 next_insn:
1889 			insn++;
1890 			i++;
1891 		}
1892 	}
1893 
1894 	/* now all pseudo BPF_LD_IMM64 instructions load valid
1895 	 * 'struct bpf_map *' into a register instead of user map_fd.
1896 	 * These pointers will be used later by verifier to validate map access.
1897 	 */
1898 	return 0;
1899 }
1900 
1901 /* drop refcnt of maps used by the rejected program */
1902 static void release_maps(struct verifier_env *env)
1903 {
1904 	int i;
1905 
1906 	for (i = 0; i < env->used_map_cnt; i++)
1907 		bpf_map_put(env->used_maps[i]);
1908 }
1909 
1910 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
1911 static void convert_pseudo_ld_imm64(struct verifier_env *env)
1912 {
1913 	struct bpf_insn *insn = env->prog->insnsi;
1914 	int insn_cnt = env->prog->len;
1915 	int i;
1916 
1917 	for (i = 0; i < insn_cnt; i++, insn++)
1918 		if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
1919 			insn->src_reg = 0;
1920 }
1921 
1922 static void adjust_branches(struct bpf_prog *prog, int pos, int delta)
1923 {
1924 	struct bpf_insn *insn = prog->insnsi;
1925 	int insn_cnt = prog->len;
1926 	int i;
1927 
1928 	for (i = 0; i < insn_cnt; i++, insn++) {
1929 		if (BPF_CLASS(insn->code) != BPF_JMP ||
1930 		    BPF_OP(insn->code) == BPF_CALL ||
1931 		    BPF_OP(insn->code) == BPF_EXIT)
1932 			continue;
1933 
1934 		/* adjust offset of jmps if necessary */
1935 		if (i < pos && i + insn->off + 1 > pos)
1936 			insn->off += delta;
1937 		else if (i > pos && i + insn->off + 1 < pos)
1938 			insn->off -= delta;
1939 	}
1940 }
1941 
1942 /* convert load instructions that access fields of 'struct __sk_buff'
1943  * into sequence of instructions that access fields of 'struct sk_buff'
1944  */
1945 static int convert_ctx_accesses(struct verifier_env *env)
1946 {
1947 	struct bpf_insn *insn = env->prog->insnsi;
1948 	int insn_cnt = env->prog->len;
1949 	struct bpf_insn insn_buf[16];
1950 	struct bpf_prog *new_prog;
1951 	u32 cnt;
1952 	int i;
1953 
1954 	if (!env->prog->aux->ops->convert_ctx_access)
1955 		return 0;
1956 
1957 	for (i = 0; i < insn_cnt; i++, insn++) {
1958 		if (insn->code != (BPF_LDX | BPF_MEM | BPF_W))
1959 			continue;
1960 
1961 		if (insn->imm != PTR_TO_CTX) {
1962 			/* clear internal mark */
1963 			insn->imm = 0;
1964 			continue;
1965 		}
1966 
1967 		cnt = env->prog->aux->ops->
1968 			convert_ctx_access(insn->dst_reg, insn->src_reg,
1969 					   insn->off, insn_buf);
1970 		if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
1971 			verbose("bpf verifier is misconfigured\n");
1972 			return -EINVAL;
1973 		}
1974 
1975 		if (cnt == 1) {
1976 			memcpy(insn, insn_buf, sizeof(*insn));
1977 			continue;
1978 		}
1979 
1980 		/* several new insns need to be inserted. Make room for them */
1981 		insn_cnt += cnt - 1;
1982 		new_prog = bpf_prog_realloc(env->prog,
1983 					    bpf_prog_size(insn_cnt),
1984 					    GFP_USER);
1985 		if (!new_prog)
1986 			return -ENOMEM;
1987 
1988 		new_prog->len = insn_cnt;
1989 
1990 		memmove(new_prog->insnsi + i + cnt, new_prog->insns + i + 1,
1991 			sizeof(*insn) * (insn_cnt - i - cnt));
1992 
1993 		/* copy substitute insns in place of load instruction */
1994 		memcpy(new_prog->insnsi + i, insn_buf, sizeof(*insn) * cnt);
1995 
1996 		/* adjust branches in the whole program */
1997 		adjust_branches(new_prog, i, cnt - 1);
1998 
1999 		/* keep walking new program and skip insns we just inserted */
2000 		env->prog = new_prog;
2001 		insn = new_prog->insnsi + i + cnt - 1;
2002 		i += cnt - 1;
2003 	}
2004 
2005 	return 0;
2006 }
2007 
2008 static void free_states(struct verifier_env *env)
2009 {
2010 	struct verifier_state_list *sl, *sln;
2011 	int i;
2012 
2013 	if (!env->explored_states)
2014 		return;
2015 
2016 	for (i = 0; i < env->prog->len; i++) {
2017 		sl = env->explored_states[i];
2018 
2019 		if (sl)
2020 			while (sl != STATE_LIST_MARK) {
2021 				sln = sl->next;
2022 				kfree(sl);
2023 				sl = sln;
2024 			}
2025 	}
2026 
2027 	kfree(env->explored_states);
2028 }
2029 
2030 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
2031 {
2032 	char __user *log_ubuf = NULL;
2033 	struct verifier_env *env;
2034 	int ret = -EINVAL;
2035 
2036 	if ((*prog)->len <= 0 || (*prog)->len > BPF_MAXINSNS)
2037 		return -E2BIG;
2038 
2039 	/* 'struct verifier_env' can be global, but since it's not small,
2040 	 * allocate/free it every time bpf_check() is called
2041 	 */
2042 	env = kzalloc(sizeof(struct verifier_env), GFP_KERNEL);
2043 	if (!env)
2044 		return -ENOMEM;
2045 
2046 	env->prog = *prog;
2047 
2048 	/* grab the mutex to protect few globals used by verifier */
2049 	mutex_lock(&bpf_verifier_lock);
2050 
2051 	if (attr->log_level || attr->log_buf || attr->log_size) {
2052 		/* user requested verbose verifier output
2053 		 * and supplied buffer to store the verification trace
2054 		 */
2055 		log_level = attr->log_level;
2056 		log_ubuf = (char __user *) (unsigned long) attr->log_buf;
2057 		log_size = attr->log_size;
2058 		log_len = 0;
2059 
2060 		ret = -EINVAL;
2061 		/* log_* values have to be sane */
2062 		if (log_size < 128 || log_size > UINT_MAX >> 8 ||
2063 		    log_level == 0 || log_ubuf == NULL)
2064 			goto free_env;
2065 
2066 		ret = -ENOMEM;
2067 		log_buf = vmalloc(log_size);
2068 		if (!log_buf)
2069 			goto free_env;
2070 	} else {
2071 		log_level = 0;
2072 	}
2073 
2074 	ret = replace_map_fd_with_map_ptr(env);
2075 	if (ret < 0)
2076 		goto skip_full_check;
2077 
2078 	env->explored_states = kcalloc(env->prog->len,
2079 				       sizeof(struct verifier_state_list *),
2080 				       GFP_USER);
2081 	ret = -ENOMEM;
2082 	if (!env->explored_states)
2083 		goto skip_full_check;
2084 
2085 	ret = check_cfg(env);
2086 	if (ret < 0)
2087 		goto skip_full_check;
2088 
2089 	ret = do_check(env);
2090 
2091 skip_full_check:
2092 	while (pop_stack(env, NULL) >= 0);
2093 	free_states(env);
2094 
2095 	if (ret == 0)
2096 		/* program is valid, convert *(u32*)(ctx + off) accesses */
2097 		ret = convert_ctx_accesses(env);
2098 
2099 	if (log_level && log_len >= log_size - 1) {
2100 		BUG_ON(log_len >= log_size);
2101 		/* verifier log exceeded user supplied buffer */
2102 		ret = -ENOSPC;
2103 		/* fall through to return what was recorded */
2104 	}
2105 
2106 	/* copy verifier log back to user space including trailing zero */
2107 	if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
2108 		ret = -EFAULT;
2109 		goto free_log_buf;
2110 	}
2111 
2112 	if (ret == 0 && env->used_map_cnt) {
2113 		/* if program passed verifier, update used_maps in bpf_prog_info */
2114 		env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
2115 							  sizeof(env->used_maps[0]),
2116 							  GFP_KERNEL);
2117 
2118 		if (!env->prog->aux->used_maps) {
2119 			ret = -ENOMEM;
2120 			goto free_log_buf;
2121 		}
2122 
2123 		memcpy(env->prog->aux->used_maps, env->used_maps,
2124 		       sizeof(env->used_maps[0]) * env->used_map_cnt);
2125 		env->prog->aux->used_map_cnt = env->used_map_cnt;
2126 
2127 		/* program is valid. Convert pseudo bpf_ld_imm64 into generic
2128 		 * bpf_ld_imm64 instructions
2129 		 */
2130 		convert_pseudo_ld_imm64(env);
2131 	}
2132 
2133 free_log_buf:
2134 	if (log_level)
2135 		vfree(log_buf);
2136 free_env:
2137 	if (!env->prog->aux->used_maps)
2138 		/* if we didn't copy map pointers into bpf_prog_info, release
2139 		 * them now. Otherwise free_bpf_prog_info() will release them.
2140 		 */
2141 		release_maps(env);
2142 	*prog = env->prog;
2143 	kfree(env);
2144 	mutex_unlock(&bpf_verifier_lock);
2145 	return ret;
2146 }
2147