1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
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.
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.
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>
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.
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
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
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
48 * At the start of BPF program the register R1 contains a pointer to bpf_context
49 * and has type PTR_TO_CTX.
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.
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)
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.
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.
72 * registers used to pass values to function calls are checked against
73 * function argument constraints.
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'
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,
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.
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)
93 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
94 * void *key = (void *) (unsigned long) r2;
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.
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
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.
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().
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.
128 /* types of values stored in eBPF registers */
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 */
142 enum bpf_reg_type type;
144 /* valid when type == CONST_IMM | PTR_TO_STACK */
147 /* valid when type == CONST_PTR_TO_MAP | PTR_TO_MAP_VALUE |
148 * PTR_TO_MAP_VALUE_OR_NULL
150 struct bpf_map *map_ptr;
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 */
160 #define BPF_REG_SIZE 8 /* size of eBPF register in bytes */
162 /* state of the program:
163 * type of all registers and stack info
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];
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;
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'
183 struct verifier_state st;
186 struct verifier_stack_elem *next;
189 #define MAX_USED_MAPS 64 /* max number of maps accessed by one eBPF program */
191 /* single container for all structs
192 * one verifier_env per bpf_check() call
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 */
204 /* verbose verifier prints what it's seeing
205 * bpf_check() is called under lock, so no race to access these global vars
207 static u32 log_level, log_size, log_len;
208 static char *log_buf;
210 static DEFINE_MUTEX(bpf_verifier_lock);
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
216 static void verbose(const char *fmt, ...)
220 if (log_level == 0 || log_len >= log_size - 1)
224 log_len += vscnprintf(log_buf + log_len, log_size - log_len, fmt, args);
228 /* string representation of 'enum bpf_reg_type' */
229 static const char * const reg_type_str[] = {
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",
237 [PTR_TO_STACK] = "fp",
241 static void print_verifier_state(struct verifier_env *env)
246 for (i = 0; i < MAX_BPF_REG; i++) {
247 t = env->cur_state.regs[i].type;
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);
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]);
267 static const char *const bpf_class_string[] = {
275 [BPF_ALU64] = "alu64",
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",
295 static const char *const bpf_ldst_string[] = {
296 [BPF_W >> 3] = "u32",
297 [BPF_H >> 3] = "u16",
299 [BPF_DW >> 3] = "u64",
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",
315 static void print_bpf_insn(struct bpf_insn *insn)
317 u8 class = BPF_CLASS(insn->code);
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) " : "",
324 bpf_alu_string[BPF_OP(insn->code) >> 4],
325 class == BPF_ALU ? "(u32) " : "",
328 verbose("(%02x) %sr%d %s %s%d\n",
329 insn->code, class == BPF_ALU ? "(u32) " : "",
331 bpf_alu_string[BPF_OP(insn->code) >> 4],
332 class == BPF_ALU ? "(u32) " : "",
334 } else if (class == BPF_STX) {
335 if (BPF_MODE(insn->code) == BPF_MEM)
336 verbose("(%02x) *(%s *)(r%d %+d) = r%d\n",
338 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
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",
344 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
345 insn->dst_reg, insn->off,
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);
354 verbose("(%02x) *(%s *)(r%d %+d) = %d\n",
356 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
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);
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",
372 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
374 } else if (BPF_MODE(insn->code) == BPF_IND) {
375 verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n",
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);
383 verbose("BUG_ld_%02x\n", insn->code);
386 } else if (class == BPF_JMP) {
387 u8 opcode = BPF_OP(insn->code);
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);
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);
408 verbose("(%02x) %s\n", insn->code, bpf_class_string[class]);
412 static int pop_stack(struct verifier_env *env, int *prev_insn_idx)
414 struct verifier_stack_elem *elem;
417 if (env->head == NULL)
420 memcpy(&env->cur_state, &env->head->st, sizeof(env->cur_state));
421 insn_idx = env->head->insn_idx;
423 *prev_insn_idx = env->head->prev_insn_idx;
424 elem = env->head->next;
431 static struct verifier_state *push_stack(struct verifier_env *env, int insn_idx,
434 struct verifier_stack_elem *elem;
436 elem = kmalloc(sizeof(struct verifier_stack_elem), GFP_KERNEL);
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;
446 if (env->stack_size > 1024) {
447 verbose("BPF program is too complex\n");
452 /* pop all elements and return */
453 while (pop_stack(env, NULL) >= 0);
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
462 static void init_reg_state(struct reg_state *regs)
466 for (i = 0; i < MAX_BPF_REG; i++) {
467 regs[i].type = NOT_INIT;
469 regs[i].map_ptr = NULL;
473 regs[BPF_REG_FP].type = FRAME_PTR;
475 /* 1st arg to a function */
476 regs[BPF_REG_1].type = PTR_TO_CTX;
479 static void mark_reg_unknown_value(struct reg_state *regs, u32 regno)
481 BUG_ON(regno >= MAX_BPF_REG);
482 regs[regno].type = UNKNOWN_VALUE;
484 regs[regno].map_ptr = NULL;
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 */
493 static int check_reg_arg(struct reg_state *regs, u32 regno,
496 if (regno >= MAX_BPF_REG) {
497 verbose("R%d is invalid\n", regno);
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);
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");
514 mark_reg_unknown_value(regs, regno);
519 static int bpf_size_to_bytes(int bpf_size)
521 if (bpf_size == BPF_W)
523 else if (bpf_size == BPF_H)
525 else if (bpf_size == BPF_B)
527 else if (bpf_size == BPF_DW)
533 /* check_stack_read/write functions track spill/fill of registers,
534 * stack boundary and alignment are checked in check_mem_access()
536 static int check_stack_write(struct verifier_state *state, int off, int size,
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
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)) {
549 /* register containing pointer is being spilled into stack */
550 if (size != BPF_REG_SIZE) {
551 verbose("invalid size of register spill\n");
555 /* save register state */
556 state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] =
557 state->regs[value_regno];
559 for (i = 0; i < BPF_REG_SIZE; i++)
560 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL;
562 /* regular write of data into stack */
563 state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] =
564 (struct reg_state) {};
566 for (i = 0; i < size; i++)
567 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC;
572 static int check_stack_read(struct verifier_state *state, int off, int size,
578 slot_type = &state->stack_slot_type[MAX_BPF_STACK + off];
580 if (slot_type[0] == STACK_SPILL) {
581 if (size != BPF_REG_SIZE) {
582 verbose("invalid size of register spill\n");
585 for (i = 1; i < BPF_REG_SIZE; i++) {
586 if (slot_type[i] != STACK_SPILL) {
587 verbose("corrupted spill memory\n");
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];
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",
605 if (value_regno >= 0)
606 /* have read misc data from the stack */
607 mark_reg_unknown_value(state->regs, value_regno);
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,
616 struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
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);
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)
630 if (env->prog->aux->ops->is_valid_access &&
631 env->prog->aux->ops->is_valid_access(off, size, t))
634 verbose("invalid bpf_context access off=%d size=%d\n", off, size);
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
644 static int check_mem_access(struct verifier_env *env, u32 regno, int off,
645 int bpf_size, enum bpf_access_type t,
648 struct verifier_state *state = &env->cur_state;
651 size = bpf_size_to_bytes(bpf_size);
655 if (off % size != 0) {
656 verbose("misaligned access off %d size %d\n", off, size);
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);
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);
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);
676 err = check_stack_write(state, off, size, value_regno);
678 err = check_stack_read(state, off, size, value_regno);
680 verbose("R%d invalid mem access '%s'\n",
681 regno, reg_type_str[state->regs[regno].type]);
687 static int check_xadd(struct verifier_env *env, struct bpf_insn *insn)
689 struct reg_state *regs = env->cur_state.regs;
692 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
694 verbose("BPF_XADD uses reserved fields\n");
698 /* check src1 operand */
699 err = check_reg_arg(regs, insn->src_reg, SRC_OP);
703 /* check src2 operand */
704 err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
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);
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);
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
723 static int check_stack_boundary(struct verifier_env *env,
724 int regno, int access_size)
726 struct verifier_state *state = &env->cur_state;
727 struct reg_state *regs = state->regs;
730 if (regs[regno].type != PTR_TO_STACK)
733 off = regs[regno].imm;
734 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
736 verbose("invalid stack type R%d off=%d access_size=%d\n",
737 regno, off, access_size);
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);
751 static int check_func_arg(struct verifier_env *env, u32 regno,
752 enum bpf_arg_type arg_type, struct bpf_map **mapp)
754 struct reg_state *reg = env->cur_state.regs + regno;
755 enum bpf_reg_type expected_type;
758 if (arg_type == ARG_DONTCARE)
761 if (reg->type == NOT_INIT) {
762 verbose("R%d !read_ok\n", regno);
766 if (arg_type == ARG_ANYTHING)
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;
777 verbose("unsupported arg_type %d\n", arg_type);
781 if (reg->type != expected_type) {
782 verbose("R%d type=%s expected=%s\n", regno,
783 reg_type_str[reg->type], reg_type_str[expected_type]);
787 if (arg_type == ARG_CONST_MAP_PTR) {
788 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
789 *mapp = reg->map_ptr;
791 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
792 /* bpf_map_xxx(..., map_ptr, ..., key) call:
793 * check that [key, key + map->key_size) are within
794 * stack limits and initialized
797 /* in function declaration map_ptr must come before
798 * map_key, so that it's verified and known before
799 * we have to check map_key here. Otherwise it means
800 * that kernel subsystem misconfigured verifier
802 verbose("invalid map_ptr to access map->key\n");
805 err = check_stack_boundary(env, regno, (*mapp)->key_size);
807 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
808 /* bpf_map_xxx(..., map_ptr, ..., value) call:
809 * check [value, value + map->value_size) validity
812 /* kernel subsystem misconfigured verifier */
813 verbose("invalid map_ptr to access map->value\n");
816 err = check_stack_boundary(env, regno, (*mapp)->value_size);
818 } else if (arg_type == ARG_CONST_STACK_SIZE) {
819 /* bpf_xxx(..., buf, len) call will access 'len' bytes
820 * from stack pointer 'buf'. Check it
821 * note: regno == len, regno - 1 == buf
824 /* kernel subsystem misconfigured verifier */
825 verbose("ARG_CONST_STACK_SIZE cannot be first argument\n");
828 err = check_stack_boundary(env, regno - 1, reg->imm);
834 static int check_call(struct verifier_env *env, int func_id)
836 struct verifier_state *state = &env->cur_state;
837 const struct bpf_func_proto *fn = NULL;
838 struct reg_state *regs = state->regs;
839 struct bpf_map *map = NULL;
840 struct reg_state *reg;
843 /* find function prototype */
844 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
845 verbose("invalid func %d\n", func_id);
849 if (env->prog->aux->ops->get_func_proto)
850 fn = env->prog->aux->ops->get_func_proto(func_id);
853 verbose("unknown func %d\n", func_id);
857 /* eBPF programs must be GPL compatible to use GPL-ed functions */
858 if (!env->prog->gpl_compatible && fn->gpl_only) {
859 verbose("cannot call GPL only function from proprietary program\n");
864 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &map);
867 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &map);
870 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &map);
873 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &map);
876 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &map);
880 /* reset caller saved regs */
881 for (i = 0; i < CALLER_SAVED_REGS; i++) {
882 reg = regs + caller_saved[i];
883 reg->type = NOT_INIT;
887 /* update return register */
888 if (fn->ret_type == RET_INTEGER) {
889 regs[BPF_REG_0].type = UNKNOWN_VALUE;
890 } else if (fn->ret_type == RET_VOID) {
891 regs[BPF_REG_0].type = NOT_INIT;
892 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
893 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
894 /* remember map_ptr, so that check_map_access()
895 * can check 'value_size' boundary of memory access
896 * to map element returned from bpf_map_lookup_elem()
899 verbose("kernel subsystem misconfigured verifier\n");
902 regs[BPF_REG_0].map_ptr = map;
904 verbose("unknown return type %d of func %d\n",
905 fn->ret_type, func_id);
911 /* check validity of 32-bit and 64-bit arithmetic operations */
912 static int check_alu_op(struct reg_state *regs, struct bpf_insn *insn)
914 u8 opcode = BPF_OP(insn->code);
917 if (opcode == BPF_END || opcode == BPF_NEG) {
918 if (opcode == BPF_NEG) {
919 if (BPF_SRC(insn->code) != 0 ||
920 insn->src_reg != BPF_REG_0 ||
921 insn->off != 0 || insn->imm != 0) {
922 verbose("BPF_NEG uses reserved fields\n");
926 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
927 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64)) {
928 verbose("BPF_END uses reserved fields\n");
933 /* check src operand */
934 err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
938 /* check dest operand */
939 err = check_reg_arg(regs, insn->dst_reg, DST_OP);
943 } else if (opcode == BPF_MOV) {
945 if (BPF_SRC(insn->code) == BPF_X) {
946 if (insn->imm != 0 || insn->off != 0) {
947 verbose("BPF_MOV uses reserved fields\n");
951 /* check src operand */
952 err = check_reg_arg(regs, insn->src_reg, SRC_OP);
956 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
957 verbose("BPF_MOV uses reserved fields\n");
962 /* check dest operand */
963 err = check_reg_arg(regs, insn->dst_reg, DST_OP);
967 if (BPF_SRC(insn->code) == BPF_X) {
968 if (BPF_CLASS(insn->code) == BPF_ALU64) {
970 * copy register state to dest reg
972 regs[insn->dst_reg] = regs[insn->src_reg];
974 regs[insn->dst_reg].type = UNKNOWN_VALUE;
975 regs[insn->dst_reg].map_ptr = NULL;
979 * remember the value we stored into this reg
981 regs[insn->dst_reg].type = CONST_IMM;
982 regs[insn->dst_reg].imm = insn->imm;
985 } else if (opcode > BPF_END) {
986 verbose("invalid BPF_ALU opcode %x\n", opcode);
989 } else { /* all other ALU ops: and, sub, xor, add, ... */
991 bool stack_relative = false;
993 if (BPF_SRC(insn->code) == BPF_X) {
994 if (insn->imm != 0 || insn->off != 0) {
995 verbose("BPF_ALU uses reserved fields\n");
998 /* check src1 operand */
999 err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1003 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
1004 verbose("BPF_ALU uses reserved fields\n");
1009 /* check src2 operand */
1010 err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1014 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
1015 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
1016 verbose("div by zero\n");
1020 /* pattern match 'bpf_add Rx, imm' instruction */
1021 if (opcode == BPF_ADD && BPF_CLASS(insn->code) == BPF_ALU64 &&
1022 regs[insn->dst_reg].type == FRAME_PTR &&
1023 BPF_SRC(insn->code) == BPF_K)
1024 stack_relative = true;
1026 /* check dest operand */
1027 err = check_reg_arg(regs, insn->dst_reg, DST_OP);
1031 if (stack_relative) {
1032 regs[insn->dst_reg].type = PTR_TO_STACK;
1033 regs[insn->dst_reg].imm = insn->imm;
1040 static int check_cond_jmp_op(struct verifier_env *env,
1041 struct bpf_insn *insn, int *insn_idx)
1043 struct reg_state *regs = env->cur_state.regs;
1044 struct verifier_state *other_branch;
1045 u8 opcode = BPF_OP(insn->code);
1048 if (opcode > BPF_EXIT) {
1049 verbose("invalid BPF_JMP opcode %x\n", opcode);
1053 if (BPF_SRC(insn->code) == BPF_X) {
1054 if (insn->imm != 0) {
1055 verbose("BPF_JMP uses reserved fields\n");
1059 /* check src1 operand */
1060 err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1064 if (insn->src_reg != BPF_REG_0) {
1065 verbose("BPF_JMP uses reserved fields\n");
1070 /* check src2 operand */
1071 err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1075 /* detect if R == 0 where R was initialized to zero earlier */
1076 if (BPF_SRC(insn->code) == BPF_K &&
1077 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
1078 regs[insn->dst_reg].type == CONST_IMM &&
1079 regs[insn->dst_reg].imm == insn->imm) {
1080 if (opcode == BPF_JEQ) {
1081 /* if (imm == imm) goto pc+off;
1082 * only follow the goto, ignore fall-through
1084 *insn_idx += insn->off;
1087 /* if (imm != imm) goto pc+off;
1088 * only follow fall-through branch, since
1089 * that's where the program will go
1095 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
1099 /* detect if R == 0 where R is returned value from bpf_map_lookup_elem() */
1100 if (BPF_SRC(insn->code) == BPF_K &&
1101 insn->imm == 0 && (opcode == BPF_JEQ ||
1102 opcode == BPF_JNE) &&
1103 regs[insn->dst_reg].type == PTR_TO_MAP_VALUE_OR_NULL) {
1104 if (opcode == BPF_JEQ) {
1105 /* next fallthrough insn can access memory via
1108 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
1109 /* branch targer cannot access it, since reg == 0 */
1110 other_branch->regs[insn->dst_reg].type = CONST_IMM;
1111 other_branch->regs[insn->dst_reg].imm = 0;
1113 other_branch->regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
1114 regs[insn->dst_reg].type = CONST_IMM;
1115 regs[insn->dst_reg].imm = 0;
1117 } else if (BPF_SRC(insn->code) == BPF_K &&
1118 (opcode == BPF_JEQ || opcode == BPF_JNE)) {
1120 if (opcode == BPF_JEQ) {
1121 /* detect if (R == imm) goto
1122 * and in the target state recognize that R = imm
1124 other_branch->regs[insn->dst_reg].type = CONST_IMM;
1125 other_branch->regs[insn->dst_reg].imm = insn->imm;
1127 /* detect if (R != imm) goto
1128 * and in the fall-through state recognize that R = imm
1130 regs[insn->dst_reg].type = CONST_IMM;
1131 regs[insn->dst_reg].imm = insn->imm;
1135 print_verifier_state(env);
1139 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
1140 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
1142 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
1144 return (struct bpf_map *) (unsigned long) imm64;
1147 /* verify BPF_LD_IMM64 instruction */
1148 static int check_ld_imm(struct verifier_env *env, struct bpf_insn *insn)
1150 struct reg_state *regs = env->cur_state.regs;
1153 if (BPF_SIZE(insn->code) != BPF_DW) {
1154 verbose("invalid BPF_LD_IMM insn\n");
1157 if (insn->off != 0) {
1158 verbose("BPF_LD_IMM64 uses reserved fields\n");
1162 err = check_reg_arg(regs, insn->dst_reg, DST_OP);
1166 if (insn->src_reg == 0)
1167 /* generic move 64-bit immediate into a register */
1170 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
1171 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
1173 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
1174 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
1178 static bool may_access_skb(enum bpf_prog_type type)
1181 case BPF_PROG_TYPE_SOCKET_FILTER:
1182 case BPF_PROG_TYPE_SCHED_CLS:
1189 /* verify safety of LD_ABS|LD_IND instructions:
1190 * - they can only appear in the programs where ctx == skb
1191 * - since they are wrappers of function calls, they scratch R1-R5 registers,
1192 * preserve R6-R9, and store return value into R0
1195 * ctx == skb == R6 == CTX
1198 * SRC == any register
1199 * IMM == 32-bit immediate
1202 * R0 - 8/16/32-bit skb data converted to cpu endianness
1204 static int check_ld_abs(struct verifier_env *env, struct bpf_insn *insn)
1206 struct reg_state *regs = env->cur_state.regs;
1207 u8 mode = BPF_MODE(insn->code);
1208 struct reg_state *reg;
1211 if (!may_access_skb(env->prog->type)) {
1212 verbose("BPF_LD_ABS|IND instructions not allowed for this program type\n");
1216 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
1217 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
1218 verbose("BPF_LD_ABS uses reserved fields\n");
1222 /* check whether implicit source operand (register R6) is readable */
1223 err = check_reg_arg(regs, BPF_REG_6, SRC_OP);
1227 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
1228 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
1232 if (mode == BPF_IND) {
1233 /* check explicit source operand */
1234 err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1239 /* reset caller saved regs to unreadable */
1240 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1241 reg = regs + caller_saved[i];
1242 reg->type = NOT_INIT;
1246 /* mark destination R0 register as readable, since it contains
1247 * the value fetched from the packet
1249 regs[BPF_REG_0].type = UNKNOWN_VALUE;
1253 /* non-recursive DFS pseudo code
1254 * 1 procedure DFS-iterative(G,v):
1255 * 2 label v as discovered
1256 * 3 let S be a stack
1258 * 5 while S is not empty
1260 * 7 if t is what we're looking for:
1262 * 9 for all edges e in G.adjacentEdges(t) do
1263 * 10 if edge e is already labelled
1264 * 11 continue with the next edge
1265 * 12 w <- G.adjacentVertex(t,e)
1266 * 13 if vertex w is not discovered and not explored
1267 * 14 label e as tree-edge
1268 * 15 label w as discovered
1271 * 18 else if vertex w is discovered
1272 * 19 label e as back-edge
1274 * 21 // vertex w is explored
1275 * 22 label e as forward- or cross-edge
1276 * 23 label t as explored
1281 * 0x11 - discovered and fall-through edge labelled
1282 * 0x12 - discovered and fall-through and branch edges labelled
1293 #define STATE_LIST_MARK ((struct verifier_state_list *) -1L)
1295 static int *insn_stack; /* stack of insns to process */
1296 static int cur_stack; /* current stack index */
1297 static int *insn_state;
1299 /* t, w, e - match pseudo-code above:
1300 * t - index of current instruction
1301 * w - next instruction
1304 static int push_insn(int t, int w, int e, struct verifier_env *env)
1306 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
1309 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
1312 if (w < 0 || w >= env->prog->len) {
1313 verbose("jump out of range from insn %d to %d\n", t, w);
1318 /* mark branch target for state pruning */
1319 env->explored_states[w] = STATE_LIST_MARK;
1321 if (insn_state[w] == 0) {
1323 insn_state[t] = DISCOVERED | e;
1324 insn_state[w] = DISCOVERED;
1325 if (cur_stack >= env->prog->len)
1327 insn_stack[cur_stack++] = w;
1329 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
1330 verbose("back-edge from insn %d to %d\n", t, w);
1332 } else if (insn_state[w] == EXPLORED) {
1333 /* forward- or cross-edge */
1334 insn_state[t] = DISCOVERED | e;
1336 verbose("insn state internal bug\n");
1342 /* non-recursive depth-first-search to detect loops in BPF program
1343 * loop == back-edge in directed graph
1345 static int check_cfg(struct verifier_env *env)
1347 struct bpf_insn *insns = env->prog->insnsi;
1348 int insn_cnt = env->prog->len;
1352 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
1356 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
1362 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
1363 insn_stack[0] = 0; /* 0 is the first instruction */
1369 t = insn_stack[cur_stack - 1];
1371 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
1372 u8 opcode = BPF_OP(insns[t].code);
1374 if (opcode == BPF_EXIT) {
1376 } else if (opcode == BPF_CALL) {
1377 ret = push_insn(t, t + 1, FALLTHROUGH, env);
1382 } else if (opcode == BPF_JA) {
1383 if (BPF_SRC(insns[t].code) != BPF_K) {
1387 /* unconditional jump with single edge */
1388 ret = push_insn(t, t + insns[t].off + 1,
1394 /* tell verifier to check for equivalent states
1395 * after every call and jump
1397 env->explored_states[t + 1] = STATE_LIST_MARK;
1399 /* conditional jump with two edges */
1400 ret = push_insn(t, t + 1, FALLTHROUGH, env);
1406 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
1413 /* all other non-branch instructions with single
1416 ret = push_insn(t, t + 1, FALLTHROUGH, env);
1424 insn_state[t] = EXPLORED;
1425 if (cur_stack-- <= 0) {
1426 verbose("pop stack internal bug\n");
1433 for (i = 0; i < insn_cnt; i++) {
1434 if (insn_state[i] != EXPLORED) {
1435 verbose("unreachable insn %d\n", i);
1440 ret = 0; /* cfg looks good */
1448 /* compare two verifier states
1450 * all states stored in state_list are known to be valid, since
1451 * verifier reached 'bpf_exit' instruction through them
1453 * this function is called when verifier exploring different branches of
1454 * execution popped from the state stack. If it sees an old state that has
1455 * more strict register state and more strict stack state then this execution
1456 * branch doesn't need to be explored further, since verifier already
1457 * concluded that more strict state leads to valid finish.
1459 * Therefore two states are equivalent if register state is more conservative
1460 * and explored stack state is more conservative than the current one.
1463 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
1464 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
1466 * In other words if current stack state (one being explored) has more
1467 * valid slots than old one that already passed validation, it means
1468 * the verifier can stop exploring and conclude that current state is valid too
1470 * Similarly with registers. If explored state has register type as invalid
1471 * whereas register type in current state is meaningful, it means that
1472 * the current state will reach 'bpf_exit' instruction safely
1474 static bool states_equal(struct verifier_state *old, struct verifier_state *cur)
1478 for (i = 0; i < MAX_BPF_REG; i++) {
1479 if (memcmp(&old->regs[i], &cur->regs[i],
1480 sizeof(old->regs[0])) != 0) {
1481 if (old->regs[i].type == NOT_INIT ||
1482 (old->regs[i].type == UNKNOWN_VALUE &&
1483 cur->regs[i].type != NOT_INIT))
1489 for (i = 0; i < MAX_BPF_STACK; i++) {
1490 if (old->stack_slot_type[i] == STACK_INVALID)
1492 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
1493 /* Ex: old explored (safe) state has STACK_SPILL in
1494 * this stack slot, but current has has STACK_MISC ->
1495 * this verifier states are not equivalent,
1496 * return false to continue verification of this path
1499 if (i % BPF_REG_SIZE)
1501 if (memcmp(&old->spilled_regs[i / BPF_REG_SIZE],
1502 &cur->spilled_regs[i / BPF_REG_SIZE],
1503 sizeof(old->spilled_regs[0])))
1504 /* when explored and current stack slot types are
1505 * the same, check that stored pointers types
1506 * are the same as well.
1507 * Ex: explored safe path could have stored
1508 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -8}
1509 * but current path has stored:
1510 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -16}
1511 * such verifier states are not equivalent.
1512 * return false to continue verification of this path
1521 static int is_state_visited(struct verifier_env *env, int insn_idx)
1523 struct verifier_state_list *new_sl;
1524 struct verifier_state_list *sl;
1526 sl = env->explored_states[insn_idx];
1528 /* this 'insn_idx' instruction wasn't marked, so we will not
1529 * be doing state search here
1533 while (sl != STATE_LIST_MARK) {
1534 if (states_equal(&sl->state, &env->cur_state))
1535 /* reached equivalent register/stack state,
1542 /* there were no equivalent states, remember current one.
1543 * technically the current state is not proven to be safe yet,
1544 * but it will either reach bpf_exit (which means it's safe) or
1545 * it will be rejected. Since there are no loops, we won't be
1546 * seeing this 'insn_idx' instruction again on the way to bpf_exit
1548 new_sl = kmalloc(sizeof(struct verifier_state_list), GFP_USER);
1552 /* add new state to the head of linked list */
1553 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
1554 new_sl->next = env->explored_states[insn_idx];
1555 env->explored_states[insn_idx] = new_sl;
1559 static int do_check(struct verifier_env *env)
1561 struct verifier_state *state = &env->cur_state;
1562 struct bpf_insn *insns = env->prog->insnsi;
1563 struct reg_state *regs = state->regs;
1564 int insn_cnt = env->prog->len;
1565 int insn_idx, prev_insn_idx = 0;
1566 int insn_processed = 0;
1567 bool do_print_state = false;
1569 init_reg_state(regs);
1572 struct bpf_insn *insn;
1576 if (insn_idx >= insn_cnt) {
1577 verbose("invalid insn idx %d insn_cnt %d\n",
1578 insn_idx, insn_cnt);
1582 insn = &insns[insn_idx];
1583 class = BPF_CLASS(insn->code);
1585 if (++insn_processed > 32768) {
1586 verbose("BPF program is too large. Proccessed %d insn\n",
1591 err = is_state_visited(env, insn_idx);
1595 /* found equivalent state, can prune the search */
1598 verbose("\nfrom %d to %d: safe\n",
1599 prev_insn_idx, insn_idx);
1601 verbose("%d: safe\n", insn_idx);
1603 goto process_bpf_exit;
1606 if (log_level && do_print_state) {
1607 verbose("\nfrom %d to %d:", prev_insn_idx, insn_idx);
1608 print_verifier_state(env);
1609 do_print_state = false;
1613 verbose("%d: ", insn_idx);
1614 print_bpf_insn(insn);
1617 if (class == BPF_ALU || class == BPF_ALU64) {
1618 err = check_alu_op(regs, insn);
1622 } else if (class == BPF_LDX) {
1623 enum bpf_reg_type src_reg_type;
1625 /* check for reserved fields is already done */
1627 /* check src operand */
1628 err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1632 err = check_reg_arg(regs, insn->dst_reg, DST_OP_NO_MARK);
1636 /* check that memory (src_reg + off) is readable,
1637 * the state of dst_reg will be updated by this func
1639 err = check_mem_access(env, insn->src_reg, insn->off,
1640 BPF_SIZE(insn->code), BPF_READ,
1645 src_reg_type = regs[insn->src_reg].type;
1647 if (insn->imm == 0 && BPF_SIZE(insn->code) == BPF_W) {
1649 * dst_reg = *(u32 *)(src_reg + off)
1650 * use reserved 'imm' field to mark this insn
1652 insn->imm = src_reg_type;
1654 } else if (src_reg_type != insn->imm &&
1655 (src_reg_type == PTR_TO_CTX ||
1656 insn->imm == PTR_TO_CTX)) {
1657 /* ABuser program is trying to use the same insn
1658 * dst_reg = *(u32*) (src_reg + off)
1659 * with different pointer types:
1660 * src_reg == ctx in one branch and
1661 * src_reg == stack|map in some other branch.
1664 verbose("same insn cannot be used with different pointers\n");
1668 } else if (class == BPF_STX) {
1669 if (BPF_MODE(insn->code) == BPF_XADD) {
1670 err = check_xadd(env, insn);
1677 if (BPF_MODE(insn->code) != BPF_MEM ||
1679 verbose("BPF_STX uses reserved fields\n");
1682 /* check src1 operand */
1683 err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1686 /* check src2 operand */
1687 err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1691 /* check that memory (dst_reg + off) is writeable */
1692 err = check_mem_access(env, insn->dst_reg, insn->off,
1693 BPF_SIZE(insn->code), BPF_WRITE,
1698 } else if (class == BPF_ST) {
1699 if (BPF_MODE(insn->code) != BPF_MEM ||
1700 insn->src_reg != BPF_REG_0) {
1701 verbose("BPF_ST uses reserved fields\n");
1704 /* check src operand */
1705 err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1709 /* check that memory (dst_reg + off) is writeable */
1710 err = check_mem_access(env, insn->dst_reg, insn->off,
1711 BPF_SIZE(insn->code), BPF_WRITE,
1716 } else if (class == BPF_JMP) {
1717 u8 opcode = BPF_OP(insn->code);
1719 if (opcode == BPF_CALL) {
1720 if (BPF_SRC(insn->code) != BPF_K ||
1722 insn->src_reg != BPF_REG_0 ||
1723 insn->dst_reg != BPF_REG_0) {
1724 verbose("BPF_CALL uses reserved fields\n");
1728 err = check_call(env, insn->imm);
1732 } else if (opcode == BPF_JA) {
1733 if (BPF_SRC(insn->code) != BPF_K ||
1735 insn->src_reg != BPF_REG_0 ||
1736 insn->dst_reg != BPF_REG_0) {
1737 verbose("BPF_JA uses reserved fields\n");
1741 insn_idx += insn->off + 1;
1744 } else if (opcode == BPF_EXIT) {
1745 if (BPF_SRC(insn->code) != BPF_K ||
1747 insn->src_reg != BPF_REG_0 ||
1748 insn->dst_reg != BPF_REG_0) {
1749 verbose("BPF_EXIT uses reserved fields\n");
1753 /* eBPF calling convetion is such that R0 is used
1754 * to return the value from eBPF program.
1755 * Make sure that it's readable at this time
1756 * of bpf_exit, which means that program wrote
1757 * something into it earlier
1759 err = check_reg_arg(regs, BPF_REG_0, SRC_OP);
1764 insn_idx = pop_stack(env, &prev_insn_idx);
1768 do_print_state = true;
1772 err = check_cond_jmp_op(env, insn, &insn_idx);
1776 } else if (class == BPF_LD) {
1777 u8 mode = BPF_MODE(insn->code);
1779 if (mode == BPF_ABS || mode == BPF_IND) {
1780 err = check_ld_abs(env, insn);
1784 } else if (mode == BPF_IMM) {
1785 err = check_ld_imm(env, insn);
1791 verbose("invalid BPF_LD mode\n");
1795 verbose("unknown insn class %d\n", class);
1805 /* look for pseudo eBPF instructions that access map FDs and
1806 * replace them with actual map pointers
1808 static int replace_map_fd_with_map_ptr(struct verifier_env *env)
1810 struct bpf_insn *insn = env->prog->insnsi;
1811 int insn_cnt = env->prog->len;
1814 for (i = 0; i < insn_cnt; i++, insn++) {
1815 if (BPF_CLASS(insn->code) == BPF_LDX &&
1816 (BPF_MODE(insn->code) != BPF_MEM ||
1818 verbose("BPF_LDX uses reserved fields\n");
1822 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
1823 struct bpf_map *map;
1826 if (i == insn_cnt - 1 || insn[1].code != 0 ||
1827 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
1829 verbose("invalid bpf_ld_imm64 insn\n");
1833 if (insn->src_reg == 0)
1834 /* valid generic load 64-bit imm */
1837 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
1838 verbose("unrecognized bpf_ld_imm64 insn\n");
1842 f = fdget(insn->imm);
1844 map = bpf_map_get(f);
1846 verbose("fd %d is not pointing to valid bpf_map\n",
1849 return PTR_ERR(map);
1852 /* store map pointer inside BPF_LD_IMM64 instruction */
1853 insn[0].imm = (u32) (unsigned long) map;
1854 insn[1].imm = ((u64) (unsigned long) map) >> 32;
1856 /* check whether we recorded this map already */
1857 for (j = 0; j < env->used_map_cnt; j++)
1858 if (env->used_maps[j] == map) {
1863 if (env->used_map_cnt >= MAX_USED_MAPS) {
1868 /* remember this map */
1869 env->used_maps[env->used_map_cnt++] = map;
1871 /* hold the map. If the program is rejected by verifier,
1872 * the map will be released by release_maps() or it
1873 * will be used by the valid program until it's unloaded
1874 * and all maps are released in free_bpf_prog_info()
1876 atomic_inc(&map->refcnt);
1885 /* now all pseudo BPF_LD_IMM64 instructions load valid
1886 * 'struct bpf_map *' into a register instead of user map_fd.
1887 * These pointers will be used later by verifier to validate map access.
1892 /* drop refcnt of maps used by the rejected program */
1893 static void release_maps(struct verifier_env *env)
1897 for (i = 0; i < env->used_map_cnt; i++)
1898 bpf_map_put(env->used_maps[i]);
1901 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
1902 static void convert_pseudo_ld_imm64(struct verifier_env *env)
1904 struct bpf_insn *insn = env->prog->insnsi;
1905 int insn_cnt = env->prog->len;
1908 for (i = 0; i < insn_cnt; i++, insn++)
1909 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
1913 static void adjust_branches(struct bpf_prog *prog, int pos, int delta)
1915 struct bpf_insn *insn = prog->insnsi;
1916 int insn_cnt = prog->len;
1919 for (i = 0; i < insn_cnt; i++, insn++) {
1920 if (BPF_CLASS(insn->code) != BPF_JMP ||
1921 BPF_OP(insn->code) == BPF_CALL ||
1922 BPF_OP(insn->code) == BPF_EXIT)
1925 /* adjust offset of jmps if necessary */
1926 if (i < pos && i + insn->off + 1 > pos)
1928 else if (i > pos && i + insn->off + 1 < pos)
1933 /* convert load instructions that access fields of 'struct __sk_buff'
1934 * into sequence of instructions that access fields of 'struct sk_buff'
1936 static int convert_ctx_accesses(struct verifier_env *env)
1938 struct bpf_insn *insn = env->prog->insnsi;
1939 int insn_cnt = env->prog->len;
1940 struct bpf_insn insn_buf[16];
1941 struct bpf_prog *new_prog;
1945 if (!env->prog->aux->ops->convert_ctx_access)
1948 for (i = 0; i < insn_cnt; i++, insn++) {
1949 if (insn->code != (BPF_LDX | BPF_MEM | BPF_W))
1952 if (insn->imm != PTR_TO_CTX) {
1953 /* clear internal mark */
1958 cnt = env->prog->aux->ops->
1959 convert_ctx_access(insn->dst_reg, insn->src_reg,
1960 insn->off, insn_buf);
1961 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
1962 verbose("bpf verifier is misconfigured\n");
1967 memcpy(insn, insn_buf, sizeof(*insn));
1971 /* several new insns need to be inserted. Make room for them */
1972 insn_cnt += cnt - 1;
1973 new_prog = bpf_prog_realloc(env->prog,
1974 bpf_prog_size(insn_cnt),
1979 new_prog->len = insn_cnt;
1981 memmove(new_prog->insnsi + i + cnt, new_prog->insns + i + 1,
1982 sizeof(*insn) * (insn_cnt - i - cnt));
1984 /* copy substitute insns in place of load instruction */
1985 memcpy(new_prog->insnsi + i, insn_buf, sizeof(*insn) * cnt);
1987 /* adjust branches in the whole program */
1988 adjust_branches(new_prog, i, cnt - 1);
1990 /* keep walking new program and skip insns we just inserted */
1991 env->prog = new_prog;
1992 insn = new_prog->insnsi + i + cnt - 1;
1999 static void free_states(struct verifier_env *env)
2001 struct verifier_state_list *sl, *sln;
2004 if (!env->explored_states)
2007 for (i = 0; i < env->prog->len; i++) {
2008 sl = env->explored_states[i];
2011 while (sl != STATE_LIST_MARK) {
2018 kfree(env->explored_states);
2021 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
2023 char __user *log_ubuf = NULL;
2024 struct verifier_env *env;
2027 if ((*prog)->len <= 0 || (*prog)->len > BPF_MAXINSNS)
2030 /* 'struct verifier_env' can be global, but since it's not small,
2031 * allocate/free it every time bpf_check() is called
2033 env = kzalloc(sizeof(struct verifier_env), GFP_KERNEL);
2039 /* grab the mutex to protect few globals used by verifier */
2040 mutex_lock(&bpf_verifier_lock);
2042 if (attr->log_level || attr->log_buf || attr->log_size) {
2043 /* user requested verbose verifier output
2044 * and supplied buffer to store the verification trace
2046 log_level = attr->log_level;
2047 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
2048 log_size = attr->log_size;
2052 /* log_* values have to be sane */
2053 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
2054 log_level == 0 || log_ubuf == NULL)
2058 log_buf = vmalloc(log_size);
2065 ret = replace_map_fd_with_map_ptr(env);
2067 goto skip_full_check;
2069 env->explored_states = kcalloc(env->prog->len,
2070 sizeof(struct verifier_state_list *),
2073 if (!env->explored_states)
2074 goto skip_full_check;
2076 ret = check_cfg(env);
2078 goto skip_full_check;
2080 ret = do_check(env);
2083 while (pop_stack(env, NULL) >= 0);
2087 /* program is valid, convert *(u32*)(ctx + off) accesses */
2088 ret = convert_ctx_accesses(env);
2090 if (log_level && log_len >= log_size - 1) {
2091 BUG_ON(log_len >= log_size);
2092 /* verifier log exceeded user supplied buffer */
2094 /* fall through to return what was recorded */
2097 /* copy verifier log back to user space including trailing zero */
2098 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
2103 if (ret == 0 && env->used_map_cnt) {
2104 /* if program passed verifier, update used_maps in bpf_prog_info */
2105 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
2106 sizeof(env->used_maps[0]),
2109 if (!env->prog->aux->used_maps) {
2114 memcpy(env->prog->aux->used_maps, env->used_maps,
2115 sizeof(env->used_maps[0]) * env->used_map_cnt);
2116 env->prog->aux->used_map_cnt = env->used_map_cnt;
2118 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
2119 * bpf_ld_imm64 instructions
2121 convert_pseudo_ld_imm64(env);
2128 if (!env->prog->aux->used_maps)
2129 /* if we didn't copy map pointers into bpf_prog_info, release
2130 * them now. Otherwise free_bpf_prog_info() will release them.
2135 mutex_unlock(&bpf_verifier_lock);