]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - kernel/bpf/verifier.c
projects / mirror_ubuntu-hirsute-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'for-linus' of git://git.kernel.dk/linux-block
[mirror_ubuntu-hirsute-kernel.git] / kernel / bpf / verifier.c
1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 */
13 #include <linux/kernel.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
16 #include <linux/bpf.h>
17 #include <linux/bpf_verifier.h>
18 #include <linux/filter.h>
19 #include <net/netlink.h>
20 #include <linux/file.h>
21 #include <linux/vmalloc.h>
22 #include <linux/stringify.h>
23
24 /* bpf_check() is a static code analyzer that walks eBPF program
25 * instruction by instruction and updates register/stack state.
26 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
27 *
28 * The first pass is depth-first-search to check that the program is a DAG.
29 * It rejects the following programs:
30 * - larger than BPF_MAXINSNS insns
31 * - if loop is present (detected via back-edge)
32 * - unreachable insns exist (shouldn't be a forest. program = one function)
33 * - out of bounds or malformed jumps
34 * The second pass is all possible path descent from the 1st insn.
35 * Since it's analyzing all pathes through the program, the length of the
36 * analysis is limited to 64k insn, which may be hit even if total number of
37 * insn is less then 4K, but there are too many branches that change stack/regs.
38 * Number of 'branches to be analyzed' is limited to 1k
39 *
40 * On entry to each instruction, each register has a type, and the instruction
41 * changes the types of the registers depending on instruction semantics.
42 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
43 * copied to R1.
44 *
45 * All registers are 64-bit.
46 * R0 - return register
47 * R1-R5 argument passing registers
48 * R6-R9 callee saved registers
49 * R10 - frame pointer read-only
50 *
51 * At the start of BPF program the register R1 contains a pointer to bpf_context
52 * and has type PTR_TO_CTX.
53 *
54 * Verifier tracks arithmetic operations on pointers in case:
55 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
56 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
57 * 1st insn copies R10 (which has FRAME_PTR) type into R1
58 * and 2nd arithmetic instruction is pattern matched to recognize
59 * that it wants to construct a pointer to some element within stack.
60 * So after 2nd insn, the register R1 has type PTR_TO_STACK
61 * (and -20 constant is saved for further stack bounds checking).
62 * Meaning that this reg is a pointer to stack plus known immediate constant.
63 *
64 * Most of the time the registers have SCALAR_VALUE type, which
65 * means the register has some value, but it's not a valid pointer.
66 * (like pointer plus pointer becomes SCALAR_VALUE type)
67 *
68 * When verifier sees load or store instructions the type of base register
69 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
70 * types recognized by check_mem_access() function.
71 *
72 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
73 * and the range of [ptr, ptr + map's value_size) is accessible.
74 *
75 * registers used to pass values to function calls are checked against
76 * function argument constraints.
77 *
78 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
79 * It means that the register type passed to this function must be
80 * PTR_TO_STACK and it will be used inside the function as
81 * 'pointer to map element key'
82 *
83 * For example the argument constraints for bpf_map_lookup_elem():
84 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
85 * .arg1_type = ARG_CONST_MAP_PTR,
86 * .arg2_type = ARG_PTR_TO_MAP_KEY,
87 *
88 * ret_type says that this function returns 'pointer to map elem value or null'
89 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
90 * 2nd argument should be a pointer to stack, which will be used inside
91 * the helper function as a pointer to map element key.
92 *
93 * On the kernel side the helper function looks like:
94 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
95 * {
96 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
97 * void *key = (void *) (unsigned long) r2;
98 * void *value;
99 *
100 * here kernel can access 'key' and 'map' pointers safely, knowing that
101 * [key, key + map->key_size) bytes are valid and were initialized on
102 * the stack of eBPF program.
103 * }
104 *
105 * Corresponding eBPF program may look like:
106 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
107 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
108 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
109 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
110 * here verifier looks at prototype of map_lookup_elem() and sees:
111 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
112 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
113 *
114 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
115 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
116 * and were initialized prior to this call.
117 * If it's ok, then verifier allows this BPF_CALL insn and looks at
118 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
119 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
120 * returns ether pointer to map value or NULL.
121 *
122 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
123 * insn, the register holding that pointer in the true branch changes state to
124 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
125 * branch. See check_cond_jmp_op().
126 *
127 * After the call R0 is set to return type of the function and registers R1-R5
128 * are set to NOT_INIT to indicate that they are no longer readable.
129 */
130
131 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
132 struct bpf_verifier_stack_elem {
133 /* verifer state is 'st'
134 * before processing instruction 'insn_idx'
135 * and after processing instruction 'prev_insn_idx'
136 */
137 struct bpf_verifier_state st;
138 int insn_idx;
139 int prev_insn_idx;
140 struct bpf_verifier_stack_elem *next;
141 };
142
143 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
144 #define BPF_COMPLEXITY_LIMIT_STACK 1024
145
146 #define BPF_MAP_PTR_POISON ((void *)0xeB9F + POISON_POINTER_DELTA)
147
148 struct bpf_call_arg_meta {
149 struct bpf_map *map_ptr;
150 bool raw_mode;
151 bool pkt_access;
152 int regno;
153 int access_size;
154 };
155
156 /* verbose verifier prints what it's seeing
157 * bpf_check() is called under lock, so no race to access these global vars
158 */
159 static u32 log_level, log_size, log_len;
160 static char *log_buf;
161
162 static DEFINE_MUTEX(bpf_verifier_lock);
163
164 /* log_level controls verbosity level of eBPF verifier.
165 * verbose() is used to dump the verification trace to the log, so the user
166 * can figure out what's wrong with the program
167 */
168 static __printf(1, 2) void verbose(const char *fmt, ...)
169 {
170 va_list args;
171
172 if (log_level == 0 || log_len >= log_size - 1)
173 return;
174
175 va_start(args, fmt);
176 log_len += vscnprintf(log_buf + log_len, log_size - log_len, fmt, args);
177 va_end(args);
178 }
179
180 /* string representation of 'enum bpf_reg_type' */
181 static const char * const reg_type_str[] = {
182 [NOT_INIT] = "?",
183 [SCALAR_VALUE] = "inv",
184 [PTR_TO_CTX] = "ctx",
185 [CONST_PTR_TO_MAP] = "map_ptr",
186 [PTR_TO_MAP_VALUE] = "map_value",
187 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
188 [PTR_TO_STACK] = "fp",
189 [PTR_TO_PACKET] = "pkt",
190 [PTR_TO_PACKET_END] = "pkt_end",
191 };
192
193 #define __BPF_FUNC_STR_FN(x) [BPF_FUNC_ ## x] = __stringify(bpf_ ## x)
194 static const char * const func_id_str[] = {
195 __BPF_FUNC_MAPPER(__BPF_FUNC_STR_FN)
196 };
197 #undef __BPF_FUNC_STR_FN
198
199 static const char *func_id_name(int id)
200 {
201 BUILD_BUG_ON(ARRAY_SIZE(func_id_str) != __BPF_FUNC_MAX_ID);
202
203 if (id >= 0 && id < __BPF_FUNC_MAX_ID && func_id_str[id])
204 return func_id_str[id];
205 else
206 return "unknown";
207 }
208
209 static void print_verifier_state(struct bpf_verifier_state *state)
210 {
211 struct bpf_reg_state *reg;
212 enum bpf_reg_type t;
213 int i;
214
215 for (i = 0; i < MAX_BPF_REG; i++) {
216 reg = &state->regs[i];
217 t = reg->type;
218 if (t == NOT_INIT)
219 continue;
220 verbose(" R%d=%s", i, reg_type_str[t]);
221 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
222 tnum_is_const(reg->var_off)) {
223 /* reg->off should be 0 for SCALAR_VALUE */
224 verbose("%lld", reg->var_off.value + reg->off);
225 } else {
226 verbose("(id=%d", reg->id);
227 if (t != SCALAR_VALUE)
228 verbose(",off=%d", reg->off);
229 if (t == PTR_TO_PACKET)
230 verbose(",r=%d", reg->range);
231 else if (t == CONST_PTR_TO_MAP ||
232 t == PTR_TO_MAP_VALUE ||
233 t == PTR_TO_MAP_VALUE_OR_NULL)
234 verbose(",ks=%d,vs=%d",
235 reg->map_ptr->key_size,
236 reg->map_ptr->value_size);
237 if (tnum_is_const(reg->var_off)) {
238 /* Typically an immediate SCALAR_VALUE, but
239 * could be a pointer whose offset is too big
240 * for reg->off
241 */
242 verbose(",imm=%llx", reg->var_off.value);
243 } else {
244 if (reg->smin_value != reg->umin_value &&
245 reg->smin_value != S64_MIN)
246 verbose(",smin_value=%lld",
247 (long long)reg->smin_value);
248 if (reg->smax_value != reg->umax_value &&
249 reg->smax_value != S64_MAX)
250 verbose(",smax_value=%lld",
251 (long long)reg->smax_value);
252 if (reg->umin_value != 0)
253 verbose(",umin_value=%llu",
254 (unsigned long long)reg->umin_value);
255 if (reg->umax_value != U64_MAX)
256 verbose(",umax_value=%llu",
257 (unsigned long long)reg->umax_value);
258 if (!tnum_is_unknown(reg->var_off)) {
259 char tn_buf[48];
260
261 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
262 verbose(",var_off=%s", tn_buf);
263 }
264 }
265 verbose(")");
266 }
267 }
268 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
269 if (state->stack_slot_type[i] == STACK_SPILL)
270 verbose(" fp%d=%s", -MAX_BPF_STACK + i,
271 reg_type_str[state->spilled_regs[i / BPF_REG_SIZE].type]);
272 }
273 verbose("\n");
274 }
275
276 static const char *const bpf_class_string[] = {
277 [BPF_LD] = "ld",
278 [BPF_LDX] = "ldx",
279 [BPF_ST] = "st",
280 [BPF_STX] = "stx",
281 [BPF_ALU] = "alu",
282 [BPF_JMP] = "jmp",
283 [BPF_RET] = "BUG",
284 [BPF_ALU64] = "alu64",
285 };
286
287 static const char *const bpf_alu_string[16] = {
288 [BPF_ADD >> 4] = "+=",
289 [BPF_SUB >> 4] = "-=",
290 [BPF_MUL >> 4] = "*=",
291 [BPF_DIV >> 4] = "/=",
292 [BPF_OR >> 4] = "|=",
293 [BPF_AND >> 4] = "&=",
294 [BPF_LSH >> 4] = "<<=",
295 [BPF_RSH >> 4] = ">>=",
296 [BPF_NEG >> 4] = "neg",
297 [BPF_MOD >> 4] = "%=",
298 [BPF_XOR >> 4] = "^=",
299 [BPF_MOV >> 4] = "=",
300 [BPF_ARSH >> 4] = "s>>=",
301 [BPF_END >> 4] = "endian",
302 };
303
304 static const char *const bpf_ldst_string[] = {
305 [BPF_W >> 3] = "u32",
306 [BPF_H >> 3] = "u16",
307 [BPF_B >> 3] = "u8",
308 [BPF_DW >> 3] = "u64",
309 };
310
311 static const char *const bpf_jmp_string[16] = {
312 [BPF_JA >> 4] = "jmp",
313 [BPF_JEQ >> 4] = "==",
314 [BPF_JGT >> 4] = ">",
315 [BPF_JLT >> 4] = "<",
316 [BPF_JGE >> 4] = ">=",
317 [BPF_JLE >> 4] = "<=",
318 [BPF_JSET >> 4] = "&",
319 [BPF_JNE >> 4] = "!=",
320 [BPF_JSGT >> 4] = "s>",
321 [BPF_JSLT >> 4] = "s<",
322 [BPF_JSGE >> 4] = "s>=",
323 [BPF_JSLE >> 4] = "s<=",
324 [BPF_CALL >> 4] = "call",
325 [BPF_EXIT >> 4] = "exit",
326 };
327
328 static void print_bpf_insn(const struct bpf_verifier_env *env,
329 const struct bpf_insn *insn)
330 {
331 u8 class = BPF_CLASS(insn->code);
332
333 if (class == BPF_ALU || class == BPF_ALU64) {
334 if (BPF_SRC(insn->code) == BPF_X)
335 verbose("(%02x) %sr%d %s %sr%d\n",
336 insn->code, class == BPF_ALU ? "(u32) " : "",
337 insn->dst_reg,
338 bpf_alu_string[BPF_OP(insn->code) >> 4],
339 class == BPF_ALU ? "(u32) " : "",
340 insn->src_reg);
341 else
342 verbose("(%02x) %sr%d %s %s%d\n",
343 insn->code, class == BPF_ALU ? "(u32) " : "",
344 insn->dst_reg,
345 bpf_alu_string[BPF_OP(insn->code) >> 4],
346 class == BPF_ALU ? "(u32) " : "",
347 insn->imm);
348 } else if (class == BPF_STX) {
349 if (BPF_MODE(insn->code) == BPF_MEM)
350 verbose("(%02x) *(%s *)(r%d %+d) = r%d\n",
351 insn->code,
352 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
353 insn->dst_reg,
354 insn->off, insn->src_reg);
355 else if (BPF_MODE(insn->code) == BPF_XADD)
356 verbose("(%02x) lock *(%s *)(r%d %+d) += r%d\n",
357 insn->code,
358 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
359 insn->dst_reg, insn->off,
360 insn->src_reg);
361 else
362 verbose("BUG_%02x\n", insn->code);
363 } else if (class == BPF_ST) {
364 if (BPF_MODE(insn->code) != BPF_MEM) {
365 verbose("BUG_st_%02x\n", insn->code);
366 return;
367 }
368 verbose("(%02x) *(%s *)(r%d %+d) = %d\n",
369 insn->code,
370 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
371 insn->dst_reg,
372 insn->off, insn->imm);
373 } else if (class == BPF_LDX) {
374 if (BPF_MODE(insn->code) != BPF_MEM) {
375 verbose("BUG_ldx_%02x\n", insn->code);
376 return;
377 }
378 verbose("(%02x) r%d = *(%s *)(r%d %+d)\n",
379 insn->code, insn->dst_reg,
380 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
381 insn->src_reg, insn->off);
382 } else if (class == BPF_LD) {
383 if (BPF_MODE(insn->code) == BPF_ABS) {
384 verbose("(%02x) r0 = *(%s *)skb[%d]\n",
385 insn->code,
386 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
387 insn->imm);
388 } else if (BPF_MODE(insn->code) == BPF_IND) {
389 verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n",
390 insn->code,
391 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
392 insn->src_reg, insn->imm);
393 } else if (BPF_MODE(insn->code) == BPF_IMM &&
394 BPF_SIZE(insn->code) == BPF_DW) {
395 /* At this point, we already made sure that the second
396 * part of the ldimm64 insn is accessible.
397 */
398 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
399 bool map_ptr = insn->src_reg == BPF_PSEUDO_MAP_FD;
400
401 if (map_ptr && !env->allow_ptr_leaks)
402 imm = 0;
403
404 verbose("(%02x) r%d = 0x%llx\n", insn->code,
405 insn->dst_reg, (unsigned long long)imm);
406 } else {
407 verbose("BUG_ld_%02x\n", insn->code);
408 return;
409 }
410 } else if (class == BPF_JMP) {
411 u8 opcode = BPF_OP(insn->code);
412
413 if (opcode == BPF_CALL) {
414 verbose("(%02x) call %s#%d\n", insn->code,
415 func_id_name(insn->imm), insn->imm);
416 } else if (insn->code == (BPF_JMP | BPF_JA)) {
417 verbose("(%02x) goto pc%+d\n",
418 insn->code, insn->off);
419 } else if (insn->code == (BPF_JMP | BPF_EXIT)) {
420 verbose("(%02x) exit\n", insn->code);
421 } else if (BPF_SRC(insn->code) == BPF_X) {
422 verbose("(%02x) if r%d %s r%d goto pc%+d\n",
423 insn->code, insn->dst_reg,
424 bpf_jmp_string[BPF_OP(insn->code) >> 4],
425 insn->src_reg, insn->off);
426 } else {
427 verbose("(%02x) if r%d %s 0x%x goto pc%+d\n",
428 insn->code, insn->dst_reg,
429 bpf_jmp_string[BPF_OP(insn->code) >> 4],
430 insn->imm, insn->off);
431 }
432 } else {
433 verbose("(%02x) %s\n", insn->code, bpf_class_string[class]);
434 }
435 }
436
437 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx)
438 {
439 struct bpf_verifier_stack_elem *elem;
440 int insn_idx;
441
442 if (env->head == NULL)
443 return -1;
444
445 memcpy(&env->cur_state, &env->head->st, sizeof(env->cur_state));
446 insn_idx = env->head->insn_idx;
447 if (prev_insn_idx)
448 *prev_insn_idx = env->head->prev_insn_idx;
449 elem = env->head->next;
450 kfree(env->head);
451 env->head = elem;
452 env->stack_size--;
453 return insn_idx;
454 }
455
456 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
457 int insn_idx, int prev_insn_idx)
458 {
459 struct bpf_verifier_stack_elem *elem;
460
461 elem = kmalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
462 if (!elem)
463 goto err;
464
465 memcpy(&elem->st, &env->cur_state, sizeof(env->cur_state));
466 elem->insn_idx = insn_idx;
467 elem->prev_insn_idx = prev_insn_idx;
468 elem->next = env->head;
469 env->head = elem;
470 env->stack_size++;
471 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
472 verbose("BPF program is too complex\n");
473 goto err;
474 }
475 return &elem->st;
476 err:
477 /* pop all elements and return */
478 while (pop_stack(env, NULL) >= 0);
479 return NULL;
480 }
481
482 #define CALLER_SAVED_REGS 6
483 static const int caller_saved[CALLER_SAVED_REGS] = {
484 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
485 };
486
487 static void __mark_reg_not_init(struct bpf_reg_state *reg);
488
489 /* Mark the unknown part of a register (variable offset or scalar value) as
490 * known to have the value @imm.
491 */
492 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
493 {
494 reg->id = 0;
495 reg->var_off = tnum_const(imm);
496 reg->smin_value = (s64)imm;
497 reg->smax_value = (s64)imm;
498 reg->umin_value = imm;
499 reg->umax_value = imm;
500 }
501
502 /* Mark the 'variable offset' part of a register as zero. This should be
503 * used only on registers holding a pointer type.
504 */
505 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
506 {
507 __mark_reg_known(reg, 0);
508 }
509
510 static void mark_reg_known_zero(struct bpf_reg_state *regs, u32 regno)
511 {
512 if (WARN_ON(regno >= MAX_BPF_REG)) {
513 verbose("mark_reg_known_zero(regs, %u)\n", regno);
514 /* Something bad happened, let's kill all regs */
515 for (regno = 0; regno < MAX_BPF_REG; regno++)
516 __mark_reg_not_init(regs + regno);
517 return;
518 }
519 __mark_reg_known_zero(regs + regno);
520 }
521
522 /* Attempts to improve min/max values based on var_off information */
523 static void __update_reg_bounds(struct bpf_reg_state *reg)
524 {
525 /* min signed is max(sign bit) | min(other bits) */
526 reg->smin_value = max_t(s64, reg->smin_value,
527 reg->var_off.value | (reg->var_off.mask & S64_MIN));
528 /* max signed is min(sign bit) | max(other bits) */
529 reg->smax_value = min_t(s64, reg->smax_value,
530 reg->var_off.value | (reg->var_off.mask & S64_MAX));
531 reg->umin_value = max(reg->umin_value, reg->var_off.value);
532 reg->umax_value = min(reg->umax_value,
533 reg->var_off.value | reg->var_off.mask);
534 }
535
536 /* Uses signed min/max values to inform unsigned, and vice-versa */
537 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
538 {
539 /* Learn sign from signed bounds.
540 * If we cannot cross the sign boundary, then signed and unsigned bounds
541 * are the same, so combine. This works even in the negative case, e.g.
542 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
543 */
544 if (reg->smin_value >= 0 || reg->smax_value < 0) {
545 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
546 reg->umin_value);
547 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
548 reg->umax_value);
549 return;
550 }
551 /* Learn sign from unsigned bounds. Signed bounds cross the sign
552 * boundary, so we must be careful.
553 */
554 if ((s64)reg->umax_value >= 0) {
555 /* Positive. We can't learn anything from the smin, but smax
556 * is positive, hence safe.
557 */
558 reg->smin_value = reg->umin_value;
559 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
560 reg->umax_value);
561 } else if ((s64)reg->umin_value < 0) {
562 /* Negative. We can't learn anything from the smax, but smin
563 * is negative, hence safe.
564 */
565 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
566 reg->umin_value);
567 reg->smax_value = reg->umax_value;
568 }
569 }
570
571 /* Attempts to improve var_off based on unsigned min/max information */
572 static void __reg_bound_offset(struct bpf_reg_state *reg)
573 {
574 reg->var_off = tnum_intersect(reg->var_off,
575 tnum_range(reg->umin_value,
576 reg->umax_value));
577 }
578
579 /* Reset the min/max bounds of a register */
580 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
581 {
582 reg->smin_value = S64_MIN;
583 reg->smax_value = S64_MAX;
584 reg->umin_value = 0;
585 reg->umax_value = U64_MAX;
586 }
587
588 /* Mark a register as having a completely unknown (scalar) value. */
589 static void __mark_reg_unknown(struct bpf_reg_state *reg)
590 {
591 reg->type = SCALAR_VALUE;
592 reg->id = 0;
593 reg->off = 0;
594 reg->var_off = tnum_unknown;
595 __mark_reg_unbounded(reg);
596 }
597
598 static void mark_reg_unknown(struct bpf_reg_state *regs, u32 regno)
599 {
600 if (WARN_ON(regno >= MAX_BPF_REG)) {
601 verbose("mark_reg_unknown(regs, %u)\n", regno);
602 /* Something bad happened, let's kill all regs */
603 for (regno = 0; regno < MAX_BPF_REG; regno++)
604 __mark_reg_not_init(regs + regno);
605 return;
606 }
607 __mark_reg_unknown(regs + regno);
608 }
609
610 static void __mark_reg_not_init(struct bpf_reg_state *reg)
611 {
612 __mark_reg_unknown(reg);
613 reg->type = NOT_INIT;
614 }
615
616 static void mark_reg_not_init(struct bpf_reg_state *regs, u32 regno)
617 {
618 if (WARN_ON(regno >= MAX_BPF_REG)) {
619 verbose("mark_reg_not_init(regs, %u)\n", regno);
620 /* Something bad happened, let's kill all regs */
621 for (regno = 0; regno < MAX_BPF_REG; regno++)
622 __mark_reg_not_init(regs + regno);
623 return;
624 }
625 __mark_reg_not_init(regs + regno);
626 }
627
628 static void init_reg_state(struct bpf_reg_state *regs)
629 {
630 int i;
631
632 for (i = 0; i < MAX_BPF_REG; i++) {
633 mark_reg_not_init(regs, i);
634 regs[i].live = REG_LIVE_NONE;
635 }
636
637 /* frame pointer */
638 regs[BPF_REG_FP].type = PTR_TO_STACK;
639 mark_reg_known_zero(regs, BPF_REG_FP);
640
641 /* 1st arg to a function */
642 regs[BPF_REG_1].type = PTR_TO_CTX;
643 mark_reg_known_zero(regs, BPF_REG_1);
644 }
645
646 enum reg_arg_type {
647 SRC_OP, /* register is used as source operand */
648 DST_OP, /* register is used as destination operand */
649 DST_OP_NO_MARK /* same as above, check only, don't mark */
650 };
651
652 static void mark_reg_read(const struct bpf_verifier_state *state, u32 regno)
653 {
654 struct bpf_verifier_state *parent = state->parent;
655
656 if (regno == BPF_REG_FP)
657 /* We don't need to worry about FP liveness because it's read-only */
658 return;
659
660 while (parent) {
661 /* if read wasn't screened by an earlier write ... */
662 if (state->regs[regno].live & REG_LIVE_WRITTEN)
663 break;
664 /* ... then we depend on parent's value */
665 parent->regs[regno].live |= REG_LIVE_READ;
666 state = parent;
667 parent = state->parent;
668 }
669 }
670
671 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
672 enum reg_arg_type t)
673 {
674 struct bpf_reg_state *regs = env->cur_state.regs;
675
676 if (regno >= MAX_BPF_REG) {
677 verbose("R%d is invalid\n", regno);
678 return -EINVAL;
679 }
680
681 if (t == SRC_OP) {
682 /* check whether register used as source operand can be read */
683 if (regs[regno].type == NOT_INIT) {
684 verbose("R%d !read_ok\n", regno);
685 return -EACCES;
686 }
687 mark_reg_read(&env->cur_state, regno);
688 } else {
689 /* check whether register used as dest operand can be written to */
690 if (regno == BPF_REG_FP) {
691 verbose("frame pointer is read only\n");
692 return -EACCES;
693 }
694 regs[regno].live |= REG_LIVE_WRITTEN;
695 if (t == DST_OP)
696 mark_reg_unknown(regs, regno);
697 }
698 return 0;
699 }
700
701 static bool is_spillable_regtype(enum bpf_reg_type type)
702 {
703 switch (type) {
704 case PTR_TO_MAP_VALUE:
705 case PTR_TO_MAP_VALUE_OR_NULL:
706 case PTR_TO_STACK:
707 case PTR_TO_CTX:
708 case PTR_TO_PACKET:
709 case PTR_TO_PACKET_END:
710 case CONST_PTR_TO_MAP:
711 return true;
712 default:
713 return false;
714 }
715 }
716
717 /* check_stack_read/write functions track spill/fill of registers,
718 * stack boundary and alignment are checked in check_mem_access()
719 */
720 static int check_stack_write(struct bpf_verifier_state *state, int off,
721 int size, int value_regno)
722 {
723 int i, spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
724 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
725 * so it's aligned access and [off, off + size) are within stack limits
726 */
727
728 if (value_regno >= 0 &&
729 is_spillable_regtype(state->regs[value_regno].type)) {
730
731 /* register containing pointer is being spilled into stack */
732 if (size != BPF_REG_SIZE) {
733 verbose("invalid size of register spill\n");
734 return -EACCES;
735 }
736
737 /* save register state */
738 state->spilled_regs[spi] = state->regs[value_regno];
739 state->spilled_regs[spi].live |= REG_LIVE_WRITTEN;
740
741 for (i = 0; i < BPF_REG_SIZE; i++)
742 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL;
743 } else {
744 /* regular write of data into stack */
745 state->spilled_regs[spi] = (struct bpf_reg_state) {};
746
747 for (i = 0; i < size; i++)
748 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC;
749 }
750 return 0;
751 }
752
753 static void mark_stack_slot_read(const struct bpf_verifier_state *state, int slot)
754 {
755 struct bpf_verifier_state *parent = state->parent;
756
757 while (parent) {
758 /* if read wasn't screened by an earlier write ... */
759 if (state->spilled_regs[slot].live & REG_LIVE_WRITTEN)
760 break;
761 /* ... then we depend on parent's value */
762 parent->spilled_regs[slot].live |= REG_LIVE_READ;
763 state = parent;
764 parent = state->parent;
765 }
766 }
767
768 static int check_stack_read(struct bpf_verifier_state *state, int off, int size,
769 int value_regno)
770 {
771 u8 *slot_type;
772 int i, spi;
773
774 slot_type = &state->stack_slot_type[MAX_BPF_STACK + off];
775
776 if (slot_type[0] == STACK_SPILL) {
777 if (size != BPF_REG_SIZE) {
778 verbose("invalid size of register spill\n");
779 return -EACCES;
780 }
781 for (i = 1; i < BPF_REG_SIZE; i++) {
782 if (slot_type[i] != STACK_SPILL) {
783 verbose("corrupted spill memory\n");
784 return -EACCES;
785 }
786 }
787
788 spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
789
790 if (value_regno >= 0) {
791 /* restore register state from stack */
792 state->regs[value_regno] = state->spilled_regs[spi];
793 mark_stack_slot_read(state, spi);
794 }
795 return 0;
796 } else {
797 for (i = 0; i < size; i++) {
798 if (slot_type[i] != STACK_MISC) {
799 verbose("invalid read from stack off %d+%d size %d\n",
800 off, i, size);
801 return -EACCES;
802 }
803 }
804 if (value_regno >= 0)
805 /* have read misc data from the stack */
806 mark_reg_unknown(state->regs, value_regno);
807 return 0;
808 }
809 }
810
811 /* check read/write into map element returned by bpf_map_lookup_elem() */
812 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
813 int size)
814 {
815 struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
816
817 if (off < 0 || size <= 0 || off + size > map->value_size) {
818 verbose("invalid access to map value, value_size=%d off=%d size=%d\n",
819 map->value_size, off, size);
820 return -EACCES;
821 }
822 return 0;
823 }
824
825 /* check read/write into a map element with possible variable offset */
826 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
827 int off, int size)
828 {
829 struct bpf_verifier_state *state = &env->cur_state;
830 struct bpf_reg_state *reg = &state->regs[regno];
831 int err;
832
833 /* We may have adjusted the register to this map value, so we
834 * need to try adding each of min_value and max_value to off
835 * to make sure our theoretical access will be safe.
836 */
837 if (log_level)
838 print_verifier_state(state);
839 /* The minimum value is only important with signed
840 * comparisons where we can't assume the floor of a
841 * value is 0. If we are using signed variables for our
842 * index'es we need to make sure that whatever we use
843 * will have a set floor within our range.
844 */
845 if (reg->smin_value < 0) {
846 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
847 regno);
848 return -EACCES;
849 }
850 err = __check_map_access(env, regno, reg->smin_value + off, size);
851 if (err) {
852 verbose("R%d min value is outside of the array range\n", regno);
853 return err;
854 }
855
856 /* If we haven't set a max value then we need to bail since we can't be
857 * sure we won't do bad things.
858 * If reg->umax_value + off could overflow, treat that as unbounded too.
859 */
860 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
861 verbose("R%d unbounded memory access, make sure to bounds check any array access into a map\n",
862 regno);
863 return -EACCES;
864 }
865 err = __check_map_access(env, regno, reg->umax_value + off, size);
866 if (err)
867 verbose("R%d max value is outside of the array range\n", regno);
868 return err;
869 }
870
871 #define MAX_PACKET_OFF 0xffff
872
873 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
874 const struct bpf_call_arg_meta *meta,
875 enum bpf_access_type t)
876 {
877 switch (env->prog->type) {
878 case BPF_PROG_TYPE_LWT_IN:
879 case BPF_PROG_TYPE_LWT_OUT:
880 /* dst_input() and dst_output() can't write for now */
881 if (t == BPF_WRITE)
882 return false;
883 /* fallthrough */
884 case BPF_PROG_TYPE_SCHED_CLS:
885 case BPF_PROG_TYPE_SCHED_ACT:
886 case BPF_PROG_TYPE_XDP:
887 case BPF_PROG_TYPE_LWT_XMIT:
888 case BPF_PROG_TYPE_SK_SKB:
889 if (meta)
890 return meta->pkt_access;
891
892 env->seen_direct_write = true;
893 return true;
894 default:
895 return false;
896 }
897 }
898
899 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
900 int off, int size)
901 {
902 struct bpf_reg_state *regs = env->cur_state.regs;
903 struct bpf_reg_state *reg = &regs[regno];
904
905 if (off < 0 || size <= 0 || (u64)off + size > reg->range) {
906 verbose("invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
907 off, size, regno, reg->id, reg->off, reg->range);
908 return -EACCES;
909 }
910 return 0;
911 }
912
913 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
914 int size)
915 {
916 struct bpf_reg_state *regs = env->cur_state.regs;
917 struct bpf_reg_state *reg = &regs[regno];
918 int err;
919
920 /* We may have added a variable offset to the packet pointer; but any
921 * reg->range we have comes after that. We are only checking the fixed
922 * offset.
923 */
924
925 /* We don't allow negative numbers, because we aren't tracking enough
926 * detail to prove they're safe.
927 */
928 if (reg->smin_value < 0) {
929 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
930 regno);
931 return -EACCES;
932 }
933 err = __check_packet_access(env, regno, off, size);
934 if (err) {
935 verbose("R%d offset is outside of the packet\n", regno);
936 return err;
937 }
938 return err;
939 }
940
941 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
942 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
943 enum bpf_access_type t, enum bpf_reg_type *reg_type)
944 {
945 struct bpf_insn_access_aux info = {
946 .reg_type = *reg_type,
947 };
948
949 /* for analyzer ctx accesses are already validated and converted */
950 if (env->analyzer_ops)
951 return 0;
952
953 if (env->prog->aux->ops->is_valid_access &&
954 env->prog->aux->ops->is_valid_access(off, size, t, &info)) {
955 /* A non zero info.ctx_field_size indicates that this field is a
956 * candidate for later verifier transformation to load the whole
957 * field and then apply a mask when accessed with a narrower
958 * access than actual ctx access size. A zero info.ctx_field_size
959 * will only allow for whole field access and rejects any other
960 * type of narrower access.
961 */
962 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
963 *reg_type = info.reg_type;
964
965 /* remember the offset of last byte accessed in ctx */
966 if (env->prog->aux->max_ctx_offset < off + size)
967 env->prog->aux->max_ctx_offset = off + size;
968 return 0;
969 }
970
971 verbose("invalid bpf_context access off=%d size=%d\n", off, size);
972 return -EACCES;
973 }
974
975 static bool __is_pointer_value(bool allow_ptr_leaks,
976 const struct bpf_reg_state *reg)
977 {
978 if (allow_ptr_leaks)
979 return false;
980
981 return reg->type != SCALAR_VALUE;
982 }
983
984 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
985 {
986 return __is_pointer_value(env->allow_ptr_leaks, &env->cur_state.regs[regno]);
987 }
988
989 static int check_pkt_ptr_alignment(const struct bpf_reg_state *reg,
990 int off, int size, bool strict)
991 {
992 struct tnum reg_off;
993 int ip_align;
994
995 /* Byte size accesses are always allowed. */
996 if (!strict || size == 1)
997 return 0;
998
999 /* For platforms that do not have a Kconfig enabling
1000 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1001 * NET_IP_ALIGN is universally set to '2'. And on platforms
1002 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1003 * to this code only in strict mode where we want to emulate
1004 * the NET_IP_ALIGN==2 checking. Therefore use an
1005 * unconditional IP align value of '2'.
1006 */
1007 ip_align = 2;
1008
1009 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1010 if (!tnum_is_aligned(reg_off, size)) {
1011 char tn_buf[48];
1012
1013 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1014 verbose("misaligned packet access off %d+%s+%d+%d size %d\n",
1015 ip_align, tn_buf, reg->off, off, size);
1016 return -EACCES;
1017 }
1018
1019 return 0;
1020 }
1021
1022 static int check_generic_ptr_alignment(const struct bpf_reg_state *reg,
1023 const char *pointer_desc,
1024 int off, int size, bool strict)
1025 {
1026 struct tnum reg_off;
1027
1028 /* Byte size accesses are always allowed. */
1029 if (!strict || size == 1)
1030 return 0;
1031
1032 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1033 if (!tnum_is_aligned(reg_off, size)) {
1034 char tn_buf[48];
1035
1036 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1037 verbose("misaligned %saccess off %s+%d+%d size %d\n",
1038 pointer_desc, tn_buf, reg->off, off, size);
1039 return -EACCES;
1040 }
1041
1042 return 0;
1043 }
1044
1045 static int check_ptr_alignment(struct bpf_verifier_env *env,
1046 const struct bpf_reg_state *reg,
1047 int off, int size)
1048 {
1049 bool strict = env->strict_alignment;
1050 const char *pointer_desc = "";
1051
1052 switch (reg->type) {
1053 case PTR_TO_PACKET:
1054 /* special case, because of NET_IP_ALIGN */
1055 return check_pkt_ptr_alignment(reg, off, size, strict);
1056 case PTR_TO_MAP_VALUE:
1057 pointer_desc = "value ";
1058 break;
1059 case PTR_TO_CTX:
1060 pointer_desc = "context ";
1061 break;
1062 case PTR_TO_STACK:
1063 pointer_desc = "stack ";
1064 break;
1065 default:
1066 break;
1067 }
1068 return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
1069 }
1070
1071 /* check whether memory at (regno + off) is accessible for t = (read | write)
1072 * if t==write, value_regno is a register which value is stored into memory
1073 * if t==read, value_regno is a register which will receive the value from memory
1074 * if t==write && value_regno==-1, some unknown value is stored into memory
1075 * if t==read && value_regno==-1, don't care what we read from memory
1076 */
1077 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1078 int bpf_size, enum bpf_access_type t,
1079 int value_regno)
1080 {
1081 struct bpf_verifier_state *state = &env->cur_state;
1082 struct bpf_reg_state *reg = &state->regs[regno];
1083 int size, err = 0;
1084
1085 size = bpf_size_to_bytes(bpf_size);
1086 if (size < 0)
1087 return size;
1088
1089 /* alignment checks will add in reg->off themselves */
1090 err = check_ptr_alignment(env, reg, off, size);
1091 if (err)
1092 return err;
1093
1094 /* for access checks, reg->off is just part of off */
1095 off += reg->off;
1096
1097 if (reg->type == PTR_TO_MAP_VALUE) {
1098 if (t == BPF_WRITE && value_regno >= 0 &&
1099 is_pointer_value(env, value_regno)) {
1100 verbose("R%d leaks addr into map\n", value_regno);
1101 return -EACCES;
1102 }
1103
1104 err = check_map_access(env, regno, off, size);
1105 if (!err && t == BPF_READ && value_regno >= 0)
1106 mark_reg_unknown(state->regs, value_regno);
1107
1108 } else if (reg->type == PTR_TO_CTX) {
1109 enum bpf_reg_type reg_type = SCALAR_VALUE;
1110
1111 if (t == BPF_WRITE && value_regno >= 0 &&
1112 is_pointer_value(env, value_regno)) {
1113 verbose("R%d leaks addr into ctx\n", value_regno);
1114 return -EACCES;
1115 }
1116 /* ctx accesses must be at a fixed offset, so that we can
1117 * determine what type of data were returned.
1118 */
1119 if (reg->off) {
1120 verbose("dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1121 regno, reg->off, off - reg->off);
1122 return -EACCES;
1123 }
1124 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1125 char tn_buf[48];
1126
1127 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1128 verbose("variable ctx access var_off=%s off=%d size=%d",
1129 tn_buf, off, size);
1130 return -EACCES;
1131 }
1132 err = check_ctx_access(env, insn_idx, off, size, t, &reg_type);
1133 if (!err && t == BPF_READ && value_regno >= 0) {
1134 /* ctx access returns either a scalar, or a
1135 * PTR_TO_PACKET[_END]. In the latter case, we know
1136 * the offset is zero.
1137 */
1138 if (reg_type == SCALAR_VALUE)
1139 mark_reg_unknown(state->regs, value_regno);
1140 else
1141 mark_reg_known_zero(state->regs, value_regno);
1142 state->regs[value_regno].id = 0;
1143 state->regs[value_regno].off = 0;
1144 state->regs[value_regno].range = 0;
1145 state->regs[value_regno].type = reg_type;
1146 }
1147
1148 } else if (reg->type == PTR_TO_STACK) {
1149 /* stack accesses must be at a fixed offset, so that we can
1150 * determine what type of data were returned.
1151 * See check_stack_read().
1152 */
1153 if (!tnum_is_const(reg->var_off)) {
1154 char tn_buf[48];
1155
1156 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1157 verbose("variable stack access var_off=%s off=%d size=%d",
1158 tn_buf, off, size);
1159 return -EACCES;
1160 }
1161 off += reg->var_off.value;
1162 if (off >= 0 || off < -MAX_BPF_STACK) {
1163 verbose("invalid stack off=%d size=%d\n", off, size);
1164 return -EACCES;
1165 }
1166
1167 if (env->prog->aux->stack_depth < -off)
1168 env->prog->aux->stack_depth = -off;
1169
1170 if (t == BPF_WRITE) {
1171 if (!env->allow_ptr_leaks &&
1172 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL &&
1173 size != BPF_REG_SIZE) {
1174 verbose("attempt to corrupt spilled pointer on stack\n");
1175 return -EACCES;
1176 }
1177 err = check_stack_write(state, off, size, value_regno);
1178 } else {
1179 err = check_stack_read(state, off, size, value_regno);
1180 }
1181 } else if (reg->type == PTR_TO_PACKET) {
1182 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1183 verbose("cannot write into packet\n");
1184 return -EACCES;
1185 }
1186 if (t == BPF_WRITE && value_regno >= 0 &&
1187 is_pointer_value(env, value_regno)) {
1188 verbose("R%d leaks addr into packet\n", value_regno);
1189 return -EACCES;
1190 }
1191 err = check_packet_access(env, regno, off, size);
1192 if (!err && t == BPF_READ && value_regno >= 0)
1193 mark_reg_unknown(state->regs, value_regno);
1194 } else {
1195 verbose("R%d invalid mem access '%s'\n",
1196 regno, reg_type_str[reg->type]);
1197 return -EACCES;
1198 }
1199
1200 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1201 state->regs[value_regno].type == SCALAR_VALUE) {
1202 /* b/h/w load zero-extends, mark upper bits as known 0 */
1203 state->regs[value_regno].var_off = tnum_cast(
1204 state->regs[value_regno].var_off, size);
1205 __update_reg_bounds(&state->regs[value_regno]);
1206 }
1207 return err;
1208 }
1209
1210 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1211 {
1212 int err;
1213
1214 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1215 insn->imm != 0) {
1216 verbose("BPF_XADD uses reserved fields\n");
1217 return -EINVAL;
1218 }
1219
1220 /* check src1 operand */
1221 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1222 if (err)
1223 return err;
1224
1225 /* check src2 operand */
1226 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1227 if (err)
1228 return err;
1229
1230 if (is_pointer_value(env, insn->src_reg)) {
1231 verbose("R%d leaks addr into mem\n", insn->src_reg);
1232 return -EACCES;
1233 }
1234
1235 /* check whether atomic_add can read the memory */
1236 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1237 BPF_SIZE(insn->code), BPF_READ, -1);
1238 if (err)
1239 return err;
1240
1241 /* check whether atomic_add can write into the same memory */
1242 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1243 BPF_SIZE(insn->code), BPF_WRITE, -1);
1244 }
1245
1246 /* Does this register contain a constant zero? */
1247 static bool register_is_null(struct bpf_reg_state reg)
1248 {
1249 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1250 }
1251
1252 /* when register 'regno' is passed into function that will read 'access_size'
1253 * bytes from that pointer, make sure that it's within stack boundary
1254 * and all elements of stack are initialized.
1255 * Unlike most pointer bounds-checking functions, this one doesn't take an
1256 * 'off' argument, so it has to add in reg->off itself.
1257 */
1258 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1259 int access_size, bool zero_size_allowed,
1260 struct bpf_call_arg_meta *meta)
1261 {
1262 struct bpf_verifier_state *state = &env->cur_state;
1263 struct bpf_reg_state *regs = state->regs;
1264 int off, i;
1265
1266 if (regs[regno].type != PTR_TO_STACK) {
1267 /* Allow zero-byte read from NULL, regardless of pointer type */
1268 if (zero_size_allowed && access_size == 0 &&
1269 register_is_null(regs[regno]))
1270 return 0;
1271
1272 verbose("R%d type=%s expected=%s\n", regno,
1273 reg_type_str[regs[regno].type],
1274 reg_type_str[PTR_TO_STACK]);
1275 return -EACCES;
1276 }
1277
1278 /* Only allow fixed-offset stack reads */
1279 if (!tnum_is_const(regs[regno].var_off)) {
1280 char tn_buf[48];
1281
1282 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1283 verbose("invalid variable stack read R%d var_off=%s\n",
1284 regno, tn_buf);
1285 }
1286 off = regs[regno].off + regs[regno].var_off.value;
1287 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1288 access_size <= 0) {
1289 verbose("invalid stack type R%d off=%d access_size=%d\n",
1290 regno, off, access_size);
1291 return -EACCES;
1292 }
1293
1294 if (env->prog->aux->stack_depth < -off)
1295 env->prog->aux->stack_depth = -off;
1296
1297 if (meta && meta->raw_mode) {
1298 meta->access_size = access_size;
1299 meta->regno = regno;
1300 return 0;
1301 }
1302
1303 for (i = 0; i < access_size; i++) {
1304 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
1305 verbose("invalid indirect read from stack off %d+%d size %d\n",
1306 off, i, access_size);
1307 return -EACCES;
1308 }
1309 }
1310 return 0;
1311 }
1312
1313 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1314 int access_size, bool zero_size_allowed,
1315 struct bpf_call_arg_meta *meta)
1316 {
1317 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1318
1319 switch (reg->type) {
1320 case PTR_TO_PACKET:
1321 return check_packet_access(env, regno, reg->off, access_size);
1322 case PTR_TO_MAP_VALUE:
1323 return check_map_access(env, regno, reg->off, access_size);
1324 default: /* scalar_value|ptr_to_stack or invalid ptr */
1325 return check_stack_boundary(env, regno, access_size,
1326 zero_size_allowed, meta);
1327 }
1328 }
1329
1330 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1331 enum bpf_arg_type arg_type,
1332 struct bpf_call_arg_meta *meta)
1333 {
1334 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1335 enum bpf_reg_type expected_type, type = reg->type;
1336 int err = 0;
1337
1338 if (arg_type == ARG_DONTCARE)
1339 return 0;
1340
1341 err = check_reg_arg(env, regno, SRC_OP);
1342 if (err)
1343 return err;
1344
1345 if (arg_type == ARG_ANYTHING) {
1346 if (is_pointer_value(env, regno)) {
1347 verbose("R%d leaks addr into helper function\n", regno);
1348 return -EACCES;
1349 }
1350 return 0;
1351 }
1352
1353 if (type == PTR_TO_PACKET &&
1354 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1355 verbose("helper access to the packet is not allowed\n");
1356 return -EACCES;
1357 }
1358
1359 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1360 arg_type == ARG_PTR_TO_MAP_VALUE) {
1361 expected_type = PTR_TO_STACK;
1362 if (type != PTR_TO_PACKET && type != expected_type)
1363 goto err_type;
1364 } else if (arg_type == ARG_CONST_SIZE ||
1365 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1366 expected_type = SCALAR_VALUE;
1367 if (type != expected_type)
1368 goto err_type;
1369 } else if (arg_type == ARG_CONST_MAP_PTR) {
1370 expected_type = CONST_PTR_TO_MAP;
1371 if (type != expected_type)
1372 goto err_type;
1373 } else if (arg_type == ARG_PTR_TO_CTX) {
1374 expected_type = PTR_TO_CTX;
1375 if (type != expected_type)
1376 goto err_type;
1377 } else if (arg_type == ARG_PTR_TO_MEM ||
1378 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1379 expected_type = PTR_TO_STACK;
1380 /* One exception here. In case function allows for NULL to be
1381 * passed in as argument, it's a SCALAR_VALUE type. Final test
1382 * happens during stack boundary checking.
1383 */
1384 if (register_is_null(*reg))
1385 /* final test in check_stack_boundary() */;
1386 else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
1387 type != expected_type)
1388 goto err_type;
1389 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1390 } else {
1391 verbose("unsupported arg_type %d\n", arg_type);
1392 return -EFAULT;
1393 }
1394
1395 if (arg_type == ARG_CONST_MAP_PTR) {
1396 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1397 meta->map_ptr = reg->map_ptr;
1398 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1399 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1400 * check that [key, key + map->key_size) are within
1401 * stack limits and initialized
1402 */
1403 if (!meta->map_ptr) {
1404 /* in function declaration map_ptr must come before
1405 * map_key, so that it's verified and known before
1406 * we have to check map_key here. Otherwise it means
1407 * that kernel subsystem misconfigured verifier
1408 */
1409 verbose("invalid map_ptr to access map->key\n");
1410 return -EACCES;
1411 }
1412 if (type == PTR_TO_PACKET)
1413 err = check_packet_access(env, regno, reg->off,
1414 meta->map_ptr->key_size);
1415 else
1416 err = check_stack_boundary(env, regno,
1417 meta->map_ptr->key_size,
1418 false, NULL);
1419 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1420 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1421 * check [value, value + map->value_size) validity
1422 */
1423 if (!meta->map_ptr) {
1424 /* kernel subsystem misconfigured verifier */
1425 verbose("invalid map_ptr to access map->value\n");
1426 return -EACCES;
1427 }
1428 if (type == PTR_TO_PACKET)
1429 err = check_packet_access(env, regno, reg->off,
1430 meta->map_ptr->value_size);
1431 else
1432 err = check_stack_boundary(env, regno,
1433 meta->map_ptr->value_size,
1434 false, NULL);
1435 } else if (arg_type == ARG_CONST_SIZE ||
1436 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1437 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1438
1439 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1440 * from stack pointer 'buf'. Check it
1441 * note: regno == len, regno - 1 == buf
1442 */
1443 if (regno == 0) {
1444 /* kernel subsystem misconfigured verifier */
1445 verbose("ARG_CONST_SIZE cannot be first argument\n");
1446 return -EACCES;
1447 }
1448
1449 /* The register is SCALAR_VALUE; the access check
1450 * happens using its boundaries.
1451 */
1452
1453 if (!tnum_is_const(reg->var_off))
1454 /* For unprivileged variable accesses, disable raw
1455 * mode so that the program is required to
1456 * initialize all the memory that the helper could
1457 * just partially fill up.
1458 */
1459 meta = NULL;
1460
1461 if (reg->smin_value < 0) {
1462 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1463 regno);
1464 return -EACCES;
1465 }
1466
1467 if (reg->umin_value == 0) {
1468 err = check_helper_mem_access(env, regno - 1, 0,
1469 zero_size_allowed,
1470 meta);
1471 if (err)
1472 return err;
1473 }
1474
1475 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1476 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1477 regno);
1478 return -EACCES;
1479 }
1480 err = check_helper_mem_access(env, regno - 1,
1481 reg->umax_value,
1482 zero_size_allowed, meta);
1483 }
1484
1485 return err;
1486 err_type:
1487 verbose("R%d type=%s expected=%s\n", regno,
1488 reg_type_str[type], reg_type_str[expected_type]);
1489 return -EACCES;
1490 }
1491
1492 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
1493 {
1494 if (!map)
1495 return 0;
1496
1497 /* We need a two way check, first is from map perspective ... */
1498 switch (map->map_type) {
1499 case BPF_MAP_TYPE_PROG_ARRAY:
1500 if (func_id != BPF_FUNC_tail_call)
1501 goto error;
1502 break;
1503 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1504 if (func_id != BPF_FUNC_perf_event_read &&
1505 func_id != BPF_FUNC_perf_event_output)
1506 goto error;
1507 break;
1508 case BPF_MAP_TYPE_STACK_TRACE:
1509 if (func_id != BPF_FUNC_get_stackid)
1510 goto error;
1511 break;
1512 case BPF_MAP_TYPE_CGROUP_ARRAY:
1513 if (func_id != BPF_FUNC_skb_under_cgroup &&
1514 func_id != BPF_FUNC_current_task_under_cgroup)
1515 goto error;
1516 break;
1517 /* devmap returns a pointer to a live net_device ifindex that we cannot
1518 * allow to be modified from bpf side. So do not allow lookup elements
1519 * for now.
1520 */
1521 case BPF_MAP_TYPE_DEVMAP:
1522 if (func_id != BPF_FUNC_redirect_map)
1523 goto error;
1524 break;
1525 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1526 case BPF_MAP_TYPE_HASH_OF_MAPS:
1527 if (func_id != BPF_FUNC_map_lookup_elem)
1528 goto error;
1529 break;
1530 case BPF_MAP_TYPE_SOCKMAP:
1531 if (func_id != BPF_FUNC_sk_redirect_map &&
1532 func_id != BPF_FUNC_sock_map_update &&
1533 func_id != BPF_FUNC_map_delete_elem)
1534 goto error;
1535 break;
1536 default:
1537 break;
1538 }
1539
1540 /* ... and second from the function itself. */
1541 switch (func_id) {
1542 case BPF_FUNC_tail_call:
1543 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1544 goto error;
1545 break;
1546 case BPF_FUNC_perf_event_read:
1547 case BPF_FUNC_perf_event_output:
1548 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1549 goto error;
1550 break;
1551 case BPF_FUNC_get_stackid:
1552 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1553 goto error;
1554 break;
1555 case BPF_FUNC_current_task_under_cgroup:
1556 case BPF_FUNC_skb_under_cgroup:
1557 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1558 goto error;
1559 break;
1560 case BPF_FUNC_redirect_map:
1561 if (map->map_type != BPF_MAP_TYPE_DEVMAP)
1562 goto error;
1563 break;
1564 case BPF_FUNC_sk_redirect_map:
1565 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1566 goto error;
1567 break;
1568 case BPF_FUNC_sock_map_update:
1569 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1570 goto error;
1571 break;
1572 default:
1573 break;
1574 }
1575
1576 return 0;
1577 error:
1578 verbose("cannot pass map_type %d into func %s#%d\n",
1579 map->map_type, func_id_name(func_id), func_id);
1580 return -EINVAL;
1581 }
1582
1583 static int check_raw_mode(const struct bpf_func_proto *fn)
1584 {
1585 int count = 0;
1586
1587 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1588 count++;
1589 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1590 count++;
1591 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1592 count++;
1593 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1594 count++;
1595 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1596 count++;
1597
1598 return count > 1 ? -EINVAL : 0;
1599 }
1600
1601 /* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
1602 * so turn them into unknown SCALAR_VALUE.
1603 */
1604 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1605 {
1606 struct bpf_verifier_state *state = &env->cur_state;
1607 struct bpf_reg_state *regs = state->regs, *reg;
1608 int i;
1609
1610 for (i = 0; i < MAX_BPF_REG; i++)
1611 if (regs[i].type == PTR_TO_PACKET ||
1612 regs[i].type == PTR_TO_PACKET_END)
1613 mark_reg_unknown(regs, i);
1614
1615 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
1616 if (state->stack_slot_type[i] != STACK_SPILL)
1617 continue;
1618 reg = &state->spilled_regs[i / BPF_REG_SIZE];
1619 if (reg->type != PTR_TO_PACKET &&
1620 reg->type != PTR_TO_PACKET_END)
1621 continue;
1622 __mark_reg_unknown(reg);
1623 }
1624 }
1625
1626 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1627 {
1628 struct bpf_verifier_state *state = &env->cur_state;
1629 const struct bpf_func_proto *fn = NULL;
1630 struct bpf_reg_state *regs = state->regs;
1631 struct bpf_call_arg_meta meta;
1632 bool changes_data;
1633 int i, err;
1634
1635 /* find function prototype */
1636 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1637 verbose("invalid func %s#%d\n", func_id_name(func_id), func_id);
1638 return -EINVAL;
1639 }
1640
1641 if (env->prog->aux->ops->get_func_proto)
1642 fn = env->prog->aux->ops->get_func_proto(func_id);
1643
1644 if (!fn) {
1645 verbose("unknown func %s#%d\n", func_id_name(func_id), func_id);
1646 return -EINVAL;
1647 }
1648
1649 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1650 if (!env->prog->gpl_compatible && fn->gpl_only) {
1651 verbose("cannot call GPL only function from proprietary program\n");
1652 return -EINVAL;
1653 }
1654
1655 changes_data = bpf_helper_changes_pkt_data(fn->func);
1656
1657 memset(&meta, 0, sizeof(meta));
1658 meta.pkt_access = fn->pkt_access;
1659
1660 /* We only support one arg being in raw mode at the moment, which
1661 * is sufficient for the helper functions we have right now.
1662 */
1663 err = check_raw_mode(fn);
1664 if (err) {
1665 verbose("kernel subsystem misconfigured func %s#%d\n",
1666 func_id_name(func_id), func_id);
1667 return err;
1668 }
1669
1670 /* check args */
1671 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1672 if (err)
1673 return err;
1674 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1675 if (err)
1676 return err;
1677 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1678 if (err)
1679 return err;
1680 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1681 if (err)
1682 return err;
1683 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1684 if (err)
1685 return err;
1686
1687 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1688 * is inferred from register state.
1689 */
1690 for (i = 0; i < meta.access_size; i++) {
1691 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1692 if (err)
1693 return err;
1694 }
1695
1696 /* reset caller saved regs */
1697 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1698 mark_reg_not_init(regs, caller_saved[i]);
1699 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1700 }
1701
1702 /* update return register (already marked as written above) */
1703 if (fn->ret_type == RET_INTEGER) {
1704 /* sets type to SCALAR_VALUE */
1705 mark_reg_unknown(regs, BPF_REG_0);
1706 } else if (fn->ret_type == RET_VOID) {
1707 regs[BPF_REG_0].type = NOT_INIT;
1708 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1709 struct bpf_insn_aux_data *insn_aux;
1710
1711 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1712 /* There is no offset yet applied, variable or fixed */
1713 mark_reg_known_zero(regs, BPF_REG_0);
1714 regs[BPF_REG_0].off = 0;
1715 /* remember map_ptr, so that check_map_access()
1716 * can check 'value_size' boundary of memory access
1717 * to map element returned from bpf_map_lookup_elem()
1718 */
1719 if (meta.map_ptr == NULL) {
1720 verbose("kernel subsystem misconfigured verifier\n");
1721 return -EINVAL;
1722 }
1723 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1724 regs[BPF_REG_0].id = ++env->id_gen;
1725 insn_aux = &env->insn_aux_data[insn_idx];
1726 if (!insn_aux->map_ptr)
1727 insn_aux->map_ptr = meta.map_ptr;
1728 else if (insn_aux->map_ptr != meta.map_ptr)
1729 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1730 } else {
1731 verbose("unknown return type %d of func %s#%d\n",
1732 fn->ret_type, func_id_name(func_id), func_id);
1733 return -EINVAL;
1734 }
1735
1736 err = check_map_func_compatibility(meta.map_ptr, func_id);
1737 if (err)
1738 return err;
1739
1740 if (changes_data)
1741 clear_all_pkt_pointers(env);
1742 return 0;
1743 }
1744
1745 static void coerce_reg_to_32(struct bpf_reg_state *reg)
1746 {
1747 /* clear high 32 bits */
1748 reg->var_off = tnum_cast(reg->var_off, 4);
1749 /* Update bounds */
1750 __update_reg_bounds(reg);
1751 }
1752
1753 static bool signed_add_overflows(s64 a, s64 b)
1754 {
1755 /* Do the add in u64, where overflow is well-defined */
1756 s64 res = (s64)((u64)a + (u64)b);
1757
1758 if (b < 0)
1759 return res > a;
1760 return res < a;
1761 }
1762
1763 static bool signed_sub_overflows(s64 a, s64 b)
1764 {
1765 /* Do the sub in u64, where overflow is well-defined */
1766 s64 res = (s64)((u64)a - (u64)b);
1767
1768 if (b < 0)
1769 return res < a;
1770 return res > a;
1771 }
1772
1773 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1774 * Caller should also handle BPF_MOV case separately.
1775 * If we return -EACCES, caller may want to try again treating pointer as a
1776 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1777 */
1778 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1779 struct bpf_insn *insn,
1780 const struct bpf_reg_state *ptr_reg,
1781 const struct bpf_reg_state *off_reg)
1782 {
1783 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
1784 bool known = tnum_is_const(off_reg->var_off);
1785 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1786 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1787 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1788 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1789 u8 opcode = BPF_OP(insn->code);
1790 u32 dst = insn->dst_reg;
1791
1792 dst_reg = &regs[dst];
1793
1794 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1795 print_verifier_state(&env->cur_state);
1796 verbose("verifier internal error: known but bad sbounds\n");
1797 return -EINVAL;
1798 }
1799 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1800 print_verifier_state(&env->cur_state);
1801 verbose("verifier internal error: known but bad ubounds\n");
1802 return -EINVAL;
1803 }
1804
1805 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1806 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1807 if (!env->allow_ptr_leaks)
1808 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1809 dst);
1810 return -EACCES;
1811 }
1812
1813 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1814 if (!env->allow_ptr_leaks)
1815 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1816 dst);
1817 return -EACCES;
1818 }
1819 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1820 if (!env->allow_ptr_leaks)
1821 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1822 dst);
1823 return -EACCES;
1824 }
1825 if (ptr_reg->type == PTR_TO_PACKET_END) {
1826 if (!env->allow_ptr_leaks)
1827 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1828 dst);
1829 return -EACCES;
1830 }
1831
1832 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1833 * The id may be overwritten later if we create a new variable offset.
1834 */
1835 dst_reg->type = ptr_reg->type;
1836 dst_reg->id = ptr_reg->id;
1837
1838 switch (opcode) {
1839 case BPF_ADD:
1840 /* We can take a fixed offset as long as it doesn't overflow
1841 * the s32 'off' field
1842 */
1843 if (known && (ptr_reg->off + smin_val ==
1844 (s64)(s32)(ptr_reg->off + smin_val))) {
1845 /* pointer += K. Accumulate it into fixed offset */
1846 dst_reg->smin_value = smin_ptr;
1847 dst_reg->smax_value = smax_ptr;
1848 dst_reg->umin_value = umin_ptr;
1849 dst_reg->umax_value = umax_ptr;
1850 dst_reg->var_off = ptr_reg->var_off;
1851 dst_reg->off = ptr_reg->off + smin_val;
1852 dst_reg->range = ptr_reg->range;
1853 break;
1854 }
1855 /* A new variable offset is created. Note that off_reg->off
1856 * == 0, since it's a scalar.
1857 * dst_reg gets the pointer type and since some positive
1858 * integer value was added to the pointer, give it a new 'id'
1859 * if it's a PTR_TO_PACKET.
1860 * this creates a new 'base' pointer, off_reg (variable) gets
1861 * added into the variable offset, and we copy the fixed offset
1862 * from ptr_reg.
1863 */
1864 if (signed_add_overflows(smin_ptr, smin_val) ||
1865 signed_add_overflows(smax_ptr, smax_val)) {
1866 dst_reg->smin_value = S64_MIN;
1867 dst_reg->smax_value = S64_MAX;
1868 } else {
1869 dst_reg->smin_value = smin_ptr + smin_val;
1870 dst_reg->smax_value = smax_ptr + smax_val;
1871 }
1872 if (umin_ptr + umin_val < umin_ptr ||
1873 umax_ptr + umax_val < umax_ptr) {
1874 dst_reg->umin_value = 0;
1875 dst_reg->umax_value = U64_MAX;
1876 } else {
1877 dst_reg->umin_value = umin_ptr + umin_val;
1878 dst_reg->umax_value = umax_ptr + umax_val;
1879 }
1880 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1881 dst_reg->off = ptr_reg->off;
1882 if (ptr_reg->type == PTR_TO_PACKET) {
1883 dst_reg->id = ++env->id_gen;
1884 /* something was added to pkt_ptr, set range to zero */
1885 dst_reg->range = 0;
1886 }
1887 break;
1888 case BPF_SUB:
1889 if (dst_reg == off_reg) {
1890 /* scalar -= pointer. Creates an unknown scalar */
1891 if (!env->allow_ptr_leaks)
1892 verbose("R%d tried to subtract pointer from scalar\n",
1893 dst);
1894 return -EACCES;
1895 }
1896 /* We don't allow subtraction from FP, because (according to
1897 * test_verifier.c test "invalid fp arithmetic", JITs might not
1898 * be able to deal with it.
1899 */
1900 if (ptr_reg->type == PTR_TO_STACK) {
1901 if (!env->allow_ptr_leaks)
1902 verbose("R%d subtraction from stack pointer prohibited\n",
1903 dst);
1904 return -EACCES;
1905 }
1906 if (known && (ptr_reg->off - smin_val ==
1907 (s64)(s32)(ptr_reg->off - smin_val))) {
1908 /* pointer -= K. Subtract it from fixed offset */
1909 dst_reg->smin_value = smin_ptr;
1910 dst_reg->smax_value = smax_ptr;
1911 dst_reg->umin_value = umin_ptr;
1912 dst_reg->umax_value = umax_ptr;
1913 dst_reg->var_off = ptr_reg->var_off;
1914 dst_reg->id = ptr_reg->id;
1915 dst_reg->off = ptr_reg->off - smin_val;
1916 dst_reg->range = ptr_reg->range;
1917 break;
1918 }
1919 /* A new variable offset is created. If the subtrahend is known
1920 * nonnegative, then any reg->range we had before is still good.
1921 */
1922 if (signed_sub_overflows(smin_ptr, smax_val) ||
1923 signed_sub_overflows(smax_ptr, smin_val)) {
1924 /* Overflow possible, we know nothing */
1925 dst_reg->smin_value = S64_MIN;
1926 dst_reg->smax_value = S64_MAX;
1927 } else {
1928 dst_reg->smin_value = smin_ptr - smax_val;
1929 dst_reg->smax_value = smax_ptr - smin_val;
1930 }
1931 if (umin_ptr < umax_val) {
1932 /* Overflow possible, we know nothing */
1933 dst_reg->umin_value = 0;
1934 dst_reg->umax_value = U64_MAX;
1935 } else {
1936 /* Cannot overflow (as long as bounds are consistent) */
1937 dst_reg->umin_value = umin_ptr - umax_val;
1938 dst_reg->umax_value = umax_ptr - umin_val;
1939 }
1940 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
1941 dst_reg->off = ptr_reg->off;
1942 if (ptr_reg->type == PTR_TO_PACKET) {
1943 dst_reg->id = ++env->id_gen;
1944 /* something was added to pkt_ptr, set range to zero */
1945 if (smin_val < 0)
1946 dst_reg->range = 0;
1947 }
1948 break;
1949 case BPF_AND:
1950 case BPF_OR:
1951 case BPF_XOR:
1952 /* bitwise ops on pointers are troublesome, prohibit for now.
1953 * (However, in principle we could allow some cases, e.g.
1954 * ptr &= ~3 which would reduce min_value by 3.)
1955 */
1956 if (!env->allow_ptr_leaks)
1957 verbose("R%d bitwise operator %s on pointer prohibited\n",
1958 dst, bpf_alu_string[opcode >> 4]);
1959 return -EACCES;
1960 default:
1961 /* other operators (e.g. MUL,LSH) produce non-pointer results */
1962 if (!env->allow_ptr_leaks)
1963 verbose("R%d pointer arithmetic with %s operator prohibited\n",
1964 dst, bpf_alu_string[opcode >> 4]);
1965 return -EACCES;
1966 }
1967
1968 __update_reg_bounds(dst_reg);
1969 __reg_deduce_bounds(dst_reg);
1970 __reg_bound_offset(dst_reg);
1971 return 0;
1972 }
1973
1974 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
1975 struct bpf_insn *insn,
1976 struct bpf_reg_state *dst_reg,
1977 struct bpf_reg_state src_reg)
1978 {
1979 struct bpf_reg_state *regs = env->cur_state.regs;
1980 u8 opcode = BPF_OP(insn->code);
1981 bool src_known, dst_known;
1982 s64 smin_val, smax_val;
1983 u64 umin_val, umax_val;
1984
1985 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1986 /* 32-bit ALU ops are (32,32)->64 */
1987 coerce_reg_to_32(dst_reg);
1988 coerce_reg_to_32(&src_reg);
1989 }
1990 smin_val = src_reg.smin_value;
1991 smax_val = src_reg.smax_value;
1992 umin_val = src_reg.umin_value;
1993 umax_val = src_reg.umax_value;
1994 src_known = tnum_is_const(src_reg.var_off);
1995 dst_known = tnum_is_const(dst_reg->var_off);
1996
1997 switch (opcode) {
1998 case BPF_ADD:
1999 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2000 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2001 dst_reg->smin_value = S64_MIN;
2002 dst_reg->smax_value = S64_MAX;
2003 } else {
2004 dst_reg->smin_value += smin_val;
2005 dst_reg->smax_value += smax_val;
2006 }
2007 if (dst_reg->umin_value + umin_val < umin_val ||
2008 dst_reg->umax_value + umax_val < umax_val) {
2009 dst_reg->umin_value = 0;
2010 dst_reg->umax_value = U64_MAX;
2011 } else {
2012 dst_reg->umin_value += umin_val;
2013 dst_reg->umax_value += umax_val;
2014 }
2015 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2016 break;
2017 case BPF_SUB:
2018 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2019 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2020 /* Overflow possible, we know nothing */
2021 dst_reg->smin_value = S64_MIN;
2022 dst_reg->smax_value = S64_MAX;
2023 } else {
2024 dst_reg->smin_value -= smax_val;
2025 dst_reg->smax_value -= smin_val;
2026 }
2027 if (dst_reg->umin_value < umax_val) {
2028 /* Overflow possible, we know nothing */
2029 dst_reg->umin_value = 0;
2030 dst_reg->umax_value = U64_MAX;
2031 } else {
2032 /* Cannot overflow (as long as bounds are consistent) */
2033 dst_reg->umin_value -= umax_val;
2034 dst_reg->umax_value -= umin_val;
2035 }
2036 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2037 break;
2038 case BPF_MUL:
2039 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2040 if (smin_val < 0 || dst_reg->smin_value < 0) {
2041 /* Ain't nobody got time to multiply that sign */
2042 __mark_reg_unbounded(dst_reg);
2043 __update_reg_bounds(dst_reg);
2044 break;
2045 }
2046 /* Both values are positive, so we can work with unsigned and
2047 * copy the result to signed (unless it exceeds S64_MAX).
2048 */
2049 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2050 /* Potential overflow, we know nothing */
2051 __mark_reg_unbounded(dst_reg);
2052 /* (except what we can learn from the var_off) */
2053 __update_reg_bounds(dst_reg);
2054 break;
2055 }
2056 dst_reg->umin_value *= umin_val;
2057 dst_reg->umax_value *= umax_val;
2058 if (dst_reg->umax_value > S64_MAX) {
2059 /* Overflow possible, we know nothing */
2060 dst_reg->smin_value = S64_MIN;
2061 dst_reg->smax_value = S64_MAX;
2062 } else {
2063 dst_reg->smin_value = dst_reg->umin_value;
2064 dst_reg->smax_value = dst_reg->umax_value;
2065 }
2066 break;
2067 case BPF_AND:
2068 if (src_known && dst_known) {
2069 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2070 src_reg.var_off.value);
2071 break;
2072 }
2073 /* We get our minimum from the var_off, since that's inherently
2074 * bitwise. Our maximum is the minimum of the operands' maxima.
2075 */
2076 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2077 dst_reg->umin_value = dst_reg->var_off.value;
2078 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2079 if (dst_reg->smin_value < 0 || smin_val < 0) {
2080 /* Lose signed bounds when ANDing negative numbers,
2081 * ain't nobody got time for that.
2082 */
2083 dst_reg->smin_value = S64_MIN;
2084 dst_reg->smax_value = S64_MAX;
2085 } else {
2086 /* ANDing two positives gives a positive, so safe to
2087 * cast result into s64.
2088 */
2089 dst_reg->smin_value = dst_reg->umin_value;
2090 dst_reg->smax_value = dst_reg->umax_value;
2091 }
2092 /* We may learn something more from the var_off */
2093 __update_reg_bounds(dst_reg);
2094 break;
2095 case BPF_OR:
2096 if (src_known && dst_known) {
2097 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2098 src_reg.var_off.value);
2099 break;
2100 }
2101 /* We get our maximum from the var_off, and our minimum is the
2102 * maximum of the operands' minima
2103 */
2104 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2105 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2106 dst_reg->umax_value = dst_reg->var_off.value |
2107 dst_reg->var_off.mask;
2108 if (dst_reg->smin_value < 0 || smin_val < 0) {
2109 /* Lose signed bounds when ORing negative numbers,
2110 * ain't nobody got time for that.
2111 */
2112 dst_reg->smin_value = S64_MIN;
2113 dst_reg->smax_value = S64_MAX;
2114 } else {
2115 /* ORing two positives gives a positive, so safe to
2116 * cast result into s64.
2117 */
2118 dst_reg->smin_value = dst_reg->umin_value;
2119 dst_reg->smax_value = dst_reg->umax_value;
2120 }
2121 /* We may learn something more from the var_off */
2122 __update_reg_bounds(dst_reg);
2123 break;
2124 case BPF_LSH:
2125 if (umax_val > 63) {
2126 /* Shifts greater than 63 are undefined. This includes
2127 * shifts by a negative number.
2128 */
2129 mark_reg_unknown(regs, insn->dst_reg);
2130 break;
2131 }
2132 /* We lose all sign bit information (except what we can pick
2133 * up from var_off)
2134 */
2135 dst_reg->smin_value = S64_MIN;
2136 dst_reg->smax_value = S64_MAX;
2137 /* If we might shift our top bit out, then we know nothing */
2138 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2139 dst_reg->umin_value = 0;
2140 dst_reg->umax_value = U64_MAX;
2141 } else {
2142 dst_reg->umin_value <<= umin_val;
2143 dst_reg->umax_value <<= umax_val;
2144 }
2145 if (src_known)
2146 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2147 else
2148 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2149 /* We may learn something more from the var_off */
2150 __update_reg_bounds(dst_reg);
2151 break;
2152 case BPF_RSH:
2153 if (umax_val > 63) {
2154 /* Shifts greater than 63 are undefined. This includes
2155 * shifts by a negative number.
2156 */
2157 mark_reg_unknown(regs, insn->dst_reg);
2158 break;
2159 }
2160 /* BPF_RSH is an unsigned shift, so make the appropriate casts */
2161 if (dst_reg->smin_value < 0) {
2162 if (umin_val) {
2163 /* Sign bit will be cleared */
2164 dst_reg->smin_value = 0;
2165 } else {
2166 /* Lost sign bit information */
2167 dst_reg->smin_value = S64_MIN;
2168 dst_reg->smax_value = S64_MAX;
2169 }
2170 } else {
2171 dst_reg->smin_value =
2172 (u64)(dst_reg->smin_value) >> umax_val;
2173 }
2174 if (src_known)
2175 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2176 umin_val);
2177 else
2178 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2179 dst_reg->umin_value >>= umax_val;
2180 dst_reg->umax_value >>= umin_val;
2181 /* We may learn something more from the var_off */
2182 __update_reg_bounds(dst_reg);
2183 break;
2184 default:
2185 mark_reg_unknown(regs, insn->dst_reg);
2186 break;
2187 }
2188
2189 __reg_deduce_bounds(dst_reg);
2190 __reg_bound_offset(dst_reg);
2191 return 0;
2192 }
2193
2194 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2195 * and var_off.
2196 */
2197 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2198 struct bpf_insn *insn)
2199 {
2200 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
2201 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2202 u8 opcode = BPF_OP(insn->code);
2203 int rc;
2204
2205 dst_reg = &regs[insn->dst_reg];
2206 src_reg = NULL;
2207 if (dst_reg->type != SCALAR_VALUE)
2208 ptr_reg = dst_reg;
2209 if (BPF_SRC(insn->code) == BPF_X) {
2210 src_reg = &regs[insn->src_reg];
2211 if (src_reg->type != SCALAR_VALUE) {
2212 if (dst_reg->type != SCALAR_VALUE) {
2213 /* Combining two pointers by any ALU op yields
2214 * an arbitrary scalar.
2215 */
2216 if (!env->allow_ptr_leaks) {
2217 verbose("R%d pointer %s pointer prohibited\n",
2218 insn->dst_reg,
2219 bpf_alu_string[opcode >> 4]);
2220 return -EACCES;
2221 }
2222 mark_reg_unknown(regs, insn->dst_reg);
2223 return 0;
2224 } else {
2225 /* scalar += pointer
2226 * This is legal, but we have to reverse our
2227 * src/dest handling in computing the range
2228 */
2229 rc = adjust_ptr_min_max_vals(env, insn,
2230 src_reg, dst_reg);
2231 if (rc == -EACCES && env->allow_ptr_leaks) {
2232 /* scalar += unknown scalar */
2233 __mark_reg_unknown(&off_reg);
2234 return adjust_scalar_min_max_vals(
2235 env, insn,
2236 dst_reg, off_reg);
2237 }
2238 return rc;
2239 }
2240 } else if (ptr_reg) {
2241 /* pointer += scalar */
2242 rc = adjust_ptr_min_max_vals(env, insn,
2243 dst_reg, src_reg);
2244 if (rc == -EACCES && env->allow_ptr_leaks) {
2245 /* unknown scalar += scalar */
2246 __mark_reg_unknown(dst_reg);
2247 return adjust_scalar_min_max_vals(
2248 env, insn, dst_reg, *src_reg);
2249 }
2250 return rc;
2251 }
2252 } else {
2253 /* Pretend the src is a reg with a known value, since we only
2254 * need to be able to read from this state.
2255 */
2256 off_reg.type = SCALAR_VALUE;
2257 __mark_reg_known(&off_reg, insn->imm);
2258 src_reg = &off_reg;
2259 if (ptr_reg) { /* pointer += K */
2260 rc = adjust_ptr_min_max_vals(env, insn,
2261 ptr_reg, src_reg);
2262 if (rc == -EACCES && env->allow_ptr_leaks) {
2263 /* unknown scalar += K */
2264 __mark_reg_unknown(dst_reg);
2265 return adjust_scalar_min_max_vals(
2266 env, insn, dst_reg, off_reg);
2267 }
2268 return rc;
2269 }
2270 }
2271
2272 /* Got here implies adding two SCALAR_VALUEs */
2273 if (WARN_ON_ONCE(ptr_reg)) {
2274 print_verifier_state(&env->cur_state);
2275 verbose("verifier internal error: unexpected ptr_reg\n");
2276 return -EINVAL;
2277 }
2278 if (WARN_ON(!src_reg)) {
2279 print_verifier_state(&env->cur_state);
2280 verbose("verifier internal error: no src_reg\n");
2281 return -EINVAL;
2282 }
2283 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2284 }
2285
2286 /* check validity of 32-bit and 64-bit arithmetic operations */
2287 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2288 {
2289 struct bpf_reg_state *regs = env->cur_state.regs;
2290 u8 opcode = BPF_OP(insn->code);
2291 int err;
2292
2293 if (opcode == BPF_END || opcode == BPF_NEG) {
2294 if (opcode == BPF_NEG) {
2295 if (BPF_SRC(insn->code) != 0 ||
2296 insn->src_reg != BPF_REG_0 ||
2297 insn->off != 0 || insn->imm != 0) {
2298 verbose("BPF_NEG uses reserved fields\n");
2299 return -EINVAL;
2300 }
2301 } else {
2302 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2303 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2304 BPF_CLASS(insn->code) == BPF_ALU64) {
2305 verbose("BPF_END uses reserved fields\n");
2306 return -EINVAL;
2307 }
2308 }
2309
2310 /* check src operand */
2311 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2312 if (err)
2313 return err;
2314
2315 if (is_pointer_value(env, insn->dst_reg)) {
2316 verbose("R%d pointer arithmetic prohibited\n",
2317 insn->dst_reg);
2318 return -EACCES;
2319 }
2320
2321 /* check dest operand */
2322 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2323 if (err)
2324 return err;
2325
2326 } else if (opcode == BPF_MOV) {
2327
2328 if (BPF_SRC(insn->code) == BPF_X) {
2329 if (insn->imm != 0 || insn->off != 0) {
2330 verbose("BPF_MOV uses reserved fields\n");
2331 return -EINVAL;
2332 }
2333
2334 /* check src operand */
2335 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2336 if (err)
2337 return err;
2338 } else {
2339 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2340 verbose("BPF_MOV uses reserved fields\n");
2341 return -EINVAL;
2342 }
2343 }
2344
2345 /* check dest operand */
2346 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2347 if (err)
2348 return err;
2349
2350 if (BPF_SRC(insn->code) == BPF_X) {
2351 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2352 /* case: R1 = R2
2353 * copy register state to dest reg
2354 */
2355 regs[insn->dst_reg] = regs[insn->src_reg];
2356 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
2357 } else {
2358 /* R1 = (u32) R2 */
2359 if (is_pointer_value(env, insn->src_reg)) {
2360 verbose("R%d partial copy of pointer\n",
2361 insn->src_reg);
2362 return -EACCES;
2363 }
2364 mark_reg_unknown(regs, insn->dst_reg);
2365 /* high 32 bits are known zero. */
2366 regs[insn->dst_reg].var_off = tnum_cast(
2367 regs[insn->dst_reg].var_off, 4);
2368 __update_reg_bounds(&regs[insn->dst_reg]);
2369 }
2370 } else {
2371 /* case: R = imm
2372 * remember the value we stored into this reg
2373 */
2374 regs[insn->dst_reg].type = SCALAR_VALUE;
2375 __mark_reg_known(regs + insn->dst_reg, insn->imm);
2376 }
2377
2378 } else if (opcode > BPF_END) {
2379 verbose("invalid BPF_ALU opcode %x\n", opcode);
2380 return -EINVAL;
2381
2382 } else { /* all other ALU ops: and, sub, xor, add, ... */
2383
2384 if (BPF_SRC(insn->code) == BPF_X) {
2385 if (insn->imm != 0 || insn->off != 0) {
2386 verbose("BPF_ALU uses reserved fields\n");
2387 return -EINVAL;
2388 }
2389 /* check src1 operand */
2390 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2391 if (err)
2392 return err;
2393 } else {
2394 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2395 verbose("BPF_ALU uses reserved fields\n");
2396 return -EINVAL;
2397 }
2398 }
2399
2400 /* check src2 operand */
2401 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2402 if (err)
2403 return err;
2404
2405 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2406 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2407 verbose("div by zero\n");
2408 return -EINVAL;
2409 }
2410
2411 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2412 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2413 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2414
2415 if (insn->imm < 0 || insn->imm >= size) {
2416 verbose("invalid shift %d\n", insn->imm);
2417 return -EINVAL;
2418 }
2419 }
2420
2421 /* check dest operand */
2422 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2423 if (err)
2424 return err;
2425
2426 return adjust_reg_min_max_vals(env, insn);
2427 }
2428
2429 return 0;
2430 }
2431
2432 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2433 struct bpf_reg_state *dst_reg,
2434 bool range_right_open)
2435 {
2436 struct bpf_reg_state *regs = state->regs, *reg;
2437 u16 new_range;
2438 int i;
2439
2440 if (dst_reg->off < 0 ||
2441 (dst_reg->off == 0 && range_right_open))
2442 /* This doesn't give us any range */
2443 return;
2444
2445 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2446 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2447 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2448 * than pkt_end, but that's because it's also less than pkt.
2449 */
2450 return;
2451
2452 new_range = dst_reg->off;
2453 if (range_right_open)
2454 new_range--;
2455
2456 /* Examples for register markings:
2457 *
2458 * pkt_data in dst register:
2459 *
2460 * r2 = r3;
2461 * r2 += 8;
2462 * if (r2 > pkt_end) goto <handle exception>
2463 * <access okay>
2464 *
2465 * r2 = r3;
2466 * r2 += 8;
2467 * if (r2 < pkt_end) goto <access okay>
2468 * <handle exception>
2469 *
2470 * Where:
2471 * r2 == dst_reg, pkt_end == src_reg
2472 * r2=pkt(id=n,off=8,r=0)
2473 * r3=pkt(id=n,off=0,r=0)
2474 *
2475 * pkt_data in src register:
2476 *
2477 * r2 = r3;
2478 * r2 += 8;
2479 * if (pkt_end >= r2) goto <access okay>
2480 * <handle exception>
2481 *
2482 * r2 = r3;
2483 * r2 += 8;
2484 * if (pkt_end <= r2) goto <handle exception>
2485 * <access okay>
2486 *
2487 * Where:
2488 * pkt_end == dst_reg, r2 == src_reg
2489 * r2=pkt(id=n,off=8,r=0)
2490 * r3=pkt(id=n,off=0,r=0)
2491 *
2492 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2493 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2494 * and [r3, r3 + 8-1) respectively is safe to access depending on
2495 * the check.
2496 */
2497
2498 /* If our ids match, then we must have the same max_value. And we
2499 * don't care about the other reg's fixed offset, since if it's too big
2500 * the range won't allow anything.
2501 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2502 */
2503 for (i = 0; i < MAX_BPF_REG; i++)
2504 if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
2505 /* keep the maximum range already checked */
2506 regs[i].range = max(regs[i].range, new_range);
2507
2508 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2509 if (state->stack_slot_type[i] != STACK_SPILL)
2510 continue;
2511 reg = &state->spilled_regs[i / BPF_REG_SIZE];
2512 if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
2513 reg->range = max(reg->range, new_range);
2514 }
2515 }
2516
2517 /* Adjusts the register min/max values in the case that the dst_reg is the
2518 * variable register that we are working on, and src_reg is a constant or we're
2519 * simply doing a BPF_K check.
2520 * In JEQ/JNE cases we also adjust the var_off values.
2521 */
2522 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2523 struct bpf_reg_state *false_reg, u64 val,
2524 u8 opcode)
2525 {
2526 /* If the dst_reg is a pointer, we can't learn anything about its
2527 * variable offset from the compare (unless src_reg were a pointer into
2528 * the same object, but we don't bother with that.
2529 * Since false_reg and true_reg have the same type by construction, we
2530 * only need to check one of them for pointerness.
2531 */
2532 if (__is_pointer_value(false, false_reg))
2533 return;
2534
2535 switch (opcode) {
2536 case BPF_JEQ:
2537 /* If this is false then we know nothing Jon Snow, but if it is
2538 * true then we know for sure.
2539 */
2540 __mark_reg_known(true_reg, val);
2541 break;
2542 case BPF_JNE:
2543 /* If this is true we know nothing Jon Snow, but if it is false
2544 * we know the value for sure;
2545 */
2546 __mark_reg_known(false_reg, val);
2547 break;
2548 case BPF_JGT:
2549 false_reg->umax_value = min(false_reg->umax_value, val);
2550 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2551 break;
2552 case BPF_JSGT:
2553 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2554 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2555 break;
2556 case BPF_JLT:
2557 false_reg->umin_value = max(false_reg->umin_value, val);
2558 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2559 break;
2560 case BPF_JSLT:
2561 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2562 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2563 break;
2564 case BPF_JGE:
2565 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2566 true_reg->umin_value = max(true_reg->umin_value, val);
2567 break;
2568 case BPF_JSGE:
2569 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2570 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2571 break;
2572 case BPF_JLE:
2573 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2574 true_reg->umax_value = min(true_reg->umax_value, val);
2575 break;
2576 case BPF_JSLE:
2577 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2578 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2579 break;
2580 default:
2581 break;
2582 }
2583
2584 __reg_deduce_bounds(false_reg);
2585 __reg_deduce_bounds(true_reg);
2586 /* We might have learned some bits from the bounds. */
2587 __reg_bound_offset(false_reg);
2588 __reg_bound_offset(true_reg);
2589 /* Intersecting with the old var_off might have improved our bounds
2590 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2591 * then new var_off is (0; 0x7f...fc) which improves our umax.
2592 */
2593 __update_reg_bounds(false_reg);
2594 __update_reg_bounds(true_reg);
2595 }
2596
2597 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2598 * the variable reg.
2599 */
2600 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2601 struct bpf_reg_state *false_reg, u64 val,
2602 u8 opcode)
2603 {
2604 if (__is_pointer_value(false, false_reg))
2605 return;
2606
2607 switch (opcode) {
2608 case BPF_JEQ:
2609 /* If this is false then we know nothing Jon Snow, but if it is
2610 * true then we know for sure.
2611 */
2612 __mark_reg_known(true_reg, val);
2613 break;
2614 case BPF_JNE:
2615 /* If this is true we know nothing Jon Snow, but if it is false
2616 * we know the value for sure;
2617 */
2618 __mark_reg_known(false_reg, val);
2619 break;
2620 case BPF_JGT:
2621 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2622 false_reg->umin_value = max(false_reg->umin_value, val);
2623 break;
2624 case BPF_JSGT:
2625 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2626 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2627 break;
2628 case BPF_JLT:
2629 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2630 false_reg->umax_value = min(false_reg->umax_value, val);
2631 break;
2632 case BPF_JSLT:
2633 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2634 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2635 break;
2636 case BPF_JGE:
2637 true_reg->umax_value = min(true_reg->umax_value, val);
2638 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2639 break;
2640 case BPF_JSGE:
2641 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2642 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2643 break;
2644 case BPF_JLE:
2645 true_reg->umin_value = max(true_reg->umin_value, val);
2646 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2647 break;
2648 case BPF_JSLE:
2649 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2650 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2651 break;
2652 default:
2653 break;
2654 }
2655
2656 __reg_deduce_bounds(false_reg);
2657 __reg_deduce_bounds(true_reg);
2658 /* We might have learned some bits from the bounds. */
2659 __reg_bound_offset(false_reg);
2660 __reg_bound_offset(true_reg);
2661 /* Intersecting with the old var_off might have improved our bounds
2662 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2663 * then new var_off is (0; 0x7f...fc) which improves our umax.
2664 */
2665 __update_reg_bounds(false_reg);
2666 __update_reg_bounds(true_reg);
2667 }
2668
2669 /* Regs are known to be equal, so intersect their min/max/var_off */
2670 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2671 struct bpf_reg_state *dst_reg)
2672 {
2673 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2674 dst_reg->umin_value);
2675 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2676 dst_reg->umax_value);
2677 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2678 dst_reg->smin_value);
2679 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2680 dst_reg->smax_value);
2681 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2682 dst_reg->var_off);
2683 /* We might have learned new bounds from the var_off. */
2684 __update_reg_bounds(src_reg);
2685 __update_reg_bounds(dst_reg);
2686 /* We might have learned something about the sign bit. */
2687 __reg_deduce_bounds(src_reg);
2688 __reg_deduce_bounds(dst_reg);
2689 /* We might have learned some bits from the bounds. */
2690 __reg_bound_offset(src_reg);
2691 __reg_bound_offset(dst_reg);
2692 /* Intersecting with the old var_off might have improved our bounds
2693 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2694 * then new var_off is (0; 0x7f...fc) which improves our umax.
2695 */
2696 __update_reg_bounds(src_reg);
2697 __update_reg_bounds(dst_reg);
2698 }
2699
2700 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2701 struct bpf_reg_state *true_dst,
2702 struct bpf_reg_state *false_src,
2703 struct bpf_reg_state *false_dst,
2704 u8 opcode)
2705 {
2706 switch (opcode) {
2707 case BPF_JEQ:
2708 __reg_combine_min_max(true_src, true_dst);
2709 break;
2710 case BPF_JNE:
2711 __reg_combine_min_max(false_src, false_dst);
2712 break;
2713 }
2714 }
2715
2716 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2717 bool is_null)
2718 {
2719 struct bpf_reg_state *reg = &regs[regno];
2720
2721 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2722 /* Old offset (both fixed and variable parts) should
2723 * have been known-zero, because we don't allow pointer
2724 * arithmetic on pointers that might be NULL.
2725 */
2726 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2727 !tnum_equals_const(reg->var_off, 0) ||
2728 reg->off)) {
2729 __mark_reg_known_zero(reg);
2730 reg->off = 0;
2731 }
2732 if (is_null) {
2733 reg->type = SCALAR_VALUE;
2734 } else if (reg->map_ptr->inner_map_meta) {
2735 reg->type = CONST_PTR_TO_MAP;
2736 reg->map_ptr = reg->map_ptr->inner_map_meta;
2737 } else {
2738 reg->type = PTR_TO_MAP_VALUE;
2739 }
2740 /* We don't need id from this point onwards anymore, thus we
2741 * should better reset it, so that state pruning has chances
2742 * to take effect.
2743 */
2744 reg->id = 0;
2745 }
2746 }
2747
2748 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2749 * be folded together at some point.
2750 */
2751 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2752 bool is_null)
2753 {
2754 struct bpf_reg_state *regs = state->regs;
2755 u32 id = regs[regno].id;
2756 int i;
2757
2758 for (i = 0; i < MAX_BPF_REG; i++)
2759 mark_map_reg(regs, i, id, is_null);
2760
2761 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2762 if (state->stack_slot_type[i] != STACK_SPILL)
2763 continue;
2764 mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
2765 }
2766 }
2767
2768 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2769 struct bpf_insn *insn, int *insn_idx)
2770 {
2771 struct bpf_verifier_state *other_branch, *this_branch = &env->cur_state;
2772 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2773 u8 opcode = BPF_OP(insn->code);
2774 int err;
2775
2776 if (opcode > BPF_JSLE) {
2777 verbose("invalid BPF_JMP opcode %x\n", opcode);
2778 return -EINVAL;
2779 }
2780
2781 if (BPF_SRC(insn->code) == BPF_X) {
2782 if (insn->imm != 0) {
2783 verbose("BPF_JMP uses reserved fields\n");
2784 return -EINVAL;
2785 }
2786
2787 /* check src1 operand */
2788 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2789 if (err)
2790 return err;
2791
2792 if (is_pointer_value(env, insn->src_reg)) {
2793 verbose("R%d pointer comparison prohibited\n",
2794 insn->src_reg);
2795 return -EACCES;
2796 }
2797 } else {
2798 if (insn->src_reg != BPF_REG_0) {
2799 verbose("BPF_JMP uses reserved fields\n");
2800 return -EINVAL;
2801 }
2802 }
2803
2804 /* check src2 operand */
2805 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2806 if (err)
2807 return err;
2808
2809 dst_reg = &regs[insn->dst_reg];
2810
2811 /* detect if R == 0 where R was initialized to zero earlier */
2812 if (BPF_SRC(insn->code) == BPF_K &&
2813 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2814 dst_reg->type == SCALAR_VALUE &&
2815 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2816 if (opcode == BPF_JEQ) {
2817 /* if (imm == imm) goto pc+off;
2818 * only follow the goto, ignore fall-through
2819 */
2820 *insn_idx += insn->off;
2821 return 0;
2822 } else {
2823 /* if (imm != imm) goto pc+off;
2824 * only follow fall-through branch, since
2825 * that's where the program will go
2826 */
2827 return 0;
2828 }
2829 }
2830
2831 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2832 if (!other_branch)
2833 return -EFAULT;
2834
2835 /* detect if we are comparing against a constant value so we can adjust
2836 * our min/max values for our dst register.
2837 * this is only legit if both are scalars (or pointers to the same
2838 * object, I suppose, but we don't support that right now), because
2839 * otherwise the different base pointers mean the offsets aren't
2840 * comparable.
2841 */
2842 if (BPF_SRC(insn->code) == BPF_X) {
2843 if (dst_reg->type == SCALAR_VALUE &&
2844 regs[insn->src_reg].type == SCALAR_VALUE) {
2845 if (tnum_is_const(regs[insn->src_reg].var_off))
2846 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2847 dst_reg, regs[insn->src_reg].var_off.value,
2848 opcode);
2849 else if (tnum_is_const(dst_reg->var_off))
2850 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
2851 &regs[insn->src_reg],
2852 dst_reg->var_off.value, opcode);
2853 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
2854 /* Comparing for equality, we can combine knowledge */
2855 reg_combine_min_max(&other_branch->regs[insn->src_reg],
2856 &other_branch->regs[insn->dst_reg],
2857 &regs[insn->src_reg],
2858 &regs[insn->dst_reg], opcode);
2859 }
2860 } else if (dst_reg->type == SCALAR_VALUE) {
2861 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2862 dst_reg, insn->imm, opcode);
2863 }
2864
2865 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2866 if (BPF_SRC(insn->code) == BPF_K &&
2867 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2868 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2869 /* Mark all identical map registers in each branch as either
2870 * safe or unknown depending R == 0 or R != 0 conditional.
2871 */
2872 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
2873 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
2874 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2875 dst_reg->type == PTR_TO_PACKET &&
2876 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2877 /* pkt_data' > pkt_end */
2878 find_good_pkt_pointers(this_branch, dst_reg, false);
2879 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2880 dst_reg->type == PTR_TO_PACKET_END &&
2881 regs[insn->src_reg].type == PTR_TO_PACKET) {
2882 /* pkt_end > pkt_data' */
2883 find_good_pkt_pointers(other_branch, &regs[insn->src_reg], true);
2884 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2885 dst_reg->type == PTR_TO_PACKET &&
2886 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2887 /* pkt_data' < pkt_end */
2888 find_good_pkt_pointers(other_branch, dst_reg, true);
2889 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2890 dst_reg->type == PTR_TO_PACKET_END &&
2891 regs[insn->src_reg].type == PTR_TO_PACKET) {
2892 /* pkt_end < pkt_data' */
2893 find_good_pkt_pointers(this_branch, &regs[insn->src_reg], false);
2894 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2895 dst_reg->type == PTR_TO_PACKET &&
2896 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2897 /* pkt_data' >= pkt_end */
2898 find_good_pkt_pointers(this_branch, dst_reg, true);
2899 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2900 dst_reg->type == PTR_TO_PACKET_END &&
2901 regs[insn->src_reg].type == PTR_TO_PACKET) {
2902 /* pkt_end >= pkt_data' */
2903 find_good_pkt_pointers(other_branch, &regs[insn->src_reg], false);
2904 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2905 dst_reg->type == PTR_TO_PACKET &&
2906 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2907 /* pkt_data' <= pkt_end */
2908 find_good_pkt_pointers(other_branch, dst_reg, false);
2909 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2910 dst_reg->type == PTR_TO_PACKET_END &&
2911 regs[insn->src_reg].type == PTR_TO_PACKET) {
2912 /* pkt_end <= pkt_data' */
2913 find_good_pkt_pointers(this_branch, &regs[insn->src_reg], true);
2914 } else if (is_pointer_value(env, insn->dst_reg)) {
2915 verbose("R%d pointer comparison prohibited\n", insn->dst_reg);
2916 return -EACCES;
2917 }
2918 if (log_level)
2919 print_verifier_state(this_branch);
2920 return 0;
2921 }
2922
2923 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
2924 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
2925 {
2926 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
2927
2928 return (struct bpf_map *) (unsigned long) imm64;
2929 }
2930
2931 /* verify BPF_LD_IMM64 instruction */
2932 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
2933 {
2934 struct bpf_reg_state *regs = env->cur_state.regs;
2935 int err;
2936
2937 if (BPF_SIZE(insn->code) != BPF_DW) {
2938 verbose("invalid BPF_LD_IMM insn\n");
2939 return -EINVAL;
2940 }
2941 if (insn->off != 0) {
2942 verbose("BPF_LD_IMM64 uses reserved fields\n");
2943 return -EINVAL;
2944 }
2945
2946 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2947 if (err)
2948 return err;
2949
2950 if (insn->src_reg == 0) {
2951 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
2952
2953 regs[insn->dst_reg].type = SCALAR_VALUE;
2954 __mark_reg_known(&regs[insn->dst_reg], imm);
2955 return 0;
2956 }
2957
2958 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
2959 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
2960
2961 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
2962 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
2963 return 0;
2964 }
2965
2966 static bool may_access_skb(enum bpf_prog_type type)
2967 {
2968 switch (type) {
2969 case BPF_PROG_TYPE_SOCKET_FILTER:
2970 case BPF_PROG_TYPE_SCHED_CLS:
2971 case BPF_PROG_TYPE_SCHED_ACT:
2972 return true;
2973 default:
2974 return false;
2975 }
2976 }
2977
2978 /* verify safety of LD_ABS|LD_IND instructions:
2979 * - they can only appear in the programs where ctx == skb
2980 * - since they are wrappers of function calls, they scratch R1-R5 registers,
2981 * preserve R6-R9, and store return value into R0
2982 *
2983 * Implicit input:
2984 * ctx == skb == R6 == CTX
2985 *
2986 * Explicit input:
2987 * SRC == any register
2988 * IMM == 32-bit immediate
2989 *
2990 * Output:
2991 * R0 - 8/16/32-bit skb data converted to cpu endianness
2992 */
2993 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
2994 {
2995 struct bpf_reg_state *regs = env->cur_state.regs;
2996 u8 mode = BPF_MODE(insn->code);
2997 int i, err;
2998
2999 if (!may_access_skb(env->prog->type)) {
3000 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3001 return -EINVAL;
3002 }
3003
3004 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3005 BPF_SIZE(insn->code) == BPF_DW ||
3006 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3007 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
3008 return -EINVAL;
3009 }
3010
3011 /* check whether implicit source operand (register R6) is readable */
3012 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3013 if (err)
3014 return err;
3015
3016 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3017 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3018 return -EINVAL;
3019 }
3020
3021 if (mode == BPF_IND) {
3022 /* check explicit source operand */
3023 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3024 if (err)
3025 return err;
3026 }
3027
3028 /* reset caller saved regs to unreadable */
3029 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3030 mark_reg_not_init(regs, caller_saved[i]);
3031 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3032 }
3033
3034 /* mark destination R0 register as readable, since it contains
3035 * the value fetched from the packet.
3036 * Already marked as written above.
3037 */
3038 mark_reg_unknown(regs, BPF_REG_0);
3039 return 0;
3040 }
3041
3042 /* non-recursive DFS pseudo code
3043 * 1 procedure DFS-iterative(G,v):
3044 * 2 label v as discovered
3045 * 3 let S be a stack
3046 * 4 S.push(v)
3047 * 5 while S is not empty
3048 * 6 t <- S.pop()
3049 * 7 if t is what we're looking for:
3050 * 8 return t
3051 * 9 for all edges e in G.adjacentEdges(t) do
3052 * 10 if edge e is already labelled
3053 * 11 continue with the next edge
3054 * 12 w <- G.adjacentVertex(t,e)
3055 * 13 if vertex w is not discovered and not explored
3056 * 14 label e as tree-edge
3057 * 15 label w as discovered
3058 * 16 S.push(w)
3059 * 17 continue at 5
3060 * 18 else if vertex w is discovered
3061 * 19 label e as back-edge
3062 * 20 else
3063 * 21 // vertex w is explored
3064 * 22 label e as forward- or cross-edge
3065 * 23 label t as explored
3066 * 24 S.pop()
3067 *
3068 * convention:
3069 * 0x10 - discovered
3070 * 0x11 - discovered and fall-through edge labelled
3071 * 0x12 - discovered and fall-through and branch edges labelled
3072 * 0x20 - explored
3073 */
3074
3075 enum {
3076 DISCOVERED = 0x10,
3077 EXPLORED = 0x20,
3078 FALLTHROUGH = 1,
3079 BRANCH = 2,
3080 };
3081
3082 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3083
3084 static int *insn_stack; /* stack of insns to process */
3085 static int cur_stack; /* current stack index */
3086 static int *insn_state;
3087
3088 /* t, w, e - match pseudo-code above:
3089 * t - index of current instruction
3090 * w - next instruction
3091 * e - edge
3092 */
3093 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3094 {
3095 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3096 return 0;
3097
3098 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3099 return 0;
3100
3101 if (w < 0 || w >= env->prog->len) {
3102 verbose("jump out of range from insn %d to %d\n", t, w);
3103 return -EINVAL;
3104 }
3105
3106 if (e == BRANCH)
3107 /* mark branch target for state pruning */
3108 env->explored_states[w] = STATE_LIST_MARK;
3109
3110 if (insn_state[w] == 0) {
3111 /* tree-edge */
3112 insn_state[t] = DISCOVERED | e;
3113 insn_state[w] = DISCOVERED;
3114 if (cur_stack >= env->prog->len)
3115 return -E2BIG;
3116 insn_stack[cur_stack++] = w;
3117 return 1;
3118 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3119 verbose("back-edge from insn %d to %d\n", t, w);
3120 return -EINVAL;
3121 } else if (insn_state[w] == EXPLORED) {
3122 /* forward- or cross-edge */
3123 insn_state[t] = DISCOVERED | e;
3124 } else {
3125 verbose("insn state internal bug\n");
3126 return -EFAULT;
3127 }
3128 return 0;
3129 }
3130
3131 /* non-recursive depth-first-search to detect loops in BPF program
3132 * loop == back-edge in directed graph
3133 */
3134 static int check_cfg(struct bpf_verifier_env *env)
3135 {
3136 struct bpf_insn *insns = env->prog->insnsi;
3137 int insn_cnt = env->prog->len;
3138 int ret = 0;
3139 int i, t;
3140
3141 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3142 if (!insn_state)
3143 return -ENOMEM;
3144
3145 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3146 if (!insn_stack) {
3147 kfree(insn_state);
3148 return -ENOMEM;
3149 }
3150
3151 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3152 insn_stack[0] = 0; /* 0 is the first instruction */
3153 cur_stack = 1;
3154
3155 peek_stack:
3156 if (cur_stack == 0)
3157 goto check_state;
3158 t = insn_stack[cur_stack - 1];
3159
3160 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3161 u8 opcode = BPF_OP(insns[t].code);
3162
3163 if (opcode == BPF_EXIT) {
3164 goto mark_explored;
3165 } else if (opcode == BPF_CALL) {
3166 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3167 if (ret == 1)
3168 goto peek_stack;
3169 else if (ret < 0)
3170 goto err_free;
3171 if (t + 1 < insn_cnt)
3172 env->explored_states[t + 1] = STATE_LIST_MARK;
3173 } else if (opcode == BPF_JA) {
3174 if (BPF_SRC(insns[t].code) != BPF_K) {
3175 ret = -EINVAL;
3176 goto err_free;
3177 }
3178 /* unconditional jump with single edge */
3179 ret = push_insn(t, t + insns[t].off + 1,
3180 FALLTHROUGH, env);
3181 if (ret == 1)
3182 goto peek_stack;
3183 else if (ret < 0)
3184 goto err_free;
3185 /* tell verifier to check for equivalent states
3186 * after every call and jump
3187 */
3188 if (t + 1 < insn_cnt)
3189 env->explored_states[t + 1] = STATE_LIST_MARK;
3190 } else {
3191 /* conditional jump with two edges */
3192 env->explored_states[t] = STATE_LIST_MARK;
3193 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3194 if (ret == 1)
3195 goto peek_stack;
3196 else if (ret < 0)
3197 goto err_free;
3198
3199 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3200 if (ret == 1)
3201 goto peek_stack;
3202 else if (ret < 0)
3203 goto err_free;
3204 }
3205 } else {
3206 /* all other non-branch instructions with single
3207 * fall-through edge
3208 */
3209 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3210 if (ret == 1)
3211 goto peek_stack;
3212 else if (ret < 0)
3213 goto err_free;
3214 }
3215
3216 mark_explored:
3217 insn_state[t] = EXPLORED;
3218 if (cur_stack-- <= 0) {
3219 verbose("pop stack internal bug\n");
3220 ret = -EFAULT;
3221 goto err_free;
3222 }
3223 goto peek_stack;
3224
3225 check_state:
3226 for (i = 0; i < insn_cnt; i++) {
3227 if (insn_state[i] != EXPLORED) {
3228 verbose("unreachable insn %d\n", i);
3229 ret = -EINVAL;
3230 goto err_free;
3231 }
3232 }
3233 ret = 0; /* cfg looks good */
3234
3235 err_free:
3236 kfree(insn_state);
3237 kfree(insn_stack);
3238 return ret;
3239 }
3240
3241 /* check %cur's range satisfies %old's */
3242 static bool range_within(struct bpf_reg_state *old,
3243 struct bpf_reg_state *cur)
3244 {
3245 return old->umin_value <= cur->umin_value &&
3246 old->umax_value >= cur->umax_value &&
3247 old->smin_value <= cur->smin_value &&
3248 old->smax_value >= cur->smax_value;
3249 }
3250
3251 /* Maximum number of register states that can exist at once */
3252 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3253 struct idpair {
3254 u32 old;
3255 u32 cur;
3256 };
3257
3258 /* If in the old state two registers had the same id, then they need to have
3259 * the same id in the new state as well. But that id could be different from
3260 * the old state, so we need to track the mapping from old to new ids.
3261 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3262 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3263 * regs with a different old id could still have new id 9, we don't care about
3264 * that.
3265 * So we look through our idmap to see if this old id has been seen before. If
3266 * so, we require the new id to match; otherwise, we add the id pair to the map.
3267 */
3268 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3269 {
3270 unsigned int i;
3271
3272 for (i = 0; i < ID_MAP_SIZE; i++) {
3273 if (!idmap[i].old) {
3274 /* Reached an empty slot; haven't seen this id before */
3275 idmap[i].old = old_id;
3276 idmap[i].cur = cur_id;
3277 return true;
3278 }
3279 if (idmap[i].old == old_id)
3280 return idmap[i].cur == cur_id;
3281 }
3282 /* We ran out of idmap slots, which should be impossible */
3283 WARN_ON_ONCE(1);
3284 return false;
3285 }
3286
3287 /* Returns true if (rold safe implies rcur safe) */
3288 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3289 struct idpair *idmap)
3290 {
3291 if (!(rold->live & REG_LIVE_READ))
3292 /* explored state didn't use this */
3293 return true;
3294
3295 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3296 return true;
3297
3298 if (rold->type == NOT_INIT)
3299 /* explored state can't have used this */
3300 return true;
3301 if (rcur->type == NOT_INIT)
3302 return false;
3303 switch (rold->type) {
3304 case SCALAR_VALUE:
3305 if (rcur->type == SCALAR_VALUE) {
3306 /* new val must satisfy old val knowledge */
3307 return range_within(rold, rcur) &&
3308 tnum_in(rold->var_off, rcur->var_off);
3309 } else {
3310 /* if we knew anything about the old value, we're not
3311 * equal, because we can't know anything about the
3312 * scalar value of the pointer in the new value.
3313 */
3314 return rold->umin_value == 0 &&
3315 rold->umax_value == U64_MAX &&
3316 rold->smin_value == S64_MIN &&
3317 rold->smax_value == S64_MAX &&
3318 tnum_is_unknown(rold->var_off);
3319 }
3320 case PTR_TO_MAP_VALUE:
3321 /* If the new min/max/var_off satisfy the old ones and
3322 * everything else matches, we are OK.
3323 * We don't care about the 'id' value, because nothing
3324 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3325 */
3326 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3327 range_within(rold, rcur) &&
3328 tnum_in(rold->var_off, rcur->var_off);
3329 case PTR_TO_MAP_VALUE_OR_NULL:
3330 /* a PTR_TO_MAP_VALUE could be safe to use as a
3331 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3332 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3333 * checked, doing so could have affected others with the same
3334 * id, and we can't check for that because we lost the id when
3335 * we converted to a PTR_TO_MAP_VALUE.
3336 */
3337 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3338 return false;
3339 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3340 return false;
3341 /* Check our ids match any regs they're supposed to */
3342 return check_ids(rold->id, rcur->id, idmap);
3343 case PTR_TO_PACKET:
3344 if (rcur->type != PTR_TO_PACKET)
3345 return false;
3346 /* We must have at least as much range as the old ptr
3347 * did, so that any accesses which were safe before are
3348 * still safe. This is true even if old range < old off,
3349 * since someone could have accessed through (ptr - k), or
3350 * even done ptr -= k in a register, to get a safe access.
3351 */
3352 if (rold->range > rcur->range)
3353 return false;
3354 /* If the offsets don't match, we can't trust our alignment;
3355 * nor can we be sure that we won't fall out of range.
3356 */
3357 if (rold->off != rcur->off)
3358 return false;
3359 /* id relations must be preserved */
3360 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3361 return false;
3362 /* new val must satisfy old val knowledge */
3363 return range_within(rold, rcur) &&
3364 tnum_in(rold->var_off, rcur->var_off);
3365 case PTR_TO_CTX:
3366 case CONST_PTR_TO_MAP:
3367 case PTR_TO_STACK:
3368 case PTR_TO_PACKET_END:
3369 /* Only valid matches are exact, which memcmp() above
3370 * would have accepted
3371 */
3372 default:
3373 /* Don't know what's going on, just say it's not safe */
3374 return false;
3375 }
3376
3377 /* Shouldn't get here; if we do, say it's not safe */
3378 WARN_ON_ONCE(1);
3379 return false;
3380 }
3381
3382 /* compare two verifier states
3383 *
3384 * all states stored in state_list are known to be valid, since
3385 * verifier reached 'bpf_exit' instruction through them
3386 *
3387 * this function is called when verifier exploring different branches of
3388 * execution popped from the state stack. If it sees an old state that has
3389 * more strict register state and more strict stack state then this execution
3390 * branch doesn't need to be explored further, since verifier already
3391 * concluded that more strict state leads to valid finish.
3392 *
3393 * Therefore two states are equivalent if register state is more conservative
3394 * and explored stack state is more conservative than the current one.
3395 * Example:
3396 * explored current
3397 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3398 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3399 *
3400 * In other words if current stack state (one being explored) has more
3401 * valid slots than old one that already passed validation, it means
3402 * the verifier can stop exploring and conclude that current state is valid too
3403 *
3404 * Similarly with registers. If explored state has register type as invalid
3405 * whereas register type in current state is meaningful, it means that
3406 * the current state will reach 'bpf_exit' instruction safely
3407 */
3408 static bool states_equal(struct bpf_verifier_env *env,
3409 struct bpf_verifier_state *old,
3410 struct bpf_verifier_state *cur)
3411 {
3412 struct idpair *idmap;
3413 bool ret = false;
3414 int i;
3415
3416 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3417 /* If we failed to allocate the idmap, just say it's not safe */
3418 if (!idmap)
3419 return false;
3420
3421 for (i = 0; i < MAX_BPF_REG; i++) {
3422 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3423 goto out_free;
3424 }
3425
3426 for (i = 0; i < MAX_BPF_STACK; i++) {
3427 if (old->stack_slot_type[i] == STACK_INVALID)
3428 continue;
3429 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
3430 /* Ex: old explored (safe) state has STACK_SPILL in
3431 * this stack slot, but current has has STACK_MISC ->
3432 * this verifier states are not equivalent,
3433 * return false to continue verification of this path
3434 */
3435 goto out_free;
3436 if (i % BPF_REG_SIZE)
3437 continue;
3438 if (old->stack_slot_type[i] != STACK_SPILL)
3439 continue;
3440 if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
3441 &cur->spilled_regs[i / BPF_REG_SIZE],
3442 idmap))
3443 /* when explored and current stack slot are both storing
3444 * spilled registers, check that stored pointers types
3445 * are the same as well.
3446 * Ex: explored safe path could have stored
3447 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3448 * but current path has stored:
3449 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3450 * such verifier states are not equivalent.
3451 * return false to continue verification of this path
3452 */
3453 goto out_free;
3454 else
3455 continue;
3456 }
3457 ret = true;
3458 out_free:
3459 kfree(idmap);
3460 return ret;
3461 }
3462
3463 /* A write screens off any subsequent reads; but write marks come from the
3464 * straight-line code between a state and its parent. When we arrive at a
3465 * jump target (in the first iteration of the propagate_liveness() loop),
3466 * we didn't arrive by the straight-line code, so read marks in state must
3467 * propagate to parent regardless of state's write marks.
3468 */
3469 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3470 struct bpf_verifier_state *parent)
3471 {
3472 bool writes = parent == state->parent; /* Observe write marks */
3473 bool touched = false; /* any changes made? */
3474 int i;
3475
3476 if (!parent)
3477 return touched;
3478 /* Propagate read liveness of registers... */
3479 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3480 /* We don't need to worry about FP liveness because it's read-only */
3481 for (i = 0; i < BPF_REG_FP; i++) {
3482 if (parent->regs[i].live & REG_LIVE_READ)
3483 continue;
3484 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3485 continue;
3486 if (state->regs[i].live & REG_LIVE_READ) {
3487 parent->regs[i].live |= REG_LIVE_READ;
3488 touched = true;
3489 }
3490 }
3491 /* ... and stack slots */
3492 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++) {
3493 if (parent->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3494 continue;
3495 if (state->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3496 continue;
3497 if (parent->spilled_regs[i].live & REG_LIVE_READ)
3498 continue;
3499 if (writes && (state->spilled_regs[i].live & REG_LIVE_WRITTEN))
3500 continue;
3501 if (state->spilled_regs[i].live & REG_LIVE_READ) {
3502 parent->spilled_regs[i].live |= REG_LIVE_READ;
3503 touched = true;
3504 }
3505 }
3506 return touched;
3507 }
3508
3509 /* "parent" is "a state from which we reach the current state", but initially
3510 * it is not the state->parent (i.e. "the state whose straight-line code leads
3511 * to the current state"), instead it is the state that happened to arrive at
3512 * a (prunable) equivalent of the current state. See comment above
3513 * do_propagate_liveness() for consequences of this.
3514 * This function is just a more efficient way of calling mark_reg_read() or
3515 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3516 * though it requires that parent != state->parent in the call arguments.
3517 */
3518 static void propagate_liveness(const struct bpf_verifier_state *state,
3519 struct bpf_verifier_state *parent)
3520 {
3521 while (do_propagate_liveness(state, parent)) {
3522 /* Something changed, so we need to feed those changes onward */
3523 state = parent;
3524 parent = state->parent;
3525 }
3526 }
3527
3528 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3529 {
3530 struct bpf_verifier_state_list *new_sl;
3531 struct bpf_verifier_state_list *sl;
3532 int i;
3533
3534 sl = env->explored_states[insn_idx];
3535 if (!sl)
3536 /* this 'insn_idx' instruction wasn't marked, so we will not
3537 * be doing state search here
3538 */
3539 return 0;
3540
3541 while (sl != STATE_LIST_MARK) {
3542 if (states_equal(env, &sl->state, &env->cur_state)) {
3543 /* reached equivalent register/stack state,
3544 * prune the search.
3545 * Registers read by the continuation are read by us.
3546 * If we have any write marks in env->cur_state, they
3547 * will prevent corresponding reads in the continuation
3548 * from reaching our parent (an explored_state). Our
3549 * own state will get the read marks recorded, but
3550 * they'll be immediately forgotten as we're pruning
3551 * this state and will pop a new one.
3552 */
3553 propagate_liveness(&sl->state, &env->cur_state);
3554 return 1;
3555 }
3556 sl = sl->next;
3557 }
3558
3559 /* there were no equivalent states, remember current one.
3560 * technically the current state is not proven to be safe yet,
3561 * but it will either reach bpf_exit (which means it's safe) or
3562 * it will be rejected. Since there are no loops, we won't be
3563 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3564 */
3565 new_sl = kmalloc(sizeof(struct bpf_verifier_state_list), GFP_USER);
3566 if (!new_sl)
3567 return -ENOMEM;
3568
3569 /* add new state to the head of linked list */
3570 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
3571 new_sl->next = env->explored_states[insn_idx];
3572 env->explored_states[insn_idx] = new_sl;
3573 /* connect new state to parentage chain */
3574 env->cur_state.parent = &new_sl->state;
3575 /* clear write marks in current state: the writes we did are not writes
3576 * our child did, so they don't screen off its reads from us.
3577 * (There are no read marks in current state, because reads always mark
3578 * their parent and current state never has children yet. Only
3579 * explored_states can get read marks.)
3580 */
3581 for (i = 0; i < BPF_REG_FP; i++)
3582 env->cur_state.regs[i].live = REG_LIVE_NONE;
3583 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++)
3584 if (env->cur_state.stack_slot_type[i * BPF_REG_SIZE] == STACK_SPILL)
3585 env->cur_state.spilled_regs[i].live = REG_LIVE_NONE;
3586 return 0;
3587 }
3588
3589 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3590 int insn_idx, int prev_insn_idx)
3591 {
3592 if (!env->analyzer_ops || !env->analyzer_ops->insn_hook)
3593 return 0;
3594
3595 return env->analyzer_ops->insn_hook(env, insn_idx, prev_insn_idx);
3596 }
3597
3598 static int do_check(struct bpf_verifier_env *env)
3599 {
3600 struct bpf_verifier_state *state = &env->cur_state;
3601 struct bpf_insn *insns = env->prog->insnsi;
3602 struct bpf_reg_state *regs = state->regs;
3603 int insn_cnt = env->prog->len;
3604 int insn_idx, prev_insn_idx = 0;
3605 int insn_processed = 0;
3606 bool do_print_state = false;
3607
3608 init_reg_state(regs);
3609 state->parent = NULL;
3610 insn_idx = 0;
3611 for (;;) {
3612 struct bpf_insn *insn;
3613 u8 class;
3614 int err;
3615
3616 if (insn_idx >= insn_cnt) {
3617 verbose("invalid insn idx %d insn_cnt %d\n",
3618 insn_idx, insn_cnt);
3619 return -EFAULT;
3620 }
3621
3622 insn = &insns[insn_idx];
3623 class = BPF_CLASS(insn->code);
3624
3625 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3626 verbose("BPF program is too large. Processed %d insn\n",
3627 insn_processed);
3628 return -E2BIG;
3629 }
3630
3631 err = is_state_visited(env, insn_idx);
3632 if (err < 0)
3633 return err;
3634 if (err == 1) {
3635 /* found equivalent state, can prune the search */
3636 if (log_level) {
3637 if (do_print_state)
3638 verbose("\nfrom %d to %d: safe\n",
3639 prev_insn_idx, insn_idx);
3640 else
3641 verbose("%d: safe\n", insn_idx);
3642 }
3643 goto process_bpf_exit;
3644 }
3645
3646 if (need_resched())
3647 cond_resched();
3648
3649 if (log_level > 1 || (log_level && do_print_state)) {
3650 if (log_level > 1)
3651 verbose("%d:", insn_idx);
3652 else
3653 verbose("\nfrom %d to %d:",
3654 prev_insn_idx, insn_idx);
3655 print_verifier_state(&env->cur_state);
3656 do_print_state = false;
3657 }
3658
3659 if (log_level) {
3660 verbose("%d: ", insn_idx);
3661 print_bpf_insn(env, insn);
3662 }
3663
3664 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3665 if (err)
3666 return err;
3667
3668 if (class == BPF_ALU || class == BPF_ALU64) {
3669 err = check_alu_op(env, insn);
3670 if (err)
3671 return err;
3672
3673 } else if (class == BPF_LDX) {
3674 enum bpf_reg_type *prev_src_type, src_reg_type;
3675
3676 /* check for reserved fields is already done */
3677
3678 /* check src operand */
3679 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3680 if (err)
3681 return err;
3682
3683 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3684 if (err)
3685 return err;
3686
3687 src_reg_type = regs[insn->src_reg].type;
3688
3689 /* check that memory (src_reg + off) is readable,
3690 * the state of dst_reg will be updated by this func
3691 */
3692 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3693 BPF_SIZE(insn->code), BPF_READ,
3694 insn->dst_reg);
3695 if (err)
3696 return err;
3697
3698 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3699
3700 if (*prev_src_type == NOT_INIT) {
3701 /* saw a valid insn
3702 * dst_reg = *(u32 *)(src_reg + off)
3703 * save type to validate intersecting paths
3704 */
3705 *prev_src_type = src_reg_type;
3706
3707 } else if (src_reg_type != *prev_src_type &&
3708 (src_reg_type == PTR_TO_CTX ||
3709 *prev_src_type == PTR_TO_CTX)) {
3710 /* ABuser program is trying to use the same insn
3711 * dst_reg = *(u32*) (src_reg + off)
3712 * with different pointer types:
3713 * src_reg == ctx in one branch and
3714 * src_reg == stack|map in some other branch.
3715 * Reject it.
3716 */
3717 verbose("same insn cannot be used with different pointers\n");
3718 return -EINVAL;
3719 }
3720
3721 } else if (class == BPF_STX) {
3722 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3723
3724 if (BPF_MODE(insn->code) == BPF_XADD) {
3725 err = check_xadd(env, insn_idx, insn);
3726 if (err)
3727 return err;
3728 insn_idx++;
3729 continue;
3730 }
3731
3732 /* check src1 operand */
3733 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3734 if (err)
3735 return err;
3736 /* check src2 operand */
3737 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3738 if (err)
3739 return err;
3740
3741 dst_reg_type = regs[insn->dst_reg].type;
3742
3743 /* check that memory (dst_reg + off) is writeable */
3744 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3745 BPF_SIZE(insn->code), BPF_WRITE,
3746 insn->src_reg);
3747 if (err)
3748 return err;
3749
3750 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3751
3752 if (*prev_dst_type == NOT_INIT) {
3753 *prev_dst_type = dst_reg_type;
3754 } else if (dst_reg_type != *prev_dst_type &&
3755 (dst_reg_type == PTR_TO_CTX ||
3756 *prev_dst_type == PTR_TO_CTX)) {
3757 verbose("same insn cannot be used with different pointers\n");
3758 return -EINVAL;
3759 }
3760
3761 } else if (class == BPF_ST) {
3762 if (BPF_MODE(insn->code) != BPF_MEM ||
3763 insn->src_reg != BPF_REG_0) {
3764 verbose("BPF_ST uses reserved fields\n");
3765 return -EINVAL;
3766 }
3767 /* check src operand */
3768 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3769 if (err)
3770 return err;
3771
3772 /* check that memory (dst_reg + off) is writeable */
3773 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3774 BPF_SIZE(insn->code), BPF_WRITE,
3775 -1);
3776 if (err)
3777 return err;
3778
3779 } else if (class == BPF_JMP) {
3780 u8 opcode = BPF_OP(insn->code);
3781
3782 if (opcode == BPF_CALL) {
3783 if (BPF_SRC(insn->code) != BPF_K ||
3784 insn->off != 0 ||
3785 insn->src_reg != BPF_REG_0 ||
3786 insn->dst_reg != BPF_REG_0) {
3787 verbose("BPF_CALL uses reserved fields\n");
3788 return -EINVAL;
3789 }
3790
3791 err = check_call(env, insn->imm, insn_idx);
3792 if (err)
3793 return err;
3794
3795 } else if (opcode == BPF_JA) {
3796 if (BPF_SRC(insn->code) != BPF_K ||
3797 insn->imm != 0 ||
3798 insn->src_reg != BPF_REG_0 ||
3799 insn->dst_reg != BPF_REG_0) {
3800 verbose("BPF_JA uses reserved fields\n");
3801 return -EINVAL;
3802 }
3803
3804 insn_idx += insn->off + 1;
3805 continue;
3806
3807 } else if (opcode == BPF_EXIT) {
3808 if (BPF_SRC(insn->code) != BPF_K ||
3809 insn->imm != 0 ||
3810 insn->src_reg != BPF_REG_0 ||
3811 insn->dst_reg != BPF_REG_0) {
3812 verbose("BPF_EXIT uses reserved fields\n");
3813 return -EINVAL;
3814 }
3815
3816 /* eBPF calling convetion is such that R0 is used
3817 * to return the value from eBPF program.
3818 * Make sure that it's readable at this time
3819 * of bpf_exit, which means that program wrote
3820 * something into it earlier
3821 */
3822 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
3823 if (err)
3824 return err;
3825
3826 if (is_pointer_value(env, BPF_REG_0)) {
3827 verbose("R0 leaks addr as return value\n");
3828 return -EACCES;
3829 }
3830
3831 process_bpf_exit:
3832 insn_idx = pop_stack(env, &prev_insn_idx);
3833 if (insn_idx < 0) {
3834 break;
3835 } else {
3836 do_print_state = true;
3837 continue;
3838 }
3839 } else {
3840 err = check_cond_jmp_op(env, insn, &insn_idx);
3841 if (err)
3842 return err;
3843 }
3844 } else if (class == BPF_LD) {
3845 u8 mode = BPF_MODE(insn->code);
3846
3847 if (mode == BPF_ABS || mode == BPF_IND) {
3848 err = check_ld_abs(env, insn);
3849 if (err)
3850 return err;
3851
3852 } else if (mode == BPF_IMM) {
3853 err = check_ld_imm(env, insn);
3854 if (err)
3855 return err;
3856
3857 insn_idx++;
3858 } else {
3859 verbose("invalid BPF_LD mode\n");
3860 return -EINVAL;
3861 }
3862 } else {
3863 verbose("unknown insn class %d\n", class);
3864 return -EINVAL;
3865 }
3866
3867 insn_idx++;
3868 }
3869
3870 verbose("processed %d insns, stack depth %d\n",
3871 insn_processed, env->prog->aux->stack_depth);
3872 return 0;
3873 }
3874
3875 static int check_map_prealloc(struct bpf_map *map)
3876 {
3877 return (map->map_type != BPF_MAP_TYPE_HASH &&
3878 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
3879 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
3880 !(map->map_flags & BPF_F_NO_PREALLOC);
3881 }
3882
3883 static int check_map_prog_compatibility(struct bpf_map *map,
3884 struct bpf_prog *prog)
3885
3886 {
3887 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3888 * preallocated hash maps, since doing memory allocation
3889 * in overflow_handler can crash depending on where nmi got
3890 * triggered.
3891 */
3892 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
3893 if (!check_map_prealloc(map)) {
3894 verbose("perf_event programs can only use preallocated hash map\n");
3895 return -EINVAL;
3896 }
3897 if (map->inner_map_meta &&
3898 !check_map_prealloc(map->inner_map_meta)) {
3899 verbose("perf_event programs can only use preallocated inner hash map\n");
3900 return -EINVAL;
3901 }
3902 }
3903 return 0;
3904 }
3905
3906 /* look for pseudo eBPF instructions that access map FDs and
3907 * replace them with actual map pointers
3908 */
3909 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
3910 {
3911 struct bpf_insn *insn = env->prog->insnsi;
3912 int insn_cnt = env->prog->len;
3913 int i, j, err;
3914
3915 err = bpf_prog_calc_tag(env->prog);
3916 if (err)
3917 return err;
3918
3919 for (i = 0; i < insn_cnt; i++, insn++) {
3920 if (BPF_CLASS(insn->code) == BPF_LDX &&
3921 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
3922 verbose("BPF_LDX uses reserved fields\n");
3923 return -EINVAL;
3924 }
3925
3926 if (BPF_CLASS(insn->code) == BPF_STX &&
3927 ((BPF_MODE(insn->code) != BPF_MEM &&
3928 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
3929 verbose("BPF_STX uses reserved fields\n");
3930 return -EINVAL;
3931 }
3932
3933 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
3934 struct bpf_map *map;
3935 struct fd f;
3936
3937 if (i == insn_cnt - 1 || insn[1].code != 0 ||
3938 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
3939 insn[1].off != 0) {
3940 verbose("invalid bpf_ld_imm64 insn\n");
3941 return -EINVAL;
3942 }
3943
3944 if (insn->src_reg == 0)
3945 /* valid generic load 64-bit imm */
3946 goto next_insn;
3947
3948 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
3949 verbose("unrecognized bpf_ld_imm64 insn\n");
3950 return -EINVAL;
3951 }
3952
3953 f = fdget(insn->imm);
3954 map = __bpf_map_get(f);
3955 if (IS_ERR(map)) {
3956 verbose("fd %d is not pointing to valid bpf_map\n",
3957 insn->imm);
3958 return PTR_ERR(map);
3959 }
3960
3961 err = check_map_prog_compatibility(map, env->prog);
3962 if (err) {
3963 fdput(f);
3964 return err;
3965 }
3966
3967 /* store map pointer inside BPF_LD_IMM64 instruction */
3968 insn[0].imm = (u32) (unsigned long) map;
3969 insn[1].imm = ((u64) (unsigned long) map) >> 32;
3970
3971 /* check whether we recorded this map already */
3972 for (j = 0; j < env->used_map_cnt; j++)
3973 if (env->used_maps[j] == map) {
3974 fdput(f);
3975 goto next_insn;
3976 }
3977
3978 if (env->used_map_cnt >= MAX_USED_MAPS) {
3979 fdput(f);
3980 return -E2BIG;
3981 }
3982
3983 /* hold the map. If the program is rejected by verifier,
3984 * the map will be released by release_maps() or it
3985 * will be used by the valid program until it's unloaded
3986 * and all maps are released in free_bpf_prog_info()
3987 */
3988 map = bpf_map_inc(map, false);
3989 if (IS_ERR(map)) {
3990 fdput(f);
3991 return PTR_ERR(map);
3992 }
3993 env->used_maps[env->used_map_cnt++] = map;
3994
3995 fdput(f);
3996 next_insn:
3997 insn++;
3998 i++;
3999 }
4000 }
4001
4002 /* now all pseudo BPF_LD_IMM64 instructions load valid
4003 * 'struct bpf_map *' into a register instead of user map_fd.
4004 * These pointers will be used later by verifier to validate map access.
4005 */
4006 return 0;
4007 }
4008
4009 /* drop refcnt of maps used by the rejected program */
4010 static void release_maps(struct bpf_verifier_env *env)
4011 {
4012 int i;
4013
4014 for (i = 0; i < env->used_map_cnt; i++)
4015 bpf_map_put(env->used_maps[i]);
4016 }
4017
4018 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4019 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
4020 {
4021 struct bpf_insn *insn = env->prog->insnsi;
4022 int insn_cnt = env->prog->len;
4023 int i;
4024
4025 for (i = 0; i < insn_cnt; i++, insn++)
4026 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
4027 insn->src_reg = 0;
4028 }
4029
4030 /* single env->prog->insni[off] instruction was replaced with the range
4031 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4032 * [0, off) and [off, end) to new locations, so the patched range stays zero
4033 */
4034 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
4035 u32 off, u32 cnt)
4036 {
4037 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
4038
4039 if (cnt == 1)
4040 return 0;
4041 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4042 if (!new_data)
4043 return -ENOMEM;
4044 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4045 memcpy(new_data + off + cnt - 1, old_data + off,
4046 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4047 env->insn_aux_data = new_data;
4048 vfree(old_data);
4049 return 0;
4050 }
4051
4052 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4053 const struct bpf_insn *patch, u32 len)
4054 {
4055 struct bpf_prog *new_prog;
4056
4057 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4058 if (!new_prog)
4059 return NULL;
4060 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4061 return NULL;
4062 return new_prog;
4063 }
4064
4065 /* convert load instructions that access fields of 'struct __sk_buff'
4066 * into sequence of instructions that access fields of 'struct sk_buff'
4067 */
4068 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4069 {
4070 const struct bpf_verifier_ops *ops = env->prog->aux->ops;
4071 int i, cnt, size, ctx_field_size, delta = 0;
4072 const int insn_cnt = env->prog->len;
4073 struct bpf_insn insn_buf[16], *insn;
4074 struct bpf_prog *new_prog;
4075 enum bpf_access_type type;
4076 bool is_narrower_load;
4077 u32 target_size;
4078
4079 if (ops->gen_prologue) {
4080 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4081 env->prog);
4082 if (cnt >= ARRAY_SIZE(insn_buf)) {
4083 verbose("bpf verifier is misconfigured\n");
4084 return -EINVAL;
4085 } else if (cnt) {
4086 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4087 if (!new_prog)
4088 return -ENOMEM;
4089
4090 env->prog = new_prog;
4091 delta += cnt - 1;
4092 }
4093 }
4094
4095 if (!ops->convert_ctx_access)
4096 return 0;
4097
4098 insn = env->prog->insnsi + delta;
4099
4100 for (i = 0; i < insn_cnt; i++, insn++) {
4101 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4102 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4103 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4104 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4105 type = BPF_READ;
4106 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4107 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4108 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4109 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4110 type = BPF_WRITE;
4111 else
4112 continue;
4113
4114 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4115 continue;
4116
4117 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4118 size = BPF_LDST_BYTES(insn);
4119
4120 /* If the read access is a narrower load of the field,
4121 * convert to a 4/8-byte load, to minimum program type specific
4122 * convert_ctx_access changes. If conversion is successful,
4123 * we will apply proper mask to the result.
4124 */
4125 is_narrower_load = size < ctx_field_size;
4126 if (is_narrower_load) {
4127 u32 off = insn->off;
4128 u8 size_code;
4129
4130 if (type == BPF_WRITE) {
4131 verbose("bpf verifier narrow ctx access misconfigured\n");
4132 return -EINVAL;
4133 }
4134
4135 size_code = BPF_H;
4136 if (ctx_field_size == 4)
4137 size_code = BPF_W;
4138 else if (ctx_field_size == 8)
4139 size_code = BPF_DW;
4140
4141 insn->off = off & ~(ctx_field_size - 1);
4142 insn->code = BPF_LDX | BPF_MEM | size_code;
4143 }
4144
4145 target_size = 0;
4146 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4147 &target_size);
4148 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4149 (ctx_field_size && !target_size)) {
4150 verbose("bpf verifier is misconfigured\n");
4151 return -EINVAL;
4152 }
4153
4154 if (is_narrower_load && size < target_size) {
4155 if (ctx_field_size <= 4)
4156 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4157 (1 << size * 8) - 1);
4158 else
4159 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4160 (1 << size * 8) - 1);
4161 }
4162
4163 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4164 if (!new_prog)
4165 return -ENOMEM;
4166
4167 delta += cnt - 1;
4168
4169 /* keep walking new program and skip insns we just inserted */
4170 env->prog = new_prog;
4171 insn = new_prog->insnsi + i + delta;
4172 }
4173
4174 return 0;
4175 }
4176
4177 /* fixup insn->imm field of bpf_call instructions
4178 * and inline eligible helpers as explicit sequence of BPF instructions
4179 *
4180 * this function is called after eBPF program passed verification
4181 */
4182 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4183 {
4184 struct bpf_prog *prog = env->prog;
4185 struct bpf_insn *insn = prog->insnsi;
4186 const struct bpf_func_proto *fn;
4187 const int insn_cnt = prog->len;
4188 struct bpf_insn insn_buf[16];
4189 struct bpf_prog *new_prog;
4190 struct bpf_map *map_ptr;
4191 int i, cnt, delta = 0;
4192
4193 for (i = 0; i < insn_cnt; i++, insn++) {
4194 if (insn->code != (BPF_JMP | BPF_CALL))
4195 continue;
4196
4197 if (insn->imm == BPF_FUNC_get_route_realm)
4198 prog->dst_needed = 1;
4199 if (insn->imm == BPF_FUNC_get_prandom_u32)
4200 bpf_user_rnd_init_once();
4201 if (insn->imm == BPF_FUNC_tail_call) {
4202 /* If we tail call into other programs, we
4203 * cannot make any assumptions since they can
4204 * be replaced dynamically during runtime in
4205 * the program array.
4206 */
4207 prog->cb_access = 1;
4208 env->prog->aux->stack_depth = MAX_BPF_STACK;
4209
4210 /* mark bpf_tail_call as different opcode to avoid
4211 * conditional branch in the interpeter for every normal
4212 * call and to prevent accidental JITing by JIT compiler
4213 * that doesn't support bpf_tail_call yet
4214 */
4215 insn->imm = 0;
4216 insn->code = BPF_JMP | BPF_TAIL_CALL;
4217 continue;
4218 }
4219
4220 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4221 * handlers are currently limited to 64 bit only.
4222 */
4223 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4224 insn->imm == BPF_FUNC_map_lookup_elem) {
4225 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4226 if (map_ptr == BPF_MAP_PTR_POISON ||
4227 !map_ptr->ops->map_gen_lookup)
4228 goto patch_call_imm;
4229
4230 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4231 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4232 verbose("bpf verifier is misconfigured\n");
4233 return -EINVAL;
4234 }
4235
4236 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4237 cnt);
4238 if (!new_prog)
4239 return -ENOMEM;
4240
4241 delta += cnt - 1;
4242
4243 /* keep walking new program and skip insns we just inserted */
4244 env->prog = prog = new_prog;
4245 insn = new_prog->insnsi + i + delta;
4246 continue;
4247 }
4248
4249 if (insn->imm == BPF_FUNC_redirect_map) {
4250 /* Note, we cannot use prog directly as imm as subsequent
4251 * rewrites would still change the prog pointer. The only
4252 * stable address we can use is aux, which also works with
4253 * prog clones during blinding.
4254 */
4255 u64 addr = (unsigned long)prog->aux;
4256 struct bpf_insn r4_ld[] = {
4257 BPF_LD_IMM64(BPF_REG_4, addr),
4258 *insn,
4259 };
4260 cnt = ARRAY_SIZE(r4_ld);
4261
4262 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4263 if (!new_prog)
4264 return -ENOMEM;
4265
4266 delta += cnt - 1;
4267 env->prog = prog = new_prog;
4268 insn = new_prog->insnsi + i + delta;
4269 }
4270 patch_call_imm:
4271 fn = prog->aux->ops->get_func_proto(insn->imm);
4272 /* all functions that have prototype and verifier allowed
4273 * programs to call them, must be real in-kernel functions
4274 */
4275 if (!fn->func) {
4276 verbose("kernel subsystem misconfigured func %s#%d\n",
4277 func_id_name(insn->imm), insn->imm);
4278 return -EFAULT;
4279 }
4280 insn->imm = fn->func - __bpf_call_base;
4281 }
4282
4283 return 0;
4284 }
4285
4286 static void free_states(struct bpf_verifier_env *env)
4287 {
4288 struct bpf_verifier_state_list *sl, *sln;
4289 int i;
4290
4291 if (!env->explored_states)
4292 return;
4293
4294 for (i = 0; i < env->prog->len; i++) {
4295 sl = env->explored_states[i];
4296
4297 if (sl)
4298 while (sl != STATE_LIST_MARK) {
4299 sln = sl->next;
4300 kfree(sl);
4301 sl = sln;
4302 }
4303 }
4304
4305 kfree(env->explored_states);
4306 }
4307
4308 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4309 {
4310 char __user *log_ubuf = NULL;
4311 struct bpf_verifier_env *env;
4312 int ret = -EINVAL;
4313
4314 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4315 * allocate/free it every time bpf_check() is called
4316 */
4317 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4318 if (!env)
4319 return -ENOMEM;
4320
4321 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4322 (*prog)->len);
4323 ret = -ENOMEM;
4324 if (!env->insn_aux_data)
4325 goto err_free_env;
4326 env->prog = *prog;
4327
4328 /* grab the mutex to protect few globals used by verifier */
4329 mutex_lock(&bpf_verifier_lock);
4330
4331 if (attr->log_level || attr->log_buf || attr->log_size) {
4332 /* user requested verbose verifier output
4333 * and supplied buffer to store the verification trace
4334 */
4335 log_level = attr->log_level;
4336 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
4337 log_size = attr->log_size;
4338 log_len = 0;
4339
4340 ret = -EINVAL;
4341 /* log_* values have to be sane */
4342 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
4343 log_level == 0 || log_ubuf == NULL)
4344 goto err_unlock;
4345
4346 ret = -ENOMEM;
4347 log_buf = vmalloc(log_size);
4348 if (!log_buf)
4349 goto err_unlock;
4350 } else {
4351 log_level = 0;
4352 }
4353
4354 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4355 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4356 env->strict_alignment = true;
4357
4358 ret = replace_map_fd_with_map_ptr(env);
4359 if (ret < 0)
4360 goto skip_full_check;
4361
4362 env->explored_states = kcalloc(env->prog->len,
4363 sizeof(struct bpf_verifier_state_list *),
4364 GFP_USER);
4365 ret = -ENOMEM;
4366 if (!env->explored_states)
4367 goto skip_full_check;
4368
4369 ret = check_cfg(env);
4370 if (ret < 0)
4371 goto skip_full_check;
4372
4373 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4374
4375 ret = do_check(env);
4376
4377 skip_full_check:
4378 while (pop_stack(env, NULL) >= 0);
4379 free_states(env);
4380
4381 if (ret == 0)
4382 /* program is valid, convert *(u32*)(ctx + off) accesses */
4383 ret = convert_ctx_accesses(env);
4384
4385 if (ret == 0)
4386 ret = fixup_bpf_calls(env);
4387
4388 if (log_level && log_len >= log_size - 1) {
4389 BUG_ON(log_len >= log_size);
4390 /* verifier log exceeded user supplied buffer */
4391 ret = -ENOSPC;
4392 /* fall through to return what was recorded */
4393 }
4394
4395 /* copy verifier log back to user space including trailing zero */
4396 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
4397 ret = -EFAULT;
4398 goto free_log_buf;
4399 }
4400
4401 if (ret == 0 && env->used_map_cnt) {
4402 /* if program passed verifier, update used_maps in bpf_prog_info */
4403 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4404 sizeof(env->used_maps[0]),
4405 GFP_KERNEL);
4406
4407 if (!env->prog->aux->used_maps) {
4408 ret = -ENOMEM;
4409 goto free_log_buf;
4410 }
4411
4412 memcpy(env->prog->aux->used_maps, env->used_maps,
4413 sizeof(env->used_maps[0]) * env->used_map_cnt);
4414 env->prog->aux->used_map_cnt = env->used_map_cnt;
4415
4416 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4417 * bpf_ld_imm64 instructions
4418 */
4419 convert_pseudo_ld_imm64(env);
4420 }
4421
4422 free_log_buf:
4423 if (log_level)
4424 vfree(log_buf);
4425 if (!env->prog->aux->used_maps)
4426 /* if we didn't copy map pointers into bpf_prog_info, release
4427 * them now. Otherwise free_bpf_prog_info() will release them.
4428 */
4429 release_maps(env);
4430 *prog = env->prog;
4431 err_unlock:
4432 mutex_unlock(&bpf_verifier_lock);
4433 vfree(env->insn_aux_data);
4434 err_free_env:
4435 kfree(env);
4436 return ret;
4437 }
4438
4439 int bpf_analyzer(struct bpf_prog *prog, const struct bpf_ext_analyzer_ops *ops,
4440 void *priv)
4441 {
4442 struct bpf_verifier_env *env;
4443 int ret;
4444
4445 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4446 if (!env)
4447 return -ENOMEM;
4448
4449 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4450 prog->len);
4451 ret = -ENOMEM;
4452 if (!env->insn_aux_data)
4453 goto err_free_env;
4454 env->prog = prog;
4455 env->analyzer_ops = ops;
4456 env->analyzer_priv = priv;
4457
4458 /* grab the mutex to protect few globals used by verifier */
4459 mutex_lock(&bpf_verifier_lock);
4460
4461 log_level = 0;
4462
4463 env->strict_alignment = false;
4464 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4465 env->strict_alignment = true;
4466
4467 env->explored_states = kcalloc(env->prog->len,
4468 sizeof(struct bpf_verifier_state_list *),
4469 GFP_KERNEL);
4470 ret = -ENOMEM;
4471 if (!env->explored_states)
4472 goto skip_full_check;
4473
4474 ret = check_cfg(env);
4475 if (ret < 0)
4476 goto skip_full_check;
4477
4478 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4479
4480 ret = do_check(env);
4481
4482 skip_full_check:
4483 while (pop_stack(env, NULL) >= 0);
4484 free_states(env);
4485
4486 mutex_unlock(&bpf_verifier_lock);
4487 vfree(env->insn_aux_data);
4488 err_free_env:
4489 kfree(env);
4490 return ret;
4491 }
4492 EXPORT_SYMBOL_GPL(bpf_analyzer);
Ubuntu Hirsute Linux kernel mirror