]> 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 tag 'trace-v4.14-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt...
[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 (!tnum_is_const(reg->var_off)) {
1120 char tn_buf[48];
1121
1122 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1123 verbose("variable ctx access var_off=%s off=%d size=%d",
1124 tn_buf, off, size);
1125 return -EACCES;
1126 }
1127 off += reg->var_off.value;
1128 err = check_ctx_access(env, insn_idx, off, size, t, &reg_type);
1129 if (!err && t == BPF_READ && value_regno >= 0) {
1130 /* ctx access returns either a scalar, or a
1131 * PTR_TO_PACKET[_END]. In the latter case, we know
1132 * the offset is zero.
1133 */
1134 if (reg_type == SCALAR_VALUE)
1135 mark_reg_unknown(state->regs, value_regno);
1136 else
1137 mark_reg_known_zero(state->regs, value_regno);
1138 state->regs[value_regno].id = 0;
1139 state->regs[value_regno].off = 0;
1140 state->regs[value_regno].range = 0;
1141 state->regs[value_regno].type = reg_type;
1142 }
1143
1144 } else if (reg->type == PTR_TO_STACK) {
1145 /* stack accesses must be at a fixed offset, so that we can
1146 * determine what type of data were returned.
1147 * See check_stack_read().
1148 */
1149 if (!tnum_is_const(reg->var_off)) {
1150 char tn_buf[48];
1151
1152 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1153 verbose("variable stack access var_off=%s off=%d size=%d",
1154 tn_buf, off, size);
1155 return -EACCES;
1156 }
1157 off += reg->var_off.value;
1158 if (off >= 0 || off < -MAX_BPF_STACK) {
1159 verbose("invalid stack off=%d size=%d\n", off, size);
1160 return -EACCES;
1161 }
1162
1163 if (env->prog->aux->stack_depth < -off)
1164 env->prog->aux->stack_depth = -off;
1165
1166 if (t == BPF_WRITE) {
1167 if (!env->allow_ptr_leaks &&
1168 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL &&
1169 size != BPF_REG_SIZE) {
1170 verbose("attempt to corrupt spilled pointer on stack\n");
1171 return -EACCES;
1172 }
1173 err = check_stack_write(state, off, size, value_regno);
1174 } else {
1175 err = check_stack_read(state, off, size, value_regno);
1176 }
1177 } else if (reg->type == PTR_TO_PACKET) {
1178 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1179 verbose("cannot write into packet\n");
1180 return -EACCES;
1181 }
1182 if (t == BPF_WRITE && value_regno >= 0 &&
1183 is_pointer_value(env, value_regno)) {
1184 verbose("R%d leaks addr into packet\n", value_regno);
1185 return -EACCES;
1186 }
1187 err = check_packet_access(env, regno, off, size);
1188 if (!err && t == BPF_READ && value_regno >= 0)
1189 mark_reg_unknown(state->regs, value_regno);
1190 } else {
1191 verbose("R%d invalid mem access '%s'\n",
1192 regno, reg_type_str[reg->type]);
1193 return -EACCES;
1194 }
1195
1196 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1197 state->regs[value_regno].type == SCALAR_VALUE) {
1198 /* b/h/w load zero-extends, mark upper bits as known 0 */
1199 state->regs[value_regno].var_off = tnum_cast(
1200 state->regs[value_regno].var_off, size);
1201 __update_reg_bounds(&state->regs[value_regno]);
1202 }
1203 return err;
1204 }
1205
1206 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1207 {
1208 int err;
1209
1210 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1211 insn->imm != 0) {
1212 verbose("BPF_XADD uses reserved fields\n");
1213 return -EINVAL;
1214 }
1215
1216 /* check src1 operand */
1217 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1218 if (err)
1219 return err;
1220
1221 /* check src2 operand */
1222 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1223 if (err)
1224 return err;
1225
1226 if (is_pointer_value(env, insn->src_reg)) {
1227 verbose("R%d leaks addr into mem\n", insn->src_reg);
1228 return -EACCES;
1229 }
1230
1231 /* check whether atomic_add can read the memory */
1232 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1233 BPF_SIZE(insn->code), BPF_READ, -1);
1234 if (err)
1235 return err;
1236
1237 /* check whether atomic_add can write into the same memory */
1238 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1239 BPF_SIZE(insn->code), BPF_WRITE, -1);
1240 }
1241
1242 /* Does this register contain a constant zero? */
1243 static bool register_is_null(struct bpf_reg_state reg)
1244 {
1245 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1246 }
1247
1248 /* when register 'regno' is passed into function that will read 'access_size'
1249 * bytes from that pointer, make sure that it's within stack boundary
1250 * and all elements of stack are initialized.
1251 * Unlike most pointer bounds-checking functions, this one doesn't take an
1252 * 'off' argument, so it has to add in reg->off itself.
1253 */
1254 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1255 int access_size, bool zero_size_allowed,
1256 struct bpf_call_arg_meta *meta)
1257 {
1258 struct bpf_verifier_state *state = &env->cur_state;
1259 struct bpf_reg_state *regs = state->regs;
1260 int off, i;
1261
1262 if (regs[regno].type != PTR_TO_STACK) {
1263 /* Allow zero-byte read from NULL, regardless of pointer type */
1264 if (zero_size_allowed && access_size == 0 &&
1265 register_is_null(regs[regno]))
1266 return 0;
1267
1268 verbose("R%d type=%s expected=%s\n", regno,
1269 reg_type_str[regs[regno].type],
1270 reg_type_str[PTR_TO_STACK]);
1271 return -EACCES;
1272 }
1273
1274 /* Only allow fixed-offset stack reads */
1275 if (!tnum_is_const(regs[regno].var_off)) {
1276 char tn_buf[48];
1277
1278 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1279 verbose("invalid variable stack read R%d var_off=%s\n",
1280 regno, tn_buf);
1281 }
1282 off = regs[regno].off + regs[regno].var_off.value;
1283 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1284 access_size <= 0) {
1285 verbose("invalid stack type R%d off=%d access_size=%d\n",
1286 regno, off, access_size);
1287 return -EACCES;
1288 }
1289
1290 if (env->prog->aux->stack_depth < -off)
1291 env->prog->aux->stack_depth = -off;
1292
1293 if (meta && meta->raw_mode) {
1294 meta->access_size = access_size;
1295 meta->regno = regno;
1296 return 0;
1297 }
1298
1299 for (i = 0; i < access_size; i++) {
1300 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
1301 verbose("invalid indirect read from stack off %d+%d size %d\n",
1302 off, i, access_size);
1303 return -EACCES;
1304 }
1305 }
1306 return 0;
1307 }
1308
1309 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1310 int access_size, bool zero_size_allowed,
1311 struct bpf_call_arg_meta *meta)
1312 {
1313 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1314
1315 switch (reg->type) {
1316 case PTR_TO_PACKET:
1317 return check_packet_access(env, regno, reg->off, access_size);
1318 case PTR_TO_MAP_VALUE:
1319 return check_map_access(env, regno, reg->off, access_size);
1320 default: /* scalar_value|ptr_to_stack or invalid ptr */
1321 return check_stack_boundary(env, regno, access_size,
1322 zero_size_allowed, meta);
1323 }
1324 }
1325
1326 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1327 enum bpf_arg_type arg_type,
1328 struct bpf_call_arg_meta *meta)
1329 {
1330 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1331 enum bpf_reg_type expected_type, type = reg->type;
1332 int err = 0;
1333
1334 if (arg_type == ARG_DONTCARE)
1335 return 0;
1336
1337 err = check_reg_arg(env, regno, SRC_OP);
1338 if (err)
1339 return err;
1340
1341 if (arg_type == ARG_ANYTHING) {
1342 if (is_pointer_value(env, regno)) {
1343 verbose("R%d leaks addr into helper function\n", regno);
1344 return -EACCES;
1345 }
1346 return 0;
1347 }
1348
1349 if (type == PTR_TO_PACKET &&
1350 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1351 verbose("helper access to the packet is not allowed\n");
1352 return -EACCES;
1353 }
1354
1355 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1356 arg_type == ARG_PTR_TO_MAP_VALUE) {
1357 expected_type = PTR_TO_STACK;
1358 if (type != PTR_TO_PACKET && type != expected_type)
1359 goto err_type;
1360 } else if (arg_type == ARG_CONST_SIZE ||
1361 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1362 expected_type = SCALAR_VALUE;
1363 if (type != expected_type)
1364 goto err_type;
1365 } else if (arg_type == ARG_CONST_MAP_PTR) {
1366 expected_type = CONST_PTR_TO_MAP;
1367 if (type != expected_type)
1368 goto err_type;
1369 } else if (arg_type == ARG_PTR_TO_CTX) {
1370 expected_type = PTR_TO_CTX;
1371 if (type != expected_type)
1372 goto err_type;
1373 } else if (arg_type == ARG_PTR_TO_MEM ||
1374 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1375 expected_type = PTR_TO_STACK;
1376 /* One exception here. In case function allows for NULL to be
1377 * passed in as argument, it's a SCALAR_VALUE type. Final test
1378 * happens during stack boundary checking.
1379 */
1380 if (register_is_null(*reg))
1381 /* final test in check_stack_boundary() */;
1382 else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
1383 type != expected_type)
1384 goto err_type;
1385 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1386 } else {
1387 verbose("unsupported arg_type %d\n", arg_type);
1388 return -EFAULT;
1389 }
1390
1391 if (arg_type == ARG_CONST_MAP_PTR) {
1392 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1393 meta->map_ptr = reg->map_ptr;
1394 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1395 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1396 * check that [key, key + map->key_size) are within
1397 * stack limits and initialized
1398 */
1399 if (!meta->map_ptr) {
1400 /* in function declaration map_ptr must come before
1401 * map_key, so that it's verified and known before
1402 * we have to check map_key here. Otherwise it means
1403 * that kernel subsystem misconfigured verifier
1404 */
1405 verbose("invalid map_ptr to access map->key\n");
1406 return -EACCES;
1407 }
1408 if (type == PTR_TO_PACKET)
1409 err = check_packet_access(env, regno, reg->off,
1410 meta->map_ptr->key_size);
1411 else
1412 err = check_stack_boundary(env, regno,
1413 meta->map_ptr->key_size,
1414 false, NULL);
1415 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1416 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1417 * check [value, value + map->value_size) validity
1418 */
1419 if (!meta->map_ptr) {
1420 /* kernel subsystem misconfigured verifier */
1421 verbose("invalid map_ptr to access map->value\n");
1422 return -EACCES;
1423 }
1424 if (type == PTR_TO_PACKET)
1425 err = check_packet_access(env, regno, reg->off,
1426 meta->map_ptr->value_size);
1427 else
1428 err = check_stack_boundary(env, regno,
1429 meta->map_ptr->value_size,
1430 false, NULL);
1431 } else if (arg_type == ARG_CONST_SIZE ||
1432 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1433 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1434
1435 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1436 * from stack pointer 'buf'. Check it
1437 * note: regno == len, regno - 1 == buf
1438 */
1439 if (regno == 0) {
1440 /* kernel subsystem misconfigured verifier */
1441 verbose("ARG_CONST_SIZE cannot be first argument\n");
1442 return -EACCES;
1443 }
1444
1445 /* The register is SCALAR_VALUE; the access check
1446 * happens using its boundaries.
1447 */
1448
1449 if (!tnum_is_const(reg->var_off))
1450 /* For unprivileged variable accesses, disable raw
1451 * mode so that the program is required to
1452 * initialize all the memory that the helper could
1453 * just partially fill up.
1454 */
1455 meta = NULL;
1456
1457 if (reg->smin_value < 0) {
1458 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1459 regno);
1460 return -EACCES;
1461 }
1462
1463 if (reg->umin_value == 0) {
1464 err = check_helper_mem_access(env, regno - 1, 0,
1465 zero_size_allowed,
1466 meta);
1467 if (err)
1468 return err;
1469 }
1470
1471 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1472 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1473 regno);
1474 return -EACCES;
1475 }
1476 err = check_helper_mem_access(env, regno - 1,
1477 reg->umax_value,
1478 zero_size_allowed, meta);
1479 }
1480
1481 return err;
1482 err_type:
1483 verbose("R%d type=%s expected=%s\n", regno,
1484 reg_type_str[type], reg_type_str[expected_type]);
1485 return -EACCES;
1486 }
1487
1488 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
1489 {
1490 if (!map)
1491 return 0;
1492
1493 /* We need a two way check, first is from map perspective ... */
1494 switch (map->map_type) {
1495 case BPF_MAP_TYPE_PROG_ARRAY:
1496 if (func_id != BPF_FUNC_tail_call)
1497 goto error;
1498 break;
1499 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1500 if (func_id != BPF_FUNC_perf_event_read &&
1501 func_id != BPF_FUNC_perf_event_output)
1502 goto error;
1503 break;
1504 case BPF_MAP_TYPE_STACK_TRACE:
1505 if (func_id != BPF_FUNC_get_stackid)
1506 goto error;
1507 break;
1508 case BPF_MAP_TYPE_CGROUP_ARRAY:
1509 if (func_id != BPF_FUNC_skb_under_cgroup &&
1510 func_id != BPF_FUNC_current_task_under_cgroup)
1511 goto error;
1512 break;
1513 /* devmap returns a pointer to a live net_device ifindex that we cannot
1514 * allow to be modified from bpf side. So do not allow lookup elements
1515 * for now.
1516 */
1517 case BPF_MAP_TYPE_DEVMAP:
1518 if (func_id != BPF_FUNC_redirect_map)
1519 goto error;
1520 break;
1521 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1522 case BPF_MAP_TYPE_HASH_OF_MAPS:
1523 if (func_id != BPF_FUNC_map_lookup_elem)
1524 goto error;
1525 break;
1526 case BPF_MAP_TYPE_SOCKMAP:
1527 if (func_id != BPF_FUNC_sk_redirect_map &&
1528 func_id != BPF_FUNC_sock_map_update &&
1529 func_id != BPF_FUNC_map_delete_elem)
1530 goto error;
1531 break;
1532 default:
1533 break;
1534 }
1535
1536 /* ... and second from the function itself. */
1537 switch (func_id) {
1538 case BPF_FUNC_tail_call:
1539 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1540 goto error;
1541 break;
1542 case BPF_FUNC_perf_event_read:
1543 case BPF_FUNC_perf_event_output:
1544 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1545 goto error;
1546 break;
1547 case BPF_FUNC_get_stackid:
1548 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1549 goto error;
1550 break;
1551 case BPF_FUNC_current_task_under_cgroup:
1552 case BPF_FUNC_skb_under_cgroup:
1553 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1554 goto error;
1555 break;
1556 case BPF_FUNC_redirect_map:
1557 if (map->map_type != BPF_MAP_TYPE_DEVMAP)
1558 goto error;
1559 break;
1560 case BPF_FUNC_sk_redirect_map:
1561 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1562 goto error;
1563 break;
1564 case BPF_FUNC_sock_map_update:
1565 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1566 goto error;
1567 break;
1568 default:
1569 break;
1570 }
1571
1572 return 0;
1573 error:
1574 verbose("cannot pass map_type %d into func %s#%d\n",
1575 map->map_type, func_id_name(func_id), func_id);
1576 return -EINVAL;
1577 }
1578
1579 static int check_raw_mode(const struct bpf_func_proto *fn)
1580 {
1581 int count = 0;
1582
1583 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1584 count++;
1585 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1586 count++;
1587 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1588 count++;
1589 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1590 count++;
1591 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1592 count++;
1593
1594 return count > 1 ? -EINVAL : 0;
1595 }
1596
1597 /* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
1598 * so turn them into unknown SCALAR_VALUE.
1599 */
1600 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1601 {
1602 struct bpf_verifier_state *state = &env->cur_state;
1603 struct bpf_reg_state *regs = state->regs, *reg;
1604 int i;
1605
1606 for (i = 0; i < MAX_BPF_REG; i++)
1607 if (regs[i].type == PTR_TO_PACKET ||
1608 regs[i].type == PTR_TO_PACKET_END)
1609 mark_reg_unknown(regs, i);
1610
1611 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
1612 if (state->stack_slot_type[i] != STACK_SPILL)
1613 continue;
1614 reg = &state->spilled_regs[i / BPF_REG_SIZE];
1615 if (reg->type != PTR_TO_PACKET &&
1616 reg->type != PTR_TO_PACKET_END)
1617 continue;
1618 __mark_reg_unknown(reg);
1619 }
1620 }
1621
1622 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1623 {
1624 struct bpf_verifier_state *state = &env->cur_state;
1625 const struct bpf_func_proto *fn = NULL;
1626 struct bpf_reg_state *regs = state->regs;
1627 struct bpf_call_arg_meta meta;
1628 bool changes_data;
1629 int i, err;
1630
1631 /* find function prototype */
1632 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1633 verbose("invalid func %s#%d\n", func_id_name(func_id), func_id);
1634 return -EINVAL;
1635 }
1636
1637 if (env->prog->aux->ops->get_func_proto)
1638 fn = env->prog->aux->ops->get_func_proto(func_id);
1639
1640 if (!fn) {
1641 verbose("unknown func %s#%d\n", func_id_name(func_id), func_id);
1642 return -EINVAL;
1643 }
1644
1645 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1646 if (!env->prog->gpl_compatible && fn->gpl_only) {
1647 verbose("cannot call GPL only function from proprietary program\n");
1648 return -EINVAL;
1649 }
1650
1651 changes_data = bpf_helper_changes_pkt_data(fn->func);
1652
1653 memset(&meta, 0, sizeof(meta));
1654 meta.pkt_access = fn->pkt_access;
1655
1656 /* We only support one arg being in raw mode at the moment, which
1657 * is sufficient for the helper functions we have right now.
1658 */
1659 err = check_raw_mode(fn);
1660 if (err) {
1661 verbose("kernel subsystem misconfigured func %s#%d\n",
1662 func_id_name(func_id), func_id);
1663 return err;
1664 }
1665
1666 /* check args */
1667 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1668 if (err)
1669 return err;
1670 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1671 if (err)
1672 return err;
1673 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1674 if (err)
1675 return err;
1676 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1677 if (err)
1678 return err;
1679 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1680 if (err)
1681 return err;
1682
1683 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1684 * is inferred from register state.
1685 */
1686 for (i = 0; i < meta.access_size; i++) {
1687 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1688 if (err)
1689 return err;
1690 }
1691
1692 /* reset caller saved regs */
1693 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1694 mark_reg_not_init(regs, caller_saved[i]);
1695 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1696 }
1697
1698 /* update return register (already marked as written above) */
1699 if (fn->ret_type == RET_INTEGER) {
1700 /* sets type to SCALAR_VALUE */
1701 mark_reg_unknown(regs, BPF_REG_0);
1702 } else if (fn->ret_type == RET_VOID) {
1703 regs[BPF_REG_0].type = NOT_INIT;
1704 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1705 struct bpf_insn_aux_data *insn_aux;
1706
1707 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1708 /* There is no offset yet applied, variable or fixed */
1709 mark_reg_known_zero(regs, BPF_REG_0);
1710 regs[BPF_REG_0].off = 0;
1711 /* remember map_ptr, so that check_map_access()
1712 * can check 'value_size' boundary of memory access
1713 * to map element returned from bpf_map_lookup_elem()
1714 */
1715 if (meta.map_ptr == NULL) {
1716 verbose("kernel subsystem misconfigured verifier\n");
1717 return -EINVAL;
1718 }
1719 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1720 regs[BPF_REG_0].id = ++env->id_gen;
1721 insn_aux = &env->insn_aux_data[insn_idx];
1722 if (!insn_aux->map_ptr)
1723 insn_aux->map_ptr = meta.map_ptr;
1724 else if (insn_aux->map_ptr != meta.map_ptr)
1725 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1726 } else {
1727 verbose("unknown return type %d of func %s#%d\n",
1728 fn->ret_type, func_id_name(func_id), func_id);
1729 return -EINVAL;
1730 }
1731
1732 err = check_map_func_compatibility(meta.map_ptr, func_id);
1733 if (err)
1734 return err;
1735
1736 if (changes_data)
1737 clear_all_pkt_pointers(env);
1738 return 0;
1739 }
1740
1741 static void coerce_reg_to_32(struct bpf_reg_state *reg)
1742 {
1743 /* clear high 32 bits */
1744 reg->var_off = tnum_cast(reg->var_off, 4);
1745 /* Update bounds */
1746 __update_reg_bounds(reg);
1747 }
1748
1749 static bool signed_add_overflows(s64 a, s64 b)
1750 {
1751 /* Do the add in u64, where overflow is well-defined */
1752 s64 res = (s64)((u64)a + (u64)b);
1753
1754 if (b < 0)
1755 return res > a;
1756 return res < a;
1757 }
1758
1759 static bool signed_sub_overflows(s64 a, s64 b)
1760 {
1761 /* Do the sub in u64, where overflow is well-defined */
1762 s64 res = (s64)((u64)a - (u64)b);
1763
1764 if (b < 0)
1765 return res < a;
1766 return res > a;
1767 }
1768
1769 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1770 * Caller should also handle BPF_MOV case separately.
1771 * If we return -EACCES, caller may want to try again treating pointer as a
1772 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1773 */
1774 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1775 struct bpf_insn *insn,
1776 const struct bpf_reg_state *ptr_reg,
1777 const struct bpf_reg_state *off_reg)
1778 {
1779 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
1780 bool known = tnum_is_const(off_reg->var_off);
1781 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1782 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1783 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1784 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1785 u8 opcode = BPF_OP(insn->code);
1786 u32 dst = insn->dst_reg;
1787
1788 dst_reg = &regs[dst];
1789
1790 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1791 print_verifier_state(&env->cur_state);
1792 verbose("verifier internal error: known but bad sbounds\n");
1793 return -EINVAL;
1794 }
1795 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1796 print_verifier_state(&env->cur_state);
1797 verbose("verifier internal error: known but bad ubounds\n");
1798 return -EINVAL;
1799 }
1800
1801 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1802 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1803 if (!env->allow_ptr_leaks)
1804 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1805 dst);
1806 return -EACCES;
1807 }
1808
1809 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1810 if (!env->allow_ptr_leaks)
1811 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1812 dst);
1813 return -EACCES;
1814 }
1815 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1816 if (!env->allow_ptr_leaks)
1817 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1818 dst);
1819 return -EACCES;
1820 }
1821 if (ptr_reg->type == PTR_TO_PACKET_END) {
1822 if (!env->allow_ptr_leaks)
1823 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1824 dst);
1825 return -EACCES;
1826 }
1827
1828 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1829 * The id may be overwritten later if we create a new variable offset.
1830 */
1831 dst_reg->type = ptr_reg->type;
1832 dst_reg->id = ptr_reg->id;
1833
1834 switch (opcode) {
1835 case BPF_ADD:
1836 /* We can take a fixed offset as long as it doesn't overflow
1837 * the s32 'off' field
1838 */
1839 if (known && (ptr_reg->off + smin_val ==
1840 (s64)(s32)(ptr_reg->off + smin_val))) {
1841 /* pointer += K. Accumulate it into fixed offset */
1842 dst_reg->smin_value = smin_ptr;
1843 dst_reg->smax_value = smax_ptr;
1844 dst_reg->umin_value = umin_ptr;
1845 dst_reg->umax_value = umax_ptr;
1846 dst_reg->var_off = ptr_reg->var_off;
1847 dst_reg->off = ptr_reg->off + smin_val;
1848 dst_reg->range = ptr_reg->range;
1849 break;
1850 }
1851 /* A new variable offset is created. Note that off_reg->off
1852 * == 0, since it's a scalar.
1853 * dst_reg gets the pointer type and since some positive
1854 * integer value was added to the pointer, give it a new 'id'
1855 * if it's a PTR_TO_PACKET.
1856 * this creates a new 'base' pointer, off_reg (variable) gets
1857 * added into the variable offset, and we copy the fixed offset
1858 * from ptr_reg.
1859 */
1860 if (signed_add_overflows(smin_ptr, smin_val) ||
1861 signed_add_overflows(smax_ptr, smax_val)) {
1862 dst_reg->smin_value = S64_MIN;
1863 dst_reg->smax_value = S64_MAX;
1864 } else {
1865 dst_reg->smin_value = smin_ptr + smin_val;
1866 dst_reg->smax_value = smax_ptr + smax_val;
1867 }
1868 if (umin_ptr + umin_val < umin_ptr ||
1869 umax_ptr + umax_val < umax_ptr) {
1870 dst_reg->umin_value = 0;
1871 dst_reg->umax_value = U64_MAX;
1872 } else {
1873 dst_reg->umin_value = umin_ptr + umin_val;
1874 dst_reg->umax_value = umax_ptr + umax_val;
1875 }
1876 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1877 dst_reg->off = ptr_reg->off;
1878 if (ptr_reg->type == PTR_TO_PACKET) {
1879 dst_reg->id = ++env->id_gen;
1880 /* something was added to pkt_ptr, set range to zero */
1881 dst_reg->range = 0;
1882 }
1883 break;
1884 case BPF_SUB:
1885 if (dst_reg == off_reg) {
1886 /* scalar -= pointer. Creates an unknown scalar */
1887 if (!env->allow_ptr_leaks)
1888 verbose("R%d tried to subtract pointer from scalar\n",
1889 dst);
1890 return -EACCES;
1891 }
1892 /* We don't allow subtraction from FP, because (according to
1893 * test_verifier.c test "invalid fp arithmetic", JITs might not
1894 * be able to deal with it.
1895 */
1896 if (ptr_reg->type == PTR_TO_STACK) {
1897 if (!env->allow_ptr_leaks)
1898 verbose("R%d subtraction from stack pointer prohibited\n",
1899 dst);
1900 return -EACCES;
1901 }
1902 if (known && (ptr_reg->off - smin_val ==
1903 (s64)(s32)(ptr_reg->off - smin_val))) {
1904 /* pointer -= K. Subtract it from fixed offset */
1905 dst_reg->smin_value = smin_ptr;
1906 dst_reg->smax_value = smax_ptr;
1907 dst_reg->umin_value = umin_ptr;
1908 dst_reg->umax_value = umax_ptr;
1909 dst_reg->var_off = ptr_reg->var_off;
1910 dst_reg->id = ptr_reg->id;
1911 dst_reg->off = ptr_reg->off - smin_val;
1912 dst_reg->range = ptr_reg->range;
1913 break;
1914 }
1915 /* A new variable offset is created. If the subtrahend is known
1916 * nonnegative, then any reg->range we had before is still good.
1917 */
1918 if (signed_sub_overflows(smin_ptr, smax_val) ||
1919 signed_sub_overflows(smax_ptr, smin_val)) {
1920 /* Overflow possible, we know nothing */
1921 dst_reg->smin_value = S64_MIN;
1922 dst_reg->smax_value = S64_MAX;
1923 } else {
1924 dst_reg->smin_value = smin_ptr - smax_val;
1925 dst_reg->smax_value = smax_ptr - smin_val;
1926 }
1927 if (umin_ptr < umax_val) {
1928 /* Overflow possible, we know nothing */
1929 dst_reg->umin_value = 0;
1930 dst_reg->umax_value = U64_MAX;
1931 } else {
1932 /* Cannot overflow (as long as bounds are consistent) */
1933 dst_reg->umin_value = umin_ptr - umax_val;
1934 dst_reg->umax_value = umax_ptr - umin_val;
1935 }
1936 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
1937 dst_reg->off = ptr_reg->off;
1938 if (ptr_reg->type == PTR_TO_PACKET) {
1939 dst_reg->id = ++env->id_gen;
1940 /* something was added to pkt_ptr, set range to zero */
1941 if (smin_val < 0)
1942 dst_reg->range = 0;
1943 }
1944 break;
1945 case BPF_AND:
1946 case BPF_OR:
1947 case BPF_XOR:
1948 /* bitwise ops on pointers are troublesome, prohibit for now.
1949 * (However, in principle we could allow some cases, e.g.
1950 * ptr &= ~3 which would reduce min_value by 3.)
1951 */
1952 if (!env->allow_ptr_leaks)
1953 verbose("R%d bitwise operator %s on pointer prohibited\n",
1954 dst, bpf_alu_string[opcode >> 4]);
1955 return -EACCES;
1956 default:
1957 /* other operators (e.g. MUL,LSH) produce non-pointer results */
1958 if (!env->allow_ptr_leaks)
1959 verbose("R%d pointer arithmetic with %s operator prohibited\n",
1960 dst, bpf_alu_string[opcode >> 4]);
1961 return -EACCES;
1962 }
1963
1964 __update_reg_bounds(dst_reg);
1965 __reg_deduce_bounds(dst_reg);
1966 __reg_bound_offset(dst_reg);
1967 return 0;
1968 }
1969
1970 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
1971 struct bpf_insn *insn,
1972 struct bpf_reg_state *dst_reg,
1973 struct bpf_reg_state src_reg)
1974 {
1975 struct bpf_reg_state *regs = env->cur_state.regs;
1976 u8 opcode = BPF_OP(insn->code);
1977 bool src_known, dst_known;
1978 s64 smin_val, smax_val;
1979 u64 umin_val, umax_val;
1980
1981 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1982 /* 32-bit ALU ops are (32,32)->64 */
1983 coerce_reg_to_32(dst_reg);
1984 coerce_reg_to_32(&src_reg);
1985 }
1986 smin_val = src_reg.smin_value;
1987 smax_val = src_reg.smax_value;
1988 umin_val = src_reg.umin_value;
1989 umax_val = src_reg.umax_value;
1990 src_known = tnum_is_const(src_reg.var_off);
1991 dst_known = tnum_is_const(dst_reg->var_off);
1992
1993 switch (opcode) {
1994 case BPF_ADD:
1995 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
1996 signed_add_overflows(dst_reg->smax_value, smax_val)) {
1997 dst_reg->smin_value = S64_MIN;
1998 dst_reg->smax_value = S64_MAX;
1999 } else {
2000 dst_reg->smin_value += smin_val;
2001 dst_reg->smax_value += smax_val;
2002 }
2003 if (dst_reg->umin_value + umin_val < umin_val ||
2004 dst_reg->umax_value + umax_val < umax_val) {
2005 dst_reg->umin_value = 0;
2006 dst_reg->umax_value = U64_MAX;
2007 } else {
2008 dst_reg->umin_value += umin_val;
2009 dst_reg->umax_value += umax_val;
2010 }
2011 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2012 break;
2013 case BPF_SUB:
2014 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2015 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2016 /* Overflow possible, we know nothing */
2017 dst_reg->smin_value = S64_MIN;
2018 dst_reg->smax_value = S64_MAX;
2019 } else {
2020 dst_reg->smin_value -= smax_val;
2021 dst_reg->smax_value -= smin_val;
2022 }
2023 if (dst_reg->umin_value < umax_val) {
2024 /* Overflow possible, we know nothing */
2025 dst_reg->umin_value = 0;
2026 dst_reg->umax_value = U64_MAX;
2027 } else {
2028 /* Cannot overflow (as long as bounds are consistent) */
2029 dst_reg->umin_value -= umax_val;
2030 dst_reg->umax_value -= umin_val;
2031 }
2032 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2033 break;
2034 case BPF_MUL:
2035 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2036 if (smin_val < 0 || dst_reg->smin_value < 0) {
2037 /* Ain't nobody got time to multiply that sign */
2038 __mark_reg_unbounded(dst_reg);
2039 __update_reg_bounds(dst_reg);
2040 break;
2041 }
2042 /* Both values are positive, so we can work with unsigned and
2043 * copy the result to signed (unless it exceeds S64_MAX).
2044 */
2045 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2046 /* Potential overflow, we know nothing */
2047 __mark_reg_unbounded(dst_reg);
2048 /* (except what we can learn from the var_off) */
2049 __update_reg_bounds(dst_reg);
2050 break;
2051 }
2052 dst_reg->umin_value *= umin_val;
2053 dst_reg->umax_value *= umax_val;
2054 if (dst_reg->umax_value > S64_MAX) {
2055 /* Overflow possible, we know nothing */
2056 dst_reg->smin_value = S64_MIN;
2057 dst_reg->smax_value = S64_MAX;
2058 } else {
2059 dst_reg->smin_value = dst_reg->umin_value;
2060 dst_reg->smax_value = dst_reg->umax_value;
2061 }
2062 break;
2063 case BPF_AND:
2064 if (src_known && dst_known) {
2065 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2066 src_reg.var_off.value);
2067 break;
2068 }
2069 /* We get our minimum from the var_off, since that's inherently
2070 * bitwise. Our maximum is the minimum of the operands' maxima.
2071 */
2072 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2073 dst_reg->umin_value = dst_reg->var_off.value;
2074 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2075 if (dst_reg->smin_value < 0 || smin_val < 0) {
2076 /* Lose signed bounds when ANDing negative numbers,
2077 * ain't nobody got time for that.
2078 */
2079 dst_reg->smin_value = S64_MIN;
2080 dst_reg->smax_value = S64_MAX;
2081 } else {
2082 /* ANDing two positives gives a positive, so safe to
2083 * cast result into s64.
2084 */
2085 dst_reg->smin_value = dst_reg->umin_value;
2086 dst_reg->smax_value = dst_reg->umax_value;
2087 }
2088 /* We may learn something more from the var_off */
2089 __update_reg_bounds(dst_reg);
2090 break;
2091 case BPF_OR:
2092 if (src_known && dst_known) {
2093 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2094 src_reg.var_off.value);
2095 break;
2096 }
2097 /* We get our maximum from the var_off, and our minimum is the
2098 * maximum of the operands' minima
2099 */
2100 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2101 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2102 dst_reg->umax_value = dst_reg->var_off.value |
2103 dst_reg->var_off.mask;
2104 if (dst_reg->smin_value < 0 || smin_val < 0) {
2105 /* Lose signed bounds when ORing negative numbers,
2106 * ain't nobody got time for that.
2107 */
2108 dst_reg->smin_value = S64_MIN;
2109 dst_reg->smax_value = S64_MAX;
2110 } else {
2111 /* ORing two positives gives a positive, so safe to
2112 * cast result into s64.
2113 */
2114 dst_reg->smin_value = dst_reg->umin_value;
2115 dst_reg->smax_value = dst_reg->umax_value;
2116 }
2117 /* We may learn something more from the var_off */
2118 __update_reg_bounds(dst_reg);
2119 break;
2120 case BPF_LSH:
2121 if (umax_val > 63) {
2122 /* Shifts greater than 63 are undefined. This includes
2123 * shifts by a negative number.
2124 */
2125 mark_reg_unknown(regs, insn->dst_reg);
2126 break;
2127 }
2128 /* We lose all sign bit information (except what we can pick
2129 * up from var_off)
2130 */
2131 dst_reg->smin_value = S64_MIN;
2132 dst_reg->smax_value = S64_MAX;
2133 /* If we might shift our top bit out, then we know nothing */
2134 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2135 dst_reg->umin_value = 0;
2136 dst_reg->umax_value = U64_MAX;
2137 } else {
2138 dst_reg->umin_value <<= umin_val;
2139 dst_reg->umax_value <<= umax_val;
2140 }
2141 if (src_known)
2142 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2143 else
2144 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2145 /* We may learn something more from the var_off */
2146 __update_reg_bounds(dst_reg);
2147 break;
2148 case BPF_RSH:
2149 if (umax_val > 63) {
2150 /* Shifts greater than 63 are undefined. This includes
2151 * shifts by a negative number.
2152 */
2153 mark_reg_unknown(regs, insn->dst_reg);
2154 break;
2155 }
2156 /* BPF_RSH is an unsigned shift, so make the appropriate casts */
2157 if (dst_reg->smin_value < 0) {
2158 if (umin_val) {
2159 /* Sign bit will be cleared */
2160 dst_reg->smin_value = 0;
2161 } else {
2162 /* Lost sign bit information */
2163 dst_reg->smin_value = S64_MIN;
2164 dst_reg->smax_value = S64_MAX;
2165 }
2166 } else {
2167 dst_reg->smin_value =
2168 (u64)(dst_reg->smin_value) >> umax_val;
2169 }
2170 if (src_known)
2171 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2172 umin_val);
2173 else
2174 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2175 dst_reg->umin_value >>= umax_val;
2176 dst_reg->umax_value >>= umin_val;
2177 /* We may learn something more from the var_off */
2178 __update_reg_bounds(dst_reg);
2179 break;
2180 default:
2181 mark_reg_unknown(regs, insn->dst_reg);
2182 break;
2183 }
2184
2185 __reg_deduce_bounds(dst_reg);
2186 __reg_bound_offset(dst_reg);
2187 return 0;
2188 }
2189
2190 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2191 * and var_off.
2192 */
2193 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2194 struct bpf_insn *insn)
2195 {
2196 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
2197 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2198 u8 opcode = BPF_OP(insn->code);
2199 int rc;
2200
2201 dst_reg = &regs[insn->dst_reg];
2202 src_reg = NULL;
2203 if (dst_reg->type != SCALAR_VALUE)
2204 ptr_reg = dst_reg;
2205 if (BPF_SRC(insn->code) == BPF_X) {
2206 src_reg = &regs[insn->src_reg];
2207 if (src_reg->type != SCALAR_VALUE) {
2208 if (dst_reg->type != SCALAR_VALUE) {
2209 /* Combining two pointers by any ALU op yields
2210 * an arbitrary scalar.
2211 */
2212 if (!env->allow_ptr_leaks) {
2213 verbose("R%d pointer %s pointer prohibited\n",
2214 insn->dst_reg,
2215 bpf_alu_string[opcode >> 4]);
2216 return -EACCES;
2217 }
2218 mark_reg_unknown(regs, insn->dst_reg);
2219 return 0;
2220 } else {
2221 /* scalar += pointer
2222 * This is legal, but we have to reverse our
2223 * src/dest handling in computing the range
2224 */
2225 rc = adjust_ptr_min_max_vals(env, insn,
2226 src_reg, dst_reg);
2227 if (rc == -EACCES && env->allow_ptr_leaks) {
2228 /* scalar += unknown scalar */
2229 __mark_reg_unknown(&off_reg);
2230 return adjust_scalar_min_max_vals(
2231 env, insn,
2232 dst_reg, off_reg);
2233 }
2234 return rc;
2235 }
2236 } else if (ptr_reg) {
2237 /* pointer += scalar */
2238 rc = adjust_ptr_min_max_vals(env, insn,
2239 dst_reg, src_reg);
2240 if (rc == -EACCES && env->allow_ptr_leaks) {
2241 /* unknown scalar += scalar */
2242 __mark_reg_unknown(dst_reg);
2243 return adjust_scalar_min_max_vals(
2244 env, insn, dst_reg, *src_reg);
2245 }
2246 return rc;
2247 }
2248 } else {
2249 /* Pretend the src is a reg with a known value, since we only
2250 * need to be able to read from this state.
2251 */
2252 off_reg.type = SCALAR_VALUE;
2253 __mark_reg_known(&off_reg, insn->imm);
2254 src_reg = &off_reg;
2255 if (ptr_reg) { /* pointer += K */
2256 rc = adjust_ptr_min_max_vals(env, insn,
2257 ptr_reg, src_reg);
2258 if (rc == -EACCES && env->allow_ptr_leaks) {
2259 /* unknown scalar += K */
2260 __mark_reg_unknown(dst_reg);
2261 return adjust_scalar_min_max_vals(
2262 env, insn, dst_reg, off_reg);
2263 }
2264 return rc;
2265 }
2266 }
2267
2268 /* Got here implies adding two SCALAR_VALUEs */
2269 if (WARN_ON_ONCE(ptr_reg)) {
2270 print_verifier_state(&env->cur_state);
2271 verbose("verifier internal error: unexpected ptr_reg\n");
2272 return -EINVAL;
2273 }
2274 if (WARN_ON(!src_reg)) {
2275 print_verifier_state(&env->cur_state);
2276 verbose("verifier internal error: no src_reg\n");
2277 return -EINVAL;
2278 }
2279 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2280 }
2281
2282 /* check validity of 32-bit and 64-bit arithmetic operations */
2283 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2284 {
2285 struct bpf_reg_state *regs = env->cur_state.regs;
2286 u8 opcode = BPF_OP(insn->code);
2287 int err;
2288
2289 if (opcode == BPF_END || opcode == BPF_NEG) {
2290 if (opcode == BPF_NEG) {
2291 if (BPF_SRC(insn->code) != 0 ||
2292 insn->src_reg != BPF_REG_0 ||
2293 insn->off != 0 || insn->imm != 0) {
2294 verbose("BPF_NEG uses reserved fields\n");
2295 return -EINVAL;
2296 }
2297 } else {
2298 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2299 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2300 BPF_CLASS(insn->code) == BPF_ALU64) {
2301 verbose("BPF_END uses reserved fields\n");
2302 return -EINVAL;
2303 }
2304 }
2305
2306 /* check src operand */
2307 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2308 if (err)
2309 return err;
2310
2311 if (is_pointer_value(env, insn->dst_reg)) {
2312 verbose("R%d pointer arithmetic prohibited\n",
2313 insn->dst_reg);
2314 return -EACCES;
2315 }
2316
2317 /* check dest operand */
2318 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2319 if (err)
2320 return err;
2321
2322 } else if (opcode == BPF_MOV) {
2323
2324 if (BPF_SRC(insn->code) == BPF_X) {
2325 if (insn->imm != 0 || insn->off != 0) {
2326 verbose("BPF_MOV uses reserved fields\n");
2327 return -EINVAL;
2328 }
2329
2330 /* check src operand */
2331 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2332 if (err)
2333 return err;
2334 } else {
2335 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2336 verbose("BPF_MOV uses reserved fields\n");
2337 return -EINVAL;
2338 }
2339 }
2340
2341 /* check dest operand */
2342 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2343 if (err)
2344 return err;
2345
2346 if (BPF_SRC(insn->code) == BPF_X) {
2347 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2348 /* case: R1 = R2
2349 * copy register state to dest reg
2350 */
2351 regs[insn->dst_reg] = regs[insn->src_reg];
2352 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
2353 } else {
2354 /* R1 = (u32) R2 */
2355 if (is_pointer_value(env, insn->src_reg)) {
2356 verbose("R%d partial copy of pointer\n",
2357 insn->src_reg);
2358 return -EACCES;
2359 }
2360 mark_reg_unknown(regs, insn->dst_reg);
2361 /* high 32 bits are known zero. */
2362 regs[insn->dst_reg].var_off = tnum_cast(
2363 regs[insn->dst_reg].var_off, 4);
2364 __update_reg_bounds(&regs[insn->dst_reg]);
2365 }
2366 } else {
2367 /* case: R = imm
2368 * remember the value we stored into this reg
2369 */
2370 regs[insn->dst_reg].type = SCALAR_VALUE;
2371 __mark_reg_known(regs + insn->dst_reg, insn->imm);
2372 }
2373
2374 } else if (opcode > BPF_END) {
2375 verbose("invalid BPF_ALU opcode %x\n", opcode);
2376 return -EINVAL;
2377
2378 } else { /* all other ALU ops: and, sub, xor, add, ... */
2379
2380 if (BPF_SRC(insn->code) == BPF_X) {
2381 if (insn->imm != 0 || insn->off != 0) {
2382 verbose("BPF_ALU uses reserved fields\n");
2383 return -EINVAL;
2384 }
2385 /* check src1 operand */
2386 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2387 if (err)
2388 return err;
2389 } else {
2390 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2391 verbose("BPF_ALU uses reserved fields\n");
2392 return -EINVAL;
2393 }
2394 }
2395
2396 /* check src2 operand */
2397 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2398 if (err)
2399 return err;
2400
2401 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2402 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2403 verbose("div by zero\n");
2404 return -EINVAL;
2405 }
2406
2407 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2408 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2409 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2410
2411 if (insn->imm < 0 || insn->imm >= size) {
2412 verbose("invalid shift %d\n", insn->imm);
2413 return -EINVAL;
2414 }
2415 }
2416
2417 /* check dest operand */
2418 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2419 if (err)
2420 return err;
2421
2422 return adjust_reg_min_max_vals(env, insn);
2423 }
2424
2425 return 0;
2426 }
2427
2428 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2429 struct bpf_reg_state *dst_reg)
2430 {
2431 struct bpf_reg_state *regs = state->regs, *reg;
2432 int i;
2433
2434 if (dst_reg->off < 0)
2435 /* This doesn't give us any range */
2436 return;
2437
2438 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2439 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2440 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2441 * than pkt_end, but that's because it's also less than pkt.
2442 */
2443 return;
2444
2445 /* LLVM can generate four kind of checks:
2446 *
2447 * Type 1/2:
2448 *
2449 * r2 = r3;
2450 * r2 += 8;
2451 * if (r2 > pkt_end) goto <handle exception>
2452 * <access okay>
2453 *
2454 * r2 = r3;
2455 * r2 += 8;
2456 * if (r2 < pkt_end) goto <access okay>
2457 * <handle exception>
2458 *
2459 * Where:
2460 * r2 == dst_reg, pkt_end == src_reg
2461 * r2=pkt(id=n,off=8,r=0)
2462 * r3=pkt(id=n,off=0,r=0)
2463 *
2464 * Type 3/4:
2465 *
2466 * r2 = r3;
2467 * r2 += 8;
2468 * if (pkt_end >= r2) goto <access okay>
2469 * <handle exception>
2470 *
2471 * r2 = r3;
2472 * r2 += 8;
2473 * if (pkt_end <= r2) goto <handle exception>
2474 * <access okay>
2475 *
2476 * Where:
2477 * pkt_end == dst_reg, r2 == src_reg
2478 * r2=pkt(id=n,off=8,r=0)
2479 * r3=pkt(id=n,off=0,r=0)
2480 *
2481 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2482 * so that range of bytes [r3, r3 + 8) is safe to access.
2483 */
2484
2485 /* If our ids match, then we must have the same max_value. And we
2486 * don't care about the other reg's fixed offset, since if it's too big
2487 * the range won't allow anything.
2488 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2489 */
2490 for (i = 0; i < MAX_BPF_REG; i++)
2491 if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
2492 /* keep the maximum range already checked */
2493 regs[i].range = max_t(u16, regs[i].range, dst_reg->off);
2494
2495 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2496 if (state->stack_slot_type[i] != STACK_SPILL)
2497 continue;
2498 reg = &state->spilled_regs[i / BPF_REG_SIZE];
2499 if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
2500 reg->range = max_t(u16, reg->range, dst_reg->off);
2501 }
2502 }
2503
2504 /* Adjusts the register min/max values in the case that the dst_reg is the
2505 * variable register that we are working on, and src_reg is a constant or we're
2506 * simply doing a BPF_K check.
2507 * In JEQ/JNE cases we also adjust the var_off values.
2508 */
2509 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2510 struct bpf_reg_state *false_reg, u64 val,
2511 u8 opcode)
2512 {
2513 /* If the dst_reg is a pointer, we can't learn anything about its
2514 * variable offset from the compare (unless src_reg were a pointer into
2515 * the same object, but we don't bother with that.
2516 * Since false_reg and true_reg have the same type by construction, we
2517 * only need to check one of them for pointerness.
2518 */
2519 if (__is_pointer_value(false, false_reg))
2520 return;
2521
2522 switch (opcode) {
2523 case BPF_JEQ:
2524 /* If this is false then we know nothing Jon Snow, but if it is
2525 * true then we know for sure.
2526 */
2527 __mark_reg_known(true_reg, val);
2528 break;
2529 case BPF_JNE:
2530 /* If this is true we know nothing Jon Snow, but if it is false
2531 * we know the value for sure;
2532 */
2533 __mark_reg_known(false_reg, val);
2534 break;
2535 case BPF_JGT:
2536 false_reg->umax_value = min(false_reg->umax_value, val);
2537 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2538 break;
2539 case BPF_JSGT:
2540 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2541 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2542 break;
2543 case BPF_JLT:
2544 false_reg->umin_value = max(false_reg->umin_value, val);
2545 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2546 break;
2547 case BPF_JSLT:
2548 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2549 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2550 break;
2551 case BPF_JGE:
2552 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2553 true_reg->umin_value = max(true_reg->umin_value, val);
2554 break;
2555 case BPF_JSGE:
2556 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2557 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2558 break;
2559 case BPF_JLE:
2560 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2561 true_reg->umax_value = min(true_reg->umax_value, val);
2562 break;
2563 case BPF_JSLE:
2564 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2565 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2566 break;
2567 default:
2568 break;
2569 }
2570
2571 __reg_deduce_bounds(false_reg);
2572 __reg_deduce_bounds(true_reg);
2573 /* We might have learned some bits from the bounds. */
2574 __reg_bound_offset(false_reg);
2575 __reg_bound_offset(true_reg);
2576 /* Intersecting with the old var_off might have improved our bounds
2577 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2578 * then new var_off is (0; 0x7f...fc) which improves our umax.
2579 */
2580 __update_reg_bounds(false_reg);
2581 __update_reg_bounds(true_reg);
2582 }
2583
2584 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2585 * the variable reg.
2586 */
2587 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2588 struct bpf_reg_state *false_reg, u64 val,
2589 u8 opcode)
2590 {
2591 if (__is_pointer_value(false, false_reg))
2592 return;
2593
2594 switch (opcode) {
2595 case BPF_JEQ:
2596 /* If this is false then we know nothing Jon Snow, but if it is
2597 * true then we know for sure.
2598 */
2599 __mark_reg_known(true_reg, val);
2600 break;
2601 case BPF_JNE:
2602 /* If this is true we know nothing Jon Snow, but if it is false
2603 * we know the value for sure;
2604 */
2605 __mark_reg_known(false_reg, val);
2606 break;
2607 case BPF_JGT:
2608 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2609 false_reg->umin_value = max(false_reg->umin_value, val);
2610 break;
2611 case BPF_JSGT:
2612 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2613 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2614 break;
2615 case BPF_JLT:
2616 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2617 false_reg->umax_value = min(false_reg->umax_value, val);
2618 break;
2619 case BPF_JSLT:
2620 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2621 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2622 break;
2623 case BPF_JGE:
2624 true_reg->umax_value = min(true_reg->umax_value, val);
2625 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2626 break;
2627 case BPF_JSGE:
2628 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2629 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2630 break;
2631 case BPF_JLE:
2632 true_reg->umin_value = max(true_reg->umin_value, val);
2633 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2634 break;
2635 case BPF_JSLE:
2636 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2637 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2638 break;
2639 default:
2640 break;
2641 }
2642
2643 __reg_deduce_bounds(false_reg);
2644 __reg_deduce_bounds(true_reg);
2645 /* We might have learned some bits from the bounds. */
2646 __reg_bound_offset(false_reg);
2647 __reg_bound_offset(true_reg);
2648 /* Intersecting with the old var_off might have improved our bounds
2649 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2650 * then new var_off is (0; 0x7f...fc) which improves our umax.
2651 */
2652 __update_reg_bounds(false_reg);
2653 __update_reg_bounds(true_reg);
2654 }
2655
2656 /* Regs are known to be equal, so intersect their min/max/var_off */
2657 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2658 struct bpf_reg_state *dst_reg)
2659 {
2660 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2661 dst_reg->umin_value);
2662 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2663 dst_reg->umax_value);
2664 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2665 dst_reg->smin_value);
2666 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2667 dst_reg->smax_value);
2668 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2669 dst_reg->var_off);
2670 /* We might have learned new bounds from the var_off. */
2671 __update_reg_bounds(src_reg);
2672 __update_reg_bounds(dst_reg);
2673 /* We might have learned something about the sign bit. */
2674 __reg_deduce_bounds(src_reg);
2675 __reg_deduce_bounds(dst_reg);
2676 /* We might have learned some bits from the bounds. */
2677 __reg_bound_offset(src_reg);
2678 __reg_bound_offset(dst_reg);
2679 /* Intersecting with the old var_off might have improved our bounds
2680 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2681 * then new var_off is (0; 0x7f...fc) which improves our umax.
2682 */
2683 __update_reg_bounds(src_reg);
2684 __update_reg_bounds(dst_reg);
2685 }
2686
2687 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2688 struct bpf_reg_state *true_dst,
2689 struct bpf_reg_state *false_src,
2690 struct bpf_reg_state *false_dst,
2691 u8 opcode)
2692 {
2693 switch (opcode) {
2694 case BPF_JEQ:
2695 __reg_combine_min_max(true_src, true_dst);
2696 break;
2697 case BPF_JNE:
2698 __reg_combine_min_max(false_src, false_dst);
2699 break;
2700 }
2701 }
2702
2703 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2704 bool is_null)
2705 {
2706 struct bpf_reg_state *reg = &regs[regno];
2707
2708 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2709 /* Old offset (both fixed and variable parts) should
2710 * have been known-zero, because we don't allow pointer
2711 * arithmetic on pointers that might be NULL.
2712 */
2713 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2714 !tnum_equals_const(reg->var_off, 0) ||
2715 reg->off)) {
2716 __mark_reg_known_zero(reg);
2717 reg->off = 0;
2718 }
2719 if (is_null) {
2720 reg->type = SCALAR_VALUE;
2721 } else if (reg->map_ptr->inner_map_meta) {
2722 reg->type = CONST_PTR_TO_MAP;
2723 reg->map_ptr = reg->map_ptr->inner_map_meta;
2724 } else {
2725 reg->type = PTR_TO_MAP_VALUE;
2726 }
2727 /* We don't need id from this point onwards anymore, thus we
2728 * should better reset it, so that state pruning has chances
2729 * to take effect.
2730 */
2731 reg->id = 0;
2732 }
2733 }
2734
2735 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2736 * be folded together at some point.
2737 */
2738 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2739 bool is_null)
2740 {
2741 struct bpf_reg_state *regs = state->regs;
2742 u32 id = regs[regno].id;
2743 int i;
2744
2745 for (i = 0; i < MAX_BPF_REG; i++)
2746 mark_map_reg(regs, i, id, is_null);
2747
2748 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2749 if (state->stack_slot_type[i] != STACK_SPILL)
2750 continue;
2751 mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
2752 }
2753 }
2754
2755 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2756 struct bpf_insn *insn, int *insn_idx)
2757 {
2758 struct bpf_verifier_state *other_branch, *this_branch = &env->cur_state;
2759 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2760 u8 opcode = BPF_OP(insn->code);
2761 int err;
2762
2763 if (opcode > BPF_JSLE) {
2764 verbose("invalid BPF_JMP opcode %x\n", opcode);
2765 return -EINVAL;
2766 }
2767
2768 if (BPF_SRC(insn->code) == BPF_X) {
2769 if (insn->imm != 0) {
2770 verbose("BPF_JMP uses reserved fields\n");
2771 return -EINVAL;
2772 }
2773
2774 /* check src1 operand */
2775 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2776 if (err)
2777 return err;
2778
2779 if (is_pointer_value(env, insn->src_reg)) {
2780 verbose("R%d pointer comparison prohibited\n",
2781 insn->src_reg);
2782 return -EACCES;
2783 }
2784 } else {
2785 if (insn->src_reg != BPF_REG_0) {
2786 verbose("BPF_JMP uses reserved fields\n");
2787 return -EINVAL;
2788 }
2789 }
2790
2791 /* check src2 operand */
2792 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2793 if (err)
2794 return err;
2795
2796 dst_reg = &regs[insn->dst_reg];
2797
2798 /* detect if R == 0 where R was initialized to zero earlier */
2799 if (BPF_SRC(insn->code) == BPF_K &&
2800 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2801 dst_reg->type == SCALAR_VALUE &&
2802 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2803 if (opcode == BPF_JEQ) {
2804 /* if (imm == imm) goto pc+off;
2805 * only follow the goto, ignore fall-through
2806 */
2807 *insn_idx += insn->off;
2808 return 0;
2809 } else {
2810 /* if (imm != imm) goto pc+off;
2811 * only follow fall-through branch, since
2812 * that's where the program will go
2813 */
2814 return 0;
2815 }
2816 }
2817
2818 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2819 if (!other_branch)
2820 return -EFAULT;
2821
2822 /* detect if we are comparing against a constant value so we can adjust
2823 * our min/max values for our dst register.
2824 * this is only legit if both are scalars (or pointers to the same
2825 * object, I suppose, but we don't support that right now), because
2826 * otherwise the different base pointers mean the offsets aren't
2827 * comparable.
2828 */
2829 if (BPF_SRC(insn->code) == BPF_X) {
2830 if (dst_reg->type == SCALAR_VALUE &&
2831 regs[insn->src_reg].type == SCALAR_VALUE) {
2832 if (tnum_is_const(regs[insn->src_reg].var_off))
2833 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2834 dst_reg, regs[insn->src_reg].var_off.value,
2835 opcode);
2836 else if (tnum_is_const(dst_reg->var_off))
2837 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
2838 &regs[insn->src_reg],
2839 dst_reg->var_off.value, opcode);
2840 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
2841 /* Comparing for equality, we can combine knowledge */
2842 reg_combine_min_max(&other_branch->regs[insn->src_reg],
2843 &other_branch->regs[insn->dst_reg],
2844 &regs[insn->src_reg],
2845 &regs[insn->dst_reg], opcode);
2846 }
2847 } else if (dst_reg->type == SCALAR_VALUE) {
2848 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2849 dst_reg, insn->imm, opcode);
2850 }
2851
2852 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2853 if (BPF_SRC(insn->code) == BPF_K &&
2854 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2855 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2856 /* Mark all identical map registers in each branch as either
2857 * safe or unknown depending R == 0 or R != 0 conditional.
2858 */
2859 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
2860 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
2861 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2862 dst_reg->type == PTR_TO_PACKET &&
2863 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2864 find_good_pkt_pointers(this_branch, dst_reg);
2865 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2866 dst_reg->type == PTR_TO_PACKET &&
2867 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2868 find_good_pkt_pointers(other_branch, dst_reg);
2869 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2870 dst_reg->type == PTR_TO_PACKET_END &&
2871 regs[insn->src_reg].type == PTR_TO_PACKET) {
2872 find_good_pkt_pointers(other_branch, &regs[insn->src_reg]);
2873 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2874 dst_reg->type == PTR_TO_PACKET_END &&
2875 regs[insn->src_reg].type == PTR_TO_PACKET) {
2876 find_good_pkt_pointers(this_branch, &regs[insn->src_reg]);
2877 } else if (is_pointer_value(env, insn->dst_reg)) {
2878 verbose("R%d pointer comparison prohibited\n", insn->dst_reg);
2879 return -EACCES;
2880 }
2881 if (log_level)
2882 print_verifier_state(this_branch);
2883 return 0;
2884 }
2885
2886 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
2887 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
2888 {
2889 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
2890
2891 return (struct bpf_map *) (unsigned long) imm64;
2892 }
2893
2894 /* verify BPF_LD_IMM64 instruction */
2895 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
2896 {
2897 struct bpf_reg_state *regs = env->cur_state.regs;
2898 int err;
2899
2900 if (BPF_SIZE(insn->code) != BPF_DW) {
2901 verbose("invalid BPF_LD_IMM insn\n");
2902 return -EINVAL;
2903 }
2904 if (insn->off != 0) {
2905 verbose("BPF_LD_IMM64 uses reserved fields\n");
2906 return -EINVAL;
2907 }
2908
2909 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2910 if (err)
2911 return err;
2912
2913 if (insn->src_reg == 0) {
2914 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
2915
2916 regs[insn->dst_reg].type = SCALAR_VALUE;
2917 __mark_reg_known(&regs[insn->dst_reg], imm);
2918 return 0;
2919 }
2920
2921 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
2922 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
2923
2924 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
2925 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
2926 return 0;
2927 }
2928
2929 static bool may_access_skb(enum bpf_prog_type type)
2930 {
2931 switch (type) {
2932 case BPF_PROG_TYPE_SOCKET_FILTER:
2933 case BPF_PROG_TYPE_SCHED_CLS:
2934 case BPF_PROG_TYPE_SCHED_ACT:
2935 return true;
2936 default:
2937 return false;
2938 }
2939 }
2940
2941 /* verify safety of LD_ABS|LD_IND instructions:
2942 * - they can only appear in the programs where ctx == skb
2943 * - since they are wrappers of function calls, they scratch R1-R5 registers,
2944 * preserve R6-R9, and store return value into R0
2945 *
2946 * Implicit input:
2947 * ctx == skb == R6 == CTX
2948 *
2949 * Explicit input:
2950 * SRC == any register
2951 * IMM == 32-bit immediate
2952 *
2953 * Output:
2954 * R0 - 8/16/32-bit skb data converted to cpu endianness
2955 */
2956 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
2957 {
2958 struct bpf_reg_state *regs = env->cur_state.regs;
2959 u8 mode = BPF_MODE(insn->code);
2960 int i, err;
2961
2962 if (!may_access_skb(env->prog->type)) {
2963 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
2964 return -EINVAL;
2965 }
2966
2967 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
2968 BPF_SIZE(insn->code) == BPF_DW ||
2969 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
2970 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
2971 return -EINVAL;
2972 }
2973
2974 /* check whether implicit source operand (register R6) is readable */
2975 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
2976 if (err)
2977 return err;
2978
2979 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
2980 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
2981 return -EINVAL;
2982 }
2983
2984 if (mode == BPF_IND) {
2985 /* check explicit source operand */
2986 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2987 if (err)
2988 return err;
2989 }
2990
2991 /* reset caller saved regs to unreadable */
2992 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2993 mark_reg_not_init(regs, caller_saved[i]);
2994 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2995 }
2996
2997 /* mark destination R0 register as readable, since it contains
2998 * the value fetched from the packet.
2999 * Already marked as written above.
3000 */
3001 mark_reg_unknown(regs, BPF_REG_0);
3002 return 0;
3003 }
3004
3005 /* non-recursive DFS pseudo code
3006 * 1 procedure DFS-iterative(G,v):
3007 * 2 label v as discovered
3008 * 3 let S be a stack
3009 * 4 S.push(v)
3010 * 5 while S is not empty
3011 * 6 t <- S.pop()
3012 * 7 if t is what we're looking for:
3013 * 8 return t
3014 * 9 for all edges e in G.adjacentEdges(t) do
3015 * 10 if edge e is already labelled
3016 * 11 continue with the next edge
3017 * 12 w <- G.adjacentVertex(t,e)
3018 * 13 if vertex w is not discovered and not explored
3019 * 14 label e as tree-edge
3020 * 15 label w as discovered
3021 * 16 S.push(w)
3022 * 17 continue at 5
3023 * 18 else if vertex w is discovered
3024 * 19 label e as back-edge
3025 * 20 else
3026 * 21 // vertex w is explored
3027 * 22 label e as forward- or cross-edge
3028 * 23 label t as explored
3029 * 24 S.pop()
3030 *
3031 * convention:
3032 * 0x10 - discovered
3033 * 0x11 - discovered and fall-through edge labelled
3034 * 0x12 - discovered and fall-through and branch edges labelled
3035 * 0x20 - explored
3036 */
3037
3038 enum {
3039 DISCOVERED = 0x10,
3040 EXPLORED = 0x20,
3041 FALLTHROUGH = 1,
3042 BRANCH = 2,
3043 };
3044
3045 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3046
3047 static int *insn_stack; /* stack of insns to process */
3048 static int cur_stack; /* current stack index */
3049 static int *insn_state;
3050
3051 /* t, w, e - match pseudo-code above:
3052 * t - index of current instruction
3053 * w - next instruction
3054 * e - edge
3055 */
3056 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3057 {
3058 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3059 return 0;
3060
3061 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3062 return 0;
3063
3064 if (w < 0 || w >= env->prog->len) {
3065 verbose("jump out of range from insn %d to %d\n", t, w);
3066 return -EINVAL;
3067 }
3068
3069 if (e == BRANCH)
3070 /* mark branch target for state pruning */
3071 env->explored_states[w] = STATE_LIST_MARK;
3072
3073 if (insn_state[w] == 0) {
3074 /* tree-edge */
3075 insn_state[t] = DISCOVERED | e;
3076 insn_state[w] = DISCOVERED;
3077 if (cur_stack >= env->prog->len)
3078 return -E2BIG;
3079 insn_stack[cur_stack++] = w;
3080 return 1;
3081 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3082 verbose("back-edge from insn %d to %d\n", t, w);
3083 return -EINVAL;
3084 } else if (insn_state[w] == EXPLORED) {
3085 /* forward- or cross-edge */
3086 insn_state[t] = DISCOVERED | e;
3087 } else {
3088 verbose("insn state internal bug\n");
3089 return -EFAULT;
3090 }
3091 return 0;
3092 }
3093
3094 /* non-recursive depth-first-search to detect loops in BPF program
3095 * loop == back-edge in directed graph
3096 */
3097 static int check_cfg(struct bpf_verifier_env *env)
3098 {
3099 struct bpf_insn *insns = env->prog->insnsi;
3100 int insn_cnt = env->prog->len;
3101 int ret = 0;
3102 int i, t;
3103
3104 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3105 if (!insn_state)
3106 return -ENOMEM;
3107
3108 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3109 if (!insn_stack) {
3110 kfree(insn_state);
3111 return -ENOMEM;
3112 }
3113
3114 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3115 insn_stack[0] = 0; /* 0 is the first instruction */
3116 cur_stack = 1;
3117
3118 peek_stack:
3119 if (cur_stack == 0)
3120 goto check_state;
3121 t = insn_stack[cur_stack - 1];
3122
3123 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3124 u8 opcode = BPF_OP(insns[t].code);
3125
3126 if (opcode == BPF_EXIT) {
3127 goto mark_explored;
3128 } else if (opcode == BPF_CALL) {
3129 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3130 if (ret == 1)
3131 goto peek_stack;
3132 else if (ret < 0)
3133 goto err_free;
3134 if (t + 1 < insn_cnt)
3135 env->explored_states[t + 1] = STATE_LIST_MARK;
3136 } else if (opcode == BPF_JA) {
3137 if (BPF_SRC(insns[t].code) != BPF_K) {
3138 ret = -EINVAL;
3139 goto err_free;
3140 }
3141 /* unconditional jump with single edge */
3142 ret = push_insn(t, t + insns[t].off + 1,
3143 FALLTHROUGH, env);
3144 if (ret == 1)
3145 goto peek_stack;
3146 else if (ret < 0)
3147 goto err_free;
3148 /* tell verifier to check for equivalent states
3149 * after every call and jump
3150 */
3151 if (t + 1 < insn_cnt)
3152 env->explored_states[t + 1] = STATE_LIST_MARK;
3153 } else {
3154 /* conditional jump with two edges */
3155 env->explored_states[t] = STATE_LIST_MARK;
3156 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3157 if (ret == 1)
3158 goto peek_stack;
3159 else if (ret < 0)
3160 goto err_free;
3161
3162 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3163 if (ret == 1)
3164 goto peek_stack;
3165 else if (ret < 0)
3166 goto err_free;
3167 }
3168 } else {
3169 /* all other non-branch instructions with single
3170 * fall-through edge
3171 */
3172 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3173 if (ret == 1)
3174 goto peek_stack;
3175 else if (ret < 0)
3176 goto err_free;
3177 }
3178
3179 mark_explored:
3180 insn_state[t] = EXPLORED;
3181 if (cur_stack-- <= 0) {
3182 verbose("pop stack internal bug\n");
3183 ret = -EFAULT;
3184 goto err_free;
3185 }
3186 goto peek_stack;
3187
3188 check_state:
3189 for (i = 0; i < insn_cnt; i++) {
3190 if (insn_state[i] != EXPLORED) {
3191 verbose("unreachable insn %d\n", i);
3192 ret = -EINVAL;
3193 goto err_free;
3194 }
3195 }
3196 ret = 0; /* cfg looks good */
3197
3198 err_free:
3199 kfree(insn_state);
3200 kfree(insn_stack);
3201 return ret;
3202 }
3203
3204 /* check %cur's range satisfies %old's */
3205 static bool range_within(struct bpf_reg_state *old,
3206 struct bpf_reg_state *cur)
3207 {
3208 return old->umin_value <= cur->umin_value &&
3209 old->umax_value >= cur->umax_value &&
3210 old->smin_value <= cur->smin_value &&
3211 old->smax_value >= cur->smax_value;
3212 }
3213
3214 /* Maximum number of register states that can exist at once */
3215 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3216 struct idpair {
3217 u32 old;
3218 u32 cur;
3219 };
3220
3221 /* If in the old state two registers had the same id, then they need to have
3222 * the same id in the new state as well. But that id could be different from
3223 * the old state, so we need to track the mapping from old to new ids.
3224 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3225 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3226 * regs with a different old id could still have new id 9, we don't care about
3227 * that.
3228 * So we look through our idmap to see if this old id has been seen before. If
3229 * so, we require the new id to match; otherwise, we add the id pair to the map.
3230 */
3231 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3232 {
3233 unsigned int i;
3234
3235 for (i = 0; i < ID_MAP_SIZE; i++) {
3236 if (!idmap[i].old) {
3237 /* Reached an empty slot; haven't seen this id before */
3238 idmap[i].old = old_id;
3239 idmap[i].cur = cur_id;
3240 return true;
3241 }
3242 if (idmap[i].old == old_id)
3243 return idmap[i].cur == cur_id;
3244 }
3245 /* We ran out of idmap slots, which should be impossible */
3246 WARN_ON_ONCE(1);
3247 return false;
3248 }
3249
3250 /* Returns true if (rold safe implies rcur safe) */
3251 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3252 struct idpair *idmap)
3253 {
3254 if (!(rold->live & REG_LIVE_READ))
3255 /* explored state didn't use this */
3256 return true;
3257
3258 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3259 return true;
3260
3261 if (rold->type == NOT_INIT)
3262 /* explored state can't have used this */
3263 return true;
3264 if (rcur->type == NOT_INIT)
3265 return false;
3266 switch (rold->type) {
3267 case SCALAR_VALUE:
3268 if (rcur->type == SCALAR_VALUE) {
3269 /* new val must satisfy old val knowledge */
3270 return range_within(rold, rcur) &&
3271 tnum_in(rold->var_off, rcur->var_off);
3272 } else {
3273 /* if we knew anything about the old value, we're not
3274 * equal, because we can't know anything about the
3275 * scalar value of the pointer in the new value.
3276 */
3277 return rold->umin_value == 0 &&
3278 rold->umax_value == U64_MAX &&
3279 rold->smin_value == S64_MIN &&
3280 rold->smax_value == S64_MAX &&
3281 tnum_is_unknown(rold->var_off);
3282 }
3283 case PTR_TO_MAP_VALUE:
3284 /* If the new min/max/var_off satisfy the old ones and
3285 * everything else matches, we are OK.
3286 * We don't care about the 'id' value, because nothing
3287 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3288 */
3289 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3290 range_within(rold, rcur) &&
3291 tnum_in(rold->var_off, rcur->var_off);
3292 case PTR_TO_MAP_VALUE_OR_NULL:
3293 /* a PTR_TO_MAP_VALUE could be safe to use as a
3294 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3295 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3296 * checked, doing so could have affected others with the same
3297 * id, and we can't check for that because we lost the id when
3298 * we converted to a PTR_TO_MAP_VALUE.
3299 */
3300 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3301 return false;
3302 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3303 return false;
3304 /* Check our ids match any regs they're supposed to */
3305 return check_ids(rold->id, rcur->id, idmap);
3306 case PTR_TO_PACKET:
3307 if (rcur->type != PTR_TO_PACKET)
3308 return false;
3309 /* We must have at least as much range as the old ptr
3310 * did, so that any accesses which were safe before are
3311 * still safe. This is true even if old range < old off,
3312 * since someone could have accessed through (ptr - k), or
3313 * even done ptr -= k in a register, to get a safe access.
3314 */
3315 if (rold->range > rcur->range)
3316 return false;
3317 /* If the offsets don't match, we can't trust our alignment;
3318 * nor can we be sure that we won't fall out of range.
3319 */
3320 if (rold->off != rcur->off)
3321 return false;
3322 /* id relations must be preserved */
3323 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3324 return false;
3325 /* new val must satisfy old val knowledge */
3326 return range_within(rold, rcur) &&
3327 tnum_in(rold->var_off, rcur->var_off);
3328 case PTR_TO_CTX:
3329 case CONST_PTR_TO_MAP:
3330 case PTR_TO_STACK:
3331 case PTR_TO_PACKET_END:
3332 /* Only valid matches are exact, which memcmp() above
3333 * would have accepted
3334 */
3335 default:
3336 /* Don't know what's going on, just say it's not safe */
3337 return false;
3338 }
3339
3340 /* Shouldn't get here; if we do, say it's not safe */
3341 WARN_ON_ONCE(1);
3342 return false;
3343 }
3344
3345 /* compare two verifier states
3346 *
3347 * all states stored in state_list are known to be valid, since
3348 * verifier reached 'bpf_exit' instruction through them
3349 *
3350 * this function is called when verifier exploring different branches of
3351 * execution popped from the state stack. If it sees an old state that has
3352 * more strict register state and more strict stack state then this execution
3353 * branch doesn't need to be explored further, since verifier already
3354 * concluded that more strict state leads to valid finish.
3355 *
3356 * Therefore two states are equivalent if register state is more conservative
3357 * and explored stack state is more conservative than the current one.
3358 * Example:
3359 * explored current
3360 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3361 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3362 *
3363 * In other words if current stack state (one being explored) has more
3364 * valid slots than old one that already passed validation, it means
3365 * the verifier can stop exploring and conclude that current state is valid too
3366 *
3367 * Similarly with registers. If explored state has register type as invalid
3368 * whereas register type in current state is meaningful, it means that
3369 * the current state will reach 'bpf_exit' instruction safely
3370 */
3371 static bool states_equal(struct bpf_verifier_env *env,
3372 struct bpf_verifier_state *old,
3373 struct bpf_verifier_state *cur)
3374 {
3375 struct idpair *idmap;
3376 bool ret = false;
3377 int i;
3378
3379 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3380 /* If we failed to allocate the idmap, just say it's not safe */
3381 if (!idmap)
3382 return false;
3383
3384 for (i = 0; i < MAX_BPF_REG; i++) {
3385 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3386 goto out_free;
3387 }
3388
3389 for (i = 0; i < MAX_BPF_STACK; i++) {
3390 if (old->stack_slot_type[i] == STACK_INVALID)
3391 continue;
3392 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
3393 /* Ex: old explored (safe) state has STACK_SPILL in
3394 * this stack slot, but current has has STACK_MISC ->
3395 * this verifier states are not equivalent,
3396 * return false to continue verification of this path
3397 */
3398 goto out_free;
3399 if (i % BPF_REG_SIZE)
3400 continue;
3401 if (old->stack_slot_type[i] != STACK_SPILL)
3402 continue;
3403 if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
3404 &cur->spilled_regs[i / BPF_REG_SIZE],
3405 idmap))
3406 /* when explored and current stack slot are both storing
3407 * spilled registers, check that stored pointers types
3408 * are the same as well.
3409 * Ex: explored safe path could have stored
3410 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3411 * but current path has stored:
3412 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3413 * such verifier states are not equivalent.
3414 * return false to continue verification of this path
3415 */
3416 goto out_free;
3417 else
3418 continue;
3419 }
3420 ret = true;
3421 out_free:
3422 kfree(idmap);
3423 return ret;
3424 }
3425
3426 /* A write screens off any subsequent reads; but write marks come from the
3427 * straight-line code between a state and its parent. When we arrive at a
3428 * jump target (in the first iteration of the propagate_liveness() loop),
3429 * we didn't arrive by the straight-line code, so read marks in state must
3430 * propagate to parent regardless of state's write marks.
3431 */
3432 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3433 struct bpf_verifier_state *parent)
3434 {
3435 bool writes = parent == state->parent; /* Observe write marks */
3436 bool touched = false; /* any changes made? */
3437 int i;
3438
3439 if (!parent)
3440 return touched;
3441 /* Propagate read liveness of registers... */
3442 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3443 /* We don't need to worry about FP liveness because it's read-only */
3444 for (i = 0; i < BPF_REG_FP; i++) {
3445 if (parent->regs[i].live & REG_LIVE_READ)
3446 continue;
3447 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3448 continue;
3449 if (state->regs[i].live & REG_LIVE_READ) {
3450 parent->regs[i].live |= REG_LIVE_READ;
3451 touched = true;
3452 }
3453 }
3454 /* ... and stack slots */
3455 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++) {
3456 if (parent->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3457 continue;
3458 if (state->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3459 continue;
3460 if (parent->spilled_regs[i].live & REG_LIVE_READ)
3461 continue;
3462 if (writes && (state->spilled_regs[i].live & REG_LIVE_WRITTEN))
3463 continue;
3464 if (state->spilled_regs[i].live & REG_LIVE_READ) {
3465 parent->spilled_regs[i].live |= REG_LIVE_READ;
3466 touched = true;
3467 }
3468 }
3469 return touched;
3470 }
3471
3472 /* "parent" is "a state from which we reach the current state", but initially
3473 * it is not the state->parent (i.e. "the state whose straight-line code leads
3474 * to the current state"), instead it is the state that happened to arrive at
3475 * a (prunable) equivalent of the current state. See comment above
3476 * do_propagate_liveness() for consequences of this.
3477 * This function is just a more efficient way of calling mark_reg_read() or
3478 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3479 * though it requires that parent != state->parent in the call arguments.
3480 */
3481 static void propagate_liveness(const struct bpf_verifier_state *state,
3482 struct bpf_verifier_state *parent)
3483 {
3484 while (do_propagate_liveness(state, parent)) {
3485 /* Something changed, so we need to feed those changes onward */
3486 state = parent;
3487 parent = state->parent;
3488 }
3489 }
3490
3491 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3492 {
3493 struct bpf_verifier_state_list *new_sl;
3494 struct bpf_verifier_state_list *sl;
3495 int i;
3496
3497 sl = env->explored_states[insn_idx];
3498 if (!sl)
3499 /* this 'insn_idx' instruction wasn't marked, so we will not
3500 * be doing state search here
3501 */
3502 return 0;
3503
3504 while (sl != STATE_LIST_MARK) {
3505 if (states_equal(env, &sl->state, &env->cur_state)) {
3506 /* reached equivalent register/stack state,
3507 * prune the search.
3508 * Registers read by the continuation are read by us.
3509 * If we have any write marks in env->cur_state, they
3510 * will prevent corresponding reads in the continuation
3511 * from reaching our parent (an explored_state). Our
3512 * own state will get the read marks recorded, but
3513 * they'll be immediately forgotten as we're pruning
3514 * this state and will pop a new one.
3515 */
3516 propagate_liveness(&sl->state, &env->cur_state);
3517 return 1;
3518 }
3519 sl = sl->next;
3520 }
3521
3522 /* there were no equivalent states, remember current one.
3523 * technically the current state is not proven to be safe yet,
3524 * but it will either reach bpf_exit (which means it's safe) or
3525 * it will be rejected. Since there are no loops, we won't be
3526 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3527 */
3528 new_sl = kmalloc(sizeof(struct bpf_verifier_state_list), GFP_USER);
3529 if (!new_sl)
3530 return -ENOMEM;
3531
3532 /* add new state to the head of linked list */
3533 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
3534 new_sl->next = env->explored_states[insn_idx];
3535 env->explored_states[insn_idx] = new_sl;
3536 /* connect new state to parentage chain */
3537 env->cur_state.parent = &new_sl->state;
3538 /* clear write marks in current state: the writes we did are not writes
3539 * our child did, so they don't screen off its reads from us.
3540 * (There are no read marks in current state, because reads always mark
3541 * their parent and current state never has children yet. Only
3542 * explored_states can get read marks.)
3543 */
3544 for (i = 0; i < BPF_REG_FP; i++)
3545 env->cur_state.regs[i].live = REG_LIVE_NONE;
3546 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++)
3547 if (env->cur_state.stack_slot_type[i * BPF_REG_SIZE] == STACK_SPILL)
3548 env->cur_state.spilled_regs[i].live = REG_LIVE_NONE;
3549 return 0;
3550 }
3551
3552 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3553 int insn_idx, int prev_insn_idx)
3554 {
3555 if (!env->analyzer_ops || !env->analyzer_ops->insn_hook)
3556 return 0;
3557
3558 return env->analyzer_ops->insn_hook(env, insn_idx, prev_insn_idx);
3559 }
3560
3561 static int do_check(struct bpf_verifier_env *env)
3562 {
3563 struct bpf_verifier_state *state = &env->cur_state;
3564 struct bpf_insn *insns = env->prog->insnsi;
3565 struct bpf_reg_state *regs = state->regs;
3566 int insn_cnt = env->prog->len;
3567 int insn_idx, prev_insn_idx = 0;
3568 int insn_processed = 0;
3569 bool do_print_state = false;
3570
3571 init_reg_state(regs);
3572 state->parent = NULL;
3573 insn_idx = 0;
3574 for (;;) {
3575 struct bpf_insn *insn;
3576 u8 class;
3577 int err;
3578
3579 if (insn_idx >= insn_cnt) {
3580 verbose("invalid insn idx %d insn_cnt %d\n",
3581 insn_idx, insn_cnt);
3582 return -EFAULT;
3583 }
3584
3585 insn = &insns[insn_idx];
3586 class = BPF_CLASS(insn->code);
3587
3588 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3589 verbose("BPF program is too large. Processed %d insn\n",
3590 insn_processed);
3591 return -E2BIG;
3592 }
3593
3594 err = is_state_visited(env, insn_idx);
3595 if (err < 0)
3596 return err;
3597 if (err == 1) {
3598 /* found equivalent state, can prune the search */
3599 if (log_level) {
3600 if (do_print_state)
3601 verbose("\nfrom %d to %d: safe\n",
3602 prev_insn_idx, insn_idx);
3603 else
3604 verbose("%d: safe\n", insn_idx);
3605 }
3606 goto process_bpf_exit;
3607 }
3608
3609 if (need_resched())
3610 cond_resched();
3611
3612 if (log_level > 1 || (log_level && do_print_state)) {
3613 if (log_level > 1)
3614 verbose("%d:", insn_idx);
3615 else
3616 verbose("\nfrom %d to %d:",
3617 prev_insn_idx, insn_idx);
3618 print_verifier_state(&env->cur_state);
3619 do_print_state = false;
3620 }
3621
3622 if (log_level) {
3623 verbose("%d: ", insn_idx);
3624 print_bpf_insn(env, insn);
3625 }
3626
3627 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3628 if (err)
3629 return err;
3630
3631 if (class == BPF_ALU || class == BPF_ALU64) {
3632 err = check_alu_op(env, insn);
3633 if (err)
3634 return err;
3635
3636 } else if (class == BPF_LDX) {
3637 enum bpf_reg_type *prev_src_type, src_reg_type;
3638
3639 /* check for reserved fields is already done */
3640
3641 /* check src operand */
3642 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3643 if (err)
3644 return err;
3645
3646 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3647 if (err)
3648 return err;
3649
3650 src_reg_type = regs[insn->src_reg].type;
3651
3652 /* check that memory (src_reg + off) is readable,
3653 * the state of dst_reg will be updated by this func
3654 */
3655 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3656 BPF_SIZE(insn->code), BPF_READ,
3657 insn->dst_reg);
3658 if (err)
3659 return err;
3660
3661 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3662
3663 if (*prev_src_type == NOT_INIT) {
3664 /* saw a valid insn
3665 * dst_reg = *(u32 *)(src_reg + off)
3666 * save type to validate intersecting paths
3667 */
3668 *prev_src_type = src_reg_type;
3669
3670 } else if (src_reg_type != *prev_src_type &&
3671 (src_reg_type == PTR_TO_CTX ||
3672 *prev_src_type == PTR_TO_CTX)) {
3673 /* ABuser program is trying to use the same insn
3674 * dst_reg = *(u32*) (src_reg + off)
3675 * with different pointer types:
3676 * src_reg == ctx in one branch and
3677 * src_reg == stack|map in some other branch.
3678 * Reject it.
3679 */
3680 verbose("same insn cannot be used with different pointers\n");
3681 return -EINVAL;
3682 }
3683
3684 } else if (class == BPF_STX) {
3685 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3686
3687 if (BPF_MODE(insn->code) == BPF_XADD) {
3688 err = check_xadd(env, insn_idx, insn);
3689 if (err)
3690 return err;
3691 insn_idx++;
3692 continue;
3693 }
3694
3695 /* check src1 operand */
3696 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3697 if (err)
3698 return err;
3699 /* check src2 operand */
3700 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3701 if (err)
3702 return err;
3703
3704 dst_reg_type = regs[insn->dst_reg].type;
3705
3706 /* check that memory (dst_reg + off) is writeable */
3707 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3708 BPF_SIZE(insn->code), BPF_WRITE,
3709 insn->src_reg);
3710 if (err)
3711 return err;
3712
3713 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3714
3715 if (*prev_dst_type == NOT_INIT) {
3716 *prev_dst_type = dst_reg_type;
3717 } else if (dst_reg_type != *prev_dst_type &&
3718 (dst_reg_type == PTR_TO_CTX ||
3719 *prev_dst_type == PTR_TO_CTX)) {
3720 verbose("same insn cannot be used with different pointers\n");
3721 return -EINVAL;
3722 }
3723
3724 } else if (class == BPF_ST) {
3725 if (BPF_MODE(insn->code) != BPF_MEM ||
3726 insn->src_reg != BPF_REG_0) {
3727 verbose("BPF_ST uses reserved fields\n");
3728 return -EINVAL;
3729 }
3730 /* check src operand */
3731 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3732 if (err)
3733 return err;
3734
3735 /* check that memory (dst_reg + off) is writeable */
3736 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3737 BPF_SIZE(insn->code), BPF_WRITE,
3738 -1);
3739 if (err)
3740 return err;
3741
3742 } else if (class == BPF_JMP) {
3743 u8 opcode = BPF_OP(insn->code);
3744
3745 if (opcode == BPF_CALL) {
3746 if (BPF_SRC(insn->code) != BPF_K ||
3747 insn->off != 0 ||
3748 insn->src_reg != BPF_REG_0 ||
3749 insn->dst_reg != BPF_REG_0) {
3750 verbose("BPF_CALL uses reserved fields\n");
3751 return -EINVAL;
3752 }
3753
3754 err = check_call(env, insn->imm, insn_idx);
3755 if (err)
3756 return err;
3757
3758 } else if (opcode == BPF_JA) {
3759 if (BPF_SRC(insn->code) != BPF_K ||
3760 insn->imm != 0 ||
3761 insn->src_reg != BPF_REG_0 ||
3762 insn->dst_reg != BPF_REG_0) {
3763 verbose("BPF_JA uses reserved fields\n");
3764 return -EINVAL;
3765 }
3766
3767 insn_idx += insn->off + 1;
3768 continue;
3769
3770 } else if (opcode == BPF_EXIT) {
3771 if (BPF_SRC(insn->code) != BPF_K ||
3772 insn->imm != 0 ||
3773 insn->src_reg != BPF_REG_0 ||
3774 insn->dst_reg != BPF_REG_0) {
3775 verbose("BPF_EXIT uses reserved fields\n");
3776 return -EINVAL;
3777 }
3778
3779 /* eBPF calling convetion is such that R0 is used
3780 * to return the value from eBPF program.
3781 * Make sure that it's readable at this time
3782 * of bpf_exit, which means that program wrote
3783 * something into it earlier
3784 */
3785 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
3786 if (err)
3787 return err;
3788
3789 if (is_pointer_value(env, BPF_REG_0)) {
3790 verbose("R0 leaks addr as return value\n");
3791 return -EACCES;
3792 }
3793
3794 process_bpf_exit:
3795 insn_idx = pop_stack(env, &prev_insn_idx);
3796 if (insn_idx < 0) {
3797 break;
3798 } else {
3799 do_print_state = true;
3800 continue;
3801 }
3802 } else {
3803 err = check_cond_jmp_op(env, insn, &insn_idx);
3804 if (err)
3805 return err;
3806 }
3807 } else if (class == BPF_LD) {
3808 u8 mode = BPF_MODE(insn->code);
3809
3810 if (mode == BPF_ABS || mode == BPF_IND) {
3811 err = check_ld_abs(env, insn);
3812 if (err)
3813 return err;
3814
3815 } else if (mode == BPF_IMM) {
3816 err = check_ld_imm(env, insn);
3817 if (err)
3818 return err;
3819
3820 insn_idx++;
3821 } else {
3822 verbose("invalid BPF_LD mode\n");
3823 return -EINVAL;
3824 }
3825 } else {
3826 verbose("unknown insn class %d\n", class);
3827 return -EINVAL;
3828 }
3829
3830 insn_idx++;
3831 }
3832
3833 verbose("processed %d insns, stack depth %d\n",
3834 insn_processed, env->prog->aux->stack_depth);
3835 return 0;
3836 }
3837
3838 static int check_map_prealloc(struct bpf_map *map)
3839 {
3840 return (map->map_type != BPF_MAP_TYPE_HASH &&
3841 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
3842 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
3843 !(map->map_flags & BPF_F_NO_PREALLOC);
3844 }
3845
3846 static int check_map_prog_compatibility(struct bpf_map *map,
3847 struct bpf_prog *prog)
3848
3849 {
3850 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3851 * preallocated hash maps, since doing memory allocation
3852 * in overflow_handler can crash depending on where nmi got
3853 * triggered.
3854 */
3855 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
3856 if (!check_map_prealloc(map)) {
3857 verbose("perf_event programs can only use preallocated hash map\n");
3858 return -EINVAL;
3859 }
3860 if (map->inner_map_meta &&
3861 !check_map_prealloc(map->inner_map_meta)) {
3862 verbose("perf_event programs can only use preallocated inner hash map\n");
3863 return -EINVAL;
3864 }
3865 }
3866 return 0;
3867 }
3868
3869 /* look for pseudo eBPF instructions that access map FDs and
3870 * replace them with actual map pointers
3871 */
3872 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
3873 {
3874 struct bpf_insn *insn = env->prog->insnsi;
3875 int insn_cnt = env->prog->len;
3876 int i, j, err;
3877
3878 err = bpf_prog_calc_tag(env->prog);
3879 if (err)
3880 return err;
3881
3882 for (i = 0; i < insn_cnt; i++, insn++) {
3883 if (BPF_CLASS(insn->code) == BPF_LDX &&
3884 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
3885 verbose("BPF_LDX uses reserved fields\n");
3886 return -EINVAL;
3887 }
3888
3889 if (BPF_CLASS(insn->code) == BPF_STX &&
3890 ((BPF_MODE(insn->code) != BPF_MEM &&
3891 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
3892 verbose("BPF_STX uses reserved fields\n");
3893 return -EINVAL;
3894 }
3895
3896 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
3897 struct bpf_map *map;
3898 struct fd f;
3899
3900 if (i == insn_cnt - 1 || insn[1].code != 0 ||
3901 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
3902 insn[1].off != 0) {
3903 verbose("invalid bpf_ld_imm64 insn\n");
3904 return -EINVAL;
3905 }
3906
3907 if (insn->src_reg == 0)
3908 /* valid generic load 64-bit imm */
3909 goto next_insn;
3910
3911 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
3912 verbose("unrecognized bpf_ld_imm64 insn\n");
3913 return -EINVAL;
3914 }
3915
3916 f = fdget(insn->imm);
3917 map = __bpf_map_get(f);
3918 if (IS_ERR(map)) {
3919 verbose("fd %d is not pointing to valid bpf_map\n",
3920 insn->imm);
3921 return PTR_ERR(map);
3922 }
3923
3924 err = check_map_prog_compatibility(map, env->prog);
3925 if (err) {
3926 fdput(f);
3927 return err;
3928 }
3929
3930 /* store map pointer inside BPF_LD_IMM64 instruction */
3931 insn[0].imm = (u32) (unsigned long) map;
3932 insn[1].imm = ((u64) (unsigned long) map) >> 32;
3933
3934 /* check whether we recorded this map already */
3935 for (j = 0; j < env->used_map_cnt; j++)
3936 if (env->used_maps[j] == map) {
3937 fdput(f);
3938 goto next_insn;
3939 }
3940
3941 if (env->used_map_cnt >= MAX_USED_MAPS) {
3942 fdput(f);
3943 return -E2BIG;
3944 }
3945
3946 /* hold the map. If the program is rejected by verifier,
3947 * the map will be released by release_maps() or it
3948 * will be used by the valid program until it's unloaded
3949 * and all maps are released in free_bpf_prog_info()
3950 */
3951 map = bpf_map_inc(map, false);
3952 if (IS_ERR(map)) {
3953 fdput(f);
3954 return PTR_ERR(map);
3955 }
3956 env->used_maps[env->used_map_cnt++] = map;
3957
3958 fdput(f);
3959 next_insn:
3960 insn++;
3961 i++;
3962 }
3963 }
3964
3965 /* now all pseudo BPF_LD_IMM64 instructions load valid
3966 * 'struct bpf_map *' into a register instead of user map_fd.
3967 * These pointers will be used later by verifier to validate map access.
3968 */
3969 return 0;
3970 }
3971
3972 /* drop refcnt of maps used by the rejected program */
3973 static void release_maps(struct bpf_verifier_env *env)
3974 {
3975 int i;
3976
3977 for (i = 0; i < env->used_map_cnt; i++)
3978 bpf_map_put(env->used_maps[i]);
3979 }
3980
3981 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
3982 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
3983 {
3984 struct bpf_insn *insn = env->prog->insnsi;
3985 int insn_cnt = env->prog->len;
3986 int i;
3987
3988 for (i = 0; i < insn_cnt; i++, insn++)
3989 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
3990 insn->src_reg = 0;
3991 }
3992
3993 /* single env->prog->insni[off] instruction was replaced with the range
3994 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
3995 * [0, off) and [off, end) to new locations, so the patched range stays zero
3996 */
3997 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
3998 u32 off, u32 cnt)
3999 {
4000 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
4001
4002 if (cnt == 1)
4003 return 0;
4004 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4005 if (!new_data)
4006 return -ENOMEM;
4007 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4008 memcpy(new_data + off + cnt - 1, old_data + off,
4009 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4010 env->insn_aux_data = new_data;
4011 vfree(old_data);
4012 return 0;
4013 }
4014
4015 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4016 const struct bpf_insn *patch, u32 len)
4017 {
4018 struct bpf_prog *new_prog;
4019
4020 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4021 if (!new_prog)
4022 return NULL;
4023 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4024 return NULL;
4025 return new_prog;
4026 }
4027
4028 /* convert load instructions that access fields of 'struct __sk_buff'
4029 * into sequence of instructions that access fields of 'struct sk_buff'
4030 */
4031 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4032 {
4033 const struct bpf_verifier_ops *ops = env->prog->aux->ops;
4034 int i, cnt, size, ctx_field_size, delta = 0;
4035 const int insn_cnt = env->prog->len;
4036 struct bpf_insn insn_buf[16], *insn;
4037 struct bpf_prog *new_prog;
4038 enum bpf_access_type type;
4039 bool is_narrower_load;
4040 u32 target_size;
4041
4042 if (ops->gen_prologue) {
4043 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4044 env->prog);
4045 if (cnt >= ARRAY_SIZE(insn_buf)) {
4046 verbose("bpf verifier is misconfigured\n");
4047 return -EINVAL;
4048 } else if (cnt) {
4049 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4050 if (!new_prog)
4051 return -ENOMEM;
4052
4053 env->prog = new_prog;
4054 delta += cnt - 1;
4055 }
4056 }
4057
4058 if (!ops->convert_ctx_access)
4059 return 0;
4060
4061 insn = env->prog->insnsi + delta;
4062
4063 for (i = 0; i < insn_cnt; i++, insn++) {
4064 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4065 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4066 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4067 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4068 type = BPF_READ;
4069 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4070 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4071 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4072 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4073 type = BPF_WRITE;
4074 else
4075 continue;
4076
4077 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4078 continue;
4079
4080 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4081 size = BPF_LDST_BYTES(insn);
4082
4083 /* If the read access is a narrower load of the field,
4084 * convert to a 4/8-byte load, to minimum program type specific
4085 * convert_ctx_access changes. If conversion is successful,
4086 * we will apply proper mask to the result.
4087 */
4088 is_narrower_load = size < ctx_field_size;
4089 if (is_narrower_load) {
4090 u32 off = insn->off;
4091 u8 size_code;
4092
4093 if (type == BPF_WRITE) {
4094 verbose("bpf verifier narrow ctx access misconfigured\n");
4095 return -EINVAL;
4096 }
4097
4098 size_code = BPF_H;
4099 if (ctx_field_size == 4)
4100 size_code = BPF_W;
4101 else if (ctx_field_size == 8)
4102 size_code = BPF_DW;
4103
4104 insn->off = off & ~(ctx_field_size - 1);
4105 insn->code = BPF_LDX | BPF_MEM | size_code;
4106 }
4107
4108 target_size = 0;
4109 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4110 &target_size);
4111 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4112 (ctx_field_size && !target_size)) {
4113 verbose("bpf verifier is misconfigured\n");
4114 return -EINVAL;
4115 }
4116
4117 if (is_narrower_load && size < target_size) {
4118 if (ctx_field_size <= 4)
4119 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4120 (1 << size * 8) - 1);
4121 else
4122 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4123 (1 << size * 8) - 1);
4124 }
4125
4126 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4127 if (!new_prog)
4128 return -ENOMEM;
4129
4130 delta += cnt - 1;
4131
4132 /* keep walking new program and skip insns we just inserted */
4133 env->prog = new_prog;
4134 insn = new_prog->insnsi + i + delta;
4135 }
4136
4137 return 0;
4138 }
4139
4140 /* fixup insn->imm field of bpf_call instructions
4141 * and inline eligible helpers as explicit sequence of BPF instructions
4142 *
4143 * this function is called after eBPF program passed verification
4144 */
4145 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4146 {
4147 struct bpf_prog *prog = env->prog;
4148 struct bpf_insn *insn = prog->insnsi;
4149 const struct bpf_func_proto *fn;
4150 const int insn_cnt = prog->len;
4151 struct bpf_insn insn_buf[16];
4152 struct bpf_prog *new_prog;
4153 struct bpf_map *map_ptr;
4154 int i, cnt, delta = 0;
4155
4156 for (i = 0; i < insn_cnt; i++, insn++) {
4157 if (insn->code != (BPF_JMP | BPF_CALL))
4158 continue;
4159
4160 if (insn->imm == BPF_FUNC_get_route_realm)
4161 prog->dst_needed = 1;
4162 if (insn->imm == BPF_FUNC_get_prandom_u32)
4163 bpf_user_rnd_init_once();
4164 if (insn->imm == BPF_FUNC_tail_call) {
4165 /* If we tail call into other programs, we
4166 * cannot make any assumptions since they can
4167 * be replaced dynamically during runtime in
4168 * the program array.
4169 */
4170 prog->cb_access = 1;
4171 env->prog->aux->stack_depth = MAX_BPF_STACK;
4172
4173 /* mark bpf_tail_call as different opcode to avoid
4174 * conditional branch in the interpeter for every normal
4175 * call and to prevent accidental JITing by JIT compiler
4176 * that doesn't support bpf_tail_call yet
4177 */
4178 insn->imm = 0;
4179 insn->code = BPF_JMP | BPF_TAIL_CALL;
4180 continue;
4181 }
4182
4183 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4184 * handlers are currently limited to 64 bit only.
4185 */
4186 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4187 insn->imm == BPF_FUNC_map_lookup_elem) {
4188 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4189 if (map_ptr == BPF_MAP_PTR_POISON ||
4190 !map_ptr->ops->map_gen_lookup)
4191 goto patch_call_imm;
4192
4193 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4194 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4195 verbose("bpf verifier is misconfigured\n");
4196 return -EINVAL;
4197 }
4198
4199 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4200 cnt);
4201 if (!new_prog)
4202 return -ENOMEM;
4203
4204 delta += cnt - 1;
4205
4206 /* keep walking new program and skip insns we just inserted */
4207 env->prog = prog = new_prog;
4208 insn = new_prog->insnsi + i + delta;
4209 continue;
4210 }
4211
4212 if (insn->imm == BPF_FUNC_redirect_map) {
4213 /* Note, we cannot use prog directly as imm as subsequent
4214 * rewrites would still change the prog pointer. The only
4215 * stable address we can use is aux, which also works with
4216 * prog clones during blinding.
4217 */
4218 u64 addr = (unsigned long)prog->aux;
4219 struct bpf_insn r4_ld[] = {
4220 BPF_LD_IMM64(BPF_REG_4, addr),
4221 *insn,
4222 };
4223 cnt = ARRAY_SIZE(r4_ld);
4224
4225 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4226 if (!new_prog)
4227 return -ENOMEM;
4228
4229 delta += cnt - 1;
4230 env->prog = prog = new_prog;
4231 insn = new_prog->insnsi + i + delta;
4232 }
4233 patch_call_imm:
4234 fn = prog->aux->ops->get_func_proto(insn->imm);
4235 /* all functions that have prototype and verifier allowed
4236 * programs to call them, must be real in-kernel functions
4237 */
4238 if (!fn->func) {
4239 verbose("kernel subsystem misconfigured func %s#%d\n",
4240 func_id_name(insn->imm), insn->imm);
4241 return -EFAULT;
4242 }
4243 insn->imm = fn->func - __bpf_call_base;
4244 }
4245
4246 return 0;
4247 }
4248
4249 static void free_states(struct bpf_verifier_env *env)
4250 {
4251 struct bpf_verifier_state_list *sl, *sln;
4252 int i;
4253
4254 if (!env->explored_states)
4255 return;
4256
4257 for (i = 0; i < env->prog->len; i++) {
4258 sl = env->explored_states[i];
4259
4260 if (sl)
4261 while (sl != STATE_LIST_MARK) {
4262 sln = sl->next;
4263 kfree(sl);
4264 sl = sln;
4265 }
4266 }
4267
4268 kfree(env->explored_states);
4269 }
4270
4271 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4272 {
4273 char __user *log_ubuf = NULL;
4274 struct bpf_verifier_env *env;
4275 int ret = -EINVAL;
4276
4277 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4278 * allocate/free it every time bpf_check() is called
4279 */
4280 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4281 if (!env)
4282 return -ENOMEM;
4283
4284 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4285 (*prog)->len);
4286 ret = -ENOMEM;
4287 if (!env->insn_aux_data)
4288 goto err_free_env;
4289 env->prog = *prog;
4290
4291 /* grab the mutex to protect few globals used by verifier */
4292 mutex_lock(&bpf_verifier_lock);
4293
4294 if (attr->log_level || attr->log_buf || attr->log_size) {
4295 /* user requested verbose verifier output
4296 * and supplied buffer to store the verification trace
4297 */
4298 log_level = attr->log_level;
4299 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
4300 log_size = attr->log_size;
4301 log_len = 0;
4302
4303 ret = -EINVAL;
4304 /* log_* values have to be sane */
4305 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
4306 log_level == 0 || log_ubuf == NULL)
4307 goto err_unlock;
4308
4309 ret = -ENOMEM;
4310 log_buf = vmalloc(log_size);
4311 if (!log_buf)
4312 goto err_unlock;
4313 } else {
4314 log_level = 0;
4315 }
4316
4317 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4318 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4319 env->strict_alignment = true;
4320
4321 ret = replace_map_fd_with_map_ptr(env);
4322 if (ret < 0)
4323 goto skip_full_check;
4324
4325 env->explored_states = kcalloc(env->prog->len,
4326 sizeof(struct bpf_verifier_state_list *),
4327 GFP_USER);
4328 ret = -ENOMEM;
4329 if (!env->explored_states)
4330 goto skip_full_check;
4331
4332 ret = check_cfg(env);
4333 if (ret < 0)
4334 goto skip_full_check;
4335
4336 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4337
4338 ret = do_check(env);
4339
4340 skip_full_check:
4341 while (pop_stack(env, NULL) >= 0);
4342 free_states(env);
4343
4344 if (ret == 0)
4345 /* program is valid, convert *(u32*)(ctx + off) accesses */
4346 ret = convert_ctx_accesses(env);
4347
4348 if (ret == 0)
4349 ret = fixup_bpf_calls(env);
4350
4351 if (log_level && log_len >= log_size - 1) {
4352 BUG_ON(log_len >= log_size);
4353 /* verifier log exceeded user supplied buffer */
4354 ret = -ENOSPC;
4355 /* fall through to return what was recorded */
4356 }
4357
4358 /* copy verifier log back to user space including trailing zero */
4359 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
4360 ret = -EFAULT;
4361 goto free_log_buf;
4362 }
4363
4364 if (ret == 0 && env->used_map_cnt) {
4365 /* if program passed verifier, update used_maps in bpf_prog_info */
4366 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4367 sizeof(env->used_maps[0]),
4368 GFP_KERNEL);
4369
4370 if (!env->prog->aux->used_maps) {
4371 ret = -ENOMEM;
4372 goto free_log_buf;
4373 }
4374
4375 memcpy(env->prog->aux->used_maps, env->used_maps,
4376 sizeof(env->used_maps[0]) * env->used_map_cnt);
4377 env->prog->aux->used_map_cnt = env->used_map_cnt;
4378
4379 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4380 * bpf_ld_imm64 instructions
4381 */
4382 convert_pseudo_ld_imm64(env);
4383 }
4384
4385 free_log_buf:
4386 if (log_level)
4387 vfree(log_buf);
4388 if (!env->prog->aux->used_maps)
4389 /* if we didn't copy map pointers into bpf_prog_info, release
4390 * them now. Otherwise free_bpf_prog_info() will release them.
4391 */
4392 release_maps(env);
4393 *prog = env->prog;
4394 err_unlock:
4395 mutex_unlock(&bpf_verifier_lock);
4396 vfree(env->insn_aux_data);
4397 err_free_env:
4398 kfree(env);
4399 return ret;
4400 }
4401
4402 int bpf_analyzer(struct bpf_prog *prog, const struct bpf_ext_analyzer_ops *ops,
4403 void *priv)
4404 {
4405 struct bpf_verifier_env *env;
4406 int ret;
4407
4408 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4409 if (!env)
4410 return -ENOMEM;
4411
4412 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4413 prog->len);
4414 ret = -ENOMEM;
4415 if (!env->insn_aux_data)
4416 goto err_free_env;
4417 env->prog = prog;
4418 env->analyzer_ops = ops;
4419 env->analyzer_priv = priv;
4420
4421 /* grab the mutex to protect few globals used by verifier */
4422 mutex_lock(&bpf_verifier_lock);
4423
4424 log_level = 0;
4425
4426 env->strict_alignment = false;
4427 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4428 env->strict_alignment = true;
4429
4430 env->explored_states = kcalloc(env->prog->len,
4431 sizeof(struct bpf_verifier_state_list *),
4432 GFP_KERNEL);
4433 ret = -ENOMEM;
4434 if (!env->explored_states)
4435 goto skip_full_check;
4436
4437 ret = check_cfg(env);
4438 if (ret < 0)
4439 goto skip_full_check;
4440
4441 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4442
4443 ret = do_check(env);
4444
4445 skip_full_check:
4446 while (pop_stack(env, NULL) >= 0);
4447 free_states(env);
4448
4449 mutex_unlock(&bpf_verifier_lock);
4450 vfree(env->insn_aux_data);
4451 err_free_env:
4452 kfree(env);
4453 return ret;
4454 }
4455 EXPORT_SYMBOL_GPL(bpf_analyzer);
Ubuntu Hirsute Linux kernel mirror