]> git.ipfire.org Git - thirdparty/kernel/stable.git/blob - kernel/bpf/verifier.c
projects / thirdparty / kernel / stable.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
bpf: don't prune branches when a scalar is replaced with a pointer
[thirdparty/kernel/stable.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 /* The stack spill tracking logic in check_stack_write()
1065 * and check_stack_read() relies on stack accesses being
1066 * aligned.
1067 */
1068 strict = true;
1069 break;
1070 default:
1071 break;
1072 }
1073 return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
1074 }
1075
1076 /* truncate register to smaller size (in bytes)
1077 * must be called with size < BPF_REG_SIZE
1078 */
1079 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1080 {
1081 u64 mask;
1082
1083 /* clear high bits in bit representation */
1084 reg->var_off = tnum_cast(reg->var_off, size);
1085
1086 /* fix arithmetic bounds */
1087 mask = ((u64)1 << (size * 8)) - 1;
1088 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1089 reg->umin_value &= mask;
1090 reg->umax_value &= mask;
1091 } else {
1092 reg->umin_value = 0;
1093 reg->umax_value = mask;
1094 }
1095 reg->smin_value = reg->umin_value;
1096 reg->smax_value = reg->umax_value;
1097 }
1098
1099 /* check whether memory at (regno + off) is accessible for t = (read | write)
1100 * if t==write, value_regno is a register which value is stored into memory
1101 * if t==read, value_regno is a register which will receive the value from memory
1102 * if t==write && value_regno==-1, some unknown value is stored into memory
1103 * if t==read && value_regno==-1, don't care what we read from memory
1104 */
1105 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1106 int bpf_size, enum bpf_access_type t,
1107 int value_regno)
1108 {
1109 struct bpf_verifier_state *state = &env->cur_state;
1110 struct bpf_reg_state *reg = &state->regs[regno];
1111 int size, err = 0;
1112
1113 size = bpf_size_to_bytes(bpf_size);
1114 if (size < 0)
1115 return size;
1116
1117 /* alignment checks will add in reg->off themselves */
1118 err = check_ptr_alignment(env, reg, off, size);
1119 if (err)
1120 return err;
1121
1122 /* for access checks, reg->off is just part of off */
1123 off += reg->off;
1124
1125 if (reg->type == PTR_TO_MAP_VALUE) {
1126 if (t == BPF_WRITE && value_regno >= 0 &&
1127 is_pointer_value(env, value_regno)) {
1128 verbose("R%d leaks addr into map\n", value_regno);
1129 return -EACCES;
1130 }
1131
1132 err = check_map_access(env, regno, off, size);
1133 if (!err && t == BPF_READ && value_regno >= 0)
1134 mark_reg_unknown(state->regs, value_regno);
1135
1136 } else if (reg->type == PTR_TO_CTX) {
1137 enum bpf_reg_type reg_type = SCALAR_VALUE;
1138
1139 if (t == BPF_WRITE && value_regno >= 0 &&
1140 is_pointer_value(env, value_regno)) {
1141 verbose("R%d leaks addr into ctx\n", value_regno);
1142 return -EACCES;
1143 }
1144 /* ctx accesses must be at a fixed offset, so that we can
1145 * determine what type of data were returned.
1146 */
1147 if (reg->off) {
1148 verbose("dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1149 regno, reg->off, off - reg->off);
1150 return -EACCES;
1151 }
1152 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1153 char tn_buf[48];
1154
1155 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1156 verbose("variable ctx access var_off=%s off=%d size=%d",
1157 tn_buf, off, size);
1158 return -EACCES;
1159 }
1160 err = check_ctx_access(env, insn_idx, off, size, t, &reg_type);
1161 if (!err && t == BPF_READ && value_regno >= 0) {
1162 /* ctx access returns either a scalar, or a
1163 * PTR_TO_PACKET[_END]. In the latter case, we know
1164 * the offset is zero.
1165 */
1166 if (reg_type == SCALAR_VALUE)
1167 mark_reg_unknown(state->regs, value_regno);
1168 else
1169 mark_reg_known_zero(state->regs, value_regno);
1170 state->regs[value_regno].id = 0;
1171 state->regs[value_regno].off = 0;
1172 state->regs[value_regno].range = 0;
1173 state->regs[value_regno].type = reg_type;
1174 }
1175
1176 } else if (reg->type == PTR_TO_STACK) {
1177 /* stack accesses must be at a fixed offset, so that we can
1178 * determine what type of data were returned.
1179 * See check_stack_read().
1180 */
1181 if (!tnum_is_const(reg->var_off)) {
1182 char tn_buf[48];
1183
1184 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1185 verbose("variable stack access var_off=%s off=%d size=%d",
1186 tn_buf, off, size);
1187 return -EACCES;
1188 }
1189 off += reg->var_off.value;
1190 if (off >= 0 || off < -MAX_BPF_STACK) {
1191 verbose("invalid stack off=%d size=%d\n", off, size);
1192 return -EACCES;
1193 }
1194
1195 if (env->prog->aux->stack_depth < -off)
1196 env->prog->aux->stack_depth = -off;
1197
1198 if (t == BPF_WRITE) {
1199 if (!env->allow_ptr_leaks &&
1200 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL &&
1201 size != BPF_REG_SIZE) {
1202 verbose("attempt to corrupt spilled pointer on stack\n");
1203 return -EACCES;
1204 }
1205 err = check_stack_write(state, off, size, value_regno);
1206 } else {
1207 err = check_stack_read(state, off, size, value_regno);
1208 }
1209 } else if (reg->type == PTR_TO_PACKET) {
1210 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1211 verbose("cannot write into packet\n");
1212 return -EACCES;
1213 }
1214 if (t == BPF_WRITE && value_regno >= 0 &&
1215 is_pointer_value(env, value_regno)) {
1216 verbose("R%d leaks addr into packet\n", value_regno);
1217 return -EACCES;
1218 }
1219 err = check_packet_access(env, regno, off, size);
1220 if (!err && t == BPF_READ && value_regno >= 0)
1221 mark_reg_unknown(state->regs, value_regno);
1222 } else {
1223 verbose("R%d invalid mem access '%s'\n",
1224 regno, reg_type_str[reg->type]);
1225 return -EACCES;
1226 }
1227
1228 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1229 state->regs[value_regno].type == SCALAR_VALUE) {
1230 /* b/h/w load zero-extends, mark upper bits as known 0 */
1231 coerce_reg_to_size(&state->regs[value_regno], size);
1232 }
1233 return err;
1234 }
1235
1236 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1237 {
1238 int err;
1239
1240 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1241 insn->imm != 0) {
1242 verbose("BPF_XADD uses reserved fields\n");
1243 return -EINVAL;
1244 }
1245
1246 /* check src1 operand */
1247 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1248 if (err)
1249 return err;
1250
1251 /* check src2 operand */
1252 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1253 if (err)
1254 return err;
1255
1256 if (is_pointer_value(env, insn->src_reg)) {
1257 verbose("R%d leaks addr into mem\n", insn->src_reg);
1258 return -EACCES;
1259 }
1260
1261 /* check whether atomic_add can read the memory */
1262 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1263 BPF_SIZE(insn->code), BPF_READ, -1);
1264 if (err)
1265 return err;
1266
1267 /* check whether atomic_add can write into the same memory */
1268 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1269 BPF_SIZE(insn->code), BPF_WRITE, -1);
1270 }
1271
1272 /* Does this register contain a constant zero? */
1273 static bool register_is_null(struct bpf_reg_state reg)
1274 {
1275 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1276 }
1277
1278 /* when register 'regno' is passed into function that will read 'access_size'
1279 * bytes from that pointer, make sure that it's within stack boundary
1280 * and all elements of stack are initialized.
1281 * Unlike most pointer bounds-checking functions, this one doesn't take an
1282 * 'off' argument, so it has to add in reg->off itself.
1283 */
1284 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1285 int access_size, bool zero_size_allowed,
1286 struct bpf_call_arg_meta *meta)
1287 {
1288 struct bpf_verifier_state *state = &env->cur_state;
1289 struct bpf_reg_state *regs = state->regs;
1290 int off, i;
1291
1292 if (regs[regno].type != PTR_TO_STACK) {
1293 /* Allow zero-byte read from NULL, regardless of pointer type */
1294 if (zero_size_allowed && access_size == 0 &&
1295 register_is_null(regs[regno]))
1296 return 0;
1297
1298 verbose("R%d type=%s expected=%s\n", regno,
1299 reg_type_str[regs[regno].type],
1300 reg_type_str[PTR_TO_STACK]);
1301 return -EACCES;
1302 }
1303
1304 /* Only allow fixed-offset stack reads */
1305 if (!tnum_is_const(regs[regno].var_off)) {
1306 char tn_buf[48];
1307
1308 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1309 verbose("invalid variable stack read R%d var_off=%s\n",
1310 regno, tn_buf);
1311 return -EACCES;
1312 }
1313 off = regs[regno].off + regs[regno].var_off.value;
1314 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1315 access_size <= 0) {
1316 verbose("invalid stack type R%d off=%d access_size=%d\n",
1317 regno, off, access_size);
1318 return -EACCES;
1319 }
1320
1321 if (env->prog->aux->stack_depth < -off)
1322 env->prog->aux->stack_depth = -off;
1323
1324 if (meta && meta->raw_mode) {
1325 meta->access_size = access_size;
1326 meta->regno = regno;
1327 return 0;
1328 }
1329
1330 for (i = 0; i < access_size; i++) {
1331 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
1332 verbose("invalid indirect read from stack off %d+%d size %d\n",
1333 off, i, access_size);
1334 return -EACCES;
1335 }
1336 }
1337 return 0;
1338 }
1339
1340 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1341 int access_size, bool zero_size_allowed,
1342 struct bpf_call_arg_meta *meta)
1343 {
1344 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1345
1346 switch (reg->type) {
1347 case PTR_TO_PACKET:
1348 return check_packet_access(env, regno, reg->off, access_size);
1349 case PTR_TO_MAP_VALUE:
1350 return check_map_access(env, regno, reg->off, access_size);
1351 default: /* scalar_value|ptr_to_stack or invalid ptr */
1352 return check_stack_boundary(env, regno, access_size,
1353 zero_size_allowed, meta);
1354 }
1355 }
1356
1357 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1358 enum bpf_arg_type arg_type,
1359 struct bpf_call_arg_meta *meta)
1360 {
1361 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1362 enum bpf_reg_type expected_type, type = reg->type;
1363 int err = 0;
1364
1365 if (arg_type == ARG_DONTCARE)
1366 return 0;
1367
1368 err = check_reg_arg(env, regno, SRC_OP);
1369 if (err)
1370 return err;
1371
1372 if (arg_type == ARG_ANYTHING) {
1373 if (is_pointer_value(env, regno)) {
1374 verbose("R%d leaks addr into helper function\n", regno);
1375 return -EACCES;
1376 }
1377 return 0;
1378 }
1379
1380 if (type == PTR_TO_PACKET &&
1381 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1382 verbose("helper access to the packet is not allowed\n");
1383 return -EACCES;
1384 }
1385
1386 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1387 arg_type == ARG_PTR_TO_MAP_VALUE) {
1388 expected_type = PTR_TO_STACK;
1389 if (type != PTR_TO_PACKET && type != expected_type)
1390 goto err_type;
1391 } else if (arg_type == ARG_CONST_SIZE ||
1392 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1393 expected_type = SCALAR_VALUE;
1394 if (type != expected_type)
1395 goto err_type;
1396 } else if (arg_type == ARG_CONST_MAP_PTR) {
1397 expected_type = CONST_PTR_TO_MAP;
1398 if (type != expected_type)
1399 goto err_type;
1400 } else if (arg_type == ARG_PTR_TO_CTX) {
1401 expected_type = PTR_TO_CTX;
1402 if (type != expected_type)
1403 goto err_type;
1404 } else if (arg_type == ARG_PTR_TO_MEM ||
1405 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1406 expected_type = PTR_TO_STACK;
1407 /* One exception here. In case function allows for NULL to be
1408 * passed in as argument, it's a SCALAR_VALUE type. Final test
1409 * happens during stack boundary checking.
1410 */
1411 if (register_is_null(*reg))
1412 /* final test in check_stack_boundary() */;
1413 else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
1414 type != expected_type)
1415 goto err_type;
1416 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1417 } else {
1418 verbose("unsupported arg_type %d\n", arg_type);
1419 return -EFAULT;
1420 }
1421
1422 if (arg_type == ARG_CONST_MAP_PTR) {
1423 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1424 meta->map_ptr = reg->map_ptr;
1425 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1426 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1427 * check that [key, key + map->key_size) are within
1428 * stack limits and initialized
1429 */
1430 if (!meta->map_ptr) {
1431 /* in function declaration map_ptr must come before
1432 * map_key, so that it's verified and known before
1433 * we have to check map_key here. Otherwise it means
1434 * that kernel subsystem misconfigured verifier
1435 */
1436 verbose("invalid map_ptr to access map->key\n");
1437 return -EACCES;
1438 }
1439 if (type == PTR_TO_PACKET)
1440 err = check_packet_access(env, regno, reg->off,
1441 meta->map_ptr->key_size);
1442 else
1443 err = check_stack_boundary(env, regno,
1444 meta->map_ptr->key_size,
1445 false, NULL);
1446 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1447 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1448 * check [value, value + map->value_size) validity
1449 */
1450 if (!meta->map_ptr) {
1451 /* kernel subsystem misconfigured verifier */
1452 verbose("invalid map_ptr to access map->value\n");
1453 return -EACCES;
1454 }
1455 if (type == PTR_TO_PACKET)
1456 err = check_packet_access(env, regno, reg->off,
1457 meta->map_ptr->value_size);
1458 else
1459 err = check_stack_boundary(env, regno,
1460 meta->map_ptr->value_size,
1461 false, NULL);
1462 } else if (arg_type == ARG_CONST_SIZE ||
1463 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1464 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1465
1466 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1467 * from stack pointer 'buf'. Check it
1468 * note: regno == len, regno - 1 == buf
1469 */
1470 if (regno == 0) {
1471 /* kernel subsystem misconfigured verifier */
1472 verbose("ARG_CONST_SIZE cannot be first argument\n");
1473 return -EACCES;
1474 }
1475
1476 /* The register is SCALAR_VALUE; the access check
1477 * happens using its boundaries.
1478 */
1479
1480 if (!tnum_is_const(reg->var_off))
1481 /* For unprivileged variable accesses, disable raw
1482 * mode so that the program is required to
1483 * initialize all the memory that the helper could
1484 * just partially fill up.
1485 */
1486 meta = NULL;
1487
1488 if (reg->smin_value < 0) {
1489 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1490 regno);
1491 return -EACCES;
1492 }
1493
1494 if (reg->umin_value == 0) {
1495 err = check_helper_mem_access(env, regno - 1, 0,
1496 zero_size_allowed,
1497 meta);
1498 if (err)
1499 return err;
1500 }
1501
1502 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1503 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1504 regno);
1505 return -EACCES;
1506 }
1507 err = check_helper_mem_access(env, regno - 1,
1508 reg->umax_value,
1509 zero_size_allowed, meta);
1510 }
1511
1512 return err;
1513 err_type:
1514 verbose("R%d type=%s expected=%s\n", regno,
1515 reg_type_str[type], reg_type_str[expected_type]);
1516 return -EACCES;
1517 }
1518
1519 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
1520 {
1521 if (!map)
1522 return 0;
1523
1524 /* We need a two way check, first is from map perspective ... */
1525 switch (map->map_type) {
1526 case BPF_MAP_TYPE_PROG_ARRAY:
1527 if (func_id != BPF_FUNC_tail_call)
1528 goto error;
1529 break;
1530 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1531 if (func_id != BPF_FUNC_perf_event_read &&
1532 func_id != BPF_FUNC_perf_event_output)
1533 goto error;
1534 break;
1535 case BPF_MAP_TYPE_STACK_TRACE:
1536 if (func_id != BPF_FUNC_get_stackid)
1537 goto error;
1538 break;
1539 case BPF_MAP_TYPE_CGROUP_ARRAY:
1540 if (func_id != BPF_FUNC_skb_under_cgroup &&
1541 func_id != BPF_FUNC_current_task_under_cgroup)
1542 goto error;
1543 break;
1544 /* devmap returns a pointer to a live net_device ifindex that we cannot
1545 * allow to be modified from bpf side. So do not allow lookup elements
1546 * for now.
1547 */
1548 case BPF_MAP_TYPE_DEVMAP:
1549 if (func_id != BPF_FUNC_redirect_map)
1550 goto error;
1551 break;
1552 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1553 case BPF_MAP_TYPE_HASH_OF_MAPS:
1554 if (func_id != BPF_FUNC_map_lookup_elem)
1555 goto error;
1556 break;
1557 case BPF_MAP_TYPE_SOCKMAP:
1558 if (func_id != BPF_FUNC_sk_redirect_map &&
1559 func_id != BPF_FUNC_sock_map_update &&
1560 func_id != BPF_FUNC_map_delete_elem)
1561 goto error;
1562 break;
1563 default:
1564 break;
1565 }
1566
1567 /* ... and second from the function itself. */
1568 switch (func_id) {
1569 case BPF_FUNC_tail_call:
1570 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1571 goto error;
1572 break;
1573 case BPF_FUNC_perf_event_read:
1574 case BPF_FUNC_perf_event_output:
1575 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1576 goto error;
1577 break;
1578 case BPF_FUNC_get_stackid:
1579 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1580 goto error;
1581 break;
1582 case BPF_FUNC_current_task_under_cgroup:
1583 case BPF_FUNC_skb_under_cgroup:
1584 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1585 goto error;
1586 break;
1587 case BPF_FUNC_redirect_map:
1588 if (map->map_type != BPF_MAP_TYPE_DEVMAP)
1589 goto error;
1590 break;
1591 case BPF_FUNC_sk_redirect_map:
1592 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1593 goto error;
1594 break;
1595 case BPF_FUNC_sock_map_update:
1596 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1597 goto error;
1598 break;
1599 default:
1600 break;
1601 }
1602
1603 return 0;
1604 error:
1605 verbose("cannot pass map_type %d into func %s#%d\n",
1606 map->map_type, func_id_name(func_id), func_id);
1607 return -EINVAL;
1608 }
1609
1610 static int check_raw_mode(const struct bpf_func_proto *fn)
1611 {
1612 int count = 0;
1613
1614 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1615 count++;
1616 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1617 count++;
1618 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1619 count++;
1620 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1621 count++;
1622 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1623 count++;
1624
1625 return count > 1 ? -EINVAL : 0;
1626 }
1627
1628 /* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
1629 * so turn them into unknown SCALAR_VALUE.
1630 */
1631 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1632 {
1633 struct bpf_verifier_state *state = &env->cur_state;
1634 struct bpf_reg_state *regs = state->regs, *reg;
1635 int i;
1636
1637 for (i = 0; i < MAX_BPF_REG; i++)
1638 if (regs[i].type == PTR_TO_PACKET ||
1639 regs[i].type == PTR_TO_PACKET_END)
1640 mark_reg_unknown(regs, i);
1641
1642 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
1643 if (state->stack_slot_type[i] != STACK_SPILL)
1644 continue;
1645 reg = &state->spilled_regs[i / BPF_REG_SIZE];
1646 if (reg->type != PTR_TO_PACKET &&
1647 reg->type != PTR_TO_PACKET_END)
1648 continue;
1649 __mark_reg_unknown(reg);
1650 }
1651 }
1652
1653 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1654 {
1655 struct bpf_verifier_state *state = &env->cur_state;
1656 const struct bpf_func_proto *fn = NULL;
1657 struct bpf_reg_state *regs = state->regs;
1658 struct bpf_call_arg_meta meta;
1659 bool changes_data;
1660 int i, err;
1661
1662 /* find function prototype */
1663 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1664 verbose("invalid func %s#%d\n", func_id_name(func_id), func_id);
1665 return -EINVAL;
1666 }
1667
1668 if (env->prog->aux->ops->get_func_proto)
1669 fn = env->prog->aux->ops->get_func_proto(func_id);
1670
1671 if (!fn) {
1672 verbose("unknown func %s#%d\n", func_id_name(func_id), func_id);
1673 return -EINVAL;
1674 }
1675
1676 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1677 if (!env->prog->gpl_compatible && fn->gpl_only) {
1678 verbose("cannot call GPL only function from proprietary program\n");
1679 return -EINVAL;
1680 }
1681
1682 changes_data = bpf_helper_changes_pkt_data(fn->func);
1683
1684 memset(&meta, 0, sizeof(meta));
1685 meta.pkt_access = fn->pkt_access;
1686
1687 /* We only support one arg being in raw mode at the moment, which
1688 * is sufficient for the helper functions we have right now.
1689 */
1690 err = check_raw_mode(fn);
1691 if (err) {
1692 verbose("kernel subsystem misconfigured func %s#%d\n",
1693 func_id_name(func_id), func_id);
1694 return err;
1695 }
1696
1697 /* check args */
1698 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1699 if (err)
1700 return err;
1701 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1702 if (err)
1703 return err;
1704 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1705 if (err)
1706 return err;
1707 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1708 if (err)
1709 return err;
1710 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1711 if (err)
1712 return err;
1713
1714 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1715 * is inferred from register state.
1716 */
1717 for (i = 0; i < meta.access_size; i++) {
1718 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1719 if (err)
1720 return err;
1721 }
1722
1723 /* reset caller saved regs */
1724 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1725 mark_reg_not_init(regs, caller_saved[i]);
1726 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1727 }
1728
1729 /* update return register (already marked as written above) */
1730 if (fn->ret_type == RET_INTEGER) {
1731 /* sets type to SCALAR_VALUE */
1732 mark_reg_unknown(regs, BPF_REG_0);
1733 } else if (fn->ret_type == RET_VOID) {
1734 regs[BPF_REG_0].type = NOT_INIT;
1735 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1736 struct bpf_insn_aux_data *insn_aux;
1737
1738 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1739 /* There is no offset yet applied, variable or fixed */
1740 mark_reg_known_zero(regs, BPF_REG_0);
1741 regs[BPF_REG_0].off = 0;
1742 /* remember map_ptr, so that check_map_access()
1743 * can check 'value_size' boundary of memory access
1744 * to map element returned from bpf_map_lookup_elem()
1745 */
1746 if (meta.map_ptr == NULL) {
1747 verbose("kernel subsystem misconfigured verifier\n");
1748 return -EINVAL;
1749 }
1750 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1751 regs[BPF_REG_0].id = ++env->id_gen;
1752 insn_aux = &env->insn_aux_data[insn_idx];
1753 if (!insn_aux->map_ptr)
1754 insn_aux->map_ptr = meta.map_ptr;
1755 else if (insn_aux->map_ptr != meta.map_ptr)
1756 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1757 } else {
1758 verbose("unknown return type %d of func %s#%d\n",
1759 fn->ret_type, func_id_name(func_id), func_id);
1760 return -EINVAL;
1761 }
1762
1763 err = check_map_func_compatibility(meta.map_ptr, func_id);
1764 if (err)
1765 return err;
1766
1767 if (changes_data)
1768 clear_all_pkt_pointers(env);
1769 return 0;
1770 }
1771
1772 static bool signed_add_overflows(s64 a, s64 b)
1773 {
1774 /* Do the add in u64, where overflow is well-defined */
1775 s64 res = (s64)((u64)a + (u64)b);
1776
1777 if (b < 0)
1778 return res > a;
1779 return res < a;
1780 }
1781
1782 static bool signed_sub_overflows(s64 a, s64 b)
1783 {
1784 /* Do the sub in u64, where overflow is well-defined */
1785 s64 res = (s64)((u64)a - (u64)b);
1786
1787 if (b < 0)
1788 return res < a;
1789 return res > a;
1790 }
1791
1792 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1793 * Caller should also handle BPF_MOV case separately.
1794 * If we return -EACCES, caller may want to try again treating pointer as a
1795 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1796 */
1797 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1798 struct bpf_insn *insn,
1799 const struct bpf_reg_state *ptr_reg,
1800 const struct bpf_reg_state *off_reg)
1801 {
1802 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
1803 bool known = tnum_is_const(off_reg->var_off);
1804 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1805 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1806 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1807 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1808 u8 opcode = BPF_OP(insn->code);
1809 u32 dst = insn->dst_reg;
1810
1811 dst_reg = &regs[dst];
1812
1813 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1814 print_verifier_state(&env->cur_state);
1815 verbose("verifier internal error: known but bad sbounds\n");
1816 return -EINVAL;
1817 }
1818 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1819 print_verifier_state(&env->cur_state);
1820 verbose("verifier internal error: known but bad ubounds\n");
1821 return -EINVAL;
1822 }
1823
1824 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1825 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1826 if (!env->allow_ptr_leaks)
1827 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1828 dst);
1829 return -EACCES;
1830 }
1831
1832 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1833 if (!env->allow_ptr_leaks)
1834 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1835 dst);
1836 return -EACCES;
1837 }
1838 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1839 if (!env->allow_ptr_leaks)
1840 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1841 dst);
1842 return -EACCES;
1843 }
1844 if (ptr_reg->type == PTR_TO_PACKET_END) {
1845 if (!env->allow_ptr_leaks)
1846 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1847 dst);
1848 return -EACCES;
1849 }
1850
1851 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1852 * The id may be overwritten later if we create a new variable offset.
1853 */
1854 dst_reg->type = ptr_reg->type;
1855 dst_reg->id = ptr_reg->id;
1856
1857 switch (opcode) {
1858 case BPF_ADD:
1859 /* We can take a fixed offset as long as it doesn't overflow
1860 * the s32 'off' field
1861 */
1862 if (known && (ptr_reg->off + smin_val ==
1863 (s64)(s32)(ptr_reg->off + smin_val))) {
1864 /* pointer += K. Accumulate it into fixed offset */
1865 dst_reg->smin_value = smin_ptr;
1866 dst_reg->smax_value = smax_ptr;
1867 dst_reg->umin_value = umin_ptr;
1868 dst_reg->umax_value = umax_ptr;
1869 dst_reg->var_off = ptr_reg->var_off;
1870 dst_reg->off = ptr_reg->off + smin_val;
1871 dst_reg->range = ptr_reg->range;
1872 break;
1873 }
1874 /* A new variable offset is created. Note that off_reg->off
1875 * == 0, since it's a scalar.
1876 * dst_reg gets the pointer type and since some positive
1877 * integer value was added to the pointer, give it a new 'id'
1878 * if it's a PTR_TO_PACKET.
1879 * this creates a new 'base' pointer, off_reg (variable) gets
1880 * added into the variable offset, and we copy the fixed offset
1881 * from ptr_reg.
1882 */
1883 if (signed_add_overflows(smin_ptr, smin_val) ||
1884 signed_add_overflows(smax_ptr, smax_val)) {
1885 dst_reg->smin_value = S64_MIN;
1886 dst_reg->smax_value = S64_MAX;
1887 } else {
1888 dst_reg->smin_value = smin_ptr + smin_val;
1889 dst_reg->smax_value = smax_ptr + smax_val;
1890 }
1891 if (umin_ptr + umin_val < umin_ptr ||
1892 umax_ptr + umax_val < umax_ptr) {
1893 dst_reg->umin_value = 0;
1894 dst_reg->umax_value = U64_MAX;
1895 } else {
1896 dst_reg->umin_value = umin_ptr + umin_val;
1897 dst_reg->umax_value = umax_ptr + umax_val;
1898 }
1899 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1900 dst_reg->off = ptr_reg->off;
1901 if (ptr_reg->type == PTR_TO_PACKET) {
1902 dst_reg->id = ++env->id_gen;
1903 /* something was added to pkt_ptr, set range to zero */
1904 dst_reg->range = 0;
1905 }
1906 break;
1907 case BPF_SUB:
1908 if (dst_reg == off_reg) {
1909 /* scalar -= pointer. Creates an unknown scalar */
1910 if (!env->allow_ptr_leaks)
1911 verbose("R%d tried to subtract pointer from scalar\n",
1912 dst);
1913 return -EACCES;
1914 }
1915 /* We don't allow subtraction from FP, because (according to
1916 * test_verifier.c test "invalid fp arithmetic", JITs might not
1917 * be able to deal with it.
1918 */
1919 if (ptr_reg->type == PTR_TO_STACK) {
1920 if (!env->allow_ptr_leaks)
1921 verbose("R%d subtraction from stack pointer prohibited\n",
1922 dst);
1923 return -EACCES;
1924 }
1925 if (known && (ptr_reg->off - smin_val ==
1926 (s64)(s32)(ptr_reg->off - smin_val))) {
1927 /* pointer -= K. Subtract it from fixed offset */
1928 dst_reg->smin_value = smin_ptr;
1929 dst_reg->smax_value = smax_ptr;
1930 dst_reg->umin_value = umin_ptr;
1931 dst_reg->umax_value = umax_ptr;
1932 dst_reg->var_off = ptr_reg->var_off;
1933 dst_reg->id = ptr_reg->id;
1934 dst_reg->off = ptr_reg->off - smin_val;
1935 dst_reg->range = ptr_reg->range;
1936 break;
1937 }
1938 /* A new variable offset is created. If the subtrahend is known
1939 * nonnegative, then any reg->range we had before is still good.
1940 */
1941 if (signed_sub_overflows(smin_ptr, smax_val) ||
1942 signed_sub_overflows(smax_ptr, smin_val)) {
1943 /* Overflow possible, we know nothing */
1944 dst_reg->smin_value = S64_MIN;
1945 dst_reg->smax_value = S64_MAX;
1946 } else {
1947 dst_reg->smin_value = smin_ptr - smax_val;
1948 dst_reg->smax_value = smax_ptr - smin_val;
1949 }
1950 if (umin_ptr < umax_val) {
1951 /* Overflow possible, we know nothing */
1952 dst_reg->umin_value = 0;
1953 dst_reg->umax_value = U64_MAX;
1954 } else {
1955 /* Cannot overflow (as long as bounds are consistent) */
1956 dst_reg->umin_value = umin_ptr - umax_val;
1957 dst_reg->umax_value = umax_ptr - umin_val;
1958 }
1959 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
1960 dst_reg->off = ptr_reg->off;
1961 if (ptr_reg->type == PTR_TO_PACKET) {
1962 dst_reg->id = ++env->id_gen;
1963 /* something was added to pkt_ptr, set range to zero */
1964 if (smin_val < 0)
1965 dst_reg->range = 0;
1966 }
1967 break;
1968 case BPF_AND:
1969 case BPF_OR:
1970 case BPF_XOR:
1971 /* bitwise ops on pointers are troublesome, prohibit for now.
1972 * (However, in principle we could allow some cases, e.g.
1973 * ptr &= ~3 which would reduce min_value by 3.)
1974 */
1975 if (!env->allow_ptr_leaks)
1976 verbose("R%d bitwise operator %s on pointer prohibited\n",
1977 dst, bpf_alu_string[opcode >> 4]);
1978 return -EACCES;
1979 default:
1980 /* other operators (e.g. MUL,LSH) produce non-pointer results */
1981 if (!env->allow_ptr_leaks)
1982 verbose("R%d pointer arithmetic with %s operator prohibited\n",
1983 dst, bpf_alu_string[opcode >> 4]);
1984 return -EACCES;
1985 }
1986
1987 __update_reg_bounds(dst_reg);
1988 __reg_deduce_bounds(dst_reg);
1989 __reg_bound_offset(dst_reg);
1990 return 0;
1991 }
1992
1993 /* WARNING: This function does calculations on 64-bit values, but the actual
1994 * execution may occur on 32-bit values. Therefore, things like bitshifts
1995 * need extra checks in the 32-bit case.
1996 */
1997 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
1998 struct bpf_insn *insn,
1999 struct bpf_reg_state *dst_reg,
2000 struct bpf_reg_state src_reg)
2001 {
2002 struct bpf_reg_state *regs = env->cur_state.regs;
2003 u8 opcode = BPF_OP(insn->code);
2004 bool src_known, dst_known;
2005 s64 smin_val, smax_val;
2006 u64 umin_val, umax_val;
2007 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2008
2009 smin_val = src_reg.smin_value;
2010 smax_val = src_reg.smax_value;
2011 umin_val = src_reg.umin_value;
2012 umax_val = src_reg.umax_value;
2013 src_known = tnum_is_const(src_reg.var_off);
2014 dst_known = tnum_is_const(dst_reg->var_off);
2015
2016 switch (opcode) {
2017 case BPF_ADD:
2018 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2019 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2020 dst_reg->smin_value = S64_MIN;
2021 dst_reg->smax_value = S64_MAX;
2022 } else {
2023 dst_reg->smin_value += smin_val;
2024 dst_reg->smax_value += smax_val;
2025 }
2026 if (dst_reg->umin_value + umin_val < umin_val ||
2027 dst_reg->umax_value + umax_val < umax_val) {
2028 dst_reg->umin_value = 0;
2029 dst_reg->umax_value = U64_MAX;
2030 } else {
2031 dst_reg->umin_value += umin_val;
2032 dst_reg->umax_value += umax_val;
2033 }
2034 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2035 break;
2036 case BPF_SUB:
2037 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2038 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2039 /* Overflow possible, we know nothing */
2040 dst_reg->smin_value = S64_MIN;
2041 dst_reg->smax_value = S64_MAX;
2042 } else {
2043 dst_reg->smin_value -= smax_val;
2044 dst_reg->smax_value -= smin_val;
2045 }
2046 if (dst_reg->umin_value < umax_val) {
2047 /* Overflow possible, we know nothing */
2048 dst_reg->umin_value = 0;
2049 dst_reg->umax_value = U64_MAX;
2050 } else {
2051 /* Cannot overflow (as long as bounds are consistent) */
2052 dst_reg->umin_value -= umax_val;
2053 dst_reg->umax_value -= umin_val;
2054 }
2055 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2056 break;
2057 case BPF_MUL:
2058 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2059 if (smin_val < 0 || dst_reg->smin_value < 0) {
2060 /* Ain't nobody got time to multiply that sign */
2061 __mark_reg_unbounded(dst_reg);
2062 __update_reg_bounds(dst_reg);
2063 break;
2064 }
2065 /* Both values are positive, so we can work with unsigned and
2066 * copy the result to signed (unless it exceeds S64_MAX).
2067 */
2068 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2069 /* Potential overflow, we know nothing */
2070 __mark_reg_unbounded(dst_reg);
2071 /* (except what we can learn from the var_off) */
2072 __update_reg_bounds(dst_reg);
2073 break;
2074 }
2075 dst_reg->umin_value *= umin_val;
2076 dst_reg->umax_value *= umax_val;
2077 if (dst_reg->umax_value > S64_MAX) {
2078 /* Overflow possible, we know nothing */
2079 dst_reg->smin_value = S64_MIN;
2080 dst_reg->smax_value = S64_MAX;
2081 } else {
2082 dst_reg->smin_value = dst_reg->umin_value;
2083 dst_reg->smax_value = dst_reg->umax_value;
2084 }
2085 break;
2086 case BPF_AND:
2087 if (src_known && dst_known) {
2088 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2089 src_reg.var_off.value);
2090 break;
2091 }
2092 /* We get our minimum from the var_off, since that's inherently
2093 * bitwise. Our maximum is the minimum of the operands' maxima.
2094 */
2095 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2096 dst_reg->umin_value = dst_reg->var_off.value;
2097 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2098 if (dst_reg->smin_value < 0 || smin_val < 0) {
2099 /* Lose signed bounds when ANDing negative numbers,
2100 * ain't nobody got time for that.
2101 */
2102 dst_reg->smin_value = S64_MIN;
2103 dst_reg->smax_value = S64_MAX;
2104 } else {
2105 /* ANDing two positives gives a positive, so safe to
2106 * cast result into s64.
2107 */
2108 dst_reg->smin_value = dst_reg->umin_value;
2109 dst_reg->smax_value = dst_reg->umax_value;
2110 }
2111 /* We may learn something more from the var_off */
2112 __update_reg_bounds(dst_reg);
2113 break;
2114 case BPF_OR:
2115 if (src_known && dst_known) {
2116 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2117 src_reg.var_off.value);
2118 break;
2119 }
2120 /* We get our maximum from the var_off, and our minimum is the
2121 * maximum of the operands' minima
2122 */
2123 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2124 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2125 dst_reg->umax_value = dst_reg->var_off.value |
2126 dst_reg->var_off.mask;
2127 if (dst_reg->smin_value < 0 || smin_val < 0) {
2128 /* Lose signed bounds when ORing negative numbers,
2129 * ain't nobody got time for that.
2130 */
2131 dst_reg->smin_value = S64_MIN;
2132 dst_reg->smax_value = S64_MAX;
2133 } else {
2134 /* ORing two positives gives a positive, so safe to
2135 * cast result into s64.
2136 */
2137 dst_reg->smin_value = dst_reg->umin_value;
2138 dst_reg->smax_value = dst_reg->umax_value;
2139 }
2140 /* We may learn something more from the var_off */
2141 __update_reg_bounds(dst_reg);
2142 break;
2143 case BPF_LSH:
2144 if (umax_val >= insn_bitness) {
2145 /* Shifts greater than 31 or 63 are undefined.
2146 * This includes shifts by a negative number.
2147 */
2148 mark_reg_unknown(regs, insn->dst_reg);
2149 break;
2150 }
2151 /* We lose all sign bit information (except what we can pick
2152 * up from var_off)
2153 */
2154 dst_reg->smin_value = S64_MIN;
2155 dst_reg->smax_value = S64_MAX;
2156 /* If we might shift our top bit out, then we know nothing */
2157 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2158 dst_reg->umin_value = 0;
2159 dst_reg->umax_value = U64_MAX;
2160 } else {
2161 dst_reg->umin_value <<= umin_val;
2162 dst_reg->umax_value <<= umax_val;
2163 }
2164 if (src_known)
2165 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2166 else
2167 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2168 /* We may learn something more from the var_off */
2169 __update_reg_bounds(dst_reg);
2170 break;
2171 case BPF_RSH:
2172 if (umax_val >= insn_bitness) {
2173 /* Shifts greater than 31 or 63 are undefined.
2174 * This includes shifts by a negative number.
2175 */
2176 mark_reg_unknown(regs, insn->dst_reg);
2177 break;
2178 }
2179 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2180 * be negative, then either:
2181 * 1) src_reg might be zero, so the sign bit of the result is
2182 * unknown, so we lose our signed bounds
2183 * 2) it's known negative, thus the unsigned bounds capture the
2184 * signed bounds
2185 * 3) the signed bounds cross zero, so they tell us nothing
2186 * about the result
2187 * If the value in dst_reg is known nonnegative, then again the
2188 * unsigned bounts capture the signed bounds.
2189 * Thus, in all cases it suffices to blow away our signed bounds
2190 * and rely on inferring new ones from the unsigned bounds and
2191 * var_off of the result.
2192 */
2193 dst_reg->smin_value = S64_MIN;
2194 dst_reg->smax_value = S64_MAX;
2195 if (src_known)
2196 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2197 umin_val);
2198 else
2199 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2200 dst_reg->umin_value >>= umax_val;
2201 dst_reg->umax_value >>= umin_val;
2202 /* We may learn something more from the var_off */
2203 __update_reg_bounds(dst_reg);
2204 break;
2205 default:
2206 mark_reg_unknown(regs, insn->dst_reg);
2207 break;
2208 }
2209
2210 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2211 /* 32-bit ALU ops are (32,32)->32 */
2212 coerce_reg_to_size(dst_reg, 4);
2213 coerce_reg_to_size(&src_reg, 4);
2214 }
2215
2216 __reg_deduce_bounds(dst_reg);
2217 __reg_bound_offset(dst_reg);
2218 return 0;
2219 }
2220
2221 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2222 * and var_off.
2223 */
2224 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2225 struct bpf_insn *insn)
2226 {
2227 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
2228 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2229 u8 opcode = BPF_OP(insn->code);
2230 int rc;
2231
2232 dst_reg = &regs[insn->dst_reg];
2233 src_reg = NULL;
2234 if (dst_reg->type != SCALAR_VALUE)
2235 ptr_reg = dst_reg;
2236 if (BPF_SRC(insn->code) == BPF_X) {
2237 src_reg = &regs[insn->src_reg];
2238 if (src_reg->type != SCALAR_VALUE) {
2239 if (dst_reg->type != SCALAR_VALUE) {
2240 /* Combining two pointers by any ALU op yields
2241 * an arbitrary scalar.
2242 */
2243 if (!env->allow_ptr_leaks) {
2244 verbose("R%d pointer %s pointer prohibited\n",
2245 insn->dst_reg,
2246 bpf_alu_string[opcode >> 4]);
2247 return -EACCES;
2248 }
2249 mark_reg_unknown(regs, insn->dst_reg);
2250 return 0;
2251 } else {
2252 /* scalar += pointer
2253 * This is legal, but we have to reverse our
2254 * src/dest handling in computing the range
2255 */
2256 rc = adjust_ptr_min_max_vals(env, insn,
2257 src_reg, dst_reg);
2258 if (rc == -EACCES && env->allow_ptr_leaks) {
2259 /* scalar += unknown scalar */
2260 __mark_reg_unknown(&off_reg);
2261 return adjust_scalar_min_max_vals(
2262 env, insn,
2263 dst_reg, off_reg);
2264 }
2265 return rc;
2266 }
2267 } else if (ptr_reg) {
2268 /* pointer += scalar */
2269 rc = adjust_ptr_min_max_vals(env, insn,
2270 dst_reg, src_reg);
2271 if (rc == -EACCES && env->allow_ptr_leaks) {
2272 /* unknown scalar += scalar */
2273 __mark_reg_unknown(dst_reg);
2274 return adjust_scalar_min_max_vals(
2275 env, insn, dst_reg, *src_reg);
2276 }
2277 return rc;
2278 }
2279 } else {
2280 /* Pretend the src is a reg with a known value, since we only
2281 * need to be able to read from this state.
2282 */
2283 off_reg.type = SCALAR_VALUE;
2284 __mark_reg_known(&off_reg, insn->imm);
2285 src_reg = &off_reg;
2286 if (ptr_reg) { /* pointer += K */
2287 rc = adjust_ptr_min_max_vals(env, insn,
2288 ptr_reg, src_reg);
2289 if (rc == -EACCES && env->allow_ptr_leaks) {
2290 /* unknown scalar += K */
2291 __mark_reg_unknown(dst_reg);
2292 return adjust_scalar_min_max_vals(
2293 env, insn, dst_reg, off_reg);
2294 }
2295 return rc;
2296 }
2297 }
2298
2299 /* Got here implies adding two SCALAR_VALUEs */
2300 if (WARN_ON_ONCE(ptr_reg)) {
2301 print_verifier_state(&env->cur_state);
2302 verbose("verifier internal error: unexpected ptr_reg\n");
2303 return -EINVAL;
2304 }
2305 if (WARN_ON(!src_reg)) {
2306 print_verifier_state(&env->cur_state);
2307 verbose("verifier internal error: no src_reg\n");
2308 return -EINVAL;
2309 }
2310 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2311 }
2312
2313 /* check validity of 32-bit and 64-bit arithmetic operations */
2314 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2315 {
2316 struct bpf_reg_state *regs = env->cur_state.regs;
2317 u8 opcode = BPF_OP(insn->code);
2318 int err;
2319
2320 if (opcode == BPF_END || opcode == BPF_NEG) {
2321 if (opcode == BPF_NEG) {
2322 if (BPF_SRC(insn->code) != 0 ||
2323 insn->src_reg != BPF_REG_0 ||
2324 insn->off != 0 || insn->imm != 0) {
2325 verbose("BPF_NEG uses reserved fields\n");
2326 return -EINVAL;
2327 }
2328 } else {
2329 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2330 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2331 BPF_CLASS(insn->code) == BPF_ALU64) {
2332 verbose("BPF_END uses reserved fields\n");
2333 return -EINVAL;
2334 }
2335 }
2336
2337 /* check src operand */
2338 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2339 if (err)
2340 return err;
2341
2342 if (is_pointer_value(env, insn->dst_reg)) {
2343 verbose("R%d pointer arithmetic prohibited\n",
2344 insn->dst_reg);
2345 return -EACCES;
2346 }
2347
2348 /* check dest operand */
2349 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2350 if (err)
2351 return err;
2352
2353 } else if (opcode == BPF_MOV) {
2354
2355 if (BPF_SRC(insn->code) == BPF_X) {
2356 if (insn->imm != 0 || insn->off != 0) {
2357 verbose("BPF_MOV uses reserved fields\n");
2358 return -EINVAL;
2359 }
2360
2361 /* check src operand */
2362 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2363 if (err)
2364 return err;
2365 } else {
2366 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2367 verbose("BPF_MOV uses reserved fields\n");
2368 return -EINVAL;
2369 }
2370 }
2371
2372 /* check dest operand */
2373 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2374 if (err)
2375 return err;
2376
2377 if (BPF_SRC(insn->code) == BPF_X) {
2378 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2379 /* case: R1 = R2
2380 * copy register state to dest reg
2381 */
2382 regs[insn->dst_reg] = regs[insn->src_reg];
2383 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
2384 } else {
2385 /* R1 = (u32) R2 */
2386 if (is_pointer_value(env, insn->src_reg)) {
2387 verbose("R%d partial copy of pointer\n",
2388 insn->src_reg);
2389 return -EACCES;
2390 }
2391 mark_reg_unknown(regs, insn->dst_reg);
2392 coerce_reg_to_size(&regs[insn->dst_reg], 4);
2393 }
2394 } else {
2395 /* case: R = imm
2396 * remember the value we stored into this reg
2397 */
2398 regs[insn->dst_reg].type = SCALAR_VALUE;
2399 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2400 __mark_reg_known(regs + insn->dst_reg,
2401 insn->imm);
2402 } else {
2403 __mark_reg_known(regs + insn->dst_reg,
2404 (u32)insn->imm);
2405 }
2406 }
2407
2408 } else if (opcode > BPF_END) {
2409 verbose("invalid BPF_ALU opcode %x\n", opcode);
2410 return -EINVAL;
2411
2412 } else { /* all other ALU ops: and, sub, xor, add, ... */
2413
2414 if (BPF_SRC(insn->code) == BPF_X) {
2415 if (insn->imm != 0 || insn->off != 0) {
2416 verbose("BPF_ALU uses reserved fields\n");
2417 return -EINVAL;
2418 }
2419 /* check src1 operand */
2420 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2421 if (err)
2422 return err;
2423 } else {
2424 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2425 verbose("BPF_ALU uses reserved fields\n");
2426 return -EINVAL;
2427 }
2428 }
2429
2430 /* check src2 operand */
2431 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2432 if (err)
2433 return err;
2434
2435 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2436 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2437 verbose("div by zero\n");
2438 return -EINVAL;
2439 }
2440
2441 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2442 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2443 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2444
2445 if (insn->imm < 0 || insn->imm >= size) {
2446 verbose("invalid shift %d\n", insn->imm);
2447 return -EINVAL;
2448 }
2449 }
2450
2451 /* check dest operand */
2452 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2453 if (err)
2454 return err;
2455
2456 return adjust_reg_min_max_vals(env, insn);
2457 }
2458
2459 return 0;
2460 }
2461
2462 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2463 struct bpf_reg_state *dst_reg,
2464 bool range_right_open)
2465 {
2466 struct bpf_reg_state *regs = state->regs, *reg;
2467 u16 new_range;
2468 int i;
2469
2470 if (dst_reg->off < 0 ||
2471 (dst_reg->off == 0 && range_right_open))
2472 /* This doesn't give us any range */
2473 return;
2474
2475 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2476 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2477 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2478 * than pkt_end, but that's because it's also less than pkt.
2479 */
2480 return;
2481
2482 new_range = dst_reg->off;
2483 if (range_right_open)
2484 new_range--;
2485
2486 /* Examples for register markings:
2487 *
2488 * pkt_data in dst register:
2489 *
2490 * r2 = r3;
2491 * r2 += 8;
2492 * if (r2 > pkt_end) goto <handle exception>
2493 * <access okay>
2494 *
2495 * r2 = r3;
2496 * r2 += 8;
2497 * if (r2 < pkt_end) goto <access okay>
2498 * <handle exception>
2499 *
2500 * Where:
2501 * r2 == dst_reg, pkt_end == src_reg
2502 * r2=pkt(id=n,off=8,r=0)
2503 * r3=pkt(id=n,off=0,r=0)
2504 *
2505 * pkt_data in src register:
2506 *
2507 * r2 = r3;
2508 * r2 += 8;
2509 * if (pkt_end >= r2) goto <access okay>
2510 * <handle exception>
2511 *
2512 * r2 = r3;
2513 * r2 += 8;
2514 * if (pkt_end <= r2) goto <handle exception>
2515 * <access okay>
2516 *
2517 * Where:
2518 * pkt_end == dst_reg, r2 == src_reg
2519 * r2=pkt(id=n,off=8,r=0)
2520 * r3=pkt(id=n,off=0,r=0)
2521 *
2522 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2523 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2524 * and [r3, r3 + 8-1) respectively is safe to access depending on
2525 * the check.
2526 */
2527
2528 /* If our ids match, then we must have the same max_value. And we
2529 * don't care about the other reg's fixed offset, since if it's too big
2530 * the range won't allow anything.
2531 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2532 */
2533 for (i = 0; i < MAX_BPF_REG; i++)
2534 if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
2535 /* keep the maximum range already checked */
2536 regs[i].range = max(regs[i].range, new_range);
2537
2538 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2539 if (state->stack_slot_type[i] != STACK_SPILL)
2540 continue;
2541 reg = &state->spilled_regs[i / BPF_REG_SIZE];
2542 if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
2543 reg->range = max(reg->range, new_range);
2544 }
2545 }
2546
2547 /* Adjusts the register min/max values in the case that the dst_reg is the
2548 * variable register that we are working on, and src_reg is a constant or we're
2549 * simply doing a BPF_K check.
2550 * In JEQ/JNE cases we also adjust the var_off values.
2551 */
2552 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2553 struct bpf_reg_state *false_reg, u64 val,
2554 u8 opcode)
2555 {
2556 /* If the dst_reg is a pointer, we can't learn anything about its
2557 * variable offset from the compare (unless src_reg were a pointer into
2558 * the same object, but we don't bother with that.
2559 * Since false_reg and true_reg have the same type by construction, we
2560 * only need to check one of them for pointerness.
2561 */
2562 if (__is_pointer_value(false, false_reg))
2563 return;
2564
2565 switch (opcode) {
2566 case BPF_JEQ:
2567 /* If this is false then we know nothing Jon Snow, but if it is
2568 * true then we know for sure.
2569 */
2570 __mark_reg_known(true_reg, val);
2571 break;
2572 case BPF_JNE:
2573 /* If this is true we know nothing Jon Snow, but if it is false
2574 * we know the value for sure;
2575 */
2576 __mark_reg_known(false_reg, val);
2577 break;
2578 case BPF_JGT:
2579 false_reg->umax_value = min(false_reg->umax_value, val);
2580 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2581 break;
2582 case BPF_JSGT:
2583 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2584 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2585 break;
2586 case BPF_JLT:
2587 false_reg->umin_value = max(false_reg->umin_value, val);
2588 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2589 break;
2590 case BPF_JSLT:
2591 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2592 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2593 break;
2594 case BPF_JGE:
2595 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2596 true_reg->umin_value = max(true_reg->umin_value, val);
2597 break;
2598 case BPF_JSGE:
2599 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2600 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2601 break;
2602 case BPF_JLE:
2603 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2604 true_reg->umax_value = min(true_reg->umax_value, val);
2605 break;
2606 case BPF_JSLE:
2607 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2608 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2609 break;
2610 default:
2611 break;
2612 }
2613
2614 __reg_deduce_bounds(false_reg);
2615 __reg_deduce_bounds(true_reg);
2616 /* We might have learned some bits from the bounds. */
2617 __reg_bound_offset(false_reg);
2618 __reg_bound_offset(true_reg);
2619 /* Intersecting with the old var_off might have improved our bounds
2620 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2621 * then new var_off is (0; 0x7f...fc) which improves our umax.
2622 */
2623 __update_reg_bounds(false_reg);
2624 __update_reg_bounds(true_reg);
2625 }
2626
2627 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2628 * the variable reg.
2629 */
2630 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2631 struct bpf_reg_state *false_reg, u64 val,
2632 u8 opcode)
2633 {
2634 if (__is_pointer_value(false, false_reg))
2635 return;
2636
2637 switch (opcode) {
2638 case BPF_JEQ:
2639 /* If this is false then we know nothing Jon Snow, but if it is
2640 * true then we know for sure.
2641 */
2642 __mark_reg_known(true_reg, val);
2643 break;
2644 case BPF_JNE:
2645 /* If this is true we know nothing Jon Snow, but if it is false
2646 * we know the value for sure;
2647 */
2648 __mark_reg_known(false_reg, val);
2649 break;
2650 case BPF_JGT:
2651 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2652 false_reg->umin_value = max(false_reg->umin_value, val);
2653 break;
2654 case BPF_JSGT:
2655 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2656 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2657 break;
2658 case BPF_JLT:
2659 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2660 false_reg->umax_value = min(false_reg->umax_value, val);
2661 break;
2662 case BPF_JSLT:
2663 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2664 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2665 break;
2666 case BPF_JGE:
2667 true_reg->umax_value = min(true_reg->umax_value, val);
2668 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2669 break;
2670 case BPF_JSGE:
2671 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2672 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2673 break;
2674 case BPF_JLE:
2675 true_reg->umin_value = max(true_reg->umin_value, val);
2676 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2677 break;
2678 case BPF_JSLE:
2679 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2680 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2681 break;
2682 default:
2683 break;
2684 }
2685
2686 __reg_deduce_bounds(false_reg);
2687 __reg_deduce_bounds(true_reg);
2688 /* We might have learned some bits from the bounds. */
2689 __reg_bound_offset(false_reg);
2690 __reg_bound_offset(true_reg);
2691 /* Intersecting with the old var_off might have improved our bounds
2692 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2693 * then new var_off is (0; 0x7f...fc) which improves our umax.
2694 */
2695 __update_reg_bounds(false_reg);
2696 __update_reg_bounds(true_reg);
2697 }
2698
2699 /* Regs are known to be equal, so intersect their min/max/var_off */
2700 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2701 struct bpf_reg_state *dst_reg)
2702 {
2703 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2704 dst_reg->umin_value);
2705 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2706 dst_reg->umax_value);
2707 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2708 dst_reg->smin_value);
2709 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2710 dst_reg->smax_value);
2711 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2712 dst_reg->var_off);
2713 /* We might have learned new bounds from the var_off. */
2714 __update_reg_bounds(src_reg);
2715 __update_reg_bounds(dst_reg);
2716 /* We might have learned something about the sign bit. */
2717 __reg_deduce_bounds(src_reg);
2718 __reg_deduce_bounds(dst_reg);
2719 /* We might have learned some bits from the bounds. */
2720 __reg_bound_offset(src_reg);
2721 __reg_bound_offset(dst_reg);
2722 /* Intersecting with the old var_off might have improved our bounds
2723 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2724 * then new var_off is (0; 0x7f...fc) which improves our umax.
2725 */
2726 __update_reg_bounds(src_reg);
2727 __update_reg_bounds(dst_reg);
2728 }
2729
2730 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2731 struct bpf_reg_state *true_dst,
2732 struct bpf_reg_state *false_src,
2733 struct bpf_reg_state *false_dst,
2734 u8 opcode)
2735 {
2736 switch (opcode) {
2737 case BPF_JEQ:
2738 __reg_combine_min_max(true_src, true_dst);
2739 break;
2740 case BPF_JNE:
2741 __reg_combine_min_max(false_src, false_dst);
2742 break;
2743 }
2744 }
2745
2746 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2747 bool is_null)
2748 {
2749 struct bpf_reg_state *reg = &regs[regno];
2750
2751 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2752 /* Old offset (both fixed and variable parts) should
2753 * have been known-zero, because we don't allow pointer
2754 * arithmetic on pointers that might be NULL.
2755 */
2756 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2757 !tnum_equals_const(reg->var_off, 0) ||
2758 reg->off)) {
2759 __mark_reg_known_zero(reg);
2760 reg->off = 0;
2761 }
2762 if (is_null) {
2763 reg->type = SCALAR_VALUE;
2764 } else if (reg->map_ptr->inner_map_meta) {
2765 reg->type = CONST_PTR_TO_MAP;
2766 reg->map_ptr = reg->map_ptr->inner_map_meta;
2767 } else {
2768 reg->type = PTR_TO_MAP_VALUE;
2769 }
2770 /* We don't need id from this point onwards anymore, thus we
2771 * should better reset it, so that state pruning has chances
2772 * to take effect.
2773 */
2774 reg->id = 0;
2775 }
2776 }
2777
2778 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2779 * be folded together at some point.
2780 */
2781 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2782 bool is_null)
2783 {
2784 struct bpf_reg_state *regs = state->regs;
2785 u32 id = regs[regno].id;
2786 int i;
2787
2788 for (i = 0; i < MAX_BPF_REG; i++)
2789 mark_map_reg(regs, i, id, is_null);
2790
2791 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2792 if (state->stack_slot_type[i] != STACK_SPILL)
2793 continue;
2794 mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
2795 }
2796 }
2797
2798 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2799 struct bpf_insn *insn, int *insn_idx)
2800 {
2801 struct bpf_verifier_state *other_branch, *this_branch = &env->cur_state;
2802 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2803 u8 opcode = BPF_OP(insn->code);
2804 int err;
2805
2806 if (opcode > BPF_JSLE) {
2807 verbose("invalid BPF_JMP opcode %x\n", opcode);
2808 return -EINVAL;
2809 }
2810
2811 if (BPF_SRC(insn->code) == BPF_X) {
2812 if (insn->imm != 0) {
2813 verbose("BPF_JMP uses reserved fields\n");
2814 return -EINVAL;
2815 }
2816
2817 /* check src1 operand */
2818 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2819 if (err)
2820 return err;
2821
2822 if (is_pointer_value(env, insn->src_reg)) {
2823 verbose("R%d pointer comparison prohibited\n",
2824 insn->src_reg);
2825 return -EACCES;
2826 }
2827 } else {
2828 if (insn->src_reg != BPF_REG_0) {
2829 verbose("BPF_JMP uses reserved fields\n");
2830 return -EINVAL;
2831 }
2832 }
2833
2834 /* check src2 operand */
2835 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2836 if (err)
2837 return err;
2838
2839 dst_reg = &regs[insn->dst_reg];
2840
2841 /* detect if R == 0 where R was initialized to zero earlier */
2842 if (BPF_SRC(insn->code) == BPF_K &&
2843 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2844 dst_reg->type == SCALAR_VALUE &&
2845 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2846 if (opcode == BPF_JEQ) {
2847 /* if (imm == imm) goto pc+off;
2848 * only follow the goto, ignore fall-through
2849 */
2850 *insn_idx += insn->off;
2851 return 0;
2852 } else {
2853 /* if (imm != imm) goto pc+off;
2854 * only follow fall-through branch, since
2855 * that's where the program will go
2856 */
2857 return 0;
2858 }
2859 }
2860
2861 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2862 if (!other_branch)
2863 return -EFAULT;
2864
2865 /* detect if we are comparing against a constant value so we can adjust
2866 * our min/max values for our dst register.
2867 * this is only legit if both are scalars (or pointers to the same
2868 * object, I suppose, but we don't support that right now), because
2869 * otherwise the different base pointers mean the offsets aren't
2870 * comparable.
2871 */
2872 if (BPF_SRC(insn->code) == BPF_X) {
2873 if (dst_reg->type == SCALAR_VALUE &&
2874 regs[insn->src_reg].type == SCALAR_VALUE) {
2875 if (tnum_is_const(regs[insn->src_reg].var_off))
2876 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2877 dst_reg, regs[insn->src_reg].var_off.value,
2878 opcode);
2879 else if (tnum_is_const(dst_reg->var_off))
2880 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
2881 &regs[insn->src_reg],
2882 dst_reg->var_off.value, opcode);
2883 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
2884 /* Comparing for equality, we can combine knowledge */
2885 reg_combine_min_max(&other_branch->regs[insn->src_reg],
2886 &other_branch->regs[insn->dst_reg],
2887 &regs[insn->src_reg],
2888 &regs[insn->dst_reg], opcode);
2889 }
2890 } else if (dst_reg->type == SCALAR_VALUE) {
2891 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2892 dst_reg, insn->imm, opcode);
2893 }
2894
2895 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2896 if (BPF_SRC(insn->code) == BPF_K &&
2897 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2898 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2899 /* Mark all identical map registers in each branch as either
2900 * safe or unknown depending R == 0 or R != 0 conditional.
2901 */
2902 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
2903 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
2904 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2905 dst_reg->type == PTR_TO_PACKET &&
2906 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2907 /* pkt_data' > pkt_end */
2908 find_good_pkt_pointers(this_branch, dst_reg, false);
2909 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2910 dst_reg->type == PTR_TO_PACKET_END &&
2911 regs[insn->src_reg].type == PTR_TO_PACKET) {
2912 /* pkt_end > pkt_data' */
2913 find_good_pkt_pointers(other_branch, &regs[insn->src_reg], true);
2914 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2915 dst_reg->type == PTR_TO_PACKET &&
2916 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2917 /* pkt_data' < pkt_end */
2918 find_good_pkt_pointers(other_branch, dst_reg, true);
2919 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2920 dst_reg->type == PTR_TO_PACKET_END &&
2921 regs[insn->src_reg].type == PTR_TO_PACKET) {
2922 /* pkt_end < pkt_data' */
2923 find_good_pkt_pointers(this_branch, &regs[insn->src_reg], false);
2924 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2925 dst_reg->type == PTR_TO_PACKET &&
2926 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2927 /* pkt_data' >= pkt_end */
2928 find_good_pkt_pointers(this_branch, dst_reg, true);
2929 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2930 dst_reg->type == PTR_TO_PACKET_END &&
2931 regs[insn->src_reg].type == PTR_TO_PACKET) {
2932 /* pkt_end >= pkt_data' */
2933 find_good_pkt_pointers(other_branch, &regs[insn->src_reg], false);
2934 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2935 dst_reg->type == PTR_TO_PACKET &&
2936 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2937 /* pkt_data' <= pkt_end */
2938 find_good_pkt_pointers(other_branch, dst_reg, false);
2939 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2940 dst_reg->type == PTR_TO_PACKET_END &&
2941 regs[insn->src_reg].type == PTR_TO_PACKET) {
2942 /* pkt_end <= pkt_data' */
2943 find_good_pkt_pointers(this_branch, &regs[insn->src_reg], true);
2944 } else if (is_pointer_value(env, insn->dst_reg)) {
2945 verbose("R%d pointer comparison prohibited\n", insn->dst_reg);
2946 return -EACCES;
2947 }
2948 if (log_level)
2949 print_verifier_state(this_branch);
2950 return 0;
2951 }
2952
2953 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
2954 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
2955 {
2956 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
2957
2958 return (struct bpf_map *) (unsigned long) imm64;
2959 }
2960
2961 /* verify BPF_LD_IMM64 instruction */
2962 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
2963 {
2964 struct bpf_reg_state *regs = env->cur_state.regs;
2965 int err;
2966
2967 if (BPF_SIZE(insn->code) != BPF_DW) {
2968 verbose("invalid BPF_LD_IMM insn\n");
2969 return -EINVAL;
2970 }
2971 if (insn->off != 0) {
2972 verbose("BPF_LD_IMM64 uses reserved fields\n");
2973 return -EINVAL;
2974 }
2975
2976 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2977 if (err)
2978 return err;
2979
2980 if (insn->src_reg == 0) {
2981 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
2982
2983 regs[insn->dst_reg].type = SCALAR_VALUE;
2984 __mark_reg_known(&regs[insn->dst_reg], imm);
2985 return 0;
2986 }
2987
2988 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
2989 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
2990
2991 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
2992 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
2993 return 0;
2994 }
2995
2996 static bool may_access_skb(enum bpf_prog_type type)
2997 {
2998 switch (type) {
2999 case BPF_PROG_TYPE_SOCKET_FILTER:
3000 case BPF_PROG_TYPE_SCHED_CLS:
3001 case BPF_PROG_TYPE_SCHED_ACT:
3002 return true;
3003 default:
3004 return false;
3005 }
3006 }
3007
3008 /* verify safety of LD_ABS|LD_IND instructions:
3009 * - they can only appear in the programs where ctx == skb
3010 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3011 * preserve R6-R9, and store return value into R0
3012 *
3013 * Implicit input:
3014 * ctx == skb == R6 == CTX
3015 *
3016 * Explicit input:
3017 * SRC == any register
3018 * IMM == 32-bit immediate
3019 *
3020 * Output:
3021 * R0 - 8/16/32-bit skb data converted to cpu endianness
3022 */
3023 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3024 {
3025 struct bpf_reg_state *regs = env->cur_state.regs;
3026 u8 mode = BPF_MODE(insn->code);
3027 int i, err;
3028
3029 if (!may_access_skb(env->prog->type)) {
3030 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3031 return -EINVAL;
3032 }
3033
3034 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3035 BPF_SIZE(insn->code) == BPF_DW ||
3036 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3037 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
3038 return -EINVAL;
3039 }
3040
3041 /* check whether implicit source operand (register R6) is readable */
3042 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3043 if (err)
3044 return err;
3045
3046 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3047 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3048 return -EINVAL;
3049 }
3050
3051 if (mode == BPF_IND) {
3052 /* check explicit source operand */
3053 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3054 if (err)
3055 return err;
3056 }
3057
3058 /* reset caller saved regs to unreadable */
3059 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3060 mark_reg_not_init(regs, caller_saved[i]);
3061 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3062 }
3063
3064 /* mark destination R0 register as readable, since it contains
3065 * the value fetched from the packet.
3066 * Already marked as written above.
3067 */
3068 mark_reg_unknown(regs, BPF_REG_0);
3069 return 0;
3070 }
3071
3072 /* non-recursive DFS pseudo code
3073 * 1 procedure DFS-iterative(G,v):
3074 * 2 label v as discovered
3075 * 3 let S be a stack
3076 * 4 S.push(v)
3077 * 5 while S is not empty
3078 * 6 t <- S.pop()
3079 * 7 if t is what we're looking for:
3080 * 8 return t
3081 * 9 for all edges e in G.adjacentEdges(t) do
3082 * 10 if edge e is already labelled
3083 * 11 continue with the next edge
3084 * 12 w <- G.adjacentVertex(t,e)
3085 * 13 if vertex w is not discovered and not explored
3086 * 14 label e as tree-edge
3087 * 15 label w as discovered
3088 * 16 S.push(w)
3089 * 17 continue at 5
3090 * 18 else if vertex w is discovered
3091 * 19 label e as back-edge
3092 * 20 else
3093 * 21 // vertex w is explored
3094 * 22 label e as forward- or cross-edge
3095 * 23 label t as explored
3096 * 24 S.pop()
3097 *
3098 * convention:
3099 * 0x10 - discovered
3100 * 0x11 - discovered and fall-through edge labelled
3101 * 0x12 - discovered and fall-through and branch edges labelled
3102 * 0x20 - explored
3103 */
3104
3105 enum {
3106 DISCOVERED = 0x10,
3107 EXPLORED = 0x20,
3108 FALLTHROUGH = 1,
3109 BRANCH = 2,
3110 };
3111
3112 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3113
3114 static int *insn_stack; /* stack of insns to process */
3115 static int cur_stack; /* current stack index */
3116 static int *insn_state;
3117
3118 /* t, w, e - match pseudo-code above:
3119 * t - index of current instruction
3120 * w - next instruction
3121 * e - edge
3122 */
3123 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3124 {
3125 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3126 return 0;
3127
3128 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3129 return 0;
3130
3131 if (w < 0 || w >= env->prog->len) {
3132 verbose("jump out of range from insn %d to %d\n", t, w);
3133 return -EINVAL;
3134 }
3135
3136 if (e == BRANCH)
3137 /* mark branch target for state pruning */
3138 env->explored_states[w] = STATE_LIST_MARK;
3139
3140 if (insn_state[w] == 0) {
3141 /* tree-edge */
3142 insn_state[t] = DISCOVERED | e;
3143 insn_state[w] = DISCOVERED;
3144 if (cur_stack >= env->prog->len)
3145 return -E2BIG;
3146 insn_stack[cur_stack++] = w;
3147 return 1;
3148 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3149 verbose("back-edge from insn %d to %d\n", t, w);
3150 return -EINVAL;
3151 } else if (insn_state[w] == EXPLORED) {
3152 /* forward- or cross-edge */
3153 insn_state[t] = DISCOVERED | e;
3154 } else {
3155 verbose("insn state internal bug\n");
3156 return -EFAULT;
3157 }
3158 return 0;
3159 }
3160
3161 /* non-recursive depth-first-search to detect loops in BPF program
3162 * loop == back-edge in directed graph
3163 */
3164 static int check_cfg(struct bpf_verifier_env *env)
3165 {
3166 struct bpf_insn *insns = env->prog->insnsi;
3167 int insn_cnt = env->prog->len;
3168 int ret = 0;
3169 int i, t;
3170
3171 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3172 if (!insn_state)
3173 return -ENOMEM;
3174
3175 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3176 if (!insn_stack) {
3177 kfree(insn_state);
3178 return -ENOMEM;
3179 }
3180
3181 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3182 insn_stack[0] = 0; /* 0 is the first instruction */
3183 cur_stack = 1;
3184
3185 peek_stack:
3186 if (cur_stack == 0)
3187 goto check_state;
3188 t = insn_stack[cur_stack - 1];
3189
3190 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3191 u8 opcode = BPF_OP(insns[t].code);
3192
3193 if (opcode == BPF_EXIT) {
3194 goto mark_explored;
3195 } else if (opcode == BPF_CALL) {
3196 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3197 if (ret == 1)
3198 goto peek_stack;
3199 else if (ret < 0)
3200 goto err_free;
3201 if (t + 1 < insn_cnt)
3202 env->explored_states[t + 1] = STATE_LIST_MARK;
3203 } else if (opcode == BPF_JA) {
3204 if (BPF_SRC(insns[t].code) != BPF_K) {
3205 ret = -EINVAL;
3206 goto err_free;
3207 }
3208 /* unconditional jump with single edge */
3209 ret = push_insn(t, t + insns[t].off + 1,
3210 FALLTHROUGH, env);
3211 if (ret == 1)
3212 goto peek_stack;
3213 else if (ret < 0)
3214 goto err_free;
3215 /* tell verifier to check for equivalent states
3216 * after every call and jump
3217 */
3218 if (t + 1 < insn_cnt)
3219 env->explored_states[t + 1] = STATE_LIST_MARK;
3220 } else {
3221 /* conditional jump with two edges */
3222 env->explored_states[t] = STATE_LIST_MARK;
3223 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3224 if (ret == 1)
3225 goto peek_stack;
3226 else if (ret < 0)
3227 goto err_free;
3228
3229 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3230 if (ret == 1)
3231 goto peek_stack;
3232 else if (ret < 0)
3233 goto err_free;
3234 }
3235 } else {
3236 /* all other non-branch instructions with single
3237 * fall-through edge
3238 */
3239 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3240 if (ret == 1)
3241 goto peek_stack;
3242 else if (ret < 0)
3243 goto err_free;
3244 }
3245
3246 mark_explored:
3247 insn_state[t] = EXPLORED;
3248 if (cur_stack-- <= 0) {
3249 verbose("pop stack internal bug\n");
3250 ret = -EFAULT;
3251 goto err_free;
3252 }
3253 goto peek_stack;
3254
3255 check_state:
3256 for (i = 0; i < insn_cnt; i++) {
3257 if (insn_state[i] != EXPLORED) {
3258 verbose("unreachable insn %d\n", i);
3259 ret = -EINVAL;
3260 goto err_free;
3261 }
3262 }
3263 ret = 0; /* cfg looks good */
3264
3265 err_free:
3266 kfree(insn_state);
3267 kfree(insn_stack);
3268 return ret;
3269 }
3270
3271 /* check %cur's range satisfies %old's */
3272 static bool range_within(struct bpf_reg_state *old,
3273 struct bpf_reg_state *cur)
3274 {
3275 return old->umin_value <= cur->umin_value &&
3276 old->umax_value >= cur->umax_value &&
3277 old->smin_value <= cur->smin_value &&
3278 old->smax_value >= cur->smax_value;
3279 }
3280
3281 /* Maximum number of register states that can exist at once */
3282 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3283 struct idpair {
3284 u32 old;
3285 u32 cur;
3286 };
3287
3288 /* If in the old state two registers had the same id, then they need to have
3289 * the same id in the new state as well. But that id could be different from
3290 * the old state, so we need to track the mapping from old to new ids.
3291 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3292 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3293 * regs with a different old id could still have new id 9, we don't care about
3294 * that.
3295 * So we look through our idmap to see if this old id has been seen before. If
3296 * so, we require the new id to match; otherwise, we add the id pair to the map.
3297 */
3298 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3299 {
3300 unsigned int i;
3301
3302 for (i = 0; i < ID_MAP_SIZE; i++) {
3303 if (!idmap[i].old) {
3304 /* Reached an empty slot; haven't seen this id before */
3305 idmap[i].old = old_id;
3306 idmap[i].cur = cur_id;
3307 return true;
3308 }
3309 if (idmap[i].old == old_id)
3310 return idmap[i].cur == cur_id;
3311 }
3312 /* We ran out of idmap slots, which should be impossible */
3313 WARN_ON_ONCE(1);
3314 return false;
3315 }
3316
3317 /* Returns true if (rold safe implies rcur safe) */
3318 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3319 struct idpair *idmap)
3320 {
3321 if (!(rold->live & REG_LIVE_READ))
3322 /* explored state didn't use this */
3323 return true;
3324
3325 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3326 return true;
3327
3328 if (rold->type == NOT_INIT)
3329 /* explored state can't have used this */
3330 return true;
3331 if (rcur->type == NOT_INIT)
3332 return false;
3333 switch (rold->type) {
3334 case SCALAR_VALUE:
3335 if (rcur->type == SCALAR_VALUE) {
3336 /* new val must satisfy old val knowledge */
3337 return range_within(rold, rcur) &&
3338 tnum_in(rold->var_off, rcur->var_off);
3339 } else {
3340 /* We're trying to use a pointer in place of a scalar.
3341 * Even if the scalar was unbounded, this could lead to
3342 * pointer leaks because scalars are allowed to leak
3343 * while pointers are not. We could make this safe in
3344 * special cases if root is calling us, but it's
3345 * probably not worth the hassle.
3346 */
3347 return false;
3348 }
3349 case PTR_TO_MAP_VALUE:
3350 /* If the new min/max/var_off satisfy the old ones and
3351 * everything else matches, we are OK.
3352 * We don't care about the 'id' value, because nothing
3353 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3354 */
3355 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3356 range_within(rold, rcur) &&
3357 tnum_in(rold->var_off, rcur->var_off);
3358 case PTR_TO_MAP_VALUE_OR_NULL:
3359 /* a PTR_TO_MAP_VALUE could be safe to use as a
3360 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3361 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3362 * checked, doing so could have affected others with the same
3363 * id, and we can't check for that because we lost the id when
3364 * we converted to a PTR_TO_MAP_VALUE.
3365 */
3366 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3367 return false;
3368 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3369 return false;
3370 /* Check our ids match any regs they're supposed to */
3371 return check_ids(rold->id, rcur->id, idmap);
3372 case PTR_TO_PACKET:
3373 if (rcur->type != PTR_TO_PACKET)
3374 return false;
3375 /* We must have at least as much range as the old ptr
3376 * did, so that any accesses which were safe before are
3377 * still safe. This is true even if old range < old off,
3378 * since someone could have accessed through (ptr - k), or
3379 * even done ptr -= k in a register, to get a safe access.
3380 */
3381 if (rold->range > rcur->range)
3382 return false;
3383 /* If the offsets don't match, we can't trust our alignment;
3384 * nor can we be sure that we won't fall out of range.
3385 */
3386 if (rold->off != rcur->off)
3387 return false;
3388 /* id relations must be preserved */
3389 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3390 return false;
3391 /* new val must satisfy old val knowledge */
3392 return range_within(rold, rcur) &&
3393 tnum_in(rold->var_off, rcur->var_off);
3394 case PTR_TO_CTX:
3395 case CONST_PTR_TO_MAP:
3396 case PTR_TO_STACK:
3397 case PTR_TO_PACKET_END:
3398 /* Only valid matches are exact, which memcmp() above
3399 * would have accepted
3400 */
3401 default:
3402 /* Don't know what's going on, just say it's not safe */
3403 return false;
3404 }
3405
3406 /* Shouldn't get here; if we do, say it's not safe */
3407 WARN_ON_ONCE(1);
3408 return false;
3409 }
3410
3411 /* compare two verifier states
3412 *
3413 * all states stored in state_list are known to be valid, since
3414 * verifier reached 'bpf_exit' instruction through them
3415 *
3416 * this function is called when verifier exploring different branches of
3417 * execution popped from the state stack. If it sees an old state that has
3418 * more strict register state and more strict stack state then this execution
3419 * branch doesn't need to be explored further, since verifier already
3420 * concluded that more strict state leads to valid finish.
3421 *
3422 * Therefore two states are equivalent if register state is more conservative
3423 * and explored stack state is more conservative than the current one.
3424 * Example:
3425 * explored current
3426 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3427 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3428 *
3429 * In other words if current stack state (one being explored) has more
3430 * valid slots than old one that already passed validation, it means
3431 * the verifier can stop exploring and conclude that current state is valid too
3432 *
3433 * Similarly with registers. If explored state has register type as invalid
3434 * whereas register type in current state is meaningful, it means that
3435 * the current state will reach 'bpf_exit' instruction safely
3436 */
3437 static bool states_equal(struct bpf_verifier_env *env,
3438 struct bpf_verifier_state *old,
3439 struct bpf_verifier_state *cur)
3440 {
3441 struct idpair *idmap;
3442 bool ret = false;
3443 int i;
3444
3445 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3446 /* If we failed to allocate the idmap, just say it's not safe */
3447 if (!idmap)
3448 return false;
3449
3450 for (i = 0; i < MAX_BPF_REG; i++) {
3451 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3452 goto out_free;
3453 }
3454
3455 for (i = 0; i < MAX_BPF_STACK; i++) {
3456 if (old->stack_slot_type[i] == STACK_INVALID)
3457 continue;
3458 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
3459 /* Ex: old explored (safe) state has STACK_SPILL in
3460 * this stack slot, but current has has STACK_MISC ->
3461 * this verifier states are not equivalent,
3462 * return false to continue verification of this path
3463 */
3464 goto out_free;
3465 if (i % BPF_REG_SIZE)
3466 continue;
3467 if (old->stack_slot_type[i] != STACK_SPILL)
3468 continue;
3469 if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
3470 &cur->spilled_regs[i / BPF_REG_SIZE],
3471 idmap))
3472 /* when explored and current stack slot are both storing
3473 * spilled registers, check that stored pointers types
3474 * are the same as well.
3475 * Ex: explored safe path could have stored
3476 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3477 * but current path has stored:
3478 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3479 * such verifier states are not equivalent.
3480 * return false to continue verification of this path
3481 */
3482 goto out_free;
3483 else
3484 continue;
3485 }
3486 ret = true;
3487 out_free:
3488 kfree(idmap);
3489 return ret;
3490 }
3491
3492 /* A write screens off any subsequent reads; but write marks come from the
3493 * straight-line code between a state and its parent. When we arrive at a
3494 * jump target (in the first iteration of the propagate_liveness() loop),
3495 * we didn't arrive by the straight-line code, so read marks in state must
3496 * propagate to parent regardless of state's write marks.
3497 */
3498 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3499 struct bpf_verifier_state *parent)
3500 {
3501 bool writes = parent == state->parent; /* Observe write marks */
3502 bool touched = false; /* any changes made? */
3503 int i;
3504
3505 if (!parent)
3506 return touched;
3507 /* Propagate read liveness of registers... */
3508 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3509 /* We don't need to worry about FP liveness because it's read-only */
3510 for (i = 0; i < BPF_REG_FP; i++) {
3511 if (parent->regs[i].live & REG_LIVE_READ)
3512 continue;
3513 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3514 continue;
3515 if (state->regs[i].live & REG_LIVE_READ) {
3516 parent->regs[i].live |= REG_LIVE_READ;
3517 touched = true;
3518 }
3519 }
3520 /* ... and stack slots */
3521 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++) {
3522 if (parent->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3523 continue;
3524 if (state->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3525 continue;
3526 if (parent->spilled_regs[i].live & REG_LIVE_READ)
3527 continue;
3528 if (writes && (state->spilled_regs[i].live & REG_LIVE_WRITTEN))
3529 continue;
3530 if (state->spilled_regs[i].live & REG_LIVE_READ) {
3531 parent->spilled_regs[i].live |= REG_LIVE_READ;
3532 touched = true;
3533 }
3534 }
3535 return touched;
3536 }
3537
3538 /* "parent" is "a state from which we reach the current state", but initially
3539 * it is not the state->parent (i.e. "the state whose straight-line code leads
3540 * to the current state"), instead it is the state that happened to arrive at
3541 * a (prunable) equivalent of the current state. See comment above
3542 * do_propagate_liveness() for consequences of this.
3543 * This function is just a more efficient way of calling mark_reg_read() or
3544 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3545 * though it requires that parent != state->parent in the call arguments.
3546 */
3547 static void propagate_liveness(const struct bpf_verifier_state *state,
3548 struct bpf_verifier_state *parent)
3549 {
3550 while (do_propagate_liveness(state, parent)) {
3551 /* Something changed, so we need to feed those changes onward */
3552 state = parent;
3553 parent = state->parent;
3554 }
3555 }
3556
3557 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3558 {
3559 struct bpf_verifier_state_list *new_sl;
3560 struct bpf_verifier_state_list *sl;
3561 int i;
3562
3563 sl = env->explored_states[insn_idx];
3564 if (!sl)
3565 /* this 'insn_idx' instruction wasn't marked, so we will not
3566 * be doing state search here
3567 */
3568 return 0;
3569
3570 while (sl != STATE_LIST_MARK) {
3571 if (states_equal(env, &sl->state, &env->cur_state)) {
3572 /* reached equivalent register/stack state,
3573 * prune the search.
3574 * Registers read by the continuation are read by us.
3575 * If we have any write marks in env->cur_state, they
3576 * will prevent corresponding reads in the continuation
3577 * from reaching our parent (an explored_state). Our
3578 * own state will get the read marks recorded, but
3579 * they'll be immediately forgotten as we're pruning
3580 * this state and will pop a new one.
3581 */
3582 propagate_liveness(&sl->state, &env->cur_state);
3583 return 1;
3584 }
3585 sl = sl->next;
3586 }
3587
3588 /* there were no equivalent states, remember current one.
3589 * technically the current state is not proven to be safe yet,
3590 * but it will either reach bpf_exit (which means it's safe) or
3591 * it will be rejected. Since there are no loops, we won't be
3592 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3593 */
3594 new_sl = kmalloc(sizeof(struct bpf_verifier_state_list), GFP_USER);
3595 if (!new_sl)
3596 return -ENOMEM;
3597
3598 /* add new state to the head of linked list */
3599 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
3600 new_sl->next = env->explored_states[insn_idx];
3601 env->explored_states[insn_idx] = new_sl;
3602 /* connect new state to parentage chain */
3603 env->cur_state.parent = &new_sl->state;
3604 /* clear write marks in current state: the writes we did are not writes
3605 * our child did, so they don't screen off its reads from us.
3606 * (There are no read marks in current state, because reads always mark
3607 * their parent and current state never has children yet. Only
3608 * explored_states can get read marks.)
3609 */
3610 for (i = 0; i < BPF_REG_FP; i++)
3611 env->cur_state.regs[i].live = REG_LIVE_NONE;
3612 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++)
3613 if (env->cur_state.stack_slot_type[i * BPF_REG_SIZE] == STACK_SPILL)
3614 env->cur_state.spilled_regs[i].live = REG_LIVE_NONE;
3615 return 0;
3616 }
3617
3618 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3619 int insn_idx, int prev_insn_idx)
3620 {
3621 if (!env->analyzer_ops || !env->analyzer_ops->insn_hook)
3622 return 0;
3623
3624 return env->analyzer_ops->insn_hook(env, insn_idx, prev_insn_idx);
3625 }
3626
3627 static int do_check(struct bpf_verifier_env *env)
3628 {
3629 struct bpf_verifier_state *state = &env->cur_state;
3630 struct bpf_insn *insns = env->prog->insnsi;
3631 struct bpf_reg_state *regs = state->regs;
3632 int insn_cnt = env->prog->len;
3633 int insn_idx, prev_insn_idx = 0;
3634 int insn_processed = 0;
3635 bool do_print_state = false;
3636
3637 init_reg_state(regs);
3638 state->parent = NULL;
3639 insn_idx = 0;
3640 for (;;) {
3641 struct bpf_insn *insn;
3642 u8 class;
3643 int err;
3644
3645 if (insn_idx >= insn_cnt) {
3646 verbose("invalid insn idx %d insn_cnt %d\n",
3647 insn_idx, insn_cnt);
3648 return -EFAULT;
3649 }
3650
3651 insn = &insns[insn_idx];
3652 class = BPF_CLASS(insn->code);
3653
3654 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3655 verbose("BPF program is too large. Processed %d insn\n",
3656 insn_processed);
3657 return -E2BIG;
3658 }
3659
3660 err = is_state_visited(env, insn_idx);
3661 if (err < 0)
3662 return err;
3663 if (err == 1) {
3664 /* found equivalent state, can prune the search */
3665 if (log_level) {
3666 if (do_print_state)
3667 verbose("\nfrom %d to %d: safe\n",
3668 prev_insn_idx, insn_idx);
3669 else
3670 verbose("%d: safe\n", insn_idx);
3671 }
3672 goto process_bpf_exit;
3673 }
3674
3675 if (need_resched())
3676 cond_resched();
3677
3678 if (log_level > 1 || (log_level && do_print_state)) {
3679 if (log_level > 1)
3680 verbose("%d:", insn_idx);
3681 else
3682 verbose("\nfrom %d to %d:",
3683 prev_insn_idx, insn_idx);
3684 print_verifier_state(&env->cur_state);
3685 do_print_state = false;
3686 }
3687
3688 if (log_level) {
3689 verbose("%d: ", insn_idx);
3690 print_bpf_insn(env, insn);
3691 }
3692
3693 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3694 if (err)
3695 return err;
3696
3697 env->insn_aux_data[insn_idx].seen = true;
3698 if (class == BPF_ALU || class == BPF_ALU64) {
3699 err = check_alu_op(env, insn);
3700 if (err)
3701 return err;
3702
3703 } else if (class == BPF_LDX) {
3704 enum bpf_reg_type *prev_src_type, src_reg_type;
3705
3706 /* check for reserved fields is already done */
3707
3708 /* check src operand */
3709 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3710 if (err)
3711 return err;
3712
3713 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3714 if (err)
3715 return err;
3716
3717 src_reg_type = regs[insn->src_reg].type;
3718
3719 /* check that memory (src_reg + off) is readable,
3720 * the state of dst_reg will be updated by this func
3721 */
3722 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3723 BPF_SIZE(insn->code), BPF_READ,
3724 insn->dst_reg);
3725 if (err)
3726 return err;
3727
3728 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3729
3730 if (*prev_src_type == NOT_INIT) {
3731 /* saw a valid insn
3732 * dst_reg = *(u32 *)(src_reg + off)
3733 * save type to validate intersecting paths
3734 */
3735 *prev_src_type = src_reg_type;
3736
3737 } else if (src_reg_type != *prev_src_type &&
3738 (src_reg_type == PTR_TO_CTX ||
3739 *prev_src_type == PTR_TO_CTX)) {
3740 /* ABuser program is trying to use the same insn
3741 * dst_reg = *(u32*) (src_reg + off)
3742 * with different pointer types:
3743 * src_reg == ctx in one branch and
3744 * src_reg == stack|map in some other branch.
3745 * Reject it.
3746 */
3747 verbose("same insn cannot be used with different pointers\n");
3748 return -EINVAL;
3749 }
3750
3751 } else if (class == BPF_STX) {
3752 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3753
3754 if (BPF_MODE(insn->code) == BPF_XADD) {
3755 err = check_xadd(env, insn_idx, insn);
3756 if (err)
3757 return err;
3758 insn_idx++;
3759 continue;
3760 }
3761
3762 /* check src1 operand */
3763 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3764 if (err)
3765 return err;
3766 /* check src2 operand */
3767 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3768 if (err)
3769 return err;
3770
3771 dst_reg_type = regs[insn->dst_reg].type;
3772
3773 /* check that memory (dst_reg + off) is writeable */
3774 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3775 BPF_SIZE(insn->code), BPF_WRITE,
3776 insn->src_reg);
3777 if (err)
3778 return err;
3779
3780 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3781
3782 if (*prev_dst_type == NOT_INIT) {
3783 *prev_dst_type = dst_reg_type;
3784 } else if (dst_reg_type != *prev_dst_type &&
3785 (dst_reg_type == PTR_TO_CTX ||
3786 *prev_dst_type == PTR_TO_CTX)) {
3787 verbose("same insn cannot be used with different pointers\n");
3788 return -EINVAL;
3789 }
3790
3791 } else if (class == BPF_ST) {
3792 if (BPF_MODE(insn->code) != BPF_MEM ||
3793 insn->src_reg != BPF_REG_0) {
3794 verbose("BPF_ST uses reserved fields\n");
3795 return -EINVAL;
3796 }
3797 /* check src operand */
3798 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3799 if (err)
3800 return err;
3801
3802 /* check that memory (dst_reg + off) is writeable */
3803 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3804 BPF_SIZE(insn->code), BPF_WRITE,
3805 -1);
3806 if (err)
3807 return err;
3808
3809 } else if (class == BPF_JMP) {
3810 u8 opcode = BPF_OP(insn->code);
3811
3812 if (opcode == BPF_CALL) {
3813 if (BPF_SRC(insn->code) != BPF_K ||
3814 insn->off != 0 ||
3815 insn->src_reg != BPF_REG_0 ||
3816 insn->dst_reg != BPF_REG_0) {
3817 verbose("BPF_CALL uses reserved fields\n");
3818 return -EINVAL;
3819 }
3820
3821 err = check_call(env, insn->imm, insn_idx);
3822 if (err)
3823 return err;
3824
3825 } else if (opcode == BPF_JA) {
3826 if (BPF_SRC(insn->code) != BPF_K ||
3827 insn->imm != 0 ||
3828 insn->src_reg != BPF_REG_0 ||
3829 insn->dst_reg != BPF_REG_0) {
3830 verbose("BPF_JA uses reserved fields\n");
3831 return -EINVAL;
3832 }
3833
3834 insn_idx += insn->off + 1;
3835 continue;
3836
3837 } else if (opcode == BPF_EXIT) {
3838 if (BPF_SRC(insn->code) != BPF_K ||
3839 insn->imm != 0 ||
3840 insn->src_reg != BPF_REG_0 ||
3841 insn->dst_reg != BPF_REG_0) {
3842 verbose("BPF_EXIT uses reserved fields\n");
3843 return -EINVAL;
3844 }
3845
3846 /* eBPF calling convetion is such that R0 is used
3847 * to return the value from eBPF program.
3848 * Make sure that it's readable at this time
3849 * of bpf_exit, which means that program wrote
3850 * something into it earlier
3851 */
3852 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
3853 if (err)
3854 return err;
3855
3856 if (is_pointer_value(env, BPF_REG_0)) {
3857 verbose("R0 leaks addr as return value\n");
3858 return -EACCES;
3859 }
3860
3861 process_bpf_exit:
3862 insn_idx = pop_stack(env, &prev_insn_idx);
3863 if (insn_idx < 0) {
3864 break;
3865 } else {
3866 do_print_state = true;
3867 continue;
3868 }
3869 } else {
3870 err = check_cond_jmp_op(env, insn, &insn_idx);
3871 if (err)
3872 return err;
3873 }
3874 } else if (class == BPF_LD) {
3875 u8 mode = BPF_MODE(insn->code);
3876
3877 if (mode == BPF_ABS || mode == BPF_IND) {
3878 err = check_ld_abs(env, insn);
3879 if (err)
3880 return err;
3881
3882 } else if (mode == BPF_IMM) {
3883 err = check_ld_imm(env, insn);
3884 if (err)
3885 return err;
3886
3887 insn_idx++;
3888 env->insn_aux_data[insn_idx].seen = true;
3889 } else {
3890 verbose("invalid BPF_LD mode\n");
3891 return -EINVAL;
3892 }
3893 } else {
3894 verbose("unknown insn class %d\n", class);
3895 return -EINVAL;
3896 }
3897
3898 insn_idx++;
3899 }
3900
3901 verbose("processed %d insns, stack depth %d\n",
3902 insn_processed, env->prog->aux->stack_depth);
3903 return 0;
3904 }
3905
3906 static int check_map_prealloc(struct bpf_map *map)
3907 {
3908 return (map->map_type != BPF_MAP_TYPE_HASH &&
3909 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
3910 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
3911 !(map->map_flags & BPF_F_NO_PREALLOC);
3912 }
3913
3914 static int check_map_prog_compatibility(struct bpf_map *map,
3915 struct bpf_prog *prog)
3916
3917 {
3918 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3919 * preallocated hash maps, since doing memory allocation
3920 * in overflow_handler can crash depending on where nmi got
3921 * triggered.
3922 */
3923 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
3924 if (!check_map_prealloc(map)) {
3925 verbose("perf_event programs can only use preallocated hash map\n");
3926 return -EINVAL;
3927 }
3928 if (map->inner_map_meta &&
3929 !check_map_prealloc(map->inner_map_meta)) {
3930 verbose("perf_event programs can only use preallocated inner hash map\n");
3931 return -EINVAL;
3932 }
3933 }
3934 return 0;
3935 }
3936
3937 /* look for pseudo eBPF instructions that access map FDs and
3938 * replace them with actual map pointers
3939 */
3940 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
3941 {
3942 struct bpf_insn *insn = env->prog->insnsi;
3943 int insn_cnt = env->prog->len;
3944 int i, j, err;
3945
3946 err = bpf_prog_calc_tag(env->prog);
3947 if (err)
3948 return err;
3949
3950 for (i = 0; i < insn_cnt; i++, insn++) {
3951 if (BPF_CLASS(insn->code) == BPF_LDX &&
3952 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
3953 verbose("BPF_LDX uses reserved fields\n");
3954 return -EINVAL;
3955 }
3956
3957 if (BPF_CLASS(insn->code) == BPF_STX &&
3958 ((BPF_MODE(insn->code) != BPF_MEM &&
3959 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
3960 verbose("BPF_STX uses reserved fields\n");
3961 return -EINVAL;
3962 }
3963
3964 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
3965 struct bpf_map *map;
3966 struct fd f;
3967
3968 if (i == insn_cnt - 1 || insn[1].code != 0 ||
3969 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
3970 insn[1].off != 0) {
3971 verbose("invalid bpf_ld_imm64 insn\n");
3972 return -EINVAL;
3973 }
3974
3975 if (insn->src_reg == 0)
3976 /* valid generic load 64-bit imm */
3977 goto next_insn;
3978
3979 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
3980 verbose("unrecognized bpf_ld_imm64 insn\n");
3981 return -EINVAL;
3982 }
3983
3984 f = fdget(insn->imm);
3985 map = __bpf_map_get(f);
3986 if (IS_ERR(map)) {
3987 verbose("fd %d is not pointing to valid bpf_map\n",
3988 insn->imm);
3989 return PTR_ERR(map);
3990 }
3991
3992 err = check_map_prog_compatibility(map, env->prog);
3993 if (err) {
3994 fdput(f);
3995 return err;
3996 }
3997
3998 /* store map pointer inside BPF_LD_IMM64 instruction */
3999 insn[0].imm = (u32) (unsigned long) map;
4000 insn[1].imm = ((u64) (unsigned long) map) >> 32;
4001
4002 /* check whether we recorded this map already */
4003 for (j = 0; j < env->used_map_cnt; j++)
4004 if (env->used_maps[j] == map) {
4005 fdput(f);
4006 goto next_insn;
4007 }
4008
4009 if (env->used_map_cnt >= MAX_USED_MAPS) {
4010 fdput(f);
4011 return -E2BIG;
4012 }
4013
4014 /* hold the map. If the program is rejected by verifier,
4015 * the map will be released by release_maps() or it
4016 * will be used by the valid program until it's unloaded
4017 * and all maps are released in free_bpf_prog_info()
4018 */
4019 map = bpf_map_inc(map, false);
4020 if (IS_ERR(map)) {
4021 fdput(f);
4022 return PTR_ERR(map);
4023 }
4024 env->used_maps[env->used_map_cnt++] = map;
4025
4026 fdput(f);
4027 next_insn:
4028 insn++;
4029 i++;
4030 }
4031 }
4032
4033 /* now all pseudo BPF_LD_IMM64 instructions load valid
4034 * 'struct bpf_map *' into a register instead of user map_fd.
4035 * These pointers will be used later by verifier to validate map access.
4036 */
4037 return 0;
4038 }
4039
4040 /* drop refcnt of maps used by the rejected program */
4041 static void release_maps(struct bpf_verifier_env *env)
4042 {
4043 int i;
4044
4045 for (i = 0; i < env->used_map_cnt; i++)
4046 bpf_map_put(env->used_maps[i]);
4047 }
4048
4049 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4050 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
4051 {
4052 struct bpf_insn *insn = env->prog->insnsi;
4053 int insn_cnt = env->prog->len;
4054 int i;
4055
4056 for (i = 0; i < insn_cnt; i++, insn++)
4057 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
4058 insn->src_reg = 0;
4059 }
4060
4061 /* single env->prog->insni[off] instruction was replaced with the range
4062 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4063 * [0, off) and [off, end) to new locations, so the patched range stays zero
4064 */
4065 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
4066 u32 off, u32 cnt)
4067 {
4068 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
4069 int i;
4070
4071 if (cnt == 1)
4072 return 0;
4073 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4074 if (!new_data)
4075 return -ENOMEM;
4076 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4077 memcpy(new_data + off + cnt - 1, old_data + off,
4078 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4079 for (i = off; i < off + cnt - 1; i++)
4080 new_data[i].seen = true;
4081 env->insn_aux_data = new_data;
4082 vfree(old_data);
4083 return 0;
4084 }
4085
4086 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4087 const struct bpf_insn *patch, u32 len)
4088 {
4089 struct bpf_prog *new_prog;
4090
4091 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4092 if (!new_prog)
4093 return NULL;
4094 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4095 return NULL;
4096 return new_prog;
4097 }
4098
4099 /* The verifier does more data flow analysis than llvm and will not explore
4100 * branches that are dead at run time. Malicious programs can have dead code
4101 * too. Therefore replace all dead at-run-time code with nops.
4102 */
4103 static void sanitize_dead_code(struct bpf_verifier_env *env)
4104 {
4105 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
4106 struct bpf_insn nop = BPF_MOV64_REG(BPF_REG_0, BPF_REG_0);
4107 struct bpf_insn *insn = env->prog->insnsi;
4108 const int insn_cnt = env->prog->len;
4109 int i;
4110
4111 for (i = 0; i < insn_cnt; i++) {
4112 if (aux_data[i].seen)
4113 continue;
4114 memcpy(insn + i, &nop, sizeof(nop));
4115 }
4116 }
4117
4118 /* convert load instructions that access fields of 'struct __sk_buff'
4119 * into sequence of instructions that access fields of 'struct sk_buff'
4120 */
4121 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4122 {
4123 const struct bpf_verifier_ops *ops = env->prog->aux->ops;
4124 int i, cnt, size, ctx_field_size, delta = 0;
4125 const int insn_cnt = env->prog->len;
4126 struct bpf_insn insn_buf[16], *insn;
4127 struct bpf_prog *new_prog;
4128 enum bpf_access_type type;
4129 bool is_narrower_load;
4130 u32 target_size;
4131
4132 if (ops->gen_prologue) {
4133 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4134 env->prog);
4135 if (cnt >= ARRAY_SIZE(insn_buf)) {
4136 verbose("bpf verifier is misconfigured\n");
4137 return -EINVAL;
4138 } else if (cnt) {
4139 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4140 if (!new_prog)
4141 return -ENOMEM;
4142
4143 env->prog = new_prog;
4144 delta += cnt - 1;
4145 }
4146 }
4147
4148 if (!ops->convert_ctx_access)
4149 return 0;
4150
4151 insn = env->prog->insnsi + delta;
4152
4153 for (i = 0; i < insn_cnt; i++, insn++) {
4154 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4155 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4156 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4157 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4158 type = BPF_READ;
4159 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4160 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4161 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4162 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4163 type = BPF_WRITE;
4164 else
4165 continue;
4166
4167 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4168 continue;
4169
4170 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4171 size = BPF_LDST_BYTES(insn);
4172
4173 /* If the read access is a narrower load of the field,
4174 * convert to a 4/8-byte load, to minimum program type specific
4175 * convert_ctx_access changes. If conversion is successful,
4176 * we will apply proper mask to the result.
4177 */
4178 is_narrower_load = size < ctx_field_size;
4179 if (is_narrower_load) {
4180 u32 off = insn->off;
4181 u8 size_code;
4182
4183 if (type == BPF_WRITE) {
4184 verbose("bpf verifier narrow ctx access misconfigured\n");
4185 return -EINVAL;
4186 }
4187
4188 size_code = BPF_H;
4189 if (ctx_field_size == 4)
4190 size_code = BPF_W;
4191 else if (ctx_field_size == 8)
4192 size_code = BPF_DW;
4193
4194 insn->off = off & ~(ctx_field_size - 1);
4195 insn->code = BPF_LDX | BPF_MEM | size_code;
4196 }
4197
4198 target_size = 0;
4199 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4200 &target_size);
4201 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4202 (ctx_field_size && !target_size)) {
4203 verbose("bpf verifier is misconfigured\n");
4204 return -EINVAL;
4205 }
4206
4207 if (is_narrower_load && size < target_size) {
4208 if (ctx_field_size <= 4)
4209 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4210 (1 << size * 8) - 1);
4211 else
4212 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4213 (1 << size * 8) - 1);
4214 }
4215
4216 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4217 if (!new_prog)
4218 return -ENOMEM;
4219
4220 delta += cnt - 1;
4221
4222 /* keep walking new program and skip insns we just inserted */
4223 env->prog = new_prog;
4224 insn = new_prog->insnsi + i + delta;
4225 }
4226
4227 return 0;
4228 }
4229
4230 /* fixup insn->imm field of bpf_call instructions
4231 * and inline eligible helpers as explicit sequence of BPF instructions
4232 *
4233 * this function is called after eBPF program passed verification
4234 */
4235 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4236 {
4237 struct bpf_prog *prog = env->prog;
4238 struct bpf_insn *insn = prog->insnsi;
4239 const struct bpf_func_proto *fn;
4240 const int insn_cnt = prog->len;
4241 struct bpf_insn insn_buf[16];
4242 struct bpf_prog *new_prog;
4243 struct bpf_map *map_ptr;
4244 int i, cnt, delta = 0;
4245
4246 for (i = 0; i < insn_cnt; i++, insn++) {
4247 if (insn->code != (BPF_JMP | BPF_CALL))
4248 continue;
4249
4250 if (insn->imm == BPF_FUNC_get_route_realm)
4251 prog->dst_needed = 1;
4252 if (insn->imm == BPF_FUNC_get_prandom_u32)
4253 bpf_user_rnd_init_once();
4254 if (insn->imm == BPF_FUNC_tail_call) {
4255 /* If we tail call into other programs, we
4256 * cannot make any assumptions since they can
4257 * be replaced dynamically during runtime in
4258 * the program array.
4259 */
4260 prog->cb_access = 1;
4261 env->prog->aux->stack_depth = MAX_BPF_STACK;
4262
4263 /* mark bpf_tail_call as different opcode to avoid
4264 * conditional branch in the interpeter for every normal
4265 * call and to prevent accidental JITing by JIT compiler
4266 * that doesn't support bpf_tail_call yet
4267 */
4268 insn->imm = 0;
4269 insn->code = BPF_JMP | BPF_TAIL_CALL;
4270 continue;
4271 }
4272
4273 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4274 * handlers are currently limited to 64 bit only.
4275 */
4276 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4277 insn->imm == BPF_FUNC_map_lookup_elem) {
4278 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4279 if (map_ptr == BPF_MAP_PTR_POISON ||
4280 !map_ptr->ops->map_gen_lookup)
4281 goto patch_call_imm;
4282
4283 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4284 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4285 verbose("bpf verifier is misconfigured\n");
4286 return -EINVAL;
4287 }
4288
4289 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4290 cnt);
4291 if (!new_prog)
4292 return -ENOMEM;
4293
4294 delta += cnt - 1;
4295
4296 /* keep walking new program and skip insns we just inserted */
4297 env->prog = prog = new_prog;
4298 insn = new_prog->insnsi + i + delta;
4299 continue;
4300 }
4301
4302 if (insn->imm == BPF_FUNC_redirect_map) {
4303 /* Note, we cannot use prog directly as imm as subsequent
4304 * rewrites would still change the prog pointer. The only
4305 * stable address we can use is aux, which also works with
4306 * prog clones during blinding.
4307 */
4308 u64 addr = (unsigned long)prog->aux;
4309 struct bpf_insn r4_ld[] = {
4310 BPF_LD_IMM64(BPF_REG_4, addr),
4311 *insn,
4312 };
4313 cnt = ARRAY_SIZE(r4_ld);
4314
4315 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4316 if (!new_prog)
4317 return -ENOMEM;
4318
4319 delta += cnt - 1;
4320 env->prog = prog = new_prog;
4321 insn = new_prog->insnsi + i + delta;
4322 }
4323 patch_call_imm:
4324 fn = prog->aux->ops->get_func_proto(insn->imm);
4325 /* all functions that have prototype and verifier allowed
4326 * programs to call them, must be real in-kernel functions
4327 */
4328 if (!fn->func) {
4329 verbose("kernel subsystem misconfigured func %s#%d\n",
4330 func_id_name(insn->imm), insn->imm);
4331 return -EFAULT;
4332 }
4333 insn->imm = fn->func - __bpf_call_base;
4334 }
4335
4336 return 0;
4337 }
4338
4339 static void free_states(struct bpf_verifier_env *env)
4340 {
4341 struct bpf_verifier_state_list *sl, *sln;
4342 int i;
4343
4344 if (!env->explored_states)
4345 return;
4346
4347 for (i = 0; i < env->prog->len; i++) {
4348 sl = env->explored_states[i];
4349
4350 if (sl)
4351 while (sl != STATE_LIST_MARK) {
4352 sln = sl->next;
4353 kfree(sl);
4354 sl = sln;
4355 }
4356 }
4357
4358 kfree(env->explored_states);
4359 }
4360
4361 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4362 {
4363 char __user *log_ubuf = NULL;
4364 struct bpf_verifier_env *env;
4365 int ret = -EINVAL;
4366
4367 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4368 * allocate/free it every time bpf_check() is called
4369 */
4370 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4371 if (!env)
4372 return -ENOMEM;
4373
4374 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4375 (*prog)->len);
4376 ret = -ENOMEM;
4377 if (!env->insn_aux_data)
4378 goto err_free_env;
4379 env->prog = *prog;
4380
4381 /* grab the mutex to protect few globals used by verifier */
4382 mutex_lock(&bpf_verifier_lock);
4383
4384 if (attr->log_level || attr->log_buf || attr->log_size) {
4385 /* user requested verbose verifier output
4386 * and supplied buffer to store the verification trace
4387 */
4388 log_level = attr->log_level;
4389 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
4390 log_size = attr->log_size;
4391 log_len = 0;
4392
4393 ret = -EINVAL;
4394 /* log_* values have to be sane */
4395 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
4396 log_level == 0 || log_ubuf == NULL)
4397 goto err_unlock;
4398
4399 ret = -ENOMEM;
4400 log_buf = vmalloc(log_size);
4401 if (!log_buf)
4402 goto err_unlock;
4403 } else {
4404 log_level = 0;
4405 }
4406
4407 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4408 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4409 env->strict_alignment = true;
4410
4411 ret = replace_map_fd_with_map_ptr(env);
4412 if (ret < 0)
4413 goto skip_full_check;
4414
4415 env->explored_states = kcalloc(env->prog->len,
4416 sizeof(struct bpf_verifier_state_list *),
4417 GFP_USER);
4418 ret = -ENOMEM;
4419 if (!env->explored_states)
4420 goto skip_full_check;
4421
4422 ret = check_cfg(env);
4423 if (ret < 0)
4424 goto skip_full_check;
4425
4426 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4427
4428 ret = do_check(env);
4429
4430 skip_full_check:
4431 while (pop_stack(env, NULL) >= 0);
4432 free_states(env);
4433
4434 if (ret == 0)
4435 sanitize_dead_code(env);
4436
4437 if (ret == 0)
4438 /* program is valid, convert *(u32*)(ctx + off) accesses */
4439 ret = convert_ctx_accesses(env);
4440
4441 if (ret == 0)
4442 ret = fixup_bpf_calls(env);
4443
4444 if (log_level && log_len >= log_size - 1) {
4445 BUG_ON(log_len >= log_size);
4446 /* verifier log exceeded user supplied buffer */
4447 ret = -ENOSPC;
4448 /* fall through to return what was recorded */
4449 }
4450
4451 /* copy verifier log back to user space including trailing zero */
4452 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
4453 ret = -EFAULT;
4454 goto free_log_buf;
4455 }
4456
4457 if (ret == 0 && env->used_map_cnt) {
4458 /* if program passed verifier, update used_maps in bpf_prog_info */
4459 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4460 sizeof(env->used_maps[0]),
4461 GFP_KERNEL);
4462
4463 if (!env->prog->aux->used_maps) {
4464 ret = -ENOMEM;
4465 goto free_log_buf;
4466 }
4467
4468 memcpy(env->prog->aux->used_maps, env->used_maps,
4469 sizeof(env->used_maps[0]) * env->used_map_cnt);
4470 env->prog->aux->used_map_cnt = env->used_map_cnt;
4471
4472 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4473 * bpf_ld_imm64 instructions
4474 */
4475 convert_pseudo_ld_imm64(env);
4476 }
4477
4478 free_log_buf:
4479 if (log_level)
4480 vfree(log_buf);
4481 if (!env->prog->aux->used_maps)
4482 /* if we didn't copy map pointers into bpf_prog_info, release
4483 * them now. Otherwise free_bpf_prog_info() will release them.
4484 */
4485 release_maps(env);
4486 *prog = env->prog;
4487 err_unlock:
4488 mutex_unlock(&bpf_verifier_lock);
4489 vfree(env->insn_aux_data);
4490 err_free_env:
4491 kfree(env);
4492 return ret;
4493 }
4494
4495 int bpf_analyzer(struct bpf_prog *prog, const struct bpf_ext_analyzer_ops *ops,
4496 void *priv)
4497 {
4498 struct bpf_verifier_env *env;
4499 int ret;
4500
4501 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4502 if (!env)
4503 return -ENOMEM;
4504
4505 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4506 prog->len);
4507 ret = -ENOMEM;
4508 if (!env->insn_aux_data)
4509 goto err_free_env;
4510 env->prog = prog;
4511 env->analyzer_ops = ops;
4512 env->analyzer_priv = priv;
4513
4514 /* grab the mutex to protect few globals used by verifier */
4515 mutex_lock(&bpf_verifier_lock);
4516
4517 log_level = 0;
4518
4519 env->strict_alignment = false;
4520 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4521 env->strict_alignment = true;
4522
4523 env->explored_states = kcalloc(env->prog->len,
4524 sizeof(struct bpf_verifier_state_list *),
4525 GFP_KERNEL);
4526 ret = -ENOMEM;
4527 if (!env->explored_states)
4528 goto skip_full_check;
4529
4530 ret = check_cfg(env);
4531 if (ret < 0)
4532 goto skip_full_check;
4533
4534 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4535
4536 ret = do_check(env);
4537
4538 skip_full_check:
4539 while (pop_stack(env, NULL) >= 0);
4540 free_states(env);
4541
4542 mutex_unlock(&bpf_verifier_lock);
4543 vfree(env->insn_aux_data);
4544 err_free_env:
4545 kfree(env);
4546 return ret;
4547 }
4548 EXPORT_SYMBOL_GPL(bpf_analyzer);
Mirror https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git