]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - kernel/bpf/verifier.c
projects / mirror_ubuntu-bionic-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge tag 'reset-fixes-for-4.14' of git://git.pengutronix.de/git/pza/linux into fixes
[mirror_ubuntu-bionic-kernel.git] / kernel / bpf / verifier.c
1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 */
13 #include <linux/kernel.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
16 #include <linux/bpf.h>
17 #include <linux/bpf_verifier.h>
18 #include <linux/filter.h>
19 #include <net/netlink.h>
20 #include <linux/file.h>
21 #include <linux/vmalloc.h>
22 #include <linux/stringify.h>
23
24 /* bpf_check() is a static code analyzer that walks eBPF program
25 * instruction by instruction and updates register/stack state.
26 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
27 *
28 * The first pass is depth-first-search to check that the program is a DAG.
29 * It rejects the following programs:
30 * - larger than BPF_MAXINSNS insns
31 * - if loop is present (detected via back-edge)
32 * - unreachable insns exist (shouldn't be a forest. program = one function)
33 * - out of bounds or malformed jumps
34 * The second pass is all possible path descent from the 1st insn.
35 * Since it's analyzing all pathes through the program, the length of the
36 * analysis is limited to 64k insn, which may be hit even if total number of
37 * insn is less then 4K, but there are too many branches that change stack/regs.
38 * Number of 'branches to be analyzed' is limited to 1k
39 *
40 * On entry to each instruction, each register has a type, and the instruction
41 * changes the types of the registers depending on instruction semantics.
42 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
43 * copied to R1.
44 *
45 * All registers are 64-bit.
46 * R0 - return register
47 * R1-R5 argument passing registers
48 * R6-R9 callee saved registers
49 * R10 - frame pointer read-only
50 *
51 * At the start of BPF program the register R1 contains a pointer to bpf_context
52 * and has type PTR_TO_CTX.
53 *
54 * Verifier tracks arithmetic operations on pointers in case:
55 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
56 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
57 * 1st insn copies R10 (which has FRAME_PTR) type into R1
58 * and 2nd arithmetic instruction is pattern matched to recognize
59 * that it wants to construct a pointer to some element within stack.
60 * So after 2nd insn, the register R1 has type PTR_TO_STACK
61 * (and -20 constant is saved for further stack bounds checking).
62 * Meaning that this reg is a pointer to stack plus known immediate constant.
63 *
64 * Most of the time the registers have SCALAR_VALUE type, which
65 * means the register has some value, but it's not a valid pointer.
66 * (like pointer plus pointer becomes SCALAR_VALUE type)
67 *
68 * When verifier sees load or store instructions the type of base register
69 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
70 * types recognized by check_mem_access() function.
71 *
72 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
73 * and the range of [ptr, ptr + map's value_size) is accessible.
74 *
75 * registers used to pass values to function calls are checked against
76 * function argument constraints.
77 *
78 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
79 * It means that the register type passed to this function must be
80 * PTR_TO_STACK and it will be used inside the function as
81 * 'pointer to map element key'
82 *
83 * For example the argument constraints for bpf_map_lookup_elem():
84 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
85 * .arg1_type = ARG_CONST_MAP_PTR,
86 * .arg2_type = ARG_PTR_TO_MAP_KEY,
87 *
88 * ret_type says that this function returns 'pointer to map elem value or null'
89 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
90 * 2nd argument should be a pointer to stack, which will be used inside
91 * the helper function as a pointer to map element key.
92 *
93 * On the kernel side the helper function looks like:
94 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
95 * {
96 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
97 * void *key = (void *) (unsigned long) r2;
98 * void *value;
99 *
100 * here kernel can access 'key' and 'map' pointers safely, knowing that
101 * [key, key + map->key_size) bytes are valid and were initialized on
102 * the stack of eBPF program.
103 * }
104 *
105 * Corresponding eBPF program may look like:
106 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
107 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
108 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
109 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
110 * here verifier looks at prototype of map_lookup_elem() and sees:
111 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
112 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
113 *
114 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
115 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
116 * and were initialized prior to this call.
117 * If it's ok, then verifier allows this BPF_CALL insn and looks at
118 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
119 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
120 * returns ether pointer to map value or NULL.
121 *
122 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
123 * insn, the register holding that pointer in the true branch changes state to
124 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
125 * branch. See check_cond_jmp_op().
126 *
127 * After the call R0 is set to return type of the function and registers R1-R5
128 * are set to NOT_INIT to indicate that they are no longer readable.
129 */
130
131 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
132 struct bpf_verifier_stack_elem {
133 /* verifer state is 'st'
134 * before processing instruction 'insn_idx'
135 * and after processing instruction 'prev_insn_idx'
136 */
137 struct bpf_verifier_state st;
138 int insn_idx;
139 int prev_insn_idx;
140 struct bpf_verifier_stack_elem *next;
141 };
142
143 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
144 #define BPF_COMPLEXITY_LIMIT_STACK 1024
145
146 #define BPF_MAP_PTR_POISON ((void *)0xeB9F + POISON_POINTER_DELTA)
147
148 struct bpf_call_arg_meta {
149 struct bpf_map *map_ptr;
150 bool raw_mode;
151 bool pkt_access;
152 int regno;
153 int access_size;
154 };
155
156 /* verbose verifier prints what it's seeing
157 * bpf_check() is called under lock, so no race to access these global vars
158 */
159 static u32 log_level, log_size, log_len;
160 static char *log_buf;
161
162 static DEFINE_MUTEX(bpf_verifier_lock);
163
164 /* log_level controls verbosity level of eBPF verifier.
165 * verbose() is used to dump the verification trace to the log, so the user
166 * can figure out what's wrong with the program
167 */
168 static __printf(1, 2) void verbose(const char *fmt, ...)
169 {
170 va_list args;
171
172 if (log_level == 0 || log_len >= log_size - 1)
173 return;
174
175 va_start(args, fmt);
176 log_len += vscnprintf(log_buf + log_len, log_size - log_len, fmt, args);
177 va_end(args);
178 }
179
180 /* string representation of 'enum bpf_reg_type' */
181 static const char * const reg_type_str[] = {
182 [NOT_INIT] = "?",
183 [SCALAR_VALUE] = "inv",
184 [PTR_TO_CTX] = "ctx",
185 [CONST_PTR_TO_MAP] = "map_ptr",
186 [PTR_TO_MAP_VALUE] = "map_value",
187 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
188 [PTR_TO_STACK] = "fp",
189 [PTR_TO_PACKET] = "pkt",
190 [PTR_TO_PACKET_END] = "pkt_end",
191 };
192
193 #define __BPF_FUNC_STR_FN(x) [BPF_FUNC_ ## x] = __stringify(bpf_ ## x)
194 static const char * const func_id_str[] = {
195 __BPF_FUNC_MAPPER(__BPF_FUNC_STR_FN)
196 };
197 #undef __BPF_FUNC_STR_FN
198
199 static const char *func_id_name(int id)
200 {
201 BUILD_BUG_ON(ARRAY_SIZE(func_id_str) != __BPF_FUNC_MAX_ID);
202
203 if (id >= 0 && id < __BPF_FUNC_MAX_ID && func_id_str[id])
204 return func_id_str[id];
205 else
206 return "unknown";
207 }
208
209 static void print_verifier_state(struct bpf_verifier_state *state)
210 {
211 struct bpf_reg_state *reg;
212 enum bpf_reg_type t;
213 int i;
214
215 for (i = 0; i < MAX_BPF_REG; i++) {
216 reg = &state->regs[i];
217 t = reg->type;
218 if (t == NOT_INIT)
219 continue;
220 verbose(" R%d=%s", i, reg_type_str[t]);
221 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
222 tnum_is_const(reg->var_off)) {
223 /* reg->off should be 0 for SCALAR_VALUE */
224 verbose("%lld", reg->var_off.value + reg->off);
225 } else {
226 verbose("(id=%d", reg->id);
227 if (t != SCALAR_VALUE)
228 verbose(",off=%d", reg->off);
229 if (t == PTR_TO_PACKET)
230 verbose(",r=%d", reg->range);
231 else if (t == CONST_PTR_TO_MAP ||
232 t == PTR_TO_MAP_VALUE ||
233 t == PTR_TO_MAP_VALUE_OR_NULL)
234 verbose(",ks=%d,vs=%d",
235 reg->map_ptr->key_size,
236 reg->map_ptr->value_size);
237 if (tnum_is_const(reg->var_off)) {
238 /* Typically an immediate SCALAR_VALUE, but
239 * could be a pointer whose offset is too big
240 * for reg->off
241 */
242 verbose(",imm=%llx", reg->var_off.value);
243 } else {
244 if (reg->smin_value != reg->umin_value &&
245 reg->smin_value != S64_MIN)
246 verbose(",smin_value=%lld",
247 (long long)reg->smin_value);
248 if (reg->smax_value != reg->umax_value &&
249 reg->smax_value != S64_MAX)
250 verbose(",smax_value=%lld",
251 (long long)reg->smax_value);
252 if (reg->umin_value != 0)
253 verbose(",umin_value=%llu",
254 (unsigned long long)reg->umin_value);
255 if (reg->umax_value != U64_MAX)
256 verbose(",umax_value=%llu",
257 (unsigned long long)reg->umax_value);
258 if (!tnum_is_unknown(reg->var_off)) {
259 char tn_buf[48];
260
261 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
262 verbose(",var_off=%s", tn_buf);
263 }
264 }
265 verbose(")");
266 }
267 }
268 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
269 if (state->stack_slot_type[i] == STACK_SPILL)
270 verbose(" fp%d=%s", -MAX_BPF_STACK + i,
271 reg_type_str[state->spilled_regs[i / BPF_REG_SIZE].type]);
272 }
273 verbose("\n");
274 }
275
276 static const char *const bpf_class_string[] = {
277 [BPF_LD] = "ld",
278 [BPF_LDX] = "ldx",
279 [BPF_ST] = "st",
280 [BPF_STX] = "stx",
281 [BPF_ALU] = "alu",
282 [BPF_JMP] = "jmp",
283 [BPF_RET] = "BUG",
284 [BPF_ALU64] = "alu64",
285 };
286
287 static const char *const bpf_alu_string[16] = {
288 [BPF_ADD >> 4] = "+=",
289 [BPF_SUB >> 4] = "-=",
290 [BPF_MUL >> 4] = "*=",
291 [BPF_DIV >> 4] = "/=",
292 [BPF_OR >> 4] = "|=",
293 [BPF_AND >> 4] = "&=",
294 [BPF_LSH >> 4] = "<<=",
295 [BPF_RSH >> 4] = ">>=",
296 [BPF_NEG >> 4] = "neg",
297 [BPF_MOD >> 4] = "%=",
298 [BPF_XOR >> 4] = "^=",
299 [BPF_MOV >> 4] = "=",
300 [BPF_ARSH >> 4] = "s>>=",
301 [BPF_END >> 4] = "endian",
302 };
303
304 static const char *const bpf_ldst_string[] = {
305 [BPF_W >> 3] = "u32",
306 [BPF_H >> 3] = "u16",
307 [BPF_B >> 3] = "u8",
308 [BPF_DW >> 3] = "u64",
309 };
310
311 static const char *const bpf_jmp_string[16] = {
312 [BPF_JA >> 4] = "jmp",
313 [BPF_JEQ >> 4] = "==",
314 [BPF_JGT >> 4] = ">",
315 [BPF_JLT >> 4] = "<",
316 [BPF_JGE >> 4] = ">=",
317 [BPF_JLE >> 4] = "<=",
318 [BPF_JSET >> 4] = "&",
319 [BPF_JNE >> 4] = "!=",
320 [BPF_JSGT >> 4] = "s>",
321 [BPF_JSLT >> 4] = "s<",
322 [BPF_JSGE >> 4] = "s>=",
323 [BPF_JSLE >> 4] = "s<=",
324 [BPF_CALL >> 4] = "call",
325 [BPF_EXIT >> 4] = "exit",
326 };
327
328 static void print_bpf_insn(const struct bpf_verifier_env *env,
329 const struct bpf_insn *insn)
330 {
331 u8 class = BPF_CLASS(insn->code);
332
333 if (class == BPF_ALU || class == BPF_ALU64) {
334 if (BPF_SRC(insn->code) == BPF_X)
335 verbose("(%02x) %sr%d %s %sr%d\n",
336 insn->code, class == BPF_ALU ? "(u32) " : "",
337 insn->dst_reg,
338 bpf_alu_string[BPF_OP(insn->code) >> 4],
339 class == BPF_ALU ? "(u32) " : "",
340 insn->src_reg);
341 else
342 verbose("(%02x) %sr%d %s %s%d\n",
343 insn->code, class == BPF_ALU ? "(u32) " : "",
344 insn->dst_reg,
345 bpf_alu_string[BPF_OP(insn->code) >> 4],
346 class == BPF_ALU ? "(u32) " : "",
347 insn->imm);
348 } else if (class == BPF_STX) {
349 if (BPF_MODE(insn->code) == BPF_MEM)
350 verbose("(%02x) *(%s *)(r%d %+d) = r%d\n",
351 insn->code,
352 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
353 insn->dst_reg,
354 insn->off, insn->src_reg);
355 else if (BPF_MODE(insn->code) == BPF_XADD)
356 verbose("(%02x) lock *(%s *)(r%d %+d) += r%d\n",
357 insn->code,
358 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
359 insn->dst_reg, insn->off,
360 insn->src_reg);
361 else
362 verbose("BUG_%02x\n", insn->code);
363 } else if (class == BPF_ST) {
364 if (BPF_MODE(insn->code) != BPF_MEM) {
365 verbose("BUG_st_%02x\n", insn->code);
366 return;
367 }
368 verbose("(%02x) *(%s *)(r%d %+d) = %d\n",
369 insn->code,
370 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
371 insn->dst_reg,
372 insn->off, insn->imm);
373 } else if (class == BPF_LDX) {
374 if (BPF_MODE(insn->code) != BPF_MEM) {
375 verbose("BUG_ldx_%02x\n", insn->code);
376 return;
377 }
378 verbose("(%02x) r%d = *(%s *)(r%d %+d)\n",
379 insn->code, insn->dst_reg,
380 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
381 insn->src_reg, insn->off);
382 } else if (class == BPF_LD) {
383 if (BPF_MODE(insn->code) == BPF_ABS) {
384 verbose("(%02x) r0 = *(%s *)skb[%d]\n",
385 insn->code,
386 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
387 insn->imm);
388 } else if (BPF_MODE(insn->code) == BPF_IND) {
389 verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n",
390 insn->code,
391 bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
392 insn->src_reg, insn->imm);
393 } else if (BPF_MODE(insn->code) == BPF_IMM &&
394 BPF_SIZE(insn->code) == BPF_DW) {
395 /* At this point, we already made sure that the second
396 * part of the ldimm64 insn is accessible.
397 */
398 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
399 bool map_ptr = insn->src_reg == BPF_PSEUDO_MAP_FD;
400
401 if (map_ptr && !env->allow_ptr_leaks)
402 imm = 0;
403
404 verbose("(%02x) r%d = 0x%llx\n", insn->code,
405 insn->dst_reg, (unsigned long long)imm);
406 } else {
407 verbose("BUG_ld_%02x\n", insn->code);
408 return;
409 }
410 } else if (class == BPF_JMP) {
411 u8 opcode = BPF_OP(insn->code);
412
413 if (opcode == BPF_CALL) {
414 verbose("(%02x) call %s#%d\n", insn->code,
415 func_id_name(insn->imm), insn->imm);
416 } else if (insn->code == (BPF_JMP | BPF_JA)) {
417 verbose("(%02x) goto pc%+d\n",
418 insn->code, insn->off);
419 } else if (insn->code == (BPF_JMP | BPF_EXIT)) {
420 verbose("(%02x) exit\n", insn->code);
421 } else if (BPF_SRC(insn->code) == BPF_X) {
422 verbose("(%02x) if r%d %s r%d goto pc%+d\n",
423 insn->code, insn->dst_reg,
424 bpf_jmp_string[BPF_OP(insn->code) >> 4],
425 insn->src_reg, insn->off);
426 } else {
427 verbose("(%02x) if r%d %s 0x%x goto pc%+d\n",
428 insn->code, insn->dst_reg,
429 bpf_jmp_string[BPF_OP(insn->code) >> 4],
430 insn->imm, insn->off);
431 }
432 } else {
433 verbose("(%02x) %s\n", insn->code, bpf_class_string[class]);
434 }
435 }
436
437 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx)
438 {
439 struct bpf_verifier_stack_elem *elem;
440 int insn_idx;
441
442 if (env->head == NULL)
443 return -1;
444
445 memcpy(&env->cur_state, &env->head->st, sizeof(env->cur_state));
446 insn_idx = env->head->insn_idx;
447 if (prev_insn_idx)
448 *prev_insn_idx = env->head->prev_insn_idx;
449 elem = env->head->next;
450 kfree(env->head);
451 env->head = elem;
452 env->stack_size--;
453 return insn_idx;
454 }
455
456 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
457 int insn_idx, int prev_insn_idx)
458 {
459 struct bpf_verifier_stack_elem *elem;
460
461 elem = kmalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
462 if (!elem)
463 goto err;
464
465 memcpy(&elem->st, &env->cur_state, sizeof(env->cur_state));
466 elem->insn_idx = insn_idx;
467 elem->prev_insn_idx = prev_insn_idx;
468 elem->next = env->head;
469 env->head = elem;
470 env->stack_size++;
471 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
472 verbose("BPF program is too complex\n");
473 goto err;
474 }
475 return &elem->st;
476 err:
477 /* pop all elements and return */
478 while (pop_stack(env, NULL) >= 0);
479 return NULL;
480 }
481
482 #define CALLER_SAVED_REGS 6
483 static const int caller_saved[CALLER_SAVED_REGS] = {
484 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
485 };
486
487 static void __mark_reg_not_init(struct bpf_reg_state *reg);
488
489 /* Mark the unknown part of a register (variable offset or scalar value) as
490 * known to have the value @imm.
491 */
492 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
493 {
494 reg->id = 0;
495 reg->var_off = tnum_const(imm);
496 reg->smin_value = (s64)imm;
497 reg->smax_value = (s64)imm;
498 reg->umin_value = imm;
499 reg->umax_value = imm;
500 }
501
502 /* Mark the 'variable offset' part of a register as zero. This should be
503 * used only on registers holding a pointer type.
504 */
505 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
506 {
507 __mark_reg_known(reg, 0);
508 }
509
510 static void mark_reg_known_zero(struct bpf_reg_state *regs, u32 regno)
511 {
512 if (WARN_ON(regno >= MAX_BPF_REG)) {
513 verbose("mark_reg_known_zero(regs, %u)\n", regno);
514 /* Something bad happened, let's kill all regs */
515 for (regno = 0; regno < MAX_BPF_REG; regno++)
516 __mark_reg_not_init(regs + regno);
517 return;
518 }
519 __mark_reg_known_zero(regs + regno);
520 }
521
522 /* Attempts to improve min/max values based on var_off information */
523 static void __update_reg_bounds(struct bpf_reg_state *reg)
524 {
525 /* min signed is max(sign bit) | min(other bits) */
526 reg->smin_value = max_t(s64, reg->smin_value,
527 reg->var_off.value | (reg->var_off.mask & S64_MIN));
528 /* max signed is min(sign bit) | max(other bits) */
529 reg->smax_value = min_t(s64, reg->smax_value,
530 reg->var_off.value | (reg->var_off.mask & S64_MAX));
531 reg->umin_value = max(reg->umin_value, reg->var_off.value);
532 reg->umax_value = min(reg->umax_value,
533 reg->var_off.value | reg->var_off.mask);
534 }
535
536 /* Uses signed min/max values to inform unsigned, and vice-versa */
537 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
538 {
539 /* Learn sign from signed bounds.
540 * If we cannot cross the sign boundary, then signed and unsigned bounds
541 * are the same, so combine. This works even in the negative case, e.g.
542 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
543 */
544 if (reg->smin_value >= 0 || reg->smax_value < 0) {
545 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
546 reg->umin_value);
547 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
548 reg->umax_value);
549 return;
550 }
551 /* Learn sign from unsigned bounds. Signed bounds cross the sign
552 * boundary, so we must be careful.
553 */
554 if ((s64)reg->umax_value >= 0) {
555 /* Positive. We can't learn anything from the smin, but smax
556 * is positive, hence safe.
557 */
558 reg->smin_value = reg->umin_value;
559 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
560 reg->umax_value);
561 } else if ((s64)reg->umin_value < 0) {
562 /* Negative. We can't learn anything from the smax, but smin
563 * is negative, hence safe.
564 */
565 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
566 reg->umin_value);
567 reg->smax_value = reg->umax_value;
568 }
569 }
570
571 /* Attempts to improve var_off based on unsigned min/max information */
572 static void __reg_bound_offset(struct bpf_reg_state *reg)
573 {
574 reg->var_off = tnum_intersect(reg->var_off,
575 tnum_range(reg->umin_value,
576 reg->umax_value));
577 }
578
579 /* Reset the min/max bounds of a register */
580 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
581 {
582 reg->smin_value = S64_MIN;
583 reg->smax_value = S64_MAX;
584 reg->umin_value = 0;
585 reg->umax_value = U64_MAX;
586 }
587
588 /* Mark a register as having a completely unknown (scalar) value. */
589 static void __mark_reg_unknown(struct bpf_reg_state *reg)
590 {
591 reg->type = SCALAR_VALUE;
592 reg->id = 0;
593 reg->off = 0;
594 reg->var_off = tnum_unknown;
595 __mark_reg_unbounded(reg);
596 }
597
598 static void mark_reg_unknown(struct bpf_reg_state *regs, u32 regno)
599 {
600 if (WARN_ON(regno >= MAX_BPF_REG)) {
601 verbose("mark_reg_unknown(regs, %u)\n", regno);
602 /* Something bad happened, let's kill all regs */
603 for (regno = 0; regno < MAX_BPF_REG; regno++)
604 __mark_reg_not_init(regs + regno);
605 return;
606 }
607 __mark_reg_unknown(regs + regno);
608 }
609
610 static void __mark_reg_not_init(struct bpf_reg_state *reg)
611 {
612 __mark_reg_unknown(reg);
613 reg->type = NOT_INIT;
614 }
615
616 static void mark_reg_not_init(struct bpf_reg_state *regs, u32 regno)
617 {
618 if (WARN_ON(regno >= MAX_BPF_REG)) {
619 verbose("mark_reg_not_init(regs, %u)\n", regno);
620 /* Something bad happened, let's kill all regs */
621 for (regno = 0; regno < MAX_BPF_REG; regno++)
622 __mark_reg_not_init(regs + regno);
623 return;
624 }
625 __mark_reg_not_init(regs + regno);
626 }
627
628 static void init_reg_state(struct bpf_reg_state *regs)
629 {
630 int i;
631
632 for (i = 0; i < MAX_BPF_REG; i++) {
633 mark_reg_not_init(regs, i);
634 regs[i].live = REG_LIVE_NONE;
635 }
636
637 /* frame pointer */
638 regs[BPF_REG_FP].type = PTR_TO_STACK;
639 mark_reg_known_zero(regs, BPF_REG_FP);
640
641 /* 1st arg to a function */
642 regs[BPF_REG_1].type = PTR_TO_CTX;
643 mark_reg_known_zero(regs, BPF_REG_1);
644 }
645
646 enum reg_arg_type {
647 SRC_OP, /* register is used as source operand */
648 DST_OP, /* register is used as destination operand */
649 DST_OP_NO_MARK /* same as above, check only, don't mark */
650 };
651
652 static void mark_reg_read(const struct bpf_verifier_state *state, u32 regno)
653 {
654 struct bpf_verifier_state *parent = state->parent;
655
656 while (parent) {
657 /* if read wasn't screened by an earlier write ... */
658 if (state->regs[regno].live & REG_LIVE_WRITTEN)
659 break;
660 /* ... then we depend on parent's value */
661 parent->regs[regno].live |= REG_LIVE_READ;
662 state = parent;
663 parent = state->parent;
664 }
665 }
666
667 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
668 enum reg_arg_type t)
669 {
670 struct bpf_reg_state *regs = env->cur_state.regs;
671
672 if (regno >= MAX_BPF_REG) {
673 verbose("R%d is invalid\n", regno);
674 return -EINVAL;
675 }
676
677 if (t == SRC_OP) {
678 /* check whether register used as source operand can be read */
679 if (regs[regno].type == NOT_INIT) {
680 verbose("R%d !read_ok\n", regno);
681 return -EACCES;
682 }
683 mark_reg_read(&env->cur_state, regno);
684 } else {
685 /* check whether register used as dest operand can be written to */
686 if (regno == BPF_REG_FP) {
687 verbose("frame pointer is read only\n");
688 return -EACCES;
689 }
690 regs[regno].live |= REG_LIVE_WRITTEN;
691 if (t == DST_OP)
692 mark_reg_unknown(regs, regno);
693 }
694 return 0;
695 }
696
697 static bool is_spillable_regtype(enum bpf_reg_type type)
698 {
699 switch (type) {
700 case PTR_TO_MAP_VALUE:
701 case PTR_TO_MAP_VALUE_OR_NULL:
702 case PTR_TO_STACK:
703 case PTR_TO_CTX:
704 case PTR_TO_PACKET:
705 case PTR_TO_PACKET_END:
706 case CONST_PTR_TO_MAP:
707 return true;
708 default:
709 return false;
710 }
711 }
712
713 /* check_stack_read/write functions track spill/fill of registers,
714 * stack boundary and alignment are checked in check_mem_access()
715 */
716 static int check_stack_write(struct bpf_verifier_state *state, int off,
717 int size, int value_regno)
718 {
719 int i, spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
720 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
721 * so it's aligned access and [off, off + size) are within stack limits
722 */
723
724 if (value_regno >= 0 &&
725 is_spillable_regtype(state->regs[value_regno].type)) {
726
727 /* register containing pointer is being spilled into stack */
728 if (size != BPF_REG_SIZE) {
729 verbose("invalid size of register spill\n");
730 return -EACCES;
731 }
732
733 /* save register state */
734 state->spilled_regs[spi] = state->regs[value_regno];
735 state->spilled_regs[spi].live |= REG_LIVE_WRITTEN;
736
737 for (i = 0; i < BPF_REG_SIZE; i++)
738 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL;
739 } else {
740 /* regular write of data into stack */
741 state->spilled_regs[spi] = (struct bpf_reg_state) {};
742
743 for (i = 0; i < size; i++)
744 state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC;
745 }
746 return 0;
747 }
748
749 static void mark_stack_slot_read(const struct bpf_verifier_state *state, int slot)
750 {
751 struct bpf_verifier_state *parent = state->parent;
752
753 while (parent) {
754 /* if read wasn't screened by an earlier write ... */
755 if (state->spilled_regs[slot].live & REG_LIVE_WRITTEN)
756 break;
757 /* ... then we depend on parent's value */
758 parent->spilled_regs[slot].live |= REG_LIVE_READ;
759 state = parent;
760 parent = state->parent;
761 }
762 }
763
764 static int check_stack_read(struct bpf_verifier_state *state, int off, int size,
765 int value_regno)
766 {
767 u8 *slot_type;
768 int i, spi;
769
770 slot_type = &state->stack_slot_type[MAX_BPF_STACK + off];
771
772 if (slot_type[0] == STACK_SPILL) {
773 if (size != BPF_REG_SIZE) {
774 verbose("invalid size of register spill\n");
775 return -EACCES;
776 }
777 for (i = 1; i < BPF_REG_SIZE; i++) {
778 if (slot_type[i] != STACK_SPILL) {
779 verbose("corrupted spill memory\n");
780 return -EACCES;
781 }
782 }
783
784 spi = (MAX_BPF_STACK + off) / BPF_REG_SIZE;
785
786 if (value_regno >= 0) {
787 /* restore register state from stack */
788 state->regs[value_regno] = state->spilled_regs[spi];
789 mark_stack_slot_read(state, spi);
790 }
791 return 0;
792 } else {
793 for (i = 0; i < size; i++) {
794 if (slot_type[i] != STACK_MISC) {
795 verbose("invalid read from stack off %d+%d size %d\n",
796 off, i, size);
797 return -EACCES;
798 }
799 }
800 if (value_regno >= 0)
801 /* have read misc data from the stack */
802 mark_reg_unknown(state->regs, value_regno);
803 return 0;
804 }
805 }
806
807 /* check read/write into map element returned by bpf_map_lookup_elem() */
808 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
809 int size)
810 {
811 struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
812
813 if (off < 0 || size <= 0 || off + size > map->value_size) {
814 verbose("invalid access to map value, value_size=%d off=%d size=%d\n",
815 map->value_size, off, size);
816 return -EACCES;
817 }
818 return 0;
819 }
820
821 /* check read/write into a map element with possible variable offset */
822 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
823 int off, int size)
824 {
825 struct bpf_verifier_state *state = &env->cur_state;
826 struct bpf_reg_state *reg = &state->regs[regno];
827 int err;
828
829 /* We may have adjusted the register to this map value, so we
830 * need to try adding each of min_value and max_value to off
831 * to make sure our theoretical access will be safe.
832 */
833 if (log_level)
834 print_verifier_state(state);
835 /* The minimum value is only important with signed
836 * comparisons where we can't assume the floor of a
837 * value is 0. If we are using signed variables for our
838 * index'es we need to make sure that whatever we use
839 * will have a set floor within our range.
840 */
841 if (reg->smin_value < 0) {
842 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
843 regno);
844 return -EACCES;
845 }
846 err = __check_map_access(env, regno, reg->smin_value + off, size);
847 if (err) {
848 verbose("R%d min value is outside of the array range\n", regno);
849 return err;
850 }
851
852 /* If we haven't set a max value then we need to bail since we can't be
853 * sure we won't do bad things.
854 * If reg->umax_value + off could overflow, treat that as unbounded too.
855 */
856 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
857 verbose("R%d unbounded memory access, make sure to bounds check any array access into a map\n",
858 regno);
859 return -EACCES;
860 }
861 err = __check_map_access(env, regno, reg->umax_value + off, size);
862 if (err)
863 verbose("R%d max value is outside of the array range\n", regno);
864 return err;
865 }
866
867 #define MAX_PACKET_OFF 0xffff
868
869 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
870 const struct bpf_call_arg_meta *meta,
871 enum bpf_access_type t)
872 {
873 switch (env->prog->type) {
874 case BPF_PROG_TYPE_LWT_IN:
875 case BPF_PROG_TYPE_LWT_OUT:
876 /* dst_input() and dst_output() can't write for now */
877 if (t == BPF_WRITE)
878 return false;
879 /* fallthrough */
880 case BPF_PROG_TYPE_SCHED_CLS:
881 case BPF_PROG_TYPE_SCHED_ACT:
882 case BPF_PROG_TYPE_XDP:
883 case BPF_PROG_TYPE_LWT_XMIT:
884 case BPF_PROG_TYPE_SK_SKB:
885 if (meta)
886 return meta->pkt_access;
887
888 env->seen_direct_write = true;
889 return true;
890 default:
891 return false;
892 }
893 }
894
895 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
896 int off, int size)
897 {
898 struct bpf_reg_state *regs = env->cur_state.regs;
899 struct bpf_reg_state *reg = &regs[regno];
900
901 if (off < 0 || size <= 0 || (u64)off + size > reg->range) {
902 verbose("invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
903 off, size, regno, reg->id, reg->off, reg->range);
904 return -EACCES;
905 }
906 return 0;
907 }
908
909 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
910 int size)
911 {
912 struct bpf_reg_state *regs = env->cur_state.regs;
913 struct bpf_reg_state *reg = &regs[regno];
914 int err;
915
916 /* We may have added a variable offset to the packet pointer; but any
917 * reg->range we have comes after that. We are only checking the fixed
918 * offset.
919 */
920
921 /* We don't allow negative numbers, because we aren't tracking enough
922 * detail to prove they're safe.
923 */
924 if (reg->smin_value < 0) {
925 verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
926 regno);
927 return -EACCES;
928 }
929 err = __check_packet_access(env, regno, off, size);
930 if (err) {
931 verbose("R%d offset is outside of the packet\n", regno);
932 return err;
933 }
934 return err;
935 }
936
937 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
938 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
939 enum bpf_access_type t, enum bpf_reg_type *reg_type)
940 {
941 struct bpf_insn_access_aux info = {
942 .reg_type = *reg_type,
943 };
944
945 /* for analyzer ctx accesses are already validated and converted */
946 if (env->analyzer_ops)
947 return 0;
948
949 if (env->prog->aux->ops->is_valid_access &&
950 env->prog->aux->ops->is_valid_access(off, size, t, &info)) {
951 /* A non zero info.ctx_field_size indicates that this field is a
952 * candidate for later verifier transformation to load the whole
953 * field and then apply a mask when accessed with a narrower
954 * access than actual ctx access size. A zero info.ctx_field_size
955 * will only allow for whole field access and rejects any other
956 * type of narrower access.
957 */
958 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
959 *reg_type = info.reg_type;
960
961 /* remember the offset of last byte accessed in ctx */
962 if (env->prog->aux->max_ctx_offset < off + size)
963 env->prog->aux->max_ctx_offset = off + size;
964 return 0;
965 }
966
967 verbose("invalid bpf_context access off=%d size=%d\n", off, size);
968 return -EACCES;
969 }
970
971 static bool __is_pointer_value(bool allow_ptr_leaks,
972 const struct bpf_reg_state *reg)
973 {
974 if (allow_ptr_leaks)
975 return false;
976
977 return reg->type != SCALAR_VALUE;
978 }
979
980 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
981 {
982 return __is_pointer_value(env->allow_ptr_leaks, &env->cur_state.regs[regno]);
983 }
984
985 static int check_pkt_ptr_alignment(const struct bpf_reg_state *reg,
986 int off, int size, bool strict)
987 {
988 struct tnum reg_off;
989 int ip_align;
990
991 /* Byte size accesses are always allowed. */
992 if (!strict || size == 1)
993 return 0;
994
995 /* For platforms that do not have a Kconfig enabling
996 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
997 * NET_IP_ALIGN is universally set to '2'. And on platforms
998 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
999 * to this code only in strict mode where we want to emulate
1000 * the NET_IP_ALIGN==2 checking. Therefore use an
1001 * unconditional IP align value of '2'.
1002 */
1003 ip_align = 2;
1004
1005 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1006 if (!tnum_is_aligned(reg_off, size)) {
1007 char tn_buf[48];
1008
1009 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1010 verbose("misaligned packet access off %d+%s+%d+%d size %d\n",
1011 ip_align, tn_buf, reg->off, off, size);
1012 return -EACCES;
1013 }
1014
1015 return 0;
1016 }
1017
1018 static int check_generic_ptr_alignment(const struct bpf_reg_state *reg,
1019 const char *pointer_desc,
1020 int off, int size, bool strict)
1021 {
1022 struct tnum reg_off;
1023
1024 /* Byte size accesses are always allowed. */
1025 if (!strict || size == 1)
1026 return 0;
1027
1028 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1029 if (!tnum_is_aligned(reg_off, size)) {
1030 char tn_buf[48];
1031
1032 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1033 verbose("misaligned %saccess off %s+%d+%d size %d\n",
1034 pointer_desc, tn_buf, reg->off, off, size);
1035 return -EACCES;
1036 }
1037
1038 return 0;
1039 }
1040
1041 static int check_ptr_alignment(struct bpf_verifier_env *env,
1042 const struct bpf_reg_state *reg,
1043 int off, int size)
1044 {
1045 bool strict = env->strict_alignment;
1046 const char *pointer_desc = "";
1047
1048 switch (reg->type) {
1049 case PTR_TO_PACKET:
1050 /* special case, because of NET_IP_ALIGN */
1051 return check_pkt_ptr_alignment(reg, off, size, strict);
1052 case PTR_TO_MAP_VALUE:
1053 pointer_desc = "value ";
1054 break;
1055 case PTR_TO_CTX:
1056 pointer_desc = "context ";
1057 break;
1058 case PTR_TO_STACK:
1059 pointer_desc = "stack ";
1060 break;
1061 default:
1062 break;
1063 }
1064 return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
1065 }
1066
1067 /* check whether memory at (regno + off) is accessible for t = (read | write)
1068 * if t==write, value_regno is a register which value is stored into memory
1069 * if t==read, value_regno is a register which will receive the value from memory
1070 * if t==write && value_regno==-1, some unknown value is stored into memory
1071 * if t==read && value_regno==-1, don't care what we read from memory
1072 */
1073 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1074 int bpf_size, enum bpf_access_type t,
1075 int value_regno)
1076 {
1077 struct bpf_verifier_state *state = &env->cur_state;
1078 struct bpf_reg_state *reg = &state->regs[regno];
1079 int size, err = 0;
1080
1081 size = bpf_size_to_bytes(bpf_size);
1082 if (size < 0)
1083 return size;
1084
1085 /* alignment checks will add in reg->off themselves */
1086 err = check_ptr_alignment(env, reg, off, size);
1087 if (err)
1088 return err;
1089
1090 /* for access checks, reg->off is just part of off */
1091 off += reg->off;
1092
1093 if (reg->type == PTR_TO_MAP_VALUE) {
1094 if (t == BPF_WRITE && value_regno >= 0 &&
1095 is_pointer_value(env, value_regno)) {
1096 verbose("R%d leaks addr into map\n", value_regno);
1097 return -EACCES;
1098 }
1099
1100 err = check_map_access(env, regno, off, size);
1101 if (!err && t == BPF_READ && value_regno >= 0)
1102 mark_reg_unknown(state->regs, value_regno);
1103
1104 } else if (reg->type == PTR_TO_CTX) {
1105 enum bpf_reg_type reg_type = SCALAR_VALUE;
1106
1107 if (t == BPF_WRITE && value_regno >= 0 &&
1108 is_pointer_value(env, value_regno)) {
1109 verbose("R%d leaks addr into ctx\n", value_regno);
1110 return -EACCES;
1111 }
1112 /* ctx accesses must be at a fixed offset, so that we can
1113 * determine what type of data were returned.
1114 */
1115 if (!tnum_is_const(reg->var_off)) {
1116 char tn_buf[48];
1117
1118 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1119 verbose("variable ctx access var_off=%s off=%d size=%d",
1120 tn_buf, off, size);
1121 return -EACCES;
1122 }
1123 off += reg->var_off.value;
1124 err = check_ctx_access(env, insn_idx, off, size, t, &reg_type);
1125 if (!err && t == BPF_READ && value_regno >= 0) {
1126 /* ctx access returns either a scalar, or a
1127 * PTR_TO_PACKET[_END]. In the latter case, we know
1128 * the offset is zero.
1129 */
1130 if (reg_type == SCALAR_VALUE)
1131 mark_reg_unknown(state->regs, value_regno);
1132 else
1133 mark_reg_known_zero(state->regs, value_regno);
1134 state->regs[value_regno].id = 0;
1135 state->regs[value_regno].off = 0;
1136 state->regs[value_regno].range = 0;
1137 state->regs[value_regno].type = reg_type;
1138 }
1139
1140 } else if (reg->type == PTR_TO_STACK) {
1141 /* stack accesses must be at a fixed offset, so that we can
1142 * determine what type of data were returned.
1143 * See check_stack_read().
1144 */
1145 if (!tnum_is_const(reg->var_off)) {
1146 char tn_buf[48];
1147
1148 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1149 verbose("variable stack access var_off=%s off=%d size=%d",
1150 tn_buf, off, size);
1151 return -EACCES;
1152 }
1153 off += reg->var_off.value;
1154 if (off >= 0 || off < -MAX_BPF_STACK) {
1155 verbose("invalid stack off=%d size=%d\n", off, size);
1156 return -EACCES;
1157 }
1158
1159 if (env->prog->aux->stack_depth < -off)
1160 env->prog->aux->stack_depth = -off;
1161
1162 if (t == BPF_WRITE) {
1163 if (!env->allow_ptr_leaks &&
1164 state->stack_slot_type[MAX_BPF_STACK + off] == STACK_SPILL &&
1165 size != BPF_REG_SIZE) {
1166 verbose("attempt to corrupt spilled pointer on stack\n");
1167 return -EACCES;
1168 }
1169 err = check_stack_write(state, off, size, value_regno);
1170 } else {
1171 err = check_stack_read(state, off, size, value_regno);
1172 }
1173 } else if (reg->type == PTR_TO_PACKET) {
1174 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1175 verbose("cannot write into packet\n");
1176 return -EACCES;
1177 }
1178 if (t == BPF_WRITE && value_regno >= 0 &&
1179 is_pointer_value(env, value_regno)) {
1180 verbose("R%d leaks addr into packet\n", value_regno);
1181 return -EACCES;
1182 }
1183 err = check_packet_access(env, regno, off, size);
1184 if (!err && t == BPF_READ && value_regno >= 0)
1185 mark_reg_unknown(state->regs, value_regno);
1186 } else {
1187 verbose("R%d invalid mem access '%s'\n",
1188 regno, reg_type_str[reg->type]);
1189 return -EACCES;
1190 }
1191
1192 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1193 state->regs[value_regno].type == SCALAR_VALUE) {
1194 /* b/h/w load zero-extends, mark upper bits as known 0 */
1195 state->regs[value_regno].var_off = tnum_cast(
1196 state->regs[value_regno].var_off, size);
1197 __update_reg_bounds(&state->regs[value_regno]);
1198 }
1199 return err;
1200 }
1201
1202 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1203 {
1204 int err;
1205
1206 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1207 insn->imm != 0) {
1208 verbose("BPF_XADD uses reserved fields\n");
1209 return -EINVAL;
1210 }
1211
1212 /* check src1 operand */
1213 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1214 if (err)
1215 return err;
1216
1217 /* check src2 operand */
1218 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1219 if (err)
1220 return err;
1221
1222 if (is_pointer_value(env, insn->src_reg)) {
1223 verbose("R%d leaks addr into mem\n", insn->src_reg);
1224 return -EACCES;
1225 }
1226
1227 /* check whether atomic_add can read the memory */
1228 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1229 BPF_SIZE(insn->code), BPF_READ, -1);
1230 if (err)
1231 return err;
1232
1233 /* check whether atomic_add can write into the same memory */
1234 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1235 BPF_SIZE(insn->code), BPF_WRITE, -1);
1236 }
1237
1238 /* Does this register contain a constant zero? */
1239 static bool register_is_null(struct bpf_reg_state reg)
1240 {
1241 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1242 }
1243
1244 /* when register 'regno' is passed into function that will read 'access_size'
1245 * bytes from that pointer, make sure that it's within stack boundary
1246 * and all elements of stack are initialized.
1247 * Unlike most pointer bounds-checking functions, this one doesn't take an
1248 * 'off' argument, so it has to add in reg->off itself.
1249 */
1250 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1251 int access_size, bool zero_size_allowed,
1252 struct bpf_call_arg_meta *meta)
1253 {
1254 struct bpf_verifier_state *state = &env->cur_state;
1255 struct bpf_reg_state *regs = state->regs;
1256 int off, i;
1257
1258 if (regs[regno].type != PTR_TO_STACK) {
1259 /* Allow zero-byte read from NULL, regardless of pointer type */
1260 if (zero_size_allowed && access_size == 0 &&
1261 register_is_null(regs[regno]))
1262 return 0;
1263
1264 verbose("R%d type=%s expected=%s\n", regno,
1265 reg_type_str[regs[regno].type],
1266 reg_type_str[PTR_TO_STACK]);
1267 return -EACCES;
1268 }
1269
1270 /* Only allow fixed-offset stack reads */
1271 if (!tnum_is_const(regs[regno].var_off)) {
1272 char tn_buf[48];
1273
1274 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1275 verbose("invalid variable stack read R%d var_off=%s\n",
1276 regno, tn_buf);
1277 }
1278 off = regs[regno].off + regs[regno].var_off.value;
1279 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1280 access_size <= 0) {
1281 verbose("invalid stack type R%d off=%d access_size=%d\n",
1282 regno, off, access_size);
1283 return -EACCES;
1284 }
1285
1286 if (env->prog->aux->stack_depth < -off)
1287 env->prog->aux->stack_depth = -off;
1288
1289 if (meta && meta->raw_mode) {
1290 meta->access_size = access_size;
1291 meta->regno = regno;
1292 return 0;
1293 }
1294
1295 for (i = 0; i < access_size; i++) {
1296 if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
1297 verbose("invalid indirect read from stack off %d+%d size %d\n",
1298 off, i, access_size);
1299 return -EACCES;
1300 }
1301 }
1302 return 0;
1303 }
1304
1305 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1306 int access_size, bool zero_size_allowed,
1307 struct bpf_call_arg_meta *meta)
1308 {
1309 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1310
1311 switch (reg->type) {
1312 case PTR_TO_PACKET:
1313 return check_packet_access(env, regno, reg->off, access_size);
1314 case PTR_TO_MAP_VALUE:
1315 return check_map_access(env, regno, reg->off, access_size);
1316 default: /* scalar_value|ptr_to_stack or invalid ptr */
1317 return check_stack_boundary(env, regno, access_size,
1318 zero_size_allowed, meta);
1319 }
1320 }
1321
1322 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1323 enum bpf_arg_type arg_type,
1324 struct bpf_call_arg_meta *meta)
1325 {
1326 struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
1327 enum bpf_reg_type expected_type, type = reg->type;
1328 int err = 0;
1329
1330 if (arg_type == ARG_DONTCARE)
1331 return 0;
1332
1333 err = check_reg_arg(env, regno, SRC_OP);
1334 if (err)
1335 return err;
1336
1337 if (arg_type == ARG_ANYTHING) {
1338 if (is_pointer_value(env, regno)) {
1339 verbose("R%d leaks addr into helper function\n", regno);
1340 return -EACCES;
1341 }
1342 return 0;
1343 }
1344
1345 if (type == PTR_TO_PACKET &&
1346 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1347 verbose("helper access to the packet is not allowed\n");
1348 return -EACCES;
1349 }
1350
1351 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1352 arg_type == ARG_PTR_TO_MAP_VALUE) {
1353 expected_type = PTR_TO_STACK;
1354 if (type != PTR_TO_PACKET && type != expected_type)
1355 goto err_type;
1356 } else if (arg_type == ARG_CONST_SIZE ||
1357 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1358 expected_type = SCALAR_VALUE;
1359 if (type != expected_type)
1360 goto err_type;
1361 } else if (arg_type == ARG_CONST_MAP_PTR) {
1362 expected_type = CONST_PTR_TO_MAP;
1363 if (type != expected_type)
1364 goto err_type;
1365 } else if (arg_type == ARG_PTR_TO_CTX) {
1366 expected_type = PTR_TO_CTX;
1367 if (type != expected_type)
1368 goto err_type;
1369 } else if (arg_type == ARG_PTR_TO_MEM ||
1370 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1371 expected_type = PTR_TO_STACK;
1372 /* One exception here. In case function allows for NULL to be
1373 * passed in as argument, it's a SCALAR_VALUE type. Final test
1374 * happens during stack boundary checking.
1375 */
1376 if (register_is_null(*reg))
1377 /* final test in check_stack_boundary() */;
1378 else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
1379 type != expected_type)
1380 goto err_type;
1381 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1382 } else {
1383 verbose("unsupported arg_type %d\n", arg_type);
1384 return -EFAULT;
1385 }
1386
1387 if (arg_type == ARG_CONST_MAP_PTR) {
1388 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1389 meta->map_ptr = reg->map_ptr;
1390 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1391 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1392 * check that [key, key + map->key_size) are within
1393 * stack limits and initialized
1394 */
1395 if (!meta->map_ptr) {
1396 /* in function declaration map_ptr must come before
1397 * map_key, so that it's verified and known before
1398 * we have to check map_key here. Otherwise it means
1399 * that kernel subsystem misconfigured verifier
1400 */
1401 verbose("invalid map_ptr to access map->key\n");
1402 return -EACCES;
1403 }
1404 if (type == PTR_TO_PACKET)
1405 err = check_packet_access(env, regno, reg->off,
1406 meta->map_ptr->key_size);
1407 else
1408 err = check_stack_boundary(env, regno,
1409 meta->map_ptr->key_size,
1410 false, NULL);
1411 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1412 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1413 * check [value, value + map->value_size) validity
1414 */
1415 if (!meta->map_ptr) {
1416 /* kernel subsystem misconfigured verifier */
1417 verbose("invalid map_ptr to access map->value\n");
1418 return -EACCES;
1419 }
1420 if (type == PTR_TO_PACKET)
1421 err = check_packet_access(env, regno, reg->off,
1422 meta->map_ptr->value_size);
1423 else
1424 err = check_stack_boundary(env, regno,
1425 meta->map_ptr->value_size,
1426 false, NULL);
1427 } else if (arg_type == ARG_CONST_SIZE ||
1428 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1429 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1430
1431 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1432 * from stack pointer 'buf'. Check it
1433 * note: regno == len, regno - 1 == buf
1434 */
1435 if (regno == 0) {
1436 /* kernel subsystem misconfigured verifier */
1437 verbose("ARG_CONST_SIZE cannot be first argument\n");
1438 return -EACCES;
1439 }
1440
1441 /* The register is SCALAR_VALUE; the access check
1442 * happens using its boundaries.
1443 */
1444
1445 if (!tnum_is_const(reg->var_off))
1446 /* For unprivileged variable accesses, disable raw
1447 * mode so that the program is required to
1448 * initialize all the memory that the helper could
1449 * just partially fill up.
1450 */
1451 meta = NULL;
1452
1453 if (reg->smin_value < 0) {
1454 verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
1455 regno);
1456 return -EACCES;
1457 }
1458
1459 if (reg->umin_value == 0) {
1460 err = check_helper_mem_access(env, regno - 1, 0,
1461 zero_size_allowed,
1462 meta);
1463 if (err)
1464 return err;
1465 }
1466
1467 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1468 verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1469 regno);
1470 return -EACCES;
1471 }
1472 err = check_helper_mem_access(env, regno - 1,
1473 reg->umax_value,
1474 zero_size_allowed, meta);
1475 }
1476
1477 return err;
1478 err_type:
1479 verbose("R%d type=%s expected=%s\n", regno,
1480 reg_type_str[type], reg_type_str[expected_type]);
1481 return -EACCES;
1482 }
1483
1484 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
1485 {
1486 if (!map)
1487 return 0;
1488
1489 /* We need a two way check, first is from map perspective ... */
1490 switch (map->map_type) {
1491 case BPF_MAP_TYPE_PROG_ARRAY:
1492 if (func_id != BPF_FUNC_tail_call)
1493 goto error;
1494 break;
1495 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1496 if (func_id != BPF_FUNC_perf_event_read &&
1497 func_id != BPF_FUNC_perf_event_output)
1498 goto error;
1499 break;
1500 case BPF_MAP_TYPE_STACK_TRACE:
1501 if (func_id != BPF_FUNC_get_stackid)
1502 goto error;
1503 break;
1504 case BPF_MAP_TYPE_CGROUP_ARRAY:
1505 if (func_id != BPF_FUNC_skb_under_cgroup &&
1506 func_id != BPF_FUNC_current_task_under_cgroup)
1507 goto error;
1508 break;
1509 /* devmap returns a pointer to a live net_device ifindex that we cannot
1510 * allow to be modified from bpf side. So do not allow lookup elements
1511 * for now.
1512 */
1513 case BPF_MAP_TYPE_DEVMAP:
1514 if (func_id != BPF_FUNC_redirect_map)
1515 goto error;
1516 break;
1517 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1518 case BPF_MAP_TYPE_HASH_OF_MAPS:
1519 if (func_id != BPF_FUNC_map_lookup_elem)
1520 goto error;
1521 break;
1522 case BPF_MAP_TYPE_SOCKMAP:
1523 if (func_id != BPF_FUNC_sk_redirect_map &&
1524 func_id != BPF_FUNC_sock_map_update &&
1525 func_id != BPF_FUNC_map_delete_elem)
1526 goto error;
1527 break;
1528 default:
1529 break;
1530 }
1531
1532 /* ... and second from the function itself. */
1533 switch (func_id) {
1534 case BPF_FUNC_tail_call:
1535 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1536 goto error;
1537 break;
1538 case BPF_FUNC_perf_event_read:
1539 case BPF_FUNC_perf_event_output:
1540 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1541 goto error;
1542 break;
1543 case BPF_FUNC_get_stackid:
1544 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1545 goto error;
1546 break;
1547 case BPF_FUNC_current_task_under_cgroup:
1548 case BPF_FUNC_skb_under_cgroup:
1549 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1550 goto error;
1551 break;
1552 case BPF_FUNC_redirect_map:
1553 if (map->map_type != BPF_MAP_TYPE_DEVMAP)
1554 goto error;
1555 break;
1556 case BPF_FUNC_sk_redirect_map:
1557 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1558 goto error;
1559 break;
1560 case BPF_FUNC_sock_map_update:
1561 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1562 goto error;
1563 break;
1564 default:
1565 break;
1566 }
1567
1568 return 0;
1569 error:
1570 verbose("cannot pass map_type %d into func %s#%d\n",
1571 map->map_type, func_id_name(func_id), func_id);
1572 return -EINVAL;
1573 }
1574
1575 static int check_raw_mode(const struct bpf_func_proto *fn)
1576 {
1577 int count = 0;
1578
1579 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1580 count++;
1581 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1582 count++;
1583 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1584 count++;
1585 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1586 count++;
1587 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1588 count++;
1589
1590 return count > 1 ? -EINVAL : 0;
1591 }
1592
1593 /* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
1594 * so turn them into unknown SCALAR_VALUE.
1595 */
1596 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1597 {
1598 struct bpf_verifier_state *state = &env->cur_state;
1599 struct bpf_reg_state *regs = state->regs, *reg;
1600 int i;
1601
1602 for (i = 0; i < MAX_BPF_REG; i++)
1603 if (regs[i].type == PTR_TO_PACKET ||
1604 regs[i].type == PTR_TO_PACKET_END)
1605 mark_reg_unknown(regs, i);
1606
1607 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
1608 if (state->stack_slot_type[i] != STACK_SPILL)
1609 continue;
1610 reg = &state->spilled_regs[i / BPF_REG_SIZE];
1611 if (reg->type != PTR_TO_PACKET &&
1612 reg->type != PTR_TO_PACKET_END)
1613 continue;
1614 __mark_reg_unknown(reg);
1615 }
1616 }
1617
1618 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1619 {
1620 struct bpf_verifier_state *state = &env->cur_state;
1621 const struct bpf_func_proto *fn = NULL;
1622 struct bpf_reg_state *regs = state->regs;
1623 struct bpf_call_arg_meta meta;
1624 bool changes_data;
1625 int i, err;
1626
1627 /* find function prototype */
1628 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1629 verbose("invalid func %s#%d\n", func_id_name(func_id), func_id);
1630 return -EINVAL;
1631 }
1632
1633 if (env->prog->aux->ops->get_func_proto)
1634 fn = env->prog->aux->ops->get_func_proto(func_id);
1635
1636 if (!fn) {
1637 verbose("unknown func %s#%d\n", func_id_name(func_id), func_id);
1638 return -EINVAL;
1639 }
1640
1641 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1642 if (!env->prog->gpl_compatible && fn->gpl_only) {
1643 verbose("cannot call GPL only function from proprietary program\n");
1644 return -EINVAL;
1645 }
1646
1647 changes_data = bpf_helper_changes_pkt_data(fn->func);
1648
1649 memset(&meta, 0, sizeof(meta));
1650 meta.pkt_access = fn->pkt_access;
1651
1652 /* We only support one arg being in raw mode at the moment, which
1653 * is sufficient for the helper functions we have right now.
1654 */
1655 err = check_raw_mode(fn);
1656 if (err) {
1657 verbose("kernel subsystem misconfigured func %s#%d\n",
1658 func_id_name(func_id), func_id);
1659 return err;
1660 }
1661
1662 /* check args */
1663 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1664 if (err)
1665 return err;
1666 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1667 if (err)
1668 return err;
1669 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1670 if (err)
1671 return err;
1672 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1673 if (err)
1674 return err;
1675 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1676 if (err)
1677 return err;
1678
1679 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1680 * is inferred from register state.
1681 */
1682 for (i = 0; i < meta.access_size; i++) {
1683 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1684 if (err)
1685 return err;
1686 }
1687
1688 /* reset caller saved regs */
1689 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1690 mark_reg_not_init(regs, caller_saved[i]);
1691 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1692 }
1693
1694 /* update return register (already marked as written above) */
1695 if (fn->ret_type == RET_INTEGER) {
1696 /* sets type to SCALAR_VALUE */
1697 mark_reg_unknown(regs, BPF_REG_0);
1698 } else if (fn->ret_type == RET_VOID) {
1699 regs[BPF_REG_0].type = NOT_INIT;
1700 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1701 struct bpf_insn_aux_data *insn_aux;
1702
1703 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1704 /* There is no offset yet applied, variable or fixed */
1705 mark_reg_known_zero(regs, BPF_REG_0);
1706 regs[BPF_REG_0].off = 0;
1707 /* remember map_ptr, so that check_map_access()
1708 * can check 'value_size' boundary of memory access
1709 * to map element returned from bpf_map_lookup_elem()
1710 */
1711 if (meta.map_ptr == NULL) {
1712 verbose("kernel subsystem misconfigured verifier\n");
1713 return -EINVAL;
1714 }
1715 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1716 regs[BPF_REG_0].id = ++env->id_gen;
1717 insn_aux = &env->insn_aux_data[insn_idx];
1718 if (!insn_aux->map_ptr)
1719 insn_aux->map_ptr = meta.map_ptr;
1720 else if (insn_aux->map_ptr != meta.map_ptr)
1721 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1722 } else {
1723 verbose("unknown return type %d of func %s#%d\n",
1724 fn->ret_type, func_id_name(func_id), func_id);
1725 return -EINVAL;
1726 }
1727
1728 err = check_map_func_compatibility(meta.map_ptr, func_id);
1729 if (err)
1730 return err;
1731
1732 if (changes_data)
1733 clear_all_pkt_pointers(env);
1734 return 0;
1735 }
1736
1737 static void coerce_reg_to_32(struct bpf_reg_state *reg)
1738 {
1739 /* clear high 32 bits */
1740 reg->var_off = tnum_cast(reg->var_off, 4);
1741 /* Update bounds */
1742 __update_reg_bounds(reg);
1743 }
1744
1745 static bool signed_add_overflows(s64 a, s64 b)
1746 {
1747 /* Do the add in u64, where overflow is well-defined */
1748 s64 res = (s64)((u64)a + (u64)b);
1749
1750 if (b < 0)
1751 return res > a;
1752 return res < a;
1753 }
1754
1755 static bool signed_sub_overflows(s64 a, s64 b)
1756 {
1757 /* Do the sub in u64, where overflow is well-defined */
1758 s64 res = (s64)((u64)a - (u64)b);
1759
1760 if (b < 0)
1761 return res < a;
1762 return res > a;
1763 }
1764
1765 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1766 * Caller should also handle BPF_MOV case separately.
1767 * If we return -EACCES, caller may want to try again treating pointer as a
1768 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1769 */
1770 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1771 struct bpf_insn *insn,
1772 const struct bpf_reg_state *ptr_reg,
1773 const struct bpf_reg_state *off_reg)
1774 {
1775 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
1776 bool known = tnum_is_const(off_reg->var_off);
1777 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1778 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1779 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1780 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1781 u8 opcode = BPF_OP(insn->code);
1782 u32 dst = insn->dst_reg;
1783
1784 dst_reg = &regs[dst];
1785
1786 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1787 print_verifier_state(&env->cur_state);
1788 verbose("verifier internal error: known but bad sbounds\n");
1789 return -EINVAL;
1790 }
1791 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1792 print_verifier_state(&env->cur_state);
1793 verbose("verifier internal error: known but bad ubounds\n");
1794 return -EINVAL;
1795 }
1796
1797 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1798 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1799 if (!env->allow_ptr_leaks)
1800 verbose("R%d 32-bit pointer arithmetic prohibited\n",
1801 dst);
1802 return -EACCES;
1803 }
1804
1805 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1806 if (!env->allow_ptr_leaks)
1807 verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1808 dst);
1809 return -EACCES;
1810 }
1811 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1812 if (!env->allow_ptr_leaks)
1813 verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1814 dst);
1815 return -EACCES;
1816 }
1817 if (ptr_reg->type == PTR_TO_PACKET_END) {
1818 if (!env->allow_ptr_leaks)
1819 verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1820 dst);
1821 return -EACCES;
1822 }
1823
1824 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1825 * The id may be overwritten later if we create a new variable offset.
1826 */
1827 dst_reg->type = ptr_reg->type;
1828 dst_reg->id = ptr_reg->id;
1829
1830 switch (opcode) {
1831 case BPF_ADD:
1832 /* We can take a fixed offset as long as it doesn't overflow
1833 * the s32 'off' field
1834 */
1835 if (known && (ptr_reg->off + smin_val ==
1836 (s64)(s32)(ptr_reg->off + smin_val))) {
1837 /* pointer += K. Accumulate it into fixed offset */
1838 dst_reg->smin_value = smin_ptr;
1839 dst_reg->smax_value = smax_ptr;
1840 dst_reg->umin_value = umin_ptr;
1841 dst_reg->umax_value = umax_ptr;
1842 dst_reg->var_off = ptr_reg->var_off;
1843 dst_reg->off = ptr_reg->off + smin_val;
1844 dst_reg->range = ptr_reg->range;
1845 break;
1846 }
1847 /* A new variable offset is created. Note that off_reg->off
1848 * == 0, since it's a scalar.
1849 * dst_reg gets the pointer type and since some positive
1850 * integer value was added to the pointer, give it a new 'id'
1851 * if it's a PTR_TO_PACKET.
1852 * this creates a new 'base' pointer, off_reg (variable) gets
1853 * added into the variable offset, and we copy the fixed offset
1854 * from ptr_reg.
1855 */
1856 if (signed_add_overflows(smin_ptr, smin_val) ||
1857 signed_add_overflows(smax_ptr, smax_val)) {
1858 dst_reg->smin_value = S64_MIN;
1859 dst_reg->smax_value = S64_MAX;
1860 } else {
1861 dst_reg->smin_value = smin_ptr + smin_val;
1862 dst_reg->smax_value = smax_ptr + smax_val;
1863 }
1864 if (umin_ptr + umin_val < umin_ptr ||
1865 umax_ptr + umax_val < umax_ptr) {
1866 dst_reg->umin_value = 0;
1867 dst_reg->umax_value = U64_MAX;
1868 } else {
1869 dst_reg->umin_value = umin_ptr + umin_val;
1870 dst_reg->umax_value = umax_ptr + umax_val;
1871 }
1872 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1873 dst_reg->off = ptr_reg->off;
1874 if (ptr_reg->type == PTR_TO_PACKET) {
1875 dst_reg->id = ++env->id_gen;
1876 /* something was added to pkt_ptr, set range to zero */
1877 dst_reg->range = 0;
1878 }
1879 break;
1880 case BPF_SUB:
1881 if (dst_reg == off_reg) {
1882 /* scalar -= pointer. Creates an unknown scalar */
1883 if (!env->allow_ptr_leaks)
1884 verbose("R%d tried to subtract pointer from scalar\n",
1885 dst);
1886 return -EACCES;
1887 }
1888 /* We don't allow subtraction from FP, because (according to
1889 * test_verifier.c test "invalid fp arithmetic", JITs might not
1890 * be able to deal with it.
1891 */
1892 if (ptr_reg->type == PTR_TO_STACK) {
1893 if (!env->allow_ptr_leaks)
1894 verbose("R%d subtraction from stack pointer prohibited\n",
1895 dst);
1896 return -EACCES;
1897 }
1898 if (known && (ptr_reg->off - smin_val ==
1899 (s64)(s32)(ptr_reg->off - smin_val))) {
1900 /* pointer -= K. Subtract it from fixed offset */
1901 dst_reg->smin_value = smin_ptr;
1902 dst_reg->smax_value = smax_ptr;
1903 dst_reg->umin_value = umin_ptr;
1904 dst_reg->umax_value = umax_ptr;
1905 dst_reg->var_off = ptr_reg->var_off;
1906 dst_reg->id = ptr_reg->id;
1907 dst_reg->off = ptr_reg->off - smin_val;
1908 dst_reg->range = ptr_reg->range;
1909 break;
1910 }
1911 /* A new variable offset is created. If the subtrahend is known
1912 * nonnegative, then any reg->range we had before is still good.
1913 */
1914 if (signed_sub_overflows(smin_ptr, smax_val) ||
1915 signed_sub_overflows(smax_ptr, smin_val)) {
1916 /* Overflow possible, we know nothing */
1917 dst_reg->smin_value = S64_MIN;
1918 dst_reg->smax_value = S64_MAX;
1919 } else {
1920 dst_reg->smin_value = smin_ptr - smax_val;
1921 dst_reg->smax_value = smax_ptr - smin_val;
1922 }
1923 if (umin_ptr < umax_val) {
1924 /* Overflow possible, we know nothing */
1925 dst_reg->umin_value = 0;
1926 dst_reg->umax_value = U64_MAX;
1927 } else {
1928 /* Cannot overflow (as long as bounds are consistent) */
1929 dst_reg->umin_value = umin_ptr - umax_val;
1930 dst_reg->umax_value = umax_ptr - umin_val;
1931 }
1932 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
1933 dst_reg->off = ptr_reg->off;
1934 if (ptr_reg->type == PTR_TO_PACKET) {
1935 dst_reg->id = ++env->id_gen;
1936 /* something was added to pkt_ptr, set range to zero */
1937 if (smin_val < 0)
1938 dst_reg->range = 0;
1939 }
1940 break;
1941 case BPF_AND:
1942 case BPF_OR:
1943 case BPF_XOR:
1944 /* bitwise ops on pointers are troublesome, prohibit for now.
1945 * (However, in principle we could allow some cases, e.g.
1946 * ptr &= ~3 which would reduce min_value by 3.)
1947 */
1948 if (!env->allow_ptr_leaks)
1949 verbose("R%d bitwise operator %s on pointer prohibited\n",
1950 dst, bpf_alu_string[opcode >> 4]);
1951 return -EACCES;
1952 default:
1953 /* other operators (e.g. MUL,LSH) produce non-pointer results */
1954 if (!env->allow_ptr_leaks)
1955 verbose("R%d pointer arithmetic with %s operator prohibited\n",
1956 dst, bpf_alu_string[opcode >> 4]);
1957 return -EACCES;
1958 }
1959
1960 __update_reg_bounds(dst_reg);
1961 __reg_deduce_bounds(dst_reg);
1962 __reg_bound_offset(dst_reg);
1963 return 0;
1964 }
1965
1966 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
1967 struct bpf_insn *insn,
1968 struct bpf_reg_state *dst_reg,
1969 struct bpf_reg_state src_reg)
1970 {
1971 struct bpf_reg_state *regs = env->cur_state.regs;
1972 u8 opcode = BPF_OP(insn->code);
1973 bool src_known, dst_known;
1974 s64 smin_val, smax_val;
1975 u64 umin_val, umax_val;
1976
1977 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1978 /* 32-bit ALU ops are (32,32)->64 */
1979 coerce_reg_to_32(dst_reg);
1980 coerce_reg_to_32(&src_reg);
1981 }
1982 smin_val = src_reg.smin_value;
1983 smax_val = src_reg.smax_value;
1984 umin_val = src_reg.umin_value;
1985 umax_val = src_reg.umax_value;
1986 src_known = tnum_is_const(src_reg.var_off);
1987 dst_known = tnum_is_const(dst_reg->var_off);
1988
1989 switch (opcode) {
1990 case BPF_ADD:
1991 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
1992 signed_add_overflows(dst_reg->smax_value, smax_val)) {
1993 dst_reg->smin_value = S64_MIN;
1994 dst_reg->smax_value = S64_MAX;
1995 } else {
1996 dst_reg->smin_value += smin_val;
1997 dst_reg->smax_value += smax_val;
1998 }
1999 if (dst_reg->umin_value + umin_val < umin_val ||
2000 dst_reg->umax_value + umax_val < umax_val) {
2001 dst_reg->umin_value = 0;
2002 dst_reg->umax_value = U64_MAX;
2003 } else {
2004 dst_reg->umin_value += umin_val;
2005 dst_reg->umax_value += umax_val;
2006 }
2007 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2008 break;
2009 case BPF_SUB:
2010 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2011 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2012 /* Overflow possible, we know nothing */
2013 dst_reg->smin_value = S64_MIN;
2014 dst_reg->smax_value = S64_MAX;
2015 } else {
2016 dst_reg->smin_value -= smax_val;
2017 dst_reg->smax_value -= smin_val;
2018 }
2019 if (dst_reg->umin_value < umax_val) {
2020 /* Overflow possible, we know nothing */
2021 dst_reg->umin_value = 0;
2022 dst_reg->umax_value = U64_MAX;
2023 } else {
2024 /* Cannot overflow (as long as bounds are consistent) */
2025 dst_reg->umin_value -= umax_val;
2026 dst_reg->umax_value -= umin_val;
2027 }
2028 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2029 break;
2030 case BPF_MUL:
2031 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2032 if (smin_val < 0 || dst_reg->smin_value < 0) {
2033 /* Ain't nobody got time to multiply that sign */
2034 __mark_reg_unbounded(dst_reg);
2035 __update_reg_bounds(dst_reg);
2036 break;
2037 }
2038 /* Both values are positive, so we can work with unsigned and
2039 * copy the result to signed (unless it exceeds S64_MAX).
2040 */
2041 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2042 /* Potential overflow, we know nothing */
2043 __mark_reg_unbounded(dst_reg);
2044 /* (except what we can learn from the var_off) */
2045 __update_reg_bounds(dst_reg);
2046 break;
2047 }
2048 dst_reg->umin_value *= umin_val;
2049 dst_reg->umax_value *= umax_val;
2050 if (dst_reg->umax_value > S64_MAX) {
2051 /* Overflow possible, we know nothing */
2052 dst_reg->smin_value = S64_MIN;
2053 dst_reg->smax_value = S64_MAX;
2054 } else {
2055 dst_reg->smin_value = dst_reg->umin_value;
2056 dst_reg->smax_value = dst_reg->umax_value;
2057 }
2058 break;
2059 case BPF_AND:
2060 if (src_known && dst_known) {
2061 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2062 src_reg.var_off.value);
2063 break;
2064 }
2065 /* We get our minimum from the var_off, since that's inherently
2066 * bitwise. Our maximum is the minimum of the operands' maxima.
2067 */
2068 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2069 dst_reg->umin_value = dst_reg->var_off.value;
2070 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2071 if (dst_reg->smin_value < 0 || smin_val < 0) {
2072 /* Lose signed bounds when ANDing negative numbers,
2073 * ain't nobody got time for that.
2074 */
2075 dst_reg->smin_value = S64_MIN;
2076 dst_reg->smax_value = S64_MAX;
2077 } else {
2078 /* ANDing two positives gives a positive, so safe to
2079 * cast result into s64.
2080 */
2081 dst_reg->smin_value = dst_reg->umin_value;
2082 dst_reg->smax_value = dst_reg->umax_value;
2083 }
2084 /* We may learn something more from the var_off */
2085 __update_reg_bounds(dst_reg);
2086 break;
2087 case BPF_OR:
2088 if (src_known && dst_known) {
2089 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2090 src_reg.var_off.value);
2091 break;
2092 }
2093 /* We get our maximum from the var_off, and our minimum is the
2094 * maximum of the operands' minima
2095 */
2096 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2097 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2098 dst_reg->umax_value = dst_reg->var_off.value |
2099 dst_reg->var_off.mask;
2100 if (dst_reg->smin_value < 0 || smin_val < 0) {
2101 /* Lose signed bounds when ORing negative numbers,
2102 * ain't nobody got time for that.
2103 */
2104 dst_reg->smin_value = S64_MIN;
2105 dst_reg->smax_value = S64_MAX;
2106 } else {
2107 /* ORing two positives gives a positive, so safe to
2108 * cast result into s64.
2109 */
2110 dst_reg->smin_value = dst_reg->umin_value;
2111 dst_reg->smax_value = dst_reg->umax_value;
2112 }
2113 /* We may learn something more from the var_off */
2114 __update_reg_bounds(dst_reg);
2115 break;
2116 case BPF_LSH:
2117 if (umax_val > 63) {
2118 /* Shifts greater than 63 are undefined. This includes
2119 * shifts by a negative number.
2120 */
2121 mark_reg_unknown(regs, insn->dst_reg);
2122 break;
2123 }
2124 /* We lose all sign bit information (except what we can pick
2125 * up from var_off)
2126 */
2127 dst_reg->smin_value = S64_MIN;
2128 dst_reg->smax_value = S64_MAX;
2129 /* If we might shift our top bit out, then we know nothing */
2130 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2131 dst_reg->umin_value = 0;
2132 dst_reg->umax_value = U64_MAX;
2133 } else {
2134 dst_reg->umin_value <<= umin_val;
2135 dst_reg->umax_value <<= umax_val;
2136 }
2137 if (src_known)
2138 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2139 else
2140 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2141 /* We may learn something more from the var_off */
2142 __update_reg_bounds(dst_reg);
2143 break;
2144 case BPF_RSH:
2145 if (umax_val > 63) {
2146 /* Shifts greater than 63 are undefined. This includes
2147 * shifts by a negative number.
2148 */
2149 mark_reg_unknown(regs, insn->dst_reg);
2150 break;
2151 }
2152 /* BPF_RSH is an unsigned shift, so make the appropriate casts */
2153 if (dst_reg->smin_value < 0) {
2154 if (umin_val) {
2155 /* Sign bit will be cleared */
2156 dst_reg->smin_value = 0;
2157 } else {
2158 /* Lost sign bit information */
2159 dst_reg->smin_value = S64_MIN;
2160 dst_reg->smax_value = S64_MAX;
2161 }
2162 } else {
2163 dst_reg->smin_value =
2164 (u64)(dst_reg->smin_value) >> umax_val;
2165 }
2166 if (src_known)
2167 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2168 umin_val);
2169 else
2170 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2171 dst_reg->umin_value >>= umax_val;
2172 dst_reg->umax_value >>= umin_val;
2173 /* We may learn something more from the var_off */
2174 __update_reg_bounds(dst_reg);
2175 break;
2176 default:
2177 mark_reg_unknown(regs, insn->dst_reg);
2178 break;
2179 }
2180
2181 __reg_deduce_bounds(dst_reg);
2182 __reg_bound_offset(dst_reg);
2183 return 0;
2184 }
2185
2186 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2187 * and var_off.
2188 */
2189 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2190 struct bpf_insn *insn)
2191 {
2192 struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
2193 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2194 u8 opcode = BPF_OP(insn->code);
2195 int rc;
2196
2197 dst_reg = &regs[insn->dst_reg];
2198 src_reg = NULL;
2199 if (dst_reg->type != SCALAR_VALUE)
2200 ptr_reg = dst_reg;
2201 if (BPF_SRC(insn->code) == BPF_X) {
2202 src_reg = &regs[insn->src_reg];
2203 if (src_reg->type != SCALAR_VALUE) {
2204 if (dst_reg->type != SCALAR_VALUE) {
2205 /* Combining two pointers by any ALU op yields
2206 * an arbitrary scalar.
2207 */
2208 if (!env->allow_ptr_leaks) {
2209 verbose("R%d pointer %s pointer prohibited\n",
2210 insn->dst_reg,
2211 bpf_alu_string[opcode >> 4]);
2212 return -EACCES;
2213 }
2214 mark_reg_unknown(regs, insn->dst_reg);
2215 return 0;
2216 } else {
2217 /* scalar += pointer
2218 * This is legal, but we have to reverse our
2219 * src/dest handling in computing the range
2220 */
2221 rc = adjust_ptr_min_max_vals(env, insn,
2222 src_reg, dst_reg);
2223 if (rc == -EACCES && env->allow_ptr_leaks) {
2224 /* scalar += unknown scalar */
2225 __mark_reg_unknown(&off_reg);
2226 return adjust_scalar_min_max_vals(
2227 env, insn,
2228 dst_reg, off_reg);
2229 }
2230 return rc;
2231 }
2232 } else if (ptr_reg) {
2233 /* pointer += scalar */
2234 rc = adjust_ptr_min_max_vals(env, insn,
2235 dst_reg, src_reg);
2236 if (rc == -EACCES && env->allow_ptr_leaks) {
2237 /* unknown scalar += scalar */
2238 __mark_reg_unknown(dst_reg);
2239 return adjust_scalar_min_max_vals(
2240 env, insn, dst_reg, *src_reg);
2241 }
2242 return rc;
2243 }
2244 } else {
2245 /* Pretend the src is a reg with a known value, since we only
2246 * need to be able to read from this state.
2247 */
2248 off_reg.type = SCALAR_VALUE;
2249 __mark_reg_known(&off_reg, insn->imm);
2250 src_reg = &off_reg;
2251 if (ptr_reg) { /* pointer += K */
2252 rc = adjust_ptr_min_max_vals(env, insn,
2253 ptr_reg, src_reg);
2254 if (rc == -EACCES && env->allow_ptr_leaks) {
2255 /* unknown scalar += K */
2256 __mark_reg_unknown(dst_reg);
2257 return adjust_scalar_min_max_vals(
2258 env, insn, dst_reg, off_reg);
2259 }
2260 return rc;
2261 }
2262 }
2263
2264 /* Got here implies adding two SCALAR_VALUEs */
2265 if (WARN_ON_ONCE(ptr_reg)) {
2266 print_verifier_state(&env->cur_state);
2267 verbose("verifier internal error: unexpected ptr_reg\n");
2268 return -EINVAL;
2269 }
2270 if (WARN_ON(!src_reg)) {
2271 print_verifier_state(&env->cur_state);
2272 verbose("verifier internal error: no src_reg\n");
2273 return -EINVAL;
2274 }
2275 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2276 }
2277
2278 /* check validity of 32-bit and 64-bit arithmetic operations */
2279 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2280 {
2281 struct bpf_reg_state *regs = env->cur_state.regs;
2282 u8 opcode = BPF_OP(insn->code);
2283 int err;
2284
2285 if (opcode == BPF_END || opcode == BPF_NEG) {
2286 if (opcode == BPF_NEG) {
2287 if (BPF_SRC(insn->code) != 0 ||
2288 insn->src_reg != BPF_REG_0 ||
2289 insn->off != 0 || insn->imm != 0) {
2290 verbose("BPF_NEG uses reserved fields\n");
2291 return -EINVAL;
2292 }
2293 } else {
2294 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2295 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2296 BPF_CLASS(insn->code) == BPF_ALU64) {
2297 verbose("BPF_END uses reserved fields\n");
2298 return -EINVAL;
2299 }
2300 }
2301
2302 /* check src operand */
2303 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2304 if (err)
2305 return err;
2306
2307 if (is_pointer_value(env, insn->dst_reg)) {
2308 verbose("R%d pointer arithmetic prohibited\n",
2309 insn->dst_reg);
2310 return -EACCES;
2311 }
2312
2313 /* check dest operand */
2314 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2315 if (err)
2316 return err;
2317
2318 } else if (opcode == BPF_MOV) {
2319
2320 if (BPF_SRC(insn->code) == BPF_X) {
2321 if (insn->imm != 0 || insn->off != 0) {
2322 verbose("BPF_MOV uses reserved fields\n");
2323 return -EINVAL;
2324 }
2325
2326 /* check src operand */
2327 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2328 if (err)
2329 return err;
2330 } else {
2331 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2332 verbose("BPF_MOV uses reserved fields\n");
2333 return -EINVAL;
2334 }
2335 }
2336
2337 /* check dest operand */
2338 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2339 if (err)
2340 return err;
2341
2342 if (BPF_SRC(insn->code) == BPF_X) {
2343 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2344 /* case: R1 = R2
2345 * copy register state to dest reg
2346 */
2347 regs[insn->dst_reg] = regs[insn->src_reg];
2348 } else {
2349 /* R1 = (u32) R2 */
2350 if (is_pointer_value(env, insn->src_reg)) {
2351 verbose("R%d partial copy of pointer\n",
2352 insn->src_reg);
2353 return -EACCES;
2354 }
2355 mark_reg_unknown(regs, insn->dst_reg);
2356 /* high 32 bits are known zero. */
2357 regs[insn->dst_reg].var_off = tnum_cast(
2358 regs[insn->dst_reg].var_off, 4);
2359 __update_reg_bounds(&regs[insn->dst_reg]);
2360 }
2361 } else {
2362 /* case: R = imm
2363 * remember the value we stored into this reg
2364 */
2365 regs[insn->dst_reg].type = SCALAR_VALUE;
2366 __mark_reg_known(regs + insn->dst_reg, insn->imm);
2367 }
2368
2369 } else if (opcode > BPF_END) {
2370 verbose("invalid BPF_ALU opcode %x\n", opcode);
2371 return -EINVAL;
2372
2373 } else { /* all other ALU ops: and, sub, xor, add, ... */
2374
2375 if (BPF_SRC(insn->code) == BPF_X) {
2376 if (insn->imm != 0 || insn->off != 0) {
2377 verbose("BPF_ALU uses reserved fields\n");
2378 return -EINVAL;
2379 }
2380 /* check src1 operand */
2381 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2382 if (err)
2383 return err;
2384 } else {
2385 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2386 verbose("BPF_ALU uses reserved fields\n");
2387 return -EINVAL;
2388 }
2389 }
2390
2391 /* check src2 operand */
2392 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2393 if (err)
2394 return err;
2395
2396 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2397 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2398 verbose("div by zero\n");
2399 return -EINVAL;
2400 }
2401
2402 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2403 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2404 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2405
2406 if (insn->imm < 0 || insn->imm >= size) {
2407 verbose("invalid shift %d\n", insn->imm);
2408 return -EINVAL;
2409 }
2410 }
2411
2412 /* check dest operand */
2413 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2414 if (err)
2415 return err;
2416
2417 return adjust_reg_min_max_vals(env, insn);
2418 }
2419
2420 return 0;
2421 }
2422
2423 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2424 struct bpf_reg_state *dst_reg)
2425 {
2426 struct bpf_reg_state *regs = state->regs, *reg;
2427 int i;
2428
2429 if (dst_reg->off < 0)
2430 /* This doesn't give us any range */
2431 return;
2432
2433 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2434 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2435 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2436 * than pkt_end, but that's because it's also less than pkt.
2437 */
2438 return;
2439
2440 /* LLVM can generate four kind of checks:
2441 *
2442 * Type 1/2:
2443 *
2444 * r2 = r3;
2445 * r2 += 8;
2446 * if (r2 > pkt_end) goto <handle exception>
2447 * <access okay>
2448 *
2449 * r2 = r3;
2450 * r2 += 8;
2451 * if (r2 < pkt_end) goto <access okay>
2452 * <handle exception>
2453 *
2454 * Where:
2455 * r2 == dst_reg, pkt_end == src_reg
2456 * r2=pkt(id=n,off=8,r=0)
2457 * r3=pkt(id=n,off=0,r=0)
2458 *
2459 * Type 3/4:
2460 *
2461 * r2 = r3;
2462 * r2 += 8;
2463 * if (pkt_end >= r2) goto <access okay>
2464 * <handle exception>
2465 *
2466 * r2 = r3;
2467 * r2 += 8;
2468 * if (pkt_end <= r2) goto <handle exception>
2469 * <access okay>
2470 *
2471 * Where:
2472 * pkt_end == dst_reg, r2 == src_reg
2473 * r2=pkt(id=n,off=8,r=0)
2474 * r3=pkt(id=n,off=0,r=0)
2475 *
2476 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2477 * so that range of bytes [r3, r3 + 8) is safe to access.
2478 */
2479
2480 /* If our ids match, then we must have the same max_value. And we
2481 * don't care about the other reg's fixed offset, since if it's too big
2482 * the range won't allow anything.
2483 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2484 */
2485 for (i = 0; i < MAX_BPF_REG; i++)
2486 if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
2487 /* keep the maximum range already checked */
2488 regs[i].range = max_t(u16, regs[i].range, dst_reg->off);
2489
2490 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2491 if (state->stack_slot_type[i] != STACK_SPILL)
2492 continue;
2493 reg = &state->spilled_regs[i / BPF_REG_SIZE];
2494 if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
2495 reg->range = max_t(u16, reg->range, dst_reg->off);
2496 }
2497 }
2498
2499 /* Adjusts the register min/max values in the case that the dst_reg is the
2500 * variable register that we are working on, and src_reg is a constant or we're
2501 * simply doing a BPF_K check.
2502 * In JEQ/JNE cases we also adjust the var_off values.
2503 */
2504 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2505 struct bpf_reg_state *false_reg, u64 val,
2506 u8 opcode)
2507 {
2508 /* If the dst_reg is a pointer, we can't learn anything about its
2509 * variable offset from the compare (unless src_reg were a pointer into
2510 * the same object, but we don't bother with that.
2511 * Since false_reg and true_reg have the same type by construction, we
2512 * only need to check one of them for pointerness.
2513 */
2514 if (__is_pointer_value(false, false_reg))
2515 return;
2516
2517 switch (opcode) {
2518 case BPF_JEQ:
2519 /* If this is false then we know nothing Jon Snow, but if it is
2520 * true then we know for sure.
2521 */
2522 __mark_reg_known(true_reg, val);
2523 break;
2524 case BPF_JNE:
2525 /* If this is true we know nothing Jon Snow, but if it is false
2526 * we know the value for sure;
2527 */
2528 __mark_reg_known(false_reg, val);
2529 break;
2530 case BPF_JGT:
2531 false_reg->umax_value = min(false_reg->umax_value, val);
2532 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2533 break;
2534 case BPF_JSGT:
2535 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2536 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2537 break;
2538 case BPF_JLT:
2539 false_reg->umin_value = max(false_reg->umin_value, val);
2540 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2541 break;
2542 case BPF_JSLT:
2543 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2544 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2545 break;
2546 case BPF_JGE:
2547 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2548 true_reg->umin_value = max(true_reg->umin_value, val);
2549 break;
2550 case BPF_JSGE:
2551 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2552 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2553 break;
2554 case BPF_JLE:
2555 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2556 true_reg->umax_value = min(true_reg->umax_value, val);
2557 break;
2558 case BPF_JSLE:
2559 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2560 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2561 break;
2562 default:
2563 break;
2564 }
2565
2566 __reg_deduce_bounds(false_reg);
2567 __reg_deduce_bounds(true_reg);
2568 /* We might have learned some bits from the bounds. */
2569 __reg_bound_offset(false_reg);
2570 __reg_bound_offset(true_reg);
2571 /* Intersecting with the old var_off might have improved our bounds
2572 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2573 * then new var_off is (0; 0x7f...fc) which improves our umax.
2574 */
2575 __update_reg_bounds(false_reg);
2576 __update_reg_bounds(true_reg);
2577 }
2578
2579 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2580 * the variable reg.
2581 */
2582 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2583 struct bpf_reg_state *false_reg, u64 val,
2584 u8 opcode)
2585 {
2586 if (__is_pointer_value(false, false_reg))
2587 return;
2588
2589 switch (opcode) {
2590 case BPF_JEQ:
2591 /* If this is false then we know nothing Jon Snow, but if it is
2592 * true then we know for sure.
2593 */
2594 __mark_reg_known(true_reg, val);
2595 break;
2596 case BPF_JNE:
2597 /* If this is true we know nothing Jon Snow, but if it is false
2598 * we know the value for sure;
2599 */
2600 __mark_reg_known(false_reg, val);
2601 break;
2602 case BPF_JGT:
2603 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2604 false_reg->umin_value = max(false_reg->umin_value, val);
2605 break;
2606 case BPF_JSGT:
2607 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2608 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2609 break;
2610 case BPF_JLT:
2611 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2612 false_reg->umax_value = min(false_reg->umax_value, val);
2613 break;
2614 case BPF_JSLT:
2615 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2616 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2617 break;
2618 case BPF_JGE:
2619 true_reg->umax_value = min(true_reg->umax_value, val);
2620 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2621 break;
2622 case BPF_JSGE:
2623 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2624 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2625 break;
2626 case BPF_JLE:
2627 true_reg->umin_value = max(true_reg->umin_value, val);
2628 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2629 break;
2630 case BPF_JSLE:
2631 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2632 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2633 break;
2634 default:
2635 break;
2636 }
2637
2638 __reg_deduce_bounds(false_reg);
2639 __reg_deduce_bounds(true_reg);
2640 /* We might have learned some bits from the bounds. */
2641 __reg_bound_offset(false_reg);
2642 __reg_bound_offset(true_reg);
2643 /* Intersecting with the old var_off might have improved our bounds
2644 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2645 * then new var_off is (0; 0x7f...fc) which improves our umax.
2646 */
2647 __update_reg_bounds(false_reg);
2648 __update_reg_bounds(true_reg);
2649 }
2650
2651 /* Regs are known to be equal, so intersect their min/max/var_off */
2652 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2653 struct bpf_reg_state *dst_reg)
2654 {
2655 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2656 dst_reg->umin_value);
2657 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2658 dst_reg->umax_value);
2659 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2660 dst_reg->smin_value);
2661 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2662 dst_reg->smax_value);
2663 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2664 dst_reg->var_off);
2665 /* We might have learned new bounds from the var_off. */
2666 __update_reg_bounds(src_reg);
2667 __update_reg_bounds(dst_reg);
2668 /* We might have learned something about the sign bit. */
2669 __reg_deduce_bounds(src_reg);
2670 __reg_deduce_bounds(dst_reg);
2671 /* We might have learned some bits from the bounds. */
2672 __reg_bound_offset(src_reg);
2673 __reg_bound_offset(dst_reg);
2674 /* Intersecting with the old var_off might have improved our bounds
2675 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2676 * then new var_off is (0; 0x7f...fc) which improves our umax.
2677 */
2678 __update_reg_bounds(src_reg);
2679 __update_reg_bounds(dst_reg);
2680 }
2681
2682 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2683 struct bpf_reg_state *true_dst,
2684 struct bpf_reg_state *false_src,
2685 struct bpf_reg_state *false_dst,
2686 u8 opcode)
2687 {
2688 switch (opcode) {
2689 case BPF_JEQ:
2690 __reg_combine_min_max(true_src, true_dst);
2691 break;
2692 case BPF_JNE:
2693 __reg_combine_min_max(false_src, false_dst);
2694 break;
2695 }
2696 }
2697
2698 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2699 bool is_null)
2700 {
2701 struct bpf_reg_state *reg = &regs[regno];
2702
2703 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2704 /* Old offset (both fixed and variable parts) should
2705 * have been known-zero, because we don't allow pointer
2706 * arithmetic on pointers that might be NULL.
2707 */
2708 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2709 !tnum_equals_const(reg->var_off, 0) ||
2710 reg->off)) {
2711 __mark_reg_known_zero(reg);
2712 reg->off = 0;
2713 }
2714 if (is_null) {
2715 reg->type = SCALAR_VALUE;
2716 } else if (reg->map_ptr->inner_map_meta) {
2717 reg->type = CONST_PTR_TO_MAP;
2718 reg->map_ptr = reg->map_ptr->inner_map_meta;
2719 } else {
2720 reg->type = PTR_TO_MAP_VALUE;
2721 }
2722 /* We don't need id from this point onwards anymore, thus we
2723 * should better reset it, so that state pruning has chances
2724 * to take effect.
2725 */
2726 reg->id = 0;
2727 }
2728 }
2729
2730 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2731 * be folded together at some point.
2732 */
2733 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2734 bool is_null)
2735 {
2736 struct bpf_reg_state *regs = state->regs;
2737 u32 id = regs[regno].id;
2738 int i;
2739
2740 for (i = 0; i < MAX_BPF_REG; i++)
2741 mark_map_reg(regs, i, id, is_null);
2742
2743 for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
2744 if (state->stack_slot_type[i] != STACK_SPILL)
2745 continue;
2746 mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
2747 }
2748 }
2749
2750 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2751 struct bpf_insn *insn, int *insn_idx)
2752 {
2753 struct bpf_verifier_state *other_branch, *this_branch = &env->cur_state;
2754 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2755 u8 opcode = BPF_OP(insn->code);
2756 int err;
2757
2758 if (opcode > BPF_JSLE) {
2759 verbose("invalid BPF_JMP opcode %x\n", opcode);
2760 return -EINVAL;
2761 }
2762
2763 if (BPF_SRC(insn->code) == BPF_X) {
2764 if (insn->imm != 0) {
2765 verbose("BPF_JMP uses reserved fields\n");
2766 return -EINVAL;
2767 }
2768
2769 /* check src1 operand */
2770 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2771 if (err)
2772 return err;
2773
2774 if (is_pointer_value(env, insn->src_reg)) {
2775 verbose("R%d pointer comparison prohibited\n",
2776 insn->src_reg);
2777 return -EACCES;
2778 }
2779 } else {
2780 if (insn->src_reg != BPF_REG_0) {
2781 verbose("BPF_JMP uses reserved fields\n");
2782 return -EINVAL;
2783 }
2784 }
2785
2786 /* check src2 operand */
2787 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2788 if (err)
2789 return err;
2790
2791 dst_reg = &regs[insn->dst_reg];
2792
2793 /* detect if R == 0 where R was initialized to zero earlier */
2794 if (BPF_SRC(insn->code) == BPF_K &&
2795 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2796 dst_reg->type == SCALAR_VALUE &&
2797 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2798 if (opcode == BPF_JEQ) {
2799 /* if (imm == imm) goto pc+off;
2800 * only follow the goto, ignore fall-through
2801 */
2802 *insn_idx += insn->off;
2803 return 0;
2804 } else {
2805 /* if (imm != imm) goto pc+off;
2806 * only follow fall-through branch, since
2807 * that's where the program will go
2808 */
2809 return 0;
2810 }
2811 }
2812
2813 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
2814 if (!other_branch)
2815 return -EFAULT;
2816
2817 /* detect if we are comparing against a constant value so we can adjust
2818 * our min/max values for our dst register.
2819 * this is only legit if both are scalars (or pointers to the same
2820 * object, I suppose, but we don't support that right now), because
2821 * otherwise the different base pointers mean the offsets aren't
2822 * comparable.
2823 */
2824 if (BPF_SRC(insn->code) == BPF_X) {
2825 if (dst_reg->type == SCALAR_VALUE &&
2826 regs[insn->src_reg].type == SCALAR_VALUE) {
2827 if (tnum_is_const(regs[insn->src_reg].var_off))
2828 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2829 dst_reg, regs[insn->src_reg].var_off.value,
2830 opcode);
2831 else if (tnum_is_const(dst_reg->var_off))
2832 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
2833 &regs[insn->src_reg],
2834 dst_reg->var_off.value, opcode);
2835 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
2836 /* Comparing for equality, we can combine knowledge */
2837 reg_combine_min_max(&other_branch->regs[insn->src_reg],
2838 &other_branch->regs[insn->dst_reg],
2839 &regs[insn->src_reg],
2840 &regs[insn->dst_reg], opcode);
2841 }
2842 } else if (dst_reg->type == SCALAR_VALUE) {
2843 reg_set_min_max(&other_branch->regs[insn->dst_reg],
2844 dst_reg, insn->imm, opcode);
2845 }
2846
2847 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
2848 if (BPF_SRC(insn->code) == BPF_K &&
2849 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2850 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2851 /* Mark all identical map registers in each branch as either
2852 * safe or unknown depending R == 0 or R != 0 conditional.
2853 */
2854 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
2855 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
2856 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
2857 dst_reg->type == PTR_TO_PACKET &&
2858 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2859 find_good_pkt_pointers(this_branch, dst_reg);
2860 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLT &&
2861 dst_reg->type == PTR_TO_PACKET &&
2862 regs[insn->src_reg].type == PTR_TO_PACKET_END) {
2863 find_good_pkt_pointers(other_branch, dst_reg);
2864 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGE &&
2865 dst_reg->type == PTR_TO_PACKET_END &&
2866 regs[insn->src_reg].type == PTR_TO_PACKET) {
2867 find_good_pkt_pointers(other_branch, &regs[insn->src_reg]);
2868 } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JLE &&
2869 dst_reg->type == PTR_TO_PACKET_END &&
2870 regs[insn->src_reg].type == PTR_TO_PACKET) {
2871 find_good_pkt_pointers(this_branch, &regs[insn->src_reg]);
2872 } else if (is_pointer_value(env, insn->dst_reg)) {
2873 verbose("R%d pointer comparison prohibited\n", insn->dst_reg);
2874 return -EACCES;
2875 }
2876 if (log_level)
2877 print_verifier_state(this_branch);
2878 return 0;
2879 }
2880
2881 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
2882 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
2883 {
2884 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
2885
2886 return (struct bpf_map *) (unsigned long) imm64;
2887 }
2888
2889 /* verify BPF_LD_IMM64 instruction */
2890 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
2891 {
2892 struct bpf_reg_state *regs = env->cur_state.regs;
2893 int err;
2894
2895 if (BPF_SIZE(insn->code) != BPF_DW) {
2896 verbose("invalid BPF_LD_IMM insn\n");
2897 return -EINVAL;
2898 }
2899 if (insn->off != 0) {
2900 verbose("BPF_LD_IMM64 uses reserved fields\n");
2901 return -EINVAL;
2902 }
2903
2904 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2905 if (err)
2906 return err;
2907
2908 if (insn->src_reg == 0) {
2909 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
2910
2911 regs[insn->dst_reg].type = SCALAR_VALUE;
2912 __mark_reg_known(&regs[insn->dst_reg], imm);
2913 return 0;
2914 }
2915
2916 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
2917 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
2918
2919 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
2920 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
2921 return 0;
2922 }
2923
2924 static bool may_access_skb(enum bpf_prog_type type)
2925 {
2926 switch (type) {
2927 case BPF_PROG_TYPE_SOCKET_FILTER:
2928 case BPF_PROG_TYPE_SCHED_CLS:
2929 case BPF_PROG_TYPE_SCHED_ACT:
2930 return true;
2931 default:
2932 return false;
2933 }
2934 }
2935
2936 /* verify safety of LD_ABS|LD_IND instructions:
2937 * - they can only appear in the programs where ctx == skb
2938 * - since they are wrappers of function calls, they scratch R1-R5 registers,
2939 * preserve R6-R9, and store return value into R0
2940 *
2941 * Implicit input:
2942 * ctx == skb == R6 == CTX
2943 *
2944 * Explicit input:
2945 * SRC == any register
2946 * IMM == 32-bit immediate
2947 *
2948 * Output:
2949 * R0 - 8/16/32-bit skb data converted to cpu endianness
2950 */
2951 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
2952 {
2953 struct bpf_reg_state *regs = env->cur_state.regs;
2954 u8 mode = BPF_MODE(insn->code);
2955 int i, err;
2956
2957 if (!may_access_skb(env->prog->type)) {
2958 verbose("BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
2959 return -EINVAL;
2960 }
2961
2962 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
2963 BPF_SIZE(insn->code) == BPF_DW ||
2964 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
2965 verbose("BPF_LD_[ABS|IND] uses reserved fields\n");
2966 return -EINVAL;
2967 }
2968
2969 /* check whether implicit source operand (register R6) is readable */
2970 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
2971 if (err)
2972 return err;
2973
2974 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
2975 verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
2976 return -EINVAL;
2977 }
2978
2979 if (mode == BPF_IND) {
2980 /* check explicit source operand */
2981 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2982 if (err)
2983 return err;
2984 }
2985
2986 /* reset caller saved regs to unreadable */
2987 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2988 mark_reg_not_init(regs, caller_saved[i]);
2989 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2990 }
2991
2992 /* mark destination R0 register as readable, since it contains
2993 * the value fetched from the packet.
2994 * Already marked as written above.
2995 */
2996 mark_reg_unknown(regs, BPF_REG_0);
2997 return 0;
2998 }
2999
3000 /* non-recursive DFS pseudo code
3001 * 1 procedure DFS-iterative(G,v):
3002 * 2 label v as discovered
3003 * 3 let S be a stack
3004 * 4 S.push(v)
3005 * 5 while S is not empty
3006 * 6 t <- S.pop()
3007 * 7 if t is what we're looking for:
3008 * 8 return t
3009 * 9 for all edges e in G.adjacentEdges(t) do
3010 * 10 if edge e is already labelled
3011 * 11 continue with the next edge
3012 * 12 w <- G.adjacentVertex(t,e)
3013 * 13 if vertex w is not discovered and not explored
3014 * 14 label e as tree-edge
3015 * 15 label w as discovered
3016 * 16 S.push(w)
3017 * 17 continue at 5
3018 * 18 else if vertex w is discovered
3019 * 19 label e as back-edge
3020 * 20 else
3021 * 21 // vertex w is explored
3022 * 22 label e as forward- or cross-edge
3023 * 23 label t as explored
3024 * 24 S.pop()
3025 *
3026 * convention:
3027 * 0x10 - discovered
3028 * 0x11 - discovered and fall-through edge labelled
3029 * 0x12 - discovered and fall-through and branch edges labelled
3030 * 0x20 - explored
3031 */
3032
3033 enum {
3034 DISCOVERED = 0x10,
3035 EXPLORED = 0x20,
3036 FALLTHROUGH = 1,
3037 BRANCH = 2,
3038 };
3039
3040 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3041
3042 static int *insn_stack; /* stack of insns to process */
3043 static int cur_stack; /* current stack index */
3044 static int *insn_state;
3045
3046 /* t, w, e - match pseudo-code above:
3047 * t - index of current instruction
3048 * w - next instruction
3049 * e - edge
3050 */
3051 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3052 {
3053 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3054 return 0;
3055
3056 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3057 return 0;
3058
3059 if (w < 0 || w >= env->prog->len) {
3060 verbose("jump out of range from insn %d to %d\n", t, w);
3061 return -EINVAL;
3062 }
3063
3064 if (e == BRANCH)
3065 /* mark branch target for state pruning */
3066 env->explored_states[w] = STATE_LIST_MARK;
3067
3068 if (insn_state[w] == 0) {
3069 /* tree-edge */
3070 insn_state[t] = DISCOVERED | e;
3071 insn_state[w] = DISCOVERED;
3072 if (cur_stack >= env->prog->len)
3073 return -E2BIG;
3074 insn_stack[cur_stack++] = w;
3075 return 1;
3076 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3077 verbose("back-edge from insn %d to %d\n", t, w);
3078 return -EINVAL;
3079 } else if (insn_state[w] == EXPLORED) {
3080 /* forward- or cross-edge */
3081 insn_state[t] = DISCOVERED | e;
3082 } else {
3083 verbose("insn state internal bug\n");
3084 return -EFAULT;
3085 }
3086 return 0;
3087 }
3088
3089 /* non-recursive depth-first-search to detect loops in BPF program
3090 * loop == back-edge in directed graph
3091 */
3092 static int check_cfg(struct bpf_verifier_env *env)
3093 {
3094 struct bpf_insn *insns = env->prog->insnsi;
3095 int insn_cnt = env->prog->len;
3096 int ret = 0;
3097 int i, t;
3098
3099 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3100 if (!insn_state)
3101 return -ENOMEM;
3102
3103 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3104 if (!insn_stack) {
3105 kfree(insn_state);
3106 return -ENOMEM;
3107 }
3108
3109 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3110 insn_stack[0] = 0; /* 0 is the first instruction */
3111 cur_stack = 1;
3112
3113 peek_stack:
3114 if (cur_stack == 0)
3115 goto check_state;
3116 t = insn_stack[cur_stack - 1];
3117
3118 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3119 u8 opcode = BPF_OP(insns[t].code);
3120
3121 if (opcode == BPF_EXIT) {
3122 goto mark_explored;
3123 } else if (opcode == BPF_CALL) {
3124 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3125 if (ret == 1)
3126 goto peek_stack;
3127 else if (ret < 0)
3128 goto err_free;
3129 if (t + 1 < insn_cnt)
3130 env->explored_states[t + 1] = STATE_LIST_MARK;
3131 } else if (opcode == BPF_JA) {
3132 if (BPF_SRC(insns[t].code) != BPF_K) {
3133 ret = -EINVAL;
3134 goto err_free;
3135 }
3136 /* unconditional jump with single edge */
3137 ret = push_insn(t, t + insns[t].off + 1,
3138 FALLTHROUGH, env);
3139 if (ret == 1)
3140 goto peek_stack;
3141 else if (ret < 0)
3142 goto err_free;
3143 /* tell verifier to check for equivalent states
3144 * after every call and jump
3145 */
3146 if (t + 1 < insn_cnt)
3147 env->explored_states[t + 1] = STATE_LIST_MARK;
3148 } else {
3149 /* conditional jump with two edges */
3150 env->explored_states[t] = STATE_LIST_MARK;
3151 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3152 if (ret == 1)
3153 goto peek_stack;
3154 else if (ret < 0)
3155 goto err_free;
3156
3157 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3158 if (ret == 1)
3159 goto peek_stack;
3160 else if (ret < 0)
3161 goto err_free;
3162 }
3163 } else {
3164 /* all other non-branch instructions with single
3165 * fall-through edge
3166 */
3167 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3168 if (ret == 1)
3169 goto peek_stack;
3170 else if (ret < 0)
3171 goto err_free;
3172 }
3173
3174 mark_explored:
3175 insn_state[t] = EXPLORED;
3176 if (cur_stack-- <= 0) {
3177 verbose("pop stack internal bug\n");
3178 ret = -EFAULT;
3179 goto err_free;
3180 }
3181 goto peek_stack;
3182
3183 check_state:
3184 for (i = 0; i < insn_cnt; i++) {
3185 if (insn_state[i] != EXPLORED) {
3186 verbose("unreachable insn %d\n", i);
3187 ret = -EINVAL;
3188 goto err_free;
3189 }
3190 }
3191 ret = 0; /* cfg looks good */
3192
3193 err_free:
3194 kfree(insn_state);
3195 kfree(insn_stack);
3196 return ret;
3197 }
3198
3199 /* check %cur's range satisfies %old's */
3200 static bool range_within(struct bpf_reg_state *old,
3201 struct bpf_reg_state *cur)
3202 {
3203 return old->umin_value <= cur->umin_value &&
3204 old->umax_value >= cur->umax_value &&
3205 old->smin_value <= cur->smin_value &&
3206 old->smax_value >= cur->smax_value;
3207 }
3208
3209 /* Maximum number of register states that can exist at once */
3210 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3211 struct idpair {
3212 u32 old;
3213 u32 cur;
3214 };
3215
3216 /* If in the old state two registers had the same id, then they need to have
3217 * the same id in the new state as well. But that id could be different from
3218 * the old state, so we need to track the mapping from old to new ids.
3219 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3220 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3221 * regs with a different old id could still have new id 9, we don't care about
3222 * that.
3223 * So we look through our idmap to see if this old id has been seen before. If
3224 * so, we require the new id to match; otherwise, we add the id pair to the map.
3225 */
3226 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3227 {
3228 unsigned int i;
3229
3230 for (i = 0; i < ID_MAP_SIZE; i++) {
3231 if (!idmap[i].old) {
3232 /* Reached an empty slot; haven't seen this id before */
3233 idmap[i].old = old_id;
3234 idmap[i].cur = cur_id;
3235 return true;
3236 }
3237 if (idmap[i].old == old_id)
3238 return idmap[i].cur == cur_id;
3239 }
3240 /* We ran out of idmap slots, which should be impossible */
3241 WARN_ON_ONCE(1);
3242 return false;
3243 }
3244
3245 /* Returns true if (rold safe implies rcur safe) */
3246 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3247 struct idpair *idmap)
3248 {
3249 if (!(rold->live & REG_LIVE_READ))
3250 /* explored state didn't use this */
3251 return true;
3252
3253 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3254 return true;
3255
3256 if (rold->type == NOT_INIT)
3257 /* explored state can't have used this */
3258 return true;
3259 if (rcur->type == NOT_INIT)
3260 return false;
3261 switch (rold->type) {
3262 case SCALAR_VALUE:
3263 if (rcur->type == SCALAR_VALUE) {
3264 /* new val must satisfy old val knowledge */
3265 return range_within(rold, rcur) &&
3266 tnum_in(rold->var_off, rcur->var_off);
3267 } else {
3268 /* if we knew anything about the old value, we're not
3269 * equal, because we can't know anything about the
3270 * scalar value of the pointer in the new value.
3271 */
3272 return rold->umin_value == 0 &&
3273 rold->umax_value == U64_MAX &&
3274 rold->smin_value == S64_MIN &&
3275 rold->smax_value == S64_MAX &&
3276 tnum_is_unknown(rold->var_off);
3277 }
3278 case PTR_TO_MAP_VALUE:
3279 /* If the new min/max/var_off satisfy the old ones and
3280 * everything else matches, we are OK.
3281 * We don't care about the 'id' value, because nothing
3282 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3283 */
3284 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3285 range_within(rold, rcur) &&
3286 tnum_in(rold->var_off, rcur->var_off);
3287 case PTR_TO_MAP_VALUE_OR_NULL:
3288 /* a PTR_TO_MAP_VALUE could be safe to use as a
3289 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3290 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3291 * checked, doing so could have affected others with the same
3292 * id, and we can't check for that because we lost the id when
3293 * we converted to a PTR_TO_MAP_VALUE.
3294 */
3295 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3296 return false;
3297 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3298 return false;
3299 /* Check our ids match any regs they're supposed to */
3300 return check_ids(rold->id, rcur->id, idmap);
3301 case PTR_TO_PACKET:
3302 if (rcur->type != PTR_TO_PACKET)
3303 return false;
3304 /* We must have at least as much range as the old ptr
3305 * did, so that any accesses which were safe before are
3306 * still safe. This is true even if old range < old off,
3307 * since someone could have accessed through (ptr - k), or
3308 * even done ptr -= k in a register, to get a safe access.
3309 */
3310 if (rold->range > rcur->range)
3311 return false;
3312 /* If the offsets don't match, we can't trust our alignment;
3313 * nor can we be sure that we won't fall out of range.
3314 */
3315 if (rold->off != rcur->off)
3316 return false;
3317 /* id relations must be preserved */
3318 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3319 return false;
3320 /* new val must satisfy old val knowledge */
3321 return range_within(rold, rcur) &&
3322 tnum_in(rold->var_off, rcur->var_off);
3323 case PTR_TO_CTX:
3324 case CONST_PTR_TO_MAP:
3325 case PTR_TO_STACK:
3326 case PTR_TO_PACKET_END:
3327 /* Only valid matches are exact, which memcmp() above
3328 * would have accepted
3329 */
3330 default:
3331 /* Don't know what's going on, just say it's not safe */
3332 return false;
3333 }
3334
3335 /* Shouldn't get here; if we do, say it's not safe */
3336 WARN_ON_ONCE(1);
3337 return false;
3338 }
3339
3340 /* compare two verifier states
3341 *
3342 * all states stored in state_list are known to be valid, since
3343 * verifier reached 'bpf_exit' instruction through them
3344 *
3345 * this function is called when verifier exploring different branches of
3346 * execution popped from the state stack. If it sees an old state that has
3347 * more strict register state and more strict stack state then this execution
3348 * branch doesn't need to be explored further, since verifier already
3349 * concluded that more strict state leads to valid finish.
3350 *
3351 * Therefore two states are equivalent if register state is more conservative
3352 * and explored stack state is more conservative than the current one.
3353 * Example:
3354 * explored current
3355 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3356 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3357 *
3358 * In other words if current stack state (one being explored) has more
3359 * valid slots than old one that already passed validation, it means
3360 * the verifier can stop exploring and conclude that current state is valid too
3361 *
3362 * Similarly with registers. If explored state has register type as invalid
3363 * whereas register type in current state is meaningful, it means that
3364 * the current state will reach 'bpf_exit' instruction safely
3365 */
3366 static bool states_equal(struct bpf_verifier_env *env,
3367 struct bpf_verifier_state *old,
3368 struct bpf_verifier_state *cur)
3369 {
3370 struct idpair *idmap;
3371 bool ret = false;
3372 int i;
3373
3374 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3375 /* If we failed to allocate the idmap, just say it's not safe */
3376 if (!idmap)
3377 return false;
3378
3379 for (i = 0; i < MAX_BPF_REG; i++) {
3380 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3381 goto out_free;
3382 }
3383
3384 for (i = 0; i < MAX_BPF_STACK; i++) {
3385 if (old->stack_slot_type[i] == STACK_INVALID)
3386 continue;
3387 if (old->stack_slot_type[i] != cur->stack_slot_type[i])
3388 /* Ex: old explored (safe) state has STACK_SPILL in
3389 * this stack slot, but current has has STACK_MISC ->
3390 * this verifier states are not equivalent,
3391 * return false to continue verification of this path
3392 */
3393 goto out_free;
3394 if (i % BPF_REG_SIZE)
3395 continue;
3396 if (old->stack_slot_type[i] != STACK_SPILL)
3397 continue;
3398 if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
3399 &cur->spilled_regs[i / BPF_REG_SIZE],
3400 idmap))
3401 /* when explored and current stack slot are both storing
3402 * spilled registers, check that stored pointers types
3403 * are the same as well.
3404 * Ex: explored safe path could have stored
3405 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3406 * but current path has stored:
3407 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3408 * such verifier states are not equivalent.
3409 * return false to continue verification of this path
3410 */
3411 goto out_free;
3412 else
3413 continue;
3414 }
3415 ret = true;
3416 out_free:
3417 kfree(idmap);
3418 return ret;
3419 }
3420
3421 /* A write screens off any subsequent reads; but write marks come from the
3422 * straight-line code between a state and its parent. When we arrive at a
3423 * jump target (in the first iteration of the propagate_liveness() loop),
3424 * we didn't arrive by the straight-line code, so read marks in state must
3425 * propagate to parent regardless of state's write marks.
3426 */
3427 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3428 struct bpf_verifier_state *parent)
3429 {
3430 bool writes = parent == state->parent; /* Observe write marks */
3431 bool touched = false; /* any changes made? */
3432 int i;
3433
3434 if (!parent)
3435 return touched;
3436 /* Propagate read liveness of registers... */
3437 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3438 /* We don't need to worry about FP liveness because it's read-only */
3439 for (i = 0; i < BPF_REG_FP; i++) {
3440 if (parent->regs[i].live & REG_LIVE_READ)
3441 continue;
3442 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3443 continue;
3444 if (state->regs[i].live & REG_LIVE_READ) {
3445 parent->regs[i].live |= REG_LIVE_READ;
3446 touched = true;
3447 }
3448 }
3449 /* ... and stack slots */
3450 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++) {
3451 if (parent->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3452 continue;
3453 if (state->stack_slot_type[i * BPF_REG_SIZE] != STACK_SPILL)
3454 continue;
3455 if (parent->spilled_regs[i].live & REG_LIVE_READ)
3456 continue;
3457 if (writes && (state->spilled_regs[i].live & REG_LIVE_WRITTEN))
3458 continue;
3459 if (state->spilled_regs[i].live & REG_LIVE_READ) {
3460 parent->spilled_regs[i].live |= REG_LIVE_READ;
3461 touched = true;
3462 }
3463 }
3464 return touched;
3465 }
3466
3467 /* "parent" is "a state from which we reach the current state", but initially
3468 * it is not the state->parent (i.e. "the state whose straight-line code leads
3469 * to the current state"), instead it is the state that happened to arrive at
3470 * a (prunable) equivalent of the current state. See comment above
3471 * do_propagate_liveness() for consequences of this.
3472 * This function is just a more efficient way of calling mark_reg_read() or
3473 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3474 * though it requires that parent != state->parent in the call arguments.
3475 */
3476 static void propagate_liveness(const struct bpf_verifier_state *state,
3477 struct bpf_verifier_state *parent)
3478 {
3479 while (do_propagate_liveness(state, parent)) {
3480 /* Something changed, so we need to feed those changes onward */
3481 state = parent;
3482 parent = state->parent;
3483 }
3484 }
3485
3486 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3487 {
3488 struct bpf_verifier_state_list *new_sl;
3489 struct bpf_verifier_state_list *sl;
3490 int i;
3491
3492 sl = env->explored_states[insn_idx];
3493 if (!sl)
3494 /* this 'insn_idx' instruction wasn't marked, so we will not
3495 * be doing state search here
3496 */
3497 return 0;
3498
3499 while (sl != STATE_LIST_MARK) {
3500 if (states_equal(env, &sl->state, &env->cur_state)) {
3501 /* reached equivalent register/stack state,
3502 * prune the search.
3503 * Registers read by the continuation are read by us.
3504 * If we have any write marks in env->cur_state, they
3505 * will prevent corresponding reads in the continuation
3506 * from reaching our parent (an explored_state). Our
3507 * own state will get the read marks recorded, but
3508 * they'll be immediately forgotten as we're pruning
3509 * this state and will pop a new one.
3510 */
3511 propagate_liveness(&sl->state, &env->cur_state);
3512 return 1;
3513 }
3514 sl = sl->next;
3515 }
3516
3517 /* there were no equivalent states, remember current one.
3518 * technically the current state is not proven to be safe yet,
3519 * but it will either reach bpf_exit (which means it's safe) or
3520 * it will be rejected. Since there are no loops, we won't be
3521 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3522 */
3523 new_sl = kmalloc(sizeof(struct bpf_verifier_state_list), GFP_USER);
3524 if (!new_sl)
3525 return -ENOMEM;
3526
3527 /* add new state to the head of linked list */
3528 memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
3529 new_sl->next = env->explored_states[insn_idx];
3530 env->explored_states[insn_idx] = new_sl;
3531 /* connect new state to parentage chain */
3532 env->cur_state.parent = &new_sl->state;
3533 /* clear write marks in current state: the writes we did are not writes
3534 * our child did, so they don't screen off its reads from us.
3535 * (There are no read marks in current state, because reads always mark
3536 * their parent and current state never has children yet. Only
3537 * explored_states can get read marks.)
3538 */
3539 for (i = 0; i < BPF_REG_FP; i++)
3540 env->cur_state.regs[i].live = REG_LIVE_NONE;
3541 for (i = 0; i < MAX_BPF_STACK / BPF_REG_SIZE; i++)
3542 if (env->cur_state.stack_slot_type[i * BPF_REG_SIZE] == STACK_SPILL)
3543 env->cur_state.spilled_regs[i].live = REG_LIVE_NONE;
3544 return 0;
3545 }
3546
3547 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3548 int insn_idx, int prev_insn_idx)
3549 {
3550 if (!env->analyzer_ops || !env->analyzer_ops->insn_hook)
3551 return 0;
3552
3553 return env->analyzer_ops->insn_hook(env, insn_idx, prev_insn_idx);
3554 }
3555
3556 static int do_check(struct bpf_verifier_env *env)
3557 {
3558 struct bpf_verifier_state *state = &env->cur_state;
3559 struct bpf_insn *insns = env->prog->insnsi;
3560 struct bpf_reg_state *regs = state->regs;
3561 int insn_cnt = env->prog->len;
3562 int insn_idx, prev_insn_idx = 0;
3563 int insn_processed = 0;
3564 bool do_print_state = false;
3565
3566 init_reg_state(regs);
3567 state->parent = NULL;
3568 insn_idx = 0;
3569 for (;;) {
3570 struct bpf_insn *insn;
3571 u8 class;
3572 int err;
3573
3574 if (insn_idx >= insn_cnt) {
3575 verbose("invalid insn idx %d insn_cnt %d\n",
3576 insn_idx, insn_cnt);
3577 return -EFAULT;
3578 }
3579
3580 insn = &insns[insn_idx];
3581 class = BPF_CLASS(insn->code);
3582
3583 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3584 verbose("BPF program is too large. Processed %d insn\n",
3585 insn_processed);
3586 return -E2BIG;
3587 }
3588
3589 err = is_state_visited(env, insn_idx);
3590 if (err < 0)
3591 return err;
3592 if (err == 1) {
3593 /* found equivalent state, can prune the search */
3594 if (log_level) {
3595 if (do_print_state)
3596 verbose("\nfrom %d to %d: safe\n",
3597 prev_insn_idx, insn_idx);
3598 else
3599 verbose("%d: safe\n", insn_idx);
3600 }
3601 goto process_bpf_exit;
3602 }
3603
3604 if (need_resched())
3605 cond_resched();
3606
3607 if (log_level > 1 || (log_level && do_print_state)) {
3608 if (log_level > 1)
3609 verbose("%d:", insn_idx);
3610 else
3611 verbose("\nfrom %d to %d:",
3612 prev_insn_idx, insn_idx);
3613 print_verifier_state(&env->cur_state);
3614 do_print_state = false;
3615 }
3616
3617 if (log_level) {
3618 verbose("%d: ", insn_idx);
3619 print_bpf_insn(env, insn);
3620 }
3621
3622 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3623 if (err)
3624 return err;
3625
3626 if (class == BPF_ALU || class == BPF_ALU64) {
3627 err = check_alu_op(env, insn);
3628 if (err)
3629 return err;
3630
3631 } else if (class == BPF_LDX) {
3632 enum bpf_reg_type *prev_src_type, src_reg_type;
3633
3634 /* check for reserved fields is already done */
3635
3636 /* check src operand */
3637 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3638 if (err)
3639 return err;
3640
3641 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3642 if (err)
3643 return err;
3644
3645 src_reg_type = regs[insn->src_reg].type;
3646
3647 /* check that memory (src_reg + off) is readable,
3648 * the state of dst_reg will be updated by this func
3649 */
3650 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3651 BPF_SIZE(insn->code), BPF_READ,
3652 insn->dst_reg);
3653 if (err)
3654 return err;
3655
3656 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3657
3658 if (*prev_src_type == NOT_INIT) {
3659 /* saw a valid insn
3660 * dst_reg = *(u32 *)(src_reg + off)
3661 * save type to validate intersecting paths
3662 */
3663 *prev_src_type = src_reg_type;
3664
3665 } else if (src_reg_type != *prev_src_type &&
3666 (src_reg_type == PTR_TO_CTX ||
3667 *prev_src_type == PTR_TO_CTX)) {
3668 /* ABuser program is trying to use the same insn
3669 * dst_reg = *(u32*) (src_reg + off)
3670 * with different pointer types:
3671 * src_reg == ctx in one branch and
3672 * src_reg == stack|map in some other branch.
3673 * Reject it.
3674 */
3675 verbose("same insn cannot be used with different pointers\n");
3676 return -EINVAL;
3677 }
3678
3679 } else if (class == BPF_STX) {
3680 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3681
3682 if (BPF_MODE(insn->code) == BPF_XADD) {
3683 err = check_xadd(env, insn_idx, insn);
3684 if (err)
3685 return err;
3686 insn_idx++;
3687 continue;
3688 }
3689
3690 /* check src1 operand */
3691 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3692 if (err)
3693 return err;
3694 /* check src2 operand */
3695 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3696 if (err)
3697 return err;
3698
3699 dst_reg_type = regs[insn->dst_reg].type;
3700
3701 /* check that memory (dst_reg + off) is writeable */
3702 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3703 BPF_SIZE(insn->code), BPF_WRITE,
3704 insn->src_reg);
3705 if (err)
3706 return err;
3707
3708 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3709
3710 if (*prev_dst_type == NOT_INIT) {
3711 *prev_dst_type = dst_reg_type;
3712 } else if (dst_reg_type != *prev_dst_type &&
3713 (dst_reg_type == PTR_TO_CTX ||
3714 *prev_dst_type == PTR_TO_CTX)) {
3715 verbose("same insn cannot be used with different pointers\n");
3716 return -EINVAL;
3717 }
3718
3719 } else if (class == BPF_ST) {
3720 if (BPF_MODE(insn->code) != BPF_MEM ||
3721 insn->src_reg != BPF_REG_0) {
3722 verbose("BPF_ST uses reserved fields\n");
3723 return -EINVAL;
3724 }
3725 /* check src operand */
3726 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3727 if (err)
3728 return err;
3729
3730 /* check that memory (dst_reg + off) is writeable */
3731 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3732 BPF_SIZE(insn->code), BPF_WRITE,
3733 -1);
3734 if (err)
3735 return err;
3736
3737 } else if (class == BPF_JMP) {
3738 u8 opcode = BPF_OP(insn->code);
3739
3740 if (opcode == BPF_CALL) {
3741 if (BPF_SRC(insn->code) != BPF_K ||
3742 insn->off != 0 ||
3743 insn->src_reg != BPF_REG_0 ||
3744 insn->dst_reg != BPF_REG_0) {
3745 verbose("BPF_CALL uses reserved fields\n");
3746 return -EINVAL;
3747 }
3748
3749 err = check_call(env, insn->imm, insn_idx);
3750 if (err)
3751 return err;
3752
3753 } else if (opcode == BPF_JA) {
3754 if (BPF_SRC(insn->code) != BPF_K ||
3755 insn->imm != 0 ||
3756 insn->src_reg != BPF_REG_0 ||
3757 insn->dst_reg != BPF_REG_0) {
3758 verbose("BPF_JA uses reserved fields\n");
3759 return -EINVAL;
3760 }
3761
3762 insn_idx += insn->off + 1;
3763 continue;
3764
3765 } else if (opcode == BPF_EXIT) {
3766 if (BPF_SRC(insn->code) != BPF_K ||
3767 insn->imm != 0 ||
3768 insn->src_reg != BPF_REG_0 ||
3769 insn->dst_reg != BPF_REG_0) {
3770 verbose("BPF_EXIT uses reserved fields\n");
3771 return -EINVAL;
3772 }
3773
3774 /* eBPF calling convetion is such that R0 is used
3775 * to return the value from eBPF program.
3776 * Make sure that it's readable at this time
3777 * of bpf_exit, which means that program wrote
3778 * something into it earlier
3779 */
3780 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
3781 if (err)
3782 return err;
3783
3784 if (is_pointer_value(env, BPF_REG_0)) {
3785 verbose("R0 leaks addr as return value\n");
3786 return -EACCES;
3787 }
3788
3789 process_bpf_exit:
3790 insn_idx = pop_stack(env, &prev_insn_idx);
3791 if (insn_idx < 0) {
3792 break;
3793 } else {
3794 do_print_state = true;
3795 continue;
3796 }
3797 } else {
3798 err = check_cond_jmp_op(env, insn, &insn_idx);
3799 if (err)
3800 return err;
3801 }
3802 } else if (class == BPF_LD) {
3803 u8 mode = BPF_MODE(insn->code);
3804
3805 if (mode == BPF_ABS || mode == BPF_IND) {
3806 err = check_ld_abs(env, insn);
3807 if (err)
3808 return err;
3809
3810 } else if (mode == BPF_IMM) {
3811 err = check_ld_imm(env, insn);
3812 if (err)
3813 return err;
3814
3815 insn_idx++;
3816 } else {
3817 verbose("invalid BPF_LD mode\n");
3818 return -EINVAL;
3819 }
3820 } else {
3821 verbose("unknown insn class %d\n", class);
3822 return -EINVAL;
3823 }
3824
3825 insn_idx++;
3826 }
3827
3828 verbose("processed %d insns, stack depth %d\n",
3829 insn_processed, env->prog->aux->stack_depth);
3830 return 0;
3831 }
3832
3833 static int check_map_prealloc(struct bpf_map *map)
3834 {
3835 return (map->map_type != BPF_MAP_TYPE_HASH &&
3836 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
3837 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
3838 !(map->map_flags & BPF_F_NO_PREALLOC);
3839 }
3840
3841 static int check_map_prog_compatibility(struct bpf_map *map,
3842 struct bpf_prog *prog)
3843
3844 {
3845 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
3846 * preallocated hash maps, since doing memory allocation
3847 * in overflow_handler can crash depending on where nmi got
3848 * triggered.
3849 */
3850 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
3851 if (!check_map_prealloc(map)) {
3852 verbose("perf_event programs can only use preallocated hash map\n");
3853 return -EINVAL;
3854 }
3855 if (map->inner_map_meta &&
3856 !check_map_prealloc(map->inner_map_meta)) {
3857 verbose("perf_event programs can only use preallocated inner hash map\n");
3858 return -EINVAL;
3859 }
3860 }
3861 return 0;
3862 }
3863
3864 /* look for pseudo eBPF instructions that access map FDs and
3865 * replace them with actual map pointers
3866 */
3867 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
3868 {
3869 struct bpf_insn *insn = env->prog->insnsi;
3870 int insn_cnt = env->prog->len;
3871 int i, j, err;
3872
3873 err = bpf_prog_calc_tag(env->prog);
3874 if (err)
3875 return err;
3876
3877 for (i = 0; i < insn_cnt; i++, insn++) {
3878 if (BPF_CLASS(insn->code) == BPF_LDX &&
3879 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
3880 verbose("BPF_LDX uses reserved fields\n");
3881 return -EINVAL;
3882 }
3883
3884 if (BPF_CLASS(insn->code) == BPF_STX &&
3885 ((BPF_MODE(insn->code) != BPF_MEM &&
3886 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
3887 verbose("BPF_STX uses reserved fields\n");
3888 return -EINVAL;
3889 }
3890
3891 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
3892 struct bpf_map *map;
3893 struct fd f;
3894
3895 if (i == insn_cnt - 1 || insn[1].code != 0 ||
3896 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
3897 insn[1].off != 0) {
3898 verbose("invalid bpf_ld_imm64 insn\n");
3899 return -EINVAL;
3900 }
3901
3902 if (insn->src_reg == 0)
3903 /* valid generic load 64-bit imm */
3904 goto next_insn;
3905
3906 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
3907 verbose("unrecognized bpf_ld_imm64 insn\n");
3908 return -EINVAL;
3909 }
3910
3911 f = fdget(insn->imm);
3912 map = __bpf_map_get(f);
3913 if (IS_ERR(map)) {
3914 verbose("fd %d is not pointing to valid bpf_map\n",
3915 insn->imm);
3916 return PTR_ERR(map);
3917 }
3918
3919 err = check_map_prog_compatibility(map, env->prog);
3920 if (err) {
3921 fdput(f);
3922 return err;
3923 }
3924
3925 /* store map pointer inside BPF_LD_IMM64 instruction */
3926 insn[0].imm = (u32) (unsigned long) map;
3927 insn[1].imm = ((u64) (unsigned long) map) >> 32;
3928
3929 /* check whether we recorded this map already */
3930 for (j = 0; j < env->used_map_cnt; j++)
3931 if (env->used_maps[j] == map) {
3932 fdput(f);
3933 goto next_insn;
3934 }
3935
3936 if (env->used_map_cnt >= MAX_USED_MAPS) {
3937 fdput(f);
3938 return -E2BIG;
3939 }
3940
3941 /* hold the map. If the program is rejected by verifier,
3942 * the map will be released by release_maps() or it
3943 * will be used by the valid program until it's unloaded
3944 * and all maps are released in free_bpf_prog_info()
3945 */
3946 map = bpf_map_inc(map, false);
3947 if (IS_ERR(map)) {
3948 fdput(f);
3949 return PTR_ERR(map);
3950 }
3951 env->used_maps[env->used_map_cnt++] = map;
3952
3953 fdput(f);
3954 next_insn:
3955 insn++;
3956 i++;
3957 }
3958 }
3959
3960 /* now all pseudo BPF_LD_IMM64 instructions load valid
3961 * 'struct bpf_map *' into a register instead of user map_fd.
3962 * These pointers will be used later by verifier to validate map access.
3963 */
3964 return 0;
3965 }
3966
3967 /* drop refcnt of maps used by the rejected program */
3968 static void release_maps(struct bpf_verifier_env *env)
3969 {
3970 int i;
3971
3972 for (i = 0; i < env->used_map_cnt; i++)
3973 bpf_map_put(env->used_maps[i]);
3974 }
3975
3976 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
3977 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
3978 {
3979 struct bpf_insn *insn = env->prog->insnsi;
3980 int insn_cnt = env->prog->len;
3981 int i;
3982
3983 for (i = 0; i < insn_cnt; i++, insn++)
3984 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
3985 insn->src_reg = 0;
3986 }
3987
3988 /* single env->prog->insni[off] instruction was replaced with the range
3989 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
3990 * [0, off) and [off, end) to new locations, so the patched range stays zero
3991 */
3992 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
3993 u32 off, u32 cnt)
3994 {
3995 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
3996
3997 if (cnt == 1)
3998 return 0;
3999 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4000 if (!new_data)
4001 return -ENOMEM;
4002 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4003 memcpy(new_data + off + cnt - 1, old_data + off,
4004 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4005 env->insn_aux_data = new_data;
4006 vfree(old_data);
4007 return 0;
4008 }
4009
4010 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4011 const struct bpf_insn *patch, u32 len)
4012 {
4013 struct bpf_prog *new_prog;
4014
4015 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4016 if (!new_prog)
4017 return NULL;
4018 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4019 return NULL;
4020 return new_prog;
4021 }
4022
4023 /* convert load instructions that access fields of 'struct __sk_buff'
4024 * into sequence of instructions that access fields of 'struct sk_buff'
4025 */
4026 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4027 {
4028 const struct bpf_verifier_ops *ops = env->prog->aux->ops;
4029 int i, cnt, size, ctx_field_size, delta = 0;
4030 const int insn_cnt = env->prog->len;
4031 struct bpf_insn insn_buf[16], *insn;
4032 struct bpf_prog *new_prog;
4033 enum bpf_access_type type;
4034 bool is_narrower_load;
4035 u32 target_size;
4036
4037 if (ops->gen_prologue) {
4038 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4039 env->prog);
4040 if (cnt >= ARRAY_SIZE(insn_buf)) {
4041 verbose("bpf verifier is misconfigured\n");
4042 return -EINVAL;
4043 } else if (cnt) {
4044 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4045 if (!new_prog)
4046 return -ENOMEM;
4047
4048 env->prog = new_prog;
4049 delta += cnt - 1;
4050 }
4051 }
4052
4053 if (!ops->convert_ctx_access)
4054 return 0;
4055
4056 insn = env->prog->insnsi + delta;
4057
4058 for (i = 0; i < insn_cnt; i++, insn++) {
4059 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4060 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4061 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4062 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4063 type = BPF_READ;
4064 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4065 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4066 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4067 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4068 type = BPF_WRITE;
4069 else
4070 continue;
4071
4072 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4073 continue;
4074
4075 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4076 size = BPF_LDST_BYTES(insn);
4077
4078 /* If the read access is a narrower load of the field,
4079 * convert to a 4/8-byte load, to minimum program type specific
4080 * convert_ctx_access changes. If conversion is successful,
4081 * we will apply proper mask to the result.
4082 */
4083 is_narrower_load = size < ctx_field_size;
4084 if (is_narrower_load) {
4085 u32 off = insn->off;
4086 u8 size_code;
4087
4088 if (type == BPF_WRITE) {
4089 verbose("bpf verifier narrow ctx access misconfigured\n");
4090 return -EINVAL;
4091 }
4092
4093 size_code = BPF_H;
4094 if (ctx_field_size == 4)
4095 size_code = BPF_W;
4096 else if (ctx_field_size == 8)
4097 size_code = BPF_DW;
4098
4099 insn->off = off & ~(ctx_field_size - 1);
4100 insn->code = BPF_LDX | BPF_MEM | size_code;
4101 }
4102
4103 target_size = 0;
4104 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4105 &target_size);
4106 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4107 (ctx_field_size && !target_size)) {
4108 verbose("bpf verifier is misconfigured\n");
4109 return -EINVAL;
4110 }
4111
4112 if (is_narrower_load && size < target_size) {
4113 if (ctx_field_size <= 4)
4114 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4115 (1 << size * 8) - 1);
4116 else
4117 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4118 (1 << size * 8) - 1);
4119 }
4120
4121 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4122 if (!new_prog)
4123 return -ENOMEM;
4124
4125 delta += cnt - 1;
4126
4127 /* keep walking new program and skip insns we just inserted */
4128 env->prog = new_prog;
4129 insn = new_prog->insnsi + i + delta;
4130 }
4131
4132 return 0;
4133 }
4134
4135 /* fixup insn->imm field of bpf_call instructions
4136 * and inline eligible helpers as explicit sequence of BPF instructions
4137 *
4138 * this function is called after eBPF program passed verification
4139 */
4140 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4141 {
4142 struct bpf_prog *prog = env->prog;
4143 struct bpf_insn *insn = prog->insnsi;
4144 const struct bpf_func_proto *fn;
4145 const int insn_cnt = prog->len;
4146 struct bpf_insn insn_buf[16];
4147 struct bpf_prog *new_prog;
4148 struct bpf_map *map_ptr;
4149 int i, cnt, delta = 0;
4150
4151 for (i = 0; i < insn_cnt; i++, insn++) {
4152 if (insn->code != (BPF_JMP | BPF_CALL))
4153 continue;
4154
4155 if (insn->imm == BPF_FUNC_get_route_realm)
4156 prog->dst_needed = 1;
4157 if (insn->imm == BPF_FUNC_get_prandom_u32)
4158 bpf_user_rnd_init_once();
4159 if (insn->imm == BPF_FUNC_tail_call) {
4160 /* If we tail call into other programs, we
4161 * cannot make any assumptions since they can
4162 * be replaced dynamically during runtime in
4163 * the program array.
4164 */
4165 prog->cb_access = 1;
4166 env->prog->aux->stack_depth = MAX_BPF_STACK;
4167
4168 /* mark bpf_tail_call as different opcode to avoid
4169 * conditional branch in the interpeter for every normal
4170 * call and to prevent accidental JITing by JIT compiler
4171 * that doesn't support bpf_tail_call yet
4172 */
4173 insn->imm = 0;
4174 insn->code = BPF_JMP | BPF_TAIL_CALL;
4175 continue;
4176 }
4177
4178 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4179 * handlers are currently limited to 64 bit only.
4180 */
4181 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4182 insn->imm == BPF_FUNC_map_lookup_elem) {
4183 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4184 if (map_ptr == BPF_MAP_PTR_POISON ||
4185 !map_ptr->ops->map_gen_lookup)
4186 goto patch_call_imm;
4187
4188 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4189 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4190 verbose("bpf verifier is misconfigured\n");
4191 return -EINVAL;
4192 }
4193
4194 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4195 cnt);
4196 if (!new_prog)
4197 return -ENOMEM;
4198
4199 delta += cnt - 1;
4200
4201 /* keep walking new program and skip insns we just inserted */
4202 env->prog = prog = new_prog;
4203 insn = new_prog->insnsi + i + delta;
4204 continue;
4205 }
4206
4207 if (insn->imm == BPF_FUNC_redirect_map) {
4208 /* Note, we cannot use prog directly as imm as subsequent
4209 * rewrites would still change the prog pointer. The only
4210 * stable address we can use is aux, which also works with
4211 * prog clones during blinding.
4212 */
4213 u64 addr = (unsigned long)prog->aux;
4214 struct bpf_insn r4_ld[] = {
4215 BPF_LD_IMM64(BPF_REG_4, addr),
4216 *insn,
4217 };
4218 cnt = ARRAY_SIZE(r4_ld);
4219
4220 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4221 if (!new_prog)
4222 return -ENOMEM;
4223
4224 delta += cnt - 1;
4225 env->prog = prog = new_prog;
4226 insn = new_prog->insnsi + i + delta;
4227 }
4228 patch_call_imm:
4229 fn = prog->aux->ops->get_func_proto(insn->imm);
4230 /* all functions that have prototype and verifier allowed
4231 * programs to call them, must be real in-kernel functions
4232 */
4233 if (!fn->func) {
4234 verbose("kernel subsystem misconfigured func %s#%d\n",
4235 func_id_name(insn->imm), insn->imm);
4236 return -EFAULT;
4237 }
4238 insn->imm = fn->func - __bpf_call_base;
4239 }
4240
4241 return 0;
4242 }
4243
4244 static void free_states(struct bpf_verifier_env *env)
4245 {
4246 struct bpf_verifier_state_list *sl, *sln;
4247 int i;
4248
4249 if (!env->explored_states)
4250 return;
4251
4252 for (i = 0; i < env->prog->len; i++) {
4253 sl = env->explored_states[i];
4254
4255 if (sl)
4256 while (sl != STATE_LIST_MARK) {
4257 sln = sl->next;
4258 kfree(sl);
4259 sl = sln;
4260 }
4261 }
4262
4263 kfree(env->explored_states);
4264 }
4265
4266 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4267 {
4268 char __user *log_ubuf = NULL;
4269 struct bpf_verifier_env *env;
4270 int ret = -EINVAL;
4271
4272 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4273 * allocate/free it every time bpf_check() is called
4274 */
4275 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4276 if (!env)
4277 return -ENOMEM;
4278
4279 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4280 (*prog)->len);
4281 ret = -ENOMEM;
4282 if (!env->insn_aux_data)
4283 goto err_free_env;
4284 env->prog = *prog;
4285
4286 /* grab the mutex to protect few globals used by verifier */
4287 mutex_lock(&bpf_verifier_lock);
4288
4289 if (attr->log_level || attr->log_buf || attr->log_size) {
4290 /* user requested verbose verifier output
4291 * and supplied buffer to store the verification trace
4292 */
4293 log_level = attr->log_level;
4294 log_ubuf = (char __user *) (unsigned long) attr->log_buf;
4295 log_size = attr->log_size;
4296 log_len = 0;
4297
4298 ret = -EINVAL;
4299 /* log_* values have to be sane */
4300 if (log_size < 128 || log_size > UINT_MAX >> 8 ||
4301 log_level == 0 || log_ubuf == NULL)
4302 goto err_unlock;
4303
4304 ret = -ENOMEM;
4305 log_buf = vmalloc(log_size);
4306 if (!log_buf)
4307 goto err_unlock;
4308 } else {
4309 log_level = 0;
4310 }
4311
4312 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4313 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4314 env->strict_alignment = true;
4315
4316 ret = replace_map_fd_with_map_ptr(env);
4317 if (ret < 0)
4318 goto skip_full_check;
4319
4320 env->explored_states = kcalloc(env->prog->len,
4321 sizeof(struct bpf_verifier_state_list *),
4322 GFP_USER);
4323 ret = -ENOMEM;
4324 if (!env->explored_states)
4325 goto skip_full_check;
4326
4327 ret = check_cfg(env);
4328 if (ret < 0)
4329 goto skip_full_check;
4330
4331 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4332
4333 ret = do_check(env);
4334
4335 skip_full_check:
4336 while (pop_stack(env, NULL) >= 0);
4337 free_states(env);
4338
4339 if (ret == 0)
4340 /* program is valid, convert *(u32*)(ctx + off) accesses */
4341 ret = convert_ctx_accesses(env);
4342
4343 if (ret == 0)
4344 ret = fixup_bpf_calls(env);
4345
4346 if (log_level && log_len >= log_size - 1) {
4347 BUG_ON(log_len >= log_size);
4348 /* verifier log exceeded user supplied buffer */
4349 ret = -ENOSPC;
4350 /* fall through to return what was recorded */
4351 }
4352
4353 /* copy verifier log back to user space including trailing zero */
4354 if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
4355 ret = -EFAULT;
4356 goto free_log_buf;
4357 }
4358
4359 if (ret == 0 && env->used_map_cnt) {
4360 /* if program passed verifier, update used_maps in bpf_prog_info */
4361 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4362 sizeof(env->used_maps[0]),
4363 GFP_KERNEL);
4364
4365 if (!env->prog->aux->used_maps) {
4366 ret = -ENOMEM;
4367 goto free_log_buf;
4368 }
4369
4370 memcpy(env->prog->aux->used_maps, env->used_maps,
4371 sizeof(env->used_maps[0]) * env->used_map_cnt);
4372 env->prog->aux->used_map_cnt = env->used_map_cnt;
4373
4374 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4375 * bpf_ld_imm64 instructions
4376 */
4377 convert_pseudo_ld_imm64(env);
4378 }
4379
4380 free_log_buf:
4381 if (log_level)
4382 vfree(log_buf);
4383 if (!env->prog->aux->used_maps)
4384 /* if we didn't copy map pointers into bpf_prog_info, release
4385 * them now. Otherwise free_bpf_prog_info() will release them.
4386 */
4387 release_maps(env);
4388 *prog = env->prog;
4389 err_unlock:
4390 mutex_unlock(&bpf_verifier_lock);
4391 vfree(env->insn_aux_data);
4392 err_free_env:
4393 kfree(env);
4394 return ret;
4395 }
4396
4397 int bpf_analyzer(struct bpf_prog *prog, const struct bpf_ext_analyzer_ops *ops,
4398 void *priv)
4399 {
4400 struct bpf_verifier_env *env;
4401 int ret;
4402
4403 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4404 if (!env)
4405 return -ENOMEM;
4406
4407 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4408 prog->len);
4409 ret = -ENOMEM;
4410 if (!env->insn_aux_data)
4411 goto err_free_env;
4412 env->prog = prog;
4413 env->analyzer_ops = ops;
4414 env->analyzer_priv = priv;
4415
4416 /* grab the mutex to protect few globals used by verifier */
4417 mutex_lock(&bpf_verifier_lock);
4418
4419 log_level = 0;
4420
4421 env->strict_alignment = false;
4422 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4423 env->strict_alignment = true;
4424
4425 env->explored_states = kcalloc(env->prog->len,
4426 sizeof(struct bpf_verifier_state_list *),
4427 GFP_KERNEL);
4428 ret = -ENOMEM;
4429 if (!env->explored_states)
4430 goto skip_full_check;
4431
4432 ret = check_cfg(env);
4433 if (ret < 0)
4434 goto skip_full_check;
4435
4436 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4437
4438 ret = do_check(env);
4439
4440 skip_full_check:
4441 while (pop_stack(env, NULL) >= 0);
4442 free_states(env);
4443
4444 mutex_unlock(&bpf_verifier_lock);
4445 vfree(env->insn_aux_data);
4446 err_free_env:
4447 kfree(env);
4448 return ret;
4449 }
4450 EXPORT_SYMBOL_GPL(bpf_analyzer);
ubuntu-bionic-kernel mirror