]> git.ipfire.org Git - thirdparty/linux.git/blob - kernel/bpf/verifier.c
Merge tag 'io_uring-5.7-2020-05-22' of git://git.kernel.dk/linux-block
[thirdparty/linux.git] / kernel / bpf / verifier.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
5 */
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24
25 #include "disasm.h"
26
27 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
28 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
29 [_id] = & _name ## _verifier_ops,
30 #define BPF_MAP_TYPE(_id, _ops)
31 #include <linux/bpf_types.h>
32 #undef BPF_PROG_TYPE
33 #undef BPF_MAP_TYPE
34 };
35
36 /* bpf_check() is a static code analyzer that walks eBPF program
37 * instruction by instruction and updates register/stack state.
38 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
39 *
40 * The first pass is depth-first-search to check that the program is a DAG.
41 * It rejects the following programs:
42 * - larger than BPF_MAXINSNS insns
43 * - if loop is present (detected via back-edge)
44 * - unreachable insns exist (shouldn't be a forest. program = one function)
45 * - out of bounds or malformed jumps
46 * The second pass is all possible path descent from the 1st insn.
47 * Since it's analyzing all pathes through the program, the length of the
48 * analysis is limited to 64k insn, which may be hit even if total number of
49 * insn is less then 4K, but there are too many branches that change stack/regs.
50 * Number of 'branches to be analyzed' is limited to 1k
51 *
52 * On entry to each instruction, each register has a type, and the instruction
53 * changes the types of the registers depending on instruction semantics.
54 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
55 * copied to R1.
56 *
57 * All registers are 64-bit.
58 * R0 - return register
59 * R1-R5 argument passing registers
60 * R6-R9 callee saved registers
61 * R10 - frame pointer read-only
62 *
63 * At the start of BPF program the register R1 contains a pointer to bpf_context
64 * and has type PTR_TO_CTX.
65 *
66 * Verifier tracks arithmetic operations on pointers in case:
67 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
68 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
69 * 1st insn copies R10 (which has FRAME_PTR) type into R1
70 * and 2nd arithmetic instruction is pattern matched to recognize
71 * that it wants to construct a pointer to some element within stack.
72 * So after 2nd insn, the register R1 has type PTR_TO_STACK
73 * (and -20 constant is saved for further stack bounds checking).
74 * Meaning that this reg is a pointer to stack plus known immediate constant.
75 *
76 * Most of the time the registers have SCALAR_VALUE type, which
77 * means the register has some value, but it's not a valid pointer.
78 * (like pointer plus pointer becomes SCALAR_VALUE type)
79 *
80 * When verifier sees load or store instructions the type of base register
81 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
82 * four pointer types recognized by check_mem_access() function.
83 *
84 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
85 * and the range of [ptr, ptr + map's value_size) is accessible.
86 *
87 * registers used to pass values to function calls are checked against
88 * function argument constraints.
89 *
90 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
91 * It means that the register type passed to this function must be
92 * PTR_TO_STACK and it will be used inside the function as
93 * 'pointer to map element key'
94 *
95 * For example the argument constraints for bpf_map_lookup_elem():
96 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
97 * .arg1_type = ARG_CONST_MAP_PTR,
98 * .arg2_type = ARG_PTR_TO_MAP_KEY,
99 *
100 * ret_type says that this function returns 'pointer to map elem value or null'
101 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
102 * 2nd argument should be a pointer to stack, which will be used inside
103 * the helper function as a pointer to map element key.
104 *
105 * On the kernel side the helper function looks like:
106 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
107 * {
108 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
109 * void *key = (void *) (unsigned long) r2;
110 * void *value;
111 *
112 * here kernel can access 'key' and 'map' pointers safely, knowing that
113 * [key, key + map->key_size) bytes are valid and were initialized on
114 * the stack of eBPF program.
115 * }
116 *
117 * Corresponding eBPF program may look like:
118 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
119 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
120 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
121 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
122 * here verifier looks at prototype of map_lookup_elem() and sees:
123 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
124 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
125 *
126 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
127 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
128 * and were initialized prior to this call.
129 * If it's ok, then verifier allows this BPF_CALL insn and looks at
130 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
131 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
132 * returns ether pointer to map value or NULL.
133 *
134 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
135 * insn, the register holding that pointer in the true branch changes state to
136 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
137 * branch. See check_cond_jmp_op().
138 *
139 * After the call R0 is set to return type of the function and registers R1-R5
140 * are set to NOT_INIT to indicate that they are no longer readable.
141 *
142 * The following reference types represent a potential reference to a kernel
143 * resource which, after first being allocated, must be checked and freed by
144 * the BPF program:
145 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
146 *
147 * When the verifier sees a helper call return a reference type, it allocates a
148 * pointer id for the reference and stores it in the current function state.
149 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
150 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
151 * passes through a NULL-check conditional. For the branch wherein the state is
152 * changed to CONST_IMM, the verifier releases the reference.
153 *
154 * For each helper function that allocates a reference, such as
155 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
156 * bpf_sk_release(). When a reference type passes into the release function,
157 * the verifier also releases the reference. If any unchecked or unreleased
158 * reference remains at the end of the program, the verifier rejects it.
159 */
160
161 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
162 struct bpf_verifier_stack_elem {
163 /* verifer state is 'st'
164 * before processing instruction 'insn_idx'
165 * and after processing instruction 'prev_insn_idx'
166 */
167 struct bpf_verifier_state st;
168 int insn_idx;
169 int prev_insn_idx;
170 struct bpf_verifier_stack_elem *next;
171 };
172
173 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
174 #define BPF_COMPLEXITY_LIMIT_STATES 64
175
176 #define BPF_MAP_KEY_POISON (1ULL << 63)
177 #define BPF_MAP_KEY_SEEN (1ULL << 62)
178
179 #define BPF_MAP_PTR_UNPRIV 1UL
180 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
181 POISON_POINTER_DELTA))
182 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
183
184 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
185 {
186 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
187 }
188
189 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
190 {
191 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
192 }
193
194 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
195 const struct bpf_map *map, bool unpriv)
196 {
197 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
198 unpriv |= bpf_map_ptr_unpriv(aux);
199 aux->map_ptr_state = (unsigned long)map |
200 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
201 }
202
203 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
204 {
205 return aux->map_key_state & BPF_MAP_KEY_POISON;
206 }
207
208 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
209 {
210 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
211 }
212
213 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
214 {
215 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
216 }
217
218 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
219 {
220 bool poisoned = bpf_map_key_poisoned(aux);
221
222 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
223 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
224 }
225
226 struct bpf_call_arg_meta {
227 struct bpf_map *map_ptr;
228 bool raw_mode;
229 bool pkt_access;
230 int regno;
231 int access_size;
232 u64 msize_max_value;
233 int ref_obj_id;
234 int func_id;
235 u32 btf_id;
236 };
237
238 struct btf *btf_vmlinux;
239
240 static DEFINE_MUTEX(bpf_verifier_lock);
241
242 static const struct bpf_line_info *
243 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
244 {
245 const struct bpf_line_info *linfo;
246 const struct bpf_prog *prog;
247 u32 i, nr_linfo;
248
249 prog = env->prog;
250 nr_linfo = prog->aux->nr_linfo;
251
252 if (!nr_linfo || insn_off >= prog->len)
253 return NULL;
254
255 linfo = prog->aux->linfo;
256 for (i = 1; i < nr_linfo; i++)
257 if (insn_off < linfo[i].insn_off)
258 break;
259
260 return &linfo[i - 1];
261 }
262
263 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
264 va_list args)
265 {
266 unsigned int n;
267
268 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
269
270 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
271 "verifier log line truncated - local buffer too short\n");
272
273 n = min(log->len_total - log->len_used - 1, n);
274 log->kbuf[n] = '\0';
275
276 if (log->level == BPF_LOG_KERNEL) {
277 pr_err("BPF:%s\n", log->kbuf);
278 return;
279 }
280 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
281 log->len_used += n;
282 else
283 log->ubuf = NULL;
284 }
285
286 /* log_level controls verbosity level of eBPF verifier.
287 * bpf_verifier_log_write() is used to dump the verification trace to the log,
288 * so the user can figure out what's wrong with the program
289 */
290 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
291 const char *fmt, ...)
292 {
293 va_list args;
294
295 if (!bpf_verifier_log_needed(&env->log))
296 return;
297
298 va_start(args, fmt);
299 bpf_verifier_vlog(&env->log, fmt, args);
300 va_end(args);
301 }
302 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
303
304 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
305 {
306 struct bpf_verifier_env *env = private_data;
307 va_list args;
308
309 if (!bpf_verifier_log_needed(&env->log))
310 return;
311
312 va_start(args, fmt);
313 bpf_verifier_vlog(&env->log, fmt, args);
314 va_end(args);
315 }
316
317 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
318 const char *fmt, ...)
319 {
320 va_list args;
321
322 if (!bpf_verifier_log_needed(log))
323 return;
324
325 va_start(args, fmt);
326 bpf_verifier_vlog(log, fmt, args);
327 va_end(args);
328 }
329
330 static const char *ltrim(const char *s)
331 {
332 while (isspace(*s))
333 s++;
334
335 return s;
336 }
337
338 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
339 u32 insn_off,
340 const char *prefix_fmt, ...)
341 {
342 const struct bpf_line_info *linfo;
343
344 if (!bpf_verifier_log_needed(&env->log))
345 return;
346
347 linfo = find_linfo(env, insn_off);
348 if (!linfo || linfo == env->prev_linfo)
349 return;
350
351 if (prefix_fmt) {
352 va_list args;
353
354 va_start(args, prefix_fmt);
355 bpf_verifier_vlog(&env->log, prefix_fmt, args);
356 va_end(args);
357 }
358
359 verbose(env, "%s\n",
360 ltrim(btf_name_by_offset(env->prog->aux->btf,
361 linfo->line_off)));
362
363 env->prev_linfo = linfo;
364 }
365
366 static bool type_is_pkt_pointer(enum bpf_reg_type type)
367 {
368 return type == PTR_TO_PACKET ||
369 type == PTR_TO_PACKET_META;
370 }
371
372 static bool type_is_sk_pointer(enum bpf_reg_type type)
373 {
374 return type == PTR_TO_SOCKET ||
375 type == PTR_TO_SOCK_COMMON ||
376 type == PTR_TO_TCP_SOCK ||
377 type == PTR_TO_XDP_SOCK;
378 }
379
380 static bool reg_type_may_be_null(enum bpf_reg_type type)
381 {
382 return type == PTR_TO_MAP_VALUE_OR_NULL ||
383 type == PTR_TO_SOCKET_OR_NULL ||
384 type == PTR_TO_SOCK_COMMON_OR_NULL ||
385 type == PTR_TO_TCP_SOCK_OR_NULL;
386 }
387
388 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
389 {
390 return reg->type == PTR_TO_MAP_VALUE &&
391 map_value_has_spin_lock(reg->map_ptr);
392 }
393
394 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
395 {
396 return type == PTR_TO_SOCKET ||
397 type == PTR_TO_SOCKET_OR_NULL ||
398 type == PTR_TO_TCP_SOCK ||
399 type == PTR_TO_TCP_SOCK_OR_NULL;
400 }
401
402 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
403 {
404 return type == ARG_PTR_TO_SOCK_COMMON;
405 }
406
407 /* Determine whether the function releases some resources allocated by another
408 * function call. The first reference type argument will be assumed to be
409 * released by release_reference().
410 */
411 static bool is_release_function(enum bpf_func_id func_id)
412 {
413 return func_id == BPF_FUNC_sk_release;
414 }
415
416 static bool is_acquire_function(enum bpf_func_id func_id)
417 {
418 return func_id == BPF_FUNC_sk_lookup_tcp ||
419 func_id == BPF_FUNC_sk_lookup_udp ||
420 func_id == BPF_FUNC_skc_lookup_tcp;
421 }
422
423 static bool is_ptr_cast_function(enum bpf_func_id func_id)
424 {
425 return func_id == BPF_FUNC_tcp_sock ||
426 func_id == BPF_FUNC_sk_fullsock;
427 }
428
429 /* string representation of 'enum bpf_reg_type' */
430 static const char * const reg_type_str[] = {
431 [NOT_INIT] = "?",
432 [SCALAR_VALUE] = "inv",
433 [PTR_TO_CTX] = "ctx",
434 [CONST_PTR_TO_MAP] = "map_ptr",
435 [PTR_TO_MAP_VALUE] = "map_value",
436 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
437 [PTR_TO_STACK] = "fp",
438 [PTR_TO_PACKET] = "pkt",
439 [PTR_TO_PACKET_META] = "pkt_meta",
440 [PTR_TO_PACKET_END] = "pkt_end",
441 [PTR_TO_FLOW_KEYS] = "flow_keys",
442 [PTR_TO_SOCKET] = "sock",
443 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
444 [PTR_TO_SOCK_COMMON] = "sock_common",
445 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
446 [PTR_TO_TCP_SOCK] = "tcp_sock",
447 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
448 [PTR_TO_TP_BUFFER] = "tp_buffer",
449 [PTR_TO_XDP_SOCK] = "xdp_sock",
450 [PTR_TO_BTF_ID] = "ptr_",
451 };
452
453 static char slot_type_char[] = {
454 [STACK_INVALID] = '?',
455 [STACK_SPILL] = 'r',
456 [STACK_MISC] = 'm',
457 [STACK_ZERO] = '0',
458 };
459
460 static void print_liveness(struct bpf_verifier_env *env,
461 enum bpf_reg_liveness live)
462 {
463 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
464 verbose(env, "_");
465 if (live & REG_LIVE_READ)
466 verbose(env, "r");
467 if (live & REG_LIVE_WRITTEN)
468 verbose(env, "w");
469 if (live & REG_LIVE_DONE)
470 verbose(env, "D");
471 }
472
473 static struct bpf_func_state *func(struct bpf_verifier_env *env,
474 const struct bpf_reg_state *reg)
475 {
476 struct bpf_verifier_state *cur = env->cur_state;
477
478 return cur->frame[reg->frameno];
479 }
480
481 const char *kernel_type_name(u32 id)
482 {
483 return btf_name_by_offset(btf_vmlinux,
484 btf_type_by_id(btf_vmlinux, id)->name_off);
485 }
486
487 static void print_verifier_state(struct bpf_verifier_env *env,
488 const struct bpf_func_state *state)
489 {
490 const struct bpf_reg_state *reg;
491 enum bpf_reg_type t;
492 int i;
493
494 if (state->frameno)
495 verbose(env, " frame%d:", state->frameno);
496 for (i = 0; i < MAX_BPF_REG; i++) {
497 reg = &state->regs[i];
498 t = reg->type;
499 if (t == NOT_INIT)
500 continue;
501 verbose(env, " R%d", i);
502 print_liveness(env, reg->live);
503 verbose(env, "=%s", reg_type_str[t]);
504 if (t == SCALAR_VALUE && reg->precise)
505 verbose(env, "P");
506 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
507 tnum_is_const(reg->var_off)) {
508 /* reg->off should be 0 for SCALAR_VALUE */
509 verbose(env, "%lld", reg->var_off.value + reg->off);
510 } else {
511 if (t == PTR_TO_BTF_ID)
512 verbose(env, "%s", kernel_type_name(reg->btf_id));
513 verbose(env, "(id=%d", reg->id);
514 if (reg_type_may_be_refcounted_or_null(t))
515 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
516 if (t != SCALAR_VALUE)
517 verbose(env, ",off=%d", reg->off);
518 if (type_is_pkt_pointer(t))
519 verbose(env, ",r=%d", reg->range);
520 else if (t == CONST_PTR_TO_MAP ||
521 t == PTR_TO_MAP_VALUE ||
522 t == PTR_TO_MAP_VALUE_OR_NULL)
523 verbose(env, ",ks=%d,vs=%d",
524 reg->map_ptr->key_size,
525 reg->map_ptr->value_size);
526 if (tnum_is_const(reg->var_off)) {
527 /* Typically an immediate SCALAR_VALUE, but
528 * could be a pointer whose offset is too big
529 * for reg->off
530 */
531 verbose(env, ",imm=%llx", reg->var_off.value);
532 } else {
533 if (reg->smin_value != reg->umin_value &&
534 reg->smin_value != S64_MIN)
535 verbose(env, ",smin_value=%lld",
536 (long long)reg->smin_value);
537 if (reg->smax_value != reg->umax_value &&
538 reg->smax_value != S64_MAX)
539 verbose(env, ",smax_value=%lld",
540 (long long)reg->smax_value);
541 if (reg->umin_value != 0)
542 verbose(env, ",umin_value=%llu",
543 (unsigned long long)reg->umin_value);
544 if (reg->umax_value != U64_MAX)
545 verbose(env, ",umax_value=%llu",
546 (unsigned long long)reg->umax_value);
547 if (!tnum_is_unknown(reg->var_off)) {
548 char tn_buf[48];
549
550 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
551 verbose(env, ",var_off=%s", tn_buf);
552 }
553 if (reg->s32_min_value != reg->smin_value &&
554 reg->s32_min_value != S32_MIN)
555 verbose(env, ",s32_min_value=%d",
556 (int)(reg->s32_min_value));
557 if (reg->s32_max_value != reg->smax_value &&
558 reg->s32_max_value != S32_MAX)
559 verbose(env, ",s32_max_value=%d",
560 (int)(reg->s32_max_value));
561 if (reg->u32_min_value != reg->umin_value &&
562 reg->u32_min_value != U32_MIN)
563 verbose(env, ",u32_min_value=%d",
564 (int)(reg->u32_min_value));
565 if (reg->u32_max_value != reg->umax_value &&
566 reg->u32_max_value != U32_MAX)
567 verbose(env, ",u32_max_value=%d",
568 (int)(reg->u32_max_value));
569 }
570 verbose(env, ")");
571 }
572 }
573 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
574 char types_buf[BPF_REG_SIZE + 1];
575 bool valid = false;
576 int j;
577
578 for (j = 0; j < BPF_REG_SIZE; j++) {
579 if (state->stack[i].slot_type[j] != STACK_INVALID)
580 valid = true;
581 types_buf[j] = slot_type_char[
582 state->stack[i].slot_type[j]];
583 }
584 types_buf[BPF_REG_SIZE] = 0;
585 if (!valid)
586 continue;
587 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
588 print_liveness(env, state->stack[i].spilled_ptr.live);
589 if (state->stack[i].slot_type[0] == STACK_SPILL) {
590 reg = &state->stack[i].spilled_ptr;
591 t = reg->type;
592 verbose(env, "=%s", reg_type_str[t]);
593 if (t == SCALAR_VALUE && reg->precise)
594 verbose(env, "P");
595 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
596 verbose(env, "%lld", reg->var_off.value + reg->off);
597 } else {
598 verbose(env, "=%s", types_buf);
599 }
600 }
601 if (state->acquired_refs && state->refs[0].id) {
602 verbose(env, " refs=%d", state->refs[0].id);
603 for (i = 1; i < state->acquired_refs; i++)
604 if (state->refs[i].id)
605 verbose(env, ",%d", state->refs[i].id);
606 }
607 verbose(env, "\n");
608 }
609
610 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
611 static int copy_##NAME##_state(struct bpf_func_state *dst, \
612 const struct bpf_func_state *src) \
613 { \
614 if (!src->FIELD) \
615 return 0; \
616 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
617 /* internal bug, make state invalid to reject the program */ \
618 memset(dst, 0, sizeof(*dst)); \
619 return -EFAULT; \
620 } \
621 memcpy(dst->FIELD, src->FIELD, \
622 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
623 return 0; \
624 }
625 /* copy_reference_state() */
626 COPY_STATE_FN(reference, acquired_refs, refs, 1)
627 /* copy_stack_state() */
628 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
629 #undef COPY_STATE_FN
630
631 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
632 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
633 bool copy_old) \
634 { \
635 u32 old_size = state->COUNT; \
636 struct bpf_##NAME##_state *new_##FIELD; \
637 int slot = size / SIZE; \
638 \
639 if (size <= old_size || !size) { \
640 if (copy_old) \
641 return 0; \
642 state->COUNT = slot * SIZE; \
643 if (!size && old_size) { \
644 kfree(state->FIELD); \
645 state->FIELD = NULL; \
646 } \
647 return 0; \
648 } \
649 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
650 GFP_KERNEL); \
651 if (!new_##FIELD) \
652 return -ENOMEM; \
653 if (copy_old) { \
654 if (state->FIELD) \
655 memcpy(new_##FIELD, state->FIELD, \
656 sizeof(*new_##FIELD) * (old_size / SIZE)); \
657 memset(new_##FIELD + old_size / SIZE, 0, \
658 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
659 } \
660 state->COUNT = slot * SIZE; \
661 kfree(state->FIELD); \
662 state->FIELD = new_##FIELD; \
663 return 0; \
664 }
665 /* realloc_reference_state() */
666 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
667 /* realloc_stack_state() */
668 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
669 #undef REALLOC_STATE_FN
670
671 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
672 * make it consume minimal amount of memory. check_stack_write() access from
673 * the program calls into realloc_func_state() to grow the stack size.
674 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
675 * which realloc_stack_state() copies over. It points to previous
676 * bpf_verifier_state which is never reallocated.
677 */
678 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
679 int refs_size, bool copy_old)
680 {
681 int err = realloc_reference_state(state, refs_size, copy_old);
682 if (err)
683 return err;
684 return realloc_stack_state(state, stack_size, copy_old);
685 }
686
687 /* Acquire a pointer id from the env and update the state->refs to include
688 * this new pointer reference.
689 * On success, returns a valid pointer id to associate with the register
690 * On failure, returns a negative errno.
691 */
692 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
693 {
694 struct bpf_func_state *state = cur_func(env);
695 int new_ofs = state->acquired_refs;
696 int id, err;
697
698 err = realloc_reference_state(state, state->acquired_refs + 1, true);
699 if (err)
700 return err;
701 id = ++env->id_gen;
702 state->refs[new_ofs].id = id;
703 state->refs[new_ofs].insn_idx = insn_idx;
704
705 return id;
706 }
707
708 /* release function corresponding to acquire_reference_state(). Idempotent. */
709 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
710 {
711 int i, last_idx;
712
713 last_idx = state->acquired_refs - 1;
714 for (i = 0; i < state->acquired_refs; i++) {
715 if (state->refs[i].id == ptr_id) {
716 if (last_idx && i != last_idx)
717 memcpy(&state->refs[i], &state->refs[last_idx],
718 sizeof(*state->refs));
719 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
720 state->acquired_refs--;
721 return 0;
722 }
723 }
724 return -EINVAL;
725 }
726
727 static int transfer_reference_state(struct bpf_func_state *dst,
728 struct bpf_func_state *src)
729 {
730 int err = realloc_reference_state(dst, src->acquired_refs, false);
731 if (err)
732 return err;
733 err = copy_reference_state(dst, src);
734 if (err)
735 return err;
736 return 0;
737 }
738
739 static void free_func_state(struct bpf_func_state *state)
740 {
741 if (!state)
742 return;
743 kfree(state->refs);
744 kfree(state->stack);
745 kfree(state);
746 }
747
748 static void clear_jmp_history(struct bpf_verifier_state *state)
749 {
750 kfree(state->jmp_history);
751 state->jmp_history = NULL;
752 state->jmp_history_cnt = 0;
753 }
754
755 static void free_verifier_state(struct bpf_verifier_state *state,
756 bool free_self)
757 {
758 int i;
759
760 for (i = 0; i <= state->curframe; i++) {
761 free_func_state(state->frame[i]);
762 state->frame[i] = NULL;
763 }
764 clear_jmp_history(state);
765 if (free_self)
766 kfree(state);
767 }
768
769 /* copy verifier state from src to dst growing dst stack space
770 * when necessary to accommodate larger src stack
771 */
772 static int copy_func_state(struct bpf_func_state *dst,
773 const struct bpf_func_state *src)
774 {
775 int err;
776
777 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
778 false);
779 if (err)
780 return err;
781 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
782 err = copy_reference_state(dst, src);
783 if (err)
784 return err;
785 return copy_stack_state(dst, src);
786 }
787
788 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
789 const struct bpf_verifier_state *src)
790 {
791 struct bpf_func_state *dst;
792 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
793 int i, err;
794
795 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
796 kfree(dst_state->jmp_history);
797 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
798 if (!dst_state->jmp_history)
799 return -ENOMEM;
800 }
801 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
802 dst_state->jmp_history_cnt = src->jmp_history_cnt;
803
804 /* if dst has more stack frames then src frame, free them */
805 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
806 free_func_state(dst_state->frame[i]);
807 dst_state->frame[i] = NULL;
808 }
809 dst_state->speculative = src->speculative;
810 dst_state->curframe = src->curframe;
811 dst_state->active_spin_lock = src->active_spin_lock;
812 dst_state->branches = src->branches;
813 dst_state->parent = src->parent;
814 dst_state->first_insn_idx = src->first_insn_idx;
815 dst_state->last_insn_idx = src->last_insn_idx;
816 for (i = 0; i <= src->curframe; i++) {
817 dst = dst_state->frame[i];
818 if (!dst) {
819 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
820 if (!dst)
821 return -ENOMEM;
822 dst_state->frame[i] = dst;
823 }
824 err = copy_func_state(dst, src->frame[i]);
825 if (err)
826 return err;
827 }
828 return 0;
829 }
830
831 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
832 {
833 while (st) {
834 u32 br = --st->branches;
835
836 /* WARN_ON(br > 1) technically makes sense here,
837 * but see comment in push_stack(), hence:
838 */
839 WARN_ONCE((int)br < 0,
840 "BUG update_branch_counts:branches_to_explore=%d\n",
841 br);
842 if (br)
843 break;
844 st = st->parent;
845 }
846 }
847
848 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
849 int *insn_idx)
850 {
851 struct bpf_verifier_state *cur = env->cur_state;
852 struct bpf_verifier_stack_elem *elem, *head = env->head;
853 int err;
854
855 if (env->head == NULL)
856 return -ENOENT;
857
858 if (cur) {
859 err = copy_verifier_state(cur, &head->st);
860 if (err)
861 return err;
862 }
863 if (insn_idx)
864 *insn_idx = head->insn_idx;
865 if (prev_insn_idx)
866 *prev_insn_idx = head->prev_insn_idx;
867 elem = head->next;
868 free_verifier_state(&head->st, false);
869 kfree(head);
870 env->head = elem;
871 env->stack_size--;
872 return 0;
873 }
874
875 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
876 int insn_idx, int prev_insn_idx,
877 bool speculative)
878 {
879 struct bpf_verifier_state *cur = env->cur_state;
880 struct bpf_verifier_stack_elem *elem;
881 int err;
882
883 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
884 if (!elem)
885 goto err;
886
887 elem->insn_idx = insn_idx;
888 elem->prev_insn_idx = prev_insn_idx;
889 elem->next = env->head;
890 env->head = elem;
891 env->stack_size++;
892 err = copy_verifier_state(&elem->st, cur);
893 if (err)
894 goto err;
895 elem->st.speculative |= speculative;
896 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
897 verbose(env, "The sequence of %d jumps is too complex.\n",
898 env->stack_size);
899 goto err;
900 }
901 if (elem->st.parent) {
902 ++elem->st.parent->branches;
903 /* WARN_ON(branches > 2) technically makes sense here,
904 * but
905 * 1. speculative states will bump 'branches' for non-branch
906 * instructions
907 * 2. is_state_visited() heuristics may decide not to create
908 * a new state for a sequence of branches and all such current
909 * and cloned states will be pointing to a single parent state
910 * which might have large 'branches' count.
911 */
912 }
913 return &elem->st;
914 err:
915 free_verifier_state(env->cur_state, true);
916 env->cur_state = NULL;
917 /* pop all elements and return */
918 while (!pop_stack(env, NULL, NULL));
919 return NULL;
920 }
921
922 #define CALLER_SAVED_REGS 6
923 static const int caller_saved[CALLER_SAVED_REGS] = {
924 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
925 };
926
927 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
928 struct bpf_reg_state *reg);
929
930 /* Mark the unknown part of a register (variable offset or scalar value) as
931 * known to have the value @imm.
932 */
933 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
934 {
935 /* Clear id, off, and union(map_ptr, range) */
936 memset(((u8 *)reg) + sizeof(reg->type), 0,
937 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
938 reg->var_off = tnum_const(imm);
939 reg->smin_value = (s64)imm;
940 reg->smax_value = (s64)imm;
941 reg->umin_value = imm;
942 reg->umax_value = imm;
943
944 reg->s32_min_value = (s32)imm;
945 reg->s32_max_value = (s32)imm;
946 reg->u32_min_value = (u32)imm;
947 reg->u32_max_value = (u32)imm;
948 }
949
950 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
951 {
952 reg->var_off = tnum_const_subreg(reg->var_off, imm);
953 reg->s32_min_value = (s32)imm;
954 reg->s32_max_value = (s32)imm;
955 reg->u32_min_value = (u32)imm;
956 reg->u32_max_value = (u32)imm;
957 }
958
959 /* Mark the 'variable offset' part of a register as zero. This should be
960 * used only on registers holding a pointer type.
961 */
962 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
963 {
964 __mark_reg_known(reg, 0);
965 }
966
967 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
968 {
969 __mark_reg_known(reg, 0);
970 reg->type = SCALAR_VALUE;
971 }
972
973 static void mark_reg_known_zero(struct bpf_verifier_env *env,
974 struct bpf_reg_state *regs, u32 regno)
975 {
976 if (WARN_ON(regno >= MAX_BPF_REG)) {
977 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
978 /* Something bad happened, let's kill all regs */
979 for (regno = 0; regno < MAX_BPF_REG; regno++)
980 __mark_reg_not_init(env, regs + regno);
981 return;
982 }
983 __mark_reg_known_zero(regs + regno);
984 }
985
986 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
987 {
988 return type_is_pkt_pointer(reg->type);
989 }
990
991 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
992 {
993 return reg_is_pkt_pointer(reg) ||
994 reg->type == PTR_TO_PACKET_END;
995 }
996
997 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
998 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
999 enum bpf_reg_type which)
1000 {
1001 /* The register can already have a range from prior markings.
1002 * This is fine as long as it hasn't been advanced from its
1003 * origin.
1004 */
1005 return reg->type == which &&
1006 reg->id == 0 &&
1007 reg->off == 0 &&
1008 tnum_equals_const(reg->var_off, 0);
1009 }
1010
1011 /* Reset the min/max bounds of a register */
1012 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1013 {
1014 reg->smin_value = S64_MIN;
1015 reg->smax_value = S64_MAX;
1016 reg->umin_value = 0;
1017 reg->umax_value = U64_MAX;
1018
1019 reg->s32_min_value = S32_MIN;
1020 reg->s32_max_value = S32_MAX;
1021 reg->u32_min_value = 0;
1022 reg->u32_max_value = U32_MAX;
1023 }
1024
1025 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1026 {
1027 reg->smin_value = S64_MIN;
1028 reg->smax_value = S64_MAX;
1029 reg->umin_value = 0;
1030 reg->umax_value = U64_MAX;
1031 }
1032
1033 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1034 {
1035 reg->s32_min_value = S32_MIN;
1036 reg->s32_max_value = S32_MAX;
1037 reg->u32_min_value = 0;
1038 reg->u32_max_value = U32_MAX;
1039 }
1040
1041 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1042 {
1043 struct tnum var32_off = tnum_subreg(reg->var_off);
1044
1045 /* min signed is max(sign bit) | min(other bits) */
1046 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1047 var32_off.value | (var32_off.mask & S32_MIN));
1048 /* max signed is min(sign bit) | max(other bits) */
1049 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1050 var32_off.value | (var32_off.mask & S32_MAX));
1051 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1052 reg->u32_max_value = min(reg->u32_max_value,
1053 (u32)(var32_off.value | var32_off.mask));
1054 }
1055
1056 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1057 {
1058 /* min signed is max(sign bit) | min(other bits) */
1059 reg->smin_value = max_t(s64, reg->smin_value,
1060 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1061 /* max signed is min(sign bit) | max(other bits) */
1062 reg->smax_value = min_t(s64, reg->smax_value,
1063 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1064 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1065 reg->umax_value = min(reg->umax_value,
1066 reg->var_off.value | reg->var_off.mask);
1067 }
1068
1069 static void __update_reg_bounds(struct bpf_reg_state *reg)
1070 {
1071 __update_reg32_bounds(reg);
1072 __update_reg64_bounds(reg);
1073 }
1074
1075 /* Uses signed min/max values to inform unsigned, and vice-versa */
1076 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1077 {
1078 /* Learn sign from signed bounds.
1079 * If we cannot cross the sign boundary, then signed and unsigned bounds
1080 * are the same, so combine. This works even in the negative case, e.g.
1081 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1082 */
1083 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1084 reg->s32_min_value = reg->u32_min_value =
1085 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1086 reg->s32_max_value = reg->u32_max_value =
1087 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1088 return;
1089 }
1090 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1091 * boundary, so we must be careful.
1092 */
1093 if ((s32)reg->u32_max_value >= 0) {
1094 /* Positive. We can't learn anything from the smin, but smax
1095 * is positive, hence safe.
1096 */
1097 reg->s32_min_value = reg->u32_min_value;
1098 reg->s32_max_value = reg->u32_max_value =
1099 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1100 } else if ((s32)reg->u32_min_value < 0) {
1101 /* Negative. We can't learn anything from the smax, but smin
1102 * is negative, hence safe.
1103 */
1104 reg->s32_min_value = reg->u32_min_value =
1105 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1106 reg->s32_max_value = reg->u32_max_value;
1107 }
1108 }
1109
1110 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1111 {
1112 /* Learn sign from signed bounds.
1113 * If we cannot cross the sign boundary, then signed and unsigned bounds
1114 * are the same, so combine. This works even in the negative case, e.g.
1115 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1116 */
1117 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1118 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1119 reg->umin_value);
1120 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1121 reg->umax_value);
1122 return;
1123 }
1124 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1125 * boundary, so we must be careful.
1126 */
1127 if ((s64)reg->umax_value >= 0) {
1128 /* Positive. We can't learn anything from the smin, but smax
1129 * is positive, hence safe.
1130 */
1131 reg->smin_value = reg->umin_value;
1132 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1133 reg->umax_value);
1134 } else if ((s64)reg->umin_value < 0) {
1135 /* Negative. We can't learn anything from the smax, but smin
1136 * is negative, hence safe.
1137 */
1138 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1139 reg->umin_value);
1140 reg->smax_value = reg->umax_value;
1141 }
1142 }
1143
1144 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1145 {
1146 __reg32_deduce_bounds(reg);
1147 __reg64_deduce_bounds(reg);
1148 }
1149
1150 /* Attempts to improve var_off based on unsigned min/max information */
1151 static void __reg_bound_offset(struct bpf_reg_state *reg)
1152 {
1153 struct tnum var64_off = tnum_intersect(reg->var_off,
1154 tnum_range(reg->umin_value,
1155 reg->umax_value));
1156 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1157 tnum_range(reg->u32_min_value,
1158 reg->u32_max_value));
1159
1160 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1161 }
1162
1163 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1164 {
1165 reg->umin_value = reg->u32_min_value;
1166 reg->umax_value = reg->u32_max_value;
1167 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1168 * but must be positive otherwise set to worse case bounds
1169 * and refine later from tnum.
1170 */
1171 if (reg->s32_min_value > 0)
1172 reg->smin_value = reg->s32_min_value;
1173 else
1174 reg->smin_value = 0;
1175 if (reg->s32_max_value > 0)
1176 reg->smax_value = reg->s32_max_value;
1177 else
1178 reg->smax_value = U32_MAX;
1179 }
1180
1181 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1182 {
1183 /* special case when 64-bit register has upper 32-bit register
1184 * zeroed. Typically happens after zext or <<32, >>32 sequence
1185 * allowing us to use 32-bit bounds directly,
1186 */
1187 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1188 __reg_assign_32_into_64(reg);
1189 } else {
1190 /* Otherwise the best we can do is push lower 32bit known and
1191 * unknown bits into register (var_off set from jmp logic)
1192 * then learn as much as possible from the 64-bit tnum
1193 * known and unknown bits. The previous smin/smax bounds are
1194 * invalid here because of jmp32 compare so mark them unknown
1195 * so they do not impact tnum bounds calculation.
1196 */
1197 __mark_reg64_unbounded(reg);
1198 __update_reg_bounds(reg);
1199 }
1200
1201 /* Intersecting with the old var_off might have improved our bounds
1202 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1203 * then new var_off is (0; 0x7f...fc) which improves our umax.
1204 */
1205 __reg_deduce_bounds(reg);
1206 __reg_bound_offset(reg);
1207 __update_reg_bounds(reg);
1208 }
1209
1210 static bool __reg64_bound_s32(s64 a)
1211 {
1212 if (a > S32_MIN && a < S32_MAX)
1213 return true;
1214 return false;
1215 }
1216
1217 static bool __reg64_bound_u32(u64 a)
1218 {
1219 if (a > U32_MIN && a < U32_MAX)
1220 return true;
1221 return false;
1222 }
1223
1224 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1225 {
1226 __mark_reg32_unbounded(reg);
1227
1228 if (__reg64_bound_s32(reg->smin_value))
1229 reg->s32_min_value = (s32)reg->smin_value;
1230 if (__reg64_bound_s32(reg->smax_value))
1231 reg->s32_max_value = (s32)reg->smax_value;
1232 if (__reg64_bound_u32(reg->umin_value))
1233 reg->u32_min_value = (u32)reg->umin_value;
1234 if (__reg64_bound_u32(reg->umax_value))
1235 reg->u32_max_value = (u32)reg->umax_value;
1236
1237 /* Intersecting with the old var_off might have improved our bounds
1238 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1239 * then new var_off is (0; 0x7f...fc) which improves our umax.
1240 */
1241 __reg_deduce_bounds(reg);
1242 __reg_bound_offset(reg);
1243 __update_reg_bounds(reg);
1244 }
1245
1246 /* Mark a register as having a completely unknown (scalar) value. */
1247 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1248 struct bpf_reg_state *reg)
1249 {
1250 /*
1251 * Clear type, id, off, and union(map_ptr, range) and
1252 * padding between 'type' and union
1253 */
1254 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1255 reg->type = SCALAR_VALUE;
1256 reg->var_off = tnum_unknown;
1257 reg->frameno = 0;
1258 reg->precise = env->subprog_cnt > 1 || !env->allow_ptr_leaks;
1259 __mark_reg_unbounded(reg);
1260 }
1261
1262 static void mark_reg_unknown(struct bpf_verifier_env *env,
1263 struct bpf_reg_state *regs, u32 regno)
1264 {
1265 if (WARN_ON(regno >= MAX_BPF_REG)) {
1266 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1267 /* Something bad happened, let's kill all regs except FP */
1268 for (regno = 0; regno < BPF_REG_FP; regno++)
1269 __mark_reg_not_init(env, regs + regno);
1270 return;
1271 }
1272 __mark_reg_unknown(env, regs + regno);
1273 }
1274
1275 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1276 struct bpf_reg_state *reg)
1277 {
1278 __mark_reg_unknown(env, reg);
1279 reg->type = NOT_INIT;
1280 }
1281
1282 static void mark_reg_not_init(struct bpf_verifier_env *env,
1283 struct bpf_reg_state *regs, u32 regno)
1284 {
1285 if (WARN_ON(regno >= MAX_BPF_REG)) {
1286 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1287 /* Something bad happened, let's kill all regs except FP */
1288 for (regno = 0; regno < BPF_REG_FP; regno++)
1289 __mark_reg_not_init(env, regs + regno);
1290 return;
1291 }
1292 __mark_reg_not_init(env, regs + regno);
1293 }
1294
1295 #define DEF_NOT_SUBREG (0)
1296 static void init_reg_state(struct bpf_verifier_env *env,
1297 struct bpf_func_state *state)
1298 {
1299 struct bpf_reg_state *regs = state->regs;
1300 int i;
1301
1302 for (i = 0; i < MAX_BPF_REG; i++) {
1303 mark_reg_not_init(env, regs, i);
1304 regs[i].live = REG_LIVE_NONE;
1305 regs[i].parent = NULL;
1306 regs[i].subreg_def = DEF_NOT_SUBREG;
1307 }
1308
1309 /* frame pointer */
1310 regs[BPF_REG_FP].type = PTR_TO_STACK;
1311 mark_reg_known_zero(env, regs, BPF_REG_FP);
1312 regs[BPF_REG_FP].frameno = state->frameno;
1313 }
1314
1315 #define BPF_MAIN_FUNC (-1)
1316 static void init_func_state(struct bpf_verifier_env *env,
1317 struct bpf_func_state *state,
1318 int callsite, int frameno, int subprogno)
1319 {
1320 state->callsite = callsite;
1321 state->frameno = frameno;
1322 state->subprogno = subprogno;
1323 init_reg_state(env, state);
1324 }
1325
1326 enum reg_arg_type {
1327 SRC_OP, /* register is used as source operand */
1328 DST_OP, /* register is used as destination operand */
1329 DST_OP_NO_MARK /* same as above, check only, don't mark */
1330 };
1331
1332 static int cmp_subprogs(const void *a, const void *b)
1333 {
1334 return ((struct bpf_subprog_info *)a)->start -
1335 ((struct bpf_subprog_info *)b)->start;
1336 }
1337
1338 static int find_subprog(struct bpf_verifier_env *env, int off)
1339 {
1340 struct bpf_subprog_info *p;
1341
1342 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1343 sizeof(env->subprog_info[0]), cmp_subprogs);
1344 if (!p)
1345 return -ENOENT;
1346 return p - env->subprog_info;
1347
1348 }
1349
1350 static int add_subprog(struct bpf_verifier_env *env, int off)
1351 {
1352 int insn_cnt = env->prog->len;
1353 int ret;
1354
1355 if (off >= insn_cnt || off < 0) {
1356 verbose(env, "call to invalid destination\n");
1357 return -EINVAL;
1358 }
1359 ret = find_subprog(env, off);
1360 if (ret >= 0)
1361 return 0;
1362 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1363 verbose(env, "too many subprograms\n");
1364 return -E2BIG;
1365 }
1366 env->subprog_info[env->subprog_cnt++].start = off;
1367 sort(env->subprog_info, env->subprog_cnt,
1368 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1369 return 0;
1370 }
1371
1372 static int check_subprogs(struct bpf_verifier_env *env)
1373 {
1374 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1375 struct bpf_subprog_info *subprog = env->subprog_info;
1376 struct bpf_insn *insn = env->prog->insnsi;
1377 int insn_cnt = env->prog->len;
1378
1379 /* Add entry function. */
1380 ret = add_subprog(env, 0);
1381 if (ret < 0)
1382 return ret;
1383
1384 /* determine subprog starts. The end is one before the next starts */
1385 for (i = 0; i < insn_cnt; i++) {
1386 if (insn[i].code != (BPF_JMP | BPF_CALL))
1387 continue;
1388 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1389 continue;
1390 if (!env->allow_ptr_leaks) {
1391 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1392 return -EPERM;
1393 }
1394 ret = add_subprog(env, i + insn[i].imm + 1);
1395 if (ret < 0)
1396 return ret;
1397 }
1398
1399 /* Add a fake 'exit' subprog which could simplify subprog iteration
1400 * logic. 'subprog_cnt' should not be increased.
1401 */
1402 subprog[env->subprog_cnt].start = insn_cnt;
1403
1404 if (env->log.level & BPF_LOG_LEVEL2)
1405 for (i = 0; i < env->subprog_cnt; i++)
1406 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1407
1408 /* now check that all jumps are within the same subprog */
1409 subprog_start = subprog[cur_subprog].start;
1410 subprog_end = subprog[cur_subprog + 1].start;
1411 for (i = 0; i < insn_cnt; i++) {
1412 u8 code = insn[i].code;
1413
1414 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1415 goto next;
1416 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1417 goto next;
1418 off = i + insn[i].off + 1;
1419 if (off < subprog_start || off >= subprog_end) {
1420 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1421 return -EINVAL;
1422 }
1423 next:
1424 if (i == subprog_end - 1) {
1425 /* to avoid fall-through from one subprog into another
1426 * the last insn of the subprog should be either exit
1427 * or unconditional jump back
1428 */
1429 if (code != (BPF_JMP | BPF_EXIT) &&
1430 code != (BPF_JMP | BPF_JA)) {
1431 verbose(env, "last insn is not an exit or jmp\n");
1432 return -EINVAL;
1433 }
1434 subprog_start = subprog_end;
1435 cur_subprog++;
1436 if (cur_subprog < env->subprog_cnt)
1437 subprog_end = subprog[cur_subprog + 1].start;
1438 }
1439 }
1440 return 0;
1441 }
1442
1443 /* Parentage chain of this register (or stack slot) should take care of all
1444 * issues like callee-saved registers, stack slot allocation time, etc.
1445 */
1446 static int mark_reg_read(struct bpf_verifier_env *env,
1447 const struct bpf_reg_state *state,
1448 struct bpf_reg_state *parent, u8 flag)
1449 {
1450 bool writes = parent == state->parent; /* Observe write marks */
1451 int cnt = 0;
1452
1453 while (parent) {
1454 /* if read wasn't screened by an earlier write ... */
1455 if (writes && state->live & REG_LIVE_WRITTEN)
1456 break;
1457 if (parent->live & REG_LIVE_DONE) {
1458 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1459 reg_type_str[parent->type],
1460 parent->var_off.value, parent->off);
1461 return -EFAULT;
1462 }
1463 /* The first condition is more likely to be true than the
1464 * second, checked it first.
1465 */
1466 if ((parent->live & REG_LIVE_READ) == flag ||
1467 parent->live & REG_LIVE_READ64)
1468 /* The parentage chain never changes and
1469 * this parent was already marked as LIVE_READ.
1470 * There is no need to keep walking the chain again and
1471 * keep re-marking all parents as LIVE_READ.
1472 * This case happens when the same register is read
1473 * multiple times without writes into it in-between.
1474 * Also, if parent has the stronger REG_LIVE_READ64 set,
1475 * then no need to set the weak REG_LIVE_READ32.
1476 */
1477 break;
1478 /* ... then we depend on parent's value */
1479 parent->live |= flag;
1480 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1481 if (flag == REG_LIVE_READ64)
1482 parent->live &= ~REG_LIVE_READ32;
1483 state = parent;
1484 parent = state->parent;
1485 writes = true;
1486 cnt++;
1487 }
1488
1489 if (env->longest_mark_read_walk < cnt)
1490 env->longest_mark_read_walk = cnt;
1491 return 0;
1492 }
1493
1494 /* This function is supposed to be used by the following 32-bit optimization
1495 * code only. It returns TRUE if the source or destination register operates
1496 * on 64-bit, otherwise return FALSE.
1497 */
1498 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1499 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1500 {
1501 u8 code, class, op;
1502
1503 code = insn->code;
1504 class = BPF_CLASS(code);
1505 op = BPF_OP(code);
1506 if (class == BPF_JMP) {
1507 /* BPF_EXIT for "main" will reach here. Return TRUE
1508 * conservatively.
1509 */
1510 if (op == BPF_EXIT)
1511 return true;
1512 if (op == BPF_CALL) {
1513 /* BPF to BPF call will reach here because of marking
1514 * caller saved clobber with DST_OP_NO_MARK for which we
1515 * don't care the register def because they are anyway
1516 * marked as NOT_INIT already.
1517 */
1518 if (insn->src_reg == BPF_PSEUDO_CALL)
1519 return false;
1520 /* Helper call will reach here because of arg type
1521 * check, conservatively return TRUE.
1522 */
1523 if (t == SRC_OP)
1524 return true;
1525
1526 return false;
1527 }
1528 }
1529
1530 if (class == BPF_ALU64 || class == BPF_JMP ||
1531 /* BPF_END always use BPF_ALU class. */
1532 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1533 return true;
1534
1535 if (class == BPF_ALU || class == BPF_JMP32)
1536 return false;
1537
1538 if (class == BPF_LDX) {
1539 if (t != SRC_OP)
1540 return BPF_SIZE(code) == BPF_DW;
1541 /* LDX source must be ptr. */
1542 return true;
1543 }
1544
1545 if (class == BPF_STX) {
1546 if (reg->type != SCALAR_VALUE)
1547 return true;
1548 return BPF_SIZE(code) == BPF_DW;
1549 }
1550
1551 if (class == BPF_LD) {
1552 u8 mode = BPF_MODE(code);
1553
1554 /* LD_IMM64 */
1555 if (mode == BPF_IMM)
1556 return true;
1557
1558 /* Both LD_IND and LD_ABS return 32-bit data. */
1559 if (t != SRC_OP)
1560 return false;
1561
1562 /* Implicit ctx ptr. */
1563 if (regno == BPF_REG_6)
1564 return true;
1565
1566 /* Explicit source could be any width. */
1567 return true;
1568 }
1569
1570 if (class == BPF_ST)
1571 /* The only source register for BPF_ST is a ptr. */
1572 return true;
1573
1574 /* Conservatively return true at default. */
1575 return true;
1576 }
1577
1578 /* Return TRUE if INSN doesn't have explicit value define. */
1579 static bool insn_no_def(struct bpf_insn *insn)
1580 {
1581 u8 class = BPF_CLASS(insn->code);
1582
1583 return (class == BPF_JMP || class == BPF_JMP32 ||
1584 class == BPF_STX || class == BPF_ST);
1585 }
1586
1587 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1588 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1589 {
1590 if (insn_no_def(insn))
1591 return false;
1592
1593 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1594 }
1595
1596 static void mark_insn_zext(struct bpf_verifier_env *env,
1597 struct bpf_reg_state *reg)
1598 {
1599 s32 def_idx = reg->subreg_def;
1600
1601 if (def_idx == DEF_NOT_SUBREG)
1602 return;
1603
1604 env->insn_aux_data[def_idx - 1].zext_dst = true;
1605 /* The dst will be zero extended, so won't be sub-register anymore. */
1606 reg->subreg_def = DEF_NOT_SUBREG;
1607 }
1608
1609 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1610 enum reg_arg_type t)
1611 {
1612 struct bpf_verifier_state *vstate = env->cur_state;
1613 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1614 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1615 struct bpf_reg_state *reg, *regs = state->regs;
1616 bool rw64;
1617
1618 if (regno >= MAX_BPF_REG) {
1619 verbose(env, "R%d is invalid\n", regno);
1620 return -EINVAL;
1621 }
1622
1623 reg = &regs[regno];
1624 rw64 = is_reg64(env, insn, regno, reg, t);
1625 if (t == SRC_OP) {
1626 /* check whether register used as source operand can be read */
1627 if (reg->type == NOT_INIT) {
1628 verbose(env, "R%d !read_ok\n", regno);
1629 return -EACCES;
1630 }
1631 /* We don't need to worry about FP liveness because it's read-only */
1632 if (regno == BPF_REG_FP)
1633 return 0;
1634
1635 if (rw64)
1636 mark_insn_zext(env, reg);
1637
1638 return mark_reg_read(env, reg, reg->parent,
1639 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1640 } else {
1641 /* check whether register used as dest operand can be written to */
1642 if (regno == BPF_REG_FP) {
1643 verbose(env, "frame pointer is read only\n");
1644 return -EACCES;
1645 }
1646 reg->live |= REG_LIVE_WRITTEN;
1647 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1648 if (t == DST_OP)
1649 mark_reg_unknown(env, regs, regno);
1650 }
1651 return 0;
1652 }
1653
1654 /* for any branch, call, exit record the history of jmps in the given state */
1655 static int push_jmp_history(struct bpf_verifier_env *env,
1656 struct bpf_verifier_state *cur)
1657 {
1658 u32 cnt = cur->jmp_history_cnt;
1659 struct bpf_idx_pair *p;
1660
1661 cnt++;
1662 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1663 if (!p)
1664 return -ENOMEM;
1665 p[cnt - 1].idx = env->insn_idx;
1666 p[cnt - 1].prev_idx = env->prev_insn_idx;
1667 cur->jmp_history = p;
1668 cur->jmp_history_cnt = cnt;
1669 return 0;
1670 }
1671
1672 /* Backtrack one insn at a time. If idx is not at the top of recorded
1673 * history then previous instruction came from straight line execution.
1674 */
1675 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1676 u32 *history)
1677 {
1678 u32 cnt = *history;
1679
1680 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1681 i = st->jmp_history[cnt - 1].prev_idx;
1682 (*history)--;
1683 } else {
1684 i--;
1685 }
1686 return i;
1687 }
1688
1689 /* For given verifier state backtrack_insn() is called from the last insn to
1690 * the first insn. Its purpose is to compute a bitmask of registers and
1691 * stack slots that needs precision in the parent verifier state.
1692 */
1693 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1694 u32 *reg_mask, u64 *stack_mask)
1695 {
1696 const struct bpf_insn_cbs cbs = {
1697 .cb_print = verbose,
1698 .private_data = env,
1699 };
1700 struct bpf_insn *insn = env->prog->insnsi + idx;
1701 u8 class = BPF_CLASS(insn->code);
1702 u8 opcode = BPF_OP(insn->code);
1703 u8 mode = BPF_MODE(insn->code);
1704 u32 dreg = 1u << insn->dst_reg;
1705 u32 sreg = 1u << insn->src_reg;
1706 u32 spi;
1707
1708 if (insn->code == 0)
1709 return 0;
1710 if (env->log.level & BPF_LOG_LEVEL) {
1711 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1712 verbose(env, "%d: ", idx);
1713 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1714 }
1715
1716 if (class == BPF_ALU || class == BPF_ALU64) {
1717 if (!(*reg_mask & dreg))
1718 return 0;
1719 if (opcode == BPF_MOV) {
1720 if (BPF_SRC(insn->code) == BPF_X) {
1721 /* dreg = sreg
1722 * dreg needs precision after this insn
1723 * sreg needs precision before this insn
1724 */
1725 *reg_mask &= ~dreg;
1726 *reg_mask |= sreg;
1727 } else {
1728 /* dreg = K
1729 * dreg needs precision after this insn.
1730 * Corresponding register is already marked
1731 * as precise=true in this verifier state.
1732 * No further markings in parent are necessary
1733 */
1734 *reg_mask &= ~dreg;
1735 }
1736 } else {
1737 if (BPF_SRC(insn->code) == BPF_X) {
1738 /* dreg += sreg
1739 * both dreg and sreg need precision
1740 * before this insn
1741 */
1742 *reg_mask |= sreg;
1743 } /* else dreg += K
1744 * dreg still needs precision before this insn
1745 */
1746 }
1747 } else if (class == BPF_LDX) {
1748 if (!(*reg_mask & dreg))
1749 return 0;
1750 *reg_mask &= ~dreg;
1751
1752 /* scalars can only be spilled into stack w/o losing precision.
1753 * Load from any other memory can be zero extended.
1754 * The desire to keep that precision is already indicated
1755 * by 'precise' mark in corresponding register of this state.
1756 * No further tracking necessary.
1757 */
1758 if (insn->src_reg != BPF_REG_FP)
1759 return 0;
1760 if (BPF_SIZE(insn->code) != BPF_DW)
1761 return 0;
1762
1763 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1764 * that [fp - off] slot contains scalar that needs to be
1765 * tracked with precision
1766 */
1767 spi = (-insn->off - 1) / BPF_REG_SIZE;
1768 if (spi >= 64) {
1769 verbose(env, "BUG spi %d\n", spi);
1770 WARN_ONCE(1, "verifier backtracking bug");
1771 return -EFAULT;
1772 }
1773 *stack_mask |= 1ull << spi;
1774 } else if (class == BPF_STX || class == BPF_ST) {
1775 if (*reg_mask & dreg)
1776 /* stx & st shouldn't be using _scalar_ dst_reg
1777 * to access memory. It means backtracking
1778 * encountered a case of pointer subtraction.
1779 */
1780 return -ENOTSUPP;
1781 /* scalars can only be spilled into stack */
1782 if (insn->dst_reg != BPF_REG_FP)
1783 return 0;
1784 if (BPF_SIZE(insn->code) != BPF_DW)
1785 return 0;
1786 spi = (-insn->off - 1) / BPF_REG_SIZE;
1787 if (spi >= 64) {
1788 verbose(env, "BUG spi %d\n", spi);
1789 WARN_ONCE(1, "verifier backtracking bug");
1790 return -EFAULT;
1791 }
1792 if (!(*stack_mask & (1ull << spi)))
1793 return 0;
1794 *stack_mask &= ~(1ull << spi);
1795 if (class == BPF_STX)
1796 *reg_mask |= sreg;
1797 } else if (class == BPF_JMP || class == BPF_JMP32) {
1798 if (opcode == BPF_CALL) {
1799 if (insn->src_reg == BPF_PSEUDO_CALL)
1800 return -ENOTSUPP;
1801 /* regular helper call sets R0 */
1802 *reg_mask &= ~1;
1803 if (*reg_mask & 0x3f) {
1804 /* if backtracing was looking for registers R1-R5
1805 * they should have been found already.
1806 */
1807 verbose(env, "BUG regs %x\n", *reg_mask);
1808 WARN_ONCE(1, "verifier backtracking bug");
1809 return -EFAULT;
1810 }
1811 } else if (opcode == BPF_EXIT) {
1812 return -ENOTSUPP;
1813 }
1814 } else if (class == BPF_LD) {
1815 if (!(*reg_mask & dreg))
1816 return 0;
1817 *reg_mask &= ~dreg;
1818 /* It's ld_imm64 or ld_abs or ld_ind.
1819 * For ld_imm64 no further tracking of precision
1820 * into parent is necessary
1821 */
1822 if (mode == BPF_IND || mode == BPF_ABS)
1823 /* to be analyzed */
1824 return -ENOTSUPP;
1825 }
1826 return 0;
1827 }
1828
1829 /* the scalar precision tracking algorithm:
1830 * . at the start all registers have precise=false.
1831 * . scalar ranges are tracked as normal through alu and jmp insns.
1832 * . once precise value of the scalar register is used in:
1833 * . ptr + scalar alu
1834 * . if (scalar cond K|scalar)
1835 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1836 * backtrack through the verifier states and mark all registers and
1837 * stack slots with spilled constants that these scalar regisers
1838 * should be precise.
1839 * . during state pruning two registers (or spilled stack slots)
1840 * are equivalent if both are not precise.
1841 *
1842 * Note the verifier cannot simply walk register parentage chain,
1843 * since many different registers and stack slots could have been
1844 * used to compute single precise scalar.
1845 *
1846 * The approach of starting with precise=true for all registers and then
1847 * backtrack to mark a register as not precise when the verifier detects
1848 * that program doesn't care about specific value (e.g., when helper
1849 * takes register as ARG_ANYTHING parameter) is not safe.
1850 *
1851 * It's ok to walk single parentage chain of the verifier states.
1852 * It's possible that this backtracking will go all the way till 1st insn.
1853 * All other branches will be explored for needing precision later.
1854 *
1855 * The backtracking needs to deal with cases like:
1856 * R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0)
1857 * r9 -= r8
1858 * r5 = r9
1859 * if r5 > 0x79f goto pc+7
1860 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1861 * r5 += 1
1862 * ...
1863 * call bpf_perf_event_output#25
1864 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1865 *
1866 * and this case:
1867 * r6 = 1
1868 * call foo // uses callee's r6 inside to compute r0
1869 * r0 += r6
1870 * if r0 == 0 goto
1871 *
1872 * to track above reg_mask/stack_mask needs to be independent for each frame.
1873 *
1874 * Also if parent's curframe > frame where backtracking started,
1875 * the verifier need to mark registers in both frames, otherwise callees
1876 * may incorrectly prune callers. This is similar to
1877 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1878 *
1879 * For now backtracking falls back into conservative marking.
1880 */
1881 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1882 struct bpf_verifier_state *st)
1883 {
1884 struct bpf_func_state *func;
1885 struct bpf_reg_state *reg;
1886 int i, j;
1887
1888 /* big hammer: mark all scalars precise in this path.
1889 * pop_stack may still get !precise scalars.
1890 */
1891 for (; st; st = st->parent)
1892 for (i = 0; i <= st->curframe; i++) {
1893 func = st->frame[i];
1894 for (j = 0; j < BPF_REG_FP; j++) {
1895 reg = &func->regs[j];
1896 if (reg->type != SCALAR_VALUE)
1897 continue;
1898 reg->precise = true;
1899 }
1900 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1901 if (func->stack[j].slot_type[0] != STACK_SPILL)
1902 continue;
1903 reg = &func->stack[j].spilled_ptr;
1904 if (reg->type != SCALAR_VALUE)
1905 continue;
1906 reg->precise = true;
1907 }
1908 }
1909 }
1910
1911 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1912 int spi)
1913 {
1914 struct bpf_verifier_state *st = env->cur_state;
1915 int first_idx = st->first_insn_idx;
1916 int last_idx = env->insn_idx;
1917 struct bpf_func_state *func;
1918 struct bpf_reg_state *reg;
1919 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1920 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1921 bool skip_first = true;
1922 bool new_marks = false;
1923 int i, err;
1924
1925 if (!env->allow_ptr_leaks)
1926 /* backtracking is root only for now */
1927 return 0;
1928
1929 func = st->frame[st->curframe];
1930 if (regno >= 0) {
1931 reg = &func->regs[regno];
1932 if (reg->type != SCALAR_VALUE) {
1933 WARN_ONCE(1, "backtracing misuse");
1934 return -EFAULT;
1935 }
1936 if (!reg->precise)
1937 new_marks = true;
1938 else
1939 reg_mask = 0;
1940 reg->precise = true;
1941 }
1942
1943 while (spi >= 0) {
1944 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
1945 stack_mask = 0;
1946 break;
1947 }
1948 reg = &func->stack[spi].spilled_ptr;
1949 if (reg->type != SCALAR_VALUE) {
1950 stack_mask = 0;
1951 break;
1952 }
1953 if (!reg->precise)
1954 new_marks = true;
1955 else
1956 stack_mask = 0;
1957 reg->precise = true;
1958 break;
1959 }
1960
1961 if (!new_marks)
1962 return 0;
1963 if (!reg_mask && !stack_mask)
1964 return 0;
1965 for (;;) {
1966 DECLARE_BITMAP(mask, 64);
1967 u32 history = st->jmp_history_cnt;
1968
1969 if (env->log.level & BPF_LOG_LEVEL)
1970 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
1971 for (i = last_idx;;) {
1972 if (skip_first) {
1973 err = 0;
1974 skip_first = false;
1975 } else {
1976 err = backtrack_insn(env, i, &reg_mask, &stack_mask);
1977 }
1978 if (err == -ENOTSUPP) {
1979 mark_all_scalars_precise(env, st);
1980 return 0;
1981 } else if (err) {
1982 return err;
1983 }
1984 if (!reg_mask && !stack_mask)
1985 /* Found assignment(s) into tracked register in this state.
1986 * Since this state is already marked, just return.
1987 * Nothing to be tracked further in the parent state.
1988 */
1989 return 0;
1990 if (i == first_idx)
1991 break;
1992 i = get_prev_insn_idx(st, i, &history);
1993 if (i >= env->prog->len) {
1994 /* This can happen if backtracking reached insn 0
1995 * and there are still reg_mask or stack_mask
1996 * to backtrack.
1997 * It means the backtracking missed the spot where
1998 * particular register was initialized with a constant.
1999 */
2000 verbose(env, "BUG backtracking idx %d\n", i);
2001 WARN_ONCE(1, "verifier backtracking bug");
2002 return -EFAULT;
2003 }
2004 }
2005 st = st->parent;
2006 if (!st)
2007 break;
2008
2009 new_marks = false;
2010 func = st->frame[st->curframe];
2011 bitmap_from_u64(mask, reg_mask);
2012 for_each_set_bit(i, mask, 32) {
2013 reg = &func->regs[i];
2014 if (reg->type != SCALAR_VALUE) {
2015 reg_mask &= ~(1u << i);
2016 continue;
2017 }
2018 if (!reg->precise)
2019 new_marks = true;
2020 reg->precise = true;
2021 }
2022
2023 bitmap_from_u64(mask, stack_mask);
2024 for_each_set_bit(i, mask, 64) {
2025 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2026 /* the sequence of instructions:
2027 * 2: (bf) r3 = r10
2028 * 3: (7b) *(u64 *)(r3 -8) = r0
2029 * 4: (79) r4 = *(u64 *)(r10 -8)
2030 * doesn't contain jmps. It's backtracked
2031 * as a single block.
2032 * During backtracking insn 3 is not recognized as
2033 * stack access, so at the end of backtracking
2034 * stack slot fp-8 is still marked in stack_mask.
2035 * However the parent state may not have accessed
2036 * fp-8 and it's "unallocated" stack space.
2037 * In such case fallback to conservative.
2038 */
2039 mark_all_scalars_precise(env, st);
2040 return 0;
2041 }
2042
2043 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2044 stack_mask &= ~(1ull << i);
2045 continue;
2046 }
2047 reg = &func->stack[i].spilled_ptr;
2048 if (reg->type != SCALAR_VALUE) {
2049 stack_mask &= ~(1ull << i);
2050 continue;
2051 }
2052 if (!reg->precise)
2053 new_marks = true;
2054 reg->precise = true;
2055 }
2056 if (env->log.level & BPF_LOG_LEVEL) {
2057 print_verifier_state(env, func);
2058 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2059 new_marks ? "didn't have" : "already had",
2060 reg_mask, stack_mask);
2061 }
2062
2063 if (!reg_mask && !stack_mask)
2064 break;
2065 if (!new_marks)
2066 break;
2067
2068 last_idx = st->last_insn_idx;
2069 first_idx = st->first_insn_idx;
2070 }
2071 return 0;
2072 }
2073
2074 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2075 {
2076 return __mark_chain_precision(env, regno, -1);
2077 }
2078
2079 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2080 {
2081 return __mark_chain_precision(env, -1, spi);
2082 }
2083
2084 static bool is_spillable_regtype(enum bpf_reg_type type)
2085 {
2086 switch (type) {
2087 case PTR_TO_MAP_VALUE:
2088 case PTR_TO_MAP_VALUE_OR_NULL:
2089 case PTR_TO_STACK:
2090 case PTR_TO_CTX:
2091 case PTR_TO_PACKET:
2092 case PTR_TO_PACKET_META:
2093 case PTR_TO_PACKET_END:
2094 case PTR_TO_FLOW_KEYS:
2095 case CONST_PTR_TO_MAP:
2096 case PTR_TO_SOCKET:
2097 case PTR_TO_SOCKET_OR_NULL:
2098 case PTR_TO_SOCK_COMMON:
2099 case PTR_TO_SOCK_COMMON_OR_NULL:
2100 case PTR_TO_TCP_SOCK:
2101 case PTR_TO_TCP_SOCK_OR_NULL:
2102 case PTR_TO_XDP_SOCK:
2103 case PTR_TO_BTF_ID:
2104 return true;
2105 default:
2106 return false;
2107 }
2108 }
2109
2110 /* Does this register contain a constant zero? */
2111 static bool register_is_null(struct bpf_reg_state *reg)
2112 {
2113 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2114 }
2115
2116 static bool register_is_const(struct bpf_reg_state *reg)
2117 {
2118 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2119 }
2120
2121 static bool __is_pointer_value(bool allow_ptr_leaks,
2122 const struct bpf_reg_state *reg)
2123 {
2124 if (allow_ptr_leaks)
2125 return false;
2126
2127 return reg->type != SCALAR_VALUE;
2128 }
2129
2130 static void save_register_state(struct bpf_func_state *state,
2131 int spi, struct bpf_reg_state *reg)
2132 {
2133 int i;
2134
2135 state->stack[spi].spilled_ptr = *reg;
2136 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2137
2138 for (i = 0; i < BPF_REG_SIZE; i++)
2139 state->stack[spi].slot_type[i] = STACK_SPILL;
2140 }
2141
2142 /* check_stack_read/write functions track spill/fill of registers,
2143 * stack boundary and alignment are checked in check_mem_access()
2144 */
2145 static int check_stack_write(struct bpf_verifier_env *env,
2146 struct bpf_func_state *state, /* func where register points to */
2147 int off, int size, int value_regno, int insn_idx)
2148 {
2149 struct bpf_func_state *cur; /* state of the current function */
2150 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2151 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2152 struct bpf_reg_state *reg = NULL;
2153
2154 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2155 state->acquired_refs, true);
2156 if (err)
2157 return err;
2158 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2159 * so it's aligned access and [off, off + size) are within stack limits
2160 */
2161 if (!env->allow_ptr_leaks &&
2162 state->stack[spi].slot_type[0] == STACK_SPILL &&
2163 size != BPF_REG_SIZE) {
2164 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2165 return -EACCES;
2166 }
2167
2168 cur = env->cur_state->frame[env->cur_state->curframe];
2169 if (value_regno >= 0)
2170 reg = &cur->regs[value_regno];
2171
2172 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
2173 !register_is_null(reg) && env->allow_ptr_leaks) {
2174 if (dst_reg != BPF_REG_FP) {
2175 /* The backtracking logic can only recognize explicit
2176 * stack slot address like [fp - 8]. Other spill of
2177 * scalar via different register has to be conervative.
2178 * Backtrack from here and mark all registers as precise
2179 * that contributed into 'reg' being a constant.
2180 */
2181 err = mark_chain_precision(env, value_regno);
2182 if (err)
2183 return err;
2184 }
2185 save_register_state(state, spi, reg);
2186 } else if (reg && is_spillable_regtype(reg->type)) {
2187 /* register containing pointer is being spilled into stack */
2188 if (size != BPF_REG_SIZE) {
2189 verbose_linfo(env, insn_idx, "; ");
2190 verbose(env, "invalid size of register spill\n");
2191 return -EACCES;
2192 }
2193
2194 if (state != cur && reg->type == PTR_TO_STACK) {
2195 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2196 return -EINVAL;
2197 }
2198
2199 if (!env->allow_ptr_leaks) {
2200 bool sanitize = false;
2201
2202 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2203 register_is_const(&state->stack[spi].spilled_ptr))
2204 sanitize = true;
2205 for (i = 0; i < BPF_REG_SIZE; i++)
2206 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2207 sanitize = true;
2208 break;
2209 }
2210 if (sanitize) {
2211 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2212 int soff = (-spi - 1) * BPF_REG_SIZE;
2213
2214 /* detected reuse of integer stack slot with a pointer
2215 * which means either llvm is reusing stack slot or
2216 * an attacker is trying to exploit CVE-2018-3639
2217 * (speculative store bypass)
2218 * Have to sanitize that slot with preemptive
2219 * store of zero.
2220 */
2221 if (*poff && *poff != soff) {
2222 /* disallow programs where single insn stores
2223 * into two different stack slots, since verifier
2224 * cannot sanitize them
2225 */
2226 verbose(env,
2227 "insn %d cannot access two stack slots fp%d and fp%d",
2228 insn_idx, *poff, soff);
2229 return -EINVAL;
2230 }
2231 *poff = soff;
2232 }
2233 }
2234 save_register_state(state, spi, reg);
2235 } else {
2236 u8 type = STACK_MISC;
2237
2238 /* regular write of data into stack destroys any spilled ptr */
2239 state->stack[spi].spilled_ptr.type = NOT_INIT;
2240 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2241 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2242 for (i = 0; i < BPF_REG_SIZE; i++)
2243 state->stack[spi].slot_type[i] = STACK_MISC;
2244
2245 /* only mark the slot as written if all 8 bytes were written
2246 * otherwise read propagation may incorrectly stop too soon
2247 * when stack slots are partially written.
2248 * This heuristic means that read propagation will be
2249 * conservative, since it will add reg_live_read marks
2250 * to stack slots all the way to first state when programs
2251 * writes+reads less than 8 bytes
2252 */
2253 if (size == BPF_REG_SIZE)
2254 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2255
2256 /* when we zero initialize stack slots mark them as such */
2257 if (reg && register_is_null(reg)) {
2258 /* backtracking doesn't work for STACK_ZERO yet. */
2259 err = mark_chain_precision(env, value_regno);
2260 if (err)
2261 return err;
2262 type = STACK_ZERO;
2263 }
2264
2265 /* Mark slots affected by this stack write. */
2266 for (i = 0; i < size; i++)
2267 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2268 type;
2269 }
2270 return 0;
2271 }
2272
2273 static int check_stack_read(struct bpf_verifier_env *env,
2274 struct bpf_func_state *reg_state /* func where register points to */,
2275 int off, int size, int value_regno)
2276 {
2277 struct bpf_verifier_state *vstate = env->cur_state;
2278 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2279 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2280 struct bpf_reg_state *reg;
2281 u8 *stype;
2282
2283 if (reg_state->allocated_stack <= slot) {
2284 verbose(env, "invalid read from stack off %d+0 size %d\n",
2285 off, size);
2286 return -EACCES;
2287 }
2288 stype = reg_state->stack[spi].slot_type;
2289 reg = &reg_state->stack[spi].spilled_ptr;
2290
2291 if (stype[0] == STACK_SPILL) {
2292 if (size != BPF_REG_SIZE) {
2293 if (reg->type != SCALAR_VALUE) {
2294 verbose_linfo(env, env->insn_idx, "; ");
2295 verbose(env, "invalid size of register fill\n");
2296 return -EACCES;
2297 }
2298 if (value_regno >= 0) {
2299 mark_reg_unknown(env, state->regs, value_regno);
2300 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2301 }
2302 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2303 return 0;
2304 }
2305 for (i = 1; i < BPF_REG_SIZE; i++) {
2306 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2307 verbose(env, "corrupted spill memory\n");
2308 return -EACCES;
2309 }
2310 }
2311
2312 if (value_regno >= 0) {
2313 /* restore register state from stack */
2314 state->regs[value_regno] = *reg;
2315 /* mark reg as written since spilled pointer state likely
2316 * has its liveness marks cleared by is_state_visited()
2317 * which resets stack/reg liveness for state transitions
2318 */
2319 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2320 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2321 /* If value_regno==-1, the caller is asking us whether
2322 * it is acceptable to use this value as a SCALAR_VALUE
2323 * (e.g. for XADD).
2324 * We must not allow unprivileged callers to do that
2325 * with spilled pointers.
2326 */
2327 verbose(env, "leaking pointer from stack off %d\n",
2328 off);
2329 return -EACCES;
2330 }
2331 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2332 } else {
2333 int zeros = 0;
2334
2335 for (i = 0; i < size; i++) {
2336 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2337 continue;
2338 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2339 zeros++;
2340 continue;
2341 }
2342 verbose(env, "invalid read from stack off %d+%d size %d\n",
2343 off, i, size);
2344 return -EACCES;
2345 }
2346 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2347 if (value_regno >= 0) {
2348 if (zeros == size) {
2349 /* any size read into register is zero extended,
2350 * so the whole register == const_zero
2351 */
2352 __mark_reg_const_zero(&state->regs[value_regno]);
2353 /* backtracking doesn't support STACK_ZERO yet,
2354 * so mark it precise here, so that later
2355 * backtracking can stop here.
2356 * Backtracking may not need this if this register
2357 * doesn't participate in pointer adjustment.
2358 * Forward propagation of precise flag is not
2359 * necessary either. This mark is only to stop
2360 * backtracking. Any register that contributed
2361 * to const 0 was marked precise before spill.
2362 */
2363 state->regs[value_regno].precise = true;
2364 } else {
2365 /* have read misc data from the stack */
2366 mark_reg_unknown(env, state->regs, value_regno);
2367 }
2368 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2369 }
2370 }
2371 return 0;
2372 }
2373
2374 static int check_stack_access(struct bpf_verifier_env *env,
2375 const struct bpf_reg_state *reg,
2376 int off, int size)
2377 {
2378 /* Stack accesses must be at a fixed offset, so that we
2379 * can determine what type of data were returned. See
2380 * check_stack_read().
2381 */
2382 if (!tnum_is_const(reg->var_off)) {
2383 char tn_buf[48];
2384
2385 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2386 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2387 tn_buf, off, size);
2388 return -EACCES;
2389 }
2390
2391 if (off >= 0 || off < -MAX_BPF_STACK) {
2392 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2393 return -EACCES;
2394 }
2395
2396 return 0;
2397 }
2398
2399 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2400 int off, int size, enum bpf_access_type type)
2401 {
2402 struct bpf_reg_state *regs = cur_regs(env);
2403 struct bpf_map *map = regs[regno].map_ptr;
2404 u32 cap = bpf_map_flags_to_cap(map);
2405
2406 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2407 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2408 map->value_size, off, size);
2409 return -EACCES;
2410 }
2411
2412 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2413 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2414 map->value_size, off, size);
2415 return -EACCES;
2416 }
2417
2418 return 0;
2419 }
2420
2421 /* check read/write into map element returned by bpf_map_lookup_elem() */
2422 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
2423 int size, bool zero_size_allowed)
2424 {
2425 struct bpf_reg_state *regs = cur_regs(env);
2426 struct bpf_map *map = regs[regno].map_ptr;
2427
2428 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2429 off + size > map->value_size) {
2430 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2431 map->value_size, off, size);
2432 return -EACCES;
2433 }
2434 return 0;
2435 }
2436
2437 /* check read/write into a map element with possible variable offset */
2438 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2439 int off, int size, bool zero_size_allowed)
2440 {
2441 struct bpf_verifier_state *vstate = env->cur_state;
2442 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2443 struct bpf_reg_state *reg = &state->regs[regno];
2444 int err;
2445
2446 /* We may have adjusted the register to this map value, so we
2447 * need to try adding each of min_value and max_value to off
2448 * to make sure our theoretical access will be safe.
2449 */
2450 if (env->log.level & BPF_LOG_LEVEL)
2451 print_verifier_state(env, state);
2452
2453 /* The minimum value is only important with signed
2454 * comparisons where we can't assume the floor of a
2455 * value is 0. If we are using signed variables for our
2456 * index'es we need to make sure that whatever we use
2457 * will have a set floor within our range.
2458 */
2459 if (reg->smin_value < 0 &&
2460 (reg->smin_value == S64_MIN ||
2461 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2462 reg->smin_value + off < 0)) {
2463 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2464 regno);
2465 return -EACCES;
2466 }
2467 err = __check_map_access(env, regno, reg->smin_value + off, size,
2468 zero_size_allowed);
2469 if (err) {
2470 verbose(env, "R%d min value is outside of the array range\n",
2471 regno);
2472 return err;
2473 }
2474
2475 /* If we haven't set a max value then we need to bail since we can't be
2476 * sure we won't do bad things.
2477 * If reg->umax_value + off could overflow, treat that as unbounded too.
2478 */
2479 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2480 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
2481 regno);
2482 return -EACCES;
2483 }
2484 err = __check_map_access(env, regno, reg->umax_value + off, size,
2485 zero_size_allowed);
2486 if (err)
2487 verbose(env, "R%d max value is outside of the array range\n",
2488 regno);
2489
2490 if (map_value_has_spin_lock(reg->map_ptr)) {
2491 u32 lock = reg->map_ptr->spin_lock_off;
2492
2493 /* if any part of struct bpf_spin_lock can be touched by
2494 * load/store reject this program.
2495 * To check that [x1, x2) overlaps with [y1, y2)
2496 * it is sufficient to check x1 < y2 && y1 < x2.
2497 */
2498 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2499 lock < reg->umax_value + off + size) {
2500 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2501 return -EACCES;
2502 }
2503 }
2504 return err;
2505 }
2506
2507 #define MAX_PACKET_OFF 0xffff
2508
2509 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2510 const struct bpf_call_arg_meta *meta,
2511 enum bpf_access_type t)
2512 {
2513 switch (env->prog->type) {
2514 /* Program types only with direct read access go here! */
2515 case BPF_PROG_TYPE_LWT_IN:
2516 case BPF_PROG_TYPE_LWT_OUT:
2517 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2518 case BPF_PROG_TYPE_SK_REUSEPORT:
2519 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2520 case BPF_PROG_TYPE_CGROUP_SKB:
2521 if (t == BPF_WRITE)
2522 return false;
2523 /* fallthrough */
2524
2525 /* Program types with direct read + write access go here! */
2526 case BPF_PROG_TYPE_SCHED_CLS:
2527 case BPF_PROG_TYPE_SCHED_ACT:
2528 case BPF_PROG_TYPE_XDP:
2529 case BPF_PROG_TYPE_LWT_XMIT:
2530 case BPF_PROG_TYPE_SK_SKB:
2531 case BPF_PROG_TYPE_SK_MSG:
2532 if (meta)
2533 return meta->pkt_access;
2534
2535 env->seen_direct_write = true;
2536 return true;
2537
2538 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2539 if (t == BPF_WRITE)
2540 env->seen_direct_write = true;
2541
2542 return true;
2543
2544 default:
2545 return false;
2546 }
2547 }
2548
2549 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
2550 int off, int size, bool zero_size_allowed)
2551 {
2552 struct bpf_reg_state *regs = cur_regs(env);
2553 struct bpf_reg_state *reg = &regs[regno];
2554
2555 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2556 (u64)off + size > reg->range) {
2557 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2558 off, size, regno, reg->id, reg->off, reg->range);
2559 return -EACCES;
2560 }
2561 return 0;
2562 }
2563
2564 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2565 int size, bool zero_size_allowed)
2566 {
2567 struct bpf_reg_state *regs = cur_regs(env);
2568 struct bpf_reg_state *reg = &regs[regno];
2569 int err;
2570
2571 /* We may have added a variable offset to the packet pointer; but any
2572 * reg->range we have comes after that. We are only checking the fixed
2573 * offset.
2574 */
2575
2576 /* We don't allow negative numbers, because we aren't tracking enough
2577 * detail to prove they're safe.
2578 */
2579 if (reg->smin_value < 0) {
2580 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2581 regno);
2582 return -EACCES;
2583 }
2584 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
2585 if (err) {
2586 verbose(env, "R%d offset is outside of the packet\n", regno);
2587 return err;
2588 }
2589
2590 /* __check_packet_access has made sure "off + size - 1" is within u16.
2591 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2592 * otherwise find_good_pkt_pointers would have refused to set range info
2593 * that __check_packet_access would have rejected this pkt access.
2594 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2595 */
2596 env->prog->aux->max_pkt_offset =
2597 max_t(u32, env->prog->aux->max_pkt_offset,
2598 off + reg->umax_value + size - 1);
2599
2600 return err;
2601 }
2602
2603 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2604 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2605 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2606 u32 *btf_id)
2607 {
2608 struct bpf_insn_access_aux info = {
2609 .reg_type = *reg_type,
2610 .log = &env->log,
2611 };
2612
2613 if (env->ops->is_valid_access &&
2614 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2615 /* A non zero info.ctx_field_size indicates that this field is a
2616 * candidate for later verifier transformation to load the whole
2617 * field and then apply a mask when accessed with a narrower
2618 * access than actual ctx access size. A zero info.ctx_field_size
2619 * will only allow for whole field access and rejects any other
2620 * type of narrower access.
2621 */
2622 *reg_type = info.reg_type;
2623
2624 if (*reg_type == PTR_TO_BTF_ID)
2625 *btf_id = info.btf_id;
2626 else
2627 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2628 /* remember the offset of last byte accessed in ctx */
2629 if (env->prog->aux->max_ctx_offset < off + size)
2630 env->prog->aux->max_ctx_offset = off + size;
2631 return 0;
2632 }
2633
2634 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2635 return -EACCES;
2636 }
2637
2638 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2639 int size)
2640 {
2641 if (size < 0 || off < 0 ||
2642 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2643 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2644 off, size);
2645 return -EACCES;
2646 }
2647 return 0;
2648 }
2649
2650 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2651 u32 regno, int off, int size,
2652 enum bpf_access_type t)
2653 {
2654 struct bpf_reg_state *regs = cur_regs(env);
2655 struct bpf_reg_state *reg = &regs[regno];
2656 struct bpf_insn_access_aux info = {};
2657 bool valid;
2658
2659 if (reg->smin_value < 0) {
2660 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2661 regno);
2662 return -EACCES;
2663 }
2664
2665 switch (reg->type) {
2666 case PTR_TO_SOCK_COMMON:
2667 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2668 break;
2669 case PTR_TO_SOCKET:
2670 valid = bpf_sock_is_valid_access(off, size, t, &info);
2671 break;
2672 case PTR_TO_TCP_SOCK:
2673 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2674 break;
2675 case PTR_TO_XDP_SOCK:
2676 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2677 break;
2678 default:
2679 valid = false;
2680 }
2681
2682
2683 if (valid) {
2684 env->insn_aux_data[insn_idx].ctx_field_size =
2685 info.ctx_field_size;
2686 return 0;
2687 }
2688
2689 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2690 regno, reg_type_str[reg->type], off, size);
2691
2692 return -EACCES;
2693 }
2694
2695 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2696 {
2697 return cur_regs(env) + regno;
2698 }
2699
2700 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2701 {
2702 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2703 }
2704
2705 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2706 {
2707 const struct bpf_reg_state *reg = reg_state(env, regno);
2708
2709 return reg->type == PTR_TO_CTX;
2710 }
2711
2712 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2713 {
2714 const struct bpf_reg_state *reg = reg_state(env, regno);
2715
2716 return type_is_sk_pointer(reg->type);
2717 }
2718
2719 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2720 {
2721 const struct bpf_reg_state *reg = reg_state(env, regno);
2722
2723 return type_is_pkt_pointer(reg->type);
2724 }
2725
2726 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2727 {
2728 const struct bpf_reg_state *reg = reg_state(env, regno);
2729
2730 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2731 return reg->type == PTR_TO_FLOW_KEYS;
2732 }
2733
2734 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2735 const struct bpf_reg_state *reg,
2736 int off, int size, bool strict)
2737 {
2738 struct tnum reg_off;
2739 int ip_align;
2740
2741 /* Byte size accesses are always allowed. */
2742 if (!strict || size == 1)
2743 return 0;
2744
2745 /* For platforms that do not have a Kconfig enabling
2746 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2747 * NET_IP_ALIGN is universally set to '2'. And on platforms
2748 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2749 * to this code only in strict mode where we want to emulate
2750 * the NET_IP_ALIGN==2 checking. Therefore use an
2751 * unconditional IP align value of '2'.
2752 */
2753 ip_align = 2;
2754
2755 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2756 if (!tnum_is_aligned(reg_off, size)) {
2757 char tn_buf[48];
2758
2759 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2760 verbose(env,
2761 "misaligned packet access off %d+%s+%d+%d size %d\n",
2762 ip_align, tn_buf, reg->off, off, size);
2763 return -EACCES;
2764 }
2765
2766 return 0;
2767 }
2768
2769 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2770 const struct bpf_reg_state *reg,
2771 const char *pointer_desc,
2772 int off, int size, bool strict)
2773 {
2774 struct tnum reg_off;
2775
2776 /* Byte size accesses are always allowed. */
2777 if (!strict || size == 1)
2778 return 0;
2779
2780 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2781 if (!tnum_is_aligned(reg_off, size)) {
2782 char tn_buf[48];
2783
2784 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2785 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2786 pointer_desc, tn_buf, reg->off, off, size);
2787 return -EACCES;
2788 }
2789
2790 return 0;
2791 }
2792
2793 static int check_ptr_alignment(struct bpf_verifier_env *env,
2794 const struct bpf_reg_state *reg, int off,
2795 int size, bool strict_alignment_once)
2796 {
2797 bool strict = env->strict_alignment || strict_alignment_once;
2798 const char *pointer_desc = "";
2799
2800 switch (reg->type) {
2801 case PTR_TO_PACKET:
2802 case PTR_TO_PACKET_META:
2803 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2804 * right in front, treat it the very same way.
2805 */
2806 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2807 case PTR_TO_FLOW_KEYS:
2808 pointer_desc = "flow keys ";
2809 break;
2810 case PTR_TO_MAP_VALUE:
2811 pointer_desc = "value ";
2812 break;
2813 case PTR_TO_CTX:
2814 pointer_desc = "context ";
2815 break;
2816 case PTR_TO_STACK:
2817 pointer_desc = "stack ";
2818 /* The stack spill tracking logic in check_stack_write()
2819 * and check_stack_read() relies on stack accesses being
2820 * aligned.
2821 */
2822 strict = true;
2823 break;
2824 case PTR_TO_SOCKET:
2825 pointer_desc = "sock ";
2826 break;
2827 case PTR_TO_SOCK_COMMON:
2828 pointer_desc = "sock_common ";
2829 break;
2830 case PTR_TO_TCP_SOCK:
2831 pointer_desc = "tcp_sock ";
2832 break;
2833 case PTR_TO_XDP_SOCK:
2834 pointer_desc = "xdp_sock ";
2835 break;
2836 default:
2837 break;
2838 }
2839 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2840 strict);
2841 }
2842
2843 static int update_stack_depth(struct bpf_verifier_env *env,
2844 const struct bpf_func_state *func,
2845 int off)
2846 {
2847 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2848
2849 if (stack >= -off)
2850 return 0;
2851
2852 /* update known max for given subprogram */
2853 env->subprog_info[func->subprogno].stack_depth = -off;
2854 return 0;
2855 }
2856
2857 /* starting from main bpf function walk all instructions of the function
2858 * and recursively walk all callees that given function can call.
2859 * Ignore jump and exit insns.
2860 * Since recursion is prevented by check_cfg() this algorithm
2861 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2862 */
2863 static int check_max_stack_depth(struct bpf_verifier_env *env)
2864 {
2865 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2866 struct bpf_subprog_info *subprog = env->subprog_info;
2867 struct bpf_insn *insn = env->prog->insnsi;
2868 int ret_insn[MAX_CALL_FRAMES];
2869 int ret_prog[MAX_CALL_FRAMES];
2870
2871 process_func:
2872 /* round up to 32-bytes, since this is granularity
2873 * of interpreter stack size
2874 */
2875 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2876 if (depth > MAX_BPF_STACK) {
2877 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2878 frame + 1, depth);
2879 return -EACCES;
2880 }
2881 continue_func:
2882 subprog_end = subprog[idx + 1].start;
2883 for (; i < subprog_end; i++) {
2884 if (insn[i].code != (BPF_JMP | BPF_CALL))
2885 continue;
2886 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2887 continue;
2888 /* remember insn and function to return to */
2889 ret_insn[frame] = i + 1;
2890 ret_prog[frame] = idx;
2891
2892 /* find the callee */
2893 i = i + insn[i].imm + 1;
2894 idx = find_subprog(env, i);
2895 if (idx < 0) {
2896 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2897 i);
2898 return -EFAULT;
2899 }
2900 frame++;
2901 if (frame >= MAX_CALL_FRAMES) {
2902 verbose(env, "the call stack of %d frames is too deep !\n",
2903 frame);
2904 return -E2BIG;
2905 }
2906 goto process_func;
2907 }
2908 /* end of for() loop means the last insn of the 'subprog'
2909 * was reached. Doesn't matter whether it was JA or EXIT
2910 */
2911 if (frame == 0)
2912 return 0;
2913 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2914 frame--;
2915 i = ret_insn[frame];
2916 idx = ret_prog[frame];
2917 goto continue_func;
2918 }
2919
2920 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2921 static int get_callee_stack_depth(struct bpf_verifier_env *env,
2922 const struct bpf_insn *insn, int idx)
2923 {
2924 int start = idx + insn->imm + 1, subprog;
2925
2926 subprog = find_subprog(env, start);
2927 if (subprog < 0) {
2928 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2929 start);
2930 return -EFAULT;
2931 }
2932 return env->subprog_info[subprog].stack_depth;
2933 }
2934 #endif
2935
2936 int check_ctx_reg(struct bpf_verifier_env *env,
2937 const struct bpf_reg_state *reg, int regno)
2938 {
2939 /* Access to ctx or passing it to a helper is only allowed in
2940 * its original, unmodified form.
2941 */
2942
2943 if (reg->off) {
2944 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
2945 regno, reg->off);
2946 return -EACCES;
2947 }
2948
2949 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2950 char tn_buf[48];
2951
2952 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2953 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
2954 return -EACCES;
2955 }
2956
2957 return 0;
2958 }
2959
2960 static int check_tp_buffer_access(struct bpf_verifier_env *env,
2961 const struct bpf_reg_state *reg,
2962 int regno, int off, int size)
2963 {
2964 if (off < 0) {
2965 verbose(env,
2966 "R%d invalid tracepoint buffer access: off=%d, size=%d",
2967 regno, off, size);
2968 return -EACCES;
2969 }
2970 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2971 char tn_buf[48];
2972
2973 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2974 verbose(env,
2975 "R%d invalid variable buffer offset: off=%d, var_off=%s",
2976 regno, off, tn_buf);
2977 return -EACCES;
2978 }
2979 if (off + size > env->prog->aux->max_tp_access)
2980 env->prog->aux->max_tp_access = off + size;
2981
2982 return 0;
2983 }
2984
2985 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
2986 static void zext_32_to_64(struct bpf_reg_state *reg)
2987 {
2988 reg->var_off = tnum_subreg(reg->var_off);
2989 __reg_assign_32_into_64(reg);
2990 }
2991
2992 /* truncate register to smaller size (in bytes)
2993 * must be called with size < BPF_REG_SIZE
2994 */
2995 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
2996 {
2997 u64 mask;
2998
2999 /* clear high bits in bit representation */
3000 reg->var_off = tnum_cast(reg->var_off, size);
3001
3002 /* fix arithmetic bounds */
3003 mask = ((u64)1 << (size * 8)) - 1;
3004 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3005 reg->umin_value &= mask;
3006 reg->umax_value &= mask;
3007 } else {
3008 reg->umin_value = 0;
3009 reg->umax_value = mask;
3010 }
3011 reg->smin_value = reg->umin_value;
3012 reg->smax_value = reg->umax_value;
3013
3014 /* If size is smaller than 32bit register the 32bit register
3015 * values are also truncated so we push 64-bit bounds into
3016 * 32-bit bounds. Above were truncated < 32-bits already.
3017 */
3018 if (size >= 4)
3019 return;
3020 __reg_combine_64_into_32(reg);
3021 }
3022
3023 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3024 {
3025 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3026 }
3027
3028 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3029 {
3030 void *ptr;
3031 u64 addr;
3032 int err;
3033
3034 err = map->ops->map_direct_value_addr(map, &addr, off);
3035 if (err)
3036 return err;
3037 ptr = (void *)(long)addr + off;
3038
3039 switch (size) {
3040 case sizeof(u8):
3041 *val = (u64)*(u8 *)ptr;
3042 break;
3043 case sizeof(u16):
3044 *val = (u64)*(u16 *)ptr;
3045 break;
3046 case sizeof(u32):
3047 *val = (u64)*(u32 *)ptr;
3048 break;
3049 case sizeof(u64):
3050 *val = *(u64 *)ptr;
3051 break;
3052 default:
3053 return -EINVAL;
3054 }
3055 return 0;
3056 }
3057
3058 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3059 struct bpf_reg_state *regs,
3060 int regno, int off, int size,
3061 enum bpf_access_type atype,
3062 int value_regno)
3063 {
3064 struct bpf_reg_state *reg = regs + regno;
3065 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
3066 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3067 u32 btf_id;
3068 int ret;
3069
3070 if (off < 0) {
3071 verbose(env,
3072 "R%d is ptr_%s invalid negative access: off=%d\n",
3073 regno, tname, off);
3074 return -EACCES;
3075 }
3076 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3077 char tn_buf[48];
3078
3079 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3080 verbose(env,
3081 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3082 regno, tname, off, tn_buf);
3083 return -EACCES;
3084 }
3085
3086 if (env->ops->btf_struct_access) {
3087 ret = env->ops->btf_struct_access(&env->log, t, off, size,
3088 atype, &btf_id);
3089 } else {
3090 if (atype != BPF_READ) {
3091 verbose(env, "only read is supported\n");
3092 return -EACCES;
3093 }
3094
3095 ret = btf_struct_access(&env->log, t, off, size, atype,
3096 &btf_id);
3097 }
3098
3099 if (ret < 0)
3100 return ret;
3101
3102 if (atype == BPF_READ && value_regno >= 0) {
3103 if (ret == SCALAR_VALUE) {
3104 mark_reg_unknown(env, regs, value_regno);
3105 return 0;
3106 }
3107 mark_reg_known_zero(env, regs, value_regno);
3108 regs[value_regno].type = PTR_TO_BTF_ID;
3109 regs[value_regno].btf_id = btf_id;
3110 }
3111
3112 return 0;
3113 }
3114
3115 /* check whether memory at (regno + off) is accessible for t = (read | write)
3116 * if t==write, value_regno is a register which value is stored into memory
3117 * if t==read, value_regno is a register which will receive the value from memory
3118 * if t==write && value_regno==-1, some unknown value is stored into memory
3119 * if t==read && value_regno==-1, don't care what we read from memory
3120 */
3121 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
3122 int off, int bpf_size, enum bpf_access_type t,
3123 int value_regno, bool strict_alignment_once)
3124 {
3125 struct bpf_reg_state *regs = cur_regs(env);
3126 struct bpf_reg_state *reg = regs + regno;
3127 struct bpf_func_state *state;
3128 int size, err = 0;
3129
3130 size = bpf_size_to_bytes(bpf_size);
3131 if (size < 0)
3132 return size;
3133
3134 /* alignment checks will add in reg->off themselves */
3135 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
3136 if (err)
3137 return err;
3138
3139 /* for access checks, reg->off is just part of off */
3140 off += reg->off;
3141
3142 if (reg->type == PTR_TO_MAP_VALUE) {
3143 if (t == BPF_WRITE && value_regno >= 0 &&
3144 is_pointer_value(env, value_regno)) {
3145 verbose(env, "R%d leaks addr into map\n", value_regno);
3146 return -EACCES;
3147 }
3148 err = check_map_access_type(env, regno, off, size, t);
3149 if (err)
3150 return err;
3151 err = check_map_access(env, regno, off, size, false);
3152 if (!err && t == BPF_READ && value_regno >= 0) {
3153 struct bpf_map *map = reg->map_ptr;
3154
3155 /* if map is read-only, track its contents as scalars */
3156 if (tnum_is_const(reg->var_off) &&
3157 bpf_map_is_rdonly(map) &&
3158 map->ops->map_direct_value_addr) {
3159 int map_off = off + reg->var_off.value;
3160 u64 val = 0;
3161
3162 err = bpf_map_direct_read(map, map_off, size,
3163 &val);
3164 if (err)
3165 return err;
3166
3167 regs[value_regno].type = SCALAR_VALUE;
3168 __mark_reg_known(&regs[value_regno], val);
3169 } else {
3170 mark_reg_unknown(env, regs, value_regno);
3171 }
3172 }
3173 } else if (reg->type == PTR_TO_CTX) {
3174 enum bpf_reg_type reg_type = SCALAR_VALUE;
3175 u32 btf_id = 0;
3176
3177 if (t == BPF_WRITE && value_regno >= 0 &&
3178 is_pointer_value(env, value_regno)) {
3179 verbose(env, "R%d leaks addr into ctx\n", value_regno);
3180 return -EACCES;
3181 }
3182
3183 err = check_ctx_reg(env, reg, regno);
3184 if (err < 0)
3185 return err;
3186
3187 err = check_ctx_access(env, insn_idx, off, size, t, &reg_type, &btf_id);
3188 if (err)
3189 verbose_linfo(env, insn_idx, "; ");
3190 if (!err && t == BPF_READ && value_regno >= 0) {
3191 /* ctx access returns either a scalar, or a
3192 * PTR_TO_PACKET[_META,_END]. In the latter
3193 * case, we know the offset is zero.
3194 */
3195 if (reg_type == SCALAR_VALUE) {
3196 mark_reg_unknown(env, regs, value_regno);
3197 } else {
3198 mark_reg_known_zero(env, regs,
3199 value_regno);
3200 if (reg_type_may_be_null(reg_type))
3201 regs[value_regno].id = ++env->id_gen;
3202 /* A load of ctx field could have different
3203 * actual load size with the one encoded in the
3204 * insn. When the dst is PTR, it is for sure not
3205 * a sub-register.
3206 */
3207 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
3208 if (reg_type == PTR_TO_BTF_ID)
3209 regs[value_regno].btf_id = btf_id;
3210 }
3211 regs[value_regno].type = reg_type;
3212 }
3213
3214 } else if (reg->type == PTR_TO_STACK) {
3215 off += reg->var_off.value;
3216 err = check_stack_access(env, reg, off, size);
3217 if (err)
3218 return err;
3219
3220 state = func(env, reg);
3221 err = update_stack_depth(env, state, off);
3222 if (err)
3223 return err;
3224
3225 if (t == BPF_WRITE)
3226 err = check_stack_write(env, state, off, size,
3227 value_regno, insn_idx);
3228 else
3229 err = check_stack_read(env, state, off, size,
3230 value_regno);
3231 } else if (reg_is_pkt_pointer(reg)) {
3232 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3233 verbose(env, "cannot write into packet\n");
3234 return -EACCES;
3235 }
3236 if (t == BPF_WRITE && value_regno >= 0 &&
3237 is_pointer_value(env, value_regno)) {
3238 verbose(env, "R%d leaks addr into packet\n",
3239 value_regno);
3240 return -EACCES;
3241 }
3242 err = check_packet_access(env, regno, off, size, false);
3243 if (!err && t == BPF_READ && value_regno >= 0)
3244 mark_reg_unknown(env, regs, value_regno);
3245 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3246 if (t == BPF_WRITE && value_regno >= 0 &&
3247 is_pointer_value(env, value_regno)) {
3248 verbose(env, "R%d leaks addr into flow keys\n",
3249 value_regno);
3250 return -EACCES;
3251 }
3252
3253 err = check_flow_keys_access(env, off, size);
3254 if (!err && t == BPF_READ && value_regno >= 0)
3255 mark_reg_unknown(env, regs, value_regno);
3256 } else if (type_is_sk_pointer(reg->type)) {
3257 if (t == BPF_WRITE) {
3258 verbose(env, "R%d cannot write into %s\n",
3259 regno, reg_type_str[reg->type]);
3260 return -EACCES;
3261 }
3262 err = check_sock_access(env, insn_idx, regno, off, size, t);
3263 if (!err && value_regno >= 0)
3264 mark_reg_unknown(env, regs, value_regno);
3265 } else if (reg->type == PTR_TO_TP_BUFFER) {
3266 err = check_tp_buffer_access(env, reg, regno, off, size);
3267 if (!err && t == BPF_READ && value_regno >= 0)
3268 mark_reg_unknown(env, regs, value_regno);
3269 } else if (reg->type == PTR_TO_BTF_ID) {
3270 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3271 value_regno);
3272 } else {
3273 verbose(env, "R%d invalid mem access '%s'\n", regno,
3274 reg_type_str[reg->type]);
3275 return -EACCES;
3276 }
3277
3278 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3279 regs[value_regno].type == SCALAR_VALUE) {
3280 /* b/h/w load zero-extends, mark upper bits as known 0 */
3281 coerce_reg_to_size(&regs[value_regno], size);
3282 }
3283 return err;
3284 }
3285
3286 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3287 {
3288 int err;
3289
3290 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3291 insn->imm != 0) {
3292 verbose(env, "BPF_XADD uses reserved fields\n");
3293 return -EINVAL;
3294 }
3295
3296 /* check src1 operand */
3297 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3298 if (err)
3299 return err;
3300
3301 /* check src2 operand */
3302 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3303 if (err)
3304 return err;
3305
3306 if (is_pointer_value(env, insn->src_reg)) {
3307 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3308 return -EACCES;
3309 }
3310
3311 if (is_ctx_reg(env, insn->dst_reg) ||
3312 is_pkt_reg(env, insn->dst_reg) ||
3313 is_flow_key_reg(env, insn->dst_reg) ||
3314 is_sk_reg(env, insn->dst_reg)) {
3315 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3316 insn->dst_reg,
3317 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3318 return -EACCES;
3319 }
3320
3321 /* check whether atomic_add can read the memory */
3322 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3323 BPF_SIZE(insn->code), BPF_READ, -1, true);
3324 if (err)
3325 return err;
3326
3327 /* check whether atomic_add can write into the same memory */
3328 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3329 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3330 }
3331
3332 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3333 int off, int access_size,
3334 bool zero_size_allowed)
3335 {
3336 struct bpf_reg_state *reg = reg_state(env, regno);
3337
3338 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3339 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3340 if (tnum_is_const(reg->var_off)) {
3341 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3342 regno, off, access_size);
3343 } else {
3344 char tn_buf[48];
3345
3346 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3347 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3348 regno, tn_buf, access_size);
3349 }
3350 return -EACCES;
3351 }
3352 return 0;
3353 }
3354
3355 /* when register 'regno' is passed into function that will read 'access_size'
3356 * bytes from that pointer, make sure that it's within stack boundary
3357 * and all elements of stack are initialized.
3358 * Unlike most pointer bounds-checking functions, this one doesn't take an
3359 * 'off' argument, so it has to add in reg->off itself.
3360 */
3361 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3362 int access_size, bool zero_size_allowed,
3363 struct bpf_call_arg_meta *meta)
3364 {
3365 struct bpf_reg_state *reg = reg_state(env, regno);
3366 struct bpf_func_state *state = func(env, reg);
3367 int err, min_off, max_off, i, j, slot, spi;
3368
3369 if (reg->type != PTR_TO_STACK) {
3370 /* Allow zero-byte read from NULL, regardless of pointer type */
3371 if (zero_size_allowed && access_size == 0 &&
3372 register_is_null(reg))
3373 return 0;
3374
3375 verbose(env, "R%d type=%s expected=%s\n", regno,
3376 reg_type_str[reg->type],
3377 reg_type_str[PTR_TO_STACK]);
3378 return -EACCES;
3379 }
3380
3381 if (tnum_is_const(reg->var_off)) {
3382 min_off = max_off = reg->var_off.value + reg->off;
3383 err = __check_stack_boundary(env, regno, min_off, access_size,
3384 zero_size_allowed);
3385 if (err)
3386 return err;
3387 } else {
3388 /* Variable offset is prohibited for unprivileged mode for
3389 * simplicity since it requires corresponding support in
3390 * Spectre masking for stack ALU.
3391 * See also retrieve_ptr_limit().
3392 */
3393 if (!env->allow_ptr_leaks) {
3394 char tn_buf[48];
3395
3396 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3397 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3398 regno, tn_buf);
3399 return -EACCES;
3400 }
3401 /* Only initialized buffer on stack is allowed to be accessed
3402 * with variable offset. With uninitialized buffer it's hard to
3403 * guarantee that whole memory is marked as initialized on
3404 * helper return since specific bounds are unknown what may
3405 * cause uninitialized stack leaking.
3406 */
3407 if (meta && meta->raw_mode)
3408 meta = NULL;
3409
3410 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3411 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3412 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3413 regno);
3414 return -EACCES;
3415 }
3416 min_off = reg->smin_value + reg->off;
3417 max_off = reg->smax_value + reg->off;
3418 err = __check_stack_boundary(env, regno, min_off, access_size,
3419 zero_size_allowed);
3420 if (err) {
3421 verbose(env, "R%d min value is outside of stack bound\n",
3422 regno);
3423 return err;
3424 }
3425 err = __check_stack_boundary(env, regno, max_off, access_size,
3426 zero_size_allowed);
3427 if (err) {
3428 verbose(env, "R%d max value is outside of stack bound\n",
3429 regno);
3430 return err;
3431 }
3432 }
3433
3434 if (meta && meta->raw_mode) {
3435 meta->access_size = access_size;
3436 meta->regno = regno;
3437 return 0;
3438 }
3439
3440 for (i = min_off; i < max_off + access_size; i++) {
3441 u8 *stype;
3442
3443 slot = -i - 1;
3444 spi = slot / BPF_REG_SIZE;
3445 if (state->allocated_stack <= slot)
3446 goto err;
3447 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3448 if (*stype == STACK_MISC)
3449 goto mark;
3450 if (*stype == STACK_ZERO) {
3451 /* helper can write anything into the stack */
3452 *stype = STACK_MISC;
3453 goto mark;
3454 }
3455 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3456 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3457 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3458 for (j = 0; j < BPF_REG_SIZE; j++)
3459 state->stack[spi].slot_type[j] = STACK_MISC;
3460 goto mark;
3461 }
3462
3463 err:
3464 if (tnum_is_const(reg->var_off)) {
3465 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3466 min_off, i - min_off, access_size);
3467 } else {
3468 char tn_buf[48];
3469
3470 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3471 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3472 tn_buf, i - min_off, access_size);
3473 }
3474 return -EACCES;
3475 mark:
3476 /* reading any byte out of 8-byte 'spill_slot' will cause
3477 * the whole slot to be marked as 'read'
3478 */
3479 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3480 state->stack[spi].spilled_ptr.parent,
3481 REG_LIVE_READ64);
3482 }
3483 return update_stack_depth(env, state, min_off);
3484 }
3485
3486 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3487 int access_size, bool zero_size_allowed,
3488 struct bpf_call_arg_meta *meta)
3489 {
3490 struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
3491
3492 switch (reg->type) {
3493 case PTR_TO_PACKET:
3494 case PTR_TO_PACKET_META:
3495 return check_packet_access(env, regno, reg->off, access_size,
3496 zero_size_allowed);
3497 case PTR_TO_MAP_VALUE:
3498 if (check_map_access_type(env, regno, reg->off, access_size,
3499 meta && meta->raw_mode ? BPF_WRITE :
3500 BPF_READ))
3501 return -EACCES;
3502 return check_map_access(env, regno, reg->off, access_size,
3503 zero_size_allowed);
3504 default: /* scalar_value|ptr_to_stack or invalid ptr */
3505 return check_stack_boundary(env, regno, access_size,
3506 zero_size_allowed, meta);
3507 }
3508 }
3509
3510 /* Implementation details:
3511 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3512 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3513 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3514 * value_or_null->value transition, since the verifier only cares about
3515 * the range of access to valid map value pointer and doesn't care about actual
3516 * address of the map element.
3517 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3518 * reg->id > 0 after value_or_null->value transition. By doing so
3519 * two bpf_map_lookups will be considered two different pointers that
3520 * point to different bpf_spin_locks.
3521 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3522 * dead-locks.
3523 * Since only one bpf_spin_lock is allowed the checks are simpler than
3524 * reg_is_refcounted() logic. The verifier needs to remember only
3525 * one spin_lock instead of array of acquired_refs.
3526 * cur_state->active_spin_lock remembers which map value element got locked
3527 * and clears it after bpf_spin_unlock.
3528 */
3529 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3530 bool is_lock)
3531 {
3532 struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
3533 struct bpf_verifier_state *cur = env->cur_state;
3534 bool is_const = tnum_is_const(reg->var_off);
3535 struct bpf_map *map = reg->map_ptr;
3536 u64 val = reg->var_off.value;
3537
3538 if (reg->type != PTR_TO_MAP_VALUE) {
3539 verbose(env, "R%d is not a pointer to map_value\n", regno);
3540 return -EINVAL;
3541 }
3542 if (!is_const) {
3543 verbose(env,
3544 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3545 regno);
3546 return -EINVAL;
3547 }
3548 if (!map->btf) {
3549 verbose(env,
3550 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3551 map->name);
3552 return -EINVAL;
3553 }
3554 if (!map_value_has_spin_lock(map)) {
3555 if (map->spin_lock_off == -E2BIG)
3556 verbose(env,
3557 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3558 map->name);
3559 else if (map->spin_lock_off == -ENOENT)
3560 verbose(env,
3561 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3562 map->name);
3563 else
3564 verbose(env,
3565 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3566 map->name);
3567 return -EINVAL;
3568 }
3569 if (map->spin_lock_off != val + reg->off) {
3570 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3571 val + reg->off);
3572 return -EINVAL;
3573 }
3574 if (is_lock) {
3575 if (cur->active_spin_lock) {
3576 verbose(env,
3577 "Locking two bpf_spin_locks are not allowed\n");
3578 return -EINVAL;
3579 }
3580 cur->active_spin_lock = reg->id;
3581 } else {
3582 if (!cur->active_spin_lock) {
3583 verbose(env, "bpf_spin_unlock without taking a lock\n");
3584 return -EINVAL;
3585 }
3586 if (cur->active_spin_lock != reg->id) {
3587 verbose(env, "bpf_spin_unlock of different lock\n");
3588 return -EINVAL;
3589 }
3590 cur->active_spin_lock = 0;
3591 }
3592 return 0;
3593 }
3594
3595 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3596 {
3597 return type == ARG_PTR_TO_MEM ||
3598 type == ARG_PTR_TO_MEM_OR_NULL ||
3599 type == ARG_PTR_TO_UNINIT_MEM;
3600 }
3601
3602 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3603 {
3604 return type == ARG_CONST_SIZE ||
3605 type == ARG_CONST_SIZE_OR_ZERO;
3606 }
3607
3608 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3609 {
3610 return type == ARG_PTR_TO_INT ||
3611 type == ARG_PTR_TO_LONG;
3612 }
3613
3614 static int int_ptr_type_to_size(enum bpf_arg_type type)
3615 {
3616 if (type == ARG_PTR_TO_INT)
3617 return sizeof(u32);
3618 else if (type == ARG_PTR_TO_LONG)
3619 return sizeof(u64);
3620
3621 return -EINVAL;
3622 }
3623
3624 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
3625 enum bpf_arg_type arg_type,
3626 struct bpf_call_arg_meta *meta)
3627 {
3628 struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
3629 enum bpf_reg_type expected_type, type = reg->type;
3630 int err = 0;
3631
3632 if (arg_type == ARG_DONTCARE)
3633 return 0;
3634
3635 err = check_reg_arg(env, regno, SRC_OP);
3636 if (err)
3637 return err;
3638
3639 if (arg_type == ARG_ANYTHING) {
3640 if (is_pointer_value(env, regno)) {
3641 verbose(env, "R%d leaks addr into helper function\n",
3642 regno);
3643 return -EACCES;
3644 }
3645 return 0;
3646 }
3647
3648 if (type_is_pkt_pointer(type) &&
3649 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3650 verbose(env, "helper access to the packet is not allowed\n");
3651 return -EACCES;
3652 }
3653
3654 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3655 arg_type == ARG_PTR_TO_MAP_VALUE ||
3656 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3657 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3658 expected_type = PTR_TO_STACK;
3659 if (register_is_null(reg) &&
3660 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3661 /* final test in check_stack_boundary() */;
3662 else if (!type_is_pkt_pointer(type) &&
3663 type != PTR_TO_MAP_VALUE &&
3664 type != expected_type)
3665 goto err_type;
3666 } else if (arg_type == ARG_CONST_SIZE ||
3667 arg_type == ARG_CONST_SIZE_OR_ZERO) {
3668 expected_type = SCALAR_VALUE;
3669 if (type != expected_type)
3670 goto err_type;
3671 } else if (arg_type == ARG_CONST_MAP_PTR) {
3672 expected_type = CONST_PTR_TO_MAP;
3673 if (type != expected_type)
3674 goto err_type;
3675 } else if (arg_type == ARG_PTR_TO_CTX ||
3676 arg_type == ARG_PTR_TO_CTX_OR_NULL) {
3677 expected_type = PTR_TO_CTX;
3678 if (!(register_is_null(reg) &&
3679 arg_type == ARG_PTR_TO_CTX_OR_NULL)) {
3680 if (type != expected_type)
3681 goto err_type;
3682 err = check_ctx_reg(env, reg, regno);
3683 if (err < 0)
3684 return err;
3685 }
3686 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3687 expected_type = PTR_TO_SOCK_COMMON;
3688 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3689 if (!type_is_sk_pointer(type))
3690 goto err_type;
3691 if (reg->ref_obj_id) {
3692 if (meta->ref_obj_id) {
3693 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3694 regno, reg->ref_obj_id,
3695 meta->ref_obj_id);
3696 return -EFAULT;
3697 }
3698 meta->ref_obj_id = reg->ref_obj_id;
3699 }
3700 } else if (arg_type == ARG_PTR_TO_SOCKET) {
3701 expected_type = PTR_TO_SOCKET;
3702 if (type != expected_type)
3703 goto err_type;
3704 } else if (arg_type == ARG_PTR_TO_BTF_ID) {
3705 expected_type = PTR_TO_BTF_ID;
3706 if (type != expected_type)
3707 goto err_type;
3708 if (reg->btf_id != meta->btf_id) {
3709 verbose(env, "Helper has type %s got %s in R%d\n",
3710 kernel_type_name(meta->btf_id),
3711 kernel_type_name(reg->btf_id), regno);
3712
3713 return -EACCES;
3714 }
3715 if (!tnum_is_const(reg->var_off) || reg->var_off.value || reg->off) {
3716 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
3717 regno);
3718 return -EACCES;
3719 }
3720 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3721 if (meta->func_id == BPF_FUNC_spin_lock) {
3722 if (process_spin_lock(env, regno, true))
3723 return -EACCES;
3724 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3725 if (process_spin_lock(env, regno, false))
3726 return -EACCES;
3727 } else {
3728 verbose(env, "verifier internal error\n");
3729 return -EFAULT;
3730 }
3731 } else if (arg_type_is_mem_ptr(arg_type)) {
3732 expected_type = PTR_TO_STACK;
3733 /* One exception here. In case function allows for NULL to be
3734 * passed in as argument, it's a SCALAR_VALUE type. Final test
3735 * happens during stack boundary checking.
3736 */
3737 if (register_is_null(reg) &&
3738 arg_type == ARG_PTR_TO_MEM_OR_NULL)
3739 /* final test in check_stack_boundary() */;
3740 else if (!type_is_pkt_pointer(type) &&
3741 type != PTR_TO_MAP_VALUE &&
3742 type != expected_type)
3743 goto err_type;
3744 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
3745 } else if (arg_type_is_int_ptr(arg_type)) {
3746 expected_type = PTR_TO_STACK;
3747 if (!type_is_pkt_pointer(type) &&
3748 type != PTR_TO_MAP_VALUE &&
3749 type != expected_type)
3750 goto err_type;
3751 } else {
3752 verbose(env, "unsupported arg_type %d\n", arg_type);
3753 return -EFAULT;
3754 }
3755
3756 if (arg_type == ARG_CONST_MAP_PTR) {
3757 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
3758 meta->map_ptr = reg->map_ptr;
3759 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
3760 /* bpf_map_xxx(..., map_ptr, ..., key) call:
3761 * check that [key, key + map->key_size) are within
3762 * stack limits and initialized
3763 */
3764 if (!meta->map_ptr) {
3765 /* in function declaration map_ptr must come before
3766 * map_key, so that it's verified and known before
3767 * we have to check map_key here. Otherwise it means
3768 * that kernel subsystem misconfigured verifier
3769 */
3770 verbose(env, "invalid map_ptr to access map->key\n");
3771 return -EACCES;
3772 }
3773 err = check_helper_mem_access(env, regno,
3774 meta->map_ptr->key_size, false,
3775 NULL);
3776 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
3777 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
3778 !register_is_null(reg)) ||
3779 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
3780 /* bpf_map_xxx(..., map_ptr, ..., value) call:
3781 * check [value, value + map->value_size) validity
3782 */
3783 if (!meta->map_ptr) {
3784 /* kernel subsystem misconfigured verifier */
3785 verbose(env, "invalid map_ptr to access map->value\n");
3786 return -EACCES;
3787 }
3788 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
3789 err = check_helper_mem_access(env, regno,
3790 meta->map_ptr->value_size, false,
3791 meta);
3792 } else if (arg_type_is_mem_size(arg_type)) {
3793 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
3794
3795 /* This is used to refine r0 return value bounds for helpers
3796 * that enforce this value as an upper bound on return values.
3797 * See do_refine_retval_range() for helpers that can refine
3798 * the return value. C type of helper is u32 so we pull register
3799 * bound from umax_value however, if negative verifier errors
3800 * out. Only upper bounds can be learned because retval is an
3801 * int type and negative retvals are allowed.
3802 */
3803 meta->msize_max_value = reg->umax_value;
3804
3805 /* The register is SCALAR_VALUE; the access check
3806 * happens using its boundaries.
3807 */
3808 if (!tnum_is_const(reg->var_off))
3809 /* For unprivileged variable accesses, disable raw
3810 * mode so that the program is required to
3811 * initialize all the memory that the helper could
3812 * just partially fill up.
3813 */
3814 meta = NULL;
3815
3816 if (reg->smin_value < 0) {
3817 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
3818 regno);
3819 return -EACCES;
3820 }
3821
3822 if (reg->umin_value == 0) {
3823 err = check_helper_mem_access(env, regno - 1, 0,
3824 zero_size_allowed,
3825 meta);
3826 if (err)
3827 return err;
3828 }
3829
3830 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
3831 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
3832 regno);
3833 return -EACCES;
3834 }
3835 err = check_helper_mem_access(env, regno - 1,
3836 reg->umax_value,
3837 zero_size_allowed, meta);
3838 if (!err)
3839 err = mark_chain_precision(env, regno);
3840 } else if (arg_type_is_int_ptr(arg_type)) {
3841 int size = int_ptr_type_to_size(arg_type);
3842
3843 err = check_helper_mem_access(env, regno, size, false, meta);
3844 if (err)
3845 return err;
3846 err = check_ptr_alignment(env, reg, 0, size, true);
3847 }
3848
3849 return err;
3850 err_type:
3851 verbose(env, "R%d type=%s expected=%s\n", regno,
3852 reg_type_str[type], reg_type_str[expected_type]);
3853 return -EACCES;
3854 }
3855
3856 static int check_map_func_compatibility(struct bpf_verifier_env *env,
3857 struct bpf_map *map, int func_id)
3858 {
3859 if (!map)
3860 return 0;
3861
3862 /* We need a two way check, first is from map perspective ... */
3863 switch (map->map_type) {
3864 case BPF_MAP_TYPE_PROG_ARRAY:
3865 if (func_id != BPF_FUNC_tail_call)
3866 goto error;
3867 break;
3868 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
3869 if (func_id != BPF_FUNC_perf_event_read &&
3870 func_id != BPF_FUNC_perf_event_output &&
3871 func_id != BPF_FUNC_skb_output &&
3872 func_id != BPF_FUNC_perf_event_read_value &&
3873 func_id != BPF_FUNC_xdp_output)
3874 goto error;
3875 break;
3876 case BPF_MAP_TYPE_STACK_TRACE:
3877 if (func_id != BPF_FUNC_get_stackid)
3878 goto error;
3879 break;
3880 case BPF_MAP_TYPE_CGROUP_ARRAY:
3881 if (func_id != BPF_FUNC_skb_under_cgroup &&
3882 func_id != BPF_FUNC_current_task_under_cgroup)
3883 goto error;
3884 break;
3885 case BPF_MAP_TYPE_CGROUP_STORAGE:
3886 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
3887 if (func_id != BPF_FUNC_get_local_storage)
3888 goto error;
3889 break;
3890 case BPF_MAP_TYPE_DEVMAP:
3891 case BPF_MAP_TYPE_DEVMAP_HASH:
3892 if (func_id != BPF_FUNC_redirect_map &&
3893 func_id != BPF_FUNC_map_lookup_elem)
3894 goto error;
3895 break;
3896 /* Restrict bpf side of cpumap and xskmap, open when use-cases
3897 * appear.
3898 */
3899 case BPF_MAP_TYPE_CPUMAP:
3900 if (func_id != BPF_FUNC_redirect_map)
3901 goto error;
3902 break;
3903 case BPF_MAP_TYPE_XSKMAP:
3904 if (func_id != BPF_FUNC_redirect_map &&
3905 func_id != BPF_FUNC_map_lookup_elem)
3906 goto error;
3907 break;
3908 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
3909 case BPF_MAP_TYPE_HASH_OF_MAPS:
3910 if (func_id != BPF_FUNC_map_lookup_elem)
3911 goto error;
3912 break;
3913 case BPF_MAP_TYPE_SOCKMAP:
3914 if (func_id != BPF_FUNC_sk_redirect_map &&
3915 func_id != BPF_FUNC_sock_map_update &&
3916 func_id != BPF_FUNC_map_delete_elem &&
3917 func_id != BPF_FUNC_msg_redirect_map &&
3918 func_id != BPF_FUNC_sk_select_reuseport)
3919 goto error;
3920 break;
3921 case BPF_MAP_TYPE_SOCKHASH:
3922 if (func_id != BPF_FUNC_sk_redirect_hash &&
3923 func_id != BPF_FUNC_sock_hash_update &&
3924 func_id != BPF_FUNC_map_delete_elem &&
3925 func_id != BPF_FUNC_msg_redirect_hash &&
3926 func_id != BPF_FUNC_sk_select_reuseport)
3927 goto error;
3928 break;
3929 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
3930 if (func_id != BPF_FUNC_sk_select_reuseport)
3931 goto error;
3932 break;
3933 case BPF_MAP_TYPE_QUEUE:
3934 case BPF_MAP_TYPE_STACK:
3935 if (func_id != BPF_FUNC_map_peek_elem &&
3936 func_id != BPF_FUNC_map_pop_elem &&
3937 func_id != BPF_FUNC_map_push_elem)
3938 goto error;
3939 break;
3940 case BPF_MAP_TYPE_SK_STORAGE:
3941 if (func_id != BPF_FUNC_sk_storage_get &&
3942 func_id != BPF_FUNC_sk_storage_delete)
3943 goto error;
3944 break;
3945 default:
3946 break;
3947 }
3948
3949 /* ... and second from the function itself. */
3950 switch (func_id) {
3951 case BPF_FUNC_tail_call:
3952 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
3953 goto error;
3954 if (env->subprog_cnt > 1) {
3955 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
3956 return -EINVAL;
3957 }
3958 break;
3959 case BPF_FUNC_perf_event_read:
3960 case BPF_FUNC_perf_event_output:
3961 case BPF_FUNC_perf_event_read_value:
3962 case BPF_FUNC_skb_output:
3963 case BPF_FUNC_xdp_output:
3964 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
3965 goto error;
3966 break;
3967 case BPF_FUNC_get_stackid:
3968 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
3969 goto error;
3970 break;
3971 case BPF_FUNC_current_task_under_cgroup:
3972 case BPF_FUNC_skb_under_cgroup:
3973 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
3974 goto error;
3975 break;
3976 case BPF_FUNC_redirect_map:
3977 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
3978 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
3979 map->map_type != BPF_MAP_TYPE_CPUMAP &&
3980 map->map_type != BPF_MAP_TYPE_XSKMAP)
3981 goto error;
3982 break;
3983 case BPF_FUNC_sk_redirect_map:
3984 case BPF_FUNC_msg_redirect_map:
3985 case BPF_FUNC_sock_map_update:
3986 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
3987 goto error;
3988 break;
3989 case BPF_FUNC_sk_redirect_hash:
3990 case BPF_FUNC_msg_redirect_hash:
3991 case BPF_FUNC_sock_hash_update:
3992 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
3993 goto error;
3994 break;
3995 case BPF_FUNC_get_local_storage:
3996 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
3997 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
3998 goto error;
3999 break;
4000 case BPF_FUNC_sk_select_reuseport:
4001 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
4002 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
4003 map->map_type != BPF_MAP_TYPE_SOCKHASH)
4004 goto error;
4005 break;
4006 case BPF_FUNC_map_peek_elem:
4007 case BPF_FUNC_map_pop_elem:
4008 case BPF_FUNC_map_push_elem:
4009 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
4010 map->map_type != BPF_MAP_TYPE_STACK)
4011 goto error;
4012 break;
4013 case BPF_FUNC_sk_storage_get:
4014 case BPF_FUNC_sk_storage_delete:
4015 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
4016 goto error;
4017 break;
4018 default:
4019 break;
4020 }
4021
4022 return 0;
4023 error:
4024 verbose(env, "cannot pass map_type %d into func %s#%d\n",
4025 map->map_type, func_id_name(func_id), func_id);
4026 return -EINVAL;
4027 }
4028
4029 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
4030 {
4031 int count = 0;
4032
4033 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
4034 count++;
4035 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
4036 count++;
4037 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
4038 count++;
4039 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
4040 count++;
4041 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
4042 count++;
4043
4044 /* We only support one arg being in raw mode at the moment,
4045 * which is sufficient for the helper functions we have
4046 * right now.
4047 */
4048 return count <= 1;
4049 }
4050
4051 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
4052 enum bpf_arg_type arg_next)
4053 {
4054 return (arg_type_is_mem_ptr(arg_curr) &&
4055 !arg_type_is_mem_size(arg_next)) ||
4056 (!arg_type_is_mem_ptr(arg_curr) &&
4057 arg_type_is_mem_size(arg_next));
4058 }
4059
4060 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
4061 {
4062 /* bpf_xxx(..., buf, len) call will access 'len'
4063 * bytes from memory 'buf'. Both arg types need
4064 * to be paired, so make sure there's no buggy
4065 * helper function specification.
4066 */
4067 if (arg_type_is_mem_size(fn->arg1_type) ||
4068 arg_type_is_mem_ptr(fn->arg5_type) ||
4069 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
4070 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
4071 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
4072 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
4073 return false;
4074
4075 return true;
4076 }
4077
4078 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
4079 {
4080 int count = 0;
4081
4082 if (arg_type_may_be_refcounted(fn->arg1_type))
4083 count++;
4084 if (arg_type_may_be_refcounted(fn->arg2_type))
4085 count++;
4086 if (arg_type_may_be_refcounted(fn->arg3_type))
4087 count++;
4088 if (arg_type_may_be_refcounted(fn->arg4_type))
4089 count++;
4090 if (arg_type_may_be_refcounted(fn->arg5_type))
4091 count++;
4092
4093 /* A reference acquiring function cannot acquire
4094 * another refcounted ptr.
4095 */
4096 if (is_acquire_function(func_id) && count)
4097 return false;
4098
4099 /* We only support one arg being unreferenced at the moment,
4100 * which is sufficient for the helper functions we have right now.
4101 */
4102 return count <= 1;
4103 }
4104
4105 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
4106 {
4107 return check_raw_mode_ok(fn) &&
4108 check_arg_pair_ok(fn) &&
4109 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
4110 }
4111
4112 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
4113 * are now invalid, so turn them into unknown SCALAR_VALUE.
4114 */
4115 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
4116 struct bpf_func_state *state)
4117 {
4118 struct bpf_reg_state *regs = state->regs, *reg;
4119 int i;
4120
4121 for (i = 0; i < MAX_BPF_REG; i++)
4122 if (reg_is_pkt_pointer_any(&regs[i]))
4123 mark_reg_unknown(env, regs, i);
4124
4125 bpf_for_each_spilled_reg(i, state, reg) {
4126 if (!reg)
4127 continue;
4128 if (reg_is_pkt_pointer_any(reg))
4129 __mark_reg_unknown(env, reg);
4130 }
4131 }
4132
4133 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
4134 {
4135 struct bpf_verifier_state *vstate = env->cur_state;
4136 int i;
4137
4138 for (i = 0; i <= vstate->curframe; i++)
4139 __clear_all_pkt_pointers(env, vstate->frame[i]);
4140 }
4141
4142 static void release_reg_references(struct bpf_verifier_env *env,
4143 struct bpf_func_state *state,
4144 int ref_obj_id)
4145 {
4146 struct bpf_reg_state *regs = state->regs, *reg;
4147 int i;
4148
4149 for (i = 0; i < MAX_BPF_REG; i++)
4150 if (regs[i].ref_obj_id == ref_obj_id)
4151 mark_reg_unknown(env, regs, i);
4152
4153 bpf_for_each_spilled_reg(i, state, reg) {
4154 if (!reg)
4155 continue;
4156 if (reg->ref_obj_id == ref_obj_id)
4157 __mark_reg_unknown(env, reg);
4158 }
4159 }
4160
4161 /* The pointer with the specified id has released its reference to kernel
4162 * resources. Identify all copies of the same pointer and clear the reference.
4163 */
4164 static int release_reference(struct bpf_verifier_env *env,
4165 int ref_obj_id)
4166 {
4167 struct bpf_verifier_state *vstate = env->cur_state;
4168 int err;
4169 int i;
4170
4171 err = release_reference_state(cur_func(env), ref_obj_id);
4172 if (err)
4173 return err;
4174
4175 for (i = 0; i <= vstate->curframe; i++)
4176 release_reg_references(env, vstate->frame[i], ref_obj_id);
4177
4178 return 0;
4179 }
4180
4181 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
4182 struct bpf_reg_state *regs)
4183 {
4184 int i;
4185
4186 /* after the call registers r0 - r5 were scratched */
4187 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4188 mark_reg_not_init(env, regs, caller_saved[i]);
4189 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4190 }
4191 }
4192
4193 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
4194 int *insn_idx)
4195 {
4196 struct bpf_verifier_state *state = env->cur_state;
4197 struct bpf_func_info_aux *func_info_aux;
4198 struct bpf_func_state *caller, *callee;
4199 int i, err, subprog, target_insn;
4200 bool is_global = false;
4201
4202 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
4203 verbose(env, "the call stack of %d frames is too deep\n",
4204 state->curframe + 2);
4205 return -E2BIG;
4206 }
4207
4208 target_insn = *insn_idx + insn->imm;
4209 subprog = find_subprog(env, target_insn + 1);
4210 if (subprog < 0) {
4211 verbose(env, "verifier bug. No program starts at insn %d\n",
4212 target_insn + 1);
4213 return -EFAULT;
4214 }
4215
4216 caller = state->frame[state->curframe];
4217 if (state->frame[state->curframe + 1]) {
4218 verbose(env, "verifier bug. Frame %d already allocated\n",
4219 state->curframe + 1);
4220 return -EFAULT;
4221 }
4222
4223 func_info_aux = env->prog->aux->func_info_aux;
4224 if (func_info_aux)
4225 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4226 err = btf_check_func_arg_match(env, subprog, caller->regs);
4227 if (err == -EFAULT)
4228 return err;
4229 if (is_global) {
4230 if (err) {
4231 verbose(env, "Caller passes invalid args into func#%d\n",
4232 subprog);
4233 return err;
4234 } else {
4235 if (env->log.level & BPF_LOG_LEVEL)
4236 verbose(env,
4237 "Func#%d is global and valid. Skipping.\n",
4238 subprog);
4239 clear_caller_saved_regs(env, caller->regs);
4240
4241 /* All global functions return SCALAR_VALUE */
4242 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4243
4244 /* continue with next insn after call */
4245 return 0;
4246 }
4247 }
4248
4249 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4250 if (!callee)
4251 return -ENOMEM;
4252 state->frame[state->curframe + 1] = callee;
4253
4254 /* callee cannot access r0, r6 - r9 for reading and has to write
4255 * into its own stack before reading from it.
4256 * callee can read/write into caller's stack
4257 */
4258 init_func_state(env, callee,
4259 /* remember the callsite, it will be used by bpf_exit */
4260 *insn_idx /* callsite */,
4261 state->curframe + 1 /* frameno within this callchain */,
4262 subprog /* subprog number within this prog */);
4263
4264 /* Transfer references to the callee */
4265 err = transfer_reference_state(callee, caller);
4266 if (err)
4267 return err;
4268
4269 /* copy r1 - r5 args that callee can access. The copy includes parent
4270 * pointers, which connects us up to the liveness chain
4271 */
4272 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4273 callee->regs[i] = caller->regs[i];
4274
4275 clear_caller_saved_regs(env, caller->regs);
4276
4277 /* only increment it after check_reg_arg() finished */
4278 state->curframe++;
4279
4280 /* and go analyze first insn of the callee */
4281 *insn_idx = target_insn;
4282
4283 if (env->log.level & BPF_LOG_LEVEL) {
4284 verbose(env, "caller:\n");
4285 print_verifier_state(env, caller);
4286 verbose(env, "callee:\n");
4287 print_verifier_state(env, callee);
4288 }
4289 return 0;
4290 }
4291
4292 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4293 {
4294 struct bpf_verifier_state *state = env->cur_state;
4295 struct bpf_func_state *caller, *callee;
4296 struct bpf_reg_state *r0;
4297 int err;
4298
4299 callee = state->frame[state->curframe];
4300 r0 = &callee->regs[BPF_REG_0];
4301 if (r0->type == PTR_TO_STACK) {
4302 /* technically it's ok to return caller's stack pointer
4303 * (or caller's caller's pointer) back to the caller,
4304 * since these pointers are valid. Only current stack
4305 * pointer will be invalid as soon as function exits,
4306 * but let's be conservative
4307 */
4308 verbose(env, "cannot return stack pointer to the caller\n");
4309 return -EINVAL;
4310 }
4311
4312 state->curframe--;
4313 caller = state->frame[state->curframe];
4314 /* return to the caller whatever r0 had in the callee */
4315 caller->regs[BPF_REG_0] = *r0;
4316
4317 /* Transfer references to the caller */
4318 err = transfer_reference_state(caller, callee);
4319 if (err)
4320 return err;
4321
4322 *insn_idx = callee->callsite + 1;
4323 if (env->log.level & BPF_LOG_LEVEL) {
4324 verbose(env, "returning from callee:\n");
4325 print_verifier_state(env, callee);
4326 verbose(env, "to caller at %d:\n", *insn_idx);
4327 print_verifier_state(env, caller);
4328 }
4329 /* clear everything in the callee */
4330 free_func_state(callee);
4331 state->frame[state->curframe + 1] = NULL;
4332 return 0;
4333 }
4334
4335 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4336 int func_id,
4337 struct bpf_call_arg_meta *meta)
4338 {
4339 struct bpf_reg_state *ret_reg = &regs[BPF_REG_0];
4340
4341 if (ret_type != RET_INTEGER ||
4342 (func_id != BPF_FUNC_get_stack &&
4343 func_id != BPF_FUNC_probe_read_str &&
4344 func_id != BPF_FUNC_probe_read_kernel_str &&
4345 func_id != BPF_FUNC_probe_read_user_str))
4346 return;
4347
4348 ret_reg->smax_value = meta->msize_max_value;
4349 ret_reg->s32_max_value = meta->msize_max_value;
4350 __reg_deduce_bounds(ret_reg);
4351 __reg_bound_offset(ret_reg);
4352 __update_reg_bounds(ret_reg);
4353 }
4354
4355 static int
4356 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4357 int func_id, int insn_idx)
4358 {
4359 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4360 struct bpf_map *map = meta->map_ptr;
4361
4362 if (func_id != BPF_FUNC_tail_call &&
4363 func_id != BPF_FUNC_map_lookup_elem &&
4364 func_id != BPF_FUNC_map_update_elem &&
4365 func_id != BPF_FUNC_map_delete_elem &&
4366 func_id != BPF_FUNC_map_push_elem &&
4367 func_id != BPF_FUNC_map_pop_elem &&
4368 func_id != BPF_FUNC_map_peek_elem)
4369 return 0;
4370
4371 if (map == NULL) {
4372 verbose(env, "kernel subsystem misconfigured verifier\n");
4373 return -EINVAL;
4374 }
4375
4376 /* In case of read-only, some additional restrictions
4377 * need to be applied in order to prevent altering the
4378 * state of the map from program side.
4379 */
4380 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4381 (func_id == BPF_FUNC_map_delete_elem ||
4382 func_id == BPF_FUNC_map_update_elem ||
4383 func_id == BPF_FUNC_map_push_elem ||
4384 func_id == BPF_FUNC_map_pop_elem)) {
4385 verbose(env, "write into map forbidden\n");
4386 return -EACCES;
4387 }
4388
4389 if (!BPF_MAP_PTR(aux->map_ptr_state))
4390 bpf_map_ptr_store(aux, meta->map_ptr,
4391 meta->map_ptr->unpriv_array);
4392 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4393 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4394 meta->map_ptr->unpriv_array);
4395 return 0;
4396 }
4397
4398 static int
4399 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4400 int func_id, int insn_idx)
4401 {
4402 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4403 struct bpf_reg_state *regs = cur_regs(env), *reg;
4404 struct bpf_map *map = meta->map_ptr;
4405 struct tnum range;
4406 u64 val;
4407 int err;
4408
4409 if (func_id != BPF_FUNC_tail_call)
4410 return 0;
4411 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4412 verbose(env, "kernel subsystem misconfigured verifier\n");
4413 return -EINVAL;
4414 }
4415
4416 range = tnum_range(0, map->max_entries - 1);
4417 reg = &regs[BPF_REG_3];
4418
4419 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4420 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4421 return 0;
4422 }
4423
4424 err = mark_chain_precision(env, BPF_REG_3);
4425 if (err)
4426 return err;
4427
4428 val = reg->var_off.value;
4429 if (bpf_map_key_unseen(aux))
4430 bpf_map_key_store(aux, val);
4431 else if (!bpf_map_key_poisoned(aux) &&
4432 bpf_map_key_immediate(aux) != val)
4433 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4434 return 0;
4435 }
4436
4437 static int check_reference_leak(struct bpf_verifier_env *env)
4438 {
4439 struct bpf_func_state *state = cur_func(env);
4440 int i;
4441
4442 for (i = 0; i < state->acquired_refs; i++) {
4443 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4444 state->refs[i].id, state->refs[i].insn_idx);
4445 }
4446 return state->acquired_refs ? -EINVAL : 0;
4447 }
4448
4449 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4450 {
4451 const struct bpf_func_proto *fn = NULL;
4452 struct bpf_reg_state *regs;
4453 struct bpf_call_arg_meta meta;
4454 bool changes_data;
4455 int i, err;
4456
4457 /* find function prototype */
4458 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4459 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4460 func_id);
4461 return -EINVAL;
4462 }
4463
4464 if (env->ops->get_func_proto)
4465 fn = env->ops->get_func_proto(func_id, env->prog);
4466 if (!fn) {
4467 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4468 func_id);
4469 return -EINVAL;
4470 }
4471
4472 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4473 if (!env->prog->gpl_compatible && fn->gpl_only) {
4474 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4475 return -EINVAL;
4476 }
4477
4478 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4479 changes_data = bpf_helper_changes_pkt_data(fn->func);
4480 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4481 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4482 func_id_name(func_id), func_id);
4483 return -EINVAL;
4484 }
4485
4486 memset(&meta, 0, sizeof(meta));
4487 meta.pkt_access = fn->pkt_access;
4488
4489 err = check_func_proto(fn, func_id);
4490 if (err) {
4491 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4492 func_id_name(func_id), func_id);
4493 return err;
4494 }
4495
4496 meta.func_id = func_id;
4497 /* check args */
4498 for (i = 0; i < 5; i++) {
4499 err = btf_resolve_helper_id(&env->log, fn, i);
4500 if (err > 0)
4501 meta.btf_id = err;
4502 err = check_func_arg(env, BPF_REG_1 + i, fn->arg_type[i], &meta);
4503 if (err)
4504 return err;
4505 }
4506
4507 err = record_func_map(env, &meta, func_id, insn_idx);
4508 if (err)
4509 return err;
4510
4511 err = record_func_key(env, &meta, func_id, insn_idx);
4512 if (err)
4513 return err;
4514
4515 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4516 * is inferred from register state.
4517 */
4518 for (i = 0; i < meta.access_size; i++) {
4519 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4520 BPF_WRITE, -1, false);
4521 if (err)
4522 return err;
4523 }
4524
4525 if (func_id == BPF_FUNC_tail_call) {
4526 err = check_reference_leak(env);
4527 if (err) {
4528 verbose(env, "tail_call would lead to reference leak\n");
4529 return err;
4530 }
4531 } else if (is_release_function(func_id)) {
4532 err = release_reference(env, meta.ref_obj_id);
4533 if (err) {
4534 verbose(env, "func %s#%d reference has not been acquired before\n",
4535 func_id_name(func_id), func_id);
4536 return err;
4537 }
4538 }
4539
4540 regs = cur_regs(env);
4541
4542 /* check that flags argument in get_local_storage(map, flags) is 0,
4543 * this is required because get_local_storage() can't return an error.
4544 */
4545 if (func_id == BPF_FUNC_get_local_storage &&
4546 !register_is_null(&regs[BPF_REG_2])) {
4547 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4548 return -EINVAL;
4549 }
4550
4551 /* reset caller saved regs */
4552 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4553 mark_reg_not_init(env, regs, caller_saved[i]);
4554 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4555 }
4556
4557 /* helper call returns 64-bit value. */
4558 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4559
4560 /* update return register (already marked as written above) */
4561 if (fn->ret_type == RET_INTEGER) {
4562 /* sets type to SCALAR_VALUE */
4563 mark_reg_unknown(env, regs, BPF_REG_0);
4564 } else if (fn->ret_type == RET_VOID) {
4565 regs[BPF_REG_0].type = NOT_INIT;
4566 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4567 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4568 /* There is no offset yet applied, variable or fixed */
4569 mark_reg_known_zero(env, regs, BPF_REG_0);
4570 /* remember map_ptr, so that check_map_access()
4571 * can check 'value_size' boundary of memory access
4572 * to map element returned from bpf_map_lookup_elem()
4573 */
4574 if (meta.map_ptr == NULL) {
4575 verbose(env,
4576 "kernel subsystem misconfigured verifier\n");
4577 return -EINVAL;
4578 }
4579 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4580 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4581 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4582 if (map_value_has_spin_lock(meta.map_ptr))
4583 regs[BPF_REG_0].id = ++env->id_gen;
4584 } else {
4585 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4586 regs[BPF_REG_0].id = ++env->id_gen;
4587 }
4588 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4589 mark_reg_known_zero(env, regs, BPF_REG_0);
4590 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4591 regs[BPF_REG_0].id = ++env->id_gen;
4592 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4593 mark_reg_known_zero(env, regs, BPF_REG_0);
4594 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4595 regs[BPF_REG_0].id = ++env->id_gen;
4596 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4597 mark_reg_known_zero(env, regs, BPF_REG_0);
4598 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4599 regs[BPF_REG_0].id = ++env->id_gen;
4600 } else {
4601 verbose(env, "unknown return type %d of func %s#%d\n",
4602 fn->ret_type, func_id_name(func_id), func_id);
4603 return -EINVAL;
4604 }
4605
4606 if (is_ptr_cast_function(func_id)) {
4607 /* For release_reference() */
4608 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4609 } else if (is_acquire_function(func_id)) {
4610 int id = acquire_reference_state(env, insn_idx);
4611
4612 if (id < 0)
4613 return id;
4614 /* For mark_ptr_or_null_reg() */
4615 regs[BPF_REG_0].id = id;
4616 /* For release_reference() */
4617 regs[BPF_REG_0].ref_obj_id = id;
4618 }
4619
4620 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
4621
4622 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4623 if (err)
4624 return err;
4625
4626 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
4627 const char *err_str;
4628
4629 #ifdef CONFIG_PERF_EVENTS
4630 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4631 err_str = "cannot get callchain buffer for func %s#%d\n";
4632 #else
4633 err = -ENOTSUPP;
4634 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4635 #endif
4636 if (err) {
4637 verbose(env, err_str, func_id_name(func_id), func_id);
4638 return err;
4639 }
4640
4641 env->prog->has_callchain_buf = true;
4642 }
4643
4644 if (changes_data)
4645 clear_all_pkt_pointers(env);
4646 return 0;
4647 }
4648
4649 static bool signed_add_overflows(s64 a, s64 b)
4650 {
4651 /* Do the add in u64, where overflow is well-defined */
4652 s64 res = (s64)((u64)a + (u64)b);
4653
4654 if (b < 0)
4655 return res > a;
4656 return res < a;
4657 }
4658
4659 static bool signed_add32_overflows(s64 a, s64 b)
4660 {
4661 /* Do the add in u32, where overflow is well-defined */
4662 s32 res = (s32)((u32)a + (u32)b);
4663
4664 if (b < 0)
4665 return res > a;
4666 return res < a;
4667 }
4668
4669 static bool signed_sub_overflows(s32 a, s32 b)
4670 {
4671 /* Do the sub in u64, where overflow is well-defined */
4672 s64 res = (s64)((u64)a - (u64)b);
4673
4674 if (b < 0)
4675 return res < a;
4676 return res > a;
4677 }
4678
4679 static bool signed_sub32_overflows(s32 a, s32 b)
4680 {
4681 /* Do the sub in u64, where overflow is well-defined */
4682 s32 res = (s32)((u32)a - (u32)b);
4683
4684 if (b < 0)
4685 return res < a;
4686 return res > a;
4687 }
4688
4689 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
4690 const struct bpf_reg_state *reg,
4691 enum bpf_reg_type type)
4692 {
4693 bool known = tnum_is_const(reg->var_off);
4694 s64 val = reg->var_off.value;
4695 s64 smin = reg->smin_value;
4696
4697 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
4698 verbose(env, "math between %s pointer and %lld is not allowed\n",
4699 reg_type_str[type], val);
4700 return false;
4701 }
4702
4703 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
4704 verbose(env, "%s pointer offset %d is not allowed\n",
4705 reg_type_str[type], reg->off);
4706 return false;
4707 }
4708
4709 if (smin == S64_MIN) {
4710 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
4711 reg_type_str[type]);
4712 return false;
4713 }
4714
4715 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
4716 verbose(env, "value %lld makes %s pointer be out of bounds\n",
4717 smin, reg_type_str[type]);
4718 return false;
4719 }
4720
4721 return true;
4722 }
4723
4724 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
4725 {
4726 return &env->insn_aux_data[env->insn_idx];
4727 }
4728
4729 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
4730 u32 *ptr_limit, u8 opcode, bool off_is_neg)
4731 {
4732 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
4733 (opcode == BPF_SUB && !off_is_neg);
4734 u32 off;
4735
4736 switch (ptr_reg->type) {
4737 case PTR_TO_STACK:
4738 /* Indirect variable offset stack access is prohibited in
4739 * unprivileged mode so it's not handled here.
4740 */
4741 off = ptr_reg->off + ptr_reg->var_off.value;
4742 if (mask_to_left)
4743 *ptr_limit = MAX_BPF_STACK + off;
4744 else
4745 *ptr_limit = -off;
4746 return 0;
4747 case PTR_TO_MAP_VALUE:
4748 if (mask_to_left) {
4749 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
4750 } else {
4751 off = ptr_reg->smin_value + ptr_reg->off;
4752 *ptr_limit = ptr_reg->map_ptr->value_size - off;
4753 }
4754 return 0;
4755 default:
4756 return -EINVAL;
4757 }
4758 }
4759
4760 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
4761 const struct bpf_insn *insn)
4762 {
4763 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
4764 }
4765
4766 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
4767 u32 alu_state, u32 alu_limit)
4768 {
4769 /* If we arrived here from different branches with different
4770 * state or limits to sanitize, then this won't work.
4771 */
4772 if (aux->alu_state &&
4773 (aux->alu_state != alu_state ||
4774 aux->alu_limit != alu_limit))
4775 return -EACCES;
4776
4777 /* Corresponding fixup done in fixup_bpf_calls(). */
4778 aux->alu_state = alu_state;
4779 aux->alu_limit = alu_limit;
4780 return 0;
4781 }
4782
4783 static int sanitize_val_alu(struct bpf_verifier_env *env,
4784 struct bpf_insn *insn)
4785 {
4786 struct bpf_insn_aux_data *aux = cur_aux(env);
4787
4788 if (can_skip_alu_sanitation(env, insn))
4789 return 0;
4790
4791 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
4792 }
4793
4794 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
4795 struct bpf_insn *insn,
4796 const struct bpf_reg_state *ptr_reg,
4797 struct bpf_reg_state *dst_reg,
4798 bool off_is_neg)
4799 {
4800 struct bpf_verifier_state *vstate = env->cur_state;
4801 struct bpf_insn_aux_data *aux = cur_aux(env);
4802 bool ptr_is_dst_reg = ptr_reg == dst_reg;
4803 u8 opcode = BPF_OP(insn->code);
4804 u32 alu_state, alu_limit;
4805 struct bpf_reg_state tmp;
4806 bool ret;
4807
4808 if (can_skip_alu_sanitation(env, insn))
4809 return 0;
4810
4811 /* We already marked aux for masking from non-speculative
4812 * paths, thus we got here in the first place. We only care
4813 * to explore bad access from here.
4814 */
4815 if (vstate->speculative)
4816 goto do_sim;
4817
4818 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
4819 alu_state |= ptr_is_dst_reg ?
4820 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
4821
4822 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
4823 return 0;
4824 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
4825 return -EACCES;
4826 do_sim:
4827 /* Simulate and find potential out-of-bounds access under
4828 * speculative execution from truncation as a result of
4829 * masking when off was not within expected range. If off
4830 * sits in dst, then we temporarily need to move ptr there
4831 * to simulate dst (== 0) +/-= ptr. Needed, for example,
4832 * for cases where we use K-based arithmetic in one direction
4833 * and truncated reg-based in the other in order to explore
4834 * bad access.
4835 */
4836 if (!ptr_is_dst_reg) {
4837 tmp = *dst_reg;
4838 *dst_reg = *ptr_reg;
4839 }
4840 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
4841 if (!ptr_is_dst_reg && ret)
4842 *dst_reg = tmp;
4843 return !ret ? -EFAULT : 0;
4844 }
4845
4846 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
4847 * Caller should also handle BPF_MOV case separately.
4848 * If we return -EACCES, caller may want to try again treating pointer as a
4849 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
4850 */
4851 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
4852 struct bpf_insn *insn,
4853 const struct bpf_reg_state *ptr_reg,
4854 const struct bpf_reg_state *off_reg)
4855 {
4856 struct bpf_verifier_state *vstate = env->cur_state;
4857 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4858 struct bpf_reg_state *regs = state->regs, *dst_reg;
4859 bool known = tnum_is_const(off_reg->var_off);
4860 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
4861 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
4862 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
4863 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
4864 u32 dst = insn->dst_reg, src = insn->src_reg;
4865 u8 opcode = BPF_OP(insn->code);
4866 int ret;
4867
4868 dst_reg = &regs[dst];
4869
4870 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
4871 smin_val > smax_val || umin_val > umax_val) {
4872 /* Taint dst register if offset had invalid bounds derived from
4873 * e.g. dead branches.
4874 */
4875 __mark_reg_unknown(env, dst_reg);
4876 return 0;
4877 }
4878
4879 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4880 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
4881 verbose(env,
4882 "R%d 32-bit pointer arithmetic prohibited\n",
4883 dst);
4884 return -EACCES;
4885 }
4886
4887 switch (ptr_reg->type) {
4888 case PTR_TO_MAP_VALUE_OR_NULL:
4889 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
4890 dst, reg_type_str[ptr_reg->type]);
4891 return -EACCES;
4892 case CONST_PTR_TO_MAP:
4893 case PTR_TO_PACKET_END:
4894 case PTR_TO_SOCKET:
4895 case PTR_TO_SOCKET_OR_NULL:
4896 case PTR_TO_SOCK_COMMON:
4897 case PTR_TO_SOCK_COMMON_OR_NULL:
4898 case PTR_TO_TCP_SOCK:
4899 case PTR_TO_TCP_SOCK_OR_NULL:
4900 case PTR_TO_XDP_SOCK:
4901 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
4902 dst, reg_type_str[ptr_reg->type]);
4903 return -EACCES;
4904 case PTR_TO_MAP_VALUE:
4905 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
4906 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
4907 off_reg == dst_reg ? dst : src);
4908 return -EACCES;
4909 }
4910 /* fall-through */
4911 default:
4912 break;
4913 }
4914
4915 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
4916 * The id may be overwritten later if we create a new variable offset.
4917 */
4918 dst_reg->type = ptr_reg->type;
4919 dst_reg->id = ptr_reg->id;
4920
4921 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
4922 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
4923 return -EINVAL;
4924
4925 /* pointer types do not carry 32-bit bounds at the moment. */
4926 __mark_reg32_unbounded(dst_reg);
4927
4928 switch (opcode) {
4929 case BPF_ADD:
4930 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4931 if (ret < 0) {
4932 verbose(env, "R%d tried to add from different maps or paths\n", dst);
4933 return ret;
4934 }
4935 /* We can take a fixed offset as long as it doesn't overflow
4936 * the s32 'off' field
4937 */
4938 if (known && (ptr_reg->off + smin_val ==
4939 (s64)(s32)(ptr_reg->off + smin_val))) {
4940 /* pointer += K. Accumulate it into fixed offset */
4941 dst_reg->smin_value = smin_ptr;
4942 dst_reg->smax_value = smax_ptr;
4943 dst_reg->umin_value = umin_ptr;
4944 dst_reg->umax_value = umax_ptr;
4945 dst_reg->var_off = ptr_reg->var_off;
4946 dst_reg->off = ptr_reg->off + smin_val;
4947 dst_reg->raw = ptr_reg->raw;
4948 break;
4949 }
4950 /* A new variable offset is created. Note that off_reg->off
4951 * == 0, since it's a scalar.
4952 * dst_reg gets the pointer type and since some positive
4953 * integer value was added to the pointer, give it a new 'id'
4954 * if it's a PTR_TO_PACKET.
4955 * this creates a new 'base' pointer, off_reg (variable) gets
4956 * added into the variable offset, and we copy the fixed offset
4957 * from ptr_reg.
4958 */
4959 if (signed_add_overflows(smin_ptr, smin_val) ||
4960 signed_add_overflows(smax_ptr, smax_val)) {
4961 dst_reg->smin_value = S64_MIN;
4962 dst_reg->smax_value = S64_MAX;
4963 } else {
4964 dst_reg->smin_value = smin_ptr + smin_val;
4965 dst_reg->smax_value = smax_ptr + smax_val;
4966 }
4967 if (umin_ptr + umin_val < umin_ptr ||
4968 umax_ptr + umax_val < umax_ptr) {
4969 dst_reg->umin_value = 0;
4970 dst_reg->umax_value = U64_MAX;
4971 } else {
4972 dst_reg->umin_value = umin_ptr + umin_val;
4973 dst_reg->umax_value = umax_ptr + umax_val;
4974 }
4975 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
4976 dst_reg->off = ptr_reg->off;
4977 dst_reg->raw = ptr_reg->raw;
4978 if (reg_is_pkt_pointer(ptr_reg)) {
4979 dst_reg->id = ++env->id_gen;
4980 /* something was added to pkt_ptr, set range to zero */
4981 dst_reg->raw = 0;
4982 }
4983 break;
4984 case BPF_SUB:
4985 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4986 if (ret < 0) {
4987 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
4988 return ret;
4989 }
4990 if (dst_reg == off_reg) {
4991 /* scalar -= pointer. Creates an unknown scalar */
4992 verbose(env, "R%d tried to subtract pointer from scalar\n",
4993 dst);
4994 return -EACCES;
4995 }
4996 /* We don't allow subtraction from FP, because (according to
4997 * test_verifier.c test "invalid fp arithmetic", JITs might not
4998 * be able to deal with it.
4999 */
5000 if (ptr_reg->type == PTR_TO_STACK) {
5001 verbose(env, "R%d subtraction from stack pointer prohibited\n",
5002 dst);
5003 return -EACCES;
5004 }
5005 if (known && (ptr_reg->off - smin_val ==
5006 (s64)(s32)(ptr_reg->off - smin_val))) {
5007 /* pointer -= K. Subtract it from fixed offset */
5008 dst_reg->smin_value = smin_ptr;
5009 dst_reg->smax_value = smax_ptr;
5010 dst_reg->umin_value = umin_ptr;
5011 dst_reg->umax_value = umax_ptr;
5012 dst_reg->var_off = ptr_reg->var_off;
5013 dst_reg->id = ptr_reg->id;
5014 dst_reg->off = ptr_reg->off - smin_val;
5015 dst_reg->raw = ptr_reg->raw;
5016 break;
5017 }
5018 /* A new variable offset is created. If the subtrahend is known
5019 * nonnegative, then any reg->range we had before is still good.
5020 */
5021 if (signed_sub_overflows(smin_ptr, smax_val) ||
5022 signed_sub_overflows(smax_ptr, smin_val)) {
5023 /* Overflow possible, we know nothing */
5024 dst_reg->smin_value = S64_MIN;
5025 dst_reg->smax_value = S64_MAX;
5026 } else {
5027 dst_reg->smin_value = smin_ptr - smax_val;
5028 dst_reg->smax_value = smax_ptr - smin_val;
5029 }
5030 if (umin_ptr < umax_val) {
5031 /* Overflow possible, we know nothing */
5032 dst_reg->umin_value = 0;
5033 dst_reg->umax_value = U64_MAX;
5034 } else {
5035 /* Cannot overflow (as long as bounds are consistent) */
5036 dst_reg->umin_value = umin_ptr - umax_val;
5037 dst_reg->umax_value = umax_ptr - umin_val;
5038 }
5039 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
5040 dst_reg->off = ptr_reg->off;
5041 dst_reg->raw = ptr_reg->raw;
5042 if (reg_is_pkt_pointer(ptr_reg)) {
5043 dst_reg->id = ++env->id_gen;
5044 /* something was added to pkt_ptr, set range to zero */
5045 if (smin_val < 0)
5046 dst_reg->raw = 0;
5047 }
5048 break;
5049 case BPF_AND:
5050 case BPF_OR:
5051 case BPF_XOR:
5052 /* bitwise ops on pointers are troublesome, prohibit. */
5053 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
5054 dst, bpf_alu_string[opcode >> 4]);
5055 return -EACCES;
5056 default:
5057 /* other operators (e.g. MUL,LSH) produce non-pointer results */
5058 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
5059 dst, bpf_alu_string[opcode >> 4]);
5060 return -EACCES;
5061 }
5062
5063 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
5064 return -EINVAL;
5065
5066 __update_reg_bounds(dst_reg);
5067 __reg_deduce_bounds(dst_reg);
5068 __reg_bound_offset(dst_reg);
5069
5070 /* For unprivileged we require that resulting offset must be in bounds
5071 * in order to be able to sanitize access later on.
5072 */
5073 if (!env->allow_ptr_leaks) {
5074 if (dst_reg->type == PTR_TO_MAP_VALUE &&
5075 check_map_access(env, dst, dst_reg->off, 1, false)) {
5076 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
5077 "prohibited for !root\n", dst);
5078 return -EACCES;
5079 } else if (dst_reg->type == PTR_TO_STACK &&
5080 check_stack_access(env, dst_reg, dst_reg->off +
5081 dst_reg->var_off.value, 1)) {
5082 verbose(env, "R%d stack pointer arithmetic goes out of range, "
5083 "prohibited for !root\n", dst);
5084 return -EACCES;
5085 }
5086 }
5087
5088 return 0;
5089 }
5090
5091 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
5092 struct bpf_reg_state *src_reg)
5093 {
5094 s32 smin_val = src_reg->s32_min_value;
5095 s32 smax_val = src_reg->s32_max_value;
5096 u32 umin_val = src_reg->u32_min_value;
5097 u32 umax_val = src_reg->u32_max_value;
5098
5099 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
5100 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
5101 dst_reg->s32_min_value = S32_MIN;
5102 dst_reg->s32_max_value = S32_MAX;
5103 } else {
5104 dst_reg->s32_min_value += smin_val;
5105 dst_reg->s32_max_value += smax_val;
5106 }
5107 if (dst_reg->u32_min_value + umin_val < umin_val ||
5108 dst_reg->u32_max_value + umax_val < umax_val) {
5109 dst_reg->u32_min_value = 0;
5110 dst_reg->u32_max_value = U32_MAX;
5111 } else {
5112 dst_reg->u32_min_value += umin_val;
5113 dst_reg->u32_max_value += umax_val;
5114 }
5115 }
5116
5117 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
5118 struct bpf_reg_state *src_reg)
5119 {
5120 s64 smin_val = src_reg->smin_value;
5121 s64 smax_val = src_reg->smax_value;
5122 u64 umin_val = src_reg->umin_value;
5123 u64 umax_val = src_reg->umax_value;
5124
5125 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
5126 signed_add_overflows(dst_reg->smax_value, smax_val)) {
5127 dst_reg->smin_value = S64_MIN;
5128 dst_reg->smax_value = S64_MAX;
5129 } else {
5130 dst_reg->smin_value += smin_val;
5131 dst_reg->smax_value += smax_val;
5132 }
5133 if (dst_reg->umin_value + umin_val < umin_val ||
5134 dst_reg->umax_value + umax_val < umax_val) {
5135 dst_reg->umin_value = 0;
5136 dst_reg->umax_value = U64_MAX;
5137 } else {
5138 dst_reg->umin_value += umin_val;
5139 dst_reg->umax_value += umax_val;
5140 }
5141 }
5142
5143 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
5144 struct bpf_reg_state *src_reg)
5145 {
5146 s32 smin_val = src_reg->s32_min_value;
5147 s32 smax_val = src_reg->s32_max_value;
5148 u32 umin_val = src_reg->u32_min_value;
5149 u32 umax_val = src_reg->u32_max_value;
5150
5151 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
5152 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
5153 /* Overflow possible, we know nothing */
5154 dst_reg->s32_min_value = S32_MIN;
5155 dst_reg->s32_max_value = S32_MAX;
5156 } else {
5157 dst_reg->s32_min_value -= smax_val;
5158 dst_reg->s32_max_value -= smin_val;
5159 }
5160 if (dst_reg->u32_min_value < umax_val) {
5161 /* Overflow possible, we know nothing */
5162 dst_reg->u32_min_value = 0;
5163 dst_reg->u32_max_value = U32_MAX;
5164 } else {
5165 /* Cannot overflow (as long as bounds are consistent) */
5166 dst_reg->u32_min_value -= umax_val;
5167 dst_reg->u32_max_value -= umin_val;
5168 }
5169 }
5170
5171 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
5172 struct bpf_reg_state *src_reg)
5173 {
5174 s64 smin_val = src_reg->smin_value;
5175 s64 smax_val = src_reg->smax_value;
5176 u64 umin_val = src_reg->umin_value;
5177 u64 umax_val = src_reg->umax_value;
5178
5179 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
5180 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
5181 /* Overflow possible, we know nothing */
5182 dst_reg->smin_value = S64_MIN;
5183 dst_reg->smax_value = S64_MAX;
5184 } else {
5185 dst_reg->smin_value -= smax_val;
5186 dst_reg->smax_value -= smin_val;
5187 }
5188 if (dst_reg->umin_value < umax_val) {
5189 /* Overflow possible, we know nothing */
5190 dst_reg->umin_value = 0;
5191 dst_reg->umax_value = U64_MAX;
5192 } else {
5193 /* Cannot overflow (as long as bounds are consistent) */
5194 dst_reg->umin_value -= umax_val;
5195 dst_reg->umax_value -= umin_val;
5196 }
5197 }
5198
5199 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
5200 struct bpf_reg_state *src_reg)
5201 {
5202 s32 smin_val = src_reg->s32_min_value;
5203 u32 umin_val = src_reg->u32_min_value;
5204 u32 umax_val = src_reg->u32_max_value;
5205
5206 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
5207 /* Ain't nobody got time to multiply that sign */
5208 __mark_reg32_unbounded(dst_reg);
5209 return;
5210 }
5211 /* Both values are positive, so we can work with unsigned and
5212 * copy the result to signed (unless it exceeds S32_MAX).
5213 */
5214 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
5215 /* Potential overflow, we know nothing */
5216 __mark_reg32_unbounded(dst_reg);
5217 return;
5218 }
5219 dst_reg->u32_min_value *= umin_val;
5220 dst_reg->u32_max_value *= umax_val;
5221 if (dst_reg->u32_max_value > S32_MAX) {
5222 /* Overflow possible, we know nothing */
5223 dst_reg->s32_min_value = S32_MIN;
5224 dst_reg->s32_max_value = S32_MAX;
5225 } else {
5226 dst_reg->s32_min_value = dst_reg->u32_min_value;
5227 dst_reg->s32_max_value = dst_reg->u32_max_value;
5228 }
5229 }
5230
5231 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
5232 struct bpf_reg_state *src_reg)
5233 {
5234 s64 smin_val = src_reg->smin_value;
5235 u64 umin_val = src_reg->umin_value;
5236 u64 umax_val = src_reg->umax_value;
5237
5238 if (smin_val < 0 || dst_reg->smin_value < 0) {
5239 /* Ain't nobody got time to multiply that sign */
5240 __mark_reg64_unbounded(dst_reg);
5241 return;
5242 }
5243 /* Both values are positive, so we can work with unsigned and
5244 * copy the result to signed (unless it exceeds S64_MAX).
5245 */
5246 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
5247 /* Potential overflow, we know nothing */
5248 __mark_reg64_unbounded(dst_reg);
5249 return;
5250 }
5251 dst_reg->umin_value *= umin_val;
5252 dst_reg->umax_value *= umax_val;
5253 if (dst_reg->umax_value > S64_MAX) {
5254 /* Overflow possible, we know nothing */
5255 dst_reg->smin_value = S64_MIN;
5256 dst_reg->smax_value = S64_MAX;
5257 } else {
5258 dst_reg->smin_value = dst_reg->umin_value;
5259 dst_reg->smax_value = dst_reg->umax_value;
5260 }
5261 }
5262
5263 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
5264 struct bpf_reg_state *src_reg)
5265 {
5266 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5267 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5268 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5269 s32 smin_val = src_reg->s32_min_value;
5270 u32 umax_val = src_reg->u32_max_value;
5271
5272 /* Assuming scalar64_min_max_and will be called so its safe
5273 * to skip updating register for known 32-bit case.
5274 */
5275 if (src_known && dst_known)
5276 return;
5277
5278 /* We get our minimum from the var_off, since that's inherently
5279 * bitwise. Our maximum is the minimum of the operands' maxima.
5280 */
5281 dst_reg->u32_min_value = var32_off.value;
5282 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
5283 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5284 /* Lose signed bounds when ANDing negative numbers,
5285 * ain't nobody got time for that.
5286 */
5287 dst_reg->s32_min_value = S32_MIN;
5288 dst_reg->s32_max_value = S32_MAX;
5289 } else {
5290 /* ANDing two positives gives a positive, so safe to
5291 * cast result into s64.
5292 */
5293 dst_reg->s32_min_value = dst_reg->u32_min_value;
5294 dst_reg->s32_max_value = dst_reg->u32_max_value;
5295 }
5296
5297 }
5298
5299 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
5300 struct bpf_reg_state *src_reg)
5301 {
5302 bool src_known = tnum_is_const(src_reg->var_off);
5303 bool dst_known = tnum_is_const(dst_reg->var_off);
5304 s64 smin_val = src_reg->smin_value;
5305 u64 umax_val = src_reg->umax_value;
5306
5307 if (src_known && dst_known) {
5308 __mark_reg_known(dst_reg, dst_reg->var_off.value &
5309 src_reg->var_off.value);
5310 return;
5311 }
5312
5313 /* We get our minimum from the var_off, since that's inherently
5314 * bitwise. Our maximum is the minimum of the operands' maxima.
5315 */
5316 dst_reg->umin_value = dst_reg->var_off.value;
5317 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
5318 if (dst_reg->smin_value < 0 || smin_val < 0) {
5319 /* Lose signed bounds when ANDing negative numbers,
5320 * ain't nobody got time for that.
5321 */
5322 dst_reg->smin_value = S64_MIN;
5323 dst_reg->smax_value = S64_MAX;
5324 } else {
5325 /* ANDing two positives gives a positive, so safe to
5326 * cast result into s64.
5327 */
5328 dst_reg->smin_value = dst_reg->umin_value;
5329 dst_reg->smax_value = dst_reg->umax_value;
5330 }
5331 /* We may learn something more from the var_off */
5332 __update_reg_bounds(dst_reg);
5333 }
5334
5335 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
5336 struct bpf_reg_state *src_reg)
5337 {
5338 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5339 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5340 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5341 s32 smin_val = src_reg->smin_value;
5342 u32 umin_val = src_reg->umin_value;
5343
5344 /* Assuming scalar64_min_max_or will be called so it is safe
5345 * to skip updating register for known case.
5346 */
5347 if (src_known && dst_known)
5348 return;
5349
5350 /* We get our maximum from the var_off, and our minimum is the
5351 * maximum of the operands' minima
5352 */
5353 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
5354 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
5355 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5356 /* Lose signed bounds when ORing negative numbers,
5357 * ain't nobody got time for that.
5358 */
5359 dst_reg->s32_min_value = S32_MIN;
5360 dst_reg->s32_max_value = S32_MAX;
5361 } else {
5362 /* ORing two positives gives a positive, so safe to
5363 * cast result into s64.
5364 */
5365 dst_reg->s32_min_value = dst_reg->umin_value;
5366 dst_reg->s32_max_value = dst_reg->umax_value;
5367 }
5368 }
5369
5370 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
5371 struct bpf_reg_state *src_reg)
5372 {
5373 bool src_known = tnum_is_const(src_reg->var_off);
5374 bool dst_known = tnum_is_const(dst_reg->var_off);
5375 s64 smin_val = src_reg->smin_value;
5376 u64 umin_val = src_reg->umin_value;
5377
5378 if (src_known && dst_known) {
5379 __mark_reg_known(dst_reg, dst_reg->var_off.value |
5380 src_reg->var_off.value);
5381 return;
5382 }
5383
5384 /* We get our maximum from the var_off, and our minimum is the
5385 * maximum of the operands' minima
5386 */
5387 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
5388 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
5389 if (dst_reg->smin_value < 0 || smin_val < 0) {
5390 /* Lose signed bounds when ORing negative numbers,
5391 * ain't nobody got time for that.
5392 */
5393 dst_reg->smin_value = S64_MIN;
5394 dst_reg->smax_value = S64_MAX;
5395 } else {
5396 /* ORing two positives gives a positive, so safe to
5397 * cast result into s64.
5398 */
5399 dst_reg->smin_value = dst_reg->umin_value;
5400 dst_reg->smax_value = dst_reg->umax_value;
5401 }
5402 /* We may learn something more from the var_off */
5403 __update_reg_bounds(dst_reg);
5404 }
5405
5406 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5407 u64 umin_val, u64 umax_val)
5408 {
5409 /* We lose all sign bit information (except what we can pick
5410 * up from var_off)
5411 */
5412 dst_reg->s32_min_value = S32_MIN;
5413 dst_reg->s32_max_value = S32_MAX;
5414 /* If we might shift our top bit out, then we know nothing */
5415 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
5416 dst_reg->u32_min_value = 0;
5417 dst_reg->u32_max_value = U32_MAX;
5418 } else {
5419 dst_reg->u32_min_value <<= umin_val;
5420 dst_reg->u32_max_value <<= umax_val;
5421 }
5422 }
5423
5424 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5425 struct bpf_reg_state *src_reg)
5426 {
5427 u32 umax_val = src_reg->u32_max_value;
5428 u32 umin_val = src_reg->u32_min_value;
5429 /* u32 alu operation will zext upper bits */
5430 struct tnum subreg = tnum_subreg(dst_reg->var_off);
5431
5432 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
5433 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
5434 /* Not required but being careful mark reg64 bounds as unknown so
5435 * that we are forced to pick them up from tnum and zext later and
5436 * if some path skips this step we are still safe.
5437 */
5438 __mark_reg64_unbounded(dst_reg);
5439 __update_reg32_bounds(dst_reg);
5440 }
5441
5442 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
5443 u64 umin_val, u64 umax_val)
5444 {
5445 /* Special case <<32 because it is a common compiler pattern to sign
5446 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
5447 * positive we know this shift will also be positive so we can track
5448 * bounds correctly. Otherwise we lose all sign bit information except
5449 * what we can pick up from var_off. Perhaps we can generalize this
5450 * later to shifts of any length.
5451 */
5452 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
5453 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
5454 else
5455 dst_reg->smax_value = S64_MAX;
5456
5457 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
5458 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
5459 else
5460 dst_reg->smin_value = S64_MIN;
5461
5462 /* If we might shift our top bit out, then we know nothing */
5463 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
5464 dst_reg->umin_value = 0;
5465 dst_reg->umax_value = U64_MAX;
5466 } else {
5467 dst_reg->umin_value <<= umin_val;
5468 dst_reg->umax_value <<= umax_val;
5469 }
5470 }
5471
5472 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
5473 struct bpf_reg_state *src_reg)
5474 {
5475 u64 umax_val = src_reg->umax_value;
5476 u64 umin_val = src_reg->umin_value;
5477
5478 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
5479 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
5480 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
5481
5482 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
5483 /* We may learn something more from the var_off */
5484 __update_reg_bounds(dst_reg);
5485 }
5486
5487 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
5488 struct bpf_reg_state *src_reg)
5489 {
5490 struct tnum subreg = tnum_subreg(dst_reg->var_off);
5491 u32 umax_val = src_reg->u32_max_value;
5492 u32 umin_val = src_reg->u32_min_value;
5493
5494 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5495 * be negative, then either:
5496 * 1) src_reg might be zero, so the sign bit of the result is
5497 * unknown, so we lose our signed bounds
5498 * 2) it's known negative, thus the unsigned bounds capture the
5499 * signed bounds
5500 * 3) the signed bounds cross zero, so they tell us nothing
5501 * about the result
5502 * If the value in dst_reg is known nonnegative, then again the
5503 * unsigned bounts capture the signed bounds.
5504 * Thus, in all cases it suffices to blow away our signed bounds
5505 * and rely on inferring new ones from the unsigned bounds and
5506 * var_off of the result.
5507 */
5508 dst_reg->s32_min_value = S32_MIN;
5509 dst_reg->s32_max_value = S32_MAX;
5510
5511 dst_reg->var_off = tnum_rshift(subreg, umin_val);
5512 dst_reg->u32_min_value >>= umax_val;
5513 dst_reg->u32_max_value >>= umin_val;
5514
5515 __mark_reg64_unbounded(dst_reg);
5516 __update_reg32_bounds(dst_reg);
5517 }
5518
5519 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
5520 struct bpf_reg_state *src_reg)
5521 {
5522 u64 umax_val = src_reg->umax_value;
5523 u64 umin_val = src_reg->umin_value;
5524
5525 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5526 * be negative, then either:
5527 * 1) src_reg might be zero, so the sign bit of the result is
5528 * unknown, so we lose our signed bounds
5529 * 2) it's known negative, thus the unsigned bounds capture the
5530 * signed bounds
5531 * 3) the signed bounds cross zero, so they tell us nothing
5532 * about the result
5533 * If the value in dst_reg is known nonnegative, then again the
5534 * unsigned bounts capture the signed bounds.
5535 * Thus, in all cases it suffices to blow away our signed bounds
5536 * and rely on inferring new ones from the unsigned bounds and
5537 * var_off of the result.
5538 */
5539 dst_reg->smin_value = S64_MIN;
5540 dst_reg->smax_value = S64_MAX;
5541 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
5542 dst_reg->umin_value >>= umax_val;
5543 dst_reg->umax_value >>= umin_val;
5544
5545 /* Its not easy to operate on alu32 bounds here because it depends
5546 * on bits being shifted in. Take easy way out and mark unbounded
5547 * so we can recalculate later from tnum.
5548 */
5549 __mark_reg32_unbounded(dst_reg);
5550 __update_reg_bounds(dst_reg);
5551 }
5552
5553 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
5554 struct bpf_reg_state *src_reg)
5555 {
5556 u64 umin_val = src_reg->u32_min_value;
5557
5558 /* Upon reaching here, src_known is true and
5559 * umax_val is equal to umin_val.
5560 */
5561 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
5562 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
5563
5564 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
5565
5566 /* blow away the dst_reg umin_value/umax_value and rely on
5567 * dst_reg var_off to refine the result.
5568 */
5569 dst_reg->u32_min_value = 0;
5570 dst_reg->u32_max_value = U32_MAX;
5571
5572 __mark_reg64_unbounded(dst_reg);
5573 __update_reg32_bounds(dst_reg);
5574 }
5575
5576 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
5577 struct bpf_reg_state *src_reg)
5578 {
5579 u64 umin_val = src_reg->umin_value;
5580
5581 /* Upon reaching here, src_known is true and umax_val is equal
5582 * to umin_val.
5583 */
5584 dst_reg->smin_value >>= umin_val;
5585 dst_reg->smax_value >>= umin_val;
5586
5587 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
5588
5589 /* blow away the dst_reg umin_value/umax_value and rely on
5590 * dst_reg var_off to refine the result.
5591 */
5592 dst_reg->umin_value = 0;
5593 dst_reg->umax_value = U64_MAX;
5594
5595 /* Its not easy to operate on alu32 bounds here because it depends
5596 * on bits being shifted in from upper 32-bits. Take easy way out
5597 * and mark unbounded so we can recalculate later from tnum.
5598 */
5599 __mark_reg32_unbounded(dst_reg);
5600 __update_reg_bounds(dst_reg);
5601 }
5602
5603 /* WARNING: This function does calculations on 64-bit values, but the actual
5604 * execution may occur on 32-bit values. Therefore, things like bitshifts
5605 * need extra checks in the 32-bit case.
5606 */
5607 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
5608 struct bpf_insn *insn,
5609 struct bpf_reg_state *dst_reg,
5610 struct bpf_reg_state src_reg)
5611 {
5612 struct bpf_reg_state *regs = cur_regs(env);
5613 u8 opcode = BPF_OP(insn->code);
5614 bool src_known, dst_known;
5615 s64 smin_val, smax_val;
5616 u64 umin_val, umax_val;
5617 s32 s32_min_val, s32_max_val;
5618 u32 u32_min_val, u32_max_val;
5619 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
5620 u32 dst = insn->dst_reg;
5621 int ret;
5622 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
5623
5624 smin_val = src_reg.smin_value;
5625 smax_val = src_reg.smax_value;
5626 umin_val = src_reg.umin_value;
5627 umax_val = src_reg.umax_value;
5628
5629 s32_min_val = src_reg.s32_min_value;
5630 s32_max_val = src_reg.s32_max_value;
5631 u32_min_val = src_reg.u32_min_value;
5632 u32_max_val = src_reg.u32_max_value;
5633
5634 if (alu32) {
5635 src_known = tnum_subreg_is_const(src_reg.var_off);
5636 dst_known = tnum_subreg_is_const(dst_reg->var_off);
5637 if ((src_known &&
5638 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
5639 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
5640 /* Taint dst register if offset had invalid bounds
5641 * derived from e.g. dead branches.
5642 */
5643 __mark_reg_unknown(env, dst_reg);
5644 return 0;
5645 }
5646 } else {
5647 src_known = tnum_is_const(src_reg.var_off);
5648 dst_known = tnum_is_const(dst_reg->var_off);
5649 if ((src_known &&
5650 (smin_val != smax_val || umin_val != umax_val)) ||
5651 smin_val > smax_val || umin_val > umax_val) {
5652 /* Taint dst register if offset had invalid bounds
5653 * derived from e.g. dead branches.
5654 */
5655 __mark_reg_unknown(env, dst_reg);
5656 return 0;
5657 }
5658 }
5659
5660 if (!src_known &&
5661 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
5662 __mark_reg_unknown(env, dst_reg);
5663 return 0;
5664 }
5665
5666 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
5667 * There are two classes of instructions: The first class we track both
5668 * alu32 and alu64 sign/unsigned bounds independently this provides the
5669 * greatest amount of precision when alu operations are mixed with jmp32
5670 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
5671 * and BPF_OR. This is possible because these ops have fairly easy to
5672 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
5673 * See alu32 verifier tests for examples. The second class of
5674 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
5675 * with regards to tracking sign/unsigned bounds because the bits may
5676 * cross subreg boundaries in the alu64 case. When this happens we mark
5677 * the reg unbounded in the subreg bound space and use the resulting
5678 * tnum to calculate an approximation of the sign/unsigned bounds.
5679 */
5680 switch (opcode) {
5681 case BPF_ADD:
5682 ret = sanitize_val_alu(env, insn);
5683 if (ret < 0) {
5684 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
5685 return ret;
5686 }
5687 scalar32_min_max_add(dst_reg, &src_reg);
5688 scalar_min_max_add(dst_reg, &src_reg);
5689 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
5690 break;
5691 case BPF_SUB:
5692 ret = sanitize_val_alu(env, insn);
5693 if (ret < 0) {
5694 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
5695 return ret;
5696 }
5697 scalar32_min_max_sub(dst_reg, &src_reg);
5698 scalar_min_max_sub(dst_reg, &src_reg);
5699 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
5700 break;
5701 case BPF_MUL:
5702 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
5703 scalar32_min_max_mul(dst_reg, &src_reg);
5704 scalar_min_max_mul(dst_reg, &src_reg);
5705 break;
5706 case BPF_AND:
5707 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
5708 scalar32_min_max_and(dst_reg, &src_reg);
5709 scalar_min_max_and(dst_reg, &src_reg);
5710 break;
5711 case BPF_OR:
5712 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
5713 scalar32_min_max_or(dst_reg, &src_reg);
5714 scalar_min_max_or(dst_reg, &src_reg);
5715 break;
5716 case BPF_LSH:
5717 if (umax_val >= insn_bitness) {
5718 /* Shifts greater than 31 or 63 are undefined.
5719 * This includes shifts by a negative number.
5720 */
5721 mark_reg_unknown(env, regs, insn->dst_reg);
5722 break;
5723 }
5724 if (alu32)
5725 scalar32_min_max_lsh(dst_reg, &src_reg);
5726 else
5727 scalar_min_max_lsh(dst_reg, &src_reg);
5728 break;
5729 case BPF_RSH:
5730 if (umax_val >= insn_bitness) {
5731 /* Shifts greater than 31 or 63 are undefined.
5732 * This includes shifts by a negative number.
5733 */
5734 mark_reg_unknown(env, regs, insn->dst_reg);
5735 break;
5736 }
5737 if (alu32)
5738 scalar32_min_max_rsh(dst_reg, &src_reg);
5739 else
5740 scalar_min_max_rsh(dst_reg, &src_reg);
5741 break;
5742 case BPF_ARSH:
5743 if (umax_val >= insn_bitness) {
5744 /* Shifts greater than 31 or 63 are undefined.
5745 * This includes shifts by a negative number.
5746 */
5747 mark_reg_unknown(env, regs, insn->dst_reg);
5748 break;
5749 }
5750 if (alu32)
5751 scalar32_min_max_arsh(dst_reg, &src_reg);
5752 else
5753 scalar_min_max_arsh(dst_reg, &src_reg);
5754 break;
5755 default:
5756 mark_reg_unknown(env, regs, insn->dst_reg);
5757 break;
5758 }
5759
5760 /* ALU32 ops are zero extended into 64bit register */
5761 if (alu32)
5762 zext_32_to_64(dst_reg);
5763
5764 __update_reg_bounds(dst_reg);
5765 __reg_deduce_bounds(dst_reg);
5766 __reg_bound_offset(dst_reg);
5767 return 0;
5768 }
5769
5770 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
5771 * and var_off.
5772 */
5773 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
5774 struct bpf_insn *insn)
5775 {
5776 struct bpf_verifier_state *vstate = env->cur_state;
5777 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5778 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
5779 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
5780 u8 opcode = BPF_OP(insn->code);
5781 int err;
5782
5783 dst_reg = &regs[insn->dst_reg];
5784 src_reg = NULL;
5785 if (dst_reg->type != SCALAR_VALUE)
5786 ptr_reg = dst_reg;
5787 if (BPF_SRC(insn->code) == BPF_X) {
5788 src_reg = &regs[insn->src_reg];
5789 if (src_reg->type != SCALAR_VALUE) {
5790 if (dst_reg->type != SCALAR_VALUE) {
5791 /* Combining two pointers by any ALU op yields
5792 * an arbitrary scalar. Disallow all math except
5793 * pointer subtraction
5794 */
5795 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5796 mark_reg_unknown(env, regs, insn->dst_reg);
5797 return 0;
5798 }
5799 verbose(env, "R%d pointer %s pointer prohibited\n",
5800 insn->dst_reg,
5801 bpf_alu_string[opcode >> 4]);
5802 return -EACCES;
5803 } else {
5804 /* scalar += pointer
5805 * This is legal, but we have to reverse our
5806 * src/dest handling in computing the range
5807 */
5808 err = mark_chain_precision(env, insn->dst_reg);
5809 if (err)
5810 return err;
5811 return adjust_ptr_min_max_vals(env, insn,
5812 src_reg, dst_reg);
5813 }
5814 } else if (ptr_reg) {
5815 /* pointer += scalar */
5816 err = mark_chain_precision(env, insn->src_reg);
5817 if (err)
5818 return err;
5819 return adjust_ptr_min_max_vals(env, insn,
5820 dst_reg, src_reg);
5821 }
5822 } else {
5823 /* Pretend the src is a reg with a known value, since we only
5824 * need to be able to read from this state.
5825 */
5826 off_reg.type = SCALAR_VALUE;
5827 __mark_reg_known(&off_reg, insn->imm);
5828 src_reg = &off_reg;
5829 if (ptr_reg) /* pointer += K */
5830 return adjust_ptr_min_max_vals(env, insn,
5831 ptr_reg, src_reg);
5832 }
5833
5834 /* Got here implies adding two SCALAR_VALUEs */
5835 if (WARN_ON_ONCE(ptr_reg)) {
5836 print_verifier_state(env, state);
5837 verbose(env, "verifier internal error: unexpected ptr_reg\n");
5838 return -EINVAL;
5839 }
5840 if (WARN_ON(!src_reg)) {
5841 print_verifier_state(env, state);
5842 verbose(env, "verifier internal error: no src_reg\n");
5843 return -EINVAL;
5844 }
5845 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
5846 }
5847
5848 /* check validity of 32-bit and 64-bit arithmetic operations */
5849 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
5850 {
5851 struct bpf_reg_state *regs = cur_regs(env);
5852 u8 opcode = BPF_OP(insn->code);
5853 int err;
5854
5855 if (opcode == BPF_END || opcode == BPF_NEG) {
5856 if (opcode == BPF_NEG) {
5857 if (BPF_SRC(insn->code) != 0 ||
5858 insn->src_reg != BPF_REG_0 ||
5859 insn->off != 0 || insn->imm != 0) {
5860 verbose(env, "BPF_NEG uses reserved fields\n");
5861 return -EINVAL;
5862 }
5863 } else {
5864 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
5865 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
5866 BPF_CLASS(insn->code) == BPF_ALU64) {
5867 verbose(env, "BPF_END uses reserved fields\n");
5868 return -EINVAL;
5869 }
5870 }
5871
5872 /* check src operand */
5873 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5874 if (err)
5875 return err;
5876
5877 if (is_pointer_value(env, insn->dst_reg)) {
5878 verbose(env, "R%d pointer arithmetic prohibited\n",
5879 insn->dst_reg);
5880 return -EACCES;
5881 }
5882
5883 /* check dest operand */
5884 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5885 if (err)
5886 return err;
5887
5888 } else if (opcode == BPF_MOV) {
5889
5890 if (BPF_SRC(insn->code) == BPF_X) {
5891 if (insn->imm != 0 || insn->off != 0) {
5892 verbose(env, "BPF_MOV uses reserved fields\n");
5893 return -EINVAL;
5894 }
5895
5896 /* check src operand */
5897 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5898 if (err)
5899 return err;
5900 } else {
5901 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5902 verbose(env, "BPF_MOV uses reserved fields\n");
5903 return -EINVAL;
5904 }
5905 }
5906
5907 /* check dest operand, mark as required later */
5908 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5909 if (err)
5910 return err;
5911
5912 if (BPF_SRC(insn->code) == BPF_X) {
5913 struct bpf_reg_state *src_reg = regs + insn->src_reg;
5914 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
5915
5916 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5917 /* case: R1 = R2
5918 * copy register state to dest reg
5919 */
5920 *dst_reg = *src_reg;
5921 dst_reg->live |= REG_LIVE_WRITTEN;
5922 dst_reg->subreg_def = DEF_NOT_SUBREG;
5923 } else {
5924 /* R1 = (u32) R2 */
5925 if (is_pointer_value(env, insn->src_reg)) {
5926 verbose(env,
5927 "R%d partial copy of pointer\n",
5928 insn->src_reg);
5929 return -EACCES;
5930 } else if (src_reg->type == SCALAR_VALUE) {
5931 *dst_reg = *src_reg;
5932 dst_reg->live |= REG_LIVE_WRITTEN;
5933 dst_reg->subreg_def = env->insn_idx + 1;
5934 } else {
5935 mark_reg_unknown(env, regs,
5936 insn->dst_reg);
5937 }
5938 zext_32_to_64(dst_reg);
5939 }
5940 } else {
5941 /* case: R = imm
5942 * remember the value we stored into this reg
5943 */
5944 /* clear any state __mark_reg_known doesn't set */
5945 mark_reg_unknown(env, regs, insn->dst_reg);
5946 regs[insn->dst_reg].type = SCALAR_VALUE;
5947 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5948 __mark_reg_known(regs + insn->dst_reg,
5949 insn->imm);
5950 } else {
5951 __mark_reg_known(regs + insn->dst_reg,
5952 (u32)insn->imm);
5953 }
5954 }
5955
5956 } else if (opcode > BPF_END) {
5957 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
5958 return -EINVAL;
5959
5960 } else { /* all other ALU ops: and, sub, xor, add, ... */
5961
5962 if (BPF_SRC(insn->code) == BPF_X) {
5963 if (insn->imm != 0 || insn->off != 0) {
5964 verbose(env, "BPF_ALU uses reserved fields\n");
5965 return -EINVAL;
5966 }
5967 /* check src1 operand */
5968 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5969 if (err)
5970 return err;
5971 } else {
5972 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5973 verbose(env, "BPF_ALU uses reserved fields\n");
5974 return -EINVAL;
5975 }
5976 }
5977
5978 /* check src2 operand */
5979 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5980 if (err)
5981 return err;
5982
5983 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
5984 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
5985 verbose(env, "div by zero\n");
5986 return -EINVAL;
5987 }
5988
5989 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
5990 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
5991 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
5992
5993 if (insn->imm < 0 || insn->imm >= size) {
5994 verbose(env, "invalid shift %d\n", insn->imm);
5995 return -EINVAL;
5996 }
5997 }
5998
5999 /* check dest operand */
6000 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6001 if (err)
6002 return err;
6003
6004 return adjust_reg_min_max_vals(env, insn);
6005 }
6006
6007 return 0;
6008 }
6009
6010 static void __find_good_pkt_pointers(struct bpf_func_state *state,
6011 struct bpf_reg_state *dst_reg,
6012 enum bpf_reg_type type, u16 new_range)
6013 {
6014 struct bpf_reg_state *reg;
6015 int i;
6016
6017 for (i = 0; i < MAX_BPF_REG; i++) {
6018 reg = &state->regs[i];
6019 if (reg->type == type && reg->id == dst_reg->id)
6020 /* keep the maximum range already checked */
6021 reg->range = max(reg->range, new_range);
6022 }
6023
6024 bpf_for_each_spilled_reg(i, state, reg) {
6025 if (!reg)
6026 continue;
6027 if (reg->type == type && reg->id == dst_reg->id)
6028 reg->range = max(reg->range, new_range);
6029 }
6030 }
6031
6032 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
6033 struct bpf_reg_state *dst_reg,
6034 enum bpf_reg_type type,
6035 bool range_right_open)
6036 {
6037 u16 new_range;
6038 int i;
6039
6040 if (dst_reg->off < 0 ||
6041 (dst_reg->off == 0 && range_right_open))
6042 /* This doesn't give us any range */
6043 return;
6044
6045 if (dst_reg->umax_value > MAX_PACKET_OFF ||
6046 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
6047 /* Risk of overflow. For instance, ptr + (1<<63) may be less
6048 * than pkt_end, but that's because it's also less than pkt.
6049 */
6050 return;
6051
6052 new_range = dst_reg->off;
6053 if (range_right_open)
6054 new_range--;
6055
6056 /* Examples for register markings:
6057 *
6058 * pkt_data in dst register:
6059 *
6060 * r2 = r3;
6061 * r2 += 8;
6062 * if (r2 > pkt_end) goto <handle exception>
6063 * <access okay>
6064 *
6065 * r2 = r3;
6066 * r2 += 8;
6067 * if (r2 < pkt_end) goto <access okay>
6068 * <handle exception>
6069 *
6070 * Where:
6071 * r2 == dst_reg, pkt_end == src_reg
6072 * r2=pkt(id=n,off=8,r=0)
6073 * r3=pkt(id=n,off=0,r=0)
6074 *
6075 * pkt_data in src register:
6076 *
6077 * r2 = r3;
6078 * r2 += 8;
6079 * if (pkt_end >= r2) goto <access okay>
6080 * <handle exception>
6081 *
6082 * r2 = r3;
6083 * r2 += 8;
6084 * if (pkt_end <= r2) goto <handle exception>
6085 * <access okay>
6086 *
6087 * Where:
6088 * pkt_end == dst_reg, r2 == src_reg
6089 * r2=pkt(id=n,off=8,r=0)
6090 * r3=pkt(id=n,off=0,r=0)
6091 *
6092 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
6093 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
6094 * and [r3, r3 + 8-1) respectively is safe to access depending on
6095 * the check.
6096 */
6097
6098 /* If our ids match, then we must have the same max_value. And we
6099 * don't care about the other reg's fixed offset, since if it's too big
6100 * the range won't allow anything.
6101 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
6102 */
6103 for (i = 0; i <= vstate->curframe; i++)
6104 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
6105 new_range);
6106 }
6107
6108 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
6109 {
6110 struct tnum subreg = tnum_subreg(reg->var_off);
6111 s32 sval = (s32)val;
6112
6113 switch (opcode) {
6114 case BPF_JEQ:
6115 if (tnum_is_const(subreg))
6116 return !!tnum_equals_const(subreg, val);
6117 break;
6118 case BPF_JNE:
6119 if (tnum_is_const(subreg))
6120 return !tnum_equals_const(subreg, val);
6121 break;
6122 case BPF_JSET:
6123 if ((~subreg.mask & subreg.value) & val)
6124 return 1;
6125 if (!((subreg.mask | subreg.value) & val))
6126 return 0;
6127 break;
6128 case BPF_JGT:
6129 if (reg->u32_min_value > val)
6130 return 1;
6131 else if (reg->u32_max_value <= val)
6132 return 0;
6133 break;
6134 case BPF_JSGT:
6135 if (reg->s32_min_value > sval)
6136 return 1;
6137 else if (reg->s32_max_value < sval)
6138 return 0;
6139 break;
6140 case BPF_JLT:
6141 if (reg->u32_max_value < val)
6142 return 1;
6143 else if (reg->u32_min_value >= val)
6144 return 0;
6145 break;
6146 case BPF_JSLT:
6147 if (reg->s32_max_value < sval)
6148 return 1;
6149 else if (reg->s32_min_value >= sval)
6150 return 0;
6151 break;
6152 case BPF_JGE:
6153 if (reg->u32_min_value >= val)
6154 return 1;
6155 else if (reg->u32_max_value < val)
6156 return 0;
6157 break;
6158 case BPF_JSGE:
6159 if (reg->s32_min_value >= sval)
6160 return 1;
6161 else if (reg->s32_max_value < sval)
6162 return 0;
6163 break;
6164 case BPF_JLE:
6165 if (reg->u32_max_value <= val)
6166 return 1;
6167 else if (reg->u32_min_value > val)
6168 return 0;
6169 break;
6170 case BPF_JSLE:
6171 if (reg->s32_max_value <= sval)
6172 return 1;
6173 else if (reg->s32_min_value > sval)
6174 return 0;
6175 break;
6176 }
6177
6178 return -1;
6179 }
6180
6181
6182 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
6183 {
6184 s64 sval = (s64)val;
6185
6186 switch (opcode) {
6187 case BPF_JEQ:
6188 if (tnum_is_const(reg->var_off))
6189 return !!tnum_equals_const(reg->var_off, val);
6190 break;
6191 case BPF_JNE:
6192 if (tnum_is_const(reg->var_off))
6193 return !tnum_equals_const(reg->var_off, val);
6194 break;
6195 case BPF_JSET:
6196 if ((~reg->var_off.mask & reg->var_off.value) & val)
6197 return 1;
6198 if (!((reg->var_off.mask | reg->var_off.value) & val))
6199 return 0;
6200 break;
6201 case BPF_JGT:
6202 if (reg->umin_value > val)
6203 return 1;
6204 else if (reg->umax_value <= val)
6205 return 0;
6206 break;
6207 case BPF_JSGT:
6208 if (reg->smin_value > sval)
6209 return 1;
6210 else if (reg->smax_value < sval)
6211 return 0;
6212 break;
6213 case BPF_JLT:
6214 if (reg->umax_value < val)
6215 return 1;
6216 else if (reg->umin_value >= val)
6217 return 0;
6218 break;
6219 case BPF_JSLT:
6220 if (reg->smax_value < sval)
6221 return 1;
6222 else if (reg->smin_value >= sval)
6223 return 0;
6224 break;
6225 case BPF_JGE:
6226 if (reg->umin_value >= val)
6227 return 1;
6228 else if (reg->umax_value < val)
6229 return 0;
6230 break;
6231 case BPF_JSGE:
6232 if (reg->smin_value >= sval)
6233 return 1;
6234 else if (reg->smax_value < sval)
6235 return 0;
6236 break;
6237 case BPF_JLE:
6238 if (reg->umax_value <= val)
6239 return 1;
6240 else if (reg->umin_value > val)
6241 return 0;
6242 break;
6243 case BPF_JSLE:
6244 if (reg->smax_value <= sval)
6245 return 1;
6246 else if (reg->smin_value > sval)
6247 return 0;
6248 break;
6249 }
6250
6251 return -1;
6252 }
6253
6254 /* compute branch direction of the expression "if (reg opcode val) goto target;"
6255 * and return:
6256 * 1 - branch will be taken and "goto target" will be executed
6257 * 0 - branch will not be taken and fall-through to next insn
6258 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
6259 * range [0,10]
6260 */
6261 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
6262 bool is_jmp32)
6263 {
6264 if (__is_pointer_value(false, reg))
6265 return -1;
6266
6267 if (is_jmp32)
6268 return is_branch32_taken(reg, val, opcode);
6269 return is_branch64_taken(reg, val, opcode);
6270 }
6271
6272 /* Adjusts the register min/max values in the case that the dst_reg is the
6273 * variable register that we are working on, and src_reg is a constant or we're
6274 * simply doing a BPF_K check.
6275 * In JEQ/JNE cases we also adjust the var_off values.
6276 */
6277 static void reg_set_min_max(struct bpf_reg_state *true_reg,
6278 struct bpf_reg_state *false_reg,
6279 u64 val, u32 val32,
6280 u8 opcode, bool is_jmp32)
6281 {
6282 struct tnum false_32off = tnum_subreg(false_reg->var_off);
6283 struct tnum false_64off = false_reg->var_off;
6284 struct tnum true_32off = tnum_subreg(true_reg->var_off);
6285 struct tnum true_64off = true_reg->var_off;
6286 s64 sval = (s64)val;
6287 s32 sval32 = (s32)val32;
6288
6289 /* If the dst_reg is a pointer, we can't learn anything about its
6290 * variable offset from the compare (unless src_reg were a pointer into
6291 * the same object, but we don't bother with that.
6292 * Since false_reg and true_reg have the same type by construction, we
6293 * only need to check one of them for pointerness.
6294 */
6295 if (__is_pointer_value(false, false_reg))
6296 return;
6297
6298 switch (opcode) {
6299 case BPF_JEQ:
6300 case BPF_JNE:
6301 {
6302 struct bpf_reg_state *reg =
6303 opcode == BPF_JEQ ? true_reg : false_reg;
6304
6305 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
6306 * if it is true we know the value for sure. Likewise for
6307 * BPF_JNE.
6308 */
6309 if (is_jmp32)
6310 __mark_reg32_known(reg, val32);
6311 else
6312 __mark_reg_known(reg, val);
6313 break;
6314 }
6315 case BPF_JSET:
6316 if (is_jmp32) {
6317 false_32off = tnum_and(false_32off, tnum_const(~val32));
6318 if (is_power_of_2(val32))
6319 true_32off = tnum_or(true_32off,
6320 tnum_const(val32));
6321 } else {
6322 false_64off = tnum_and(false_64off, tnum_const(~val));
6323 if (is_power_of_2(val))
6324 true_64off = tnum_or(true_64off,
6325 tnum_const(val));
6326 }
6327 break;
6328 case BPF_JGE:
6329 case BPF_JGT:
6330 {
6331 if (is_jmp32) {
6332 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
6333 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
6334
6335 false_reg->u32_max_value = min(false_reg->u32_max_value,
6336 false_umax);
6337 true_reg->u32_min_value = max(true_reg->u32_min_value,
6338 true_umin);
6339 } else {
6340 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
6341 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
6342
6343 false_reg->umax_value = min(false_reg->umax_value, false_umax);
6344 true_reg->umin_value = max(true_reg->umin_value, true_umin);
6345 }
6346 break;
6347 }
6348 case BPF_JSGE:
6349 case BPF_JSGT:
6350 {
6351 if (is_jmp32) {
6352 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
6353 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
6354
6355 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
6356 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
6357 } else {
6358 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
6359 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
6360
6361 false_reg->smax_value = min(false_reg->smax_value, false_smax);
6362 true_reg->smin_value = max(true_reg->smin_value, true_smin);
6363 }
6364 break;
6365 }
6366 case BPF_JLE:
6367 case BPF_JLT:
6368 {
6369 if (is_jmp32) {
6370 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
6371 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
6372
6373 false_reg->u32_min_value = max(false_reg->u32_min_value,
6374 false_umin);
6375 true_reg->u32_max_value = min(true_reg->u32_max_value,
6376 true_umax);
6377 } else {
6378 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
6379 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
6380
6381 false_reg->umin_value = max(false_reg->umin_value, false_umin);
6382 true_reg->umax_value = min(true_reg->umax_value, true_umax);
6383 }
6384 break;
6385 }
6386 case BPF_JSLE:
6387 case BPF_JSLT:
6388 {
6389 if (is_jmp32) {
6390 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
6391 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
6392
6393 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
6394 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
6395 } else {
6396 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
6397 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
6398
6399 false_reg->smin_value = max(false_reg->smin_value, false_smin);
6400 true_reg->smax_value = min(true_reg->smax_value, true_smax);
6401 }
6402 break;
6403 }
6404 default:
6405 return;
6406 }
6407
6408 if (is_jmp32) {
6409 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
6410 tnum_subreg(false_32off));
6411 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
6412 tnum_subreg(true_32off));
6413 __reg_combine_32_into_64(false_reg);
6414 __reg_combine_32_into_64(true_reg);
6415 } else {
6416 false_reg->var_off = false_64off;
6417 true_reg->var_off = true_64off;
6418 __reg_combine_64_into_32(false_reg);
6419 __reg_combine_64_into_32(true_reg);
6420 }
6421 }
6422
6423 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
6424 * the variable reg.
6425 */
6426 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
6427 struct bpf_reg_state *false_reg,
6428 u64 val, u32 val32,
6429 u8 opcode, bool is_jmp32)
6430 {
6431 /* How can we transform "a <op> b" into "b <op> a"? */
6432 static const u8 opcode_flip[16] = {
6433 /* these stay the same */
6434 [BPF_JEQ >> 4] = BPF_JEQ,
6435 [BPF_JNE >> 4] = BPF_JNE,
6436 [BPF_JSET >> 4] = BPF_JSET,
6437 /* these swap "lesser" and "greater" (L and G in the opcodes) */
6438 [BPF_JGE >> 4] = BPF_JLE,
6439 [BPF_JGT >> 4] = BPF_JLT,
6440 [BPF_JLE >> 4] = BPF_JGE,
6441 [BPF_JLT >> 4] = BPF_JGT,
6442 [BPF_JSGE >> 4] = BPF_JSLE,
6443 [BPF_JSGT >> 4] = BPF_JSLT,
6444 [BPF_JSLE >> 4] = BPF_JSGE,
6445 [BPF_JSLT >> 4] = BPF_JSGT
6446 };
6447 opcode = opcode_flip[opcode >> 4];
6448 /* This uses zero as "not present in table"; luckily the zero opcode,
6449 * BPF_JA, can't get here.
6450 */
6451 if (opcode)
6452 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
6453 }
6454
6455 /* Regs are known to be equal, so intersect their min/max/var_off */
6456 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
6457 struct bpf_reg_state *dst_reg)
6458 {
6459 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
6460 dst_reg->umin_value);
6461 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
6462 dst_reg->umax_value);
6463 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
6464 dst_reg->smin_value);
6465 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
6466 dst_reg->smax_value);
6467 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
6468 dst_reg->var_off);
6469 /* We might have learned new bounds from the var_off. */
6470 __update_reg_bounds(src_reg);
6471 __update_reg_bounds(dst_reg);
6472 /* We might have learned something about the sign bit. */
6473 __reg_deduce_bounds(src_reg);
6474 __reg_deduce_bounds(dst_reg);
6475 /* We might have learned some bits from the bounds. */
6476 __reg_bound_offset(src_reg);
6477 __reg_bound_offset(dst_reg);
6478 /* Intersecting with the old var_off might have improved our bounds
6479 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
6480 * then new var_off is (0; 0x7f...fc) which improves our umax.
6481 */
6482 __update_reg_bounds(src_reg);
6483 __update_reg_bounds(dst_reg);
6484 }
6485
6486 static void reg_combine_min_max(struct bpf_reg_state *true_src,
6487 struct bpf_reg_state *true_dst,
6488 struct bpf_reg_state *false_src,
6489 struct bpf_reg_state *false_dst,
6490 u8 opcode)
6491 {
6492 switch (opcode) {
6493 case BPF_JEQ:
6494 __reg_combine_min_max(true_src, true_dst);
6495 break;
6496 case BPF_JNE:
6497 __reg_combine_min_max(false_src, false_dst);
6498 break;
6499 }
6500 }
6501
6502 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
6503 struct bpf_reg_state *reg, u32 id,
6504 bool is_null)
6505 {
6506 if (reg_type_may_be_null(reg->type) && reg->id == id) {
6507 /* Old offset (both fixed and variable parts) should
6508 * have been known-zero, because we don't allow pointer
6509 * arithmetic on pointers that might be NULL.
6510 */
6511 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
6512 !tnum_equals_const(reg->var_off, 0) ||
6513 reg->off)) {
6514 __mark_reg_known_zero(reg);
6515 reg->off = 0;
6516 }
6517 if (is_null) {
6518 reg->type = SCALAR_VALUE;
6519 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
6520 if (reg->map_ptr->inner_map_meta) {
6521 reg->type = CONST_PTR_TO_MAP;
6522 reg->map_ptr = reg->map_ptr->inner_map_meta;
6523 } else if (reg->map_ptr->map_type ==
6524 BPF_MAP_TYPE_XSKMAP) {
6525 reg->type = PTR_TO_XDP_SOCK;
6526 } else {
6527 reg->type = PTR_TO_MAP_VALUE;
6528 }
6529 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
6530 reg->type = PTR_TO_SOCKET;
6531 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
6532 reg->type = PTR_TO_SOCK_COMMON;
6533 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
6534 reg->type = PTR_TO_TCP_SOCK;
6535 }
6536 if (is_null) {
6537 /* We don't need id and ref_obj_id from this point
6538 * onwards anymore, thus we should better reset it,
6539 * so that state pruning has chances to take effect.
6540 */
6541 reg->id = 0;
6542 reg->ref_obj_id = 0;
6543 } else if (!reg_may_point_to_spin_lock(reg)) {
6544 /* For not-NULL ptr, reg->ref_obj_id will be reset
6545 * in release_reg_references().
6546 *
6547 * reg->id is still used by spin_lock ptr. Other
6548 * than spin_lock ptr type, reg->id can be reset.
6549 */
6550 reg->id = 0;
6551 }
6552 }
6553 }
6554
6555 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
6556 bool is_null)
6557 {
6558 struct bpf_reg_state *reg;
6559 int i;
6560
6561 for (i = 0; i < MAX_BPF_REG; i++)
6562 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
6563
6564 bpf_for_each_spilled_reg(i, state, reg) {
6565 if (!reg)
6566 continue;
6567 mark_ptr_or_null_reg(state, reg, id, is_null);
6568 }
6569 }
6570
6571 /* The logic is similar to find_good_pkt_pointers(), both could eventually
6572 * be folded together at some point.
6573 */
6574 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
6575 bool is_null)
6576 {
6577 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6578 struct bpf_reg_state *regs = state->regs;
6579 u32 ref_obj_id = regs[regno].ref_obj_id;
6580 u32 id = regs[regno].id;
6581 int i;
6582
6583 if (ref_obj_id && ref_obj_id == id && is_null)
6584 /* regs[regno] is in the " == NULL" branch.
6585 * No one could have freed the reference state before
6586 * doing the NULL check.
6587 */
6588 WARN_ON_ONCE(release_reference_state(state, id));
6589
6590 for (i = 0; i <= vstate->curframe; i++)
6591 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
6592 }
6593
6594 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
6595 struct bpf_reg_state *dst_reg,
6596 struct bpf_reg_state *src_reg,
6597 struct bpf_verifier_state *this_branch,
6598 struct bpf_verifier_state *other_branch)
6599 {
6600 if (BPF_SRC(insn->code) != BPF_X)
6601 return false;
6602
6603 /* Pointers are always 64-bit. */
6604 if (BPF_CLASS(insn->code) == BPF_JMP32)
6605 return false;
6606
6607 switch (BPF_OP(insn->code)) {
6608 case BPF_JGT:
6609 if ((dst_reg->type == PTR_TO_PACKET &&
6610 src_reg->type == PTR_TO_PACKET_END) ||
6611 (dst_reg->type == PTR_TO_PACKET_META &&
6612 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6613 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
6614 find_good_pkt_pointers(this_branch, dst_reg,
6615 dst_reg->type, false);
6616 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6617 src_reg->type == PTR_TO_PACKET) ||
6618 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6619 src_reg->type == PTR_TO_PACKET_META)) {
6620 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
6621 find_good_pkt_pointers(other_branch, src_reg,
6622 src_reg->type, true);
6623 } else {
6624 return false;
6625 }
6626 break;
6627 case BPF_JLT:
6628 if ((dst_reg->type == PTR_TO_PACKET &&
6629 src_reg->type == PTR_TO_PACKET_END) ||
6630 (dst_reg->type == PTR_TO_PACKET_META &&
6631 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6632 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
6633 find_good_pkt_pointers(other_branch, dst_reg,
6634 dst_reg->type, true);
6635 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6636 src_reg->type == PTR_TO_PACKET) ||
6637 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6638 src_reg->type == PTR_TO_PACKET_META)) {
6639 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
6640 find_good_pkt_pointers(this_branch, src_reg,
6641 src_reg->type, false);
6642 } else {
6643 return false;
6644 }
6645 break;
6646 case BPF_JGE:
6647 if ((dst_reg->type == PTR_TO_PACKET &&
6648 src_reg->type == PTR_TO_PACKET_END) ||
6649 (dst_reg->type == PTR_TO_PACKET_META &&
6650 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6651 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
6652 find_good_pkt_pointers(this_branch, dst_reg,
6653 dst_reg->type, true);
6654 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6655 src_reg->type == PTR_TO_PACKET) ||
6656 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6657 src_reg->type == PTR_TO_PACKET_META)) {
6658 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
6659 find_good_pkt_pointers(other_branch, src_reg,
6660 src_reg->type, false);
6661 } else {
6662 return false;
6663 }
6664 break;
6665 case BPF_JLE:
6666 if ((dst_reg->type == PTR_TO_PACKET &&
6667 src_reg->type == PTR_TO_PACKET_END) ||
6668 (dst_reg->type == PTR_TO_PACKET_META &&
6669 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6670 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
6671 find_good_pkt_pointers(other_branch, dst_reg,
6672 dst_reg->type, false);
6673 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6674 src_reg->type == PTR_TO_PACKET) ||
6675 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6676 src_reg->type == PTR_TO_PACKET_META)) {
6677 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
6678 find_good_pkt_pointers(this_branch, src_reg,
6679 src_reg->type, true);
6680 } else {
6681 return false;
6682 }
6683 break;
6684 default:
6685 return false;
6686 }
6687
6688 return true;
6689 }
6690
6691 static int check_cond_jmp_op(struct bpf_verifier_env *env,
6692 struct bpf_insn *insn, int *insn_idx)
6693 {
6694 struct bpf_verifier_state *this_branch = env->cur_state;
6695 struct bpf_verifier_state *other_branch;
6696 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
6697 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
6698 u8 opcode = BPF_OP(insn->code);
6699 bool is_jmp32;
6700 int pred = -1;
6701 int err;
6702
6703 /* Only conditional jumps are expected to reach here. */
6704 if (opcode == BPF_JA || opcode > BPF_JSLE) {
6705 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
6706 return -EINVAL;
6707 }
6708
6709 if (BPF_SRC(insn->code) == BPF_X) {
6710 if (insn->imm != 0) {
6711 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6712 return -EINVAL;
6713 }
6714
6715 /* check src1 operand */
6716 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6717 if (err)
6718 return err;
6719
6720 if (is_pointer_value(env, insn->src_reg)) {
6721 verbose(env, "R%d pointer comparison prohibited\n",
6722 insn->src_reg);
6723 return -EACCES;
6724 }
6725 src_reg = &regs[insn->src_reg];
6726 } else {
6727 if (insn->src_reg != BPF_REG_0) {
6728 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6729 return -EINVAL;
6730 }
6731 }
6732
6733 /* check src2 operand */
6734 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6735 if (err)
6736 return err;
6737
6738 dst_reg = &regs[insn->dst_reg];
6739 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
6740
6741 if (BPF_SRC(insn->code) == BPF_K) {
6742 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
6743 } else if (src_reg->type == SCALAR_VALUE &&
6744 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
6745 pred = is_branch_taken(dst_reg,
6746 tnum_subreg(src_reg->var_off).value,
6747 opcode,
6748 is_jmp32);
6749 } else if (src_reg->type == SCALAR_VALUE &&
6750 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
6751 pred = is_branch_taken(dst_reg,
6752 src_reg->var_off.value,
6753 opcode,
6754 is_jmp32);
6755 }
6756
6757 if (pred >= 0) {
6758 err = mark_chain_precision(env, insn->dst_reg);
6759 if (BPF_SRC(insn->code) == BPF_X && !err)
6760 err = mark_chain_precision(env, insn->src_reg);
6761 if (err)
6762 return err;
6763 }
6764 if (pred == 1) {
6765 /* only follow the goto, ignore fall-through */
6766 *insn_idx += insn->off;
6767 return 0;
6768 } else if (pred == 0) {
6769 /* only follow fall-through branch, since
6770 * that's where the program will go
6771 */
6772 return 0;
6773 }
6774
6775 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
6776 false);
6777 if (!other_branch)
6778 return -EFAULT;
6779 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
6780
6781 /* detect if we are comparing against a constant value so we can adjust
6782 * our min/max values for our dst register.
6783 * this is only legit if both are scalars (or pointers to the same
6784 * object, I suppose, but we don't support that right now), because
6785 * otherwise the different base pointers mean the offsets aren't
6786 * comparable.
6787 */
6788 if (BPF_SRC(insn->code) == BPF_X) {
6789 struct bpf_reg_state *src_reg = &regs[insn->src_reg];
6790
6791 if (dst_reg->type == SCALAR_VALUE &&
6792 src_reg->type == SCALAR_VALUE) {
6793 if (tnum_is_const(src_reg->var_off) ||
6794 (is_jmp32 &&
6795 tnum_is_const(tnum_subreg(src_reg->var_off))))
6796 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6797 dst_reg,
6798 src_reg->var_off.value,
6799 tnum_subreg(src_reg->var_off).value,
6800 opcode, is_jmp32);
6801 else if (tnum_is_const(dst_reg->var_off) ||
6802 (is_jmp32 &&
6803 tnum_is_const(tnum_subreg(dst_reg->var_off))))
6804 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
6805 src_reg,
6806 dst_reg->var_off.value,
6807 tnum_subreg(dst_reg->var_off).value,
6808 opcode, is_jmp32);
6809 else if (!is_jmp32 &&
6810 (opcode == BPF_JEQ || opcode == BPF_JNE))
6811 /* Comparing for equality, we can combine knowledge */
6812 reg_combine_min_max(&other_branch_regs[insn->src_reg],
6813 &other_branch_regs[insn->dst_reg],
6814 src_reg, dst_reg, opcode);
6815 }
6816 } else if (dst_reg->type == SCALAR_VALUE) {
6817 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6818 dst_reg, insn->imm, (u32)insn->imm,
6819 opcode, is_jmp32);
6820 }
6821
6822 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
6823 * NOTE: these optimizations below are related with pointer comparison
6824 * which will never be JMP32.
6825 */
6826 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
6827 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
6828 reg_type_may_be_null(dst_reg->type)) {
6829 /* Mark all identical registers in each branch as either
6830 * safe or unknown depending R == 0 or R != 0 conditional.
6831 */
6832 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
6833 opcode == BPF_JNE);
6834 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
6835 opcode == BPF_JEQ);
6836 } else if (!try_match_pkt_pointers(insn, dst_reg, &regs[insn->src_reg],
6837 this_branch, other_branch) &&
6838 is_pointer_value(env, insn->dst_reg)) {
6839 verbose(env, "R%d pointer comparison prohibited\n",
6840 insn->dst_reg);
6841 return -EACCES;
6842 }
6843 if (env->log.level & BPF_LOG_LEVEL)
6844 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
6845 return 0;
6846 }
6847
6848 /* verify BPF_LD_IMM64 instruction */
6849 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
6850 {
6851 struct bpf_insn_aux_data *aux = cur_aux(env);
6852 struct bpf_reg_state *regs = cur_regs(env);
6853 struct bpf_map *map;
6854 int err;
6855
6856 if (BPF_SIZE(insn->code) != BPF_DW) {
6857 verbose(env, "invalid BPF_LD_IMM insn\n");
6858 return -EINVAL;
6859 }
6860 if (insn->off != 0) {
6861 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
6862 return -EINVAL;
6863 }
6864
6865 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6866 if (err)
6867 return err;
6868
6869 if (insn->src_reg == 0) {
6870 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
6871
6872 regs[insn->dst_reg].type = SCALAR_VALUE;
6873 __mark_reg_known(&regs[insn->dst_reg], imm);
6874 return 0;
6875 }
6876
6877 map = env->used_maps[aux->map_index];
6878 mark_reg_known_zero(env, regs, insn->dst_reg);
6879 regs[insn->dst_reg].map_ptr = map;
6880
6881 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
6882 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
6883 regs[insn->dst_reg].off = aux->map_off;
6884 if (map_value_has_spin_lock(map))
6885 regs[insn->dst_reg].id = ++env->id_gen;
6886 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
6887 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
6888 } else {
6889 verbose(env, "bpf verifier is misconfigured\n");
6890 return -EINVAL;
6891 }
6892
6893 return 0;
6894 }
6895
6896 static bool may_access_skb(enum bpf_prog_type type)
6897 {
6898 switch (type) {
6899 case BPF_PROG_TYPE_SOCKET_FILTER:
6900 case BPF_PROG_TYPE_SCHED_CLS:
6901 case BPF_PROG_TYPE_SCHED_ACT:
6902 return true;
6903 default:
6904 return false;
6905 }
6906 }
6907
6908 /* verify safety of LD_ABS|LD_IND instructions:
6909 * - they can only appear in the programs where ctx == skb
6910 * - since they are wrappers of function calls, they scratch R1-R5 registers,
6911 * preserve R6-R9, and store return value into R0
6912 *
6913 * Implicit input:
6914 * ctx == skb == R6 == CTX
6915 *
6916 * Explicit input:
6917 * SRC == any register
6918 * IMM == 32-bit immediate
6919 *
6920 * Output:
6921 * R0 - 8/16/32-bit skb data converted to cpu endianness
6922 */
6923 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
6924 {
6925 struct bpf_reg_state *regs = cur_regs(env);
6926 static const int ctx_reg = BPF_REG_6;
6927 u8 mode = BPF_MODE(insn->code);
6928 int i, err;
6929
6930 if (!may_access_skb(env->prog->type)) {
6931 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
6932 return -EINVAL;
6933 }
6934
6935 if (!env->ops->gen_ld_abs) {
6936 verbose(env, "bpf verifier is misconfigured\n");
6937 return -EINVAL;
6938 }
6939
6940 if (env->subprog_cnt > 1) {
6941 /* when program has LD_ABS insn JITs and interpreter assume
6942 * that r1 == ctx == skb which is not the case for callees
6943 * that can have arbitrary arguments. It's problematic
6944 * for main prog as well since JITs would need to analyze
6945 * all functions in order to make proper register save/restore
6946 * decisions in the main prog. Hence disallow LD_ABS with calls
6947 */
6948 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
6949 return -EINVAL;
6950 }
6951
6952 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
6953 BPF_SIZE(insn->code) == BPF_DW ||
6954 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
6955 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
6956 return -EINVAL;
6957 }
6958
6959 /* check whether implicit source operand (register R6) is readable */
6960 err = check_reg_arg(env, ctx_reg, SRC_OP);
6961 if (err)
6962 return err;
6963
6964 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
6965 * gen_ld_abs() may terminate the program at runtime, leading to
6966 * reference leak.
6967 */
6968 err = check_reference_leak(env);
6969 if (err) {
6970 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
6971 return err;
6972 }
6973
6974 if (env->cur_state->active_spin_lock) {
6975 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
6976 return -EINVAL;
6977 }
6978
6979 if (regs[ctx_reg].type != PTR_TO_CTX) {
6980 verbose(env,
6981 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
6982 return -EINVAL;
6983 }
6984
6985 if (mode == BPF_IND) {
6986 /* check explicit source operand */
6987 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6988 if (err)
6989 return err;
6990 }
6991
6992 err = check_ctx_reg(env, &regs[ctx_reg], ctx_reg);
6993 if (err < 0)
6994 return err;
6995
6996 /* reset caller saved regs to unreadable */
6997 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6998 mark_reg_not_init(env, regs, caller_saved[i]);
6999 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7000 }
7001
7002 /* mark destination R0 register as readable, since it contains
7003 * the value fetched from the packet.
7004 * Already marked as written above.
7005 */
7006 mark_reg_unknown(env, regs, BPF_REG_0);
7007 /* ld_abs load up to 32-bit skb data. */
7008 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
7009 return 0;
7010 }
7011
7012 static int check_return_code(struct bpf_verifier_env *env)
7013 {
7014 struct tnum enforce_attach_type_range = tnum_unknown;
7015 const struct bpf_prog *prog = env->prog;
7016 struct bpf_reg_state *reg;
7017 struct tnum range = tnum_range(0, 1);
7018 int err;
7019
7020 /* LSM and struct_ops func-ptr's return type could be "void" */
7021 if ((env->prog->type == BPF_PROG_TYPE_STRUCT_OPS ||
7022 env->prog->type == BPF_PROG_TYPE_LSM) &&
7023 !prog->aux->attach_func_proto->type)
7024 return 0;
7025
7026 /* eBPF calling convetion is such that R0 is used
7027 * to return the value from eBPF program.
7028 * Make sure that it's readable at this time
7029 * of bpf_exit, which means that program wrote
7030 * something into it earlier
7031 */
7032 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
7033 if (err)
7034 return err;
7035
7036 if (is_pointer_value(env, BPF_REG_0)) {
7037 verbose(env, "R0 leaks addr as return value\n");
7038 return -EACCES;
7039 }
7040
7041 switch (env->prog->type) {
7042 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
7043 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
7044 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG)
7045 range = tnum_range(1, 1);
7046 break;
7047 case BPF_PROG_TYPE_CGROUP_SKB:
7048 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
7049 range = tnum_range(0, 3);
7050 enforce_attach_type_range = tnum_range(2, 3);
7051 }
7052 break;
7053 case BPF_PROG_TYPE_CGROUP_SOCK:
7054 case BPF_PROG_TYPE_SOCK_OPS:
7055 case BPF_PROG_TYPE_CGROUP_DEVICE:
7056 case BPF_PROG_TYPE_CGROUP_SYSCTL:
7057 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
7058 break;
7059 case BPF_PROG_TYPE_RAW_TRACEPOINT:
7060 if (!env->prog->aux->attach_btf_id)
7061 return 0;
7062 range = tnum_const(0);
7063 break;
7064 case BPF_PROG_TYPE_TRACING:
7065 switch (env->prog->expected_attach_type) {
7066 case BPF_TRACE_FENTRY:
7067 case BPF_TRACE_FEXIT:
7068 range = tnum_const(0);
7069 break;
7070 case BPF_TRACE_RAW_TP:
7071 case BPF_MODIFY_RETURN:
7072 return 0;
7073 default:
7074 return -ENOTSUPP;
7075 }
7076 break;
7077 case BPF_PROG_TYPE_EXT:
7078 /* freplace program can return anything as its return value
7079 * depends on the to-be-replaced kernel func or bpf program.
7080 */
7081 default:
7082 return 0;
7083 }
7084
7085 reg = cur_regs(env) + BPF_REG_0;
7086 if (reg->type != SCALAR_VALUE) {
7087 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
7088 reg_type_str[reg->type]);
7089 return -EINVAL;
7090 }
7091
7092 if (!tnum_in(range, reg->var_off)) {
7093 char tn_buf[48];
7094
7095 verbose(env, "At program exit the register R0 ");
7096 if (!tnum_is_unknown(reg->var_off)) {
7097 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7098 verbose(env, "has value %s", tn_buf);
7099 } else {
7100 verbose(env, "has unknown scalar value");
7101 }
7102 tnum_strn(tn_buf, sizeof(tn_buf), range);
7103 verbose(env, " should have been in %s\n", tn_buf);
7104 return -EINVAL;
7105 }
7106
7107 if (!tnum_is_unknown(enforce_attach_type_range) &&
7108 tnum_in(enforce_attach_type_range, reg->var_off))
7109 env->prog->enforce_expected_attach_type = 1;
7110 return 0;
7111 }
7112
7113 /* non-recursive DFS pseudo code
7114 * 1 procedure DFS-iterative(G,v):
7115 * 2 label v as discovered
7116 * 3 let S be a stack
7117 * 4 S.push(v)
7118 * 5 while S is not empty
7119 * 6 t <- S.pop()
7120 * 7 if t is what we're looking for:
7121 * 8 return t
7122 * 9 for all edges e in G.adjacentEdges(t) do
7123 * 10 if edge e is already labelled
7124 * 11 continue with the next edge
7125 * 12 w <- G.adjacentVertex(t,e)
7126 * 13 if vertex w is not discovered and not explored
7127 * 14 label e as tree-edge
7128 * 15 label w as discovered
7129 * 16 S.push(w)
7130 * 17 continue at 5
7131 * 18 else if vertex w is discovered
7132 * 19 label e as back-edge
7133 * 20 else
7134 * 21 // vertex w is explored
7135 * 22 label e as forward- or cross-edge
7136 * 23 label t as explored
7137 * 24 S.pop()
7138 *
7139 * convention:
7140 * 0x10 - discovered
7141 * 0x11 - discovered and fall-through edge labelled
7142 * 0x12 - discovered and fall-through and branch edges labelled
7143 * 0x20 - explored
7144 */
7145
7146 enum {
7147 DISCOVERED = 0x10,
7148 EXPLORED = 0x20,
7149 FALLTHROUGH = 1,
7150 BRANCH = 2,
7151 };
7152
7153 static u32 state_htab_size(struct bpf_verifier_env *env)
7154 {
7155 return env->prog->len;
7156 }
7157
7158 static struct bpf_verifier_state_list **explored_state(
7159 struct bpf_verifier_env *env,
7160 int idx)
7161 {
7162 struct bpf_verifier_state *cur = env->cur_state;
7163 struct bpf_func_state *state = cur->frame[cur->curframe];
7164
7165 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
7166 }
7167
7168 static void init_explored_state(struct bpf_verifier_env *env, int idx)
7169 {
7170 env->insn_aux_data[idx].prune_point = true;
7171 }
7172
7173 /* t, w, e - match pseudo-code above:
7174 * t - index of current instruction
7175 * w - next instruction
7176 * e - edge
7177 */
7178 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
7179 bool loop_ok)
7180 {
7181 int *insn_stack = env->cfg.insn_stack;
7182 int *insn_state = env->cfg.insn_state;
7183
7184 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
7185 return 0;
7186
7187 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
7188 return 0;
7189
7190 if (w < 0 || w >= env->prog->len) {
7191 verbose_linfo(env, t, "%d: ", t);
7192 verbose(env, "jump out of range from insn %d to %d\n", t, w);
7193 return -EINVAL;
7194 }
7195
7196 if (e == BRANCH)
7197 /* mark branch target for state pruning */
7198 init_explored_state(env, w);
7199
7200 if (insn_state[w] == 0) {
7201 /* tree-edge */
7202 insn_state[t] = DISCOVERED | e;
7203 insn_state[w] = DISCOVERED;
7204 if (env->cfg.cur_stack >= env->prog->len)
7205 return -E2BIG;
7206 insn_stack[env->cfg.cur_stack++] = w;
7207 return 1;
7208 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
7209 if (loop_ok && env->allow_ptr_leaks)
7210 return 0;
7211 verbose_linfo(env, t, "%d: ", t);
7212 verbose_linfo(env, w, "%d: ", w);
7213 verbose(env, "back-edge from insn %d to %d\n", t, w);
7214 return -EINVAL;
7215 } else if (insn_state[w] == EXPLORED) {
7216 /* forward- or cross-edge */
7217 insn_state[t] = DISCOVERED | e;
7218 } else {
7219 verbose(env, "insn state internal bug\n");
7220 return -EFAULT;
7221 }
7222 return 0;
7223 }
7224
7225 /* non-recursive depth-first-search to detect loops in BPF program
7226 * loop == back-edge in directed graph
7227 */
7228 static int check_cfg(struct bpf_verifier_env *env)
7229 {
7230 struct bpf_insn *insns = env->prog->insnsi;
7231 int insn_cnt = env->prog->len;
7232 int *insn_stack, *insn_state;
7233 int ret = 0;
7234 int i, t;
7235
7236 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7237 if (!insn_state)
7238 return -ENOMEM;
7239
7240 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7241 if (!insn_stack) {
7242 kvfree(insn_state);
7243 return -ENOMEM;
7244 }
7245
7246 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
7247 insn_stack[0] = 0; /* 0 is the first instruction */
7248 env->cfg.cur_stack = 1;
7249
7250 peek_stack:
7251 if (env->cfg.cur_stack == 0)
7252 goto check_state;
7253 t = insn_stack[env->cfg.cur_stack - 1];
7254
7255 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
7256 BPF_CLASS(insns[t].code) == BPF_JMP32) {
7257 u8 opcode = BPF_OP(insns[t].code);
7258
7259 if (opcode == BPF_EXIT) {
7260 goto mark_explored;
7261 } else if (opcode == BPF_CALL) {
7262 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7263 if (ret == 1)
7264 goto peek_stack;
7265 else if (ret < 0)
7266 goto err_free;
7267 if (t + 1 < insn_cnt)
7268 init_explored_state(env, t + 1);
7269 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
7270 init_explored_state(env, t);
7271 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
7272 env, false);
7273 if (ret == 1)
7274 goto peek_stack;
7275 else if (ret < 0)
7276 goto err_free;
7277 }
7278 } else if (opcode == BPF_JA) {
7279 if (BPF_SRC(insns[t].code) != BPF_K) {
7280 ret = -EINVAL;
7281 goto err_free;
7282 }
7283 /* unconditional jump with single edge */
7284 ret = push_insn(t, t + insns[t].off + 1,
7285 FALLTHROUGH, env, true);
7286 if (ret == 1)
7287 goto peek_stack;
7288 else if (ret < 0)
7289 goto err_free;
7290 /* unconditional jmp is not a good pruning point,
7291 * but it's marked, since backtracking needs
7292 * to record jmp history in is_state_visited().
7293 */
7294 init_explored_state(env, t + insns[t].off + 1);
7295 /* tell verifier to check for equivalent states
7296 * after every call and jump
7297 */
7298 if (t + 1 < insn_cnt)
7299 init_explored_state(env, t + 1);
7300 } else {
7301 /* conditional jump with two edges */
7302 init_explored_state(env, t);
7303 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
7304 if (ret == 1)
7305 goto peek_stack;
7306 else if (ret < 0)
7307 goto err_free;
7308
7309 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
7310 if (ret == 1)
7311 goto peek_stack;
7312 else if (ret < 0)
7313 goto err_free;
7314 }
7315 } else {
7316 /* all other non-branch instructions with single
7317 * fall-through edge
7318 */
7319 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7320 if (ret == 1)
7321 goto peek_stack;
7322 else if (ret < 0)
7323 goto err_free;
7324 }
7325
7326 mark_explored:
7327 insn_state[t] = EXPLORED;
7328 if (env->cfg.cur_stack-- <= 0) {
7329 verbose(env, "pop stack internal bug\n");
7330 ret = -EFAULT;
7331 goto err_free;
7332 }
7333 goto peek_stack;
7334
7335 check_state:
7336 for (i = 0; i < insn_cnt; i++) {
7337 if (insn_state[i] != EXPLORED) {
7338 verbose(env, "unreachable insn %d\n", i);
7339 ret = -EINVAL;
7340 goto err_free;
7341 }
7342 }
7343 ret = 0; /* cfg looks good */
7344
7345 err_free:
7346 kvfree(insn_state);
7347 kvfree(insn_stack);
7348 env->cfg.insn_state = env->cfg.insn_stack = NULL;
7349 return ret;
7350 }
7351
7352 /* The minimum supported BTF func info size */
7353 #define MIN_BPF_FUNCINFO_SIZE 8
7354 #define MAX_FUNCINFO_REC_SIZE 252
7355
7356 static int check_btf_func(struct bpf_verifier_env *env,
7357 const union bpf_attr *attr,
7358 union bpf_attr __user *uattr)
7359 {
7360 u32 i, nfuncs, urec_size, min_size;
7361 u32 krec_size = sizeof(struct bpf_func_info);
7362 struct bpf_func_info *krecord;
7363 struct bpf_func_info_aux *info_aux = NULL;
7364 const struct btf_type *type;
7365 struct bpf_prog *prog;
7366 const struct btf *btf;
7367 void __user *urecord;
7368 u32 prev_offset = 0;
7369 int ret = 0;
7370
7371 nfuncs = attr->func_info_cnt;
7372 if (!nfuncs)
7373 return 0;
7374
7375 if (nfuncs != env->subprog_cnt) {
7376 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
7377 return -EINVAL;
7378 }
7379
7380 urec_size = attr->func_info_rec_size;
7381 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
7382 urec_size > MAX_FUNCINFO_REC_SIZE ||
7383 urec_size % sizeof(u32)) {
7384 verbose(env, "invalid func info rec size %u\n", urec_size);
7385 return -EINVAL;
7386 }
7387
7388 prog = env->prog;
7389 btf = prog->aux->btf;
7390
7391 urecord = u64_to_user_ptr(attr->func_info);
7392 min_size = min_t(u32, krec_size, urec_size);
7393
7394 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
7395 if (!krecord)
7396 return -ENOMEM;
7397 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
7398 if (!info_aux)
7399 goto err_free;
7400
7401 for (i = 0; i < nfuncs; i++) {
7402 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
7403 if (ret) {
7404 if (ret == -E2BIG) {
7405 verbose(env, "nonzero tailing record in func info");
7406 /* set the size kernel expects so loader can zero
7407 * out the rest of the record.
7408 */
7409 if (put_user(min_size, &uattr->func_info_rec_size))
7410 ret = -EFAULT;
7411 }
7412 goto err_free;
7413 }
7414
7415 if (copy_from_user(&krecord[i], urecord, min_size)) {
7416 ret = -EFAULT;
7417 goto err_free;
7418 }
7419
7420 /* check insn_off */
7421 if (i == 0) {
7422 if (krecord[i].insn_off) {
7423 verbose(env,
7424 "nonzero insn_off %u for the first func info record",
7425 krecord[i].insn_off);
7426 ret = -EINVAL;
7427 goto err_free;
7428 }
7429 } else if (krecord[i].insn_off <= prev_offset) {
7430 verbose(env,
7431 "same or smaller insn offset (%u) than previous func info record (%u)",
7432 krecord[i].insn_off, prev_offset);
7433 ret = -EINVAL;
7434 goto err_free;
7435 }
7436
7437 if (env->subprog_info[i].start != krecord[i].insn_off) {
7438 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
7439 ret = -EINVAL;
7440 goto err_free;
7441 }
7442
7443 /* check type_id */
7444 type = btf_type_by_id(btf, krecord[i].type_id);
7445 if (!type || !btf_type_is_func(type)) {
7446 verbose(env, "invalid type id %d in func info",
7447 krecord[i].type_id);
7448 ret = -EINVAL;
7449 goto err_free;
7450 }
7451 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
7452 prev_offset = krecord[i].insn_off;
7453 urecord += urec_size;
7454 }
7455
7456 prog->aux->func_info = krecord;
7457 prog->aux->func_info_cnt = nfuncs;
7458 prog->aux->func_info_aux = info_aux;
7459 return 0;
7460
7461 err_free:
7462 kvfree(krecord);
7463 kfree(info_aux);
7464 return ret;
7465 }
7466
7467 static void adjust_btf_func(struct bpf_verifier_env *env)
7468 {
7469 struct bpf_prog_aux *aux = env->prog->aux;
7470 int i;
7471
7472 if (!aux->func_info)
7473 return;
7474
7475 for (i = 0; i < env->subprog_cnt; i++)
7476 aux->func_info[i].insn_off = env->subprog_info[i].start;
7477 }
7478
7479 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
7480 sizeof(((struct bpf_line_info *)(0))->line_col))
7481 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
7482
7483 static int check_btf_line(struct bpf_verifier_env *env,
7484 const union bpf_attr *attr,
7485 union bpf_attr __user *uattr)
7486 {
7487 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
7488 struct bpf_subprog_info *sub;
7489 struct bpf_line_info *linfo;
7490 struct bpf_prog *prog;
7491 const struct btf *btf;
7492 void __user *ulinfo;
7493 int err;
7494
7495 nr_linfo = attr->line_info_cnt;
7496 if (!nr_linfo)
7497 return 0;
7498
7499 rec_size = attr->line_info_rec_size;
7500 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
7501 rec_size > MAX_LINEINFO_REC_SIZE ||
7502 rec_size & (sizeof(u32) - 1))
7503 return -EINVAL;
7504
7505 /* Need to zero it in case the userspace may
7506 * pass in a smaller bpf_line_info object.
7507 */
7508 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
7509 GFP_KERNEL | __GFP_NOWARN);
7510 if (!linfo)
7511 return -ENOMEM;
7512
7513 prog = env->prog;
7514 btf = prog->aux->btf;
7515
7516 s = 0;
7517 sub = env->subprog_info;
7518 ulinfo = u64_to_user_ptr(attr->line_info);
7519 expected_size = sizeof(struct bpf_line_info);
7520 ncopy = min_t(u32, expected_size, rec_size);
7521 for (i = 0; i < nr_linfo; i++) {
7522 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
7523 if (err) {
7524 if (err == -E2BIG) {
7525 verbose(env, "nonzero tailing record in line_info");
7526 if (put_user(expected_size,
7527 &uattr->line_info_rec_size))
7528 err = -EFAULT;
7529 }
7530 goto err_free;
7531 }
7532
7533 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
7534 err = -EFAULT;
7535 goto err_free;
7536 }
7537
7538 /*
7539 * Check insn_off to ensure
7540 * 1) strictly increasing AND
7541 * 2) bounded by prog->len
7542 *
7543 * The linfo[0].insn_off == 0 check logically falls into
7544 * the later "missing bpf_line_info for func..." case
7545 * because the first linfo[0].insn_off must be the
7546 * first sub also and the first sub must have
7547 * subprog_info[0].start == 0.
7548 */
7549 if ((i && linfo[i].insn_off <= prev_offset) ||
7550 linfo[i].insn_off >= prog->len) {
7551 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
7552 i, linfo[i].insn_off, prev_offset,
7553 prog->len);
7554 err = -EINVAL;
7555 goto err_free;
7556 }
7557
7558 if (!prog->insnsi[linfo[i].insn_off].code) {
7559 verbose(env,
7560 "Invalid insn code at line_info[%u].insn_off\n",
7561 i);
7562 err = -EINVAL;
7563 goto err_free;
7564 }
7565
7566 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
7567 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
7568 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
7569 err = -EINVAL;
7570 goto err_free;
7571 }
7572
7573 if (s != env->subprog_cnt) {
7574 if (linfo[i].insn_off == sub[s].start) {
7575 sub[s].linfo_idx = i;
7576 s++;
7577 } else if (sub[s].start < linfo[i].insn_off) {
7578 verbose(env, "missing bpf_line_info for func#%u\n", s);
7579 err = -EINVAL;
7580 goto err_free;
7581 }
7582 }
7583
7584 prev_offset = linfo[i].insn_off;
7585 ulinfo += rec_size;
7586 }
7587
7588 if (s != env->subprog_cnt) {
7589 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
7590 env->subprog_cnt - s, s);
7591 err = -EINVAL;
7592 goto err_free;
7593 }
7594
7595 prog->aux->linfo = linfo;
7596 prog->aux->nr_linfo = nr_linfo;
7597
7598 return 0;
7599
7600 err_free:
7601 kvfree(linfo);
7602 return err;
7603 }
7604
7605 static int check_btf_info(struct bpf_verifier_env *env,
7606 const union bpf_attr *attr,
7607 union bpf_attr __user *uattr)
7608 {
7609 struct btf *btf;
7610 int err;
7611
7612 if (!attr->func_info_cnt && !attr->line_info_cnt)
7613 return 0;
7614
7615 btf = btf_get_by_fd(attr->prog_btf_fd);
7616 if (IS_ERR(btf))
7617 return PTR_ERR(btf);
7618 env->prog->aux->btf = btf;
7619
7620 err = check_btf_func(env, attr, uattr);
7621 if (err)
7622 return err;
7623
7624 err = check_btf_line(env, attr, uattr);
7625 if (err)
7626 return err;
7627
7628 return 0;
7629 }
7630
7631 /* check %cur's range satisfies %old's */
7632 static bool range_within(struct bpf_reg_state *old,
7633 struct bpf_reg_state *cur)
7634 {
7635 return old->umin_value <= cur->umin_value &&
7636 old->umax_value >= cur->umax_value &&
7637 old->smin_value <= cur->smin_value &&
7638 old->smax_value >= cur->smax_value;
7639 }
7640
7641 /* Maximum number of register states that can exist at once */
7642 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
7643 struct idpair {
7644 u32 old;
7645 u32 cur;
7646 };
7647
7648 /* If in the old state two registers had the same id, then they need to have
7649 * the same id in the new state as well. But that id could be different from
7650 * the old state, so we need to track the mapping from old to new ids.
7651 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
7652 * regs with old id 5 must also have new id 9 for the new state to be safe. But
7653 * regs with a different old id could still have new id 9, we don't care about
7654 * that.
7655 * So we look through our idmap to see if this old id has been seen before. If
7656 * so, we require the new id to match; otherwise, we add the id pair to the map.
7657 */
7658 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
7659 {
7660 unsigned int i;
7661
7662 for (i = 0; i < ID_MAP_SIZE; i++) {
7663 if (!idmap[i].old) {
7664 /* Reached an empty slot; haven't seen this id before */
7665 idmap[i].old = old_id;
7666 idmap[i].cur = cur_id;
7667 return true;
7668 }
7669 if (idmap[i].old == old_id)
7670 return idmap[i].cur == cur_id;
7671 }
7672 /* We ran out of idmap slots, which should be impossible */
7673 WARN_ON_ONCE(1);
7674 return false;
7675 }
7676
7677 static void clean_func_state(struct bpf_verifier_env *env,
7678 struct bpf_func_state *st)
7679 {
7680 enum bpf_reg_liveness live;
7681 int i, j;
7682
7683 for (i = 0; i < BPF_REG_FP; i++) {
7684 live = st->regs[i].live;
7685 /* liveness must not touch this register anymore */
7686 st->regs[i].live |= REG_LIVE_DONE;
7687 if (!(live & REG_LIVE_READ))
7688 /* since the register is unused, clear its state
7689 * to make further comparison simpler
7690 */
7691 __mark_reg_not_init(env, &st->regs[i]);
7692 }
7693
7694 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
7695 live = st->stack[i].spilled_ptr.live;
7696 /* liveness must not touch this stack slot anymore */
7697 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
7698 if (!(live & REG_LIVE_READ)) {
7699 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
7700 for (j = 0; j < BPF_REG_SIZE; j++)
7701 st->stack[i].slot_type[j] = STACK_INVALID;
7702 }
7703 }
7704 }
7705
7706 static void clean_verifier_state(struct bpf_verifier_env *env,
7707 struct bpf_verifier_state *st)
7708 {
7709 int i;
7710
7711 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
7712 /* all regs in this state in all frames were already marked */
7713 return;
7714
7715 for (i = 0; i <= st->curframe; i++)
7716 clean_func_state(env, st->frame[i]);
7717 }
7718
7719 /* the parentage chains form a tree.
7720 * the verifier states are added to state lists at given insn and
7721 * pushed into state stack for future exploration.
7722 * when the verifier reaches bpf_exit insn some of the verifer states
7723 * stored in the state lists have their final liveness state already,
7724 * but a lot of states will get revised from liveness point of view when
7725 * the verifier explores other branches.
7726 * Example:
7727 * 1: r0 = 1
7728 * 2: if r1 == 100 goto pc+1
7729 * 3: r0 = 2
7730 * 4: exit
7731 * when the verifier reaches exit insn the register r0 in the state list of
7732 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
7733 * of insn 2 and goes exploring further. At the insn 4 it will walk the
7734 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
7735 *
7736 * Since the verifier pushes the branch states as it sees them while exploring
7737 * the program the condition of walking the branch instruction for the second
7738 * time means that all states below this branch were already explored and
7739 * their final liveness markes are already propagated.
7740 * Hence when the verifier completes the search of state list in is_state_visited()
7741 * we can call this clean_live_states() function to mark all liveness states
7742 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
7743 * will not be used.
7744 * This function also clears the registers and stack for states that !READ
7745 * to simplify state merging.
7746 *
7747 * Important note here that walking the same branch instruction in the callee
7748 * doesn't meant that the states are DONE. The verifier has to compare
7749 * the callsites
7750 */
7751 static void clean_live_states(struct bpf_verifier_env *env, int insn,
7752 struct bpf_verifier_state *cur)
7753 {
7754 struct bpf_verifier_state_list *sl;
7755 int i;
7756
7757 sl = *explored_state(env, insn);
7758 while (sl) {
7759 if (sl->state.branches)
7760 goto next;
7761 if (sl->state.insn_idx != insn ||
7762 sl->state.curframe != cur->curframe)
7763 goto next;
7764 for (i = 0; i <= cur->curframe; i++)
7765 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
7766 goto next;
7767 clean_verifier_state(env, &sl->state);
7768 next:
7769 sl = sl->next;
7770 }
7771 }
7772
7773 /* Returns true if (rold safe implies rcur safe) */
7774 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
7775 struct idpair *idmap)
7776 {
7777 bool equal;
7778
7779 if (!(rold->live & REG_LIVE_READ))
7780 /* explored state didn't use this */
7781 return true;
7782
7783 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
7784
7785 if (rold->type == PTR_TO_STACK)
7786 /* two stack pointers are equal only if they're pointing to
7787 * the same stack frame, since fp-8 in foo != fp-8 in bar
7788 */
7789 return equal && rold->frameno == rcur->frameno;
7790
7791 if (equal)
7792 return true;
7793
7794 if (rold->type == NOT_INIT)
7795 /* explored state can't have used this */
7796 return true;
7797 if (rcur->type == NOT_INIT)
7798 return false;
7799 switch (rold->type) {
7800 case SCALAR_VALUE:
7801 if (rcur->type == SCALAR_VALUE) {
7802 if (!rold->precise && !rcur->precise)
7803 return true;
7804 /* new val must satisfy old val knowledge */
7805 return range_within(rold, rcur) &&
7806 tnum_in(rold->var_off, rcur->var_off);
7807 } else {
7808 /* We're trying to use a pointer in place of a scalar.
7809 * Even if the scalar was unbounded, this could lead to
7810 * pointer leaks because scalars are allowed to leak
7811 * while pointers are not. We could make this safe in
7812 * special cases if root is calling us, but it's
7813 * probably not worth the hassle.
7814 */
7815 return false;
7816 }
7817 case PTR_TO_MAP_VALUE:
7818 /* If the new min/max/var_off satisfy the old ones and
7819 * everything else matches, we are OK.
7820 * 'id' is not compared, since it's only used for maps with
7821 * bpf_spin_lock inside map element and in such cases if
7822 * the rest of the prog is valid for one map element then
7823 * it's valid for all map elements regardless of the key
7824 * used in bpf_map_lookup()
7825 */
7826 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
7827 range_within(rold, rcur) &&
7828 tnum_in(rold->var_off, rcur->var_off);
7829 case PTR_TO_MAP_VALUE_OR_NULL:
7830 /* a PTR_TO_MAP_VALUE could be safe to use as a
7831 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
7832 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
7833 * checked, doing so could have affected others with the same
7834 * id, and we can't check for that because we lost the id when
7835 * we converted to a PTR_TO_MAP_VALUE.
7836 */
7837 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
7838 return false;
7839 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
7840 return false;
7841 /* Check our ids match any regs they're supposed to */
7842 return check_ids(rold->id, rcur->id, idmap);
7843 case PTR_TO_PACKET_META:
7844 case PTR_TO_PACKET:
7845 if (rcur->type != rold->type)
7846 return false;
7847 /* We must have at least as much range as the old ptr
7848 * did, so that any accesses which were safe before are
7849 * still safe. This is true even if old range < old off,
7850 * since someone could have accessed through (ptr - k), or
7851 * even done ptr -= k in a register, to get a safe access.
7852 */
7853 if (rold->range > rcur->range)
7854 return false;
7855 /* If the offsets don't match, we can't trust our alignment;
7856 * nor can we be sure that we won't fall out of range.
7857 */
7858 if (rold->off != rcur->off)
7859 return false;
7860 /* id relations must be preserved */
7861 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
7862 return false;
7863 /* new val must satisfy old val knowledge */
7864 return range_within(rold, rcur) &&
7865 tnum_in(rold->var_off, rcur->var_off);
7866 case PTR_TO_CTX:
7867 case CONST_PTR_TO_MAP:
7868 case PTR_TO_PACKET_END:
7869 case PTR_TO_FLOW_KEYS:
7870 case PTR_TO_SOCKET:
7871 case PTR_TO_SOCKET_OR_NULL:
7872 case PTR_TO_SOCK_COMMON:
7873 case PTR_TO_SOCK_COMMON_OR_NULL:
7874 case PTR_TO_TCP_SOCK:
7875 case PTR_TO_TCP_SOCK_OR_NULL:
7876 case PTR_TO_XDP_SOCK:
7877 /* Only valid matches are exact, which memcmp() above
7878 * would have accepted
7879 */
7880 default:
7881 /* Don't know what's going on, just say it's not safe */
7882 return false;
7883 }
7884
7885 /* Shouldn't get here; if we do, say it's not safe */
7886 WARN_ON_ONCE(1);
7887 return false;
7888 }
7889
7890 static bool stacksafe(struct bpf_func_state *old,
7891 struct bpf_func_state *cur,
7892 struct idpair *idmap)
7893 {
7894 int i, spi;
7895
7896 /* walk slots of the explored stack and ignore any additional
7897 * slots in the current stack, since explored(safe) state
7898 * didn't use them
7899 */
7900 for (i = 0; i < old->allocated_stack; i++) {
7901 spi = i / BPF_REG_SIZE;
7902
7903 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
7904 i += BPF_REG_SIZE - 1;
7905 /* explored state didn't use this */
7906 continue;
7907 }
7908
7909 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
7910 continue;
7911
7912 /* explored stack has more populated slots than current stack
7913 * and these slots were used
7914 */
7915 if (i >= cur->allocated_stack)
7916 return false;
7917
7918 /* if old state was safe with misc data in the stack
7919 * it will be safe with zero-initialized stack.
7920 * The opposite is not true
7921 */
7922 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
7923 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
7924 continue;
7925 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
7926 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
7927 /* Ex: old explored (safe) state has STACK_SPILL in
7928 * this stack slot, but current has has STACK_MISC ->
7929 * this verifier states are not equivalent,
7930 * return false to continue verification of this path
7931 */
7932 return false;
7933 if (i % BPF_REG_SIZE)
7934 continue;
7935 if (old->stack[spi].slot_type[0] != STACK_SPILL)
7936 continue;
7937 if (!regsafe(&old->stack[spi].spilled_ptr,
7938 &cur->stack[spi].spilled_ptr,
7939 idmap))
7940 /* when explored and current stack slot are both storing
7941 * spilled registers, check that stored pointers types
7942 * are the same as well.
7943 * Ex: explored safe path could have stored
7944 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
7945 * but current path has stored:
7946 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
7947 * such verifier states are not equivalent.
7948 * return false to continue verification of this path
7949 */
7950 return false;
7951 }
7952 return true;
7953 }
7954
7955 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
7956 {
7957 if (old->acquired_refs != cur->acquired_refs)
7958 return false;
7959 return !memcmp(old->refs, cur->refs,
7960 sizeof(*old->refs) * old->acquired_refs);
7961 }
7962
7963 /* compare two verifier states
7964 *
7965 * all states stored in state_list are known to be valid, since
7966 * verifier reached 'bpf_exit' instruction through them
7967 *
7968 * this function is called when verifier exploring different branches of
7969 * execution popped from the state stack. If it sees an old state that has
7970 * more strict register state and more strict stack state then this execution
7971 * branch doesn't need to be explored further, since verifier already
7972 * concluded that more strict state leads to valid finish.
7973 *
7974 * Therefore two states are equivalent if register state is more conservative
7975 * and explored stack state is more conservative than the current one.
7976 * Example:
7977 * explored current
7978 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
7979 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
7980 *
7981 * In other words if current stack state (one being explored) has more
7982 * valid slots than old one that already passed validation, it means
7983 * the verifier can stop exploring and conclude that current state is valid too
7984 *
7985 * Similarly with registers. If explored state has register type as invalid
7986 * whereas register type in current state is meaningful, it means that
7987 * the current state will reach 'bpf_exit' instruction safely
7988 */
7989 static bool func_states_equal(struct bpf_func_state *old,
7990 struct bpf_func_state *cur)
7991 {
7992 struct idpair *idmap;
7993 bool ret = false;
7994 int i;
7995
7996 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
7997 /* If we failed to allocate the idmap, just say it's not safe */
7998 if (!idmap)
7999 return false;
8000
8001 for (i = 0; i < MAX_BPF_REG; i++) {
8002 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
8003 goto out_free;
8004 }
8005
8006 if (!stacksafe(old, cur, idmap))
8007 goto out_free;
8008
8009 if (!refsafe(old, cur))
8010 goto out_free;
8011 ret = true;
8012 out_free:
8013 kfree(idmap);
8014 return ret;
8015 }
8016
8017 static bool states_equal(struct bpf_verifier_env *env,
8018 struct bpf_verifier_state *old,
8019 struct bpf_verifier_state *cur)
8020 {
8021 int i;
8022
8023 if (old->curframe != cur->curframe)
8024 return false;
8025
8026 /* Verification state from speculative execution simulation
8027 * must never prune a non-speculative execution one.
8028 */
8029 if (old->speculative && !cur->speculative)
8030 return false;
8031
8032 if (old->active_spin_lock != cur->active_spin_lock)
8033 return false;
8034
8035 /* for states to be equal callsites have to be the same
8036 * and all frame states need to be equivalent
8037 */
8038 for (i = 0; i <= old->curframe; i++) {
8039 if (old->frame[i]->callsite != cur->frame[i]->callsite)
8040 return false;
8041 if (!func_states_equal(old->frame[i], cur->frame[i]))
8042 return false;
8043 }
8044 return true;
8045 }
8046
8047 /* Return 0 if no propagation happened. Return negative error code if error
8048 * happened. Otherwise, return the propagated bit.
8049 */
8050 static int propagate_liveness_reg(struct bpf_verifier_env *env,
8051 struct bpf_reg_state *reg,
8052 struct bpf_reg_state *parent_reg)
8053 {
8054 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
8055 u8 flag = reg->live & REG_LIVE_READ;
8056 int err;
8057
8058 /* When comes here, read flags of PARENT_REG or REG could be any of
8059 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
8060 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
8061 */
8062 if (parent_flag == REG_LIVE_READ64 ||
8063 /* Or if there is no read flag from REG. */
8064 !flag ||
8065 /* Or if the read flag from REG is the same as PARENT_REG. */
8066 parent_flag == flag)
8067 return 0;
8068
8069 err = mark_reg_read(env, reg, parent_reg, flag);
8070 if (err)
8071 return err;
8072
8073 return flag;
8074 }
8075
8076 /* A write screens off any subsequent reads; but write marks come from the
8077 * straight-line code between a state and its parent. When we arrive at an
8078 * equivalent state (jump target or such) we didn't arrive by the straight-line
8079 * code, so read marks in the state must propagate to the parent regardless
8080 * of the state's write marks. That's what 'parent == state->parent' comparison
8081 * in mark_reg_read() is for.
8082 */
8083 static int propagate_liveness(struct bpf_verifier_env *env,
8084 const struct bpf_verifier_state *vstate,
8085 struct bpf_verifier_state *vparent)
8086 {
8087 struct bpf_reg_state *state_reg, *parent_reg;
8088 struct bpf_func_state *state, *parent;
8089 int i, frame, err = 0;
8090
8091 if (vparent->curframe != vstate->curframe) {
8092 WARN(1, "propagate_live: parent frame %d current frame %d\n",
8093 vparent->curframe, vstate->curframe);
8094 return -EFAULT;
8095 }
8096 /* Propagate read liveness of registers... */
8097 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
8098 for (frame = 0; frame <= vstate->curframe; frame++) {
8099 parent = vparent->frame[frame];
8100 state = vstate->frame[frame];
8101 parent_reg = parent->regs;
8102 state_reg = state->regs;
8103 /* We don't need to worry about FP liveness, it's read-only */
8104 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
8105 err = propagate_liveness_reg(env, &state_reg[i],
8106 &parent_reg[i]);
8107 if (err < 0)
8108 return err;
8109 if (err == REG_LIVE_READ64)
8110 mark_insn_zext(env, &parent_reg[i]);
8111 }
8112
8113 /* Propagate stack slots. */
8114 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
8115 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
8116 parent_reg = &parent->stack[i].spilled_ptr;
8117 state_reg = &state->stack[i].spilled_ptr;
8118 err = propagate_liveness_reg(env, state_reg,
8119 parent_reg);
8120 if (err < 0)
8121 return err;
8122 }
8123 }
8124 return 0;
8125 }
8126
8127 /* find precise scalars in the previous equivalent state and
8128 * propagate them into the current state
8129 */
8130 static int propagate_precision(struct bpf_verifier_env *env,
8131 const struct bpf_verifier_state *old)
8132 {
8133 struct bpf_reg_state *state_reg;
8134 struct bpf_func_state *state;
8135 int i, err = 0;
8136
8137 state = old->frame[old->curframe];
8138 state_reg = state->regs;
8139 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
8140 if (state_reg->type != SCALAR_VALUE ||
8141 !state_reg->precise)
8142 continue;
8143 if (env->log.level & BPF_LOG_LEVEL2)
8144 verbose(env, "propagating r%d\n", i);
8145 err = mark_chain_precision(env, i);
8146 if (err < 0)
8147 return err;
8148 }
8149
8150 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
8151 if (state->stack[i].slot_type[0] != STACK_SPILL)
8152 continue;
8153 state_reg = &state->stack[i].spilled_ptr;
8154 if (state_reg->type != SCALAR_VALUE ||
8155 !state_reg->precise)
8156 continue;
8157 if (env->log.level & BPF_LOG_LEVEL2)
8158 verbose(env, "propagating fp%d\n",
8159 (-i - 1) * BPF_REG_SIZE);
8160 err = mark_chain_precision_stack(env, i);
8161 if (err < 0)
8162 return err;
8163 }
8164 return 0;
8165 }
8166
8167 static bool states_maybe_looping(struct bpf_verifier_state *old,
8168 struct bpf_verifier_state *cur)
8169 {
8170 struct bpf_func_state *fold, *fcur;
8171 int i, fr = cur->curframe;
8172
8173 if (old->curframe != fr)
8174 return false;
8175
8176 fold = old->frame[fr];
8177 fcur = cur->frame[fr];
8178 for (i = 0; i < MAX_BPF_REG; i++)
8179 if (memcmp(&fold->regs[i], &fcur->regs[i],
8180 offsetof(struct bpf_reg_state, parent)))
8181 return false;
8182 return true;
8183 }
8184
8185
8186 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
8187 {
8188 struct bpf_verifier_state_list *new_sl;
8189 struct bpf_verifier_state_list *sl, **pprev;
8190 struct bpf_verifier_state *cur = env->cur_state, *new;
8191 int i, j, err, states_cnt = 0;
8192 bool add_new_state = env->test_state_freq ? true : false;
8193
8194 cur->last_insn_idx = env->prev_insn_idx;
8195 if (!env->insn_aux_data[insn_idx].prune_point)
8196 /* this 'insn_idx' instruction wasn't marked, so we will not
8197 * be doing state search here
8198 */
8199 return 0;
8200
8201 /* bpf progs typically have pruning point every 4 instructions
8202 * http://vger.kernel.org/bpfconf2019.html#session-1
8203 * Do not add new state for future pruning if the verifier hasn't seen
8204 * at least 2 jumps and at least 8 instructions.
8205 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
8206 * In tests that amounts to up to 50% reduction into total verifier
8207 * memory consumption and 20% verifier time speedup.
8208 */
8209 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
8210 env->insn_processed - env->prev_insn_processed >= 8)
8211 add_new_state = true;
8212
8213 pprev = explored_state(env, insn_idx);
8214 sl = *pprev;
8215
8216 clean_live_states(env, insn_idx, cur);
8217
8218 while (sl) {
8219 states_cnt++;
8220 if (sl->state.insn_idx != insn_idx)
8221 goto next;
8222 if (sl->state.branches) {
8223 if (states_maybe_looping(&sl->state, cur) &&
8224 states_equal(env, &sl->state, cur)) {
8225 verbose_linfo(env, insn_idx, "; ");
8226 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
8227 return -EINVAL;
8228 }
8229 /* if the verifier is processing a loop, avoid adding new state
8230 * too often, since different loop iterations have distinct
8231 * states and may not help future pruning.
8232 * This threshold shouldn't be too low to make sure that
8233 * a loop with large bound will be rejected quickly.
8234 * The most abusive loop will be:
8235 * r1 += 1
8236 * if r1 < 1000000 goto pc-2
8237 * 1M insn_procssed limit / 100 == 10k peak states.
8238 * This threshold shouldn't be too high either, since states
8239 * at the end of the loop are likely to be useful in pruning.
8240 */
8241 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
8242 env->insn_processed - env->prev_insn_processed < 100)
8243 add_new_state = false;
8244 goto miss;
8245 }
8246 if (states_equal(env, &sl->state, cur)) {
8247 sl->hit_cnt++;
8248 /* reached equivalent register/stack state,
8249 * prune the search.
8250 * Registers read by the continuation are read by us.
8251 * If we have any write marks in env->cur_state, they
8252 * will prevent corresponding reads in the continuation
8253 * from reaching our parent (an explored_state). Our
8254 * own state will get the read marks recorded, but
8255 * they'll be immediately forgotten as we're pruning
8256 * this state and will pop a new one.
8257 */
8258 err = propagate_liveness(env, &sl->state, cur);
8259
8260 /* if previous state reached the exit with precision and
8261 * current state is equivalent to it (except precsion marks)
8262 * the precision needs to be propagated back in
8263 * the current state.
8264 */
8265 err = err ? : push_jmp_history(env, cur);
8266 err = err ? : propagate_precision(env, &sl->state);
8267 if (err)
8268 return err;
8269 return 1;
8270 }
8271 miss:
8272 /* when new state is not going to be added do not increase miss count.
8273 * Otherwise several loop iterations will remove the state
8274 * recorded earlier. The goal of these heuristics is to have
8275 * states from some iterations of the loop (some in the beginning
8276 * and some at the end) to help pruning.
8277 */
8278 if (add_new_state)
8279 sl->miss_cnt++;
8280 /* heuristic to determine whether this state is beneficial
8281 * to keep checking from state equivalence point of view.
8282 * Higher numbers increase max_states_per_insn and verification time,
8283 * but do not meaningfully decrease insn_processed.
8284 */
8285 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
8286 /* the state is unlikely to be useful. Remove it to
8287 * speed up verification
8288 */
8289 *pprev = sl->next;
8290 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
8291 u32 br = sl->state.branches;
8292
8293 WARN_ONCE(br,
8294 "BUG live_done but branches_to_explore %d\n",
8295 br);
8296 free_verifier_state(&sl->state, false);
8297 kfree(sl);
8298 env->peak_states--;
8299 } else {
8300 /* cannot free this state, since parentage chain may
8301 * walk it later. Add it for free_list instead to
8302 * be freed at the end of verification
8303 */
8304 sl->next = env->free_list;
8305 env->free_list = sl;
8306 }
8307 sl = *pprev;
8308 continue;
8309 }
8310 next:
8311 pprev = &sl->next;
8312 sl = *pprev;
8313 }
8314
8315 if (env->max_states_per_insn < states_cnt)
8316 env->max_states_per_insn = states_cnt;
8317
8318 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
8319 return push_jmp_history(env, cur);
8320
8321 if (!add_new_state)
8322 return push_jmp_history(env, cur);
8323
8324 /* There were no equivalent states, remember the current one.
8325 * Technically the current state is not proven to be safe yet,
8326 * but it will either reach outer most bpf_exit (which means it's safe)
8327 * or it will be rejected. When there are no loops the verifier won't be
8328 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
8329 * again on the way to bpf_exit.
8330 * When looping the sl->state.branches will be > 0 and this state
8331 * will not be considered for equivalence until branches == 0.
8332 */
8333 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
8334 if (!new_sl)
8335 return -ENOMEM;
8336 env->total_states++;
8337 env->peak_states++;
8338 env->prev_jmps_processed = env->jmps_processed;
8339 env->prev_insn_processed = env->insn_processed;
8340
8341 /* add new state to the head of linked list */
8342 new = &new_sl->state;
8343 err = copy_verifier_state(new, cur);
8344 if (err) {
8345 free_verifier_state(new, false);
8346 kfree(new_sl);
8347 return err;
8348 }
8349 new->insn_idx = insn_idx;
8350 WARN_ONCE(new->branches != 1,
8351 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
8352
8353 cur->parent = new;
8354 cur->first_insn_idx = insn_idx;
8355 clear_jmp_history(cur);
8356 new_sl->next = *explored_state(env, insn_idx);
8357 *explored_state(env, insn_idx) = new_sl;
8358 /* connect new state to parentage chain. Current frame needs all
8359 * registers connected. Only r6 - r9 of the callers are alive (pushed
8360 * to the stack implicitly by JITs) so in callers' frames connect just
8361 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
8362 * the state of the call instruction (with WRITTEN set), and r0 comes
8363 * from callee with its full parentage chain, anyway.
8364 */
8365 /* clear write marks in current state: the writes we did are not writes
8366 * our child did, so they don't screen off its reads from us.
8367 * (There are no read marks in current state, because reads always mark
8368 * their parent and current state never has children yet. Only
8369 * explored_states can get read marks.)
8370 */
8371 for (j = 0; j <= cur->curframe; j++) {
8372 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
8373 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
8374 for (i = 0; i < BPF_REG_FP; i++)
8375 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
8376 }
8377
8378 /* all stack frames are accessible from callee, clear them all */
8379 for (j = 0; j <= cur->curframe; j++) {
8380 struct bpf_func_state *frame = cur->frame[j];
8381 struct bpf_func_state *newframe = new->frame[j];
8382
8383 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
8384 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
8385 frame->stack[i].spilled_ptr.parent =
8386 &newframe->stack[i].spilled_ptr;
8387 }
8388 }
8389 return 0;
8390 }
8391
8392 /* Return true if it's OK to have the same insn return a different type. */
8393 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
8394 {
8395 switch (type) {
8396 case PTR_TO_CTX:
8397 case PTR_TO_SOCKET:
8398 case PTR_TO_SOCKET_OR_NULL:
8399 case PTR_TO_SOCK_COMMON:
8400 case PTR_TO_SOCK_COMMON_OR_NULL:
8401 case PTR_TO_TCP_SOCK:
8402 case PTR_TO_TCP_SOCK_OR_NULL:
8403 case PTR_TO_XDP_SOCK:
8404 case PTR_TO_BTF_ID:
8405 return false;
8406 default:
8407 return true;
8408 }
8409 }
8410
8411 /* If an instruction was previously used with particular pointer types, then we
8412 * need to be careful to avoid cases such as the below, where it may be ok
8413 * for one branch accessing the pointer, but not ok for the other branch:
8414 *
8415 * R1 = sock_ptr
8416 * goto X;
8417 * ...
8418 * R1 = some_other_valid_ptr;
8419 * goto X;
8420 * ...
8421 * R2 = *(u32 *)(R1 + 0);
8422 */
8423 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
8424 {
8425 return src != prev && (!reg_type_mismatch_ok(src) ||
8426 !reg_type_mismatch_ok(prev));
8427 }
8428
8429 static int do_check(struct bpf_verifier_env *env)
8430 {
8431 struct bpf_verifier_state *state = env->cur_state;
8432 struct bpf_insn *insns = env->prog->insnsi;
8433 struct bpf_reg_state *regs;
8434 int insn_cnt = env->prog->len;
8435 bool do_print_state = false;
8436 int prev_insn_idx = -1;
8437
8438 for (;;) {
8439 struct bpf_insn *insn;
8440 u8 class;
8441 int err;
8442
8443 env->prev_insn_idx = prev_insn_idx;
8444 if (env->insn_idx >= insn_cnt) {
8445 verbose(env, "invalid insn idx %d insn_cnt %d\n",
8446 env->insn_idx, insn_cnt);
8447 return -EFAULT;
8448 }
8449
8450 insn = &insns[env->insn_idx];
8451 class = BPF_CLASS(insn->code);
8452
8453 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
8454 verbose(env,
8455 "BPF program is too large. Processed %d insn\n",
8456 env->insn_processed);
8457 return -E2BIG;
8458 }
8459
8460 err = is_state_visited(env, env->insn_idx);
8461 if (err < 0)
8462 return err;
8463 if (err == 1) {
8464 /* found equivalent state, can prune the search */
8465 if (env->log.level & BPF_LOG_LEVEL) {
8466 if (do_print_state)
8467 verbose(env, "\nfrom %d to %d%s: safe\n",
8468 env->prev_insn_idx, env->insn_idx,
8469 env->cur_state->speculative ?
8470 " (speculative execution)" : "");
8471 else
8472 verbose(env, "%d: safe\n", env->insn_idx);
8473 }
8474 goto process_bpf_exit;
8475 }
8476
8477 if (signal_pending(current))
8478 return -EAGAIN;
8479
8480 if (need_resched())
8481 cond_resched();
8482
8483 if (env->log.level & BPF_LOG_LEVEL2 ||
8484 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
8485 if (env->log.level & BPF_LOG_LEVEL2)
8486 verbose(env, "%d:", env->insn_idx);
8487 else
8488 verbose(env, "\nfrom %d to %d%s:",
8489 env->prev_insn_idx, env->insn_idx,
8490 env->cur_state->speculative ?
8491 " (speculative execution)" : "");
8492 print_verifier_state(env, state->frame[state->curframe]);
8493 do_print_state = false;
8494 }
8495
8496 if (env->log.level & BPF_LOG_LEVEL) {
8497 const struct bpf_insn_cbs cbs = {
8498 .cb_print = verbose,
8499 .private_data = env,
8500 };
8501
8502 verbose_linfo(env, env->insn_idx, "; ");
8503 verbose(env, "%d: ", env->insn_idx);
8504 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
8505 }
8506
8507 if (bpf_prog_is_dev_bound(env->prog->aux)) {
8508 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
8509 env->prev_insn_idx);
8510 if (err)
8511 return err;
8512 }
8513
8514 regs = cur_regs(env);
8515 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8516 prev_insn_idx = env->insn_idx;
8517
8518 if (class == BPF_ALU || class == BPF_ALU64) {
8519 err = check_alu_op(env, insn);
8520 if (err)
8521 return err;
8522
8523 } else if (class == BPF_LDX) {
8524 enum bpf_reg_type *prev_src_type, src_reg_type;
8525
8526 /* check for reserved fields is already done */
8527
8528 /* check src operand */
8529 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8530 if (err)
8531 return err;
8532
8533 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8534 if (err)
8535 return err;
8536
8537 src_reg_type = regs[insn->src_reg].type;
8538
8539 /* check that memory (src_reg + off) is readable,
8540 * the state of dst_reg will be updated by this func
8541 */
8542 err = check_mem_access(env, env->insn_idx, insn->src_reg,
8543 insn->off, BPF_SIZE(insn->code),
8544 BPF_READ, insn->dst_reg, false);
8545 if (err)
8546 return err;
8547
8548 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
8549
8550 if (*prev_src_type == NOT_INIT) {
8551 /* saw a valid insn
8552 * dst_reg = *(u32 *)(src_reg + off)
8553 * save type to validate intersecting paths
8554 */
8555 *prev_src_type = src_reg_type;
8556
8557 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
8558 /* ABuser program is trying to use the same insn
8559 * dst_reg = *(u32*) (src_reg + off)
8560 * with different pointer types:
8561 * src_reg == ctx in one branch and
8562 * src_reg == stack|map in some other branch.
8563 * Reject it.
8564 */
8565 verbose(env, "same insn cannot be used with different pointers\n");
8566 return -EINVAL;
8567 }
8568
8569 } else if (class == BPF_STX) {
8570 enum bpf_reg_type *prev_dst_type, dst_reg_type;
8571
8572 if (BPF_MODE(insn->code) == BPF_XADD) {
8573 err = check_xadd(env, env->insn_idx, insn);
8574 if (err)
8575 return err;
8576 env->insn_idx++;
8577 continue;
8578 }
8579
8580 /* check src1 operand */
8581 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8582 if (err)
8583 return err;
8584 /* check src2 operand */
8585 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8586 if (err)
8587 return err;
8588
8589 dst_reg_type = regs[insn->dst_reg].type;
8590
8591 /* check that memory (dst_reg + off) is writeable */
8592 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8593 insn->off, BPF_SIZE(insn->code),
8594 BPF_WRITE, insn->src_reg, false);
8595 if (err)
8596 return err;
8597
8598 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
8599
8600 if (*prev_dst_type == NOT_INIT) {
8601 *prev_dst_type = dst_reg_type;
8602 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
8603 verbose(env, "same insn cannot be used with different pointers\n");
8604 return -EINVAL;
8605 }
8606
8607 } else if (class == BPF_ST) {
8608 if (BPF_MODE(insn->code) != BPF_MEM ||
8609 insn->src_reg != BPF_REG_0) {
8610 verbose(env, "BPF_ST uses reserved fields\n");
8611 return -EINVAL;
8612 }
8613 /* check src operand */
8614 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8615 if (err)
8616 return err;
8617
8618 if (is_ctx_reg(env, insn->dst_reg)) {
8619 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
8620 insn->dst_reg,
8621 reg_type_str[reg_state(env, insn->dst_reg)->type]);
8622 return -EACCES;
8623 }
8624
8625 /* check that memory (dst_reg + off) is writeable */
8626 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8627 insn->off, BPF_SIZE(insn->code),
8628 BPF_WRITE, -1, false);
8629 if (err)
8630 return err;
8631
8632 } else if (class == BPF_JMP || class == BPF_JMP32) {
8633 u8 opcode = BPF_OP(insn->code);
8634
8635 env->jmps_processed++;
8636 if (opcode == BPF_CALL) {
8637 if (BPF_SRC(insn->code) != BPF_K ||
8638 insn->off != 0 ||
8639 (insn->src_reg != BPF_REG_0 &&
8640 insn->src_reg != BPF_PSEUDO_CALL) ||
8641 insn->dst_reg != BPF_REG_0 ||
8642 class == BPF_JMP32) {
8643 verbose(env, "BPF_CALL uses reserved fields\n");
8644 return -EINVAL;
8645 }
8646
8647 if (env->cur_state->active_spin_lock &&
8648 (insn->src_reg == BPF_PSEUDO_CALL ||
8649 insn->imm != BPF_FUNC_spin_unlock)) {
8650 verbose(env, "function calls are not allowed while holding a lock\n");
8651 return -EINVAL;
8652 }
8653 if (insn->src_reg == BPF_PSEUDO_CALL)
8654 err = check_func_call(env, insn, &env->insn_idx);
8655 else
8656 err = check_helper_call(env, insn->imm, env->insn_idx);
8657 if (err)
8658 return err;
8659
8660 } else if (opcode == BPF_JA) {
8661 if (BPF_SRC(insn->code) != BPF_K ||
8662 insn->imm != 0 ||
8663 insn->src_reg != BPF_REG_0 ||
8664 insn->dst_reg != BPF_REG_0 ||
8665 class == BPF_JMP32) {
8666 verbose(env, "BPF_JA uses reserved fields\n");
8667 return -EINVAL;
8668 }
8669
8670 env->insn_idx += insn->off + 1;
8671 continue;
8672
8673 } else if (opcode == BPF_EXIT) {
8674 if (BPF_SRC(insn->code) != BPF_K ||
8675 insn->imm != 0 ||
8676 insn->src_reg != BPF_REG_0 ||
8677 insn->dst_reg != BPF_REG_0 ||
8678 class == BPF_JMP32) {
8679 verbose(env, "BPF_EXIT uses reserved fields\n");
8680 return -EINVAL;
8681 }
8682
8683 if (env->cur_state->active_spin_lock) {
8684 verbose(env, "bpf_spin_unlock is missing\n");
8685 return -EINVAL;
8686 }
8687
8688 if (state->curframe) {
8689 /* exit from nested function */
8690 err = prepare_func_exit(env, &env->insn_idx);
8691 if (err)
8692 return err;
8693 do_print_state = true;
8694 continue;
8695 }
8696
8697 err = check_reference_leak(env);
8698 if (err)
8699 return err;
8700
8701 err = check_return_code(env);
8702 if (err)
8703 return err;
8704 process_bpf_exit:
8705 update_branch_counts(env, env->cur_state);
8706 err = pop_stack(env, &prev_insn_idx,
8707 &env->insn_idx);
8708 if (err < 0) {
8709 if (err != -ENOENT)
8710 return err;
8711 break;
8712 } else {
8713 do_print_state = true;
8714 continue;
8715 }
8716 } else {
8717 err = check_cond_jmp_op(env, insn, &env->insn_idx);
8718 if (err)
8719 return err;
8720 }
8721 } else if (class == BPF_LD) {
8722 u8 mode = BPF_MODE(insn->code);
8723
8724 if (mode == BPF_ABS || mode == BPF_IND) {
8725 err = check_ld_abs(env, insn);
8726 if (err)
8727 return err;
8728
8729 } else if (mode == BPF_IMM) {
8730 err = check_ld_imm(env, insn);
8731 if (err)
8732 return err;
8733
8734 env->insn_idx++;
8735 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8736 } else {
8737 verbose(env, "invalid BPF_LD mode\n");
8738 return -EINVAL;
8739 }
8740 } else {
8741 verbose(env, "unknown insn class %d\n", class);
8742 return -EINVAL;
8743 }
8744
8745 env->insn_idx++;
8746 }
8747
8748 return 0;
8749 }
8750
8751 static int check_map_prealloc(struct bpf_map *map)
8752 {
8753 return (map->map_type != BPF_MAP_TYPE_HASH &&
8754 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8755 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
8756 !(map->map_flags & BPF_F_NO_PREALLOC);
8757 }
8758
8759 static bool is_tracing_prog_type(enum bpf_prog_type type)
8760 {
8761 switch (type) {
8762 case BPF_PROG_TYPE_KPROBE:
8763 case BPF_PROG_TYPE_TRACEPOINT:
8764 case BPF_PROG_TYPE_PERF_EVENT:
8765 case BPF_PROG_TYPE_RAW_TRACEPOINT:
8766 return true;
8767 default:
8768 return false;
8769 }
8770 }
8771
8772 static bool is_preallocated_map(struct bpf_map *map)
8773 {
8774 if (!check_map_prealloc(map))
8775 return false;
8776 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
8777 return false;
8778 return true;
8779 }
8780
8781 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
8782 struct bpf_map *map,
8783 struct bpf_prog *prog)
8784
8785 {
8786 /*
8787 * Validate that trace type programs use preallocated hash maps.
8788 *
8789 * For programs attached to PERF events this is mandatory as the
8790 * perf NMI can hit any arbitrary code sequence.
8791 *
8792 * All other trace types using preallocated hash maps are unsafe as
8793 * well because tracepoint or kprobes can be inside locked regions
8794 * of the memory allocator or at a place where a recursion into the
8795 * memory allocator would see inconsistent state.
8796 *
8797 * On RT enabled kernels run-time allocation of all trace type
8798 * programs is strictly prohibited due to lock type constraints. On
8799 * !RT kernels it is allowed for backwards compatibility reasons for
8800 * now, but warnings are emitted so developers are made aware of
8801 * the unsafety and can fix their programs before this is enforced.
8802 */
8803 if (is_tracing_prog_type(prog->type) && !is_preallocated_map(map)) {
8804 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
8805 verbose(env, "perf_event programs can only use preallocated hash map\n");
8806 return -EINVAL;
8807 }
8808 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
8809 verbose(env, "trace type programs can only use preallocated hash map\n");
8810 return -EINVAL;
8811 }
8812 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
8813 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
8814 }
8815
8816 if ((is_tracing_prog_type(prog->type) ||
8817 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
8818 map_value_has_spin_lock(map)) {
8819 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
8820 return -EINVAL;
8821 }
8822
8823 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
8824 !bpf_offload_prog_map_match(prog, map)) {
8825 verbose(env, "offload device mismatch between prog and map\n");
8826 return -EINVAL;
8827 }
8828
8829 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
8830 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
8831 return -EINVAL;
8832 }
8833
8834 return 0;
8835 }
8836
8837 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
8838 {
8839 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
8840 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
8841 }
8842
8843 /* look for pseudo eBPF instructions that access map FDs and
8844 * replace them with actual map pointers
8845 */
8846 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
8847 {
8848 struct bpf_insn *insn = env->prog->insnsi;
8849 int insn_cnt = env->prog->len;
8850 int i, j, err;
8851
8852 err = bpf_prog_calc_tag(env->prog);
8853 if (err)
8854 return err;
8855
8856 for (i = 0; i < insn_cnt; i++, insn++) {
8857 if (BPF_CLASS(insn->code) == BPF_LDX &&
8858 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
8859 verbose(env, "BPF_LDX uses reserved fields\n");
8860 return -EINVAL;
8861 }
8862
8863 if (BPF_CLASS(insn->code) == BPF_STX &&
8864 ((BPF_MODE(insn->code) != BPF_MEM &&
8865 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
8866 verbose(env, "BPF_STX uses reserved fields\n");
8867 return -EINVAL;
8868 }
8869
8870 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
8871 struct bpf_insn_aux_data *aux;
8872 struct bpf_map *map;
8873 struct fd f;
8874 u64 addr;
8875
8876 if (i == insn_cnt - 1 || insn[1].code != 0 ||
8877 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
8878 insn[1].off != 0) {
8879 verbose(env, "invalid bpf_ld_imm64 insn\n");
8880 return -EINVAL;
8881 }
8882
8883 if (insn[0].src_reg == 0)
8884 /* valid generic load 64-bit imm */
8885 goto next_insn;
8886
8887 /* In final convert_pseudo_ld_imm64() step, this is
8888 * converted into regular 64-bit imm load insn.
8889 */
8890 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
8891 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
8892 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
8893 insn[1].imm != 0)) {
8894 verbose(env,
8895 "unrecognized bpf_ld_imm64 insn\n");
8896 return -EINVAL;
8897 }
8898
8899 f = fdget(insn[0].imm);
8900 map = __bpf_map_get(f);
8901 if (IS_ERR(map)) {
8902 verbose(env, "fd %d is not pointing to valid bpf_map\n",
8903 insn[0].imm);
8904 return PTR_ERR(map);
8905 }
8906
8907 err = check_map_prog_compatibility(env, map, env->prog);
8908 if (err) {
8909 fdput(f);
8910 return err;
8911 }
8912
8913 aux = &env->insn_aux_data[i];
8914 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8915 addr = (unsigned long)map;
8916 } else {
8917 u32 off = insn[1].imm;
8918
8919 if (off >= BPF_MAX_VAR_OFF) {
8920 verbose(env, "direct value offset of %u is not allowed\n", off);
8921 fdput(f);
8922 return -EINVAL;
8923 }
8924
8925 if (!map->ops->map_direct_value_addr) {
8926 verbose(env, "no direct value access support for this map type\n");
8927 fdput(f);
8928 return -EINVAL;
8929 }
8930
8931 err = map->ops->map_direct_value_addr(map, &addr, off);
8932 if (err) {
8933 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
8934 map->value_size, off);
8935 fdput(f);
8936 return err;
8937 }
8938
8939 aux->map_off = off;
8940 addr += off;
8941 }
8942
8943 insn[0].imm = (u32)addr;
8944 insn[1].imm = addr >> 32;
8945
8946 /* check whether we recorded this map already */
8947 for (j = 0; j < env->used_map_cnt; j++) {
8948 if (env->used_maps[j] == map) {
8949 aux->map_index = j;
8950 fdput(f);
8951 goto next_insn;
8952 }
8953 }
8954
8955 if (env->used_map_cnt >= MAX_USED_MAPS) {
8956 fdput(f);
8957 return -E2BIG;
8958 }
8959
8960 /* hold the map. If the program is rejected by verifier,
8961 * the map will be released by release_maps() or it
8962 * will be used by the valid program until it's unloaded
8963 * and all maps are released in free_used_maps()
8964 */
8965 bpf_map_inc(map);
8966
8967 aux->map_index = env->used_map_cnt;
8968 env->used_maps[env->used_map_cnt++] = map;
8969
8970 if (bpf_map_is_cgroup_storage(map) &&
8971 bpf_cgroup_storage_assign(env->prog->aux, map)) {
8972 verbose(env, "only one cgroup storage of each type is allowed\n");
8973 fdput(f);
8974 return -EBUSY;
8975 }
8976
8977 fdput(f);
8978 next_insn:
8979 insn++;
8980 i++;
8981 continue;
8982 }
8983
8984 /* Basic sanity check before we invest more work here. */
8985 if (!bpf_opcode_in_insntable(insn->code)) {
8986 verbose(env, "unknown opcode %02x\n", insn->code);
8987 return -EINVAL;
8988 }
8989 }
8990
8991 /* now all pseudo BPF_LD_IMM64 instructions load valid
8992 * 'struct bpf_map *' into a register instead of user map_fd.
8993 * These pointers will be used later by verifier to validate map access.
8994 */
8995 return 0;
8996 }
8997
8998 /* drop refcnt of maps used by the rejected program */
8999 static void release_maps(struct bpf_verifier_env *env)
9000 {
9001 __bpf_free_used_maps(env->prog->aux, env->used_maps,
9002 env->used_map_cnt);
9003 }
9004
9005 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
9006 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
9007 {
9008 struct bpf_insn *insn = env->prog->insnsi;
9009 int insn_cnt = env->prog->len;
9010 int i;
9011
9012 for (i = 0; i < insn_cnt; i++, insn++)
9013 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
9014 insn->src_reg = 0;
9015 }
9016
9017 /* single env->prog->insni[off] instruction was replaced with the range
9018 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
9019 * [0, off) and [off, end) to new locations, so the patched range stays zero
9020 */
9021 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
9022 struct bpf_prog *new_prog, u32 off, u32 cnt)
9023 {
9024 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
9025 struct bpf_insn *insn = new_prog->insnsi;
9026 u32 prog_len;
9027 int i;
9028
9029 /* aux info at OFF always needs adjustment, no matter fast path
9030 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
9031 * original insn at old prog.
9032 */
9033 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
9034
9035 if (cnt == 1)
9036 return 0;
9037 prog_len = new_prog->len;
9038 new_data = vzalloc(array_size(prog_len,
9039 sizeof(struct bpf_insn_aux_data)));
9040 if (!new_data)
9041 return -ENOMEM;
9042 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
9043 memcpy(new_data + off + cnt - 1, old_data + off,
9044 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
9045 for (i = off; i < off + cnt - 1; i++) {
9046 new_data[i].seen = env->pass_cnt;
9047 new_data[i].zext_dst = insn_has_def32(env, insn + i);
9048 }
9049 env->insn_aux_data = new_data;
9050 vfree(old_data);
9051 return 0;
9052 }
9053
9054 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
9055 {
9056 int i;
9057
9058 if (len == 1)
9059 return;
9060 /* NOTE: fake 'exit' subprog should be updated as well. */
9061 for (i = 0; i <= env->subprog_cnt; i++) {
9062 if (env->subprog_info[i].start <= off)
9063 continue;
9064 env->subprog_info[i].start += len - 1;
9065 }
9066 }
9067
9068 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
9069 const struct bpf_insn *patch, u32 len)
9070 {
9071 struct bpf_prog *new_prog;
9072
9073 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
9074 if (IS_ERR(new_prog)) {
9075 if (PTR_ERR(new_prog) == -ERANGE)
9076 verbose(env,
9077 "insn %d cannot be patched due to 16-bit range\n",
9078 env->insn_aux_data[off].orig_idx);
9079 return NULL;
9080 }
9081 if (adjust_insn_aux_data(env, new_prog, off, len))
9082 return NULL;
9083 adjust_subprog_starts(env, off, len);
9084 return new_prog;
9085 }
9086
9087 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
9088 u32 off, u32 cnt)
9089 {
9090 int i, j;
9091
9092 /* find first prog starting at or after off (first to remove) */
9093 for (i = 0; i < env->subprog_cnt; i++)
9094 if (env->subprog_info[i].start >= off)
9095 break;
9096 /* find first prog starting at or after off + cnt (first to stay) */
9097 for (j = i; j < env->subprog_cnt; j++)
9098 if (env->subprog_info[j].start >= off + cnt)
9099 break;
9100 /* if j doesn't start exactly at off + cnt, we are just removing
9101 * the front of previous prog
9102 */
9103 if (env->subprog_info[j].start != off + cnt)
9104 j--;
9105
9106 if (j > i) {
9107 struct bpf_prog_aux *aux = env->prog->aux;
9108 int move;
9109
9110 /* move fake 'exit' subprog as well */
9111 move = env->subprog_cnt + 1 - j;
9112
9113 memmove(env->subprog_info + i,
9114 env->subprog_info + j,
9115 sizeof(*env->subprog_info) * move);
9116 env->subprog_cnt -= j - i;
9117
9118 /* remove func_info */
9119 if (aux->func_info) {
9120 move = aux->func_info_cnt - j;
9121
9122 memmove(aux->func_info + i,
9123 aux->func_info + j,
9124 sizeof(*aux->func_info) * move);
9125 aux->func_info_cnt -= j - i;
9126 /* func_info->insn_off is set after all code rewrites,
9127 * in adjust_btf_func() - no need to adjust
9128 */
9129 }
9130 } else {
9131 /* convert i from "first prog to remove" to "first to adjust" */
9132 if (env->subprog_info[i].start == off)
9133 i++;
9134 }
9135
9136 /* update fake 'exit' subprog as well */
9137 for (; i <= env->subprog_cnt; i++)
9138 env->subprog_info[i].start -= cnt;
9139
9140 return 0;
9141 }
9142
9143 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
9144 u32 cnt)
9145 {
9146 struct bpf_prog *prog = env->prog;
9147 u32 i, l_off, l_cnt, nr_linfo;
9148 struct bpf_line_info *linfo;
9149
9150 nr_linfo = prog->aux->nr_linfo;
9151 if (!nr_linfo)
9152 return 0;
9153
9154 linfo = prog->aux->linfo;
9155
9156 /* find first line info to remove, count lines to be removed */
9157 for (i = 0; i < nr_linfo; i++)
9158 if (linfo[i].insn_off >= off)
9159 break;
9160
9161 l_off = i;
9162 l_cnt = 0;
9163 for (; i < nr_linfo; i++)
9164 if (linfo[i].insn_off < off + cnt)
9165 l_cnt++;
9166 else
9167 break;
9168
9169 /* First live insn doesn't match first live linfo, it needs to "inherit"
9170 * last removed linfo. prog is already modified, so prog->len == off
9171 * means no live instructions after (tail of the program was removed).
9172 */
9173 if (prog->len != off && l_cnt &&
9174 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
9175 l_cnt--;
9176 linfo[--i].insn_off = off + cnt;
9177 }
9178
9179 /* remove the line info which refer to the removed instructions */
9180 if (l_cnt) {
9181 memmove(linfo + l_off, linfo + i,
9182 sizeof(*linfo) * (nr_linfo - i));
9183
9184 prog->aux->nr_linfo -= l_cnt;
9185 nr_linfo = prog->aux->nr_linfo;
9186 }
9187
9188 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
9189 for (i = l_off; i < nr_linfo; i++)
9190 linfo[i].insn_off -= cnt;
9191
9192 /* fix up all subprogs (incl. 'exit') which start >= off */
9193 for (i = 0; i <= env->subprog_cnt; i++)
9194 if (env->subprog_info[i].linfo_idx > l_off) {
9195 /* program may have started in the removed region but
9196 * may not be fully removed
9197 */
9198 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
9199 env->subprog_info[i].linfo_idx -= l_cnt;
9200 else
9201 env->subprog_info[i].linfo_idx = l_off;
9202 }
9203
9204 return 0;
9205 }
9206
9207 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
9208 {
9209 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9210 unsigned int orig_prog_len = env->prog->len;
9211 int err;
9212
9213 if (bpf_prog_is_dev_bound(env->prog->aux))
9214 bpf_prog_offload_remove_insns(env, off, cnt);
9215
9216 err = bpf_remove_insns(env->prog, off, cnt);
9217 if (err)
9218 return err;
9219
9220 err = adjust_subprog_starts_after_remove(env, off, cnt);
9221 if (err)
9222 return err;
9223
9224 err = bpf_adj_linfo_after_remove(env, off, cnt);
9225 if (err)
9226 return err;
9227
9228 memmove(aux_data + off, aux_data + off + cnt,
9229 sizeof(*aux_data) * (orig_prog_len - off - cnt));
9230
9231 return 0;
9232 }
9233
9234 /* The verifier does more data flow analysis than llvm and will not
9235 * explore branches that are dead at run time. Malicious programs can
9236 * have dead code too. Therefore replace all dead at-run-time code
9237 * with 'ja -1'.
9238 *
9239 * Just nops are not optimal, e.g. if they would sit at the end of the
9240 * program and through another bug we would manage to jump there, then
9241 * we'd execute beyond program memory otherwise. Returning exception
9242 * code also wouldn't work since we can have subprogs where the dead
9243 * code could be located.
9244 */
9245 static void sanitize_dead_code(struct bpf_verifier_env *env)
9246 {
9247 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9248 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
9249 struct bpf_insn *insn = env->prog->insnsi;
9250 const int insn_cnt = env->prog->len;
9251 int i;
9252
9253 for (i = 0; i < insn_cnt; i++) {
9254 if (aux_data[i].seen)
9255 continue;
9256 memcpy(insn + i, &trap, sizeof(trap));
9257 }
9258 }
9259
9260 static bool insn_is_cond_jump(u8 code)
9261 {
9262 u8 op;
9263
9264 if (BPF_CLASS(code) == BPF_JMP32)
9265 return true;
9266
9267 if (BPF_CLASS(code) != BPF_JMP)
9268 return false;
9269
9270 op = BPF_OP(code);
9271 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
9272 }
9273
9274 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
9275 {
9276 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9277 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
9278 struct bpf_insn *insn = env->prog->insnsi;
9279 const int insn_cnt = env->prog->len;
9280 int i;
9281
9282 for (i = 0; i < insn_cnt; i++, insn++) {
9283 if (!insn_is_cond_jump(insn->code))
9284 continue;
9285
9286 if (!aux_data[i + 1].seen)
9287 ja.off = insn->off;
9288 else if (!aux_data[i + 1 + insn->off].seen)
9289 ja.off = 0;
9290 else
9291 continue;
9292
9293 if (bpf_prog_is_dev_bound(env->prog->aux))
9294 bpf_prog_offload_replace_insn(env, i, &ja);
9295
9296 memcpy(insn, &ja, sizeof(ja));
9297 }
9298 }
9299
9300 static int opt_remove_dead_code(struct bpf_verifier_env *env)
9301 {
9302 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9303 int insn_cnt = env->prog->len;
9304 int i, err;
9305
9306 for (i = 0; i < insn_cnt; i++) {
9307 int j;
9308
9309 j = 0;
9310 while (i + j < insn_cnt && !aux_data[i + j].seen)
9311 j++;
9312 if (!j)
9313 continue;
9314
9315 err = verifier_remove_insns(env, i, j);
9316 if (err)
9317 return err;
9318 insn_cnt = env->prog->len;
9319 }
9320
9321 return 0;
9322 }
9323
9324 static int opt_remove_nops(struct bpf_verifier_env *env)
9325 {
9326 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
9327 struct bpf_insn *insn = env->prog->insnsi;
9328 int insn_cnt = env->prog->len;
9329 int i, err;
9330
9331 for (i = 0; i < insn_cnt; i++) {
9332 if (memcmp(&insn[i], &ja, sizeof(ja)))
9333 continue;
9334
9335 err = verifier_remove_insns(env, i, 1);
9336 if (err)
9337 return err;
9338 insn_cnt--;
9339 i--;
9340 }
9341
9342 return 0;
9343 }
9344
9345 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
9346 const union bpf_attr *attr)
9347 {
9348 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
9349 struct bpf_insn_aux_data *aux = env->insn_aux_data;
9350 int i, patch_len, delta = 0, len = env->prog->len;
9351 struct bpf_insn *insns = env->prog->insnsi;
9352 struct bpf_prog *new_prog;
9353 bool rnd_hi32;
9354
9355 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
9356 zext_patch[1] = BPF_ZEXT_REG(0);
9357 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
9358 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
9359 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
9360 for (i = 0; i < len; i++) {
9361 int adj_idx = i + delta;
9362 struct bpf_insn insn;
9363
9364 insn = insns[adj_idx];
9365 if (!aux[adj_idx].zext_dst) {
9366 u8 code, class;
9367 u32 imm_rnd;
9368
9369 if (!rnd_hi32)
9370 continue;
9371
9372 code = insn.code;
9373 class = BPF_CLASS(code);
9374 if (insn_no_def(&insn))
9375 continue;
9376
9377 /* NOTE: arg "reg" (the fourth one) is only used for
9378 * BPF_STX which has been ruled out in above
9379 * check, it is safe to pass NULL here.
9380 */
9381 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
9382 if (class == BPF_LD &&
9383 BPF_MODE(code) == BPF_IMM)
9384 i++;
9385 continue;
9386 }
9387
9388 /* ctx load could be transformed into wider load. */
9389 if (class == BPF_LDX &&
9390 aux[adj_idx].ptr_type == PTR_TO_CTX)
9391 continue;
9392
9393 imm_rnd = get_random_int();
9394 rnd_hi32_patch[0] = insn;
9395 rnd_hi32_patch[1].imm = imm_rnd;
9396 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
9397 patch = rnd_hi32_patch;
9398 patch_len = 4;
9399 goto apply_patch_buffer;
9400 }
9401
9402 if (!bpf_jit_needs_zext())
9403 continue;
9404
9405 zext_patch[0] = insn;
9406 zext_patch[1].dst_reg = insn.dst_reg;
9407 zext_patch[1].src_reg = insn.dst_reg;
9408 patch = zext_patch;
9409 patch_len = 2;
9410 apply_patch_buffer:
9411 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
9412 if (!new_prog)
9413 return -ENOMEM;
9414 env->prog = new_prog;
9415 insns = new_prog->insnsi;
9416 aux = env->insn_aux_data;
9417 delta += patch_len - 1;
9418 }
9419
9420 return 0;
9421 }
9422
9423 /* convert load instructions that access fields of a context type into a
9424 * sequence of instructions that access fields of the underlying structure:
9425 * struct __sk_buff -> struct sk_buff
9426 * struct bpf_sock_ops -> struct sock
9427 */
9428 static int convert_ctx_accesses(struct bpf_verifier_env *env)
9429 {
9430 const struct bpf_verifier_ops *ops = env->ops;
9431 int i, cnt, size, ctx_field_size, delta = 0;
9432 const int insn_cnt = env->prog->len;
9433 struct bpf_insn insn_buf[16], *insn;
9434 u32 target_size, size_default, off;
9435 struct bpf_prog *new_prog;
9436 enum bpf_access_type type;
9437 bool is_narrower_load;
9438
9439 if (ops->gen_prologue || env->seen_direct_write) {
9440 if (!ops->gen_prologue) {
9441 verbose(env, "bpf verifier is misconfigured\n");
9442 return -EINVAL;
9443 }
9444 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
9445 env->prog);
9446 if (cnt >= ARRAY_SIZE(insn_buf)) {
9447 verbose(env, "bpf verifier is misconfigured\n");
9448 return -EINVAL;
9449 } else if (cnt) {
9450 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
9451 if (!new_prog)
9452 return -ENOMEM;
9453
9454 env->prog = new_prog;
9455 delta += cnt - 1;
9456 }
9457 }
9458
9459 if (bpf_prog_is_dev_bound(env->prog->aux))
9460 return 0;
9461
9462 insn = env->prog->insnsi + delta;
9463
9464 for (i = 0; i < insn_cnt; i++, insn++) {
9465 bpf_convert_ctx_access_t convert_ctx_access;
9466
9467 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
9468 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
9469 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
9470 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
9471 type = BPF_READ;
9472 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
9473 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
9474 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
9475 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
9476 type = BPF_WRITE;
9477 else
9478 continue;
9479
9480 if (type == BPF_WRITE &&
9481 env->insn_aux_data[i + delta].sanitize_stack_off) {
9482 struct bpf_insn patch[] = {
9483 /* Sanitize suspicious stack slot with zero.
9484 * There are no memory dependencies for this store,
9485 * since it's only using frame pointer and immediate
9486 * constant of zero
9487 */
9488 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
9489 env->insn_aux_data[i + delta].sanitize_stack_off,
9490 0),
9491 /* the original STX instruction will immediately
9492 * overwrite the same stack slot with appropriate value
9493 */
9494 *insn,
9495 };
9496
9497 cnt = ARRAY_SIZE(patch);
9498 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
9499 if (!new_prog)
9500 return -ENOMEM;
9501
9502 delta += cnt - 1;
9503 env->prog = new_prog;
9504 insn = new_prog->insnsi + i + delta;
9505 continue;
9506 }
9507
9508 switch (env->insn_aux_data[i + delta].ptr_type) {
9509 case PTR_TO_CTX:
9510 if (!ops->convert_ctx_access)
9511 continue;
9512 convert_ctx_access = ops->convert_ctx_access;
9513 break;
9514 case PTR_TO_SOCKET:
9515 case PTR_TO_SOCK_COMMON:
9516 convert_ctx_access = bpf_sock_convert_ctx_access;
9517 break;
9518 case PTR_TO_TCP_SOCK:
9519 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
9520 break;
9521 case PTR_TO_XDP_SOCK:
9522 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
9523 break;
9524 case PTR_TO_BTF_ID:
9525 if (type == BPF_READ) {
9526 insn->code = BPF_LDX | BPF_PROBE_MEM |
9527 BPF_SIZE((insn)->code);
9528 env->prog->aux->num_exentries++;
9529 } else if (env->prog->type != BPF_PROG_TYPE_STRUCT_OPS) {
9530 verbose(env, "Writes through BTF pointers are not allowed\n");
9531 return -EINVAL;
9532 }
9533 continue;
9534 default:
9535 continue;
9536 }
9537
9538 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
9539 size = BPF_LDST_BYTES(insn);
9540
9541 /* If the read access is a narrower load of the field,
9542 * convert to a 4/8-byte load, to minimum program type specific
9543 * convert_ctx_access changes. If conversion is successful,
9544 * we will apply proper mask to the result.
9545 */
9546 is_narrower_load = size < ctx_field_size;
9547 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
9548 off = insn->off;
9549 if (is_narrower_load) {
9550 u8 size_code;
9551
9552 if (type == BPF_WRITE) {
9553 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
9554 return -EINVAL;
9555 }
9556
9557 size_code = BPF_H;
9558 if (ctx_field_size == 4)
9559 size_code = BPF_W;
9560 else if (ctx_field_size == 8)
9561 size_code = BPF_DW;
9562
9563 insn->off = off & ~(size_default - 1);
9564 insn->code = BPF_LDX | BPF_MEM | size_code;
9565 }
9566
9567 target_size = 0;
9568 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
9569 &target_size);
9570 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
9571 (ctx_field_size && !target_size)) {
9572 verbose(env, "bpf verifier is misconfigured\n");
9573 return -EINVAL;
9574 }
9575
9576 if (is_narrower_load && size < target_size) {
9577 u8 shift = bpf_ctx_narrow_access_offset(
9578 off, size, size_default) * 8;
9579 if (ctx_field_size <= 4) {
9580 if (shift)
9581 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
9582 insn->dst_reg,
9583 shift);
9584 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
9585 (1 << size * 8) - 1);
9586 } else {
9587 if (shift)
9588 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
9589 insn->dst_reg,
9590 shift);
9591 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
9592 (1ULL << size * 8) - 1);
9593 }
9594 }
9595
9596 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9597 if (!new_prog)
9598 return -ENOMEM;
9599
9600 delta += cnt - 1;
9601
9602 /* keep walking new program and skip insns we just inserted */
9603 env->prog = new_prog;
9604 insn = new_prog->insnsi + i + delta;
9605 }
9606
9607 return 0;
9608 }
9609
9610 static int jit_subprogs(struct bpf_verifier_env *env)
9611 {
9612 struct bpf_prog *prog = env->prog, **func, *tmp;
9613 int i, j, subprog_start, subprog_end = 0, len, subprog;
9614 struct bpf_insn *insn;
9615 void *old_bpf_func;
9616 int err;
9617
9618 if (env->subprog_cnt <= 1)
9619 return 0;
9620
9621 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9622 if (insn->code != (BPF_JMP | BPF_CALL) ||
9623 insn->src_reg != BPF_PSEUDO_CALL)
9624 continue;
9625 /* Upon error here we cannot fall back to interpreter but
9626 * need a hard reject of the program. Thus -EFAULT is
9627 * propagated in any case.
9628 */
9629 subprog = find_subprog(env, i + insn->imm + 1);
9630 if (subprog < 0) {
9631 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
9632 i + insn->imm + 1);
9633 return -EFAULT;
9634 }
9635 /* temporarily remember subprog id inside insn instead of
9636 * aux_data, since next loop will split up all insns into funcs
9637 */
9638 insn->off = subprog;
9639 /* remember original imm in case JIT fails and fallback
9640 * to interpreter will be needed
9641 */
9642 env->insn_aux_data[i].call_imm = insn->imm;
9643 /* point imm to __bpf_call_base+1 from JITs point of view */
9644 insn->imm = 1;
9645 }
9646
9647 err = bpf_prog_alloc_jited_linfo(prog);
9648 if (err)
9649 goto out_undo_insn;
9650
9651 err = -ENOMEM;
9652 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
9653 if (!func)
9654 goto out_undo_insn;
9655
9656 for (i = 0; i < env->subprog_cnt; i++) {
9657 subprog_start = subprog_end;
9658 subprog_end = env->subprog_info[i + 1].start;
9659
9660 len = subprog_end - subprog_start;
9661 /* BPF_PROG_RUN doesn't call subprogs directly,
9662 * hence main prog stats include the runtime of subprogs.
9663 * subprogs don't have IDs and not reachable via prog_get_next_id
9664 * func[i]->aux->stats will never be accessed and stays NULL
9665 */
9666 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
9667 if (!func[i])
9668 goto out_free;
9669 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
9670 len * sizeof(struct bpf_insn));
9671 func[i]->type = prog->type;
9672 func[i]->len = len;
9673 if (bpf_prog_calc_tag(func[i]))
9674 goto out_free;
9675 func[i]->is_func = 1;
9676 func[i]->aux->func_idx = i;
9677 /* the btf and func_info will be freed only at prog->aux */
9678 func[i]->aux->btf = prog->aux->btf;
9679 func[i]->aux->func_info = prog->aux->func_info;
9680
9681 /* Use bpf_prog_F_tag to indicate functions in stack traces.
9682 * Long term would need debug info to populate names
9683 */
9684 func[i]->aux->name[0] = 'F';
9685 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
9686 func[i]->jit_requested = 1;
9687 func[i]->aux->linfo = prog->aux->linfo;
9688 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
9689 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
9690 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
9691 func[i] = bpf_int_jit_compile(func[i]);
9692 if (!func[i]->jited) {
9693 err = -ENOTSUPP;
9694 goto out_free;
9695 }
9696 cond_resched();
9697 }
9698 /* at this point all bpf functions were successfully JITed
9699 * now populate all bpf_calls with correct addresses and
9700 * run last pass of JIT
9701 */
9702 for (i = 0; i < env->subprog_cnt; i++) {
9703 insn = func[i]->insnsi;
9704 for (j = 0; j < func[i]->len; j++, insn++) {
9705 if (insn->code != (BPF_JMP | BPF_CALL) ||
9706 insn->src_reg != BPF_PSEUDO_CALL)
9707 continue;
9708 subprog = insn->off;
9709 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
9710 __bpf_call_base;
9711 }
9712
9713 /* we use the aux data to keep a list of the start addresses
9714 * of the JITed images for each function in the program
9715 *
9716 * for some architectures, such as powerpc64, the imm field
9717 * might not be large enough to hold the offset of the start
9718 * address of the callee's JITed image from __bpf_call_base
9719 *
9720 * in such cases, we can lookup the start address of a callee
9721 * by using its subprog id, available from the off field of
9722 * the call instruction, as an index for this list
9723 */
9724 func[i]->aux->func = func;
9725 func[i]->aux->func_cnt = env->subprog_cnt;
9726 }
9727 for (i = 0; i < env->subprog_cnt; i++) {
9728 old_bpf_func = func[i]->bpf_func;
9729 tmp = bpf_int_jit_compile(func[i]);
9730 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
9731 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
9732 err = -ENOTSUPP;
9733 goto out_free;
9734 }
9735 cond_resched();
9736 }
9737
9738 /* finally lock prog and jit images for all functions and
9739 * populate kallsysm
9740 */
9741 for (i = 0; i < env->subprog_cnt; i++) {
9742 bpf_prog_lock_ro(func[i]);
9743 bpf_prog_kallsyms_add(func[i]);
9744 }
9745
9746 /* Last step: make now unused interpreter insns from main
9747 * prog consistent for later dump requests, so they can
9748 * later look the same as if they were interpreted only.
9749 */
9750 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9751 if (insn->code != (BPF_JMP | BPF_CALL) ||
9752 insn->src_reg != BPF_PSEUDO_CALL)
9753 continue;
9754 insn->off = env->insn_aux_data[i].call_imm;
9755 subprog = find_subprog(env, i + insn->off + 1);
9756 insn->imm = subprog;
9757 }
9758
9759 prog->jited = 1;
9760 prog->bpf_func = func[0]->bpf_func;
9761 prog->aux->func = func;
9762 prog->aux->func_cnt = env->subprog_cnt;
9763 bpf_prog_free_unused_jited_linfo(prog);
9764 return 0;
9765 out_free:
9766 for (i = 0; i < env->subprog_cnt; i++)
9767 if (func[i])
9768 bpf_jit_free(func[i]);
9769 kfree(func);
9770 out_undo_insn:
9771 /* cleanup main prog to be interpreted */
9772 prog->jit_requested = 0;
9773 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9774 if (insn->code != (BPF_JMP | BPF_CALL) ||
9775 insn->src_reg != BPF_PSEUDO_CALL)
9776 continue;
9777 insn->off = 0;
9778 insn->imm = env->insn_aux_data[i].call_imm;
9779 }
9780 bpf_prog_free_jited_linfo(prog);
9781 return err;
9782 }
9783
9784 static int fixup_call_args(struct bpf_verifier_env *env)
9785 {
9786 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9787 struct bpf_prog *prog = env->prog;
9788 struct bpf_insn *insn = prog->insnsi;
9789 int i, depth;
9790 #endif
9791 int err = 0;
9792
9793 if (env->prog->jit_requested &&
9794 !bpf_prog_is_dev_bound(env->prog->aux)) {
9795 err = jit_subprogs(env);
9796 if (err == 0)
9797 return 0;
9798 if (err == -EFAULT)
9799 return err;
9800 }
9801 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9802 for (i = 0; i < prog->len; i++, insn++) {
9803 if (insn->code != (BPF_JMP | BPF_CALL) ||
9804 insn->src_reg != BPF_PSEUDO_CALL)
9805 continue;
9806 depth = get_callee_stack_depth(env, insn, i);
9807 if (depth < 0)
9808 return depth;
9809 bpf_patch_call_args(insn, depth);
9810 }
9811 err = 0;
9812 #endif
9813 return err;
9814 }
9815
9816 /* fixup insn->imm field of bpf_call instructions
9817 * and inline eligible helpers as explicit sequence of BPF instructions
9818 *
9819 * this function is called after eBPF program passed verification
9820 */
9821 static int fixup_bpf_calls(struct bpf_verifier_env *env)
9822 {
9823 struct bpf_prog *prog = env->prog;
9824 bool expect_blinding = bpf_jit_blinding_enabled(prog);
9825 struct bpf_insn *insn = prog->insnsi;
9826 const struct bpf_func_proto *fn;
9827 const int insn_cnt = prog->len;
9828 const struct bpf_map_ops *ops;
9829 struct bpf_insn_aux_data *aux;
9830 struct bpf_insn insn_buf[16];
9831 struct bpf_prog *new_prog;
9832 struct bpf_map *map_ptr;
9833 int i, ret, cnt, delta = 0;
9834
9835 for (i = 0; i < insn_cnt; i++, insn++) {
9836 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
9837 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9838 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
9839 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9840 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
9841 struct bpf_insn mask_and_div[] = {
9842 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9843 /* Rx div 0 -> 0 */
9844 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
9845 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
9846 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
9847 *insn,
9848 };
9849 struct bpf_insn mask_and_mod[] = {
9850 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9851 /* Rx mod 0 -> Rx */
9852 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
9853 *insn,
9854 };
9855 struct bpf_insn *patchlet;
9856
9857 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9858 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9859 patchlet = mask_and_div + (is64 ? 1 : 0);
9860 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
9861 } else {
9862 patchlet = mask_and_mod + (is64 ? 1 : 0);
9863 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
9864 }
9865
9866 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
9867 if (!new_prog)
9868 return -ENOMEM;
9869
9870 delta += cnt - 1;
9871 env->prog = prog = new_prog;
9872 insn = new_prog->insnsi + i + delta;
9873 continue;
9874 }
9875
9876 if (BPF_CLASS(insn->code) == BPF_LD &&
9877 (BPF_MODE(insn->code) == BPF_ABS ||
9878 BPF_MODE(insn->code) == BPF_IND)) {
9879 cnt = env->ops->gen_ld_abs(insn, insn_buf);
9880 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9881 verbose(env, "bpf verifier is misconfigured\n");
9882 return -EINVAL;
9883 }
9884
9885 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9886 if (!new_prog)
9887 return -ENOMEM;
9888
9889 delta += cnt - 1;
9890 env->prog = prog = new_prog;
9891 insn = new_prog->insnsi + i + delta;
9892 continue;
9893 }
9894
9895 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
9896 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
9897 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
9898 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
9899 struct bpf_insn insn_buf[16];
9900 struct bpf_insn *patch = &insn_buf[0];
9901 bool issrc, isneg;
9902 u32 off_reg;
9903
9904 aux = &env->insn_aux_data[i + delta];
9905 if (!aux->alu_state ||
9906 aux->alu_state == BPF_ALU_NON_POINTER)
9907 continue;
9908
9909 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
9910 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
9911 BPF_ALU_SANITIZE_SRC;
9912
9913 off_reg = issrc ? insn->src_reg : insn->dst_reg;
9914 if (isneg)
9915 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9916 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
9917 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
9918 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
9919 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
9920 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
9921 if (issrc) {
9922 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
9923 off_reg);
9924 insn->src_reg = BPF_REG_AX;
9925 } else {
9926 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
9927 BPF_REG_AX);
9928 }
9929 if (isneg)
9930 insn->code = insn->code == code_add ?
9931 code_sub : code_add;
9932 *patch++ = *insn;
9933 if (issrc && isneg)
9934 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9935 cnt = patch - insn_buf;
9936
9937 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9938 if (!new_prog)
9939 return -ENOMEM;
9940
9941 delta += cnt - 1;
9942 env->prog = prog = new_prog;
9943 insn = new_prog->insnsi + i + delta;
9944 continue;
9945 }
9946
9947 if (insn->code != (BPF_JMP | BPF_CALL))
9948 continue;
9949 if (insn->src_reg == BPF_PSEUDO_CALL)
9950 continue;
9951
9952 if (insn->imm == BPF_FUNC_get_route_realm)
9953 prog->dst_needed = 1;
9954 if (insn->imm == BPF_FUNC_get_prandom_u32)
9955 bpf_user_rnd_init_once();
9956 if (insn->imm == BPF_FUNC_override_return)
9957 prog->kprobe_override = 1;
9958 if (insn->imm == BPF_FUNC_tail_call) {
9959 /* If we tail call into other programs, we
9960 * cannot make any assumptions since they can
9961 * be replaced dynamically during runtime in
9962 * the program array.
9963 */
9964 prog->cb_access = 1;
9965 env->prog->aux->stack_depth = MAX_BPF_STACK;
9966 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
9967
9968 /* mark bpf_tail_call as different opcode to avoid
9969 * conditional branch in the interpeter for every normal
9970 * call and to prevent accidental JITing by JIT compiler
9971 * that doesn't support bpf_tail_call yet
9972 */
9973 insn->imm = 0;
9974 insn->code = BPF_JMP | BPF_TAIL_CALL;
9975
9976 aux = &env->insn_aux_data[i + delta];
9977 if (env->allow_ptr_leaks && !expect_blinding &&
9978 prog->jit_requested &&
9979 !bpf_map_key_poisoned(aux) &&
9980 !bpf_map_ptr_poisoned(aux) &&
9981 !bpf_map_ptr_unpriv(aux)) {
9982 struct bpf_jit_poke_descriptor desc = {
9983 .reason = BPF_POKE_REASON_TAIL_CALL,
9984 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
9985 .tail_call.key = bpf_map_key_immediate(aux),
9986 };
9987
9988 ret = bpf_jit_add_poke_descriptor(prog, &desc);
9989 if (ret < 0) {
9990 verbose(env, "adding tail call poke descriptor failed\n");
9991 return ret;
9992 }
9993
9994 insn->imm = ret + 1;
9995 continue;
9996 }
9997
9998 if (!bpf_map_ptr_unpriv(aux))
9999 continue;
10000
10001 /* instead of changing every JIT dealing with tail_call
10002 * emit two extra insns:
10003 * if (index >= max_entries) goto out;
10004 * index &= array->index_mask;
10005 * to avoid out-of-bounds cpu speculation
10006 */
10007 if (bpf_map_ptr_poisoned(aux)) {
10008 verbose(env, "tail_call abusing map_ptr\n");
10009 return -EINVAL;
10010 }
10011
10012 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
10013 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
10014 map_ptr->max_entries, 2);
10015 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
10016 container_of(map_ptr,
10017 struct bpf_array,
10018 map)->index_mask);
10019 insn_buf[2] = *insn;
10020 cnt = 3;
10021 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10022 if (!new_prog)
10023 return -ENOMEM;
10024
10025 delta += cnt - 1;
10026 env->prog = prog = new_prog;
10027 insn = new_prog->insnsi + i + delta;
10028 continue;
10029 }
10030
10031 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
10032 * and other inlining handlers are currently limited to 64 bit
10033 * only.
10034 */
10035 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10036 (insn->imm == BPF_FUNC_map_lookup_elem ||
10037 insn->imm == BPF_FUNC_map_update_elem ||
10038 insn->imm == BPF_FUNC_map_delete_elem ||
10039 insn->imm == BPF_FUNC_map_push_elem ||
10040 insn->imm == BPF_FUNC_map_pop_elem ||
10041 insn->imm == BPF_FUNC_map_peek_elem)) {
10042 aux = &env->insn_aux_data[i + delta];
10043 if (bpf_map_ptr_poisoned(aux))
10044 goto patch_call_imm;
10045
10046 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
10047 ops = map_ptr->ops;
10048 if (insn->imm == BPF_FUNC_map_lookup_elem &&
10049 ops->map_gen_lookup) {
10050 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
10051 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
10052 verbose(env, "bpf verifier is misconfigured\n");
10053 return -EINVAL;
10054 }
10055
10056 new_prog = bpf_patch_insn_data(env, i + delta,
10057 insn_buf, cnt);
10058 if (!new_prog)
10059 return -ENOMEM;
10060
10061 delta += cnt - 1;
10062 env->prog = prog = new_prog;
10063 insn = new_prog->insnsi + i + delta;
10064 continue;
10065 }
10066
10067 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
10068 (void *(*)(struct bpf_map *map, void *key))NULL));
10069 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
10070 (int (*)(struct bpf_map *map, void *key))NULL));
10071 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
10072 (int (*)(struct bpf_map *map, void *key, void *value,
10073 u64 flags))NULL));
10074 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
10075 (int (*)(struct bpf_map *map, void *value,
10076 u64 flags))NULL));
10077 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
10078 (int (*)(struct bpf_map *map, void *value))NULL));
10079 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
10080 (int (*)(struct bpf_map *map, void *value))NULL));
10081
10082 switch (insn->imm) {
10083 case BPF_FUNC_map_lookup_elem:
10084 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
10085 __bpf_call_base;
10086 continue;
10087 case BPF_FUNC_map_update_elem:
10088 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
10089 __bpf_call_base;
10090 continue;
10091 case BPF_FUNC_map_delete_elem:
10092 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
10093 __bpf_call_base;
10094 continue;
10095 case BPF_FUNC_map_push_elem:
10096 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
10097 __bpf_call_base;
10098 continue;
10099 case BPF_FUNC_map_pop_elem:
10100 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
10101 __bpf_call_base;
10102 continue;
10103 case BPF_FUNC_map_peek_elem:
10104 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
10105 __bpf_call_base;
10106 continue;
10107 }
10108
10109 goto patch_call_imm;
10110 }
10111
10112 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10113 insn->imm == BPF_FUNC_jiffies64) {
10114 struct bpf_insn ld_jiffies_addr[2] = {
10115 BPF_LD_IMM64(BPF_REG_0,
10116 (unsigned long)&jiffies),
10117 };
10118
10119 insn_buf[0] = ld_jiffies_addr[0];
10120 insn_buf[1] = ld_jiffies_addr[1];
10121 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
10122 BPF_REG_0, 0);
10123 cnt = 3;
10124
10125 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
10126 cnt);
10127 if (!new_prog)
10128 return -ENOMEM;
10129
10130 delta += cnt - 1;
10131 env->prog = prog = new_prog;
10132 insn = new_prog->insnsi + i + delta;
10133 continue;
10134 }
10135
10136 patch_call_imm:
10137 fn = env->ops->get_func_proto(insn->imm, env->prog);
10138 /* all functions that have prototype and verifier allowed
10139 * programs to call them, must be real in-kernel functions
10140 */
10141 if (!fn->func) {
10142 verbose(env,
10143 "kernel subsystem misconfigured func %s#%d\n",
10144 func_id_name(insn->imm), insn->imm);
10145 return -EFAULT;
10146 }
10147 insn->imm = fn->func - __bpf_call_base;
10148 }
10149
10150 /* Since poke tab is now finalized, publish aux to tracker. */
10151 for (i = 0; i < prog->aux->size_poke_tab; i++) {
10152 map_ptr = prog->aux->poke_tab[i].tail_call.map;
10153 if (!map_ptr->ops->map_poke_track ||
10154 !map_ptr->ops->map_poke_untrack ||
10155 !map_ptr->ops->map_poke_run) {
10156 verbose(env, "bpf verifier is misconfigured\n");
10157 return -EINVAL;
10158 }
10159
10160 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
10161 if (ret < 0) {
10162 verbose(env, "tracking tail call prog failed\n");
10163 return ret;
10164 }
10165 }
10166
10167 return 0;
10168 }
10169
10170 static void free_states(struct bpf_verifier_env *env)
10171 {
10172 struct bpf_verifier_state_list *sl, *sln;
10173 int i;
10174
10175 sl = env->free_list;
10176 while (sl) {
10177 sln = sl->next;
10178 free_verifier_state(&sl->state, false);
10179 kfree(sl);
10180 sl = sln;
10181 }
10182 env->free_list = NULL;
10183
10184 if (!env->explored_states)
10185 return;
10186
10187 for (i = 0; i < state_htab_size(env); i++) {
10188 sl = env->explored_states[i];
10189
10190 while (sl) {
10191 sln = sl->next;
10192 free_verifier_state(&sl->state, false);
10193 kfree(sl);
10194 sl = sln;
10195 }
10196 env->explored_states[i] = NULL;
10197 }
10198 }
10199
10200 /* The verifier is using insn_aux_data[] to store temporary data during
10201 * verification and to store information for passes that run after the
10202 * verification like dead code sanitization. do_check_common() for subprogram N
10203 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
10204 * temporary data after do_check_common() finds that subprogram N cannot be
10205 * verified independently. pass_cnt counts the number of times
10206 * do_check_common() was run and insn->aux->seen tells the pass number
10207 * insn_aux_data was touched. These variables are compared to clear temporary
10208 * data from failed pass. For testing and experiments do_check_common() can be
10209 * run multiple times even when prior attempt to verify is unsuccessful.
10210 */
10211 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
10212 {
10213 struct bpf_insn *insn = env->prog->insnsi;
10214 struct bpf_insn_aux_data *aux;
10215 int i, class;
10216
10217 for (i = 0; i < env->prog->len; i++) {
10218 class = BPF_CLASS(insn[i].code);
10219 if (class != BPF_LDX && class != BPF_STX)
10220 continue;
10221 aux = &env->insn_aux_data[i];
10222 if (aux->seen != env->pass_cnt)
10223 continue;
10224 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
10225 }
10226 }
10227
10228 static int do_check_common(struct bpf_verifier_env *env, int subprog)
10229 {
10230 struct bpf_verifier_state *state;
10231 struct bpf_reg_state *regs;
10232 int ret, i;
10233
10234 env->prev_linfo = NULL;
10235 env->pass_cnt++;
10236
10237 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
10238 if (!state)
10239 return -ENOMEM;
10240 state->curframe = 0;
10241 state->speculative = false;
10242 state->branches = 1;
10243 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
10244 if (!state->frame[0]) {
10245 kfree(state);
10246 return -ENOMEM;
10247 }
10248 env->cur_state = state;
10249 init_func_state(env, state->frame[0],
10250 BPF_MAIN_FUNC /* callsite */,
10251 0 /* frameno */,
10252 subprog);
10253
10254 regs = state->frame[state->curframe]->regs;
10255 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
10256 ret = btf_prepare_func_args(env, subprog, regs);
10257 if (ret)
10258 goto out;
10259 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
10260 if (regs[i].type == PTR_TO_CTX)
10261 mark_reg_known_zero(env, regs, i);
10262 else if (regs[i].type == SCALAR_VALUE)
10263 mark_reg_unknown(env, regs, i);
10264 }
10265 } else {
10266 /* 1st arg to a function */
10267 regs[BPF_REG_1].type = PTR_TO_CTX;
10268 mark_reg_known_zero(env, regs, BPF_REG_1);
10269 ret = btf_check_func_arg_match(env, subprog, regs);
10270 if (ret == -EFAULT)
10271 /* unlikely verifier bug. abort.
10272 * ret == 0 and ret < 0 are sadly acceptable for
10273 * main() function due to backward compatibility.
10274 * Like socket filter program may be written as:
10275 * int bpf_prog(struct pt_regs *ctx)
10276 * and never dereference that ctx in the program.
10277 * 'struct pt_regs' is a type mismatch for socket
10278 * filter that should be using 'struct __sk_buff'.
10279 */
10280 goto out;
10281 }
10282
10283 ret = do_check(env);
10284 out:
10285 /* check for NULL is necessary, since cur_state can be freed inside
10286 * do_check() under memory pressure.
10287 */
10288 if (env->cur_state) {
10289 free_verifier_state(env->cur_state, true);
10290 env->cur_state = NULL;
10291 }
10292 while (!pop_stack(env, NULL, NULL));
10293 free_states(env);
10294 if (ret)
10295 /* clean aux data in case subprog was rejected */
10296 sanitize_insn_aux_data(env);
10297 return ret;
10298 }
10299
10300 /* Verify all global functions in a BPF program one by one based on their BTF.
10301 * All global functions must pass verification. Otherwise the whole program is rejected.
10302 * Consider:
10303 * int bar(int);
10304 * int foo(int f)
10305 * {
10306 * return bar(f);
10307 * }
10308 * int bar(int b)
10309 * {
10310 * ...
10311 * }
10312 * foo() will be verified first for R1=any_scalar_value. During verification it
10313 * will be assumed that bar() already verified successfully and call to bar()
10314 * from foo() will be checked for type match only. Later bar() will be verified
10315 * independently to check that it's safe for R1=any_scalar_value.
10316 */
10317 static int do_check_subprogs(struct bpf_verifier_env *env)
10318 {
10319 struct bpf_prog_aux *aux = env->prog->aux;
10320 int i, ret;
10321
10322 if (!aux->func_info)
10323 return 0;
10324
10325 for (i = 1; i < env->subprog_cnt; i++) {
10326 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
10327 continue;
10328 env->insn_idx = env->subprog_info[i].start;
10329 WARN_ON_ONCE(env->insn_idx == 0);
10330 ret = do_check_common(env, i);
10331 if (ret) {
10332 return ret;
10333 } else if (env->log.level & BPF_LOG_LEVEL) {
10334 verbose(env,
10335 "Func#%d is safe for any args that match its prototype\n",
10336 i);
10337 }
10338 }
10339 return 0;
10340 }
10341
10342 static int do_check_main(struct bpf_verifier_env *env)
10343 {
10344 int ret;
10345
10346 env->insn_idx = 0;
10347 ret = do_check_common(env, 0);
10348 if (!ret)
10349 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
10350 return ret;
10351 }
10352
10353
10354 static void print_verification_stats(struct bpf_verifier_env *env)
10355 {
10356 int i;
10357
10358 if (env->log.level & BPF_LOG_STATS) {
10359 verbose(env, "verification time %lld usec\n",
10360 div_u64(env->verification_time, 1000));
10361 verbose(env, "stack depth ");
10362 for (i = 0; i < env->subprog_cnt; i++) {
10363 u32 depth = env->subprog_info[i].stack_depth;
10364
10365 verbose(env, "%d", depth);
10366 if (i + 1 < env->subprog_cnt)
10367 verbose(env, "+");
10368 }
10369 verbose(env, "\n");
10370 }
10371 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
10372 "total_states %d peak_states %d mark_read %d\n",
10373 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
10374 env->max_states_per_insn, env->total_states,
10375 env->peak_states, env->longest_mark_read_walk);
10376 }
10377
10378 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
10379 {
10380 const struct btf_type *t, *func_proto;
10381 const struct bpf_struct_ops *st_ops;
10382 const struct btf_member *member;
10383 struct bpf_prog *prog = env->prog;
10384 u32 btf_id, member_idx;
10385 const char *mname;
10386
10387 btf_id = prog->aux->attach_btf_id;
10388 st_ops = bpf_struct_ops_find(btf_id);
10389 if (!st_ops) {
10390 verbose(env, "attach_btf_id %u is not a supported struct\n",
10391 btf_id);
10392 return -ENOTSUPP;
10393 }
10394
10395 t = st_ops->type;
10396 member_idx = prog->expected_attach_type;
10397 if (member_idx >= btf_type_vlen(t)) {
10398 verbose(env, "attach to invalid member idx %u of struct %s\n",
10399 member_idx, st_ops->name);
10400 return -EINVAL;
10401 }
10402
10403 member = &btf_type_member(t)[member_idx];
10404 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
10405 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
10406 NULL);
10407 if (!func_proto) {
10408 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
10409 mname, member_idx, st_ops->name);
10410 return -EINVAL;
10411 }
10412
10413 if (st_ops->check_member) {
10414 int err = st_ops->check_member(t, member);
10415
10416 if (err) {
10417 verbose(env, "attach to unsupported member %s of struct %s\n",
10418 mname, st_ops->name);
10419 return err;
10420 }
10421 }
10422
10423 prog->aux->attach_func_proto = func_proto;
10424 prog->aux->attach_func_name = mname;
10425 env->ops = st_ops->verifier_ops;
10426
10427 return 0;
10428 }
10429 #define SECURITY_PREFIX "security_"
10430
10431 static int check_attach_modify_return(struct bpf_verifier_env *env)
10432 {
10433 struct bpf_prog *prog = env->prog;
10434 unsigned long addr = (unsigned long) prog->aux->trampoline->func.addr;
10435
10436 /* This is expected to be cleaned up in the future with the KRSI effort
10437 * introducing the LSM_HOOK macro for cleaning up lsm_hooks.h.
10438 */
10439 if (within_error_injection_list(addr) ||
10440 !strncmp(SECURITY_PREFIX, prog->aux->attach_func_name,
10441 sizeof(SECURITY_PREFIX) - 1))
10442 return 0;
10443
10444 verbose(env, "fmod_ret attach_btf_id %u (%s) is not modifiable\n",
10445 prog->aux->attach_btf_id, prog->aux->attach_func_name);
10446
10447 return -EINVAL;
10448 }
10449
10450 static int check_attach_btf_id(struct bpf_verifier_env *env)
10451 {
10452 struct bpf_prog *prog = env->prog;
10453 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
10454 struct bpf_prog *tgt_prog = prog->aux->linked_prog;
10455 u32 btf_id = prog->aux->attach_btf_id;
10456 const char prefix[] = "btf_trace_";
10457 int ret = 0, subprog = -1, i;
10458 struct bpf_trampoline *tr;
10459 const struct btf_type *t;
10460 bool conservative = true;
10461 const char *tname;
10462 struct btf *btf;
10463 long addr;
10464 u64 key;
10465
10466 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
10467 return check_struct_ops_btf_id(env);
10468
10469 if (prog->type != BPF_PROG_TYPE_TRACING &&
10470 prog->type != BPF_PROG_TYPE_LSM &&
10471 !prog_extension)
10472 return 0;
10473
10474 if (!btf_id) {
10475 verbose(env, "Tracing programs must provide btf_id\n");
10476 return -EINVAL;
10477 }
10478 btf = bpf_prog_get_target_btf(prog);
10479 if (!btf) {
10480 verbose(env,
10481 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
10482 return -EINVAL;
10483 }
10484 t = btf_type_by_id(btf, btf_id);
10485 if (!t) {
10486 verbose(env, "attach_btf_id %u is invalid\n", btf_id);
10487 return -EINVAL;
10488 }
10489 tname = btf_name_by_offset(btf, t->name_off);
10490 if (!tname) {
10491 verbose(env, "attach_btf_id %u doesn't have a name\n", btf_id);
10492 return -EINVAL;
10493 }
10494 if (tgt_prog) {
10495 struct bpf_prog_aux *aux = tgt_prog->aux;
10496
10497 for (i = 0; i < aux->func_info_cnt; i++)
10498 if (aux->func_info[i].type_id == btf_id) {
10499 subprog = i;
10500 break;
10501 }
10502 if (subprog == -1) {
10503 verbose(env, "Subprog %s doesn't exist\n", tname);
10504 return -EINVAL;
10505 }
10506 conservative = aux->func_info_aux[subprog].unreliable;
10507 if (prog_extension) {
10508 if (conservative) {
10509 verbose(env,
10510 "Cannot replace static functions\n");
10511 return -EINVAL;
10512 }
10513 if (!prog->jit_requested) {
10514 verbose(env,
10515 "Extension programs should be JITed\n");
10516 return -EINVAL;
10517 }
10518 env->ops = bpf_verifier_ops[tgt_prog->type];
10519 prog->expected_attach_type = tgt_prog->expected_attach_type;
10520 }
10521 if (!tgt_prog->jited) {
10522 verbose(env, "Can attach to only JITed progs\n");
10523 return -EINVAL;
10524 }
10525 if (tgt_prog->type == prog->type) {
10526 /* Cannot fentry/fexit another fentry/fexit program.
10527 * Cannot attach program extension to another extension.
10528 * It's ok to attach fentry/fexit to extension program.
10529 */
10530 verbose(env, "Cannot recursively attach\n");
10531 return -EINVAL;
10532 }
10533 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
10534 prog_extension &&
10535 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
10536 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
10537 /* Program extensions can extend all program types
10538 * except fentry/fexit. The reason is the following.
10539 * The fentry/fexit programs are used for performance
10540 * analysis, stats and can be attached to any program
10541 * type except themselves. When extension program is
10542 * replacing XDP function it is necessary to allow
10543 * performance analysis of all functions. Both original
10544 * XDP program and its program extension. Hence
10545 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
10546 * allowed. If extending of fentry/fexit was allowed it
10547 * would be possible to create long call chain
10548 * fentry->extension->fentry->extension beyond
10549 * reasonable stack size. Hence extending fentry is not
10550 * allowed.
10551 */
10552 verbose(env, "Cannot extend fentry/fexit\n");
10553 return -EINVAL;
10554 }
10555 key = ((u64)aux->id) << 32 | btf_id;
10556 } else {
10557 if (prog_extension) {
10558 verbose(env, "Cannot replace kernel functions\n");
10559 return -EINVAL;
10560 }
10561 key = btf_id;
10562 }
10563
10564 switch (prog->expected_attach_type) {
10565 case BPF_TRACE_RAW_TP:
10566 if (tgt_prog) {
10567 verbose(env,
10568 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
10569 return -EINVAL;
10570 }
10571 if (!btf_type_is_typedef(t)) {
10572 verbose(env, "attach_btf_id %u is not a typedef\n",
10573 btf_id);
10574 return -EINVAL;
10575 }
10576 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
10577 verbose(env, "attach_btf_id %u points to wrong type name %s\n",
10578 btf_id, tname);
10579 return -EINVAL;
10580 }
10581 tname += sizeof(prefix) - 1;
10582 t = btf_type_by_id(btf, t->type);
10583 if (!btf_type_is_ptr(t))
10584 /* should never happen in valid vmlinux build */
10585 return -EINVAL;
10586 t = btf_type_by_id(btf, t->type);
10587 if (!btf_type_is_func_proto(t))
10588 /* should never happen in valid vmlinux build */
10589 return -EINVAL;
10590
10591 /* remember two read only pointers that are valid for
10592 * the life time of the kernel
10593 */
10594 prog->aux->attach_func_name = tname;
10595 prog->aux->attach_func_proto = t;
10596 prog->aux->attach_btf_trace = true;
10597 return 0;
10598 default:
10599 if (!prog_extension)
10600 return -EINVAL;
10601 /* fallthrough */
10602 case BPF_MODIFY_RETURN:
10603 case BPF_LSM_MAC:
10604 case BPF_TRACE_FENTRY:
10605 case BPF_TRACE_FEXIT:
10606 prog->aux->attach_func_name = tname;
10607 if (prog->type == BPF_PROG_TYPE_LSM) {
10608 ret = bpf_lsm_verify_prog(&env->log, prog);
10609 if (ret < 0)
10610 return ret;
10611 }
10612
10613 if (!btf_type_is_func(t)) {
10614 verbose(env, "attach_btf_id %u is not a function\n",
10615 btf_id);
10616 return -EINVAL;
10617 }
10618 if (prog_extension &&
10619 btf_check_type_match(env, prog, btf, t))
10620 return -EINVAL;
10621 t = btf_type_by_id(btf, t->type);
10622 if (!btf_type_is_func_proto(t))
10623 return -EINVAL;
10624 tr = bpf_trampoline_lookup(key);
10625 if (!tr)
10626 return -ENOMEM;
10627 /* t is either vmlinux type or another program's type */
10628 prog->aux->attach_func_proto = t;
10629 mutex_lock(&tr->mutex);
10630 if (tr->func.addr) {
10631 prog->aux->trampoline = tr;
10632 goto out;
10633 }
10634 if (tgt_prog && conservative) {
10635 prog->aux->attach_func_proto = NULL;
10636 t = NULL;
10637 }
10638 ret = btf_distill_func_proto(&env->log, btf, t,
10639 tname, &tr->func.model);
10640 if (ret < 0)
10641 goto out;
10642 if (tgt_prog) {
10643 if (subprog == 0)
10644 addr = (long) tgt_prog->bpf_func;
10645 else
10646 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
10647 } else {
10648 addr = kallsyms_lookup_name(tname);
10649 if (!addr) {
10650 verbose(env,
10651 "The address of function %s cannot be found\n",
10652 tname);
10653 ret = -ENOENT;
10654 goto out;
10655 }
10656 }
10657 tr->func.addr = (void *)addr;
10658 prog->aux->trampoline = tr;
10659
10660 if (prog->expected_attach_type == BPF_MODIFY_RETURN)
10661 ret = check_attach_modify_return(env);
10662 out:
10663 mutex_unlock(&tr->mutex);
10664 if (ret)
10665 bpf_trampoline_put(tr);
10666 return ret;
10667 }
10668 }
10669
10670 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
10671 union bpf_attr __user *uattr)
10672 {
10673 u64 start_time = ktime_get_ns();
10674 struct bpf_verifier_env *env;
10675 struct bpf_verifier_log *log;
10676 int i, len, ret = -EINVAL;
10677 bool is_priv;
10678
10679 /* no program is valid */
10680 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
10681 return -EINVAL;
10682
10683 /* 'struct bpf_verifier_env' can be global, but since it's not small,
10684 * allocate/free it every time bpf_check() is called
10685 */
10686 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
10687 if (!env)
10688 return -ENOMEM;
10689 log = &env->log;
10690
10691 len = (*prog)->len;
10692 env->insn_aux_data =
10693 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
10694 ret = -ENOMEM;
10695 if (!env->insn_aux_data)
10696 goto err_free_env;
10697 for (i = 0; i < len; i++)
10698 env->insn_aux_data[i].orig_idx = i;
10699 env->prog = *prog;
10700 env->ops = bpf_verifier_ops[env->prog->type];
10701 is_priv = capable(CAP_SYS_ADMIN);
10702
10703 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
10704 mutex_lock(&bpf_verifier_lock);
10705 if (!btf_vmlinux)
10706 btf_vmlinux = btf_parse_vmlinux();
10707 mutex_unlock(&bpf_verifier_lock);
10708 }
10709
10710 /* grab the mutex to protect few globals used by verifier */
10711 if (!is_priv)
10712 mutex_lock(&bpf_verifier_lock);
10713
10714 if (attr->log_level || attr->log_buf || attr->log_size) {
10715 /* user requested verbose verifier output
10716 * and supplied buffer to store the verification trace
10717 */
10718 log->level = attr->log_level;
10719 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
10720 log->len_total = attr->log_size;
10721
10722 ret = -EINVAL;
10723 /* log attributes have to be sane */
10724 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
10725 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
10726 goto err_unlock;
10727 }
10728
10729 if (IS_ERR(btf_vmlinux)) {
10730 /* Either gcc or pahole or kernel are broken. */
10731 verbose(env, "in-kernel BTF is malformed\n");
10732 ret = PTR_ERR(btf_vmlinux);
10733 goto skip_full_check;
10734 }
10735
10736 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
10737 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
10738 env->strict_alignment = true;
10739 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
10740 env->strict_alignment = false;
10741
10742 env->allow_ptr_leaks = is_priv;
10743
10744 if (is_priv)
10745 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
10746
10747 ret = replace_map_fd_with_map_ptr(env);
10748 if (ret < 0)
10749 goto skip_full_check;
10750
10751 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10752 ret = bpf_prog_offload_verifier_prep(env->prog);
10753 if (ret)
10754 goto skip_full_check;
10755 }
10756
10757 env->explored_states = kvcalloc(state_htab_size(env),
10758 sizeof(struct bpf_verifier_state_list *),
10759 GFP_USER);
10760 ret = -ENOMEM;
10761 if (!env->explored_states)
10762 goto skip_full_check;
10763
10764 ret = check_subprogs(env);
10765 if (ret < 0)
10766 goto skip_full_check;
10767
10768 ret = check_btf_info(env, attr, uattr);
10769 if (ret < 0)
10770 goto skip_full_check;
10771
10772 ret = check_attach_btf_id(env);
10773 if (ret)
10774 goto skip_full_check;
10775
10776 ret = check_cfg(env);
10777 if (ret < 0)
10778 goto skip_full_check;
10779
10780 ret = do_check_subprogs(env);
10781 ret = ret ?: do_check_main(env);
10782
10783 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
10784 ret = bpf_prog_offload_finalize(env);
10785
10786 skip_full_check:
10787 kvfree(env->explored_states);
10788
10789 if (ret == 0)
10790 ret = check_max_stack_depth(env);
10791
10792 /* instruction rewrites happen after this point */
10793 if (is_priv) {
10794 if (ret == 0)
10795 opt_hard_wire_dead_code_branches(env);
10796 if (ret == 0)
10797 ret = opt_remove_dead_code(env);
10798 if (ret == 0)
10799 ret = opt_remove_nops(env);
10800 } else {
10801 if (ret == 0)
10802 sanitize_dead_code(env);
10803 }
10804
10805 if (ret == 0)
10806 /* program is valid, convert *(u32*)(ctx + off) accesses */
10807 ret = convert_ctx_accesses(env);
10808
10809 if (ret == 0)
10810 ret = fixup_bpf_calls(env);
10811
10812 /* do 32-bit optimization after insn patching has done so those patched
10813 * insns could be handled correctly.
10814 */
10815 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
10816 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
10817 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
10818 : false;
10819 }
10820
10821 if (ret == 0)
10822 ret = fixup_call_args(env);
10823
10824 env->verification_time = ktime_get_ns() - start_time;
10825 print_verification_stats(env);
10826
10827 if (log->level && bpf_verifier_log_full(log))
10828 ret = -ENOSPC;
10829 if (log->level && !log->ubuf) {
10830 ret = -EFAULT;
10831 goto err_release_maps;
10832 }
10833
10834 if (ret == 0 && env->used_map_cnt) {
10835 /* if program passed verifier, update used_maps in bpf_prog_info */
10836 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
10837 sizeof(env->used_maps[0]),
10838 GFP_KERNEL);
10839
10840 if (!env->prog->aux->used_maps) {
10841 ret = -ENOMEM;
10842 goto err_release_maps;
10843 }
10844
10845 memcpy(env->prog->aux->used_maps, env->used_maps,
10846 sizeof(env->used_maps[0]) * env->used_map_cnt);
10847 env->prog->aux->used_map_cnt = env->used_map_cnt;
10848
10849 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
10850 * bpf_ld_imm64 instructions
10851 */
10852 convert_pseudo_ld_imm64(env);
10853 }
10854
10855 if (ret == 0)
10856 adjust_btf_func(env);
10857
10858 err_release_maps:
10859 if (!env->prog->aux->used_maps)
10860 /* if we didn't copy map pointers into bpf_prog_info, release
10861 * them now. Otherwise free_used_maps() will release them.
10862 */
10863 release_maps(env);
10864
10865 /* extension progs temporarily inherit the attach_type of their targets
10866 for verification purposes, so set it back to zero before returning
10867 */
10868 if (env->prog->type == BPF_PROG_TYPE_EXT)
10869 env->prog->expected_attach_type = 0;
10870
10871 *prog = env->prog;
10872 err_unlock:
10873 if (!is_priv)
10874 mutex_unlock(&bpf_verifier_lock);
10875 vfree(env->insn_aux_data);
10876 err_free_env:
10877 kfree(env);
10878 return ret;
10879 }