]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - kernel/bpf/verifier.c
bpf: Use a table to drive helper arg type checks
[mirror_ubuntu-hirsute-kernel.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 #include <linux/btf_ids.h>
25
26 #include "disasm.h"
27
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
34 #undef BPF_PROG_TYPE
35 #undef BPF_MAP_TYPE
36 #undef BPF_LINK_TYPE
37 };
38
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
42 *
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all pathes through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
54 *
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
58 * copied to R1.
59 *
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
65 *
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
68 *
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
78 *
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
82 *
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
86 *
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
89 *
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
92 *
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
97 *
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
102 *
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
107 *
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
110 * {
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
113 * void *value;
114 *
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
118 * }
119 *
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
128 *
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns ether pointer to map value or NULL.
136 *
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
141 *
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
144 *
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
147 * the BPF program:
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
149 *
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
156 *
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
162 */
163
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
169 */
170 struct bpf_verifier_state st;
171 int insn_idx;
172 int prev_insn_idx;
173 struct bpf_verifier_stack_elem *next;
174 /* length of verifier log at the time this state was pushed on stack */
175 u32 log_pos;
176 };
177
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
180
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
183
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
188
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
190 {
191 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
192 }
193
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
195 {
196 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
197 }
198
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
200 const struct bpf_map *map, bool unpriv)
201 {
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
203 unpriv |= bpf_map_ptr_unpriv(aux);
204 aux->map_ptr_state = (unsigned long)map |
205 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
206 }
207
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
209 {
210 return aux->map_key_state & BPF_MAP_KEY_POISON;
211 }
212
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
214 {
215 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
216 }
217
218 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
219 {
220 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
221 }
222
223 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
224 {
225 bool poisoned = bpf_map_key_poisoned(aux);
226
227 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
228 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
229 }
230
231 struct bpf_call_arg_meta {
232 struct bpf_map *map_ptr;
233 bool raw_mode;
234 bool pkt_access;
235 int regno;
236 int access_size;
237 int mem_size;
238 u64 msize_max_value;
239 int ref_obj_id;
240 int func_id;
241 };
242
243 struct btf *btf_vmlinux;
244
245 static DEFINE_MUTEX(bpf_verifier_lock);
246
247 static const struct bpf_line_info *
248 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
249 {
250 const struct bpf_line_info *linfo;
251 const struct bpf_prog *prog;
252 u32 i, nr_linfo;
253
254 prog = env->prog;
255 nr_linfo = prog->aux->nr_linfo;
256
257 if (!nr_linfo || insn_off >= prog->len)
258 return NULL;
259
260 linfo = prog->aux->linfo;
261 for (i = 1; i < nr_linfo; i++)
262 if (insn_off < linfo[i].insn_off)
263 break;
264
265 return &linfo[i - 1];
266 }
267
268 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
269 va_list args)
270 {
271 unsigned int n;
272
273 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
274
275 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
276 "verifier log line truncated - local buffer too short\n");
277
278 n = min(log->len_total - log->len_used - 1, n);
279 log->kbuf[n] = '\0';
280
281 if (log->level == BPF_LOG_KERNEL) {
282 pr_err("BPF:%s\n", log->kbuf);
283 return;
284 }
285 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
286 log->len_used += n;
287 else
288 log->ubuf = NULL;
289 }
290
291 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
292 {
293 char zero = 0;
294
295 if (!bpf_verifier_log_needed(log))
296 return;
297
298 log->len_used = new_pos;
299 if (put_user(zero, log->ubuf + new_pos))
300 log->ubuf = NULL;
301 }
302
303 /* log_level controls verbosity level of eBPF verifier.
304 * bpf_verifier_log_write() is used to dump the verification trace to the log,
305 * so the user can figure out what's wrong with the program
306 */
307 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
308 const char *fmt, ...)
309 {
310 va_list args;
311
312 if (!bpf_verifier_log_needed(&env->log))
313 return;
314
315 va_start(args, fmt);
316 bpf_verifier_vlog(&env->log, fmt, args);
317 va_end(args);
318 }
319 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
320
321 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
322 {
323 struct bpf_verifier_env *env = private_data;
324 va_list args;
325
326 if (!bpf_verifier_log_needed(&env->log))
327 return;
328
329 va_start(args, fmt);
330 bpf_verifier_vlog(&env->log, fmt, args);
331 va_end(args);
332 }
333
334 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
335 const char *fmt, ...)
336 {
337 va_list args;
338
339 if (!bpf_verifier_log_needed(log))
340 return;
341
342 va_start(args, fmt);
343 bpf_verifier_vlog(log, fmt, args);
344 va_end(args);
345 }
346
347 static const char *ltrim(const char *s)
348 {
349 while (isspace(*s))
350 s++;
351
352 return s;
353 }
354
355 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
356 u32 insn_off,
357 const char *prefix_fmt, ...)
358 {
359 const struct bpf_line_info *linfo;
360
361 if (!bpf_verifier_log_needed(&env->log))
362 return;
363
364 linfo = find_linfo(env, insn_off);
365 if (!linfo || linfo == env->prev_linfo)
366 return;
367
368 if (prefix_fmt) {
369 va_list args;
370
371 va_start(args, prefix_fmt);
372 bpf_verifier_vlog(&env->log, prefix_fmt, args);
373 va_end(args);
374 }
375
376 verbose(env, "%s\n",
377 ltrim(btf_name_by_offset(env->prog->aux->btf,
378 linfo->line_off)));
379
380 env->prev_linfo = linfo;
381 }
382
383 static bool type_is_pkt_pointer(enum bpf_reg_type type)
384 {
385 return type == PTR_TO_PACKET ||
386 type == PTR_TO_PACKET_META;
387 }
388
389 static bool type_is_sk_pointer(enum bpf_reg_type type)
390 {
391 return type == PTR_TO_SOCKET ||
392 type == PTR_TO_SOCK_COMMON ||
393 type == PTR_TO_TCP_SOCK ||
394 type == PTR_TO_XDP_SOCK;
395 }
396
397 static bool reg_type_not_null(enum bpf_reg_type type)
398 {
399 return type == PTR_TO_SOCKET ||
400 type == PTR_TO_TCP_SOCK ||
401 type == PTR_TO_MAP_VALUE ||
402 type == PTR_TO_SOCK_COMMON;
403 }
404
405 static bool reg_type_may_be_null(enum bpf_reg_type type)
406 {
407 return type == PTR_TO_MAP_VALUE_OR_NULL ||
408 type == PTR_TO_SOCKET_OR_NULL ||
409 type == PTR_TO_SOCK_COMMON_OR_NULL ||
410 type == PTR_TO_TCP_SOCK_OR_NULL ||
411 type == PTR_TO_BTF_ID_OR_NULL ||
412 type == PTR_TO_MEM_OR_NULL ||
413 type == PTR_TO_RDONLY_BUF_OR_NULL ||
414 type == PTR_TO_RDWR_BUF_OR_NULL;
415 }
416
417 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
418 {
419 return reg->type == PTR_TO_MAP_VALUE &&
420 map_value_has_spin_lock(reg->map_ptr);
421 }
422
423 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
424 {
425 return type == PTR_TO_SOCKET ||
426 type == PTR_TO_SOCKET_OR_NULL ||
427 type == PTR_TO_TCP_SOCK ||
428 type == PTR_TO_TCP_SOCK_OR_NULL ||
429 type == PTR_TO_MEM ||
430 type == PTR_TO_MEM_OR_NULL;
431 }
432
433 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
434 {
435 return type == ARG_PTR_TO_SOCK_COMMON;
436 }
437
438 static bool arg_type_may_be_null(enum bpf_arg_type type)
439 {
440 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
441 type == ARG_PTR_TO_MEM_OR_NULL ||
442 type == ARG_PTR_TO_CTX_OR_NULL ||
443 type == ARG_PTR_TO_SOCKET_OR_NULL ||
444 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL;
445 }
446
447 /* Determine whether the function releases some resources allocated by another
448 * function call. The first reference type argument will be assumed to be
449 * released by release_reference().
450 */
451 static bool is_release_function(enum bpf_func_id func_id)
452 {
453 return func_id == BPF_FUNC_sk_release ||
454 func_id == BPF_FUNC_ringbuf_submit ||
455 func_id == BPF_FUNC_ringbuf_discard;
456 }
457
458 static bool may_be_acquire_function(enum bpf_func_id func_id)
459 {
460 return func_id == BPF_FUNC_sk_lookup_tcp ||
461 func_id == BPF_FUNC_sk_lookup_udp ||
462 func_id == BPF_FUNC_skc_lookup_tcp ||
463 func_id == BPF_FUNC_map_lookup_elem ||
464 func_id == BPF_FUNC_ringbuf_reserve;
465 }
466
467 static bool is_acquire_function(enum bpf_func_id func_id,
468 const struct bpf_map *map)
469 {
470 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
471
472 if (func_id == BPF_FUNC_sk_lookup_tcp ||
473 func_id == BPF_FUNC_sk_lookup_udp ||
474 func_id == BPF_FUNC_skc_lookup_tcp ||
475 func_id == BPF_FUNC_ringbuf_reserve)
476 return true;
477
478 if (func_id == BPF_FUNC_map_lookup_elem &&
479 (map_type == BPF_MAP_TYPE_SOCKMAP ||
480 map_type == BPF_MAP_TYPE_SOCKHASH))
481 return true;
482
483 return false;
484 }
485
486 static bool is_ptr_cast_function(enum bpf_func_id func_id)
487 {
488 return func_id == BPF_FUNC_tcp_sock ||
489 func_id == BPF_FUNC_sk_fullsock;
490 }
491
492 /* string representation of 'enum bpf_reg_type' */
493 static const char * const reg_type_str[] = {
494 [NOT_INIT] = "?",
495 [SCALAR_VALUE] = "inv",
496 [PTR_TO_CTX] = "ctx",
497 [CONST_PTR_TO_MAP] = "map_ptr",
498 [PTR_TO_MAP_VALUE] = "map_value",
499 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
500 [PTR_TO_STACK] = "fp",
501 [PTR_TO_PACKET] = "pkt",
502 [PTR_TO_PACKET_META] = "pkt_meta",
503 [PTR_TO_PACKET_END] = "pkt_end",
504 [PTR_TO_FLOW_KEYS] = "flow_keys",
505 [PTR_TO_SOCKET] = "sock",
506 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
507 [PTR_TO_SOCK_COMMON] = "sock_common",
508 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
509 [PTR_TO_TCP_SOCK] = "tcp_sock",
510 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
511 [PTR_TO_TP_BUFFER] = "tp_buffer",
512 [PTR_TO_XDP_SOCK] = "xdp_sock",
513 [PTR_TO_BTF_ID] = "ptr_",
514 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
515 [PTR_TO_MEM] = "mem",
516 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
517 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
518 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
519 [PTR_TO_RDWR_BUF] = "rdwr_buf",
520 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
521 };
522
523 static char slot_type_char[] = {
524 [STACK_INVALID] = '?',
525 [STACK_SPILL] = 'r',
526 [STACK_MISC] = 'm',
527 [STACK_ZERO] = '0',
528 };
529
530 static void print_liveness(struct bpf_verifier_env *env,
531 enum bpf_reg_liveness live)
532 {
533 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
534 verbose(env, "_");
535 if (live & REG_LIVE_READ)
536 verbose(env, "r");
537 if (live & REG_LIVE_WRITTEN)
538 verbose(env, "w");
539 if (live & REG_LIVE_DONE)
540 verbose(env, "D");
541 }
542
543 static struct bpf_func_state *func(struct bpf_verifier_env *env,
544 const struct bpf_reg_state *reg)
545 {
546 struct bpf_verifier_state *cur = env->cur_state;
547
548 return cur->frame[reg->frameno];
549 }
550
551 const char *kernel_type_name(u32 id)
552 {
553 return btf_name_by_offset(btf_vmlinux,
554 btf_type_by_id(btf_vmlinux, id)->name_off);
555 }
556
557 static void print_verifier_state(struct bpf_verifier_env *env,
558 const struct bpf_func_state *state)
559 {
560 const struct bpf_reg_state *reg;
561 enum bpf_reg_type t;
562 int i;
563
564 if (state->frameno)
565 verbose(env, " frame%d:", state->frameno);
566 for (i = 0; i < MAX_BPF_REG; i++) {
567 reg = &state->regs[i];
568 t = reg->type;
569 if (t == NOT_INIT)
570 continue;
571 verbose(env, " R%d", i);
572 print_liveness(env, reg->live);
573 verbose(env, "=%s", reg_type_str[t]);
574 if (t == SCALAR_VALUE && reg->precise)
575 verbose(env, "P");
576 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
577 tnum_is_const(reg->var_off)) {
578 /* reg->off should be 0 for SCALAR_VALUE */
579 verbose(env, "%lld", reg->var_off.value + reg->off);
580 } else {
581 if (t == PTR_TO_BTF_ID || t == PTR_TO_BTF_ID_OR_NULL)
582 verbose(env, "%s", kernel_type_name(reg->btf_id));
583 verbose(env, "(id=%d", reg->id);
584 if (reg_type_may_be_refcounted_or_null(t))
585 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
586 if (t != SCALAR_VALUE)
587 verbose(env, ",off=%d", reg->off);
588 if (type_is_pkt_pointer(t))
589 verbose(env, ",r=%d", reg->range);
590 else if (t == CONST_PTR_TO_MAP ||
591 t == PTR_TO_MAP_VALUE ||
592 t == PTR_TO_MAP_VALUE_OR_NULL)
593 verbose(env, ",ks=%d,vs=%d",
594 reg->map_ptr->key_size,
595 reg->map_ptr->value_size);
596 if (tnum_is_const(reg->var_off)) {
597 /* Typically an immediate SCALAR_VALUE, but
598 * could be a pointer whose offset is too big
599 * for reg->off
600 */
601 verbose(env, ",imm=%llx", reg->var_off.value);
602 } else {
603 if (reg->smin_value != reg->umin_value &&
604 reg->smin_value != S64_MIN)
605 verbose(env, ",smin_value=%lld",
606 (long long)reg->smin_value);
607 if (reg->smax_value != reg->umax_value &&
608 reg->smax_value != S64_MAX)
609 verbose(env, ",smax_value=%lld",
610 (long long)reg->smax_value);
611 if (reg->umin_value != 0)
612 verbose(env, ",umin_value=%llu",
613 (unsigned long long)reg->umin_value);
614 if (reg->umax_value != U64_MAX)
615 verbose(env, ",umax_value=%llu",
616 (unsigned long long)reg->umax_value);
617 if (!tnum_is_unknown(reg->var_off)) {
618 char tn_buf[48];
619
620 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
621 verbose(env, ",var_off=%s", tn_buf);
622 }
623 if (reg->s32_min_value != reg->smin_value &&
624 reg->s32_min_value != S32_MIN)
625 verbose(env, ",s32_min_value=%d",
626 (int)(reg->s32_min_value));
627 if (reg->s32_max_value != reg->smax_value &&
628 reg->s32_max_value != S32_MAX)
629 verbose(env, ",s32_max_value=%d",
630 (int)(reg->s32_max_value));
631 if (reg->u32_min_value != reg->umin_value &&
632 reg->u32_min_value != U32_MIN)
633 verbose(env, ",u32_min_value=%d",
634 (int)(reg->u32_min_value));
635 if (reg->u32_max_value != reg->umax_value &&
636 reg->u32_max_value != U32_MAX)
637 verbose(env, ",u32_max_value=%d",
638 (int)(reg->u32_max_value));
639 }
640 verbose(env, ")");
641 }
642 }
643 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
644 char types_buf[BPF_REG_SIZE + 1];
645 bool valid = false;
646 int j;
647
648 for (j = 0; j < BPF_REG_SIZE; j++) {
649 if (state->stack[i].slot_type[j] != STACK_INVALID)
650 valid = true;
651 types_buf[j] = slot_type_char[
652 state->stack[i].slot_type[j]];
653 }
654 types_buf[BPF_REG_SIZE] = 0;
655 if (!valid)
656 continue;
657 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
658 print_liveness(env, state->stack[i].spilled_ptr.live);
659 if (state->stack[i].slot_type[0] == STACK_SPILL) {
660 reg = &state->stack[i].spilled_ptr;
661 t = reg->type;
662 verbose(env, "=%s", reg_type_str[t]);
663 if (t == SCALAR_VALUE && reg->precise)
664 verbose(env, "P");
665 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
666 verbose(env, "%lld", reg->var_off.value + reg->off);
667 } else {
668 verbose(env, "=%s", types_buf);
669 }
670 }
671 if (state->acquired_refs && state->refs[0].id) {
672 verbose(env, " refs=%d", state->refs[0].id);
673 for (i = 1; i < state->acquired_refs; i++)
674 if (state->refs[i].id)
675 verbose(env, ",%d", state->refs[i].id);
676 }
677 verbose(env, "\n");
678 }
679
680 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
681 static int copy_##NAME##_state(struct bpf_func_state *dst, \
682 const struct bpf_func_state *src) \
683 { \
684 if (!src->FIELD) \
685 return 0; \
686 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
687 /* internal bug, make state invalid to reject the program */ \
688 memset(dst, 0, sizeof(*dst)); \
689 return -EFAULT; \
690 } \
691 memcpy(dst->FIELD, src->FIELD, \
692 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
693 return 0; \
694 }
695 /* copy_reference_state() */
696 COPY_STATE_FN(reference, acquired_refs, refs, 1)
697 /* copy_stack_state() */
698 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
699 #undef COPY_STATE_FN
700
701 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
702 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
703 bool copy_old) \
704 { \
705 u32 old_size = state->COUNT; \
706 struct bpf_##NAME##_state *new_##FIELD; \
707 int slot = size / SIZE; \
708 \
709 if (size <= old_size || !size) { \
710 if (copy_old) \
711 return 0; \
712 state->COUNT = slot * SIZE; \
713 if (!size && old_size) { \
714 kfree(state->FIELD); \
715 state->FIELD = NULL; \
716 } \
717 return 0; \
718 } \
719 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
720 GFP_KERNEL); \
721 if (!new_##FIELD) \
722 return -ENOMEM; \
723 if (copy_old) { \
724 if (state->FIELD) \
725 memcpy(new_##FIELD, state->FIELD, \
726 sizeof(*new_##FIELD) * (old_size / SIZE)); \
727 memset(new_##FIELD + old_size / SIZE, 0, \
728 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
729 } \
730 state->COUNT = slot * SIZE; \
731 kfree(state->FIELD); \
732 state->FIELD = new_##FIELD; \
733 return 0; \
734 }
735 /* realloc_reference_state() */
736 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
737 /* realloc_stack_state() */
738 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
739 #undef REALLOC_STATE_FN
740
741 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
742 * make it consume minimal amount of memory. check_stack_write() access from
743 * the program calls into realloc_func_state() to grow the stack size.
744 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
745 * which realloc_stack_state() copies over. It points to previous
746 * bpf_verifier_state which is never reallocated.
747 */
748 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
749 int refs_size, bool copy_old)
750 {
751 int err = realloc_reference_state(state, refs_size, copy_old);
752 if (err)
753 return err;
754 return realloc_stack_state(state, stack_size, copy_old);
755 }
756
757 /* Acquire a pointer id from the env and update the state->refs to include
758 * this new pointer reference.
759 * On success, returns a valid pointer id to associate with the register
760 * On failure, returns a negative errno.
761 */
762 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
763 {
764 struct bpf_func_state *state = cur_func(env);
765 int new_ofs = state->acquired_refs;
766 int id, err;
767
768 err = realloc_reference_state(state, state->acquired_refs + 1, true);
769 if (err)
770 return err;
771 id = ++env->id_gen;
772 state->refs[new_ofs].id = id;
773 state->refs[new_ofs].insn_idx = insn_idx;
774
775 return id;
776 }
777
778 /* release function corresponding to acquire_reference_state(). Idempotent. */
779 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
780 {
781 int i, last_idx;
782
783 last_idx = state->acquired_refs - 1;
784 for (i = 0; i < state->acquired_refs; i++) {
785 if (state->refs[i].id == ptr_id) {
786 if (last_idx && i != last_idx)
787 memcpy(&state->refs[i], &state->refs[last_idx],
788 sizeof(*state->refs));
789 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
790 state->acquired_refs--;
791 return 0;
792 }
793 }
794 return -EINVAL;
795 }
796
797 static int transfer_reference_state(struct bpf_func_state *dst,
798 struct bpf_func_state *src)
799 {
800 int err = realloc_reference_state(dst, src->acquired_refs, false);
801 if (err)
802 return err;
803 err = copy_reference_state(dst, src);
804 if (err)
805 return err;
806 return 0;
807 }
808
809 static void free_func_state(struct bpf_func_state *state)
810 {
811 if (!state)
812 return;
813 kfree(state->refs);
814 kfree(state->stack);
815 kfree(state);
816 }
817
818 static void clear_jmp_history(struct bpf_verifier_state *state)
819 {
820 kfree(state->jmp_history);
821 state->jmp_history = NULL;
822 state->jmp_history_cnt = 0;
823 }
824
825 static void free_verifier_state(struct bpf_verifier_state *state,
826 bool free_self)
827 {
828 int i;
829
830 for (i = 0; i <= state->curframe; i++) {
831 free_func_state(state->frame[i]);
832 state->frame[i] = NULL;
833 }
834 clear_jmp_history(state);
835 if (free_self)
836 kfree(state);
837 }
838
839 /* copy verifier state from src to dst growing dst stack space
840 * when necessary to accommodate larger src stack
841 */
842 static int copy_func_state(struct bpf_func_state *dst,
843 const struct bpf_func_state *src)
844 {
845 int err;
846
847 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
848 false);
849 if (err)
850 return err;
851 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
852 err = copy_reference_state(dst, src);
853 if (err)
854 return err;
855 return copy_stack_state(dst, src);
856 }
857
858 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
859 const struct bpf_verifier_state *src)
860 {
861 struct bpf_func_state *dst;
862 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
863 int i, err;
864
865 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
866 kfree(dst_state->jmp_history);
867 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
868 if (!dst_state->jmp_history)
869 return -ENOMEM;
870 }
871 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
872 dst_state->jmp_history_cnt = src->jmp_history_cnt;
873
874 /* if dst has more stack frames then src frame, free them */
875 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
876 free_func_state(dst_state->frame[i]);
877 dst_state->frame[i] = NULL;
878 }
879 dst_state->speculative = src->speculative;
880 dst_state->curframe = src->curframe;
881 dst_state->active_spin_lock = src->active_spin_lock;
882 dst_state->branches = src->branches;
883 dst_state->parent = src->parent;
884 dst_state->first_insn_idx = src->first_insn_idx;
885 dst_state->last_insn_idx = src->last_insn_idx;
886 for (i = 0; i <= src->curframe; i++) {
887 dst = dst_state->frame[i];
888 if (!dst) {
889 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
890 if (!dst)
891 return -ENOMEM;
892 dst_state->frame[i] = dst;
893 }
894 err = copy_func_state(dst, src->frame[i]);
895 if (err)
896 return err;
897 }
898 return 0;
899 }
900
901 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
902 {
903 while (st) {
904 u32 br = --st->branches;
905
906 /* WARN_ON(br > 1) technically makes sense here,
907 * but see comment in push_stack(), hence:
908 */
909 WARN_ONCE((int)br < 0,
910 "BUG update_branch_counts:branches_to_explore=%d\n",
911 br);
912 if (br)
913 break;
914 st = st->parent;
915 }
916 }
917
918 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
919 int *insn_idx, bool pop_log)
920 {
921 struct bpf_verifier_state *cur = env->cur_state;
922 struct bpf_verifier_stack_elem *elem, *head = env->head;
923 int err;
924
925 if (env->head == NULL)
926 return -ENOENT;
927
928 if (cur) {
929 err = copy_verifier_state(cur, &head->st);
930 if (err)
931 return err;
932 }
933 if (pop_log)
934 bpf_vlog_reset(&env->log, head->log_pos);
935 if (insn_idx)
936 *insn_idx = head->insn_idx;
937 if (prev_insn_idx)
938 *prev_insn_idx = head->prev_insn_idx;
939 elem = head->next;
940 free_verifier_state(&head->st, false);
941 kfree(head);
942 env->head = elem;
943 env->stack_size--;
944 return 0;
945 }
946
947 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
948 int insn_idx, int prev_insn_idx,
949 bool speculative)
950 {
951 struct bpf_verifier_state *cur = env->cur_state;
952 struct bpf_verifier_stack_elem *elem;
953 int err;
954
955 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
956 if (!elem)
957 goto err;
958
959 elem->insn_idx = insn_idx;
960 elem->prev_insn_idx = prev_insn_idx;
961 elem->next = env->head;
962 elem->log_pos = env->log.len_used;
963 env->head = elem;
964 env->stack_size++;
965 err = copy_verifier_state(&elem->st, cur);
966 if (err)
967 goto err;
968 elem->st.speculative |= speculative;
969 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
970 verbose(env, "The sequence of %d jumps is too complex.\n",
971 env->stack_size);
972 goto err;
973 }
974 if (elem->st.parent) {
975 ++elem->st.parent->branches;
976 /* WARN_ON(branches > 2) technically makes sense here,
977 * but
978 * 1. speculative states will bump 'branches' for non-branch
979 * instructions
980 * 2. is_state_visited() heuristics may decide not to create
981 * a new state for a sequence of branches and all such current
982 * and cloned states will be pointing to a single parent state
983 * which might have large 'branches' count.
984 */
985 }
986 return &elem->st;
987 err:
988 free_verifier_state(env->cur_state, true);
989 env->cur_state = NULL;
990 /* pop all elements and return */
991 while (!pop_stack(env, NULL, NULL, false));
992 return NULL;
993 }
994
995 #define CALLER_SAVED_REGS 6
996 static const int caller_saved[CALLER_SAVED_REGS] = {
997 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
998 };
999
1000 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1001 struct bpf_reg_state *reg);
1002
1003 /* Mark the unknown part of a register (variable offset or scalar value) as
1004 * known to have the value @imm.
1005 */
1006 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1007 {
1008 /* Clear id, off, and union(map_ptr, range) */
1009 memset(((u8 *)reg) + sizeof(reg->type), 0,
1010 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1011 reg->var_off = tnum_const(imm);
1012 reg->smin_value = (s64)imm;
1013 reg->smax_value = (s64)imm;
1014 reg->umin_value = imm;
1015 reg->umax_value = imm;
1016
1017 reg->s32_min_value = (s32)imm;
1018 reg->s32_max_value = (s32)imm;
1019 reg->u32_min_value = (u32)imm;
1020 reg->u32_max_value = (u32)imm;
1021 }
1022
1023 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1024 {
1025 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1026 reg->s32_min_value = (s32)imm;
1027 reg->s32_max_value = (s32)imm;
1028 reg->u32_min_value = (u32)imm;
1029 reg->u32_max_value = (u32)imm;
1030 }
1031
1032 /* Mark the 'variable offset' part of a register as zero. This should be
1033 * used only on registers holding a pointer type.
1034 */
1035 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1036 {
1037 __mark_reg_known(reg, 0);
1038 }
1039
1040 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1041 {
1042 __mark_reg_known(reg, 0);
1043 reg->type = SCALAR_VALUE;
1044 }
1045
1046 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1047 struct bpf_reg_state *regs, u32 regno)
1048 {
1049 if (WARN_ON(regno >= MAX_BPF_REG)) {
1050 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1051 /* Something bad happened, let's kill all regs */
1052 for (regno = 0; regno < MAX_BPF_REG; regno++)
1053 __mark_reg_not_init(env, regs + regno);
1054 return;
1055 }
1056 __mark_reg_known_zero(regs + regno);
1057 }
1058
1059 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1060 {
1061 return type_is_pkt_pointer(reg->type);
1062 }
1063
1064 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1065 {
1066 return reg_is_pkt_pointer(reg) ||
1067 reg->type == PTR_TO_PACKET_END;
1068 }
1069
1070 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1071 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1072 enum bpf_reg_type which)
1073 {
1074 /* The register can already have a range from prior markings.
1075 * This is fine as long as it hasn't been advanced from its
1076 * origin.
1077 */
1078 return reg->type == which &&
1079 reg->id == 0 &&
1080 reg->off == 0 &&
1081 tnum_equals_const(reg->var_off, 0);
1082 }
1083
1084 /* Reset the min/max bounds of a register */
1085 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1086 {
1087 reg->smin_value = S64_MIN;
1088 reg->smax_value = S64_MAX;
1089 reg->umin_value = 0;
1090 reg->umax_value = U64_MAX;
1091
1092 reg->s32_min_value = S32_MIN;
1093 reg->s32_max_value = S32_MAX;
1094 reg->u32_min_value = 0;
1095 reg->u32_max_value = U32_MAX;
1096 }
1097
1098 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1099 {
1100 reg->smin_value = S64_MIN;
1101 reg->smax_value = S64_MAX;
1102 reg->umin_value = 0;
1103 reg->umax_value = U64_MAX;
1104 }
1105
1106 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1107 {
1108 reg->s32_min_value = S32_MIN;
1109 reg->s32_max_value = S32_MAX;
1110 reg->u32_min_value = 0;
1111 reg->u32_max_value = U32_MAX;
1112 }
1113
1114 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1115 {
1116 struct tnum var32_off = tnum_subreg(reg->var_off);
1117
1118 /* min signed is max(sign bit) | min(other bits) */
1119 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1120 var32_off.value | (var32_off.mask & S32_MIN));
1121 /* max signed is min(sign bit) | max(other bits) */
1122 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1123 var32_off.value | (var32_off.mask & S32_MAX));
1124 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1125 reg->u32_max_value = min(reg->u32_max_value,
1126 (u32)(var32_off.value | var32_off.mask));
1127 }
1128
1129 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1130 {
1131 /* min signed is max(sign bit) | min(other bits) */
1132 reg->smin_value = max_t(s64, reg->smin_value,
1133 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1134 /* max signed is min(sign bit) | max(other bits) */
1135 reg->smax_value = min_t(s64, reg->smax_value,
1136 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1137 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1138 reg->umax_value = min(reg->umax_value,
1139 reg->var_off.value | reg->var_off.mask);
1140 }
1141
1142 static void __update_reg_bounds(struct bpf_reg_state *reg)
1143 {
1144 __update_reg32_bounds(reg);
1145 __update_reg64_bounds(reg);
1146 }
1147
1148 /* Uses signed min/max values to inform unsigned, and vice-versa */
1149 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1150 {
1151 /* Learn sign from signed bounds.
1152 * If we cannot cross the sign boundary, then signed and unsigned bounds
1153 * are the same, so combine. This works even in the negative case, e.g.
1154 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1155 */
1156 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1157 reg->s32_min_value = reg->u32_min_value =
1158 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1159 reg->s32_max_value = reg->u32_max_value =
1160 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1161 return;
1162 }
1163 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1164 * boundary, so we must be careful.
1165 */
1166 if ((s32)reg->u32_max_value >= 0) {
1167 /* Positive. We can't learn anything from the smin, but smax
1168 * is positive, hence safe.
1169 */
1170 reg->s32_min_value = reg->u32_min_value;
1171 reg->s32_max_value = reg->u32_max_value =
1172 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1173 } else if ((s32)reg->u32_min_value < 0) {
1174 /* Negative. We can't learn anything from the smax, but smin
1175 * is negative, hence safe.
1176 */
1177 reg->s32_min_value = reg->u32_min_value =
1178 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1179 reg->s32_max_value = reg->u32_max_value;
1180 }
1181 }
1182
1183 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1184 {
1185 /* Learn sign from signed bounds.
1186 * If we cannot cross the sign boundary, then signed and unsigned bounds
1187 * are the same, so combine. This works even in the negative case, e.g.
1188 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1189 */
1190 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1191 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1192 reg->umin_value);
1193 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1194 reg->umax_value);
1195 return;
1196 }
1197 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1198 * boundary, so we must be careful.
1199 */
1200 if ((s64)reg->umax_value >= 0) {
1201 /* Positive. We can't learn anything from the smin, but smax
1202 * is positive, hence safe.
1203 */
1204 reg->smin_value = reg->umin_value;
1205 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1206 reg->umax_value);
1207 } else if ((s64)reg->umin_value < 0) {
1208 /* Negative. We can't learn anything from the smax, but smin
1209 * is negative, hence safe.
1210 */
1211 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1212 reg->umin_value);
1213 reg->smax_value = reg->umax_value;
1214 }
1215 }
1216
1217 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1218 {
1219 __reg32_deduce_bounds(reg);
1220 __reg64_deduce_bounds(reg);
1221 }
1222
1223 /* Attempts to improve var_off based on unsigned min/max information */
1224 static void __reg_bound_offset(struct bpf_reg_state *reg)
1225 {
1226 struct tnum var64_off = tnum_intersect(reg->var_off,
1227 tnum_range(reg->umin_value,
1228 reg->umax_value));
1229 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1230 tnum_range(reg->u32_min_value,
1231 reg->u32_max_value));
1232
1233 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1234 }
1235
1236 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1237 {
1238 reg->umin_value = reg->u32_min_value;
1239 reg->umax_value = reg->u32_max_value;
1240 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1241 * but must be positive otherwise set to worse case bounds
1242 * and refine later from tnum.
1243 */
1244 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1245 reg->smax_value = reg->s32_max_value;
1246 else
1247 reg->smax_value = U32_MAX;
1248 if (reg->s32_min_value >= 0)
1249 reg->smin_value = reg->s32_min_value;
1250 else
1251 reg->smin_value = 0;
1252 }
1253
1254 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1255 {
1256 /* special case when 64-bit register has upper 32-bit register
1257 * zeroed. Typically happens after zext or <<32, >>32 sequence
1258 * allowing us to use 32-bit bounds directly,
1259 */
1260 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1261 __reg_assign_32_into_64(reg);
1262 } else {
1263 /* Otherwise the best we can do is push lower 32bit known and
1264 * unknown bits into register (var_off set from jmp logic)
1265 * then learn as much as possible from the 64-bit tnum
1266 * known and unknown bits. The previous smin/smax bounds are
1267 * invalid here because of jmp32 compare so mark them unknown
1268 * so they do not impact tnum bounds calculation.
1269 */
1270 __mark_reg64_unbounded(reg);
1271 __update_reg_bounds(reg);
1272 }
1273
1274 /* Intersecting with the old var_off might have improved our bounds
1275 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1276 * then new var_off is (0; 0x7f...fc) which improves our umax.
1277 */
1278 __reg_deduce_bounds(reg);
1279 __reg_bound_offset(reg);
1280 __update_reg_bounds(reg);
1281 }
1282
1283 static bool __reg64_bound_s32(s64 a)
1284 {
1285 if (a > S32_MIN && a < S32_MAX)
1286 return true;
1287 return false;
1288 }
1289
1290 static bool __reg64_bound_u32(u64 a)
1291 {
1292 if (a > U32_MIN && a < U32_MAX)
1293 return true;
1294 return false;
1295 }
1296
1297 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1298 {
1299 __mark_reg32_unbounded(reg);
1300
1301 if (__reg64_bound_s32(reg->smin_value))
1302 reg->s32_min_value = (s32)reg->smin_value;
1303 if (__reg64_bound_s32(reg->smax_value))
1304 reg->s32_max_value = (s32)reg->smax_value;
1305 if (__reg64_bound_u32(reg->umin_value))
1306 reg->u32_min_value = (u32)reg->umin_value;
1307 if (__reg64_bound_u32(reg->umax_value))
1308 reg->u32_max_value = (u32)reg->umax_value;
1309
1310 /* Intersecting with the old var_off might have improved our bounds
1311 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1312 * then new var_off is (0; 0x7f...fc) which improves our umax.
1313 */
1314 __reg_deduce_bounds(reg);
1315 __reg_bound_offset(reg);
1316 __update_reg_bounds(reg);
1317 }
1318
1319 /* Mark a register as having a completely unknown (scalar) value. */
1320 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1321 struct bpf_reg_state *reg)
1322 {
1323 /*
1324 * Clear type, id, off, and union(map_ptr, range) and
1325 * padding between 'type' and union
1326 */
1327 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1328 reg->type = SCALAR_VALUE;
1329 reg->var_off = tnum_unknown;
1330 reg->frameno = 0;
1331 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1332 __mark_reg_unbounded(reg);
1333 }
1334
1335 static void mark_reg_unknown(struct bpf_verifier_env *env,
1336 struct bpf_reg_state *regs, u32 regno)
1337 {
1338 if (WARN_ON(regno >= MAX_BPF_REG)) {
1339 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1340 /* Something bad happened, let's kill all regs except FP */
1341 for (regno = 0; regno < BPF_REG_FP; regno++)
1342 __mark_reg_not_init(env, regs + regno);
1343 return;
1344 }
1345 __mark_reg_unknown(env, regs + regno);
1346 }
1347
1348 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1349 struct bpf_reg_state *reg)
1350 {
1351 __mark_reg_unknown(env, reg);
1352 reg->type = NOT_INIT;
1353 }
1354
1355 static void mark_reg_not_init(struct bpf_verifier_env *env,
1356 struct bpf_reg_state *regs, u32 regno)
1357 {
1358 if (WARN_ON(regno >= MAX_BPF_REG)) {
1359 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1360 /* Something bad happened, let's kill all regs except FP */
1361 for (regno = 0; regno < BPF_REG_FP; regno++)
1362 __mark_reg_not_init(env, regs + regno);
1363 return;
1364 }
1365 __mark_reg_not_init(env, regs + regno);
1366 }
1367
1368 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1369 struct bpf_reg_state *regs, u32 regno,
1370 enum bpf_reg_type reg_type, u32 btf_id)
1371 {
1372 if (reg_type == SCALAR_VALUE) {
1373 mark_reg_unknown(env, regs, regno);
1374 return;
1375 }
1376 mark_reg_known_zero(env, regs, regno);
1377 regs[regno].type = PTR_TO_BTF_ID;
1378 regs[regno].btf_id = btf_id;
1379 }
1380
1381 #define DEF_NOT_SUBREG (0)
1382 static void init_reg_state(struct bpf_verifier_env *env,
1383 struct bpf_func_state *state)
1384 {
1385 struct bpf_reg_state *regs = state->regs;
1386 int i;
1387
1388 for (i = 0; i < MAX_BPF_REG; i++) {
1389 mark_reg_not_init(env, regs, i);
1390 regs[i].live = REG_LIVE_NONE;
1391 regs[i].parent = NULL;
1392 regs[i].subreg_def = DEF_NOT_SUBREG;
1393 }
1394
1395 /* frame pointer */
1396 regs[BPF_REG_FP].type = PTR_TO_STACK;
1397 mark_reg_known_zero(env, regs, BPF_REG_FP);
1398 regs[BPF_REG_FP].frameno = state->frameno;
1399 }
1400
1401 #define BPF_MAIN_FUNC (-1)
1402 static void init_func_state(struct bpf_verifier_env *env,
1403 struct bpf_func_state *state,
1404 int callsite, int frameno, int subprogno)
1405 {
1406 state->callsite = callsite;
1407 state->frameno = frameno;
1408 state->subprogno = subprogno;
1409 init_reg_state(env, state);
1410 }
1411
1412 enum reg_arg_type {
1413 SRC_OP, /* register is used as source operand */
1414 DST_OP, /* register is used as destination operand */
1415 DST_OP_NO_MARK /* same as above, check only, don't mark */
1416 };
1417
1418 static int cmp_subprogs(const void *a, const void *b)
1419 {
1420 return ((struct bpf_subprog_info *)a)->start -
1421 ((struct bpf_subprog_info *)b)->start;
1422 }
1423
1424 static int find_subprog(struct bpf_verifier_env *env, int off)
1425 {
1426 struct bpf_subprog_info *p;
1427
1428 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1429 sizeof(env->subprog_info[0]), cmp_subprogs);
1430 if (!p)
1431 return -ENOENT;
1432 return p - env->subprog_info;
1433
1434 }
1435
1436 static int add_subprog(struct bpf_verifier_env *env, int off)
1437 {
1438 int insn_cnt = env->prog->len;
1439 int ret;
1440
1441 if (off >= insn_cnt || off < 0) {
1442 verbose(env, "call to invalid destination\n");
1443 return -EINVAL;
1444 }
1445 ret = find_subprog(env, off);
1446 if (ret >= 0)
1447 return 0;
1448 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1449 verbose(env, "too many subprograms\n");
1450 return -E2BIG;
1451 }
1452 env->subprog_info[env->subprog_cnt++].start = off;
1453 sort(env->subprog_info, env->subprog_cnt,
1454 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1455 return 0;
1456 }
1457
1458 static int check_subprogs(struct bpf_verifier_env *env)
1459 {
1460 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1461 struct bpf_subprog_info *subprog = env->subprog_info;
1462 struct bpf_insn *insn = env->prog->insnsi;
1463 int insn_cnt = env->prog->len;
1464
1465 /* Add entry function. */
1466 ret = add_subprog(env, 0);
1467 if (ret < 0)
1468 return ret;
1469
1470 /* determine subprog starts. The end is one before the next starts */
1471 for (i = 0; i < insn_cnt; i++) {
1472 if (insn[i].code != (BPF_JMP | BPF_CALL))
1473 continue;
1474 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1475 continue;
1476 if (!env->bpf_capable) {
1477 verbose(env,
1478 "function calls to other bpf functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1479 return -EPERM;
1480 }
1481 ret = add_subprog(env, i + insn[i].imm + 1);
1482 if (ret < 0)
1483 return ret;
1484 }
1485
1486 /* Add a fake 'exit' subprog which could simplify subprog iteration
1487 * logic. 'subprog_cnt' should not be increased.
1488 */
1489 subprog[env->subprog_cnt].start = insn_cnt;
1490
1491 if (env->log.level & BPF_LOG_LEVEL2)
1492 for (i = 0; i < env->subprog_cnt; i++)
1493 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1494
1495 /* now check that all jumps are within the same subprog */
1496 subprog_start = subprog[cur_subprog].start;
1497 subprog_end = subprog[cur_subprog + 1].start;
1498 for (i = 0; i < insn_cnt; i++) {
1499 u8 code = insn[i].code;
1500
1501 if (code == (BPF_JMP | BPF_CALL) &&
1502 insn[i].imm == BPF_FUNC_tail_call &&
1503 insn[i].src_reg != BPF_PSEUDO_CALL)
1504 subprog[cur_subprog].has_tail_call = true;
1505 if (BPF_CLASS(code) == BPF_LD &&
1506 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1507 subprog[cur_subprog].has_ld_abs = true;
1508 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1509 goto next;
1510 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1511 goto next;
1512 off = i + insn[i].off + 1;
1513 if (off < subprog_start || off >= subprog_end) {
1514 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1515 return -EINVAL;
1516 }
1517 next:
1518 if (i == subprog_end - 1) {
1519 /* to avoid fall-through from one subprog into another
1520 * the last insn of the subprog should be either exit
1521 * or unconditional jump back
1522 */
1523 if (code != (BPF_JMP | BPF_EXIT) &&
1524 code != (BPF_JMP | BPF_JA)) {
1525 verbose(env, "last insn is not an exit or jmp\n");
1526 return -EINVAL;
1527 }
1528 subprog_start = subprog_end;
1529 cur_subprog++;
1530 if (cur_subprog < env->subprog_cnt)
1531 subprog_end = subprog[cur_subprog + 1].start;
1532 }
1533 }
1534 return 0;
1535 }
1536
1537 /* Parentage chain of this register (or stack slot) should take care of all
1538 * issues like callee-saved registers, stack slot allocation time, etc.
1539 */
1540 static int mark_reg_read(struct bpf_verifier_env *env,
1541 const struct bpf_reg_state *state,
1542 struct bpf_reg_state *parent, u8 flag)
1543 {
1544 bool writes = parent == state->parent; /* Observe write marks */
1545 int cnt = 0;
1546
1547 while (parent) {
1548 /* if read wasn't screened by an earlier write ... */
1549 if (writes && state->live & REG_LIVE_WRITTEN)
1550 break;
1551 if (parent->live & REG_LIVE_DONE) {
1552 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1553 reg_type_str[parent->type],
1554 parent->var_off.value, parent->off);
1555 return -EFAULT;
1556 }
1557 /* The first condition is more likely to be true than the
1558 * second, checked it first.
1559 */
1560 if ((parent->live & REG_LIVE_READ) == flag ||
1561 parent->live & REG_LIVE_READ64)
1562 /* The parentage chain never changes and
1563 * this parent was already marked as LIVE_READ.
1564 * There is no need to keep walking the chain again and
1565 * keep re-marking all parents as LIVE_READ.
1566 * This case happens when the same register is read
1567 * multiple times without writes into it in-between.
1568 * Also, if parent has the stronger REG_LIVE_READ64 set,
1569 * then no need to set the weak REG_LIVE_READ32.
1570 */
1571 break;
1572 /* ... then we depend on parent's value */
1573 parent->live |= flag;
1574 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1575 if (flag == REG_LIVE_READ64)
1576 parent->live &= ~REG_LIVE_READ32;
1577 state = parent;
1578 parent = state->parent;
1579 writes = true;
1580 cnt++;
1581 }
1582
1583 if (env->longest_mark_read_walk < cnt)
1584 env->longest_mark_read_walk = cnt;
1585 return 0;
1586 }
1587
1588 /* This function is supposed to be used by the following 32-bit optimization
1589 * code only. It returns TRUE if the source or destination register operates
1590 * on 64-bit, otherwise return FALSE.
1591 */
1592 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1593 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1594 {
1595 u8 code, class, op;
1596
1597 code = insn->code;
1598 class = BPF_CLASS(code);
1599 op = BPF_OP(code);
1600 if (class == BPF_JMP) {
1601 /* BPF_EXIT for "main" will reach here. Return TRUE
1602 * conservatively.
1603 */
1604 if (op == BPF_EXIT)
1605 return true;
1606 if (op == BPF_CALL) {
1607 /* BPF to BPF call will reach here because of marking
1608 * caller saved clobber with DST_OP_NO_MARK for which we
1609 * don't care the register def because they are anyway
1610 * marked as NOT_INIT already.
1611 */
1612 if (insn->src_reg == BPF_PSEUDO_CALL)
1613 return false;
1614 /* Helper call will reach here because of arg type
1615 * check, conservatively return TRUE.
1616 */
1617 if (t == SRC_OP)
1618 return true;
1619
1620 return false;
1621 }
1622 }
1623
1624 if (class == BPF_ALU64 || class == BPF_JMP ||
1625 /* BPF_END always use BPF_ALU class. */
1626 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1627 return true;
1628
1629 if (class == BPF_ALU || class == BPF_JMP32)
1630 return false;
1631
1632 if (class == BPF_LDX) {
1633 if (t != SRC_OP)
1634 return BPF_SIZE(code) == BPF_DW;
1635 /* LDX source must be ptr. */
1636 return true;
1637 }
1638
1639 if (class == BPF_STX) {
1640 if (reg->type != SCALAR_VALUE)
1641 return true;
1642 return BPF_SIZE(code) == BPF_DW;
1643 }
1644
1645 if (class == BPF_LD) {
1646 u8 mode = BPF_MODE(code);
1647
1648 /* LD_IMM64 */
1649 if (mode == BPF_IMM)
1650 return true;
1651
1652 /* Both LD_IND and LD_ABS return 32-bit data. */
1653 if (t != SRC_OP)
1654 return false;
1655
1656 /* Implicit ctx ptr. */
1657 if (regno == BPF_REG_6)
1658 return true;
1659
1660 /* Explicit source could be any width. */
1661 return true;
1662 }
1663
1664 if (class == BPF_ST)
1665 /* The only source register for BPF_ST is a ptr. */
1666 return true;
1667
1668 /* Conservatively return true at default. */
1669 return true;
1670 }
1671
1672 /* Return TRUE if INSN doesn't have explicit value define. */
1673 static bool insn_no_def(struct bpf_insn *insn)
1674 {
1675 u8 class = BPF_CLASS(insn->code);
1676
1677 return (class == BPF_JMP || class == BPF_JMP32 ||
1678 class == BPF_STX || class == BPF_ST);
1679 }
1680
1681 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1682 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1683 {
1684 if (insn_no_def(insn))
1685 return false;
1686
1687 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1688 }
1689
1690 static void mark_insn_zext(struct bpf_verifier_env *env,
1691 struct bpf_reg_state *reg)
1692 {
1693 s32 def_idx = reg->subreg_def;
1694
1695 if (def_idx == DEF_NOT_SUBREG)
1696 return;
1697
1698 env->insn_aux_data[def_idx - 1].zext_dst = true;
1699 /* The dst will be zero extended, so won't be sub-register anymore. */
1700 reg->subreg_def = DEF_NOT_SUBREG;
1701 }
1702
1703 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1704 enum reg_arg_type t)
1705 {
1706 struct bpf_verifier_state *vstate = env->cur_state;
1707 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1708 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1709 struct bpf_reg_state *reg, *regs = state->regs;
1710 bool rw64;
1711
1712 if (regno >= MAX_BPF_REG) {
1713 verbose(env, "R%d is invalid\n", regno);
1714 return -EINVAL;
1715 }
1716
1717 reg = &regs[regno];
1718 rw64 = is_reg64(env, insn, regno, reg, t);
1719 if (t == SRC_OP) {
1720 /* check whether register used as source operand can be read */
1721 if (reg->type == NOT_INIT) {
1722 verbose(env, "R%d !read_ok\n", regno);
1723 return -EACCES;
1724 }
1725 /* We don't need to worry about FP liveness because it's read-only */
1726 if (regno == BPF_REG_FP)
1727 return 0;
1728
1729 if (rw64)
1730 mark_insn_zext(env, reg);
1731
1732 return mark_reg_read(env, reg, reg->parent,
1733 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1734 } else {
1735 /* check whether register used as dest operand can be written to */
1736 if (regno == BPF_REG_FP) {
1737 verbose(env, "frame pointer is read only\n");
1738 return -EACCES;
1739 }
1740 reg->live |= REG_LIVE_WRITTEN;
1741 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1742 if (t == DST_OP)
1743 mark_reg_unknown(env, regs, regno);
1744 }
1745 return 0;
1746 }
1747
1748 /* for any branch, call, exit record the history of jmps in the given state */
1749 static int push_jmp_history(struct bpf_verifier_env *env,
1750 struct bpf_verifier_state *cur)
1751 {
1752 u32 cnt = cur->jmp_history_cnt;
1753 struct bpf_idx_pair *p;
1754
1755 cnt++;
1756 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1757 if (!p)
1758 return -ENOMEM;
1759 p[cnt - 1].idx = env->insn_idx;
1760 p[cnt - 1].prev_idx = env->prev_insn_idx;
1761 cur->jmp_history = p;
1762 cur->jmp_history_cnt = cnt;
1763 return 0;
1764 }
1765
1766 /* Backtrack one insn at a time. If idx is not at the top of recorded
1767 * history then previous instruction came from straight line execution.
1768 */
1769 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1770 u32 *history)
1771 {
1772 u32 cnt = *history;
1773
1774 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1775 i = st->jmp_history[cnt - 1].prev_idx;
1776 (*history)--;
1777 } else {
1778 i--;
1779 }
1780 return i;
1781 }
1782
1783 /* For given verifier state backtrack_insn() is called from the last insn to
1784 * the first insn. Its purpose is to compute a bitmask of registers and
1785 * stack slots that needs precision in the parent verifier state.
1786 */
1787 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1788 u32 *reg_mask, u64 *stack_mask)
1789 {
1790 const struct bpf_insn_cbs cbs = {
1791 .cb_print = verbose,
1792 .private_data = env,
1793 };
1794 struct bpf_insn *insn = env->prog->insnsi + idx;
1795 u8 class = BPF_CLASS(insn->code);
1796 u8 opcode = BPF_OP(insn->code);
1797 u8 mode = BPF_MODE(insn->code);
1798 u32 dreg = 1u << insn->dst_reg;
1799 u32 sreg = 1u << insn->src_reg;
1800 u32 spi;
1801
1802 if (insn->code == 0)
1803 return 0;
1804 if (env->log.level & BPF_LOG_LEVEL) {
1805 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1806 verbose(env, "%d: ", idx);
1807 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1808 }
1809
1810 if (class == BPF_ALU || class == BPF_ALU64) {
1811 if (!(*reg_mask & dreg))
1812 return 0;
1813 if (opcode == BPF_MOV) {
1814 if (BPF_SRC(insn->code) == BPF_X) {
1815 /* dreg = sreg
1816 * dreg needs precision after this insn
1817 * sreg needs precision before this insn
1818 */
1819 *reg_mask &= ~dreg;
1820 *reg_mask |= sreg;
1821 } else {
1822 /* dreg = K
1823 * dreg needs precision after this insn.
1824 * Corresponding register is already marked
1825 * as precise=true in this verifier state.
1826 * No further markings in parent are necessary
1827 */
1828 *reg_mask &= ~dreg;
1829 }
1830 } else {
1831 if (BPF_SRC(insn->code) == BPF_X) {
1832 /* dreg += sreg
1833 * both dreg and sreg need precision
1834 * before this insn
1835 */
1836 *reg_mask |= sreg;
1837 } /* else dreg += K
1838 * dreg still needs precision before this insn
1839 */
1840 }
1841 } else if (class == BPF_LDX) {
1842 if (!(*reg_mask & dreg))
1843 return 0;
1844 *reg_mask &= ~dreg;
1845
1846 /* scalars can only be spilled into stack w/o losing precision.
1847 * Load from any other memory can be zero extended.
1848 * The desire to keep that precision is already indicated
1849 * by 'precise' mark in corresponding register of this state.
1850 * No further tracking necessary.
1851 */
1852 if (insn->src_reg != BPF_REG_FP)
1853 return 0;
1854 if (BPF_SIZE(insn->code) != BPF_DW)
1855 return 0;
1856
1857 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1858 * that [fp - off] slot contains scalar that needs to be
1859 * tracked with precision
1860 */
1861 spi = (-insn->off - 1) / BPF_REG_SIZE;
1862 if (spi >= 64) {
1863 verbose(env, "BUG spi %d\n", spi);
1864 WARN_ONCE(1, "verifier backtracking bug");
1865 return -EFAULT;
1866 }
1867 *stack_mask |= 1ull << spi;
1868 } else if (class == BPF_STX || class == BPF_ST) {
1869 if (*reg_mask & dreg)
1870 /* stx & st shouldn't be using _scalar_ dst_reg
1871 * to access memory. It means backtracking
1872 * encountered a case of pointer subtraction.
1873 */
1874 return -ENOTSUPP;
1875 /* scalars can only be spilled into stack */
1876 if (insn->dst_reg != BPF_REG_FP)
1877 return 0;
1878 if (BPF_SIZE(insn->code) != BPF_DW)
1879 return 0;
1880 spi = (-insn->off - 1) / BPF_REG_SIZE;
1881 if (spi >= 64) {
1882 verbose(env, "BUG spi %d\n", spi);
1883 WARN_ONCE(1, "verifier backtracking bug");
1884 return -EFAULT;
1885 }
1886 if (!(*stack_mask & (1ull << spi)))
1887 return 0;
1888 *stack_mask &= ~(1ull << spi);
1889 if (class == BPF_STX)
1890 *reg_mask |= sreg;
1891 } else if (class == BPF_JMP || class == BPF_JMP32) {
1892 if (opcode == BPF_CALL) {
1893 if (insn->src_reg == BPF_PSEUDO_CALL)
1894 return -ENOTSUPP;
1895 /* regular helper call sets R0 */
1896 *reg_mask &= ~1;
1897 if (*reg_mask & 0x3f) {
1898 /* if backtracing was looking for registers R1-R5
1899 * they should have been found already.
1900 */
1901 verbose(env, "BUG regs %x\n", *reg_mask);
1902 WARN_ONCE(1, "verifier backtracking bug");
1903 return -EFAULT;
1904 }
1905 } else if (opcode == BPF_EXIT) {
1906 return -ENOTSUPP;
1907 }
1908 } else if (class == BPF_LD) {
1909 if (!(*reg_mask & dreg))
1910 return 0;
1911 *reg_mask &= ~dreg;
1912 /* It's ld_imm64 or ld_abs or ld_ind.
1913 * For ld_imm64 no further tracking of precision
1914 * into parent is necessary
1915 */
1916 if (mode == BPF_IND || mode == BPF_ABS)
1917 /* to be analyzed */
1918 return -ENOTSUPP;
1919 }
1920 return 0;
1921 }
1922
1923 /* the scalar precision tracking algorithm:
1924 * . at the start all registers have precise=false.
1925 * . scalar ranges are tracked as normal through alu and jmp insns.
1926 * . once precise value of the scalar register is used in:
1927 * . ptr + scalar alu
1928 * . if (scalar cond K|scalar)
1929 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1930 * backtrack through the verifier states and mark all registers and
1931 * stack slots with spilled constants that these scalar regisers
1932 * should be precise.
1933 * . during state pruning two registers (or spilled stack slots)
1934 * are equivalent if both are not precise.
1935 *
1936 * Note the verifier cannot simply walk register parentage chain,
1937 * since many different registers and stack slots could have been
1938 * used to compute single precise scalar.
1939 *
1940 * The approach of starting with precise=true for all registers and then
1941 * backtrack to mark a register as not precise when the verifier detects
1942 * that program doesn't care about specific value (e.g., when helper
1943 * takes register as ARG_ANYTHING parameter) is not safe.
1944 *
1945 * It's ok to walk single parentage chain of the verifier states.
1946 * It's possible that this backtracking will go all the way till 1st insn.
1947 * All other branches will be explored for needing precision later.
1948 *
1949 * The backtracking needs to deal with cases like:
1950 * 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)
1951 * r9 -= r8
1952 * r5 = r9
1953 * if r5 > 0x79f goto pc+7
1954 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1955 * r5 += 1
1956 * ...
1957 * call bpf_perf_event_output#25
1958 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1959 *
1960 * and this case:
1961 * r6 = 1
1962 * call foo // uses callee's r6 inside to compute r0
1963 * r0 += r6
1964 * if r0 == 0 goto
1965 *
1966 * to track above reg_mask/stack_mask needs to be independent for each frame.
1967 *
1968 * Also if parent's curframe > frame where backtracking started,
1969 * the verifier need to mark registers in both frames, otherwise callees
1970 * may incorrectly prune callers. This is similar to
1971 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1972 *
1973 * For now backtracking falls back into conservative marking.
1974 */
1975 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1976 struct bpf_verifier_state *st)
1977 {
1978 struct bpf_func_state *func;
1979 struct bpf_reg_state *reg;
1980 int i, j;
1981
1982 /* big hammer: mark all scalars precise in this path.
1983 * pop_stack may still get !precise scalars.
1984 */
1985 for (; st; st = st->parent)
1986 for (i = 0; i <= st->curframe; i++) {
1987 func = st->frame[i];
1988 for (j = 0; j < BPF_REG_FP; j++) {
1989 reg = &func->regs[j];
1990 if (reg->type != SCALAR_VALUE)
1991 continue;
1992 reg->precise = true;
1993 }
1994 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1995 if (func->stack[j].slot_type[0] != STACK_SPILL)
1996 continue;
1997 reg = &func->stack[j].spilled_ptr;
1998 if (reg->type != SCALAR_VALUE)
1999 continue;
2000 reg->precise = true;
2001 }
2002 }
2003 }
2004
2005 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2006 int spi)
2007 {
2008 struct bpf_verifier_state *st = env->cur_state;
2009 int first_idx = st->first_insn_idx;
2010 int last_idx = env->insn_idx;
2011 struct bpf_func_state *func;
2012 struct bpf_reg_state *reg;
2013 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2014 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2015 bool skip_first = true;
2016 bool new_marks = false;
2017 int i, err;
2018
2019 if (!env->bpf_capable)
2020 return 0;
2021
2022 func = st->frame[st->curframe];
2023 if (regno >= 0) {
2024 reg = &func->regs[regno];
2025 if (reg->type != SCALAR_VALUE) {
2026 WARN_ONCE(1, "backtracing misuse");
2027 return -EFAULT;
2028 }
2029 if (!reg->precise)
2030 new_marks = true;
2031 else
2032 reg_mask = 0;
2033 reg->precise = true;
2034 }
2035
2036 while (spi >= 0) {
2037 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2038 stack_mask = 0;
2039 break;
2040 }
2041 reg = &func->stack[spi].spilled_ptr;
2042 if (reg->type != SCALAR_VALUE) {
2043 stack_mask = 0;
2044 break;
2045 }
2046 if (!reg->precise)
2047 new_marks = true;
2048 else
2049 stack_mask = 0;
2050 reg->precise = true;
2051 break;
2052 }
2053
2054 if (!new_marks)
2055 return 0;
2056 if (!reg_mask && !stack_mask)
2057 return 0;
2058 for (;;) {
2059 DECLARE_BITMAP(mask, 64);
2060 u32 history = st->jmp_history_cnt;
2061
2062 if (env->log.level & BPF_LOG_LEVEL)
2063 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2064 for (i = last_idx;;) {
2065 if (skip_first) {
2066 err = 0;
2067 skip_first = false;
2068 } else {
2069 err = backtrack_insn(env, i, &reg_mask, &stack_mask);
2070 }
2071 if (err == -ENOTSUPP) {
2072 mark_all_scalars_precise(env, st);
2073 return 0;
2074 } else if (err) {
2075 return err;
2076 }
2077 if (!reg_mask && !stack_mask)
2078 /* Found assignment(s) into tracked register in this state.
2079 * Since this state is already marked, just return.
2080 * Nothing to be tracked further in the parent state.
2081 */
2082 return 0;
2083 if (i == first_idx)
2084 break;
2085 i = get_prev_insn_idx(st, i, &history);
2086 if (i >= env->prog->len) {
2087 /* This can happen if backtracking reached insn 0
2088 * and there are still reg_mask or stack_mask
2089 * to backtrack.
2090 * It means the backtracking missed the spot where
2091 * particular register was initialized with a constant.
2092 */
2093 verbose(env, "BUG backtracking idx %d\n", i);
2094 WARN_ONCE(1, "verifier backtracking bug");
2095 return -EFAULT;
2096 }
2097 }
2098 st = st->parent;
2099 if (!st)
2100 break;
2101
2102 new_marks = false;
2103 func = st->frame[st->curframe];
2104 bitmap_from_u64(mask, reg_mask);
2105 for_each_set_bit(i, mask, 32) {
2106 reg = &func->regs[i];
2107 if (reg->type != SCALAR_VALUE) {
2108 reg_mask &= ~(1u << i);
2109 continue;
2110 }
2111 if (!reg->precise)
2112 new_marks = true;
2113 reg->precise = true;
2114 }
2115
2116 bitmap_from_u64(mask, stack_mask);
2117 for_each_set_bit(i, mask, 64) {
2118 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2119 /* the sequence of instructions:
2120 * 2: (bf) r3 = r10
2121 * 3: (7b) *(u64 *)(r3 -8) = r0
2122 * 4: (79) r4 = *(u64 *)(r10 -8)
2123 * doesn't contain jmps. It's backtracked
2124 * as a single block.
2125 * During backtracking insn 3 is not recognized as
2126 * stack access, so at the end of backtracking
2127 * stack slot fp-8 is still marked in stack_mask.
2128 * However the parent state may not have accessed
2129 * fp-8 and it's "unallocated" stack space.
2130 * In such case fallback to conservative.
2131 */
2132 mark_all_scalars_precise(env, st);
2133 return 0;
2134 }
2135
2136 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2137 stack_mask &= ~(1ull << i);
2138 continue;
2139 }
2140 reg = &func->stack[i].spilled_ptr;
2141 if (reg->type != SCALAR_VALUE) {
2142 stack_mask &= ~(1ull << i);
2143 continue;
2144 }
2145 if (!reg->precise)
2146 new_marks = true;
2147 reg->precise = true;
2148 }
2149 if (env->log.level & BPF_LOG_LEVEL) {
2150 print_verifier_state(env, func);
2151 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2152 new_marks ? "didn't have" : "already had",
2153 reg_mask, stack_mask);
2154 }
2155
2156 if (!reg_mask && !stack_mask)
2157 break;
2158 if (!new_marks)
2159 break;
2160
2161 last_idx = st->last_insn_idx;
2162 first_idx = st->first_insn_idx;
2163 }
2164 return 0;
2165 }
2166
2167 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2168 {
2169 return __mark_chain_precision(env, regno, -1);
2170 }
2171
2172 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2173 {
2174 return __mark_chain_precision(env, -1, spi);
2175 }
2176
2177 static bool is_spillable_regtype(enum bpf_reg_type type)
2178 {
2179 switch (type) {
2180 case PTR_TO_MAP_VALUE:
2181 case PTR_TO_MAP_VALUE_OR_NULL:
2182 case PTR_TO_STACK:
2183 case PTR_TO_CTX:
2184 case PTR_TO_PACKET:
2185 case PTR_TO_PACKET_META:
2186 case PTR_TO_PACKET_END:
2187 case PTR_TO_FLOW_KEYS:
2188 case CONST_PTR_TO_MAP:
2189 case PTR_TO_SOCKET:
2190 case PTR_TO_SOCKET_OR_NULL:
2191 case PTR_TO_SOCK_COMMON:
2192 case PTR_TO_SOCK_COMMON_OR_NULL:
2193 case PTR_TO_TCP_SOCK:
2194 case PTR_TO_TCP_SOCK_OR_NULL:
2195 case PTR_TO_XDP_SOCK:
2196 case PTR_TO_BTF_ID:
2197 case PTR_TO_BTF_ID_OR_NULL:
2198 case PTR_TO_RDONLY_BUF:
2199 case PTR_TO_RDONLY_BUF_OR_NULL:
2200 case PTR_TO_RDWR_BUF:
2201 case PTR_TO_RDWR_BUF_OR_NULL:
2202 return true;
2203 default:
2204 return false;
2205 }
2206 }
2207
2208 /* Does this register contain a constant zero? */
2209 static bool register_is_null(struct bpf_reg_state *reg)
2210 {
2211 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2212 }
2213
2214 static bool register_is_const(struct bpf_reg_state *reg)
2215 {
2216 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2217 }
2218
2219 static bool __is_pointer_value(bool allow_ptr_leaks,
2220 const struct bpf_reg_state *reg)
2221 {
2222 if (allow_ptr_leaks)
2223 return false;
2224
2225 return reg->type != SCALAR_VALUE;
2226 }
2227
2228 static void save_register_state(struct bpf_func_state *state,
2229 int spi, struct bpf_reg_state *reg)
2230 {
2231 int i;
2232
2233 state->stack[spi].spilled_ptr = *reg;
2234 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2235
2236 for (i = 0; i < BPF_REG_SIZE; i++)
2237 state->stack[spi].slot_type[i] = STACK_SPILL;
2238 }
2239
2240 /* check_stack_read/write functions track spill/fill of registers,
2241 * stack boundary and alignment are checked in check_mem_access()
2242 */
2243 static int check_stack_write(struct bpf_verifier_env *env,
2244 struct bpf_func_state *state, /* func where register points to */
2245 int off, int size, int value_regno, int insn_idx)
2246 {
2247 struct bpf_func_state *cur; /* state of the current function */
2248 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2249 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2250 struct bpf_reg_state *reg = NULL;
2251
2252 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2253 state->acquired_refs, true);
2254 if (err)
2255 return err;
2256 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2257 * so it's aligned access and [off, off + size) are within stack limits
2258 */
2259 if (!env->allow_ptr_leaks &&
2260 state->stack[spi].slot_type[0] == STACK_SPILL &&
2261 size != BPF_REG_SIZE) {
2262 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2263 return -EACCES;
2264 }
2265
2266 cur = env->cur_state->frame[env->cur_state->curframe];
2267 if (value_regno >= 0)
2268 reg = &cur->regs[value_regno];
2269
2270 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
2271 !register_is_null(reg) && env->bpf_capable) {
2272 if (dst_reg != BPF_REG_FP) {
2273 /* The backtracking logic can only recognize explicit
2274 * stack slot address like [fp - 8]. Other spill of
2275 * scalar via different register has to be conervative.
2276 * Backtrack from here and mark all registers as precise
2277 * that contributed into 'reg' being a constant.
2278 */
2279 err = mark_chain_precision(env, value_regno);
2280 if (err)
2281 return err;
2282 }
2283 save_register_state(state, spi, reg);
2284 } else if (reg && is_spillable_regtype(reg->type)) {
2285 /* register containing pointer is being spilled into stack */
2286 if (size != BPF_REG_SIZE) {
2287 verbose_linfo(env, insn_idx, "; ");
2288 verbose(env, "invalid size of register spill\n");
2289 return -EACCES;
2290 }
2291
2292 if (state != cur && reg->type == PTR_TO_STACK) {
2293 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2294 return -EINVAL;
2295 }
2296
2297 if (!env->bypass_spec_v4) {
2298 bool sanitize = false;
2299
2300 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2301 register_is_const(&state->stack[spi].spilled_ptr))
2302 sanitize = true;
2303 for (i = 0; i < BPF_REG_SIZE; i++)
2304 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2305 sanitize = true;
2306 break;
2307 }
2308 if (sanitize) {
2309 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2310 int soff = (-spi - 1) * BPF_REG_SIZE;
2311
2312 /* detected reuse of integer stack slot with a pointer
2313 * which means either llvm is reusing stack slot or
2314 * an attacker is trying to exploit CVE-2018-3639
2315 * (speculative store bypass)
2316 * Have to sanitize that slot with preemptive
2317 * store of zero.
2318 */
2319 if (*poff && *poff != soff) {
2320 /* disallow programs where single insn stores
2321 * into two different stack slots, since verifier
2322 * cannot sanitize them
2323 */
2324 verbose(env,
2325 "insn %d cannot access two stack slots fp%d and fp%d",
2326 insn_idx, *poff, soff);
2327 return -EINVAL;
2328 }
2329 *poff = soff;
2330 }
2331 }
2332 save_register_state(state, spi, reg);
2333 } else {
2334 u8 type = STACK_MISC;
2335
2336 /* regular write of data into stack destroys any spilled ptr */
2337 state->stack[spi].spilled_ptr.type = NOT_INIT;
2338 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2339 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2340 for (i = 0; i < BPF_REG_SIZE; i++)
2341 state->stack[spi].slot_type[i] = STACK_MISC;
2342
2343 /* only mark the slot as written if all 8 bytes were written
2344 * otherwise read propagation may incorrectly stop too soon
2345 * when stack slots are partially written.
2346 * This heuristic means that read propagation will be
2347 * conservative, since it will add reg_live_read marks
2348 * to stack slots all the way to first state when programs
2349 * writes+reads less than 8 bytes
2350 */
2351 if (size == BPF_REG_SIZE)
2352 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2353
2354 /* when we zero initialize stack slots mark them as such */
2355 if (reg && register_is_null(reg)) {
2356 /* backtracking doesn't work for STACK_ZERO yet. */
2357 err = mark_chain_precision(env, value_regno);
2358 if (err)
2359 return err;
2360 type = STACK_ZERO;
2361 }
2362
2363 /* Mark slots affected by this stack write. */
2364 for (i = 0; i < size; i++)
2365 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2366 type;
2367 }
2368 return 0;
2369 }
2370
2371 static int check_stack_read(struct bpf_verifier_env *env,
2372 struct bpf_func_state *reg_state /* func where register points to */,
2373 int off, int size, int value_regno)
2374 {
2375 struct bpf_verifier_state *vstate = env->cur_state;
2376 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2377 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2378 struct bpf_reg_state *reg;
2379 u8 *stype;
2380
2381 if (reg_state->allocated_stack <= slot) {
2382 verbose(env, "invalid read from stack off %d+0 size %d\n",
2383 off, size);
2384 return -EACCES;
2385 }
2386 stype = reg_state->stack[spi].slot_type;
2387 reg = &reg_state->stack[spi].spilled_ptr;
2388
2389 if (stype[0] == STACK_SPILL) {
2390 if (size != BPF_REG_SIZE) {
2391 if (reg->type != SCALAR_VALUE) {
2392 verbose_linfo(env, env->insn_idx, "; ");
2393 verbose(env, "invalid size of register fill\n");
2394 return -EACCES;
2395 }
2396 if (value_regno >= 0) {
2397 mark_reg_unknown(env, state->regs, value_regno);
2398 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2399 }
2400 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2401 return 0;
2402 }
2403 for (i = 1; i < BPF_REG_SIZE; i++) {
2404 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2405 verbose(env, "corrupted spill memory\n");
2406 return -EACCES;
2407 }
2408 }
2409
2410 if (value_regno >= 0) {
2411 /* restore register state from stack */
2412 state->regs[value_regno] = *reg;
2413 /* mark reg as written since spilled pointer state likely
2414 * has its liveness marks cleared by is_state_visited()
2415 * which resets stack/reg liveness for state transitions
2416 */
2417 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2418 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2419 /* If value_regno==-1, the caller is asking us whether
2420 * it is acceptable to use this value as a SCALAR_VALUE
2421 * (e.g. for XADD).
2422 * We must not allow unprivileged callers to do that
2423 * with spilled pointers.
2424 */
2425 verbose(env, "leaking pointer from stack off %d\n",
2426 off);
2427 return -EACCES;
2428 }
2429 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2430 } else {
2431 int zeros = 0;
2432
2433 for (i = 0; i < size; i++) {
2434 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2435 continue;
2436 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2437 zeros++;
2438 continue;
2439 }
2440 verbose(env, "invalid read from stack off %d+%d size %d\n",
2441 off, i, size);
2442 return -EACCES;
2443 }
2444 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2445 if (value_regno >= 0) {
2446 if (zeros == size) {
2447 /* any size read into register is zero extended,
2448 * so the whole register == const_zero
2449 */
2450 __mark_reg_const_zero(&state->regs[value_regno]);
2451 /* backtracking doesn't support STACK_ZERO yet,
2452 * so mark it precise here, so that later
2453 * backtracking can stop here.
2454 * Backtracking may not need this if this register
2455 * doesn't participate in pointer adjustment.
2456 * Forward propagation of precise flag is not
2457 * necessary either. This mark is only to stop
2458 * backtracking. Any register that contributed
2459 * to const 0 was marked precise before spill.
2460 */
2461 state->regs[value_regno].precise = true;
2462 } else {
2463 /* have read misc data from the stack */
2464 mark_reg_unknown(env, state->regs, value_regno);
2465 }
2466 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2467 }
2468 }
2469 return 0;
2470 }
2471
2472 static int check_stack_access(struct bpf_verifier_env *env,
2473 const struct bpf_reg_state *reg,
2474 int off, int size)
2475 {
2476 /* Stack accesses must be at a fixed offset, so that we
2477 * can determine what type of data were returned. See
2478 * check_stack_read().
2479 */
2480 if (!tnum_is_const(reg->var_off)) {
2481 char tn_buf[48];
2482
2483 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2484 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2485 tn_buf, off, size);
2486 return -EACCES;
2487 }
2488
2489 if (off >= 0 || off < -MAX_BPF_STACK) {
2490 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2491 return -EACCES;
2492 }
2493
2494 return 0;
2495 }
2496
2497 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2498 int off, int size, enum bpf_access_type type)
2499 {
2500 struct bpf_reg_state *regs = cur_regs(env);
2501 struct bpf_map *map = regs[regno].map_ptr;
2502 u32 cap = bpf_map_flags_to_cap(map);
2503
2504 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2505 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2506 map->value_size, off, size);
2507 return -EACCES;
2508 }
2509
2510 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2511 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2512 map->value_size, off, size);
2513 return -EACCES;
2514 }
2515
2516 return 0;
2517 }
2518
2519 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
2520 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
2521 int off, int size, u32 mem_size,
2522 bool zero_size_allowed)
2523 {
2524 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
2525 struct bpf_reg_state *reg;
2526
2527 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
2528 return 0;
2529
2530 reg = &cur_regs(env)[regno];
2531 switch (reg->type) {
2532 case PTR_TO_MAP_VALUE:
2533 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2534 mem_size, off, size);
2535 break;
2536 case PTR_TO_PACKET:
2537 case PTR_TO_PACKET_META:
2538 case PTR_TO_PACKET_END:
2539 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2540 off, size, regno, reg->id, off, mem_size);
2541 break;
2542 case PTR_TO_MEM:
2543 default:
2544 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
2545 mem_size, off, size);
2546 }
2547
2548 return -EACCES;
2549 }
2550
2551 /* check read/write into a memory region with possible variable offset */
2552 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
2553 int off, int size, u32 mem_size,
2554 bool zero_size_allowed)
2555 {
2556 struct bpf_verifier_state *vstate = env->cur_state;
2557 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2558 struct bpf_reg_state *reg = &state->regs[regno];
2559 int err;
2560
2561 /* We may have adjusted the register pointing to memory region, so we
2562 * need to try adding each of min_value and max_value to off
2563 * to make sure our theoretical access will be safe.
2564 */
2565 if (env->log.level & BPF_LOG_LEVEL)
2566 print_verifier_state(env, state);
2567
2568 /* The minimum value is only important with signed
2569 * comparisons where we can't assume the floor of a
2570 * value is 0. If we are using signed variables for our
2571 * index'es we need to make sure that whatever we use
2572 * will have a set floor within our range.
2573 */
2574 if (reg->smin_value < 0 &&
2575 (reg->smin_value == S64_MIN ||
2576 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2577 reg->smin_value + off < 0)) {
2578 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2579 regno);
2580 return -EACCES;
2581 }
2582 err = __check_mem_access(env, regno, reg->smin_value + off, size,
2583 mem_size, zero_size_allowed);
2584 if (err) {
2585 verbose(env, "R%d min value is outside of the allowed memory range\n",
2586 regno);
2587 return err;
2588 }
2589
2590 /* If we haven't set a max value then we need to bail since we can't be
2591 * sure we won't do bad things.
2592 * If reg->umax_value + off could overflow, treat that as unbounded too.
2593 */
2594 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2595 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
2596 regno);
2597 return -EACCES;
2598 }
2599 err = __check_mem_access(env, regno, reg->umax_value + off, size,
2600 mem_size, zero_size_allowed);
2601 if (err) {
2602 verbose(env, "R%d max value is outside of the allowed memory range\n",
2603 regno);
2604 return err;
2605 }
2606
2607 return 0;
2608 }
2609
2610 /* check read/write into a map element with possible variable offset */
2611 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2612 int off, int size, bool zero_size_allowed)
2613 {
2614 struct bpf_verifier_state *vstate = env->cur_state;
2615 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2616 struct bpf_reg_state *reg = &state->regs[regno];
2617 struct bpf_map *map = reg->map_ptr;
2618 int err;
2619
2620 err = check_mem_region_access(env, regno, off, size, map->value_size,
2621 zero_size_allowed);
2622 if (err)
2623 return err;
2624
2625 if (map_value_has_spin_lock(map)) {
2626 u32 lock = map->spin_lock_off;
2627
2628 /* if any part of struct bpf_spin_lock can be touched by
2629 * load/store reject this program.
2630 * To check that [x1, x2) overlaps with [y1, y2)
2631 * it is sufficient to check x1 < y2 && y1 < x2.
2632 */
2633 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2634 lock < reg->umax_value + off + size) {
2635 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2636 return -EACCES;
2637 }
2638 }
2639 return err;
2640 }
2641
2642 #define MAX_PACKET_OFF 0xffff
2643
2644 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
2645 {
2646 return prog->aux->linked_prog ? prog->aux->linked_prog->type
2647 : prog->type;
2648 }
2649
2650 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2651 const struct bpf_call_arg_meta *meta,
2652 enum bpf_access_type t)
2653 {
2654 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
2655
2656 switch (prog_type) {
2657 /* Program types only with direct read access go here! */
2658 case BPF_PROG_TYPE_LWT_IN:
2659 case BPF_PROG_TYPE_LWT_OUT:
2660 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2661 case BPF_PROG_TYPE_SK_REUSEPORT:
2662 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2663 case BPF_PROG_TYPE_CGROUP_SKB:
2664 if (t == BPF_WRITE)
2665 return false;
2666 /* fallthrough */
2667
2668 /* Program types with direct read + write access go here! */
2669 case BPF_PROG_TYPE_SCHED_CLS:
2670 case BPF_PROG_TYPE_SCHED_ACT:
2671 case BPF_PROG_TYPE_XDP:
2672 case BPF_PROG_TYPE_LWT_XMIT:
2673 case BPF_PROG_TYPE_SK_SKB:
2674 case BPF_PROG_TYPE_SK_MSG:
2675 if (meta)
2676 return meta->pkt_access;
2677
2678 env->seen_direct_write = true;
2679 return true;
2680
2681 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2682 if (t == BPF_WRITE)
2683 env->seen_direct_write = true;
2684
2685 return true;
2686
2687 default:
2688 return false;
2689 }
2690 }
2691
2692 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2693 int size, bool zero_size_allowed)
2694 {
2695 struct bpf_reg_state *regs = cur_regs(env);
2696 struct bpf_reg_state *reg = &regs[regno];
2697 int err;
2698
2699 /* We may have added a variable offset to the packet pointer; but any
2700 * reg->range we have comes after that. We are only checking the fixed
2701 * offset.
2702 */
2703
2704 /* We don't allow negative numbers, because we aren't tracking enough
2705 * detail to prove they're safe.
2706 */
2707 if (reg->smin_value < 0) {
2708 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2709 regno);
2710 return -EACCES;
2711 }
2712 err = __check_mem_access(env, regno, off, size, reg->range,
2713 zero_size_allowed);
2714 if (err) {
2715 verbose(env, "R%d offset is outside of the packet\n", regno);
2716 return err;
2717 }
2718
2719 /* __check_mem_access has made sure "off + size - 1" is within u16.
2720 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2721 * otherwise find_good_pkt_pointers would have refused to set range info
2722 * that __check_mem_access would have rejected this pkt access.
2723 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2724 */
2725 env->prog->aux->max_pkt_offset =
2726 max_t(u32, env->prog->aux->max_pkt_offset,
2727 off + reg->umax_value + size - 1);
2728
2729 return err;
2730 }
2731
2732 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2733 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2734 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2735 u32 *btf_id)
2736 {
2737 struct bpf_insn_access_aux info = {
2738 .reg_type = *reg_type,
2739 .log = &env->log,
2740 };
2741
2742 if (env->ops->is_valid_access &&
2743 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2744 /* A non zero info.ctx_field_size indicates that this field is a
2745 * candidate for later verifier transformation to load the whole
2746 * field and then apply a mask when accessed with a narrower
2747 * access than actual ctx access size. A zero info.ctx_field_size
2748 * will only allow for whole field access and rejects any other
2749 * type of narrower access.
2750 */
2751 *reg_type = info.reg_type;
2752
2753 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL)
2754 *btf_id = info.btf_id;
2755 else
2756 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2757 /* remember the offset of last byte accessed in ctx */
2758 if (env->prog->aux->max_ctx_offset < off + size)
2759 env->prog->aux->max_ctx_offset = off + size;
2760 return 0;
2761 }
2762
2763 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2764 return -EACCES;
2765 }
2766
2767 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2768 int size)
2769 {
2770 if (size < 0 || off < 0 ||
2771 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2772 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2773 off, size);
2774 return -EACCES;
2775 }
2776 return 0;
2777 }
2778
2779 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2780 u32 regno, int off, int size,
2781 enum bpf_access_type t)
2782 {
2783 struct bpf_reg_state *regs = cur_regs(env);
2784 struct bpf_reg_state *reg = &regs[regno];
2785 struct bpf_insn_access_aux info = {};
2786 bool valid;
2787
2788 if (reg->smin_value < 0) {
2789 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2790 regno);
2791 return -EACCES;
2792 }
2793
2794 switch (reg->type) {
2795 case PTR_TO_SOCK_COMMON:
2796 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2797 break;
2798 case PTR_TO_SOCKET:
2799 valid = bpf_sock_is_valid_access(off, size, t, &info);
2800 break;
2801 case PTR_TO_TCP_SOCK:
2802 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2803 break;
2804 case PTR_TO_XDP_SOCK:
2805 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2806 break;
2807 default:
2808 valid = false;
2809 }
2810
2811
2812 if (valid) {
2813 env->insn_aux_data[insn_idx].ctx_field_size =
2814 info.ctx_field_size;
2815 return 0;
2816 }
2817
2818 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2819 regno, reg_type_str[reg->type], off, size);
2820
2821 return -EACCES;
2822 }
2823
2824 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2825 {
2826 return cur_regs(env) + regno;
2827 }
2828
2829 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2830 {
2831 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2832 }
2833
2834 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2835 {
2836 const struct bpf_reg_state *reg = reg_state(env, regno);
2837
2838 return reg->type == PTR_TO_CTX;
2839 }
2840
2841 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2842 {
2843 const struct bpf_reg_state *reg = reg_state(env, regno);
2844
2845 return type_is_sk_pointer(reg->type);
2846 }
2847
2848 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2849 {
2850 const struct bpf_reg_state *reg = reg_state(env, regno);
2851
2852 return type_is_pkt_pointer(reg->type);
2853 }
2854
2855 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2856 {
2857 const struct bpf_reg_state *reg = reg_state(env, regno);
2858
2859 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2860 return reg->type == PTR_TO_FLOW_KEYS;
2861 }
2862
2863 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2864 const struct bpf_reg_state *reg,
2865 int off, int size, bool strict)
2866 {
2867 struct tnum reg_off;
2868 int ip_align;
2869
2870 /* Byte size accesses are always allowed. */
2871 if (!strict || size == 1)
2872 return 0;
2873
2874 /* For platforms that do not have a Kconfig enabling
2875 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2876 * NET_IP_ALIGN is universally set to '2'. And on platforms
2877 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2878 * to this code only in strict mode where we want to emulate
2879 * the NET_IP_ALIGN==2 checking. Therefore use an
2880 * unconditional IP align value of '2'.
2881 */
2882 ip_align = 2;
2883
2884 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2885 if (!tnum_is_aligned(reg_off, size)) {
2886 char tn_buf[48];
2887
2888 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2889 verbose(env,
2890 "misaligned packet access off %d+%s+%d+%d size %d\n",
2891 ip_align, tn_buf, reg->off, off, size);
2892 return -EACCES;
2893 }
2894
2895 return 0;
2896 }
2897
2898 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2899 const struct bpf_reg_state *reg,
2900 const char *pointer_desc,
2901 int off, int size, bool strict)
2902 {
2903 struct tnum reg_off;
2904
2905 /* Byte size accesses are always allowed. */
2906 if (!strict || size == 1)
2907 return 0;
2908
2909 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2910 if (!tnum_is_aligned(reg_off, size)) {
2911 char tn_buf[48];
2912
2913 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2914 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2915 pointer_desc, tn_buf, reg->off, off, size);
2916 return -EACCES;
2917 }
2918
2919 return 0;
2920 }
2921
2922 static int check_ptr_alignment(struct bpf_verifier_env *env,
2923 const struct bpf_reg_state *reg, int off,
2924 int size, bool strict_alignment_once)
2925 {
2926 bool strict = env->strict_alignment || strict_alignment_once;
2927 const char *pointer_desc = "";
2928
2929 switch (reg->type) {
2930 case PTR_TO_PACKET:
2931 case PTR_TO_PACKET_META:
2932 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2933 * right in front, treat it the very same way.
2934 */
2935 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2936 case PTR_TO_FLOW_KEYS:
2937 pointer_desc = "flow keys ";
2938 break;
2939 case PTR_TO_MAP_VALUE:
2940 pointer_desc = "value ";
2941 break;
2942 case PTR_TO_CTX:
2943 pointer_desc = "context ";
2944 break;
2945 case PTR_TO_STACK:
2946 pointer_desc = "stack ";
2947 /* The stack spill tracking logic in check_stack_write()
2948 * and check_stack_read() relies on stack accesses being
2949 * aligned.
2950 */
2951 strict = true;
2952 break;
2953 case PTR_TO_SOCKET:
2954 pointer_desc = "sock ";
2955 break;
2956 case PTR_TO_SOCK_COMMON:
2957 pointer_desc = "sock_common ";
2958 break;
2959 case PTR_TO_TCP_SOCK:
2960 pointer_desc = "tcp_sock ";
2961 break;
2962 case PTR_TO_XDP_SOCK:
2963 pointer_desc = "xdp_sock ";
2964 break;
2965 default:
2966 break;
2967 }
2968 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2969 strict);
2970 }
2971
2972 static int update_stack_depth(struct bpf_verifier_env *env,
2973 const struct bpf_func_state *func,
2974 int off)
2975 {
2976 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2977
2978 if (stack >= -off)
2979 return 0;
2980
2981 /* update known max for given subprogram */
2982 env->subprog_info[func->subprogno].stack_depth = -off;
2983 return 0;
2984 }
2985
2986 /* starting from main bpf function walk all instructions of the function
2987 * and recursively walk all callees that given function can call.
2988 * Ignore jump and exit insns.
2989 * Since recursion is prevented by check_cfg() this algorithm
2990 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2991 */
2992 static int check_max_stack_depth(struct bpf_verifier_env *env)
2993 {
2994 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2995 struct bpf_subprog_info *subprog = env->subprog_info;
2996 struct bpf_insn *insn = env->prog->insnsi;
2997 bool tail_call_reachable = false;
2998 int ret_insn[MAX_CALL_FRAMES];
2999 int ret_prog[MAX_CALL_FRAMES];
3000 int j;
3001
3002 process_func:
3003 /* protect against potential stack overflow that might happen when
3004 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3005 * depth for such case down to 256 so that the worst case scenario
3006 * would result in 8k stack size (32 which is tailcall limit * 256 =
3007 * 8k).
3008 *
3009 * To get the idea what might happen, see an example:
3010 * func1 -> sub rsp, 128
3011 * subfunc1 -> sub rsp, 256
3012 * tailcall1 -> add rsp, 256
3013 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3014 * subfunc2 -> sub rsp, 64
3015 * subfunc22 -> sub rsp, 128
3016 * tailcall2 -> add rsp, 128
3017 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3018 *
3019 * tailcall will unwind the current stack frame but it will not get rid
3020 * of caller's stack as shown on the example above.
3021 */
3022 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3023 verbose(env,
3024 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3025 depth);
3026 return -EACCES;
3027 }
3028 /* round up to 32-bytes, since this is granularity
3029 * of interpreter stack size
3030 */
3031 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3032 if (depth > MAX_BPF_STACK) {
3033 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3034 frame + 1, depth);
3035 return -EACCES;
3036 }
3037 continue_func:
3038 subprog_end = subprog[idx + 1].start;
3039 for (; i < subprog_end; i++) {
3040 if (insn[i].code != (BPF_JMP | BPF_CALL))
3041 continue;
3042 if (insn[i].src_reg != BPF_PSEUDO_CALL)
3043 continue;
3044 /* remember insn and function to return to */
3045 ret_insn[frame] = i + 1;
3046 ret_prog[frame] = idx;
3047
3048 /* find the callee */
3049 i = i + insn[i].imm + 1;
3050 idx = find_subprog(env, i);
3051 if (idx < 0) {
3052 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3053 i);
3054 return -EFAULT;
3055 }
3056
3057 if (subprog[idx].has_tail_call)
3058 tail_call_reachable = true;
3059
3060 frame++;
3061 if (frame >= MAX_CALL_FRAMES) {
3062 verbose(env, "the call stack of %d frames is too deep !\n",
3063 frame);
3064 return -E2BIG;
3065 }
3066 goto process_func;
3067 }
3068 /* if tail call got detected across bpf2bpf calls then mark each of the
3069 * currently present subprog frames as tail call reachable subprogs;
3070 * this info will be utilized by JIT so that we will be preserving the
3071 * tail call counter throughout bpf2bpf calls combined with tailcalls
3072 */
3073 if (tail_call_reachable)
3074 for (j = 0; j < frame; j++)
3075 subprog[ret_prog[j]].tail_call_reachable = true;
3076
3077 /* end of for() loop means the last insn of the 'subprog'
3078 * was reached. Doesn't matter whether it was JA or EXIT
3079 */
3080 if (frame == 0)
3081 return 0;
3082 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3083 frame--;
3084 i = ret_insn[frame];
3085 idx = ret_prog[frame];
3086 goto continue_func;
3087 }
3088
3089 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3090 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3091 const struct bpf_insn *insn, int idx)
3092 {
3093 int start = idx + insn->imm + 1, subprog;
3094
3095 subprog = find_subprog(env, start);
3096 if (subprog < 0) {
3097 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3098 start);
3099 return -EFAULT;
3100 }
3101 return env->subprog_info[subprog].stack_depth;
3102 }
3103 #endif
3104
3105 int check_ctx_reg(struct bpf_verifier_env *env,
3106 const struct bpf_reg_state *reg, int regno)
3107 {
3108 /* Access to ctx or passing it to a helper is only allowed in
3109 * its original, unmodified form.
3110 */
3111
3112 if (reg->off) {
3113 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3114 regno, reg->off);
3115 return -EACCES;
3116 }
3117
3118 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3119 char tn_buf[48];
3120
3121 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3122 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3123 return -EACCES;
3124 }
3125
3126 return 0;
3127 }
3128
3129 static int __check_buffer_access(struct bpf_verifier_env *env,
3130 const char *buf_info,
3131 const struct bpf_reg_state *reg,
3132 int regno, int off, int size)
3133 {
3134 if (off < 0) {
3135 verbose(env,
3136 "R%d invalid %s buffer access: off=%d, size=%d\n",
3137 regno, buf_info, off, size);
3138 return -EACCES;
3139 }
3140 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3141 char tn_buf[48];
3142
3143 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3144 verbose(env,
3145 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3146 regno, off, tn_buf);
3147 return -EACCES;
3148 }
3149
3150 return 0;
3151 }
3152
3153 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3154 const struct bpf_reg_state *reg,
3155 int regno, int off, int size)
3156 {
3157 int err;
3158
3159 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3160 if (err)
3161 return err;
3162
3163 if (off + size > env->prog->aux->max_tp_access)
3164 env->prog->aux->max_tp_access = off + size;
3165
3166 return 0;
3167 }
3168
3169 static int check_buffer_access(struct bpf_verifier_env *env,
3170 const struct bpf_reg_state *reg,
3171 int regno, int off, int size,
3172 bool zero_size_allowed,
3173 const char *buf_info,
3174 u32 *max_access)
3175 {
3176 int err;
3177
3178 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3179 if (err)
3180 return err;
3181
3182 if (off + size > *max_access)
3183 *max_access = off + size;
3184
3185 return 0;
3186 }
3187
3188 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3189 static void zext_32_to_64(struct bpf_reg_state *reg)
3190 {
3191 reg->var_off = tnum_subreg(reg->var_off);
3192 __reg_assign_32_into_64(reg);
3193 }
3194
3195 /* truncate register to smaller size (in bytes)
3196 * must be called with size < BPF_REG_SIZE
3197 */
3198 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3199 {
3200 u64 mask;
3201
3202 /* clear high bits in bit representation */
3203 reg->var_off = tnum_cast(reg->var_off, size);
3204
3205 /* fix arithmetic bounds */
3206 mask = ((u64)1 << (size * 8)) - 1;
3207 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3208 reg->umin_value &= mask;
3209 reg->umax_value &= mask;
3210 } else {
3211 reg->umin_value = 0;
3212 reg->umax_value = mask;
3213 }
3214 reg->smin_value = reg->umin_value;
3215 reg->smax_value = reg->umax_value;
3216
3217 /* If size is smaller than 32bit register the 32bit register
3218 * values are also truncated so we push 64-bit bounds into
3219 * 32-bit bounds. Above were truncated < 32-bits already.
3220 */
3221 if (size >= 4)
3222 return;
3223 __reg_combine_64_into_32(reg);
3224 }
3225
3226 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3227 {
3228 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3229 }
3230
3231 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3232 {
3233 void *ptr;
3234 u64 addr;
3235 int err;
3236
3237 err = map->ops->map_direct_value_addr(map, &addr, off);
3238 if (err)
3239 return err;
3240 ptr = (void *)(long)addr + off;
3241
3242 switch (size) {
3243 case sizeof(u8):
3244 *val = (u64)*(u8 *)ptr;
3245 break;
3246 case sizeof(u16):
3247 *val = (u64)*(u16 *)ptr;
3248 break;
3249 case sizeof(u32):
3250 *val = (u64)*(u32 *)ptr;
3251 break;
3252 case sizeof(u64):
3253 *val = *(u64 *)ptr;
3254 break;
3255 default:
3256 return -EINVAL;
3257 }
3258 return 0;
3259 }
3260
3261 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3262 struct bpf_reg_state *regs,
3263 int regno, int off, int size,
3264 enum bpf_access_type atype,
3265 int value_regno)
3266 {
3267 struct bpf_reg_state *reg = regs + regno;
3268 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
3269 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3270 u32 btf_id;
3271 int ret;
3272
3273 if (off < 0) {
3274 verbose(env,
3275 "R%d is ptr_%s invalid negative access: off=%d\n",
3276 regno, tname, off);
3277 return -EACCES;
3278 }
3279 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3280 char tn_buf[48];
3281
3282 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3283 verbose(env,
3284 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3285 regno, tname, off, tn_buf);
3286 return -EACCES;
3287 }
3288
3289 if (env->ops->btf_struct_access) {
3290 ret = env->ops->btf_struct_access(&env->log, t, off, size,
3291 atype, &btf_id);
3292 } else {
3293 if (atype != BPF_READ) {
3294 verbose(env, "only read is supported\n");
3295 return -EACCES;
3296 }
3297
3298 ret = btf_struct_access(&env->log, t, off, size, atype,
3299 &btf_id);
3300 }
3301
3302 if (ret < 0)
3303 return ret;
3304
3305 if (atype == BPF_READ && value_regno >= 0)
3306 mark_btf_ld_reg(env, regs, value_regno, ret, btf_id);
3307
3308 return 0;
3309 }
3310
3311 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3312 struct bpf_reg_state *regs,
3313 int regno, int off, int size,
3314 enum bpf_access_type atype,
3315 int value_regno)
3316 {
3317 struct bpf_reg_state *reg = regs + regno;
3318 struct bpf_map *map = reg->map_ptr;
3319 const struct btf_type *t;
3320 const char *tname;
3321 u32 btf_id;
3322 int ret;
3323
3324 if (!btf_vmlinux) {
3325 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3326 return -ENOTSUPP;
3327 }
3328
3329 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3330 verbose(env, "map_ptr access not supported for map type %d\n",
3331 map->map_type);
3332 return -ENOTSUPP;
3333 }
3334
3335 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3336 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3337
3338 if (!env->allow_ptr_to_map_access) {
3339 verbose(env,
3340 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3341 tname);
3342 return -EPERM;
3343 }
3344
3345 if (off < 0) {
3346 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3347 regno, tname, off);
3348 return -EACCES;
3349 }
3350
3351 if (atype != BPF_READ) {
3352 verbose(env, "only read from %s is supported\n", tname);
3353 return -EACCES;
3354 }
3355
3356 ret = btf_struct_access(&env->log, t, off, size, atype, &btf_id);
3357 if (ret < 0)
3358 return ret;
3359
3360 if (value_regno >= 0)
3361 mark_btf_ld_reg(env, regs, value_regno, ret, btf_id);
3362
3363 return 0;
3364 }
3365
3366
3367 /* check whether memory at (regno + off) is accessible for t = (read | write)
3368 * if t==write, value_regno is a register which value is stored into memory
3369 * if t==read, value_regno is a register which will receive the value from memory
3370 * if t==write && value_regno==-1, some unknown value is stored into memory
3371 * if t==read && value_regno==-1, don't care what we read from memory
3372 */
3373 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
3374 int off, int bpf_size, enum bpf_access_type t,
3375 int value_regno, bool strict_alignment_once)
3376 {
3377 struct bpf_reg_state *regs = cur_regs(env);
3378 struct bpf_reg_state *reg = regs + regno;
3379 struct bpf_func_state *state;
3380 int size, err = 0;
3381
3382 size = bpf_size_to_bytes(bpf_size);
3383 if (size < 0)
3384 return size;
3385
3386 /* alignment checks will add in reg->off themselves */
3387 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
3388 if (err)
3389 return err;
3390
3391 /* for access checks, reg->off is just part of off */
3392 off += reg->off;
3393
3394 if (reg->type == PTR_TO_MAP_VALUE) {
3395 if (t == BPF_WRITE && value_regno >= 0 &&
3396 is_pointer_value(env, value_regno)) {
3397 verbose(env, "R%d leaks addr into map\n", value_regno);
3398 return -EACCES;
3399 }
3400 err = check_map_access_type(env, regno, off, size, t);
3401 if (err)
3402 return err;
3403 err = check_map_access(env, regno, off, size, false);
3404 if (!err && t == BPF_READ && value_regno >= 0) {
3405 struct bpf_map *map = reg->map_ptr;
3406
3407 /* if map is read-only, track its contents as scalars */
3408 if (tnum_is_const(reg->var_off) &&
3409 bpf_map_is_rdonly(map) &&
3410 map->ops->map_direct_value_addr) {
3411 int map_off = off + reg->var_off.value;
3412 u64 val = 0;
3413
3414 err = bpf_map_direct_read(map, map_off, size,
3415 &val);
3416 if (err)
3417 return err;
3418
3419 regs[value_regno].type = SCALAR_VALUE;
3420 __mark_reg_known(&regs[value_regno], val);
3421 } else {
3422 mark_reg_unknown(env, regs, value_regno);
3423 }
3424 }
3425 } else if (reg->type == PTR_TO_MEM) {
3426 if (t == BPF_WRITE && value_regno >= 0 &&
3427 is_pointer_value(env, value_regno)) {
3428 verbose(env, "R%d leaks addr into mem\n", value_regno);
3429 return -EACCES;
3430 }
3431 err = check_mem_region_access(env, regno, off, size,
3432 reg->mem_size, false);
3433 if (!err && t == BPF_READ && value_regno >= 0)
3434 mark_reg_unknown(env, regs, value_regno);
3435 } else if (reg->type == PTR_TO_CTX) {
3436 enum bpf_reg_type reg_type = SCALAR_VALUE;
3437 u32 btf_id = 0;
3438
3439 if (t == BPF_WRITE && value_regno >= 0 &&
3440 is_pointer_value(env, value_regno)) {
3441 verbose(env, "R%d leaks addr into ctx\n", value_regno);
3442 return -EACCES;
3443 }
3444
3445 err = check_ctx_reg(env, reg, regno);
3446 if (err < 0)
3447 return err;
3448
3449 err = check_ctx_access(env, insn_idx, off, size, t, &reg_type, &btf_id);
3450 if (err)
3451 verbose_linfo(env, insn_idx, "; ");
3452 if (!err && t == BPF_READ && value_regno >= 0) {
3453 /* ctx access returns either a scalar, or a
3454 * PTR_TO_PACKET[_META,_END]. In the latter
3455 * case, we know the offset is zero.
3456 */
3457 if (reg_type == SCALAR_VALUE) {
3458 mark_reg_unknown(env, regs, value_regno);
3459 } else {
3460 mark_reg_known_zero(env, regs,
3461 value_regno);
3462 if (reg_type_may_be_null(reg_type))
3463 regs[value_regno].id = ++env->id_gen;
3464 /* A load of ctx field could have different
3465 * actual load size with the one encoded in the
3466 * insn. When the dst is PTR, it is for sure not
3467 * a sub-register.
3468 */
3469 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
3470 if (reg_type == PTR_TO_BTF_ID ||
3471 reg_type == PTR_TO_BTF_ID_OR_NULL)
3472 regs[value_regno].btf_id = btf_id;
3473 }
3474 regs[value_regno].type = reg_type;
3475 }
3476
3477 } else if (reg->type == PTR_TO_STACK) {
3478 off += reg->var_off.value;
3479 err = check_stack_access(env, reg, off, size);
3480 if (err)
3481 return err;
3482
3483 state = func(env, reg);
3484 err = update_stack_depth(env, state, off);
3485 if (err)
3486 return err;
3487
3488 if (t == BPF_WRITE)
3489 err = check_stack_write(env, state, off, size,
3490 value_regno, insn_idx);
3491 else
3492 err = check_stack_read(env, state, off, size,
3493 value_regno);
3494 } else if (reg_is_pkt_pointer(reg)) {
3495 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3496 verbose(env, "cannot write into packet\n");
3497 return -EACCES;
3498 }
3499 if (t == BPF_WRITE && value_regno >= 0 &&
3500 is_pointer_value(env, value_regno)) {
3501 verbose(env, "R%d leaks addr into packet\n",
3502 value_regno);
3503 return -EACCES;
3504 }
3505 err = check_packet_access(env, regno, off, size, false);
3506 if (!err && t == BPF_READ && value_regno >= 0)
3507 mark_reg_unknown(env, regs, value_regno);
3508 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3509 if (t == BPF_WRITE && value_regno >= 0 &&
3510 is_pointer_value(env, value_regno)) {
3511 verbose(env, "R%d leaks addr into flow keys\n",
3512 value_regno);
3513 return -EACCES;
3514 }
3515
3516 err = check_flow_keys_access(env, off, size);
3517 if (!err && t == BPF_READ && value_regno >= 0)
3518 mark_reg_unknown(env, regs, value_regno);
3519 } else if (type_is_sk_pointer(reg->type)) {
3520 if (t == BPF_WRITE) {
3521 verbose(env, "R%d cannot write into %s\n",
3522 regno, reg_type_str[reg->type]);
3523 return -EACCES;
3524 }
3525 err = check_sock_access(env, insn_idx, regno, off, size, t);
3526 if (!err && value_regno >= 0)
3527 mark_reg_unknown(env, regs, value_regno);
3528 } else if (reg->type == PTR_TO_TP_BUFFER) {
3529 err = check_tp_buffer_access(env, reg, regno, off, size);
3530 if (!err && t == BPF_READ && value_regno >= 0)
3531 mark_reg_unknown(env, regs, value_regno);
3532 } else if (reg->type == PTR_TO_BTF_ID) {
3533 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3534 value_regno);
3535 } else if (reg->type == CONST_PTR_TO_MAP) {
3536 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
3537 value_regno);
3538 } else if (reg->type == PTR_TO_RDONLY_BUF) {
3539 if (t == BPF_WRITE) {
3540 verbose(env, "R%d cannot write into %s\n",
3541 regno, reg_type_str[reg->type]);
3542 return -EACCES;
3543 }
3544 err = check_buffer_access(env, reg, regno, off, size, false,
3545 "rdonly",
3546 &env->prog->aux->max_rdonly_access);
3547 if (!err && value_regno >= 0)
3548 mark_reg_unknown(env, regs, value_regno);
3549 } else if (reg->type == PTR_TO_RDWR_BUF) {
3550 err = check_buffer_access(env, reg, regno, off, size, false,
3551 "rdwr",
3552 &env->prog->aux->max_rdwr_access);
3553 if (!err && t == BPF_READ && value_regno >= 0)
3554 mark_reg_unknown(env, regs, value_regno);
3555 } else {
3556 verbose(env, "R%d invalid mem access '%s'\n", regno,
3557 reg_type_str[reg->type]);
3558 return -EACCES;
3559 }
3560
3561 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3562 regs[value_regno].type == SCALAR_VALUE) {
3563 /* b/h/w load zero-extends, mark upper bits as known 0 */
3564 coerce_reg_to_size(&regs[value_regno], size);
3565 }
3566 return err;
3567 }
3568
3569 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3570 {
3571 int err;
3572
3573 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3574 insn->imm != 0) {
3575 verbose(env, "BPF_XADD uses reserved fields\n");
3576 return -EINVAL;
3577 }
3578
3579 /* check src1 operand */
3580 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3581 if (err)
3582 return err;
3583
3584 /* check src2 operand */
3585 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3586 if (err)
3587 return err;
3588
3589 if (is_pointer_value(env, insn->src_reg)) {
3590 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3591 return -EACCES;
3592 }
3593
3594 if (is_ctx_reg(env, insn->dst_reg) ||
3595 is_pkt_reg(env, insn->dst_reg) ||
3596 is_flow_key_reg(env, insn->dst_reg) ||
3597 is_sk_reg(env, insn->dst_reg)) {
3598 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3599 insn->dst_reg,
3600 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3601 return -EACCES;
3602 }
3603
3604 /* check whether atomic_add can read the memory */
3605 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3606 BPF_SIZE(insn->code), BPF_READ, -1, true);
3607 if (err)
3608 return err;
3609
3610 /* check whether atomic_add can write into the same memory */
3611 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3612 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3613 }
3614
3615 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3616 int off, int access_size,
3617 bool zero_size_allowed)
3618 {
3619 struct bpf_reg_state *reg = reg_state(env, regno);
3620
3621 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3622 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3623 if (tnum_is_const(reg->var_off)) {
3624 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3625 regno, off, access_size);
3626 } else {
3627 char tn_buf[48];
3628
3629 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3630 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3631 regno, tn_buf, access_size);
3632 }
3633 return -EACCES;
3634 }
3635 return 0;
3636 }
3637
3638 /* when register 'regno' is passed into function that will read 'access_size'
3639 * bytes from that pointer, make sure that it's within stack boundary
3640 * and all elements of stack are initialized.
3641 * Unlike most pointer bounds-checking functions, this one doesn't take an
3642 * 'off' argument, so it has to add in reg->off itself.
3643 */
3644 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3645 int access_size, bool zero_size_allowed,
3646 struct bpf_call_arg_meta *meta)
3647 {
3648 struct bpf_reg_state *reg = reg_state(env, regno);
3649 struct bpf_func_state *state = func(env, reg);
3650 int err, min_off, max_off, i, j, slot, spi;
3651
3652 if (tnum_is_const(reg->var_off)) {
3653 min_off = max_off = reg->var_off.value + reg->off;
3654 err = __check_stack_boundary(env, regno, min_off, access_size,
3655 zero_size_allowed);
3656 if (err)
3657 return err;
3658 } else {
3659 /* Variable offset is prohibited for unprivileged mode for
3660 * simplicity since it requires corresponding support in
3661 * Spectre masking for stack ALU.
3662 * See also retrieve_ptr_limit().
3663 */
3664 if (!env->bypass_spec_v1) {
3665 char tn_buf[48];
3666
3667 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3668 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3669 regno, tn_buf);
3670 return -EACCES;
3671 }
3672 /* Only initialized buffer on stack is allowed to be accessed
3673 * with variable offset. With uninitialized buffer it's hard to
3674 * guarantee that whole memory is marked as initialized on
3675 * helper return since specific bounds are unknown what may
3676 * cause uninitialized stack leaking.
3677 */
3678 if (meta && meta->raw_mode)
3679 meta = NULL;
3680
3681 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3682 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3683 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3684 regno);
3685 return -EACCES;
3686 }
3687 min_off = reg->smin_value + reg->off;
3688 max_off = reg->smax_value + reg->off;
3689 err = __check_stack_boundary(env, regno, min_off, access_size,
3690 zero_size_allowed);
3691 if (err) {
3692 verbose(env, "R%d min value is outside of stack bound\n",
3693 regno);
3694 return err;
3695 }
3696 err = __check_stack_boundary(env, regno, max_off, access_size,
3697 zero_size_allowed);
3698 if (err) {
3699 verbose(env, "R%d max value is outside of stack bound\n",
3700 regno);
3701 return err;
3702 }
3703 }
3704
3705 if (meta && meta->raw_mode) {
3706 meta->access_size = access_size;
3707 meta->regno = regno;
3708 return 0;
3709 }
3710
3711 for (i = min_off; i < max_off + access_size; i++) {
3712 u8 *stype;
3713
3714 slot = -i - 1;
3715 spi = slot / BPF_REG_SIZE;
3716 if (state->allocated_stack <= slot)
3717 goto err;
3718 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3719 if (*stype == STACK_MISC)
3720 goto mark;
3721 if (*stype == STACK_ZERO) {
3722 /* helper can write anything into the stack */
3723 *stype = STACK_MISC;
3724 goto mark;
3725 }
3726
3727 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3728 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
3729 goto mark;
3730
3731 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3732 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3733 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3734 for (j = 0; j < BPF_REG_SIZE; j++)
3735 state->stack[spi].slot_type[j] = STACK_MISC;
3736 goto mark;
3737 }
3738
3739 err:
3740 if (tnum_is_const(reg->var_off)) {
3741 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3742 min_off, i - min_off, access_size);
3743 } else {
3744 char tn_buf[48];
3745
3746 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3747 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3748 tn_buf, i - min_off, access_size);
3749 }
3750 return -EACCES;
3751 mark:
3752 /* reading any byte out of 8-byte 'spill_slot' will cause
3753 * the whole slot to be marked as 'read'
3754 */
3755 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3756 state->stack[spi].spilled_ptr.parent,
3757 REG_LIVE_READ64);
3758 }
3759 return update_stack_depth(env, state, min_off);
3760 }
3761
3762 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3763 int access_size, bool zero_size_allowed,
3764 struct bpf_call_arg_meta *meta)
3765 {
3766 struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
3767
3768 switch (reg->type) {
3769 case PTR_TO_PACKET:
3770 case PTR_TO_PACKET_META:
3771 return check_packet_access(env, regno, reg->off, access_size,
3772 zero_size_allowed);
3773 case PTR_TO_MAP_VALUE:
3774 if (check_map_access_type(env, regno, reg->off, access_size,
3775 meta && meta->raw_mode ? BPF_WRITE :
3776 BPF_READ))
3777 return -EACCES;
3778 return check_map_access(env, regno, reg->off, access_size,
3779 zero_size_allowed);
3780 case PTR_TO_MEM:
3781 return check_mem_region_access(env, regno, reg->off,
3782 access_size, reg->mem_size,
3783 zero_size_allowed);
3784 case PTR_TO_RDONLY_BUF:
3785 if (meta && meta->raw_mode)
3786 return -EACCES;
3787 return check_buffer_access(env, reg, regno, reg->off,
3788 access_size, zero_size_allowed,
3789 "rdonly",
3790 &env->prog->aux->max_rdonly_access);
3791 case PTR_TO_RDWR_BUF:
3792 return check_buffer_access(env, reg, regno, reg->off,
3793 access_size, zero_size_allowed,
3794 "rdwr",
3795 &env->prog->aux->max_rdwr_access);
3796 case PTR_TO_STACK:
3797 return check_stack_boundary(env, regno, access_size,
3798 zero_size_allowed, meta);
3799 default: /* scalar_value or invalid ptr */
3800 /* Allow zero-byte read from NULL, regardless of pointer type */
3801 if (zero_size_allowed && access_size == 0 &&
3802 register_is_null(reg))
3803 return 0;
3804
3805 verbose(env, "R%d type=%s expected=%s\n", regno,
3806 reg_type_str[reg->type],
3807 reg_type_str[PTR_TO_STACK]);
3808 return -EACCES;
3809 }
3810 }
3811
3812 /* Implementation details:
3813 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3814 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3815 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3816 * value_or_null->value transition, since the verifier only cares about
3817 * the range of access to valid map value pointer and doesn't care about actual
3818 * address of the map element.
3819 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3820 * reg->id > 0 after value_or_null->value transition. By doing so
3821 * two bpf_map_lookups will be considered two different pointers that
3822 * point to different bpf_spin_locks.
3823 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3824 * dead-locks.
3825 * Since only one bpf_spin_lock is allowed the checks are simpler than
3826 * reg_is_refcounted() logic. The verifier needs to remember only
3827 * one spin_lock instead of array of acquired_refs.
3828 * cur_state->active_spin_lock remembers which map value element got locked
3829 * and clears it after bpf_spin_unlock.
3830 */
3831 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3832 bool is_lock)
3833 {
3834 struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
3835 struct bpf_verifier_state *cur = env->cur_state;
3836 bool is_const = tnum_is_const(reg->var_off);
3837 struct bpf_map *map = reg->map_ptr;
3838 u64 val = reg->var_off.value;
3839
3840 if (!is_const) {
3841 verbose(env,
3842 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3843 regno);
3844 return -EINVAL;
3845 }
3846 if (!map->btf) {
3847 verbose(env,
3848 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3849 map->name);
3850 return -EINVAL;
3851 }
3852 if (!map_value_has_spin_lock(map)) {
3853 if (map->spin_lock_off == -E2BIG)
3854 verbose(env,
3855 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3856 map->name);
3857 else if (map->spin_lock_off == -ENOENT)
3858 verbose(env,
3859 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3860 map->name);
3861 else
3862 verbose(env,
3863 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3864 map->name);
3865 return -EINVAL;
3866 }
3867 if (map->spin_lock_off != val + reg->off) {
3868 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3869 val + reg->off);
3870 return -EINVAL;
3871 }
3872 if (is_lock) {
3873 if (cur->active_spin_lock) {
3874 verbose(env,
3875 "Locking two bpf_spin_locks are not allowed\n");
3876 return -EINVAL;
3877 }
3878 cur->active_spin_lock = reg->id;
3879 } else {
3880 if (!cur->active_spin_lock) {
3881 verbose(env, "bpf_spin_unlock without taking a lock\n");
3882 return -EINVAL;
3883 }
3884 if (cur->active_spin_lock != reg->id) {
3885 verbose(env, "bpf_spin_unlock of different lock\n");
3886 return -EINVAL;
3887 }
3888 cur->active_spin_lock = 0;
3889 }
3890 return 0;
3891 }
3892
3893 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3894 {
3895 return type == ARG_PTR_TO_MEM ||
3896 type == ARG_PTR_TO_MEM_OR_NULL ||
3897 type == ARG_PTR_TO_UNINIT_MEM;
3898 }
3899
3900 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3901 {
3902 return type == ARG_CONST_SIZE ||
3903 type == ARG_CONST_SIZE_OR_ZERO;
3904 }
3905
3906 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
3907 {
3908 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
3909 }
3910
3911 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3912 {
3913 return type == ARG_PTR_TO_INT ||
3914 type == ARG_PTR_TO_LONG;
3915 }
3916
3917 static int int_ptr_type_to_size(enum bpf_arg_type type)
3918 {
3919 if (type == ARG_PTR_TO_INT)
3920 return sizeof(u32);
3921 else if (type == ARG_PTR_TO_LONG)
3922 return sizeof(u64);
3923
3924 return -EINVAL;
3925 }
3926
3927 static int resolve_map_arg_type(struct bpf_verifier_env *env,
3928 const struct bpf_call_arg_meta *meta,
3929 enum bpf_arg_type *arg_type)
3930 {
3931 if (!meta->map_ptr) {
3932 /* kernel subsystem misconfigured verifier */
3933 verbose(env, "invalid map_ptr to access map->type\n");
3934 return -EACCES;
3935 }
3936
3937 switch (meta->map_ptr->map_type) {
3938 case BPF_MAP_TYPE_SOCKMAP:
3939 case BPF_MAP_TYPE_SOCKHASH:
3940 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
3941 *arg_type = ARG_PTR_TO_SOCKET;
3942 } else {
3943 verbose(env, "invalid arg_type for sockmap/sockhash\n");
3944 return -EINVAL;
3945 }
3946 break;
3947
3948 default:
3949 break;
3950 }
3951 return 0;
3952 }
3953
3954 struct bpf_reg_types {
3955 const enum bpf_reg_type types[10];
3956 };
3957
3958 static const struct bpf_reg_types map_key_value_types = {
3959 .types = {
3960 PTR_TO_STACK,
3961 PTR_TO_PACKET,
3962 PTR_TO_PACKET_META,
3963 PTR_TO_MAP_VALUE,
3964 },
3965 };
3966
3967 static const struct bpf_reg_types sock_types = {
3968 .types = {
3969 PTR_TO_SOCK_COMMON,
3970 PTR_TO_SOCKET,
3971 PTR_TO_TCP_SOCK,
3972 PTR_TO_XDP_SOCK,
3973 },
3974 };
3975
3976 static const struct bpf_reg_types mem_types = {
3977 .types = {
3978 PTR_TO_STACK,
3979 PTR_TO_PACKET,
3980 PTR_TO_PACKET_META,
3981 PTR_TO_MAP_VALUE,
3982 PTR_TO_MEM,
3983 PTR_TO_RDONLY_BUF,
3984 PTR_TO_RDWR_BUF,
3985 },
3986 };
3987
3988 static const struct bpf_reg_types int_ptr_types = {
3989 .types = {
3990 PTR_TO_STACK,
3991 PTR_TO_PACKET,
3992 PTR_TO_PACKET_META,
3993 PTR_TO_MAP_VALUE,
3994 },
3995 };
3996
3997 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
3998 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
3999 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4000 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4001 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4002 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4003 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4004
4005 static const struct bpf_reg_types *compatible_reg_types[] = {
4006 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4007 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4008 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4009 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4010 [ARG_CONST_SIZE] = &scalar_types,
4011 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4012 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4013 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4014 [ARG_PTR_TO_CTX] = &context_types,
4015 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4016 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4017 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4018 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4019 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4020 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4021 [ARG_PTR_TO_MEM] = &mem_types,
4022 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4023 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4024 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4025 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4026 [ARG_PTR_TO_INT] = &int_ptr_types,
4027 [ARG_PTR_TO_LONG] = &int_ptr_types,
4028 [__BPF_ARG_TYPE_MAX] = NULL,
4029 };
4030
4031 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4032 const struct bpf_reg_types *compatible)
4033 {
4034 struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
4035 enum bpf_reg_type expected, type = reg->type;
4036 int i, j;
4037
4038 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4039 expected = compatible->types[i];
4040 if (expected == NOT_INIT)
4041 break;
4042
4043 if (type == expected)
4044 return 0;
4045 }
4046
4047 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4048 for (j = 0; j + 1 < i; j++)
4049 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4050 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4051 return -EACCES;
4052 }
4053
4054 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4055 struct bpf_call_arg_meta *meta,
4056 const struct bpf_func_proto *fn)
4057 {
4058 u32 regno = BPF_REG_1 + arg;
4059 struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
4060 enum bpf_arg_type arg_type = fn->arg_type[arg];
4061 const struct bpf_reg_types *compatible;
4062 enum bpf_reg_type type = reg->type;
4063 int err = 0;
4064
4065 if (arg_type == ARG_DONTCARE)
4066 return 0;
4067
4068 err = check_reg_arg(env, regno, SRC_OP);
4069 if (err)
4070 return err;
4071
4072 if (arg_type == ARG_ANYTHING) {
4073 if (is_pointer_value(env, regno)) {
4074 verbose(env, "R%d leaks addr into helper function\n",
4075 regno);
4076 return -EACCES;
4077 }
4078 return 0;
4079 }
4080
4081 if (type_is_pkt_pointer(type) &&
4082 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4083 verbose(env, "helper access to the packet is not allowed\n");
4084 return -EACCES;
4085 }
4086
4087 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4088 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4089 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4090 err = resolve_map_arg_type(env, meta, &arg_type);
4091 if (err)
4092 return err;
4093 }
4094
4095 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4096 /* A NULL register has a SCALAR_VALUE type, so skip
4097 * type checking.
4098 */
4099 goto skip_type_check;
4100
4101 compatible = compatible_reg_types[arg_type];
4102 if (!compatible) {
4103 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4104 return -EFAULT;
4105 }
4106
4107 err = check_reg_type(env, regno, compatible);
4108 if (err)
4109 return err;
4110
4111 if (type == PTR_TO_BTF_ID) {
4112 const u32 *btf_id = fn->arg_btf_id[arg];
4113
4114 if (!btf_id) {
4115 verbose(env, "verifier internal error: missing BTF ID\n");
4116 return -EFAULT;
4117 }
4118
4119 if (!btf_struct_ids_match(&env->log, reg->off, reg->btf_id, *btf_id)) {
4120 verbose(env, "R%d is of type %s but %s is expected\n",
4121 regno, kernel_type_name(reg->btf_id), kernel_type_name(*btf_id));
4122 return -EACCES;
4123 }
4124 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4125 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4126 regno);
4127 return -EACCES;
4128 }
4129 } else if (type == PTR_TO_CTX) {
4130 err = check_ctx_reg(env, reg, regno);
4131 if (err < 0)
4132 return err;
4133 }
4134
4135 skip_type_check:
4136 if (reg->ref_obj_id) {
4137 if (meta->ref_obj_id) {
4138 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4139 regno, reg->ref_obj_id,
4140 meta->ref_obj_id);
4141 return -EFAULT;
4142 }
4143 meta->ref_obj_id = reg->ref_obj_id;
4144 }
4145
4146 if (arg_type == ARG_CONST_MAP_PTR) {
4147 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4148 meta->map_ptr = reg->map_ptr;
4149 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4150 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4151 * check that [key, key + map->key_size) are within
4152 * stack limits and initialized
4153 */
4154 if (!meta->map_ptr) {
4155 /* in function declaration map_ptr must come before
4156 * map_key, so that it's verified and known before
4157 * we have to check map_key here. Otherwise it means
4158 * that kernel subsystem misconfigured verifier
4159 */
4160 verbose(env, "invalid map_ptr to access map->key\n");
4161 return -EACCES;
4162 }
4163 err = check_helper_mem_access(env, regno,
4164 meta->map_ptr->key_size, false,
4165 NULL);
4166 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4167 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4168 !register_is_null(reg)) ||
4169 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4170 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4171 * check [value, value + map->value_size) validity
4172 */
4173 if (!meta->map_ptr) {
4174 /* kernel subsystem misconfigured verifier */
4175 verbose(env, "invalid map_ptr to access map->value\n");
4176 return -EACCES;
4177 }
4178 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4179 err = check_helper_mem_access(env, regno,
4180 meta->map_ptr->value_size, false,
4181 meta);
4182 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4183 if (meta->func_id == BPF_FUNC_spin_lock) {
4184 if (process_spin_lock(env, regno, true))
4185 return -EACCES;
4186 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4187 if (process_spin_lock(env, regno, false))
4188 return -EACCES;
4189 } else {
4190 verbose(env, "verifier internal error\n");
4191 return -EFAULT;
4192 }
4193 } else if (arg_type_is_mem_ptr(arg_type)) {
4194 /* The access to this pointer is only checked when we hit the
4195 * next is_mem_size argument below.
4196 */
4197 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
4198 } else if (arg_type_is_mem_size(arg_type)) {
4199 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
4200
4201 /* This is used to refine r0 return value bounds for helpers
4202 * that enforce this value as an upper bound on return values.
4203 * See do_refine_retval_range() for helpers that can refine
4204 * the return value. C type of helper is u32 so we pull register
4205 * bound from umax_value however, if negative verifier errors
4206 * out. Only upper bounds can be learned because retval is an
4207 * int type and negative retvals are allowed.
4208 */
4209 meta->msize_max_value = reg->umax_value;
4210
4211 /* The register is SCALAR_VALUE; the access check
4212 * happens using its boundaries.
4213 */
4214 if (!tnum_is_const(reg->var_off))
4215 /* For unprivileged variable accesses, disable raw
4216 * mode so that the program is required to
4217 * initialize all the memory that the helper could
4218 * just partially fill up.
4219 */
4220 meta = NULL;
4221
4222 if (reg->smin_value < 0) {
4223 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
4224 regno);
4225 return -EACCES;
4226 }
4227
4228 if (reg->umin_value == 0) {
4229 err = check_helper_mem_access(env, regno - 1, 0,
4230 zero_size_allowed,
4231 meta);
4232 if (err)
4233 return err;
4234 }
4235
4236 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
4237 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
4238 regno);
4239 return -EACCES;
4240 }
4241 err = check_helper_mem_access(env, regno - 1,
4242 reg->umax_value,
4243 zero_size_allowed, meta);
4244 if (!err)
4245 err = mark_chain_precision(env, regno);
4246 } else if (arg_type_is_alloc_size(arg_type)) {
4247 if (!tnum_is_const(reg->var_off)) {
4248 verbose(env, "R%d unbounded size, use 'var &= const' or 'if (var < const)'\n",
4249 regno);
4250 return -EACCES;
4251 }
4252 meta->mem_size = reg->var_off.value;
4253 } else if (arg_type_is_int_ptr(arg_type)) {
4254 int size = int_ptr_type_to_size(arg_type);
4255
4256 err = check_helper_mem_access(env, regno, size, false, meta);
4257 if (err)
4258 return err;
4259 err = check_ptr_alignment(env, reg, 0, size, true);
4260 }
4261
4262 return err;
4263 }
4264
4265 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
4266 {
4267 enum bpf_attach_type eatype = env->prog->expected_attach_type;
4268 enum bpf_prog_type type = resolve_prog_type(env->prog);
4269
4270 if (func_id != BPF_FUNC_map_update_elem)
4271 return false;
4272
4273 /* It's not possible to get access to a locked struct sock in these
4274 * contexts, so updating is safe.
4275 */
4276 switch (type) {
4277 case BPF_PROG_TYPE_TRACING:
4278 if (eatype == BPF_TRACE_ITER)
4279 return true;
4280 break;
4281 case BPF_PROG_TYPE_SOCKET_FILTER:
4282 case BPF_PROG_TYPE_SCHED_CLS:
4283 case BPF_PROG_TYPE_SCHED_ACT:
4284 case BPF_PROG_TYPE_XDP:
4285 case BPF_PROG_TYPE_SK_REUSEPORT:
4286 case BPF_PROG_TYPE_FLOW_DISSECTOR:
4287 case BPF_PROG_TYPE_SK_LOOKUP:
4288 return true;
4289 default:
4290 break;
4291 }
4292
4293 verbose(env, "cannot update sockmap in this context\n");
4294 return false;
4295 }
4296
4297 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
4298 {
4299 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
4300 }
4301
4302 static int check_map_func_compatibility(struct bpf_verifier_env *env,
4303 struct bpf_map *map, int func_id)
4304 {
4305 if (!map)
4306 return 0;
4307
4308 /* We need a two way check, first is from map perspective ... */
4309 switch (map->map_type) {
4310 case BPF_MAP_TYPE_PROG_ARRAY:
4311 if (func_id != BPF_FUNC_tail_call)
4312 goto error;
4313 break;
4314 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
4315 if (func_id != BPF_FUNC_perf_event_read &&
4316 func_id != BPF_FUNC_perf_event_output &&
4317 func_id != BPF_FUNC_skb_output &&
4318 func_id != BPF_FUNC_perf_event_read_value &&
4319 func_id != BPF_FUNC_xdp_output)
4320 goto error;
4321 break;
4322 case BPF_MAP_TYPE_RINGBUF:
4323 if (func_id != BPF_FUNC_ringbuf_output &&
4324 func_id != BPF_FUNC_ringbuf_reserve &&
4325 func_id != BPF_FUNC_ringbuf_submit &&
4326 func_id != BPF_FUNC_ringbuf_discard &&
4327 func_id != BPF_FUNC_ringbuf_query)
4328 goto error;
4329 break;
4330 case BPF_MAP_TYPE_STACK_TRACE:
4331 if (func_id != BPF_FUNC_get_stackid)
4332 goto error;
4333 break;
4334 case BPF_MAP_TYPE_CGROUP_ARRAY:
4335 if (func_id != BPF_FUNC_skb_under_cgroup &&
4336 func_id != BPF_FUNC_current_task_under_cgroup)
4337 goto error;
4338 break;
4339 case BPF_MAP_TYPE_CGROUP_STORAGE:
4340 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
4341 if (func_id != BPF_FUNC_get_local_storage)
4342 goto error;
4343 break;
4344 case BPF_MAP_TYPE_DEVMAP:
4345 case BPF_MAP_TYPE_DEVMAP_HASH:
4346 if (func_id != BPF_FUNC_redirect_map &&
4347 func_id != BPF_FUNC_map_lookup_elem)
4348 goto error;
4349 break;
4350 /* Restrict bpf side of cpumap and xskmap, open when use-cases
4351 * appear.
4352 */
4353 case BPF_MAP_TYPE_CPUMAP:
4354 if (func_id != BPF_FUNC_redirect_map)
4355 goto error;
4356 break;
4357 case BPF_MAP_TYPE_XSKMAP:
4358 if (func_id != BPF_FUNC_redirect_map &&
4359 func_id != BPF_FUNC_map_lookup_elem)
4360 goto error;
4361 break;
4362 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
4363 case BPF_MAP_TYPE_HASH_OF_MAPS:
4364 if (func_id != BPF_FUNC_map_lookup_elem)
4365 goto error;
4366 break;
4367 case BPF_MAP_TYPE_SOCKMAP:
4368 if (func_id != BPF_FUNC_sk_redirect_map &&
4369 func_id != BPF_FUNC_sock_map_update &&
4370 func_id != BPF_FUNC_map_delete_elem &&
4371 func_id != BPF_FUNC_msg_redirect_map &&
4372 func_id != BPF_FUNC_sk_select_reuseport &&
4373 func_id != BPF_FUNC_map_lookup_elem &&
4374 !may_update_sockmap(env, func_id))
4375 goto error;
4376 break;
4377 case BPF_MAP_TYPE_SOCKHASH:
4378 if (func_id != BPF_FUNC_sk_redirect_hash &&
4379 func_id != BPF_FUNC_sock_hash_update &&
4380 func_id != BPF_FUNC_map_delete_elem &&
4381 func_id != BPF_FUNC_msg_redirect_hash &&
4382 func_id != BPF_FUNC_sk_select_reuseport &&
4383 func_id != BPF_FUNC_map_lookup_elem &&
4384 !may_update_sockmap(env, func_id))
4385 goto error;
4386 break;
4387 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
4388 if (func_id != BPF_FUNC_sk_select_reuseport)
4389 goto error;
4390 break;
4391 case BPF_MAP_TYPE_QUEUE:
4392 case BPF_MAP_TYPE_STACK:
4393 if (func_id != BPF_FUNC_map_peek_elem &&
4394 func_id != BPF_FUNC_map_pop_elem &&
4395 func_id != BPF_FUNC_map_push_elem)
4396 goto error;
4397 break;
4398 case BPF_MAP_TYPE_SK_STORAGE:
4399 if (func_id != BPF_FUNC_sk_storage_get &&
4400 func_id != BPF_FUNC_sk_storage_delete)
4401 goto error;
4402 break;
4403 case BPF_MAP_TYPE_INODE_STORAGE:
4404 if (func_id != BPF_FUNC_inode_storage_get &&
4405 func_id != BPF_FUNC_inode_storage_delete)
4406 goto error;
4407 break;
4408 default:
4409 break;
4410 }
4411
4412 /* ... and second from the function itself. */
4413 switch (func_id) {
4414 case BPF_FUNC_tail_call:
4415 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
4416 goto error;
4417 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
4418 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
4419 return -EINVAL;
4420 }
4421 break;
4422 case BPF_FUNC_perf_event_read:
4423 case BPF_FUNC_perf_event_output:
4424 case BPF_FUNC_perf_event_read_value:
4425 case BPF_FUNC_skb_output:
4426 case BPF_FUNC_xdp_output:
4427 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
4428 goto error;
4429 break;
4430 case BPF_FUNC_get_stackid:
4431 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
4432 goto error;
4433 break;
4434 case BPF_FUNC_current_task_under_cgroup:
4435 case BPF_FUNC_skb_under_cgroup:
4436 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
4437 goto error;
4438 break;
4439 case BPF_FUNC_redirect_map:
4440 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
4441 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
4442 map->map_type != BPF_MAP_TYPE_CPUMAP &&
4443 map->map_type != BPF_MAP_TYPE_XSKMAP)
4444 goto error;
4445 break;
4446 case BPF_FUNC_sk_redirect_map:
4447 case BPF_FUNC_msg_redirect_map:
4448 case BPF_FUNC_sock_map_update:
4449 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
4450 goto error;
4451 break;
4452 case BPF_FUNC_sk_redirect_hash:
4453 case BPF_FUNC_msg_redirect_hash:
4454 case BPF_FUNC_sock_hash_update:
4455 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
4456 goto error;
4457 break;
4458 case BPF_FUNC_get_local_storage:
4459 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
4460 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
4461 goto error;
4462 break;
4463 case BPF_FUNC_sk_select_reuseport:
4464 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
4465 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
4466 map->map_type != BPF_MAP_TYPE_SOCKHASH)
4467 goto error;
4468 break;
4469 case BPF_FUNC_map_peek_elem:
4470 case BPF_FUNC_map_pop_elem:
4471 case BPF_FUNC_map_push_elem:
4472 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
4473 map->map_type != BPF_MAP_TYPE_STACK)
4474 goto error;
4475 break;
4476 case BPF_FUNC_sk_storage_get:
4477 case BPF_FUNC_sk_storage_delete:
4478 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
4479 goto error;
4480 break;
4481 case BPF_FUNC_inode_storage_get:
4482 case BPF_FUNC_inode_storage_delete:
4483 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
4484 goto error;
4485 break;
4486 default:
4487 break;
4488 }
4489
4490 return 0;
4491 error:
4492 verbose(env, "cannot pass map_type %d into func %s#%d\n",
4493 map->map_type, func_id_name(func_id), func_id);
4494 return -EINVAL;
4495 }
4496
4497 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
4498 {
4499 int count = 0;
4500
4501 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
4502 count++;
4503 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
4504 count++;
4505 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
4506 count++;
4507 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
4508 count++;
4509 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
4510 count++;
4511
4512 /* We only support one arg being in raw mode at the moment,
4513 * which is sufficient for the helper functions we have
4514 * right now.
4515 */
4516 return count <= 1;
4517 }
4518
4519 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
4520 enum bpf_arg_type arg_next)
4521 {
4522 return (arg_type_is_mem_ptr(arg_curr) &&
4523 !arg_type_is_mem_size(arg_next)) ||
4524 (!arg_type_is_mem_ptr(arg_curr) &&
4525 arg_type_is_mem_size(arg_next));
4526 }
4527
4528 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
4529 {
4530 /* bpf_xxx(..., buf, len) call will access 'len'
4531 * bytes from memory 'buf'. Both arg types need
4532 * to be paired, so make sure there's no buggy
4533 * helper function specification.
4534 */
4535 if (arg_type_is_mem_size(fn->arg1_type) ||
4536 arg_type_is_mem_ptr(fn->arg5_type) ||
4537 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
4538 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
4539 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
4540 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
4541 return false;
4542
4543 return true;
4544 }
4545
4546 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
4547 {
4548 int count = 0;
4549
4550 if (arg_type_may_be_refcounted(fn->arg1_type))
4551 count++;
4552 if (arg_type_may_be_refcounted(fn->arg2_type))
4553 count++;
4554 if (arg_type_may_be_refcounted(fn->arg3_type))
4555 count++;
4556 if (arg_type_may_be_refcounted(fn->arg4_type))
4557 count++;
4558 if (arg_type_may_be_refcounted(fn->arg5_type))
4559 count++;
4560
4561 /* A reference acquiring function cannot acquire
4562 * another refcounted ptr.
4563 */
4564 if (may_be_acquire_function(func_id) && count)
4565 return false;
4566
4567 /* We only support one arg being unreferenced at the moment,
4568 * which is sufficient for the helper functions we have right now.
4569 */
4570 return count <= 1;
4571 }
4572
4573 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
4574 {
4575 int i;
4576
4577 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++)
4578 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
4579 return false;
4580
4581 return true;
4582 }
4583
4584 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
4585 {
4586 return check_raw_mode_ok(fn) &&
4587 check_arg_pair_ok(fn) &&
4588 check_btf_id_ok(fn) &&
4589 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
4590 }
4591
4592 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
4593 * are now invalid, so turn them into unknown SCALAR_VALUE.
4594 */
4595 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
4596 struct bpf_func_state *state)
4597 {
4598 struct bpf_reg_state *regs = state->regs, *reg;
4599 int i;
4600
4601 for (i = 0; i < MAX_BPF_REG; i++)
4602 if (reg_is_pkt_pointer_any(&regs[i]))
4603 mark_reg_unknown(env, regs, i);
4604
4605 bpf_for_each_spilled_reg(i, state, reg) {
4606 if (!reg)
4607 continue;
4608 if (reg_is_pkt_pointer_any(reg))
4609 __mark_reg_unknown(env, reg);
4610 }
4611 }
4612
4613 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
4614 {
4615 struct bpf_verifier_state *vstate = env->cur_state;
4616 int i;
4617
4618 for (i = 0; i <= vstate->curframe; i++)
4619 __clear_all_pkt_pointers(env, vstate->frame[i]);
4620 }
4621
4622 static void release_reg_references(struct bpf_verifier_env *env,
4623 struct bpf_func_state *state,
4624 int ref_obj_id)
4625 {
4626 struct bpf_reg_state *regs = state->regs, *reg;
4627 int i;
4628
4629 for (i = 0; i < MAX_BPF_REG; i++)
4630 if (regs[i].ref_obj_id == ref_obj_id)
4631 mark_reg_unknown(env, regs, i);
4632
4633 bpf_for_each_spilled_reg(i, state, reg) {
4634 if (!reg)
4635 continue;
4636 if (reg->ref_obj_id == ref_obj_id)
4637 __mark_reg_unknown(env, reg);
4638 }
4639 }
4640
4641 /* The pointer with the specified id has released its reference to kernel
4642 * resources. Identify all copies of the same pointer and clear the reference.
4643 */
4644 static int release_reference(struct bpf_verifier_env *env,
4645 int ref_obj_id)
4646 {
4647 struct bpf_verifier_state *vstate = env->cur_state;
4648 int err;
4649 int i;
4650
4651 err = release_reference_state(cur_func(env), ref_obj_id);
4652 if (err)
4653 return err;
4654
4655 for (i = 0; i <= vstate->curframe; i++)
4656 release_reg_references(env, vstate->frame[i], ref_obj_id);
4657
4658 return 0;
4659 }
4660
4661 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
4662 struct bpf_reg_state *regs)
4663 {
4664 int i;
4665
4666 /* after the call registers r0 - r5 were scratched */
4667 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4668 mark_reg_not_init(env, regs, caller_saved[i]);
4669 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4670 }
4671 }
4672
4673 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
4674 int *insn_idx)
4675 {
4676 struct bpf_verifier_state *state = env->cur_state;
4677 struct bpf_func_info_aux *func_info_aux;
4678 struct bpf_func_state *caller, *callee;
4679 int i, err, subprog, target_insn;
4680 bool is_global = false;
4681
4682 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
4683 verbose(env, "the call stack of %d frames is too deep\n",
4684 state->curframe + 2);
4685 return -E2BIG;
4686 }
4687
4688 target_insn = *insn_idx + insn->imm;
4689 subprog = find_subprog(env, target_insn + 1);
4690 if (subprog < 0) {
4691 verbose(env, "verifier bug. No program starts at insn %d\n",
4692 target_insn + 1);
4693 return -EFAULT;
4694 }
4695
4696 caller = state->frame[state->curframe];
4697 if (state->frame[state->curframe + 1]) {
4698 verbose(env, "verifier bug. Frame %d already allocated\n",
4699 state->curframe + 1);
4700 return -EFAULT;
4701 }
4702
4703 func_info_aux = env->prog->aux->func_info_aux;
4704 if (func_info_aux)
4705 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4706 err = btf_check_func_arg_match(env, subprog, caller->regs);
4707 if (err == -EFAULT)
4708 return err;
4709 if (is_global) {
4710 if (err) {
4711 verbose(env, "Caller passes invalid args into func#%d\n",
4712 subprog);
4713 return err;
4714 } else {
4715 if (env->log.level & BPF_LOG_LEVEL)
4716 verbose(env,
4717 "Func#%d is global and valid. Skipping.\n",
4718 subprog);
4719 clear_caller_saved_regs(env, caller->regs);
4720
4721 /* All global functions return SCALAR_VALUE */
4722 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4723
4724 /* continue with next insn after call */
4725 return 0;
4726 }
4727 }
4728
4729 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4730 if (!callee)
4731 return -ENOMEM;
4732 state->frame[state->curframe + 1] = callee;
4733
4734 /* callee cannot access r0, r6 - r9 for reading and has to write
4735 * into its own stack before reading from it.
4736 * callee can read/write into caller's stack
4737 */
4738 init_func_state(env, callee,
4739 /* remember the callsite, it will be used by bpf_exit */
4740 *insn_idx /* callsite */,
4741 state->curframe + 1 /* frameno within this callchain */,
4742 subprog /* subprog number within this prog */);
4743
4744 /* Transfer references to the callee */
4745 err = transfer_reference_state(callee, caller);
4746 if (err)
4747 return err;
4748
4749 /* copy r1 - r5 args that callee can access. The copy includes parent
4750 * pointers, which connects us up to the liveness chain
4751 */
4752 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4753 callee->regs[i] = caller->regs[i];
4754
4755 clear_caller_saved_regs(env, caller->regs);
4756
4757 /* only increment it after check_reg_arg() finished */
4758 state->curframe++;
4759
4760 /* and go analyze first insn of the callee */
4761 *insn_idx = target_insn;
4762
4763 if (env->log.level & BPF_LOG_LEVEL) {
4764 verbose(env, "caller:\n");
4765 print_verifier_state(env, caller);
4766 verbose(env, "callee:\n");
4767 print_verifier_state(env, callee);
4768 }
4769 return 0;
4770 }
4771
4772 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4773 {
4774 struct bpf_verifier_state *state = env->cur_state;
4775 struct bpf_func_state *caller, *callee;
4776 struct bpf_reg_state *r0;
4777 int err;
4778
4779 callee = state->frame[state->curframe];
4780 r0 = &callee->regs[BPF_REG_0];
4781 if (r0->type == PTR_TO_STACK) {
4782 /* technically it's ok to return caller's stack pointer
4783 * (or caller's caller's pointer) back to the caller,
4784 * since these pointers are valid. Only current stack
4785 * pointer will be invalid as soon as function exits,
4786 * but let's be conservative
4787 */
4788 verbose(env, "cannot return stack pointer to the caller\n");
4789 return -EINVAL;
4790 }
4791
4792 state->curframe--;
4793 caller = state->frame[state->curframe];
4794 /* return to the caller whatever r0 had in the callee */
4795 caller->regs[BPF_REG_0] = *r0;
4796
4797 /* Transfer references to the caller */
4798 err = transfer_reference_state(caller, callee);
4799 if (err)
4800 return err;
4801
4802 *insn_idx = callee->callsite + 1;
4803 if (env->log.level & BPF_LOG_LEVEL) {
4804 verbose(env, "returning from callee:\n");
4805 print_verifier_state(env, callee);
4806 verbose(env, "to caller at %d:\n", *insn_idx);
4807 print_verifier_state(env, caller);
4808 }
4809 /* clear everything in the callee */
4810 free_func_state(callee);
4811 state->frame[state->curframe + 1] = NULL;
4812 return 0;
4813 }
4814
4815 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4816 int func_id,
4817 struct bpf_call_arg_meta *meta)
4818 {
4819 struct bpf_reg_state *ret_reg = &regs[BPF_REG_0];
4820
4821 if (ret_type != RET_INTEGER ||
4822 (func_id != BPF_FUNC_get_stack &&
4823 func_id != BPF_FUNC_probe_read_str &&
4824 func_id != BPF_FUNC_probe_read_kernel_str &&
4825 func_id != BPF_FUNC_probe_read_user_str))
4826 return;
4827
4828 ret_reg->smax_value = meta->msize_max_value;
4829 ret_reg->s32_max_value = meta->msize_max_value;
4830 __reg_deduce_bounds(ret_reg);
4831 __reg_bound_offset(ret_reg);
4832 __update_reg_bounds(ret_reg);
4833 }
4834
4835 static int
4836 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4837 int func_id, int insn_idx)
4838 {
4839 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4840 struct bpf_map *map = meta->map_ptr;
4841
4842 if (func_id != BPF_FUNC_tail_call &&
4843 func_id != BPF_FUNC_map_lookup_elem &&
4844 func_id != BPF_FUNC_map_update_elem &&
4845 func_id != BPF_FUNC_map_delete_elem &&
4846 func_id != BPF_FUNC_map_push_elem &&
4847 func_id != BPF_FUNC_map_pop_elem &&
4848 func_id != BPF_FUNC_map_peek_elem)
4849 return 0;
4850
4851 if (map == NULL) {
4852 verbose(env, "kernel subsystem misconfigured verifier\n");
4853 return -EINVAL;
4854 }
4855
4856 /* In case of read-only, some additional restrictions
4857 * need to be applied in order to prevent altering the
4858 * state of the map from program side.
4859 */
4860 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4861 (func_id == BPF_FUNC_map_delete_elem ||
4862 func_id == BPF_FUNC_map_update_elem ||
4863 func_id == BPF_FUNC_map_push_elem ||
4864 func_id == BPF_FUNC_map_pop_elem)) {
4865 verbose(env, "write into map forbidden\n");
4866 return -EACCES;
4867 }
4868
4869 if (!BPF_MAP_PTR(aux->map_ptr_state))
4870 bpf_map_ptr_store(aux, meta->map_ptr,
4871 !meta->map_ptr->bypass_spec_v1);
4872 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4873 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4874 !meta->map_ptr->bypass_spec_v1);
4875 return 0;
4876 }
4877
4878 static int
4879 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4880 int func_id, int insn_idx)
4881 {
4882 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4883 struct bpf_reg_state *regs = cur_regs(env), *reg;
4884 struct bpf_map *map = meta->map_ptr;
4885 struct tnum range;
4886 u64 val;
4887 int err;
4888
4889 if (func_id != BPF_FUNC_tail_call)
4890 return 0;
4891 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4892 verbose(env, "kernel subsystem misconfigured verifier\n");
4893 return -EINVAL;
4894 }
4895
4896 range = tnum_range(0, map->max_entries - 1);
4897 reg = &regs[BPF_REG_3];
4898
4899 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4900 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4901 return 0;
4902 }
4903
4904 err = mark_chain_precision(env, BPF_REG_3);
4905 if (err)
4906 return err;
4907
4908 val = reg->var_off.value;
4909 if (bpf_map_key_unseen(aux))
4910 bpf_map_key_store(aux, val);
4911 else if (!bpf_map_key_poisoned(aux) &&
4912 bpf_map_key_immediate(aux) != val)
4913 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4914 return 0;
4915 }
4916
4917 static int check_reference_leak(struct bpf_verifier_env *env)
4918 {
4919 struct bpf_func_state *state = cur_func(env);
4920 int i;
4921
4922 for (i = 0; i < state->acquired_refs; i++) {
4923 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4924 state->refs[i].id, state->refs[i].insn_idx);
4925 }
4926 return state->acquired_refs ? -EINVAL : 0;
4927 }
4928
4929 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4930 {
4931 const struct bpf_func_proto *fn = NULL;
4932 struct bpf_reg_state *regs;
4933 struct bpf_call_arg_meta meta;
4934 bool changes_data;
4935 int i, err;
4936
4937 /* find function prototype */
4938 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4939 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4940 func_id);
4941 return -EINVAL;
4942 }
4943
4944 if (env->ops->get_func_proto)
4945 fn = env->ops->get_func_proto(func_id, env->prog);
4946 if (!fn) {
4947 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4948 func_id);
4949 return -EINVAL;
4950 }
4951
4952 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4953 if (!env->prog->gpl_compatible && fn->gpl_only) {
4954 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4955 return -EINVAL;
4956 }
4957
4958 if (fn->allowed && !fn->allowed(env->prog)) {
4959 verbose(env, "helper call is not allowed in probe\n");
4960 return -EINVAL;
4961 }
4962
4963 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4964 changes_data = bpf_helper_changes_pkt_data(fn->func);
4965 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4966 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4967 func_id_name(func_id), func_id);
4968 return -EINVAL;
4969 }
4970
4971 memset(&meta, 0, sizeof(meta));
4972 meta.pkt_access = fn->pkt_access;
4973
4974 err = check_func_proto(fn, func_id);
4975 if (err) {
4976 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4977 func_id_name(func_id), func_id);
4978 return err;
4979 }
4980
4981 meta.func_id = func_id;
4982 /* check args */
4983 for (i = 0; i < 5; i++) {
4984 err = check_func_arg(env, i, &meta, fn);
4985 if (err)
4986 return err;
4987 }
4988
4989 err = record_func_map(env, &meta, func_id, insn_idx);
4990 if (err)
4991 return err;
4992
4993 err = record_func_key(env, &meta, func_id, insn_idx);
4994 if (err)
4995 return err;
4996
4997 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4998 * is inferred from register state.
4999 */
5000 for (i = 0; i < meta.access_size; i++) {
5001 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
5002 BPF_WRITE, -1, false);
5003 if (err)
5004 return err;
5005 }
5006
5007 if (func_id == BPF_FUNC_tail_call) {
5008 err = check_reference_leak(env);
5009 if (err) {
5010 verbose(env, "tail_call would lead to reference leak\n");
5011 return err;
5012 }
5013 } else if (is_release_function(func_id)) {
5014 err = release_reference(env, meta.ref_obj_id);
5015 if (err) {
5016 verbose(env, "func %s#%d reference has not been acquired before\n",
5017 func_id_name(func_id), func_id);
5018 return err;
5019 }
5020 }
5021
5022 regs = cur_regs(env);
5023
5024 /* check that flags argument in get_local_storage(map, flags) is 0,
5025 * this is required because get_local_storage() can't return an error.
5026 */
5027 if (func_id == BPF_FUNC_get_local_storage &&
5028 !register_is_null(&regs[BPF_REG_2])) {
5029 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
5030 return -EINVAL;
5031 }
5032
5033 /* reset caller saved regs */
5034 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5035 mark_reg_not_init(env, regs, caller_saved[i]);
5036 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5037 }
5038
5039 /* helper call returns 64-bit value. */
5040 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5041
5042 /* update return register (already marked as written above) */
5043 if (fn->ret_type == RET_INTEGER) {
5044 /* sets type to SCALAR_VALUE */
5045 mark_reg_unknown(env, regs, BPF_REG_0);
5046 } else if (fn->ret_type == RET_VOID) {
5047 regs[BPF_REG_0].type = NOT_INIT;
5048 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
5049 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5050 /* There is no offset yet applied, variable or fixed */
5051 mark_reg_known_zero(env, regs, BPF_REG_0);
5052 /* remember map_ptr, so that check_map_access()
5053 * can check 'value_size' boundary of memory access
5054 * to map element returned from bpf_map_lookup_elem()
5055 */
5056 if (meta.map_ptr == NULL) {
5057 verbose(env,
5058 "kernel subsystem misconfigured verifier\n");
5059 return -EINVAL;
5060 }
5061 regs[BPF_REG_0].map_ptr = meta.map_ptr;
5062 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5063 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
5064 if (map_value_has_spin_lock(meta.map_ptr))
5065 regs[BPF_REG_0].id = ++env->id_gen;
5066 } else {
5067 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
5068 regs[BPF_REG_0].id = ++env->id_gen;
5069 }
5070 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
5071 mark_reg_known_zero(env, regs, BPF_REG_0);
5072 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
5073 regs[BPF_REG_0].id = ++env->id_gen;
5074 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
5075 mark_reg_known_zero(env, regs, BPF_REG_0);
5076 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
5077 regs[BPF_REG_0].id = ++env->id_gen;
5078 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
5079 mark_reg_known_zero(env, regs, BPF_REG_0);
5080 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
5081 regs[BPF_REG_0].id = ++env->id_gen;
5082 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
5083 mark_reg_known_zero(env, regs, BPF_REG_0);
5084 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
5085 regs[BPF_REG_0].id = ++env->id_gen;
5086 regs[BPF_REG_0].mem_size = meta.mem_size;
5087 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL) {
5088 int ret_btf_id;
5089
5090 mark_reg_known_zero(env, regs, BPF_REG_0);
5091 regs[BPF_REG_0].type = PTR_TO_BTF_ID_OR_NULL;
5092 ret_btf_id = *fn->ret_btf_id;
5093 if (ret_btf_id == 0) {
5094 verbose(env, "invalid return type %d of func %s#%d\n",
5095 fn->ret_type, func_id_name(func_id), func_id);
5096 return -EINVAL;
5097 }
5098 regs[BPF_REG_0].btf_id = ret_btf_id;
5099 } else {
5100 verbose(env, "unknown return type %d of func %s#%d\n",
5101 fn->ret_type, func_id_name(func_id), func_id);
5102 return -EINVAL;
5103 }
5104
5105 if (is_ptr_cast_function(func_id)) {
5106 /* For release_reference() */
5107 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
5108 } else if (is_acquire_function(func_id, meta.map_ptr)) {
5109 int id = acquire_reference_state(env, insn_idx);
5110
5111 if (id < 0)
5112 return id;
5113 /* For mark_ptr_or_null_reg() */
5114 regs[BPF_REG_0].id = id;
5115 /* For release_reference() */
5116 regs[BPF_REG_0].ref_obj_id = id;
5117 }
5118
5119 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
5120
5121 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
5122 if (err)
5123 return err;
5124
5125 if ((func_id == BPF_FUNC_get_stack ||
5126 func_id == BPF_FUNC_get_task_stack) &&
5127 !env->prog->has_callchain_buf) {
5128 const char *err_str;
5129
5130 #ifdef CONFIG_PERF_EVENTS
5131 err = get_callchain_buffers(sysctl_perf_event_max_stack);
5132 err_str = "cannot get callchain buffer for func %s#%d\n";
5133 #else
5134 err = -ENOTSUPP;
5135 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
5136 #endif
5137 if (err) {
5138 verbose(env, err_str, func_id_name(func_id), func_id);
5139 return err;
5140 }
5141
5142 env->prog->has_callchain_buf = true;
5143 }
5144
5145 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
5146 env->prog->call_get_stack = true;
5147
5148 if (changes_data)
5149 clear_all_pkt_pointers(env);
5150 return 0;
5151 }
5152
5153 static bool signed_add_overflows(s64 a, s64 b)
5154 {
5155 /* Do the add in u64, where overflow is well-defined */
5156 s64 res = (s64)((u64)a + (u64)b);
5157
5158 if (b < 0)
5159 return res > a;
5160 return res < a;
5161 }
5162
5163 static bool signed_add32_overflows(s64 a, s64 b)
5164 {
5165 /* Do the add in u32, where overflow is well-defined */
5166 s32 res = (s32)((u32)a + (u32)b);
5167
5168 if (b < 0)
5169 return res > a;
5170 return res < a;
5171 }
5172
5173 static bool signed_sub_overflows(s32 a, s32 b)
5174 {
5175 /* Do the sub in u64, where overflow is well-defined */
5176 s64 res = (s64)((u64)a - (u64)b);
5177
5178 if (b < 0)
5179 return res < a;
5180 return res > a;
5181 }
5182
5183 static bool signed_sub32_overflows(s32 a, s32 b)
5184 {
5185 /* Do the sub in u64, where overflow is well-defined */
5186 s32 res = (s32)((u32)a - (u32)b);
5187
5188 if (b < 0)
5189 return res < a;
5190 return res > a;
5191 }
5192
5193 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
5194 const struct bpf_reg_state *reg,
5195 enum bpf_reg_type type)
5196 {
5197 bool known = tnum_is_const(reg->var_off);
5198 s64 val = reg->var_off.value;
5199 s64 smin = reg->smin_value;
5200
5201 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
5202 verbose(env, "math between %s pointer and %lld is not allowed\n",
5203 reg_type_str[type], val);
5204 return false;
5205 }
5206
5207 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
5208 verbose(env, "%s pointer offset %d is not allowed\n",
5209 reg_type_str[type], reg->off);
5210 return false;
5211 }
5212
5213 if (smin == S64_MIN) {
5214 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
5215 reg_type_str[type]);
5216 return false;
5217 }
5218
5219 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
5220 verbose(env, "value %lld makes %s pointer be out of bounds\n",
5221 smin, reg_type_str[type]);
5222 return false;
5223 }
5224
5225 return true;
5226 }
5227
5228 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
5229 {
5230 return &env->insn_aux_data[env->insn_idx];
5231 }
5232
5233 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
5234 u32 *ptr_limit, u8 opcode, bool off_is_neg)
5235 {
5236 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
5237 (opcode == BPF_SUB && !off_is_neg);
5238 u32 off;
5239
5240 switch (ptr_reg->type) {
5241 case PTR_TO_STACK:
5242 /* Indirect variable offset stack access is prohibited in
5243 * unprivileged mode so it's not handled here.
5244 */
5245 off = ptr_reg->off + ptr_reg->var_off.value;
5246 if (mask_to_left)
5247 *ptr_limit = MAX_BPF_STACK + off;
5248 else
5249 *ptr_limit = -off;
5250 return 0;
5251 case PTR_TO_MAP_VALUE:
5252 if (mask_to_left) {
5253 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
5254 } else {
5255 off = ptr_reg->smin_value + ptr_reg->off;
5256 *ptr_limit = ptr_reg->map_ptr->value_size - off;
5257 }
5258 return 0;
5259 default:
5260 return -EINVAL;
5261 }
5262 }
5263
5264 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
5265 const struct bpf_insn *insn)
5266 {
5267 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
5268 }
5269
5270 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
5271 u32 alu_state, u32 alu_limit)
5272 {
5273 /* If we arrived here from different branches with different
5274 * state or limits to sanitize, then this won't work.
5275 */
5276 if (aux->alu_state &&
5277 (aux->alu_state != alu_state ||
5278 aux->alu_limit != alu_limit))
5279 return -EACCES;
5280
5281 /* Corresponding fixup done in fixup_bpf_calls(). */
5282 aux->alu_state = alu_state;
5283 aux->alu_limit = alu_limit;
5284 return 0;
5285 }
5286
5287 static int sanitize_val_alu(struct bpf_verifier_env *env,
5288 struct bpf_insn *insn)
5289 {
5290 struct bpf_insn_aux_data *aux = cur_aux(env);
5291
5292 if (can_skip_alu_sanitation(env, insn))
5293 return 0;
5294
5295 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
5296 }
5297
5298 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
5299 struct bpf_insn *insn,
5300 const struct bpf_reg_state *ptr_reg,
5301 struct bpf_reg_state *dst_reg,
5302 bool off_is_neg)
5303 {
5304 struct bpf_verifier_state *vstate = env->cur_state;
5305 struct bpf_insn_aux_data *aux = cur_aux(env);
5306 bool ptr_is_dst_reg = ptr_reg == dst_reg;
5307 u8 opcode = BPF_OP(insn->code);
5308 u32 alu_state, alu_limit;
5309 struct bpf_reg_state tmp;
5310 bool ret;
5311
5312 if (can_skip_alu_sanitation(env, insn))
5313 return 0;
5314
5315 /* We already marked aux for masking from non-speculative
5316 * paths, thus we got here in the first place. We only care
5317 * to explore bad access from here.
5318 */
5319 if (vstate->speculative)
5320 goto do_sim;
5321
5322 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
5323 alu_state |= ptr_is_dst_reg ?
5324 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
5325
5326 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
5327 return 0;
5328 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
5329 return -EACCES;
5330 do_sim:
5331 /* Simulate and find potential out-of-bounds access under
5332 * speculative execution from truncation as a result of
5333 * masking when off was not within expected range. If off
5334 * sits in dst, then we temporarily need to move ptr there
5335 * to simulate dst (== 0) +/-= ptr. Needed, for example,
5336 * for cases where we use K-based arithmetic in one direction
5337 * and truncated reg-based in the other in order to explore
5338 * bad access.
5339 */
5340 if (!ptr_is_dst_reg) {
5341 tmp = *dst_reg;
5342 *dst_reg = *ptr_reg;
5343 }
5344 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
5345 if (!ptr_is_dst_reg && ret)
5346 *dst_reg = tmp;
5347 return !ret ? -EFAULT : 0;
5348 }
5349
5350 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
5351 * Caller should also handle BPF_MOV case separately.
5352 * If we return -EACCES, caller may want to try again treating pointer as a
5353 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
5354 */
5355 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
5356 struct bpf_insn *insn,
5357 const struct bpf_reg_state *ptr_reg,
5358 const struct bpf_reg_state *off_reg)
5359 {
5360 struct bpf_verifier_state *vstate = env->cur_state;
5361 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5362 struct bpf_reg_state *regs = state->regs, *dst_reg;
5363 bool known = tnum_is_const(off_reg->var_off);
5364 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
5365 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
5366 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
5367 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
5368 u32 dst = insn->dst_reg, src = insn->src_reg;
5369 u8 opcode = BPF_OP(insn->code);
5370 int ret;
5371
5372 dst_reg = &regs[dst];
5373
5374 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
5375 smin_val > smax_val || umin_val > umax_val) {
5376 /* Taint dst register if offset had invalid bounds derived from
5377 * e.g. dead branches.
5378 */
5379 __mark_reg_unknown(env, dst_reg);
5380 return 0;
5381 }
5382
5383 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5384 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
5385 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5386 __mark_reg_unknown(env, dst_reg);
5387 return 0;
5388 }
5389
5390 verbose(env,
5391 "R%d 32-bit pointer arithmetic prohibited\n",
5392 dst);
5393 return -EACCES;
5394 }
5395
5396 switch (ptr_reg->type) {
5397 case PTR_TO_MAP_VALUE_OR_NULL:
5398 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
5399 dst, reg_type_str[ptr_reg->type]);
5400 return -EACCES;
5401 case CONST_PTR_TO_MAP:
5402 /* smin_val represents the known value */
5403 if (known && smin_val == 0 && opcode == BPF_ADD)
5404 break;
5405 /* fall-through */
5406 case PTR_TO_PACKET_END:
5407 case PTR_TO_SOCKET:
5408 case PTR_TO_SOCKET_OR_NULL:
5409 case PTR_TO_SOCK_COMMON:
5410 case PTR_TO_SOCK_COMMON_OR_NULL:
5411 case PTR_TO_TCP_SOCK:
5412 case PTR_TO_TCP_SOCK_OR_NULL:
5413 case PTR_TO_XDP_SOCK:
5414 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
5415 dst, reg_type_str[ptr_reg->type]);
5416 return -EACCES;
5417 case PTR_TO_MAP_VALUE:
5418 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
5419 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
5420 off_reg == dst_reg ? dst : src);
5421 return -EACCES;
5422 }
5423 /* fall-through */
5424 default:
5425 break;
5426 }
5427
5428 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
5429 * The id may be overwritten later if we create a new variable offset.
5430 */
5431 dst_reg->type = ptr_reg->type;
5432 dst_reg->id = ptr_reg->id;
5433
5434 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
5435 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
5436 return -EINVAL;
5437
5438 /* pointer types do not carry 32-bit bounds at the moment. */
5439 __mark_reg32_unbounded(dst_reg);
5440
5441 switch (opcode) {
5442 case BPF_ADD:
5443 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
5444 if (ret < 0) {
5445 verbose(env, "R%d tried to add from different maps or paths\n", dst);
5446 return ret;
5447 }
5448 /* We can take a fixed offset as long as it doesn't overflow
5449 * the s32 'off' field
5450 */
5451 if (known && (ptr_reg->off + smin_val ==
5452 (s64)(s32)(ptr_reg->off + smin_val))) {
5453 /* pointer += K. Accumulate it into fixed offset */
5454 dst_reg->smin_value = smin_ptr;
5455 dst_reg->smax_value = smax_ptr;
5456 dst_reg->umin_value = umin_ptr;
5457 dst_reg->umax_value = umax_ptr;
5458 dst_reg->var_off = ptr_reg->var_off;
5459 dst_reg->off = ptr_reg->off + smin_val;
5460 dst_reg->raw = ptr_reg->raw;
5461 break;
5462 }
5463 /* A new variable offset is created. Note that off_reg->off
5464 * == 0, since it's a scalar.
5465 * dst_reg gets the pointer type and since some positive
5466 * integer value was added to the pointer, give it a new 'id'
5467 * if it's a PTR_TO_PACKET.
5468 * this creates a new 'base' pointer, off_reg (variable) gets
5469 * added into the variable offset, and we copy the fixed offset
5470 * from ptr_reg.
5471 */
5472 if (signed_add_overflows(smin_ptr, smin_val) ||
5473 signed_add_overflows(smax_ptr, smax_val)) {
5474 dst_reg->smin_value = S64_MIN;
5475 dst_reg->smax_value = S64_MAX;
5476 } else {
5477 dst_reg->smin_value = smin_ptr + smin_val;
5478 dst_reg->smax_value = smax_ptr + smax_val;
5479 }
5480 if (umin_ptr + umin_val < umin_ptr ||
5481 umax_ptr + umax_val < umax_ptr) {
5482 dst_reg->umin_value = 0;
5483 dst_reg->umax_value = U64_MAX;
5484 } else {
5485 dst_reg->umin_value = umin_ptr + umin_val;
5486 dst_reg->umax_value = umax_ptr + umax_val;
5487 }
5488 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
5489 dst_reg->off = ptr_reg->off;
5490 dst_reg->raw = ptr_reg->raw;
5491 if (reg_is_pkt_pointer(ptr_reg)) {
5492 dst_reg->id = ++env->id_gen;
5493 /* something was added to pkt_ptr, set range to zero */
5494 dst_reg->raw = 0;
5495 }
5496 break;
5497 case BPF_SUB:
5498 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
5499 if (ret < 0) {
5500 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
5501 return ret;
5502 }
5503 if (dst_reg == off_reg) {
5504 /* scalar -= pointer. Creates an unknown scalar */
5505 verbose(env, "R%d tried to subtract pointer from scalar\n",
5506 dst);
5507 return -EACCES;
5508 }
5509 /* We don't allow subtraction from FP, because (according to
5510 * test_verifier.c test "invalid fp arithmetic", JITs might not
5511 * be able to deal with it.
5512 */
5513 if (ptr_reg->type == PTR_TO_STACK) {
5514 verbose(env, "R%d subtraction from stack pointer prohibited\n",
5515 dst);
5516 return -EACCES;
5517 }
5518 if (known && (ptr_reg->off - smin_val ==
5519 (s64)(s32)(ptr_reg->off - smin_val))) {
5520 /* pointer -= K. Subtract it from fixed offset */
5521 dst_reg->smin_value = smin_ptr;
5522 dst_reg->smax_value = smax_ptr;
5523 dst_reg->umin_value = umin_ptr;
5524 dst_reg->umax_value = umax_ptr;
5525 dst_reg->var_off = ptr_reg->var_off;
5526 dst_reg->id = ptr_reg->id;
5527 dst_reg->off = ptr_reg->off - smin_val;
5528 dst_reg->raw = ptr_reg->raw;
5529 break;
5530 }
5531 /* A new variable offset is created. If the subtrahend is known
5532 * nonnegative, then any reg->range we had before is still good.
5533 */
5534 if (signed_sub_overflows(smin_ptr, smax_val) ||
5535 signed_sub_overflows(smax_ptr, smin_val)) {
5536 /* Overflow possible, we know nothing */
5537 dst_reg->smin_value = S64_MIN;
5538 dst_reg->smax_value = S64_MAX;
5539 } else {
5540 dst_reg->smin_value = smin_ptr - smax_val;
5541 dst_reg->smax_value = smax_ptr - smin_val;
5542 }
5543 if (umin_ptr < umax_val) {
5544 /* Overflow possible, we know nothing */
5545 dst_reg->umin_value = 0;
5546 dst_reg->umax_value = U64_MAX;
5547 } else {
5548 /* Cannot overflow (as long as bounds are consistent) */
5549 dst_reg->umin_value = umin_ptr - umax_val;
5550 dst_reg->umax_value = umax_ptr - umin_val;
5551 }
5552 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
5553 dst_reg->off = ptr_reg->off;
5554 dst_reg->raw = ptr_reg->raw;
5555 if (reg_is_pkt_pointer(ptr_reg)) {
5556 dst_reg->id = ++env->id_gen;
5557 /* something was added to pkt_ptr, set range to zero */
5558 if (smin_val < 0)
5559 dst_reg->raw = 0;
5560 }
5561 break;
5562 case BPF_AND:
5563 case BPF_OR:
5564 case BPF_XOR:
5565 /* bitwise ops on pointers are troublesome, prohibit. */
5566 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
5567 dst, bpf_alu_string[opcode >> 4]);
5568 return -EACCES;
5569 default:
5570 /* other operators (e.g. MUL,LSH) produce non-pointer results */
5571 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
5572 dst, bpf_alu_string[opcode >> 4]);
5573 return -EACCES;
5574 }
5575
5576 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
5577 return -EINVAL;
5578
5579 __update_reg_bounds(dst_reg);
5580 __reg_deduce_bounds(dst_reg);
5581 __reg_bound_offset(dst_reg);
5582
5583 /* For unprivileged we require that resulting offset must be in bounds
5584 * in order to be able to sanitize access later on.
5585 */
5586 if (!env->bypass_spec_v1) {
5587 if (dst_reg->type == PTR_TO_MAP_VALUE &&
5588 check_map_access(env, dst, dst_reg->off, 1, false)) {
5589 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
5590 "prohibited for !root\n", dst);
5591 return -EACCES;
5592 } else if (dst_reg->type == PTR_TO_STACK &&
5593 check_stack_access(env, dst_reg, dst_reg->off +
5594 dst_reg->var_off.value, 1)) {
5595 verbose(env, "R%d stack pointer arithmetic goes out of range, "
5596 "prohibited for !root\n", dst);
5597 return -EACCES;
5598 }
5599 }
5600
5601 return 0;
5602 }
5603
5604 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
5605 struct bpf_reg_state *src_reg)
5606 {
5607 s32 smin_val = src_reg->s32_min_value;
5608 s32 smax_val = src_reg->s32_max_value;
5609 u32 umin_val = src_reg->u32_min_value;
5610 u32 umax_val = src_reg->u32_max_value;
5611
5612 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
5613 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
5614 dst_reg->s32_min_value = S32_MIN;
5615 dst_reg->s32_max_value = S32_MAX;
5616 } else {
5617 dst_reg->s32_min_value += smin_val;
5618 dst_reg->s32_max_value += smax_val;
5619 }
5620 if (dst_reg->u32_min_value + umin_val < umin_val ||
5621 dst_reg->u32_max_value + umax_val < umax_val) {
5622 dst_reg->u32_min_value = 0;
5623 dst_reg->u32_max_value = U32_MAX;
5624 } else {
5625 dst_reg->u32_min_value += umin_val;
5626 dst_reg->u32_max_value += umax_val;
5627 }
5628 }
5629
5630 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
5631 struct bpf_reg_state *src_reg)
5632 {
5633 s64 smin_val = src_reg->smin_value;
5634 s64 smax_val = src_reg->smax_value;
5635 u64 umin_val = src_reg->umin_value;
5636 u64 umax_val = src_reg->umax_value;
5637
5638 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
5639 signed_add_overflows(dst_reg->smax_value, smax_val)) {
5640 dst_reg->smin_value = S64_MIN;
5641 dst_reg->smax_value = S64_MAX;
5642 } else {
5643 dst_reg->smin_value += smin_val;
5644 dst_reg->smax_value += smax_val;
5645 }
5646 if (dst_reg->umin_value + umin_val < umin_val ||
5647 dst_reg->umax_value + umax_val < umax_val) {
5648 dst_reg->umin_value = 0;
5649 dst_reg->umax_value = U64_MAX;
5650 } else {
5651 dst_reg->umin_value += umin_val;
5652 dst_reg->umax_value += umax_val;
5653 }
5654 }
5655
5656 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
5657 struct bpf_reg_state *src_reg)
5658 {
5659 s32 smin_val = src_reg->s32_min_value;
5660 s32 smax_val = src_reg->s32_max_value;
5661 u32 umin_val = src_reg->u32_min_value;
5662 u32 umax_val = src_reg->u32_max_value;
5663
5664 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
5665 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
5666 /* Overflow possible, we know nothing */
5667 dst_reg->s32_min_value = S32_MIN;
5668 dst_reg->s32_max_value = S32_MAX;
5669 } else {
5670 dst_reg->s32_min_value -= smax_val;
5671 dst_reg->s32_max_value -= smin_val;
5672 }
5673 if (dst_reg->u32_min_value < umax_val) {
5674 /* Overflow possible, we know nothing */
5675 dst_reg->u32_min_value = 0;
5676 dst_reg->u32_max_value = U32_MAX;
5677 } else {
5678 /* Cannot overflow (as long as bounds are consistent) */
5679 dst_reg->u32_min_value -= umax_val;
5680 dst_reg->u32_max_value -= umin_val;
5681 }
5682 }
5683
5684 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
5685 struct bpf_reg_state *src_reg)
5686 {
5687 s64 smin_val = src_reg->smin_value;
5688 s64 smax_val = src_reg->smax_value;
5689 u64 umin_val = src_reg->umin_value;
5690 u64 umax_val = src_reg->umax_value;
5691
5692 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
5693 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
5694 /* Overflow possible, we know nothing */
5695 dst_reg->smin_value = S64_MIN;
5696 dst_reg->smax_value = S64_MAX;
5697 } else {
5698 dst_reg->smin_value -= smax_val;
5699 dst_reg->smax_value -= smin_val;
5700 }
5701 if (dst_reg->umin_value < umax_val) {
5702 /* Overflow possible, we know nothing */
5703 dst_reg->umin_value = 0;
5704 dst_reg->umax_value = U64_MAX;
5705 } else {
5706 /* Cannot overflow (as long as bounds are consistent) */
5707 dst_reg->umin_value -= umax_val;
5708 dst_reg->umax_value -= umin_val;
5709 }
5710 }
5711
5712 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
5713 struct bpf_reg_state *src_reg)
5714 {
5715 s32 smin_val = src_reg->s32_min_value;
5716 u32 umin_val = src_reg->u32_min_value;
5717 u32 umax_val = src_reg->u32_max_value;
5718
5719 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
5720 /* Ain't nobody got time to multiply that sign */
5721 __mark_reg32_unbounded(dst_reg);
5722 return;
5723 }
5724 /* Both values are positive, so we can work with unsigned and
5725 * copy the result to signed (unless it exceeds S32_MAX).
5726 */
5727 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
5728 /* Potential overflow, we know nothing */
5729 __mark_reg32_unbounded(dst_reg);
5730 return;
5731 }
5732 dst_reg->u32_min_value *= umin_val;
5733 dst_reg->u32_max_value *= umax_val;
5734 if (dst_reg->u32_max_value > S32_MAX) {
5735 /* Overflow possible, we know nothing */
5736 dst_reg->s32_min_value = S32_MIN;
5737 dst_reg->s32_max_value = S32_MAX;
5738 } else {
5739 dst_reg->s32_min_value = dst_reg->u32_min_value;
5740 dst_reg->s32_max_value = dst_reg->u32_max_value;
5741 }
5742 }
5743
5744 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
5745 struct bpf_reg_state *src_reg)
5746 {
5747 s64 smin_val = src_reg->smin_value;
5748 u64 umin_val = src_reg->umin_value;
5749 u64 umax_val = src_reg->umax_value;
5750
5751 if (smin_val < 0 || dst_reg->smin_value < 0) {
5752 /* Ain't nobody got time to multiply that sign */
5753 __mark_reg64_unbounded(dst_reg);
5754 return;
5755 }
5756 /* Both values are positive, so we can work with unsigned and
5757 * copy the result to signed (unless it exceeds S64_MAX).
5758 */
5759 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
5760 /* Potential overflow, we know nothing */
5761 __mark_reg64_unbounded(dst_reg);
5762 return;
5763 }
5764 dst_reg->umin_value *= umin_val;
5765 dst_reg->umax_value *= umax_val;
5766 if (dst_reg->umax_value > S64_MAX) {
5767 /* Overflow possible, we know nothing */
5768 dst_reg->smin_value = S64_MIN;
5769 dst_reg->smax_value = S64_MAX;
5770 } else {
5771 dst_reg->smin_value = dst_reg->umin_value;
5772 dst_reg->smax_value = dst_reg->umax_value;
5773 }
5774 }
5775
5776 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
5777 struct bpf_reg_state *src_reg)
5778 {
5779 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5780 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5781 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5782 s32 smin_val = src_reg->s32_min_value;
5783 u32 umax_val = src_reg->u32_max_value;
5784
5785 /* Assuming scalar64_min_max_and will be called so its safe
5786 * to skip updating register for known 32-bit case.
5787 */
5788 if (src_known && dst_known)
5789 return;
5790
5791 /* We get our minimum from the var_off, since that's inherently
5792 * bitwise. Our maximum is the minimum of the operands' maxima.
5793 */
5794 dst_reg->u32_min_value = var32_off.value;
5795 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
5796 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5797 /* Lose signed bounds when ANDing negative numbers,
5798 * ain't nobody got time for that.
5799 */
5800 dst_reg->s32_min_value = S32_MIN;
5801 dst_reg->s32_max_value = S32_MAX;
5802 } else {
5803 /* ANDing two positives gives a positive, so safe to
5804 * cast result into s64.
5805 */
5806 dst_reg->s32_min_value = dst_reg->u32_min_value;
5807 dst_reg->s32_max_value = dst_reg->u32_max_value;
5808 }
5809
5810 }
5811
5812 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
5813 struct bpf_reg_state *src_reg)
5814 {
5815 bool src_known = tnum_is_const(src_reg->var_off);
5816 bool dst_known = tnum_is_const(dst_reg->var_off);
5817 s64 smin_val = src_reg->smin_value;
5818 u64 umax_val = src_reg->umax_value;
5819
5820 if (src_known && dst_known) {
5821 __mark_reg_known(dst_reg, dst_reg->var_off.value &
5822 src_reg->var_off.value);
5823 return;
5824 }
5825
5826 /* We get our minimum from the var_off, since that's inherently
5827 * bitwise. Our maximum is the minimum of the operands' maxima.
5828 */
5829 dst_reg->umin_value = dst_reg->var_off.value;
5830 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
5831 if (dst_reg->smin_value < 0 || smin_val < 0) {
5832 /* Lose signed bounds when ANDing negative numbers,
5833 * ain't nobody got time for that.
5834 */
5835 dst_reg->smin_value = S64_MIN;
5836 dst_reg->smax_value = S64_MAX;
5837 } else {
5838 /* ANDing two positives gives a positive, so safe to
5839 * cast result into s64.
5840 */
5841 dst_reg->smin_value = dst_reg->umin_value;
5842 dst_reg->smax_value = dst_reg->umax_value;
5843 }
5844 /* We may learn something more from the var_off */
5845 __update_reg_bounds(dst_reg);
5846 }
5847
5848 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
5849 struct bpf_reg_state *src_reg)
5850 {
5851 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5852 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5853 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5854 s32 smin_val = src_reg->smin_value;
5855 u32 umin_val = src_reg->umin_value;
5856
5857 /* Assuming scalar64_min_max_or will be called so it is safe
5858 * to skip updating register for known case.
5859 */
5860 if (src_known && dst_known)
5861 return;
5862
5863 /* We get our maximum from the var_off, and our minimum is the
5864 * maximum of the operands' minima
5865 */
5866 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
5867 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
5868 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5869 /* Lose signed bounds when ORing negative numbers,
5870 * ain't nobody got time for that.
5871 */
5872 dst_reg->s32_min_value = S32_MIN;
5873 dst_reg->s32_max_value = S32_MAX;
5874 } else {
5875 /* ORing two positives gives a positive, so safe to
5876 * cast result into s64.
5877 */
5878 dst_reg->s32_min_value = dst_reg->umin_value;
5879 dst_reg->s32_max_value = dst_reg->umax_value;
5880 }
5881 }
5882
5883 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
5884 struct bpf_reg_state *src_reg)
5885 {
5886 bool src_known = tnum_is_const(src_reg->var_off);
5887 bool dst_known = tnum_is_const(dst_reg->var_off);
5888 s64 smin_val = src_reg->smin_value;
5889 u64 umin_val = src_reg->umin_value;
5890
5891 if (src_known && dst_known) {
5892 __mark_reg_known(dst_reg, dst_reg->var_off.value |
5893 src_reg->var_off.value);
5894 return;
5895 }
5896
5897 /* We get our maximum from the var_off, and our minimum is the
5898 * maximum of the operands' minima
5899 */
5900 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
5901 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
5902 if (dst_reg->smin_value < 0 || smin_val < 0) {
5903 /* Lose signed bounds when ORing negative numbers,
5904 * ain't nobody got time for that.
5905 */
5906 dst_reg->smin_value = S64_MIN;
5907 dst_reg->smax_value = S64_MAX;
5908 } else {
5909 /* ORing two positives gives a positive, so safe to
5910 * cast result into s64.
5911 */
5912 dst_reg->smin_value = dst_reg->umin_value;
5913 dst_reg->smax_value = dst_reg->umax_value;
5914 }
5915 /* We may learn something more from the var_off */
5916 __update_reg_bounds(dst_reg);
5917 }
5918
5919 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
5920 struct bpf_reg_state *src_reg)
5921 {
5922 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5923 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5924 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5925 s32 smin_val = src_reg->s32_min_value;
5926
5927 /* Assuming scalar64_min_max_xor will be called so it is safe
5928 * to skip updating register for known case.
5929 */
5930 if (src_known && dst_known)
5931 return;
5932
5933 /* We get both minimum and maximum from the var32_off. */
5934 dst_reg->u32_min_value = var32_off.value;
5935 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
5936
5937 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
5938 /* XORing two positive sign numbers gives a positive,
5939 * so safe to cast u32 result into s32.
5940 */
5941 dst_reg->s32_min_value = dst_reg->u32_min_value;
5942 dst_reg->s32_max_value = dst_reg->u32_max_value;
5943 } else {
5944 dst_reg->s32_min_value = S32_MIN;
5945 dst_reg->s32_max_value = S32_MAX;
5946 }
5947 }
5948
5949 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
5950 struct bpf_reg_state *src_reg)
5951 {
5952 bool src_known = tnum_is_const(src_reg->var_off);
5953 bool dst_known = tnum_is_const(dst_reg->var_off);
5954 s64 smin_val = src_reg->smin_value;
5955
5956 if (src_known && dst_known) {
5957 /* dst_reg->var_off.value has been updated earlier */
5958 __mark_reg_known(dst_reg, dst_reg->var_off.value);
5959 return;
5960 }
5961
5962 /* We get both minimum and maximum from the var_off. */
5963 dst_reg->umin_value = dst_reg->var_off.value;
5964 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
5965
5966 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
5967 /* XORing two positive sign numbers gives a positive,
5968 * so safe to cast u64 result into s64.
5969 */
5970 dst_reg->smin_value = dst_reg->umin_value;
5971 dst_reg->smax_value = dst_reg->umax_value;
5972 } else {
5973 dst_reg->smin_value = S64_MIN;
5974 dst_reg->smax_value = S64_MAX;
5975 }
5976
5977 __update_reg_bounds(dst_reg);
5978 }
5979
5980 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5981 u64 umin_val, u64 umax_val)
5982 {
5983 /* We lose all sign bit information (except what we can pick
5984 * up from var_off)
5985 */
5986 dst_reg->s32_min_value = S32_MIN;
5987 dst_reg->s32_max_value = S32_MAX;
5988 /* If we might shift our top bit out, then we know nothing */
5989 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
5990 dst_reg->u32_min_value = 0;
5991 dst_reg->u32_max_value = U32_MAX;
5992 } else {
5993 dst_reg->u32_min_value <<= umin_val;
5994 dst_reg->u32_max_value <<= umax_val;
5995 }
5996 }
5997
5998 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
5999 struct bpf_reg_state *src_reg)
6000 {
6001 u32 umax_val = src_reg->u32_max_value;
6002 u32 umin_val = src_reg->u32_min_value;
6003 /* u32 alu operation will zext upper bits */
6004 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6005
6006 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6007 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
6008 /* Not required but being careful mark reg64 bounds as unknown so
6009 * that we are forced to pick them up from tnum and zext later and
6010 * if some path skips this step we are still safe.
6011 */
6012 __mark_reg64_unbounded(dst_reg);
6013 __update_reg32_bounds(dst_reg);
6014 }
6015
6016 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
6017 u64 umin_val, u64 umax_val)
6018 {
6019 /* Special case <<32 because it is a common compiler pattern to sign
6020 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
6021 * positive we know this shift will also be positive so we can track
6022 * bounds correctly. Otherwise we lose all sign bit information except
6023 * what we can pick up from var_off. Perhaps we can generalize this
6024 * later to shifts of any length.
6025 */
6026 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
6027 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
6028 else
6029 dst_reg->smax_value = S64_MAX;
6030
6031 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
6032 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
6033 else
6034 dst_reg->smin_value = S64_MIN;
6035
6036 /* If we might shift our top bit out, then we know nothing */
6037 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
6038 dst_reg->umin_value = 0;
6039 dst_reg->umax_value = U64_MAX;
6040 } else {
6041 dst_reg->umin_value <<= umin_val;
6042 dst_reg->umax_value <<= umax_val;
6043 }
6044 }
6045
6046 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
6047 struct bpf_reg_state *src_reg)
6048 {
6049 u64 umax_val = src_reg->umax_value;
6050 u64 umin_val = src_reg->umin_value;
6051
6052 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
6053 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
6054 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6055
6056 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
6057 /* We may learn something more from the var_off */
6058 __update_reg_bounds(dst_reg);
6059 }
6060
6061 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
6062 struct bpf_reg_state *src_reg)
6063 {
6064 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6065 u32 umax_val = src_reg->u32_max_value;
6066 u32 umin_val = src_reg->u32_min_value;
6067
6068 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6069 * be negative, then either:
6070 * 1) src_reg might be zero, so the sign bit of the result is
6071 * unknown, so we lose our signed bounds
6072 * 2) it's known negative, thus the unsigned bounds capture the
6073 * signed bounds
6074 * 3) the signed bounds cross zero, so they tell us nothing
6075 * about the result
6076 * If the value in dst_reg is known nonnegative, then again the
6077 * unsigned bounts capture the signed bounds.
6078 * Thus, in all cases it suffices to blow away our signed bounds
6079 * and rely on inferring new ones from the unsigned bounds and
6080 * var_off of the result.
6081 */
6082 dst_reg->s32_min_value = S32_MIN;
6083 dst_reg->s32_max_value = S32_MAX;
6084
6085 dst_reg->var_off = tnum_rshift(subreg, umin_val);
6086 dst_reg->u32_min_value >>= umax_val;
6087 dst_reg->u32_max_value >>= umin_val;
6088
6089 __mark_reg64_unbounded(dst_reg);
6090 __update_reg32_bounds(dst_reg);
6091 }
6092
6093 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
6094 struct bpf_reg_state *src_reg)
6095 {
6096 u64 umax_val = src_reg->umax_value;
6097 u64 umin_val = src_reg->umin_value;
6098
6099 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6100 * be negative, then either:
6101 * 1) src_reg might be zero, so the sign bit of the result is
6102 * unknown, so we lose our signed bounds
6103 * 2) it's known negative, thus the unsigned bounds capture the
6104 * signed bounds
6105 * 3) the signed bounds cross zero, so they tell us nothing
6106 * about the result
6107 * If the value in dst_reg is known nonnegative, then again the
6108 * unsigned bounts capture the signed bounds.
6109 * Thus, in all cases it suffices to blow away our signed bounds
6110 * and rely on inferring new ones from the unsigned bounds and
6111 * var_off of the result.
6112 */
6113 dst_reg->smin_value = S64_MIN;
6114 dst_reg->smax_value = S64_MAX;
6115 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
6116 dst_reg->umin_value >>= umax_val;
6117 dst_reg->umax_value >>= umin_val;
6118
6119 /* Its not easy to operate on alu32 bounds here because it depends
6120 * on bits being shifted in. Take easy way out and mark unbounded
6121 * so we can recalculate later from tnum.
6122 */
6123 __mark_reg32_unbounded(dst_reg);
6124 __update_reg_bounds(dst_reg);
6125 }
6126
6127 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
6128 struct bpf_reg_state *src_reg)
6129 {
6130 u64 umin_val = src_reg->u32_min_value;
6131
6132 /* Upon reaching here, src_known is true and
6133 * umax_val is equal to umin_val.
6134 */
6135 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
6136 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
6137
6138 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
6139
6140 /* blow away the dst_reg umin_value/umax_value and rely on
6141 * dst_reg var_off to refine the result.
6142 */
6143 dst_reg->u32_min_value = 0;
6144 dst_reg->u32_max_value = U32_MAX;
6145
6146 __mark_reg64_unbounded(dst_reg);
6147 __update_reg32_bounds(dst_reg);
6148 }
6149
6150 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
6151 struct bpf_reg_state *src_reg)
6152 {
6153 u64 umin_val = src_reg->umin_value;
6154
6155 /* Upon reaching here, src_known is true and umax_val is equal
6156 * to umin_val.
6157 */
6158 dst_reg->smin_value >>= umin_val;
6159 dst_reg->smax_value >>= umin_val;
6160
6161 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
6162
6163 /* blow away the dst_reg umin_value/umax_value and rely on
6164 * dst_reg var_off to refine the result.
6165 */
6166 dst_reg->umin_value = 0;
6167 dst_reg->umax_value = U64_MAX;
6168
6169 /* Its not easy to operate on alu32 bounds here because it depends
6170 * on bits being shifted in from upper 32-bits. Take easy way out
6171 * and mark unbounded so we can recalculate later from tnum.
6172 */
6173 __mark_reg32_unbounded(dst_reg);
6174 __update_reg_bounds(dst_reg);
6175 }
6176
6177 /* WARNING: This function does calculations on 64-bit values, but the actual
6178 * execution may occur on 32-bit values. Therefore, things like bitshifts
6179 * need extra checks in the 32-bit case.
6180 */
6181 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
6182 struct bpf_insn *insn,
6183 struct bpf_reg_state *dst_reg,
6184 struct bpf_reg_state src_reg)
6185 {
6186 struct bpf_reg_state *regs = cur_regs(env);
6187 u8 opcode = BPF_OP(insn->code);
6188 bool src_known;
6189 s64 smin_val, smax_val;
6190 u64 umin_val, umax_val;
6191 s32 s32_min_val, s32_max_val;
6192 u32 u32_min_val, u32_max_val;
6193 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
6194 u32 dst = insn->dst_reg;
6195 int ret;
6196 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
6197
6198 smin_val = src_reg.smin_value;
6199 smax_val = src_reg.smax_value;
6200 umin_val = src_reg.umin_value;
6201 umax_val = src_reg.umax_value;
6202
6203 s32_min_val = src_reg.s32_min_value;
6204 s32_max_val = src_reg.s32_max_value;
6205 u32_min_val = src_reg.u32_min_value;
6206 u32_max_val = src_reg.u32_max_value;
6207
6208 if (alu32) {
6209 src_known = tnum_subreg_is_const(src_reg.var_off);
6210 if ((src_known &&
6211 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
6212 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
6213 /* Taint dst register if offset had invalid bounds
6214 * derived from e.g. dead branches.
6215 */
6216 __mark_reg_unknown(env, dst_reg);
6217 return 0;
6218 }
6219 } else {
6220 src_known = tnum_is_const(src_reg.var_off);
6221 if ((src_known &&
6222 (smin_val != smax_val || umin_val != umax_val)) ||
6223 smin_val > smax_val || umin_val > umax_val) {
6224 /* Taint dst register if offset had invalid bounds
6225 * derived from e.g. dead branches.
6226 */
6227 __mark_reg_unknown(env, dst_reg);
6228 return 0;
6229 }
6230 }
6231
6232 if (!src_known &&
6233 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
6234 __mark_reg_unknown(env, dst_reg);
6235 return 0;
6236 }
6237
6238 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
6239 * There are two classes of instructions: The first class we track both
6240 * alu32 and alu64 sign/unsigned bounds independently this provides the
6241 * greatest amount of precision when alu operations are mixed with jmp32
6242 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
6243 * and BPF_OR. This is possible because these ops have fairly easy to
6244 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
6245 * See alu32 verifier tests for examples. The second class of
6246 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
6247 * with regards to tracking sign/unsigned bounds because the bits may
6248 * cross subreg boundaries in the alu64 case. When this happens we mark
6249 * the reg unbounded in the subreg bound space and use the resulting
6250 * tnum to calculate an approximation of the sign/unsigned bounds.
6251 */
6252 switch (opcode) {
6253 case BPF_ADD:
6254 ret = sanitize_val_alu(env, insn);
6255 if (ret < 0) {
6256 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
6257 return ret;
6258 }
6259 scalar32_min_max_add(dst_reg, &src_reg);
6260 scalar_min_max_add(dst_reg, &src_reg);
6261 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
6262 break;
6263 case BPF_SUB:
6264 ret = sanitize_val_alu(env, insn);
6265 if (ret < 0) {
6266 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
6267 return ret;
6268 }
6269 scalar32_min_max_sub(dst_reg, &src_reg);
6270 scalar_min_max_sub(dst_reg, &src_reg);
6271 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
6272 break;
6273 case BPF_MUL:
6274 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
6275 scalar32_min_max_mul(dst_reg, &src_reg);
6276 scalar_min_max_mul(dst_reg, &src_reg);
6277 break;
6278 case BPF_AND:
6279 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
6280 scalar32_min_max_and(dst_reg, &src_reg);
6281 scalar_min_max_and(dst_reg, &src_reg);
6282 break;
6283 case BPF_OR:
6284 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
6285 scalar32_min_max_or(dst_reg, &src_reg);
6286 scalar_min_max_or(dst_reg, &src_reg);
6287 break;
6288 case BPF_XOR:
6289 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
6290 scalar32_min_max_xor(dst_reg, &src_reg);
6291 scalar_min_max_xor(dst_reg, &src_reg);
6292 break;
6293 case BPF_LSH:
6294 if (umax_val >= insn_bitness) {
6295 /* Shifts greater than 31 or 63 are undefined.
6296 * This includes shifts by a negative number.
6297 */
6298 mark_reg_unknown(env, regs, insn->dst_reg);
6299 break;
6300 }
6301 if (alu32)
6302 scalar32_min_max_lsh(dst_reg, &src_reg);
6303 else
6304 scalar_min_max_lsh(dst_reg, &src_reg);
6305 break;
6306 case BPF_RSH:
6307 if (umax_val >= insn_bitness) {
6308 /* Shifts greater than 31 or 63 are undefined.
6309 * This includes shifts by a negative number.
6310 */
6311 mark_reg_unknown(env, regs, insn->dst_reg);
6312 break;
6313 }
6314 if (alu32)
6315 scalar32_min_max_rsh(dst_reg, &src_reg);
6316 else
6317 scalar_min_max_rsh(dst_reg, &src_reg);
6318 break;
6319 case BPF_ARSH:
6320 if (umax_val >= insn_bitness) {
6321 /* Shifts greater than 31 or 63 are undefined.
6322 * This includes shifts by a negative number.
6323 */
6324 mark_reg_unknown(env, regs, insn->dst_reg);
6325 break;
6326 }
6327 if (alu32)
6328 scalar32_min_max_arsh(dst_reg, &src_reg);
6329 else
6330 scalar_min_max_arsh(dst_reg, &src_reg);
6331 break;
6332 default:
6333 mark_reg_unknown(env, regs, insn->dst_reg);
6334 break;
6335 }
6336
6337 /* ALU32 ops are zero extended into 64bit register */
6338 if (alu32)
6339 zext_32_to_64(dst_reg);
6340
6341 __update_reg_bounds(dst_reg);
6342 __reg_deduce_bounds(dst_reg);
6343 __reg_bound_offset(dst_reg);
6344 return 0;
6345 }
6346
6347 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
6348 * and var_off.
6349 */
6350 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
6351 struct bpf_insn *insn)
6352 {
6353 struct bpf_verifier_state *vstate = env->cur_state;
6354 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6355 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
6356 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
6357 u8 opcode = BPF_OP(insn->code);
6358 int err;
6359
6360 dst_reg = &regs[insn->dst_reg];
6361 src_reg = NULL;
6362 if (dst_reg->type != SCALAR_VALUE)
6363 ptr_reg = dst_reg;
6364 if (BPF_SRC(insn->code) == BPF_X) {
6365 src_reg = &regs[insn->src_reg];
6366 if (src_reg->type != SCALAR_VALUE) {
6367 if (dst_reg->type != SCALAR_VALUE) {
6368 /* Combining two pointers by any ALU op yields
6369 * an arbitrary scalar. Disallow all math except
6370 * pointer subtraction
6371 */
6372 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6373 mark_reg_unknown(env, regs, insn->dst_reg);
6374 return 0;
6375 }
6376 verbose(env, "R%d pointer %s pointer prohibited\n",
6377 insn->dst_reg,
6378 bpf_alu_string[opcode >> 4]);
6379 return -EACCES;
6380 } else {
6381 /* scalar += pointer
6382 * This is legal, but we have to reverse our
6383 * src/dest handling in computing the range
6384 */
6385 err = mark_chain_precision(env, insn->dst_reg);
6386 if (err)
6387 return err;
6388 return adjust_ptr_min_max_vals(env, insn,
6389 src_reg, dst_reg);
6390 }
6391 } else if (ptr_reg) {
6392 /* pointer += scalar */
6393 err = mark_chain_precision(env, insn->src_reg);
6394 if (err)
6395 return err;
6396 return adjust_ptr_min_max_vals(env, insn,
6397 dst_reg, src_reg);
6398 }
6399 } else {
6400 /* Pretend the src is a reg with a known value, since we only
6401 * need to be able to read from this state.
6402 */
6403 off_reg.type = SCALAR_VALUE;
6404 __mark_reg_known(&off_reg, insn->imm);
6405 src_reg = &off_reg;
6406 if (ptr_reg) /* pointer += K */
6407 return adjust_ptr_min_max_vals(env, insn,
6408 ptr_reg, src_reg);
6409 }
6410
6411 /* Got here implies adding two SCALAR_VALUEs */
6412 if (WARN_ON_ONCE(ptr_reg)) {
6413 print_verifier_state(env, state);
6414 verbose(env, "verifier internal error: unexpected ptr_reg\n");
6415 return -EINVAL;
6416 }
6417 if (WARN_ON(!src_reg)) {
6418 print_verifier_state(env, state);
6419 verbose(env, "verifier internal error: no src_reg\n");
6420 return -EINVAL;
6421 }
6422 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
6423 }
6424
6425 /* check validity of 32-bit and 64-bit arithmetic operations */
6426 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
6427 {
6428 struct bpf_reg_state *regs = cur_regs(env);
6429 u8 opcode = BPF_OP(insn->code);
6430 int err;
6431
6432 if (opcode == BPF_END || opcode == BPF_NEG) {
6433 if (opcode == BPF_NEG) {
6434 if (BPF_SRC(insn->code) != 0 ||
6435 insn->src_reg != BPF_REG_0 ||
6436 insn->off != 0 || insn->imm != 0) {
6437 verbose(env, "BPF_NEG uses reserved fields\n");
6438 return -EINVAL;
6439 }
6440 } else {
6441 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
6442 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
6443 BPF_CLASS(insn->code) == BPF_ALU64) {
6444 verbose(env, "BPF_END uses reserved fields\n");
6445 return -EINVAL;
6446 }
6447 }
6448
6449 /* check src operand */
6450 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6451 if (err)
6452 return err;
6453
6454 if (is_pointer_value(env, insn->dst_reg)) {
6455 verbose(env, "R%d pointer arithmetic prohibited\n",
6456 insn->dst_reg);
6457 return -EACCES;
6458 }
6459
6460 /* check dest operand */
6461 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6462 if (err)
6463 return err;
6464
6465 } else if (opcode == BPF_MOV) {
6466
6467 if (BPF_SRC(insn->code) == BPF_X) {
6468 if (insn->imm != 0 || insn->off != 0) {
6469 verbose(env, "BPF_MOV uses reserved fields\n");
6470 return -EINVAL;
6471 }
6472
6473 /* check src operand */
6474 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6475 if (err)
6476 return err;
6477 } else {
6478 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6479 verbose(env, "BPF_MOV uses reserved fields\n");
6480 return -EINVAL;
6481 }
6482 }
6483
6484 /* check dest operand, mark as required later */
6485 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6486 if (err)
6487 return err;
6488
6489 if (BPF_SRC(insn->code) == BPF_X) {
6490 struct bpf_reg_state *src_reg = regs + insn->src_reg;
6491 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
6492
6493 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6494 /* case: R1 = R2
6495 * copy register state to dest reg
6496 */
6497 *dst_reg = *src_reg;
6498 dst_reg->live |= REG_LIVE_WRITTEN;
6499 dst_reg->subreg_def = DEF_NOT_SUBREG;
6500 } else {
6501 /* R1 = (u32) R2 */
6502 if (is_pointer_value(env, insn->src_reg)) {
6503 verbose(env,
6504 "R%d partial copy of pointer\n",
6505 insn->src_reg);
6506 return -EACCES;
6507 } else if (src_reg->type == SCALAR_VALUE) {
6508 *dst_reg = *src_reg;
6509 dst_reg->live |= REG_LIVE_WRITTEN;
6510 dst_reg->subreg_def = env->insn_idx + 1;
6511 } else {
6512 mark_reg_unknown(env, regs,
6513 insn->dst_reg);
6514 }
6515 zext_32_to_64(dst_reg);
6516 }
6517 } else {
6518 /* case: R = imm
6519 * remember the value we stored into this reg
6520 */
6521 /* clear any state __mark_reg_known doesn't set */
6522 mark_reg_unknown(env, regs, insn->dst_reg);
6523 regs[insn->dst_reg].type = SCALAR_VALUE;
6524 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6525 __mark_reg_known(regs + insn->dst_reg,
6526 insn->imm);
6527 } else {
6528 __mark_reg_known(regs + insn->dst_reg,
6529 (u32)insn->imm);
6530 }
6531 }
6532
6533 } else if (opcode > BPF_END) {
6534 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
6535 return -EINVAL;
6536
6537 } else { /* all other ALU ops: and, sub, xor, add, ... */
6538
6539 if (BPF_SRC(insn->code) == BPF_X) {
6540 if (insn->imm != 0 || insn->off != 0) {
6541 verbose(env, "BPF_ALU uses reserved fields\n");
6542 return -EINVAL;
6543 }
6544 /* check src1 operand */
6545 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6546 if (err)
6547 return err;
6548 } else {
6549 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6550 verbose(env, "BPF_ALU uses reserved fields\n");
6551 return -EINVAL;
6552 }
6553 }
6554
6555 /* check src2 operand */
6556 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6557 if (err)
6558 return err;
6559
6560 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
6561 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
6562 verbose(env, "div by zero\n");
6563 return -EINVAL;
6564 }
6565
6566 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
6567 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
6568 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
6569
6570 if (insn->imm < 0 || insn->imm >= size) {
6571 verbose(env, "invalid shift %d\n", insn->imm);
6572 return -EINVAL;
6573 }
6574 }
6575
6576 /* check dest operand */
6577 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6578 if (err)
6579 return err;
6580
6581 return adjust_reg_min_max_vals(env, insn);
6582 }
6583
6584 return 0;
6585 }
6586
6587 static void __find_good_pkt_pointers(struct bpf_func_state *state,
6588 struct bpf_reg_state *dst_reg,
6589 enum bpf_reg_type type, u16 new_range)
6590 {
6591 struct bpf_reg_state *reg;
6592 int i;
6593
6594 for (i = 0; i < MAX_BPF_REG; i++) {
6595 reg = &state->regs[i];
6596 if (reg->type == type && reg->id == dst_reg->id)
6597 /* keep the maximum range already checked */
6598 reg->range = max(reg->range, new_range);
6599 }
6600
6601 bpf_for_each_spilled_reg(i, state, reg) {
6602 if (!reg)
6603 continue;
6604 if (reg->type == type && reg->id == dst_reg->id)
6605 reg->range = max(reg->range, new_range);
6606 }
6607 }
6608
6609 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
6610 struct bpf_reg_state *dst_reg,
6611 enum bpf_reg_type type,
6612 bool range_right_open)
6613 {
6614 u16 new_range;
6615 int i;
6616
6617 if (dst_reg->off < 0 ||
6618 (dst_reg->off == 0 && range_right_open))
6619 /* This doesn't give us any range */
6620 return;
6621
6622 if (dst_reg->umax_value > MAX_PACKET_OFF ||
6623 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
6624 /* Risk of overflow. For instance, ptr + (1<<63) may be less
6625 * than pkt_end, but that's because it's also less than pkt.
6626 */
6627 return;
6628
6629 new_range = dst_reg->off;
6630 if (range_right_open)
6631 new_range--;
6632
6633 /* Examples for register markings:
6634 *
6635 * pkt_data in dst register:
6636 *
6637 * r2 = r3;
6638 * r2 += 8;
6639 * if (r2 > pkt_end) goto <handle exception>
6640 * <access okay>
6641 *
6642 * r2 = r3;
6643 * r2 += 8;
6644 * if (r2 < pkt_end) goto <access okay>
6645 * <handle exception>
6646 *
6647 * Where:
6648 * r2 == dst_reg, pkt_end == src_reg
6649 * r2=pkt(id=n,off=8,r=0)
6650 * r3=pkt(id=n,off=0,r=0)
6651 *
6652 * pkt_data in src register:
6653 *
6654 * r2 = r3;
6655 * r2 += 8;
6656 * if (pkt_end >= r2) goto <access okay>
6657 * <handle exception>
6658 *
6659 * r2 = r3;
6660 * r2 += 8;
6661 * if (pkt_end <= r2) goto <handle exception>
6662 * <access okay>
6663 *
6664 * Where:
6665 * pkt_end == dst_reg, r2 == src_reg
6666 * r2=pkt(id=n,off=8,r=0)
6667 * r3=pkt(id=n,off=0,r=0)
6668 *
6669 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
6670 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
6671 * and [r3, r3 + 8-1) respectively is safe to access depending on
6672 * the check.
6673 */
6674
6675 /* If our ids match, then we must have the same max_value. And we
6676 * don't care about the other reg's fixed offset, since if it's too big
6677 * the range won't allow anything.
6678 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
6679 */
6680 for (i = 0; i <= vstate->curframe; i++)
6681 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
6682 new_range);
6683 }
6684
6685 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
6686 {
6687 struct tnum subreg = tnum_subreg(reg->var_off);
6688 s32 sval = (s32)val;
6689
6690 switch (opcode) {
6691 case BPF_JEQ:
6692 if (tnum_is_const(subreg))
6693 return !!tnum_equals_const(subreg, val);
6694 break;
6695 case BPF_JNE:
6696 if (tnum_is_const(subreg))
6697 return !tnum_equals_const(subreg, val);
6698 break;
6699 case BPF_JSET:
6700 if ((~subreg.mask & subreg.value) & val)
6701 return 1;
6702 if (!((subreg.mask | subreg.value) & val))
6703 return 0;
6704 break;
6705 case BPF_JGT:
6706 if (reg->u32_min_value > val)
6707 return 1;
6708 else if (reg->u32_max_value <= val)
6709 return 0;
6710 break;
6711 case BPF_JSGT:
6712 if (reg->s32_min_value > sval)
6713 return 1;
6714 else if (reg->s32_max_value < sval)
6715 return 0;
6716 break;
6717 case BPF_JLT:
6718 if (reg->u32_max_value < val)
6719 return 1;
6720 else if (reg->u32_min_value >= val)
6721 return 0;
6722 break;
6723 case BPF_JSLT:
6724 if (reg->s32_max_value < sval)
6725 return 1;
6726 else if (reg->s32_min_value >= sval)
6727 return 0;
6728 break;
6729 case BPF_JGE:
6730 if (reg->u32_min_value >= val)
6731 return 1;
6732 else if (reg->u32_max_value < val)
6733 return 0;
6734 break;
6735 case BPF_JSGE:
6736 if (reg->s32_min_value >= sval)
6737 return 1;
6738 else if (reg->s32_max_value < sval)
6739 return 0;
6740 break;
6741 case BPF_JLE:
6742 if (reg->u32_max_value <= val)
6743 return 1;
6744 else if (reg->u32_min_value > val)
6745 return 0;
6746 break;
6747 case BPF_JSLE:
6748 if (reg->s32_max_value <= sval)
6749 return 1;
6750 else if (reg->s32_min_value > sval)
6751 return 0;
6752 break;
6753 }
6754
6755 return -1;
6756 }
6757
6758
6759 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
6760 {
6761 s64 sval = (s64)val;
6762
6763 switch (opcode) {
6764 case BPF_JEQ:
6765 if (tnum_is_const(reg->var_off))
6766 return !!tnum_equals_const(reg->var_off, val);
6767 break;
6768 case BPF_JNE:
6769 if (tnum_is_const(reg->var_off))
6770 return !tnum_equals_const(reg->var_off, val);
6771 break;
6772 case BPF_JSET:
6773 if ((~reg->var_off.mask & reg->var_off.value) & val)
6774 return 1;
6775 if (!((reg->var_off.mask | reg->var_off.value) & val))
6776 return 0;
6777 break;
6778 case BPF_JGT:
6779 if (reg->umin_value > val)
6780 return 1;
6781 else if (reg->umax_value <= val)
6782 return 0;
6783 break;
6784 case BPF_JSGT:
6785 if (reg->smin_value > sval)
6786 return 1;
6787 else if (reg->smax_value < sval)
6788 return 0;
6789 break;
6790 case BPF_JLT:
6791 if (reg->umax_value < val)
6792 return 1;
6793 else if (reg->umin_value >= val)
6794 return 0;
6795 break;
6796 case BPF_JSLT:
6797 if (reg->smax_value < sval)
6798 return 1;
6799 else if (reg->smin_value >= sval)
6800 return 0;
6801 break;
6802 case BPF_JGE:
6803 if (reg->umin_value >= val)
6804 return 1;
6805 else if (reg->umax_value < val)
6806 return 0;
6807 break;
6808 case BPF_JSGE:
6809 if (reg->smin_value >= sval)
6810 return 1;
6811 else if (reg->smax_value < sval)
6812 return 0;
6813 break;
6814 case BPF_JLE:
6815 if (reg->umax_value <= val)
6816 return 1;
6817 else if (reg->umin_value > val)
6818 return 0;
6819 break;
6820 case BPF_JSLE:
6821 if (reg->smax_value <= sval)
6822 return 1;
6823 else if (reg->smin_value > sval)
6824 return 0;
6825 break;
6826 }
6827
6828 return -1;
6829 }
6830
6831 /* compute branch direction of the expression "if (reg opcode val) goto target;"
6832 * and return:
6833 * 1 - branch will be taken and "goto target" will be executed
6834 * 0 - branch will not be taken and fall-through to next insn
6835 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
6836 * range [0,10]
6837 */
6838 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
6839 bool is_jmp32)
6840 {
6841 if (__is_pointer_value(false, reg)) {
6842 if (!reg_type_not_null(reg->type))
6843 return -1;
6844
6845 /* If pointer is valid tests against zero will fail so we can
6846 * use this to direct branch taken.
6847 */
6848 if (val != 0)
6849 return -1;
6850
6851 switch (opcode) {
6852 case BPF_JEQ:
6853 return 0;
6854 case BPF_JNE:
6855 return 1;
6856 default:
6857 return -1;
6858 }
6859 }
6860
6861 if (is_jmp32)
6862 return is_branch32_taken(reg, val, opcode);
6863 return is_branch64_taken(reg, val, opcode);
6864 }
6865
6866 /* Adjusts the register min/max values in the case that the dst_reg is the
6867 * variable register that we are working on, and src_reg is a constant or we're
6868 * simply doing a BPF_K check.
6869 * In JEQ/JNE cases we also adjust the var_off values.
6870 */
6871 static void reg_set_min_max(struct bpf_reg_state *true_reg,
6872 struct bpf_reg_state *false_reg,
6873 u64 val, u32 val32,
6874 u8 opcode, bool is_jmp32)
6875 {
6876 struct tnum false_32off = tnum_subreg(false_reg->var_off);
6877 struct tnum false_64off = false_reg->var_off;
6878 struct tnum true_32off = tnum_subreg(true_reg->var_off);
6879 struct tnum true_64off = true_reg->var_off;
6880 s64 sval = (s64)val;
6881 s32 sval32 = (s32)val32;
6882
6883 /* If the dst_reg is a pointer, we can't learn anything about its
6884 * variable offset from the compare (unless src_reg were a pointer into
6885 * the same object, but we don't bother with that.
6886 * Since false_reg and true_reg have the same type by construction, we
6887 * only need to check one of them for pointerness.
6888 */
6889 if (__is_pointer_value(false, false_reg))
6890 return;
6891
6892 switch (opcode) {
6893 case BPF_JEQ:
6894 case BPF_JNE:
6895 {
6896 struct bpf_reg_state *reg =
6897 opcode == BPF_JEQ ? true_reg : false_reg;
6898
6899 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
6900 * if it is true we know the value for sure. Likewise for
6901 * BPF_JNE.
6902 */
6903 if (is_jmp32)
6904 __mark_reg32_known(reg, val32);
6905 else
6906 __mark_reg_known(reg, val);
6907 break;
6908 }
6909 case BPF_JSET:
6910 if (is_jmp32) {
6911 false_32off = tnum_and(false_32off, tnum_const(~val32));
6912 if (is_power_of_2(val32))
6913 true_32off = tnum_or(true_32off,
6914 tnum_const(val32));
6915 } else {
6916 false_64off = tnum_and(false_64off, tnum_const(~val));
6917 if (is_power_of_2(val))
6918 true_64off = tnum_or(true_64off,
6919 tnum_const(val));
6920 }
6921 break;
6922 case BPF_JGE:
6923 case BPF_JGT:
6924 {
6925 if (is_jmp32) {
6926 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
6927 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
6928
6929 false_reg->u32_max_value = min(false_reg->u32_max_value,
6930 false_umax);
6931 true_reg->u32_min_value = max(true_reg->u32_min_value,
6932 true_umin);
6933 } else {
6934 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
6935 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
6936
6937 false_reg->umax_value = min(false_reg->umax_value, false_umax);
6938 true_reg->umin_value = max(true_reg->umin_value, true_umin);
6939 }
6940 break;
6941 }
6942 case BPF_JSGE:
6943 case BPF_JSGT:
6944 {
6945 if (is_jmp32) {
6946 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
6947 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
6948
6949 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
6950 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
6951 } else {
6952 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
6953 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
6954
6955 false_reg->smax_value = min(false_reg->smax_value, false_smax);
6956 true_reg->smin_value = max(true_reg->smin_value, true_smin);
6957 }
6958 break;
6959 }
6960 case BPF_JLE:
6961 case BPF_JLT:
6962 {
6963 if (is_jmp32) {
6964 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
6965 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
6966
6967 false_reg->u32_min_value = max(false_reg->u32_min_value,
6968 false_umin);
6969 true_reg->u32_max_value = min(true_reg->u32_max_value,
6970 true_umax);
6971 } else {
6972 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
6973 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
6974
6975 false_reg->umin_value = max(false_reg->umin_value, false_umin);
6976 true_reg->umax_value = min(true_reg->umax_value, true_umax);
6977 }
6978 break;
6979 }
6980 case BPF_JSLE:
6981 case BPF_JSLT:
6982 {
6983 if (is_jmp32) {
6984 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
6985 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
6986
6987 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
6988 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
6989 } else {
6990 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
6991 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
6992
6993 false_reg->smin_value = max(false_reg->smin_value, false_smin);
6994 true_reg->smax_value = min(true_reg->smax_value, true_smax);
6995 }
6996 break;
6997 }
6998 default:
6999 return;
7000 }
7001
7002 if (is_jmp32) {
7003 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
7004 tnum_subreg(false_32off));
7005 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
7006 tnum_subreg(true_32off));
7007 __reg_combine_32_into_64(false_reg);
7008 __reg_combine_32_into_64(true_reg);
7009 } else {
7010 false_reg->var_off = false_64off;
7011 true_reg->var_off = true_64off;
7012 __reg_combine_64_into_32(false_reg);
7013 __reg_combine_64_into_32(true_reg);
7014 }
7015 }
7016
7017 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
7018 * the variable reg.
7019 */
7020 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
7021 struct bpf_reg_state *false_reg,
7022 u64 val, u32 val32,
7023 u8 opcode, bool is_jmp32)
7024 {
7025 /* How can we transform "a <op> b" into "b <op> a"? */
7026 static const u8 opcode_flip[16] = {
7027 /* these stay the same */
7028 [BPF_JEQ >> 4] = BPF_JEQ,
7029 [BPF_JNE >> 4] = BPF_JNE,
7030 [BPF_JSET >> 4] = BPF_JSET,
7031 /* these swap "lesser" and "greater" (L and G in the opcodes) */
7032 [BPF_JGE >> 4] = BPF_JLE,
7033 [BPF_JGT >> 4] = BPF_JLT,
7034 [BPF_JLE >> 4] = BPF_JGE,
7035 [BPF_JLT >> 4] = BPF_JGT,
7036 [BPF_JSGE >> 4] = BPF_JSLE,
7037 [BPF_JSGT >> 4] = BPF_JSLT,
7038 [BPF_JSLE >> 4] = BPF_JSGE,
7039 [BPF_JSLT >> 4] = BPF_JSGT
7040 };
7041 opcode = opcode_flip[opcode >> 4];
7042 /* This uses zero as "not present in table"; luckily the zero opcode,
7043 * BPF_JA, can't get here.
7044 */
7045 if (opcode)
7046 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
7047 }
7048
7049 /* Regs are known to be equal, so intersect their min/max/var_off */
7050 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
7051 struct bpf_reg_state *dst_reg)
7052 {
7053 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
7054 dst_reg->umin_value);
7055 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
7056 dst_reg->umax_value);
7057 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
7058 dst_reg->smin_value);
7059 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
7060 dst_reg->smax_value);
7061 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
7062 dst_reg->var_off);
7063 /* We might have learned new bounds from the var_off. */
7064 __update_reg_bounds(src_reg);
7065 __update_reg_bounds(dst_reg);
7066 /* We might have learned something about the sign bit. */
7067 __reg_deduce_bounds(src_reg);
7068 __reg_deduce_bounds(dst_reg);
7069 /* We might have learned some bits from the bounds. */
7070 __reg_bound_offset(src_reg);
7071 __reg_bound_offset(dst_reg);
7072 /* Intersecting with the old var_off might have improved our bounds
7073 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
7074 * then new var_off is (0; 0x7f...fc) which improves our umax.
7075 */
7076 __update_reg_bounds(src_reg);
7077 __update_reg_bounds(dst_reg);
7078 }
7079
7080 static void reg_combine_min_max(struct bpf_reg_state *true_src,
7081 struct bpf_reg_state *true_dst,
7082 struct bpf_reg_state *false_src,
7083 struct bpf_reg_state *false_dst,
7084 u8 opcode)
7085 {
7086 switch (opcode) {
7087 case BPF_JEQ:
7088 __reg_combine_min_max(true_src, true_dst);
7089 break;
7090 case BPF_JNE:
7091 __reg_combine_min_max(false_src, false_dst);
7092 break;
7093 }
7094 }
7095
7096 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
7097 struct bpf_reg_state *reg, u32 id,
7098 bool is_null)
7099 {
7100 if (reg_type_may_be_null(reg->type) && reg->id == id) {
7101 /* Old offset (both fixed and variable parts) should
7102 * have been known-zero, because we don't allow pointer
7103 * arithmetic on pointers that might be NULL.
7104 */
7105 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
7106 !tnum_equals_const(reg->var_off, 0) ||
7107 reg->off)) {
7108 __mark_reg_known_zero(reg);
7109 reg->off = 0;
7110 }
7111 if (is_null) {
7112 reg->type = SCALAR_VALUE;
7113 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
7114 const struct bpf_map *map = reg->map_ptr;
7115
7116 if (map->inner_map_meta) {
7117 reg->type = CONST_PTR_TO_MAP;
7118 reg->map_ptr = map->inner_map_meta;
7119 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
7120 reg->type = PTR_TO_XDP_SOCK;
7121 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
7122 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
7123 reg->type = PTR_TO_SOCKET;
7124 } else {
7125 reg->type = PTR_TO_MAP_VALUE;
7126 }
7127 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
7128 reg->type = PTR_TO_SOCKET;
7129 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
7130 reg->type = PTR_TO_SOCK_COMMON;
7131 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
7132 reg->type = PTR_TO_TCP_SOCK;
7133 } else if (reg->type == PTR_TO_BTF_ID_OR_NULL) {
7134 reg->type = PTR_TO_BTF_ID;
7135 } else if (reg->type == PTR_TO_MEM_OR_NULL) {
7136 reg->type = PTR_TO_MEM;
7137 } else if (reg->type == PTR_TO_RDONLY_BUF_OR_NULL) {
7138 reg->type = PTR_TO_RDONLY_BUF;
7139 } else if (reg->type == PTR_TO_RDWR_BUF_OR_NULL) {
7140 reg->type = PTR_TO_RDWR_BUF;
7141 }
7142 if (is_null) {
7143 /* We don't need id and ref_obj_id from this point
7144 * onwards anymore, thus we should better reset it,
7145 * so that state pruning has chances to take effect.
7146 */
7147 reg->id = 0;
7148 reg->ref_obj_id = 0;
7149 } else if (!reg_may_point_to_spin_lock(reg)) {
7150 /* For not-NULL ptr, reg->ref_obj_id will be reset
7151 * in release_reg_references().
7152 *
7153 * reg->id is still used by spin_lock ptr. Other
7154 * than spin_lock ptr type, reg->id can be reset.
7155 */
7156 reg->id = 0;
7157 }
7158 }
7159 }
7160
7161 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
7162 bool is_null)
7163 {
7164 struct bpf_reg_state *reg;
7165 int i;
7166
7167 for (i = 0; i < MAX_BPF_REG; i++)
7168 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
7169
7170 bpf_for_each_spilled_reg(i, state, reg) {
7171 if (!reg)
7172 continue;
7173 mark_ptr_or_null_reg(state, reg, id, is_null);
7174 }
7175 }
7176
7177 /* The logic is similar to find_good_pkt_pointers(), both could eventually
7178 * be folded together at some point.
7179 */
7180 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
7181 bool is_null)
7182 {
7183 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7184 struct bpf_reg_state *regs = state->regs;
7185 u32 ref_obj_id = regs[regno].ref_obj_id;
7186 u32 id = regs[regno].id;
7187 int i;
7188
7189 if (ref_obj_id && ref_obj_id == id && is_null)
7190 /* regs[regno] is in the " == NULL" branch.
7191 * No one could have freed the reference state before
7192 * doing the NULL check.
7193 */
7194 WARN_ON_ONCE(release_reference_state(state, id));
7195
7196 for (i = 0; i <= vstate->curframe; i++)
7197 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
7198 }
7199
7200 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
7201 struct bpf_reg_state *dst_reg,
7202 struct bpf_reg_state *src_reg,
7203 struct bpf_verifier_state *this_branch,
7204 struct bpf_verifier_state *other_branch)
7205 {
7206 if (BPF_SRC(insn->code) != BPF_X)
7207 return false;
7208
7209 /* Pointers are always 64-bit. */
7210 if (BPF_CLASS(insn->code) == BPF_JMP32)
7211 return false;
7212
7213 switch (BPF_OP(insn->code)) {
7214 case BPF_JGT:
7215 if ((dst_reg->type == PTR_TO_PACKET &&
7216 src_reg->type == PTR_TO_PACKET_END) ||
7217 (dst_reg->type == PTR_TO_PACKET_META &&
7218 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7219 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
7220 find_good_pkt_pointers(this_branch, dst_reg,
7221 dst_reg->type, false);
7222 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7223 src_reg->type == PTR_TO_PACKET) ||
7224 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7225 src_reg->type == PTR_TO_PACKET_META)) {
7226 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
7227 find_good_pkt_pointers(other_branch, src_reg,
7228 src_reg->type, true);
7229 } else {
7230 return false;
7231 }
7232 break;
7233 case BPF_JLT:
7234 if ((dst_reg->type == PTR_TO_PACKET &&
7235 src_reg->type == PTR_TO_PACKET_END) ||
7236 (dst_reg->type == PTR_TO_PACKET_META &&
7237 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7238 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
7239 find_good_pkt_pointers(other_branch, dst_reg,
7240 dst_reg->type, true);
7241 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7242 src_reg->type == PTR_TO_PACKET) ||
7243 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7244 src_reg->type == PTR_TO_PACKET_META)) {
7245 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
7246 find_good_pkt_pointers(this_branch, src_reg,
7247 src_reg->type, false);
7248 } else {
7249 return false;
7250 }
7251 break;
7252 case BPF_JGE:
7253 if ((dst_reg->type == PTR_TO_PACKET &&
7254 src_reg->type == PTR_TO_PACKET_END) ||
7255 (dst_reg->type == PTR_TO_PACKET_META &&
7256 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7257 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
7258 find_good_pkt_pointers(this_branch, dst_reg,
7259 dst_reg->type, true);
7260 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7261 src_reg->type == PTR_TO_PACKET) ||
7262 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7263 src_reg->type == PTR_TO_PACKET_META)) {
7264 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
7265 find_good_pkt_pointers(other_branch, src_reg,
7266 src_reg->type, false);
7267 } else {
7268 return false;
7269 }
7270 break;
7271 case BPF_JLE:
7272 if ((dst_reg->type == PTR_TO_PACKET &&
7273 src_reg->type == PTR_TO_PACKET_END) ||
7274 (dst_reg->type == PTR_TO_PACKET_META &&
7275 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7276 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
7277 find_good_pkt_pointers(other_branch, dst_reg,
7278 dst_reg->type, false);
7279 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7280 src_reg->type == PTR_TO_PACKET) ||
7281 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7282 src_reg->type == PTR_TO_PACKET_META)) {
7283 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
7284 find_good_pkt_pointers(this_branch, src_reg,
7285 src_reg->type, true);
7286 } else {
7287 return false;
7288 }
7289 break;
7290 default:
7291 return false;
7292 }
7293
7294 return true;
7295 }
7296
7297 static int check_cond_jmp_op(struct bpf_verifier_env *env,
7298 struct bpf_insn *insn, int *insn_idx)
7299 {
7300 struct bpf_verifier_state *this_branch = env->cur_state;
7301 struct bpf_verifier_state *other_branch;
7302 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
7303 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
7304 u8 opcode = BPF_OP(insn->code);
7305 bool is_jmp32;
7306 int pred = -1;
7307 int err;
7308
7309 /* Only conditional jumps are expected to reach here. */
7310 if (opcode == BPF_JA || opcode > BPF_JSLE) {
7311 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
7312 return -EINVAL;
7313 }
7314
7315 if (BPF_SRC(insn->code) == BPF_X) {
7316 if (insn->imm != 0) {
7317 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7318 return -EINVAL;
7319 }
7320
7321 /* check src1 operand */
7322 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7323 if (err)
7324 return err;
7325
7326 if (is_pointer_value(env, insn->src_reg)) {
7327 verbose(env, "R%d pointer comparison prohibited\n",
7328 insn->src_reg);
7329 return -EACCES;
7330 }
7331 src_reg = &regs[insn->src_reg];
7332 } else {
7333 if (insn->src_reg != BPF_REG_0) {
7334 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7335 return -EINVAL;
7336 }
7337 }
7338
7339 /* check src2 operand */
7340 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7341 if (err)
7342 return err;
7343
7344 dst_reg = &regs[insn->dst_reg];
7345 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
7346
7347 if (BPF_SRC(insn->code) == BPF_K) {
7348 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
7349 } else if (src_reg->type == SCALAR_VALUE &&
7350 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
7351 pred = is_branch_taken(dst_reg,
7352 tnum_subreg(src_reg->var_off).value,
7353 opcode,
7354 is_jmp32);
7355 } else if (src_reg->type == SCALAR_VALUE &&
7356 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
7357 pred = is_branch_taken(dst_reg,
7358 src_reg->var_off.value,
7359 opcode,
7360 is_jmp32);
7361 }
7362
7363 if (pred >= 0) {
7364 /* If we get here with a dst_reg pointer type it is because
7365 * above is_branch_taken() special cased the 0 comparison.
7366 */
7367 if (!__is_pointer_value(false, dst_reg))
7368 err = mark_chain_precision(env, insn->dst_reg);
7369 if (BPF_SRC(insn->code) == BPF_X && !err)
7370 err = mark_chain_precision(env, insn->src_reg);
7371 if (err)
7372 return err;
7373 }
7374 if (pred == 1) {
7375 /* only follow the goto, ignore fall-through */
7376 *insn_idx += insn->off;
7377 return 0;
7378 } else if (pred == 0) {
7379 /* only follow fall-through branch, since
7380 * that's where the program will go
7381 */
7382 return 0;
7383 }
7384
7385 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
7386 false);
7387 if (!other_branch)
7388 return -EFAULT;
7389 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
7390
7391 /* detect if we are comparing against a constant value so we can adjust
7392 * our min/max values for our dst register.
7393 * this is only legit if both are scalars (or pointers to the same
7394 * object, I suppose, but we don't support that right now), because
7395 * otherwise the different base pointers mean the offsets aren't
7396 * comparable.
7397 */
7398 if (BPF_SRC(insn->code) == BPF_X) {
7399 struct bpf_reg_state *src_reg = &regs[insn->src_reg];
7400
7401 if (dst_reg->type == SCALAR_VALUE &&
7402 src_reg->type == SCALAR_VALUE) {
7403 if (tnum_is_const(src_reg->var_off) ||
7404 (is_jmp32 &&
7405 tnum_is_const(tnum_subreg(src_reg->var_off))))
7406 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7407 dst_reg,
7408 src_reg->var_off.value,
7409 tnum_subreg(src_reg->var_off).value,
7410 opcode, is_jmp32);
7411 else if (tnum_is_const(dst_reg->var_off) ||
7412 (is_jmp32 &&
7413 tnum_is_const(tnum_subreg(dst_reg->var_off))))
7414 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
7415 src_reg,
7416 dst_reg->var_off.value,
7417 tnum_subreg(dst_reg->var_off).value,
7418 opcode, is_jmp32);
7419 else if (!is_jmp32 &&
7420 (opcode == BPF_JEQ || opcode == BPF_JNE))
7421 /* Comparing for equality, we can combine knowledge */
7422 reg_combine_min_max(&other_branch_regs[insn->src_reg],
7423 &other_branch_regs[insn->dst_reg],
7424 src_reg, dst_reg, opcode);
7425 }
7426 } else if (dst_reg->type == SCALAR_VALUE) {
7427 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7428 dst_reg, insn->imm, (u32)insn->imm,
7429 opcode, is_jmp32);
7430 }
7431
7432 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
7433 * NOTE: these optimizations below are related with pointer comparison
7434 * which will never be JMP32.
7435 */
7436 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
7437 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
7438 reg_type_may_be_null(dst_reg->type)) {
7439 /* Mark all identical registers in each branch as either
7440 * safe or unknown depending R == 0 or R != 0 conditional.
7441 */
7442 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
7443 opcode == BPF_JNE);
7444 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
7445 opcode == BPF_JEQ);
7446 } else if (!try_match_pkt_pointers(insn, dst_reg, &regs[insn->src_reg],
7447 this_branch, other_branch) &&
7448 is_pointer_value(env, insn->dst_reg)) {
7449 verbose(env, "R%d pointer comparison prohibited\n",
7450 insn->dst_reg);
7451 return -EACCES;
7452 }
7453 if (env->log.level & BPF_LOG_LEVEL)
7454 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
7455 return 0;
7456 }
7457
7458 /* verify BPF_LD_IMM64 instruction */
7459 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
7460 {
7461 struct bpf_insn_aux_data *aux = cur_aux(env);
7462 struct bpf_reg_state *regs = cur_regs(env);
7463 struct bpf_map *map;
7464 int err;
7465
7466 if (BPF_SIZE(insn->code) != BPF_DW) {
7467 verbose(env, "invalid BPF_LD_IMM insn\n");
7468 return -EINVAL;
7469 }
7470 if (insn->off != 0) {
7471 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
7472 return -EINVAL;
7473 }
7474
7475 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7476 if (err)
7477 return err;
7478
7479 if (insn->src_reg == 0) {
7480 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
7481
7482 regs[insn->dst_reg].type = SCALAR_VALUE;
7483 __mark_reg_known(&regs[insn->dst_reg], imm);
7484 return 0;
7485 }
7486
7487 map = env->used_maps[aux->map_index];
7488 mark_reg_known_zero(env, regs, insn->dst_reg);
7489 regs[insn->dst_reg].map_ptr = map;
7490
7491 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
7492 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
7493 regs[insn->dst_reg].off = aux->map_off;
7494 if (map_value_has_spin_lock(map))
7495 regs[insn->dst_reg].id = ++env->id_gen;
7496 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
7497 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
7498 } else {
7499 verbose(env, "bpf verifier is misconfigured\n");
7500 return -EINVAL;
7501 }
7502
7503 return 0;
7504 }
7505
7506 static bool may_access_skb(enum bpf_prog_type type)
7507 {
7508 switch (type) {
7509 case BPF_PROG_TYPE_SOCKET_FILTER:
7510 case BPF_PROG_TYPE_SCHED_CLS:
7511 case BPF_PROG_TYPE_SCHED_ACT:
7512 return true;
7513 default:
7514 return false;
7515 }
7516 }
7517
7518 /* verify safety of LD_ABS|LD_IND instructions:
7519 * - they can only appear in the programs where ctx == skb
7520 * - since they are wrappers of function calls, they scratch R1-R5 registers,
7521 * preserve R6-R9, and store return value into R0
7522 *
7523 * Implicit input:
7524 * ctx == skb == R6 == CTX
7525 *
7526 * Explicit input:
7527 * SRC == any register
7528 * IMM == 32-bit immediate
7529 *
7530 * Output:
7531 * R0 - 8/16/32-bit skb data converted to cpu endianness
7532 */
7533 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
7534 {
7535 struct bpf_reg_state *regs = cur_regs(env);
7536 static const int ctx_reg = BPF_REG_6;
7537 u8 mode = BPF_MODE(insn->code);
7538 int i, err;
7539
7540 if (!may_access_skb(resolve_prog_type(env->prog))) {
7541 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
7542 return -EINVAL;
7543 }
7544
7545 if (!env->ops->gen_ld_abs) {
7546 verbose(env, "bpf verifier is misconfigured\n");
7547 return -EINVAL;
7548 }
7549
7550 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
7551 BPF_SIZE(insn->code) == BPF_DW ||
7552 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
7553 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
7554 return -EINVAL;
7555 }
7556
7557 /* check whether implicit source operand (register R6) is readable */
7558 err = check_reg_arg(env, ctx_reg, SRC_OP);
7559 if (err)
7560 return err;
7561
7562 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
7563 * gen_ld_abs() may terminate the program at runtime, leading to
7564 * reference leak.
7565 */
7566 err = check_reference_leak(env);
7567 if (err) {
7568 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
7569 return err;
7570 }
7571
7572 if (env->cur_state->active_spin_lock) {
7573 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
7574 return -EINVAL;
7575 }
7576
7577 if (regs[ctx_reg].type != PTR_TO_CTX) {
7578 verbose(env,
7579 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
7580 return -EINVAL;
7581 }
7582
7583 if (mode == BPF_IND) {
7584 /* check explicit source operand */
7585 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7586 if (err)
7587 return err;
7588 }
7589
7590 err = check_ctx_reg(env, &regs[ctx_reg], ctx_reg);
7591 if (err < 0)
7592 return err;
7593
7594 /* reset caller saved regs to unreadable */
7595 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7596 mark_reg_not_init(env, regs, caller_saved[i]);
7597 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7598 }
7599
7600 /* mark destination R0 register as readable, since it contains
7601 * the value fetched from the packet.
7602 * Already marked as written above.
7603 */
7604 mark_reg_unknown(env, regs, BPF_REG_0);
7605 /* ld_abs load up to 32-bit skb data. */
7606 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
7607 return 0;
7608 }
7609
7610 static int check_return_code(struct bpf_verifier_env *env)
7611 {
7612 struct tnum enforce_attach_type_range = tnum_unknown;
7613 const struct bpf_prog *prog = env->prog;
7614 struct bpf_reg_state *reg;
7615 struct tnum range = tnum_range(0, 1);
7616 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
7617 int err;
7618
7619 /* LSM and struct_ops func-ptr's return type could be "void" */
7620 if ((prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
7621 prog_type == BPF_PROG_TYPE_LSM) &&
7622 !prog->aux->attach_func_proto->type)
7623 return 0;
7624
7625 /* eBPF calling convetion is such that R0 is used
7626 * to return the value from eBPF program.
7627 * Make sure that it's readable at this time
7628 * of bpf_exit, which means that program wrote
7629 * something into it earlier
7630 */
7631 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
7632 if (err)
7633 return err;
7634
7635 if (is_pointer_value(env, BPF_REG_0)) {
7636 verbose(env, "R0 leaks addr as return value\n");
7637 return -EACCES;
7638 }
7639
7640 switch (prog_type) {
7641 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
7642 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
7643 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
7644 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
7645 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
7646 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
7647 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
7648 range = tnum_range(1, 1);
7649 break;
7650 case BPF_PROG_TYPE_CGROUP_SKB:
7651 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
7652 range = tnum_range(0, 3);
7653 enforce_attach_type_range = tnum_range(2, 3);
7654 }
7655 break;
7656 case BPF_PROG_TYPE_CGROUP_SOCK:
7657 case BPF_PROG_TYPE_SOCK_OPS:
7658 case BPF_PROG_TYPE_CGROUP_DEVICE:
7659 case BPF_PROG_TYPE_CGROUP_SYSCTL:
7660 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
7661 break;
7662 case BPF_PROG_TYPE_RAW_TRACEPOINT:
7663 if (!env->prog->aux->attach_btf_id)
7664 return 0;
7665 range = tnum_const(0);
7666 break;
7667 case BPF_PROG_TYPE_TRACING:
7668 switch (env->prog->expected_attach_type) {
7669 case BPF_TRACE_FENTRY:
7670 case BPF_TRACE_FEXIT:
7671 range = tnum_const(0);
7672 break;
7673 case BPF_TRACE_RAW_TP:
7674 case BPF_MODIFY_RETURN:
7675 return 0;
7676 case BPF_TRACE_ITER:
7677 break;
7678 default:
7679 return -ENOTSUPP;
7680 }
7681 break;
7682 case BPF_PROG_TYPE_SK_LOOKUP:
7683 range = tnum_range(SK_DROP, SK_PASS);
7684 break;
7685 case BPF_PROG_TYPE_EXT:
7686 /* freplace program can return anything as its return value
7687 * depends on the to-be-replaced kernel func or bpf program.
7688 */
7689 default:
7690 return 0;
7691 }
7692
7693 reg = cur_regs(env) + BPF_REG_0;
7694 if (reg->type != SCALAR_VALUE) {
7695 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
7696 reg_type_str[reg->type]);
7697 return -EINVAL;
7698 }
7699
7700 if (!tnum_in(range, reg->var_off)) {
7701 char tn_buf[48];
7702
7703 verbose(env, "At program exit the register R0 ");
7704 if (!tnum_is_unknown(reg->var_off)) {
7705 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7706 verbose(env, "has value %s", tn_buf);
7707 } else {
7708 verbose(env, "has unknown scalar value");
7709 }
7710 tnum_strn(tn_buf, sizeof(tn_buf), range);
7711 verbose(env, " should have been in %s\n", tn_buf);
7712 return -EINVAL;
7713 }
7714
7715 if (!tnum_is_unknown(enforce_attach_type_range) &&
7716 tnum_in(enforce_attach_type_range, reg->var_off))
7717 env->prog->enforce_expected_attach_type = 1;
7718 return 0;
7719 }
7720
7721 /* non-recursive DFS pseudo code
7722 * 1 procedure DFS-iterative(G,v):
7723 * 2 label v as discovered
7724 * 3 let S be a stack
7725 * 4 S.push(v)
7726 * 5 while S is not empty
7727 * 6 t <- S.pop()
7728 * 7 if t is what we're looking for:
7729 * 8 return t
7730 * 9 for all edges e in G.adjacentEdges(t) do
7731 * 10 if edge e is already labelled
7732 * 11 continue with the next edge
7733 * 12 w <- G.adjacentVertex(t,e)
7734 * 13 if vertex w is not discovered and not explored
7735 * 14 label e as tree-edge
7736 * 15 label w as discovered
7737 * 16 S.push(w)
7738 * 17 continue at 5
7739 * 18 else if vertex w is discovered
7740 * 19 label e as back-edge
7741 * 20 else
7742 * 21 // vertex w is explored
7743 * 22 label e as forward- or cross-edge
7744 * 23 label t as explored
7745 * 24 S.pop()
7746 *
7747 * convention:
7748 * 0x10 - discovered
7749 * 0x11 - discovered and fall-through edge labelled
7750 * 0x12 - discovered and fall-through and branch edges labelled
7751 * 0x20 - explored
7752 */
7753
7754 enum {
7755 DISCOVERED = 0x10,
7756 EXPLORED = 0x20,
7757 FALLTHROUGH = 1,
7758 BRANCH = 2,
7759 };
7760
7761 static u32 state_htab_size(struct bpf_verifier_env *env)
7762 {
7763 return env->prog->len;
7764 }
7765
7766 static struct bpf_verifier_state_list **explored_state(
7767 struct bpf_verifier_env *env,
7768 int idx)
7769 {
7770 struct bpf_verifier_state *cur = env->cur_state;
7771 struct bpf_func_state *state = cur->frame[cur->curframe];
7772
7773 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
7774 }
7775
7776 static void init_explored_state(struct bpf_verifier_env *env, int idx)
7777 {
7778 env->insn_aux_data[idx].prune_point = true;
7779 }
7780
7781 /* t, w, e - match pseudo-code above:
7782 * t - index of current instruction
7783 * w - next instruction
7784 * e - edge
7785 */
7786 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
7787 bool loop_ok)
7788 {
7789 int *insn_stack = env->cfg.insn_stack;
7790 int *insn_state = env->cfg.insn_state;
7791
7792 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
7793 return 0;
7794
7795 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
7796 return 0;
7797
7798 if (w < 0 || w >= env->prog->len) {
7799 verbose_linfo(env, t, "%d: ", t);
7800 verbose(env, "jump out of range from insn %d to %d\n", t, w);
7801 return -EINVAL;
7802 }
7803
7804 if (e == BRANCH)
7805 /* mark branch target for state pruning */
7806 init_explored_state(env, w);
7807
7808 if (insn_state[w] == 0) {
7809 /* tree-edge */
7810 insn_state[t] = DISCOVERED | e;
7811 insn_state[w] = DISCOVERED;
7812 if (env->cfg.cur_stack >= env->prog->len)
7813 return -E2BIG;
7814 insn_stack[env->cfg.cur_stack++] = w;
7815 return 1;
7816 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
7817 if (loop_ok && env->bpf_capable)
7818 return 0;
7819 verbose_linfo(env, t, "%d: ", t);
7820 verbose_linfo(env, w, "%d: ", w);
7821 verbose(env, "back-edge from insn %d to %d\n", t, w);
7822 return -EINVAL;
7823 } else if (insn_state[w] == EXPLORED) {
7824 /* forward- or cross-edge */
7825 insn_state[t] = DISCOVERED | e;
7826 } else {
7827 verbose(env, "insn state internal bug\n");
7828 return -EFAULT;
7829 }
7830 return 0;
7831 }
7832
7833 /* non-recursive depth-first-search to detect loops in BPF program
7834 * loop == back-edge in directed graph
7835 */
7836 static int check_cfg(struct bpf_verifier_env *env)
7837 {
7838 struct bpf_insn *insns = env->prog->insnsi;
7839 int insn_cnt = env->prog->len;
7840 int *insn_stack, *insn_state;
7841 int ret = 0;
7842 int i, t;
7843
7844 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7845 if (!insn_state)
7846 return -ENOMEM;
7847
7848 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
7849 if (!insn_stack) {
7850 kvfree(insn_state);
7851 return -ENOMEM;
7852 }
7853
7854 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
7855 insn_stack[0] = 0; /* 0 is the first instruction */
7856 env->cfg.cur_stack = 1;
7857
7858 peek_stack:
7859 if (env->cfg.cur_stack == 0)
7860 goto check_state;
7861 t = insn_stack[env->cfg.cur_stack - 1];
7862
7863 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
7864 BPF_CLASS(insns[t].code) == BPF_JMP32) {
7865 u8 opcode = BPF_OP(insns[t].code);
7866
7867 if (opcode == BPF_EXIT) {
7868 goto mark_explored;
7869 } else if (opcode == BPF_CALL) {
7870 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7871 if (ret == 1)
7872 goto peek_stack;
7873 else if (ret < 0)
7874 goto err_free;
7875 if (t + 1 < insn_cnt)
7876 init_explored_state(env, t + 1);
7877 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
7878 init_explored_state(env, t);
7879 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
7880 env, false);
7881 if (ret == 1)
7882 goto peek_stack;
7883 else if (ret < 0)
7884 goto err_free;
7885 }
7886 } else if (opcode == BPF_JA) {
7887 if (BPF_SRC(insns[t].code) != BPF_K) {
7888 ret = -EINVAL;
7889 goto err_free;
7890 }
7891 /* unconditional jump with single edge */
7892 ret = push_insn(t, t + insns[t].off + 1,
7893 FALLTHROUGH, env, true);
7894 if (ret == 1)
7895 goto peek_stack;
7896 else if (ret < 0)
7897 goto err_free;
7898 /* unconditional jmp is not a good pruning point,
7899 * but it's marked, since backtracking needs
7900 * to record jmp history in is_state_visited().
7901 */
7902 init_explored_state(env, t + insns[t].off + 1);
7903 /* tell verifier to check for equivalent states
7904 * after every call and jump
7905 */
7906 if (t + 1 < insn_cnt)
7907 init_explored_state(env, t + 1);
7908 } else {
7909 /* conditional jump with two edges */
7910 init_explored_state(env, t);
7911 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
7912 if (ret == 1)
7913 goto peek_stack;
7914 else if (ret < 0)
7915 goto err_free;
7916
7917 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
7918 if (ret == 1)
7919 goto peek_stack;
7920 else if (ret < 0)
7921 goto err_free;
7922 }
7923 } else {
7924 /* all other non-branch instructions with single
7925 * fall-through edge
7926 */
7927 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
7928 if (ret == 1)
7929 goto peek_stack;
7930 else if (ret < 0)
7931 goto err_free;
7932 }
7933
7934 mark_explored:
7935 insn_state[t] = EXPLORED;
7936 if (env->cfg.cur_stack-- <= 0) {
7937 verbose(env, "pop stack internal bug\n");
7938 ret = -EFAULT;
7939 goto err_free;
7940 }
7941 goto peek_stack;
7942
7943 check_state:
7944 for (i = 0; i < insn_cnt; i++) {
7945 if (insn_state[i] != EXPLORED) {
7946 verbose(env, "unreachable insn %d\n", i);
7947 ret = -EINVAL;
7948 goto err_free;
7949 }
7950 }
7951 ret = 0; /* cfg looks good */
7952
7953 err_free:
7954 kvfree(insn_state);
7955 kvfree(insn_stack);
7956 env->cfg.insn_state = env->cfg.insn_stack = NULL;
7957 return ret;
7958 }
7959
7960 static int check_abnormal_return(struct bpf_verifier_env *env)
7961 {
7962 int i;
7963
7964 for (i = 1; i < env->subprog_cnt; i++) {
7965 if (env->subprog_info[i].has_ld_abs) {
7966 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
7967 return -EINVAL;
7968 }
7969 if (env->subprog_info[i].has_tail_call) {
7970 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
7971 return -EINVAL;
7972 }
7973 }
7974 return 0;
7975 }
7976
7977 /* The minimum supported BTF func info size */
7978 #define MIN_BPF_FUNCINFO_SIZE 8
7979 #define MAX_FUNCINFO_REC_SIZE 252
7980
7981 static int check_btf_func(struct bpf_verifier_env *env,
7982 const union bpf_attr *attr,
7983 union bpf_attr __user *uattr)
7984 {
7985 const struct btf_type *type, *func_proto, *ret_type;
7986 u32 i, nfuncs, urec_size, min_size;
7987 u32 krec_size = sizeof(struct bpf_func_info);
7988 struct bpf_func_info *krecord;
7989 struct bpf_func_info_aux *info_aux = NULL;
7990 struct bpf_prog *prog;
7991 const struct btf *btf;
7992 void __user *urecord;
7993 u32 prev_offset = 0;
7994 bool scalar_return;
7995 int ret = -ENOMEM;
7996
7997 nfuncs = attr->func_info_cnt;
7998 if (!nfuncs) {
7999 if (check_abnormal_return(env))
8000 return -EINVAL;
8001 return 0;
8002 }
8003
8004 if (nfuncs != env->subprog_cnt) {
8005 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
8006 return -EINVAL;
8007 }
8008
8009 urec_size = attr->func_info_rec_size;
8010 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
8011 urec_size > MAX_FUNCINFO_REC_SIZE ||
8012 urec_size % sizeof(u32)) {
8013 verbose(env, "invalid func info rec size %u\n", urec_size);
8014 return -EINVAL;
8015 }
8016
8017 prog = env->prog;
8018 btf = prog->aux->btf;
8019
8020 urecord = u64_to_user_ptr(attr->func_info);
8021 min_size = min_t(u32, krec_size, urec_size);
8022
8023 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
8024 if (!krecord)
8025 return -ENOMEM;
8026 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
8027 if (!info_aux)
8028 goto err_free;
8029
8030 for (i = 0; i < nfuncs; i++) {
8031 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
8032 if (ret) {
8033 if (ret == -E2BIG) {
8034 verbose(env, "nonzero tailing record in func info");
8035 /* set the size kernel expects so loader can zero
8036 * out the rest of the record.
8037 */
8038 if (put_user(min_size, &uattr->func_info_rec_size))
8039 ret = -EFAULT;
8040 }
8041 goto err_free;
8042 }
8043
8044 if (copy_from_user(&krecord[i], urecord, min_size)) {
8045 ret = -EFAULT;
8046 goto err_free;
8047 }
8048
8049 /* check insn_off */
8050 ret = -EINVAL;
8051 if (i == 0) {
8052 if (krecord[i].insn_off) {
8053 verbose(env,
8054 "nonzero insn_off %u for the first func info record",
8055 krecord[i].insn_off);
8056 goto err_free;
8057 }
8058 } else if (krecord[i].insn_off <= prev_offset) {
8059 verbose(env,
8060 "same or smaller insn offset (%u) than previous func info record (%u)",
8061 krecord[i].insn_off, prev_offset);
8062 goto err_free;
8063 }
8064
8065 if (env->subprog_info[i].start != krecord[i].insn_off) {
8066 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
8067 goto err_free;
8068 }
8069
8070 /* check type_id */
8071 type = btf_type_by_id(btf, krecord[i].type_id);
8072 if (!type || !btf_type_is_func(type)) {
8073 verbose(env, "invalid type id %d in func info",
8074 krecord[i].type_id);
8075 goto err_free;
8076 }
8077 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
8078
8079 func_proto = btf_type_by_id(btf, type->type);
8080 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
8081 /* btf_func_check() already verified it during BTF load */
8082 goto err_free;
8083 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
8084 scalar_return =
8085 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
8086 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
8087 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
8088 goto err_free;
8089 }
8090 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
8091 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
8092 goto err_free;
8093 }
8094
8095 prev_offset = krecord[i].insn_off;
8096 urecord += urec_size;
8097 }
8098
8099 prog->aux->func_info = krecord;
8100 prog->aux->func_info_cnt = nfuncs;
8101 prog->aux->func_info_aux = info_aux;
8102 return 0;
8103
8104 err_free:
8105 kvfree(krecord);
8106 kfree(info_aux);
8107 return ret;
8108 }
8109
8110 static void adjust_btf_func(struct bpf_verifier_env *env)
8111 {
8112 struct bpf_prog_aux *aux = env->prog->aux;
8113 int i;
8114
8115 if (!aux->func_info)
8116 return;
8117
8118 for (i = 0; i < env->subprog_cnt; i++)
8119 aux->func_info[i].insn_off = env->subprog_info[i].start;
8120 }
8121
8122 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
8123 sizeof(((struct bpf_line_info *)(0))->line_col))
8124 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
8125
8126 static int check_btf_line(struct bpf_verifier_env *env,
8127 const union bpf_attr *attr,
8128 union bpf_attr __user *uattr)
8129 {
8130 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
8131 struct bpf_subprog_info *sub;
8132 struct bpf_line_info *linfo;
8133 struct bpf_prog *prog;
8134 const struct btf *btf;
8135 void __user *ulinfo;
8136 int err;
8137
8138 nr_linfo = attr->line_info_cnt;
8139 if (!nr_linfo)
8140 return 0;
8141
8142 rec_size = attr->line_info_rec_size;
8143 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
8144 rec_size > MAX_LINEINFO_REC_SIZE ||
8145 rec_size & (sizeof(u32) - 1))
8146 return -EINVAL;
8147
8148 /* Need to zero it in case the userspace may
8149 * pass in a smaller bpf_line_info object.
8150 */
8151 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
8152 GFP_KERNEL | __GFP_NOWARN);
8153 if (!linfo)
8154 return -ENOMEM;
8155
8156 prog = env->prog;
8157 btf = prog->aux->btf;
8158
8159 s = 0;
8160 sub = env->subprog_info;
8161 ulinfo = u64_to_user_ptr(attr->line_info);
8162 expected_size = sizeof(struct bpf_line_info);
8163 ncopy = min_t(u32, expected_size, rec_size);
8164 for (i = 0; i < nr_linfo; i++) {
8165 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
8166 if (err) {
8167 if (err == -E2BIG) {
8168 verbose(env, "nonzero tailing record in line_info");
8169 if (put_user(expected_size,
8170 &uattr->line_info_rec_size))
8171 err = -EFAULT;
8172 }
8173 goto err_free;
8174 }
8175
8176 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
8177 err = -EFAULT;
8178 goto err_free;
8179 }
8180
8181 /*
8182 * Check insn_off to ensure
8183 * 1) strictly increasing AND
8184 * 2) bounded by prog->len
8185 *
8186 * The linfo[0].insn_off == 0 check logically falls into
8187 * the later "missing bpf_line_info for func..." case
8188 * because the first linfo[0].insn_off must be the
8189 * first sub also and the first sub must have
8190 * subprog_info[0].start == 0.
8191 */
8192 if ((i && linfo[i].insn_off <= prev_offset) ||
8193 linfo[i].insn_off >= prog->len) {
8194 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
8195 i, linfo[i].insn_off, prev_offset,
8196 prog->len);
8197 err = -EINVAL;
8198 goto err_free;
8199 }
8200
8201 if (!prog->insnsi[linfo[i].insn_off].code) {
8202 verbose(env,
8203 "Invalid insn code at line_info[%u].insn_off\n",
8204 i);
8205 err = -EINVAL;
8206 goto err_free;
8207 }
8208
8209 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
8210 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
8211 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
8212 err = -EINVAL;
8213 goto err_free;
8214 }
8215
8216 if (s != env->subprog_cnt) {
8217 if (linfo[i].insn_off == sub[s].start) {
8218 sub[s].linfo_idx = i;
8219 s++;
8220 } else if (sub[s].start < linfo[i].insn_off) {
8221 verbose(env, "missing bpf_line_info for func#%u\n", s);
8222 err = -EINVAL;
8223 goto err_free;
8224 }
8225 }
8226
8227 prev_offset = linfo[i].insn_off;
8228 ulinfo += rec_size;
8229 }
8230
8231 if (s != env->subprog_cnt) {
8232 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
8233 env->subprog_cnt - s, s);
8234 err = -EINVAL;
8235 goto err_free;
8236 }
8237
8238 prog->aux->linfo = linfo;
8239 prog->aux->nr_linfo = nr_linfo;
8240
8241 return 0;
8242
8243 err_free:
8244 kvfree(linfo);
8245 return err;
8246 }
8247
8248 static int check_btf_info(struct bpf_verifier_env *env,
8249 const union bpf_attr *attr,
8250 union bpf_attr __user *uattr)
8251 {
8252 struct btf *btf;
8253 int err;
8254
8255 if (!attr->func_info_cnt && !attr->line_info_cnt) {
8256 if (check_abnormal_return(env))
8257 return -EINVAL;
8258 return 0;
8259 }
8260
8261 btf = btf_get_by_fd(attr->prog_btf_fd);
8262 if (IS_ERR(btf))
8263 return PTR_ERR(btf);
8264 env->prog->aux->btf = btf;
8265
8266 err = check_btf_func(env, attr, uattr);
8267 if (err)
8268 return err;
8269
8270 err = check_btf_line(env, attr, uattr);
8271 if (err)
8272 return err;
8273
8274 return 0;
8275 }
8276
8277 /* check %cur's range satisfies %old's */
8278 static bool range_within(struct bpf_reg_state *old,
8279 struct bpf_reg_state *cur)
8280 {
8281 return old->umin_value <= cur->umin_value &&
8282 old->umax_value >= cur->umax_value &&
8283 old->smin_value <= cur->smin_value &&
8284 old->smax_value >= cur->smax_value;
8285 }
8286
8287 /* Maximum number of register states that can exist at once */
8288 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
8289 struct idpair {
8290 u32 old;
8291 u32 cur;
8292 };
8293
8294 /* If in the old state two registers had the same id, then they need to have
8295 * the same id in the new state as well. But that id could be different from
8296 * the old state, so we need to track the mapping from old to new ids.
8297 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
8298 * regs with old id 5 must also have new id 9 for the new state to be safe. But
8299 * regs with a different old id could still have new id 9, we don't care about
8300 * that.
8301 * So we look through our idmap to see if this old id has been seen before. If
8302 * so, we require the new id to match; otherwise, we add the id pair to the map.
8303 */
8304 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
8305 {
8306 unsigned int i;
8307
8308 for (i = 0; i < ID_MAP_SIZE; i++) {
8309 if (!idmap[i].old) {
8310 /* Reached an empty slot; haven't seen this id before */
8311 idmap[i].old = old_id;
8312 idmap[i].cur = cur_id;
8313 return true;
8314 }
8315 if (idmap[i].old == old_id)
8316 return idmap[i].cur == cur_id;
8317 }
8318 /* We ran out of idmap slots, which should be impossible */
8319 WARN_ON_ONCE(1);
8320 return false;
8321 }
8322
8323 static void clean_func_state(struct bpf_verifier_env *env,
8324 struct bpf_func_state *st)
8325 {
8326 enum bpf_reg_liveness live;
8327 int i, j;
8328
8329 for (i = 0; i < BPF_REG_FP; i++) {
8330 live = st->regs[i].live;
8331 /* liveness must not touch this register anymore */
8332 st->regs[i].live |= REG_LIVE_DONE;
8333 if (!(live & REG_LIVE_READ))
8334 /* since the register is unused, clear its state
8335 * to make further comparison simpler
8336 */
8337 __mark_reg_not_init(env, &st->regs[i]);
8338 }
8339
8340 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
8341 live = st->stack[i].spilled_ptr.live;
8342 /* liveness must not touch this stack slot anymore */
8343 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
8344 if (!(live & REG_LIVE_READ)) {
8345 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
8346 for (j = 0; j < BPF_REG_SIZE; j++)
8347 st->stack[i].slot_type[j] = STACK_INVALID;
8348 }
8349 }
8350 }
8351
8352 static void clean_verifier_state(struct bpf_verifier_env *env,
8353 struct bpf_verifier_state *st)
8354 {
8355 int i;
8356
8357 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
8358 /* all regs in this state in all frames were already marked */
8359 return;
8360
8361 for (i = 0; i <= st->curframe; i++)
8362 clean_func_state(env, st->frame[i]);
8363 }
8364
8365 /* the parentage chains form a tree.
8366 * the verifier states are added to state lists at given insn and
8367 * pushed into state stack for future exploration.
8368 * when the verifier reaches bpf_exit insn some of the verifer states
8369 * stored in the state lists have their final liveness state already,
8370 * but a lot of states will get revised from liveness point of view when
8371 * the verifier explores other branches.
8372 * Example:
8373 * 1: r0 = 1
8374 * 2: if r1 == 100 goto pc+1
8375 * 3: r0 = 2
8376 * 4: exit
8377 * when the verifier reaches exit insn the register r0 in the state list of
8378 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
8379 * of insn 2 and goes exploring further. At the insn 4 it will walk the
8380 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
8381 *
8382 * Since the verifier pushes the branch states as it sees them while exploring
8383 * the program the condition of walking the branch instruction for the second
8384 * time means that all states below this branch were already explored and
8385 * their final liveness markes are already propagated.
8386 * Hence when the verifier completes the search of state list in is_state_visited()
8387 * we can call this clean_live_states() function to mark all liveness states
8388 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
8389 * will not be used.
8390 * This function also clears the registers and stack for states that !READ
8391 * to simplify state merging.
8392 *
8393 * Important note here that walking the same branch instruction in the callee
8394 * doesn't meant that the states are DONE. The verifier has to compare
8395 * the callsites
8396 */
8397 static void clean_live_states(struct bpf_verifier_env *env, int insn,
8398 struct bpf_verifier_state *cur)
8399 {
8400 struct bpf_verifier_state_list *sl;
8401 int i;
8402
8403 sl = *explored_state(env, insn);
8404 while (sl) {
8405 if (sl->state.branches)
8406 goto next;
8407 if (sl->state.insn_idx != insn ||
8408 sl->state.curframe != cur->curframe)
8409 goto next;
8410 for (i = 0; i <= cur->curframe; i++)
8411 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
8412 goto next;
8413 clean_verifier_state(env, &sl->state);
8414 next:
8415 sl = sl->next;
8416 }
8417 }
8418
8419 /* Returns true if (rold safe implies rcur safe) */
8420 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
8421 struct idpair *idmap)
8422 {
8423 bool equal;
8424
8425 if (!(rold->live & REG_LIVE_READ))
8426 /* explored state didn't use this */
8427 return true;
8428
8429 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
8430
8431 if (rold->type == PTR_TO_STACK)
8432 /* two stack pointers are equal only if they're pointing to
8433 * the same stack frame, since fp-8 in foo != fp-8 in bar
8434 */
8435 return equal && rold->frameno == rcur->frameno;
8436
8437 if (equal)
8438 return true;
8439
8440 if (rold->type == NOT_INIT)
8441 /* explored state can't have used this */
8442 return true;
8443 if (rcur->type == NOT_INIT)
8444 return false;
8445 switch (rold->type) {
8446 case SCALAR_VALUE:
8447 if (rcur->type == SCALAR_VALUE) {
8448 if (!rold->precise && !rcur->precise)
8449 return true;
8450 /* new val must satisfy old val knowledge */
8451 return range_within(rold, rcur) &&
8452 tnum_in(rold->var_off, rcur->var_off);
8453 } else {
8454 /* We're trying to use a pointer in place of a scalar.
8455 * Even if the scalar was unbounded, this could lead to
8456 * pointer leaks because scalars are allowed to leak
8457 * while pointers are not. We could make this safe in
8458 * special cases if root is calling us, but it's
8459 * probably not worth the hassle.
8460 */
8461 return false;
8462 }
8463 case PTR_TO_MAP_VALUE:
8464 /* If the new min/max/var_off satisfy the old ones and
8465 * everything else matches, we are OK.
8466 * 'id' is not compared, since it's only used for maps with
8467 * bpf_spin_lock inside map element and in such cases if
8468 * the rest of the prog is valid for one map element then
8469 * it's valid for all map elements regardless of the key
8470 * used in bpf_map_lookup()
8471 */
8472 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
8473 range_within(rold, rcur) &&
8474 tnum_in(rold->var_off, rcur->var_off);
8475 case PTR_TO_MAP_VALUE_OR_NULL:
8476 /* a PTR_TO_MAP_VALUE could be safe to use as a
8477 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
8478 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
8479 * checked, doing so could have affected others with the same
8480 * id, and we can't check for that because we lost the id when
8481 * we converted to a PTR_TO_MAP_VALUE.
8482 */
8483 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
8484 return false;
8485 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
8486 return false;
8487 /* Check our ids match any regs they're supposed to */
8488 return check_ids(rold->id, rcur->id, idmap);
8489 case PTR_TO_PACKET_META:
8490 case PTR_TO_PACKET:
8491 if (rcur->type != rold->type)
8492 return false;
8493 /* We must have at least as much range as the old ptr
8494 * did, so that any accesses which were safe before are
8495 * still safe. This is true even if old range < old off,
8496 * since someone could have accessed through (ptr - k), or
8497 * even done ptr -= k in a register, to get a safe access.
8498 */
8499 if (rold->range > rcur->range)
8500 return false;
8501 /* If the offsets don't match, we can't trust our alignment;
8502 * nor can we be sure that we won't fall out of range.
8503 */
8504 if (rold->off != rcur->off)
8505 return false;
8506 /* id relations must be preserved */
8507 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
8508 return false;
8509 /* new val must satisfy old val knowledge */
8510 return range_within(rold, rcur) &&
8511 tnum_in(rold->var_off, rcur->var_off);
8512 case PTR_TO_CTX:
8513 case CONST_PTR_TO_MAP:
8514 case PTR_TO_PACKET_END:
8515 case PTR_TO_FLOW_KEYS:
8516 case PTR_TO_SOCKET:
8517 case PTR_TO_SOCKET_OR_NULL:
8518 case PTR_TO_SOCK_COMMON:
8519 case PTR_TO_SOCK_COMMON_OR_NULL:
8520 case PTR_TO_TCP_SOCK:
8521 case PTR_TO_TCP_SOCK_OR_NULL:
8522 case PTR_TO_XDP_SOCK:
8523 /* Only valid matches are exact, which memcmp() above
8524 * would have accepted
8525 */
8526 default:
8527 /* Don't know what's going on, just say it's not safe */
8528 return false;
8529 }
8530
8531 /* Shouldn't get here; if we do, say it's not safe */
8532 WARN_ON_ONCE(1);
8533 return false;
8534 }
8535
8536 static bool stacksafe(struct bpf_func_state *old,
8537 struct bpf_func_state *cur,
8538 struct idpair *idmap)
8539 {
8540 int i, spi;
8541
8542 /* walk slots of the explored stack and ignore any additional
8543 * slots in the current stack, since explored(safe) state
8544 * didn't use them
8545 */
8546 for (i = 0; i < old->allocated_stack; i++) {
8547 spi = i / BPF_REG_SIZE;
8548
8549 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
8550 i += BPF_REG_SIZE - 1;
8551 /* explored state didn't use this */
8552 continue;
8553 }
8554
8555 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
8556 continue;
8557
8558 /* explored stack has more populated slots than current stack
8559 * and these slots were used
8560 */
8561 if (i >= cur->allocated_stack)
8562 return false;
8563
8564 /* if old state was safe with misc data in the stack
8565 * it will be safe with zero-initialized stack.
8566 * The opposite is not true
8567 */
8568 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
8569 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
8570 continue;
8571 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
8572 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
8573 /* Ex: old explored (safe) state has STACK_SPILL in
8574 * this stack slot, but current has STACK_MISC ->
8575 * this verifier states are not equivalent,
8576 * return false to continue verification of this path
8577 */
8578 return false;
8579 if (i % BPF_REG_SIZE)
8580 continue;
8581 if (old->stack[spi].slot_type[0] != STACK_SPILL)
8582 continue;
8583 if (!regsafe(&old->stack[spi].spilled_ptr,
8584 &cur->stack[spi].spilled_ptr,
8585 idmap))
8586 /* when explored and current stack slot are both storing
8587 * spilled registers, check that stored pointers types
8588 * are the same as well.
8589 * Ex: explored safe path could have stored
8590 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
8591 * but current path has stored:
8592 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
8593 * such verifier states are not equivalent.
8594 * return false to continue verification of this path
8595 */
8596 return false;
8597 }
8598 return true;
8599 }
8600
8601 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
8602 {
8603 if (old->acquired_refs != cur->acquired_refs)
8604 return false;
8605 return !memcmp(old->refs, cur->refs,
8606 sizeof(*old->refs) * old->acquired_refs);
8607 }
8608
8609 /* compare two verifier states
8610 *
8611 * all states stored in state_list are known to be valid, since
8612 * verifier reached 'bpf_exit' instruction through them
8613 *
8614 * this function is called when verifier exploring different branches of
8615 * execution popped from the state stack. If it sees an old state that has
8616 * more strict register state and more strict stack state then this execution
8617 * branch doesn't need to be explored further, since verifier already
8618 * concluded that more strict state leads to valid finish.
8619 *
8620 * Therefore two states are equivalent if register state is more conservative
8621 * and explored stack state is more conservative than the current one.
8622 * Example:
8623 * explored current
8624 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
8625 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
8626 *
8627 * In other words if current stack state (one being explored) has more
8628 * valid slots than old one that already passed validation, it means
8629 * the verifier can stop exploring and conclude that current state is valid too
8630 *
8631 * Similarly with registers. If explored state has register type as invalid
8632 * whereas register type in current state is meaningful, it means that
8633 * the current state will reach 'bpf_exit' instruction safely
8634 */
8635 static bool func_states_equal(struct bpf_func_state *old,
8636 struct bpf_func_state *cur)
8637 {
8638 struct idpair *idmap;
8639 bool ret = false;
8640 int i;
8641
8642 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
8643 /* If we failed to allocate the idmap, just say it's not safe */
8644 if (!idmap)
8645 return false;
8646
8647 for (i = 0; i < MAX_BPF_REG; i++) {
8648 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
8649 goto out_free;
8650 }
8651
8652 if (!stacksafe(old, cur, idmap))
8653 goto out_free;
8654
8655 if (!refsafe(old, cur))
8656 goto out_free;
8657 ret = true;
8658 out_free:
8659 kfree(idmap);
8660 return ret;
8661 }
8662
8663 static bool states_equal(struct bpf_verifier_env *env,
8664 struct bpf_verifier_state *old,
8665 struct bpf_verifier_state *cur)
8666 {
8667 int i;
8668
8669 if (old->curframe != cur->curframe)
8670 return false;
8671
8672 /* Verification state from speculative execution simulation
8673 * must never prune a non-speculative execution one.
8674 */
8675 if (old->speculative && !cur->speculative)
8676 return false;
8677
8678 if (old->active_spin_lock != cur->active_spin_lock)
8679 return false;
8680
8681 /* for states to be equal callsites have to be the same
8682 * and all frame states need to be equivalent
8683 */
8684 for (i = 0; i <= old->curframe; i++) {
8685 if (old->frame[i]->callsite != cur->frame[i]->callsite)
8686 return false;
8687 if (!func_states_equal(old->frame[i], cur->frame[i]))
8688 return false;
8689 }
8690 return true;
8691 }
8692
8693 /* Return 0 if no propagation happened. Return negative error code if error
8694 * happened. Otherwise, return the propagated bit.
8695 */
8696 static int propagate_liveness_reg(struct bpf_verifier_env *env,
8697 struct bpf_reg_state *reg,
8698 struct bpf_reg_state *parent_reg)
8699 {
8700 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
8701 u8 flag = reg->live & REG_LIVE_READ;
8702 int err;
8703
8704 /* When comes here, read flags of PARENT_REG or REG could be any of
8705 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
8706 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
8707 */
8708 if (parent_flag == REG_LIVE_READ64 ||
8709 /* Or if there is no read flag from REG. */
8710 !flag ||
8711 /* Or if the read flag from REG is the same as PARENT_REG. */
8712 parent_flag == flag)
8713 return 0;
8714
8715 err = mark_reg_read(env, reg, parent_reg, flag);
8716 if (err)
8717 return err;
8718
8719 return flag;
8720 }
8721
8722 /* A write screens off any subsequent reads; but write marks come from the
8723 * straight-line code between a state and its parent. When we arrive at an
8724 * equivalent state (jump target or such) we didn't arrive by the straight-line
8725 * code, so read marks in the state must propagate to the parent regardless
8726 * of the state's write marks. That's what 'parent == state->parent' comparison
8727 * in mark_reg_read() is for.
8728 */
8729 static int propagate_liveness(struct bpf_verifier_env *env,
8730 const struct bpf_verifier_state *vstate,
8731 struct bpf_verifier_state *vparent)
8732 {
8733 struct bpf_reg_state *state_reg, *parent_reg;
8734 struct bpf_func_state *state, *parent;
8735 int i, frame, err = 0;
8736
8737 if (vparent->curframe != vstate->curframe) {
8738 WARN(1, "propagate_live: parent frame %d current frame %d\n",
8739 vparent->curframe, vstate->curframe);
8740 return -EFAULT;
8741 }
8742 /* Propagate read liveness of registers... */
8743 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
8744 for (frame = 0; frame <= vstate->curframe; frame++) {
8745 parent = vparent->frame[frame];
8746 state = vstate->frame[frame];
8747 parent_reg = parent->regs;
8748 state_reg = state->regs;
8749 /* We don't need to worry about FP liveness, it's read-only */
8750 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
8751 err = propagate_liveness_reg(env, &state_reg[i],
8752 &parent_reg[i]);
8753 if (err < 0)
8754 return err;
8755 if (err == REG_LIVE_READ64)
8756 mark_insn_zext(env, &parent_reg[i]);
8757 }
8758
8759 /* Propagate stack slots. */
8760 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
8761 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
8762 parent_reg = &parent->stack[i].spilled_ptr;
8763 state_reg = &state->stack[i].spilled_ptr;
8764 err = propagate_liveness_reg(env, state_reg,
8765 parent_reg);
8766 if (err < 0)
8767 return err;
8768 }
8769 }
8770 return 0;
8771 }
8772
8773 /* find precise scalars in the previous equivalent state and
8774 * propagate them into the current state
8775 */
8776 static int propagate_precision(struct bpf_verifier_env *env,
8777 const struct bpf_verifier_state *old)
8778 {
8779 struct bpf_reg_state *state_reg;
8780 struct bpf_func_state *state;
8781 int i, err = 0;
8782
8783 state = old->frame[old->curframe];
8784 state_reg = state->regs;
8785 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
8786 if (state_reg->type != SCALAR_VALUE ||
8787 !state_reg->precise)
8788 continue;
8789 if (env->log.level & BPF_LOG_LEVEL2)
8790 verbose(env, "propagating r%d\n", i);
8791 err = mark_chain_precision(env, i);
8792 if (err < 0)
8793 return err;
8794 }
8795
8796 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
8797 if (state->stack[i].slot_type[0] != STACK_SPILL)
8798 continue;
8799 state_reg = &state->stack[i].spilled_ptr;
8800 if (state_reg->type != SCALAR_VALUE ||
8801 !state_reg->precise)
8802 continue;
8803 if (env->log.level & BPF_LOG_LEVEL2)
8804 verbose(env, "propagating fp%d\n",
8805 (-i - 1) * BPF_REG_SIZE);
8806 err = mark_chain_precision_stack(env, i);
8807 if (err < 0)
8808 return err;
8809 }
8810 return 0;
8811 }
8812
8813 static bool states_maybe_looping(struct bpf_verifier_state *old,
8814 struct bpf_verifier_state *cur)
8815 {
8816 struct bpf_func_state *fold, *fcur;
8817 int i, fr = cur->curframe;
8818
8819 if (old->curframe != fr)
8820 return false;
8821
8822 fold = old->frame[fr];
8823 fcur = cur->frame[fr];
8824 for (i = 0; i < MAX_BPF_REG; i++)
8825 if (memcmp(&fold->regs[i], &fcur->regs[i],
8826 offsetof(struct bpf_reg_state, parent)))
8827 return false;
8828 return true;
8829 }
8830
8831
8832 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
8833 {
8834 struct bpf_verifier_state_list *new_sl;
8835 struct bpf_verifier_state_list *sl, **pprev;
8836 struct bpf_verifier_state *cur = env->cur_state, *new;
8837 int i, j, err, states_cnt = 0;
8838 bool add_new_state = env->test_state_freq ? true : false;
8839
8840 cur->last_insn_idx = env->prev_insn_idx;
8841 if (!env->insn_aux_data[insn_idx].prune_point)
8842 /* this 'insn_idx' instruction wasn't marked, so we will not
8843 * be doing state search here
8844 */
8845 return 0;
8846
8847 /* bpf progs typically have pruning point every 4 instructions
8848 * http://vger.kernel.org/bpfconf2019.html#session-1
8849 * Do not add new state for future pruning if the verifier hasn't seen
8850 * at least 2 jumps and at least 8 instructions.
8851 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
8852 * In tests that amounts to up to 50% reduction into total verifier
8853 * memory consumption and 20% verifier time speedup.
8854 */
8855 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
8856 env->insn_processed - env->prev_insn_processed >= 8)
8857 add_new_state = true;
8858
8859 pprev = explored_state(env, insn_idx);
8860 sl = *pprev;
8861
8862 clean_live_states(env, insn_idx, cur);
8863
8864 while (sl) {
8865 states_cnt++;
8866 if (sl->state.insn_idx != insn_idx)
8867 goto next;
8868 if (sl->state.branches) {
8869 if (states_maybe_looping(&sl->state, cur) &&
8870 states_equal(env, &sl->state, cur)) {
8871 verbose_linfo(env, insn_idx, "; ");
8872 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
8873 return -EINVAL;
8874 }
8875 /* if the verifier is processing a loop, avoid adding new state
8876 * too often, since different loop iterations have distinct
8877 * states and may not help future pruning.
8878 * This threshold shouldn't be too low to make sure that
8879 * a loop with large bound will be rejected quickly.
8880 * The most abusive loop will be:
8881 * r1 += 1
8882 * if r1 < 1000000 goto pc-2
8883 * 1M insn_procssed limit / 100 == 10k peak states.
8884 * This threshold shouldn't be too high either, since states
8885 * at the end of the loop are likely to be useful in pruning.
8886 */
8887 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
8888 env->insn_processed - env->prev_insn_processed < 100)
8889 add_new_state = false;
8890 goto miss;
8891 }
8892 if (states_equal(env, &sl->state, cur)) {
8893 sl->hit_cnt++;
8894 /* reached equivalent register/stack state,
8895 * prune the search.
8896 * Registers read by the continuation are read by us.
8897 * If we have any write marks in env->cur_state, they
8898 * will prevent corresponding reads in the continuation
8899 * from reaching our parent (an explored_state). Our
8900 * own state will get the read marks recorded, but
8901 * they'll be immediately forgotten as we're pruning
8902 * this state and will pop a new one.
8903 */
8904 err = propagate_liveness(env, &sl->state, cur);
8905
8906 /* if previous state reached the exit with precision and
8907 * current state is equivalent to it (except precsion marks)
8908 * the precision needs to be propagated back in
8909 * the current state.
8910 */
8911 err = err ? : push_jmp_history(env, cur);
8912 err = err ? : propagate_precision(env, &sl->state);
8913 if (err)
8914 return err;
8915 return 1;
8916 }
8917 miss:
8918 /* when new state is not going to be added do not increase miss count.
8919 * Otherwise several loop iterations will remove the state
8920 * recorded earlier. The goal of these heuristics is to have
8921 * states from some iterations of the loop (some in the beginning
8922 * and some at the end) to help pruning.
8923 */
8924 if (add_new_state)
8925 sl->miss_cnt++;
8926 /* heuristic to determine whether this state is beneficial
8927 * to keep checking from state equivalence point of view.
8928 * Higher numbers increase max_states_per_insn and verification time,
8929 * but do not meaningfully decrease insn_processed.
8930 */
8931 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
8932 /* the state is unlikely to be useful. Remove it to
8933 * speed up verification
8934 */
8935 *pprev = sl->next;
8936 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
8937 u32 br = sl->state.branches;
8938
8939 WARN_ONCE(br,
8940 "BUG live_done but branches_to_explore %d\n",
8941 br);
8942 free_verifier_state(&sl->state, false);
8943 kfree(sl);
8944 env->peak_states--;
8945 } else {
8946 /* cannot free this state, since parentage chain may
8947 * walk it later. Add it for free_list instead to
8948 * be freed at the end of verification
8949 */
8950 sl->next = env->free_list;
8951 env->free_list = sl;
8952 }
8953 sl = *pprev;
8954 continue;
8955 }
8956 next:
8957 pprev = &sl->next;
8958 sl = *pprev;
8959 }
8960
8961 if (env->max_states_per_insn < states_cnt)
8962 env->max_states_per_insn = states_cnt;
8963
8964 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
8965 return push_jmp_history(env, cur);
8966
8967 if (!add_new_state)
8968 return push_jmp_history(env, cur);
8969
8970 /* There were no equivalent states, remember the current one.
8971 * Technically the current state is not proven to be safe yet,
8972 * but it will either reach outer most bpf_exit (which means it's safe)
8973 * or it will be rejected. When there are no loops the verifier won't be
8974 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
8975 * again on the way to bpf_exit.
8976 * When looping the sl->state.branches will be > 0 and this state
8977 * will not be considered for equivalence until branches == 0.
8978 */
8979 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
8980 if (!new_sl)
8981 return -ENOMEM;
8982 env->total_states++;
8983 env->peak_states++;
8984 env->prev_jmps_processed = env->jmps_processed;
8985 env->prev_insn_processed = env->insn_processed;
8986
8987 /* add new state to the head of linked list */
8988 new = &new_sl->state;
8989 err = copy_verifier_state(new, cur);
8990 if (err) {
8991 free_verifier_state(new, false);
8992 kfree(new_sl);
8993 return err;
8994 }
8995 new->insn_idx = insn_idx;
8996 WARN_ONCE(new->branches != 1,
8997 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
8998
8999 cur->parent = new;
9000 cur->first_insn_idx = insn_idx;
9001 clear_jmp_history(cur);
9002 new_sl->next = *explored_state(env, insn_idx);
9003 *explored_state(env, insn_idx) = new_sl;
9004 /* connect new state to parentage chain. Current frame needs all
9005 * registers connected. Only r6 - r9 of the callers are alive (pushed
9006 * to the stack implicitly by JITs) so in callers' frames connect just
9007 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
9008 * the state of the call instruction (with WRITTEN set), and r0 comes
9009 * from callee with its full parentage chain, anyway.
9010 */
9011 /* clear write marks in current state: the writes we did are not writes
9012 * our child did, so they don't screen off its reads from us.
9013 * (There are no read marks in current state, because reads always mark
9014 * their parent and current state never has children yet. Only
9015 * explored_states can get read marks.)
9016 */
9017 for (j = 0; j <= cur->curframe; j++) {
9018 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
9019 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
9020 for (i = 0; i < BPF_REG_FP; i++)
9021 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
9022 }
9023
9024 /* all stack frames are accessible from callee, clear them all */
9025 for (j = 0; j <= cur->curframe; j++) {
9026 struct bpf_func_state *frame = cur->frame[j];
9027 struct bpf_func_state *newframe = new->frame[j];
9028
9029 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
9030 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
9031 frame->stack[i].spilled_ptr.parent =
9032 &newframe->stack[i].spilled_ptr;
9033 }
9034 }
9035 return 0;
9036 }
9037
9038 /* Return true if it's OK to have the same insn return a different type. */
9039 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
9040 {
9041 switch (type) {
9042 case PTR_TO_CTX:
9043 case PTR_TO_SOCKET:
9044 case PTR_TO_SOCKET_OR_NULL:
9045 case PTR_TO_SOCK_COMMON:
9046 case PTR_TO_SOCK_COMMON_OR_NULL:
9047 case PTR_TO_TCP_SOCK:
9048 case PTR_TO_TCP_SOCK_OR_NULL:
9049 case PTR_TO_XDP_SOCK:
9050 case PTR_TO_BTF_ID:
9051 case PTR_TO_BTF_ID_OR_NULL:
9052 return false;
9053 default:
9054 return true;
9055 }
9056 }
9057
9058 /* If an instruction was previously used with particular pointer types, then we
9059 * need to be careful to avoid cases such as the below, where it may be ok
9060 * for one branch accessing the pointer, but not ok for the other branch:
9061 *
9062 * R1 = sock_ptr
9063 * goto X;
9064 * ...
9065 * R1 = some_other_valid_ptr;
9066 * goto X;
9067 * ...
9068 * R2 = *(u32 *)(R1 + 0);
9069 */
9070 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
9071 {
9072 return src != prev && (!reg_type_mismatch_ok(src) ||
9073 !reg_type_mismatch_ok(prev));
9074 }
9075
9076 static int do_check(struct bpf_verifier_env *env)
9077 {
9078 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
9079 struct bpf_verifier_state *state = env->cur_state;
9080 struct bpf_insn *insns = env->prog->insnsi;
9081 struct bpf_reg_state *regs;
9082 int insn_cnt = env->prog->len;
9083 bool do_print_state = false;
9084 int prev_insn_idx = -1;
9085
9086 for (;;) {
9087 struct bpf_insn *insn;
9088 u8 class;
9089 int err;
9090
9091 env->prev_insn_idx = prev_insn_idx;
9092 if (env->insn_idx >= insn_cnt) {
9093 verbose(env, "invalid insn idx %d insn_cnt %d\n",
9094 env->insn_idx, insn_cnt);
9095 return -EFAULT;
9096 }
9097
9098 insn = &insns[env->insn_idx];
9099 class = BPF_CLASS(insn->code);
9100
9101 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
9102 verbose(env,
9103 "BPF program is too large. Processed %d insn\n",
9104 env->insn_processed);
9105 return -E2BIG;
9106 }
9107
9108 err = is_state_visited(env, env->insn_idx);
9109 if (err < 0)
9110 return err;
9111 if (err == 1) {
9112 /* found equivalent state, can prune the search */
9113 if (env->log.level & BPF_LOG_LEVEL) {
9114 if (do_print_state)
9115 verbose(env, "\nfrom %d to %d%s: safe\n",
9116 env->prev_insn_idx, env->insn_idx,
9117 env->cur_state->speculative ?
9118 " (speculative execution)" : "");
9119 else
9120 verbose(env, "%d: safe\n", env->insn_idx);
9121 }
9122 goto process_bpf_exit;
9123 }
9124
9125 if (signal_pending(current))
9126 return -EAGAIN;
9127
9128 if (need_resched())
9129 cond_resched();
9130
9131 if (env->log.level & BPF_LOG_LEVEL2 ||
9132 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
9133 if (env->log.level & BPF_LOG_LEVEL2)
9134 verbose(env, "%d:", env->insn_idx);
9135 else
9136 verbose(env, "\nfrom %d to %d%s:",
9137 env->prev_insn_idx, env->insn_idx,
9138 env->cur_state->speculative ?
9139 " (speculative execution)" : "");
9140 print_verifier_state(env, state->frame[state->curframe]);
9141 do_print_state = false;
9142 }
9143
9144 if (env->log.level & BPF_LOG_LEVEL) {
9145 const struct bpf_insn_cbs cbs = {
9146 .cb_print = verbose,
9147 .private_data = env,
9148 };
9149
9150 verbose_linfo(env, env->insn_idx, "; ");
9151 verbose(env, "%d: ", env->insn_idx);
9152 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
9153 }
9154
9155 if (bpf_prog_is_dev_bound(env->prog->aux)) {
9156 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
9157 env->prev_insn_idx);
9158 if (err)
9159 return err;
9160 }
9161
9162 regs = cur_regs(env);
9163 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9164 prev_insn_idx = env->insn_idx;
9165
9166 if (class == BPF_ALU || class == BPF_ALU64) {
9167 err = check_alu_op(env, insn);
9168 if (err)
9169 return err;
9170
9171 } else if (class == BPF_LDX) {
9172 enum bpf_reg_type *prev_src_type, src_reg_type;
9173
9174 /* check for reserved fields is already done */
9175
9176 /* check src operand */
9177 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9178 if (err)
9179 return err;
9180
9181 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
9182 if (err)
9183 return err;
9184
9185 src_reg_type = regs[insn->src_reg].type;
9186
9187 /* check that memory (src_reg + off) is readable,
9188 * the state of dst_reg will be updated by this func
9189 */
9190 err = check_mem_access(env, env->insn_idx, insn->src_reg,
9191 insn->off, BPF_SIZE(insn->code),
9192 BPF_READ, insn->dst_reg, false);
9193 if (err)
9194 return err;
9195
9196 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
9197
9198 if (*prev_src_type == NOT_INIT) {
9199 /* saw a valid insn
9200 * dst_reg = *(u32 *)(src_reg + off)
9201 * save type to validate intersecting paths
9202 */
9203 *prev_src_type = src_reg_type;
9204
9205 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
9206 /* ABuser program is trying to use the same insn
9207 * dst_reg = *(u32*) (src_reg + off)
9208 * with different pointer types:
9209 * src_reg == ctx in one branch and
9210 * src_reg == stack|map in some other branch.
9211 * Reject it.
9212 */
9213 verbose(env, "same insn cannot be used with different pointers\n");
9214 return -EINVAL;
9215 }
9216
9217 } else if (class == BPF_STX) {
9218 enum bpf_reg_type *prev_dst_type, dst_reg_type;
9219
9220 if (BPF_MODE(insn->code) == BPF_XADD) {
9221 err = check_xadd(env, env->insn_idx, insn);
9222 if (err)
9223 return err;
9224 env->insn_idx++;
9225 continue;
9226 }
9227
9228 /* check src1 operand */
9229 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9230 if (err)
9231 return err;
9232 /* check src2 operand */
9233 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9234 if (err)
9235 return err;
9236
9237 dst_reg_type = regs[insn->dst_reg].type;
9238
9239 /* check that memory (dst_reg + off) is writeable */
9240 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
9241 insn->off, BPF_SIZE(insn->code),
9242 BPF_WRITE, insn->src_reg, false);
9243 if (err)
9244 return err;
9245
9246 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
9247
9248 if (*prev_dst_type == NOT_INIT) {
9249 *prev_dst_type = dst_reg_type;
9250 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
9251 verbose(env, "same insn cannot be used with different pointers\n");
9252 return -EINVAL;
9253 }
9254
9255 } else if (class == BPF_ST) {
9256 if (BPF_MODE(insn->code) != BPF_MEM ||
9257 insn->src_reg != BPF_REG_0) {
9258 verbose(env, "BPF_ST uses reserved fields\n");
9259 return -EINVAL;
9260 }
9261 /* check src operand */
9262 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9263 if (err)
9264 return err;
9265
9266 if (is_ctx_reg(env, insn->dst_reg)) {
9267 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
9268 insn->dst_reg,
9269 reg_type_str[reg_state(env, insn->dst_reg)->type]);
9270 return -EACCES;
9271 }
9272
9273 /* check that memory (dst_reg + off) is writeable */
9274 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
9275 insn->off, BPF_SIZE(insn->code),
9276 BPF_WRITE, -1, false);
9277 if (err)
9278 return err;
9279
9280 } else if (class == BPF_JMP || class == BPF_JMP32) {
9281 u8 opcode = BPF_OP(insn->code);
9282
9283 env->jmps_processed++;
9284 if (opcode == BPF_CALL) {
9285 if (BPF_SRC(insn->code) != BPF_K ||
9286 insn->off != 0 ||
9287 (insn->src_reg != BPF_REG_0 &&
9288 insn->src_reg != BPF_PSEUDO_CALL) ||
9289 insn->dst_reg != BPF_REG_0 ||
9290 class == BPF_JMP32) {
9291 verbose(env, "BPF_CALL uses reserved fields\n");
9292 return -EINVAL;
9293 }
9294
9295 if (env->cur_state->active_spin_lock &&
9296 (insn->src_reg == BPF_PSEUDO_CALL ||
9297 insn->imm != BPF_FUNC_spin_unlock)) {
9298 verbose(env, "function calls are not allowed while holding a lock\n");
9299 return -EINVAL;
9300 }
9301 if (insn->src_reg == BPF_PSEUDO_CALL)
9302 err = check_func_call(env, insn, &env->insn_idx);
9303 else
9304 err = check_helper_call(env, insn->imm, env->insn_idx);
9305 if (err)
9306 return err;
9307
9308 } else if (opcode == BPF_JA) {
9309 if (BPF_SRC(insn->code) != BPF_K ||
9310 insn->imm != 0 ||
9311 insn->src_reg != BPF_REG_0 ||
9312 insn->dst_reg != BPF_REG_0 ||
9313 class == BPF_JMP32) {
9314 verbose(env, "BPF_JA uses reserved fields\n");
9315 return -EINVAL;
9316 }
9317
9318 env->insn_idx += insn->off + 1;
9319 continue;
9320
9321 } else if (opcode == BPF_EXIT) {
9322 if (BPF_SRC(insn->code) != BPF_K ||
9323 insn->imm != 0 ||
9324 insn->src_reg != BPF_REG_0 ||
9325 insn->dst_reg != BPF_REG_0 ||
9326 class == BPF_JMP32) {
9327 verbose(env, "BPF_EXIT uses reserved fields\n");
9328 return -EINVAL;
9329 }
9330
9331 if (env->cur_state->active_spin_lock) {
9332 verbose(env, "bpf_spin_unlock is missing\n");
9333 return -EINVAL;
9334 }
9335
9336 if (state->curframe) {
9337 /* exit from nested function */
9338 err = prepare_func_exit(env, &env->insn_idx);
9339 if (err)
9340 return err;
9341 do_print_state = true;
9342 continue;
9343 }
9344
9345 err = check_reference_leak(env);
9346 if (err)
9347 return err;
9348
9349 err = check_return_code(env);
9350 if (err)
9351 return err;
9352 process_bpf_exit:
9353 update_branch_counts(env, env->cur_state);
9354 err = pop_stack(env, &prev_insn_idx,
9355 &env->insn_idx, pop_log);
9356 if (err < 0) {
9357 if (err != -ENOENT)
9358 return err;
9359 break;
9360 } else {
9361 do_print_state = true;
9362 continue;
9363 }
9364 } else {
9365 err = check_cond_jmp_op(env, insn, &env->insn_idx);
9366 if (err)
9367 return err;
9368 }
9369 } else if (class == BPF_LD) {
9370 u8 mode = BPF_MODE(insn->code);
9371
9372 if (mode == BPF_ABS || mode == BPF_IND) {
9373 err = check_ld_abs(env, insn);
9374 if (err)
9375 return err;
9376
9377 } else if (mode == BPF_IMM) {
9378 err = check_ld_imm(env, insn);
9379 if (err)
9380 return err;
9381
9382 env->insn_idx++;
9383 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9384 } else {
9385 verbose(env, "invalid BPF_LD mode\n");
9386 return -EINVAL;
9387 }
9388 } else {
9389 verbose(env, "unknown insn class %d\n", class);
9390 return -EINVAL;
9391 }
9392
9393 env->insn_idx++;
9394 }
9395
9396 return 0;
9397 }
9398
9399 static int check_map_prealloc(struct bpf_map *map)
9400 {
9401 return (map->map_type != BPF_MAP_TYPE_HASH &&
9402 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
9403 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
9404 !(map->map_flags & BPF_F_NO_PREALLOC);
9405 }
9406
9407 static bool is_tracing_prog_type(enum bpf_prog_type type)
9408 {
9409 switch (type) {
9410 case BPF_PROG_TYPE_KPROBE:
9411 case BPF_PROG_TYPE_TRACEPOINT:
9412 case BPF_PROG_TYPE_PERF_EVENT:
9413 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9414 return true;
9415 default:
9416 return false;
9417 }
9418 }
9419
9420 static bool is_preallocated_map(struct bpf_map *map)
9421 {
9422 if (!check_map_prealloc(map))
9423 return false;
9424 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
9425 return false;
9426 return true;
9427 }
9428
9429 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
9430 struct bpf_map *map,
9431 struct bpf_prog *prog)
9432
9433 {
9434 enum bpf_prog_type prog_type = resolve_prog_type(prog);
9435 /*
9436 * Validate that trace type programs use preallocated hash maps.
9437 *
9438 * For programs attached to PERF events this is mandatory as the
9439 * perf NMI can hit any arbitrary code sequence.
9440 *
9441 * All other trace types using preallocated hash maps are unsafe as
9442 * well because tracepoint or kprobes can be inside locked regions
9443 * of the memory allocator or at a place where a recursion into the
9444 * memory allocator would see inconsistent state.
9445 *
9446 * On RT enabled kernels run-time allocation of all trace type
9447 * programs is strictly prohibited due to lock type constraints. On
9448 * !RT kernels it is allowed for backwards compatibility reasons for
9449 * now, but warnings are emitted so developers are made aware of
9450 * the unsafety and can fix their programs before this is enforced.
9451 */
9452 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
9453 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
9454 verbose(env, "perf_event programs can only use preallocated hash map\n");
9455 return -EINVAL;
9456 }
9457 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
9458 verbose(env, "trace type programs can only use preallocated hash map\n");
9459 return -EINVAL;
9460 }
9461 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
9462 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
9463 }
9464
9465 if ((is_tracing_prog_type(prog_type) ||
9466 prog_type == BPF_PROG_TYPE_SOCKET_FILTER) &&
9467 map_value_has_spin_lock(map)) {
9468 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
9469 return -EINVAL;
9470 }
9471
9472 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
9473 !bpf_offload_prog_map_match(prog, map)) {
9474 verbose(env, "offload device mismatch between prog and map\n");
9475 return -EINVAL;
9476 }
9477
9478 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
9479 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
9480 return -EINVAL;
9481 }
9482
9483 if (prog->aux->sleepable)
9484 switch (map->map_type) {
9485 case BPF_MAP_TYPE_HASH:
9486 case BPF_MAP_TYPE_LRU_HASH:
9487 case BPF_MAP_TYPE_ARRAY:
9488 if (!is_preallocated_map(map)) {
9489 verbose(env,
9490 "Sleepable programs can only use preallocated hash maps\n");
9491 return -EINVAL;
9492 }
9493 break;
9494 default:
9495 verbose(env,
9496 "Sleepable programs can only use array and hash maps\n");
9497 return -EINVAL;
9498 }
9499
9500 return 0;
9501 }
9502
9503 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
9504 {
9505 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
9506 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
9507 }
9508
9509 /* look for pseudo eBPF instructions that access map FDs and
9510 * replace them with actual map pointers
9511 */
9512 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
9513 {
9514 struct bpf_insn *insn = env->prog->insnsi;
9515 int insn_cnt = env->prog->len;
9516 int i, j, err;
9517
9518 err = bpf_prog_calc_tag(env->prog);
9519 if (err)
9520 return err;
9521
9522 for (i = 0; i < insn_cnt; i++, insn++) {
9523 if (BPF_CLASS(insn->code) == BPF_LDX &&
9524 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
9525 verbose(env, "BPF_LDX uses reserved fields\n");
9526 return -EINVAL;
9527 }
9528
9529 if (BPF_CLASS(insn->code) == BPF_STX &&
9530 ((BPF_MODE(insn->code) != BPF_MEM &&
9531 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
9532 verbose(env, "BPF_STX uses reserved fields\n");
9533 return -EINVAL;
9534 }
9535
9536 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
9537 struct bpf_insn_aux_data *aux;
9538 struct bpf_map *map;
9539 struct fd f;
9540 u64 addr;
9541
9542 if (i == insn_cnt - 1 || insn[1].code != 0 ||
9543 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
9544 insn[1].off != 0) {
9545 verbose(env, "invalid bpf_ld_imm64 insn\n");
9546 return -EINVAL;
9547 }
9548
9549 if (insn[0].src_reg == 0)
9550 /* valid generic load 64-bit imm */
9551 goto next_insn;
9552
9553 /* In final convert_pseudo_ld_imm64() step, this is
9554 * converted into regular 64-bit imm load insn.
9555 */
9556 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
9557 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
9558 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
9559 insn[1].imm != 0)) {
9560 verbose(env,
9561 "unrecognized bpf_ld_imm64 insn\n");
9562 return -EINVAL;
9563 }
9564
9565 f = fdget(insn[0].imm);
9566 map = __bpf_map_get(f);
9567 if (IS_ERR(map)) {
9568 verbose(env, "fd %d is not pointing to valid bpf_map\n",
9569 insn[0].imm);
9570 return PTR_ERR(map);
9571 }
9572
9573 err = check_map_prog_compatibility(env, map, env->prog);
9574 if (err) {
9575 fdput(f);
9576 return err;
9577 }
9578
9579 aux = &env->insn_aux_data[i];
9580 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
9581 addr = (unsigned long)map;
9582 } else {
9583 u32 off = insn[1].imm;
9584
9585 if (off >= BPF_MAX_VAR_OFF) {
9586 verbose(env, "direct value offset of %u is not allowed\n", off);
9587 fdput(f);
9588 return -EINVAL;
9589 }
9590
9591 if (!map->ops->map_direct_value_addr) {
9592 verbose(env, "no direct value access support for this map type\n");
9593 fdput(f);
9594 return -EINVAL;
9595 }
9596
9597 err = map->ops->map_direct_value_addr(map, &addr, off);
9598 if (err) {
9599 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
9600 map->value_size, off);
9601 fdput(f);
9602 return err;
9603 }
9604
9605 aux->map_off = off;
9606 addr += off;
9607 }
9608
9609 insn[0].imm = (u32)addr;
9610 insn[1].imm = addr >> 32;
9611
9612 /* check whether we recorded this map already */
9613 for (j = 0; j < env->used_map_cnt; j++) {
9614 if (env->used_maps[j] == map) {
9615 aux->map_index = j;
9616 fdput(f);
9617 goto next_insn;
9618 }
9619 }
9620
9621 if (env->used_map_cnt >= MAX_USED_MAPS) {
9622 fdput(f);
9623 return -E2BIG;
9624 }
9625
9626 /* hold the map. If the program is rejected by verifier,
9627 * the map will be released by release_maps() or it
9628 * will be used by the valid program until it's unloaded
9629 * and all maps are released in free_used_maps()
9630 */
9631 bpf_map_inc(map);
9632
9633 aux->map_index = env->used_map_cnt;
9634 env->used_maps[env->used_map_cnt++] = map;
9635
9636 if (bpf_map_is_cgroup_storage(map) &&
9637 bpf_cgroup_storage_assign(env->prog->aux, map)) {
9638 verbose(env, "only one cgroup storage of each type is allowed\n");
9639 fdput(f);
9640 return -EBUSY;
9641 }
9642
9643 fdput(f);
9644 next_insn:
9645 insn++;
9646 i++;
9647 continue;
9648 }
9649
9650 /* Basic sanity check before we invest more work here. */
9651 if (!bpf_opcode_in_insntable(insn->code)) {
9652 verbose(env, "unknown opcode %02x\n", insn->code);
9653 return -EINVAL;
9654 }
9655 }
9656
9657 /* now all pseudo BPF_LD_IMM64 instructions load valid
9658 * 'struct bpf_map *' into a register instead of user map_fd.
9659 * These pointers will be used later by verifier to validate map access.
9660 */
9661 return 0;
9662 }
9663
9664 /* drop refcnt of maps used by the rejected program */
9665 static void release_maps(struct bpf_verifier_env *env)
9666 {
9667 __bpf_free_used_maps(env->prog->aux, env->used_maps,
9668 env->used_map_cnt);
9669 }
9670
9671 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
9672 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
9673 {
9674 struct bpf_insn *insn = env->prog->insnsi;
9675 int insn_cnt = env->prog->len;
9676 int i;
9677
9678 for (i = 0; i < insn_cnt; i++, insn++)
9679 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
9680 insn->src_reg = 0;
9681 }
9682
9683 /* single env->prog->insni[off] instruction was replaced with the range
9684 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
9685 * [0, off) and [off, end) to new locations, so the patched range stays zero
9686 */
9687 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
9688 struct bpf_prog *new_prog, u32 off, u32 cnt)
9689 {
9690 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
9691 struct bpf_insn *insn = new_prog->insnsi;
9692 u32 prog_len;
9693 int i;
9694
9695 /* aux info at OFF always needs adjustment, no matter fast path
9696 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
9697 * original insn at old prog.
9698 */
9699 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
9700
9701 if (cnt == 1)
9702 return 0;
9703 prog_len = new_prog->len;
9704 new_data = vzalloc(array_size(prog_len,
9705 sizeof(struct bpf_insn_aux_data)));
9706 if (!new_data)
9707 return -ENOMEM;
9708 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
9709 memcpy(new_data + off + cnt - 1, old_data + off,
9710 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
9711 for (i = off; i < off + cnt - 1; i++) {
9712 new_data[i].seen = env->pass_cnt;
9713 new_data[i].zext_dst = insn_has_def32(env, insn + i);
9714 }
9715 env->insn_aux_data = new_data;
9716 vfree(old_data);
9717 return 0;
9718 }
9719
9720 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
9721 {
9722 int i;
9723
9724 if (len == 1)
9725 return;
9726 /* NOTE: fake 'exit' subprog should be updated as well. */
9727 for (i = 0; i <= env->subprog_cnt; i++) {
9728 if (env->subprog_info[i].start <= off)
9729 continue;
9730 env->subprog_info[i].start += len - 1;
9731 }
9732 }
9733
9734 static void adjust_poke_descs(struct bpf_prog *prog, u32 len)
9735 {
9736 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
9737 int i, sz = prog->aux->size_poke_tab;
9738 struct bpf_jit_poke_descriptor *desc;
9739
9740 for (i = 0; i < sz; i++) {
9741 desc = &tab[i];
9742 desc->insn_idx += len - 1;
9743 }
9744 }
9745
9746 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
9747 const struct bpf_insn *patch, u32 len)
9748 {
9749 struct bpf_prog *new_prog;
9750
9751 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
9752 if (IS_ERR(new_prog)) {
9753 if (PTR_ERR(new_prog) == -ERANGE)
9754 verbose(env,
9755 "insn %d cannot be patched due to 16-bit range\n",
9756 env->insn_aux_data[off].orig_idx);
9757 return NULL;
9758 }
9759 if (adjust_insn_aux_data(env, new_prog, off, len))
9760 return NULL;
9761 adjust_subprog_starts(env, off, len);
9762 adjust_poke_descs(new_prog, len);
9763 return new_prog;
9764 }
9765
9766 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
9767 u32 off, u32 cnt)
9768 {
9769 int i, j;
9770
9771 /* find first prog starting at or after off (first to remove) */
9772 for (i = 0; i < env->subprog_cnt; i++)
9773 if (env->subprog_info[i].start >= off)
9774 break;
9775 /* find first prog starting at or after off + cnt (first to stay) */
9776 for (j = i; j < env->subprog_cnt; j++)
9777 if (env->subprog_info[j].start >= off + cnt)
9778 break;
9779 /* if j doesn't start exactly at off + cnt, we are just removing
9780 * the front of previous prog
9781 */
9782 if (env->subprog_info[j].start != off + cnt)
9783 j--;
9784
9785 if (j > i) {
9786 struct bpf_prog_aux *aux = env->prog->aux;
9787 int move;
9788
9789 /* move fake 'exit' subprog as well */
9790 move = env->subprog_cnt + 1 - j;
9791
9792 memmove(env->subprog_info + i,
9793 env->subprog_info + j,
9794 sizeof(*env->subprog_info) * move);
9795 env->subprog_cnt -= j - i;
9796
9797 /* remove func_info */
9798 if (aux->func_info) {
9799 move = aux->func_info_cnt - j;
9800
9801 memmove(aux->func_info + i,
9802 aux->func_info + j,
9803 sizeof(*aux->func_info) * move);
9804 aux->func_info_cnt -= j - i;
9805 /* func_info->insn_off is set after all code rewrites,
9806 * in adjust_btf_func() - no need to adjust
9807 */
9808 }
9809 } else {
9810 /* convert i from "first prog to remove" to "first to adjust" */
9811 if (env->subprog_info[i].start == off)
9812 i++;
9813 }
9814
9815 /* update fake 'exit' subprog as well */
9816 for (; i <= env->subprog_cnt; i++)
9817 env->subprog_info[i].start -= cnt;
9818
9819 return 0;
9820 }
9821
9822 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
9823 u32 cnt)
9824 {
9825 struct bpf_prog *prog = env->prog;
9826 u32 i, l_off, l_cnt, nr_linfo;
9827 struct bpf_line_info *linfo;
9828
9829 nr_linfo = prog->aux->nr_linfo;
9830 if (!nr_linfo)
9831 return 0;
9832
9833 linfo = prog->aux->linfo;
9834
9835 /* find first line info to remove, count lines to be removed */
9836 for (i = 0; i < nr_linfo; i++)
9837 if (linfo[i].insn_off >= off)
9838 break;
9839
9840 l_off = i;
9841 l_cnt = 0;
9842 for (; i < nr_linfo; i++)
9843 if (linfo[i].insn_off < off + cnt)
9844 l_cnt++;
9845 else
9846 break;
9847
9848 /* First live insn doesn't match first live linfo, it needs to "inherit"
9849 * last removed linfo. prog is already modified, so prog->len == off
9850 * means no live instructions after (tail of the program was removed).
9851 */
9852 if (prog->len != off && l_cnt &&
9853 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
9854 l_cnt--;
9855 linfo[--i].insn_off = off + cnt;
9856 }
9857
9858 /* remove the line info which refer to the removed instructions */
9859 if (l_cnt) {
9860 memmove(linfo + l_off, linfo + i,
9861 sizeof(*linfo) * (nr_linfo - i));
9862
9863 prog->aux->nr_linfo -= l_cnt;
9864 nr_linfo = prog->aux->nr_linfo;
9865 }
9866
9867 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
9868 for (i = l_off; i < nr_linfo; i++)
9869 linfo[i].insn_off -= cnt;
9870
9871 /* fix up all subprogs (incl. 'exit') which start >= off */
9872 for (i = 0; i <= env->subprog_cnt; i++)
9873 if (env->subprog_info[i].linfo_idx > l_off) {
9874 /* program may have started in the removed region but
9875 * may not be fully removed
9876 */
9877 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
9878 env->subprog_info[i].linfo_idx -= l_cnt;
9879 else
9880 env->subprog_info[i].linfo_idx = l_off;
9881 }
9882
9883 return 0;
9884 }
9885
9886 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
9887 {
9888 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9889 unsigned int orig_prog_len = env->prog->len;
9890 int err;
9891
9892 if (bpf_prog_is_dev_bound(env->prog->aux))
9893 bpf_prog_offload_remove_insns(env, off, cnt);
9894
9895 err = bpf_remove_insns(env->prog, off, cnt);
9896 if (err)
9897 return err;
9898
9899 err = adjust_subprog_starts_after_remove(env, off, cnt);
9900 if (err)
9901 return err;
9902
9903 err = bpf_adj_linfo_after_remove(env, off, cnt);
9904 if (err)
9905 return err;
9906
9907 memmove(aux_data + off, aux_data + off + cnt,
9908 sizeof(*aux_data) * (orig_prog_len - off - cnt));
9909
9910 return 0;
9911 }
9912
9913 /* The verifier does more data flow analysis than llvm and will not
9914 * explore branches that are dead at run time. Malicious programs can
9915 * have dead code too. Therefore replace all dead at-run-time code
9916 * with 'ja -1'.
9917 *
9918 * Just nops are not optimal, e.g. if they would sit at the end of the
9919 * program and through another bug we would manage to jump there, then
9920 * we'd execute beyond program memory otherwise. Returning exception
9921 * code also wouldn't work since we can have subprogs where the dead
9922 * code could be located.
9923 */
9924 static void sanitize_dead_code(struct bpf_verifier_env *env)
9925 {
9926 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9927 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
9928 struct bpf_insn *insn = env->prog->insnsi;
9929 const int insn_cnt = env->prog->len;
9930 int i;
9931
9932 for (i = 0; i < insn_cnt; i++) {
9933 if (aux_data[i].seen)
9934 continue;
9935 memcpy(insn + i, &trap, sizeof(trap));
9936 }
9937 }
9938
9939 static bool insn_is_cond_jump(u8 code)
9940 {
9941 u8 op;
9942
9943 if (BPF_CLASS(code) == BPF_JMP32)
9944 return true;
9945
9946 if (BPF_CLASS(code) != BPF_JMP)
9947 return false;
9948
9949 op = BPF_OP(code);
9950 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
9951 }
9952
9953 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
9954 {
9955 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9956 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
9957 struct bpf_insn *insn = env->prog->insnsi;
9958 const int insn_cnt = env->prog->len;
9959 int i;
9960
9961 for (i = 0; i < insn_cnt; i++, insn++) {
9962 if (!insn_is_cond_jump(insn->code))
9963 continue;
9964
9965 if (!aux_data[i + 1].seen)
9966 ja.off = insn->off;
9967 else if (!aux_data[i + 1 + insn->off].seen)
9968 ja.off = 0;
9969 else
9970 continue;
9971
9972 if (bpf_prog_is_dev_bound(env->prog->aux))
9973 bpf_prog_offload_replace_insn(env, i, &ja);
9974
9975 memcpy(insn, &ja, sizeof(ja));
9976 }
9977 }
9978
9979 static int opt_remove_dead_code(struct bpf_verifier_env *env)
9980 {
9981 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
9982 int insn_cnt = env->prog->len;
9983 int i, err;
9984
9985 for (i = 0; i < insn_cnt; i++) {
9986 int j;
9987
9988 j = 0;
9989 while (i + j < insn_cnt && !aux_data[i + j].seen)
9990 j++;
9991 if (!j)
9992 continue;
9993
9994 err = verifier_remove_insns(env, i, j);
9995 if (err)
9996 return err;
9997 insn_cnt = env->prog->len;
9998 }
9999
10000 return 0;
10001 }
10002
10003 static int opt_remove_nops(struct bpf_verifier_env *env)
10004 {
10005 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
10006 struct bpf_insn *insn = env->prog->insnsi;
10007 int insn_cnt = env->prog->len;
10008 int i, err;
10009
10010 for (i = 0; i < insn_cnt; i++) {
10011 if (memcmp(&insn[i], &ja, sizeof(ja)))
10012 continue;
10013
10014 err = verifier_remove_insns(env, i, 1);
10015 if (err)
10016 return err;
10017 insn_cnt--;
10018 i--;
10019 }
10020
10021 return 0;
10022 }
10023
10024 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
10025 const union bpf_attr *attr)
10026 {
10027 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
10028 struct bpf_insn_aux_data *aux = env->insn_aux_data;
10029 int i, patch_len, delta = 0, len = env->prog->len;
10030 struct bpf_insn *insns = env->prog->insnsi;
10031 struct bpf_prog *new_prog;
10032 bool rnd_hi32;
10033
10034 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
10035 zext_patch[1] = BPF_ZEXT_REG(0);
10036 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
10037 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
10038 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
10039 for (i = 0; i < len; i++) {
10040 int adj_idx = i + delta;
10041 struct bpf_insn insn;
10042
10043 insn = insns[adj_idx];
10044 if (!aux[adj_idx].zext_dst) {
10045 u8 code, class;
10046 u32 imm_rnd;
10047
10048 if (!rnd_hi32)
10049 continue;
10050
10051 code = insn.code;
10052 class = BPF_CLASS(code);
10053 if (insn_no_def(&insn))
10054 continue;
10055
10056 /* NOTE: arg "reg" (the fourth one) is only used for
10057 * BPF_STX which has been ruled out in above
10058 * check, it is safe to pass NULL here.
10059 */
10060 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
10061 if (class == BPF_LD &&
10062 BPF_MODE(code) == BPF_IMM)
10063 i++;
10064 continue;
10065 }
10066
10067 /* ctx load could be transformed into wider load. */
10068 if (class == BPF_LDX &&
10069 aux[adj_idx].ptr_type == PTR_TO_CTX)
10070 continue;
10071
10072 imm_rnd = get_random_int();
10073 rnd_hi32_patch[0] = insn;
10074 rnd_hi32_patch[1].imm = imm_rnd;
10075 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
10076 patch = rnd_hi32_patch;
10077 patch_len = 4;
10078 goto apply_patch_buffer;
10079 }
10080
10081 if (!bpf_jit_needs_zext())
10082 continue;
10083
10084 zext_patch[0] = insn;
10085 zext_patch[1].dst_reg = insn.dst_reg;
10086 zext_patch[1].src_reg = insn.dst_reg;
10087 patch = zext_patch;
10088 patch_len = 2;
10089 apply_patch_buffer:
10090 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
10091 if (!new_prog)
10092 return -ENOMEM;
10093 env->prog = new_prog;
10094 insns = new_prog->insnsi;
10095 aux = env->insn_aux_data;
10096 delta += patch_len - 1;
10097 }
10098
10099 return 0;
10100 }
10101
10102 /* convert load instructions that access fields of a context type into a
10103 * sequence of instructions that access fields of the underlying structure:
10104 * struct __sk_buff -> struct sk_buff
10105 * struct bpf_sock_ops -> struct sock
10106 */
10107 static int convert_ctx_accesses(struct bpf_verifier_env *env)
10108 {
10109 const struct bpf_verifier_ops *ops = env->ops;
10110 int i, cnt, size, ctx_field_size, delta = 0;
10111 const int insn_cnt = env->prog->len;
10112 struct bpf_insn insn_buf[16], *insn;
10113 u32 target_size, size_default, off;
10114 struct bpf_prog *new_prog;
10115 enum bpf_access_type type;
10116 bool is_narrower_load;
10117
10118 if (ops->gen_prologue || env->seen_direct_write) {
10119 if (!ops->gen_prologue) {
10120 verbose(env, "bpf verifier is misconfigured\n");
10121 return -EINVAL;
10122 }
10123 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
10124 env->prog);
10125 if (cnt >= ARRAY_SIZE(insn_buf)) {
10126 verbose(env, "bpf verifier is misconfigured\n");
10127 return -EINVAL;
10128 } else if (cnt) {
10129 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
10130 if (!new_prog)
10131 return -ENOMEM;
10132
10133 env->prog = new_prog;
10134 delta += cnt - 1;
10135 }
10136 }
10137
10138 if (bpf_prog_is_dev_bound(env->prog->aux))
10139 return 0;
10140
10141 insn = env->prog->insnsi + delta;
10142
10143 for (i = 0; i < insn_cnt; i++, insn++) {
10144 bpf_convert_ctx_access_t convert_ctx_access;
10145
10146 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
10147 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
10148 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
10149 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
10150 type = BPF_READ;
10151 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
10152 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
10153 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
10154 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
10155 type = BPF_WRITE;
10156 else
10157 continue;
10158
10159 if (type == BPF_WRITE &&
10160 env->insn_aux_data[i + delta].sanitize_stack_off) {
10161 struct bpf_insn patch[] = {
10162 /* Sanitize suspicious stack slot with zero.
10163 * There are no memory dependencies for this store,
10164 * since it's only using frame pointer and immediate
10165 * constant of zero
10166 */
10167 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
10168 env->insn_aux_data[i + delta].sanitize_stack_off,
10169 0),
10170 /* the original STX instruction will immediately
10171 * overwrite the same stack slot with appropriate value
10172 */
10173 *insn,
10174 };
10175
10176 cnt = ARRAY_SIZE(patch);
10177 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
10178 if (!new_prog)
10179 return -ENOMEM;
10180
10181 delta += cnt - 1;
10182 env->prog = new_prog;
10183 insn = new_prog->insnsi + i + delta;
10184 continue;
10185 }
10186
10187 switch (env->insn_aux_data[i + delta].ptr_type) {
10188 case PTR_TO_CTX:
10189 if (!ops->convert_ctx_access)
10190 continue;
10191 convert_ctx_access = ops->convert_ctx_access;
10192 break;
10193 case PTR_TO_SOCKET:
10194 case PTR_TO_SOCK_COMMON:
10195 convert_ctx_access = bpf_sock_convert_ctx_access;
10196 break;
10197 case PTR_TO_TCP_SOCK:
10198 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
10199 break;
10200 case PTR_TO_XDP_SOCK:
10201 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
10202 break;
10203 case PTR_TO_BTF_ID:
10204 if (type == BPF_READ) {
10205 insn->code = BPF_LDX | BPF_PROBE_MEM |
10206 BPF_SIZE((insn)->code);
10207 env->prog->aux->num_exentries++;
10208 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
10209 verbose(env, "Writes through BTF pointers are not allowed\n");
10210 return -EINVAL;
10211 }
10212 continue;
10213 default:
10214 continue;
10215 }
10216
10217 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
10218 size = BPF_LDST_BYTES(insn);
10219
10220 /* If the read access is a narrower load of the field,
10221 * convert to a 4/8-byte load, to minimum program type specific
10222 * convert_ctx_access changes. If conversion is successful,
10223 * we will apply proper mask to the result.
10224 */
10225 is_narrower_load = size < ctx_field_size;
10226 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
10227 off = insn->off;
10228 if (is_narrower_load) {
10229 u8 size_code;
10230
10231 if (type == BPF_WRITE) {
10232 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
10233 return -EINVAL;
10234 }
10235
10236 size_code = BPF_H;
10237 if (ctx_field_size == 4)
10238 size_code = BPF_W;
10239 else if (ctx_field_size == 8)
10240 size_code = BPF_DW;
10241
10242 insn->off = off & ~(size_default - 1);
10243 insn->code = BPF_LDX | BPF_MEM | size_code;
10244 }
10245
10246 target_size = 0;
10247 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
10248 &target_size);
10249 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
10250 (ctx_field_size && !target_size)) {
10251 verbose(env, "bpf verifier is misconfigured\n");
10252 return -EINVAL;
10253 }
10254
10255 if (is_narrower_load && size < target_size) {
10256 u8 shift = bpf_ctx_narrow_access_offset(
10257 off, size, size_default) * 8;
10258 if (ctx_field_size <= 4) {
10259 if (shift)
10260 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
10261 insn->dst_reg,
10262 shift);
10263 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
10264 (1 << size * 8) - 1);
10265 } else {
10266 if (shift)
10267 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
10268 insn->dst_reg,
10269 shift);
10270 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
10271 (1ULL << size * 8) - 1);
10272 }
10273 }
10274
10275 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10276 if (!new_prog)
10277 return -ENOMEM;
10278
10279 delta += cnt - 1;
10280
10281 /* keep walking new program and skip insns we just inserted */
10282 env->prog = new_prog;
10283 insn = new_prog->insnsi + i + delta;
10284 }
10285
10286 return 0;
10287 }
10288
10289 static int jit_subprogs(struct bpf_verifier_env *env)
10290 {
10291 struct bpf_prog *prog = env->prog, **func, *tmp;
10292 int i, j, subprog_start, subprog_end = 0, len, subprog;
10293 struct bpf_map *map_ptr;
10294 struct bpf_insn *insn;
10295 void *old_bpf_func;
10296 int err, num_exentries;
10297
10298 if (env->subprog_cnt <= 1)
10299 return 0;
10300
10301 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10302 if (insn->code != (BPF_JMP | BPF_CALL) ||
10303 insn->src_reg != BPF_PSEUDO_CALL)
10304 continue;
10305 /* Upon error here we cannot fall back to interpreter but
10306 * need a hard reject of the program. Thus -EFAULT is
10307 * propagated in any case.
10308 */
10309 subprog = find_subprog(env, i + insn->imm + 1);
10310 if (subprog < 0) {
10311 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
10312 i + insn->imm + 1);
10313 return -EFAULT;
10314 }
10315 /* temporarily remember subprog id inside insn instead of
10316 * aux_data, since next loop will split up all insns into funcs
10317 */
10318 insn->off = subprog;
10319 /* remember original imm in case JIT fails and fallback
10320 * to interpreter will be needed
10321 */
10322 env->insn_aux_data[i].call_imm = insn->imm;
10323 /* point imm to __bpf_call_base+1 from JITs point of view */
10324 insn->imm = 1;
10325 }
10326
10327 err = bpf_prog_alloc_jited_linfo(prog);
10328 if (err)
10329 goto out_undo_insn;
10330
10331 err = -ENOMEM;
10332 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
10333 if (!func)
10334 goto out_undo_insn;
10335
10336 for (i = 0; i < env->subprog_cnt; i++) {
10337 subprog_start = subprog_end;
10338 subprog_end = env->subprog_info[i + 1].start;
10339
10340 len = subprog_end - subprog_start;
10341 /* BPF_PROG_RUN doesn't call subprogs directly,
10342 * hence main prog stats include the runtime of subprogs.
10343 * subprogs don't have IDs and not reachable via prog_get_next_id
10344 * func[i]->aux->stats will never be accessed and stays NULL
10345 */
10346 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
10347 if (!func[i])
10348 goto out_free;
10349 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
10350 len * sizeof(struct bpf_insn));
10351 func[i]->type = prog->type;
10352 func[i]->len = len;
10353 if (bpf_prog_calc_tag(func[i]))
10354 goto out_free;
10355 func[i]->is_func = 1;
10356 func[i]->aux->func_idx = i;
10357 /* the btf and func_info will be freed only at prog->aux */
10358 func[i]->aux->btf = prog->aux->btf;
10359 func[i]->aux->func_info = prog->aux->func_info;
10360
10361 for (j = 0; j < prog->aux->size_poke_tab; j++) {
10362 u32 insn_idx = prog->aux->poke_tab[j].insn_idx;
10363 int ret;
10364
10365 if (!(insn_idx >= subprog_start &&
10366 insn_idx <= subprog_end))
10367 continue;
10368
10369 ret = bpf_jit_add_poke_descriptor(func[i],
10370 &prog->aux->poke_tab[j]);
10371 if (ret < 0) {
10372 verbose(env, "adding tail call poke descriptor failed\n");
10373 goto out_free;
10374 }
10375
10376 func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
10377
10378 map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
10379 ret = map_ptr->ops->map_poke_track(map_ptr, func[i]->aux);
10380 if (ret < 0) {
10381 verbose(env, "tracking tail call prog failed\n");
10382 goto out_free;
10383 }
10384 }
10385
10386 /* Use bpf_prog_F_tag to indicate functions in stack traces.
10387 * Long term would need debug info to populate names
10388 */
10389 func[i]->aux->name[0] = 'F';
10390 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
10391 func[i]->jit_requested = 1;
10392 func[i]->aux->linfo = prog->aux->linfo;
10393 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
10394 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
10395 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
10396 num_exentries = 0;
10397 insn = func[i]->insnsi;
10398 for (j = 0; j < func[i]->len; j++, insn++) {
10399 if (BPF_CLASS(insn->code) == BPF_LDX &&
10400 BPF_MODE(insn->code) == BPF_PROBE_MEM)
10401 num_exentries++;
10402 }
10403 func[i]->aux->num_exentries = num_exentries;
10404 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
10405 func[i] = bpf_int_jit_compile(func[i]);
10406 if (!func[i]->jited) {
10407 err = -ENOTSUPP;
10408 goto out_free;
10409 }
10410 cond_resched();
10411 }
10412
10413 /* Untrack main program's aux structs so that during map_poke_run()
10414 * we will not stumble upon the unfilled poke descriptors; each
10415 * of the main program's poke descs got distributed across subprogs
10416 * and got tracked onto map, so we are sure that none of them will
10417 * be missed after the operation below
10418 */
10419 for (i = 0; i < prog->aux->size_poke_tab; i++) {
10420 map_ptr = prog->aux->poke_tab[i].tail_call.map;
10421
10422 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
10423 }
10424
10425 /* at this point all bpf functions were successfully JITed
10426 * now populate all bpf_calls with correct addresses and
10427 * run last pass of JIT
10428 */
10429 for (i = 0; i < env->subprog_cnt; i++) {
10430 insn = func[i]->insnsi;
10431 for (j = 0; j < func[i]->len; j++, insn++) {
10432 if (insn->code != (BPF_JMP | BPF_CALL) ||
10433 insn->src_reg != BPF_PSEUDO_CALL)
10434 continue;
10435 subprog = insn->off;
10436 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
10437 __bpf_call_base;
10438 }
10439
10440 /* we use the aux data to keep a list of the start addresses
10441 * of the JITed images for each function in the program
10442 *
10443 * for some architectures, such as powerpc64, the imm field
10444 * might not be large enough to hold the offset of the start
10445 * address of the callee's JITed image from __bpf_call_base
10446 *
10447 * in such cases, we can lookup the start address of a callee
10448 * by using its subprog id, available from the off field of
10449 * the call instruction, as an index for this list
10450 */
10451 func[i]->aux->func = func;
10452 func[i]->aux->func_cnt = env->subprog_cnt;
10453 }
10454 for (i = 0; i < env->subprog_cnt; i++) {
10455 old_bpf_func = func[i]->bpf_func;
10456 tmp = bpf_int_jit_compile(func[i]);
10457 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
10458 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
10459 err = -ENOTSUPP;
10460 goto out_free;
10461 }
10462 cond_resched();
10463 }
10464
10465 /* finally lock prog and jit images for all functions and
10466 * populate kallsysm
10467 */
10468 for (i = 0; i < env->subprog_cnt; i++) {
10469 bpf_prog_lock_ro(func[i]);
10470 bpf_prog_kallsyms_add(func[i]);
10471 }
10472
10473 /* Last step: make now unused interpreter insns from main
10474 * prog consistent for later dump requests, so they can
10475 * later look the same as if they were interpreted only.
10476 */
10477 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10478 if (insn->code != (BPF_JMP | BPF_CALL) ||
10479 insn->src_reg != BPF_PSEUDO_CALL)
10480 continue;
10481 insn->off = env->insn_aux_data[i].call_imm;
10482 subprog = find_subprog(env, i + insn->off + 1);
10483 insn->imm = subprog;
10484 }
10485
10486 prog->jited = 1;
10487 prog->bpf_func = func[0]->bpf_func;
10488 prog->aux->func = func;
10489 prog->aux->func_cnt = env->subprog_cnt;
10490 bpf_prog_free_unused_jited_linfo(prog);
10491 return 0;
10492 out_free:
10493 for (i = 0; i < env->subprog_cnt; i++) {
10494 if (!func[i])
10495 continue;
10496
10497 for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
10498 map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
10499 map_ptr->ops->map_poke_untrack(map_ptr, func[i]->aux);
10500 }
10501 bpf_jit_free(func[i]);
10502 }
10503 kfree(func);
10504 out_undo_insn:
10505 /* cleanup main prog to be interpreted */
10506 prog->jit_requested = 0;
10507 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10508 if (insn->code != (BPF_JMP | BPF_CALL) ||
10509 insn->src_reg != BPF_PSEUDO_CALL)
10510 continue;
10511 insn->off = 0;
10512 insn->imm = env->insn_aux_data[i].call_imm;
10513 }
10514 bpf_prog_free_jited_linfo(prog);
10515 return err;
10516 }
10517
10518 static int fixup_call_args(struct bpf_verifier_env *env)
10519 {
10520 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10521 struct bpf_prog *prog = env->prog;
10522 struct bpf_insn *insn = prog->insnsi;
10523 int i, depth;
10524 #endif
10525 int err = 0;
10526
10527 if (env->prog->jit_requested &&
10528 !bpf_prog_is_dev_bound(env->prog->aux)) {
10529 err = jit_subprogs(env);
10530 if (err == 0)
10531 return 0;
10532 if (err == -EFAULT)
10533 return err;
10534 }
10535 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10536 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
10537 /* When JIT fails the progs with bpf2bpf calls and tail_calls
10538 * have to be rejected, since interpreter doesn't support them yet.
10539 */
10540 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
10541 return -EINVAL;
10542 }
10543 for (i = 0; i < prog->len; i++, insn++) {
10544 if (insn->code != (BPF_JMP | BPF_CALL) ||
10545 insn->src_reg != BPF_PSEUDO_CALL)
10546 continue;
10547 depth = get_callee_stack_depth(env, insn, i);
10548 if (depth < 0)
10549 return depth;
10550 bpf_patch_call_args(insn, depth);
10551 }
10552 err = 0;
10553 #endif
10554 return err;
10555 }
10556
10557 /* fixup insn->imm field of bpf_call instructions
10558 * and inline eligible helpers as explicit sequence of BPF instructions
10559 *
10560 * this function is called after eBPF program passed verification
10561 */
10562 static int fixup_bpf_calls(struct bpf_verifier_env *env)
10563 {
10564 struct bpf_prog *prog = env->prog;
10565 bool expect_blinding = bpf_jit_blinding_enabled(prog);
10566 struct bpf_insn *insn = prog->insnsi;
10567 const struct bpf_func_proto *fn;
10568 const int insn_cnt = prog->len;
10569 const struct bpf_map_ops *ops;
10570 struct bpf_insn_aux_data *aux;
10571 struct bpf_insn insn_buf[16];
10572 struct bpf_prog *new_prog;
10573 struct bpf_map *map_ptr;
10574 int i, ret, cnt, delta = 0;
10575
10576 for (i = 0; i < insn_cnt; i++, insn++) {
10577 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
10578 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
10579 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
10580 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
10581 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
10582 struct bpf_insn mask_and_div[] = {
10583 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
10584 /* Rx div 0 -> 0 */
10585 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
10586 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
10587 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
10588 *insn,
10589 };
10590 struct bpf_insn mask_and_mod[] = {
10591 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
10592 /* Rx mod 0 -> Rx */
10593 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
10594 *insn,
10595 };
10596 struct bpf_insn *patchlet;
10597
10598 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
10599 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
10600 patchlet = mask_and_div + (is64 ? 1 : 0);
10601 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
10602 } else {
10603 patchlet = mask_and_mod + (is64 ? 1 : 0);
10604 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
10605 }
10606
10607 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
10608 if (!new_prog)
10609 return -ENOMEM;
10610
10611 delta += cnt - 1;
10612 env->prog = prog = new_prog;
10613 insn = new_prog->insnsi + i + delta;
10614 continue;
10615 }
10616
10617 if (BPF_CLASS(insn->code) == BPF_LD &&
10618 (BPF_MODE(insn->code) == BPF_ABS ||
10619 BPF_MODE(insn->code) == BPF_IND)) {
10620 cnt = env->ops->gen_ld_abs(insn, insn_buf);
10621 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
10622 verbose(env, "bpf verifier is misconfigured\n");
10623 return -EINVAL;
10624 }
10625
10626 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10627 if (!new_prog)
10628 return -ENOMEM;
10629
10630 delta += cnt - 1;
10631 env->prog = prog = new_prog;
10632 insn = new_prog->insnsi + i + delta;
10633 continue;
10634 }
10635
10636 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
10637 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
10638 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
10639 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
10640 struct bpf_insn insn_buf[16];
10641 struct bpf_insn *patch = &insn_buf[0];
10642 bool issrc, isneg;
10643 u32 off_reg;
10644
10645 aux = &env->insn_aux_data[i + delta];
10646 if (!aux->alu_state ||
10647 aux->alu_state == BPF_ALU_NON_POINTER)
10648 continue;
10649
10650 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
10651 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
10652 BPF_ALU_SANITIZE_SRC;
10653
10654 off_reg = issrc ? insn->src_reg : insn->dst_reg;
10655 if (isneg)
10656 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
10657 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
10658 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
10659 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
10660 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
10661 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
10662 if (issrc) {
10663 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
10664 off_reg);
10665 insn->src_reg = BPF_REG_AX;
10666 } else {
10667 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
10668 BPF_REG_AX);
10669 }
10670 if (isneg)
10671 insn->code = insn->code == code_add ?
10672 code_sub : code_add;
10673 *patch++ = *insn;
10674 if (issrc && isneg)
10675 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
10676 cnt = patch - insn_buf;
10677
10678 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10679 if (!new_prog)
10680 return -ENOMEM;
10681
10682 delta += cnt - 1;
10683 env->prog = prog = new_prog;
10684 insn = new_prog->insnsi + i + delta;
10685 continue;
10686 }
10687
10688 if (insn->code != (BPF_JMP | BPF_CALL))
10689 continue;
10690 if (insn->src_reg == BPF_PSEUDO_CALL)
10691 continue;
10692
10693 if (insn->imm == BPF_FUNC_get_route_realm)
10694 prog->dst_needed = 1;
10695 if (insn->imm == BPF_FUNC_get_prandom_u32)
10696 bpf_user_rnd_init_once();
10697 if (insn->imm == BPF_FUNC_override_return)
10698 prog->kprobe_override = 1;
10699 if (insn->imm == BPF_FUNC_tail_call) {
10700 /* If we tail call into other programs, we
10701 * cannot make any assumptions since they can
10702 * be replaced dynamically during runtime in
10703 * the program array.
10704 */
10705 prog->cb_access = 1;
10706 if (!allow_tail_call_in_subprogs(env))
10707 prog->aux->stack_depth = MAX_BPF_STACK;
10708 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
10709
10710 /* mark bpf_tail_call as different opcode to avoid
10711 * conditional branch in the interpeter for every normal
10712 * call and to prevent accidental JITing by JIT compiler
10713 * that doesn't support bpf_tail_call yet
10714 */
10715 insn->imm = 0;
10716 insn->code = BPF_JMP | BPF_TAIL_CALL;
10717
10718 aux = &env->insn_aux_data[i + delta];
10719 if (env->bpf_capable && !expect_blinding &&
10720 prog->jit_requested &&
10721 !bpf_map_key_poisoned(aux) &&
10722 !bpf_map_ptr_poisoned(aux) &&
10723 !bpf_map_ptr_unpriv(aux)) {
10724 struct bpf_jit_poke_descriptor desc = {
10725 .reason = BPF_POKE_REASON_TAIL_CALL,
10726 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
10727 .tail_call.key = bpf_map_key_immediate(aux),
10728 .insn_idx = i + delta,
10729 };
10730
10731 ret = bpf_jit_add_poke_descriptor(prog, &desc);
10732 if (ret < 0) {
10733 verbose(env, "adding tail call poke descriptor failed\n");
10734 return ret;
10735 }
10736
10737 insn->imm = ret + 1;
10738 continue;
10739 }
10740
10741 if (!bpf_map_ptr_unpriv(aux))
10742 continue;
10743
10744 /* instead of changing every JIT dealing with tail_call
10745 * emit two extra insns:
10746 * if (index >= max_entries) goto out;
10747 * index &= array->index_mask;
10748 * to avoid out-of-bounds cpu speculation
10749 */
10750 if (bpf_map_ptr_poisoned(aux)) {
10751 verbose(env, "tail_call abusing map_ptr\n");
10752 return -EINVAL;
10753 }
10754
10755 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
10756 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
10757 map_ptr->max_entries, 2);
10758 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
10759 container_of(map_ptr,
10760 struct bpf_array,
10761 map)->index_mask);
10762 insn_buf[2] = *insn;
10763 cnt = 3;
10764 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10765 if (!new_prog)
10766 return -ENOMEM;
10767
10768 delta += cnt - 1;
10769 env->prog = prog = new_prog;
10770 insn = new_prog->insnsi + i + delta;
10771 continue;
10772 }
10773
10774 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
10775 * and other inlining handlers are currently limited to 64 bit
10776 * only.
10777 */
10778 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10779 (insn->imm == BPF_FUNC_map_lookup_elem ||
10780 insn->imm == BPF_FUNC_map_update_elem ||
10781 insn->imm == BPF_FUNC_map_delete_elem ||
10782 insn->imm == BPF_FUNC_map_push_elem ||
10783 insn->imm == BPF_FUNC_map_pop_elem ||
10784 insn->imm == BPF_FUNC_map_peek_elem)) {
10785 aux = &env->insn_aux_data[i + delta];
10786 if (bpf_map_ptr_poisoned(aux))
10787 goto patch_call_imm;
10788
10789 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
10790 ops = map_ptr->ops;
10791 if (insn->imm == BPF_FUNC_map_lookup_elem &&
10792 ops->map_gen_lookup) {
10793 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
10794 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
10795 verbose(env, "bpf verifier is misconfigured\n");
10796 return -EINVAL;
10797 }
10798
10799 new_prog = bpf_patch_insn_data(env, i + delta,
10800 insn_buf, cnt);
10801 if (!new_prog)
10802 return -ENOMEM;
10803
10804 delta += cnt - 1;
10805 env->prog = prog = new_prog;
10806 insn = new_prog->insnsi + i + delta;
10807 continue;
10808 }
10809
10810 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
10811 (void *(*)(struct bpf_map *map, void *key))NULL));
10812 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
10813 (int (*)(struct bpf_map *map, void *key))NULL));
10814 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
10815 (int (*)(struct bpf_map *map, void *key, void *value,
10816 u64 flags))NULL));
10817 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
10818 (int (*)(struct bpf_map *map, void *value,
10819 u64 flags))NULL));
10820 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
10821 (int (*)(struct bpf_map *map, void *value))NULL));
10822 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
10823 (int (*)(struct bpf_map *map, void *value))NULL));
10824
10825 switch (insn->imm) {
10826 case BPF_FUNC_map_lookup_elem:
10827 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
10828 __bpf_call_base;
10829 continue;
10830 case BPF_FUNC_map_update_elem:
10831 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
10832 __bpf_call_base;
10833 continue;
10834 case BPF_FUNC_map_delete_elem:
10835 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
10836 __bpf_call_base;
10837 continue;
10838 case BPF_FUNC_map_push_elem:
10839 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
10840 __bpf_call_base;
10841 continue;
10842 case BPF_FUNC_map_pop_elem:
10843 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
10844 __bpf_call_base;
10845 continue;
10846 case BPF_FUNC_map_peek_elem:
10847 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
10848 __bpf_call_base;
10849 continue;
10850 }
10851
10852 goto patch_call_imm;
10853 }
10854
10855 if (prog->jit_requested && BITS_PER_LONG == 64 &&
10856 insn->imm == BPF_FUNC_jiffies64) {
10857 struct bpf_insn ld_jiffies_addr[2] = {
10858 BPF_LD_IMM64(BPF_REG_0,
10859 (unsigned long)&jiffies),
10860 };
10861
10862 insn_buf[0] = ld_jiffies_addr[0];
10863 insn_buf[1] = ld_jiffies_addr[1];
10864 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
10865 BPF_REG_0, 0);
10866 cnt = 3;
10867
10868 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
10869 cnt);
10870 if (!new_prog)
10871 return -ENOMEM;
10872
10873 delta += cnt - 1;
10874 env->prog = prog = new_prog;
10875 insn = new_prog->insnsi + i + delta;
10876 continue;
10877 }
10878
10879 patch_call_imm:
10880 fn = env->ops->get_func_proto(insn->imm, env->prog);
10881 /* all functions that have prototype and verifier allowed
10882 * programs to call them, must be real in-kernel functions
10883 */
10884 if (!fn->func) {
10885 verbose(env,
10886 "kernel subsystem misconfigured func %s#%d\n",
10887 func_id_name(insn->imm), insn->imm);
10888 return -EFAULT;
10889 }
10890 insn->imm = fn->func - __bpf_call_base;
10891 }
10892
10893 /* Since poke tab is now finalized, publish aux to tracker. */
10894 for (i = 0; i < prog->aux->size_poke_tab; i++) {
10895 map_ptr = prog->aux->poke_tab[i].tail_call.map;
10896 if (!map_ptr->ops->map_poke_track ||
10897 !map_ptr->ops->map_poke_untrack ||
10898 !map_ptr->ops->map_poke_run) {
10899 verbose(env, "bpf verifier is misconfigured\n");
10900 return -EINVAL;
10901 }
10902
10903 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
10904 if (ret < 0) {
10905 verbose(env, "tracking tail call prog failed\n");
10906 return ret;
10907 }
10908 }
10909
10910 return 0;
10911 }
10912
10913 static void free_states(struct bpf_verifier_env *env)
10914 {
10915 struct bpf_verifier_state_list *sl, *sln;
10916 int i;
10917
10918 sl = env->free_list;
10919 while (sl) {
10920 sln = sl->next;
10921 free_verifier_state(&sl->state, false);
10922 kfree(sl);
10923 sl = sln;
10924 }
10925 env->free_list = NULL;
10926
10927 if (!env->explored_states)
10928 return;
10929
10930 for (i = 0; i < state_htab_size(env); i++) {
10931 sl = env->explored_states[i];
10932
10933 while (sl) {
10934 sln = sl->next;
10935 free_verifier_state(&sl->state, false);
10936 kfree(sl);
10937 sl = sln;
10938 }
10939 env->explored_states[i] = NULL;
10940 }
10941 }
10942
10943 /* The verifier is using insn_aux_data[] to store temporary data during
10944 * verification and to store information for passes that run after the
10945 * verification like dead code sanitization. do_check_common() for subprogram N
10946 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
10947 * temporary data after do_check_common() finds that subprogram N cannot be
10948 * verified independently. pass_cnt counts the number of times
10949 * do_check_common() was run and insn->aux->seen tells the pass number
10950 * insn_aux_data was touched. These variables are compared to clear temporary
10951 * data from failed pass. For testing and experiments do_check_common() can be
10952 * run multiple times even when prior attempt to verify is unsuccessful.
10953 */
10954 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
10955 {
10956 struct bpf_insn *insn = env->prog->insnsi;
10957 struct bpf_insn_aux_data *aux;
10958 int i, class;
10959
10960 for (i = 0; i < env->prog->len; i++) {
10961 class = BPF_CLASS(insn[i].code);
10962 if (class != BPF_LDX && class != BPF_STX)
10963 continue;
10964 aux = &env->insn_aux_data[i];
10965 if (aux->seen != env->pass_cnt)
10966 continue;
10967 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
10968 }
10969 }
10970
10971 static int do_check_common(struct bpf_verifier_env *env, int subprog)
10972 {
10973 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10974 struct bpf_verifier_state *state;
10975 struct bpf_reg_state *regs;
10976 int ret, i;
10977
10978 env->prev_linfo = NULL;
10979 env->pass_cnt++;
10980
10981 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
10982 if (!state)
10983 return -ENOMEM;
10984 state->curframe = 0;
10985 state->speculative = false;
10986 state->branches = 1;
10987 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
10988 if (!state->frame[0]) {
10989 kfree(state);
10990 return -ENOMEM;
10991 }
10992 env->cur_state = state;
10993 init_func_state(env, state->frame[0],
10994 BPF_MAIN_FUNC /* callsite */,
10995 0 /* frameno */,
10996 subprog);
10997
10998 regs = state->frame[state->curframe]->regs;
10999 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
11000 ret = btf_prepare_func_args(env, subprog, regs);
11001 if (ret)
11002 goto out;
11003 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
11004 if (regs[i].type == PTR_TO_CTX)
11005 mark_reg_known_zero(env, regs, i);
11006 else if (regs[i].type == SCALAR_VALUE)
11007 mark_reg_unknown(env, regs, i);
11008 }
11009 } else {
11010 /* 1st arg to a function */
11011 regs[BPF_REG_1].type = PTR_TO_CTX;
11012 mark_reg_known_zero(env, regs, BPF_REG_1);
11013 ret = btf_check_func_arg_match(env, subprog, regs);
11014 if (ret == -EFAULT)
11015 /* unlikely verifier bug. abort.
11016 * ret == 0 and ret < 0 are sadly acceptable for
11017 * main() function due to backward compatibility.
11018 * Like socket filter program may be written as:
11019 * int bpf_prog(struct pt_regs *ctx)
11020 * and never dereference that ctx in the program.
11021 * 'struct pt_regs' is a type mismatch for socket
11022 * filter that should be using 'struct __sk_buff'.
11023 */
11024 goto out;
11025 }
11026
11027 ret = do_check(env);
11028 out:
11029 /* check for NULL is necessary, since cur_state can be freed inside
11030 * do_check() under memory pressure.
11031 */
11032 if (env->cur_state) {
11033 free_verifier_state(env->cur_state, true);
11034 env->cur_state = NULL;
11035 }
11036 while (!pop_stack(env, NULL, NULL, false));
11037 if (!ret && pop_log)
11038 bpf_vlog_reset(&env->log, 0);
11039 free_states(env);
11040 if (ret)
11041 /* clean aux data in case subprog was rejected */
11042 sanitize_insn_aux_data(env);
11043 return ret;
11044 }
11045
11046 /* Verify all global functions in a BPF program one by one based on their BTF.
11047 * All global functions must pass verification. Otherwise the whole program is rejected.
11048 * Consider:
11049 * int bar(int);
11050 * int foo(int f)
11051 * {
11052 * return bar(f);
11053 * }
11054 * int bar(int b)
11055 * {
11056 * ...
11057 * }
11058 * foo() will be verified first for R1=any_scalar_value. During verification it
11059 * will be assumed that bar() already verified successfully and call to bar()
11060 * from foo() will be checked for type match only. Later bar() will be verified
11061 * independently to check that it's safe for R1=any_scalar_value.
11062 */
11063 static int do_check_subprogs(struct bpf_verifier_env *env)
11064 {
11065 struct bpf_prog_aux *aux = env->prog->aux;
11066 int i, ret;
11067
11068 if (!aux->func_info)
11069 return 0;
11070
11071 for (i = 1; i < env->subprog_cnt; i++) {
11072 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
11073 continue;
11074 env->insn_idx = env->subprog_info[i].start;
11075 WARN_ON_ONCE(env->insn_idx == 0);
11076 ret = do_check_common(env, i);
11077 if (ret) {
11078 return ret;
11079 } else if (env->log.level & BPF_LOG_LEVEL) {
11080 verbose(env,
11081 "Func#%d is safe for any args that match its prototype\n",
11082 i);
11083 }
11084 }
11085 return 0;
11086 }
11087
11088 static int do_check_main(struct bpf_verifier_env *env)
11089 {
11090 int ret;
11091
11092 env->insn_idx = 0;
11093 ret = do_check_common(env, 0);
11094 if (!ret)
11095 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
11096 return ret;
11097 }
11098
11099
11100 static void print_verification_stats(struct bpf_verifier_env *env)
11101 {
11102 int i;
11103
11104 if (env->log.level & BPF_LOG_STATS) {
11105 verbose(env, "verification time %lld usec\n",
11106 div_u64(env->verification_time, 1000));
11107 verbose(env, "stack depth ");
11108 for (i = 0; i < env->subprog_cnt; i++) {
11109 u32 depth = env->subprog_info[i].stack_depth;
11110
11111 verbose(env, "%d", depth);
11112 if (i + 1 < env->subprog_cnt)
11113 verbose(env, "+");
11114 }
11115 verbose(env, "\n");
11116 }
11117 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
11118 "total_states %d peak_states %d mark_read %d\n",
11119 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
11120 env->max_states_per_insn, env->total_states,
11121 env->peak_states, env->longest_mark_read_walk);
11122 }
11123
11124 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
11125 {
11126 const struct btf_type *t, *func_proto;
11127 const struct bpf_struct_ops *st_ops;
11128 const struct btf_member *member;
11129 struct bpf_prog *prog = env->prog;
11130 u32 btf_id, member_idx;
11131 const char *mname;
11132
11133 btf_id = prog->aux->attach_btf_id;
11134 st_ops = bpf_struct_ops_find(btf_id);
11135 if (!st_ops) {
11136 verbose(env, "attach_btf_id %u is not a supported struct\n",
11137 btf_id);
11138 return -ENOTSUPP;
11139 }
11140
11141 t = st_ops->type;
11142 member_idx = prog->expected_attach_type;
11143 if (member_idx >= btf_type_vlen(t)) {
11144 verbose(env, "attach to invalid member idx %u of struct %s\n",
11145 member_idx, st_ops->name);
11146 return -EINVAL;
11147 }
11148
11149 member = &btf_type_member(t)[member_idx];
11150 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
11151 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
11152 NULL);
11153 if (!func_proto) {
11154 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
11155 mname, member_idx, st_ops->name);
11156 return -EINVAL;
11157 }
11158
11159 if (st_ops->check_member) {
11160 int err = st_ops->check_member(t, member);
11161
11162 if (err) {
11163 verbose(env, "attach to unsupported member %s of struct %s\n",
11164 mname, st_ops->name);
11165 return err;
11166 }
11167 }
11168
11169 prog->aux->attach_func_proto = func_proto;
11170 prog->aux->attach_func_name = mname;
11171 env->ops = st_ops->verifier_ops;
11172
11173 return 0;
11174 }
11175 #define SECURITY_PREFIX "security_"
11176
11177 static int check_attach_modify_return(struct bpf_prog *prog, unsigned long addr)
11178 {
11179 if (within_error_injection_list(addr) ||
11180 !strncmp(SECURITY_PREFIX, prog->aux->attach_func_name,
11181 sizeof(SECURITY_PREFIX) - 1))
11182 return 0;
11183
11184 return -EINVAL;
11185 }
11186
11187 /* non exhaustive list of sleepable bpf_lsm_*() functions */
11188 BTF_SET_START(btf_sleepable_lsm_hooks)
11189 #ifdef CONFIG_BPF_LSM
11190 BTF_ID(func, bpf_lsm_bprm_committed_creds)
11191 #else
11192 BTF_ID_UNUSED
11193 #endif
11194 BTF_SET_END(btf_sleepable_lsm_hooks)
11195
11196 static int check_sleepable_lsm_hook(u32 btf_id)
11197 {
11198 return btf_id_set_contains(&btf_sleepable_lsm_hooks, btf_id);
11199 }
11200
11201 /* list of non-sleepable functions that are otherwise on
11202 * ALLOW_ERROR_INJECTION list
11203 */
11204 BTF_SET_START(btf_non_sleepable_error_inject)
11205 /* Three functions below can be called from sleepable and non-sleepable context.
11206 * Assume non-sleepable from bpf safety point of view.
11207 */
11208 BTF_ID(func, __add_to_page_cache_locked)
11209 BTF_ID(func, should_fail_alloc_page)
11210 BTF_ID(func, should_failslab)
11211 BTF_SET_END(btf_non_sleepable_error_inject)
11212
11213 static int check_non_sleepable_error_inject(u32 btf_id)
11214 {
11215 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
11216 }
11217
11218 static int check_attach_btf_id(struct bpf_verifier_env *env)
11219 {
11220 struct bpf_prog *prog = env->prog;
11221 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
11222 struct bpf_prog *tgt_prog = prog->aux->linked_prog;
11223 u32 btf_id = prog->aux->attach_btf_id;
11224 const char prefix[] = "btf_trace_";
11225 struct btf_func_model fmodel;
11226 int ret = 0, subprog = -1, i;
11227 struct bpf_trampoline *tr;
11228 const struct btf_type *t;
11229 bool conservative = true;
11230 const char *tname;
11231 struct btf *btf;
11232 long addr;
11233 u64 key;
11234
11235 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
11236 prog->type != BPF_PROG_TYPE_LSM) {
11237 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
11238 return -EINVAL;
11239 }
11240
11241 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
11242 return check_struct_ops_btf_id(env);
11243
11244 if (prog->type != BPF_PROG_TYPE_TRACING &&
11245 prog->type != BPF_PROG_TYPE_LSM &&
11246 !prog_extension)
11247 return 0;
11248
11249 if (!btf_id) {
11250 verbose(env, "Tracing programs must provide btf_id\n");
11251 return -EINVAL;
11252 }
11253 btf = bpf_prog_get_target_btf(prog);
11254 if (!btf) {
11255 verbose(env,
11256 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
11257 return -EINVAL;
11258 }
11259 t = btf_type_by_id(btf, btf_id);
11260 if (!t) {
11261 verbose(env, "attach_btf_id %u is invalid\n", btf_id);
11262 return -EINVAL;
11263 }
11264 tname = btf_name_by_offset(btf, t->name_off);
11265 if (!tname) {
11266 verbose(env, "attach_btf_id %u doesn't have a name\n", btf_id);
11267 return -EINVAL;
11268 }
11269 if (tgt_prog) {
11270 struct bpf_prog_aux *aux = tgt_prog->aux;
11271
11272 for (i = 0; i < aux->func_info_cnt; i++)
11273 if (aux->func_info[i].type_id == btf_id) {
11274 subprog = i;
11275 break;
11276 }
11277 if (subprog == -1) {
11278 verbose(env, "Subprog %s doesn't exist\n", tname);
11279 return -EINVAL;
11280 }
11281 conservative = aux->func_info_aux[subprog].unreliable;
11282 if (prog_extension) {
11283 if (conservative) {
11284 verbose(env,
11285 "Cannot replace static functions\n");
11286 return -EINVAL;
11287 }
11288 if (!prog->jit_requested) {
11289 verbose(env,
11290 "Extension programs should be JITed\n");
11291 return -EINVAL;
11292 }
11293 env->ops = bpf_verifier_ops[tgt_prog->type];
11294 prog->expected_attach_type = tgt_prog->expected_attach_type;
11295 }
11296 if (!tgt_prog->jited) {
11297 verbose(env, "Can attach to only JITed progs\n");
11298 return -EINVAL;
11299 }
11300 if (tgt_prog->type == prog->type) {
11301 /* Cannot fentry/fexit another fentry/fexit program.
11302 * Cannot attach program extension to another extension.
11303 * It's ok to attach fentry/fexit to extension program.
11304 */
11305 verbose(env, "Cannot recursively attach\n");
11306 return -EINVAL;
11307 }
11308 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
11309 prog_extension &&
11310 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
11311 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
11312 /* Program extensions can extend all program types
11313 * except fentry/fexit. The reason is the following.
11314 * The fentry/fexit programs are used for performance
11315 * analysis, stats and can be attached to any program
11316 * type except themselves. When extension program is
11317 * replacing XDP function it is necessary to allow
11318 * performance analysis of all functions. Both original
11319 * XDP program and its program extension. Hence
11320 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
11321 * allowed. If extending of fentry/fexit was allowed it
11322 * would be possible to create long call chain
11323 * fentry->extension->fentry->extension beyond
11324 * reasonable stack size. Hence extending fentry is not
11325 * allowed.
11326 */
11327 verbose(env, "Cannot extend fentry/fexit\n");
11328 return -EINVAL;
11329 }
11330 key = ((u64)aux->id) << 32 | btf_id;
11331 } else {
11332 if (prog_extension) {
11333 verbose(env, "Cannot replace kernel functions\n");
11334 return -EINVAL;
11335 }
11336 key = btf_id;
11337 }
11338
11339 switch (prog->expected_attach_type) {
11340 case BPF_TRACE_RAW_TP:
11341 if (tgt_prog) {
11342 verbose(env,
11343 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
11344 return -EINVAL;
11345 }
11346 if (!btf_type_is_typedef(t)) {
11347 verbose(env, "attach_btf_id %u is not a typedef\n",
11348 btf_id);
11349 return -EINVAL;
11350 }
11351 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
11352 verbose(env, "attach_btf_id %u points to wrong type name %s\n",
11353 btf_id, tname);
11354 return -EINVAL;
11355 }
11356 tname += sizeof(prefix) - 1;
11357 t = btf_type_by_id(btf, t->type);
11358 if (!btf_type_is_ptr(t))
11359 /* should never happen in valid vmlinux build */
11360 return -EINVAL;
11361 t = btf_type_by_id(btf, t->type);
11362 if (!btf_type_is_func_proto(t))
11363 /* should never happen in valid vmlinux build */
11364 return -EINVAL;
11365
11366 /* remember two read only pointers that are valid for
11367 * the life time of the kernel
11368 */
11369 prog->aux->attach_func_name = tname;
11370 prog->aux->attach_func_proto = t;
11371 prog->aux->attach_btf_trace = true;
11372 return 0;
11373 case BPF_TRACE_ITER:
11374 if (!btf_type_is_func(t)) {
11375 verbose(env, "attach_btf_id %u is not a function\n",
11376 btf_id);
11377 return -EINVAL;
11378 }
11379 t = btf_type_by_id(btf, t->type);
11380 if (!btf_type_is_func_proto(t))
11381 return -EINVAL;
11382 prog->aux->attach_func_name = tname;
11383 prog->aux->attach_func_proto = t;
11384 if (!bpf_iter_prog_supported(prog))
11385 return -EINVAL;
11386 ret = btf_distill_func_proto(&env->log, btf, t,
11387 tname, &fmodel);
11388 return ret;
11389 default:
11390 if (!prog_extension)
11391 return -EINVAL;
11392 /* fallthrough */
11393 case BPF_MODIFY_RETURN:
11394 case BPF_LSM_MAC:
11395 case BPF_TRACE_FENTRY:
11396 case BPF_TRACE_FEXIT:
11397 prog->aux->attach_func_name = tname;
11398 if (prog->type == BPF_PROG_TYPE_LSM) {
11399 ret = bpf_lsm_verify_prog(&env->log, prog);
11400 if (ret < 0)
11401 return ret;
11402 }
11403
11404 if (!btf_type_is_func(t)) {
11405 verbose(env, "attach_btf_id %u is not a function\n",
11406 btf_id);
11407 return -EINVAL;
11408 }
11409 if (prog_extension &&
11410 btf_check_type_match(env, prog, btf, t))
11411 return -EINVAL;
11412 t = btf_type_by_id(btf, t->type);
11413 if (!btf_type_is_func_proto(t))
11414 return -EINVAL;
11415 tr = bpf_trampoline_lookup(key);
11416 if (!tr)
11417 return -ENOMEM;
11418 /* t is either vmlinux type or another program's type */
11419 prog->aux->attach_func_proto = t;
11420 mutex_lock(&tr->mutex);
11421 if (tr->func.addr) {
11422 prog->aux->trampoline = tr;
11423 goto out;
11424 }
11425 if (tgt_prog && conservative) {
11426 prog->aux->attach_func_proto = NULL;
11427 t = NULL;
11428 }
11429 ret = btf_distill_func_proto(&env->log, btf, t,
11430 tname, &tr->func.model);
11431 if (ret < 0)
11432 goto out;
11433 if (tgt_prog) {
11434 if (subprog == 0)
11435 addr = (long) tgt_prog->bpf_func;
11436 else
11437 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
11438 } else {
11439 addr = kallsyms_lookup_name(tname);
11440 if (!addr) {
11441 verbose(env,
11442 "The address of function %s cannot be found\n",
11443 tname);
11444 ret = -ENOENT;
11445 goto out;
11446 }
11447 }
11448
11449 if (prog->aux->sleepable) {
11450 ret = -EINVAL;
11451 switch (prog->type) {
11452 case BPF_PROG_TYPE_TRACING:
11453 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
11454 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
11455 */
11456 if (!check_non_sleepable_error_inject(btf_id) &&
11457 within_error_injection_list(addr))
11458 ret = 0;
11459 break;
11460 case BPF_PROG_TYPE_LSM:
11461 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
11462 * Only some of them are sleepable.
11463 */
11464 if (check_sleepable_lsm_hook(btf_id))
11465 ret = 0;
11466 break;
11467 default:
11468 break;
11469 }
11470 if (ret)
11471 verbose(env, "%s is not sleepable\n",
11472 prog->aux->attach_func_name);
11473 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
11474 ret = check_attach_modify_return(prog, addr);
11475 if (ret)
11476 verbose(env, "%s() is not modifiable\n",
11477 prog->aux->attach_func_name);
11478 }
11479 if (ret)
11480 goto out;
11481 tr->func.addr = (void *)addr;
11482 prog->aux->trampoline = tr;
11483 out:
11484 mutex_unlock(&tr->mutex);
11485 if (ret)
11486 bpf_trampoline_put(tr);
11487 return ret;
11488 }
11489 }
11490
11491 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
11492 union bpf_attr __user *uattr)
11493 {
11494 u64 start_time = ktime_get_ns();
11495 struct bpf_verifier_env *env;
11496 struct bpf_verifier_log *log;
11497 int i, len, ret = -EINVAL;
11498 bool is_priv;
11499
11500 /* no program is valid */
11501 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
11502 return -EINVAL;
11503
11504 /* 'struct bpf_verifier_env' can be global, but since it's not small,
11505 * allocate/free it every time bpf_check() is called
11506 */
11507 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
11508 if (!env)
11509 return -ENOMEM;
11510 log = &env->log;
11511
11512 len = (*prog)->len;
11513 env->insn_aux_data =
11514 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
11515 ret = -ENOMEM;
11516 if (!env->insn_aux_data)
11517 goto err_free_env;
11518 for (i = 0; i < len; i++)
11519 env->insn_aux_data[i].orig_idx = i;
11520 env->prog = *prog;
11521 env->ops = bpf_verifier_ops[env->prog->type];
11522 is_priv = bpf_capable();
11523
11524 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
11525 mutex_lock(&bpf_verifier_lock);
11526 if (!btf_vmlinux)
11527 btf_vmlinux = btf_parse_vmlinux();
11528 mutex_unlock(&bpf_verifier_lock);
11529 }
11530
11531 /* grab the mutex to protect few globals used by verifier */
11532 if (!is_priv)
11533 mutex_lock(&bpf_verifier_lock);
11534
11535 if (attr->log_level || attr->log_buf || attr->log_size) {
11536 /* user requested verbose verifier output
11537 * and supplied buffer to store the verification trace
11538 */
11539 log->level = attr->log_level;
11540 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
11541 log->len_total = attr->log_size;
11542
11543 ret = -EINVAL;
11544 /* log attributes have to be sane */
11545 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
11546 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
11547 goto err_unlock;
11548 }
11549
11550 if (IS_ERR(btf_vmlinux)) {
11551 /* Either gcc or pahole or kernel are broken. */
11552 verbose(env, "in-kernel BTF is malformed\n");
11553 ret = PTR_ERR(btf_vmlinux);
11554 goto skip_full_check;
11555 }
11556
11557 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
11558 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
11559 env->strict_alignment = true;
11560 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
11561 env->strict_alignment = false;
11562
11563 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
11564 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
11565 env->bypass_spec_v1 = bpf_bypass_spec_v1();
11566 env->bypass_spec_v4 = bpf_bypass_spec_v4();
11567 env->bpf_capable = bpf_capable();
11568
11569 if (is_priv)
11570 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
11571
11572 ret = replace_map_fd_with_map_ptr(env);
11573 if (ret < 0)
11574 goto skip_full_check;
11575
11576 if (bpf_prog_is_dev_bound(env->prog->aux)) {
11577 ret = bpf_prog_offload_verifier_prep(env->prog);
11578 if (ret)
11579 goto skip_full_check;
11580 }
11581
11582 env->explored_states = kvcalloc(state_htab_size(env),
11583 sizeof(struct bpf_verifier_state_list *),
11584 GFP_USER);
11585 ret = -ENOMEM;
11586 if (!env->explored_states)
11587 goto skip_full_check;
11588
11589 ret = check_subprogs(env);
11590 if (ret < 0)
11591 goto skip_full_check;
11592
11593 ret = check_btf_info(env, attr, uattr);
11594 if (ret < 0)
11595 goto skip_full_check;
11596
11597 ret = check_attach_btf_id(env);
11598 if (ret)
11599 goto skip_full_check;
11600
11601 ret = check_cfg(env);
11602 if (ret < 0)
11603 goto skip_full_check;
11604
11605 ret = do_check_subprogs(env);
11606 ret = ret ?: do_check_main(env);
11607
11608 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
11609 ret = bpf_prog_offload_finalize(env);
11610
11611 skip_full_check:
11612 kvfree(env->explored_states);
11613
11614 if (ret == 0)
11615 ret = check_max_stack_depth(env);
11616
11617 /* instruction rewrites happen after this point */
11618 if (is_priv) {
11619 if (ret == 0)
11620 opt_hard_wire_dead_code_branches(env);
11621 if (ret == 0)
11622 ret = opt_remove_dead_code(env);
11623 if (ret == 0)
11624 ret = opt_remove_nops(env);
11625 } else {
11626 if (ret == 0)
11627 sanitize_dead_code(env);
11628 }
11629
11630 if (ret == 0)
11631 /* program is valid, convert *(u32*)(ctx + off) accesses */
11632 ret = convert_ctx_accesses(env);
11633
11634 if (ret == 0)
11635 ret = fixup_bpf_calls(env);
11636
11637 /* do 32-bit optimization after insn patching has done so those patched
11638 * insns could be handled correctly.
11639 */
11640 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
11641 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
11642 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
11643 : false;
11644 }
11645
11646 if (ret == 0)
11647 ret = fixup_call_args(env);
11648
11649 env->verification_time = ktime_get_ns() - start_time;
11650 print_verification_stats(env);
11651
11652 if (log->level && bpf_verifier_log_full(log))
11653 ret = -ENOSPC;
11654 if (log->level && !log->ubuf) {
11655 ret = -EFAULT;
11656 goto err_release_maps;
11657 }
11658
11659 if (ret == 0 && env->used_map_cnt) {
11660 /* if program passed verifier, update used_maps in bpf_prog_info */
11661 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
11662 sizeof(env->used_maps[0]),
11663 GFP_KERNEL);
11664
11665 if (!env->prog->aux->used_maps) {
11666 ret = -ENOMEM;
11667 goto err_release_maps;
11668 }
11669
11670 memcpy(env->prog->aux->used_maps, env->used_maps,
11671 sizeof(env->used_maps[0]) * env->used_map_cnt);
11672 env->prog->aux->used_map_cnt = env->used_map_cnt;
11673
11674 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
11675 * bpf_ld_imm64 instructions
11676 */
11677 convert_pseudo_ld_imm64(env);
11678 }
11679
11680 if (ret == 0)
11681 adjust_btf_func(env);
11682
11683 err_release_maps:
11684 if (!env->prog->aux->used_maps)
11685 /* if we didn't copy map pointers into bpf_prog_info, release
11686 * them now. Otherwise free_used_maps() will release them.
11687 */
11688 release_maps(env);
11689
11690 /* extension progs temporarily inherit the attach_type of their targets
11691 for verification purposes, so set it back to zero before returning
11692 */
11693 if (env->prog->type == BPF_PROG_TYPE_EXT)
11694 env->prog->expected_attach_type = 0;
11695
11696 *prog = env->prog;
11697 err_unlock:
11698 if (!is_priv)
11699 mutex_unlock(&bpf_verifier_lock);
11700 vfree(env->insn_aux_data);
11701 err_free_env:
11702 kfree(env);
11703 return ret;
11704 }