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
bpf_map_ptr_poisoned(const struct bpf_insn_aux_data * aux)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
bpf_map_ptr_unpriv(const struct bpf_insn_aux_data * aux)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
bpf_map_ptr_store(struct bpf_insn_aux_data * aux,const struct bpf_map * map,bool unpriv)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
bpf_map_key_poisoned(const struct bpf_insn_aux_data * aux)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
bpf_map_key_unseen(const struct bpf_insn_aux_data * aux)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
bpf_map_key_immediate(const struct bpf_insn_aux_data * aux)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
bpf_map_key_store(struct bpf_insn_aux_data * aux,u64 state)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
bpf_pseudo_call(const struct bpf_insn * insn)231 static bool bpf_pseudo_call(const struct bpf_insn *insn)
232 {
233 return insn->code == (BPF_JMP | BPF_CALL) &&
234 insn->src_reg == BPF_PSEUDO_CALL;
235 }
236
bpf_pseudo_kfunc_call(const struct bpf_insn * insn)237 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
238 {
239 return insn->code == (BPF_JMP | BPF_CALL) &&
240 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
241 }
242
bpf_pseudo_func(const struct bpf_insn * insn)243 static bool bpf_pseudo_func(const struct bpf_insn *insn)
244 {
245 return insn->code == (BPF_LD | BPF_IMM | BPF_DW) &&
246 insn->src_reg == BPF_PSEUDO_FUNC;
247 }
248
249 struct bpf_call_arg_meta {
250 struct bpf_map *map_ptr;
251 bool raw_mode;
252 bool pkt_access;
253 int regno;
254 int access_size;
255 int mem_size;
256 u64 msize_max_value;
257 int ref_obj_id;
258 int func_id;
259 struct btf *btf;
260 u32 btf_id;
261 struct btf *ret_btf;
262 u32 ret_btf_id;
263 u32 subprogno;
264 };
265
266 struct btf *btf_vmlinux;
267
268 static DEFINE_MUTEX(bpf_verifier_lock);
269
270 static const struct bpf_line_info *
find_linfo(const struct bpf_verifier_env * env,u32 insn_off)271 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
272 {
273 const struct bpf_line_info *linfo;
274 const struct bpf_prog *prog;
275 u32 i, nr_linfo;
276
277 prog = env->prog;
278 nr_linfo = prog->aux->nr_linfo;
279
280 if (!nr_linfo || insn_off >= prog->len)
281 return NULL;
282
283 linfo = prog->aux->linfo;
284 for (i = 1; i < nr_linfo; i++)
285 if (insn_off < linfo[i].insn_off)
286 break;
287
288 return &linfo[i - 1];
289 }
290
bpf_verifier_vlog(struct bpf_verifier_log * log,const char * fmt,va_list args)291 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
292 va_list args)
293 {
294 unsigned int n;
295
296 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
297
298 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
299 "verifier log line truncated - local buffer too short\n");
300
301 n = min(log->len_total - log->len_used - 1, n);
302 log->kbuf[n] = '\0';
303
304 if (log->level == BPF_LOG_KERNEL) {
305 pr_err("BPF:%s\n", log->kbuf);
306 return;
307 }
308 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
309 log->len_used += n;
310 else
311 log->ubuf = NULL;
312 }
313
bpf_vlog_reset(struct bpf_verifier_log * log,u32 new_pos)314 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
315 {
316 char zero = 0;
317
318 if (!bpf_verifier_log_needed(log))
319 return;
320
321 log->len_used = new_pos;
322 if (put_user(zero, log->ubuf + new_pos))
323 log->ubuf = NULL;
324 }
325
326 /* log_level controls verbosity level of eBPF verifier.
327 * bpf_verifier_log_write() is used to dump the verification trace to the log,
328 * so the user can figure out what's wrong with the program
329 */
bpf_verifier_log_write(struct bpf_verifier_env * env,const char * fmt,...)330 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
331 const char *fmt, ...)
332 {
333 va_list args;
334
335 if (!bpf_verifier_log_needed(&env->log))
336 return;
337
338 va_start(args, fmt);
339 bpf_verifier_vlog(&env->log, fmt, args);
340 va_end(args);
341 }
342 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
343
verbose(void * private_data,const char * fmt,...)344 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
345 {
346 struct bpf_verifier_env *env = private_data;
347 va_list args;
348
349 if (!bpf_verifier_log_needed(&env->log))
350 return;
351
352 va_start(args, fmt);
353 bpf_verifier_vlog(&env->log, fmt, args);
354 va_end(args);
355 }
356
bpf_log(struct bpf_verifier_log * log,const char * fmt,...)357 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
358 const char *fmt, ...)
359 {
360 va_list args;
361
362 if (!bpf_verifier_log_needed(log))
363 return;
364
365 va_start(args, fmt);
366 bpf_verifier_vlog(log, fmt, args);
367 va_end(args);
368 }
369
ltrim(const char * s)370 static const char *ltrim(const char *s)
371 {
372 while (isspace(*s))
373 s++;
374
375 return s;
376 }
377
verbose_linfo(struct bpf_verifier_env * env,u32 insn_off,const char * prefix_fmt,...)378 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
379 u32 insn_off,
380 const char *prefix_fmt, ...)
381 {
382 const struct bpf_line_info *linfo;
383
384 if (!bpf_verifier_log_needed(&env->log))
385 return;
386
387 linfo = find_linfo(env, insn_off);
388 if (!linfo || linfo == env->prev_linfo)
389 return;
390
391 if (prefix_fmt) {
392 va_list args;
393
394 va_start(args, prefix_fmt);
395 bpf_verifier_vlog(&env->log, prefix_fmt, args);
396 va_end(args);
397 }
398
399 verbose(env, "%s\n",
400 ltrim(btf_name_by_offset(env->prog->aux->btf,
401 linfo->line_off)));
402
403 env->prev_linfo = linfo;
404 }
405
verbose_invalid_scalar(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct tnum * range,const char * ctx,const char * reg_name)406 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
407 struct bpf_reg_state *reg,
408 struct tnum *range, const char *ctx,
409 const char *reg_name)
410 {
411 char tn_buf[48];
412
413 verbose(env, "At %s the register %s ", ctx, reg_name);
414 if (!tnum_is_unknown(reg->var_off)) {
415 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
416 verbose(env, "has value %s", tn_buf);
417 } else {
418 verbose(env, "has unknown scalar value");
419 }
420 tnum_strn(tn_buf, sizeof(tn_buf), *range);
421 verbose(env, " should have been in %s\n", tn_buf);
422 }
423
type_is_pkt_pointer(enum bpf_reg_type type)424 static bool type_is_pkt_pointer(enum bpf_reg_type type)
425 {
426 return type == PTR_TO_PACKET ||
427 type == PTR_TO_PACKET_META;
428 }
429
type_is_sk_pointer(enum bpf_reg_type type)430 static bool type_is_sk_pointer(enum bpf_reg_type type)
431 {
432 return type == PTR_TO_SOCKET ||
433 type == PTR_TO_SOCK_COMMON ||
434 type == PTR_TO_TCP_SOCK ||
435 type == PTR_TO_XDP_SOCK;
436 }
437
reg_type_not_null(enum bpf_reg_type type)438 static bool reg_type_not_null(enum bpf_reg_type type)
439 {
440 return type == PTR_TO_SOCKET ||
441 type == PTR_TO_TCP_SOCK ||
442 type == PTR_TO_MAP_VALUE ||
443 type == PTR_TO_MAP_KEY ||
444 type == PTR_TO_SOCK_COMMON;
445 }
446
reg_type_may_be_null(enum bpf_reg_type type)447 static bool reg_type_may_be_null(enum bpf_reg_type type)
448 {
449 return type == PTR_TO_MAP_VALUE_OR_NULL ||
450 type == PTR_TO_SOCKET_OR_NULL ||
451 type == PTR_TO_SOCK_COMMON_OR_NULL ||
452 type == PTR_TO_TCP_SOCK_OR_NULL ||
453 type == PTR_TO_BTF_ID_OR_NULL ||
454 type == PTR_TO_MEM_OR_NULL ||
455 type == PTR_TO_RDONLY_BUF_OR_NULL ||
456 type == PTR_TO_RDWR_BUF_OR_NULL;
457 }
458
reg_may_point_to_spin_lock(const struct bpf_reg_state * reg)459 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
460 {
461 return reg->type == PTR_TO_MAP_VALUE &&
462 map_value_has_spin_lock(reg->map_ptr);
463 }
464
reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)465 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
466 {
467 return type == PTR_TO_SOCKET ||
468 type == PTR_TO_SOCKET_OR_NULL ||
469 type == PTR_TO_TCP_SOCK ||
470 type == PTR_TO_TCP_SOCK_OR_NULL ||
471 type == PTR_TO_MEM ||
472 type == PTR_TO_MEM_OR_NULL;
473 }
474
arg_type_may_be_refcounted(enum bpf_arg_type type)475 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
476 {
477 return type == ARG_PTR_TO_SOCK_COMMON;
478 }
479
arg_type_may_be_null(enum bpf_arg_type type)480 static bool arg_type_may_be_null(enum bpf_arg_type type)
481 {
482 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
483 type == ARG_PTR_TO_MEM_OR_NULL ||
484 type == ARG_PTR_TO_CTX_OR_NULL ||
485 type == ARG_PTR_TO_SOCKET_OR_NULL ||
486 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL ||
487 type == ARG_PTR_TO_STACK_OR_NULL;
488 }
489
490 /* Determine whether the function releases some resources allocated by another
491 * function call. The first reference type argument will be assumed to be
492 * released by release_reference().
493 */
is_release_function(enum bpf_func_id func_id)494 static bool is_release_function(enum bpf_func_id func_id)
495 {
496 return func_id == BPF_FUNC_sk_release ||
497 func_id == BPF_FUNC_ringbuf_submit ||
498 func_id == BPF_FUNC_ringbuf_discard;
499 }
500
may_be_acquire_function(enum bpf_func_id func_id)501 static bool may_be_acquire_function(enum bpf_func_id func_id)
502 {
503 return func_id == BPF_FUNC_sk_lookup_tcp ||
504 func_id == BPF_FUNC_sk_lookup_udp ||
505 func_id == BPF_FUNC_skc_lookup_tcp ||
506 func_id == BPF_FUNC_map_lookup_elem ||
507 func_id == BPF_FUNC_ringbuf_reserve;
508 }
509
is_acquire_function(enum bpf_func_id func_id,const struct bpf_map * map)510 static bool is_acquire_function(enum bpf_func_id func_id,
511 const struct bpf_map *map)
512 {
513 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
514
515 if (func_id == BPF_FUNC_sk_lookup_tcp ||
516 func_id == BPF_FUNC_sk_lookup_udp ||
517 func_id == BPF_FUNC_skc_lookup_tcp ||
518 func_id == BPF_FUNC_ringbuf_reserve)
519 return true;
520
521 if (func_id == BPF_FUNC_map_lookup_elem &&
522 (map_type == BPF_MAP_TYPE_SOCKMAP ||
523 map_type == BPF_MAP_TYPE_SOCKHASH))
524 return true;
525
526 return false;
527 }
528
is_ptr_cast_function(enum bpf_func_id func_id)529 static bool is_ptr_cast_function(enum bpf_func_id func_id)
530 {
531 return func_id == BPF_FUNC_tcp_sock ||
532 func_id == BPF_FUNC_sk_fullsock ||
533 func_id == BPF_FUNC_skc_to_tcp_sock ||
534 func_id == BPF_FUNC_skc_to_tcp6_sock ||
535 func_id == BPF_FUNC_skc_to_udp6_sock ||
536 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
537 func_id == BPF_FUNC_skc_to_tcp_request_sock;
538 }
539
is_cmpxchg_insn(const struct bpf_insn * insn)540 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
541 {
542 return BPF_CLASS(insn->code) == BPF_STX &&
543 BPF_MODE(insn->code) == BPF_ATOMIC &&
544 insn->imm == BPF_CMPXCHG;
545 }
546
547 /* string representation of 'enum bpf_reg_type' */
548 static const char * const reg_type_str[] = {
549 [NOT_INIT] = "?",
550 [SCALAR_VALUE] = "inv",
551 [PTR_TO_CTX] = "ctx",
552 [CONST_PTR_TO_MAP] = "map_ptr",
553 [PTR_TO_MAP_VALUE] = "map_value",
554 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
555 [PTR_TO_STACK] = "fp",
556 [PTR_TO_PACKET] = "pkt",
557 [PTR_TO_PACKET_META] = "pkt_meta",
558 [PTR_TO_PACKET_END] = "pkt_end",
559 [PTR_TO_FLOW_KEYS] = "flow_keys",
560 [PTR_TO_SOCKET] = "sock",
561 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
562 [PTR_TO_SOCK_COMMON] = "sock_common",
563 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
564 [PTR_TO_TCP_SOCK] = "tcp_sock",
565 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
566 [PTR_TO_TP_BUFFER] = "tp_buffer",
567 [PTR_TO_XDP_SOCK] = "xdp_sock",
568 [PTR_TO_BTF_ID] = "ptr_",
569 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
570 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
571 [PTR_TO_MEM] = "mem",
572 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
573 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
574 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
575 [PTR_TO_RDWR_BUF] = "rdwr_buf",
576 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
577 [PTR_TO_FUNC] = "func",
578 [PTR_TO_MAP_KEY] = "map_key",
579 };
580
581 static char slot_type_char[] = {
582 [STACK_INVALID] = '?',
583 [STACK_SPILL] = 'r',
584 [STACK_MISC] = 'm',
585 [STACK_ZERO] = '0',
586 };
587
print_liveness(struct bpf_verifier_env * env,enum bpf_reg_liveness live)588 static void print_liveness(struct bpf_verifier_env *env,
589 enum bpf_reg_liveness live)
590 {
591 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
592 verbose(env, "_");
593 if (live & REG_LIVE_READ)
594 verbose(env, "r");
595 if (live & REG_LIVE_WRITTEN)
596 verbose(env, "w");
597 if (live & REG_LIVE_DONE)
598 verbose(env, "D");
599 }
600
func(struct bpf_verifier_env * env,const struct bpf_reg_state * reg)601 static struct bpf_func_state *func(struct bpf_verifier_env *env,
602 const struct bpf_reg_state *reg)
603 {
604 struct bpf_verifier_state *cur = env->cur_state;
605
606 return cur->frame[reg->frameno];
607 }
608
kernel_type_name(const struct btf * btf,u32 id)609 static const char *kernel_type_name(const struct btf* btf, u32 id)
610 {
611 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
612 }
613
print_verifier_state(struct bpf_verifier_env * env,const struct bpf_func_state * state)614 static void print_verifier_state(struct bpf_verifier_env *env,
615 const struct bpf_func_state *state)
616 {
617 const struct bpf_reg_state *reg;
618 enum bpf_reg_type t;
619 int i;
620
621 if (state->frameno)
622 verbose(env, " frame%d:", state->frameno);
623 for (i = 0; i < MAX_BPF_REG; i++) {
624 reg = &state->regs[i];
625 t = reg->type;
626 if (t == NOT_INIT)
627 continue;
628 verbose(env, " R%d", i);
629 print_liveness(env, reg->live);
630 verbose(env, "=%s", reg_type_str[t]);
631 if (t == SCALAR_VALUE && reg->precise)
632 verbose(env, "P");
633 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
634 tnum_is_const(reg->var_off)) {
635 /* reg->off should be 0 for SCALAR_VALUE */
636 verbose(env, "%lld", reg->var_off.value + reg->off);
637 } else {
638 if (t == PTR_TO_BTF_ID ||
639 t == PTR_TO_BTF_ID_OR_NULL ||
640 t == PTR_TO_PERCPU_BTF_ID)
641 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
642 verbose(env, "(id=%d", reg->id);
643 if (reg_type_may_be_refcounted_or_null(t))
644 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
645 if (t != SCALAR_VALUE)
646 verbose(env, ",off=%d", reg->off);
647 if (type_is_pkt_pointer(t))
648 verbose(env, ",r=%d", reg->range);
649 else if (t == CONST_PTR_TO_MAP ||
650 t == PTR_TO_MAP_KEY ||
651 t == PTR_TO_MAP_VALUE ||
652 t == PTR_TO_MAP_VALUE_OR_NULL)
653 verbose(env, ",ks=%d,vs=%d",
654 reg->map_ptr->key_size,
655 reg->map_ptr->value_size);
656 if (tnum_is_const(reg->var_off)) {
657 /* Typically an immediate SCALAR_VALUE, but
658 * could be a pointer whose offset is too big
659 * for reg->off
660 */
661 verbose(env, ",imm=%llx", reg->var_off.value);
662 } else {
663 if (reg->smin_value != reg->umin_value &&
664 reg->smin_value != S64_MIN)
665 verbose(env, ",smin_value=%lld",
666 (long long)reg->smin_value);
667 if (reg->smax_value != reg->umax_value &&
668 reg->smax_value != S64_MAX)
669 verbose(env, ",smax_value=%lld",
670 (long long)reg->smax_value);
671 if (reg->umin_value != 0)
672 verbose(env, ",umin_value=%llu",
673 (unsigned long long)reg->umin_value);
674 if (reg->umax_value != U64_MAX)
675 verbose(env, ",umax_value=%llu",
676 (unsigned long long)reg->umax_value);
677 if (!tnum_is_unknown(reg->var_off)) {
678 char tn_buf[48];
679
680 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
681 verbose(env, ",var_off=%s", tn_buf);
682 }
683 if (reg->s32_min_value != reg->smin_value &&
684 reg->s32_min_value != S32_MIN)
685 verbose(env, ",s32_min_value=%d",
686 (int)(reg->s32_min_value));
687 if (reg->s32_max_value != reg->smax_value &&
688 reg->s32_max_value != S32_MAX)
689 verbose(env, ",s32_max_value=%d",
690 (int)(reg->s32_max_value));
691 if (reg->u32_min_value != reg->umin_value &&
692 reg->u32_min_value != U32_MIN)
693 verbose(env, ",u32_min_value=%d",
694 (int)(reg->u32_min_value));
695 if (reg->u32_max_value != reg->umax_value &&
696 reg->u32_max_value != U32_MAX)
697 verbose(env, ",u32_max_value=%d",
698 (int)(reg->u32_max_value));
699 }
700 verbose(env, ")");
701 }
702 }
703 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
704 char types_buf[BPF_REG_SIZE + 1];
705 bool valid = false;
706 int j;
707
708 for (j = 0; j < BPF_REG_SIZE; j++) {
709 if (state->stack[i].slot_type[j] != STACK_INVALID)
710 valid = true;
711 types_buf[j] = slot_type_char[
712 state->stack[i].slot_type[j]];
713 }
714 types_buf[BPF_REG_SIZE] = 0;
715 if (!valid)
716 continue;
717 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
718 print_liveness(env, state->stack[i].spilled_ptr.live);
719 if (state->stack[i].slot_type[0] == STACK_SPILL) {
720 reg = &state->stack[i].spilled_ptr;
721 t = reg->type;
722 verbose(env, "=%s", reg_type_str[t]);
723 if (t == SCALAR_VALUE && reg->precise)
724 verbose(env, "P");
725 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
726 verbose(env, "%lld", reg->var_off.value + reg->off);
727 } else {
728 verbose(env, "=%s", types_buf);
729 }
730 }
731 if (state->acquired_refs && state->refs[0].id) {
732 verbose(env, " refs=%d", state->refs[0].id);
733 for (i = 1; i < state->acquired_refs; i++)
734 if (state->refs[i].id)
735 verbose(env, ",%d", state->refs[i].id);
736 }
737 verbose(env, "\n");
738 }
739
740 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
741 static int copy_##NAME##_state(struct bpf_func_state *dst, \
742 const struct bpf_func_state *src) \
743 { \
744 if (!src->FIELD) \
745 return 0; \
746 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
747 /* internal bug, make state invalid to reject the program */ \
748 memset(dst, 0, sizeof(*dst)); \
749 return -EFAULT; \
750 } \
751 memcpy(dst->FIELD, src->FIELD, \
752 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
753 return 0; \
754 }
755 /* copy_reference_state() */
756 COPY_STATE_FN(reference, acquired_refs, refs, 1)
757 /* copy_stack_state() */
COPY_STATE_FN(stack,allocated_stack,stack,BPF_REG_SIZE)758 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
759 #undef COPY_STATE_FN
760
761 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
762 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
763 bool copy_old) \
764 { \
765 u32 old_size = state->COUNT; \
766 struct bpf_##NAME##_state *new_##FIELD; \
767 int slot = size / SIZE; \
768 \
769 if (size <= old_size || !size) { \
770 if (copy_old) \
771 return 0; \
772 state->COUNT = slot * SIZE; \
773 if (!size && old_size) { \
774 kfree(state->FIELD); \
775 state->FIELD = NULL; \
776 } \
777 return 0; \
778 } \
779 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
780 GFP_KERNEL); \
781 if (!new_##FIELD) \
782 return -ENOMEM; \
783 if (copy_old) { \
784 if (state->FIELD) \
785 memcpy(new_##FIELD, state->FIELD, \
786 sizeof(*new_##FIELD) * (old_size / SIZE)); \
787 memset(new_##FIELD + old_size / SIZE, 0, \
788 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
789 } \
790 state->COUNT = slot * SIZE; \
791 kfree(state->FIELD); \
792 state->FIELD = new_##FIELD; \
793 return 0; \
794 }
795 /* realloc_reference_state() */
796 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
797 /* realloc_stack_state() */
798 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
799 #undef REALLOC_STATE_FN
800
801 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
802 * make it consume minimal amount of memory. check_stack_write() access from
803 * the program calls into realloc_func_state() to grow the stack size.
804 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
805 * which realloc_stack_state() copies over. It points to previous
806 * bpf_verifier_state which is never reallocated.
807 */
808 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
809 int refs_size, bool copy_old)
810 {
811 int err = realloc_reference_state(state, refs_size, copy_old);
812 if (err)
813 return err;
814 return realloc_stack_state(state, stack_size, copy_old);
815 }
816
817 /* Acquire a pointer id from the env and update the state->refs to include
818 * this new pointer reference.
819 * On success, returns a valid pointer id to associate with the register
820 * On failure, returns a negative errno.
821 */
acquire_reference_state(struct bpf_verifier_env * env,int insn_idx)822 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
823 {
824 struct bpf_func_state *state = cur_func(env);
825 int new_ofs = state->acquired_refs;
826 int id, err;
827
828 err = realloc_reference_state(state, state->acquired_refs + 1, true);
829 if (err)
830 return err;
831 id = ++env->id_gen;
832 state->refs[new_ofs].id = id;
833 state->refs[new_ofs].insn_idx = insn_idx;
834
835 return id;
836 }
837
838 /* release function corresponding to acquire_reference_state(). Idempotent. */
release_reference_state(struct bpf_func_state * state,int ptr_id)839 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
840 {
841 int i, last_idx;
842
843 last_idx = state->acquired_refs - 1;
844 for (i = 0; i < state->acquired_refs; i++) {
845 if (state->refs[i].id == ptr_id) {
846 if (last_idx && i != last_idx)
847 memcpy(&state->refs[i], &state->refs[last_idx],
848 sizeof(*state->refs));
849 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
850 state->acquired_refs--;
851 return 0;
852 }
853 }
854 return -EINVAL;
855 }
856
transfer_reference_state(struct bpf_func_state * dst,struct bpf_func_state * src)857 static int transfer_reference_state(struct bpf_func_state *dst,
858 struct bpf_func_state *src)
859 {
860 int err = realloc_reference_state(dst, src->acquired_refs, false);
861 if (err)
862 return err;
863 err = copy_reference_state(dst, src);
864 if (err)
865 return err;
866 return 0;
867 }
868
free_func_state(struct bpf_func_state * state)869 static void free_func_state(struct bpf_func_state *state)
870 {
871 if (!state)
872 return;
873 kfree(state->refs);
874 kfree(state->stack);
875 kfree(state);
876 }
877
clear_jmp_history(struct bpf_verifier_state * state)878 static void clear_jmp_history(struct bpf_verifier_state *state)
879 {
880 kfree(state->jmp_history);
881 state->jmp_history = NULL;
882 state->jmp_history_cnt = 0;
883 }
884
free_verifier_state(struct bpf_verifier_state * state,bool free_self)885 static void free_verifier_state(struct bpf_verifier_state *state,
886 bool free_self)
887 {
888 int i;
889
890 for (i = 0; i <= state->curframe; i++) {
891 free_func_state(state->frame[i]);
892 state->frame[i] = NULL;
893 }
894 clear_jmp_history(state);
895 if (free_self)
896 kfree(state);
897 }
898
899 /* copy verifier state from src to dst growing dst stack space
900 * when necessary to accommodate larger src stack
901 */
copy_func_state(struct bpf_func_state * dst,const struct bpf_func_state * src)902 static int copy_func_state(struct bpf_func_state *dst,
903 const struct bpf_func_state *src)
904 {
905 int err;
906
907 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
908 false);
909 if (err)
910 return err;
911 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
912 err = copy_reference_state(dst, src);
913 if (err)
914 return err;
915 return copy_stack_state(dst, src);
916 }
917
copy_verifier_state(struct bpf_verifier_state * dst_state,const struct bpf_verifier_state * src)918 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
919 const struct bpf_verifier_state *src)
920 {
921 struct bpf_func_state *dst;
922 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
923 int i, err;
924
925 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
926 kfree(dst_state->jmp_history);
927 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
928 if (!dst_state->jmp_history)
929 return -ENOMEM;
930 }
931 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
932 dst_state->jmp_history_cnt = src->jmp_history_cnt;
933
934 /* if dst has more stack frames then src frame, free them */
935 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
936 free_func_state(dst_state->frame[i]);
937 dst_state->frame[i] = NULL;
938 }
939 dst_state->speculative = src->speculative;
940 dst_state->curframe = src->curframe;
941 dst_state->active_spin_lock = src->active_spin_lock;
942 dst_state->branches = src->branches;
943 dst_state->parent = src->parent;
944 dst_state->first_insn_idx = src->first_insn_idx;
945 dst_state->last_insn_idx = src->last_insn_idx;
946 for (i = 0; i <= src->curframe; i++) {
947 dst = dst_state->frame[i];
948 if (!dst) {
949 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
950 if (!dst)
951 return -ENOMEM;
952 dst_state->frame[i] = dst;
953 }
954 err = copy_func_state(dst, src->frame[i]);
955 if (err)
956 return err;
957 }
958 return 0;
959 }
960
update_branch_counts(struct bpf_verifier_env * env,struct bpf_verifier_state * st)961 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
962 {
963 while (st) {
964 u32 br = --st->branches;
965
966 /* WARN_ON(br > 1) technically makes sense here,
967 * but see comment in push_stack(), hence:
968 */
969 WARN_ONCE((int)br < 0,
970 "BUG update_branch_counts:branches_to_explore=%d\n",
971 br);
972 if (br)
973 break;
974 st = st->parent;
975 }
976 }
977
pop_stack(struct bpf_verifier_env * env,int * prev_insn_idx,int * insn_idx,bool pop_log)978 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
979 int *insn_idx, bool pop_log)
980 {
981 struct bpf_verifier_state *cur = env->cur_state;
982 struct bpf_verifier_stack_elem *elem, *head = env->head;
983 int err;
984
985 if (env->head == NULL)
986 return -ENOENT;
987
988 if (cur) {
989 err = copy_verifier_state(cur, &head->st);
990 if (err)
991 return err;
992 }
993 if (pop_log)
994 bpf_vlog_reset(&env->log, head->log_pos);
995 if (insn_idx)
996 *insn_idx = head->insn_idx;
997 if (prev_insn_idx)
998 *prev_insn_idx = head->prev_insn_idx;
999 elem = head->next;
1000 free_verifier_state(&head->st, false);
1001 kfree(head);
1002 env->head = elem;
1003 env->stack_size--;
1004 return 0;
1005 }
1006
push_stack(struct bpf_verifier_env * env,int insn_idx,int prev_insn_idx,bool speculative)1007 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1008 int insn_idx, int prev_insn_idx,
1009 bool speculative)
1010 {
1011 struct bpf_verifier_state *cur = env->cur_state;
1012 struct bpf_verifier_stack_elem *elem;
1013 int err;
1014
1015 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1016 if (!elem)
1017 goto err;
1018
1019 elem->insn_idx = insn_idx;
1020 elem->prev_insn_idx = prev_insn_idx;
1021 elem->next = env->head;
1022 elem->log_pos = env->log.len_used;
1023 env->head = elem;
1024 env->stack_size++;
1025 err = copy_verifier_state(&elem->st, cur);
1026 if (err)
1027 goto err;
1028 elem->st.speculative |= speculative;
1029 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1030 verbose(env, "The sequence of %d jumps is too complex.\n",
1031 env->stack_size);
1032 goto err;
1033 }
1034 if (elem->st.parent) {
1035 ++elem->st.parent->branches;
1036 /* WARN_ON(branches > 2) technically makes sense here,
1037 * but
1038 * 1. speculative states will bump 'branches' for non-branch
1039 * instructions
1040 * 2. is_state_visited() heuristics may decide not to create
1041 * a new state for a sequence of branches and all such current
1042 * and cloned states will be pointing to a single parent state
1043 * which might have large 'branches' count.
1044 */
1045 }
1046 return &elem->st;
1047 err:
1048 free_verifier_state(env->cur_state, true);
1049 env->cur_state = NULL;
1050 /* pop all elements and return */
1051 while (!pop_stack(env, NULL, NULL, false));
1052 return NULL;
1053 }
1054
1055 #define CALLER_SAVED_REGS 6
1056 static const int caller_saved[CALLER_SAVED_REGS] = {
1057 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1058 };
1059
1060 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1061 struct bpf_reg_state *reg);
1062
1063 /* This helper doesn't clear reg->id */
___mark_reg_known(struct bpf_reg_state * reg,u64 imm)1064 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1065 {
1066 reg->var_off = tnum_const(imm);
1067 reg->smin_value = (s64)imm;
1068 reg->smax_value = (s64)imm;
1069 reg->umin_value = imm;
1070 reg->umax_value = imm;
1071
1072 reg->s32_min_value = (s32)imm;
1073 reg->s32_max_value = (s32)imm;
1074 reg->u32_min_value = (u32)imm;
1075 reg->u32_max_value = (u32)imm;
1076 }
1077
1078 /* Mark the unknown part of a register (variable offset or scalar value) as
1079 * known to have the value @imm.
1080 */
__mark_reg_known(struct bpf_reg_state * reg,u64 imm)1081 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1082 {
1083 /* Clear id, off, and union(map_ptr, range) */
1084 memset(((u8 *)reg) + sizeof(reg->type), 0,
1085 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1086 ___mark_reg_known(reg, imm);
1087 }
1088
__mark_reg32_known(struct bpf_reg_state * reg,u64 imm)1089 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1090 {
1091 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1092 reg->s32_min_value = (s32)imm;
1093 reg->s32_max_value = (s32)imm;
1094 reg->u32_min_value = (u32)imm;
1095 reg->u32_max_value = (u32)imm;
1096 }
1097
1098 /* Mark the 'variable offset' part of a register as zero. This should be
1099 * used only on registers holding a pointer type.
1100 */
__mark_reg_known_zero(struct bpf_reg_state * reg)1101 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1102 {
1103 __mark_reg_known(reg, 0);
1104 }
1105
__mark_reg_const_zero(struct bpf_reg_state * reg)1106 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1107 {
1108 __mark_reg_known(reg, 0);
1109 reg->type = SCALAR_VALUE;
1110 }
1111
mark_reg_known_zero(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)1112 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1113 struct bpf_reg_state *regs, u32 regno)
1114 {
1115 if (WARN_ON(regno >= MAX_BPF_REG)) {
1116 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1117 /* Something bad happened, let's kill all regs */
1118 for (regno = 0; regno < MAX_BPF_REG; regno++)
1119 __mark_reg_not_init(env, regs + regno);
1120 return;
1121 }
1122 __mark_reg_known_zero(regs + regno);
1123 }
1124
mark_ptr_not_null_reg(struct bpf_reg_state * reg)1125 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1126 {
1127 switch (reg->type) {
1128 case PTR_TO_MAP_VALUE_OR_NULL: {
1129 const struct bpf_map *map = reg->map_ptr;
1130
1131 if (map->inner_map_meta) {
1132 reg->type = CONST_PTR_TO_MAP;
1133 reg->map_ptr = map->inner_map_meta;
1134 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1135 reg->type = PTR_TO_XDP_SOCK;
1136 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1137 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1138 reg->type = PTR_TO_SOCKET;
1139 } else {
1140 reg->type = PTR_TO_MAP_VALUE;
1141 }
1142 break;
1143 }
1144 case PTR_TO_SOCKET_OR_NULL:
1145 reg->type = PTR_TO_SOCKET;
1146 break;
1147 case PTR_TO_SOCK_COMMON_OR_NULL:
1148 reg->type = PTR_TO_SOCK_COMMON;
1149 break;
1150 case PTR_TO_TCP_SOCK_OR_NULL:
1151 reg->type = PTR_TO_TCP_SOCK;
1152 break;
1153 case PTR_TO_BTF_ID_OR_NULL:
1154 reg->type = PTR_TO_BTF_ID;
1155 break;
1156 case PTR_TO_MEM_OR_NULL:
1157 reg->type = PTR_TO_MEM;
1158 break;
1159 case PTR_TO_RDONLY_BUF_OR_NULL:
1160 reg->type = PTR_TO_RDONLY_BUF;
1161 break;
1162 case PTR_TO_RDWR_BUF_OR_NULL:
1163 reg->type = PTR_TO_RDWR_BUF;
1164 break;
1165 default:
1166 WARN_ONCE(1, "unknown nullable register type");
1167 }
1168 }
1169
reg_is_pkt_pointer(const struct bpf_reg_state * reg)1170 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1171 {
1172 return type_is_pkt_pointer(reg->type);
1173 }
1174
reg_is_pkt_pointer_any(const struct bpf_reg_state * reg)1175 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1176 {
1177 return reg_is_pkt_pointer(reg) ||
1178 reg->type == PTR_TO_PACKET_END;
1179 }
1180
1181 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
reg_is_init_pkt_pointer(const struct bpf_reg_state * reg,enum bpf_reg_type which)1182 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1183 enum bpf_reg_type which)
1184 {
1185 /* The register can already have a range from prior markings.
1186 * This is fine as long as it hasn't been advanced from its
1187 * origin.
1188 */
1189 return reg->type == which &&
1190 reg->id == 0 &&
1191 reg->off == 0 &&
1192 tnum_equals_const(reg->var_off, 0);
1193 }
1194
1195 /* Reset the min/max bounds of a register */
__mark_reg_unbounded(struct bpf_reg_state * reg)1196 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1197 {
1198 reg->smin_value = S64_MIN;
1199 reg->smax_value = S64_MAX;
1200 reg->umin_value = 0;
1201 reg->umax_value = U64_MAX;
1202
1203 reg->s32_min_value = S32_MIN;
1204 reg->s32_max_value = S32_MAX;
1205 reg->u32_min_value = 0;
1206 reg->u32_max_value = U32_MAX;
1207 }
1208
__mark_reg64_unbounded(struct bpf_reg_state * reg)1209 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1210 {
1211 reg->smin_value = S64_MIN;
1212 reg->smax_value = S64_MAX;
1213 reg->umin_value = 0;
1214 reg->umax_value = U64_MAX;
1215 }
1216
__mark_reg32_unbounded(struct bpf_reg_state * reg)1217 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1218 {
1219 reg->s32_min_value = S32_MIN;
1220 reg->s32_max_value = S32_MAX;
1221 reg->u32_min_value = 0;
1222 reg->u32_max_value = U32_MAX;
1223 }
1224
__update_reg32_bounds(struct bpf_reg_state * reg)1225 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1226 {
1227 struct tnum var32_off = tnum_subreg(reg->var_off);
1228
1229 /* min signed is max(sign bit) | min(other bits) */
1230 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1231 var32_off.value | (var32_off.mask & S32_MIN));
1232 /* max signed is min(sign bit) | max(other bits) */
1233 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1234 var32_off.value | (var32_off.mask & S32_MAX));
1235 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1236 reg->u32_max_value = min(reg->u32_max_value,
1237 (u32)(var32_off.value | var32_off.mask));
1238 }
1239
__update_reg64_bounds(struct bpf_reg_state * reg)1240 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1241 {
1242 /* min signed is max(sign bit) | min(other bits) */
1243 reg->smin_value = max_t(s64, reg->smin_value,
1244 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1245 /* max signed is min(sign bit) | max(other bits) */
1246 reg->smax_value = min_t(s64, reg->smax_value,
1247 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1248 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1249 reg->umax_value = min(reg->umax_value,
1250 reg->var_off.value | reg->var_off.mask);
1251 }
1252
__update_reg_bounds(struct bpf_reg_state * reg)1253 static void __update_reg_bounds(struct bpf_reg_state *reg)
1254 {
1255 __update_reg32_bounds(reg);
1256 __update_reg64_bounds(reg);
1257 }
1258
1259 /* Uses signed min/max values to inform unsigned, and vice-versa */
__reg32_deduce_bounds(struct bpf_reg_state * reg)1260 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1261 {
1262 /* Learn sign from signed bounds.
1263 * If we cannot cross the sign boundary, then signed and unsigned bounds
1264 * are the same, so combine. This works even in the negative case, e.g.
1265 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1266 */
1267 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1268 reg->s32_min_value = reg->u32_min_value =
1269 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1270 reg->s32_max_value = reg->u32_max_value =
1271 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1272 return;
1273 }
1274 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1275 * boundary, so we must be careful.
1276 */
1277 if ((s32)reg->u32_max_value >= 0) {
1278 /* Positive. We can't learn anything from the smin, but smax
1279 * is positive, hence safe.
1280 */
1281 reg->s32_min_value = reg->u32_min_value;
1282 reg->s32_max_value = reg->u32_max_value =
1283 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1284 } else if ((s32)reg->u32_min_value < 0) {
1285 /* Negative. We can't learn anything from the smax, but smin
1286 * is negative, hence safe.
1287 */
1288 reg->s32_min_value = reg->u32_min_value =
1289 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1290 reg->s32_max_value = reg->u32_max_value;
1291 }
1292 }
1293
__reg64_deduce_bounds(struct bpf_reg_state * reg)1294 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1295 {
1296 /* Learn sign from signed bounds.
1297 * If we cannot cross the sign boundary, then signed and unsigned bounds
1298 * are the same, so combine. This works even in the negative case, e.g.
1299 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1300 */
1301 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1302 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1303 reg->umin_value);
1304 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1305 reg->umax_value);
1306 return;
1307 }
1308 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1309 * boundary, so we must be careful.
1310 */
1311 if ((s64)reg->umax_value >= 0) {
1312 /* Positive. We can't learn anything from the smin, but smax
1313 * is positive, hence safe.
1314 */
1315 reg->smin_value = reg->umin_value;
1316 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1317 reg->umax_value);
1318 } else if ((s64)reg->umin_value < 0) {
1319 /* Negative. We can't learn anything from the smax, but smin
1320 * is negative, hence safe.
1321 */
1322 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1323 reg->umin_value);
1324 reg->smax_value = reg->umax_value;
1325 }
1326 }
1327
__reg_deduce_bounds(struct bpf_reg_state * reg)1328 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1329 {
1330 __reg32_deduce_bounds(reg);
1331 __reg64_deduce_bounds(reg);
1332 }
1333
1334 /* Attempts to improve var_off based on unsigned min/max information */
__reg_bound_offset(struct bpf_reg_state * reg)1335 static void __reg_bound_offset(struct bpf_reg_state *reg)
1336 {
1337 struct tnum var64_off = tnum_intersect(reg->var_off,
1338 tnum_range(reg->umin_value,
1339 reg->umax_value));
1340 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1341 tnum_range(reg->u32_min_value,
1342 reg->u32_max_value));
1343
1344 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1345 }
1346
__reg_assign_32_into_64(struct bpf_reg_state * reg)1347 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1348 {
1349 reg->umin_value = reg->u32_min_value;
1350 reg->umax_value = reg->u32_max_value;
1351 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1352 * but must be positive otherwise set to worse case bounds
1353 * and refine later from tnum.
1354 */
1355 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1356 reg->smax_value = reg->s32_max_value;
1357 else
1358 reg->smax_value = U32_MAX;
1359 if (reg->s32_min_value >= 0)
1360 reg->smin_value = reg->s32_min_value;
1361 else
1362 reg->smin_value = 0;
1363 }
1364
__reg_combine_32_into_64(struct bpf_reg_state * reg)1365 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1366 {
1367 /* special case when 64-bit register has upper 32-bit register
1368 * zeroed. Typically happens after zext or <<32, >>32 sequence
1369 * allowing us to use 32-bit bounds directly,
1370 */
1371 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1372 __reg_assign_32_into_64(reg);
1373 } else {
1374 /* Otherwise the best we can do is push lower 32bit known and
1375 * unknown bits into register (var_off set from jmp logic)
1376 * then learn as much as possible from the 64-bit tnum
1377 * known and unknown bits. The previous smin/smax bounds are
1378 * invalid here because of jmp32 compare so mark them unknown
1379 * so they do not impact tnum bounds calculation.
1380 */
1381 __mark_reg64_unbounded(reg);
1382 __update_reg_bounds(reg);
1383 }
1384
1385 /* Intersecting with the old var_off might have improved our bounds
1386 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1387 * then new var_off is (0; 0x7f...fc) which improves our umax.
1388 */
1389 __reg_deduce_bounds(reg);
1390 __reg_bound_offset(reg);
1391 __update_reg_bounds(reg);
1392 }
1393
__reg64_bound_s32(s64 a)1394 static bool __reg64_bound_s32(s64 a)
1395 {
1396 return a > S32_MIN && a < S32_MAX;
1397 }
1398
__reg64_bound_u32(u64 a)1399 static bool __reg64_bound_u32(u64 a)
1400 {
1401 return a > U32_MIN && a < U32_MAX;
1402 }
1403
__reg_combine_64_into_32(struct bpf_reg_state * reg)1404 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1405 {
1406 __mark_reg32_unbounded(reg);
1407
1408 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1409 reg->s32_min_value = (s32)reg->smin_value;
1410 reg->s32_max_value = (s32)reg->smax_value;
1411 }
1412 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1413 reg->u32_min_value = (u32)reg->umin_value;
1414 reg->u32_max_value = (u32)reg->umax_value;
1415 }
1416
1417 /* Intersecting with the old var_off might have improved our bounds
1418 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1419 * then new var_off is (0; 0x7f...fc) which improves our umax.
1420 */
1421 __reg_deduce_bounds(reg);
1422 __reg_bound_offset(reg);
1423 __update_reg_bounds(reg);
1424 }
1425
1426 /* Mark a register as having a completely unknown (scalar) value. */
__mark_reg_unknown(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)1427 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1428 struct bpf_reg_state *reg)
1429 {
1430 /*
1431 * Clear type, id, off, and union(map_ptr, range) and
1432 * padding between 'type' and union
1433 */
1434 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1435 reg->type = SCALAR_VALUE;
1436 reg->var_off = tnum_unknown;
1437 reg->frameno = 0;
1438 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1439 __mark_reg_unbounded(reg);
1440 }
1441
mark_reg_unknown(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)1442 static void mark_reg_unknown(struct bpf_verifier_env *env,
1443 struct bpf_reg_state *regs, u32 regno)
1444 {
1445 if (WARN_ON(regno >= MAX_BPF_REG)) {
1446 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1447 /* Something bad happened, let's kill all regs except FP */
1448 for (regno = 0; regno < BPF_REG_FP; regno++)
1449 __mark_reg_not_init(env, regs + regno);
1450 return;
1451 }
1452 __mark_reg_unknown(env, regs + regno);
1453 }
1454
__mark_reg_not_init(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)1455 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1456 struct bpf_reg_state *reg)
1457 {
1458 __mark_reg_unknown(env, reg);
1459 reg->type = NOT_INIT;
1460 }
1461
mark_reg_not_init(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)1462 static void mark_reg_not_init(struct bpf_verifier_env *env,
1463 struct bpf_reg_state *regs, u32 regno)
1464 {
1465 if (WARN_ON(regno >= MAX_BPF_REG)) {
1466 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1467 /* Something bad happened, let's kill all regs except FP */
1468 for (regno = 0; regno < BPF_REG_FP; regno++)
1469 __mark_reg_not_init(env, regs + regno);
1470 return;
1471 }
1472 __mark_reg_not_init(env, regs + regno);
1473 }
1474
mark_btf_ld_reg(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,enum bpf_reg_type reg_type,struct btf * btf,u32 btf_id)1475 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1476 struct bpf_reg_state *regs, u32 regno,
1477 enum bpf_reg_type reg_type,
1478 struct btf *btf, u32 btf_id)
1479 {
1480 if (reg_type == SCALAR_VALUE) {
1481 mark_reg_unknown(env, regs, regno);
1482 return;
1483 }
1484 mark_reg_known_zero(env, regs, regno);
1485 regs[regno].type = PTR_TO_BTF_ID;
1486 regs[regno].btf = btf;
1487 regs[regno].btf_id = btf_id;
1488 }
1489
1490 #define DEF_NOT_SUBREG (0)
init_reg_state(struct bpf_verifier_env * env,struct bpf_func_state * state)1491 static void init_reg_state(struct bpf_verifier_env *env,
1492 struct bpf_func_state *state)
1493 {
1494 struct bpf_reg_state *regs = state->regs;
1495 int i;
1496
1497 for (i = 0; i < MAX_BPF_REG; i++) {
1498 mark_reg_not_init(env, regs, i);
1499 regs[i].live = REG_LIVE_NONE;
1500 regs[i].parent = NULL;
1501 regs[i].subreg_def = DEF_NOT_SUBREG;
1502 }
1503
1504 /* frame pointer */
1505 regs[BPF_REG_FP].type = PTR_TO_STACK;
1506 mark_reg_known_zero(env, regs, BPF_REG_FP);
1507 regs[BPF_REG_FP].frameno = state->frameno;
1508 }
1509
1510 #define BPF_MAIN_FUNC (-1)
init_func_state(struct bpf_verifier_env * env,struct bpf_func_state * state,int callsite,int frameno,int subprogno)1511 static void init_func_state(struct bpf_verifier_env *env,
1512 struct bpf_func_state *state,
1513 int callsite, int frameno, int subprogno)
1514 {
1515 state->callsite = callsite;
1516 state->frameno = frameno;
1517 state->subprogno = subprogno;
1518 init_reg_state(env, state);
1519 }
1520
1521 enum reg_arg_type {
1522 SRC_OP, /* register is used as source operand */
1523 DST_OP, /* register is used as destination operand */
1524 DST_OP_NO_MARK /* same as above, check only, don't mark */
1525 };
1526
cmp_subprogs(const void * a,const void * b)1527 static int cmp_subprogs(const void *a, const void *b)
1528 {
1529 return ((struct bpf_subprog_info *)a)->start -
1530 ((struct bpf_subprog_info *)b)->start;
1531 }
1532
find_subprog(struct bpf_verifier_env * env,int off)1533 static int find_subprog(struct bpf_verifier_env *env, int off)
1534 {
1535 struct bpf_subprog_info *p;
1536
1537 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1538 sizeof(env->subprog_info[0]), cmp_subprogs);
1539 if (!p)
1540 return -ENOENT;
1541 return p - env->subprog_info;
1542
1543 }
1544
add_subprog(struct bpf_verifier_env * env,int off)1545 static int add_subprog(struct bpf_verifier_env *env, int off)
1546 {
1547 int insn_cnt = env->prog->len;
1548 int ret;
1549
1550 if (off >= insn_cnt || off < 0) {
1551 verbose(env, "call to invalid destination\n");
1552 return -EINVAL;
1553 }
1554 ret = find_subprog(env, off);
1555 if (ret >= 0)
1556 return ret;
1557 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1558 verbose(env, "too many subprograms\n");
1559 return -E2BIG;
1560 }
1561 /* determine subprog starts. The end is one before the next starts */
1562 env->subprog_info[env->subprog_cnt++].start = off;
1563 sort(env->subprog_info, env->subprog_cnt,
1564 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1565 return env->subprog_cnt - 1;
1566 }
1567
1568 struct bpf_kfunc_desc {
1569 struct btf_func_model func_model;
1570 u32 func_id;
1571 s32 imm;
1572 };
1573
1574 #define MAX_KFUNC_DESCS 256
1575 struct bpf_kfunc_desc_tab {
1576 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1577 u32 nr_descs;
1578 };
1579
kfunc_desc_cmp_by_id(const void * a,const void * b)1580 static int kfunc_desc_cmp_by_id(const void *a, const void *b)
1581 {
1582 const struct bpf_kfunc_desc *d0 = a;
1583 const struct bpf_kfunc_desc *d1 = b;
1584
1585 /* func_id is not greater than BTF_MAX_TYPE */
1586 return d0->func_id - d1->func_id;
1587 }
1588
1589 static const struct bpf_kfunc_desc *
find_kfunc_desc(const struct bpf_prog * prog,u32 func_id)1590 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
1591 {
1592 struct bpf_kfunc_desc desc = {
1593 .func_id = func_id,
1594 };
1595 struct bpf_kfunc_desc_tab *tab;
1596
1597 tab = prog->aux->kfunc_tab;
1598 return bsearch(&desc, tab->descs, tab->nr_descs,
1599 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
1600 }
1601
add_kfunc_call(struct bpf_verifier_env * env,u32 func_id)1602 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
1603 {
1604 const struct btf_type *func, *func_proto;
1605 struct bpf_kfunc_desc_tab *tab;
1606 struct bpf_prog_aux *prog_aux;
1607 struct bpf_kfunc_desc *desc;
1608 const char *func_name;
1609 unsigned long addr;
1610 int err;
1611
1612 prog_aux = env->prog->aux;
1613 tab = prog_aux->kfunc_tab;
1614 if (!tab) {
1615 if (!btf_vmlinux) {
1616 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1617 return -ENOTSUPP;
1618 }
1619
1620 if (!env->prog->jit_requested) {
1621 verbose(env, "JIT is required for calling kernel function\n");
1622 return -ENOTSUPP;
1623 }
1624
1625 if (!bpf_jit_supports_kfunc_call()) {
1626 verbose(env, "JIT does not support calling kernel function\n");
1627 return -ENOTSUPP;
1628 }
1629
1630 if (!env->prog->gpl_compatible) {
1631 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1632 return -EINVAL;
1633 }
1634
1635 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1636 if (!tab)
1637 return -ENOMEM;
1638 prog_aux->kfunc_tab = tab;
1639 }
1640
1641 if (find_kfunc_desc(env->prog, func_id))
1642 return 0;
1643
1644 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1645 verbose(env, "too many different kernel function calls\n");
1646 return -E2BIG;
1647 }
1648
1649 func = btf_type_by_id(btf_vmlinux, func_id);
1650 if (!func || !btf_type_is_func(func)) {
1651 verbose(env, "kernel btf_id %u is not a function\n",
1652 func_id);
1653 return -EINVAL;
1654 }
1655 func_proto = btf_type_by_id(btf_vmlinux, func->type);
1656 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1657 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1658 func_id);
1659 return -EINVAL;
1660 }
1661
1662 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
1663 addr = kallsyms_lookup_name(func_name);
1664 if (!addr) {
1665 verbose(env, "cannot find address for kernel function %s\n",
1666 func_name);
1667 return -EINVAL;
1668 }
1669
1670 desc = &tab->descs[tab->nr_descs++];
1671 desc->func_id = func_id;
1672 desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
1673 err = btf_distill_func_proto(&env->log, btf_vmlinux,
1674 func_proto, func_name,
1675 &desc->func_model);
1676 if (!err)
1677 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1678 kfunc_desc_cmp_by_id, NULL);
1679 return err;
1680 }
1681
kfunc_desc_cmp_by_imm(const void * a,const void * b)1682 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1683 {
1684 const struct bpf_kfunc_desc *d0 = a;
1685 const struct bpf_kfunc_desc *d1 = b;
1686
1687 if (d0->imm > d1->imm)
1688 return 1;
1689 else if (d0->imm < d1->imm)
1690 return -1;
1691 return 0;
1692 }
1693
sort_kfunc_descs_by_imm(struct bpf_prog * prog)1694 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1695 {
1696 struct bpf_kfunc_desc_tab *tab;
1697
1698 tab = prog->aux->kfunc_tab;
1699 if (!tab)
1700 return;
1701
1702 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1703 kfunc_desc_cmp_by_imm, NULL);
1704 }
1705
bpf_prog_has_kfunc_call(const struct bpf_prog * prog)1706 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1707 {
1708 return !!prog->aux->kfunc_tab;
1709 }
1710
1711 const struct btf_func_model *
bpf_jit_find_kfunc_model(const struct bpf_prog * prog,const struct bpf_insn * insn)1712 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1713 const struct bpf_insn *insn)
1714 {
1715 const struct bpf_kfunc_desc desc = {
1716 .imm = insn->imm,
1717 };
1718 const struct bpf_kfunc_desc *res;
1719 struct bpf_kfunc_desc_tab *tab;
1720
1721 tab = prog->aux->kfunc_tab;
1722 res = bsearch(&desc, tab->descs, tab->nr_descs,
1723 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1724
1725 return res ? &res->func_model : NULL;
1726 }
1727
add_subprog_and_kfunc(struct bpf_verifier_env * env)1728 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1729 {
1730 struct bpf_subprog_info *subprog = env->subprog_info;
1731 struct bpf_insn *insn = env->prog->insnsi;
1732 int i, ret, insn_cnt = env->prog->len;
1733
1734 /* Add entry function. */
1735 ret = add_subprog(env, 0);
1736 if (ret)
1737 return ret;
1738
1739 for (i = 0; i < insn_cnt; i++, insn++) {
1740 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1741 !bpf_pseudo_kfunc_call(insn))
1742 continue;
1743
1744 if (!env->bpf_capable) {
1745 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1746 return -EPERM;
1747 }
1748
1749 if (bpf_pseudo_func(insn)) {
1750 ret = add_subprog(env, i + insn->imm + 1);
1751 if (ret >= 0)
1752 /* remember subprog */
1753 insn[1].imm = ret;
1754 } else if (bpf_pseudo_call(insn)) {
1755 ret = add_subprog(env, i + insn->imm + 1);
1756 } else {
1757 ret = add_kfunc_call(env, insn->imm);
1758 }
1759
1760 if (ret < 0)
1761 return ret;
1762 }
1763
1764 /* Add a fake 'exit' subprog which could simplify subprog iteration
1765 * logic. 'subprog_cnt' should not be increased.
1766 */
1767 subprog[env->subprog_cnt].start = insn_cnt;
1768
1769 if (env->log.level & BPF_LOG_LEVEL2)
1770 for (i = 0; i < env->subprog_cnt; i++)
1771 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1772
1773 return 0;
1774 }
1775
check_subprogs(struct bpf_verifier_env * env)1776 static int check_subprogs(struct bpf_verifier_env *env)
1777 {
1778 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1779 struct bpf_subprog_info *subprog = env->subprog_info;
1780 struct bpf_insn *insn = env->prog->insnsi;
1781 int insn_cnt = env->prog->len;
1782
1783 /* now check that all jumps are within the same subprog */
1784 subprog_start = subprog[cur_subprog].start;
1785 subprog_end = subprog[cur_subprog + 1].start;
1786 for (i = 0; i < insn_cnt; i++) {
1787 u8 code = insn[i].code;
1788
1789 if (code == (BPF_JMP | BPF_CALL) &&
1790 insn[i].imm == BPF_FUNC_tail_call &&
1791 insn[i].src_reg != BPF_PSEUDO_CALL)
1792 subprog[cur_subprog].has_tail_call = true;
1793 if (BPF_CLASS(code) == BPF_LD &&
1794 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1795 subprog[cur_subprog].has_ld_abs = true;
1796 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1797 goto next;
1798 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1799 goto next;
1800 off = i + insn[i].off + 1;
1801 if (off < subprog_start || off >= subprog_end) {
1802 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1803 return -EINVAL;
1804 }
1805 next:
1806 if (i == subprog_end - 1) {
1807 /* to avoid fall-through from one subprog into another
1808 * the last insn of the subprog should be either exit
1809 * or unconditional jump back
1810 */
1811 if (code != (BPF_JMP | BPF_EXIT) &&
1812 code != (BPF_JMP | BPF_JA)) {
1813 verbose(env, "last insn is not an exit or jmp\n");
1814 return -EINVAL;
1815 }
1816 subprog_start = subprog_end;
1817 cur_subprog++;
1818 if (cur_subprog < env->subprog_cnt)
1819 subprog_end = subprog[cur_subprog + 1].start;
1820 }
1821 }
1822 return 0;
1823 }
1824
1825 /* Parentage chain of this register (or stack slot) should take care of all
1826 * issues like callee-saved registers, stack slot allocation time, etc.
1827 */
mark_reg_read(struct bpf_verifier_env * env,const struct bpf_reg_state * state,struct bpf_reg_state * parent,u8 flag)1828 static int mark_reg_read(struct bpf_verifier_env *env,
1829 const struct bpf_reg_state *state,
1830 struct bpf_reg_state *parent, u8 flag)
1831 {
1832 bool writes = parent == state->parent; /* Observe write marks */
1833 int cnt = 0;
1834
1835 while (parent) {
1836 /* if read wasn't screened by an earlier write ... */
1837 if (writes && state->live & REG_LIVE_WRITTEN)
1838 break;
1839 if (parent->live & REG_LIVE_DONE) {
1840 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1841 reg_type_str[parent->type],
1842 parent->var_off.value, parent->off);
1843 return -EFAULT;
1844 }
1845 /* The first condition is more likely to be true than the
1846 * second, checked it first.
1847 */
1848 if ((parent->live & REG_LIVE_READ) == flag ||
1849 parent->live & REG_LIVE_READ64)
1850 /* The parentage chain never changes and
1851 * this parent was already marked as LIVE_READ.
1852 * There is no need to keep walking the chain again and
1853 * keep re-marking all parents as LIVE_READ.
1854 * This case happens when the same register is read
1855 * multiple times without writes into it in-between.
1856 * Also, if parent has the stronger REG_LIVE_READ64 set,
1857 * then no need to set the weak REG_LIVE_READ32.
1858 */
1859 break;
1860 /* ... then we depend on parent's value */
1861 parent->live |= flag;
1862 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1863 if (flag == REG_LIVE_READ64)
1864 parent->live &= ~REG_LIVE_READ32;
1865 state = parent;
1866 parent = state->parent;
1867 writes = true;
1868 cnt++;
1869 }
1870
1871 if (env->longest_mark_read_walk < cnt)
1872 env->longest_mark_read_walk = cnt;
1873 return 0;
1874 }
1875
1876 /* This function is supposed to be used by the following 32-bit optimization
1877 * code only. It returns TRUE if the source or destination register operates
1878 * on 64-bit, otherwise return FALSE.
1879 */
is_reg64(struct bpf_verifier_env * env,struct bpf_insn * insn,u32 regno,struct bpf_reg_state * reg,enum reg_arg_type t)1880 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1881 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1882 {
1883 u8 code, class, op;
1884
1885 code = insn->code;
1886 class = BPF_CLASS(code);
1887 op = BPF_OP(code);
1888 if (class == BPF_JMP) {
1889 /* BPF_EXIT for "main" will reach here. Return TRUE
1890 * conservatively.
1891 */
1892 if (op == BPF_EXIT)
1893 return true;
1894 if (op == BPF_CALL) {
1895 /* BPF to BPF call will reach here because of marking
1896 * caller saved clobber with DST_OP_NO_MARK for which we
1897 * don't care the register def because they are anyway
1898 * marked as NOT_INIT already.
1899 */
1900 if (insn->src_reg == BPF_PSEUDO_CALL)
1901 return false;
1902 /* Helper call will reach here because of arg type
1903 * check, conservatively return TRUE.
1904 */
1905 if (t == SRC_OP)
1906 return true;
1907
1908 return false;
1909 }
1910 }
1911
1912 if (class == BPF_ALU64 || class == BPF_JMP ||
1913 /* BPF_END always use BPF_ALU class. */
1914 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1915 return true;
1916
1917 if (class == BPF_ALU || class == BPF_JMP32)
1918 return false;
1919
1920 if (class == BPF_LDX) {
1921 if (t != SRC_OP)
1922 return BPF_SIZE(code) == BPF_DW;
1923 /* LDX source must be ptr. */
1924 return true;
1925 }
1926
1927 if (class == BPF_STX) {
1928 /* BPF_STX (including atomic variants) has multiple source
1929 * operands, one of which is a ptr. Check whether the caller is
1930 * asking about it.
1931 */
1932 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1933 return true;
1934 return BPF_SIZE(code) == BPF_DW;
1935 }
1936
1937 if (class == BPF_LD) {
1938 u8 mode = BPF_MODE(code);
1939
1940 /* LD_IMM64 */
1941 if (mode == BPF_IMM)
1942 return true;
1943
1944 /* Both LD_IND and LD_ABS return 32-bit data. */
1945 if (t != SRC_OP)
1946 return false;
1947
1948 /* Implicit ctx ptr. */
1949 if (regno == BPF_REG_6)
1950 return true;
1951
1952 /* Explicit source could be any width. */
1953 return true;
1954 }
1955
1956 if (class == BPF_ST)
1957 /* The only source register for BPF_ST is a ptr. */
1958 return true;
1959
1960 /* Conservatively return true at default. */
1961 return true;
1962 }
1963
1964 /* Return the regno defined by the insn, or -1. */
insn_def_regno(const struct bpf_insn * insn)1965 static int insn_def_regno(const struct bpf_insn *insn)
1966 {
1967 switch (BPF_CLASS(insn->code)) {
1968 case BPF_JMP:
1969 case BPF_JMP32:
1970 case BPF_ST:
1971 return -1;
1972 case BPF_STX:
1973 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
1974 (insn->imm & BPF_FETCH)) {
1975 if (insn->imm == BPF_CMPXCHG)
1976 return BPF_REG_0;
1977 else
1978 return insn->src_reg;
1979 } else {
1980 return -1;
1981 }
1982 default:
1983 return insn->dst_reg;
1984 }
1985 }
1986
1987 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
insn_has_def32(struct bpf_verifier_env * env,struct bpf_insn * insn)1988 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1989 {
1990 int dst_reg = insn_def_regno(insn);
1991
1992 if (dst_reg == -1)
1993 return false;
1994
1995 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
1996 }
1997
mark_insn_zext(struct bpf_verifier_env * env,struct bpf_reg_state * reg)1998 static void mark_insn_zext(struct bpf_verifier_env *env,
1999 struct bpf_reg_state *reg)
2000 {
2001 s32 def_idx = reg->subreg_def;
2002
2003 if (def_idx == DEF_NOT_SUBREG)
2004 return;
2005
2006 env->insn_aux_data[def_idx - 1].zext_dst = true;
2007 /* The dst will be zero extended, so won't be sub-register anymore. */
2008 reg->subreg_def = DEF_NOT_SUBREG;
2009 }
2010
check_reg_arg(struct bpf_verifier_env * env,u32 regno,enum reg_arg_type t)2011 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2012 enum reg_arg_type t)
2013 {
2014 struct bpf_verifier_state *vstate = env->cur_state;
2015 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2016 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2017 struct bpf_reg_state *reg, *regs = state->regs;
2018 bool rw64;
2019
2020 if (regno >= MAX_BPF_REG) {
2021 verbose(env, "R%d is invalid\n", regno);
2022 return -EINVAL;
2023 }
2024
2025 reg = ®s[regno];
2026 rw64 = is_reg64(env, insn, regno, reg, t);
2027 if (t == SRC_OP) {
2028 /* check whether register used as source operand can be read */
2029 if (reg->type == NOT_INIT) {
2030 verbose(env, "R%d !read_ok\n", regno);
2031 return -EACCES;
2032 }
2033 /* We don't need to worry about FP liveness because it's read-only */
2034 if (regno == BPF_REG_FP)
2035 return 0;
2036
2037 if (rw64)
2038 mark_insn_zext(env, reg);
2039
2040 return mark_reg_read(env, reg, reg->parent,
2041 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2042 } else {
2043 /* check whether register used as dest operand can be written to */
2044 if (regno == BPF_REG_FP) {
2045 verbose(env, "frame pointer is read only\n");
2046 return -EACCES;
2047 }
2048 reg->live |= REG_LIVE_WRITTEN;
2049 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2050 if (t == DST_OP)
2051 mark_reg_unknown(env, regs, regno);
2052 }
2053 return 0;
2054 }
2055
2056 /* for any branch, call, exit record the history of jmps in the given state */
push_jmp_history(struct bpf_verifier_env * env,struct bpf_verifier_state * cur)2057 static int push_jmp_history(struct bpf_verifier_env *env,
2058 struct bpf_verifier_state *cur)
2059 {
2060 u32 cnt = cur->jmp_history_cnt;
2061 struct bpf_idx_pair *p;
2062
2063 cnt++;
2064 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2065 if (!p)
2066 return -ENOMEM;
2067 p[cnt - 1].idx = env->insn_idx;
2068 p[cnt - 1].prev_idx = env->prev_insn_idx;
2069 cur->jmp_history = p;
2070 cur->jmp_history_cnt = cnt;
2071 return 0;
2072 }
2073
2074 /* Backtrack one insn at a time. If idx is not at the top of recorded
2075 * history then previous instruction came from straight line execution.
2076 */
get_prev_insn_idx(struct bpf_verifier_state * st,int i,u32 * history)2077 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2078 u32 *history)
2079 {
2080 u32 cnt = *history;
2081
2082 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2083 i = st->jmp_history[cnt - 1].prev_idx;
2084 (*history)--;
2085 } else {
2086 i--;
2087 }
2088 return i;
2089 }
2090
disasm_kfunc_name(void * data,const struct bpf_insn * insn)2091 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2092 {
2093 const struct btf_type *func;
2094
2095 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2096 return NULL;
2097
2098 func = btf_type_by_id(btf_vmlinux, insn->imm);
2099 return btf_name_by_offset(btf_vmlinux, func->name_off);
2100 }
2101
2102 /* For given verifier state backtrack_insn() is called from the last insn to
2103 * the first insn. Its purpose is to compute a bitmask of registers and
2104 * stack slots that needs precision in the parent verifier state.
2105 */
backtrack_insn(struct bpf_verifier_env * env,int idx,u32 * reg_mask,u64 * stack_mask)2106 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2107 u32 *reg_mask, u64 *stack_mask)
2108 {
2109 const struct bpf_insn_cbs cbs = {
2110 .cb_call = disasm_kfunc_name,
2111 .cb_print = verbose,
2112 .private_data = env,
2113 };
2114 struct bpf_insn *insn = env->prog->insnsi + idx;
2115 u8 class = BPF_CLASS(insn->code);
2116 u8 opcode = BPF_OP(insn->code);
2117 u8 mode = BPF_MODE(insn->code);
2118 u32 dreg = 1u << insn->dst_reg;
2119 u32 sreg = 1u << insn->src_reg;
2120 u32 spi;
2121
2122 if (insn->code == 0)
2123 return 0;
2124 if (env->log.level & BPF_LOG_LEVEL) {
2125 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2126 verbose(env, "%d: ", idx);
2127 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2128 }
2129
2130 if (class == BPF_ALU || class == BPF_ALU64) {
2131 if (!(*reg_mask & dreg))
2132 return 0;
2133 if (opcode == BPF_MOV) {
2134 if (BPF_SRC(insn->code) == BPF_X) {
2135 /* dreg = sreg
2136 * dreg needs precision after this insn
2137 * sreg needs precision before this insn
2138 */
2139 *reg_mask &= ~dreg;
2140 *reg_mask |= sreg;
2141 } else {
2142 /* dreg = K
2143 * dreg needs precision after this insn.
2144 * Corresponding register is already marked
2145 * as precise=true in this verifier state.
2146 * No further markings in parent are necessary
2147 */
2148 *reg_mask &= ~dreg;
2149 }
2150 } else {
2151 if (BPF_SRC(insn->code) == BPF_X) {
2152 /* dreg += sreg
2153 * both dreg and sreg need precision
2154 * before this insn
2155 */
2156 *reg_mask |= sreg;
2157 } /* else dreg += K
2158 * dreg still needs precision before this insn
2159 */
2160 }
2161 } else if (class == BPF_LDX) {
2162 if (!(*reg_mask & dreg))
2163 return 0;
2164 *reg_mask &= ~dreg;
2165
2166 /* scalars can only be spilled into stack w/o losing precision.
2167 * Load from any other memory can be zero extended.
2168 * The desire to keep that precision is already indicated
2169 * by 'precise' mark in corresponding register of this state.
2170 * No further tracking necessary.
2171 */
2172 if (insn->src_reg != BPF_REG_FP)
2173 return 0;
2174 if (BPF_SIZE(insn->code) != BPF_DW)
2175 return 0;
2176
2177 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2178 * that [fp - off] slot contains scalar that needs to be
2179 * tracked with precision
2180 */
2181 spi = (-insn->off - 1) / BPF_REG_SIZE;
2182 if (spi >= 64) {
2183 verbose(env, "BUG spi %d\n", spi);
2184 WARN_ONCE(1, "verifier backtracking bug");
2185 return -EFAULT;
2186 }
2187 *stack_mask |= 1ull << spi;
2188 } else if (class == BPF_STX || class == BPF_ST) {
2189 if (*reg_mask & dreg)
2190 /* stx & st shouldn't be using _scalar_ dst_reg
2191 * to access memory. It means backtracking
2192 * encountered a case of pointer subtraction.
2193 */
2194 return -ENOTSUPP;
2195 /* scalars can only be spilled into stack */
2196 if (insn->dst_reg != BPF_REG_FP)
2197 return 0;
2198 if (BPF_SIZE(insn->code) != BPF_DW)
2199 return 0;
2200 spi = (-insn->off - 1) / BPF_REG_SIZE;
2201 if (spi >= 64) {
2202 verbose(env, "BUG spi %d\n", spi);
2203 WARN_ONCE(1, "verifier backtracking bug");
2204 return -EFAULT;
2205 }
2206 if (!(*stack_mask & (1ull << spi)))
2207 return 0;
2208 *stack_mask &= ~(1ull << spi);
2209 if (class == BPF_STX)
2210 *reg_mask |= sreg;
2211 } else if (class == BPF_JMP || class == BPF_JMP32) {
2212 if (opcode == BPF_CALL) {
2213 if (insn->src_reg == BPF_PSEUDO_CALL)
2214 return -ENOTSUPP;
2215 /* regular helper call sets R0 */
2216 *reg_mask &= ~1;
2217 if (*reg_mask & 0x3f) {
2218 /* if backtracing was looking for registers R1-R5
2219 * they should have been found already.
2220 */
2221 verbose(env, "BUG regs %x\n", *reg_mask);
2222 WARN_ONCE(1, "verifier backtracking bug");
2223 return -EFAULT;
2224 }
2225 } else if (opcode == BPF_EXIT) {
2226 return -ENOTSUPP;
2227 }
2228 } else if (class == BPF_LD) {
2229 if (!(*reg_mask & dreg))
2230 return 0;
2231 *reg_mask &= ~dreg;
2232 /* It's ld_imm64 or ld_abs or ld_ind.
2233 * For ld_imm64 no further tracking of precision
2234 * into parent is necessary
2235 */
2236 if (mode == BPF_IND || mode == BPF_ABS)
2237 /* to be analyzed */
2238 return -ENOTSUPP;
2239 }
2240 return 0;
2241 }
2242
2243 /* the scalar precision tracking algorithm:
2244 * . at the start all registers have precise=false.
2245 * . scalar ranges are tracked as normal through alu and jmp insns.
2246 * . once precise value of the scalar register is used in:
2247 * . ptr + scalar alu
2248 * . if (scalar cond K|scalar)
2249 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2250 * backtrack through the verifier states and mark all registers and
2251 * stack slots with spilled constants that these scalar regisers
2252 * should be precise.
2253 * . during state pruning two registers (or spilled stack slots)
2254 * are equivalent if both are not precise.
2255 *
2256 * Note the verifier cannot simply walk register parentage chain,
2257 * since many different registers and stack slots could have been
2258 * used to compute single precise scalar.
2259 *
2260 * The approach of starting with precise=true for all registers and then
2261 * backtrack to mark a register as not precise when the verifier detects
2262 * that program doesn't care about specific value (e.g., when helper
2263 * takes register as ARG_ANYTHING parameter) is not safe.
2264 *
2265 * It's ok to walk single parentage chain of the verifier states.
2266 * It's possible that this backtracking will go all the way till 1st insn.
2267 * All other branches will be explored for needing precision later.
2268 *
2269 * The backtracking needs to deal with cases like:
2270 * 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)
2271 * r9 -= r8
2272 * r5 = r9
2273 * if r5 > 0x79f goto pc+7
2274 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2275 * r5 += 1
2276 * ...
2277 * call bpf_perf_event_output#25
2278 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2279 *
2280 * and this case:
2281 * r6 = 1
2282 * call foo // uses callee's r6 inside to compute r0
2283 * r0 += r6
2284 * if r0 == 0 goto
2285 *
2286 * to track above reg_mask/stack_mask needs to be independent for each frame.
2287 *
2288 * Also if parent's curframe > frame where backtracking started,
2289 * the verifier need to mark registers in both frames, otherwise callees
2290 * may incorrectly prune callers. This is similar to
2291 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2292 *
2293 * For now backtracking falls back into conservative marking.
2294 */
mark_all_scalars_precise(struct bpf_verifier_env * env,struct bpf_verifier_state * st)2295 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2296 struct bpf_verifier_state *st)
2297 {
2298 struct bpf_func_state *func;
2299 struct bpf_reg_state *reg;
2300 int i, j;
2301
2302 /* big hammer: mark all scalars precise in this path.
2303 * pop_stack may still get !precise scalars.
2304 */
2305 for (; st; st = st->parent)
2306 for (i = 0; i <= st->curframe; i++) {
2307 func = st->frame[i];
2308 for (j = 0; j < BPF_REG_FP; j++) {
2309 reg = &func->regs[j];
2310 if (reg->type != SCALAR_VALUE)
2311 continue;
2312 reg->precise = true;
2313 }
2314 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2315 if (func->stack[j].slot_type[0] != STACK_SPILL)
2316 continue;
2317 reg = &func->stack[j].spilled_ptr;
2318 if (reg->type != SCALAR_VALUE)
2319 continue;
2320 reg->precise = true;
2321 }
2322 }
2323 }
2324
__mark_chain_precision(struct bpf_verifier_env * env,int regno,int spi)2325 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2326 int spi)
2327 {
2328 struct bpf_verifier_state *st = env->cur_state;
2329 int first_idx = st->first_insn_idx;
2330 int last_idx = env->insn_idx;
2331 struct bpf_func_state *func;
2332 struct bpf_reg_state *reg;
2333 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2334 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2335 bool skip_first = true;
2336 bool new_marks = false;
2337 int i, err;
2338
2339 if (!env->bpf_capable)
2340 return 0;
2341
2342 func = st->frame[st->curframe];
2343 if (regno >= 0) {
2344 reg = &func->regs[regno];
2345 if (reg->type != SCALAR_VALUE) {
2346 WARN_ONCE(1, "backtracing misuse");
2347 return -EFAULT;
2348 }
2349 if (!reg->precise)
2350 new_marks = true;
2351 else
2352 reg_mask = 0;
2353 reg->precise = true;
2354 }
2355
2356 while (spi >= 0) {
2357 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2358 stack_mask = 0;
2359 break;
2360 }
2361 reg = &func->stack[spi].spilled_ptr;
2362 if (reg->type != SCALAR_VALUE) {
2363 stack_mask = 0;
2364 break;
2365 }
2366 if (!reg->precise)
2367 new_marks = true;
2368 else
2369 stack_mask = 0;
2370 reg->precise = true;
2371 break;
2372 }
2373
2374 if (!new_marks)
2375 return 0;
2376 if (!reg_mask && !stack_mask)
2377 return 0;
2378 for (;;) {
2379 DECLARE_BITMAP(mask, 64);
2380 u32 history = st->jmp_history_cnt;
2381
2382 if (env->log.level & BPF_LOG_LEVEL)
2383 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2384 for (i = last_idx;;) {
2385 if (skip_first) {
2386 err = 0;
2387 skip_first = false;
2388 } else {
2389 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2390 }
2391 if (err == -ENOTSUPP) {
2392 mark_all_scalars_precise(env, st);
2393 return 0;
2394 } else if (err) {
2395 return err;
2396 }
2397 if (!reg_mask && !stack_mask)
2398 /* Found assignment(s) into tracked register in this state.
2399 * Since this state is already marked, just return.
2400 * Nothing to be tracked further in the parent state.
2401 */
2402 return 0;
2403 if (i == first_idx)
2404 break;
2405 i = get_prev_insn_idx(st, i, &history);
2406 if (i >= env->prog->len) {
2407 /* This can happen if backtracking reached insn 0
2408 * and there are still reg_mask or stack_mask
2409 * to backtrack.
2410 * It means the backtracking missed the spot where
2411 * particular register was initialized with a constant.
2412 */
2413 verbose(env, "BUG backtracking idx %d\n", i);
2414 WARN_ONCE(1, "verifier backtracking bug");
2415 return -EFAULT;
2416 }
2417 }
2418 st = st->parent;
2419 if (!st)
2420 break;
2421
2422 new_marks = false;
2423 func = st->frame[st->curframe];
2424 bitmap_from_u64(mask, reg_mask);
2425 for_each_set_bit(i, mask, 32) {
2426 reg = &func->regs[i];
2427 if (reg->type != SCALAR_VALUE) {
2428 reg_mask &= ~(1u << i);
2429 continue;
2430 }
2431 if (!reg->precise)
2432 new_marks = true;
2433 reg->precise = true;
2434 }
2435
2436 bitmap_from_u64(mask, stack_mask);
2437 for_each_set_bit(i, mask, 64) {
2438 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2439 /* the sequence of instructions:
2440 * 2: (bf) r3 = r10
2441 * 3: (7b) *(u64 *)(r3 -8) = r0
2442 * 4: (79) r4 = *(u64 *)(r10 -8)
2443 * doesn't contain jmps. It's backtracked
2444 * as a single block.
2445 * During backtracking insn 3 is not recognized as
2446 * stack access, so at the end of backtracking
2447 * stack slot fp-8 is still marked in stack_mask.
2448 * However the parent state may not have accessed
2449 * fp-8 and it's "unallocated" stack space.
2450 * In such case fallback to conservative.
2451 */
2452 mark_all_scalars_precise(env, st);
2453 return 0;
2454 }
2455
2456 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2457 stack_mask &= ~(1ull << i);
2458 continue;
2459 }
2460 reg = &func->stack[i].spilled_ptr;
2461 if (reg->type != SCALAR_VALUE) {
2462 stack_mask &= ~(1ull << i);
2463 continue;
2464 }
2465 if (!reg->precise)
2466 new_marks = true;
2467 reg->precise = true;
2468 }
2469 if (env->log.level & BPF_LOG_LEVEL) {
2470 print_verifier_state(env, func);
2471 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2472 new_marks ? "didn't have" : "already had",
2473 reg_mask, stack_mask);
2474 }
2475
2476 if (!reg_mask && !stack_mask)
2477 break;
2478 if (!new_marks)
2479 break;
2480
2481 last_idx = st->last_insn_idx;
2482 first_idx = st->first_insn_idx;
2483 }
2484 return 0;
2485 }
2486
mark_chain_precision(struct bpf_verifier_env * env,int regno)2487 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2488 {
2489 return __mark_chain_precision(env, regno, -1);
2490 }
2491
mark_chain_precision_stack(struct bpf_verifier_env * env,int spi)2492 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2493 {
2494 return __mark_chain_precision(env, -1, spi);
2495 }
2496
is_spillable_regtype(enum bpf_reg_type type)2497 static bool is_spillable_regtype(enum bpf_reg_type type)
2498 {
2499 switch (type) {
2500 case PTR_TO_MAP_VALUE:
2501 case PTR_TO_MAP_VALUE_OR_NULL:
2502 case PTR_TO_STACK:
2503 case PTR_TO_CTX:
2504 case PTR_TO_PACKET:
2505 case PTR_TO_PACKET_META:
2506 case PTR_TO_PACKET_END:
2507 case PTR_TO_FLOW_KEYS:
2508 case CONST_PTR_TO_MAP:
2509 case PTR_TO_SOCKET:
2510 case PTR_TO_SOCKET_OR_NULL:
2511 case PTR_TO_SOCK_COMMON:
2512 case PTR_TO_SOCK_COMMON_OR_NULL:
2513 case PTR_TO_TCP_SOCK:
2514 case PTR_TO_TCP_SOCK_OR_NULL:
2515 case PTR_TO_XDP_SOCK:
2516 case PTR_TO_BTF_ID:
2517 case PTR_TO_BTF_ID_OR_NULL:
2518 case PTR_TO_RDONLY_BUF:
2519 case PTR_TO_RDONLY_BUF_OR_NULL:
2520 case PTR_TO_RDWR_BUF:
2521 case PTR_TO_RDWR_BUF_OR_NULL:
2522 case PTR_TO_PERCPU_BTF_ID:
2523 case PTR_TO_MEM:
2524 case PTR_TO_MEM_OR_NULL:
2525 case PTR_TO_FUNC:
2526 case PTR_TO_MAP_KEY:
2527 return true;
2528 default:
2529 return false;
2530 }
2531 }
2532
2533 /* Does this register contain a constant zero? */
register_is_null(struct bpf_reg_state * reg)2534 static bool register_is_null(struct bpf_reg_state *reg)
2535 {
2536 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2537 }
2538
register_is_const(struct bpf_reg_state * reg)2539 static bool register_is_const(struct bpf_reg_state *reg)
2540 {
2541 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2542 }
2543
__is_scalar_unbounded(struct bpf_reg_state * reg)2544 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2545 {
2546 return tnum_is_unknown(reg->var_off) &&
2547 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2548 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2549 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2550 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2551 }
2552
register_is_bounded(struct bpf_reg_state * reg)2553 static bool register_is_bounded(struct bpf_reg_state *reg)
2554 {
2555 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2556 }
2557
__is_pointer_value(bool allow_ptr_leaks,const struct bpf_reg_state * reg)2558 static bool __is_pointer_value(bool allow_ptr_leaks,
2559 const struct bpf_reg_state *reg)
2560 {
2561 if (allow_ptr_leaks)
2562 return false;
2563
2564 return reg->type != SCALAR_VALUE;
2565 }
2566
save_register_state(struct bpf_func_state * state,int spi,struct bpf_reg_state * reg)2567 static void save_register_state(struct bpf_func_state *state,
2568 int spi, struct bpf_reg_state *reg)
2569 {
2570 int i;
2571
2572 state->stack[spi].spilled_ptr = *reg;
2573 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2574
2575 for (i = 0; i < BPF_REG_SIZE; i++)
2576 state->stack[spi].slot_type[i] = STACK_SPILL;
2577 }
2578
2579 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2580 * stack boundary and alignment are checked in check_mem_access()
2581 */
check_stack_write_fixed_off(struct bpf_verifier_env * env,struct bpf_func_state * state,int off,int size,int value_regno,int insn_idx)2582 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2583 /* stack frame we're writing to */
2584 struct bpf_func_state *state,
2585 int off, int size, int value_regno,
2586 int insn_idx)
2587 {
2588 struct bpf_func_state *cur; /* state of the current function */
2589 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2590 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2591 struct bpf_reg_state *reg = NULL;
2592
2593 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2594 state->acquired_refs, true);
2595 if (err)
2596 return err;
2597 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2598 * so it's aligned access and [off, off + size) are within stack limits
2599 */
2600 if (!env->allow_ptr_leaks &&
2601 state->stack[spi].slot_type[0] == STACK_SPILL &&
2602 size != BPF_REG_SIZE) {
2603 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2604 return -EACCES;
2605 }
2606
2607 cur = env->cur_state->frame[env->cur_state->curframe];
2608 if (value_regno >= 0)
2609 reg = &cur->regs[value_regno];
2610
2611 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2612 !register_is_null(reg) && env->bpf_capable) {
2613 if (dst_reg != BPF_REG_FP) {
2614 /* The backtracking logic can only recognize explicit
2615 * stack slot address like [fp - 8]. Other spill of
2616 * scalar via different register has to be conervative.
2617 * Backtrack from here and mark all registers as precise
2618 * that contributed into 'reg' being a constant.
2619 */
2620 err = mark_chain_precision(env, value_regno);
2621 if (err)
2622 return err;
2623 }
2624 save_register_state(state, spi, reg);
2625 } else if (reg && is_spillable_regtype(reg->type)) {
2626 /* register containing pointer is being spilled into stack */
2627 if (size != BPF_REG_SIZE) {
2628 verbose_linfo(env, insn_idx, "; ");
2629 verbose(env, "invalid size of register spill\n");
2630 return -EACCES;
2631 }
2632
2633 if (state != cur && reg->type == PTR_TO_STACK) {
2634 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2635 return -EINVAL;
2636 }
2637
2638 if (!env->bypass_spec_v4) {
2639 bool sanitize = false;
2640
2641 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2642 register_is_const(&state->stack[spi].spilled_ptr))
2643 sanitize = true;
2644 for (i = 0; i < BPF_REG_SIZE; i++)
2645 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2646 sanitize = true;
2647 break;
2648 }
2649 if (sanitize) {
2650 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2651 int soff = (-spi - 1) * BPF_REG_SIZE;
2652
2653 /* detected reuse of integer stack slot with a pointer
2654 * which means either llvm is reusing stack slot or
2655 * an attacker is trying to exploit CVE-2018-3639
2656 * (speculative store bypass)
2657 * Have to sanitize that slot with preemptive
2658 * store of zero.
2659 */
2660 if (*poff && *poff != soff) {
2661 /* disallow programs where single insn stores
2662 * into two different stack slots, since verifier
2663 * cannot sanitize them
2664 */
2665 verbose(env,
2666 "insn %d cannot access two stack slots fp%d and fp%d",
2667 insn_idx, *poff, soff);
2668 return -EINVAL;
2669 }
2670 *poff = soff;
2671 }
2672 }
2673 save_register_state(state, spi, reg);
2674 } else {
2675 u8 type = STACK_MISC;
2676
2677 /* regular write of data into stack destroys any spilled ptr */
2678 state->stack[spi].spilled_ptr.type = NOT_INIT;
2679 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2680 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2681 for (i = 0; i < BPF_REG_SIZE; i++)
2682 state->stack[spi].slot_type[i] = STACK_MISC;
2683
2684 /* only mark the slot as written if all 8 bytes were written
2685 * otherwise read propagation may incorrectly stop too soon
2686 * when stack slots are partially written.
2687 * This heuristic means that read propagation will be
2688 * conservative, since it will add reg_live_read marks
2689 * to stack slots all the way to first state when programs
2690 * writes+reads less than 8 bytes
2691 */
2692 if (size == BPF_REG_SIZE)
2693 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2694
2695 /* when we zero initialize stack slots mark them as such */
2696 if (reg && register_is_null(reg)) {
2697 /* backtracking doesn't work for STACK_ZERO yet. */
2698 err = mark_chain_precision(env, value_regno);
2699 if (err)
2700 return err;
2701 type = STACK_ZERO;
2702 }
2703
2704 /* Mark slots affected by this stack write. */
2705 for (i = 0; i < size; i++)
2706 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2707 type;
2708 }
2709 return 0;
2710 }
2711
2712 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2713 * known to contain a variable offset.
2714 * This function checks whether the write is permitted and conservatively
2715 * tracks the effects of the write, considering that each stack slot in the
2716 * dynamic range is potentially written to.
2717 *
2718 * 'off' includes 'regno->off'.
2719 * 'value_regno' can be -1, meaning that an unknown value is being written to
2720 * the stack.
2721 *
2722 * Spilled pointers in range are not marked as written because we don't know
2723 * what's going to be actually written. This means that read propagation for
2724 * future reads cannot be terminated by this write.
2725 *
2726 * For privileged programs, uninitialized stack slots are considered
2727 * initialized by this write (even though we don't know exactly what offsets
2728 * are going to be written to). The idea is that we don't want the verifier to
2729 * reject future reads that access slots written to through variable offsets.
2730 */
check_stack_write_var_off(struct bpf_verifier_env * env,struct bpf_func_state * state,int ptr_regno,int off,int size,int value_regno,int insn_idx)2731 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2732 /* func where register points to */
2733 struct bpf_func_state *state,
2734 int ptr_regno, int off, int size,
2735 int value_regno, int insn_idx)
2736 {
2737 struct bpf_func_state *cur; /* state of the current function */
2738 int min_off, max_off;
2739 int i, err;
2740 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2741 bool writing_zero = false;
2742 /* set if the fact that we're writing a zero is used to let any
2743 * stack slots remain STACK_ZERO
2744 */
2745 bool zero_used = false;
2746
2747 cur = env->cur_state->frame[env->cur_state->curframe];
2748 ptr_reg = &cur->regs[ptr_regno];
2749 min_off = ptr_reg->smin_value + off;
2750 max_off = ptr_reg->smax_value + off + size;
2751 if (value_regno >= 0)
2752 value_reg = &cur->regs[value_regno];
2753 if (value_reg && register_is_null(value_reg))
2754 writing_zero = true;
2755
2756 err = realloc_func_state(state, round_up(-min_off, BPF_REG_SIZE),
2757 state->acquired_refs, true);
2758 if (err)
2759 return err;
2760
2761
2762 /* Variable offset writes destroy any spilled pointers in range. */
2763 for (i = min_off; i < max_off; i++) {
2764 u8 new_type, *stype;
2765 int slot, spi;
2766
2767 slot = -i - 1;
2768 spi = slot / BPF_REG_SIZE;
2769 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2770
2771 if (!env->allow_ptr_leaks
2772 && *stype != NOT_INIT
2773 && *stype != SCALAR_VALUE) {
2774 /* Reject the write if there's are spilled pointers in
2775 * range. If we didn't reject here, the ptr status
2776 * would be erased below (even though not all slots are
2777 * actually overwritten), possibly opening the door to
2778 * leaks.
2779 */
2780 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2781 insn_idx, i);
2782 return -EINVAL;
2783 }
2784
2785 /* Erase all spilled pointers. */
2786 state->stack[spi].spilled_ptr.type = NOT_INIT;
2787
2788 /* Update the slot type. */
2789 new_type = STACK_MISC;
2790 if (writing_zero && *stype == STACK_ZERO) {
2791 new_type = STACK_ZERO;
2792 zero_used = true;
2793 }
2794 /* If the slot is STACK_INVALID, we check whether it's OK to
2795 * pretend that it will be initialized by this write. The slot
2796 * might not actually be written to, and so if we mark it as
2797 * initialized future reads might leak uninitialized memory.
2798 * For privileged programs, we will accept such reads to slots
2799 * that may or may not be written because, if we're reject
2800 * them, the error would be too confusing.
2801 */
2802 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2803 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2804 insn_idx, i);
2805 return -EINVAL;
2806 }
2807 *stype = new_type;
2808 }
2809 if (zero_used) {
2810 /* backtracking doesn't work for STACK_ZERO yet. */
2811 err = mark_chain_precision(env, value_regno);
2812 if (err)
2813 return err;
2814 }
2815 return 0;
2816 }
2817
2818 /* When register 'dst_regno' is assigned some values from stack[min_off,
2819 * max_off), we set the register's type according to the types of the
2820 * respective stack slots. If all the stack values are known to be zeros, then
2821 * so is the destination reg. Otherwise, the register is considered to be
2822 * SCALAR. This function does not deal with register filling; the caller must
2823 * ensure that all spilled registers in the stack range have been marked as
2824 * read.
2825 */
mark_reg_stack_read(struct bpf_verifier_env * env,struct bpf_func_state * ptr_state,int min_off,int max_off,int dst_regno)2826 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2827 /* func where src register points to */
2828 struct bpf_func_state *ptr_state,
2829 int min_off, int max_off, int dst_regno)
2830 {
2831 struct bpf_verifier_state *vstate = env->cur_state;
2832 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2833 int i, slot, spi;
2834 u8 *stype;
2835 int zeros = 0;
2836
2837 for (i = min_off; i < max_off; i++) {
2838 slot = -i - 1;
2839 spi = slot / BPF_REG_SIZE;
2840 stype = ptr_state->stack[spi].slot_type;
2841 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2842 break;
2843 zeros++;
2844 }
2845 if (zeros == max_off - min_off) {
2846 /* any access_size read into register is zero extended,
2847 * so the whole register == const_zero
2848 */
2849 __mark_reg_const_zero(&state->regs[dst_regno]);
2850 /* backtracking doesn't support STACK_ZERO yet,
2851 * so mark it precise here, so that later
2852 * backtracking can stop here.
2853 * Backtracking may not need this if this register
2854 * doesn't participate in pointer adjustment.
2855 * Forward propagation of precise flag is not
2856 * necessary either. This mark is only to stop
2857 * backtracking. Any register that contributed
2858 * to const 0 was marked precise before spill.
2859 */
2860 state->regs[dst_regno].precise = true;
2861 } else {
2862 /* have read misc data from the stack */
2863 mark_reg_unknown(env, state->regs, dst_regno);
2864 }
2865 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2866 }
2867
2868 /* Read the stack at 'off' and put the results into the register indicated by
2869 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2870 * spilled reg.
2871 *
2872 * 'dst_regno' can be -1, meaning that the read value is not going to a
2873 * register.
2874 *
2875 * The access is assumed to be within the current stack bounds.
2876 */
check_stack_read_fixed_off(struct bpf_verifier_env * env,struct bpf_func_state * reg_state,int off,int size,int dst_regno)2877 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2878 /* func where src register points to */
2879 struct bpf_func_state *reg_state,
2880 int off, int size, int dst_regno)
2881 {
2882 struct bpf_verifier_state *vstate = env->cur_state;
2883 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2884 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2885 struct bpf_reg_state *reg;
2886 u8 *stype;
2887
2888 stype = reg_state->stack[spi].slot_type;
2889 reg = ®_state->stack[spi].spilled_ptr;
2890
2891 if (stype[0] == STACK_SPILL) {
2892 if (size != BPF_REG_SIZE) {
2893 if (reg->type != SCALAR_VALUE) {
2894 verbose_linfo(env, env->insn_idx, "; ");
2895 verbose(env, "invalid size of register fill\n");
2896 return -EACCES;
2897 }
2898 if (dst_regno >= 0) {
2899 mark_reg_unknown(env, state->regs, dst_regno);
2900 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2901 }
2902 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2903 return 0;
2904 }
2905 for (i = 1; i < BPF_REG_SIZE; i++) {
2906 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2907 verbose(env, "corrupted spill memory\n");
2908 return -EACCES;
2909 }
2910 }
2911
2912 if (dst_regno >= 0) {
2913 /* restore register state from stack */
2914 state->regs[dst_regno] = *reg;
2915 /* mark reg as written since spilled pointer state likely
2916 * has its liveness marks cleared by is_state_visited()
2917 * which resets stack/reg liveness for state transitions
2918 */
2919 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2920 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2921 /* If dst_regno==-1, the caller is asking us whether
2922 * it is acceptable to use this value as a SCALAR_VALUE
2923 * (e.g. for XADD).
2924 * We must not allow unprivileged callers to do that
2925 * with spilled pointers.
2926 */
2927 verbose(env, "leaking pointer from stack off %d\n",
2928 off);
2929 return -EACCES;
2930 }
2931 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2932 } else {
2933 u8 type;
2934
2935 for (i = 0; i < size; i++) {
2936 type = stype[(slot - i) % BPF_REG_SIZE];
2937 if (type == STACK_MISC)
2938 continue;
2939 if (type == STACK_ZERO)
2940 continue;
2941 verbose(env, "invalid read from stack off %d+%d size %d\n",
2942 off, i, size);
2943 return -EACCES;
2944 }
2945 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2946 if (dst_regno >= 0)
2947 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2948 }
2949 return 0;
2950 }
2951
2952 enum stack_access_src {
2953 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2954 ACCESS_HELPER = 2, /* the access is performed by a helper */
2955 };
2956
2957 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2958 int regno, int off, int access_size,
2959 bool zero_size_allowed,
2960 enum stack_access_src type,
2961 struct bpf_call_arg_meta *meta);
2962
reg_state(struct bpf_verifier_env * env,int regno)2963 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2964 {
2965 return cur_regs(env) + regno;
2966 }
2967
2968 /* Read the stack at 'ptr_regno + off' and put the result into the register
2969 * 'dst_regno'.
2970 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2971 * but not its variable offset.
2972 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2973 *
2974 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2975 * filling registers (i.e. reads of spilled register cannot be detected when
2976 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2977 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2978 * offset; for a fixed offset check_stack_read_fixed_off should be used
2979 * instead.
2980 */
check_stack_read_var_off(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)2981 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2982 int ptr_regno, int off, int size, int dst_regno)
2983 {
2984 /* The state of the source register. */
2985 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2986 struct bpf_func_state *ptr_state = func(env, reg);
2987 int err;
2988 int min_off, max_off;
2989
2990 /* Note that we pass a NULL meta, so raw access will not be permitted.
2991 */
2992 err = check_stack_range_initialized(env, ptr_regno, off, size,
2993 false, ACCESS_DIRECT, NULL);
2994 if (err)
2995 return err;
2996
2997 min_off = reg->smin_value + off;
2998 max_off = reg->smax_value + off;
2999 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3000 return 0;
3001 }
3002
3003 /* check_stack_read dispatches to check_stack_read_fixed_off or
3004 * check_stack_read_var_off.
3005 *
3006 * The caller must ensure that the offset falls within the allocated stack
3007 * bounds.
3008 *
3009 * 'dst_regno' is a register which will receive the value from the stack. It
3010 * can be -1, meaning that the read value is not going to a register.
3011 */
check_stack_read(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)3012 static int check_stack_read(struct bpf_verifier_env *env,
3013 int ptr_regno, int off, int size,
3014 int dst_regno)
3015 {
3016 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3017 struct bpf_func_state *state = func(env, reg);
3018 int err;
3019 /* Some accesses are only permitted with a static offset. */
3020 bool var_off = !tnum_is_const(reg->var_off);
3021
3022 /* The offset is required to be static when reads don't go to a
3023 * register, in order to not leak pointers (see
3024 * check_stack_read_fixed_off).
3025 */
3026 if (dst_regno < 0 && var_off) {
3027 char tn_buf[48];
3028
3029 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3030 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3031 tn_buf, off, size);
3032 return -EACCES;
3033 }
3034 /* Variable offset is prohibited for unprivileged mode for simplicity
3035 * since it requires corresponding support in Spectre masking for stack
3036 * ALU. See also retrieve_ptr_limit().
3037 */
3038 if (!env->bypass_spec_v1 && var_off) {
3039 char tn_buf[48];
3040
3041 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3042 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3043 ptr_regno, tn_buf);
3044 return -EACCES;
3045 }
3046
3047 if (!var_off) {
3048 off += reg->var_off.value;
3049 err = check_stack_read_fixed_off(env, state, off, size,
3050 dst_regno);
3051 } else {
3052 /* Variable offset stack reads need more conservative handling
3053 * than fixed offset ones. Note that dst_regno >= 0 on this
3054 * branch.
3055 */
3056 err = check_stack_read_var_off(env, ptr_regno, off, size,
3057 dst_regno);
3058 }
3059 return err;
3060 }
3061
3062
3063 /* check_stack_write dispatches to check_stack_write_fixed_off or
3064 * check_stack_write_var_off.
3065 *
3066 * 'ptr_regno' is the register used as a pointer into the stack.
3067 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3068 * 'value_regno' is the register whose value we're writing to the stack. It can
3069 * be -1, meaning that we're not writing from a register.
3070 *
3071 * The caller must ensure that the offset falls within the maximum stack size.
3072 */
check_stack_write(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int value_regno,int insn_idx)3073 static int check_stack_write(struct bpf_verifier_env *env,
3074 int ptr_regno, int off, int size,
3075 int value_regno, int insn_idx)
3076 {
3077 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3078 struct bpf_func_state *state = func(env, reg);
3079 int err;
3080
3081 if (tnum_is_const(reg->var_off)) {
3082 off += reg->var_off.value;
3083 err = check_stack_write_fixed_off(env, state, off, size,
3084 value_regno, insn_idx);
3085 } else {
3086 /* Variable offset stack reads need more conservative handling
3087 * than fixed offset ones.
3088 */
3089 err = check_stack_write_var_off(env, state,
3090 ptr_regno, off, size,
3091 value_regno, insn_idx);
3092 }
3093 return err;
3094 }
3095
check_map_access_type(struct bpf_verifier_env * env,u32 regno,int off,int size,enum bpf_access_type type)3096 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3097 int off, int size, enum bpf_access_type type)
3098 {
3099 struct bpf_reg_state *regs = cur_regs(env);
3100 struct bpf_map *map = regs[regno].map_ptr;
3101 u32 cap = bpf_map_flags_to_cap(map);
3102
3103 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3104 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3105 map->value_size, off, size);
3106 return -EACCES;
3107 }
3108
3109 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3110 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3111 map->value_size, off, size);
3112 return -EACCES;
3113 }
3114
3115 return 0;
3116 }
3117
3118 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
__check_mem_access(struct bpf_verifier_env * env,int regno,int off,int size,u32 mem_size,bool zero_size_allowed)3119 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3120 int off, int size, u32 mem_size,
3121 bool zero_size_allowed)
3122 {
3123 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3124 struct bpf_reg_state *reg;
3125
3126 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3127 return 0;
3128
3129 reg = &cur_regs(env)[regno];
3130 switch (reg->type) {
3131 case PTR_TO_MAP_KEY:
3132 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3133 mem_size, off, size);
3134 break;
3135 case PTR_TO_MAP_VALUE:
3136 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3137 mem_size, off, size);
3138 break;
3139 case PTR_TO_PACKET:
3140 case PTR_TO_PACKET_META:
3141 case PTR_TO_PACKET_END:
3142 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3143 off, size, regno, reg->id, off, mem_size);
3144 break;
3145 case PTR_TO_MEM:
3146 default:
3147 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3148 mem_size, off, size);
3149 }
3150
3151 return -EACCES;
3152 }
3153
3154 /* check read/write into a memory region with possible variable offset */
check_mem_region_access(struct bpf_verifier_env * env,u32 regno,int off,int size,u32 mem_size,bool zero_size_allowed)3155 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3156 int off, int size, u32 mem_size,
3157 bool zero_size_allowed)
3158 {
3159 struct bpf_verifier_state *vstate = env->cur_state;
3160 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3161 struct bpf_reg_state *reg = &state->regs[regno];
3162 int err;
3163
3164 /* We may have adjusted the register pointing to memory region, so we
3165 * need to try adding each of min_value and max_value to off
3166 * to make sure our theoretical access will be safe.
3167 */
3168 if (env->log.level & BPF_LOG_LEVEL)
3169 print_verifier_state(env, state);
3170
3171 /* The minimum value is only important with signed
3172 * comparisons where we can't assume the floor of a
3173 * value is 0. If we are using signed variables for our
3174 * index'es we need to make sure that whatever we use
3175 * will have a set floor within our range.
3176 */
3177 if (reg->smin_value < 0 &&
3178 (reg->smin_value == S64_MIN ||
3179 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3180 reg->smin_value + off < 0)) {
3181 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3182 regno);
3183 return -EACCES;
3184 }
3185 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3186 mem_size, zero_size_allowed);
3187 if (err) {
3188 verbose(env, "R%d min value is outside of the allowed memory range\n",
3189 regno);
3190 return err;
3191 }
3192
3193 /* If we haven't set a max value then we need to bail since we can't be
3194 * sure we won't do bad things.
3195 * If reg->umax_value + off could overflow, treat that as unbounded too.
3196 */
3197 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3198 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3199 regno);
3200 return -EACCES;
3201 }
3202 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3203 mem_size, zero_size_allowed);
3204 if (err) {
3205 verbose(env, "R%d max value is outside of the allowed memory range\n",
3206 regno);
3207 return err;
3208 }
3209
3210 return 0;
3211 }
3212
3213 /* check read/write into a map element with possible variable offset */
check_map_access(struct bpf_verifier_env * env,u32 regno,int off,int size,bool zero_size_allowed)3214 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3215 int off, int size, bool zero_size_allowed)
3216 {
3217 struct bpf_verifier_state *vstate = env->cur_state;
3218 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3219 struct bpf_reg_state *reg = &state->regs[regno];
3220 struct bpf_map *map = reg->map_ptr;
3221 int err;
3222
3223 err = check_mem_region_access(env, regno, off, size, map->value_size,
3224 zero_size_allowed);
3225 if (err)
3226 return err;
3227
3228 if (map_value_has_spin_lock(map)) {
3229 u32 lock = map->spin_lock_off;
3230
3231 /* if any part of struct bpf_spin_lock can be touched by
3232 * load/store reject this program.
3233 * To check that [x1, x2) overlaps with [y1, y2)
3234 * it is sufficient to check x1 < y2 && y1 < x2.
3235 */
3236 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3237 lock < reg->umax_value + off + size) {
3238 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3239 return -EACCES;
3240 }
3241 }
3242 return err;
3243 }
3244
3245 #define MAX_PACKET_OFF 0xffff
3246
resolve_prog_type(struct bpf_prog * prog)3247 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3248 {
3249 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3250 }
3251
may_access_direct_pkt_data(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_access_type t)3252 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3253 const struct bpf_call_arg_meta *meta,
3254 enum bpf_access_type t)
3255 {
3256 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3257
3258 switch (prog_type) {
3259 /* Program types only with direct read access go here! */
3260 case BPF_PROG_TYPE_LWT_IN:
3261 case BPF_PROG_TYPE_LWT_OUT:
3262 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3263 case BPF_PROG_TYPE_SK_REUSEPORT:
3264 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3265 case BPF_PROG_TYPE_CGROUP_SKB:
3266 if (t == BPF_WRITE)
3267 return false;
3268 fallthrough;
3269
3270 /* Program types with direct read + write access go here! */
3271 case BPF_PROG_TYPE_SCHED_CLS:
3272 case BPF_PROG_TYPE_SCHED_ACT:
3273 case BPF_PROG_TYPE_XDP:
3274 case BPF_PROG_TYPE_LWT_XMIT:
3275 case BPF_PROG_TYPE_SK_SKB:
3276 case BPF_PROG_TYPE_SK_MSG:
3277 if (meta)
3278 return meta->pkt_access;
3279
3280 env->seen_direct_write = true;
3281 return true;
3282
3283 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3284 if (t == BPF_WRITE)
3285 env->seen_direct_write = true;
3286
3287 return true;
3288
3289 default:
3290 return false;
3291 }
3292 }
3293
check_packet_access(struct bpf_verifier_env * env,u32 regno,int off,int size,bool zero_size_allowed)3294 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3295 int size, bool zero_size_allowed)
3296 {
3297 struct bpf_reg_state *regs = cur_regs(env);
3298 struct bpf_reg_state *reg = ®s[regno];
3299 int err;
3300
3301 /* We may have added a variable offset to the packet pointer; but any
3302 * reg->range we have comes after that. We are only checking the fixed
3303 * offset.
3304 */
3305
3306 /* We don't allow negative numbers, because we aren't tracking enough
3307 * detail to prove they're safe.
3308 */
3309 if (reg->smin_value < 0) {
3310 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3311 regno);
3312 return -EACCES;
3313 }
3314
3315 err = reg->range < 0 ? -EINVAL :
3316 __check_mem_access(env, regno, off, size, reg->range,
3317 zero_size_allowed);
3318 if (err) {
3319 verbose(env, "R%d offset is outside of the packet\n", regno);
3320 return err;
3321 }
3322
3323 /* __check_mem_access has made sure "off + size - 1" is within u16.
3324 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3325 * otherwise find_good_pkt_pointers would have refused to set range info
3326 * that __check_mem_access would have rejected this pkt access.
3327 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3328 */
3329 env->prog->aux->max_pkt_offset =
3330 max_t(u32, env->prog->aux->max_pkt_offset,
3331 off + reg->umax_value + size - 1);
3332
3333 return err;
3334 }
3335
3336 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
check_ctx_access(struct bpf_verifier_env * env,int insn_idx,int off,int size,enum bpf_access_type t,enum bpf_reg_type * reg_type,struct btf ** btf,u32 * btf_id)3337 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3338 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3339 struct btf **btf, u32 *btf_id)
3340 {
3341 struct bpf_insn_access_aux info = {
3342 .reg_type = *reg_type,
3343 .log = &env->log,
3344 };
3345
3346 if (env->ops->is_valid_access &&
3347 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3348 /* A non zero info.ctx_field_size indicates that this field is a
3349 * candidate for later verifier transformation to load the whole
3350 * field and then apply a mask when accessed with a narrower
3351 * access than actual ctx access size. A zero info.ctx_field_size
3352 * will only allow for whole field access and rejects any other
3353 * type of narrower access.
3354 */
3355 *reg_type = info.reg_type;
3356
3357 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3358 *btf = info.btf;
3359 *btf_id = info.btf_id;
3360 } else {
3361 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3362 }
3363 /* remember the offset of last byte accessed in ctx */
3364 if (env->prog->aux->max_ctx_offset < off + size)
3365 env->prog->aux->max_ctx_offset = off + size;
3366 return 0;
3367 }
3368
3369 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3370 return -EACCES;
3371 }
3372
check_flow_keys_access(struct bpf_verifier_env * env,int off,int size)3373 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3374 int size)
3375 {
3376 if (size < 0 || off < 0 ||
3377 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3378 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3379 off, size);
3380 return -EACCES;
3381 }
3382 return 0;
3383 }
3384
check_sock_access(struct bpf_verifier_env * env,int insn_idx,u32 regno,int off,int size,enum bpf_access_type t)3385 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3386 u32 regno, int off, int size,
3387 enum bpf_access_type t)
3388 {
3389 struct bpf_reg_state *regs = cur_regs(env);
3390 struct bpf_reg_state *reg = ®s[regno];
3391 struct bpf_insn_access_aux info = {};
3392 bool valid;
3393
3394 if (reg->smin_value < 0) {
3395 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3396 regno);
3397 return -EACCES;
3398 }
3399
3400 switch (reg->type) {
3401 case PTR_TO_SOCK_COMMON:
3402 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3403 break;
3404 case PTR_TO_SOCKET:
3405 valid = bpf_sock_is_valid_access(off, size, t, &info);
3406 break;
3407 case PTR_TO_TCP_SOCK:
3408 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3409 break;
3410 case PTR_TO_XDP_SOCK:
3411 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3412 break;
3413 default:
3414 valid = false;
3415 }
3416
3417
3418 if (valid) {
3419 env->insn_aux_data[insn_idx].ctx_field_size =
3420 info.ctx_field_size;
3421 return 0;
3422 }
3423
3424 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3425 regno, reg_type_str[reg->type], off, size);
3426
3427 return -EACCES;
3428 }
3429
is_pointer_value(struct bpf_verifier_env * env,int regno)3430 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3431 {
3432 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3433 }
3434
is_ctx_reg(struct bpf_verifier_env * env,int regno)3435 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3436 {
3437 const struct bpf_reg_state *reg = reg_state(env, regno);
3438
3439 return reg->type == PTR_TO_CTX;
3440 }
3441
is_sk_reg(struct bpf_verifier_env * env,int regno)3442 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3443 {
3444 const struct bpf_reg_state *reg = reg_state(env, regno);
3445
3446 return type_is_sk_pointer(reg->type);
3447 }
3448
is_pkt_reg(struct bpf_verifier_env * env,int regno)3449 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3450 {
3451 const struct bpf_reg_state *reg = reg_state(env, regno);
3452
3453 return type_is_pkt_pointer(reg->type);
3454 }
3455
is_flow_key_reg(struct bpf_verifier_env * env,int regno)3456 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3457 {
3458 const struct bpf_reg_state *reg = reg_state(env, regno);
3459
3460 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3461 return reg->type == PTR_TO_FLOW_KEYS;
3462 }
3463
check_pkt_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict)3464 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3465 const struct bpf_reg_state *reg,
3466 int off, int size, bool strict)
3467 {
3468 struct tnum reg_off;
3469 int ip_align;
3470
3471 /* Byte size accesses are always allowed. */
3472 if (!strict || size == 1)
3473 return 0;
3474
3475 /* For platforms that do not have a Kconfig enabling
3476 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3477 * NET_IP_ALIGN is universally set to '2'. And on platforms
3478 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3479 * to this code only in strict mode where we want to emulate
3480 * the NET_IP_ALIGN==2 checking. Therefore use an
3481 * unconditional IP align value of '2'.
3482 */
3483 ip_align = 2;
3484
3485 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3486 if (!tnum_is_aligned(reg_off, size)) {
3487 char tn_buf[48];
3488
3489 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3490 verbose(env,
3491 "misaligned packet access off %d+%s+%d+%d size %d\n",
3492 ip_align, tn_buf, reg->off, off, size);
3493 return -EACCES;
3494 }
3495
3496 return 0;
3497 }
3498
check_generic_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,const char * pointer_desc,int off,int size,bool strict)3499 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3500 const struct bpf_reg_state *reg,
3501 const char *pointer_desc,
3502 int off, int size, bool strict)
3503 {
3504 struct tnum reg_off;
3505
3506 /* Byte size accesses are always allowed. */
3507 if (!strict || size == 1)
3508 return 0;
3509
3510 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3511 if (!tnum_is_aligned(reg_off, size)) {
3512 char tn_buf[48];
3513
3514 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3515 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3516 pointer_desc, tn_buf, reg->off, off, size);
3517 return -EACCES;
3518 }
3519
3520 return 0;
3521 }
3522
check_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict_alignment_once)3523 static int check_ptr_alignment(struct bpf_verifier_env *env,
3524 const struct bpf_reg_state *reg, int off,
3525 int size, bool strict_alignment_once)
3526 {
3527 bool strict = env->strict_alignment || strict_alignment_once;
3528 const char *pointer_desc = "";
3529
3530 switch (reg->type) {
3531 case PTR_TO_PACKET:
3532 case PTR_TO_PACKET_META:
3533 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3534 * right in front, treat it the very same way.
3535 */
3536 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3537 case PTR_TO_FLOW_KEYS:
3538 pointer_desc = "flow keys ";
3539 break;
3540 case PTR_TO_MAP_KEY:
3541 pointer_desc = "key ";
3542 break;
3543 case PTR_TO_MAP_VALUE:
3544 pointer_desc = "value ";
3545 break;
3546 case PTR_TO_CTX:
3547 pointer_desc = "context ";
3548 break;
3549 case PTR_TO_STACK:
3550 pointer_desc = "stack ";
3551 /* The stack spill tracking logic in check_stack_write_fixed_off()
3552 * and check_stack_read_fixed_off() relies on stack accesses being
3553 * aligned.
3554 */
3555 strict = true;
3556 break;
3557 case PTR_TO_SOCKET:
3558 pointer_desc = "sock ";
3559 break;
3560 case PTR_TO_SOCK_COMMON:
3561 pointer_desc = "sock_common ";
3562 break;
3563 case PTR_TO_TCP_SOCK:
3564 pointer_desc = "tcp_sock ";
3565 break;
3566 case PTR_TO_XDP_SOCK:
3567 pointer_desc = "xdp_sock ";
3568 break;
3569 default:
3570 break;
3571 }
3572 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3573 strict);
3574 }
3575
update_stack_depth(struct bpf_verifier_env * env,const struct bpf_func_state * func,int off)3576 static int update_stack_depth(struct bpf_verifier_env *env,
3577 const struct bpf_func_state *func,
3578 int off)
3579 {
3580 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3581
3582 if (stack >= -off)
3583 return 0;
3584
3585 /* update known max for given subprogram */
3586 env->subprog_info[func->subprogno].stack_depth = -off;
3587 return 0;
3588 }
3589
3590 /* starting from main bpf function walk all instructions of the function
3591 * and recursively walk all callees that given function can call.
3592 * Ignore jump and exit insns.
3593 * Since recursion is prevented by check_cfg() this algorithm
3594 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3595 */
check_max_stack_depth(struct bpf_verifier_env * env)3596 static int check_max_stack_depth(struct bpf_verifier_env *env)
3597 {
3598 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3599 struct bpf_subprog_info *subprog = env->subprog_info;
3600 struct bpf_insn *insn = env->prog->insnsi;
3601 bool tail_call_reachable = false;
3602 int ret_insn[MAX_CALL_FRAMES];
3603 int ret_prog[MAX_CALL_FRAMES];
3604 int j;
3605
3606 process_func:
3607 /* protect against potential stack overflow that might happen when
3608 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3609 * depth for such case down to 256 so that the worst case scenario
3610 * would result in 8k stack size (32 which is tailcall limit * 256 =
3611 * 8k).
3612 *
3613 * To get the idea what might happen, see an example:
3614 * func1 -> sub rsp, 128
3615 * subfunc1 -> sub rsp, 256
3616 * tailcall1 -> add rsp, 256
3617 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3618 * subfunc2 -> sub rsp, 64
3619 * subfunc22 -> sub rsp, 128
3620 * tailcall2 -> add rsp, 128
3621 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3622 *
3623 * tailcall will unwind the current stack frame but it will not get rid
3624 * of caller's stack as shown on the example above.
3625 */
3626 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3627 verbose(env,
3628 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3629 depth);
3630 return -EACCES;
3631 }
3632 /* round up to 32-bytes, since this is granularity
3633 * of interpreter stack size
3634 */
3635 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3636 if (depth > MAX_BPF_STACK) {
3637 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3638 frame + 1, depth);
3639 return -EACCES;
3640 }
3641 continue_func:
3642 subprog_end = subprog[idx + 1].start;
3643 for (; i < subprog_end; i++) {
3644 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3645 continue;
3646 /* remember insn and function to return to */
3647 ret_insn[frame] = i + 1;
3648 ret_prog[frame] = idx;
3649
3650 /* find the callee */
3651 i = i + insn[i].imm + 1;
3652 idx = find_subprog(env, i);
3653 if (idx < 0) {
3654 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3655 i);
3656 return -EFAULT;
3657 }
3658
3659 if (subprog[idx].has_tail_call)
3660 tail_call_reachable = true;
3661
3662 frame++;
3663 if (frame >= MAX_CALL_FRAMES) {
3664 verbose(env, "the call stack of %d frames is too deep !\n",
3665 frame);
3666 return -E2BIG;
3667 }
3668 goto process_func;
3669 }
3670 /* if tail call got detected across bpf2bpf calls then mark each of the
3671 * currently present subprog frames as tail call reachable subprogs;
3672 * this info will be utilized by JIT so that we will be preserving the
3673 * tail call counter throughout bpf2bpf calls combined with tailcalls
3674 */
3675 if (tail_call_reachable)
3676 for (j = 0; j < frame; j++)
3677 subprog[ret_prog[j]].tail_call_reachable = true;
3678
3679 /* end of for() loop means the last insn of the 'subprog'
3680 * was reached. Doesn't matter whether it was JA or EXIT
3681 */
3682 if (frame == 0)
3683 return 0;
3684 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3685 frame--;
3686 i = ret_insn[frame];
3687 idx = ret_prog[frame];
3688 goto continue_func;
3689 }
3690
3691 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
get_callee_stack_depth(struct bpf_verifier_env * env,const struct bpf_insn * insn,int idx)3692 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3693 const struct bpf_insn *insn, int idx)
3694 {
3695 int start = idx + insn->imm + 1, subprog;
3696
3697 subprog = find_subprog(env, start);
3698 if (subprog < 0) {
3699 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3700 start);
3701 return -EFAULT;
3702 }
3703 return env->subprog_info[subprog].stack_depth;
3704 }
3705 #endif
3706
check_ctx_reg(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno)3707 int check_ctx_reg(struct bpf_verifier_env *env,
3708 const struct bpf_reg_state *reg, int regno)
3709 {
3710 /* Access to ctx or passing it to a helper is only allowed in
3711 * its original, unmodified form.
3712 */
3713
3714 if (reg->off) {
3715 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3716 regno, reg->off);
3717 return -EACCES;
3718 }
3719
3720 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3721 char tn_buf[48];
3722
3723 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3724 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3725 return -EACCES;
3726 }
3727
3728 return 0;
3729 }
3730
__check_buffer_access(struct bpf_verifier_env * env,const char * buf_info,const struct bpf_reg_state * reg,int regno,int off,int size)3731 static int __check_buffer_access(struct bpf_verifier_env *env,
3732 const char *buf_info,
3733 const struct bpf_reg_state *reg,
3734 int regno, int off, int size)
3735 {
3736 if (off < 0) {
3737 verbose(env,
3738 "R%d invalid %s buffer access: off=%d, size=%d\n",
3739 regno, buf_info, off, size);
3740 return -EACCES;
3741 }
3742 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3743 char tn_buf[48];
3744
3745 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3746 verbose(env,
3747 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3748 regno, off, tn_buf);
3749 return -EACCES;
3750 }
3751
3752 return 0;
3753 }
3754
check_tp_buffer_access(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,int off,int size)3755 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3756 const struct bpf_reg_state *reg,
3757 int regno, int off, int size)
3758 {
3759 int err;
3760
3761 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3762 if (err)
3763 return err;
3764
3765 if (off + size > env->prog->aux->max_tp_access)
3766 env->prog->aux->max_tp_access = off + size;
3767
3768 return 0;
3769 }
3770
check_buffer_access(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,int off,int size,bool zero_size_allowed,const char * buf_info,u32 * max_access)3771 static int check_buffer_access(struct bpf_verifier_env *env,
3772 const struct bpf_reg_state *reg,
3773 int regno, int off, int size,
3774 bool zero_size_allowed,
3775 const char *buf_info,
3776 u32 *max_access)
3777 {
3778 int err;
3779
3780 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3781 if (err)
3782 return err;
3783
3784 if (off + size > *max_access)
3785 *max_access = off + size;
3786
3787 return 0;
3788 }
3789
3790 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
zext_32_to_64(struct bpf_reg_state * reg)3791 static void zext_32_to_64(struct bpf_reg_state *reg)
3792 {
3793 reg->var_off = tnum_subreg(reg->var_off);
3794 __reg_assign_32_into_64(reg);
3795 }
3796
3797 /* truncate register to smaller size (in bytes)
3798 * must be called with size < BPF_REG_SIZE
3799 */
coerce_reg_to_size(struct bpf_reg_state * reg,int size)3800 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3801 {
3802 u64 mask;
3803
3804 /* clear high bits in bit representation */
3805 reg->var_off = tnum_cast(reg->var_off, size);
3806
3807 /* fix arithmetic bounds */
3808 mask = ((u64)1 << (size * 8)) - 1;
3809 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3810 reg->umin_value &= mask;
3811 reg->umax_value &= mask;
3812 } else {
3813 reg->umin_value = 0;
3814 reg->umax_value = mask;
3815 }
3816 reg->smin_value = reg->umin_value;
3817 reg->smax_value = reg->umax_value;
3818
3819 /* If size is smaller than 32bit register the 32bit register
3820 * values are also truncated so we push 64-bit bounds into
3821 * 32-bit bounds. Above were truncated < 32-bits already.
3822 */
3823 if (size >= 4)
3824 return;
3825 __reg_combine_64_into_32(reg);
3826 }
3827
bpf_map_is_rdonly(const struct bpf_map * map)3828 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3829 {
3830 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3831 }
3832
bpf_map_direct_read(struct bpf_map * map,int off,int size,u64 * val)3833 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3834 {
3835 void *ptr;
3836 u64 addr;
3837 int err;
3838
3839 err = map->ops->map_direct_value_addr(map, &addr, off);
3840 if (err)
3841 return err;
3842 ptr = (void *)(long)addr + off;
3843
3844 switch (size) {
3845 case sizeof(u8):
3846 *val = (u64)*(u8 *)ptr;
3847 break;
3848 case sizeof(u16):
3849 *val = (u64)*(u16 *)ptr;
3850 break;
3851 case sizeof(u32):
3852 *val = (u64)*(u32 *)ptr;
3853 break;
3854 case sizeof(u64):
3855 *val = *(u64 *)ptr;
3856 break;
3857 default:
3858 return -EINVAL;
3859 }
3860 return 0;
3861 }
3862
check_ptr_to_btf_access(struct bpf_verifier_env * env,struct bpf_reg_state * regs,int regno,int off,int size,enum bpf_access_type atype,int value_regno)3863 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3864 struct bpf_reg_state *regs,
3865 int regno, int off, int size,
3866 enum bpf_access_type atype,
3867 int value_regno)
3868 {
3869 struct bpf_reg_state *reg = regs + regno;
3870 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3871 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3872 u32 btf_id;
3873 int ret;
3874
3875 if (off < 0) {
3876 verbose(env,
3877 "R%d is ptr_%s invalid negative access: off=%d\n",
3878 regno, tname, off);
3879 return -EACCES;
3880 }
3881 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3882 char tn_buf[48];
3883
3884 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3885 verbose(env,
3886 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3887 regno, tname, off, tn_buf);
3888 return -EACCES;
3889 }
3890
3891 if (env->ops->btf_struct_access) {
3892 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3893 off, size, atype, &btf_id);
3894 } else {
3895 if (atype != BPF_READ) {
3896 verbose(env, "only read is supported\n");
3897 return -EACCES;
3898 }
3899
3900 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3901 atype, &btf_id);
3902 }
3903
3904 if (ret < 0)
3905 return ret;
3906
3907 if (atype == BPF_READ && value_regno >= 0)
3908 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3909
3910 return 0;
3911 }
3912
check_ptr_to_map_access(struct bpf_verifier_env * env,struct bpf_reg_state * regs,int regno,int off,int size,enum bpf_access_type atype,int value_regno)3913 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3914 struct bpf_reg_state *regs,
3915 int regno, int off, int size,
3916 enum bpf_access_type atype,
3917 int value_regno)
3918 {
3919 struct bpf_reg_state *reg = regs + regno;
3920 struct bpf_map *map = reg->map_ptr;
3921 const struct btf_type *t;
3922 const char *tname;
3923 u32 btf_id;
3924 int ret;
3925
3926 if (!btf_vmlinux) {
3927 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3928 return -ENOTSUPP;
3929 }
3930
3931 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3932 verbose(env, "map_ptr access not supported for map type %d\n",
3933 map->map_type);
3934 return -ENOTSUPP;
3935 }
3936
3937 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3938 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3939
3940 if (!env->allow_ptr_to_map_access) {
3941 verbose(env,
3942 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3943 tname);
3944 return -EPERM;
3945 }
3946
3947 if (off < 0) {
3948 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3949 regno, tname, off);
3950 return -EACCES;
3951 }
3952
3953 if (atype != BPF_READ) {
3954 verbose(env, "only read from %s is supported\n", tname);
3955 return -EACCES;
3956 }
3957
3958 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
3959 if (ret < 0)
3960 return ret;
3961
3962 if (value_regno >= 0)
3963 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
3964
3965 return 0;
3966 }
3967
3968 /* Check that the stack access at the given offset is within bounds. The
3969 * maximum valid offset is -1.
3970 *
3971 * The minimum valid offset is -MAX_BPF_STACK for writes, and
3972 * -state->allocated_stack for reads.
3973 */
check_stack_slot_within_bounds(int off,struct bpf_func_state * state,enum bpf_access_type t)3974 static int check_stack_slot_within_bounds(int off,
3975 struct bpf_func_state *state,
3976 enum bpf_access_type t)
3977 {
3978 int min_valid_off;
3979
3980 if (t == BPF_WRITE)
3981 min_valid_off = -MAX_BPF_STACK;
3982 else
3983 min_valid_off = -state->allocated_stack;
3984
3985 if (off < min_valid_off || off > -1)
3986 return -EACCES;
3987 return 0;
3988 }
3989
3990 /* Check that the stack access at 'regno + off' falls within the maximum stack
3991 * bounds.
3992 *
3993 * 'off' includes `regno->offset`, but not its dynamic part (if any).
3994 */
check_stack_access_within_bounds(struct bpf_verifier_env * env,int regno,int off,int access_size,enum stack_access_src src,enum bpf_access_type type)3995 static int check_stack_access_within_bounds(
3996 struct bpf_verifier_env *env,
3997 int regno, int off, int access_size,
3998 enum stack_access_src src, enum bpf_access_type type)
3999 {
4000 struct bpf_reg_state *regs = cur_regs(env);
4001 struct bpf_reg_state *reg = regs + regno;
4002 struct bpf_func_state *state = func(env, reg);
4003 int min_off, max_off;
4004 int err;
4005 char *err_extra;
4006
4007 if (src == ACCESS_HELPER)
4008 /* We don't know if helpers are reading or writing (or both). */
4009 err_extra = " indirect access to";
4010 else if (type == BPF_READ)
4011 err_extra = " read from";
4012 else
4013 err_extra = " write to";
4014
4015 if (tnum_is_const(reg->var_off)) {
4016 min_off = reg->var_off.value + off;
4017 if (access_size > 0)
4018 max_off = min_off + access_size - 1;
4019 else
4020 max_off = min_off;
4021 } else {
4022 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4023 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4024 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4025 err_extra, regno);
4026 return -EACCES;
4027 }
4028 min_off = reg->smin_value + off;
4029 if (access_size > 0)
4030 max_off = reg->smax_value + off + access_size - 1;
4031 else
4032 max_off = min_off;
4033 }
4034
4035 err = check_stack_slot_within_bounds(min_off, state, type);
4036 if (!err)
4037 err = check_stack_slot_within_bounds(max_off, state, type);
4038
4039 if (err) {
4040 if (tnum_is_const(reg->var_off)) {
4041 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4042 err_extra, regno, off, access_size);
4043 } else {
4044 char tn_buf[48];
4045
4046 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4047 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4048 err_extra, regno, tn_buf, access_size);
4049 }
4050 }
4051 return err;
4052 }
4053
4054 /* check whether memory at (regno + off) is accessible for t = (read | write)
4055 * if t==write, value_regno is a register which value is stored into memory
4056 * if t==read, value_regno is a register which will receive the value from memory
4057 * if t==write && value_regno==-1, some unknown value is stored into memory
4058 * if t==read && value_regno==-1, don't care what we read from memory
4059 */
check_mem_access(struct bpf_verifier_env * env,int insn_idx,u32 regno,int off,int bpf_size,enum bpf_access_type t,int value_regno,bool strict_alignment_once)4060 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4061 int off, int bpf_size, enum bpf_access_type t,
4062 int value_regno, bool strict_alignment_once)
4063 {
4064 struct bpf_reg_state *regs = cur_regs(env);
4065 struct bpf_reg_state *reg = regs + regno;
4066 struct bpf_func_state *state;
4067 int size, err = 0;
4068
4069 size = bpf_size_to_bytes(bpf_size);
4070 if (size < 0)
4071 return size;
4072
4073 /* alignment checks will add in reg->off themselves */
4074 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4075 if (err)
4076 return err;
4077
4078 /* for access checks, reg->off is just part of off */
4079 off += reg->off;
4080
4081 if (reg->type == PTR_TO_MAP_KEY) {
4082 if (t == BPF_WRITE) {
4083 verbose(env, "write to change key R%d not allowed\n", regno);
4084 return -EACCES;
4085 }
4086
4087 err = check_mem_region_access(env, regno, off, size,
4088 reg->map_ptr->key_size, false);
4089 if (err)
4090 return err;
4091 if (value_regno >= 0)
4092 mark_reg_unknown(env, regs, value_regno);
4093 } else if (reg->type == PTR_TO_MAP_VALUE) {
4094 if (t == BPF_WRITE && value_regno >= 0 &&
4095 is_pointer_value(env, value_regno)) {
4096 verbose(env, "R%d leaks addr into map\n", value_regno);
4097 return -EACCES;
4098 }
4099 err = check_map_access_type(env, regno, off, size, t);
4100 if (err)
4101 return err;
4102 err = check_map_access(env, regno, off, size, false);
4103 if (!err && t == BPF_READ && value_regno >= 0) {
4104 struct bpf_map *map = reg->map_ptr;
4105
4106 /* if map is read-only, track its contents as scalars */
4107 if (tnum_is_const(reg->var_off) &&
4108 bpf_map_is_rdonly(map) &&
4109 map->ops->map_direct_value_addr) {
4110 int map_off = off + reg->var_off.value;
4111 u64 val = 0;
4112
4113 err = bpf_map_direct_read(map, map_off, size,
4114 &val);
4115 if (err)
4116 return err;
4117
4118 regs[value_regno].type = SCALAR_VALUE;
4119 __mark_reg_known(®s[value_regno], val);
4120 } else {
4121 mark_reg_unknown(env, regs, value_regno);
4122 }
4123 }
4124 } else if (reg->type == PTR_TO_MEM) {
4125 if (t == BPF_WRITE && value_regno >= 0 &&
4126 is_pointer_value(env, value_regno)) {
4127 verbose(env, "R%d leaks addr into mem\n", value_regno);
4128 return -EACCES;
4129 }
4130 err = check_mem_region_access(env, regno, off, size,
4131 reg->mem_size, false);
4132 if (!err && t == BPF_READ && value_regno >= 0)
4133 mark_reg_unknown(env, regs, value_regno);
4134 } else if (reg->type == PTR_TO_CTX) {
4135 enum bpf_reg_type reg_type = SCALAR_VALUE;
4136 struct btf *btf = NULL;
4137 u32 btf_id = 0;
4138
4139 if (t == BPF_WRITE && value_regno >= 0 &&
4140 is_pointer_value(env, value_regno)) {
4141 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4142 return -EACCES;
4143 }
4144
4145 err = check_ctx_reg(env, reg, regno);
4146 if (err < 0)
4147 return err;
4148
4149 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4150 if (err)
4151 verbose_linfo(env, insn_idx, "; ");
4152 if (!err && t == BPF_READ && value_regno >= 0) {
4153 /* ctx access returns either a scalar, or a
4154 * PTR_TO_PACKET[_META,_END]. In the latter
4155 * case, we know the offset is zero.
4156 */
4157 if (reg_type == SCALAR_VALUE) {
4158 mark_reg_unknown(env, regs, value_regno);
4159 } else {
4160 mark_reg_known_zero(env, regs,
4161 value_regno);
4162 if (reg_type_may_be_null(reg_type))
4163 regs[value_regno].id = ++env->id_gen;
4164 /* A load of ctx field could have different
4165 * actual load size with the one encoded in the
4166 * insn. When the dst is PTR, it is for sure not
4167 * a sub-register.
4168 */
4169 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4170 if (reg_type == PTR_TO_BTF_ID ||
4171 reg_type == PTR_TO_BTF_ID_OR_NULL) {
4172 regs[value_regno].btf = btf;
4173 regs[value_regno].btf_id = btf_id;
4174 }
4175 }
4176 regs[value_regno].type = reg_type;
4177 }
4178
4179 } else if (reg->type == PTR_TO_STACK) {
4180 /* Basic bounds checks. */
4181 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4182 if (err)
4183 return err;
4184
4185 state = func(env, reg);
4186 err = update_stack_depth(env, state, off);
4187 if (err)
4188 return err;
4189
4190 if (t == BPF_READ)
4191 err = check_stack_read(env, regno, off, size,
4192 value_regno);
4193 else
4194 err = check_stack_write(env, regno, off, size,
4195 value_regno, insn_idx);
4196 } else if (reg_is_pkt_pointer(reg)) {
4197 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4198 verbose(env, "cannot write into packet\n");
4199 return -EACCES;
4200 }
4201 if (t == BPF_WRITE && value_regno >= 0 &&
4202 is_pointer_value(env, value_regno)) {
4203 verbose(env, "R%d leaks addr into packet\n",
4204 value_regno);
4205 return -EACCES;
4206 }
4207 err = check_packet_access(env, regno, off, size, false);
4208 if (!err && t == BPF_READ && value_regno >= 0)
4209 mark_reg_unknown(env, regs, value_regno);
4210 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4211 if (t == BPF_WRITE && value_regno >= 0 &&
4212 is_pointer_value(env, value_regno)) {
4213 verbose(env, "R%d leaks addr into flow keys\n",
4214 value_regno);
4215 return -EACCES;
4216 }
4217
4218 err = check_flow_keys_access(env, off, size);
4219 if (!err && t == BPF_READ && value_regno >= 0)
4220 mark_reg_unknown(env, regs, value_regno);
4221 } else if (type_is_sk_pointer(reg->type)) {
4222 if (t == BPF_WRITE) {
4223 verbose(env, "R%d cannot write into %s\n",
4224 regno, reg_type_str[reg->type]);
4225 return -EACCES;
4226 }
4227 err = check_sock_access(env, insn_idx, regno, off, size, t);
4228 if (!err && value_regno >= 0)
4229 mark_reg_unknown(env, regs, value_regno);
4230 } else if (reg->type == PTR_TO_TP_BUFFER) {
4231 err = check_tp_buffer_access(env, reg, regno, off, size);
4232 if (!err && t == BPF_READ && value_regno >= 0)
4233 mark_reg_unknown(env, regs, value_regno);
4234 } else if (reg->type == PTR_TO_BTF_ID) {
4235 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4236 value_regno);
4237 } else if (reg->type == CONST_PTR_TO_MAP) {
4238 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4239 value_regno);
4240 } else if (reg->type == PTR_TO_RDONLY_BUF) {
4241 if (t == BPF_WRITE) {
4242 verbose(env, "R%d cannot write into %s\n",
4243 regno, reg_type_str[reg->type]);
4244 return -EACCES;
4245 }
4246 err = check_buffer_access(env, reg, regno, off, size, false,
4247 "rdonly",
4248 &env->prog->aux->max_rdonly_access);
4249 if (!err && value_regno >= 0)
4250 mark_reg_unknown(env, regs, value_regno);
4251 } else if (reg->type == PTR_TO_RDWR_BUF) {
4252 err = check_buffer_access(env, reg, regno, off, size, false,
4253 "rdwr",
4254 &env->prog->aux->max_rdwr_access);
4255 if (!err && t == BPF_READ && value_regno >= 0)
4256 mark_reg_unknown(env, regs, value_regno);
4257 } else {
4258 verbose(env, "R%d invalid mem access '%s'\n", regno,
4259 reg_type_str[reg->type]);
4260 return -EACCES;
4261 }
4262
4263 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4264 regs[value_regno].type == SCALAR_VALUE) {
4265 /* b/h/w load zero-extends, mark upper bits as known 0 */
4266 coerce_reg_to_size(®s[value_regno], size);
4267 }
4268 return err;
4269 }
4270
check_atomic(struct bpf_verifier_env * env,int insn_idx,struct bpf_insn * insn)4271 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4272 {
4273 int load_reg;
4274 int err;
4275
4276 switch (insn->imm) {
4277 case BPF_ADD:
4278 case BPF_ADD | BPF_FETCH:
4279 case BPF_AND:
4280 case BPF_AND | BPF_FETCH:
4281 case BPF_OR:
4282 case BPF_OR | BPF_FETCH:
4283 case BPF_XOR:
4284 case BPF_XOR | BPF_FETCH:
4285 case BPF_XCHG:
4286 case BPF_CMPXCHG:
4287 break;
4288 default:
4289 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4290 return -EINVAL;
4291 }
4292
4293 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4294 verbose(env, "invalid atomic operand size\n");
4295 return -EINVAL;
4296 }
4297
4298 /* check src1 operand */
4299 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4300 if (err)
4301 return err;
4302
4303 /* check src2 operand */
4304 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4305 if (err)
4306 return err;
4307
4308 if (insn->imm == BPF_CMPXCHG) {
4309 /* Check comparison of R0 with memory location */
4310 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4311 if (err)
4312 return err;
4313 }
4314
4315 if (is_pointer_value(env, insn->src_reg)) {
4316 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4317 return -EACCES;
4318 }
4319
4320 if (is_ctx_reg(env, insn->dst_reg) ||
4321 is_pkt_reg(env, insn->dst_reg) ||
4322 is_flow_key_reg(env, insn->dst_reg) ||
4323 is_sk_reg(env, insn->dst_reg)) {
4324 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4325 insn->dst_reg,
4326 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4327 return -EACCES;
4328 }
4329
4330 if (insn->imm & BPF_FETCH) {
4331 if (insn->imm == BPF_CMPXCHG)
4332 load_reg = BPF_REG_0;
4333 else
4334 load_reg = insn->src_reg;
4335
4336 /* check and record load of old value */
4337 err = check_reg_arg(env, load_reg, DST_OP);
4338 if (err)
4339 return err;
4340 } else {
4341 /* This instruction accesses a memory location but doesn't
4342 * actually load it into a register.
4343 */
4344 load_reg = -1;
4345 }
4346
4347 /* check whether we can read the memory */
4348 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4349 BPF_SIZE(insn->code), BPF_READ, load_reg, true);
4350 if (err)
4351 return err;
4352
4353 /* check whether we can write into the same memory */
4354 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4355 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4356 if (err)
4357 return err;
4358
4359 return 0;
4360 }
4361
4362 /* When register 'regno' is used to read the stack (either directly or through
4363 * a helper function) make sure that it's within stack boundary and, depending
4364 * on the access type, that all elements of the stack are initialized.
4365 *
4366 * 'off' includes 'regno->off', but not its dynamic part (if any).
4367 *
4368 * All registers that have been spilled on the stack in the slots within the
4369 * read offsets are marked as read.
4370 */
check_stack_range_initialized(struct bpf_verifier_env * env,int regno,int off,int access_size,bool zero_size_allowed,enum stack_access_src type,struct bpf_call_arg_meta * meta)4371 static int check_stack_range_initialized(
4372 struct bpf_verifier_env *env, int regno, int off,
4373 int access_size, bool zero_size_allowed,
4374 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4375 {
4376 struct bpf_reg_state *reg = reg_state(env, regno);
4377 struct bpf_func_state *state = func(env, reg);
4378 int err, min_off, max_off, i, j, slot, spi;
4379 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4380 enum bpf_access_type bounds_check_type;
4381 /* Some accesses can write anything into the stack, others are
4382 * read-only.
4383 */
4384 bool clobber = false;
4385
4386 if (access_size == 0 && !zero_size_allowed) {
4387 verbose(env, "invalid zero-sized read\n");
4388 return -EACCES;
4389 }
4390
4391 if (type == ACCESS_HELPER) {
4392 /* The bounds checks for writes are more permissive than for
4393 * reads. However, if raw_mode is not set, we'll do extra
4394 * checks below.
4395 */
4396 bounds_check_type = BPF_WRITE;
4397 clobber = true;
4398 } else {
4399 bounds_check_type = BPF_READ;
4400 }
4401 err = check_stack_access_within_bounds(env, regno, off, access_size,
4402 type, bounds_check_type);
4403 if (err)
4404 return err;
4405
4406
4407 if (tnum_is_const(reg->var_off)) {
4408 min_off = max_off = reg->var_off.value + off;
4409 } else {
4410 /* Variable offset is prohibited for unprivileged mode for
4411 * simplicity since it requires corresponding support in
4412 * Spectre masking for stack ALU.
4413 * See also retrieve_ptr_limit().
4414 */
4415 if (!env->bypass_spec_v1) {
4416 char tn_buf[48];
4417
4418 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4419 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4420 regno, err_extra, tn_buf);
4421 return -EACCES;
4422 }
4423 /* Only initialized buffer on stack is allowed to be accessed
4424 * with variable offset. With uninitialized buffer it's hard to
4425 * guarantee that whole memory is marked as initialized on
4426 * helper return since specific bounds are unknown what may
4427 * cause uninitialized stack leaking.
4428 */
4429 if (meta && meta->raw_mode)
4430 meta = NULL;
4431
4432 min_off = reg->smin_value + off;
4433 max_off = reg->smax_value + off;
4434 }
4435
4436 if (meta && meta->raw_mode) {
4437 meta->access_size = access_size;
4438 meta->regno = regno;
4439 return 0;
4440 }
4441
4442 for (i = min_off; i < max_off + access_size; i++) {
4443 u8 *stype;
4444
4445 slot = -i - 1;
4446 spi = slot / BPF_REG_SIZE;
4447 if (state->allocated_stack <= slot)
4448 goto err;
4449 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4450 if (*stype == STACK_MISC)
4451 goto mark;
4452 if (*stype == STACK_ZERO) {
4453 if (clobber) {
4454 /* helper can write anything into the stack */
4455 *stype = STACK_MISC;
4456 }
4457 goto mark;
4458 }
4459
4460 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4461 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4462 goto mark;
4463
4464 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4465 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4466 env->allow_ptr_leaks)) {
4467 if (clobber) {
4468 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4469 for (j = 0; j < BPF_REG_SIZE; j++)
4470 state->stack[spi].slot_type[j] = STACK_MISC;
4471 }
4472 goto mark;
4473 }
4474
4475 err:
4476 if (tnum_is_const(reg->var_off)) {
4477 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4478 err_extra, regno, min_off, i - min_off, access_size);
4479 } else {
4480 char tn_buf[48];
4481
4482 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4483 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4484 err_extra, regno, tn_buf, i - min_off, access_size);
4485 }
4486 return -EACCES;
4487 mark:
4488 /* reading any byte out of 8-byte 'spill_slot' will cause
4489 * the whole slot to be marked as 'read'
4490 */
4491 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4492 state->stack[spi].spilled_ptr.parent,
4493 REG_LIVE_READ64);
4494 }
4495 return update_stack_depth(env, state, min_off);
4496 }
4497
check_helper_mem_access(struct bpf_verifier_env * env,int regno,int access_size,bool zero_size_allowed,struct bpf_call_arg_meta * meta)4498 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4499 int access_size, bool zero_size_allowed,
4500 struct bpf_call_arg_meta *meta)
4501 {
4502 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4503
4504 switch (reg->type) {
4505 case PTR_TO_PACKET:
4506 case PTR_TO_PACKET_META:
4507 return check_packet_access(env, regno, reg->off, access_size,
4508 zero_size_allowed);
4509 case PTR_TO_MAP_KEY:
4510 return check_mem_region_access(env, regno, reg->off, access_size,
4511 reg->map_ptr->key_size, false);
4512 case PTR_TO_MAP_VALUE:
4513 if (check_map_access_type(env, regno, reg->off, access_size,
4514 meta && meta->raw_mode ? BPF_WRITE :
4515 BPF_READ))
4516 return -EACCES;
4517 return check_map_access(env, regno, reg->off, access_size,
4518 zero_size_allowed);
4519 case PTR_TO_MEM:
4520 return check_mem_region_access(env, regno, reg->off,
4521 access_size, reg->mem_size,
4522 zero_size_allowed);
4523 case PTR_TO_RDONLY_BUF:
4524 if (meta && meta->raw_mode)
4525 return -EACCES;
4526 return check_buffer_access(env, reg, regno, reg->off,
4527 access_size, zero_size_allowed,
4528 "rdonly",
4529 &env->prog->aux->max_rdonly_access);
4530 case PTR_TO_RDWR_BUF:
4531 return check_buffer_access(env, reg, regno, reg->off,
4532 access_size, zero_size_allowed,
4533 "rdwr",
4534 &env->prog->aux->max_rdwr_access);
4535 case PTR_TO_STACK:
4536 return check_stack_range_initialized(
4537 env,
4538 regno, reg->off, access_size,
4539 zero_size_allowed, ACCESS_HELPER, meta);
4540 default: /* scalar_value or invalid ptr */
4541 /* Allow zero-byte read from NULL, regardless of pointer type */
4542 if (zero_size_allowed && access_size == 0 &&
4543 register_is_null(reg))
4544 return 0;
4545
4546 verbose(env, "R%d type=%s expected=%s\n", regno,
4547 reg_type_str[reg->type],
4548 reg_type_str[PTR_TO_STACK]);
4549 return -EACCES;
4550 }
4551 }
4552
check_mem_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,u32 mem_size)4553 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4554 u32 regno, u32 mem_size)
4555 {
4556 if (register_is_null(reg))
4557 return 0;
4558
4559 if (reg_type_may_be_null(reg->type)) {
4560 /* Assuming that the register contains a value check if the memory
4561 * access is safe. Temporarily save and restore the register's state as
4562 * the conversion shouldn't be visible to a caller.
4563 */
4564 const struct bpf_reg_state saved_reg = *reg;
4565 int rv;
4566
4567 mark_ptr_not_null_reg(reg);
4568 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4569 *reg = saved_reg;
4570 return rv;
4571 }
4572
4573 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4574 }
4575
4576 /* Implementation details:
4577 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4578 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4579 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4580 * value_or_null->value transition, since the verifier only cares about
4581 * the range of access to valid map value pointer and doesn't care about actual
4582 * address of the map element.
4583 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4584 * reg->id > 0 after value_or_null->value transition. By doing so
4585 * two bpf_map_lookups will be considered two different pointers that
4586 * point to different bpf_spin_locks.
4587 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4588 * dead-locks.
4589 * Since only one bpf_spin_lock is allowed the checks are simpler than
4590 * reg_is_refcounted() logic. The verifier needs to remember only
4591 * one spin_lock instead of array of acquired_refs.
4592 * cur_state->active_spin_lock remembers which map value element got locked
4593 * and clears it after bpf_spin_unlock.
4594 */
process_spin_lock(struct bpf_verifier_env * env,int regno,bool is_lock)4595 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4596 bool is_lock)
4597 {
4598 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4599 struct bpf_verifier_state *cur = env->cur_state;
4600 bool is_const = tnum_is_const(reg->var_off);
4601 struct bpf_map *map = reg->map_ptr;
4602 u64 val = reg->var_off.value;
4603
4604 if (!is_const) {
4605 verbose(env,
4606 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4607 regno);
4608 return -EINVAL;
4609 }
4610 if (!map->btf) {
4611 verbose(env,
4612 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4613 map->name);
4614 return -EINVAL;
4615 }
4616 if (!map_value_has_spin_lock(map)) {
4617 if (map->spin_lock_off == -E2BIG)
4618 verbose(env,
4619 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4620 map->name);
4621 else if (map->spin_lock_off == -ENOENT)
4622 verbose(env,
4623 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4624 map->name);
4625 else
4626 verbose(env,
4627 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4628 map->name);
4629 return -EINVAL;
4630 }
4631 if (map->spin_lock_off != val + reg->off) {
4632 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4633 val + reg->off);
4634 return -EINVAL;
4635 }
4636 if (is_lock) {
4637 if (cur->active_spin_lock) {
4638 verbose(env,
4639 "Locking two bpf_spin_locks are not allowed\n");
4640 return -EINVAL;
4641 }
4642 cur->active_spin_lock = reg->id;
4643 } else {
4644 if (!cur->active_spin_lock) {
4645 verbose(env, "bpf_spin_unlock without taking a lock\n");
4646 return -EINVAL;
4647 }
4648 if (cur->active_spin_lock != reg->id) {
4649 verbose(env, "bpf_spin_unlock of different lock\n");
4650 return -EINVAL;
4651 }
4652 cur->active_spin_lock = 0;
4653 }
4654 return 0;
4655 }
4656
arg_type_is_mem_ptr(enum bpf_arg_type type)4657 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4658 {
4659 return type == ARG_PTR_TO_MEM ||
4660 type == ARG_PTR_TO_MEM_OR_NULL ||
4661 type == ARG_PTR_TO_UNINIT_MEM;
4662 }
4663
arg_type_is_mem_size(enum bpf_arg_type type)4664 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4665 {
4666 return type == ARG_CONST_SIZE ||
4667 type == ARG_CONST_SIZE_OR_ZERO;
4668 }
4669
arg_type_is_alloc_size(enum bpf_arg_type type)4670 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4671 {
4672 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4673 }
4674
arg_type_is_int_ptr(enum bpf_arg_type type)4675 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4676 {
4677 return type == ARG_PTR_TO_INT ||
4678 type == ARG_PTR_TO_LONG;
4679 }
4680
int_ptr_type_to_size(enum bpf_arg_type type)4681 static int int_ptr_type_to_size(enum bpf_arg_type type)
4682 {
4683 if (type == ARG_PTR_TO_INT)
4684 return sizeof(u32);
4685 else if (type == ARG_PTR_TO_LONG)
4686 return sizeof(u64);
4687
4688 return -EINVAL;
4689 }
4690
resolve_map_arg_type(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_arg_type * arg_type)4691 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4692 const struct bpf_call_arg_meta *meta,
4693 enum bpf_arg_type *arg_type)
4694 {
4695 if (!meta->map_ptr) {
4696 /* kernel subsystem misconfigured verifier */
4697 verbose(env, "invalid map_ptr to access map->type\n");
4698 return -EACCES;
4699 }
4700
4701 switch (meta->map_ptr->map_type) {
4702 case BPF_MAP_TYPE_SOCKMAP:
4703 case BPF_MAP_TYPE_SOCKHASH:
4704 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4705 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4706 } else {
4707 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4708 return -EINVAL;
4709 }
4710 break;
4711
4712 default:
4713 break;
4714 }
4715 return 0;
4716 }
4717
4718 struct bpf_reg_types {
4719 const enum bpf_reg_type types[10];
4720 u32 *btf_id;
4721 };
4722
4723 static const struct bpf_reg_types map_key_value_types = {
4724 .types = {
4725 PTR_TO_STACK,
4726 PTR_TO_PACKET,
4727 PTR_TO_PACKET_META,
4728 PTR_TO_MAP_KEY,
4729 PTR_TO_MAP_VALUE,
4730 },
4731 };
4732
4733 static const struct bpf_reg_types sock_types = {
4734 .types = {
4735 PTR_TO_SOCK_COMMON,
4736 PTR_TO_SOCKET,
4737 PTR_TO_TCP_SOCK,
4738 PTR_TO_XDP_SOCK,
4739 },
4740 };
4741
4742 #ifdef CONFIG_NET
4743 static const struct bpf_reg_types btf_id_sock_common_types = {
4744 .types = {
4745 PTR_TO_SOCK_COMMON,
4746 PTR_TO_SOCKET,
4747 PTR_TO_TCP_SOCK,
4748 PTR_TO_XDP_SOCK,
4749 PTR_TO_BTF_ID,
4750 },
4751 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4752 };
4753 #endif
4754
4755 static const struct bpf_reg_types mem_types = {
4756 .types = {
4757 PTR_TO_STACK,
4758 PTR_TO_PACKET,
4759 PTR_TO_PACKET_META,
4760 PTR_TO_MAP_KEY,
4761 PTR_TO_MAP_VALUE,
4762 PTR_TO_MEM,
4763 PTR_TO_RDONLY_BUF,
4764 PTR_TO_RDWR_BUF,
4765 },
4766 };
4767
4768 static const struct bpf_reg_types int_ptr_types = {
4769 .types = {
4770 PTR_TO_STACK,
4771 PTR_TO_PACKET,
4772 PTR_TO_PACKET_META,
4773 PTR_TO_MAP_KEY,
4774 PTR_TO_MAP_VALUE,
4775 },
4776 };
4777
4778 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4779 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4780 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4781 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4782 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4783 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4784 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4785 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4786 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
4787 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
4788 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
4789
4790 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4791 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4792 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4793 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4794 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4795 [ARG_CONST_SIZE] = &scalar_types,
4796 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4797 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4798 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4799 [ARG_PTR_TO_CTX] = &context_types,
4800 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4801 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4802 #ifdef CONFIG_NET
4803 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4804 #endif
4805 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4806 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4807 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4808 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4809 [ARG_PTR_TO_MEM] = &mem_types,
4810 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4811 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4812 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4813 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4814 [ARG_PTR_TO_INT] = &int_ptr_types,
4815 [ARG_PTR_TO_LONG] = &int_ptr_types,
4816 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4817 [ARG_PTR_TO_FUNC] = &func_ptr_types,
4818 [ARG_PTR_TO_STACK_OR_NULL] = &stack_ptr_types,
4819 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
4820 };
4821
check_reg_type(struct bpf_verifier_env * env,u32 regno,enum bpf_arg_type arg_type,const u32 * arg_btf_id)4822 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4823 enum bpf_arg_type arg_type,
4824 const u32 *arg_btf_id)
4825 {
4826 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4827 enum bpf_reg_type expected, type = reg->type;
4828 const struct bpf_reg_types *compatible;
4829 int i, j;
4830
4831 compatible = compatible_reg_types[arg_type];
4832 if (!compatible) {
4833 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4834 return -EFAULT;
4835 }
4836
4837 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4838 expected = compatible->types[i];
4839 if (expected == NOT_INIT)
4840 break;
4841
4842 if (type == expected)
4843 goto found;
4844 }
4845
4846 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4847 for (j = 0; j + 1 < i; j++)
4848 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4849 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4850 return -EACCES;
4851
4852 found:
4853 if (type == PTR_TO_BTF_ID) {
4854 if (!arg_btf_id) {
4855 if (!compatible->btf_id) {
4856 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4857 return -EFAULT;
4858 }
4859 arg_btf_id = compatible->btf_id;
4860 }
4861
4862 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
4863 btf_vmlinux, *arg_btf_id)) {
4864 verbose(env, "R%d is of type %s but %s is expected\n",
4865 regno, kernel_type_name(reg->btf, reg->btf_id),
4866 kernel_type_name(btf_vmlinux, *arg_btf_id));
4867 return -EACCES;
4868 }
4869
4870 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4871 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4872 regno);
4873 return -EACCES;
4874 }
4875 }
4876
4877 return 0;
4878 }
4879
check_func_arg(struct bpf_verifier_env * env,u32 arg,struct bpf_call_arg_meta * meta,const struct bpf_func_proto * fn)4880 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4881 struct bpf_call_arg_meta *meta,
4882 const struct bpf_func_proto *fn)
4883 {
4884 u32 regno = BPF_REG_1 + arg;
4885 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4886 enum bpf_arg_type arg_type = fn->arg_type[arg];
4887 enum bpf_reg_type type = reg->type;
4888 int err = 0;
4889
4890 if (arg_type == ARG_DONTCARE)
4891 return 0;
4892
4893 err = check_reg_arg(env, regno, SRC_OP);
4894 if (err)
4895 return err;
4896
4897 if (arg_type == ARG_ANYTHING) {
4898 if (is_pointer_value(env, regno)) {
4899 verbose(env, "R%d leaks addr into helper function\n",
4900 regno);
4901 return -EACCES;
4902 }
4903 return 0;
4904 }
4905
4906 if (type_is_pkt_pointer(type) &&
4907 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4908 verbose(env, "helper access to the packet is not allowed\n");
4909 return -EACCES;
4910 }
4911
4912 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4913 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4914 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4915 err = resolve_map_arg_type(env, meta, &arg_type);
4916 if (err)
4917 return err;
4918 }
4919
4920 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4921 /* A NULL register has a SCALAR_VALUE type, so skip
4922 * type checking.
4923 */
4924 goto skip_type_check;
4925
4926 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4927 if (err)
4928 return err;
4929
4930 if (type == PTR_TO_CTX) {
4931 err = check_ctx_reg(env, reg, regno);
4932 if (err < 0)
4933 return err;
4934 }
4935
4936 skip_type_check:
4937 if (reg->ref_obj_id) {
4938 if (meta->ref_obj_id) {
4939 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4940 regno, reg->ref_obj_id,
4941 meta->ref_obj_id);
4942 return -EFAULT;
4943 }
4944 meta->ref_obj_id = reg->ref_obj_id;
4945 }
4946
4947 if (arg_type == ARG_CONST_MAP_PTR) {
4948 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4949 meta->map_ptr = reg->map_ptr;
4950 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4951 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4952 * check that [key, key + map->key_size) are within
4953 * stack limits and initialized
4954 */
4955 if (!meta->map_ptr) {
4956 /* in function declaration map_ptr must come before
4957 * map_key, so that it's verified and known before
4958 * we have to check map_key here. Otherwise it means
4959 * that kernel subsystem misconfigured verifier
4960 */
4961 verbose(env, "invalid map_ptr to access map->key\n");
4962 return -EACCES;
4963 }
4964 err = check_helper_mem_access(env, regno,
4965 meta->map_ptr->key_size, false,
4966 NULL);
4967 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4968 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4969 !register_is_null(reg)) ||
4970 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4971 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4972 * check [value, value + map->value_size) validity
4973 */
4974 if (!meta->map_ptr) {
4975 /* kernel subsystem misconfigured verifier */
4976 verbose(env, "invalid map_ptr to access map->value\n");
4977 return -EACCES;
4978 }
4979 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4980 err = check_helper_mem_access(env, regno,
4981 meta->map_ptr->value_size, false,
4982 meta);
4983 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
4984 if (!reg->btf_id) {
4985 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
4986 return -EACCES;
4987 }
4988 meta->ret_btf = reg->btf;
4989 meta->ret_btf_id = reg->btf_id;
4990 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4991 if (meta->func_id == BPF_FUNC_spin_lock) {
4992 if (process_spin_lock(env, regno, true))
4993 return -EACCES;
4994 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4995 if (process_spin_lock(env, regno, false))
4996 return -EACCES;
4997 } else {
4998 verbose(env, "verifier internal error\n");
4999 return -EFAULT;
5000 }
5001 } else if (arg_type == ARG_PTR_TO_FUNC) {
5002 meta->subprogno = reg->subprogno;
5003 } else if (arg_type_is_mem_ptr(arg_type)) {
5004 /* The access to this pointer is only checked when we hit the
5005 * next is_mem_size argument below.
5006 */
5007 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5008 } else if (arg_type_is_mem_size(arg_type)) {
5009 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5010
5011 /* This is used to refine r0 return value bounds for helpers
5012 * that enforce this value as an upper bound on return values.
5013 * See do_refine_retval_range() for helpers that can refine
5014 * the return value. C type of helper is u32 so we pull register
5015 * bound from umax_value however, if negative verifier errors
5016 * out. Only upper bounds can be learned because retval is an
5017 * int type and negative retvals are allowed.
5018 */
5019 meta->msize_max_value = reg->umax_value;
5020
5021 /* The register is SCALAR_VALUE; the access check
5022 * happens using its boundaries.
5023 */
5024 if (!tnum_is_const(reg->var_off))
5025 /* For unprivileged variable accesses, disable raw
5026 * mode so that the program is required to
5027 * initialize all the memory that the helper could
5028 * just partially fill up.
5029 */
5030 meta = NULL;
5031
5032 if (reg->smin_value < 0) {
5033 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5034 regno);
5035 return -EACCES;
5036 }
5037
5038 if (reg->umin_value == 0) {
5039 err = check_helper_mem_access(env, regno - 1, 0,
5040 zero_size_allowed,
5041 meta);
5042 if (err)
5043 return err;
5044 }
5045
5046 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5047 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5048 regno);
5049 return -EACCES;
5050 }
5051 err = check_helper_mem_access(env, regno - 1,
5052 reg->umax_value,
5053 zero_size_allowed, meta);
5054 if (!err)
5055 err = mark_chain_precision(env, regno);
5056 } else if (arg_type_is_alloc_size(arg_type)) {
5057 if (!tnum_is_const(reg->var_off)) {
5058 verbose(env, "R%d is not a known constant'\n",
5059 regno);
5060 return -EACCES;
5061 }
5062 meta->mem_size = reg->var_off.value;
5063 } else if (arg_type_is_int_ptr(arg_type)) {
5064 int size = int_ptr_type_to_size(arg_type);
5065
5066 err = check_helper_mem_access(env, regno, size, false, meta);
5067 if (err)
5068 return err;
5069 err = check_ptr_alignment(env, reg, 0, size, true);
5070 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5071 struct bpf_map *map = reg->map_ptr;
5072 int map_off;
5073 u64 map_addr;
5074 char *str_ptr;
5075
5076 if (!bpf_map_is_rdonly(map)) {
5077 verbose(env, "R%d does not point to a readonly map'\n", regno);
5078 return -EACCES;
5079 }
5080
5081 if (!tnum_is_const(reg->var_off)) {
5082 verbose(env, "R%d is not a constant address'\n", regno);
5083 return -EACCES;
5084 }
5085
5086 if (!map->ops->map_direct_value_addr) {
5087 verbose(env, "no direct value access support for this map type\n");
5088 return -EACCES;
5089 }
5090
5091 err = check_map_access(env, regno, reg->off,
5092 map->value_size - reg->off, false);
5093 if (err)
5094 return err;
5095
5096 map_off = reg->off + reg->var_off.value;
5097 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5098 if (err) {
5099 verbose(env, "direct value access on string failed\n");
5100 return err;
5101 }
5102
5103 str_ptr = (char *)(long)(map_addr);
5104 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5105 verbose(env, "string is not zero-terminated\n");
5106 return -EINVAL;
5107 }
5108 }
5109
5110 return err;
5111 }
5112
may_update_sockmap(struct bpf_verifier_env * env,int func_id)5113 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5114 {
5115 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5116 enum bpf_prog_type type = resolve_prog_type(env->prog);
5117
5118 if (func_id != BPF_FUNC_map_update_elem)
5119 return false;
5120
5121 /* It's not possible to get access to a locked struct sock in these
5122 * contexts, so updating is safe.
5123 */
5124 switch (type) {
5125 case BPF_PROG_TYPE_TRACING:
5126 if (eatype == BPF_TRACE_ITER)
5127 return true;
5128 break;
5129 case BPF_PROG_TYPE_SOCKET_FILTER:
5130 case BPF_PROG_TYPE_SCHED_CLS:
5131 case BPF_PROG_TYPE_SCHED_ACT:
5132 case BPF_PROG_TYPE_XDP:
5133 case BPF_PROG_TYPE_SK_REUSEPORT:
5134 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5135 case BPF_PROG_TYPE_SK_LOOKUP:
5136 return true;
5137 default:
5138 break;
5139 }
5140
5141 verbose(env, "cannot update sockmap in this context\n");
5142 return false;
5143 }
5144
allow_tail_call_in_subprogs(struct bpf_verifier_env * env)5145 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5146 {
5147 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5148 }
5149
check_map_func_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,int func_id)5150 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5151 struct bpf_map *map, int func_id)
5152 {
5153 if (!map)
5154 return 0;
5155
5156 /* We need a two way check, first is from map perspective ... */
5157 switch (map->map_type) {
5158 case BPF_MAP_TYPE_PROG_ARRAY:
5159 if (func_id != BPF_FUNC_tail_call)
5160 goto error;
5161 break;
5162 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5163 if (func_id != BPF_FUNC_perf_event_read &&
5164 func_id != BPF_FUNC_perf_event_output &&
5165 func_id != BPF_FUNC_skb_output &&
5166 func_id != BPF_FUNC_perf_event_read_value &&
5167 func_id != BPF_FUNC_xdp_output)
5168 goto error;
5169 break;
5170 case BPF_MAP_TYPE_RINGBUF:
5171 if (func_id != BPF_FUNC_ringbuf_output &&
5172 func_id != BPF_FUNC_ringbuf_reserve &&
5173 func_id != BPF_FUNC_ringbuf_submit &&
5174 func_id != BPF_FUNC_ringbuf_discard &&
5175 func_id != BPF_FUNC_ringbuf_query)
5176 goto error;
5177 break;
5178 case BPF_MAP_TYPE_STACK_TRACE:
5179 if (func_id != BPF_FUNC_get_stackid)
5180 goto error;
5181 break;
5182 case BPF_MAP_TYPE_CGROUP_ARRAY:
5183 if (func_id != BPF_FUNC_skb_under_cgroup &&
5184 func_id != BPF_FUNC_current_task_under_cgroup)
5185 goto error;
5186 break;
5187 case BPF_MAP_TYPE_CGROUP_STORAGE:
5188 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5189 if (func_id != BPF_FUNC_get_local_storage)
5190 goto error;
5191 break;
5192 case BPF_MAP_TYPE_DEVMAP:
5193 case BPF_MAP_TYPE_DEVMAP_HASH:
5194 if (func_id != BPF_FUNC_redirect_map &&
5195 func_id != BPF_FUNC_map_lookup_elem)
5196 goto error;
5197 break;
5198 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5199 * appear.
5200 */
5201 case BPF_MAP_TYPE_CPUMAP:
5202 if (func_id != BPF_FUNC_redirect_map)
5203 goto error;
5204 break;
5205 case BPF_MAP_TYPE_XSKMAP:
5206 if (func_id != BPF_FUNC_redirect_map &&
5207 func_id != BPF_FUNC_map_lookup_elem)
5208 goto error;
5209 break;
5210 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5211 case BPF_MAP_TYPE_HASH_OF_MAPS:
5212 if (func_id != BPF_FUNC_map_lookup_elem)
5213 goto error;
5214 break;
5215 case BPF_MAP_TYPE_SOCKMAP:
5216 if (func_id != BPF_FUNC_sk_redirect_map &&
5217 func_id != BPF_FUNC_sock_map_update &&
5218 func_id != BPF_FUNC_map_delete_elem &&
5219 func_id != BPF_FUNC_msg_redirect_map &&
5220 func_id != BPF_FUNC_sk_select_reuseport &&
5221 func_id != BPF_FUNC_map_lookup_elem &&
5222 !may_update_sockmap(env, func_id))
5223 goto error;
5224 break;
5225 case BPF_MAP_TYPE_SOCKHASH:
5226 if (func_id != BPF_FUNC_sk_redirect_hash &&
5227 func_id != BPF_FUNC_sock_hash_update &&
5228 func_id != BPF_FUNC_map_delete_elem &&
5229 func_id != BPF_FUNC_msg_redirect_hash &&
5230 func_id != BPF_FUNC_sk_select_reuseport &&
5231 func_id != BPF_FUNC_map_lookup_elem &&
5232 !may_update_sockmap(env, func_id))
5233 goto error;
5234 break;
5235 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5236 if (func_id != BPF_FUNC_sk_select_reuseport)
5237 goto error;
5238 break;
5239 case BPF_MAP_TYPE_QUEUE:
5240 case BPF_MAP_TYPE_STACK:
5241 if (func_id != BPF_FUNC_map_peek_elem &&
5242 func_id != BPF_FUNC_map_pop_elem &&
5243 func_id != BPF_FUNC_map_push_elem)
5244 goto error;
5245 break;
5246 case BPF_MAP_TYPE_SK_STORAGE:
5247 if (func_id != BPF_FUNC_sk_storage_get &&
5248 func_id != BPF_FUNC_sk_storage_delete)
5249 goto error;
5250 break;
5251 case BPF_MAP_TYPE_INODE_STORAGE:
5252 if (func_id != BPF_FUNC_inode_storage_get &&
5253 func_id != BPF_FUNC_inode_storage_delete)
5254 goto error;
5255 break;
5256 case BPF_MAP_TYPE_TASK_STORAGE:
5257 if (func_id != BPF_FUNC_task_storage_get &&
5258 func_id != BPF_FUNC_task_storage_delete)
5259 goto error;
5260 break;
5261 default:
5262 break;
5263 }
5264
5265 /* ... and second from the function itself. */
5266 switch (func_id) {
5267 case BPF_FUNC_tail_call:
5268 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5269 goto error;
5270 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5271 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5272 return -EINVAL;
5273 }
5274 break;
5275 case BPF_FUNC_perf_event_read:
5276 case BPF_FUNC_perf_event_output:
5277 case BPF_FUNC_perf_event_read_value:
5278 case BPF_FUNC_skb_output:
5279 case BPF_FUNC_xdp_output:
5280 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5281 goto error;
5282 break;
5283 case BPF_FUNC_get_stackid:
5284 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5285 goto error;
5286 break;
5287 case BPF_FUNC_current_task_under_cgroup:
5288 case BPF_FUNC_skb_under_cgroup:
5289 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5290 goto error;
5291 break;
5292 case BPF_FUNC_redirect_map:
5293 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5294 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5295 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5296 map->map_type != BPF_MAP_TYPE_XSKMAP)
5297 goto error;
5298 break;
5299 case BPF_FUNC_sk_redirect_map:
5300 case BPF_FUNC_msg_redirect_map:
5301 case BPF_FUNC_sock_map_update:
5302 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5303 goto error;
5304 break;
5305 case BPF_FUNC_sk_redirect_hash:
5306 case BPF_FUNC_msg_redirect_hash:
5307 case BPF_FUNC_sock_hash_update:
5308 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5309 goto error;
5310 break;
5311 case BPF_FUNC_get_local_storage:
5312 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5313 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5314 goto error;
5315 break;
5316 case BPF_FUNC_sk_select_reuseport:
5317 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5318 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5319 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5320 goto error;
5321 break;
5322 case BPF_FUNC_map_peek_elem:
5323 case BPF_FUNC_map_pop_elem:
5324 case BPF_FUNC_map_push_elem:
5325 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5326 map->map_type != BPF_MAP_TYPE_STACK)
5327 goto error;
5328 break;
5329 case BPF_FUNC_sk_storage_get:
5330 case BPF_FUNC_sk_storage_delete:
5331 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5332 goto error;
5333 break;
5334 case BPF_FUNC_inode_storage_get:
5335 case BPF_FUNC_inode_storage_delete:
5336 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5337 goto error;
5338 break;
5339 case BPF_FUNC_task_storage_get:
5340 case BPF_FUNC_task_storage_delete:
5341 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5342 goto error;
5343 break;
5344 default:
5345 break;
5346 }
5347
5348 return 0;
5349 error:
5350 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5351 map->map_type, func_id_name(func_id), func_id);
5352 return -EINVAL;
5353 }
5354
check_raw_mode_ok(const struct bpf_func_proto * fn)5355 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5356 {
5357 int count = 0;
5358
5359 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5360 count++;
5361 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5362 count++;
5363 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5364 count++;
5365 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5366 count++;
5367 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5368 count++;
5369
5370 /* We only support one arg being in raw mode at the moment,
5371 * which is sufficient for the helper functions we have
5372 * right now.
5373 */
5374 return count <= 1;
5375 }
5376
check_args_pair_invalid(enum bpf_arg_type arg_curr,enum bpf_arg_type arg_next)5377 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5378 enum bpf_arg_type arg_next)
5379 {
5380 return (arg_type_is_mem_ptr(arg_curr) &&
5381 !arg_type_is_mem_size(arg_next)) ||
5382 (!arg_type_is_mem_ptr(arg_curr) &&
5383 arg_type_is_mem_size(arg_next));
5384 }
5385
check_arg_pair_ok(const struct bpf_func_proto * fn)5386 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5387 {
5388 /* bpf_xxx(..., buf, len) call will access 'len'
5389 * bytes from memory 'buf'. Both arg types need
5390 * to be paired, so make sure there's no buggy
5391 * helper function specification.
5392 */
5393 if (arg_type_is_mem_size(fn->arg1_type) ||
5394 arg_type_is_mem_ptr(fn->arg5_type) ||
5395 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5396 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5397 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5398 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5399 return false;
5400
5401 return true;
5402 }
5403
check_refcount_ok(const struct bpf_func_proto * fn,int func_id)5404 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5405 {
5406 int count = 0;
5407
5408 if (arg_type_may_be_refcounted(fn->arg1_type))
5409 count++;
5410 if (arg_type_may_be_refcounted(fn->arg2_type))
5411 count++;
5412 if (arg_type_may_be_refcounted(fn->arg3_type))
5413 count++;
5414 if (arg_type_may_be_refcounted(fn->arg4_type))
5415 count++;
5416 if (arg_type_may_be_refcounted(fn->arg5_type))
5417 count++;
5418
5419 /* A reference acquiring function cannot acquire
5420 * another refcounted ptr.
5421 */
5422 if (may_be_acquire_function(func_id) && count)
5423 return false;
5424
5425 /* We only support one arg being unreferenced at the moment,
5426 * which is sufficient for the helper functions we have right now.
5427 */
5428 return count <= 1;
5429 }
5430
check_btf_id_ok(const struct bpf_func_proto * fn)5431 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5432 {
5433 int i;
5434
5435 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5436 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5437 return false;
5438
5439 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5440 return false;
5441 }
5442
5443 return true;
5444 }
5445
check_func_proto(const struct bpf_func_proto * fn,int func_id)5446 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5447 {
5448 return check_raw_mode_ok(fn) &&
5449 check_arg_pair_ok(fn) &&
5450 check_btf_id_ok(fn) &&
5451 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5452 }
5453
5454 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5455 * are now invalid, so turn them into unknown SCALAR_VALUE.
5456 */
__clear_all_pkt_pointers(struct bpf_verifier_env * env,struct bpf_func_state * state)5457 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5458 struct bpf_func_state *state)
5459 {
5460 struct bpf_reg_state *regs = state->regs, *reg;
5461 int i;
5462
5463 for (i = 0; i < MAX_BPF_REG; i++)
5464 if (reg_is_pkt_pointer_any(®s[i]))
5465 mark_reg_unknown(env, regs, i);
5466
5467 bpf_for_each_spilled_reg(i, state, reg) {
5468 if (!reg)
5469 continue;
5470 if (reg_is_pkt_pointer_any(reg))
5471 __mark_reg_unknown(env, reg);
5472 }
5473 }
5474
clear_all_pkt_pointers(struct bpf_verifier_env * env)5475 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5476 {
5477 struct bpf_verifier_state *vstate = env->cur_state;
5478 int i;
5479
5480 for (i = 0; i <= vstate->curframe; i++)
5481 __clear_all_pkt_pointers(env, vstate->frame[i]);
5482 }
5483
5484 enum {
5485 AT_PKT_END = -1,
5486 BEYOND_PKT_END = -2,
5487 };
5488
mark_pkt_end(struct bpf_verifier_state * vstate,int regn,bool range_open)5489 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5490 {
5491 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5492 struct bpf_reg_state *reg = &state->regs[regn];
5493
5494 if (reg->type != PTR_TO_PACKET)
5495 /* PTR_TO_PACKET_META is not supported yet */
5496 return;
5497
5498 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5499 * How far beyond pkt_end it goes is unknown.
5500 * if (!range_open) it's the case of pkt >= pkt_end
5501 * if (range_open) it's the case of pkt > pkt_end
5502 * hence this pointer is at least 1 byte bigger than pkt_end
5503 */
5504 if (range_open)
5505 reg->range = BEYOND_PKT_END;
5506 else
5507 reg->range = AT_PKT_END;
5508 }
5509
release_reg_references(struct bpf_verifier_env * env,struct bpf_func_state * state,int ref_obj_id)5510 static void release_reg_references(struct bpf_verifier_env *env,
5511 struct bpf_func_state *state,
5512 int ref_obj_id)
5513 {
5514 struct bpf_reg_state *regs = state->regs, *reg;
5515 int i;
5516
5517 for (i = 0; i < MAX_BPF_REG; i++)
5518 if (regs[i].ref_obj_id == ref_obj_id)
5519 mark_reg_unknown(env, regs, i);
5520
5521 bpf_for_each_spilled_reg(i, state, reg) {
5522 if (!reg)
5523 continue;
5524 if (reg->ref_obj_id == ref_obj_id)
5525 __mark_reg_unknown(env, reg);
5526 }
5527 }
5528
5529 /* The pointer with the specified id has released its reference to kernel
5530 * resources. Identify all copies of the same pointer and clear the reference.
5531 */
release_reference(struct bpf_verifier_env * env,int ref_obj_id)5532 static int release_reference(struct bpf_verifier_env *env,
5533 int ref_obj_id)
5534 {
5535 struct bpf_verifier_state *vstate = env->cur_state;
5536 int err;
5537 int i;
5538
5539 err = release_reference_state(cur_func(env), ref_obj_id);
5540 if (err)
5541 return err;
5542
5543 for (i = 0; i <= vstate->curframe; i++)
5544 release_reg_references(env, vstate->frame[i], ref_obj_id);
5545
5546 return 0;
5547 }
5548
clear_caller_saved_regs(struct bpf_verifier_env * env,struct bpf_reg_state * regs)5549 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5550 struct bpf_reg_state *regs)
5551 {
5552 int i;
5553
5554 /* after the call registers r0 - r5 were scratched */
5555 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5556 mark_reg_not_init(env, regs, caller_saved[i]);
5557 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5558 }
5559 }
5560
5561 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5562 struct bpf_func_state *caller,
5563 struct bpf_func_state *callee,
5564 int insn_idx);
5565
__check_func_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx,int subprog,set_callee_state_fn set_callee_state_cb)5566 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5567 int *insn_idx, int subprog,
5568 set_callee_state_fn set_callee_state_cb)
5569 {
5570 struct bpf_verifier_state *state = env->cur_state;
5571 struct bpf_func_info_aux *func_info_aux;
5572 struct bpf_func_state *caller, *callee;
5573 int err;
5574 bool is_global = false;
5575
5576 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5577 verbose(env, "the call stack of %d frames is too deep\n",
5578 state->curframe + 2);
5579 return -E2BIG;
5580 }
5581
5582 caller = state->frame[state->curframe];
5583 if (state->frame[state->curframe + 1]) {
5584 verbose(env, "verifier bug. Frame %d already allocated\n",
5585 state->curframe + 1);
5586 return -EFAULT;
5587 }
5588
5589 func_info_aux = env->prog->aux->func_info_aux;
5590 if (func_info_aux)
5591 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5592 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5593 if (err == -EFAULT)
5594 return err;
5595 if (is_global) {
5596 if (err) {
5597 verbose(env, "Caller passes invalid args into func#%d\n",
5598 subprog);
5599 return err;
5600 } else {
5601 if (env->log.level & BPF_LOG_LEVEL)
5602 verbose(env,
5603 "Func#%d is global and valid. Skipping.\n",
5604 subprog);
5605 clear_caller_saved_regs(env, caller->regs);
5606
5607 /* All global functions return a 64-bit SCALAR_VALUE */
5608 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5609 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5610
5611 /* continue with next insn after call */
5612 return 0;
5613 }
5614 }
5615
5616 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5617 if (!callee)
5618 return -ENOMEM;
5619 state->frame[state->curframe + 1] = callee;
5620
5621 /* callee cannot access r0, r6 - r9 for reading and has to write
5622 * into its own stack before reading from it.
5623 * callee can read/write into caller's stack
5624 */
5625 init_func_state(env, callee,
5626 /* remember the callsite, it will be used by bpf_exit */
5627 *insn_idx /* callsite */,
5628 state->curframe + 1 /* frameno within this callchain */,
5629 subprog /* subprog number within this prog */);
5630
5631 /* Transfer references to the callee */
5632 err = transfer_reference_state(callee, caller);
5633 if (err)
5634 return err;
5635
5636 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5637 if (err)
5638 return err;
5639
5640 clear_caller_saved_regs(env, caller->regs);
5641
5642 /* only increment it after check_reg_arg() finished */
5643 state->curframe++;
5644
5645 /* and go analyze first insn of the callee */
5646 *insn_idx = env->subprog_info[subprog].start - 1;
5647
5648 if (env->log.level & BPF_LOG_LEVEL) {
5649 verbose(env, "caller:\n");
5650 print_verifier_state(env, caller);
5651 verbose(env, "callee:\n");
5652 print_verifier_state(env, callee);
5653 }
5654 return 0;
5655 }
5656
map_set_for_each_callback_args(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee)5657 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
5658 struct bpf_func_state *caller,
5659 struct bpf_func_state *callee)
5660 {
5661 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
5662 * void *callback_ctx, u64 flags);
5663 * callback_fn(struct bpf_map *map, void *key, void *value,
5664 * void *callback_ctx);
5665 */
5666 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
5667
5668 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5669 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5670 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5671
5672 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5673 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5674 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5675
5676 /* pointer to stack or null */
5677 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
5678
5679 /* unused */
5680 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5681 return 0;
5682 }
5683
set_callee_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)5684 static int set_callee_state(struct bpf_verifier_env *env,
5685 struct bpf_func_state *caller,
5686 struct bpf_func_state *callee, int insn_idx)
5687 {
5688 int i;
5689
5690 /* copy r1 - r5 args that callee can access. The copy includes parent
5691 * pointers, which connects us up to the liveness chain
5692 */
5693 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5694 callee->regs[i] = caller->regs[i];
5695 return 0;
5696 }
5697
check_func_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)5698 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5699 int *insn_idx)
5700 {
5701 int subprog, target_insn;
5702
5703 target_insn = *insn_idx + insn->imm + 1;
5704 subprog = find_subprog(env, target_insn);
5705 if (subprog < 0) {
5706 verbose(env, "verifier bug. No program starts at insn %d\n",
5707 target_insn);
5708 return -EFAULT;
5709 }
5710
5711 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
5712 }
5713
set_map_elem_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)5714 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
5715 struct bpf_func_state *caller,
5716 struct bpf_func_state *callee,
5717 int insn_idx)
5718 {
5719 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
5720 struct bpf_map *map;
5721 int err;
5722
5723 if (bpf_map_ptr_poisoned(insn_aux)) {
5724 verbose(env, "tail_call abusing map_ptr\n");
5725 return -EINVAL;
5726 }
5727
5728 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
5729 if (!map->ops->map_set_for_each_callback_args ||
5730 !map->ops->map_for_each_callback) {
5731 verbose(env, "callback function not allowed for map\n");
5732 return -ENOTSUPP;
5733 }
5734
5735 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
5736 if (err)
5737 return err;
5738
5739 callee->in_callback_fn = true;
5740 return 0;
5741 }
5742
prepare_func_exit(struct bpf_verifier_env * env,int * insn_idx)5743 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5744 {
5745 struct bpf_verifier_state *state = env->cur_state;
5746 struct bpf_func_state *caller, *callee;
5747 struct bpf_reg_state *r0;
5748 int err;
5749
5750 callee = state->frame[state->curframe];
5751 r0 = &callee->regs[BPF_REG_0];
5752 if (r0->type == PTR_TO_STACK) {
5753 /* technically it's ok to return caller's stack pointer
5754 * (or caller's caller's pointer) back to the caller,
5755 * since these pointers are valid. Only current stack
5756 * pointer will be invalid as soon as function exits,
5757 * but let's be conservative
5758 */
5759 verbose(env, "cannot return stack pointer to the caller\n");
5760 return -EINVAL;
5761 }
5762
5763 state->curframe--;
5764 caller = state->frame[state->curframe];
5765 if (callee->in_callback_fn) {
5766 /* enforce R0 return value range [0, 1]. */
5767 struct tnum range = tnum_range(0, 1);
5768
5769 if (r0->type != SCALAR_VALUE) {
5770 verbose(env, "R0 not a scalar value\n");
5771 return -EACCES;
5772 }
5773 if (!tnum_in(range, r0->var_off)) {
5774 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
5775 return -EINVAL;
5776 }
5777 } else {
5778 /* return to the caller whatever r0 had in the callee */
5779 caller->regs[BPF_REG_0] = *r0;
5780 }
5781
5782 /* Transfer references to the caller */
5783 err = transfer_reference_state(caller, callee);
5784 if (err)
5785 return err;
5786
5787 *insn_idx = callee->callsite + 1;
5788 if (env->log.level & BPF_LOG_LEVEL) {
5789 verbose(env, "returning from callee:\n");
5790 print_verifier_state(env, callee);
5791 verbose(env, "to caller at %d:\n", *insn_idx);
5792 print_verifier_state(env, caller);
5793 }
5794 /* clear everything in the callee */
5795 free_func_state(callee);
5796 state->frame[state->curframe + 1] = NULL;
5797 return 0;
5798 }
5799
do_refine_retval_range(struct bpf_reg_state * regs,int ret_type,int func_id,struct bpf_call_arg_meta * meta)5800 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
5801 int func_id,
5802 struct bpf_call_arg_meta *meta)
5803 {
5804 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
5805
5806 if (ret_type != RET_INTEGER ||
5807 (func_id != BPF_FUNC_get_stack &&
5808 func_id != BPF_FUNC_get_task_stack &&
5809 func_id != BPF_FUNC_probe_read_str &&
5810 func_id != BPF_FUNC_probe_read_kernel_str &&
5811 func_id != BPF_FUNC_probe_read_user_str))
5812 return;
5813
5814 ret_reg->smax_value = meta->msize_max_value;
5815 ret_reg->s32_max_value = meta->msize_max_value;
5816 ret_reg->smin_value = -MAX_ERRNO;
5817 ret_reg->s32_min_value = -MAX_ERRNO;
5818 __reg_deduce_bounds(ret_reg);
5819 __reg_bound_offset(ret_reg);
5820 __update_reg_bounds(ret_reg);
5821 }
5822
5823 static int
record_func_map(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)5824 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5825 int func_id, int insn_idx)
5826 {
5827 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5828 struct bpf_map *map = meta->map_ptr;
5829
5830 if (func_id != BPF_FUNC_tail_call &&
5831 func_id != BPF_FUNC_map_lookup_elem &&
5832 func_id != BPF_FUNC_map_update_elem &&
5833 func_id != BPF_FUNC_map_delete_elem &&
5834 func_id != BPF_FUNC_map_push_elem &&
5835 func_id != BPF_FUNC_map_pop_elem &&
5836 func_id != BPF_FUNC_map_peek_elem &&
5837 func_id != BPF_FUNC_for_each_map_elem &&
5838 func_id != BPF_FUNC_redirect_map)
5839 return 0;
5840
5841 if (map == NULL) {
5842 verbose(env, "kernel subsystem misconfigured verifier\n");
5843 return -EINVAL;
5844 }
5845
5846 /* In case of read-only, some additional restrictions
5847 * need to be applied in order to prevent altering the
5848 * state of the map from program side.
5849 */
5850 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
5851 (func_id == BPF_FUNC_map_delete_elem ||
5852 func_id == BPF_FUNC_map_update_elem ||
5853 func_id == BPF_FUNC_map_push_elem ||
5854 func_id == BPF_FUNC_map_pop_elem)) {
5855 verbose(env, "write into map forbidden\n");
5856 return -EACCES;
5857 }
5858
5859 if (!BPF_MAP_PTR(aux->map_ptr_state))
5860 bpf_map_ptr_store(aux, meta->map_ptr,
5861 !meta->map_ptr->bypass_spec_v1);
5862 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
5863 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
5864 !meta->map_ptr->bypass_spec_v1);
5865 return 0;
5866 }
5867
5868 static int
record_func_key(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)5869 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5870 int func_id, int insn_idx)
5871 {
5872 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5873 struct bpf_reg_state *regs = cur_regs(env), *reg;
5874 struct bpf_map *map = meta->map_ptr;
5875 struct tnum range;
5876 u64 val;
5877 int err;
5878
5879 if (func_id != BPF_FUNC_tail_call)
5880 return 0;
5881 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
5882 verbose(env, "kernel subsystem misconfigured verifier\n");
5883 return -EINVAL;
5884 }
5885
5886 range = tnum_range(0, map->max_entries - 1);
5887 reg = ®s[BPF_REG_3];
5888
5889 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
5890 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5891 return 0;
5892 }
5893
5894 err = mark_chain_precision(env, BPF_REG_3);
5895 if (err)
5896 return err;
5897
5898 val = reg->var_off.value;
5899 if (bpf_map_key_unseen(aux))
5900 bpf_map_key_store(aux, val);
5901 else if (!bpf_map_key_poisoned(aux) &&
5902 bpf_map_key_immediate(aux) != val)
5903 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5904 return 0;
5905 }
5906
check_reference_leak(struct bpf_verifier_env * env)5907 static int check_reference_leak(struct bpf_verifier_env *env)
5908 {
5909 struct bpf_func_state *state = cur_func(env);
5910 int i;
5911
5912 for (i = 0; i < state->acquired_refs; i++) {
5913 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
5914 state->refs[i].id, state->refs[i].insn_idx);
5915 }
5916 return state->acquired_refs ? -EINVAL : 0;
5917 }
5918
check_bpf_snprintf_call(struct bpf_verifier_env * env,struct bpf_reg_state * regs)5919 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
5920 struct bpf_reg_state *regs)
5921 {
5922 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
5923 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
5924 struct bpf_map *fmt_map = fmt_reg->map_ptr;
5925 int err, fmt_map_off, num_args;
5926 u64 fmt_addr;
5927 char *fmt;
5928
5929 /* data must be an array of u64 */
5930 if (data_len_reg->var_off.value % 8)
5931 return -EINVAL;
5932 num_args = data_len_reg->var_off.value / 8;
5933
5934 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
5935 * and map_direct_value_addr is set.
5936 */
5937 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
5938 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
5939 fmt_map_off);
5940 if (err) {
5941 verbose(env, "verifier bug\n");
5942 return -EFAULT;
5943 }
5944 fmt = (char *)(long)fmt_addr + fmt_map_off;
5945
5946 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
5947 * can focus on validating the format specifiers.
5948 */
5949 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
5950 if (err < 0)
5951 verbose(env, "Invalid format string\n");
5952
5953 return err;
5954 }
5955
check_helper_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx_p)5956 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5957 int *insn_idx_p)
5958 {
5959 const struct bpf_func_proto *fn = NULL;
5960 struct bpf_reg_state *regs;
5961 struct bpf_call_arg_meta meta;
5962 int insn_idx = *insn_idx_p;
5963 bool changes_data;
5964 int i, err, func_id;
5965
5966 /* find function prototype */
5967 func_id = insn->imm;
5968 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
5969 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
5970 func_id);
5971 return -EINVAL;
5972 }
5973
5974 if (env->ops->get_func_proto)
5975 fn = env->ops->get_func_proto(func_id, env->prog);
5976 if (!fn) {
5977 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
5978 func_id);
5979 return -EINVAL;
5980 }
5981
5982 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5983 if (!env->prog->gpl_compatible && fn->gpl_only) {
5984 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
5985 return -EINVAL;
5986 }
5987
5988 if (fn->allowed && !fn->allowed(env->prog)) {
5989 verbose(env, "helper call is not allowed in probe\n");
5990 return -EINVAL;
5991 }
5992
5993 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5994 changes_data = bpf_helper_changes_pkt_data(fn->func);
5995 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
5996 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5997 func_id_name(func_id), func_id);
5998 return -EINVAL;
5999 }
6000
6001 memset(&meta, 0, sizeof(meta));
6002 meta.pkt_access = fn->pkt_access;
6003
6004 err = check_func_proto(fn, func_id);
6005 if (err) {
6006 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6007 func_id_name(func_id), func_id);
6008 return err;
6009 }
6010
6011 meta.func_id = func_id;
6012 /* check args */
6013 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6014 err = check_func_arg(env, i, &meta, fn);
6015 if (err)
6016 return err;
6017 }
6018
6019 err = record_func_map(env, &meta, func_id, insn_idx);
6020 if (err)
6021 return err;
6022
6023 err = record_func_key(env, &meta, func_id, insn_idx);
6024 if (err)
6025 return err;
6026
6027 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6028 * is inferred from register state.
6029 */
6030 for (i = 0; i < meta.access_size; i++) {
6031 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6032 BPF_WRITE, -1, false);
6033 if (err)
6034 return err;
6035 }
6036
6037 if (func_id == BPF_FUNC_tail_call) {
6038 err = check_reference_leak(env);
6039 if (err) {
6040 verbose(env, "tail_call would lead to reference leak\n");
6041 return err;
6042 }
6043 } else if (is_release_function(func_id)) {
6044 err = release_reference(env, meta.ref_obj_id);
6045 if (err) {
6046 verbose(env, "func %s#%d reference has not been acquired before\n",
6047 func_id_name(func_id), func_id);
6048 return err;
6049 }
6050 }
6051
6052 regs = cur_regs(env);
6053
6054 /* check that flags argument in get_local_storage(map, flags) is 0,
6055 * this is required because get_local_storage() can't return an error.
6056 */
6057 if (func_id == BPF_FUNC_get_local_storage &&
6058 !register_is_null(®s[BPF_REG_2])) {
6059 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6060 return -EINVAL;
6061 }
6062
6063 if (func_id == BPF_FUNC_for_each_map_elem) {
6064 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6065 set_map_elem_callback_state);
6066 if (err < 0)
6067 return -EINVAL;
6068 }
6069
6070 if (func_id == BPF_FUNC_snprintf) {
6071 err = check_bpf_snprintf_call(env, regs);
6072 if (err < 0)
6073 return err;
6074 }
6075
6076 /* reset caller saved regs */
6077 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6078 mark_reg_not_init(env, regs, caller_saved[i]);
6079 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6080 }
6081
6082 /* helper call returns 64-bit value. */
6083 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6084
6085 /* update return register (already marked as written above) */
6086 if (fn->ret_type == RET_INTEGER) {
6087 /* sets type to SCALAR_VALUE */
6088 mark_reg_unknown(env, regs, BPF_REG_0);
6089 } else if (fn->ret_type == RET_VOID) {
6090 regs[BPF_REG_0].type = NOT_INIT;
6091 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
6092 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6093 /* There is no offset yet applied, variable or fixed */
6094 mark_reg_known_zero(env, regs, BPF_REG_0);
6095 /* remember map_ptr, so that check_map_access()
6096 * can check 'value_size' boundary of memory access
6097 * to map element returned from bpf_map_lookup_elem()
6098 */
6099 if (meta.map_ptr == NULL) {
6100 verbose(env,
6101 "kernel subsystem misconfigured verifier\n");
6102 return -EINVAL;
6103 }
6104 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6105 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6106 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
6107 if (map_value_has_spin_lock(meta.map_ptr))
6108 regs[BPF_REG_0].id = ++env->id_gen;
6109 } else {
6110 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
6111 }
6112 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
6113 mark_reg_known_zero(env, regs, BPF_REG_0);
6114 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
6115 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
6116 mark_reg_known_zero(env, regs, BPF_REG_0);
6117 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
6118 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
6119 mark_reg_known_zero(env, regs, BPF_REG_0);
6120 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
6121 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
6122 mark_reg_known_zero(env, regs, BPF_REG_0);
6123 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
6124 regs[BPF_REG_0].mem_size = meta.mem_size;
6125 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
6126 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
6127 const struct btf_type *t;
6128
6129 mark_reg_known_zero(env, regs, BPF_REG_0);
6130 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6131 if (!btf_type_is_struct(t)) {
6132 u32 tsize;
6133 const struct btf_type *ret;
6134 const char *tname;
6135
6136 /* resolve the type size of ksym. */
6137 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6138 if (IS_ERR(ret)) {
6139 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6140 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6141 tname, PTR_ERR(ret));
6142 return -EINVAL;
6143 }
6144 regs[BPF_REG_0].type =
6145 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6146 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
6147 regs[BPF_REG_0].mem_size = tsize;
6148 } else {
6149 regs[BPF_REG_0].type =
6150 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6151 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
6152 regs[BPF_REG_0].btf = meta.ret_btf;
6153 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6154 }
6155 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
6156 fn->ret_type == RET_PTR_TO_BTF_ID) {
6157 int ret_btf_id;
6158
6159 mark_reg_known_zero(env, regs, BPF_REG_0);
6160 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
6161 PTR_TO_BTF_ID :
6162 PTR_TO_BTF_ID_OR_NULL;
6163 ret_btf_id = *fn->ret_btf_id;
6164 if (ret_btf_id == 0) {
6165 verbose(env, "invalid return type %d of func %s#%d\n",
6166 fn->ret_type, func_id_name(func_id), func_id);
6167 return -EINVAL;
6168 }
6169 /* current BPF helper definitions are only coming from
6170 * built-in code with type IDs from vmlinux BTF
6171 */
6172 regs[BPF_REG_0].btf = btf_vmlinux;
6173 regs[BPF_REG_0].btf_id = ret_btf_id;
6174 } else {
6175 verbose(env, "unknown return type %d of func %s#%d\n",
6176 fn->ret_type, func_id_name(func_id), func_id);
6177 return -EINVAL;
6178 }
6179
6180 if (reg_type_may_be_null(regs[BPF_REG_0].type))
6181 regs[BPF_REG_0].id = ++env->id_gen;
6182
6183 if (is_ptr_cast_function(func_id)) {
6184 /* For release_reference() */
6185 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6186 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6187 int id = acquire_reference_state(env, insn_idx);
6188
6189 if (id < 0)
6190 return id;
6191 /* For mark_ptr_or_null_reg() */
6192 regs[BPF_REG_0].id = id;
6193 /* For release_reference() */
6194 regs[BPF_REG_0].ref_obj_id = id;
6195 }
6196
6197 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6198
6199 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6200 if (err)
6201 return err;
6202
6203 if ((func_id == BPF_FUNC_get_stack ||
6204 func_id == BPF_FUNC_get_task_stack) &&
6205 !env->prog->has_callchain_buf) {
6206 const char *err_str;
6207
6208 #ifdef CONFIG_PERF_EVENTS
6209 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6210 err_str = "cannot get callchain buffer for func %s#%d\n";
6211 #else
6212 err = -ENOTSUPP;
6213 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6214 #endif
6215 if (err) {
6216 verbose(env, err_str, func_id_name(func_id), func_id);
6217 return err;
6218 }
6219
6220 env->prog->has_callchain_buf = true;
6221 }
6222
6223 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6224 env->prog->call_get_stack = true;
6225
6226 if (changes_data)
6227 clear_all_pkt_pointers(env);
6228 return 0;
6229 }
6230
6231 /* mark_btf_func_reg_size() is used when the reg size is determined by
6232 * the BTF func_proto's return value size and argument.
6233 */
mark_btf_func_reg_size(struct bpf_verifier_env * env,u32 regno,size_t reg_size)6234 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6235 size_t reg_size)
6236 {
6237 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6238
6239 if (regno == BPF_REG_0) {
6240 /* Function return value */
6241 reg->live |= REG_LIVE_WRITTEN;
6242 reg->subreg_def = reg_size == sizeof(u64) ?
6243 DEF_NOT_SUBREG : env->insn_idx + 1;
6244 } else {
6245 /* Function argument */
6246 if (reg_size == sizeof(u64)) {
6247 mark_insn_zext(env, reg);
6248 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6249 } else {
6250 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6251 }
6252 }
6253 }
6254
check_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn)6255 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6256 {
6257 const struct btf_type *t, *func, *func_proto, *ptr_type;
6258 struct bpf_reg_state *regs = cur_regs(env);
6259 const char *func_name, *ptr_type_name;
6260 u32 i, nargs, func_id, ptr_type_id;
6261 const struct btf_param *args;
6262 int err;
6263
6264 func_id = insn->imm;
6265 func = btf_type_by_id(btf_vmlinux, func_id);
6266 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
6267 func_proto = btf_type_by_id(btf_vmlinux, func->type);
6268
6269 if (!env->ops->check_kfunc_call ||
6270 !env->ops->check_kfunc_call(func_id)) {
6271 verbose(env, "calling kernel function %s is not allowed\n",
6272 func_name);
6273 return -EACCES;
6274 }
6275
6276 /* Check the arguments */
6277 err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
6278 if (err)
6279 return err;
6280
6281 for (i = 0; i < CALLER_SAVED_REGS; i++)
6282 mark_reg_not_init(env, regs, caller_saved[i]);
6283
6284 /* Check return type */
6285 t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
6286 if (btf_type_is_scalar(t)) {
6287 mark_reg_unknown(env, regs, BPF_REG_0);
6288 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6289 } else if (btf_type_is_ptr(t)) {
6290 ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
6291 &ptr_type_id);
6292 if (!btf_type_is_struct(ptr_type)) {
6293 ptr_type_name = btf_name_by_offset(btf_vmlinux,
6294 ptr_type->name_off);
6295 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6296 func_name, btf_type_str(ptr_type),
6297 ptr_type_name);
6298 return -EINVAL;
6299 }
6300 mark_reg_known_zero(env, regs, BPF_REG_0);
6301 regs[BPF_REG_0].btf = btf_vmlinux;
6302 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6303 regs[BPF_REG_0].btf_id = ptr_type_id;
6304 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6305 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6306
6307 nargs = btf_type_vlen(func_proto);
6308 args = (const struct btf_param *)(func_proto + 1);
6309 for (i = 0; i < nargs; i++) {
6310 u32 regno = i + 1;
6311
6312 t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
6313 if (btf_type_is_ptr(t))
6314 mark_btf_func_reg_size(env, regno, sizeof(void *));
6315 else
6316 /* scalar. ensured by btf_check_kfunc_arg_match() */
6317 mark_btf_func_reg_size(env, regno, t->size);
6318 }
6319
6320 return 0;
6321 }
6322
signed_add_overflows(s64 a,s64 b)6323 static bool signed_add_overflows(s64 a, s64 b)
6324 {
6325 /* Do the add in u64, where overflow is well-defined */
6326 s64 res = (s64)((u64)a + (u64)b);
6327
6328 if (b < 0)
6329 return res > a;
6330 return res < a;
6331 }
6332
signed_add32_overflows(s32 a,s32 b)6333 static bool signed_add32_overflows(s32 a, s32 b)
6334 {
6335 /* Do the add in u32, where overflow is well-defined */
6336 s32 res = (s32)((u32)a + (u32)b);
6337
6338 if (b < 0)
6339 return res > a;
6340 return res < a;
6341 }
6342
signed_sub_overflows(s64 a,s64 b)6343 static bool signed_sub_overflows(s64 a, s64 b)
6344 {
6345 /* Do the sub in u64, where overflow is well-defined */
6346 s64 res = (s64)((u64)a - (u64)b);
6347
6348 if (b < 0)
6349 return res < a;
6350 return res > a;
6351 }
6352
signed_sub32_overflows(s32 a,s32 b)6353 static bool signed_sub32_overflows(s32 a, s32 b)
6354 {
6355 /* Do the sub in u32, where overflow is well-defined */
6356 s32 res = (s32)((u32)a - (u32)b);
6357
6358 if (b < 0)
6359 return res < a;
6360 return res > a;
6361 }
6362
check_reg_sane_offset(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,enum bpf_reg_type type)6363 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6364 const struct bpf_reg_state *reg,
6365 enum bpf_reg_type type)
6366 {
6367 bool known = tnum_is_const(reg->var_off);
6368 s64 val = reg->var_off.value;
6369 s64 smin = reg->smin_value;
6370
6371 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6372 verbose(env, "math between %s pointer and %lld is not allowed\n",
6373 reg_type_str[type], val);
6374 return false;
6375 }
6376
6377 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6378 verbose(env, "%s pointer offset %d is not allowed\n",
6379 reg_type_str[type], reg->off);
6380 return false;
6381 }
6382
6383 if (smin == S64_MIN) {
6384 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6385 reg_type_str[type]);
6386 return false;
6387 }
6388
6389 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6390 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6391 smin, reg_type_str[type]);
6392 return false;
6393 }
6394
6395 return true;
6396 }
6397
cur_aux(struct bpf_verifier_env * env)6398 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6399 {
6400 return &env->insn_aux_data[env->insn_idx];
6401 }
6402
6403 enum {
6404 REASON_BOUNDS = -1,
6405 REASON_TYPE = -2,
6406 REASON_PATHS = -3,
6407 REASON_LIMIT = -4,
6408 REASON_STACK = -5,
6409 };
6410
retrieve_ptr_limit(const struct bpf_reg_state * ptr_reg,const struct bpf_reg_state * off_reg,u32 * alu_limit,u8 opcode)6411 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6412 const struct bpf_reg_state *off_reg,
6413 u32 *alu_limit, u8 opcode)
6414 {
6415 bool off_is_neg = off_reg->smin_value < 0;
6416 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6417 (opcode == BPF_SUB && !off_is_neg);
6418 u32 max = 0, ptr_limit = 0;
6419
6420 if (!tnum_is_const(off_reg->var_off) &&
6421 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6422 return REASON_BOUNDS;
6423
6424 switch (ptr_reg->type) {
6425 case PTR_TO_STACK:
6426 /* Offset 0 is out-of-bounds, but acceptable start for the
6427 * left direction, see BPF_REG_FP. Also, unknown scalar
6428 * offset where we would need to deal with min/max bounds is
6429 * currently prohibited for unprivileged.
6430 */
6431 max = MAX_BPF_STACK + mask_to_left;
6432 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6433 break;
6434 case PTR_TO_MAP_VALUE:
6435 max = ptr_reg->map_ptr->value_size;
6436 ptr_limit = (mask_to_left ?
6437 ptr_reg->smin_value :
6438 ptr_reg->umax_value) + ptr_reg->off;
6439 break;
6440 default:
6441 return REASON_TYPE;
6442 }
6443
6444 if (ptr_limit >= max)
6445 return REASON_LIMIT;
6446 *alu_limit = ptr_limit;
6447 return 0;
6448 }
6449
can_skip_alu_sanitation(const struct bpf_verifier_env * env,const struct bpf_insn * insn)6450 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6451 const struct bpf_insn *insn)
6452 {
6453 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6454 }
6455
update_alu_sanitation_state(struct bpf_insn_aux_data * aux,u32 alu_state,u32 alu_limit)6456 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6457 u32 alu_state, u32 alu_limit)
6458 {
6459 /* If we arrived here from different branches with different
6460 * state or limits to sanitize, then this won't work.
6461 */
6462 if (aux->alu_state &&
6463 (aux->alu_state != alu_state ||
6464 aux->alu_limit != alu_limit))
6465 return REASON_PATHS;
6466
6467 /* Corresponding fixup done in do_misc_fixups(). */
6468 aux->alu_state = alu_state;
6469 aux->alu_limit = alu_limit;
6470 return 0;
6471 }
6472
sanitize_val_alu(struct bpf_verifier_env * env,struct bpf_insn * insn)6473 static int sanitize_val_alu(struct bpf_verifier_env *env,
6474 struct bpf_insn *insn)
6475 {
6476 struct bpf_insn_aux_data *aux = cur_aux(env);
6477
6478 if (can_skip_alu_sanitation(env, insn))
6479 return 0;
6480
6481 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6482 }
6483
sanitize_needed(u8 opcode)6484 static bool sanitize_needed(u8 opcode)
6485 {
6486 return opcode == BPF_ADD || opcode == BPF_SUB;
6487 }
6488
sanitize_ptr_alu(struct bpf_verifier_env * env,struct bpf_insn * insn,const struct bpf_reg_state * ptr_reg,const struct bpf_reg_state * off_reg,struct bpf_reg_state * dst_reg,struct bpf_insn_aux_data * tmp_aux,const bool commit_window)6489 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6490 struct bpf_insn *insn,
6491 const struct bpf_reg_state *ptr_reg,
6492 const struct bpf_reg_state *off_reg,
6493 struct bpf_reg_state *dst_reg,
6494 struct bpf_insn_aux_data *tmp_aux,
6495 const bool commit_window)
6496 {
6497 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : tmp_aux;
6498 struct bpf_verifier_state *vstate = env->cur_state;
6499 bool off_is_imm = tnum_is_const(off_reg->var_off);
6500 bool off_is_neg = off_reg->smin_value < 0;
6501 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6502 u8 opcode = BPF_OP(insn->code);
6503 u32 alu_state, alu_limit;
6504 struct bpf_reg_state tmp;
6505 bool ret;
6506 int err;
6507
6508 if (can_skip_alu_sanitation(env, insn))
6509 return 0;
6510
6511 /* We already marked aux for masking from non-speculative
6512 * paths, thus we got here in the first place. We only care
6513 * to explore bad access from here.
6514 */
6515 if (vstate->speculative)
6516 goto do_sim;
6517
6518 err = retrieve_ptr_limit(ptr_reg, off_reg, &alu_limit, opcode);
6519 if (err < 0)
6520 return err;
6521
6522 if (commit_window) {
6523 /* In commit phase we narrow the masking window based on
6524 * the observed pointer move after the simulated operation.
6525 */
6526 alu_state = tmp_aux->alu_state;
6527 alu_limit = abs(tmp_aux->alu_limit - alu_limit);
6528 } else {
6529 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6530 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6531 alu_state |= ptr_is_dst_reg ?
6532 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
6533 }
6534
6535 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
6536 if (err < 0)
6537 return err;
6538 do_sim:
6539 /* If we're in commit phase, we're done here given we already
6540 * pushed the truncated dst_reg into the speculative verification
6541 * stack.
6542 */
6543 if (commit_window)
6544 return 0;
6545
6546 /* Simulate and find potential out-of-bounds access under
6547 * speculative execution from truncation as a result of
6548 * masking when off was not within expected range. If off
6549 * sits in dst, then we temporarily need to move ptr there
6550 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6551 * for cases where we use K-based arithmetic in one direction
6552 * and truncated reg-based in the other in order to explore
6553 * bad access.
6554 */
6555 if (!ptr_is_dst_reg) {
6556 tmp = *dst_reg;
6557 *dst_reg = *ptr_reg;
6558 }
6559 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
6560 if (!ptr_is_dst_reg && ret)
6561 *dst_reg = tmp;
6562 return !ret ? REASON_STACK : 0;
6563 }
6564
sanitize_err(struct bpf_verifier_env * env,const struct bpf_insn * insn,int reason,const struct bpf_reg_state * off_reg,const struct bpf_reg_state * dst_reg)6565 static int sanitize_err(struct bpf_verifier_env *env,
6566 const struct bpf_insn *insn, int reason,
6567 const struct bpf_reg_state *off_reg,
6568 const struct bpf_reg_state *dst_reg)
6569 {
6570 static const char *err = "pointer arithmetic with it prohibited for !root";
6571 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6572 u32 dst = insn->dst_reg, src = insn->src_reg;
6573
6574 switch (reason) {
6575 case REASON_BOUNDS:
6576 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6577 off_reg == dst_reg ? dst : src, err);
6578 break;
6579 case REASON_TYPE:
6580 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6581 off_reg == dst_reg ? src : dst, err);
6582 break;
6583 case REASON_PATHS:
6584 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6585 dst, op, err);
6586 break;
6587 case REASON_LIMIT:
6588 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6589 dst, op, err);
6590 break;
6591 case REASON_STACK:
6592 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6593 dst, err);
6594 break;
6595 default:
6596 verbose(env, "verifier internal error: unknown reason (%d)\n",
6597 reason);
6598 break;
6599 }
6600
6601 return -EACCES;
6602 }
6603
6604 /* check that stack access falls within stack limits and that 'reg' doesn't
6605 * have a variable offset.
6606 *
6607 * Variable offset is prohibited for unprivileged mode for simplicity since it
6608 * requires corresponding support in Spectre masking for stack ALU. See also
6609 * retrieve_ptr_limit().
6610 *
6611 *
6612 * 'off' includes 'reg->off'.
6613 */
check_stack_access_for_ptr_arithmetic(struct bpf_verifier_env * env,int regno,const struct bpf_reg_state * reg,int off)6614 static int check_stack_access_for_ptr_arithmetic(
6615 struct bpf_verifier_env *env,
6616 int regno,
6617 const struct bpf_reg_state *reg,
6618 int off)
6619 {
6620 if (!tnum_is_const(reg->var_off)) {
6621 char tn_buf[48];
6622
6623 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6624 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6625 regno, tn_buf, off);
6626 return -EACCES;
6627 }
6628
6629 if (off >= 0 || off < -MAX_BPF_STACK) {
6630 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6631 "prohibited for !root; off=%d\n", regno, off);
6632 return -EACCES;
6633 }
6634
6635 return 0;
6636 }
6637
sanitize_check_bounds(struct bpf_verifier_env * env,const struct bpf_insn * insn,const struct bpf_reg_state * dst_reg)6638 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6639 const struct bpf_insn *insn,
6640 const struct bpf_reg_state *dst_reg)
6641 {
6642 u32 dst = insn->dst_reg;
6643
6644 /* For unprivileged we require that resulting offset must be in bounds
6645 * in order to be able to sanitize access later on.
6646 */
6647 if (env->bypass_spec_v1)
6648 return 0;
6649
6650 switch (dst_reg->type) {
6651 case PTR_TO_STACK:
6652 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6653 dst_reg->off + dst_reg->var_off.value))
6654 return -EACCES;
6655 break;
6656 case PTR_TO_MAP_VALUE:
6657 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6658 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6659 "prohibited for !root\n", dst);
6660 return -EACCES;
6661 }
6662 break;
6663 default:
6664 break;
6665 }
6666
6667 return 0;
6668 }
6669
6670 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6671 * Caller should also handle BPF_MOV case separately.
6672 * If we return -EACCES, caller may want to try again treating pointer as a
6673 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
6674 */
adjust_ptr_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn,const struct bpf_reg_state * ptr_reg,const struct bpf_reg_state * off_reg)6675 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
6676 struct bpf_insn *insn,
6677 const struct bpf_reg_state *ptr_reg,
6678 const struct bpf_reg_state *off_reg)
6679 {
6680 struct bpf_verifier_state *vstate = env->cur_state;
6681 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6682 struct bpf_reg_state *regs = state->regs, *dst_reg;
6683 bool known = tnum_is_const(off_reg->var_off);
6684 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
6685 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
6686 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
6687 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
6688 struct bpf_insn_aux_data tmp_aux = {};
6689 u8 opcode = BPF_OP(insn->code);
6690 u32 dst = insn->dst_reg;
6691 int ret;
6692
6693 dst_reg = ®s[dst];
6694
6695 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
6696 smin_val > smax_val || umin_val > umax_val) {
6697 /* Taint dst register if offset had invalid bounds derived from
6698 * e.g. dead branches.
6699 */
6700 __mark_reg_unknown(env, dst_reg);
6701 return 0;
6702 }
6703
6704 if (BPF_CLASS(insn->code) != BPF_ALU64) {
6705 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
6706 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6707 __mark_reg_unknown(env, dst_reg);
6708 return 0;
6709 }
6710
6711 verbose(env,
6712 "R%d 32-bit pointer arithmetic prohibited\n",
6713 dst);
6714 return -EACCES;
6715 }
6716
6717 switch (ptr_reg->type) {
6718 case PTR_TO_MAP_VALUE_OR_NULL:
6719 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
6720 dst, reg_type_str[ptr_reg->type]);
6721 return -EACCES;
6722 case CONST_PTR_TO_MAP:
6723 /* smin_val represents the known value */
6724 if (known && smin_val == 0 && opcode == BPF_ADD)
6725 break;
6726 fallthrough;
6727 case PTR_TO_PACKET_END:
6728 case PTR_TO_SOCKET:
6729 case PTR_TO_SOCKET_OR_NULL:
6730 case PTR_TO_SOCK_COMMON:
6731 case PTR_TO_SOCK_COMMON_OR_NULL:
6732 case PTR_TO_TCP_SOCK:
6733 case PTR_TO_TCP_SOCK_OR_NULL:
6734 case PTR_TO_XDP_SOCK:
6735 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
6736 dst, reg_type_str[ptr_reg->type]);
6737 return -EACCES;
6738 default:
6739 break;
6740 }
6741
6742 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
6743 * The id may be overwritten later if we create a new variable offset.
6744 */
6745 dst_reg->type = ptr_reg->type;
6746 dst_reg->id = ptr_reg->id;
6747
6748 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
6749 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
6750 return -EINVAL;
6751
6752 /* pointer types do not carry 32-bit bounds at the moment. */
6753 __mark_reg32_unbounded(dst_reg);
6754
6755 if (sanitize_needed(opcode)) {
6756 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
6757 &tmp_aux, false);
6758 if (ret < 0)
6759 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6760 }
6761
6762 switch (opcode) {
6763 case BPF_ADD:
6764 /* We can take a fixed offset as long as it doesn't overflow
6765 * the s32 'off' field
6766 */
6767 if (known && (ptr_reg->off + smin_val ==
6768 (s64)(s32)(ptr_reg->off + smin_val))) {
6769 /* pointer += K. Accumulate it into fixed offset */
6770 dst_reg->smin_value = smin_ptr;
6771 dst_reg->smax_value = smax_ptr;
6772 dst_reg->umin_value = umin_ptr;
6773 dst_reg->umax_value = umax_ptr;
6774 dst_reg->var_off = ptr_reg->var_off;
6775 dst_reg->off = ptr_reg->off + smin_val;
6776 dst_reg->raw = ptr_reg->raw;
6777 break;
6778 }
6779 /* A new variable offset is created. Note that off_reg->off
6780 * == 0, since it's a scalar.
6781 * dst_reg gets the pointer type and since some positive
6782 * integer value was added to the pointer, give it a new 'id'
6783 * if it's a PTR_TO_PACKET.
6784 * this creates a new 'base' pointer, off_reg (variable) gets
6785 * added into the variable offset, and we copy the fixed offset
6786 * from ptr_reg.
6787 */
6788 if (signed_add_overflows(smin_ptr, smin_val) ||
6789 signed_add_overflows(smax_ptr, smax_val)) {
6790 dst_reg->smin_value = S64_MIN;
6791 dst_reg->smax_value = S64_MAX;
6792 } else {
6793 dst_reg->smin_value = smin_ptr + smin_val;
6794 dst_reg->smax_value = smax_ptr + smax_val;
6795 }
6796 if (umin_ptr + umin_val < umin_ptr ||
6797 umax_ptr + umax_val < umax_ptr) {
6798 dst_reg->umin_value = 0;
6799 dst_reg->umax_value = U64_MAX;
6800 } else {
6801 dst_reg->umin_value = umin_ptr + umin_val;
6802 dst_reg->umax_value = umax_ptr + umax_val;
6803 }
6804 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
6805 dst_reg->off = ptr_reg->off;
6806 dst_reg->raw = ptr_reg->raw;
6807 if (reg_is_pkt_pointer(ptr_reg)) {
6808 dst_reg->id = ++env->id_gen;
6809 /* something was added to pkt_ptr, set range to zero */
6810 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6811 }
6812 break;
6813 case BPF_SUB:
6814 if (dst_reg == off_reg) {
6815 /* scalar -= pointer. Creates an unknown scalar */
6816 verbose(env, "R%d tried to subtract pointer from scalar\n",
6817 dst);
6818 return -EACCES;
6819 }
6820 /* We don't allow subtraction from FP, because (according to
6821 * test_verifier.c test "invalid fp arithmetic", JITs might not
6822 * be able to deal with it.
6823 */
6824 if (ptr_reg->type == PTR_TO_STACK) {
6825 verbose(env, "R%d subtraction from stack pointer prohibited\n",
6826 dst);
6827 return -EACCES;
6828 }
6829 if (known && (ptr_reg->off - smin_val ==
6830 (s64)(s32)(ptr_reg->off - smin_val))) {
6831 /* pointer -= K. Subtract it from fixed offset */
6832 dst_reg->smin_value = smin_ptr;
6833 dst_reg->smax_value = smax_ptr;
6834 dst_reg->umin_value = umin_ptr;
6835 dst_reg->umax_value = umax_ptr;
6836 dst_reg->var_off = ptr_reg->var_off;
6837 dst_reg->id = ptr_reg->id;
6838 dst_reg->off = ptr_reg->off - smin_val;
6839 dst_reg->raw = ptr_reg->raw;
6840 break;
6841 }
6842 /* A new variable offset is created. If the subtrahend is known
6843 * nonnegative, then any reg->range we had before is still good.
6844 */
6845 if (signed_sub_overflows(smin_ptr, smax_val) ||
6846 signed_sub_overflows(smax_ptr, smin_val)) {
6847 /* Overflow possible, we know nothing */
6848 dst_reg->smin_value = S64_MIN;
6849 dst_reg->smax_value = S64_MAX;
6850 } else {
6851 dst_reg->smin_value = smin_ptr - smax_val;
6852 dst_reg->smax_value = smax_ptr - smin_val;
6853 }
6854 if (umin_ptr < umax_val) {
6855 /* Overflow possible, we know nothing */
6856 dst_reg->umin_value = 0;
6857 dst_reg->umax_value = U64_MAX;
6858 } else {
6859 /* Cannot overflow (as long as bounds are consistent) */
6860 dst_reg->umin_value = umin_ptr - umax_val;
6861 dst_reg->umax_value = umax_ptr - umin_val;
6862 }
6863 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
6864 dst_reg->off = ptr_reg->off;
6865 dst_reg->raw = ptr_reg->raw;
6866 if (reg_is_pkt_pointer(ptr_reg)) {
6867 dst_reg->id = ++env->id_gen;
6868 /* something was added to pkt_ptr, set range to zero */
6869 if (smin_val < 0)
6870 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6871 }
6872 break;
6873 case BPF_AND:
6874 case BPF_OR:
6875 case BPF_XOR:
6876 /* bitwise ops on pointers are troublesome, prohibit. */
6877 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
6878 dst, bpf_alu_string[opcode >> 4]);
6879 return -EACCES;
6880 default:
6881 /* other operators (e.g. MUL,LSH) produce non-pointer results */
6882 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
6883 dst, bpf_alu_string[opcode >> 4]);
6884 return -EACCES;
6885 }
6886
6887 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
6888 return -EINVAL;
6889
6890 __update_reg_bounds(dst_reg);
6891 __reg_deduce_bounds(dst_reg);
6892 __reg_bound_offset(dst_reg);
6893
6894 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
6895 return -EACCES;
6896 if (sanitize_needed(opcode)) {
6897 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
6898 &tmp_aux, true);
6899 if (ret < 0)
6900 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6901 }
6902
6903 return 0;
6904 }
6905
scalar32_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)6906 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
6907 struct bpf_reg_state *src_reg)
6908 {
6909 s32 smin_val = src_reg->s32_min_value;
6910 s32 smax_val = src_reg->s32_max_value;
6911 u32 umin_val = src_reg->u32_min_value;
6912 u32 umax_val = src_reg->u32_max_value;
6913
6914 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
6915 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
6916 dst_reg->s32_min_value = S32_MIN;
6917 dst_reg->s32_max_value = S32_MAX;
6918 } else {
6919 dst_reg->s32_min_value += smin_val;
6920 dst_reg->s32_max_value += smax_val;
6921 }
6922 if (dst_reg->u32_min_value + umin_val < umin_val ||
6923 dst_reg->u32_max_value + umax_val < umax_val) {
6924 dst_reg->u32_min_value = 0;
6925 dst_reg->u32_max_value = U32_MAX;
6926 } else {
6927 dst_reg->u32_min_value += umin_val;
6928 dst_reg->u32_max_value += umax_val;
6929 }
6930 }
6931
scalar_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)6932 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
6933 struct bpf_reg_state *src_reg)
6934 {
6935 s64 smin_val = src_reg->smin_value;
6936 s64 smax_val = src_reg->smax_value;
6937 u64 umin_val = src_reg->umin_value;
6938 u64 umax_val = src_reg->umax_value;
6939
6940 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
6941 signed_add_overflows(dst_reg->smax_value, smax_val)) {
6942 dst_reg->smin_value = S64_MIN;
6943 dst_reg->smax_value = S64_MAX;
6944 } else {
6945 dst_reg->smin_value += smin_val;
6946 dst_reg->smax_value += smax_val;
6947 }
6948 if (dst_reg->umin_value + umin_val < umin_val ||
6949 dst_reg->umax_value + umax_val < umax_val) {
6950 dst_reg->umin_value = 0;
6951 dst_reg->umax_value = U64_MAX;
6952 } else {
6953 dst_reg->umin_value += umin_val;
6954 dst_reg->umax_value += umax_val;
6955 }
6956 }
6957
scalar32_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)6958 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
6959 struct bpf_reg_state *src_reg)
6960 {
6961 s32 smin_val = src_reg->s32_min_value;
6962 s32 smax_val = src_reg->s32_max_value;
6963 u32 umin_val = src_reg->u32_min_value;
6964 u32 umax_val = src_reg->u32_max_value;
6965
6966 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
6967 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
6968 /* Overflow possible, we know nothing */
6969 dst_reg->s32_min_value = S32_MIN;
6970 dst_reg->s32_max_value = S32_MAX;
6971 } else {
6972 dst_reg->s32_min_value -= smax_val;
6973 dst_reg->s32_max_value -= smin_val;
6974 }
6975 if (dst_reg->u32_min_value < umax_val) {
6976 /* Overflow possible, we know nothing */
6977 dst_reg->u32_min_value = 0;
6978 dst_reg->u32_max_value = U32_MAX;
6979 } else {
6980 /* Cannot overflow (as long as bounds are consistent) */
6981 dst_reg->u32_min_value -= umax_val;
6982 dst_reg->u32_max_value -= umin_val;
6983 }
6984 }
6985
scalar_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)6986 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
6987 struct bpf_reg_state *src_reg)
6988 {
6989 s64 smin_val = src_reg->smin_value;
6990 s64 smax_val = src_reg->smax_value;
6991 u64 umin_val = src_reg->umin_value;
6992 u64 umax_val = src_reg->umax_value;
6993
6994 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
6995 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
6996 /* Overflow possible, we know nothing */
6997 dst_reg->smin_value = S64_MIN;
6998 dst_reg->smax_value = S64_MAX;
6999 } else {
7000 dst_reg->smin_value -= smax_val;
7001 dst_reg->smax_value -= smin_val;
7002 }
7003 if (dst_reg->umin_value < umax_val) {
7004 /* Overflow possible, we know nothing */
7005 dst_reg->umin_value = 0;
7006 dst_reg->umax_value = U64_MAX;
7007 } else {
7008 /* Cannot overflow (as long as bounds are consistent) */
7009 dst_reg->umin_value -= umax_val;
7010 dst_reg->umax_value -= umin_val;
7011 }
7012 }
7013
scalar32_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7014 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7015 struct bpf_reg_state *src_reg)
7016 {
7017 s32 smin_val = src_reg->s32_min_value;
7018 u32 umin_val = src_reg->u32_min_value;
7019 u32 umax_val = src_reg->u32_max_value;
7020
7021 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7022 /* Ain't nobody got time to multiply that sign */
7023 __mark_reg32_unbounded(dst_reg);
7024 return;
7025 }
7026 /* Both values are positive, so we can work with unsigned and
7027 * copy the result to signed (unless it exceeds S32_MAX).
7028 */
7029 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7030 /* Potential overflow, we know nothing */
7031 __mark_reg32_unbounded(dst_reg);
7032 return;
7033 }
7034 dst_reg->u32_min_value *= umin_val;
7035 dst_reg->u32_max_value *= umax_val;
7036 if (dst_reg->u32_max_value > S32_MAX) {
7037 /* Overflow possible, we know nothing */
7038 dst_reg->s32_min_value = S32_MIN;
7039 dst_reg->s32_max_value = S32_MAX;
7040 } else {
7041 dst_reg->s32_min_value = dst_reg->u32_min_value;
7042 dst_reg->s32_max_value = dst_reg->u32_max_value;
7043 }
7044 }
7045
scalar_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7046 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7047 struct bpf_reg_state *src_reg)
7048 {
7049 s64 smin_val = src_reg->smin_value;
7050 u64 umin_val = src_reg->umin_value;
7051 u64 umax_val = src_reg->umax_value;
7052
7053 if (smin_val < 0 || dst_reg->smin_value < 0) {
7054 /* Ain't nobody got time to multiply that sign */
7055 __mark_reg64_unbounded(dst_reg);
7056 return;
7057 }
7058 /* Both values are positive, so we can work with unsigned and
7059 * copy the result to signed (unless it exceeds S64_MAX).
7060 */
7061 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7062 /* Potential overflow, we know nothing */
7063 __mark_reg64_unbounded(dst_reg);
7064 return;
7065 }
7066 dst_reg->umin_value *= umin_val;
7067 dst_reg->umax_value *= umax_val;
7068 if (dst_reg->umax_value > S64_MAX) {
7069 /* Overflow possible, we know nothing */
7070 dst_reg->smin_value = S64_MIN;
7071 dst_reg->smax_value = S64_MAX;
7072 } else {
7073 dst_reg->smin_value = dst_reg->umin_value;
7074 dst_reg->smax_value = dst_reg->umax_value;
7075 }
7076 }
7077
scalar32_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7078 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7079 struct bpf_reg_state *src_reg)
7080 {
7081 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7082 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7083 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7084 s32 smin_val = src_reg->s32_min_value;
7085 u32 umax_val = src_reg->u32_max_value;
7086
7087 /* Assuming scalar64_min_max_and will be called so its safe
7088 * to skip updating register for known 32-bit case.
7089 */
7090 if (src_known && dst_known)
7091 return;
7092
7093 /* We get our minimum from the var_off, since that's inherently
7094 * bitwise. Our maximum is the minimum of the operands' maxima.
7095 */
7096 dst_reg->u32_min_value = var32_off.value;
7097 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7098 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7099 /* Lose signed bounds when ANDing negative numbers,
7100 * ain't nobody got time for that.
7101 */
7102 dst_reg->s32_min_value = S32_MIN;
7103 dst_reg->s32_max_value = S32_MAX;
7104 } else {
7105 /* ANDing two positives gives a positive, so safe to
7106 * cast result into s64.
7107 */
7108 dst_reg->s32_min_value = dst_reg->u32_min_value;
7109 dst_reg->s32_max_value = dst_reg->u32_max_value;
7110 }
7111
7112 }
7113
scalar_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7114 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7115 struct bpf_reg_state *src_reg)
7116 {
7117 bool src_known = tnum_is_const(src_reg->var_off);
7118 bool dst_known = tnum_is_const(dst_reg->var_off);
7119 s64 smin_val = src_reg->smin_value;
7120 u64 umax_val = src_reg->umax_value;
7121
7122 if (src_known && dst_known) {
7123 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7124 return;
7125 }
7126
7127 /* We get our minimum from the var_off, since that's inherently
7128 * bitwise. Our maximum is the minimum of the operands' maxima.
7129 */
7130 dst_reg->umin_value = dst_reg->var_off.value;
7131 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7132 if (dst_reg->smin_value < 0 || smin_val < 0) {
7133 /* Lose signed bounds when ANDing negative numbers,
7134 * ain't nobody got time for that.
7135 */
7136 dst_reg->smin_value = S64_MIN;
7137 dst_reg->smax_value = S64_MAX;
7138 } else {
7139 /* ANDing two positives gives a positive, so safe to
7140 * cast result into s64.
7141 */
7142 dst_reg->smin_value = dst_reg->umin_value;
7143 dst_reg->smax_value = dst_reg->umax_value;
7144 }
7145 /* We may learn something more from the var_off */
7146 __update_reg_bounds(dst_reg);
7147 }
7148
scalar32_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7149 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7150 struct bpf_reg_state *src_reg)
7151 {
7152 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7153 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7154 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7155 s32 smin_val = src_reg->s32_min_value;
7156 u32 umin_val = src_reg->u32_min_value;
7157
7158 /* Assuming scalar64_min_max_or will be called so it is safe
7159 * to skip updating register for known case.
7160 */
7161 if (src_known && dst_known)
7162 return;
7163
7164 /* We get our maximum from the var_off, and our minimum is the
7165 * maximum of the operands' minima
7166 */
7167 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7168 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7169 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7170 /* Lose signed bounds when ORing negative numbers,
7171 * ain't nobody got time for that.
7172 */
7173 dst_reg->s32_min_value = S32_MIN;
7174 dst_reg->s32_max_value = S32_MAX;
7175 } else {
7176 /* ORing two positives gives a positive, so safe to
7177 * cast result into s64.
7178 */
7179 dst_reg->s32_min_value = dst_reg->u32_min_value;
7180 dst_reg->s32_max_value = dst_reg->u32_max_value;
7181 }
7182 }
7183
scalar_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7184 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7185 struct bpf_reg_state *src_reg)
7186 {
7187 bool src_known = tnum_is_const(src_reg->var_off);
7188 bool dst_known = tnum_is_const(dst_reg->var_off);
7189 s64 smin_val = src_reg->smin_value;
7190 u64 umin_val = src_reg->umin_value;
7191
7192 if (src_known && dst_known) {
7193 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7194 return;
7195 }
7196
7197 /* We get our maximum from the var_off, and our minimum is the
7198 * maximum of the operands' minima
7199 */
7200 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7201 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7202 if (dst_reg->smin_value < 0 || smin_val < 0) {
7203 /* Lose signed bounds when ORing negative numbers,
7204 * ain't nobody got time for that.
7205 */
7206 dst_reg->smin_value = S64_MIN;
7207 dst_reg->smax_value = S64_MAX;
7208 } else {
7209 /* ORing two positives gives a positive, so safe to
7210 * cast result into s64.
7211 */
7212 dst_reg->smin_value = dst_reg->umin_value;
7213 dst_reg->smax_value = dst_reg->umax_value;
7214 }
7215 /* We may learn something more from the var_off */
7216 __update_reg_bounds(dst_reg);
7217 }
7218
scalar32_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7219 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7220 struct bpf_reg_state *src_reg)
7221 {
7222 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7223 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7224 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7225 s32 smin_val = src_reg->s32_min_value;
7226
7227 /* Assuming scalar64_min_max_xor will be called so it is safe
7228 * to skip updating register for known case.
7229 */
7230 if (src_known && dst_known)
7231 return;
7232
7233 /* We get both minimum and maximum from the var32_off. */
7234 dst_reg->u32_min_value = var32_off.value;
7235 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7236
7237 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7238 /* XORing two positive sign numbers gives a positive,
7239 * so safe to cast u32 result into s32.
7240 */
7241 dst_reg->s32_min_value = dst_reg->u32_min_value;
7242 dst_reg->s32_max_value = dst_reg->u32_max_value;
7243 } else {
7244 dst_reg->s32_min_value = S32_MIN;
7245 dst_reg->s32_max_value = S32_MAX;
7246 }
7247 }
7248
scalar_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7249 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7250 struct bpf_reg_state *src_reg)
7251 {
7252 bool src_known = tnum_is_const(src_reg->var_off);
7253 bool dst_known = tnum_is_const(dst_reg->var_off);
7254 s64 smin_val = src_reg->smin_value;
7255
7256 if (src_known && dst_known) {
7257 /* dst_reg->var_off.value has been updated earlier */
7258 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7259 return;
7260 }
7261
7262 /* We get both minimum and maximum from the var_off. */
7263 dst_reg->umin_value = dst_reg->var_off.value;
7264 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7265
7266 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7267 /* XORing two positive sign numbers gives a positive,
7268 * so safe to cast u64 result into s64.
7269 */
7270 dst_reg->smin_value = dst_reg->umin_value;
7271 dst_reg->smax_value = dst_reg->umax_value;
7272 } else {
7273 dst_reg->smin_value = S64_MIN;
7274 dst_reg->smax_value = S64_MAX;
7275 }
7276
7277 __update_reg_bounds(dst_reg);
7278 }
7279
__scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)7280 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7281 u64 umin_val, u64 umax_val)
7282 {
7283 /* We lose all sign bit information (except what we can pick
7284 * up from var_off)
7285 */
7286 dst_reg->s32_min_value = S32_MIN;
7287 dst_reg->s32_max_value = S32_MAX;
7288 /* If we might shift our top bit out, then we know nothing */
7289 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7290 dst_reg->u32_min_value = 0;
7291 dst_reg->u32_max_value = U32_MAX;
7292 } else {
7293 dst_reg->u32_min_value <<= umin_val;
7294 dst_reg->u32_max_value <<= umax_val;
7295 }
7296 }
7297
scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7298 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7299 struct bpf_reg_state *src_reg)
7300 {
7301 u32 umax_val = src_reg->u32_max_value;
7302 u32 umin_val = src_reg->u32_min_value;
7303 /* u32 alu operation will zext upper bits */
7304 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7305
7306 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7307 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7308 /* Not required but being careful mark reg64 bounds as unknown so
7309 * that we are forced to pick them up from tnum and zext later and
7310 * if some path skips this step we are still safe.
7311 */
7312 __mark_reg64_unbounded(dst_reg);
7313 __update_reg32_bounds(dst_reg);
7314 }
7315
__scalar64_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)7316 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7317 u64 umin_val, u64 umax_val)
7318 {
7319 /* Special case <<32 because it is a common compiler pattern to sign
7320 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7321 * positive we know this shift will also be positive so we can track
7322 * bounds correctly. Otherwise we lose all sign bit information except
7323 * what we can pick up from var_off. Perhaps we can generalize this
7324 * later to shifts of any length.
7325 */
7326 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7327 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7328 else
7329 dst_reg->smax_value = S64_MAX;
7330
7331 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7332 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7333 else
7334 dst_reg->smin_value = S64_MIN;
7335
7336 /* If we might shift our top bit out, then we know nothing */
7337 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7338 dst_reg->umin_value = 0;
7339 dst_reg->umax_value = U64_MAX;
7340 } else {
7341 dst_reg->umin_value <<= umin_val;
7342 dst_reg->umax_value <<= umax_val;
7343 }
7344 }
7345
scalar_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7346 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7347 struct bpf_reg_state *src_reg)
7348 {
7349 u64 umax_val = src_reg->umax_value;
7350 u64 umin_val = src_reg->umin_value;
7351
7352 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7353 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7354 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7355
7356 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7357 /* We may learn something more from the var_off */
7358 __update_reg_bounds(dst_reg);
7359 }
7360
scalar32_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7361 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7362 struct bpf_reg_state *src_reg)
7363 {
7364 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7365 u32 umax_val = src_reg->u32_max_value;
7366 u32 umin_val = src_reg->u32_min_value;
7367
7368 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7369 * be negative, then either:
7370 * 1) src_reg might be zero, so the sign bit of the result is
7371 * unknown, so we lose our signed bounds
7372 * 2) it's known negative, thus the unsigned bounds capture the
7373 * signed bounds
7374 * 3) the signed bounds cross zero, so they tell us nothing
7375 * about the result
7376 * If the value in dst_reg is known nonnegative, then again the
7377 * unsigned bounds capture the signed bounds.
7378 * Thus, in all cases it suffices to blow away our signed bounds
7379 * and rely on inferring new ones from the unsigned bounds and
7380 * var_off of the result.
7381 */
7382 dst_reg->s32_min_value = S32_MIN;
7383 dst_reg->s32_max_value = S32_MAX;
7384
7385 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7386 dst_reg->u32_min_value >>= umax_val;
7387 dst_reg->u32_max_value >>= umin_val;
7388
7389 __mark_reg64_unbounded(dst_reg);
7390 __update_reg32_bounds(dst_reg);
7391 }
7392
scalar_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7393 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7394 struct bpf_reg_state *src_reg)
7395 {
7396 u64 umax_val = src_reg->umax_value;
7397 u64 umin_val = src_reg->umin_value;
7398
7399 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7400 * be negative, then either:
7401 * 1) src_reg might be zero, so the sign bit of the result is
7402 * unknown, so we lose our signed bounds
7403 * 2) it's known negative, thus the unsigned bounds capture the
7404 * signed bounds
7405 * 3) the signed bounds cross zero, so they tell us nothing
7406 * about the result
7407 * If the value in dst_reg is known nonnegative, then again the
7408 * unsigned bounds capture the signed bounds.
7409 * Thus, in all cases it suffices to blow away our signed bounds
7410 * and rely on inferring new ones from the unsigned bounds and
7411 * var_off of the result.
7412 */
7413 dst_reg->smin_value = S64_MIN;
7414 dst_reg->smax_value = S64_MAX;
7415 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7416 dst_reg->umin_value >>= umax_val;
7417 dst_reg->umax_value >>= umin_val;
7418
7419 /* Its not easy to operate on alu32 bounds here because it depends
7420 * on bits being shifted in. Take easy way out and mark unbounded
7421 * so we can recalculate later from tnum.
7422 */
7423 __mark_reg32_unbounded(dst_reg);
7424 __update_reg_bounds(dst_reg);
7425 }
7426
scalar32_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7427 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7428 struct bpf_reg_state *src_reg)
7429 {
7430 u64 umin_val = src_reg->u32_min_value;
7431
7432 /* Upon reaching here, src_known is true and
7433 * umax_val is equal to umin_val.
7434 */
7435 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7436 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7437
7438 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7439
7440 /* blow away the dst_reg umin_value/umax_value and rely on
7441 * dst_reg var_off to refine the result.
7442 */
7443 dst_reg->u32_min_value = 0;
7444 dst_reg->u32_max_value = U32_MAX;
7445
7446 __mark_reg64_unbounded(dst_reg);
7447 __update_reg32_bounds(dst_reg);
7448 }
7449
scalar_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)7450 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7451 struct bpf_reg_state *src_reg)
7452 {
7453 u64 umin_val = src_reg->umin_value;
7454
7455 /* Upon reaching here, src_known is true and umax_val is equal
7456 * to umin_val.
7457 */
7458 dst_reg->smin_value >>= umin_val;
7459 dst_reg->smax_value >>= umin_val;
7460
7461 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7462
7463 /* blow away the dst_reg umin_value/umax_value and rely on
7464 * dst_reg var_off to refine the result.
7465 */
7466 dst_reg->umin_value = 0;
7467 dst_reg->umax_value = U64_MAX;
7468
7469 /* Its not easy to operate on alu32 bounds here because it depends
7470 * on bits being shifted in from upper 32-bits. Take easy way out
7471 * and mark unbounded so we can recalculate later from tnum.
7472 */
7473 __mark_reg32_unbounded(dst_reg);
7474 __update_reg_bounds(dst_reg);
7475 }
7476
7477 /* WARNING: This function does calculations on 64-bit values, but the actual
7478 * execution may occur on 32-bit values. Therefore, things like bitshifts
7479 * need extra checks in the 32-bit case.
7480 */
adjust_scalar_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_reg_state * dst_reg,struct bpf_reg_state src_reg)7481 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7482 struct bpf_insn *insn,
7483 struct bpf_reg_state *dst_reg,
7484 struct bpf_reg_state src_reg)
7485 {
7486 struct bpf_reg_state *regs = cur_regs(env);
7487 u8 opcode = BPF_OP(insn->code);
7488 bool src_known;
7489 s64 smin_val, smax_val;
7490 u64 umin_val, umax_val;
7491 s32 s32_min_val, s32_max_val;
7492 u32 u32_min_val, u32_max_val;
7493 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7494 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7495 int ret;
7496
7497 smin_val = src_reg.smin_value;
7498 smax_val = src_reg.smax_value;
7499 umin_val = src_reg.umin_value;
7500 umax_val = src_reg.umax_value;
7501
7502 s32_min_val = src_reg.s32_min_value;
7503 s32_max_val = src_reg.s32_max_value;
7504 u32_min_val = src_reg.u32_min_value;
7505 u32_max_val = src_reg.u32_max_value;
7506
7507 if (alu32) {
7508 src_known = tnum_subreg_is_const(src_reg.var_off);
7509 if ((src_known &&
7510 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7511 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7512 /* Taint dst register if offset had invalid bounds
7513 * derived from e.g. dead branches.
7514 */
7515 __mark_reg_unknown(env, dst_reg);
7516 return 0;
7517 }
7518 } else {
7519 src_known = tnum_is_const(src_reg.var_off);
7520 if ((src_known &&
7521 (smin_val != smax_val || umin_val != umax_val)) ||
7522 smin_val > smax_val || umin_val > umax_val) {
7523 /* Taint dst register if offset had invalid bounds
7524 * derived from e.g. dead branches.
7525 */
7526 __mark_reg_unknown(env, dst_reg);
7527 return 0;
7528 }
7529 }
7530
7531 if (!src_known &&
7532 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
7533 __mark_reg_unknown(env, dst_reg);
7534 return 0;
7535 }
7536
7537 if (sanitize_needed(opcode)) {
7538 ret = sanitize_val_alu(env, insn);
7539 if (ret < 0)
7540 return sanitize_err(env, insn, ret, NULL, NULL);
7541 }
7542
7543 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
7544 * There are two classes of instructions: The first class we track both
7545 * alu32 and alu64 sign/unsigned bounds independently this provides the
7546 * greatest amount of precision when alu operations are mixed with jmp32
7547 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7548 * and BPF_OR. This is possible because these ops have fairly easy to
7549 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7550 * See alu32 verifier tests for examples. The second class of
7551 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7552 * with regards to tracking sign/unsigned bounds because the bits may
7553 * cross subreg boundaries in the alu64 case. When this happens we mark
7554 * the reg unbounded in the subreg bound space and use the resulting
7555 * tnum to calculate an approximation of the sign/unsigned bounds.
7556 */
7557 switch (opcode) {
7558 case BPF_ADD:
7559 scalar32_min_max_add(dst_reg, &src_reg);
7560 scalar_min_max_add(dst_reg, &src_reg);
7561 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7562 break;
7563 case BPF_SUB:
7564 scalar32_min_max_sub(dst_reg, &src_reg);
7565 scalar_min_max_sub(dst_reg, &src_reg);
7566 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7567 break;
7568 case BPF_MUL:
7569 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7570 scalar32_min_max_mul(dst_reg, &src_reg);
7571 scalar_min_max_mul(dst_reg, &src_reg);
7572 break;
7573 case BPF_AND:
7574 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7575 scalar32_min_max_and(dst_reg, &src_reg);
7576 scalar_min_max_and(dst_reg, &src_reg);
7577 break;
7578 case BPF_OR:
7579 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7580 scalar32_min_max_or(dst_reg, &src_reg);
7581 scalar_min_max_or(dst_reg, &src_reg);
7582 break;
7583 case BPF_XOR:
7584 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7585 scalar32_min_max_xor(dst_reg, &src_reg);
7586 scalar_min_max_xor(dst_reg, &src_reg);
7587 break;
7588 case BPF_LSH:
7589 if (umax_val >= insn_bitness) {
7590 /* Shifts greater than 31 or 63 are undefined.
7591 * This includes shifts by a negative number.
7592 */
7593 mark_reg_unknown(env, regs, insn->dst_reg);
7594 break;
7595 }
7596 if (alu32)
7597 scalar32_min_max_lsh(dst_reg, &src_reg);
7598 else
7599 scalar_min_max_lsh(dst_reg, &src_reg);
7600 break;
7601 case BPF_RSH:
7602 if (umax_val >= insn_bitness) {
7603 /* Shifts greater than 31 or 63 are undefined.
7604 * This includes shifts by a negative number.
7605 */
7606 mark_reg_unknown(env, regs, insn->dst_reg);
7607 break;
7608 }
7609 if (alu32)
7610 scalar32_min_max_rsh(dst_reg, &src_reg);
7611 else
7612 scalar_min_max_rsh(dst_reg, &src_reg);
7613 break;
7614 case BPF_ARSH:
7615 if (umax_val >= insn_bitness) {
7616 /* Shifts greater than 31 or 63 are undefined.
7617 * This includes shifts by a negative number.
7618 */
7619 mark_reg_unknown(env, regs, insn->dst_reg);
7620 break;
7621 }
7622 if (alu32)
7623 scalar32_min_max_arsh(dst_reg, &src_reg);
7624 else
7625 scalar_min_max_arsh(dst_reg, &src_reg);
7626 break;
7627 default:
7628 mark_reg_unknown(env, regs, insn->dst_reg);
7629 break;
7630 }
7631
7632 /* ALU32 ops are zero extended into 64bit register */
7633 if (alu32)
7634 zext_32_to_64(dst_reg);
7635
7636 __update_reg_bounds(dst_reg);
7637 __reg_deduce_bounds(dst_reg);
7638 __reg_bound_offset(dst_reg);
7639 return 0;
7640 }
7641
7642 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7643 * and var_off.
7644 */
adjust_reg_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn)7645 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7646 struct bpf_insn *insn)
7647 {
7648 struct bpf_verifier_state *vstate = env->cur_state;
7649 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7650 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7651 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7652 u8 opcode = BPF_OP(insn->code);
7653 int err;
7654
7655 dst_reg = ®s[insn->dst_reg];
7656 src_reg = NULL;
7657 if (dst_reg->type != SCALAR_VALUE)
7658 ptr_reg = dst_reg;
7659 else
7660 /* Make sure ID is cleared otherwise dst_reg min/max could be
7661 * incorrectly propagated into other registers by find_equal_scalars()
7662 */
7663 dst_reg->id = 0;
7664 if (BPF_SRC(insn->code) == BPF_X) {
7665 src_reg = ®s[insn->src_reg];
7666 if (src_reg->type != SCALAR_VALUE) {
7667 if (dst_reg->type != SCALAR_VALUE) {
7668 /* Combining two pointers by any ALU op yields
7669 * an arbitrary scalar. Disallow all math except
7670 * pointer subtraction
7671 */
7672 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7673 mark_reg_unknown(env, regs, insn->dst_reg);
7674 return 0;
7675 }
7676 verbose(env, "R%d pointer %s pointer prohibited\n",
7677 insn->dst_reg,
7678 bpf_alu_string[opcode >> 4]);
7679 return -EACCES;
7680 } else {
7681 /* scalar += pointer
7682 * This is legal, but we have to reverse our
7683 * src/dest handling in computing the range
7684 */
7685 err = mark_chain_precision(env, insn->dst_reg);
7686 if (err)
7687 return err;
7688 return adjust_ptr_min_max_vals(env, insn,
7689 src_reg, dst_reg);
7690 }
7691 } else if (ptr_reg) {
7692 /* pointer += scalar */
7693 err = mark_chain_precision(env, insn->src_reg);
7694 if (err)
7695 return err;
7696 return adjust_ptr_min_max_vals(env, insn,
7697 dst_reg, src_reg);
7698 }
7699 } else {
7700 /* Pretend the src is a reg with a known value, since we only
7701 * need to be able to read from this state.
7702 */
7703 off_reg.type = SCALAR_VALUE;
7704 __mark_reg_known(&off_reg, insn->imm);
7705 src_reg = &off_reg;
7706 if (ptr_reg) /* pointer += K */
7707 return adjust_ptr_min_max_vals(env, insn,
7708 ptr_reg, src_reg);
7709 }
7710
7711 /* Got here implies adding two SCALAR_VALUEs */
7712 if (WARN_ON_ONCE(ptr_reg)) {
7713 print_verifier_state(env, state);
7714 verbose(env, "verifier internal error: unexpected ptr_reg\n");
7715 return -EINVAL;
7716 }
7717 if (WARN_ON(!src_reg)) {
7718 print_verifier_state(env, state);
7719 verbose(env, "verifier internal error: no src_reg\n");
7720 return -EINVAL;
7721 }
7722 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
7723 }
7724
7725 /* check validity of 32-bit and 64-bit arithmetic operations */
check_alu_op(struct bpf_verifier_env * env,struct bpf_insn * insn)7726 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
7727 {
7728 struct bpf_reg_state *regs = cur_regs(env);
7729 u8 opcode = BPF_OP(insn->code);
7730 int err;
7731
7732 if (opcode == BPF_END || opcode == BPF_NEG) {
7733 if (opcode == BPF_NEG) {
7734 if (BPF_SRC(insn->code) != 0 ||
7735 insn->src_reg != BPF_REG_0 ||
7736 insn->off != 0 || insn->imm != 0) {
7737 verbose(env, "BPF_NEG uses reserved fields\n");
7738 return -EINVAL;
7739 }
7740 } else {
7741 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
7742 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
7743 BPF_CLASS(insn->code) == BPF_ALU64) {
7744 verbose(env, "BPF_END uses reserved fields\n");
7745 return -EINVAL;
7746 }
7747 }
7748
7749 /* check src operand */
7750 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7751 if (err)
7752 return err;
7753
7754 if (is_pointer_value(env, insn->dst_reg)) {
7755 verbose(env, "R%d pointer arithmetic prohibited\n",
7756 insn->dst_reg);
7757 return -EACCES;
7758 }
7759
7760 /* check dest operand */
7761 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7762 if (err)
7763 return err;
7764
7765 } else if (opcode == BPF_MOV) {
7766
7767 if (BPF_SRC(insn->code) == BPF_X) {
7768 if (insn->imm != 0 || insn->off != 0) {
7769 verbose(env, "BPF_MOV uses reserved fields\n");
7770 return -EINVAL;
7771 }
7772
7773 /* check src operand */
7774 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7775 if (err)
7776 return err;
7777 } else {
7778 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7779 verbose(env, "BPF_MOV uses reserved fields\n");
7780 return -EINVAL;
7781 }
7782 }
7783
7784 /* check dest operand, mark as required later */
7785 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7786 if (err)
7787 return err;
7788
7789 if (BPF_SRC(insn->code) == BPF_X) {
7790 struct bpf_reg_state *src_reg = regs + insn->src_reg;
7791 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
7792
7793 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7794 /* case: R1 = R2
7795 * copy register state to dest reg
7796 */
7797 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
7798 /* Assign src and dst registers the same ID
7799 * that will be used by find_equal_scalars()
7800 * to propagate min/max range.
7801 */
7802 src_reg->id = ++env->id_gen;
7803 *dst_reg = *src_reg;
7804 dst_reg->live |= REG_LIVE_WRITTEN;
7805 dst_reg->subreg_def = DEF_NOT_SUBREG;
7806 } else {
7807 /* R1 = (u32) R2 */
7808 if (is_pointer_value(env, insn->src_reg)) {
7809 verbose(env,
7810 "R%d partial copy of pointer\n",
7811 insn->src_reg);
7812 return -EACCES;
7813 } else if (src_reg->type == SCALAR_VALUE) {
7814 *dst_reg = *src_reg;
7815 /* Make sure ID is cleared otherwise
7816 * dst_reg min/max could be incorrectly
7817 * propagated into src_reg by find_equal_scalars()
7818 */
7819 dst_reg->id = 0;
7820 dst_reg->live |= REG_LIVE_WRITTEN;
7821 dst_reg->subreg_def = env->insn_idx + 1;
7822 } else {
7823 mark_reg_unknown(env, regs,
7824 insn->dst_reg);
7825 }
7826 zext_32_to_64(dst_reg);
7827 }
7828 } else {
7829 /* case: R = imm
7830 * remember the value we stored into this reg
7831 */
7832 /* clear any state __mark_reg_known doesn't set */
7833 mark_reg_unknown(env, regs, insn->dst_reg);
7834 regs[insn->dst_reg].type = SCALAR_VALUE;
7835 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7836 __mark_reg_known(regs + insn->dst_reg,
7837 insn->imm);
7838 } else {
7839 __mark_reg_known(regs + insn->dst_reg,
7840 (u32)insn->imm);
7841 }
7842 }
7843
7844 } else if (opcode > BPF_END) {
7845 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
7846 return -EINVAL;
7847
7848 } else { /* all other ALU ops: and, sub, xor, add, ... */
7849
7850 if (BPF_SRC(insn->code) == BPF_X) {
7851 if (insn->imm != 0 || insn->off != 0) {
7852 verbose(env, "BPF_ALU uses reserved fields\n");
7853 return -EINVAL;
7854 }
7855 /* check src1 operand */
7856 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7857 if (err)
7858 return err;
7859 } else {
7860 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7861 verbose(env, "BPF_ALU uses reserved fields\n");
7862 return -EINVAL;
7863 }
7864 }
7865
7866 /* check src2 operand */
7867 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7868 if (err)
7869 return err;
7870
7871 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
7872 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
7873 verbose(env, "div by zero\n");
7874 return -EINVAL;
7875 }
7876
7877 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
7878 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
7879 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
7880
7881 if (insn->imm < 0 || insn->imm >= size) {
7882 verbose(env, "invalid shift %d\n", insn->imm);
7883 return -EINVAL;
7884 }
7885 }
7886
7887 /* check dest operand */
7888 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7889 if (err)
7890 return err;
7891
7892 return adjust_reg_min_max_vals(env, insn);
7893 }
7894
7895 return 0;
7896 }
7897
__find_good_pkt_pointers(struct bpf_func_state * state,struct bpf_reg_state * dst_reg,enum bpf_reg_type type,int new_range)7898 static void __find_good_pkt_pointers(struct bpf_func_state *state,
7899 struct bpf_reg_state *dst_reg,
7900 enum bpf_reg_type type, int new_range)
7901 {
7902 struct bpf_reg_state *reg;
7903 int i;
7904
7905 for (i = 0; i < MAX_BPF_REG; i++) {
7906 reg = &state->regs[i];
7907 if (reg->type == type && reg->id == dst_reg->id)
7908 /* keep the maximum range already checked */
7909 reg->range = max(reg->range, new_range);
7910 }
7911
7912 bpf_for_each_spilled_reg(i, state, reg) {
7913 if (!reg)
7914 continue;
7915 if (reg->type == type && reg->id == dst_reg->id)
7916 reg->range = max(reg->range, new_range);
7917 }
7918 }
7919
find_good_pkt_pointers(struct bpf_verifier_state * vstate,struct bpf_reg_state * dst_reg,enum bpf_reg_type type,bool range_right_open)7920 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
7921 struct bpf_reg_state *dst_reg,
7922 enum bpf_reg_type type,
7923 bool range_right_open)
7924 {
7925 int new_range, i;
7926
7927 if (dst_reg->off < 0 ||
7928 (dst_reg->off == 0 && range_right_open))
7929 /* This doesn't give us any range */
7930 return;
7931
7932 if (dst_reg->umax_value > MAX_PACKET_OFF ||
7933 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
7934 /* Risk of overflow. For instance, ptr + (1<<63) may be less
7935 * than pkt_end, but that's because it's also less than pkt.
7936 */
7937 return;
7938
7939 new_range = dst_reg->off;
7940 if (range_right_open)
7941 new_range--;
7942
7943 /* Examples for register markings:
7944 *
7945 * pkt_data in dst register:
7946 *
7947 * r2 = r3;
7948 * r2 += 8;
7949 * if (r2 > pkt_end) goto <handle exception>
7950 * <access okay>
7951 *
7952 * r2 = r3;
7953 * r2 += 8;
7954 * if (r2 < pkt_end) goto <access okay>
7955 * <handle exception>
7956 *
7957 * Where:
7958 * r2 == dst_reg, pkt_end == src_reg
7959 * r2=pkt(id=n,off=8,r=0)
7960 * r3=pkt(id=n,off=0,r=0)
7961 *
7962 * pkt_data in src register:
7963 *
7964 * r2 = r3;
7965 * r2 += 8;
7966 * if (pkt_end >= r2) goto <access okay>
7967 * <handle exception>
7968 *
7969 * r2 = r3;
7970 * r2 += 8;
7971 * if (pkt_end <= r2) goto <handle exception>
7972 * <access okay>
7973 *
7974 * Where:
7975 * pkt_end == dst_reg, r2 == src_reg
7976 * r2=pkt(id=n,off=8,r=0)
7977 * r3=pkt(id=n,off=0,r=0)
7978 *
7979 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
7980 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
7981 * and [r3, r3 + 8-1) respectively is safe to access depending on
7982 * the check.
7983 */
7984
7985 /* If our ids match, then we must have the same max_value. And we
7986 * don't care about the other reg's fixed offset, since if it's too big
7987 * the range won't allow anything.
7988 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
7989 */
7990 for (i = 0; i <= vstate->curframe; i++)
7991 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
7992 new_range);
7993 }
7994
is_branch32_taken(struct bpf_reg_state * reg,u32 val,u8 opcode)7995 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
7996 {
7997 struct tnum subreg = tnum_subreg(reg->var_off);
7998 s32 sval = (s32)val;
7999
8000 switch (opcode) {
8001 case BPF_JEQ:
8002 if (tnum_is_const(subreg))
8003 return !!tnum_equals_const(subreg, val);
8004 break;
8005 case BPF_JNE:
8006 if (tnum_is_const(subreg))
8007 return !tnum_equals_const(subreg, val);
8008 break;
8009 case BPF_JSET:
8010 if ((~subreg.mask & subreg.value) & val)
8011 return 1;
8012 if (!((subreg.mask | subreg.value) & val))
8013 return 0;
8014 break;
8015 case BPF_JGT:
8016 if (reg->u32_min_value > val)
8017 return 1;
8018 else if (reg->u32_max_value <= val)
8019 return 0;
8020 break;
8021 case BPF_JSGT:
8022 if (reg->s32_min_value > sval)
8023 return 1;
8024 else if (reg->s32_max_value <= sval)
8025 return 0;
8026 break;
8027 case BPF_JLT:
8028 if (reg->u32_max_value < val)
8029 return 1;
8030 else if (reg->u32_min_value >= val)
8031 return 0;
8032 break;
8033 case BPF_JSLT:
8034 if (reg->s32_max_value < sval)
8035 return 1;
8036 else if (reg->s32_min_value >= sval)
8037 return 0;
8038 break;
8039 case BPF_JGE:
8040 if (reg->u32_min_value >= val)
8041 return 1;
8042 else if (reg->u32_max_value < val)
8043 return 0;
8044 break;
8045 case BPF_JSGE:
8046 if (reg->s32_min_value >= sval)
8047 return 1;
8048 else if (reg->s32_max_value < sval)
8049 return 0;
8050 break;
8051 case BPF_JLE:
8052 if (reg->u32_max_value <= val)
8053 return 1;
8054 else if (reg->u32_min_value > val)
8055 return 0;
8056 break;
8057 case BPF_JSLE:
8058 if (reg->s32_max_value <= sval)
8059 return 1;
8060 else if (reg->s32_min_value > sval)
8061 return 0;
8062 break;
8063 }
8064
8065 return -1;
8066 }
8067
8068
is_branch64_taken(struct bpf_reg_state * reg,u64 val,u8 opcode)8069 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8070 {
8071 s64 sval = (s64)val;
8072
8073 switch (opcode) {
8074 case BPF_JEQ:
8075 if (tnum_is_const(reg->var_off))
8076 return !!tnum_equals_const(reg->var_off, val);
8077 break;
8078 case BPF_JNE:
8079 if (tnum_is_const(reg->var_off))
8080 return !tnum_equals_const(reg->var_off, val);
8081 break;
8082 case BPF_JSET:
8083 if ((~reg->var_off.mask & reg->var_off.value) & val)
8084 return 1;
8085 if (!((reg->var_off.mask | reg->var_off.value) & val))
8086 return 0;
8087 break;
8088 case BPF_JGT:
8089 if (reg->umin_value > val)
8090 return 1;
8091 else if (reg->umax_value <= val)
8092 return 0;
8093 break;
8094 case BPF_JSGT:
8095 if (reg->smin_value > sval)
8096 return 1;
8097 else if (reg->smax_value <= sval)
8098 return 0;
8099 break;
8100 case BPF_JLT:
8101 if (reg->umax_value < val)
8102 return 1;
8103 else if (reg->umin_value >= val)
8104 return 0;
8105 break;
8106 case BPF_JSLT:
8107 if (reg->smax_value < sval)
8108 return 1;
8109 else if (reg->smin_value >= sval)
8110 return 0;
8111 break;
8112 case BPF_JGE:
8113 if (reg->umin_value >= val)
8114 return 1;
8115 else if (reg->umax_value < val)
8116 return 0;
8117 break;
8118 case BPF_JSGE:
8119 if (reg->smin_value >= sval)
8120 return 1;
8121 else if (reg->smax_value < sval)
8122 return 0;
8123 break;
8124 case BPF_JLE:
8125 if (reg->umax_value <= val)
8126 return 1;
8127 else if (reg->umin_value > val)
8128 return 0;
8129 break;
8130 case BPF_JSLE:
8131 if (reg->smax_value <= sval)
8132 return 1;
8133 else if (reg->smin_value > sval)
8134 return 0;
8135 break;
8136 }
8137
8138 return -1;
8139 }
8140
8141 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8142 * and return:
8143 * 1 - branch will be taken and "goto target" will be executed
8144 * 0 - branch will not be taken and fall-through to next insn
8145 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8146 * range [0,10]
8147 */
is_branch_taken(struct bpf_reg_state * reg,u64 val,u8 opcode,bool is_jmp32)8148 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8149 bool is_jmp32)
8150 {
8151 if (__is_pointer_value(false, reg)) {
8152 if (!reg_type_not_null(reg->type))
8153 return -1;
8154
8155 /* If pointer is valid tests against zero will fail so we can
8156 * use this to direct branch taken.
8157 */
8158 if (val != 0)
8159 return -1;
8160
8161 switch (opcode) {
8162 case BPF_JEQ:
8163 return 0;
8164 case BPF_JNE:
8165 return 1;
8166 default:
8167 return -1;
8168 }
8169 }
8170
8171 if (is_jmp32)
8172 return is_branch32_taken(reg, val, opcode);
8173 return is_branch64_taken(reg, val, opcode);
8174 }
8175
flip_opcode(u32 opcode)8176 static int flip_opcode(u32 opcode)
8177 {
8178 /* How can we transform "a <op> b" into "b <op> a"? */
8179 static const u8 opcode_flip[16] = {
8180 /* these stay the same */
8181 [BPF_JEQ >> 4] = BPF_JEQ,
8182 [BPF_JNE >> 4] = BPF_JNE,
8183 [BPF_JSET >> 4] = BPF_JSET,
8184 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8185 [BPF_JGE >> 4] = BPF_JLE,
8186 [BPF_JGT >> 4] = BPF_JLT,
8187 [BPF_JLE >> 4] = BPF_JGE,
8188 [BPF_JLT >> 4] = BPF_JGT,
8189 [BPF_JSGE >> 4] = BPF_JSLE,
8190 [BPF_JSGT >> 4] = BPF_JSLT,
8191 [BPF_JSLE >> 4] = BPF_JSGE,
8192 [BPF_JSLT >> 4] = BPF_JSGT
8193 };
8194 return opcode_flip[opcode >> 4];
8195 }
8196
is_pkt_ptr_branch_taken(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg,u8 opcode)8197 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8198 struct bpf_reg_state *src_reg,
8199 u8 opcode)
8200 {
8201 struct bpf_reg_state *pkt;
8202
8203 if (src_reg->type == PTR_TO_PACKET_END) {
8204 pkt = dst_reg;
8205 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8206 pkt = src_reg;
8207 opcode = flip_opcode(opcode);
8208 } else {
8209 return -1;
8210 }
8211
8212 if (pkt->range >= 0)
8213 return -1;
8214
8215 switch (opcode) {
8216 case BPF_JLE:
8217 /* pkt <= pkt_end */
8218 fallthrough;
8219 case BPF_JGT:
8220 /* pkt > pkt_end */
8221 if (pkt->range == BEYOND_PKT_END)
8222 /* pkt has at last one extra byte beyond pkt_end */
8223 return opcode == BPF_JGT;
8224 break;
8225 case BPF_JLT:
8226 /* pkt < pkt_end */
8227 fallthrough;
8228 case BPF_JGE:
8229 /* pkt >= pkt_end */
8230 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8231 return opcode == BPF_JGE;
8232 break;
8233 }
8234 return -1;
8235 }
8236
8237 /* Adjusts the register min/max values in the case that the dst_reg is the
8238 * variable register that we are working on, and src_reg is a constant or we're
8239 * simply doing a BPF_K check.
8240 * In JEQ/JNE cases we also adjust the var_off values.
8241 */
reg_set_min_max(struct bpf_reg_state * true_reg,struct bpf_reg_state * false_reg,u64 val,u32 val32,u8 opcode,bool is_jmp32)8242 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8243 struct bpf_reg_state *false_reg,
8244 u64 val, u32 val32,
8245 u8 opcode, bool is_jmp32)
8246 {
8247 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8248 struct tnum false_64off = false_reg->var_off;
8249 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8250 struct tnum true_64off = true_reg->var_off;
8251 s64 sval = (s64)val;
8252 s32 sval32 = (s32)val32;
8253
8254 /* If the dst_reg is a pointer, we can't learn anything about its
8255 * variable offset from the compare (unless src_reg were a pointer into
8256 * the same object, but we don't bother with that.
8257 * Since false_reg and true_reg have the same type by construction, we
8258 * only need to check one of them for pointerness.
8259 */
8260 if (__is_pointer_value(false, false_reg))
8261 return;
8262
8263 switch (opcode) {
8264 case BPF_JEQ:
8265 case BPF_JNE:
8266 {
8267 struct bpf_reg_state *reg =
8268 opcode == BPF_JEQ ? true_reg : false_reg;
8269
8270 /* JEQ/JNE comparison doesn't change the register equivalence.
8271 * r1 = r2;
8272 * if (r1 == 42) goto label;
8273 * ...
8274 * label: // here both r1 and r2 are known to be 42.
8275 *
8276 * Hence when marking register as known preserve it's ID.
8277 */
8278 if (is_jmp32)
8279 __mark_reg32_known(reg, val32);
8280 else
8281 ___mark_reg_known(reg, val);
8282 break;
8283 }
8284 case BPF_JSET:
8285 if (is_jmp32) {
8286 false_32off = tnum_and(false_32off, tnum_const(~val32));
8287 if (is_power_of_2(val32))
8288 true_32off = tnum_or(true_32off,
8289 tnum_const(val32));
8290 } else {
8291 false_64off = tnum_and(false_64off, tnum_const(~val));
8292 if (is_power_of_2(val))
8293 true_64off = tnum_or(true_64off,
8294 tnum_const(val));
8295 }
8296 break;
8297 case BPF_JGE:
8298 case BPF_JGT:
8299 {
8300 if (is_jmp32) {
8301 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8302 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8303
8304 false_reg->u32_max_value = min(false_reg->u32_max_value,
8305 false_umax);
8306 true_reg->u32_min_value = max(true_reg->u32_min_value,
8307 true_umin);
8308 } else {
8309 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8310 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8311
8312 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8313 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8314 }
8315 break;
8316 }
8317 case BPF_JSGE:
8318 case BPF_JSGT:
8319 {
8320 if (is_jmp32) {
8321 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8322 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8323
8324 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8325 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8326 } else {
8327 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8328 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8329
8330 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8331 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8332 }
8333 break;
8334 }
8335 case BPF_JLE:
8336 case BPF_JLT:
8337 {
8338 if (is_jmp32) {
8339 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8340 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8341
8342 false_reg->u32_min_value = max(false_reg->u32_min_value,
8343 false_umin);
8344 true_reg->u32_max_value = min(true_reg->u32_max_value,
8345 true_umax);
8346 } else {
8347 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8348 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8349
8350 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8351 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8352 }
8353 break;
8354 }
8355 case BPF_JSLE:
8356 case BPF_JSLT:
8357 {
8358 if (is_jmp32) {
8359 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8360 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8361
8362 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8363 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8364 } else {
8365 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8366 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8367
8368 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8369 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8370 }
8371 break;
8372 }
8373 default:
8374 return;
8375 }
8376
8377 if (is_jmp32) {
8378 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8379 tnum_subreg(false_32off));
8380 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8381 tnum_subreg(true_32off));
8382 __reg_combine_32_into_64(false_reg);
8383 __reg_combine_32_into_64(true_reg);
8384 } else {
8385 false_reg->var_off = false_64off;
8386 true_reg->var_off = true_64off;
8387 __reg_combine_64_into_32(false_reg);
8388 __reg_combine_64_into_32(true_reg);
8389 }
8390 }
8391
8392 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8393 * the variable reg.
8394 */
reg_set_min_max_inv(struct bpf_reg_state * true_reg,struct bpf_reg_state * false_reg,u64 val,u32 val32,u8 opcode,bool is_jmp32)8395 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8396 struct bpf_reg_state *false_reg,
8397 u64 val, u32 val32,
8398 u8 opcode, bool is_jmp32)
8399 {
8400 opcode = flip_opcode(opcode);
8401 /* This uses zero as "not present in table"; luckily the zero opcode,
8402 * BPF_JA, can't get here.
8403 */
8404 if (opcode)
8405 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8406 }
8407
8408 /* Regs are known to be equal, so intersect their min/max/var_off */
__reg_combine_min_max(struct bpf_reg_state * src_reg,struct bpf_reg_state * dst_reg)8409 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8410 struct bpf_reg_state *dst_reg)
8411 {
8412 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8413 dst_reg->umin_value);
8414 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8415 dst_reg->umax_value);
8416 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8417 dst_reg->smin_value);
8418 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8419 dst_reg->smax_value);
8420 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8421 dst_reg->var_off);
8422 /* We might have learned new bounds from the var_off. */
8423 __update_reg_bounds(src_reg);
8424 __update_reg_bounds(dst_reg);
8425 /* We might have learned something about the sign bit. */
8426 __reg_deduce_bounds(src_reg);
8427 __reg_deduce_bounds(dst_reg);
8428 /* We might have learned some bits from the bounds. */
8429 __reg_bound_offset(src_reg);
8430 __reg_bound_offset(dst_reg);
8431 /* Intersecting with the old var_off might have improved our bounds
8432 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8433 * then new var_off is (0; 0x7f...fc) which improves our umax.
8434 */
8435 __update_reg_bounds(src_reg);
8436 __update_reg_bounds(dst_reg);
8437 }
8438
reg_combine_min_max(struct bpf_reg_state * true_src,struct bpf_reg_state * true_dst,struct bpf_reg_state * false_src,struct bpf_reg_state * false_dst,u8 opcode)8439 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8440 struct bpf_reg_state *true_dst,
8441 struct bpf_reg_state *false_src,
8442 struct bpf_reg_state *false_dst,
8443 u8 opcode)
8444 {
8445 switch (opcode) {
8446 case BPF_JEQ:
8447 __reg_combine_min_max(true_src, true_dst);
8448 break;
8449 case BPF_JNE:
8450 __reg_combine_min_max(false_src, false_dst);
8451 break;
8452 }
8453 }
8454
mark_ptr_or_null_reg(struct bpf_func_state * state,struct bpf_reg_state * reg,u32 id,bool is_null)8455 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8456 struct bpf_reg_state *reg, u32 id,
8457 bool is_null)
8458 {
8459 if (reg_type_may_be_null(reg->type) && reg->id == id &&
8460 !WARN_ON_ONCE(!reg->id)) {
8461 /* Old offset (both fixed and variable parts) should
8462 * have been known-zero, because we don't allow pointer
8463 * arithmetic on pointers that might be NULL.
8464 */
8465 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8466 !tnum_equals_const(reg->var_off, 0) ||
8467 reg->off)) {
8468 __mark_reg_known_zero(reg);
8469 reg->off = 0;
8470 }
8471 if (is_null) {
8472 reg->type = SCALAR_VALUE;
8473 /* We don't need id and ref_obj_id from this point
8474 * onwards anymore, thus we should better reset it,
8475 * so that state pruning has chances to take effect.
8476 */
8477 reg->id = 0;
8478 reg->ref_obj_id = 0;
8479
8480 return;
8481 }
8482
8483 mark_ptr_not_null_reg(reg);
8484
8485 if (!reg_may_point_to_spin_lock(reg)) {
8486 /* For not-NULL ptr, reg->ref_obj_id will be reset
8487 * in release_reg_references().
8488 *
8489 * reg->id is still used by spin_lock ptr. Other
8490 * than spin_lock ptr type, reg->id can be reset.
8491 */
8492 reg->id = 0;
8493 }
8494 }
8495 }
8496
__mark_ptr_or_null_regs(struct bpf_func_state * state,u32 id,bool is_null)8497 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
8498 bool is_null)
8499 {
8500 struct bpf_reg_state *reg;
8501 int i;
8502
8503 for (i = 0; i < MAX_BPF_REG; i++)
8504 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
8505
8506 bpf_for_each_spilled_reg(i, state, reg) {
8507 if (!reg)
8508 continue;
8509 mark_ptr_or_null_reg(state, reg, id, is_null);
8510 }
8511 }
8512
8513 /* The logic is similar to find_good_pkt_pointers(), both could eventually
8514 * be folded together at some point.
8515 */
mark_ptr_or_null_regs(struct bpf_verifier_state * vstate,u32 regno,bool is_null)8516 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
8517 bool is_null)
8518 {
8519 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8520 struct bpf_reg_state *regs = state->regs;
8521 u32 ref_obj_id = regs[regno].ref_obj_id;
8522 u32 id = regs[regno].id;
8523 int i;
8524
8525 if (ref_obj_id && ref_obj_id == id && is_null)
8526 /* regs[regno] is in the " == NULL" branch.
8527 * No one could have freed the reference state before
8528 * doing the NULL check.
8529 */
8530 WARN_ON_ONCE(release_reference_state(state, id));
8531
8532 for (i = 0; i <= vstate->curframe; i++)
8533 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
8534 }
8535
try_match_pkt_pointers(const struct bpf_insn * insn,struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg,struct bpf_verifier_state * this_branch,struct bpf_verifier_state * other_branch)8536 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
8537 struct bpf_reg_state *dst_reg,
8538 struct bpf_reg_state *src_reg,
8539 struct bpf_verifier_state *this_branch,
8540 struct bpf_verifier_state *other_branch)
8541 {
8542 if (BPF_SRC(insn->code) != BPF_X)
8543 return false;
8544
8545 /* Pointers are always 64-bit. */
8546 if (BPF_CLASS(insn->code) == BPF_JMP32)
8547 return false;
8548
8549 switch (BPF_OP(insn->code)) {
8550 case BPF_JGT:
8551 if ((dst_reg->type == PTR_TO_PACKET &&
8552 src_reg->type == PTR_TO_PACKET_END) ||
8553 (dst_reg->type == PTR_TO_PACKET_META &&
8554 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8555 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8556 find_good_pkt_pointers(this_branch, dst_reg,
8557 dst_reg->type, false);
8558 mark_pkt_end(other_branch, insn->dst_reg, true);
8559 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8560 src_reg->type == PTR_TO_PACKET) ||
8561 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8562 src_reg->type == PTR_TO_PACKET_META)) {
8563 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8564 find_good_pkt_pointers(other_branch, src_reg,
8565 src_reg->type, true);
8566 mark_pkt_end(this_branch, insn->src_reg, false);
8567 } else {
8568 return false;
8569 }
8570 break;
8571 case BPF_JLT:
8572 if ((dst_reg->type == PTR_TO_PACKET &&
8573 src_reg->type == PTR_TO_PACKET_END) ||
8574 (dst_reg->type == PTR_TO_PACKET_META &&
8575 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8576 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8577 find_good_pkt_pointers(other_branch, dst_reg,
8578 dst_reg->type, true);
8579 mark_pkt_end(this_branch, insn->dst_reg, false);
8580 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8581 src_reg->type == PTR_TO_PACKET) ||
8582 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8583 src_reg->type == PTR_TO_PACKET_META)) {
8584 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8585 find_good_pkt_pointers(this_branch, src_reg,
8586 src_reg->type, false);
8587 mark_pkt_end(other_branch, insn->src_reg, true);
8588 } else {
8589 return false;
8590 }
8591 break;
8592 case BPF_JGE:
8593 if ((dst_reg->type == PTR_TO_PACKET &&
8594 src_reg->type == PTR_TO_PACKET_END) ||
8595 (dst_reg->type == PTR_TO_PACKET_META &&
8596 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8597 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8598 find_good_pkt_pointers(this_branch, dst_reg,
8599 dst_reg->type, true);
8600 mark_pkt_end(other_branch, insn->dst_reg, false);
8601 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8602 src_reg->type == PTR_TO_PACKET) ||
8603 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8604 src_reg->type == PTR_TO_PACKET_META)) {
8605 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8606 find_good_pkt_pointers(other_branch, src_reg,
8607 src_reg->type, false);
8608 mark_pkt_end(this_branch, insn->src_reg, true);
8609 } else {
8610 return false;
8611 }
8612 break;
8613 case BPF_JLE:
8614 if ((dst_reg->type == PTR_TO_PACKET &&
8615 src_reg->type == PTR_TO_PACKET_END) ||
8616 (dst_reg->type == PTR_TO_PACKET_META &&
8617 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8618 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8619 find_good_pkt_pointers(other_branch, dst_reg,
8620 dst_reg->type, false);
8621 mark_pkt_end(this_branch, insn->dst_reg, true);
8622 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8623 src_reg->type == PTR_TO_PACKET) ||
8624 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8625 src_reg->type == PTR_TO_PACKET_META)) {
8626 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8627 find_good_pkt_pointers(this_branch, src_reg,
8628 src_reg->type, true);
8629 mark_pkt_end(other_branch, insn->src_reg, false);
8630 } else {
8631 return false;
8632 }
8633 break;
8634 default:
8635 return false;
8636 }
8637
8638 return true;
8639 }
8640
find_equal_scalars(struct bpf_verifier_state * vstate,struct bpf_reg_state * known_reg)8641 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8642 struct bpf_reg_state *known_reg)
8643 {
8644 struct bpf_func_state *state;
8645 struct bpf_reg_state *reg;
8646 int i, j;
8647
8648 for (i = 0; i <= vstate->curframe; i++) {
8649 state = vstate->frame[i];
8650 for (j = 0; j < MAX_BPF_REG; j++) {
8651 reg = &state->regs[j];
8652 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8653 *reg = *known_reg;
8654 }
8655
8656 bpf_for_each_spilled_reg(j, state, reg) {
8657 if (!reg)
8658 continue;
8659 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8660 *reg = *known_reg;
8661 }
8662 }
8663 }
8664
check_cond_jmp_op(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)8665 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8666 struct bpf_insn *insn, int *insn_idx)
8667 {
8668 struct bpf_verifier_state *this_branch = env->cur_state;
8669 struct bpf_verifier_state *other_branch;
8670 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
8671 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
8672 u8 opcode = BPF_OP(insn->code);
8673 bool is_jmp32;
8674 int pred = -1;
8675 int err;
8676
8677 /* Only conditional jumps are expected to reach here. */
8678 if (opcode == BPF_JA || opcode > BPF_JSLE) {
8679 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
8680 return -EINVAL;
8681 }
8682
8683 if (BPF_SRC(insn->code) == BPF_X) {
8684 if (insn->imm != 0) {
8685 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8686 return -EINVAL;
8687 }
8688
8689 /* check src1 operand */
8690 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8691 if (err)
8692 return err;
8693
8694 if (is_pointer_value(env, insn->src_reg)) {
8695 verbose(env, "R%d pointer comparison prohibited\n",
8696 insn->src_reg);
8697 return -EACCES;
8698 }
8699 src_reg = ®s[insn->src_reg];
8700 } else {
8701 if (insn->src_reg != BPF_REG_0) {
8702 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8703 return -EINVAL;
8704 }
8705 }
8706
8707 /* check src2 operand */
8708 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8709 if (err)
8710 return err;
8711
8712 dst_reg = ®s[insn->dst_reg];
8713 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
8714
8715 if (BPF_SRC(insn->code) == BPF_K) {
8716 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
8717 } else if (src_reg->type == SCALAR_VALUE &&
8718 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
8719 pred = is_branch_taken(dst_reg,
8720 tnum_subreg(src_reg->var_off).value,
8721 opcode,
8722 is_jmp32);
8723 } else if (src_reg->type == SCALAR_VALUE &&
8724 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
8725 pred = is_branch_taken(dst_reg,
8726 src_reg->var_off.value,
8727 opcode,
8728 is_jmp32);
8729 } else if (reg_is_pkt_pointer_any(dst_reg) &&
8730 reg_is_pkt_pointer_any(src_reg) &&
8731 !is_jmp32) {
8732 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
8733 }
8734
8735 if (pred >= 0) {
8736 /* If we get here with a dst_reg pointer type it is because
8737 * above is_branch_taken() special cased the 0 comparison.
8738 */
8739 if (!__is_pointer_value(false, dst_reg))
8740 err = mark_chain_precision(env, insn->dst_reg);
8741 if (BPF_SRC(insn->code) == BPF_X && !err &&
8742 !__is_pointer_value(false, src_reg))
8743 err = mark_chain_precision(env, insn->src_reg);
8744 if (err)
8745 return err;
8746 }
8747 if (pred == 1) {
8748 /* only follow the goto, ignore fall-through */
8749 *insn_idx += insn->off;
8750 return 0;
8751 } else if (pred == 0) {
8752 /* only follow fall-through branch, since
8753 * that's where the program will go
8754 */
8755 return 0;
8756 }
8757
8758 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
8759 false);
8760 if (!other_branch)
8761 return -EFAULT;
8762 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
8763
8764 /* detect if we are comparing against a constant value so we can adjust
8765 * our min/max values for our dst register.
8766 * this is only legit if both are scalars (or pointers to the same
8767 * object, I suppose, but we don't support that right now), because
8768 * otherwise the different base pointers mean the offsets aren't
8769 * comparable.
8770 */
8771 if (BPF_SRC(insn->code) == BPF_X) {
8772 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
8773
8774 if (dst_reg->type == SCALAR_VALUE &&
8775 src_reg->type == SCALAR_VALUE) {
8776 if (tnum_is_const(src_reg->var_off) ||
8777 (is_jmp32 &&
8778 tnum_is_const(tnum_subreg(src_reg->var_off))))
8779 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8780 dst_reg,
8781 src_reg->var_off.value,
8782 tnum_subreg(src_reg->var_off).value,
8783 opcode, is_jmp32);
8784 else if (tnum_is_const(dst_reg->var_off) ||
8785 (is_jmp32 &&
8786 tnum_is_const(tnum_subreg(dst_reg->var_off))))
8787 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
8788 src_reg,
8789 dst_reg->var_off.value,
8790 tnum_subreg(dst_reg->var_off).value,
8791 opcode, is_jmp32);
8792 else if (!is_jmp32 &&
8793 (opcode == BPF_JEQ || opcode == BPF_JNE))
8794 /* Comparing for equality, we can combine knowledge */
8795 reg_combine_min_max(&other_branch_regs[insn->src_reg],
8796 &other_branch_regs[insn->dst_reg],
8797 src_reg, dst_reg, opcode);
8798 if (src_reg->id &&
8799 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
8800 find_equal_scalars(this_branch, src_reg);
8801 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
8802 }
8803
8804 }
8805 } else if (dst_reg->type == SCALAR_VALUE) {
8806 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8807 dst_reg, insn->imm, (u32)insn->imm,
8808 opcode, is_jmp32);
8809 }
8810
8811 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
8812 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
8813 find_equal_scalars(this_branch, dst_reg);
8814 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
8815 }
8816
8817 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
8818 * NOTE: these optimizations below are related with pointer comparison
8819 * which will never be JMP32.
8820 */
8821 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
8822 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
8823 reg_type_may_be_null(dst_reg->type)) {
8824 /* Mark all identical registers in each branch as either
8825 * safe or unknown depending R == 0 or R != 0 conditional.
8826 */
8827 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
8828 opcode == BPF_JNE);
8829 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
8830 opcode == BPF_JEQ);
8831 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
8832 this_branch, other_branch) &&
8833 is_pointer_value(env, insn->dst_reg)) {
8834 verbose(env, "R%d pointer comparison prohibited\n",
8835 insn->dst_reg);
8836 return -EACCES;
8837 }
8838 if (env->log.level & BPF_LOG_LEVEL)
8839 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
8840 return 0;
8841 }
8842
8843 /* verify BPF_LD_IMM64 instruction */
check_ld_imm(struct bpf_verifier_env * env,struct bpf_insn * insn)8844 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
8845 {
8846 struct bpf_insn_aux_data *aux = cur_aux(env);
8847 struct bpf_reg_state *regs = cur_regs(env);
8848 struct bpf_reg_state *dst_reg;
8849 struct bpf_map *map;
8850 int err;
8851
8852 if (BPF_SIZE(insn->code) != BPF_DW) {
8853 verbose(env, "invalid BPF_LD_IMM insn\n");
8854 return -EINVAL;
8855 }
8856 if (insn->off != 0) {
8857 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
8858 return -EINVAL;
8859 }
8860
8861 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8862 if (err)
8863 return err;
8864
8865 dst_reg = ®s[insn->dst_reg];
8866 if (insn->src_reg == 0) {
8867 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
8868
8869 dst_reg->type = SCALAR_VALUE;
8870 __mark_reg_known(®s[insn->dst_reg], imm);
8871 return 0;
8872 }
8873
8874 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
8875 mark_reg_known_zero(env, regs, insn->dst_reg);
8876
8877 dst_reg->type = aux->btf_var.reg_type;
8878 switch (dst_reg->type) {
8879 case PTR_TO_MEM:
8880 dst_reg->mem_size = aux->btf_var.mem_size;
8881 break;
8882 case PTR_TO_BTF_ID:
8883 case PTR_TO_PERCPU_BTF_ID:
8884 dst_reg->btf = aux->btf_var.btf;
8885 dst_reg->btf_id = aux->btf_var.btf_id;
8886 break;
8887 default:
8888 verbose(env, "bpf verifier is misconfigured\n");
8889 return -EFAULT;
8890 }
8891 return 0;
8892 }
8893
8894 if (insn->src_reg == BPF_PSEUDO_FUNC) {
8895 struct bpf_prog_aux *aux = env->prog->aux;
8896 u32 subprogno = insn[1].imm;
8897
8898 if (!aux->func_info) {
8899 verbose(env, "missing btf func_info\n");
8900 return -EINVAL;
8901 }
8902 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
8903 verbose(env, "callback function not static\n");
8904 return -EINVAL;
8905 }
8906
8907 dst_reg->type = PTR_TO_FUNC;
8908 dst_reg->subprogno = subprogno;
8909 return 0;
8910 }
8911
8912 map = env->used_maps[aux->map_index];
8913 mark_reg_known_zero(env, regs, insn->dst_reg);
8914 dst_reg->map_ptr = map;
8915
8916 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
8917 dst_reg->type = PTR_TO_MAP_VALUE;
8918 dst_reg->off = aux->map_off;
8919 if (map_value_has_spin_lock(map))
8920 dst_reg->id = ++env->id_gen;
8921 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8922 dst_reg->type = CONST_PTR_TO_MAP;
8923 } else {
8924 verbose(env, "bpf verifier is misconfigured\n");
8925 return -EINVAL;
8926 }
8927
8928 return 0;
8929 }
8930
may_access_skb(enum bpf_prog_type type)8931 static bool may_access_skb(enum bpf_prog_type type)
8932 {
8933 switch (type) {
8934 case BPF_PROG_TYPE_SOCKET_FILTER:
8935 case BPF_PROG_TYPE_SCHED_CLS:
8936 case BPF_PROG_TYPE_SCHED_ACT:
8937 return true;
8938 default:
8939 return false;
8940 }
8941 }
8942
8943 /* verify safety of LD_ABS|LD_IND instructions:
8944 * - they can only appear in the programs where ctx == skb
8945 * - since they are wrappers of function calls, they scratch R1-R5 registers,
8946 * preserve R6-R9, and store return value into R0
8947 *
8948 * Implicit input:
8949 * ctx == skb == R6 == CTX
8950 *
8951 * Explicit input:
8952 * SRC == any register
8953 * IMM == 32-bit immediate
8954 *
8955 * Output:
8956 * R0 - 8/16/32-bit skb data converted to cpu endianness
8957 */
check_ld_abs(struct bpf_verifier_env * env,struct bpf_insn * insn)8958 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
8959 {
8960 struct bpf_reg_state *regs = cur_regs(env);
8961 static const int ctx_reg = BPF_REG_6;
8962 u8 mode = BPF_MODE(insn->code);
8963 int i, err;
8964
8965 if (!may_access_skb(resolve_prog_type(env->prog))) {
8966 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
8967 return -EINVAL;
8968 }
8969
8970 if (!env->ops->gen_ld_abs) {
8971 verbose(env, "bpf verifier is misconfigured\n");
8972 return -EINVAL;
8973 }
8974
8975 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
8976 BPF_SIZE(insn->code) == BPF_DW ||
8977 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
8978 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
8979 return -EINVAL;
8980 }
8981
8982 /* check whether implicit source operand (register R6) is readable */
8983 err = check_reg_arg(env, ctx_reg, SRC_OP);
8984 if (err)
8985 return err;
8986
8987 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
8988 * gen_ld_abs() may terminate the program at runtime, leading to
8989 * reference leak.
8990 */
8991 err = check_reference_leak(env);
8992 if (err) {
8993 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
8994 return err;
8995 }
8996
8997 if (env->cur_state->active_spin_lock) {
8998 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
8999 return -EINVAL;
9000 }
9001
9002 if (regs[ctx_reg].type != PTR_TO_CTX) {
9003 verbose(env,
9004 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9005 return -EINVAL;
9006 }
9007
9008 if (mode == BPF_IND) {
9009 /* check explicit source operand */
9010 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9011 if (err)
9012 return err;
9013 }
9014
9015 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9016 if (err < 0)
9017 return err;
9018
9019 /* reset caller saved regs to unreadable */
9020 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9021 mark_reg_not_init(env, regs, caller_saved[i]);
9022 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9023 }
9024
9025 /* mark destination R0 register as readable, since it contains
9026 * the value fetched from the packet.
9027 * Already marked as written above.
9028 */
9029 mark_reg_unknown(env, regs, BPF_REG_0);
9030 /* ld_abs load up to 32-bit skb data. */
9031 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9032 return 0;
9033 }
9034
check_return_code(struct bpf_verifier_env * env)9035 static int check_return_code(struct bpf_verifier_env *env)
9036 {
9037 struct tnum enforce_attach_type_range = tnum_unknown;
9038 const struct bpf_prog *prog = env->prog;
9039 struct bpf_reg_state *reg;
9040 struct tnum range = tnum_range(0, 1);
9041 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9042 int err;
9043 const bool is_subprog = env->cur_state->frame[0]->subprogno;
9044
9045 /* LSM and struct_ops func-ptr's return type could be "void" */
9046 if (!is_subprog &&
9047 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9048 prog_type == BPF_PROG_TYPE_LSM) &&
9049 !prog->aux->attach_func_proto->type)
9050 return 0;
9051
9052 /* eBPF calling convetion is such that R0 is used
9053 * to return the value from eBPF program.
9054 * Make sure that it's readable at this time
9055 * of bpf_exit, which means that program wrote
9056 * something into it earlier
9057 */
9058 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9059 if (err)
9060 return err;
9061
9062 if (is_pointer_value(env, BPF_REG_0)) {
9063 verbose(env, "R0 leaks addr as return value\n");
9064 return -EACCES;
9065 }
9066
9067 reg = cur_regs(env) + BPF_REG_0;
9068 if (is_subprog) {
9069 if (reg->type != SCALAR_VALUE) {
9070 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9071 reg_type_str[reg->type]);
9072 return -EINVAL;
9073 }
9074 return 0;
9075 }
9076
9077 switch (prog_type) {
9078 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9079 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9080 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9081 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9082 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9083 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9084 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9085 range = tnum_range(1, 1);
9086 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9087 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9088 range = tnum_range(0, 3);
9089 break;
9090 case BPF_PROG_TYPE_CGROUP_SKB:
9091 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9092 range = tnum_range(0, 3);
9093 enforce_attach_type_range = tnum_range(2, 3);
9094 }
9095 break;
9096 case BPF_PROG_TYPE_CGROUP_SOCK:
9097 case BPF_PROG_TYPE_SOCK_OPS:
9098 case BPF_PROG_TYPE_CGROUP_DEVICE:
9099 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9100 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9101 break;
9102 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9103 if (!env->prog->aux->attach_btf_id)
9104 return 0;
9105 range = tnum_const(0);
9106 break;
9107 case BPF_PROG_TYPE_TRACING:
9108 switch (env->prog->expected_attach_type) {
9109 case BPF_TRACE_FENTRY:
9110 case BPF_TRACE_FEXIT:
9111 range = tnum_const(0);
9112 break;
9113 case BPF_TRACE_RAW_TP:
9114 case BPF_MODIFY_RETURN:
9115 return 0;
9116 case BPF_TRACE_ITER:
9117 break;
9118 default:
9119 return -ENOTSUPP;
9120 }
9121 break;
9122 case BPF_PROG_TYPE_SK_LOOKUP:
9123 range = tnum_range(SK_DROP, SK_PASS);
9124 break;
9125 case BPF_PROG_TYPE_EXT:
9126 /* freplace program can return anything as its return value
9127 * depends on the to-be-replaced kernel func or bpf program.
9128 */
9129 default:
9130 return 0;
9131 }
9132
9133 if (reg->type != SCALAR_VALUE) {
9134 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9135 reg_type_str[reg->type]);
9136 return -EINVAL;
9137 }
9138
9139 if (!tnum_in(range, reg->var_off)) {
9140 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9141 return -EINVAL;
9142 }
9143
9144 if (!tnum_is_unknown(enforce_attach_type_range) &&
9145 tnum_in(enforce_attach_type_range, reg->var_off))
9146 env->prog->enforce_expected_attach_type = 1;
9147 return 0;
9148 }
9149
9150 /* non-recursive DFS pseudo code
9151 * 1 procedure DFS-iterative(G,v):
9152 * 2 label v as discovered
9153 * 3 let S be a stack
9154 * 4 S.push(v)
9155 * 5 while S is not empty
9156 * 6 t <- S.pop()
9157 * 7 if t is what we're looking for:
9158 * 8 return t
9159 * 9 for all edges e in G.adjacentEdges(t) do
9160 * 10 if edge e is already labelled
9161 * 11 continue with the next edge
9162 * 12 w <- G.adjacentVertex(t,e)
9163 * 13 if vertex w is not discovered and not explored
9164 * 14 label e as tree-edge
9165 * 15 label w as discovered
9166 * 16 S.push(w)
9167 * 17 continue at 5
9168 * 18 else if vertex w is discovered
9169 * 19 label e as back-edge
9170 * 20 else
9171 * 21 // vertex w is explored
9172 * 22 label e as forward- or cross-edge
9173 * 23 label t as explored
9174 * 24 S.pop()
9175 *
9176 * convention:
9177 * 0x10 - discovered
9178 * 0x11 - discovered and fall-through edge labelled
9179 * 0x12 - discovered and fall-through and branch edges labelled
9180 * 0x20 - explored
9181 */
9182
9183 enum {
9184 DISCOVERED = 0x10,
9185 EXPLORED = 0x20,
9186 FALLTHROUGH = 1,
9187 BRANCH = 2,
9188 };
9189
state_htab_size(struct bpf_verifier_env * env)9190 static u32 state_htab_size(struct bpf_verifier_env *env)
9191 {
9192 return env->prog->len;
9193 }
9194
explored_state(struct bpf_verifier_env * env,int idx)9195 static struct bpf_verifier_state_list **explored_state(
9196 struct bpf_verifier_env *env,
9197 int idx)
9198 {
9199 struct bpf_verifier_state *cur = env->cur_state;
9200 struct bpf_func_state *state = cur->frame[cur->curframe];
9201
9202 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9203 }
9204
init_explored_state(struct bpf_verifier_env * env,int idx)9205 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9206 {
9207 env->insn_aux_data[idx].prune_point = true;
9208 }
9209
9210 enum {
9211 DONE_EXPLORING = 0,
9212 KEEP_EXPLORING = 1,
9213 };
9214
9215 /* t, w, e - match pseudo-code above:
9216 * t - index of current instruction
9217 * w - next instruction
9218 * e - edge
9219 */
push_insn(int t,int w,int e,struct bpf_verifier_env * env,bool loop_ok)9220 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9221 bool loop_ok)
9222 {
9223 int *insn_stack = env->cfg.insn_stack;
9224 int *insn_state = env->cfg.insn_state;
9225
9226 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9227 return DONE_EXPLORING;
9228
9229 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9230 return DONE_EXPLORING;
9231
9232 if (w < 0 || w >= env->prog->len) {
9233 verbose_linfo(env, t, "%d: ", t);
9234 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9235 return -EINVAL;
9236 }
9237
9238 if (e == BRANCH)
9239 /* mark branch target for state pruning */
9240 init_explored_state(env, w);
9241
9242 if (insn_state[w] == 0) {
9243 /* tree-edge */
9244 insn_state[t] = DISCOVERED | e;
9245 insn_state[w] = DISCOVERED;
9246 if (env->cfg.cur_stack >= env->prog->len)
9247 return -E2BIG;
9248 insn_stack[env->cfg.cur_stack++] = w;
9249 return KEEP_EXPLORING;
9250 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9251 if (loop_ok && env->bpf_capable)
9252 return DONE_EXPLORING;
9253 verbose_linfo(env, t, "%d: ", t);
9254 verbose_linfo(env, w, "%d: ", w);
9255 verbose(env, "back-edge from insn %d to %d\n", t, w);
9256 return -EINVAL;
9257 } else if (insn_state[w] == EXPLORED) {
9258 /* forward- or cross-edge */
9259 insn_state[t] = DISCOVERED | e;
9260 } else {
9261 verbose(env, "insn state internal bug\n");
9262 return -EFAULT;
9263 }
9264 return DONE_EXPLORING;
9265 }
9266
visit_func_call_insn(int t,int insn_cnt,struct bpf_insn * insns,struct bpf_verifier_env * env,bool visit_callee)9267 static int visit_func_call_insn(int t, int insn_cnt,
9268 struct bpf_insn *insns,
9269 struct bpf_verifier_env *env,
9270 bool visit_callee)
9271 {
9272 int ret;
9273
9274 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9275 if (ret)
9276 return ret;
9277
9278 if (t + 1 < insn_cnt)
9279 init_explored_state(env, t + 1);
9280 if (visit_callee) {
9281 init_explored_state(env, t);
9282 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
9283 env, false);
9284 }
9285 return ret;
9286 }
9287
9288 /* Visits the instruction at index t and returns one of the following:
9289 * < 0 - an error occurred
9290 * DONE_EXPLORING - the instruction was fully explored
9291 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9292 */
visit_insn(int t,int insn_cnt,struct bpf_verifier_env * env)9293 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9294 {
9295 struct bpf_insn *insns = env->prog->insnsi;
9296 int ret;
9297
9298 if (bpf_pseudo_func(insns + t))
9299 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9300
9301 /* All non-branch instructions have a single fall-through edge. */
9302 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9303 BPF_CLASS(insns[t].code) != BPF_JMP32)
9304 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9305
9306 switch (BPF_OP(insns[t].code)) {
9307 case BPF_EXIT:
9308 return DONE_EXPLORING;
9309
9310 case BPF_CALL:
9311 return visit_func_call_insn(t, insn_cnt, insns, env,
9312 insns[t].src_reg == BPF_PSEUDO_CALL);
9313
9314 case BPF_JA:
9315 if (BPF_SRC(insns[t].code) != BPF_K)
9316 return -EINVAL;
9317
9318 /* unconditional jump with single edge */
9319 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9320 true);
9321 if (ret)
9322 return ret;
9323
9324 /* unconditional jmp is not a good pruning point,
9325 * but it's marked, since backtracking needs
9326 * to record jmp history in is_state_visited().
9327 */
9328 init_explored_state(env, t + insns[t].off + 1);
9329 /* tell verifier to check for equivalent states
9330 * after every call and jump
9331 */
9332 if (t + 1 < insn_cnt)
9333 init_explored_state(env, t + 1);
9334
9335 return ret;
9336
9337 default:
9338 /* conditional jump with two edges */
9339 init_explored_state(env, t);
9340 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9341 if (ret)
9342 return ret;
9343
9344 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9345 }
9346 }
9347
9348 /* non-recursive depth-first-search to detect loops in BPF program
9349 * loop == back-edge in directed graph
9350 */
check_cfg(struct bpf_verifier_env * env)9351 static int check_cfg(struct bpf_verifier_env *env)
9352 {
9353 int insn_cnt = env->prog->len;
9354 int *insn_stack, *insn_state;
9355 int ret = 0;
9356 int i;
9357
9358 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9359 if (!insn_state)
9360 return -ENOMEM;
9361
9362 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9363 if (!insn_stack) {
9364 kvfree(insn_state);
9365 return -ENOMEM;
9366 }
9367
9368 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9369 insn_stack[0] = 0; /* 0 is the first instruction */
9370 env->cfg.cur_stack = 1;
9371
9372 while (env->cfg.cur_stack > 0) {
9373 int t = insn_stack[env->cfg.cur_stack - 1];
9374
9375 ret = visit_insn(t, insn_cnt, env);
9376 switch (ret) {
9377 case DONE_EXPLORING:
9378 insn_state[t] = EXPLORED;
9379 env->cfg.cur_stack--;
9380 break;
9381 case KEEP_EXPLORING:
9382 break;
9383 default:
9384 if (ret > 0) {
9385 verbose(env, "visit_insn internal bug\n");
9386 ret = -EFAULT;
9387 }
9388 goto err_free;
9389 }
9390 }
9391
9392 if (env->cfg.cur_stack < 0) {
9393 verbose(env, "pop stack internal bug\n");
9394 ret = -EFAULT;
9395 goto err_free;
9396 }
9397
9398 for (i = 0; i < insn_cnt; i++) {
9399 if (insn_state[i] != EXPLORED) {
9400 verbose(env, "unreachable insn %d\n", i);
9401 ret = -EINVAL;
9402 goto err_free;
9403 }
9404 }
9405 ret = 0; /* cfg looks good */
9406
9407 err_free:
9408 kvfree(insn_state);
9409 kvfree(insn_stack);
9410 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9411 return ret;
9412 }
9413
check_abnormal_return(struct bpf_verifier_env * env)9414 static int check_abnormal_return(struct bpf_verifier_env *env)
9415 {
9416 int i;
9417
9418 for (i = 1; i < env->subprog_cnt; i++) {
9419 if (env->subprog_info[i].has_ld_abs) {
9420 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
9421 return -EINVAL;
9422 }
9423 if (env->subprog_info[i].has_tail_call) {
9424 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
9425 return -EINVAL;
9426 }
9427 }
9428 return 0;
9429 }
9430
9431 /* The minimum supported BTF func info size */
9432 #define MIN_BPF_FUNCINFO_SIZE 8
9433 #define MAX_FUNCINFO_REC_SIZE 252
9434
check_btf_func(struct bpf_verifier_env * env,const union bpf_attr * attr,union bpf_attr __user * uattr)9435 static int check_btf_func(struct bpf_verifier_env *env,
9436 const union bpf_attr *attr,
9437 union bpf_attr __user *uattr)
9438 {
9439 const struct btf_type *type, *func_proto, *ret_type;
9440 u32 i, nfuncs, urec_size, min_size;
9441 u32 krec_size = sizeof(struct bpf_func_info);
9442 struct bpf_func_info *krecord;
9443 struct bpf_func_info_aux *info_aux = NULL;
9444 struct bpf_prog *prog;
9445 const struct btf *btf;
9446 void __user *urecord;
9447 u32 prev_offset = 0;
9448 bool scalar_return;
9449 int ret = -ENOMEM;
9450
9451 nfuncs = attr->func_info_cnt;
9452 if (!nfuncs) {
9453 if (check_abnormal_return(env))
9454 return -EINVAL;
9455 return 0;
9456 }
9457
9458 if (nfuncs != env->subprog_cnt) {
9459 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
9460 return -EINVAL;
9461 }
9462
9463 urec_size = attr->func_info_rec_size;
9464 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
9465 urec_size > MAX_FUNCINFO_REC_SIZE ||
9466 urec_size % sizeof(u32)) {
9467 verbose(env, "invalid func info rec size %u\n", urec_size);
9468 return -EINVAL;
9469 }
9470
9471 prog = env->prog;
9472 btf = prog->aux->btf;
9473
9474 urecord = u64_to_user_ptr(attr->func_info);
9475 min_size = min_t(u32, krec_size, urec_size);
9476
9477 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
9478 if (!krecord)
9479 return -ENOMEM;
9480 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
9481 if (!info_aux)
9482 goto err_free;
9483
9484 for (i = 0; i < nfuncs; i++) {
9485 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
9486 if (ret) {
9487 if (ret == -E2BIG) {
9488 verbose(env, "nonzero tailing record in func info");
9489 /* set the size kernel expects so loader can zero
9490 * out the rest of the record.
9491 */
9492 if (put_user(min_size, &uattr->func_info_rec_size))
9493 ret = -EFAULT;
9494 }
9495 goto err_free;
9496 }
9497
9498 if (copy_from_user(&krecord[i], urecord, min_size)) {
9499 ret = -EFAULT;
9500 goto err_free;
9501 }
9502
9503 /* check insn_off */
9504 ret = -EINVAL;
9505 if (i == 0) {
9506 if (krecord[i].insn_off) {
9507 verbose(env,
9508 "nonzero insn_off %u for the first func info record",
9509 krecord[i].insn_off);
9510 goto err_free;
9511 }
9512 } else if (krecord[i].insn_off <= prev_offset) {
9513 verbose(env,
9514 "same or smaller insn offset (%u) than previous func info record (%u)",
9515 krecord[i].insn_off, prev_offset);
9516 goto err_free;
9517 }
9518
9519 if (env->subprog_info[i].start != krecord[i].insn_off) {
9520 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
9521 goto err_free;
9522 }
9523
9524 /* check type_id */
9525 type = btf_type_by_id(btf, krecord[i].type_id);
9526 if (!type || !btf_type_is_func(type)) {
9527 verbose(env, "invalid type id %d in func info",
9528 krecord[i].type_id);
9529 goto err_free;
9530 }
9531 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
9532
9533 func_proto = btf_type_by_id(btf, type->type);
9534 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
9535 /* btf_func_check() already verified it during BTF load */
9536 goto err_free;
9537 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
9538 scalar_return =
9539 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
9540 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
9541 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
9542 goto err_free;
9543 }
9544 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
9545 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
9546 goto err_free;
9547 }
9548
9549 prev_offset = krecord[i].insn_off;
9550 urecord += urec_size;
9551 }
9552
9553 prog->aux->func_info = krecord;
9554 prog->aux->func_info_cnt = nfuncs;
9555 prog->aux->func_info_aux = info_aux;
9556 return 0;
9557
9558 err_free:
9559 kvfree(krecord);
9560 kfree(info_aux);
9561 return ret;
9562 }
9563
adjust_btf_func(struct bpf_verifier_env * env)9564 static void adjust_btf_func(struct bpf_verifier_env *env)
9565 {
9566 struct bpf_prog_aux *aux = env->prog->aux;
9567 int i;
9568
9569 if (!aux->func_info)
9570 return;
9571
9572 for (i = 0; i < env->subprog_cnt; i++)
9573 aux->func_info[i].insn_off = env->subprog_info[i].start;
9574 }
9575
9576 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9577 sizeof(((struct bpf_line_info *)(0))->line_col))
9578 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9579
check_btf_line(struct bpf_verifier_env * env,const union bpf_attr * attr,union bpf_attr __user * uattr)9580 static int check_btf_line(struct bpf_verifier_env *env,
9581 const union bpf_attr *attr,
9582 union bpf_attr __user *uattr)
9583 {
9584 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9585 struct bpf_subprog_info *sub;
9586 struct bpf_line_info *linfo;
9587 struct bpf_prog *prog;
9588 const struct btf *btf;
9589 void __user *ulinfo;
9590 int err;
9591
9592 nr_linfo = attr->line_info_cnt;
9593 if (!nr_linfo)
9594 return 0;
9595
9596 rec_size = attr->line_info_rec_size;
9597 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9598 rec_size > MAX_LINEINFO_REC_SIZE ||
9599 rec_size & (sizeof(u32) - 1))
9600 return -EINVAL;
9601
9602 /* Need to zero it in case the userspace may
9603 * pass in a smaller bpf_line_info object.
9604 */
9605 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9606 GFP_KERNEL | __GFP_NOWARN);
9607 if (!linfo)
9608 return -ENOMEM;
9609
9610 prog = env->prog;
9611 btf = prog->aux->btf;
9612
9613 s = 0;
9614 sub = env->subprog_info;
9615 ulinfo = u64_to_user_ptr(attr->line_info);
9616 expected_size = sizeof(struct bpf_line_info);
9617 ncopy = min_t(u32, expected_size, rec_size);
9618 for (i = 0; i < nr_linfo; i++) {
9619 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
9620 if (err) {
9621 if (err == -E2BIG) {
9622 verbose(env, "nonzero tailing record in line_info");
9623 if (put_user(expected_size,
9624 &uattr->line_info_rec_size))
9625 err = -EFAULT;
9626 }
9627 goto err_free;
9628 }
9629
9630 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
9631 err = -EFAULT;
9632 goto err_free;
9633 }
9634
9635 /*
9636 * Check insn_off to ensure
9637 * 1) strictly increasing AND
9638 * 2) bounded by prog->len
9639 *
9640 * The linfo[0].insn_off == 0 check logically falls into
9641 * the later "missing bpf_line_info for func..." case
9642 * because the first linfo[0].insn_off must be the
9643 * first sub also and the first sub must have
9644 * subprog_info[0].start == 0.
9645 */
9646 if ((i && linfo[i].insn_off <= prev_offset) ||
9647 linfo[i].insn_off >= prog->len) {
9648 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
9649 i, linfo[i].insn_off, prev_offset,
9650 prog->len);
9651 err = -EINVAL;
9652 goto err_free;
9653 }
9654
9655 if (!prog->insnsi[linfo[i].insn_off].code) {
9656 verbose(env,
9657 "Invalid insn code at line_info[%u].insn_off\n",
9658 i);
9659 err = -EINVAL;
9660 goto err_free;
9661 }
9662
9663 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
9664 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
9665 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
9666 err = -EINVAL;
9667 goto err_free;
9668 }
9669
9670 if (s != env->subprog_cnt) {
9671 if (linfo[i].insn_off == sub[s].start) {
9672 sub[s].linfo_idx = i;
9673 s++;
9674 } else if (sub[s].start < linfo[i].insn_off) {
9675 verbose(env, "missing bpf_line_info for func#%u\n", s);
9676 err = -EINVAL;
9677 goto err_free;
9678 }
9679 }
9680
9681 prev_offset = linfo[i].insn_off;
9682 ulinfo += rec_size;
9683 }
9684
9685 if (s != env->subprog_cnt) {
9686 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
9687 env->subprog_cnt - s, s);
9688 err = -EINVAL;
9689 goto err_free;
9690 }
9691
9692 prog->aux->linfo = linfo;
9693 prog->aux->nr_linfo = nr_linfo;
9694
9695 return 0;
9696
9697 err_free:
9698 kvfree(linfo);
9699 return err;
9700 }
9701
check_btf_info(struct bpf_verifier_env * env,const union bpf_attr * attr,union bpf_attr __user * uattr)9702 static int check_btf_info(struct bpf_verifier_env *env,
9703 const union bpf_attr *attr,
9704 union bpf_attr __user *uattr)
9705 {
9706 struct btf *btf;
9707 int err;
9708
9709 if (!attr->func_info_cnt && !attr->line_info_cnt) {
9710 if (check_abnormal_return(env))
9711 return -EINVAL;
9712 return 0;
9713 }
9714
9715 btf = btf_get_by_fd(attr->prog_btf_fd);
9716 if (IS_ERR(btf))
9717 return PTR_ERR(btf);
9718 if (btf_is_kernel(btf)) {
9719 btf_put(btf);
9720 return -EACCES;
9721 }
9722 env->prog->aux->btf = btf;
9723
9724 err = check_btf_func(env, attr, uattr);
9725 if (err)
9726 return err;
9727
9728 err = check_btf_line(env, attr, uattr);
9729 if (err)
9730 return err;
9731
9732 return 0;
9733 }
9734
9735 /* check %cur's range satisfies %old's */
range_within(struct bpf_reg_state * old,struct bpf_reg_state * cur)9736 static bool range_within(struct bpf_reg_state *old,
9737 struct bpf_reg_state *cur)
9738 {
9739 return old->umin_value <= cur->umin_value &&
9740 old->umax_value >= cur->umax_value &&
9741 old->smin_value <= cur->smin_value &&
9742 old->smax_value >= cur->smax_value &&
9743 old->u32_min_value <= cur->u32_min_value &&
9744 old->u32_max_value >= cur->u32_max_value &&
9745 old->s32_min_value <= cur->s32_min_value &&
9746 old->s32_max_value >= cur->s32_max_value;
9747 }
9748
9749 /* Maximum number of register states that can exist at once */
9750 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
9751 struct idpair {
9752 u32 old;
9753 u32 cur;
9754 };
9755
9756 /* If in the old state two registers had the same id, then they need to have
9757 * the same id in the new state as well. But that id could be different from
9758 * the old state, so we need to track the mapping from old to new ids.
9759 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
9760 * regs with old id 5 must also have new id 9 for the new state to be safe. But
9761 * regs with a different old id could still have new id 9, we don't care about
9762 * that.
9763 * So we look through our idmap to see if this old id has been seen before. If
9764 * so, we require the new id to match; otherwise, we add the id pair to the map.
9765 */
check_ids(u32 old_id,u32 cur_id,struct idpair * idmap)9766 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
9767 {
9768 unsigned int i;
9769
9770 for (i = 0; i < ID_MAP_SIZE; i++) {
9771 if (!idmap[i].old) {
9772 /* Reached an empty slot; haven't seen this id before */
9773 idmap[i].old = old_id;
9774 idmap[i].cur = cur_id;
9775 return true;
9776 }
9777 if (idmap[i].old == old_id)
9778 return idmap[i].cur == cur_id;
9779 }
9780 /* We ran out of idmap slots, which should be impossible */
9781 WARN_ON_ONCE(1);
9782 return false;
9783 }
9784
clean_func_state(struct bpf_verifier_env * env,struct bpf_func_state * st)9785 static void clean_func_state(struct bpf_verifier_env *env,
9786 struct bpf_func_state *st)
9787 {
9788 enum bpf_reg_liveness live;
9789 int i, j;
9790
9791 for (i = 0; i < BPF_REG_FP; i++) {
9792 live = st->regs[i].live;
9793 /* liveness must not touch this register anymore */
9794 st->regs[i].live |= REG_LIVE_DONE;
9795 if (!(live & REG_LIVE_READ))
9796 /* since the register is unused, clear its state
9797 * to make further comparison simpler
9798 */
9799 __mark_reg_not_init(env, &st->regs[i]);
9800 }
9801
9802 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
9803 live = st->stack[i].spilled_ptr.live;
9804 /* liveness must not touch this stack slot anymore */
9805 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
9806 if (!(live & REG_LIVE_READ)) {
9807 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
9808 for (j = 0; j < BPF_REG_SIZE; j++)
9809 st->stack[i].slot_type[j] = STACK_INVALID;
9810 }
9811 }
9812 }
9813
clean_verifier_state(struct bpf_verifier_env * env,struct bpf_verifier_state * st)9814 static void clean_verifier_state(struct bpf_verifier_env *env,
9815 struct bpf_verifier_state *st)
9816 {
9817 int i;
9818
9819 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
9820 /* all regs in this state in all frames were already marked */
9821 return;
9822
9823 for (i = 0; i <= st->curframe; i++)
9824 clean_func_state(env, st->frame[i]);
9825 }
9826
9827 /* the parentage chains form a tree.
9828 * the verifier states are added to state lists at given insn and
9829 * pushed into state stack for future exploration.
9830 * when the verifier reaches bpf_exit insn some of the verifer states
9831 * stored in the state lists have their final liveness state already,
9832 * but a lot of states will get revised from liveness point of view when
9833 * the verifier explores other branches.
9834 * Example:
9835 * 1: r0 = 1
9836 * 2: if r1 == 100 goto pc+1
9837 * 3: r0 = 2
9838 * 4: exit
9839 * when the verifier reaches exit insn the register r0 in the state list of
9840 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
9841 * of insn 2 and goes exploring further. At the insn 4 it will walk the
9842 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
9843 *
9844 * Since the verifier pushes the branch states as it sees them while exploring
9845 * the program the condition of walking the branch instruction for the second
9846 * time means that all states below this branch were already explored and
9847 * their final liveness markes are already propagated.
9848 * Hence when the verifier completes the search of state list in is_state_visited()
9849 * we can call this clean_live_states() function to mark all liveness states
9850 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
9851 * will not be used.
9852 * This function also clears the registers and stack for states that !READ
9853 * to simplify state merging.
9854 *
9855 * Important note here that walking the same branch instruction in the callee
9856 * doesn't meant that the states are DONE. The verifier has to compare
9857 * the callsites
9858 */
clean_live_states(struct bpf_verifier_env * env,int insn,struct bpf_verifier_state * cur)9859 static void clean_live_states(struct bpf_verifier_env *env, int insn,
9860 struct bpf_verifier_state *cur)
9861 {
9862 struct bpf_verifier_state_list *sl;
9863 int i;
9864
9865 sl = *explored_state(env, insn);
9866 while (sl) {
9867 if (sl->state.branches)
9868 goto next;
9869 if (sl->state.insn_idx != insn ||
9870 sl->state.curframe != cur->curframe)
9871 goto next;
9872 for (i = 0; i <= cur->curframe; i++)
9873 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
9874 goto next;
9875 clean_verifier_state(env, &sl->state);
9876 next:
9877 sl = sl->next;
9878 }
9879 }
9880
9881 /* Returns true if (rold safe implies rcur safe) */
regsafe(struct bpf_reg_state * rold,struct bpf_reg_state * rcur,struct idpair * idmap)9882 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
9883 struct idpair *idmap)
9884 {
9885 bool equal;
9886
9887 if (!(rold->live & REG_LIVE_READ))
9888 /* explored state didn't use this */
9889 return true;
9890
9891 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
9892
9893 if (rold->type == PTR_TO_STACK)
9894 /* two stack pointers are equal only if they're pointing to
9895 * the same stack frame, since fp-8 in foo != fp-8 in bar
9896 */
9897 return equal && rold->frameno == rcur->frameno;
9898
9899 if (equal)
9900 return true;
9901
9902 if (rold->type == NOT_INIT)
9903 /* explored state can't have used this */
9904 return true;
9905 if (rcur->type == NOT_INIT)
9906 return false;
9907 switch (rold->type) {
9908 case SCALAR_VALUE:
9909 if (rcur->type == SCALAR_VALUE) {
9910 if (!rold->precise && !rcur->precise)
9911 return true;
9912 /* new val must satisfy old val knowledge */
9913 return range_within(rold, rcur) &&
9914 tnum_in(rold->var_off, rcur->var_off);
9915 } else {
9916 /* We're trying to use a pointer in place of a scalar.
9917 * Even if the scalar was unbounded, this could lead to
9918 * pointer leaks because scalars are allowed to leak
9919 * while pointers are not. We could make this safe in
9920 * special cases if root is calling us, but it's
9921 * probably not worth the hassle.
9922 */
9923 return false;
9924 }
9925 case PTR_TO_MAP_KEY:
9926 case PTR_TO_MAP_VALUE:
9927 /* If the new min/max/var_off satisfy the old ones and
9928 * everything else matches, we are OK.
9929 * 'id' is not compared, since it's only used for maps with
9930 * bpf_spin_lock inside map element and in such cases if
9931 * the rest of the prog is valid for one map element then
9932 * it's valid for all map elements regardless of the key
9933 * used in bpf_map_lookup()
9934 */
9935 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
9936 range_within(rold, rcur) &&
9937 tnum_in(rold->var_off, rcur->var_off);
9938 case PTR_TO_MAP_VALUE_OR_NULL:
9939 /* a PTR_TO_MAP_VALUE could be safe to use as a
9940 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
9941 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
9942 * checked, doing so could have affected others with the same
9943 * id, and we can't check for that because we lost the id when
9944 * we converted to a PTR_TO_MAP_VALUE.
9945 */
9946 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
9947 return false;
9948 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
9949 return false;
9950 /* Check our ids match any regs they're supposed to */
9951 return check_ids(rold->id, rcur->id, idmap);
9952 case PTR_TO_PACKET_META:
9953 case PTR_TO_PACKET:
9954 if (rcur->type != rold->type)
9955 return false;
9956 /* We must have at least as much range as the old ptr
9957 * did, so that any accesses which were safe before are
9958 * still safe. This is true even if old range < old off,
9959 * since someone could have accessed through (ptr - k), or
9960 * even done ptr -= k in a register, to get a safe access.
9961 */
9962 if (rold->range > rcur->range)
9963 return false;
9964 /* If the offsets don't match, we can't trust our alignment;
9965 * nor can we be sure that we won't fall out of range.
9966 */
9967 if (rold->off != rcur->off)
9968 return false;
9969 /* id relations must be preserved */
9970 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
9971 return false;
9972 /* new val must satisfy old val knowledge */
9973 return range_within(rold, rcur) &&
9974 tnum_in(rold->var_off, rcur->var_off);
9975 case PTR_TO_CTX:
9976 case CONST_PTR_TO_MAP:
9977 case PTR_TO_PACKET_END:
9978 case PTR_TO_FLOW_KEYS:
9979 case PTR_TO_SOCKET:
9980 case PTR_TO_SOCKET_OR_NULL:
9981 case PTR_TO_SOCK_COMMON:
9982 case PTR_TO_SOCK_COMMON_OR_NULL:
9983 case PTR_TO_TCP_SOCK:
9984 case PTR_TO_TCP_SOCK_OR_NULL:
9985 case PTR_TO_XDP_SOCK:
9986 /* Only valid matches are exact, which memcmp() above
9987 * would have accepted
9988 */
9989 default:
9990 /* Don't know what's going on, just say it's not safe */
9991 return false;
9992 }
9993
9994 /* Shouldn't get here; if we do, say it's not safe */
9995 WARN_ON_ONCE(1);
9996 return false;
9997 }
9998
stacksafe(struct bpf_func_state * old,struct bpf_func_state * cur,struct idpair * idmap)9999 static bool stacksafe(struct bpf_func_state *old,
10000 struct bpf_func_state *cur,
10001 struct idpair *idmap)
10002 {
10003 int i, spi;
10004
10005 /* walk slots of the explored stack and ignore any additional
10006 * slots in the current stack, since explored(safe) state
10007 * didn't use them
10008 */
10009 for (i = 0; i < old->allocated_stack; i++) {
10010 spi = i / BPF_REG_SIZE;
10011
10012 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10013 i += BPF_REG_SIZE - 1;
10014 /* explored state didn't use this */
10015 continue;
10016 }
10017
10018 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10019 continue;
10020
10021 /* explored stack has more populated slots than current stack
10022 * and these slots were used
10023 */
10024 if (i >= cur->allocated_stack)
10025 return false;
10026
10027 /* if old state was safe with misc data in the stack
10028 * it will be safe with zero-initialized stack.
10029 * The opposite is not true
10030 */
10031 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10032 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10033 continue;
10034 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10035 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10036 /* Ex: old explored (safe) state has STACK_SPILL in
10037 * this stack slot, but current has STACK_MISC ->
10038 * this verifier states are not equivalent,
10039 * return false to continue verification of this path
10040 */
10041 return false;
10042 if (i % BPF_REG_SIZE)
10043 continue;
10044 if (old->stack[spi].slot_type[0] != STACK_SPILL)
10045 continue;
10046 if (!regsafe(&old->stack[spi].spilled_ptr,
10047 &cur->stack[spi].spilled_ptr,
10048 idmap))
10049 /* when explored and current stack slot are both storing
10050 * spilled registers, check that stored pointers types
10051 * are the same as well.
10052 * Ex: explored safe path could have stored
10053 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10054 * but current path has stored:
10055 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10056 * such verifier states are not equivalent.
10057 * return false to continue verification of this path
10058 */
10059 return false;
10060 }
10061 return true;
10062 }
10063
refsafe(struct bpf_func_state * old,struct bpf_func_state * cur)10064 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10065 {
10066 if (old->acquired_refs != cur->acquired_refs)
10067 return false;
10068 return !memcmp(old->refs, cur->refs,
10069 sizeof(*old->refs) * old->acquired_refs);
10070 }
10071
10072 /* compare two verifier states
10073 *
10074 * all states stored in state_list are known to be valid, since
10075 * verifier reached 'bpf_exit' instruction through them
10076 *
10077 * this function is called when verifier exploring different branches of
10078 * execution popped from the state stack. If it sees an old state that has
10079 * more strict register state and more strict stack state then this execution
10080 * branch doesn't need to be explored further, since verifier already
10081 * concluded that more strict state leads to valid finish.
10082 *
10083 * Therefore two states are equivalent if register state is more conservative
10084 * and explored stack state is more conservative than the current one.
10085 * Example:
10086 * explored current
10087 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10088 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10089 *
10090 * In other words if current stack state (one being explored) has more
10091 * valid slots than old one that already passed validation, it means
10092 * the verifier can stop exploring and conclude that current state is valid too
10093 *
10094 * Similarly with registers. If explored state has register type as invalid
10095 * whereas register type in current state is meaningful, it means that
10096 * the current state will reach 'bpf_exit' instruction safely
10097 */
func_states_equal(struct bpf_func_state * old,struct bpf_func_state * cur)10098 static bool func_states_equal(struct bpf_func_state *old,
10099 struct bpf_func_state *cur)
10100 {
10101 struct idpair *idmap;
10102 bool ret = false;
10103 int i;
10104
10105 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
10106 /* If we failed to allocate the idmap, just say it's not safe */
10107 if (!idmap)
10108 return false;
10109
10110 for (i = 0; i < MAX_BPF_REG; i++) {
10111 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
10112 goto out_free;
10113 }
10114
10115 if (!stacksafe(old, cur, idmap))
10116 goto out_free;
10117
10118 if (!refsafe(old, cur))
10119 goto out_free;
10120 ret = true;
10121 out_free:
10122 kfree(idmap);
10123 return ret;
10124 }
10125
states_equal(struct bpf_verifier_env * env,struct bpf_verifier_state * old,struct bpf_verifier_state * cur)10126 static bool states_equal(struct bpf_verifier_env *env,
10127 struct bpf_verifier_state *old,
10128 struct bpf_verifier_state *cur)
10129 {
10130 int i;
10131
10132 if (old->curframe != cur->curframe)
10133 return false;
10134
10135 /* Verification state from speculative execution simulation
10136 * must never prune a non-speculative execution one.
10137 */
10138 if (old->speculative && !cur->speculative)
10139 return false;
10140
10141 if (old->active_spin_lock != cur->active_spin_lock)
10142 return false;
10143
10144 /* for states to be equal callsites have to be the same
10145 * and all frame states need to be equivalent
10146 */
10147 for (i = 0; i <= old->curframe; i++) {
10148 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10149 return false;
10150 if (!func_states_equal(old->frame[i], cur->frame[i]))
10151 return false;
10152 }
10153 return true;
10154 }
10155
10156 /* Return 0 if no propagation happened. Return negative error code if error
10157 * happened. Otherwise, return the propagated bit.
10158 */
propagate_liveness_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct bpf_reg_state * parent_reg)10159 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10160 struct bpf_reg_state *reg,
10161 struct bpf_reg_state *parent_reg)
10162 {
10163 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10164 u8 flag = reg->live & REG_LIVE_READ;
10165 int err;
10166
10167 /* When comes here, read flags of PARENT_REG or REG could be any of
10168 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10169 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10170 */
10171 if (parent_flag == REG_LIVE_READ64 ||
10172 /* Or if there is no read flag from REG. */
10173 !flag ||
10174 /* Or if the read flag from REG is the same as PARENT_REG. */
10175 parent_flag == flag)
10176 return 0;
10177
10178 err = mark_reg_read(env, reg, parent_reg, flag);
10179 if (err)
10180 return err;
10181
10182 return flag;
10183 }
10184
10185 /* A write screens off any subsequent reads; but write marks come from the
10186 * straight-line code between a state and its parent. When we arrive at an
10187 * equivalent state (jump target or such) we didn't arrive by the straight-line
10188 * code, so read marks in the state must propagate to the parent regardless
10189 * of the state's write marks. That's what 'parent == state->parent' comparison
10190 * in mark_reg_read() is for.
10191 */
propagate_liveness(struct bpf_verifier_env * env,const struct bpf_verifier_state * vstate,struct bpf_verifier_state * vparent)10192 static int propagate_liveness(struct bpf_verifier_env *env,
10193 const struct bpf_verifier_state *vstate,
10194 struct bpf_verifier_state *vparent)
10195 {
10196 struct bpf_reg_state *state_reg, *parent_reg;
10197 struct bpf_func_state *state, *parent;
10198 int i, frame, err = 0;
10199
10200 if (vparent->curframe != vstate->curframe) {
10201 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10202 vparent->curframe, vstate->curframe);
10203 return -EFAULT;
10204 }
10205 /* Propagate read liveness of registers... */
10206 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10207 for (frame = 0; frame <= vstate->curframe; frame++) {
10208 parent = vparent->frame[frame];
10209 state = vstate->frame[frame];
10210 parent_reg = parent->regs;
10211 state_reg = state->regs;
10212 /* We don't need to worry about FP liveness, it's read-only */
10213 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10214 err = propagate_liveness_reg(env, &state_reg[i],
10215 &parent_reg[i]);
10216 if (err < 0)
10217 return err;
10218 if (err == REG_LIVE_READ64)
10219 mark_insn_zext(env, &parent_reg[i]);
10220 }
10221
10222 /* Propagate stack slots. */
10223 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10224 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10225 parent_reg = &parent->stack[i].spilled_ptr;
10226 state_reg = &state->stack[i].spilled_ptr;
10227 err = propagate_liveness_reg(env, state_reg,
10228 parent_reg);
10229 if (err < 0)
10230 return err;
10231 }
10232 }
10233 return 0;
10234 }
10235
10236 /* find precise scalars in the previous equivalent state and
10237 * propagate them into the current state
10238 */
propagate_precision(struct bpf_verifier_env * env,const struct bpf_verifier_state * old)10239 static int propagate_precision(struct bpf_verifier_env *env,
10240 const struct bpf_verifier_state *old)
10241 {
10242 struct bpf_reg_state *state_reg;
10243 struct bpf_func_state *state;
10244 int i, err = 0;
10245
10246 state = old->frame[old->curframe];
10247 state_reg = state->regs;
10248 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10249 if (state_reg->type != SCALAR_VALUE ||
10250 !state_reg->precise)
10251 continue;
10252 if (env->log.level & BPF_LOG_LEVEL2)
10253 verbose(env, "propagating r%d\n", i);
10254 err = mark_chain_precision(env, i);
10255 if (err < 0)
10256 return err;
10257 }
10258
10259 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10260 if (state->stack[i].slot_type[0] != STACK_SPILL)
10261 continue;
10262 state_reg = &state->stack[i].spilled_ptr;
10263 if (state_reg->type != SCALAR_VALUE ||
10264 !state_reg->precise)
10265 continue;
10266 if (env->log.level & BPF_LOG_LEVEL2)
10267 verbose(env, "propagating fp%d\n",
10268 (-i - 1) * BPF_REG_SIZE);
10269 err = mark_chain_precision_stack(env, i);
10270 if (err < 0)
10271 return err;
10272 }
10273 return 0;
10274 }
10275
states_maybe_looping(struct bpf_verifier_state * old,struct bpf_verifier_state * cur)10276 static bool states_maybe_looping(struct bpf_verifier_state *old,
10277 struct bpf_verifier_state *cur)
10278 {
10279 struct bpf_func_state *fold, *fcur;
10280 int i, fr = cur->curframe;
10281
10282 if (old->curframe != fr)
10283 return false;
10284
10285 fold = old->frame[fr];
10286 fcur = cur->frame[fr];
10287 for (i = 0; i < MAX_BPF_REG; i++)
10288 if (memcmp(&fold->regs[i], &fcur->regs[i],
10289 offsetof(struct bpf_reg_state, parent)))
10290 return false;
10291 return true;
10292 }
10293
10294
is_state_visited(struct bpf_verifier_env * env,int insn_idx)10295 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10296 {
10297 struct bpf_verifier_state_list *new_sl;
10298 struct bpf_verifier_state_list *sl, **pprev;
10299 struct bpf_verifier_state *cur = env->cur_state, *new;
10300 int i, j, err, states_cnt = 0;
10301 bool add_new_state = env->test_state_freq ? true : false;
10302
10303 cur->last_insn_idx = env->prev_insn_idx;
10304 if (!env->insn_aux_data[insn_idx].prune_point)
10305 /* this 'insn_idx' instruction wasn't marked, so we will not
10306 * be doing state search here
10307 */
10308 return 0;
10309
10310 /* bpf progs typically have pruning point every 4 instructions
10311 * http://vger.kernel.org/bpfconf2019.html#session-1
10312 * Do not add new state for future pruning if the verifier hasn't seen
10313 * at least 2 jumps and at least 8 instructions.
10314 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10315 * In tests that amounts to up to 50% reduction into total verifier
10316 * memory consumption and 20% verifier time speedup.
10317 */
10318 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10319 env->insn_processed - env->prev_insn_processed >= 8)
10320 add_new_state = true;
10321
10322 pprev = explored_state(env, insn_idx);
10323 sl = *pprev;
10324
10325 clean_live_states(env, insn_idx, cur);
10326
10327 while (sl) {
10328 states_cnt++;
10329 if (sl->state.insn_idx != insn_idx)
10330 goto next;
10331 if (sl->state.branches) {
10332 if (states_maybe_looping(&sl->state, cur) &&
10333 states_equal(env, &sl->state, cur)) {
10334 verbose_linfo(env, insn_idx, "; ");
10335 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10336 return -EINVAL;
10337 }
10338 /* if the verifier is processing a loop, avoid adding new state
10339 * too often, since different loop iterations have distinct
10340 * states and may not help future pruning.
10341 * This threshold shouldn't be too low to make sure that
10342 * a loop with large bound will be rejected quickly.
10343 * The most abusive loop will be:
10344 * r1 += 1
10345 * if r1 < 1000000 goto pc-2
10346 * 1M insn_procssed limit / 100 == 10k peak states.
10347 * This threshold shouldn't be too high either, since states
10348 * at the end of the loop are likely to be useful in pruning.
10349 */
10350 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10351 env->insn_processed - env->prev_insn_processed < 100)
10352 add_new_state = false;
10353 goto miss;
10354 }
10355 if (states_equal(env, &sl->state, cur)) {
10356 sl->hit_cnt++;
10357 /* reached equivalent register/stack state,
10358 * prune the search.
10359 * Registers read by the continuation are read by us.
10360 * If we have any write marks in env->cur_state, they
10361 * will prevent corresponding reads in the continuation
10362 * from reaching our parent (an explored_state). Our
10363 * own state will get the read marks recorded, but
10364 * they'll be immediately forgotten as we're pruning
10365 * this state and will pop a new one.
10366 */
10367 err = propagate_liveness(env, &sl->state, cur);
10368
10369 /* if previous state reached the exit with precision and
10370 * current state is equivalent to it (except precsion marks)
10371 * the precision needs to be propagated back in
10372 * the current state.
10373 */
10374 err = err ? : push_jmp_history(env, cur);
10375 err = err ? : propagate_precision(env, &sl->state);
10376 if (err)
10377 return err;
10378 return 1;
10379 }
10380 miss:
10381 /* when new state is not going to be added do not increase miss count.
10382 * Otherwise several loop iterations will remove the state
10383 * recorded earlier. The goal of these heuristics is to have
10384 * states from some iterations of the loop (some in the beginning
10385 * and some at the end) to help pruning.
10386 */
10387 if (add_new_state)
10388 sl->miss_cnt++;
10389 /* heuristic to determine whether this state is beneficial
10390 * to keep checking from state equivalence point of view.
10391 * Higher numbers increase max_states_per_insn and verification time,
10392 * but do not meaningfully decrease insn_processed.
10393 */
10394 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10395 /* the state is unlikely to be useful. Remove it to
10396 * speed up verification
10397 */
10398 *pprev = sl->next;
10399 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10400 u32 br = sl->state.branches;
10401
10402 WARN_ONCE(br,
10403 "BUG live_done but branches_to_explore %d\n",
10404 br);
10405 free_verifier_state(&sl->state, false);
10406 kfree(sl);
10407 env->peak_states--;
10408 } else {
10409 /* cannot free this state, since parentage chain may
10410 * walk it later. Add it for free_list instead to
10411 * be freed at the end of verification
10412 */
10413 sl->next = env->free_list;
10414 env->free_list = sl;
10415 }
10416 sl = *pprev;
10417 continue;
10418 }
10419 next:
10420 pprev = &sl->next;
10421 sl = *pprev;
10422 }
10423
10424 if (env->max_states_per_insn < states_cnt)
10425 env->max_states_per_insn = states_cnt;
10426
10427 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10428 return push_jmp_history(env, cur);
10429
10430 if (!add_new_state)
10431 return push_jmp_history(env, cur);
10432
10433 /* There were no equivalent states, remember the current one.
10434 * Technically the current state is not proven to be safe yet,
10435 * but it will either reach outer most bpf_exit (which means it's safe)
10436 * or it will be rejected. When there are no loops the verifier won't be
10437 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10438 * again on the way to bpf_exit.
10439 * When looping the sl->state.branches will be > 0 and this state
10440 * will not be considered for equivalence until branches == 0.
10441 */
10442 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10443 if (!new_sl)
10444 return -ENOMEM;
10445 env->total_states++;
10446 env->peak_states++;
10447 env->prev_jmps_processed = env->jmps_processed;
10448 env->prev_insn_processed = env->insn_processed;
10449
10450 /* add new state to the head of linked list */
10451 new = &new_sl->state;
10452 err = copy_verifier_state(new, cur);
10453 if (err) {
10454 free_verifier_state(new, false);
10455 kfree(new_sl);
10456 return err;
10457 }
10458 new->insn_idx = insn_idx;
10459 WARN_ONCE(new->branches != 1,
10460 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10461
10462 cur->parent = new;
10463 cur->first_insn_idx = insn_idx;
10464 clear_jmp_history(cur);
10465 new_sl->next = *explored_state(env, insn_idx);
10466 *explored_state(env, insn_idx) = new_sl;
10467 /* connect new state to parentage chain. Current frame needs all
10468 * registers connected. Only r6 - r9 of the callers are alive (pushed
10469 * to the stack implicitly by JITs) so in callers' frames connect just
10470 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10471 * the state of the call instruction (with WRITTEN set), and r0 comes
10472 * from callee with its full parentage chain, anyway.
10473 */
10474 /* clear write marks in current state: the writes we did are not writes
10475 * our child did, so they don't screen off its reads from us.
10476 * (There are no read marks in current state, because reads always mark
10477 * their parent and current state never has children yet. Only
10478 * explored_states can get read marks.)
10479 */
10480 for (j = 0; j <= cur->curframe; j++) {
10481 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10482 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10483 for (i = 0; i < BPF_REG_FP; i++)
10484 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10485 }
10486
10487 /* all stack frames are accessible from callee, clear them all */
10488 for (j = 0; j <= cur->curframe; j++) {
10489 struct bpf_func_state *frame = cur->frame[j];
10490 struct bpf_func_state *newframe = new->frame[j];
10491
10492 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10493 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10494 frame->stack[i].spilled_ptr.parent =
10495 &newframe->stack[i].spilled_ptr;
10496 }
10497 }
10498 return 0;
10499 }
10500
10501 /* Return true if it's OK to have the same insn return a different type. */
reg_type_mismatch_ok(enum bpf_reg_type type)10502 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10503 {
10504 switch (type) {
10505 case PTR_TO_CTX:
10506 case PTR_TO_SOCKET:
10507 case PTR_TO_SOCKET_OR_NULL:
10508 case PTR_TO_SOCK_COMMON:
10509 case PTR_TO_SOCK_COMMON_OR_NULL:
10510 case PTR_TO_TCP_SOCK:
10511 case PTR_TO_TCP_SOCK_OR_NULL:
10512 case PTR_TO_XDP_SOCK:
10513 case PTR_TO_BTF_ID:
10514 case PTR_TO_BTF_ID_OR_NULL:
10515 return false;
10516 default:
10517 return true;
10518 }
10519 }
10520
10521 /* If an instruction was previously used with particular pointer types, then we
10522 * need to be careful to avoid cases such as the below, where it may be ok
10523 * for one branch accessing the pointer, but not ok for the other branch:
10524 *
10525 * R1 = sock_ptr
10526 * goto X;
10527 * ...
10528 * R1 = some_other_valid_ptr;
10529 * goto X;
10530 * ...
10531 * R2 = *(u32 *)(R1 + 0);
10532 */
reg_type_mismatch(enum bpf_reg_type src,enum bpf_reg_type prev)10533 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10534 {
10535 return src != prev && (!reg_type_mismatch_ok(src) ||
10536 !reg_type_mismatch_ok(prev));
10537 }
10538
do_check(struct bpf_verifier_env * env)10539 static int do_check(struct bpf_verifier_env *env)
10540 {
10541 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10542 struct bpf_verifier_state *state = env->cur_state;
10543 struct bpf_insn *insns = env->prog->insnsi;
10544 struct bpf_reg_state *regs;
10545 int insn_cnt = env->prog->len;
10546 bool do_print_state = false;
10547 int prev_insn_idx = -1;
10548
10549 for (;;) {
10550 struct bpf_insn *insn;
10551 u8 class;
10552 int err;
10553
10554 env->prev_insn_idx = prev_insn_idx;
10555 if (env->insn_idx >= insn_cnt) {
10556 verbose(env, "invalid insn idx %d insn_cnt %d\n",
10557 env->insn_idx, insn_cnt);
10558 return -EFAULT;
10559 }
10560
10561 insn = &insns[env->insn_idx];
10562 class = BPF_CLASS(insn->code);
10563
10564 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10565 verbose(env,
10566 "BPF program is too large. Processed %d insn\n",
10567 env->insn_processed);
10568 return -E2BIG;
10569 }
10570
10571 err = is_state_visited(env, env->insn_idx);
10572 if (err < 0)
10573 return err;
10574 if (err == 1) {
10575 /* found equivalent state, can prune the search */
10576 if (env->log.level & BPF_LOG_LEVEL) {
10577 if (do_print_state)
10578 verbose(env, "\nfrom %d to %d%s: safe\n",
10579 env->prev_insn_idx, env->insn_idx,
10580 env->cur_state->speculative ?
10581 " (speculative execution)" : "");
10582 else
10583 verbose(env, "%d: safe\n", env->insn_idx);
10584 }
10585 goto process_bpf_exit;
10586 }
10587
10588 if (signal_pending(current))
10589 return -EAGAIN;
10590
10591 if (need_resched())
10592 cond_resched();
10593
10594 if (env->log.level & BPF_LOG_LEVEL2 ||
10595 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10596 if (env->log.level & BPF_LOG_LEVEL2)
10597 verbose(env, "%d:", env->insn_idx);
10598 else
10599 verbose(env, "\nfrom %d to %d%s:",
10600 env->prev_insn_idx, env->insn_idx,
10601 env->cur_state->speculative ?
10602 " (speculative execution)" : "");
10603 print_verifier_state(env, state->frame[state->curframe]);
10604 do_print_state = false;
10605 }
10606
10607 if (env->log.level & BPF_LOG_LEVEL) {
10608 const struct bpf_insn_cbs cbs = {
10609 .cb_call = disasm_kfunc_name,
10610 .cb_print = verbose,
10611 .private_data = env,
10612 };
10613
10614 verbose_linfo(env, env->insn_idx, "; ");
10615 verbose(env, "%d: ", env->insn_idx);
10616 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10617 }
10618
10619 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10620 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10621 env->prev_insn_idx);
10622 if (err)
10623 return err;
10624 }
10625
10626 regs = cur_regs(env);
10627 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
10628 prev_insn_idx = env->insn_idx;
10629
10630 if (class == BPF_ALU || class == BPF_ALU64) {
10631 err = check_alu_op(env, insn);
10632 if (err)
10633 return err;
10634
10635 } else if (class == BPF_LDX) {
10636 enum bpf_reg_type *prev_src_type, src_reg_type;
10637
10638 /* check for reserved fields is already done */
10639
10640 /* check src operand */
10641 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10642 if (err)
10643 return err;
10644
10645 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10646 if (err)
10647 return err;
10648
10649 src_reg_type = regs[insn->src_reg].type;
10650
10651 /* check that memory (src_reg + off) is readable,
10652 * the state of dst_reg will be updated by this func
10653 */
10654 err = check_mem_access(env, env->insn_idx, insn->src_reg,
10655 insn->off, BPF_SIZE(insn->code),
10656 BPF_READ, insn->dst_reg, false);
10657 if (err)
10658 return err;
10659
10660 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10661
10662 if (*prev_src_type == NOT_INIT) {
10663 /* saw a valid insn
10664 * dst_reg = *(u32 *)(src_reg + off)
10665 * save type to validate intersecting paths
10666 */
10667 *prev_src_type = src_reg_type;
10668
10669 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10670 /* ABuser program is trying to use the same insn
10671 * dst_reg = *(u32*) (src_reg + off)
10672 * with different pointer types:
10673 * src_reg == ctx in one branch and
10674 * src_reg == stack|map in some other branch.
10675 * Reject it.
10676 */
10677 verbose(env, "same insn cannot be used with different pointers\n");
10678 return -EINVAL;
10679 }
10680
10681 } else if (class == BPF_STX) {
10682 enum bpf_reg_type *prev_dst_type, dst_reg_type;
10683
10684 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10685 err = check_atomic(env, env->insn_idx, insn);
10686 if (err)
10687 return err;
10688 env->insn_idx++;
10689 continue;
10690 }
10691
10692 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10693 verbose(env, "BPF_STX uses reserved fields\n");
10694 return -EINVAL;
10695 }
10696
10697 /* check src1 operand */
10698 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10699 if (err)
10700 return err;
10701 /* check src2 operand */
10702 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10703 if (err)
10704 return err;
10705
10706 dst_reg_type = regs[insn->dst_reg].type;
10707
10708 /* check that memory (dst_reg + off) is writeable */
10709 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10710 insn->off, BPF_SIZE(insn->code),
10711 BPF_WRITE, insn->src_reg, false);
10712 if (err)
10713 return err;
10714
10715 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10716
10717 if (*prev_dst_type == NOT_INIT) {
10718 *prev_dst_type = dst_reg_type;
10719 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
10720 verbose(env, "same insn cannot be used with different pointers\n");
10721 return -EINVAL;
10722 }
10723
10724 } else if (class == BPF_ST) {
10725 if (BPF_MODE(insn->code) != BPF_MEM ||
10726 insn->src_reg != BPF_REG_0) {
10727 verbose(env, "BPF_ST uses reserved fields\n");
10728 return -EINVAL;
10729 }
10730 /* check src operand */
10731 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10732 if (err)
10733 return err;
10734
10735 if (is_ctx_reg(env, insn->dst_reg)) {
10736 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
10737 insn->dst_reg,
10738 reg_type_str[reg_state(env, insn->dst_reg)->type]);
10739 return -EACCES;
10740 }
10741
10742 /* check that memory (dst_reg + off) is writeable */
10743 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10744 insn->off, BPF_SIZE(insn->code),
10745 BPF_WRITE, -1, false);
10746 if (err)
10747 return err;
10748
10749 } else if (class == BPF_JMP || class == BPF_JMP32) {
10750 u8 opcode = BPF_OP(insn->code);
10751
10752 env->jmps_processed++;
10753 if (opcode == BPF_CALL) {
10754 if (BPF_SRC(insn->code) != BPF_K ||
10755 insn->off != 0 ||
10756 (insn->src_reg != BPF_REG_0 &&
10757 insn->src_reg != BPF_PSEUDO_CALL &&
10758 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
10759 insn->dst_reg != BPF_REG_0 ||
10760 class == BPF_JMP32) {
10761 verbose(env, "BPF_CALL uses reserved fields\n");
10762 return -EINVAL;
10763 }
10764
10765 if (env->cur_state->active_spin_lock &&
10766 (insn->src_reg == BPF_PSEUDO_CALL ||
10767 insn->imm != BPF_FUNC_spin_unlock)) {
10768 verbose(env, "function calls are not allowed while holding a lock\n");
10769 return -EINVAL;
10770 }
10771 if (insn->src_reg == BPF_PSEUDO_CALL)
10772 err = check_func_call(env, insn, &env->insn_idx);
10773 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
10774 err = check_kfunc_call(env, insn);
10775 else
10776 err = check_helper_call(env, insn, &env->insn_idx);
10777 if (err)
10778 return err;
10779 } else if (opcode == BPF_JA) {
10780 if (BPF_SRC(insn->code) != BPF_K ||
10781 insn->imm != 0 ||
10782 insn->src_reg != BPF_REG_0 ||
10783 insn->dst_reg != BPF_REG_0 ||
10784 class == BPF_JMP32) {
10785 verbose(env, "BPF_JA uses reserved fields\n");
10786 return -EINVAL;
10787 }
10788
10789 env->insn_idx += insn->off + 1;
10790 continue;
10791
10792 } else if (opcode == BPF_EXIT) {
10793 if (BPF_SRC(insn->code) != BPF_K ||
10794 insn->imm != 0 ||
10795 insn->src_reg != BPF_REG_0 ||
10796 insn->dst_reg != BPF_REG_0 ||
10797 class == BPF_JMP32) {
10798 verbose(env, "BPF_EXIT uses reserved fields\n");
10799 return -EINVAL;
10800 }
10801
10802 if (env->cur_state->active_spin_lock) {
10803 verbose(env, "bpf_spin_unlock is missing\n");
10804 return -EINVAL;
10805 }
10806
10807 if (state->curframe) {
10808 /* exit from nested function */
10809 err = prepare_func_exit(env, &env->insn_idx);
10810 if (err)
10811 return err;
10812 do_print_state = true;
10813 continue;
10814 }
10815
10816 err = check_reference_leak(env);
10817 if (err)
10818 return err;
10819
10820 err = check_return_code(env);
10821 if (err)
10822 return err;
10823 process_bpf_exit:
10824 update_branch_counts(env, env->cur_state);
10825 err = pop_stack(env, &prev_insn_idx,
10826 &env->insn_idx, pop_log);
10827 if (err < 0) {
10828 if (err != -ENOENT)
10829 return err;
10830 break;
10831 } else {
10832 do_print_state = true;
10833 continue;
10834 }
10835 } else {
10836 err = check_cond_jmp_op(env, insn, &env->insn_idx);
10837 if (err)
10838 return err;
10839 }
10840 } else if (class == BPF_LD) {
10841 u8 mode = BPF_MODE(insn->code);
10842
10843 if (mode == BPF_ABS || mode == BPF_IND) {
10844 err = check_ld_abs(env, insn);
10845 if (err)
10846 return err;
10847
10848 } else if (mode == BPF_IMM) {
10849 err = check_ld_imm(env, insn);
10850 if (err)
10851 return err;
10852
10853 env->insn_idx++;
10854 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
10855 } else {
10856 verbose(env, "invalid BPF_LD mode\n");
10857 return -EINVAL;
10858 }
10859 } else {
10860 verbose(env, "unknown insn class %d\n", class);
10861 return -EINVAL;
10862 }
10863
10864 env->insn_idx++;
10865 }
10866
10867 return 0;
10868 }
10869
find_btf_percpu_datasec(struct btf * btf)10870 static int find_btf_percpu_datasec(struct btf *btf)
10871 {
10872 const struct btf_type *t;
10873 const char *tname;
10874 int i, n;
10875
10876 /*
10877 * Both vmlinux and module each have their own ".data..percpu"
10878 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
10879 * types to look at only module's own BTF types.
10880 */
10881 n = btf_nr_types(btf);
10882 if (btf_is_module(btf))
10883 i = btf_nr_types(btf_vmlinux);
10884 else
10885 i = 1;
10886
10887 for(; i < n; i++) {
10888 t = btf_type_by_id(btf, i);
10889 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
10890 continue;
10891
10892 tname = btf_name_by_offset(btf, t->name_off);
10893 if (!strcmp(tname, ".data..percpu"))
10894 return i;
10895 }
10896
10897 return -ENOENT;
10898 }
10899
10900 /* replace pseudo btf_id with kernel symbol address */
check_pseudo_btf_id(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_insn_aux_data * aux)10901 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
10902 struct bpf_insn *insn,
10903 struct bpf_insn_aux_data *aux)
10904 {
10905 const struct btf_var_secinfo *vsi;
10906 const struct btf_type *datasec;
10907 struct btf_mod_pair *btf_mod;
10908 const struct btf_type *t;
10909 const char *sym_name;
10910 bool percpu = false;
10911 u32 type, id = insn->imm;
10912 struct btf *btf;
10913 s32 datasec_id;
10914 u64 addr;
10915 int i, btf_fd, err;
10916
10917 btf_fd = insn[1].imm;
10918 if (btf_fd) {
10919 btf = btf_get_by_fd(btf_fd);
10920 if (IS_ERR(btf)) {
10921 verbose(env, "invalid module BTF object FD specified.\n");
10922 return -EINVAL;
10923 }
10924 } else {
10925 if (!btf_vmlinux) {
10926 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
10927 return -EINVAL;
10928 }
10929 btf = btf_vmlinux;
10930 btf_get(btf);
10931 }
10932
10933 t = btf_type_by_id(btf, id);
10934 if (!t) {
10935 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
10936 err = -ENOENT;
10937 goto err_put;
10938 }
10939
10940 if (!btf_type_is_var(t)) {
10941 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
10942 err = -EINVAL;
10943 goto err_put;
10944 }
10945
10946 sym_name = btf_name_by_offset(btf, t->name_off);
10947 addr = kallsyms_lookup_name(sym_name);
10948 if (!addr) {
10949 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
10950 sym_name);
10951 err = -ENOENT;
10952 goto err_put;
10953 }
10954
10955 datasec_id = find_btf_percpu_datasec(btf);
10956 if (datasec_id > 0) {
10957 datasec = btf_type_by_id(btf, datasec_id);
10958 for_each_vsi(i, datasec, vsi) {
10959 if (vsi->type == id) {
10960 percpu = true;
10961 break;
10962 }
10963 }
10964 }
10965
10966 insn[0].imm = (u32)addr;
10967 insn[1].imm = addr >> 32;
10968
10969 type = t->type;
10970 t = btf_type_skip_modifiers(btf, type, NULL);
10971 if (percpu) {
10972 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
10973 aux->btf_var.btf = btf;
10974 aux->btf_var.btf_id = type;
10975 } else if (!btf_type_is_struct(t)) {
10976 const struct btf_type *ret;
10977 const char *tname;
10978 u32 tsize;
10979
10980 /* resolve the type size of ksym. */
10981 ret = btf_resolve_size(btf, t, &tsize);
10982 if (IS_ERR(ret)) {
10983 tname = btf_name_by_offset(btf, t->name_off);
10984 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
10985 tname, PTR_ERR(ret));
10986 err = -EINVAL;
10987 goto err_put;
10988 }
10989 aux->btf_var.reg_type = PTR_TO_MEM;
10990 aux->btf_var.mem_size = tsize;
10991 } else {
10992 aux->btf_var.reg_type = PTR_TO_BTF_ID;
10993 aux->btf_var.btf = btf;
10994 aux->btf_var.btf_id = type;
10995 }
10996
10997 /* check whether we recorded this BTF (and maybe module) already */
10998 for (i = 0; i < env->used_btf_cnt; i++) {
10999 if (env->used_btfs[i].btf == btf) {
11000 btf_put(btf);
11001 return 0;
11002 }
11003 }
11004
11005 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11006 err = -E2BIG;
11007 goto err_put;
11008 }
11009
11010 btf_mod = &env->used_btfs[env->used_btf_cnt];
11011 btf_mod->btf = btf;
11012 btf_mod->module = NULL;
11013
11014 /* if we reference variables from kernel module, bump its refcount */
11015 if (btf_is_module(btf)) {
11016 btf_mod->module = btf_try_get_module(btf);
11017 if (!btf_mod->module) {
11018 err = -ENXIO;
11019 goto err_put;
11020 }
11021 }
11022
11023 env->used_btf_cnt++;
11024
11025 return 0;
11026 err_put:
11027 btf_put(btf);
11028 return err;
11029 }
11030
check_map_prealloc(struct bpf_map * map)11031 static int check_map_prealloc(struct bpf_map *map)
11032 {
11033 return (map->map_type != BPF_MAP_TYPE_HASH &&
11034 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11035 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11036 !(map->map_flags & BPF_F_NO_PREALLOC);
11037 }
11038
is_tracing_prog_type(enum bpf_prog_type type)11039 static bool is_tracing_prog_type(enum bpf_prog_type type)
11040 {
11041 switch (type) {
11042 case BPF_PROG_TYPE_KPROBE:
11043 case BPF_PROG_TYPE_TRACEPOINT:
11044 case BPF_PROG_TYPE_PERF_EVENT:
11045 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11046 return true;
11047 default:
11048 return false;
11049 }
11050 }
11051
is_preallocated_map(struct bpf_map * map)11052 static bool is_preallocated_map(struct bpf_map *map)
11053 {
11054 if (!check_map_prealloc(map))
11055 return false;
11056 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11057 return false;
11058 return true;
11059 }
11060
check_map_prog_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,struct bpf_prog * prog)11061 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11062 struct bpf_map *map,
11063 struct bpf_prog *prog)
11064
11065 {
11066 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11067 /*
11068 * Validate that trace type programs use preallocated hash maps.
11069 *
11070 * For programs attached to PERF events this is mandatory as the
11071 * perf NMI can hit any arbitrary code sequence.
11072 *
11073 * All other trace types using preallocated hash maps are unsafe as
11074 * well because tracepoint or kprobes can be inside locked regions
11075 * of the memory allocator or at a place where a recursion into the
11076 * memory allocator would see inconsistent state.
11077 *
11078 * On RT enabled kernels run-time allocation of all trace type
11079 * programs is strictly prohibited due to lock type constraints. On
11080 * !RT kernels it is allowed for backwards compatibility reasons for
11081 * now, but warnings are emitted so developers are made aware of
11082 * the unsafety and can fix their programs before this is enforced.
11083 */
11084 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11085 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11086 verbose(env, "perf_event programs can only use preallocated hash map\n");
11087 return -EINVAL;
11088 }
11089 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11090 verbose(env, "trace type programs can only use preallocated hash map\n");
11091 return -EINVAL;
11092 }
11093 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11094 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11095 }
11096
11097 if (map_value_has_spin_lock(map)) {
11098 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11099 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11100 return -EINVAL;
11101 }
11102
11103 if (is_tracing_prog_type(prog_type)) {
11104 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11105 return -EINVAL;
11106 }
11107
11108 if (prog->aux->sleepable) {
11109 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11110 return -EINVAL;
11111 }
11112 }
11113
11114 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11115 !bpf_offload_prog_map_match(prog, map)) {
11116 verbose(env, "offload device mismatch between prog and map\n");
11117 return -EINVAL;
11118 }
11119
11120 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11121 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11122 return -EINVAL;
11123 }
11124
11125 if (prog->aux->sleepable)
11126 switch (map->map_type) {
11127 case BPF_MAP_TYPE_HASH:
11128 case BPF_MAP_TYPE_LRU_HASH:
11129 case BPF_MAP_TYPE_ARRAY:
11130 case BPF_MAP_TYPE_PERCPU_HASH:
11131 case BPF_MAP_TYPE_PERCPU_ARRAY:
11132 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11133 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11134 case BPF_MAP_TYPE_HASH_OF_MAPS:
11135 if (!is_preallocated_map(map)) {
11136 verbose(env,
11137 "Sleepable programs can only use preallocated maps\n");
11138 return -EINVAL;
11139 }
11140 break;
11141 case BPF_MAP_TYPE_RINGBUF:
11142 break;
11143 default:
11144 verbose(env,
11145 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11146 return -EINVAL;
11147 }
11148
11149 return 0;
11150 }
11151
bpf_map_is_cgroup_storage(struct bpf_map * map)11152 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11153 {
11154 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11155 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11156 }
11157
11158 /* find and rewrite pseudo imm in ld_imm64 instructions:
11159 *
11160 * 1. if it accesses map FD, replace it with actual map pointer.
11161 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11162 *
11163 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11164 */
resolve_pseudo_ldimm64(struct bpf_verifier_env * env)11165 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11166 {
11167 struct bpf_insn *insn = env->prog->insnsi;
11168 int insn_cnt = env->prog->len;
11169 int i, j, err;
11170
11171 err = bpf_prog_calc_tag(env->prog);
11172 if (err)
11173 return err;
11174
11175 for (i = 0; i < insn_cnt; i++, insn++) {
11176 if (BPF_CLASS(insn->code) == BPF_LDX &&
11177 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11178 verbose(env, "BPF_LDX uses reserved fields\n");
11179 return -EINVAL;
11180 }
11181
11182 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11183 struct bpf_insn_aux_data *aux;
11184 struct bpf_map *map;
11185 struct fd f;
11186 u64 addr;
11187
11188 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11189 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11190 insn[1].off != 0) {
11191 verbose(env, "invalid bpf_ld_imm64 insn\n");
11192 return -EINVAL;
11193 }
11194
11195 if (insn[0].src_reg == 0)
11196 /* valid generic load 64-bit imm */
11197 goto next_insn;
11198
11199 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11200 aux = &env->insn_aux_data[i];
11201 err = check_pseudo_btf_id(env, insn, aux);
11202 if (err)
11203 return err;
11204 goto next_insn;
11205 }
11206
11207 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11208 aux = &env->insn_aux_data[i];
11209 aux->ptr_type = PTR_TO_FUNC;
11210 goto next_insn;
11211 }
11212
11213 /* In final convert_pseudo_ld_imm64() step, this is
11214 * converted into regular 64-bit imm load insn.
11215 */
11216 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
11217 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
11218 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
11219 insn[1].imm != 0)) {
11220 verbose(env,
11221 "unrecognized bpf_ld_imm64 insn\n");
11222 return -EINVAL;
11223 }
11224
11225 f = fdget(insn[0].imm);
11226 map = __bpf_map_get(f);
11227 if (IS_ERR(map)) {
11228 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11229 insn[0].imm);
11230 return PTR_ERR(map);
11231 }
11232
11233 err = check_map_prog_compatibility(env, map, env->prog);
11234 if (err) {
11235 fdput(f);
11236 return err;
11237 }
11238
11239 aux = &env->insn_aux_data[i];
11240 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
11241 addr = (unsigned long)map;
11242 } else {
11243 u32 off = insn[1].imm;
11244
11245 if (off >= BPF_MAX_VAR_OFF) {
11246 verbose(env, "direct value offset of %u is not allowed\n", off);
11247 fdput(f);
11248 return -EINVAL;
11249 }
11250
11251 if (!map->ops->map_direct_value_addr) {
11252 verbose(env, "no direct value access support for this map type\n");
11253 fdput(f);
11254 return -EINVAL;
11255 }
11256
11257 err = map->ops->map_direct_value_addr(map, &addr, off);
11258 if (err) {
11259 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11260 map->value_size, off);
11261 fdput(f);
11262 return err;
11263 }
11264
11265 aux->map_off = off;
11266 addr += off;
11267 }
11268
11269 insn[0].imm = (u32)addr;
11270 insn[1].imm = addr >> 32;
11271
11272 /* check whether we recorded this map already */
11273 for (j = 0; j < env->used_map_cnt; j++) {
11274 if (env->used_maps[j] == map) {
11275 aux->map_index = j;
11276 fdput(f);
11277 goto next_insn;
11278 }
11279 }
11280
11281 if (env->used_map_cnt >= MAX_USED_MAPS) {
11282 fdput(f);
11283 return -E2BIG;
11284 }
11285
11286 /* hold the map. If the program is rejected by verifier,
11287 * the map will be released by release_maps() or it
11288 * will be used by the valid program until it's unloaded
11289 * and all maps are released in free_used_maps()
11290 */
11291 bpf_map_inc(map);
11292
11293 aux->map_index = env->used_map_cnt;
11294 env->used_maps[env->used_map_cnt++] = map;
11295
11296 if (bpf_map_is_cgroup_storage(map) &&
11297 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11298 verbose(env, "only one cgroup storage of each type is allowed\n");
11299 fdput(f);
11300 return -EBUSY;
11301 }
11302
11303 fdput(f);
11304 next_insn:
11305 insn++;
11306 i++;
11307 continue;
11308 }
11309
11310 /* Basic sanity check before we invest more work here. */
11311 if (!bpf_opcode_in_insntable(insn->code)) {
11312 verbose(env, "unknown opcode %02x\n", insn->code);
11313 return -EINVAL;
11314 }
11315 }
11316
11317 /* now all pseudo BPF_LD_IMM64 instructions load valid
11318 * 'struct bpf_map *' into a register instead of user map_fd.
11319 * These pointers will be used later by verifier to validate map access.
11320 */
11321 return 0;
11322 }
11323
11324 /* drop refcnt of maps used by the rejected program */
release_maps(struct bpf_verifier_env * env)11325 static void release_maps(struct bpf_verifier_env *env)
11326 {
11327 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11328 env->used_map_cnt);
11329 }
11330
11331 /* drop refcnt of maps used by the rejected program */
release_btfs(struct bpf_verifier_env * env)11332 static void release_btfs(struct bpf_verifier_env *env)
11333 {
11334 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11335 env->used_btf_cnt);
11336 }
11337
11338 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
convert_pseudo_ld_imm64(struct bpf_verifier_env * env)11339 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11340 {
11341 struct bpf_insn *insn = env->prog->insnsi;
11342 int insn_cnt = env->prog->len;
11343 int i;
11344
11345 for (i = 0; i < insn_cnt; i++, insn++) {
11346 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11347 continue;
11348 if (insn->src_reg == BPF_PSEUDO_FUNC)
11349 continue;
11350 insn->src_reg = 0;
11351 }
11352 }
11353
11354 /* single env->prog->insni[off] instruction was replaced with the range
11355 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11356 * [0, off) and [off, end) to new locations, so the patched range stays zero
11357 */
adjust_insn_aux_data(struct bpf_verifier_env * env,struct bpf_prog * new_prog,u32 off,u32 cnt)11358 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
11359 struct bpf_prog *new_prog, u32 off, u32 cnt)
11360 {
11361 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
11362 struct bpf_insn *insn = new_prog->insnsi;
11363 u32 prog_len;
11364 int i;
11365
11366 /* aux info at OFF always needs adjustment, no matter fast path
11367 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11368 * original insn at old prog.
11369 */
11370 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11371
11372 if (cnt == 1)
11373 return 0;
11374 prog_len = new_prog->len;
11375 new_data = vzalloc(array_size(prog_len,
11376 sizeof(struct bpf_insn_aux_data)));
11377 if (!new_data)
11378 return -ENOMEM;
11379 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11380 memcpy(new_data + off + cnt - 1, old_data + off,
11381 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11382 for (i = off; i < off + cnt - 1; i++) {
11383 new_data[i].seen = env->pass_cnt;
11384 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11385 }
11386 env->insn_aux_data = new_data;
11387 vfree(old_data);
11388 return 0;
11389 }
11390
adjust_subprog_starts(struct bpf_verifier_env * env,u32 off,u32 len)11391 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11392 {
11393 int i;
11394
11395 if (len == 1)
11396 return;
11397 /* NOTE: fake 'exit' subprog should be updated as well. */
11398 for (i = 0; i <= env->subprog_cnt; i++) {
11399 if (env->subprog_info[i].start <= off)
11400 continue;
11401 env->subprog_info[i].start += len - 1;
11402 }
11403 }
11404
adjust_poke_descs(struct bpf_prog * prog,u32 len)11405 static void adjust_poke_descs(struct bpf_prog *prog, u32 len)
11406 {
11407 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11408 int i, sz = prog->aux->size_poke_tab;
11409 struct bpf_jit_poke_descriptor *desc;
11410
11411 for (i = 0; i < sz; i++) {
11412 desc = &tab[i];
11413 desc->insn_idx += len - 1;
11414 }
11415 }
11416
bpf_patch_insn_data(struct bpf_verifier_env * env,u32 off,const struct bpf_insn * patch,u32 len)11417 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11418 const struct bpf_insn *patch, u32 len)
11419 {
11420 struct bpf_prog *new_prog;
11421
11422 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11423 if (IS_ERR(new_prog)) {
11424 if (PTR_ERR(new_prog) == -ERANGE)
11425 verbose(env,
11426 "insn %d cannot be patched due to 16-bit range\n",
11427 env->insn_aux_data[off].orig_idx);
11428 return NULL;
11429 }
11430 if (adjust_insn_aux_data(env, new_prog, off, len))
11431 return NULL;
11432 adjust_subprog_starts(env, off, len);
11433 adjust_poke_descs(new_prog, len);
11434 return new_prog;
11435 }
11436
adjust_subprog_starts_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)11437 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11438 u32 off, u32 cnt)
11439 {
11440 int i, j;
11441
11442 /* find first prog starting at or after off (first to remove) */
11443 for (i = 0; i < env->subprog_cnt; i++)
11444 if (env->subprog_info[i].start >= off)
11445 break;
11446 /* find first prog starting at or after off + cnt (first to stay) */
11447 for (j = i; j < env->subprog_cnt; j++)
11448 if (env->subprog_info[j].start >= off + cnt)
11449 break;
11450 /* if j doesn't start exactly at off + cnt, we are just removing
11451 * the front of previous prog
11452 */
11453 if (env->subprog_info[j].start != off + cnt)
11454 j--;
11455
11456 if (j > i) {
11457 struct bpf_prog_aux *aux = env->prog->aux;
11458 int move;
11459
11460 /* move fake 'exit' subprog as well */
11461 move = env->subprog_cnt + 1 - j;
11462
11463 memmove(env->subprog_info + i,
11464 env->subprog_info + j,
11465 sizeof(*env->subprog_info) * move);
11466 env->subprog_cnt -= j - i;
11467
11468 /* remove func_info */
11469 if (aux->func_info) {
11470 move = aux->func_info_cnt - j;
11471
11472 memmove(aux->func_info + i,
11473 aux->func_info + j,
11474 sizeof(*aux->func_info) * move);
11475 aux->func_info_cnt -= j - i;
11476 /* func_info->insn_off is set after all code rewrites,
11477 * in adjust_btf_func() - no need to adjust
11478 */
11479 }
11480 } else {
11481 /* convert i from "first prog to remove" to "first to adjust" */
11482 if (env->subprog_info[i].start == off)
11483 i++;
11484 }
11485
11486 /* update fake 'exit' subprog as well */
11487 for (; i <= env->subprog_cnt; i++)
11488 env->subprog_info[i].start -= cnt;
11489
11490 return 0;
11491 }
11492
bpf_adj_linfo_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)11493 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11494 u32 cnt)
11495 {
11496 struct bpf_prog *prog = env->prog;
11497 u32 i, l_off, l_cnt, nr_linfo;
11498 struct bpf_line_info *linfo;
11499
11500 nr_linfo = prog->aux->nr_linfo;
11501 if (!nr_linfo)
11502 return 0;
11503
11504 linfo = prog->aux->linfo;
11505
11506 /* find first line info to remove, count lines to be removed */
11507 for (i = 0; i < nr_linfo; i++)
11508 if (linfo[i].insn_off >= off)
11509 break;
11510
11511 l_off = i;
11512 l_cnt = 0;
11513 for (; i < nr_linfo; i++)
11514 if (linfo[i].insn_off < off + cnt)
11515 l_cnt++;
11516 else
11517 break;
11518
11519 /* First live insn doesn't match first live linfo, it needs to "inherit"
11520 * last removed linfo. prog is already modified, so prog->len == off
11521 * means no live instructions after (tail of the program was removed).
11522 */
11523 if (prog->len != off && l_cnt &&
11524 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11525 l_cnt--;
11526 linfo[--i].insn_off = off + cnt;
11527 }
11528
11529 /* remove the line info which refer to the removed instructions */
11530 if (l_cnt) {
11531 memmove(linfo + l_off, linfo + i,
11532 sizeof(*linfo) * (nr_linfo - i));
11533
11534 prog->aux->nr_linfo -= l_cnt;
11535 nr_linfo = prog->aux->nr_linfo;
11536 }
11537
11538 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
11539 for (i = l_off; i < nr_linfo; i++)
11540 linfo[i].insn_off -= cnt;
11541
11542 /* fix up all subprogs (incl. 'exit') which start >= off */
11543 for (i = 0; i <= env->subprog_cnt; i++)
11544 if (env->subprog_info[i].linfo_idx > l_off) {
11545 /* program may have started in the removed region but
11546 * may not be fully removed
11547 */
11548 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11549 env->subprog_info[i].linfo_idx -= l_cnt;
11550 else
11551 env->subprog_info[i].linfo_idx = l_off;
11552 }
11553
11554 return 0;
11555 }
11556
verifier_remove_insns(struct bpf_verifier_env * env,u32 off,u32 cnt)11557 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11558 {
11559 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11560 unsigned int orig_prog_len = env->prog->len;
11561 int err;
11562
11563 if (bpf_prog_is_dev_bound(env->prog->aux))
11564 bpf_prog_offload_remove_insns(env, off, cnt);
11565
11566 err = bpf_remove_insns(env->prog, off, cnt);
11567 if (err)
11568 return err;
11569
11570 err = adjust_subprog_starts_after_remove(env, off, cnt);
11571 if (err)
11572 return err;
11573
11574 err = bpf_adj_linfo_after_remove(env, off, cnt);
11575 if (err)
11576 return err;
11577
11578 memmove(aux_data + off, aux_data + off + cnt,
11579 sizeof(*aux_data) * (orig_prog_len - off - cnt));
11580
11581 return 0;
11582 }
11583
11584 /* The verifier does more data flow analysis than llvm and will not
11585 * explore branches that are dead at run time. Malicious programs can
11586 * have dead code too. Therefore replace all dead at-run-time code
11587 * with 'ja -1'.
11588 *
11589 * Just nops are not optimal, e.g. if they would sit at the end of the
11590 * program and through another bug we would manage to jump there, then
11591 * we'd execute beyond program memory otherwise. Returning exception
11592 * code also wouldn't work since we can have subprogs where the dead
11593 * code could be located.
11594 */
sanitize_dead_code(struct bpf_verifier_env * env)11595 static void sanitize_dead_code(struct bpf_verifier_env *env)
11596 {
11597 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11598 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11599 struct bpf_insn *insn = env->prog->insnsi;
11600 const int insn_cnt = env->prog->len;
11601 int i;
11602
11603 for (i = 0; i < insn_cnt; i++) {
11604 if (aux_data[i].seen)
11605 continue;
11606 memcpy(insn + i, &trap, sizeof(trap));
11607 }
11608 }
11609
insn_is_cond_jump(u8 code)11610 static bool insn_is_cond_jump(u8 code)
11611 {
11612 u8 op;
11613
11614 if (BPF_CLASS(code) == BPF_JMP32)
11615 return true;
11616
11617 if (BPF_CLASS(code) != BPF_JMP)
11618 return false;
11619
11620 op = BPF_OP(code);
11621 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11622 }
11623
opt_hard_wire_dead_code_branches(struct bpf_verifier_env * env)11624 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11625 {
11626 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11627 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11628 struct bpf_insn *insn = env->prog->insnsi;
11629 const int insn_cnt = env->prog->len;
11630 int i;
11631
11632 for (i = 0; i < insn_cnt; i++, insn++) {
11633 if (!insn_is_cond_jump(insn->code))
11634 continue;
11635
11636 if (!aux_data[i + 1].seen)
11637 ja.off = insn->off;
11638 else if (!aux_data[i + 1 + insn->off].seen)
11639 ja.off = 0;
11640 else
11641 continue;
11642
11643 if (bpf_prog_is_dev_bound(env->prog->aux))
11644 bpf_prog_offload_replace_insn(env, i, &ja);
11645
11646 memcpy(insn, &ja, sizeof(ja));
11647 }
11648 }
11649
opt_remove_dead_code(struct bpf_verifier_env * env)11650 static int opt_remove_dead_code(struct bpf_verifier_env *env)
11651 {
11652 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11653 int insn_cnt = env->prog->len;
11654 int i, err;
11655
11656 for (i = 0; i < insn_cnt; i++) {
11657 int j;
11658
11659 j = 0;
11660 while (i + j < insn_cnt && !aux_data[i + j].seen)
11661 j++;
11662 if (!j)
11663 continue;
11664
11665 err = verifier_remove_insns(env, i, j);
11666 if (err)
11667 return err;
11668 insn_cnt = env->prog->len;
11669 }
11670
11671 return 0;
11672 }
11673
opt_remove_nops(struct bpf_verifier_env * env)11674 static int opt_remove_nops(struct bpf_verifier_env *env)
11675 {
11676 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11677 struct bpf_insn *insn = env->prog->insnsi;
11678 int insn_cnt = env->prog->len;
11679 int i, err;
11680
11681 for (i = 0; i < insn_cnt; i++) {
11682 if (memcmp(&insn[i], &ja, sizeof(ja)))
11683 continue;
11684
11685 err = verifier_remove_insns(env, i, 1);
11686 if (err)
11687 return err;
11688 insn_cnt--;
11689 i--;
11690 }
11691
11692 return 0;
11693 }
11694
opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env * env,const union bpf_attr * attr)11695 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
11696 const union bpf_attr *attr)
11697 {
11698 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
11699 struct bpf_insn_aux_data *aux = env->insn_aux_data;
11700 int i, patch_len, delta = 0, len = env->prog->len;
11701 struct bpf_insn *insns = env->prog->insnsi;
11702 struct bpf_prog *new_prog;
11703 bool rnd_hi32;
11704
11705 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
11706 zext_patch[1] = BPF_ZEXT_REG(0);
11707 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
11708 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
11709 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
11710 for (i = 0; i < len; i++) {
11711 int adj_idx = i + delta;
11712 struct bpf_insn insn;
11713 int load_reg;
11714
11715 insn = insns[adj_idx];
11716 load_reg = insn_def_regno(&insn);
11717 if (!aux[adj_idx].zext_dst) {
11718 u8 code, class;
11719 u32 imm_rnd;
11720
11721 if (!rnd_hi32)
11722 continue;
11723
11724 code = insn.code;
11725 class = BPF_CLASS(code);
11726 if (load_reg == -1)
11727 continue;
11728
11729 /* NOTE: arg "reg" (the fourth one) is only used for
11730 * BPF_STX + SRC_OP, so it is safe to pass NULL
11731 * here.
11732 */
11733 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
11734 if (class == BPF_LD &&
11735 BPF_MODE(code) == BPF_IMM)
11736 i++;
11737 continue;
11738 }
11739
11740 /* ctx load could be transformed into wider load. */
11741 if (class == BPF_LDX &&
11742 aux[adj_idx].ptr_type == PTR_TO_CTX)
11743 continue;
11744
11745 imm_rnd = get_random_int();
11746 rnd_hi32_patch[0] = insn;
11747 rnd_hi32_patch[1].imm = imm_rnd;
11748 rnd_hi32_patch[3].dst_reg = load_reg;
11749 patch = rnd_hi32_patch;
11750 patch_len = 4;
11751 goto apply_patch_buffer;
11752 }
11753
11754 /* Add in an zero-extend instruction if a) the JIT has requested
11755 * it or b) it's a CMPXCHG.
11756 *
11757 * The latter is because: BPF_CMPXCHG always loads a value into
11758 * R0, therefore always zero-extends. However some archs'
11759 * equivalent instruction only does this load when the
11760 * comparison is successful. This detail of CMPXCHG is
11761 * orthogonal to the general zero-extension behaviour of the
11762 * CPU, so it's treated independently of bpf_jit_needs_zext.
11763 */
11764 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
11765 continue;
11766
11767 if (WARN_ON(load_reg == -1)) {
11768 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
11769 return -EFAULT;
11770 }
11771
11772 zext_patch[0] = insn;
11773 zext_patch[1].dst_reg = load_reg;
11774 zext_patch[1].src_reg = load_reg;
11775 patch = zext_patch;
11776 patch_len = 2;
11777 apply_patch_buffer:
11778 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
11779 if (!new_prog)
11780 return -ENOMEM;
11781 env->prog = new_prog;
11782 insns = new_prog->insnsi;
11783 aux = env->insn_aux_data;
11784 delta += patch_len - 1;
11785 }
11786
11787 return 0;
11788 }
11789
11790 /* convert load instructions that access fields of a context type into a
11791 * sequence of instructions that access fields of the underlying structure:
11792 * struct __sk_buff -> struct sk_buff
11793 * struct bpf_sock_ops -> struct sock
11794 */
convert_ctx_accesses(struct bpf_verifier_env * env)11795 static int convert_ctx_accesses(struct bpf_verifier_env *env)
11796 {
11797 const struct bpf_verifier_ops *ops = env->ops;
11798 int i, cnt, size, ctx_field_size, delta = 0;
11799 const int insn_cnt = env->prog->len;
11800 struct bpf_insn insn_buf[16], *insn;
11801 u32 target_size, size_default, off;
11802 struct bpf_prog *new_prog;
11803 enum bpf_access_type type;
11804 bool is_narrower_load;
11805
11806 if (ops->gen_prologue || env->seen_direct_write) {
11807 if (!ops->gen_prologue) {
11808 verbose(env, "bpf verifier is misconfigured\n");
11809 return -EINVAL;
11810 }
11811 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
11812 env->prog);
11813 if (cnt >= ARRAY_SIZE(insn_buf)) {
11814 verbose(env, "bpf verifier is misconfigured\n");
11815 return -EINVAL;
11816 } else if (cnt) {
11817 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
11818 if (!new_prog)
11819 return -ENOMEM;
11820
11821 env->prog = new_prog;
11822 delta += cnt - 1;
11823 }
11824 }
11825
11826 if (bpf_prog_is_dev_bound(env->prog->aux))
11827 return 0;
11828
11829 insn = env->prog->insnsi + delta;
11830
11831 for (i = 0; i < insn_cnt; i++, insn++) {
11832 bpf_convert_ctx_access_t convert_ctx_access;
11833
11834 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
11835 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
11836 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
11837 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
11838 type = BPF_READ;
11839 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
11840 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
11841 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
11842 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
11843 type = BPF_WRITE;
11844 else
11845 continue;
11846
11847 if (type == BPF_WRITE &&
11848 env->insn_aux_data[i + delta].sanitize_stack_off) {
11849 struct bpf_insn patch[] = {
11850 /* Sanitize suspicious stack slot with zero.
11851 * There are no memory dependencies for this store,
11852 * since it's only using frame pointer and immediate
11853 * constant of zero
11854 */
11855 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
11856 env->insn_aux_data[i + delta].sanitize_stack_off,
11857 0),
11858 /* the original STX instruction will immediately
11859 * overwrite the same stack slot with appropriate value
11860 */
11861 *insn,
11862 };
11863
11864 cnt = ARRAY_SIZE(patch);
11865 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
11866 if (!new_prog)
11867 return -ENOMEM;
11868
11869 delta += cnt - 1;
11870 env->prog = new_prog;
11871 insn = new_prog->insnsi + i + delta;
11872 continue;
11873 }
11874
11875 switch (env->insn_aux_data[i + delta].ptr_type) {
11876 case PTR_TO_CTX:
11877 if (!ops->convert_ctx_access)
11878 continue;
11879 convert_ctx_access = ops->convert_ctx_access;
11880 break;
11881 case PTR_TO_SOCKET:
11882 case PTR_TO_SOCK_COMMON:
11883 convert_ctx_access = bpf_sock_convert_ctx_access;
11884 break;
11885 case PTR_TO_TCP_SOCK:
11886 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
11887 break;
11888 case PTR_TO_XDP_SOCK:
11889 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
11890 break;
11891 case PTR_TO_BTF_ID:
11892 if (type == BPF_READ) {
11893 insn->code = BPF_LDX | BPF_PROBE_MEM |
11894 BPF_SIZE((insn)->code);
11895 env->prog->aux->num_exentries++;
11896 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
11897 verbose(env, "Writes through BTF pointers are not allowed\n");
11898 return -EINVAL;
11899 }
11900 continue;
11901 default:
11902 continue;
11903 }
11904
11905 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
11906 size = BPF_LDST_BYTES(insn);
11907
11908 /* If the read access is a narrower load of the field,
11909 * convert to a 4/8-byte load, to minimum program type specific
11910 * convert_ctx_access changes. If conversion is successful,
11911 * we will apply proper mask to the result.
11912 */
11913 is_narrower_load = size < ctx_field_size;
11914 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
11915 off = insn->off;
11916 if (is_narrower_load) {
11917 u8 size_code;
11918
11919 if (type == BPF_WRITE) {
11920 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
11921 return -EINVAL;
11922 }
11923
11924 size_code = BPF_H;
11925 if (ctx_field_size == 4)
11926 size_code = BPF_W;
11927 else if (ctx_field_size == 8)
11928 size_code = BPF_DW;
11929
11930 insn->off = off & ~(size_default - 1);
11931 insn->code = BPF_LDX | BPF_MEM | size_code;
11932 }
11933
11934 target_size = 0;
11935 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
11936 &target_size);
11937 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
11938 (ctx_field_size && !target_size)) {
11939 verbose(env, "bpf verifier is misconfigured\n");
11940 return -EINVAL;
11941 }
11942
11943 if (is_narrower_load && size < target_size) {
11944 u8 shift = bpf_ctx_narrow_access_offset(
11945 off, size, size_default) * 8;
11946 if (ctx_field_size <= 4) {
11947 if (shift)
11948 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
11949 insn->dst_reg,
11950 shift);
11951 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
11952 (1 << size * 8) - 1);
11953 } else {
11954 if (shift)
11955 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
11956 insn->dst_reg,
11957 shift);
11958 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
11959 (1ULL << size * 8) - 1);
11960 }
11961 }
11962
11963 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11964 if (!new_prog)
11965 return -ENOMEM;
11966
11967 delta += cnt - 1;
11968
11969 /* keep walking new program and skip insns we just inserted */
11970 env->prog = new_prog;
11971 insn = new_prog->insnsi + i + delta;
11972 }
11973
11974 return 0;
11975 }
11976
jit_subprogs(struct bpf_verifier_env * env)11977 static int jit_subprogs(struct bpf_verifier_env *env)
11978 {
11979 struct bpf_prog *prog = env->prog, **func, *tmp;
11980 int i, j, subprog_start, subprog_end = 0, len, subprog;
11981 struct bpf_map *map_ptr;
11982 struct bpf_insn *insn;
11983 void *old_bpf_func;
11984 int err, num_exentries;
11985
11986 if (env->subprog_cnt <= 1)
11987 return 0;
11988
11989 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
11990 if (bpf_pseudo_func(insn)) {
11991 env->insn_aux_data[i].call_imm = insn->imm;
11992 /* subprog is encoded in insn[1].imm */
11993 continue;
11994 }
11995
11996 if (!bpf_pseudo_call(insn))
11997 continue;
11998 /* Upon error here we cannot fall back to interpreter but
11999 * need a hard reject of the program. Thus -EFAULT is
12000 * propagated in any case.
12001 */
12002 subprog = find_subprog(env, i + insn->imm + 1);
12003 if (subprog < 0) {
12004 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12005 i + insn->imm + 1);
12006 return -EFAULT;
12007 }
12008 /* temporarily remember subprog id inside insn instead of
12009 * aux_data, since next loop will split up all insns into funcs
12010 */
12011 insn->off = subprog;
12012 /* remember original imm in case JIT fails and fallback
12013 * to interpreter will be needed
12014 */
12015 env->insn_aux_data[i].call_imm = insn->imm;
12016 /* point imm to __bpf_call_base+1 from JITs point of view */
12017 insn->imm = 1;
12018 }
12019
12020 err = bpf_prog_alloc_jited_linfo(prog);
12021 if (err)
12022 goto out_undo_insn;
12023
12024 err = -ENOMEM;
12025 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12026 if (!func)
12027 goto out_undo_insn;
12028
12029 for (i = 0; i < env->subprog_cnt; i++) {
12030 subprog_start = subprog_end;
12031 subprog_end = env->subprog_info[i + 1].start;
12032
12033 len = subprog_end - subprog_start;
12034 /* BPF_PROG_RUN doesn't call subprogs directly,
12035 * hence main prog stats include the runtime of subprogs.
12036 * subprogs don't have IDs and not reachable via prog_get_next_id
12037 * func[i]->stats will never be accessed and stays NULL
12038 */
12039 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12040 if (!func[i])
12041 goto out_free;
12042 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12043 len * sizeof(struct bpf_insn));
12044 func[i]->type = prog->type;
12045 func[i]->len = len;
12046 if (bpf_prog_calc_tag(func[i]))
12047 goto out_free;
12048 func[i]->is_func = 1;
12049 func[i]->aux->func_idx = i;
12050 /* the btf and func_info will be freed only at prog->aux */
12051 func[i]->aux->btf = prog->aux->btf;
12052 func[i]->aux->func_info = prog->aux->func_info;
12053
12054 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12055 u32 insn_idx = prog->aux->poke_tab[j].insn_idx;
12056 int ret;
12057
12058 if (!(insn_idx >= subprog_start &&
12059 insn_idx <= subprog_end))
12060 continue;
12061
12062 ret = bpf_jit_add_poke_descriptor(func[i],
12063 &prog->aux->poke_tab[j]);
12064 if (ret < 0) {
12065 verbose(env, "adding tail call poke descriptor failed\n");
12066 goto out_free;
12067 }
12068
12069 func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
12070
12071 map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
12072 ret = map_ptr->ops->map_poke_track(map_ptr, func[i]->aux);
12073 if (ret < 0) {
12074 verbose(env, "tracking tail call prog failed\n");
12075 goto out_free;
12076 }
12077 }
12078
12079 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12080 * Long term would need debug info to populate names
12081 */
12082 func[i]->aux->name[0] = 'F';
12083 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12084 func[i]->jit_requested = 1;
12085 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12086 func[i]->aux->linfo = prog->aux->linfo;
12087 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12088 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12089 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12090 num_exentries = 0;
12091 insn = func[i]->insnsi;
12092 for (j = 0; j < func[i]->len; j++, insn++) {
12093 if (BPF_CLASS(insn->code) == BPF_LDX &&
12094 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12095 num_exentries++;
12096 }
12097 func[i]->aux->num_exentries = num_exentries;
12098 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12099 func[i] = bpf_int_jit_compile(func[i]);
12100 if (!func[i]->jited) {
12101 err = -ENOTSUPP;
12102 goto out_free;
12103 }
12104 cond_resched();
12105 }
12106
12107 /* Untrack main program's aux structs so that during map_poke_run()
12108 * we will not stumble upon the unfilled poke descriptors; each
12109 * of the main program's poke descs got distributed across subprogs
12110 * and got tracked onto map, so we are sure that none of them will
12111 * be missed after the operation below
12112 */
12113 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12114 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12115
12116 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12117 }
12118
12119 /* at this point all bpf functions were successfully JITed
12120 * now populate all bpf_calls with correct addresses and
12121 * run last pass of JIT
12122 */
12123 for (i = 0; i < env->subprog_cnt; i++) {
12124 insn = func[i]->insnsi;
12125 for (j = 0; j < func[i]->len; j++, insn++) {
12126 if (bpf_pseudo_func(insn)) {
12127 subprog = insn[1].imm;
12128 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12129 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12130 continue;
12131 }
12132 if (!bpf_pseudo_call(insn))
12133 continue;
12134 subprog = insn->off;
12135 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12136 __bpf_call_base;
12137 }
12138
12139 /* we use the aux data to keep a list of the start addresses
12140 * of the JITed images for each function in the program
12141 *
12142 * for some architectures, such as powerpc64, the imm field
12143 * might not be large enough to hold the offset of the start
12144 * address of the callee's JITed image from __bpf_call_base
12145 *
12146 * in such cases, we can lookup the start address of a callee
12147 * by using its subprog id, available from the off field of
12148 * the call instruction, as an index for this list
12149 */
12150 func[i]->aux->func = func;
12151 func[i]->aux->func_cnt = env->subprog_cnt;
12152 }
12153 for (i = 0; i < env->subprog_cnt; i++) {
12154 old_bpf_func = func[i]->bpf_func;
12155 tmp = bpf_int_jit_compile(func[i]);
12156 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12157 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12158 err = -ENOTSUPP;
12159 goto out_free;
12160 }
12161 cond_resched();
12162 }
12163
12164 /* finally lock prog and jit images for all functions and
12165 * populate kallsysm
12166 */
12167 for (i = 0; i < env->subprog_cnt; i++) {
12168 bpf_prog_lock_ro(func[i]);
12169 bpf_prog_kallsyms_add(func[i]);
12170 }
12171
12172 /* Last step: make now unused interpreter insns from main
12173 * prog consistent for later dump requests, so they can
12174 * later look the same as if they were interpreted only.
12175 */
12176 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12177 if (bpf_pseudo_func(insn)) {
12178 insn[0].imm = env->insn_aux_data[i].call_imm;
12179 insn[1].imm = find_subprog(env, i + insn[0].imm + 1);
12180 continue;
12181 }
12182 if (!bpf_pseudo_call(insn))
12183 continue;
12184 insn->off = env->insn_aux_data[i].call_imm;
12185 subprog = find_subprog(env, i + insn->off + 1);
12186 insn->imm = subprog;
12187 }
12188
12189 prog->jited = 1;
12190 prog->bpf_func = func[0]->bpf_func;
12191 prog->aux->func = func;
12192 prog->aux->func_cnt = env->subprog_cnt;
12193 bpf_prog_jit_attempt_done(prog);
12194 return 0;
12195 out_free:
12196 for (i = 0; i < env->subprog_cnt; i++) {
12197 if (!func[i])
12198 continue;
12199
12200 for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
12201 map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
12202 map_ptr->ops->map_poke_untrack(map_ptr, func[i]->aux);
12203 }
12204 bpf_jit_free(func[i]);
12205 }
12206 kfree(func);
12207 out_undo_insn:
12208 /* cleanup main prog to be interpreted */
12209 prog->jit_requested = 0;
12210 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12211 if (!bpf_pseudo_call(insn))
12212 continue;
12213 insn->off = 0;
12214 insn->imm = env->insn_aux_data[i].call_imm;
12215 }
12216 bpf_prog_jit_attempt_done(prog);
12217 return err;
12218 }
12219
fixup_call_args(struct bpf_verifier_env * env)12220 static int fixup_call_args(struct bpf_verifier_env *env)
12221 {
12222 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12223 struct bpf_prog *prog = env->prog;
12224 struct bpf_insn *insn = prog->insnsi;
12225 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12226 int i, depth;
12227 #endif
12228 int err = 0;
12229
12230 if (env->prog->jit_requested &&
12231 !bpf_prog_is_dev_bound(env->prog->aux)) {
12232 err = jit_subprogs(env);
12233 if (err == 0)
12234 return 0;
12235 if (err == -EFAULT)
12236 return err;
12237 }
12238 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12239 if (has_kfunc_call) {
12240 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12241 return -EINVAL;
12242 }
12243 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12244 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12245 * have to be rejected, since interpreter doesn't support them yet.
12246 */
12247 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12248 return -EINVAL;
12249 }
12250 for (i = 0; i < prog->len; i++, insn++) {
12251 if (bpf_pseudo_func(insn)) {
12252 /* When JIT fails the progs with callback calls
12253 * have to be rejected, since interpreter doesn't support them yet.
12254 */
12255 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12256 return -EINVAL;
12257 }
12258
12259 if (!bpf_pseudo_call(insn))
12260 continue;
12261 depth = get_callee_stack_depth(env, insn, i);
12262 if (depth < 0)
12263 return depth;
12264 bpf_patch_call_args(insn, depth);
12265 }
12266 err = 0;
12267 #endif
12268 return err;
12269 }
12270
fixup_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn)12271 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12272 struct bpf_insn *insn)
12273 {
12274 const struct bpf_kfunc_desc *desc;
12275
12276 /* insn->imm has the btf func_id. Replace it with
12277 * an address (relative to __bpf_base_call).
12278 */
12279 desc = find_kfunc_desc(env->prog, insn->imm);
12280 if (!desc) {
12281 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12282 insn->imm);
12283 return -EFAULT;
12284 }
12285
12286 insn->imm = desc->imm;
12287
12288 return 0;
12289 }
12290
12291 /* Do various post-verification rewrites in a single program pass.
12292 * These rewrites simplify JIT and interpreter implementations.
12293 */
do_misc_fixups(struct bpf_verifier_env * env)12294 static int do_misc_fixups(struct bpf_verifier_env *env)
12295 {
12296 struct bpf_prog *prog = env->prog;
12297 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12298 struct bpf_insn *insn = prog->insnsi;
12299 const struct bpf_func_proto *fn;
12300 const int insn_cnt = prog->len;
12301 const struct bpf_map_ops *ops;
12302 struct bpf_insn_aux_data *aux;
12303 struct bpf_insn insn_buf[16];
12304 struct bpf_prog *new_prog;
12305 struct bpf_map *map_ptr;
12306 int i, ret, cnt, delta = 0;
12307
12308 for (i = 0; i < insn_cnt; i++, insn++) {
12309 /* Make divide-by-zero exceptions impossible. */
12310 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12311 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12312 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12313 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12314 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12315 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12316 struct bpf_insn *patchlet;
12317 struct bpf_insn chk_and_div[] = {
12318 /* [R,W]x div 0 -> 0 */
12319 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12320 BPF_JNE | BPF_K, insn->src_reg,
12321 0, 2, 0),
12322 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12323 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12324 *insn,
12325 };
12326 struct bpf_insn chk_and_mod[] = {
12327 /* [R,W]x mod 0 -> [R,W]x */
12328 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12329 BPF_JEQ | BPF_K, insn->src_reg,
12330 0, 1 + (is64 ? 0 : 1), 0),
12331 *insn,
12332 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12333 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12334 };
12335
12336 patchlet = isdiv ? chk_and_div : chk_and_mod;
12337 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12338 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12339
12340 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12341 if (!new_prog)
12342 return -ENOMEM;
12343
12344 delta += cnt - 1;
12345 env->prog = prog = new_prog;
12346 insn = new_prog->insnsi + i + delta;
12347 continue;
12348 }
12349
12350 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12351 if (BPF_CLASS(insn->code) == BPF_LD &&
12352 (BPF_MODE(insn->code) == BPF_ABS ||
12353 BPF_MODE(insn->code) == BPF_IND)) {
12354 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12355 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12356 verbose(env, "bpf verifier is misconfigured\n");
12357 return -EINVAL;
12358 }
12359
12360 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12361 if (!new_prog)
12362 return -ENOMEM;
12363
12364 delta += cnt - 1;
12365 env->prog = prog = new_prog;
12366 insn = new_prog->insnsi + i + delta;
12367 continue;
12368 }
12369
12370 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12371 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12372 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12373 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12374 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12375 struct bpf_insn *patch = &insn_buf[0];
12376 bool issrc, isneg, isimm;
12377 u32 off_reg;
12378
12379 aux = &env->insn_aux_data[i + delta];
12380 if (!aux->alu_state ||
12381 aux->alu_state == BPF_ALU_NON_POINTER)
12382 continue;
12383
12384 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12385 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12386 BPF_ALU_SANITIZE_SRC;
12387 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12388
12389 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12390 if (isimm) {
12391 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12392 } else {
12393 if (isneg)
12394 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12395 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12396 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12397 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12398 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12399 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12400 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12401 }
12402 if (!issrc)
12403 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12404 insn->src_reg = BPF_REG_AX;
12405 if (isneg)
12406 insn->code = insn->code == code_add ?
12407 code_sub : code_add;
12408 *patch++ = *insn;
12409 if (issrc && isneg && !isimm)
12410 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12411 cnt = patch - insn_buf;
12412
12413 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12414 if (!new_prog)
12415 return -ENOMEM;
12416
12417 delta += cnt - 1;
12418 env->prog = prog = new_prog;
12419 insn = new_prog->insnsi + i + delta;
12420 continue;
12421 }
12422
12423 if (insn->code != (BPF_JMP | BPF_CALL))
12424 continue;
12425 if (insn->src_reg == BPF_PSEUDO_CALL)
12426 continue;
12427 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12428 ret = fixup_kfunc_call(env, insn);
12429 if (ret)
12430 return ret;
12431 continue;
12432 }
12433
12434 if (insn->imm == BPF_FUNC_get_route_realm)
12435 prog->dst_needed = 1;
12436 if (insn->imm == BPF_FUNC_get_prandom_u32)
12437 bpf_user_rnd_init_once();
12438 if (insn->imm == BPF_FUNC_override_return)
12439 prog->kprobe_override = 1;
12440 if (insn->imm == BPF_FUNC_tail_call) {
12441 /* If we tail call into other programs, we
12442 * cannot make any assumptions since they can
12443 * be replaced dynamically during runtime in
12444 * the program array.
12445 */
12446 prog->cb_access = 1;
12447 if (!allow_tail_call_in_subprogs(env))
12448 prog->aux->stack_depth = MAX_BPF_STACK;
12449 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12450
12451 /* mark bpf_tail_call as different opcode to avoid
12452 * conditional branch in the interpeter for every normal
12453 * call and to prevent accidental JITing by JIT compiler
12454 * that doesn't support bpf_tail_call yet
12455 */
12456 insn->imm = 0;
12457 insn->code = BPF_JMP | BPF_TAIL_CALL;
12458
12459 aux = &env->insn_aux_data[i + delta];
12460 if (env->bpf_capable && !expect_blinding &&
12461 prog->jit_requested &&
12462 !bpf_map_key_poisoned(aux) &&
12463 !bpf_map_ptr_poisoned(aux) &&
12464 !bpf_map_ptr_unpriv(aux)) {
12465 struct bpf_jit_poke_descriptor desc = {
12466 .reason = BPF_POKE_REASON_TAIL_CALL,
12467 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12468 .tail_call.key = bpf_map_key_immediate(aux),
12469 .insn_idx = i + delta,
12470 };
12471
12472 ret = bpf_jit_add_poke_descriptor(prog, &desc);
12473 if (ret < 0) {
12474 verbose(env, "adding tail call poke descriptor failed\n");
12475 return ret;
12476 }
12477
12478 insn->imm = ret + 1;
12479 continue;
12480 }
12481
12482 if (!bpf_map_ptr_unpriv(aux))
12483 continue;
12484
12485 /* instead of changing every JIT dealing with tail_call
12486 * emit two extra insns:
12487 * if (index >= max_entries) goto out;
12488 * index &= array->index_mask;
12489 * to avoid out-of-bounds cpu speculation
12490 */
12491 if (bpf_map_ptr_poisoned(aux)) {
12492 verbose(env, "tail_call abusing map_ptr\n");
12493 return -EINVAL;
12494 }
12495
12496 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12497 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12498 map_ptr->max_entries, 2);
12499 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12500 container_of(map_ptr,
12501 struct bpf_array,
12502 map)->index_mask);
12503 insn_buf[2] = *insn;
12504 cnt = 3;
12505 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12506 if (!new_prog)
12507 return -ENOMEM;
12508
12509 delta += cnt - 1;
12510 env->prog = prog = new_prog;
12511 insn = new_prog->insnsi + i + delta;
12512 continue;
12513 }
12514
12515 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12516 * and other inlining handlers are currently limited to 64 bit
12517 * only.
12518 */
12519 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12520 (insn->imm == BPF_FUNC_map_lookup_elem ||
12521 insn->imm == BPF_FUNC_map_update_elem ||
12522 insn->imm == BPF_FUNC_map_delete_elem ||
12523 insn->imm == BPF_FUNC_map_push_elem ||
12524 insn->imm == BPF_FUNC_map_pop_elem ||
12525 insn->imm == BPF_FUNC_map_peek_elem ||
12526 insn->imm == BPF_FUNC_redirect_map)) {
12527 aux = &env->insn_aux_data[i + delta];
12528 if (bpf_map_ptr_poisoned(aux))
12529 goto patch_call_imm;
12530
12531 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12532 ops = map_ptr->ops;
12533 if (insn->imm == BPF_FUNC_map_lookup_elem &&
12534 ops->map_gen_lookup) {
12535 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12536 if (cnt == -EOPNOTSUPP)
12537 goto patch_map_ops_generic;
12538 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12539 verbose(env, "bpf verifier is misconfigured\n");
12540 return -EINVAL;
12541 }
12542
12543 new_prog = bpf_patch_insn_data(env, i + delta,
12544 insn_buf, cnt);
12545 if (!new_prog)
12546 return -ENOMEM;
12547
12548 delta += cnt - 1;
12549 env->prog = prog = new_prog;
12550 insn = new_prog->insnsi + i + delta;
12551 continue;
12552 }
12553
12554 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12555 (void *(*)(struct bpf_map *map, void *key))NULL));
12556 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12557 (int (*)(struct bpf_map *map, void *key))NULL));
12558 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12559 (int (*)(struct bpf_map *map, void *key, void *value,
12560 u64 flags))NULL));
12561 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
12562 (int (*)(struct bpf_map *map, void *value,
12563 u64 flags))NULL));
12564 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
12565 (int (*)(struct bpf_map *map, void *value))NULL));
12566 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
12567 (int (*)(struct bpf_map *map, void *value))NULL));
12568 BUILD_BUG_ON(!__same_type(ops->map_redirect,
12569 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
12570
12571 patch_map_ops_generic:
12572 switch (insn->imm) {
12573 case BPF_FUNC_map_lookup_elem:
12574 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
12575 __bpf_call_base;
12576 continue;
12577 case BPF_FUNC_map_update_elem:
12578 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
12579 __bpf_call_base;
12580 continue;
12581 case BPF_FUNC_map_delete_elem:
12582 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
12583 __bpf_call_base;
12584 continue;
12585 case BPF_FUNC_map_push_elem:
12586 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
12587 __bpf_call_base;
12588 continue;
12589 case BPF_FUNC_map_pop_elem:
12590 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
12591 __bpf_call_base;
12592 continue;
12593 case BPF_FUNC_map_peek_elem:
12594 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
12595 __bpf_call_base;
12596 continue;
12597 case BPF_FUNC_redirect_map:
12598 insn->imm = BPF_CAST_CALL(ops->map_redirect) -
12599 __bpf_call_base;
12600 continue;
12601 }
12602
12603 goto patch_call_imm;
12604 }
12605
12606 /* Implement bpf_jiffies64 inline. */
12607 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12608 insn->imm == BPF_FUNC_jiffies64) {
12609 struct bpf_insn ld_jiffies_addr[2] = {
12610 BPF_LD_IMM64(BPF_REG_0,
12611 (unsigned long)&jiffies),
12612 };
12613
12614 insn_buf[0] = ld_jiffies_addr[0];
12615 insn_buf[1] = ld_jiffies_addr[1];
12616 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
12617 BPF_REG_0, 0);
12618 cnt = 3;
12619
12620 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
12621 cnt);
12622 if (!new_prog)
12623 return -ENOMEM;
12624
12625 delta += cnt - 1;
12626 env->prog = prog = new_prog;
12627 insn = new_prog->insnsi + i + delta;
12628 continue;
12629 }
12630
12631 patch_call_imm:
12632 fn = env->ops->get_func_proto(insn->imm, env->prog);
12633 /* all functions that have prototype and verifier allowed
12634 * programs to call them, must be real in-kernel functions
12635 */
12636 if (!fn->func) {
12637 verbose(env,
12638 "kernel subsystem misconfigured func %s#%d\n",
12639 func_id_name(insn->imm), insn->imm);
12640 return -EFAULT;
12641 }
12642 insn->imm = fn->func - __bpf_call_base;
12643 }
12644
12645 /* Since poke tab is now finalized, publish aux to tracker. */
12646 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12647 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12648 if (!map_ptr->ops->map_poke_track ||
12649 !map_ptr->ops->map_poke_untrack ||
12650 !map_ptr->ops->map_poke_run) {
12651 verbose(env, "bpf verifier is misconfigured\n");
12652 return -EINVAL;
12653 }
12654
12655 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
12656 if (ret < 0) {
12657 verbose(env, "tracking tail call prog failed\n");
12658 return ret;
12659 }
12660 }
12661
12662 sort_kfunc_descs_by_imm(env->prog);
12663
12664 return 0;
12665 }
12666
free_states(struct bpf_verifier_env * env)12667 static void free_states(struct bpf_verifier_env *env)
12668 {
12669 struct bpf_verifier_state_list *sl, *sln;
12670 int i;
12671
12672 sl = env->free_list;
12673 while (sl) {
12674 sln = sl->next;
12675 free_verifier_state(&sl->state, false);
12676 kfree(sl);
12677 sl = sln;
12678 }
12679 env->free_list = NULL;
12680
12681 if (!env->explored_states)
12682 return;
12683
12684 for (i = 0; i < state_htab_size(env); i++) {
12685 sl = env->explored_states[i];
12686
12687 while (sl) {
12688 sln = sl->next;
12689 free_verifier_state(&sl->state, false);
12690 kfree(sl);
12691 sl = sln;
12692 }
12693 env->explored_states[i] = NULL;
12694 }
12695 }
12696
12697 /* The verifier is using insn_aux_data[] to store temporary data during
12698 * verification and to store information for passes that run after the
12699 * verification like dead code sanitization. do_check_common() for subprogram N
12700 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
12701 * temporary data after do_check_common() finds that subprogram N cannot be
12702 * verified independently. pass_cnt counts the number of times
12703 * do_check_common() was run and insn->aux->seen tells the pass number
12704 * insn_aux_data was touched. These variables are compared to clear temporary
12705 * data from failed pass. For testing and experiments do_check_common() can be
12706 * run multiple times even when prior attempt to verify is unsuccessful.
12707 */
sanitize_insn_aux_data(struct bpf_verifier_env * env)12708 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
12709 {
12710 struct bpf_insn *insn = env->prog->insnsi;
12711 struct bpf_insn_aux_data *aux;
12712 int i, class;
12713
12714 for (i = 0; i < env->prog->len; i++) {
12715 class = BPF_CLASS(insn[i].code);
12716 if (class != BPF_LDX && class != BPF_STX)
12717 continue;
12718 aux = &env->insn_aux_data[i];
12719 if (aux->seen != env->pass_cnt)
12720 continue;
12721 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
12722 }
12723 }
12724
do_check_common(struct bpf_verifier_env * env,int subprog)12725 static int do_check_common(struct bpf_verifier_env *env, int subprog)
12726 {
12727 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12728 struct bpf_verifier_state *state;
12729 struct bpf_reg_state *regs;
12730 int ret, i;
12731
12732 env->prev_linfo = NULL;
12733 env->pass_cnt++;
12734
12735 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
12736 if (!state)
12737 return -ENOMEM;
12738 state->curframe = 0;
12739 state->speculative = false;
12740 state->branches = 1;
12741 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
12742 if (!state->frame[0]) {
12743 kfree(state);
12744 return -ENOMEM;
12745 }
12746 env->cur_state = state;
12747 init_func_state(env, state->frame[0],
12748 BPF_MAIN_FUNC /* callsite */,
12749 0 /* frameno */,
12750 subprog);
12751
12752 regs = state->frame[state->curframe]->regs;
12753 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
12754 ret = btf_prepare_func_args(env, subprog, regs);
12755 if (ret)
12756 goto out;
12757 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
12758 if (regs[i].type == PTR_TO_CTX)
12759 mark_reg_known_zero(env, regs, i);
12760 else if (regs[i].type == SCALAR_VALUE)
12761 mark_reg_unknown(env, regs, i);
12762 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
12763 const u32 mem_size = regs[i].mem_size;
12764
12765 mark_reg_known_zero(env, regs, i);
12766 regs[i].mem_size = mem_size;
12767 regs[i].id = ++env->id_gen;
12768 }
12769 }
12770 } else {
12771 /* 1st arg to a function */
12772 regs[BPF_REG_1].type = PTR_TO_CTX;
12773 mark_reg_known_zero(env, regs, BPF_REG_1);
12774 ret = btf_check_subprog_arg_match(env, subprog, regs);
12775 if (ret == -EFAULT)
12776 /* unlikely verifier bug. abort.
12777 * ret == 0 and ret < 0 are sadly acceptable for
12778 * main() function due to backward compatibility.
12779 * Like socket filter program may be written as:
12780 * int bpf_prog(struct pt_regs *ctx)
12781 * and never dereference that ctx in the program.
12782 * 'struct pt_regs' is a type mismatch for socket
12783 * filter that should be using 'struct __sk_buff'.
12784 */
12785 goto out;
12786 }
12787
12788 ret = do_check(env);
12789 out:
12790 /* check for NULL is necessary, since cur_state can be freed inside
12791 * do_check() under memory pressure.
12792 */
12793 if (env->cur_state) {
12794 free_verifier_state(env->cur_state, true);
12795 env->cur_state = NULL;
12796 }
12797 while (!pop_stack(env, NULL, NULL, false));
12798 if (!ret && pop_log)
12799 bpf_vlog_reset(&env->log, 0);
12800 free_states(env);
12801 if (ret)
12802 /* clean aux data in case subprog was rejected */
12803 sanitize_insn_aux_data(env);
12804 return ret;
12805 }
12806
12807 /* Verify all global functions in a BPF program one by one based on their BTF.
12808 * All global functions must pass verification. Otherwise the whole program is rejected.
12809 * Consider:
12810 * int bar(int);
12811 * int foo(int f)
12812 * {
12813 * return bar(f);
12814 * }
12815 * int bar(int b)
12816 * {
12817 * ...
12818 * }
12819 * foo() will be verified first for R1=any_scalar_value. During verification it
12820 * will be assumed that bar() already verified successfully and call to bar()
12821 * from foo() will be checked for type match only. Later bar() will be verified
12822 * independently to check that it's safe for R1=any_scalar_value.
12823 */
do_check_subprogs(struct bpf_verifier_env * env)12824 static int do_check_subprogs(struct bpf_verifier_env *env)
12825 {
12826 struct bpf_prog_aux *aux = env->prog->aux;
12827 int i, ret;
12828
12829 if (!aux->func_info)
12830 return 0;
12831
12832 for (i = 1; i < env->subprog_cnt; i++) {
12833 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
12834 continue;
12835 env->insn_idx = env->subprog_info[i].start;
12836 WARN_ON_ONCE(env->insn_idx == 0);
12837 ret = do_check_common(env, i);
12838 if (ret) {
12839 return ret;
12840 } else if (env->log.level & BPF_LOG_LEVEL) {
12841 verbose(env,
12842 "Func#%d is safe for any args that match its prototype\n",
12843 i);
12844 }
12845 }
12846 return 0;
12847 }
12848
do_check_main(struct bpf_verifier_env * env)12849 static int do_check_main(struct bpf_verifier_env *env)
12850 {
12851 int ret;
12852
12853 env->insn_idx = 0;
12854 ret = do_check_common(env, 0);
12855 if (!ret)
12856 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
12857 return ret;
12858 }
12859
12860
print_verification_stats(struct bpf_verifier_env * env)12861 static void print_verification_stats(struct bpf_verifier_env *env)
12862 {
12863 int i;
12864
12865 if (env->log.level & BPF_LOG_STATS) {
12866 verbose(env, "verification time %lld usec\n",
12867 div_u64(env->verification_time, 1000));
12868 verbose(env, "stack depth ");
12869 for (i = 0; i < env->subprog_cnt; i++) {
12870 u32 depth = env->subprog_info[i].stack_depth;
12871
12872 verbose(env, "%d", depth);
12873 if (i + 1 < env->subprog_cnt)
12874 verbose(env, "+");
12875 }
12876 verbose(env, "\n");
12877 }
12878 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
12879 "total_states %d peak_states %d mark_read %d\n",
12880 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
12881 env->max_states_per_insn, env->total_states,
12882 env->peak_states, env->longest_mark_read_walk);
12883 }
12884
check_struct_ops_btf_id(struct bpf_verifier_env * env)12885 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
12886 {
12887 const struct btf_type *t, *func_proto;
12888 const struct bpf_struct_ops *st_ops;
12889 const struct btf_member *member;
12890 struct bpf_prog *prog = env->prog;
12891 u32 btf_id, member_idx;
12892 const char *mname;
12893
12894 if (!prog->gpl_compatible) {
12895 verbose(env, "struct ops programs must have a GPL compatible license\n");
12896 return -EINVAL;
12897 }
12898
12899 btf_id = prog->aux->attach_btf_id;
12900 st_ops = bpf_struct_ops_find(btf_id);
12901 if (!st_ops) {
12902 verbose(env, "attach_btf_id %u is not a supported struct\n",
12903 btf_id);
12904 return -ENOTSUPP;
12905 }
12906
12907 t = st_ops->type;
12908 member_idx = prog->expected_attach_type;
12909 if (member_idx >= btf_type_vlen(t)) {
12910 verbose(env, "attach to invalid member idx %u of struct %s\n",
12911 member_idx, st_ops->name);
12912 return -EINVAL;
12913 }
12914
12915 member = &btf_type_member(t)[member_idx];
12916 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
12917 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
12918 NULL);
12919 if (!func_proto) {
12920 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
12921 mname, member_idx, st_ops->name);
12922 return -EINVAL;
12923 }
12924
12925 if (st_ops->check_member) {
12926 int err = st_ops->check_member(t, member);
12927
12928 if (err) {
12929 verbose(env, "attach to unsupported member %s of struct %s\n",
12930 mname, st_ops->name);
12931 return err;
12932 }
12933 }
12934
12935 prog->aux->attach_func_proto = func_proto;
12936 prog->aux->attach_func_name = mname;
12937 env->ops = st_ops->verifier_ops;
12938
12939 return 0;
12940 }
12941 #define SECURITY_PREFIX "security_"
12942
check_attach_modify_return(unsigned long addr,const char * func_name)12943 static int check_attach_modify_return(unsigned long addr, const char *func_name)
12944 {
12945 if (within_error_injection_list(addr) ||
12946 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
12947 return 0;
12948
12949 return -EINVAL;
12950 }
12951
12952 /* list of non-sleepable functions that are otherwise on
12953 * ALLOW_ERROR_INJECTION list
12954 */
12955 BTF_SET_START(btf_non_sleepable_error_inject)
12956 /* Three functions below can be called from sleepable and non-sleepable context.
12957 * Assume non-sleepable from bpf safety point of view.
12958 */
BTF_ID(func,__add_to_page_cache_locked)12959 BTF_ID(func, __add_to_page_cache_locked)
12960 BTF_ID(func, should_fail_alloc_page)
12961 BTF_ID(func, should_failslab)
12962 BTF_SET_END(btf_non_sleepable_error_inject)
12963
12964 static int check_non_sleepable_error_inject(u32 btf_id)
12965 {
12966 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
12967 }
12968
bpf_check_attach_target(struct bpf_verifier_log * log,const struct bpf_prog * prog,const struct bpf_prog * tgt_prog,u32 btf_id,struct bpf_attach_target_info * tgt_info)12969 int bpf_check_attach_target(struct bpf_verifier_log *log,
12970 const struct bpf_prog *prog,
12971 const struct bpf_prog *tgt_prog,
12972 u32 btf_id,
12973 struct bpf_attach_target_info *tgt_info)
12974 {
12975 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
12976 const char prefix[] = "btf_trace_";
12977 int ret = 0, subprog = -1, i;
12978 const struct btf_type *t;
12979 bool conservative = true;
12980 const char *tname;
12981 struct btf *btf;
12982 long addr = 0;
12983
12984 if (!btf_id) {
12985 bpf_log(log, "Tracing programs must provide btf_id\n");
12986 return -EINVAL;
12987 }
12988 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
12989 if (!btf) {
12990 bpf_log(log,
12991 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
12992 return -EINVAL;
12993 }
12994 t = btf_type_by_id(btf, btf_id);
12995 if (!t) {
12996 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
12997 return -EINVAL;
12998 }
12999 tname = btf_name_by_offset(btf, t->name_off);
13000 if (!tname) {
13001 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13002 return -EINVAL;
13003 }
13004 if (tgt_prog) {
13005 struct bpf_prog_aux *aux = tgt_prog->aux;
13006
13007 for (i = 0; i < aux->func_info_cnt; i++)
13008 if (aux->func_info[i].type_id == btf_id) {
13009 subprog = i;
13010 break;
13011 }
13012 if (subprog == -1) {
13013 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13014 return -EINVAL;
13015 }
13016 conservative = aux->func_info_aux[subprog].unreliable;
13017 if (prog_extension) {
13018 if (conservative) {
13019 bpf_log(log,
13020 "Cannot replace static functions\n");
13021 return -EINVAL;
13022 }
13023 if (!prog->jit_requested) {
13024 bpf_log(log,
13025 "Extension programs should be JITed\n");
13026 return -EINVAL;
13027 }
13028 }
13029 if (!tgt_prog->jited) {
13030 bpf_log(log, "Can attach to only JITed progs\n");
13031 return -EINVAL;
13032 }
13033 if (tgt_prog->type == prog->type) {
13034 /* Cannot fentry/fexit another fentry/fexit program.
13035 * Cannot attach program extension to another extension.
13036 * It's ok to attach fentry/fexit to extension program.
13037 */
13038 bpf_log(log, "Cannot recursively attach\n");
13039 return -EINVAL;
13040 }
13041 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13042 prog_extension &&
13043 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13044 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13045 /* Program extensions can extend all program types
13046 * except fentry/fexit. The reason is the following.
13047 * The fentry/fexit programs are used for performance
13048 * analysis, stats and can be attached to any program
13049 * type except themselves. When extension program is
13050 * replacing XDP function it is necessary to allow
13051 * performance analysis of all functions. Both original
13052 * XDP program and its program extension. Hence
13053 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13054 * allowed. If extending of fentry/fexit was allowed it
13055 * would be possible to create long call chain
13056 * fentry->extension->fentry->extension beyond
13057 * reasonable stack size. Hence extending fentry is not
13058 * allowed.
13059 */
13060 bpf_log(log, "Cannot extend fentry/fexit\n");
13061 return -EINVAL;
13062 }
13063 } else {
13064 if (prog_extension) {
13065 bpf_log(log, "Cannot replace kernel functions\n");
13066 return -EINVAL;
13067 }
13068 }
13069
13070 switch (prog->expected_attach_type) {
13071 case BPF_TRACE_RAW_TP:
13072 if (tgt_prog) {
13073 bpf_log(log,
13074 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13075 return -EINVAL;
13076 }
13077 if (!btf_type_is_typedef(t)) {
13078 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13079 btf_id);
13080 return -EINVAL;
13081 }
13082 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13083 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13084 btf_id, tname);
13085 return -EINVAL;
13086 }
13087 tname += sizeof(prefix) - 1;
13088 t = btf_type_by_id(btf, t->type);
13089 if (!btf_type_is_ptr(t))
13090 /* should never happen in valid vmlinux build */
13091 return -EINVAL;
13092 t = btf_type_by_id(btf, t->type);
13093 if (!btf_type_is_func_proto(t))
13094 /* should never happen in valid vmlinux build */
13095 return -EINVAL;
13096
13097 break;
13098 case BPF_TRACE_ITER:
13099 if (!btf_type_is_func(t)) {
13100 bpf_log(log, "attach_btf_id %u is not a function\n",
13101 btf_id);
13102 return -EINVAL;
13103 }
13104 t = btf_type_by_id(btf, t->type);
13105 if (!btf_type_is_func_proto(t))
13106 return -EINVAL;
13107 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13108 if (ret)
13109 return ret;
13110 break;
13111 default:
13112 if (!prog_extension)
13113 return -EINVAL;
13114 fallthrough;
13115 case BPF_MODIFY_RETURN:
13116 case BPF_LSM_MAC:
13117 case BPF_TRACE_FENTRY:
13118 case BPF_TRACE_FEXIT:
13119 if (!btf_type_is_func(t)) {
13120 bpf_log(log, "attach_btf_id %u is not a function\n",
13121 btf_id);
13122 return -EINVAL;
13123 }
13124 if (prog_extension &&
13125 btf_check_type_match(log, prog, btf, t))
13126 return -EINVAL;
13127 t = btf_type_by_id(btf, t->type);
13128 if (!btf_type_is_func_proto(t))
13129 return -EINVAL;
13130
13131 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13132 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13133 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13134 return -EINVAL;
13135
13136 if (tgt_prog && conservative)
13137 t = NULL;
13138
13139 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13140 if (ret < 0)
13141 return ret;
13142
13143 if (tgt_prog) {
13144 if (subprog == 0)
13145 addr = (long) tgt_prog->bpf_func;
13146 else
13147 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13148 } else {
13149 addr = kallsyms_lookup_name(tname);
13150 if (!addr) {
13151 bpf_log(log,
13152 "The address of function %s cannot be found\n",
13153 tname);
13154 return -ENOENT;
13155 }
13156 }
13157
13158 if (prog->aux->sleepable) {
13159 ret = -EINVAL;
13160 switch (prog->type) {
13161 case BPF_PROG_TYPE_TRACING:
13162 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13163 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13164 */
13165 if (!check_non_sleepable_error_inject(btf_id) &&
13166 within_error_injection_list(addr))
13167 ret = 0;
13168 break;
13169 case BPF_PROG_TYPE_LSM:
13170 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13171 * Only some of them are sleepable.
13172 */
13173 if (bpf_lsm_is_sleepable_hook(btf_id))
13174 ret = 0;
13175 break;
13176 default:
13177 break;
13178 }
13179 if (ret) {
13180 bpf_log(log, "%s is not sleepable\n", tname);
13181 return ret;
13182 }
13183 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13184 if (tgt_prog) {
13185 bpf_log(log, "can't modify return codes of BPF programs\n");
13186 return -EINVAL;
13187 }
13188 ret = check_attach_modify_return(addr, tname);
13189 if (ret) {
13190 bpf_log(log, "%s() is not modifiable\n", tname);
13191 return ret;
13192 }
13193 }
13194
13195 break;
13196 }
13197 tgt_info->tgt_addr = addr;
13198 tgt_info->tgt_name = tname;
13199 tgt_info->tgt_type = t;
13200 return 0;
13201 }
13202
check_attach_btf_id(struct bpf_verifier_env * env)13203 static int check_attach_btf_id(struct bpf_verifier_env *env)
13204 {
13205 struct bpf_prog *prog = env->prog;
13206 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13207 struct bpf_attach_target_info tgt_info = {};
13208 u32 btf_id = prog->aux->attach_btf_id;
13209 struct bpf_trampoline *tr;
13210 int ret;
13211 u64 key;
13212
13213 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13214 prog->type != BPF_PROG_TYPE_LSM) {
13215 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13216 return -EINVAL;
13217 }
13218
13219 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13220 return check_struct_ops_btf_id(env);
13221
13222 if (prog->type != BPF_PROG_TYPE_TRACING &&
13223 prog->type != BPF_PROG_TYPE_LSM &&
13224 prog->type != BPF_PROG_TYPE_EXT)
13225 return 0;
13226
13227 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13228 if (ret)
13229 return ret;
13230
13231 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13232 /* to make freplace equivalent to their targets, they need to
13233 * inherit env->ops and expected_attach_type for the rest of the
13234 * verification
13235 */
13236 env->ops = bpf_verifier_ops[tgt_prog->type];
13237 prog->expected_attach_type = tgt_prog->expected_attach_type;
13238 }
13239
13240 /* store info about the attachment target that will be used later */
13241 prog->aux->attach_func_proto = tgt_info.tgt_type;
13242 prog->aux->attach_func_name = tgt_info.tgt_name;
13243
13244 if (tgt_prog) {
13245 prog->aux->saved_dst_prog_type = tgt_prog->type;
13246 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13247 }
13248
13249 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13250 prog->aux->attach_btf_trace = true;
13251 return 0;
13252 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13253 if (!bpf_iter_prog_supported(prog))
13254 return -EINVAL;
13255 return 0;
13256 }
13257
13258 if (prog->type == BPF_PROG_TYPE_LSM) {
13259 ret = bpf_lsm_verify_prog(&env->log, prog);
13260 if (ret < 0)
13261 return ret;
13262 }
13263
13264 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13265 tr = bpf_trampoline_get(key, &tgt_info);
13266 if (!tr)
13267 return -ENOMEM;
13268
13269 prog->aux->dst_trampoline = tr;
13270 return 0;
13271 }
13272
bpf_get_btf_vmlinux(void)13273 struct btf *bpf_get_btf_vmlinux(void)
13274 {
13275 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13276 mutex_lock(&bpf_verifier_lock);
13277 if (!btf_vmlinux)
13278 btf_vmlinux = btf_parse_vmlinux();
13279 mutex_unlock(&bpf_verifier_lock);
13280 }
13281 return btf_vmlinux;
13282 }
13283
bpf_check(struct bpf_prog ** prog,union bpf_attr * attr,union bpf_attr __user * uattr)13284 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
13285 union bpf_attr __user *uattr)
13286 {
13287 u64 start_time = ktime_get_ns();
13288 struct bpf_verifier_env *env;
13289 struct bpf_verifier_log *log;
13290 int i, len, ret = -EINVAL;
13291 bool is_priv;
13292
13293 /* no program is valid */
13294 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13295 return -EINVAL;
13296
13297 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13298 * allocate/free it every time bpf_check() is called
13299 */
13300 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13301 if (!env)
13302 return -ENOMEM;
13303 log = &env->log;
13304
13305 len = (*prog)->len;
13306 env->insn_aux_data =
13307 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13308 ret = -ENOMEM;
13309 if (!env->insn_aux_data)
13310 goto err_free_env;
13311 for (i = 0; i < len; i++)
13312 env->insn_aux_data[i].orig_idx = i;
13313 env->prog = *prog;
13314 env->ops = bpf_verifier_ops[env->prog->type];
13315 is_priv = bpf_capable();
13316
13317 bpf_get_btf_vmlinux();
13318
13319 /* grab the mutex to protect few globals used by verifier */
13320 if (!is_priv)
13321 mutex_lock(&bpf_verifier_lock);
13322
13323 if (attr->log_level || attr->log_buf || attr->log_size) {
13324 /* user requested verbose verifier output
13325 * and supplied buffer to store the verification trace
13326 */
13327 log->level = attr->log_level;
13328 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13329 log->len_total = attr->log_size;
13330
13331 ret = -EINVAL;
13332 /* log attributes have to be sane */
13333 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
13334 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
13335 goto err_unlock;
13336 }
13337
13338 if (IS_ERR(btf_vmlinux)) {
13339 /* Either gcc or pahole or kernel are broken. */
13340 verbose(env, "in-kernel BTF is malformed\n");
13341 ret = PTR_ERR(btf_vmlinux);
13342 goto skip_full_check;
13343 }
13344
13345 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13346 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13347 env->strict_alignment = true;
13348 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13349 env->strict_alignment = false;
13350
13351 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13352 env->allow_uninit_stack = bpf_allow_uninit_stack();
13353 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13354 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13355 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13356 env->bpf_capable = bpf_capable();
13357
13358 if (is_priv)
13359 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13360
13361 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13362 ret = bpf_prog_offload_verifier_prep(env->prog);
13363 if (ret)
13364 goto skip_full_check;
13365 }
13366
13367 env->explored_states = kvcalloc(state_htab_size(env),
13368 sizeof(struct bpf_verifier_state_list *),
13369 GFP_USER);
13370 ret = -ENOMEM;
13371 if (!env->explored_states)
13372 goto skip_full_check;
13373
13374 ret = add_subprog_and_kfunc(env);
13375 if (ret < 0)
13376 goto skip_full_check;
13377
13378 ret = check_subprogs(env);
13379 if (ret < 0)
13380 goto skip_full_check;
13381
13382 ret = check_btf_info(env, attr, uattr);
13383 if (ret < 0)
13384 goto skip_full_check;
13385
13386 ret = check_attach_btf_id(env);
13387 if (ret)
13388 goto skip_full_check;
13389
13390 ret = resolve_pseudo_ldimm64(env);
13391 if (ret < 0)
13392 goto skip_full_check;
13393
13394 ret = check_cfg(env);
13395 if (ret < 0)
13396 goto skip_full_check;
13397
13398 ret = do_check_subprogs(env);
13399 ret = ret ?: do_check_main(env);
13400
13401 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13402 ret = bpf_prog_offload_finalize(env);
13403
13404 skip_full_check:
13405 kvfree(env->explored_states);
13406
13407 if (ret == 0)
13408 ret = check_max_stack_depth(env);
13409
13410 /* instruction rewrites happen after this point */
13411 if (is_priv) {
13412 if (ret == 0)
13413 opt_hard_wire_dead_code_branches(env);
13414 if (ret == 0)
13415 ret = opt_remove_dead_code(env);
13416 if (ret == 0)
13417 ret = opt_remove_nops(env);
13418 } else {
13419 if (ret == 0)
13420 sanitize_dead_code(env);
13421 }
13422
13423 if (ret == 0)
13424 /* program is valid, convert *(u32*)(ctx + off) accesses */
13425 ret = convert_ctx_accesses(env);
13426
13427 if (ret == 0)
13428 ret = do_misc_fixups(env);
13429
13430 /* do 32-bit optimization after insn patching has done so those patched
13431 * insns could be handled correctly.
13432 */
13433 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13434 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13435 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13436 : false;
13437 }
13438
13439 if (ret == 0)
13440 ret = fixup_call_args(env);
13441
13442 env->verification_time = ktime_get_ns() - start_time;
13443 print_verification_stats(env);
13444
13445 if (log->level && bpf_verifier_log_full(log))
13446 ret = -ENOSPC;
13447 if (log->level && !log->ubuf) {
13448 ret = -EFAULT;
13449 goto err_release_maps;
13450 }
13451
13452 if (ret)
13453 goto err_release_maps;
13454
13455 if (env->used_map_cnt) {
13456 /* if program passed verifier, update used_maps in bpf_prog_info */
13457 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13458 sizeof(env->used_maps[0]),
13459 GFP_KERNEL);
13460
13461 if (!env->prog->aux->used_maps) {
13462 ret = -ENOMEM;
13463 goto err_release_maps;
13464 }
13465
13466 memcpy(env->prog->aux->used_maps, env->used_maps,
13467 sizeof(env->used_maps[0]) * env->used_map_cnt);
13468 env->prog->aux->used_map_cnt = env->used_map_cnt;
13469 }
13470 if (env->used_btf_cnt) {
13471 /* if program passed verifier, update used_btfs in bpf_prog_aux */
13472 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13473 sizeof(env->used_btfs[0]),
13474 GFP_KERNEL);
13475 if (!env->prog->aux->used_btfs) {
13476 ret = -ENOMEM;
13477 goto err_release_maps;
13478 }
13479
13480 memcpy(env->prog->aux->used_btfs, env->used_btfs,
13481 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13482 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13483 }
13484 if (env->used_map_cnt || env->used_btf_cnt) {
13485 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
13486 * bpf_ld_imm64 instructions
13487 */
13488 convert_pseudo_ld_imm64(env);
13489 }
13490
13491 adjust_btf_func(env);
13492
13493 err_release_maps:
13494 if (!env->prog->aux->used_maps)
13495 /* if we didn't copy map pointers into bpf_prog_info, release
13496 * them now. Otherwise free_used_maps() will release them.
13497 */
13498 release_maps(env);
13499 if (!env->prog->aux->used_btfs)
13500 release_btfs(env);
13501
13502 /* extension progs temporarily inherit the attach_type of their targets
13503 for verification purposes, so set it back to zero before returning
13504 */
13505 if (env->prog->type == BPF_PROG_TYPE_EXT)
13506 env->prog->expected_attach_type = 0;
13507
13508 *prog = env->prog;
13509 err_unlock:
13510 if (!is_priv)
13511 mutex_unlock(&bpf_verifier_lock);
13512 vfree(env->insn_aux_data);
13513 err_free_env:
13514 kfree(env);
13515 return ret;
13516 }
13517