?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
accel/amdxdna: use modern PM helpers
[drm/drm-misc.git] / kernel / bpf / core.c
bloba2327c4fdc8bc67d42bb5518f6499430f660676a
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Linux Socket Filter - Kernel level socket filtering
4 *
5 * Based on the design of the Berkeley Packet Filter. The new
6 * internal format has been designed by PLUMgrid:
7 *
8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
9 *
10 * Authors:
11 *
12 * Jay Schulist <jschlst@samba.org>
13 * Alexei Starovoitov <ast@plumgrid.com>
14 * Daniel Borkmann <dborkman@redhat.com>
15 *
16 * Andi Kleen - Fix a few bad bugs and races.
17 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
18 */
19
20 #include <uapi/linux/btf.h>
21 #include <linux/filter.h>
22 #include <linux/skbuff.h>
23 #include <linux/vmalloc.h>
24 #include <linux/prandom.h>
25 #include <linux/bpf.h>
26 #include <linux/btf.h>
27 #include <linux/objtool.h>
28 #include <linux/overflow.h>
29 #include <linux/rbtree_latch.h>
30 #include <linux/kallsyms.h>
31 #include <linux/rcupdate.h>
32 #include <linux/perf_event.h>
33 #include <linux/extable.h>
34 #include <linux/log2.h>
35 #include <linux/bpf_verifier.h>
36 #include <linux/nodemask.h>
37 #include <linux/nospec.h>
38 #include <linux/bpf_mem_alloc.h>
39 #include <linux/memcontrol.h>
40 #include <linux/execmem.h>
41
42 #include <asm/barrier.h>
43 #include <linux/unaligned.h>
44
45 /* Registers */
46 #define BPF_R0 regs[BPF_REG_0]
47 #define BPF_R1 regs[BPF_REG_1]
48 #define BPF_R2 regs[BPF_REG_2]
49 #define BPF_R3 regs[BPF_REG_3]
50 #define BPF_R4 regs[BPF_REG_4]
51 #define BPF_R5 regs[BPF_REG_5]
52 #define BPF_R6 regs[BPF_REG_6]
53 #define BPF_R7 regs[BPF_REG_7]
54 #define BPF_R8 regs[BPF_REG_8]
55 #define BPF_R9 regs[BPF_REG_9]
56 #define BPF_R10 regs[BPF_REG_10]
57
58 /* Named registers */
59 #define DST regs[insn->dst_reg]
60 #define SRC regs[insn->src_reg]
61 #define FP regs[BPF_REG_FP]
62 #define AX regs[BPF_REG_AX]
63 #define ARG1 regs[BPF_REG_ARG1]
64 #define CTX regs[BPF_REG_CTX]
65 #define OFF insn->off
66 #define IMM insn->imm
67
68 struct bpf_mem_alloc bpf_global_ma;
69 bool bpf_global_ma_set;
70
71 /* No hurry in this branch
72 *
73 * Exported for the bpf jit load helper.
74 */
75 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
76 {
77 u8 *ptr = NULL;
78
79 if (k >= SKF_NET_OFF) {
80 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
81 } else if (k >= SKF_LL_OFF) {
82 if (unlikely(!skb_mac_header_was_set(skb)))
83 return NULL;
84 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
85 }
86 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
87 return ptr;
88
89 return NULL;
90 }
91
92 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */
93 enum page_size_enum {
94 __PAGE_SIZE = PAGE_SIZE
95 };
96
97 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
98 {
99 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
100 struct bpf_prog_aux *aux;
101 struct bpf_prog *fp;
102
103 size = round_up(size, __PAGE_SIZE);
104 fp = __vmalloc(size, gfp_flags);
105 if (fp == NULL)
106 return NULL;
107
108 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
109 if (aux == NULL) {
110 vfree(fp);
111 return NULL;
112 }
113 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
114 if (!fp->active) {
115 vfree(fp);
116 kfree(aux);
117 return NULL;
118 }
119
120 fp->pages = size / PAGE_SIZE;
121 fp->aux = aux;
122 fp->aux->prog = fp;
123 fp->jit_requested = ebpf_jit_enabled();
124 fp->blinding_requested = bpf_jit_blinding_enabled(fp);
125 #ifdef CONFIG_CGROUP_BPF
126 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
127 #endif
128
129 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
130 #ifdef CONFIG_FINEIBT
131 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode);
132 #endif
133 mutex_init(&fp->aux->used_maps_mutex);
134 mutex_init(&fp->aux->ext_mutex);
135 mutex_init(&fp->aux->dst_mutex);
136
137 return fp;
138 }
139
140 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
141 {
142 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
143 struct bpf_prog *prog;
144 int cpu;
145
146 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
147 if (!prog)
148 return NULL;
149
150 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
151 if (!prog->stats) {
152 free_percpu(prog->active);
153 kfree(prog->aux);
154 vfree(prog);
155 return NULL;
156 }
157
158 for_each_possible_cpu(cpu) {
159 struct bpf_prog_stats *pstats;
160
161 pstats = per_cpu_ptr(prog->stats, cpu);
162 u64_stats_init(&pstats->syncp);
163 }
164 return prog;
165 }
166 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
167
168 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
169 {
170 if (!prog->aux->nr_linfo || !prog->jit_requested)
171 return 0;
172
173 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
174 sizeof(*prog->aux->jited_linfo),
175 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN));
176 if (!prog->aux->jited_linfo)
177 return -ENOMEM;
178
179 return 0;
180 }
181
182 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
183 {
184 if (prog->aux->jited_linfo &&
185 (!prog->jited || !prog->aux->jited_linfo[0])) {
186 kvfree(prog->aux->jited_linfo);
187 prog->aux->jited_linfo = NULL;
188 }
189
190 kfree(prog->aux->kfunc_tab);
191 prog->aux->kfunc_tab = NULL;
192 }
193
194 /* The jit engine is responsible to provide an array
195 * for insn_off to the jited_off mapping (insn_to_jit_off).
196 *
197 * The idx to this array is the insn_off. Hence, the insn_off
198 * here is relative to the prog itself instead of the main prog.
199 * This array has one entry for each xlated bpf insn.
200 *
201 * jited_off is the byte off to the end of the jited insn.
202 *
203 * Hence, with
204 * insn_start:
205 * The first bpf insn off of the prog. The insn off
206 * here is relative to the main prog.
207 * e.g. if prog is a subprog, insn_start > 0
208 * linfo_idx:
209 * The prog's idx to prog->aux->linfo and jited_linfo
210 *
211 * jited_linfo[linfo_idx] = prog->bpf_func
212 *
213 * For i > linfo_idx,
214 *
215 * jited_linfo[i] = prog->bpf_func +
216 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
217 */
218 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
219 const u32 *insn_to_jit_off)
220 {
221 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
222 const struct bpf_line_info *linfo;
223 void **jited_linfo;
224
225 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt)
226 /* Userspace did not provide linfo */
227 return;
228
229 linfo_idx = prog->aux->linfo_idx;
230 linfo = &prog->aux->linfo[linfo_idx];
231 insn_start = linfo[0].insn_off;
232 insn_end = insn_start + prog->len;
233
234 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
235 jited_linfo[0] = prog->bpf_func;
236
237 nr_linfo = prog->aux->nr_linfo - linfo_idx;
238
239 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
240 /* The verifier ensures that linfo[i].insn_off is
241 * strictly increasing
242 */
243 jited_linfo[i] = prog->bpf_func +
244 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
245 }
246
247 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
248 gfp_t gfp_extra_flags)
249 {
250 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
251 struct bpf_prog *fp;
252 u32 pages;
253
254 size = round_up(size, PAGE_SIZE);
255 pages = size / PAGE_SIZE;
256 if (pages <= fp_old->pages)
257 return fp_old;
258
259 fp = __vmalloc(size, gfp_flags);
260 if (fp) {
261 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
262 fp->pages = pages;
263 fp->aux->prog = fp;
264
265 /* We keep fp->aux from fp_old around in the new
266 * reallocated structure.
267 */
268 fp_old->aux = NULL;
269 fp_old->stats = NULL;
270 fp_old->active = NULL;
271 __bpf_prog_free(fp_old);
272 }
273
274 return fp;
275 }
276
277 void __bpf_prog_free(struct bpf_prog *fp)
278 {
279 if (fp->aux) {
280 mutex_destroy(&fp->aux->used_maps_mutex);
281 mutex_destroy(&fp->aux->dst_mutex);
282 kfree(fp->aux->poke_tab);
283 kfree(fp->aux);
284 }
285 free_percpu(fp->stats);
286 free_percpu(fp->active);
287 vfree(fp);
288 }
289
290 int bpf_prog_calc_tag(struct bpf_prog *fp)
291 {
292 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
293 u32 raw_size = bpf_prog_tag_scratch_size(fp);
294 u32 digest[SHA1_DIGEST_WORDS];
295 u32 ws[SHA1_WORKSPACE_WORDS];
296 u32 i, bsize, psize, blocks;
297 struct bpf_insn *dst;
298 bool was_ld_map;
299 u8 *raw, *todo;
300 __be32 *result;
301 __be64 *bits;
302
303 raw = vmalloc(raw_size);
304 if (!raw)
305 return -ENOMEM;
306
307 sha1_init(digest);
308 memset(ws, 0, sizeof(ws));
309
310 /* We need to take out the map fd for the digest calculation
311 * since they are unstable from user space side.
312 */
313 dst = (void *)raw;
314 for (i = 0, was_ld_map = false; i < fp->len; i++) {
315 dst[i] = fp->insnsi[i];
316 if (!was_ld_map &&
317 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
318 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
319 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
320 was_ld_map = true;
321 dst[i].imm = 0;
322 } else if (was_ld_map &&
323 dst[i].code == 0 &&
324 dst[i].dst_reg == 0 &&
325 dst[i].src_reg == 0 &&
326 dst[i].off == 0) {
327 was_ld_map = false;
328 dst[i].imm = 0;
329 } else {
330 was_ld_map = false;
331 }
332 }
333
334 psize = bpf_prog_insn_size(fp);
335 memset(&raw[psize], 0, raw_size - psize);
336 raw[psize++] = 0x80;
337
338 bsize = round_up(psize, SHA1_BLOCK_SIZE);
339 blocks = bsize / SHA1_BLOCK_SIZE;
340 todo = raw;
341 if (bsize - psize >= sizeof(__be64)) {
342 bits = (__be64 *)(todo + bsize - sizeof(__be64));
343 } else {
344 bits = (__be64 *)(todo + bsize + bits_offset);
345 blocks++;
346 }
347 *bits = cpu_to_be64((psize - 1) << 3);
348
349 while (blocks--) {
350 sha1_transform(digest, todo, ws);
351 todo += SHA1_BLOCK_SIZE;
352 }
353
354 result = (__force __be32 *)digest;
355 for (i = 0; i < SHA1_DIGEST_WORDS; i++)
356 result[i] = cpu_to_be32(digest[i]);
357 memcpy(fp->tag, result, sizeof(fp->tag));
358
359 vfree(raw);
360 return 0;
361 }
362
363 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
364 s32 end_new, s32 curr, const bool probe_pass)
365 {
366 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
367 s32 delta = end_new - end_old;
368 s64 imm = insn->imm;
369
370 if (curr < pos && curr + imm + 1 >= end_old)
371 imm += delta;
372 else if (curr >= end_new && curr + imm + 1 < end_new)
373 imm -= delta;
374 if (imm < imm_min || imm > imm_max)
375 return -ERANGE;
376 if (!probe_pass)
377 insn->imm = imm;
378 return 0;
379 }
380
381 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
382 s32 end_new, s32 curr, const bool probe_pass)
383 {
384 s64 off_min, off_max, off;
385 s32 delta = end_new - end_old;
386
387 if (insn->code == (BPF_JMP32 | BPF_JA)) {
388 off = insn->imm;
389 off_min = S32_MIN;
390 off_max = S32_MAX;
391 } else {
392 off = insn->off;
393 off_min = S16_MIN;
394 off_max = S16_MAX;
395 }
396
397 if (curr < pos && curr + off + 1 >= end_old)
398 off += delta;
399 else if (curr >= end_new && curr + off + 1 < end_new)
400 off -= delta;
401 if (off < off_min || off > off_max)
402 return -ERANGE;
403 if (!probe_pass) {
404 if (insn->code == (BPF_JMP32 | BPF_JA))
405 insn->imm = off;
406 else
407 insn->off = off;
408 }
409 return 0;
410 }
411
412 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
413 s32 end_new, const bool probe_pass)
414 {
415 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
416 struct bpf_insn *insn = prog->insnsi;
417 int ret = 0;
418
419 for (i = 0; i < insn_cnt; i++, insn++) {
420 u8 code;
421
422 /* In the probing pass we still operate on the original,
423 * unpatched image in order to check overflows before we
424 * do any other adjustments. Therefore skip the patchlet.
425 */
426 if (probe_pass && i == pos) {
427 i = end_new;
428 insn = prog->insnsi + end_old;
429 }
430 if (bpf_pseudo_func(insn)) {
431 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
432 end_new, i, probe_pass);
433 if (ret)
434 return ret;
435 continue;
436 }
437 code = insn->code;
438 if ((BPF_CLASS(code) != BPF_JMP &&
439 BPF_CLASS(code) != BPF_JMP32) ||
440 BPF_OP(code) == BPF_EXIT)
441 continue;
442 /* Adjust offset of jmps if we cross patch boundaries. */
443 if (BPF_OP(code) == BPF_CALL) {
444 if (insn->src_reg != BPF_PSEUDO_CALL)
445 continue;
446 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
447 end_new, i, probe_pass);
448 } else {
449 ret = bpf_adj_delta_to_off(insn, pos, end_old,
450 end_new, i, probe_pass);
451 }
452 if (ret)
453 break;
454 }
455
456 return ret;
457 }
458
459 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
460 {
461 struct bpf_line_info *linfo;
462 u32 i, nr_linfo;
463
464 nr_linfo = prog->aux->nr_linfo;
465 if (!nr_linfo || !delta)
466 return;
467
468 linfo = prog->aux->linfo;
469
470 for (i = 0; i < nr_linfo; i++)
471 if (off < linfo[i].insn_off)
472 break;
473
474 /* Push all off < linfo[i].insn_off by delta */
475 for (; i < nr_linfo; i++)
476 linfo[i].insn_off += delta;
477 }
478
479 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
480 const struct bpf_insn *patch, u32 len)
481 {
482 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
483 const u32 cnt_max = S16_MAX;
484 struct bpf_prog *prog_adj;
485 int err;
486
487 /* Since our patchlet doesn't expand the image, we're done. */
488 if (insn_delta == 0) {
489 memcpy(prog->insnsi + off, patch, sizeof(*patch));
490 return prog;
491 }
492
493 insn_adj_cnt = prog->len + insn_delta;
494
495 /* Reject anything that would potentially let the insn->off
496 * target overflow when we have excessive program expansions.
497 * We need to probe here before we do any reallocation where
498 * we afterwards may not fail anymore.
499 */
500 if (insn_adj_cnt > cnt_max &&
501 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
502 return ERR_PTR(err);
503
504 /* Several new instructions need to be inserted. Make room
505 * for them. Likely, there's no need for a new allocation as
506 * last page could have large enough tailroom.
507 */
508 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
509 GFP_USER);
510 if (!prog_adj)
511 return ERR_PTR(-ENOMEM);
512
513 prog_adj->len = insn_adj_cnt;
514
515 /* Patching happens in 3 steps:
516 *
517 * 1) Move over tail of insnsi from next instruction onwards,
518 * so we can patch the single target insn with one or more
519 * new ones (patching is always from 1 to n insns, n > 0).
520 * 2) Inject new instructions at the target location.
521 * 3) Adjust branch offsets if necessary.
522 */
523 insn_rest = insn_adj_cnt - off - len;
524
525 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
526 sizeof(*patch) * insn_rest);
527 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
528
529 /* We are guaranteed to not fail at this point, otherwise
530 * the ship has sailed to reverse to the original state. An
531 * overflow cannot happen at this point.
532 */
533 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
534
535 bpf_adj_linfo(prog_adj, off, insn_delta);
536
537 return prog_adj;
538 }
539
540 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
541 {
542 /* Branch offsets can't overflow when program is shrinking, no need
543 * to call bpf_adj_branches(..., true) here
544 */
545 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
546 sizeof(struct bpf_insn) * (prog->len - off - cnt));
547 prog->len -= cnt;
548
549 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
550 }
551
552 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
553 {
554 int i;
555
556 for (i = 0; i < fp->aux->real_func_cnt; i++)
557 bpf_prog_kallsyms_del(fp->aux->func[i]);
558 }
559
560 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
561 {
562 bpf_prog_kallsyms_del_subprogs(fp);
563 bpf_prog_kallsyms_del(fp);
564 }
565
566 #ifdef CONFIG_BPF_JIT
567 /* All BPF JIT sysctl knobs here. */
568 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
569 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
570 int bpf_jit_harden __read_mostly;
571 long bpf_jit_limit __read_mostly;
572 long bpf_jit_limit_max __read_mostly;
573
574 static void
575 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
576 {
577 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
578
579 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
580 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len;
581 }
582
583 static void
584 bpf_prog_ksym_set_name(struct bpf_prog *prog)
585 {
586 char *sym = prog->aux->ksym.name;
587 const char *end = sym + KSYM_NAME_LEN;
588 const struct btf_type *type;
589 const char *func_name;
590
591 BUILD_BUG_ON(sizeof("bpf_prog_") +
592 sizeof(prog->tag) * 2 +
593 /* name has been null terminated.
594 * We should need +1 for the '_' preceding
595 * the name. However, the null character
596 * is double counted between the name and the
597 * sizeof("bpf_prog_") above, so we omit
598 * the +1 here.
599 */
600 sizeof(prog->aux->name) > KSYM_NAME_LEN);
601
602 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
603 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
604
605 /* prog->aux->name will be ignored if full btf name is available */
606 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) {
607 type = btf_type_by_id(prog->aux->btf,
608 prog->aux->func_info[prog->aux->func_idx].type_id);
609 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
610 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
611 return;
612 }
613
614 if (prog->aux->name[0])
615 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
616 else
617 *sym = 0;
618 }
619
620 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
621 {
622 return container_of(n, struct bpf_ksym, tnode)->start;
623 }
624
625 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
626 struct latch_tree_node *b)
627 {
628 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
629 }
630
631 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
632 {
633 unsigned long val = (unsigned long)key;
634 const struct bpf_ksym *ksym;
635
636 ksym = container_of(n, struct bpf_ksym, tnode);
637
638 if (val < ksym->start)
639 return -1;
640 /* Ensure that we detect return addresses as part of the program, when
641 * the final instruction is a call for a program part of the stack
642 * trace. Therefore, do val > ksym->end instead of val >= ksym->end.
643 */
644 if (val > ksym->end)
645 return 1;
646
647 return 0;
648 }
649
650 static const struct latch_tree_ops bpf_tree_ops = {
651 .less = bpf_tree_less,
652 .comp = bpf_tree_comp,
653 };
654
655 static DEFINE_SPINLOCK(bpf_lock);
656 static LIST_HEAD(bpf_kallsyms);
657 static struct latch_tree_root bpf_tree __cacheline_aligned;
658
659 void bpf_ksym_add(struct bpf_ksym *ksym)
660 {
661 spin_lock_bh(&bpf_lock);
662 WARN_ON_ONCE(!list_empty(&ksym->lnode));
663 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
664 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
665 spin_unlock_bh(&bpf_lock);
666 }
667
668 static void __bpf_ksym_del(struct bpf_ksym *ksym)
669 {
670 if (list_empty(&ksym->lnode))
671 return;
672
673 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
674 list_del_rcu(&ksym->lnode);
675 }
676
677 void bpf_ksym_del(struct bpf_ksym *ksym)
678 {
679 spin_lock_bh(&bpf_lock);
680 __bpf_ksym_del(ksym);
681 spin_unlock_bh(&bpf_lock);
682 }
683
684 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
685 {
686 return fp->jited && !bpf_prog_was_classic(fp);
687 }
688
689 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
690 {
691 if (!bpf_prog_kallsyms_candidate(fp) ||
692 !bpf_token_capable(fp->aux->token, CAP_BPF))
693 return;
694
695 bpf_prog_ksym_set_addr(fp);
696 bpf_prog_ksym_set_name(fp);
697 fp->aux->ksym.prog = true;
698
699 bpf_ksym_add(&fp->aux->ksym);
700
701 #ifdef CONFIG_FINEIBT
702 /*
703 * When FineIBT, code in the __cfi_foo() symbols can get executed
704 * and hence unwinder needs help.
705 */
706 if (cfi_mode != CFI_FINEIBT)
707 return;
708
709 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN,
710 "__cfi_%s", fp->aux->ksym.name);
711
712 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16;
713 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func;
714
715 bpf_ksym_add(&fp->aux->ksym_prefix);
716 #endif
717 }
718
719 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
720 {
721 if (!bpf_prog_kallsyms_candidate(fp))
722 return;
723
724 bpf_ksym_del(&fp->aux->ksym);
725 #ifdef CONFIG_FINEIBT
726 if (cfi_mode != CFI_FINEIBT)
727 return;
728 bpf_ksym_del(&fp->aux->ksym_prefix);
729 #endif
730 }
731
732 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
733 {
734 struct latch_tree_node *n;
735
736 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
737 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
738 }
739
740 int __bpf_address_lookup(unsigned long addr, unsigned long *size,
741 unsigned long *off, char *sym)
742 {
743 struct bpf_ksym *ksym;
744 int ret = 0;
745
746 rcu_read_lock();
747 ksym = bpf_ksym_find(addr);
748 if (ksym) {
749 unsigned long symbol_start = ksym->start;
750 unsigned long symbol_end = ksym->end;
751
752 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN);
753
754 if (size)
755 *size = symbol_end - symbol_start;
756 if (off)
757 *off = addr - symbol_start;
758 }
759 rcu_read_unlock();
760
761 return ret;
762 }
763
764 bool is_bpf_text_address(unsigned long addr)
765 {
766 bool ret;
767
768 rcu_read_lock();
769 ret = bpf_ksym_find(addr) != NULL;
770 rcu_read_unlock();
771
772 return ret;
773 }
774
775 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
776 {
777 struct bpf_ksym *ksym = bpf_ksym_find(addr);
778
779 return ksym && ksym->prog ?
780 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
781 NULL;
782 }
783
784 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
785 {
786 const struct exception_table_entry *e = NULL;
787 struct bpf_prog *prog;
788
789 rcu_read_lock();
790 prog = bpf_prog_ksym_find(addr);
791 if (!prog)
792 goto out;
793 if (!prog->aux->num_exentries)
794 goto out;
795
796 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
797 out:
798 rcu_read_unlock();
799 return e;
800 }
801
802 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
803 char *sym)
804 {
805 struct bpf_ksym *ksym;
806 unsigned int it = 0;
807 int ret = -ERANGE;
808
809 if (!bpf_jit_kallsyms_enabled())
810 return ret;
811
812 rcu_read_lock();
813 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
814 if (it++ != symnum)
815 continue;
816
817 strscpy(sym, ksym->name, KSYM_NAME_LEN);
818
819 *value = ksym->start;
820 *type = BPF_SYM_ELF_TYPE;
821
822 ret = 0;
823 break;
824 }
825 rcu_read_unlock();
826
827 return ret;
828 }
829
830 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
831 struct bpf_jit_poke_descriptor *poke)
832 {
833 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
834 static const u32 poke_tab_max = 1024;
835 u32 slot = prog->aux->size_poke_tab;
836 u32 size = slot + 1;
837
838 if (size > poke_tab_max)
839 return -ENOSPC;
840 if (poke->tailcall_target || poke->tailcall_target_stable ||
841 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
842 return -EINVAL;
843
844 switch (poke->reason) {
845 case BPF_POKE_REASON_TAIL_CALL:
846 if (!poke->tail_call.map)
847 return -EINVAL;
848 break;
849 default:
850 return -EINVAL;
851 }
852
853 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL);
854 if (!tab)
855 return -ENOMEM;
856
857 memcpy(&tab[slot], poke, sizeof(*poke));
858 prog->aux->size_poke_tab = size;
859 prog->aux->poke_tab = tab;
860
861 return slot;
862 }
863
864 /*
865 * BPF program pack allocator.
866 *
867 * Most BPF programs are pretty small. Allocating a hole page for each
868 * program is sometime a waste. Many small bpf program also adds pressure
869 * to instruction TLB. To solve this issue, we introduce a BPF program pack
870 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
871 * to host BPF programs.
872 */
873 #define BPF_PROG_CHUNK_SHIFT 6
874 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT)
875 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1))
876
877 struct bpf_prog_pack {
878 struct list_head list;
879 void *ptr;
880 unsigned long bitmap[];
881 };
882
883 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size)
884 {
885 memset(area, 0, size);
886 }
887
888 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
889
890 static DEFINE_MUTEX(pack_mutex);
891 static LIST_HEAD(pack_list);
892
893 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
894 * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
895 */
896 #ifdef PMD_SIZE
897 /* PMD_SIZE is really big for some archs. It doesn't make sense to
898 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to
899 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be
900 * greater than or equal to 2MB.
901 */
902 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes())
903 #else
904 #define BPF_PROG_PACK_SIZE PAGE_SIZE
905 #endif
906
907 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
908
909 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
910 {
911 struct bpf_prog_pack *pack;
912 int err;
913
914 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
915 GFP_KERNEL);
916 if (!pack)
917 return NULL;
918 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE);
919 if (!pack->ptr)
920 goto out;
921 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
922 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
923
924 set_vm_flush_reset_perms(pack->ptr);
925 err = set_memory_rox((unsigned long)pack->ptr,
926 BPF_PROG_PACK_SIZE / PAGE_SIZE);
927 if (err)
928 goto out;
929 list_add_tail(&pack->list, &pack_list);
930 return pack;
931
932 out:
933 bpf_jit_free_exec(pack->ptr);
934 kfree(pack);
935 return NULL;
936 }
937
938 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
939 {
940 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
941 struct bpf_prog_pack *pack;
942 unsigned long pos;
943 void *ptr = NULL;
944
945 mutex_lock(&pack_mutex);
946 if (size > BPF_PROG_PACK_SIZE) {
947 size = round_up(size, PAGE_SIZE);
948 ptr = bpf_jit_alloc_exec(size);
949 if (ptr) {
950 int err;
951
952 bpf_fill_ill_insns(ptr, size);
953 set_vm_flush_reset_perms(ptr);
954 err = set_memory_rox((unsigned long)ptr,
955 size / PAGE_SIZE);
956 if (err) {
957 bpf_jit_free_exec(ptr);
958 ptr = NULL;
959 }
960 }
961 goto out;
962 }
963 list_for_each_entry(pack, &pack_list, list) {
964 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
965 nbits, 0);
966 if (pos < BPF_PROG_CHUNK_COUNT)
967 goto found_free_area;
968 }
969
970 pack = alloc_new_pack(bpf_fill_ill_insns);
971 if (!pack)
972 goto out;
973
974 pos = 0;
975
976 found_free_area:
977 bitmap_set(pack->bitmap, pos, nbits);
978 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
979
980 out:
981 mutex_unlock(&pack_mutex);
982 return ptr;
983 }
984
985 void bpf_prog_pack_free(void *ptr, u32 size)
986 {
987 struct bpf_prog_pack *pack = NULL, *tmp;
988 unsigned int nbits;
989 unsigned long pos;
990
991 mutex_lock(&pack_mutex);
992 if (size > BPF_PROG_PACK_SIZE) {
993 bpf_jit_free_exec(ptr);
994 goto out;
995 }
996
997 list_for_each_entry(tmp, &pack_list, list) {
998 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) {
999 pack = tmp;
1000 break;
1001 }
1002 }
1003
1004 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
1005 goto out;
1006
1007 nbits = BPF_PROG_SIZE_TO_NBITS(size);
1008 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
1009
1010 WARN_ONCE(bpf_arch_text_invalidate(ptr, size),
1011 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
1012
1013 bitmap_clear(pack->bitmap, pos, nbits);
1014 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
1015 BPF_PROG_CHUNK_COUNT, 0) == 0) {
1016 list_del(&pack->list);
1017 bpf_jit_free_exec(pack->ptr);
1018 kfree(pack);
1019 }
1020 out:
1021 mutex_unlock(&pack_mutex);
1022 }
1023
1024 static atomic_long_t bpf_jit_current;
1025
1026 /* Can be overridden by an arch's JIT compiler if it has a custom,
1027 * dedicated BPF backend memory area, or if neither of the two
1028 * below apply.
1029 */
1030 u64 __weak bpf_jit_alloc_exec_limit(void)
1031 {
1032 #if defined(MODULES_VADDR)
1033 return MODULES_END - MODULES_VADDR;
1034 #else
1035 return VMALLOC_END - VMALLOC_START;
1036 #endif
1037 }
1038
1039 static int __init bpf_jit_charge_init(void)
1040 {
1041 /* Only used as heuristic here to derive limit. */
1042 bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
1043 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1,
1044 PAGE_SIZE), LONG_MAX);
1045 return 0;
1046 }
1047 pure_initcall(bpf_jit_charge_init);
1048
1049 int bpf_jit_charge_modmem(u32 size)
1050 {
1051 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
1052 if (!bpf_capable()) {
1053 atomic_long_sub(size, &bpf_jit_current);
1054 return -EPERM;
1055 }
1056 }
1057
1058 return 0;
1059 }
1060
1061 void bpf_jit_uncharge_modmem(u32 size)
1062 {
1063 atomic_long_sub(size, &bpf_jit_current);
1064 }
1065
1066 void *__weak bpf_jit_alloc_exec(unsigned long size)
1067 {
1068 return execmem_alloc(EXECMEM_BPF, size);
1069 }
1070
1071 void __weak bpf_jit_free_exec(void *addr)
1072 {
1073 execmem_free(addr);
1074 }
1075
1076 struct bpf_binary_header *
1077 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1078 unsigned int alignment,
1079 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1080 {
1081 struct bpf_binary_header *hdr;
1082 u32 size, hole, start;
1083
1084 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1085 alignment > BPF_IMAGE_ALIGNMENT);
1086
1087 /* Most of BPF filters are really small, but if some of them
1088 * fill a page, allow at least 128 extra bytes to insert a
1089 * random section of illegal instructions.
1090 */
1091 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1092
1093 if (bpf_jit_charge_modmem(size))
1094 return NULL;
1095 hdr = bpf_jit_alloc_exec(size);
1096 if (!hdr) {
1097 bpf_jit_uncharge_modmem(size);
1098 return NULL;
1099 }
1100
1101 /* Fill space with illegal/arch-dep instructions. */
1102 bpf_fill_ill_insns(hdr, size);
1103
1104 hdr->size = size;
1105 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1106 PAGE_SIZE - sizeof(*hdr));
1107 start = get_random_u32_below(hole) & ~(alignment - 1);
1108
1109 /* Leave a random number of instructions before BPF code. */
1110 *image_ptr = &hdr->image[start];
1111
1112 return hdr;
1113 }
1114
1115 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1116 {
1117 u32 size = hdr->size;
1118
1119 bpf_jit_free_exec(hdr);
1120 bpf_jit_uncharge_modmem(size);
1121 }
1122
1123 /* Allocate jit binary from bpf_prog_pack allocator.
1124 * Since the allocated memory is RO+X, the JIT engine cannot write directly
1125 * to the memory. To solve this problem, a RW buffer is also allocated at
1126 * as the same time. The JIT engine should calculate offsets based on the
1127 * RO memory address, but write JITed program to the RW buffer. Once the
1128 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1129 * the JITed program to the RO memory.
1130 */
1131 struct bpf_binary_header *
1132 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1133 unsigned int alignment,
1134 struct bpf_binary_header **rw_header,
1135 u8 **rw_image,
1136 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1137 {
1138 struct bpf_binary_header *ro_header;
1139 u32 size, hole, start;
1140
1141 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1142 alignment > BPF_IMAGE_ALIGNMENT);
1143
1144 /* add 16 bytes for a random section of illegal instructions */
1145 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1146
1147 if (bpf_jit_charge_modmem(size))
1148 return NULL;
1149 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1150 if (!ro_header) {
1151 bpf_jit_uncharge_modmem(size);
1152 return NULL;
1153 }
1154
1155 *rw_header = kvmalloc(size, GFP_KERNEL);
1156 if (!*rw_header) {
1157 bpf_prog_pack_free(ro_header, size);
1158 bpf_jit_uncharge_modmem(size);
1159 return NULL;
1160 }
1161
1162 /* Fill space with illegal/arch-dep instructions. */
1163 bpf_fill_ill_insns(*rw_header, size);
1164 (*rw_header)->size = size;
1165
1166 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1167 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1168 start = get_random_u32_below(hole) & ~(alignment - 1);
1169
1170 *image_ptr = &ro_header->image[start];
1171 *rw_image = &(*rw_header)->image[start];
1172
1173 return ro_header;
1174 }
1175
1176 /* Copy JITed text from rw_header to its final location, the ro_header. */
1177 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header,
1178 struct bpf_binary_header *rw_header)
1179 {
1180 void *ptr;
1181
1182 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1183
1184 kvfree(rw_header);
1185
1186 if (IS_ERR(ptr)) {
1187 bpf_prog_pack_free(ro_header, ro_header->size);
1188 return PTR_ERR(ptr);
1189 }
1190 return 0;
1191 }
1192
1193 /* bpf_jit_binary_pack_free is called in two different scenarios:
1194 * 1) when the program is freed after;
1195 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1196 * For case 2), we need to free both the RO memory and the RW buffer.
1197 *
1198 * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1199 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1200 * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1201 * bpf_arch_text_copy (when jit fails).
1202 */
1203 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1204 struct bpf_binary_header *rw_header)
1205 {
1206 u32 size = ro_header->size;
1207
1208 bpf_prog_pack_free(ro_header, size);
1209 kvfree(rw_header);
1210 bpf_jit_uncharge_modmem(size);
1211 }
1212
1213 struct bpf_binary_header *
1214 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1215 {
1216 unsigned long real_start = (unsigned long)fp->bpf_func;
1217 unsigned long addr;
1218
1219 addr = real_start & BPF_PROG_CHUNK_MASK;
1220 return (void *)addr;
1221 }
1222
1223 static inline struct bpf_binary_header *
1224 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1225 {
1226 unsigned long real_start = (unsigned long)fp->bpf_func;
1227 unsigned long addr;
1228
1229 addr = real_start & PAGE_MASK;
1230 return (void *)addr;
1231 }
1232
1233 /* This symbol is only overridden by archs that have different
1234 * requirements than the usual eBPF JITs, f.e. when they only
1235 * implement cBPF JIT, do not set images read-only, etc.
1236 */
1237 void __weak bpf_jit_free(struct bpf_prog *fp)
1238 {
1239 if (fp->jited) {
1240 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1241
1242 bpf_jit_binary_free(hdr);
1243 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1244 }
1245
1246 bpf_prog_unlock_free(fp);
1247 }
1248
1249 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1250 const struct bpf_insn *insn, bool extra_pass,
1251 u64 *func_addr, bool *func_addr_fixed)
1252 {
1253 s16 off = insn->off;
1254 s32 imm = insn->imm;
1255 u8 *addr;
1256 int err;
1257
1258 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1259 if (!*func_addr_fixed) {
1260 /* Place-holder address till the last pass has collected
1261 * all addresses for JITed subprograms in which case we
1262 * can pick them up from prog->aux.
1263 */
1264 if (!extra_pass)
1265 addr = NULL;
1266 else if (prog->aux->func &&
1267 off >= 0 && off < prog->aux->real_func_cnt)
1268 addr = (u8 *)prog->aux->func[off]->bpf_func;
1269 else
1270 return -EINVAL;
1271 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
1272 bpf_jit_supports_far_kfunc_call()) {
1273 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr);
1274 if (err)
1275 return err;
1276 } else {
1277 /* Address of a BPF helper call. Since part of the core
1278 * kernel, it's always at a fixed location. __bpf_call_base
1279 * and the helper with imm relative to it are both in core
1280 * kernel.
1281 */
1282 addr = (u8 *)__bpf_call_base + imm;
1283 }
1284
1285 *func_addr = (unsigned long)addr;
1286 return 0;
1287 }
1288
1289 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1290 const struct bpf_insn *aux,
1291 struct bpf_insn *to_buff,
1292 bool emit_zext)
1293 {
1294 struct bpf_insn *to = to_buff;
1295 u32 imm_rnd = get_random_u32();
1296 s16 off;
1297
1298 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
1299 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1300
1301 /* Constraints on AX register:
1302 *
1303 * AX register is inaccessible from user space. It is mapped in
1304 * all JITs, and used here for constant blinding rewrites. It is
1305 * typically "stateless" meaning its contents are only valid within
1306 * the executed instruction, but not across several instructions.
1307 * There are a few exceptions however which are further detailed
1308 * below.
1309 *
1310 * Constant blinding is only used by JITs, not in the interpreter.
1311 * The interpreter uses AX in some occasions as a local temporary
1312 * register e.g. in DIV or MOD instructions.
1313 *
1314 * In restricted circumstances, the verifier can also use the AX
1315 * register for rewrites as long as they do not interfere with
1316 * the above cases!
1317 */
1318 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1319 goto out;
1320
1321 if (from->imm == 0 &&
1322 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
1323 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1324 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1325 goto out;
1326 }
1327
1328 switch (from->code) {
1329 case BPF_ALU | BPF_ADD | BPF_K:
1330 case BPF_ALU | BPF_SUB | BPF_K:
1331 case BPF_ALU | BPF_AND | BPF_K:
1332 case BPF_ALU | BPF_OR | BPF_K:
1333 case BPF_ALU | BPF_XOR | BPF_K:
1334 case BPF_ALU | BPF_MUL | BPF_K:
1335 case BPF_ALU | BPF_MOV | BPF_K:
1336 case BPF_ALU | BPF_DIV | BPF_K:
1337 case BPF_ALU | BPF_MOD | BPF_K:
1338 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1339 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1340 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off);
1341 break;
1342
1343 case BPF_ALU64 | BPF_ADD | BPF_K:
1344 case BPF_ALU64 | BPF_SUB | BPF_K:
1345 case BPF_ALU64 | BPF_AND | BPF_K:
1346 case BPF_ALU64 | BPF_OR | BPF_K:
1347 case BPF_ALU64 | BPF_XOR | BPF_K:
1348 case BPF_ALU64 | BPF_MUL | BPF_K:
1349 case BPF_ALU64 | BPF_MOV | BPF_K:
1350 case BPF_ALU64 | BPF_DIV | BPF_K:
1351 case BPF_ALU64 | BPF_MOD | BPF_K:
1352 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1353 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1354 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off);
1355 break;
1356
1357 case BPF_JMP | BPF_JEQ | BPF_K:
1358 case BPF_JMP | BPF_JNE | BPF_K:
1359 case BPF_JMP | BPF_JGT | BPF_K:
1360 case BPF_JMP | BPF_JLT | BPF_K:
1361 case BPF_JMP | BPF_JGE | BPF_K:
1362 case BPF_JMP | BPF_JLE | BPF_K:
1363 case BPF_JMP | BPF_JSGT | BPF_K:
1364 case BPF_JMP | BPF_JSLT | BPF_K:
1365 case BPF_JMP | BPF_JSGE | BPF_K:
1366 case BPF_JMP | BPF_JSLE | BPF_K:
1367 case BPF_JMP | BPF_JSET | BPF_K:
1368 /* Accommodate for extra offset in case of a backjump. */
1369 off = from->off;
1370 if (off < 0)
1371 off -= 2;
1372 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1373 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1374 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1375 break;
1376
1377 case BPF_JMP32 | BPF_JEQ | BPF_K:
1378 case BPF_JMP32 | BPF_JNE | BPF_K:
1379 case BPF_JMP32 | BPF_JGT | BPF_K:
1380 case BPF_JMP32 | BPF_JLT | BPF_K:
1381 case BPF_JMP32 | BPF_JGE | BPF_K:
1382 case BPF_JMP32 | BPF_JLE | BPF_K:
1383 case BPF_JMP32 | BPF_JSGT | BPF_K:
1384 case BPF_JMP32 | BPF_JSLT | BPF_K:
1385 case BPF_JMP32 | BPF_JSGE | BPF_K:
1386 case BPF_JMP32 | BPF_JSLE | BPF_K:
1387 case BPF_JMP32 | BPF_JSET | BPF_K:
1388 /* Accommodate for extra offset in case of a backjump. */
1389 off = from->off;
1390 if (off < 0)
1391 off -= 2;
1392 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1393 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1394 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1395 off);
1396 break;
1397
1398 case BPF_LD | BPF_IMM | BPF_DW:
1399 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1400 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1401 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1402 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1403 break;
1404 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1405 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1406 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1407 if (emit_zext)
1408 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1409 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1410 break;
1411
1412 case BPF_ST | BPF_MEM | BPF_DW:
1413 case BPF_ST | BPF_MEM | BPF_W:
1414 case BPF_ST | BPF_MEM | BPF_H:
1415 case BPF_ST | BPF_MEM | BPF_B:
1416 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1417 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1418 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1419 break;
1420 }
1421 out:
1422 return to - to_buff;
1423 }
1424
1425 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1426 gfp_t gfp_extra_flags)
1427 {
1428 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1429 struct bpf_prog *fp;
1430
1431 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1432 if (fp != NULL) {
1433 /* aux->prog still points to the fp_other one, so
1434 * when promoting the clone to the real program,
1435 * this still needs to be adapted.
1436 */
1437 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1438 }
1439
1440 return fp;
1441 }
1442
1443 static void bpf_prog_clone_free(struct bpf_prog *fp)
1444 {
1445 /* aux was stolen by the other clone, so we cannot free
1446 * it from this path! It will be freed eventually by the
1447 * other program on release.
1448 *
1449 * At this point, we don't need a deferred release since
1450 * clone is guaranteed to not be locked.
1451 */
1452 fp->aux = NULL;
1453 fp->stats = NULL;
1454 fp->active = NULL;
1455 __bpf_prog_free(fp);
1456 }
1457
1458 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1459 {
1460 /* We have to repoint aux->prog to self, as we don't
1461 * know whether fp here is the clone or the original.
1462 */
1463 fp->aux->prog = fp;
1464 bpf_prog_clone_free(fp_other);
1465 }
1466
1467 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1468 {
1469 struct bpf_insn insn_buff[16], aux[2];
1470 struct bpf_prog *clone, *tmp;
1471 int insn_delta, insn_cnt;
1472 struct bpf_insn *insn;
1473 int i, rewritten;
1474
1475 if (!prog->blinding_requested || prog->blinded)
1476 return prog;
1477
1478 clone = bpf_prog_clone_create(prog, GFP_USER);
1479 if (!clone)
1480 return ERR_PTR(-ENOMEM);
1481
1482 insn_cnt = clone->len;
1483 insn = clone->insnsi;
1484
1485 for (i = 0; i < insn_cnt; i++, insn++) {
1486 if (bpf_pseudo_func(insn)) {
1487 /* ld_imm64 with an address of bpf subprog is not
1488 * a user controlled constant. Don't randomize it,
1489 * since it will conflict with jit_subprogs() logic.
1490 */
1491 insn++;
1492 i++;
1493 continue;
1494 }
1495
1496 /* We temporarily need to hold the original ld64 insn
1497 * so that we can still access the first part in the
1498 * second blinding run.
1499 */
1500 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1501 insn[1].code == 0)
1502 memcpy(aux, insn, sizeof(aux));
1503
1504 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1505 clone->aux->verifier_zext);
1506 if (!rewritten)
1507 continue;
1508
1509 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1510 if (IS_ERR(tmp)) {
1511 /* Patching may have repointed aux->prog during
1512 * realloc from the original one, so we need to
1513 * fix it up here on error.
1514 */
1515 bpf_jit_prog_release_other(prog, clone);
1516 return tmp;
1517 }
1518
1519 clone = tmp;
1520 insn_delta = rewritten - 1;
1521
1522 /* Walk new program and skip insns we just inserted. */
1523 insn = clone->insnsi + i + insn_delta;
1524 insn_cnt += insn_delta;
1525 i += insn_delta;
1526 }
1527
1528 clone->blinded = 1;
1529 return clone;
1530 }
1531 #endif /* CONFIG_BPF_JIT */
1532
1533 /* Base function for offset calculation. Needs to go into .text section,
1534 * therefore keeping it non-static as well; will also be used by JITs
1535 * anyway later on, so do not let the compiler omit it. This also needs
1536 * to go into kallsyms for correlation from e.g. bpftool, so naming
1537 * must not change.
1538 */
1539 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1540 {
1541 return 0;
1542 }
1543 EXPORT_SYMBOL_GPL(__bpf_call_base);
1544
1545 /* All UAPI available opcodes. */
1546 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1547 /* 32 bit ALU operations. */ \
1548 /* Register based. */ \
1549 INSN_3(ALU, ADD, X), \
1550 INSN_3(ALU, SUB, X), \
1551 INSN_3(ALU, AND, X), \
1552 INSN_3(ALU, OR, X), \
1553 INSN_3(ALU, LSH, X), \
1554 INSN_3(ALU, RSH, X), \
1555 INSN_3(ALU, XOR, X), \
1556 INSN_3(ALU, MUL, X), \
1557 INSN_3(ALU, MOV, X), \
1558 INSN_3(ALU, ARSH, X), \
1559 INSN_3(ALU, DIV, X), \
1560 INSN_3(ALU, MOD, X), \
1561 INSN_2(ALU, NEG), \
1562 INSN_3(ALU, END, TO_BE), \
1563 INSN_3(ALU, END, TO_LE), \
1564 /* Immediate based. */ \
1565 INSN_3(ALU, ADD, K), \
1566 INSN_3(ALU, SUB, K), \
1567 INSN_3(ALU, AND, K), \
1568 INSN_3(ALU, OR, K), \
1569 INSN_3(ALU, LSH, K), \
1570 INSN_3(ALU, RSH, K), \
1571 INSN_3(ALU, XOR, K), \
1572 INSN_3(ALU, MUL, K), \
1573 INSN_3(ALU, MOV, K), \
1574 INSN_3(ALU, ARSH, K), \
1575 INSN_3(ALU, DIV, K), \
1576 INSN_3(ALU, MOD, K), \
1577 /* 64 bit ALU operations. */ \
1578 /* Register based. */ \
1579 INSN_3(ALU64, ADD, X), \
1580 INSN_3(ALU64, SUB, X), \
1581 INSN_3(ALU64, AND, X), \
1582 INSN_3(ALU64, OR, X), \
1583 INSN_3(ALU64, LSH, X), \
1584 INSN_3(ALU64, RSH, X), \
1585 INSN_3(ALU64, XOR, X), \
1586 INSN_3(ALU64, MUL, X), \
1587 INSN_3(ALU64, MOV, X), \
1588 INSN_3(ALU64, ARSH, X), \
1589 INSN_3(ALU64, DIV, X), \
1590 INSN_3(ALU64, MOD, X), \
1591 INSN_2(ALU64, NEG), \
1592 INSN_3(ALU64, END, TO_LE), \
1593 /* Immediate based. */ \
1594 INSN_3(ALU64, ADD, K), \
1595 INSN_3(ALU64, SUB, K), \
1596 INSN_3(ALU64, AND, K), \
1597 INSN_3(ALU64, OR, K), \
1598 INSN_3(ALU64, LSH, K), \
1599 INSN_3(ALU64, RSH, K), \
1600 INSN_3(ALU64, XOR, K), \
1601 INSN_3(ALU64, MUL, K), \
1602 INSN_3(ALU64, MOV, K), \
1603 INSN_3(ALU64, ARSH, K), \
1604 INSN_3(ALU64, DIV, K), \
1605 INSN_3(ALU64, MOD, K), \
1606 /* Call instruction. */ \
1607 INSN_2(JMP, CALL), \
1608 /* Exit instruction. */ \
1609 INSN_2(JMP, EXIT), \
1610 /* 32-bit Jump instructions. */ \
1611 /* Register based. */ \
1612 INSN_3(JMP32, JEQ, X), \
1613 INSN_3(JMP32, JNE, X), \
1614 INSN_3(JMP32, JGT, X), \
1615 INSN_3(JMP32, JLT, X), \
1616 INSN_3(JMP32, JGE, X), \
1617 INSN_3(JMP32, JLE, X), \
1618 INSN_3(JMP32, JSGT, X), \
1619 INSN_3(JMP32, JSLT, X), \
1620 INSN_3(JMP32, JSGE, X), \
1621 INSN_3(JMP32, JSLE, X), \
1622 INSN_3(JMP32, JSET, X), \
1623 /* Immediate based. */ \
1624 INSN_3(JMP32, JEQ, K), \
1625 INSN_3(JMP32, JNE, K), \
1626 INSN_3(JMP32, JGT, K), \
1627 INSN_3(JMP32, JLT, K), \
1628 INSN_3(JMP32, JGE, K), \
1629 INSN_3(JMP32, JLE, K), \
1630 INSN_3(JMP32, JSGT, K), \
1631 INSN_3(JMP32, JSLT, K), \
1632 INSN_3(JMP32, JSGE, K), \
1633 INSN_3(JMP32, JSLE, K), \
1634 INSN_3(JMP32, JSET, K), \
1635 /* Jump instructions. */ \
1636 /* Register based. */ \
1637 INSN_3(JMP, JEQ, X), \
1638 INSN_3(JMP, JNE, X), \
1639 INSN_3(JMP, JGT, X), \
1640 INSN_3(JMP, JLT, X), \
1641 INSN_3(JMP, JGE, X), \
1642 INSN_3(JMP, JLE, X), \
1643 INSN_3(JMP, JSGT, X), \
1644 INSN_3(JMP, JSLT, X), \
1645 INSN_3(JMP, JSGE, X), \
1646 INSN_3(JMP, JSLE, X), \
1647 INSN_3(JMP, JSET, X), \
1648 /* Immediate based. */ \
1649 INSN_3(JMP, JEQ, K), \
1650 INSN_3(JMP, JNE, K), \
1651 INSN_3(JMP, JGT, K), \
1652 INSN_3(JMP, JLT, K), \
1653 INSN_3(JMP, JGE, K), \
1654 INSN_3(JMP, JLE, K), \
1655 INSN_3(JMP, JSGT, K), \
1656 INSN_3(JMP, JSLT, K), \
1657 INSN_3(JMP, JSGE, K), \
1658 INSN_3(JMP, JSLE, K), \
1659 INSN_3(JMP, JSET, K), \
1660 INSN_2(JMP, JA), \
1661 INSN_2(JMP32, JA), \
1662 /* Store instructions. */ \
1663 /* Register based. */ \
1664 INSN_3(STX, MEM, B), \
1665 INSN_3(STX, MEM, H), \
1666 INSN_3(STX, MEM, W), \
1667 INSN_3(STX, MEM, DW), \
1668 INSN_3(STX, ATOMIC, W), \
1669 INSN_3(STX, ATOMIC, DW), \
1670 /* Immediate based. */ \
1671 INSN_3(ST, MEM, B), \
1672 INSN_3(ST, MEM, H), \
1673 INSN_3(ST, MEM, W), \
1674 INSN_3(ST, MEM, DW), \
1675 /* Load instructions. */ \
1676 /* Register based. */ \
1677 INSN_3(LDX, MEM, B), \
1678 INSN_3(LDX, MEM, H), \
1679 INSN_3(LDX, MEM, W), \
1680 INSN_3(LDX, MEM, DW), \
1681 INSN_3(LDX, MEMSX, B), \
1682 INSN_3(LDX, MEMSX, H), \
1683 INSN_3(LDX, MEMSX, W), \
1684 /* Immediate based. */ \
1685 INSN_3(LD, IMM, DW)
1686
1687 bool bpf_opcode_in_insntable(u8 code)
1688 {
1689 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1690 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1691 static const bool public_insntable[256] = {
1692 [0 ... 255] = false,
1693 /* Now overwrite non-defaults ... */
1694 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1695 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1696 [BPF_LD | BPF_ABS | BPF_B] = true,
1697 [BPF_LD | BPF_ABS | BPF_H] = true,
1698 [BPF_LD | BPF_ABS | BPF_W] = true,
1699 [BPF_LD | BPF_IND | BPF_B] = true,
1700 [BPF_LD | BPF_IND | BPF_H] = true,
1701 [BPF_LD | BPF_IND | BPF_W] = true,
1702 [BPF_JMP | BPF_JCOND] = true,
1703 };
1704 #undef BPF_INSN_3_TBL
1705 #undef BPF_INSN_2_TBL
1706 return public_insntable[code];
1707 }
1708
1709 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1710 /**
1711 * ___bpf_prog_run - run eBPF program on a given context
1712 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1713 * @insn: is the array of eBPF instructions
1714 *
1715 * Decode and execute eBPF instructions.
1716 *
1717 * Return: whatever value is in %BPF_R0 at program exit
1718 */
1719 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1720 {
1721 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1722 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1723 static const void * const jumptable[256] __annotate_jump_table = {
1724 [0 ... 255] = &&default_label,
1725 /* Now overwrite non-defaults ... */
1726 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1727 /* Non-UAPI available opcodes. */
1728 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1729 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1730 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1731 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1732 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1733 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1734 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1735 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B,
1736 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H,
1737 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W,
1738 };
1739 #undef BPF_INSN_3_LBL
1740 #undef BPF_INSN_2_LBL
1741 u32 tail_call_cnt = 0;
1742
1743 #define CONT ({ insn++; goto select_insn; })
1744 #define CONT_JMP ({ insn++; goto select_insn; })
1745
1746 select_insn:
1747 goto *jumptable[insn->code];
1748
1749 /* Explicitly mask the register-based shift amounts with 63 or 31
1750 * to avoid undefined behavior. Normally this won't affect the
1751 * generated code, for example, in case of native 64 bit archs such
1752 * as x86-64 or arm64, the compiler is optimizing the AND away for
1753 * the interpreter. In case of JITs, each of the JIT backends compiles
1754 * the BPF shift operations to machine instructions which produce
1755 * implementation-defined results in such a case; the resulting
1756 * contents of the register may be arbitrary, but program behaviour
1757 * as a whole remains defined. In other words, in case of JIT backends,
1758 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1759 */
1760 /* ALU (shifts) */
1761 #define SHT(OPCODE, OP) \
1762 ALU64_##OPCODE##_X: \
1763 DST = DST OP (SRC & 63); \
1764 CONT; \
1765 ALU_##OPCODE##_X: \
1766 DST = (u32) DST OP ((u32) SRC & 31); \
1767 CONT; \
1768 ALU64_##OPCODE##_K: \
1769 DST = DST OP IMM; \
1770 CONT; \
1771 ALU_##OPCODE##_K: \
1772 DST = (u32) DST OP (u32) IMM; \
1773 CONT;
1774 /* ALU (rest) */
1775 #define ALU(OPCODE, OP) \
1776 ALU64_##OPCODE##_X: \
1777 DST = DST OP SRC; \
1778 CONT; \
1779 ALU_##OPCODE##_X: \
1780 DST = (u32) DST OP (u32) SRC; \
1781 CONT; \
1782 ALU64_##OPCODE##_K: \
1783 DST = DST OP IMM; \
1784 CONT; \
1785 ALU_##OPCODE##_K: \
1786 DST = (u32) DST OP (u32) IMM; \
1787 CONT;
1788 ALU(ADD, +)
1789 ALU(SUB, -)
1790 ALU(AND, &)
1791 ALU(OR, |)
1792 ALU(XOR, ^)
1793 ALU(MUL, *)
1794 SHT(LSH, <<)
1795 SHT(RSH, >>)
1796 #undef SHT
1797 #undef ALU
1798 ALU_NEG:
1799 DST = (u32) -DST;
1800 CONT;
1801 ALU64_NEG:
1802 DST = -DST;
1803 CONT;
1804 ALU_MOV_X:
1805 switch (OFF) {
1806 case 0:
1807 DST = (u32) SRC;
1808 break;
1809 case 8:
1810 DST = (u32)(s8) SRC;
1811 break;
1812 case 16:
1813 DST = (u32)(s16) SRC;
1814 break;
1815 }
1816 CONT;
1817 ALU_MOV_K:
1818 DST = (u32) IMM;
1819 CONT;
1820 ALU64_MOV_X:
1821 switch (OFF) {
1822 case 0:
1823 DST = SRC;
1824 break;
1825 case 8:
1826 DST = (s8) SRC;
1827 break;
1828 case 16:
1829 DST = (s16) SRC;
1830 break;
1831 case 32:
1832 DST = (s32) SRC;
1833 break;
1834 }
1835 CONT;
1836 ALU64_MOV_K:
1837 DST = IMM;
1838 CONT;
1839 LD_IMM_DW:
1840 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1841 insn++;
1842 CONT;
1843 ALU_ARSH_X:
1844 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1845 CONT;
1846 ALU_ARSH_K:
1847 DST = (u64) (u32) (((s32) DST) >> IMM);
1848 CONT;
1849 ALU64_ARSH_X:
1850 (*(s64 *) &DST) >>= (SRC & 63);
1851 CONT;
1852 ALU64_ARSH_K:
1853 (*(s64 *) &DST) >>= IMM;
1854 CONT;
1855 ALU64_MOD_X:
1856 switch (OFF) {
1857 case 0:
1858 div64_u64_rem(DST, SRC, &AX);
1859 DST = AX;
1860 break;
1861 case 1:
1862 AX = div64_s64(DST, SRC);
1863 DST = DST - AX * SRC;
1864 break;
1865 }
1866 CONT;
1867 ALU_MOD_X:
1868 switch (OFF) {
1869 case 0:
1870 AX = (u32) DST;
1871 DST = do_div(AX, (u32) SRC);
1872 break;
1873 case 1:
1874 AX = abs((s32)DST);
1875 AX = do_div(AX, abs((s32)SRC));
1876 if ((s32)DST < 0)
1877 DST = (u32)-AX;
1878 else
1879 DST = (u32)AX;
1880 break;
1881 }
1882 CONT;
1883 ALU64_MOD_K:
1884 switch (OFF) {
1885 case 0:
1886 div64_u64_rem(DST, IMM, &AX);
1887 DST = AX;
1888 break;
1889 case 1:
1890 AX = div64_s64(DST, IMM);
1891 DST = DST - AX * IMM;
1892 break;
1893 }
1894 CONT;
1895 ALU_MOD_K:
1896 switch (OFF) {
1897 case 0:
1898 AX = (u32) DST;
1899 DST = do_div(AX, (u32) IMM);
1900 break;
1901 case 1:
1902 AX = abs((s32)DST);
1903 AX = do_div(AX, abs((s32)IMM));
1904 if ((s32)DST < 0)
1905 DST = (u32)-AX;
1906 else
1907 DST = (u32)AX;
1908 break;
1909 }
1910 CONT;
1911 ALU64_DIV_X:
1912 switch (OFF) {
1913 case 0:
1914 DST = div64_u64(DST, SRC);
1915 break;
1916 case 1:
1917 DST = div64_s64(DST, SRC);
1918 break;
1919 }
1920 CONT;
1921 ALU_DIV_X:
1922 switch (OFF) {
1923 case 0:
1924 AX = (u32) DST;
1925 do_div(AX, (u32) SRC);
1926 DST = (u32) AX;
1927 break;
1928 case 1:
1929 AX = abs((s32)DST);
1930 do_div(AX, abs((s32)SRC));
1931 if (((s32)DST < 0) == ((s32)SRC < 0))
1932 DST = (u32)AX;
1933 else
1934 DST = (u32)-AX;
1935 break;
1936 }
1937 CONT;
1938 ALU64_DIV_K:
1939 switch (OFF) {
1940 case 0:
1941 DST = div64_u64(DST, IMM);
1942 break;
1943 case 1:
1944 DST = div64_s64(DST, IMM);
1945 break;
1946 }
1947 CONT;
1948 ALU_DIV_K:
1949 switch (OFF) {
1950 case 0:
1951 AX = (u32) DST;
1952 do_div(AX, (u32) IMM);
1953 DST = (u32) AX;
1954 break;
1955 case 1:
1956 AX = abs((s32)DST);
1957 do_div(AX, abs((s32)IMM));
1958 if (((s32)DST < 0) == ((s32)IMM < 0))
1959 DST = (u32)AX;
1960 else
1961 DST = (u32)-AX;
1962 break;
1963 }
1964 CONT;
1965 ALU_END_TO_BE:
1966 switch (IMM) {
1967 case 16:
1968 DST = (__force u16) cpu_to_be16(DST);
1969 break;
1970 case 32:
1971 DST = (__force u32) cpu_to_be32(DST);
1972 break;
1973 case 64:
1974 DST = (__force u64) cpu_to_be64(DST);
1975 break;
1976 }
1977 CONT;
1978 ALU_END_TO_LE:
1979 switch (IMM) {
1980 case 16:
1981 DST = (__force u16) cpu_to_le16(DST);
1982 break;
1983 case 32:
1984 DST = (__force u32) cpu_to_le32(DST);
1985 break;
1986 case 64:
1987 DST = (__force u64) cpu_to_le64(DST);
1988 break;
1989 }
1990 CONT;
1991 ALU64_END_TO_LE:
1992 switch (IMM) {
1993 case 16:
1994 DST = (__force u16) __swab16(DST);
1995 break;
1996 case 32:
1997 DST = (__force u32) __swab32(DST);
1998 break;
1999 case 64:
2000 DST = (__force u64) __swab64(DST);
2001 break;
2002 }
2003 CONT;
2004
2005 /* CALL */
2006 JMP_CALL:
2007 /* Function call scratches BPF_R1-BPF_R5 registers,
2008 * preserves BPF_R6-BPF_R9, and stores return value
2009 * into BPF_R0.
2010 */
2011 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
2012 BPF_R4, BPF_R5);
2013 CONT;
2014
2015 JMP_CALL_ARGS:
2016 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
2017 BPF_R3, BPF_R4,
2018 BPF_R5,
2019 insn + insn->off + 1);
2020 CONT;
2021
2022 JMP_TAIL_CALL: {
2023 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
2024 struct bpf_array *array = container_of(map, struct bpf_array, map);
2025 struct bpf_prog *prog;
2026 u32 index = BPF_R3;
2027
2028 if (unlikely(index >= array->map.max_entries))
2029 goto out;
2030
2031 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
2032 goto out;
2033
2034 tail_call_cnt++;
2035
2036 prog = READ_ONCE(array->ptrs[index]);
2037 if (!prog)
2038 goto out;
2039
2040 /* ARG1 at this point is guaranteed to point to CTX from
2041 * the verifier side due to the fact that the tail call is
2042 * handled like a helper, that is, bpf_tail_call_proto,
2043 * where arg1_type is ARG_PTR_TO_CTX.
2044 */
2045 insn = prog->insnsi;
2046 goto select_insn;
2047 out:
2048 CONT;
2049 }
2050 JMP_JA:
2051 insn += insn->off;
2052 CONT;
2053 JMP32_JA:
2054 insn += insn->imm;
2055 CONT;
2056 JMP_EXIT:
2057 return BPF_R0;
2058 /* JMP */
2059 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
2060 JMP_##OPCODE##_X: \
2061 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
2062 insn += insn->off; \
2063 CONT_JMP; \
2064 } \
2065 CONT; \
2066 JMP32_##OPCODE##_X: \
2067 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
2068 insn += insn->off; \
2069 CONT_JMP; \
2070 } \
2071 CONT; \
2072 JMP_##OPCODE##_K: \
2073 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
2074 insn += insn->off; \
2075 CONT_JMP; \
2076 } \
2077 CONT; \
2078 JMP32_##OPCODE##_K: \
2079 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
2080 insn += insn->off; \
2081 CONT_JMP; \
2082 } \
2083 CONT;
2084 COND_JMP(u, JEQ, ==)
2085 COND_JMP(u, JNE, !=)
2086 COND_JMP(u, JGT, >)
2087 COND_JMP(u, JLT, <)
2088 COND_JMP(u, JGE, >=)
2089 COND_JMP(u, JLE, <=)
2090 COND_JMP(u, JSET, &)
2091 COND_JMP(s, JSGT, >)
2092 COND_JMP(s, JSLT, <)
2093 COND_JMP(s, JSGE, >=)
2094 COND_JMP(s, JSLE, <=)
2095 #undef COND_JMP
2096 /* ST, STX and LDX*/
2097 ST_NOSPEC:
2098 /* Speculation barrier for mitigating Speculative Store Bypass.
2099 * In case of arm64, we rely on the firmware mitigation as
2100 * controlled via the ssbd kernel parameter. Whenever the
2101 * mitigation is enabled, it works for all of the kernel code
2102 * with no need to provide any additional instructions here.
2103 * In case of x86, we use 'lfence' insn for mitigation. We
2104 * reuse preexisting logic from Spectre v1 mitigation that
2105 * happens to produce the required code on x86 for v4 as well.
2106 */
2107 barrier_nospec();
2108 CONT;
2109 #define LDST(SIZEOP, SIZE) \
2110 STX_MEM_##SIZEOP: \
2111 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
2112 CONT; \
2113 ST_MEM_##SIZEOP: \
2114 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
2115 CONT; \
2116 LDX_MEM_##SIZEOP: \
2117 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
2118 CONT; \
2119 LDX_PROBE_MEM_##SIZEOP: \
2120 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \
2121 (const void *)(long) (SRC + insn->off)); \
2122 DST = *((SIZE *)&DST); \
2123 CONT;
2124
2125 LDST(B, u8)
2126 LDST(H, u16)
2127 LDST(W, u32)
2128 LDST(DW, u64)
2129 #undef LDST
2130
2131 #define LDSX(SIZEOP, SIZE) \
2132 LDX_MEMSX_##SIZEOP: \
2133 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
2134 CONT; \
2135 LDX_PROBE_MEMSX_##SIZEOP: \
2136 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \
2137 (const void *)(long) (SRC + insn->off)); \
2138 DST = *((SIZE *)&DST); \
2139 CONT;
2140
2141 LDSX(B, s8)
2142 LDSX(H, s16)
2143 LDSX(W, s32)
2144 #undef LDSX
2145
2146 #define ATOMIC_ALU_OP(BOP, KOP) \
2147 case BOP: \
2148 if (BPF_SIZE(insn->code) == BPF_W) \
2149 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
2150 (DST + insn->off)); \
2151 else \
2152 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
2153 (DST + insn->off)); \
2154 break; \
2155 case BOP | BPF_FETCH: \
2156 if (BPF_SIZE(insn->code) == BPF_W) \
2157 SRC = (u32) atomic_fetch_##KOP( \
2158 (u32) SRC, \
2159 (atomic_t *)(unsigned long) (DST + insn->off)); \
2160 else \
2161 SRC = (u64) atomic64_fetch_##KOP( \
2162 (u64) SRC, \
2163 (atomic64_t *)(unsigned long) (DST + insn->off)); \
2164 break;
2165
2166 STX_ATOMIC_DW:
2167 STX_ATOMIC_W:
2168 switch (IMM) {
2169 ATOMIC_ALU_OP(BPF_ADD, add)
2170 ATOMIC_ALU_OP(BPF_AND, and)
2171 ATOMIC_ALU_OP(BPF_OR, or)
2172 ATOMIC_ALU_OP(BPF_XOR, xor)
2173 #undef ATOMIC_ALU_OP
2174
2175 case BPF_XCHG:
2176 if (BPF_SIZE(insn->code) == BPF_W)
2177 SRC = (u32) atomic_xchg(
2178 (atomic_t *)(unsigned long) (DST + insn->off),
2179 (u32) SRC);
2180 else
2181 SRC = (u64) atomic64_xchg(
2182 (atomic64_t *)(unsigned long) (DST + insn->off),
2183 (u64) SRC);
2184 break;
2185 case BPF_CMPXCHG:
2186 if (BPF_SIZE(insn->code) == BPF_W)
2187 BPF_R0 = (u32) atomic_cmpxchg(
2188 (atomic_t *)(unsigned long) (DST + insn->off),
2189 (u32) BPF_R0, (u32) SRC);
2190 else
2191 BPF_R0 = (u64) atomic64_cmpxchg(
2192 (atomic64_t *)(unsigned long) (DST + insn->off),
2193 (u64) BPF_R0, (u64) SRC);
2194 break;
2195
2196 default:
2197 goto default_label;
2198 }
2199 CONT;
2200
2201 default_label:
2202 /* If we ever reach this, we have a bug somewhere. Die hard here
2203 * instead of just returning 0; we could be somewhere in a subprog,
2204 * so execution could continue otherwise which we do /not/ want.
2205 *
2206 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
2207 */
2208 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
2209 insn->code, insn->imm);
2210 BUG_ON(1);
2211 return 0;
2212 }
2213
2214 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2215 #define DEFINE_BPF_PROG_RUN(stack_size) \
2216 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2217 { \
2218 u64 stack[stack_size / sizeof(u64)]; \
2219 u64 regs[MAX_BPF_EXT_REG] = {}; \
2220 \
2221 kmsan_unpoison_memory(stack, sizeof(stack)); \
2222 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2223 ARG1 = (u64) (unsigned long) ctx; \
2224 return ___bpf_prog_run(regs, insn); \
2225 }
2226
2227 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2228 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2229 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2230 const struct bpf_insn *insn) \
2231 { \
2232 u64 stack[stack_size / sizeof(u64)]; \
2233 u64 regs[MAX_BPF_EXT_REG]; \
2234 \
2235 kmsan_unpoison_memory(stack, sizeof(stack)); \
2236 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2237 BPF_R1 = r1; \
2238 BPF_R2 = r2; \
2239 BPF_R3 = r3; \
2240 BPF_R4 = r4; \
2241 BPF_R5 = r5; \
2242 return ___bpf_prog_run(regs, insn); \
2243 }
2244
2245 #define EVAL1(FN, X) FN(X)
2246 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2247 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2248 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2249 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2250 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2251
2252 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2253 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2254 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2255
2256 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2257 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2258 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2259
2260 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2261
2262 static unsigned int (*interpreters[])(const void *ctx,
2263 const struct bpf_insn *insn) = {
2264 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2265 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2266 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2267 };
2268 #undef PROG_NAME_LIST
2269 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2270 static __maybe_unused
2271 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2272 const struct bpf_insn *insn) = {
2273 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2274 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2275 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2276 };
2277 #undef PROG_NAME_LIST
2278
2279 #ifdef CONFIG_BPF_SYSCALL
2280 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2281 {
2282 stack_depth = max_t(u32, stack_depth, 1);
2283 insn->off = (s16) insn->imm;
2284 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2285 __bpf_call_base_args;
2286 insn->code = BPF_JMP | BPF_CALL_ARGS;
2287 }
2288 #endif
2289 #else
2290 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2291 const struct bpf_insn *insn)
2292 {
2293 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2294 * is not working properly, so warn about it!
2295 */
2296 WARN_ON_ONCE(1);
2297 return 0;
2298 }
2299 #endif
2300
2301 bool bpf_prog_map_compatible(struct bpf_map *map,
2302 const struct bpf_prog *fp)
2303 {
2304 enum bpf_prog_type prog_type = resolve_prog_type(fp);
2305 bool ret;
2306 struct bpf_prog_aux *aux = fp->aux;
2307
2308 if (fp->kprobe_override)
2309 return false;
2310
2311 /* XDP programs inserted into maps are not guaranteed to run on
2312 * a particular netdev (and can run outside driver context entirely
2313 * in the case of devmap and cpumap). Until device checks
2314 * are implemented, prohibit adding dev-bound programs to program maps.
2315 */
2316 if (bpf_prog_is_dev_bound(aux))
2317 return false;
2318
2319 spin_lock(&map->owner.lock);
2320 if (!map->owner.type) {
2321 /* There's no owner yet where we could check for
2322 * compatibility.
2323 */
2324 map->owner.type = prog_type;
2325 map->owner.jited = fp->jited;
2326 map->owner.xdp_has_frags = aux->xdp_has_frags;
2327 map->owner.attach_func_proto = aux->attach_func_proto;
2328 ret = true;
2329 } else {
2330 ret = map->owner.type == prog_type &&
2331 map->owner.jited == fp->jited &&
2332 map->owner.xdp_has_frags == aux->xdp_has_frags;
2333 if (ret &&
2334 map->owner.attach_func_proto != aux->attach_func_proto) {
2335 switch (prog_type) {
2336 case BPF_PROG_TYPE_TRACING:
2337 case BPF_PROG_TYPE_LSM:
2338 case BPF_PROG_TYPE_EXT:
2339 case BPF_PROG_TYPE_STRUCT_OPS:
2340 ret = false;
2341 break;
2342 default:
2343 break;
2344 }
2345 }
2346 }
2347 spin_unlock(&map->owner.lock);
2348
2349 return ret;
2350 }
2351
2352 static int bpf_check_tail_call(const struct bpf_prog *fp)
2353 {
2354 struct bpf_prog_aux *aux = fp->aux;
2355 int i, ret = 0;
2356
2357 mutex_lock(&aux->used_maps_mutex);
2358 for (i = 0; i < aux->used_map_cnt; i++) {
2359 struct bpf_map *map = aux->used_maps[i];
2360
2361 if (!map_type_contains_progs(map))
2362 continue;
2363
2364 if (!bpf_prog_map_compatible(map, fp)) {
2365 ret = -EINVAL;
2366 goto out;
2367 }
2368 }
2369
2370 out:
2371 mutex_unlock(&aux->used_maps_mutex);
2372 return ret;
2373 }
2374
2375 static void bpf_prog_select_func(struct bpf_prog *fp)
2376 {
2377 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2378 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2379
2380 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2381 #else
2382 fp->bpf_func = __bpf_prog_ret0_warn;
2383 #endif
2384 }
2385
2386 /**
2387 * bpf_prog_select_runtime - select exec runtime for BPF program
2388 * @fp: bpf_prog populated with BPF program
2389 * @err: pointer to error variable
2390 *
2391 * Try to JIT eBPF program, if JIT is not available, use interpreter.
2392 * The BPF program will be executed via bpf_prog_run() function.
2393 *
2394 * Return: the &fp argument along with &err set to 0 for success or
2395 * a negative errno code on failure
2396 */
2397 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2398 {
2399 /* In case of BPF to BPF calls, verifier did all the prep
2400 * work with regards to JITing, etc.
2401 */
2402 bool jit_needed = false;
2403
2404 if (fp->bpf_func)
2405 goto finalize;
2406
2407 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2408 bpf_prog_has_kfunc_call(fp))
2409 jit_needed = true;
2410
2411 bpf_prog_select_func(fp);
2412
2413 /* eBPF JITs can rewrite the program in case constant
2414 * blinding is active. However, in case of error during
2415 * blinding, bpf_int_jit_compile() must always return a
2416 * valid program, which in this case would simply not
2417 * be JITed, but falls back to the interpreter.
2418 */
2419 if (!bpf_prog_is_offloaded(fp->aux)) {
2420 *err = bpf_prog_alloc_jited_linfo(fp);
2421 if (*err)
2422 return fp;
2423
2424 fp = bpf_int_jit_compile(fp);
2425 bpf_prog_jit_attempt_done(fp);
2426 if (!fp->jited && jit_needed) {
2427 *err = -ENOTSUPP;
2428 return fp;
2429 }
2430 } else {
2431 *err = bpf_prog_offload_compile(fp);
2432 if (*err)
2433 return fp;
2434 }
2435
2436 finalize:
2437 *err = bpf_prog_lock_ro(fp);
2438 if (*err)
2439 return fp;
2440
2441 /* The tail call compatibility check can only be done at
2442 * this late stage as we need to determine, if we deal
2443 * with JITed or non JITed program concatenations and not
2444 * all eBPF JITs might immediately support all features.
2445 */
2446 *err = bpf_check_tail_call(fp);
2447
2448 return fp;
2449 }
2450 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2451
2452 static unsigned int __bpf_prog_ret1(const void *ctx,
2453 const struct bpf_insn *insn)
2454 {
2455 return 1;
2456 }
2457
2458 static struct bpf_prog_dummy {
2459 struct bpf_prog prog;
2460 } dummy_bpf_prog = {
2461 .prog = {
2462 .bpf_func = __bpf_prog_ret1,
2463 },
2464 };
2465
2466 struct bpf_empty_prog_array bpf_empty_prog_array = {
2467 .null_prog = NULL,
2468 };
2469 EXPORT_SYMBOL(bpf_empty_prog_array);
2470
2471 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2472 {
2473 struct bpf_prog_array *p;
2474
2475 if (prog_cnt)
2476 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags);
2477 else
2478 p = &bpf_empty_prog_array.hdr;
2479
2480 return p;
2481 }
2482
2483 void bpf_prog_array_free(struct bpf_prog_array *progs)
2484 {
2485 if (!progs || progs == &bpf_empty_prog_array.hdr)
2486 return;
2487 kfree_rcu(progs, rcu);
2488 }
2489
2490 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2491 {
2492 struct bpf_prog_array *progs;
2493
2494 /* If RCU Tasks Trace grace period implies RCU grace period, there is
2495 * no need to call kfree_rcu(), just call kfree() directly.
2496 */
2497 progs = container_of(rcu, struct bpf_prog_array, rcu);
2498 if (rcu_trace_implies_rcu_gp())
2499 kfree(progs);
2500 else
2501 kfree_rcu(progs, rcu);
2502 }
2503
2504 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2505 {
2506 if (!progs || progs == &bpf_empty_prog_array.hdr)
2507 return;
2508 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2509 }
2510
2511 int bpf_prog_array_length(struct bpf_prog_array *array)
2512 {
2513 struct bpf_prog_array_item *item;
2514 u32 cnt = 0;
2515
2516 for (item = array->items; item->prog; item++)
2517 if (item->prog != &dummy_bpf_prog.prog)
2518 cnt++;
2519 return cnt;
2520 }
2521
2522 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2523 {
2524 struct bpf_prog_array_item *item;
2525
2526 for (item = array->items; item->prog; item++)
2527 if (item->prog != &dummy_bpf_prog.prog)
2528 return false;
2529 return true;
2530 }
2531
2532 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2533 u32 *prog_ids,
2534 u32 request_cnt)
2535 {
2536 struct bpf_prog_array_item *item;
2537 int i = 0;
2538
2539 for (item = array->items; item->prog; item++) {
2540 if (item->prog == &dummy_bpf_prog.prog)
2541 continue;
2542 prog_ids[i] = item->prog->aux->id;
2543 if (++i == request_cnt) {
2544 item++;
2545 break;
2546 }
2547 }
2548
2549 return !!(item->prog);
2550 }
2551
2552 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2553 __u32 __user *prog_ids, u32 cnt)
2554 {
2555 unsigned long err = 0;
2556 bool nospc;
2557 u32 *ids;
2558
2559 /* users of this function are doing:
2560 * cnt = bpf_prog_array_length();
2561 * if (cnt > 0)
2562 * bpf_prog_array_copy_to_user(..., cnt);
2563 * so below kcalloc doesn't need extra cnt > 0 check.
2564 */
2565 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2566 if (!ids)
2567 return -ENOMEM;
2568 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2569 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2570 kfree(ids);
2571 if (err)
2572 return -EFAULT;
2573 if (nospc)
2574 return -ENOSPC;
2575 return 0;
2576 }
2577
2578 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2579 struct bpf_prog *old_prog)
2580 {
2581 struct bpf_prog_array_item *item;
2582
2583 for (item = array->items; item->prog; item++)
2584 if (item->prog == old_prog) {
2585 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2586 break;
2587 }
2588 }
2589
2590 /**
2591 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2592 * index into the program array with
2593 * a dummy no-op program.
2594 * @array: a bpf_prog_array
2595 * @index: the index of the program to replace
2596 *
2597 * Skips over dummy programs, by not counting them, when calculating
2598 * the position of the program to replace.
2599 *
2600 * Return:
2601 * * 0 - Success
2602 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2603 * * -ENOENT - Index out of range
2604 */
2605 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2606 {
2607 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2608 }
2609
2610 /**
2611 * bpf_prog_array_update_at() - Updates the program at the given index
2612 * into the program array.
2613 * @array: a bpf_prog_array
2614 * @index: the index of the program to update
2615 * @prog: the program to insert into the array
2616 *
2617 * Skips over dummy programs, by not counting them, when calculating
2618 * the position of the program to update.
2619 *
2620 * Return:
2621 * * 0 - Success
2622 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2623 * * -ENOENT - Index out of range
2624 */
2625 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2626 struct bpf_prog *prog)
2627 {
2628 struct bpf_prog_array_item *item;
2629
2630 if (unlikely(index < 0))
2631 return -EINVAL;
2632
2633 for (item = array->items; item->prog; item++) {
2634 if (item->prog == &dummy_bpf_prog.prog)
2635 continue;
2636 if (!index) {
2637 WRITE_ONCE(item->prog, prog);
2638 return 0;
2639 }
2640 index--;
2641 }
2642 return -ENOENT;
2643 }
2644
2645 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2646 struct bpf_prog *exclude_prog,
2647 struct bpf_prog *include_prog,
2648 u64 bpf_cookie,
2649 struct bpf_prog_array **new_array)
2650 {
2651 int new_prog_cnt, carry_prog_cnt = 0;
2652 struct bpf_prog_array_item *existing, *new;
2653 struct bpf_prog_array *array;
2654 bool found_exclude = false;
2655
2656 /* Figure out how many existing progs we need to carry over to
2657 * the new array.
2658 */
2659 if (old_array) {
2660 existing = old_array->items;
2661 for (; existing->prog; existing++) {
2662 if (existing->prog == exclude_prog) {
2663 found_exclude = true;
2664 continue;
2665 }
2666 if (existing->prog != &dummy_bpf_prog.prog)
2667 carry_prog_cnt++;
2668 if (existing->prog == include_prog)
2669 return -EEXIST;
2670 }
2671 }
2672
2673 if (exclude_prog && !found_exclude)
2674 return -ENOENT;
2675
2676 /* How many progs (not NULL) will be in the new array? */
2677 new_prog_cnt = carry_prog_cnt;
2678 if (include_prog)
2679 new_prog_cnt += 1;
2680
2681 /* Do we have any prog (not NULL) in the new array? */
2682 if (!new_prog_cnt) {
2683 *new_array = NULL;
2684 return 0;
2685 }
2686
2687 /* +1 as the end of prog_array is marked with NULL */
2688 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2689 if (!array)
2690 return -ENOMEM;
2691 new = array->items;
2692
2693 /* Fill in the new prog array */
2694 if (carry_prog_cnt) {
2695 existing = old_array->items;
2696 for (; existing->prog; existing++) {
2697 if (existing->prog == exclude_prog ||
2698 existing->prog == &dummy_bpf_prog.prog)
2699 continue;
2700
2701 new->prog = existing->prog;
2702 new->bpf_cookie = existing->bpf_cookie;
2703 new++;
2704 }
2705 }
2706 if (include_prog) {
2707 new->prog = include_prog;
2708 new->bpf_cookie = bpf_cookie;
2709 new++;
2710 }
2711 new->prog = NULL;
2712 *new_array = array;
2713 return 0;
2714 }
2715
2716 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2717 u32 *prog_ids, u32 request_cnt,
2718 u32 *prog_cnt)
2719 {
2720 u32 cnt = 0;
2721
2722 if (array)
2723 cnt = bpf_prog_array_length(array);
2724
2725 *prog_cnt = cnt;
2726
2727 /* return early if user requested only program count or nothing to copy */
2728 if (!request_cnt || !cnt)
2729 return 0;
2730
2731 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2732 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2733 : 0;
2734 }
2735
2736 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2737 struct bpf_map **used_maps, u32 len)
2738 {
2739 struct bpf_map *map;
2740 bool sleepable;
2741 u32 i;
2742
2743 sleepable = aux->prog->sleepable;
2744 for (i = 0; i < len; i++) {
2745 map = used_maps[i];
2746 if (map->ops->map_poke_untrack)
2747 map->ops->map_poke_untrack(map, aux);
2748 if (sleepable)
2749 atomic64_dec(&map->sleepable_refcnt);
2750 bpf_map_put(map);
2751 }
2752 }
2753
2754 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2755 {
2756 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2757 kfree(aux->used_maps);
2758 }
2759
2760 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len)
2761 {
2762 #ifdef CONFIG_BPF_SYSCALL
2763 struct btf_mod_pair *btf_mod;
2764 u32 i;
2765
2766 for (i = 0; i < len; i++) {
2767 btf_mod = &used_btfs[i];
2768 if (btf_mod->module)
2769 module_put(btf_mod->module);
2770 btf_put(btf_mod->btf);
2771 }
2772 #endif
2773 }
2774
2775 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2776 {
2777 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt);
2778 kfree(aux->used_btfs);
2779 }
2780
2781 static void bpf_prog_free_deferred(struct work_struct *work)
2782 {
2783 struct bpf_prog_aux *aux;
2784 int i;
2785
2786 aux = container_of(work, struct bpf_prog_aux, work);
2787 #ifdef CONFIG_BPF_SYSCALL
2788 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2789 #endif
2790 #ifdef CONFIG_CGROUP_BPF
2791 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2792 bpf_cgroup_atype_put(aux->cgroup_atype);
2793 #endif
2794 bpf_free_used_maps(aux);
2795 bpf_free_used_btfs(aux);
2796 if (bpf_prog_is_dev_bound(aux))
2797 bpf_prog_dev_bound_destroy(aux->prog);
2798 #ifdef CONFIG_PERF_EVENTS
2799 if (aux->prog->has_callchain_buf)
2800 put_callchain_buffers();
2801 #endif
2802 if (aux->dst_trampoline)
2803 bpf_trampoline_put(aux->dst_trampoline);
2804 for (i = 0; i < aux->real_func_cnt; i++) {
2805 /* We can just unlink the subprog poke descriptor table as
2806 * it was originally linked to the main program and is also
2807 * released along with it.
2808 */
2809 aux->func[i]->aux->poke_tab = NULL;
2810 bpf_jit_free(aux->func[i]);
2811 }
2812 if (aux->real_func_cnt) {
2813 kfree(aux->func);
2814 bpf_prog_unlock_free(aux->prog);
2815 } else {
2816 bpf_jit_free(aux->prog);
2817 }
2818 }
2819
2820 void bpf_prog_free(struct bpf_prog *fp)
2821 {
2822 struct bpf_prog_aux *aux = fp->aux;
2823
2824 if (aux->dst_prog)
2825 bpf_prog_put(aux->dst_prog);
2826 bpf_token_put(aux->token);
2827 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2828 schedule_work(&aux->work);
2829 }
2830 EXPORT_SYMBOL_GPL(bpf_prog_free);
2831
2832 /* RNG for unprivileged user space with separated state from prandom_u32(). */
2833 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2834
2835 void bpf_user_rnd_init_once(void)
2836 {
2837 prandom_init_once(&bpf_user_rnd_state);
2838 }
2839
2840 BPF_CALL_0(bpf_user_rnd_u32)
2841 {
2842 /* Should someone ever have the rather unwise idea to use some
2843 * of the registers passed into this function, then note that
2844 * this function is called from native eBPF and classic-to-eBPF
2845 * transformations. Register assignments from both sides are
2846 * different, f.e. classic always sets fn(ctx, A, X) here.
2847 */
2848 struct rnd_state *state;
2849 u32 res;
2850
2851 state = &get_cpu_var(bpf_user_rnd_state);
2852 res = prandom_u32_state(state);
2853 put_cpu_var(bpf_user_rnd_state);
2854
2855 return res;
2856 }
2857
2858 BPF_CALL_0(bpf_get_raw_cpu_id)
2859 {
2860 return raw_smp_processor_id();
2861 }
2862
2863 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2864 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2865 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2866 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2867 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2868 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2869 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2870 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2871 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2872 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2873 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2874
2875 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2876 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2877 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2878 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2879 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2880 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2881 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak;
2882
2883 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2884 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2885 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2886 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2887 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2888 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2889 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2890 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2891 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2892 const struct bpf_func_proto bpf_set_retval_proto __weak;
2893 const struct bpf_func_proto bpf_get_retval_proto __weak;
2894
2895 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2896 {
2897 return NULL;
2898 }
2899
2900 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2901 {
2902 return NULL;
2903 }
2904
2905 u64 __weak
2906 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2907 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2908 {
2909 return -ENOTSUPP;
2910 }
2911 EXPORT_SYMBOL_GPL(bpf_event_output);
2912
2913 /* Always built-in helper functions. */
2914 const struct bpf_func_proto bpf_tail_call_proto = {
2915 .func = NULL,
2916 .gpl_only = false,
2917 .ret_type = RET_VOID,
2918 .arg1_type = ARG_PTR_TO_CTX,
2919 .arg2_type = ARG_CONST_MAP_PTR,
2920 .arg3_type = ARG_ANYTHING,
2921 };
2922
2923 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2924 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2925 * eBPF and implicitly also cBPF can get JITed!
2926 */
2927 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2928 {
2929 return prog;
2930 }
2931
2932 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2933 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2934 */
2935 void __weak bpf_jit_compile(struct bpf_prog *prog)
2936 {
2937 }
2938
2939 bool __weak bpf_helper_changes_pkt_data(void *func)
2940 {
2941 return false;
2942 }
2943
2944 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2945 * analysis code and wants explicit zero extension inserted by verifier.
2946 * Otherwise, return FALSE.
2947 *
2948 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2949 * you don't override this. JITs that don't want these extra insns can detect
2950 * them using insn_is_zext.
2951 */
2952 bool __weak bpf_jit_needs_zext(void)
2953 {
2954 return false;
2955 }
2956
2957 /* Return true if the JIT inlines the call to the helper corresponding to
2958 * the imm.
2959 *
2960 * The verifier will not patch the insn->imm for the call to the helper if
2961 * this returns true.
2962 */
2963 bool __weak bpf_jit_inlines_helper_call(s32 imm)
2964 {
2965 return false;
2966 }
2967
2968 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2969 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2970 {
2971 return false;
2972 }
2973
2974 bool __weak bpf_jit_supports_percpu_insn(void)
2975 {
2976 return false;
2977 }
2978
2979 bool __weak bpf_jit_supports_kfunc_call(void)
2980 {
2981 return false;
2982 }
2983
2984 bool __weak bpf_jit_supports_far_kfunc_call(void)
2985 {
2986 return false;
2987 }
2988
2989 bool __weak bpf_jit_supports_arena(void)
2990 {
2991 return false;
2992 }
2993
2994 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
2995 {
2996 return false;
2997 }
2998
2999 u64 __weak bpf_arch_uaddress_limit(void)
3000 {
3001 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE)
3002 return TASK_SIZE;
3003 #else
3004 return 0;
3005 #endif
3006 }
3007
3008 /* Return TRUE if the JIT backend satisfies the following two conditions:
3009 * 1) JIT backend supports atomic_xchg() on pointer-sized words.
3010 * 2) Under the specific arch, the implementation of xchg() is the same
3011 * as atomic_xchg() on pointer-sized words.
3012 */
3013 bool __weak bpf_jit_supports_ptr_xchg(void)
3014 {
3015 return false;
3016 }
3017
3018 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
3019 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
3020 */
3021 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
3022 int len)
3023 {
3024 return -EFAULT;
3025 }
3026
3027 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
3028 void *addr1, void *addr2)
3029 {
3030 return -ENOTSUPP;
3031 }
3032
3033 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
3034 {
3035 return ERR_PTR(-ENOTSUPP);
3036 }
3037
3038 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
3039 {
3040 return -ENOTSUPP;
3041 }
3042
3043 bool __weak bpf_jit_supports_exceptions(void)
3044 {
3045 return false;
3046 }
3047
3048 bool __weak bpf_jit_supports_private_stack(void)
3049 {
3050 return false;
3051 }
3052
3053 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
3054 {
3055 }
3056
3057 /* for configs without MMU or 32-bit */
3058 __weak const struct bpf_map_ops arena_map_ops;
3059 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena)
3060 {
3061 return 0;
3062 }
3063 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena)
3064 {
3065 return 0;
3066 }
3067
3068 #ifdef CONFIG_BPF_SYSCALL
3069 static int __init bpf_global_ma_init(void)
3070 {
3071 int ret;
3072
3073 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false);
3074 bpf_global_ma_set = !ret;
3075 return ret;
3076 }
3077 late_initcall(bpf_global_ma_init);
3078 #endif
3079
3080 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
3081 EXPORT_SYMBOL(bpf_stats_enabled_key);
3082
3083 /* All definitions of tracepoints related to BPF. */
3084 #define CREATE_TRACE_POINTS
3085 #include <linux/bpf_trace.h>
3086
3087 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
3088 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);
Kernel DRM miscellaneous fixes and cross-tree changes