c37c454b09ebb8713f3d34612dddcfb1d25dcd7f
[platform/kernel/linux-rpi.git] / kernel / bpf / core.c
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/random.h>
25 #include <linux/moduleloader.h>
26 #include <linux/bpf.h>
27 #include <linux/btf.h>
28 #include <linux/objtool.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
41 #include <asm/barrier.h>
42 #include <asm/unaligned.h>
43
44 /* Registers */
45 #define BPF_R0  regs[BPF_REG_0]
46 #define BPF_R1  regs[BPF_REG_1]
47 #define BPF_R2  regs[BPF_REG_2]
48 #define BPF_R3  regs[BPF_REG_3]
49 #define BPF_R4  regs[BPF_REG_4]
50 #define BPF_R5  regs[BPF_REG_5]
51 #define BPF_R6  regs[BPF_REG_6]
52 #define BPF_R7  regs[BPF_REG_7]
53 #define BPF_R8  regs[BPF_REG_8]
54 #define BPF_R9  regs[BPF_REG_9]
55 #define BPF_R10 regs[BPF_REG_10]
56
57 /* Named registers */
58 #define DST     regs[insn->dst_reg]
59 #define SRC     regs[insn->src_reg]
60 #define FP      regs[BPF_REG_FP]
61 #define AX      regs[BPF_REG_AX]
62 #define ARG1    regs[BPF_REG_ARG1]
63 #define CTX     regs[BPF_REG_CTX]
64 #define OFF     insn->off
65 #define IMM     insn->imm
66
67 struct bpf_mem_alloc bpf_global_ma;
68 bool bpf_global_ma_set;
69
70 /* No hurry in this branch
71  *
72  * Exported for the bpf jit load helper.
73  */
74 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
75 {
76         u8 *ptr = NULL;
77
78         if (k >= SKF_NET_OFF) {
79                 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
80         } else if (k >= SKF_LL_OFF) {
81                 if (unlikely(!skb_mac_header_was_set(skb)))
82                         return NULL;
83                 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
84         }
85         if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
86                 return ptr;
87
88         return NULL;
89 }
90
91 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
92 {
93         gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
94         struct bpf_prog_aux *aux;
95         struct bpf_prog *fp;
96
97         size = round_up(size, PAGE_SIZE);
98         fp = __vmalloc(size, gfp_flags);
99         if (fp == NULL)
100                 return NULL;
101
102         aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
103         if (aux == NULL) {
104                 vfree(fp);
105                 return NULL;
106         }
107         fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
108         if (!fp->active) {
109                 vfree(fp);
110                 kfree(aux);
111                 return NULL;
112         }
113
114         fp->pages = size / PAGE_SIZE;
115         fp->aux = aux;
116         fp->aux->prog = fp;
117         fp->jit_requested = ebpf_jit_enabled();
118         fp->blinding_requested = bpf_jit_blinding_enabled(fp);
119 #ifdef CONFIG_CGROUP_BPF
120         aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
121 #endif
122
123         INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
124         mutex_init(&fp->aux->used_maps_mutex);
125         mutex_init(&fp->aux->dst_mutex);
126
127         return fp;
128 }
129
130 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
131 {
132         gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
133         struct bpf_prog *prog;
134         int cpu;
135
136         prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
137         if (!prog)
138                 return NULL;
139
140         prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
141         if (!prog->stats) {
142                 free_percpu(prog->active);
143                 kfree(prog->aux);
144                 vfree(prog);
145                 return NULL;
146         }
147
148         for_each_possible_cpu(cpu) {
149                 struct bpf_prog_stats *pstats;
150
151                 pstats = per_cpu_ptr(prog->stats, cpu);
152                 u64_stats_init(&pstats->syncp);
153         }
154         return prog;
155 }
156 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
157
158 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
159 {
160         if (!prog->aux->nr_linfo || !prog->jit_requested)
161                 return 0;
162
163         prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
164                                           sizeof(*prog->aux->jited_linfo),
165                                           bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN));
166         if (!prog->aux->jited_linfo)
167                 return -ENOMEM;
168
169         return 0;
170 }
171
172 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
173 {
174         if (prog->aux->jited_linfo &&
175             (!prog->jited || !prog->aux->jited_linfo[0])) {
176                 kvfree(prog->aux->jited_linfo);
177                 prog->aux->jited_linfo = NULL;
178         }
179
180         kfree(prog->aux->kfunc_tab);
181         prog->aux->kfunc_tab = NULL;
182 }
183
184 /* The jit engine is responsible to provide an array
185  * for insn_off to the jited_off mapping (insn_to_jit_off).
186  *
187  * The idx to this array is the insn_off.  Hence, the insn_off
188  * here is relative to the prog itself instead of the main prog.
189  * This array has one entry for each xlated bpf insn.
190  *
191  * jited_off is the byte off to the end of the jited insn.
192  *
193  * Hence, with
194  * insn_start:
195  *      The first bpf insn off of the prog.  The insn off
196  *      here is relative to the main prog.
197  *      e.g. if prog is a subprog, insn_start > 0
198  * linfo_idx:
199  *      The prog's idx to prog->aux->linfo and jited_linfo
200  *
201  * jited_linfo[linfo_idx] = prog->bpf_func
202  *
203  * For i > linfo_idx,
204  *
205  * jited_linfo[i] = prog->bpf_func +
206  *      insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
207  */
208 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
209                                const u32 *insn_to_jit_off)
210 {
211         u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
212         const struct bpf_line_info *linfo;
213         void **jited_linfo;
214
215         if (!prog->aux->jited_linfo)
216                 /* Userspace did not provide linfo */
217                 return;
218
219         linfo_idx = prog->aux->linfo_idx;
220         linfo = &prog->aux->linfo[linfo_idx];
221         insn_start = linfo[0].insn_off;
222         insn_end = insn_start + prog->len;
223
224         jited_linfo = &prog->aux->jited_linfo[linfo_idx];
225         jited_linfo[0] = prog->bpf_func;
226
227         nr_linfo = prog->aux->nr_linfo - linfo_idx;
228
229         for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
230                 /* The verifier ensures that linfo[i].insn_off is
231                  * strictly increasing
232                  */
233                 jited_linfo[i] = prog->bpf_func +
234                         insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
235 }
236
237 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
238                                   gfp_t gfp_extra_flags)
239 {
240         gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
241         struct bpf_prog *fp;
242         u32 pages;
243
244         size = round_up(size, PAGE_SIZE);
245         pages = size / PAGE_SIZE;
246         if (pages <= fp_old->pages)
247                 return fp_old;
248
249         fp = __vmalloc(size, gfp_flags);
250         if (fp) {
251                 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
252                 fp->pages = pages;
253                 fp->aux->prog = fp;
254
255                 /* We keep fp->aux from fp_old around in the new
256                  * reallocated structure.
257                  */
258                 fp_old->aux = NULL;
259                 fp_old->stats = NULL;
260                 fp_old->active = NULL;
261                 __bpf_prog_free(fp_old);
262         }
263
264         return fp;
265 }
266
267 void __bpf_prog_free(struct bpf_prog *fp)
268 {
269         if (fp->aux) {
270                 mutex_destroy(&fp->aux->used_maps_mutex);
271                 mutex_destroy(&fp->aux->dst_mutex);
272                 kfree(fp->aux->poke_tab);
273                 kfree(fp->aux);
274         }
275         free_percpu(fp->stats);
276         free_percpu(fp->active);
277         vfree(fp);
278 }
279
280 int bpf_prog_calc_tag(struct bpf_prog *fp)
281 {
282         const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
283         u32 raw_size = bpf_prog_tag_scratch_size(fp);
284         u32 digest[SHA1_DIGEST_WORDS];
285         u32 ws[SHA1_WORKSPACE_WORDS];
286         u32 i, bsize, psize, blocks;
287         struct bpf_insn *dst;
288         bool was_ld_map;
289         u8 *raw, *todo;
290         __be32 *result;
291         __be64 *bits;
292
293         raw = vmalloc(raw_size);
294         if (!raw)
295                 return -ENOMEM;
296
297         sha1_init(digest);
298         memset(ws, 0, sizeof(ws));
299
300         /* We need to take out the map fd for the digest calculation
301          * since they are unstable from user space side.
302          */
303         dst = (void *)raw;
304         for (i = 0, was_ld_map = false; i < fp->len; i++) {
305                 dst[i] = fp->insnsi[i];
306                 if (!was_ld_map &&
307                     dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
308                     (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
309                      dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
310                         was_ld_map = true;
311                         dst[i].imm = 0;
312                 } else if (was_ld_map &&
313                            dst[i].code == 0 &&
314                            dst[i].dst_reg == 0 &&
315                            dst[i].src_reg == 0 &&
316                            dst[i].off == 0) {
317                         was_ld_map = false;
318                         dst[i].imm = 0;
319                 } else {
320                         was_ld_map = false;
321                 }
322         }
323
324         psize = bpf_prog_insn_size(fp);
325         memset(&raw[psize], 0, raw_size - psize);
326         raw[psize++] = 0x80;
327
328         bsize  = round_up(psize, SHA1_BLOCK_SIZE);
329         blocks = bsize / SHA1_BLOCK_SIZE;
330         todo   = raw;
331         if (bsize - psize >= sizeof(__be64)) {
332                 bits = (__be64 *)(todo + bsize - sizeof(__be64));
333         } else {
334                 bits = (__be64 *)(todo + bsize + bits_offset);
335                 blocks++;
336         }
337         *bits = cpu_to_be64((psize - 1) << 3);
338
339         while (blocks--) {
340                 sha1_transform(digest, todo, ws);
341                 todo += SHA1_BLOCK_SIZE;
342         }
343
344         result = (__force __be32 *)digest;
345         for (i = 0; i < SHA1_DIGEST_WORDS; i++)
346                 result[i] = cpu_to_be32(digest[i]);
347         memcpy(fp->tag, result, sizeof(fp->tag));
348
349         vfree(raw);
350         return 0;
351 }
352
353 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
354                                 s32 end_new, s32 curr, const bool probe_pass)
355 {
356         const s64 imm_min = S32_MIN, imm_max = S32_MAX;
357         s32 delta = end_new - end_old;
358         s64 imm = insn->imm;
359
360         if (curr < pos && curr + imm + 1 >= end_old)
361                 imm += delta;
362         else if (curr >= end_new && curr + imm + 1 < end_new)
363                 imm -= delta;
364         if (imm < imm_min || imm > imm_max)
365                 return -ERANGE;
366         if (!probe_pass)
367                 insn->imm = imm;
368         return 0;
369 }
370
371 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
372                                 s32 end_new, s32 curr, const bool probe_pass)
373 {
374         const s32 off_min = S16_MIN, off_max = S16_MAX;
375         s32 delta = end_new - end_old;
376         s32 off = insn->off;
377
378         if (curr < pos && curr + off + 1 >= end_old)
379                 off += delta;
380         else if (curr >= end_new && curr + off + 1 < end_new)
381                 off -= delta;
382         if (off < off_min || off > off_max)
383                 return -ERANGE;
384         if (!probe_pass)
385                 insn->off = off;
386         return 0;
387 }
388
389 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
390                             s32 end_new, const bool probe_pass)
391 {
392         u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
393         struct bpf_insn *insn = prog->insnsi;
394         int ret = 0;
395
396         for (i = 0; i < insn_cnt; i++, insn++) {
397                 u8 code;
398
399                 /* In the probing pass we still operate on the original,
400                  * unpatched image in order to check overflows before we
401                  * do any other adjustments. Therefore skip the patchlet.
402                  */
403                 if (probe_pass && i == pos) {
404                         i = end_new;
405                         insn = prog->insnsi + end_old;
406                 }
407                 if (bpf_pseudo_func(insn)) {
408                         ret = bpf_adj_delta_to_imm(insn, pos, end_old,
409                                                    end_new, i, probe_pass);
410                         if (ret)
411                                 return ret;
412                         continue;
413                 }
414                 code = insn->code;
415                 if ((BPF_CLASS(code) != BPF_JMP &&
416                      BPF_CLASS(code) != BPF_JMP32) ||
417                     BPF_OP(code) == BPF_EXIT)
418                         continue;
419                 /* Adjust offset of jmps if we cross patch boundaries. */
420                 if (BPF_OP(code) == BPF_CALL) {
421                         if (insn->src_reg != BPF_PSEUDO_CALL)
422                                 continue;
423                         ret = bpf_adj_delta_to_imm(insn, pos, end_old,
424                                                    end_new, i, probe_pass);
425                 } else {
426                         ret = bpf_adj_delta_to_off(insn, pos, end_old,
427                                                    end_new, i, probe_pass);
428                 }
429                 if (ret)
430                         break;
431         }
432
433         return ret;
434 }
435
436 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
437 {
438         struct bpf_line_info *linfo;
439         u32 i, nr_linfo;
440
441         nr_linfo = prog->aux->nr_linfo;
442         if (!nr_linfo || !delta)
443                 return;
444
445         linfo = prog->aux->linfo;
446
447         for (i = 0; i < nr_linfo; i++)
448                 if (off < linfo[i].insn_off)
449                         break;
450
451         /* Push all off < linfo[i].insn_off by delta */
452         for (; i < nr_linfo; i++)
453                 linfo[i].insn_off += delta;
454 }
455
456 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
457                                        const struct bpf_insn *patch, u32 len)
458 {
459         u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
460         const u32 cnt_max = S16_MAX;
461         struct bpf_prog *prog_adj;
462         int err;
463
464         /* Since our patchlet doesn't expand the image, we're done. */
465         if (insn_delta == 0) {
466                 memcpy(prog->insnsi + off, patch, sizeof(*patch));
467                 return prog;
468         }
469
470         insn_adj_cnt = prog->len + insn_delta;
471
472         /* Reject anything that would potentially let the insn->off
473          * target overflow when we have excessive program expansions.
474          * We need to probe here before we do any reallocation where
475          * we afterwards may not fail anymore.
476          */
477         if (insn_adj_cnt > cnt_max &&
478             (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
479                 return ERR_PTR(err);
480
481         /* Several new instructions need to be inserted. Make room
482          * for them. Likely, there's no need for a new allocation as
483          * last page could have large enough tailroom.
484          */
485         prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
486                                     GFP_USER);
487         if (!prog_adj)
488                 return ERR_PTR(-ENOMEM);
489
490         prog_adj->len = insn_adj_cnt;
491
492         /* Patching happens in 3 steps:
493          *
494          * 1) Move over tail of insnsi from next instruction onwards,
495          *    so we can patch the single target insn with one or more
496          *    new ones (patching is always from 1 to n insns, n > 0).
497          * 2) Inject new instructions at the target location.
498          * 3) Adjust branch offsets if necessary.
499          */
500         insn_rest = insn_adj_cnt - off - len;
501
502         memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
503                 sizeof(*patch) * insn_rest);
504         memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
505
506         /* We are guaranteed to not fail at this point, otherwise
507          * the ship has sailed to reverse to the original state. An
508          * overflow cannot happen at this point.
509          */
510         BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
511
512         bpf_adj_linfo(prog_adj, off, insn_delta);
513
514         return prog_adj;
515 }
516
517 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
518 {
519         /* Branch offsets can't overflow when program is shrinking, no need
520          * to call bpf_adj_branches(..., true) here
521          */
522         memmove(prog->insnsi + off, prog->insnsi + off + cnt,
523                 sizeof(struct bpf_insn) * (prog->len - off - cnt));
524         prog->len -= cnt;
525
526         return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
527 }
528
529 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
530 {
531         int i;
532
533         for (i = 0; i < fp->aux->func_cnt; i++)
534                 bpf_prog_kallsyms_del(fp->aux->func[i]);
535 }
536
537 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
538 {
539         bpf_prog_kallsyms_del_subprogs(fp);
540         bpf_prog_kallsyms_del(fp);
541 }
542
543 #ifdef CONFIG_BPF_JIT
544 /* All BPF JIT sysctl knobs here. */
545 int bpf_jit_enable   __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
546 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
547 int bpf_jit_harden   __read_mostly;
548 long bpf_jit_limit   __read_mostly;
549 long bpf_jit_limit_max __read_mostly;
550
551 static void
552 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
553 {
554         WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
555
556         prog->aux->ksym.start = (unsigned long) prog->bpf_func;
557         prog->aux->ksym.end   = prog->aux->ksym.start + prog->jited_len;
558 }
559
560 static void
561 bpf_prog_ksym_set_name(struct bpf_prog *prog)
562 {
563         char *sym = prog->aux->ksym.name;
564         const char *end = sym + KSYM_NAME_LEN;
565         const struct btf_type *type;
566         const char *func_name;
567
568         BUILD_BUG_ON(sizeof("bpf_prog_") +
569                      sizeof(prog->tag) * 2 +
570                      /* name has been null terminated.
571                       * We should need +1 for the '_' preceding
572                       * the name.  However, the null character
573                       * is double counted between the name and the
574                       * sizeof("bpf_prog_") above, so we omit
575                       * the +1 here.
576                       */
577                      sizeof(prog->aux->name) > KSYM_NAME_LEN);
578
579         sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
580         sym  = bin2hex(sym, prog->tag, sizeof(prog->tag));
581
582         /* prog->aux->name will be ignored if full btf name is available */
583         if (prog->aux->func_info_cnt) {
584                 type = btf_type_by_id(prog->aux->btf,
585                                       prog->aux->func_info[prog->aux->func_idx].type_id);
586                 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
587                 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
588                 return;
589         }
590
591         if (prog->aux->name[0])
592                 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
593         else
594                 *sym = 0;
595 }
596
597 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
598 {
599         return container_of(n, struct bpf_ksym, tnode)->start;
600 }
601
602 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
603                                           struct latch_tree_node *b)
604 {
605         return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
606 }
607
608 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
609 {
610         unsigned long val = (unsigned long)key;
611         const struct bpf_ksym *ksym;
612
613         ksym = container_of(n, struct bpf_ksym, tnode);
614
615         if (val < ksym->start)
616                 return -1;
617         if (val >= ksym->end)
618                 return  1;
619
620         return 0;
621 }
622
623 static const struct latch_tree_ops bpf_tree_ops = {
624         .less   = bpf_tree_less,
625         .comp   = bpf_tree_comp,
626 };
627
628 static DEFINE_SPINLOCK(bpf_lock);
629 static LIST_HEAD(bpf_kallsyms);
630 static struct latch_tree_root bpf_tree __cacheline_aligned;
631
632 void bpf_ksym_add(struct bpf_ksym *ksym)
633 {
634         spin_lock_bh(&bpf_lock);
635         WARN_ON_ONCE(!list_empty(&ksym->lnode));
636         list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
637         latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
638         spin_unlock_bh(&bpf_lock);
639 }
640
641 static void __bpf_ksym_del(struct bpf_ksym *ksym)
642 {
643         if (list_empty(&ksym->lnode))
644                 return;
645
646         latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
647         list_del_rcu(&ksym->lnode);
648 }
649
650 void bpf_ksym_del(struct bpf_ksym *ksym)
651 {
652         spin_lock_bh(&bpf_lock);
653         __bpf_ksym_del(ksym);
654         spin_unlock_bh(&bpf_lock);
655 }
656
657 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
658 {
659         return fp->jited && !bpf_prog_was_classic(fp);
660 }
661
662 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
663 {
664         if (!bpf_prog_kallsyms_candidate(fp) ||
665             !bpf_capable())
666                 return;
667
668         bpf_prog_ksym_set_addr(fp);
669         bpf_prog_ksym_set_name(fp);
670         fp->aux->ksym.prog = true;
671
672         bpf_ksym_add(&fp->aux->ksym);
673 }
674
675 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
676 {
677         if (!bpf_prog_kallsyms_candidate(fp))
678                 return;
679
680         bpf_ksym_del(&fp->aux->ksym);
681 }
682
683 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
684 {
685         struct latch_tree_node *n;
686
687         n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
688         return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
689 }
690
691 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
692                                  unsigned long *off, char *sym)
693 {
694         struct bpf_ksym *ksym;
695         char *ret = NULL;
696
697         rcu_read_lock();
698         ksym = bpf_ksym_find(addr);
699         if (ksym) {
700                 unsigned long symbol_start = ksym->start;
701                 unsigned long symbol_end = ksym->end;
702
703                 strncpy(sym, ksym->name, KSYM_NAME_LEN);
704
705                 ret = sym;
706                 if (size)
707                         *size = symbol_end - symbol_start;
708                 if (off)
709                         *off  = addr - symbol_start;
710         }
711         rcu_read_unlock();
712
713         return ret;
714 }
715
716 bool is_bpf_text_address(unsigned long addr)
717 {
718         bool ret;
719
720         rcu_read_lock();
721         ret = bpf_ksym_find(addr) != NULL;
722         rcu_read_unlock();
723
724         return ret;
725 }
726
727 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
728 {
729         struct bpf_ksym *ksym = bpf_ksym_find(addr);
730
731         return ksym && ksym->prog ?
732                container_of(ksym, struct bpf_prog_aux, ksym)->prog :
733                NULL;
734 }
735
736 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
737 {
738         const struct exception_table_entry *e = NULL;
739         struct bpf_prog *prog;
740
741         rcu_read_lock();
742         prog = bpf_prog_ksym_find(addr);
743         if (!prog)
744                 goto out;
745         if (!prog->aux->num_exentries)
746                 goto out;
747
748         e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
749 out:
750         rcu_read_unlock();
751         return e;
752 }
753
754 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
755                     char *sym)
756 {
757         struct bpf_ksym *ksym;
758         unsigned int it = 0;
759         int ret = -ERANGE;
760
761         if (!bpf_jit_kallsyms_enabled())
762                 return ret;
763
764         rcu_read_lock();
765         list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
766                 if (it++ != symnum)
767                         continue;
768
769                 strncpy(sym, ksym->name, KSYM_NAME_LEN);
770
771                 *value = ksym->start;
772                 *type  = BPF_SYM_ELF_TYPE;
773
774                 ret = 0;
775                 break;
776         }
777         rcu_read_unlock();
778
779         return ret;
780 }
781
782 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
783                                 struct bpf_jit_poke_descriptor *poke)
784 {
785         struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
786         static const u32 poke_tab_max = 1024;
787         u32 slot = prog->aux->size_poke_tab;
788         u32 size = slot + 1;
789
790         if (size > poke_tab_max)
791                 return -ENOSPC;
792         if (poke->tailcall_target || poke->tailcall_target_stable ||
793             poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
794                 return -EINVAL;
795
796         switch (poke->reason) {
797         case BPF_POKE_REASON_TAIL_CALL:
798                 if (!poke->tail_call.map)
799                         return -EINVAL;
800                 break;
801         default:
802                 return -EINVAL;
803         }
804
805         tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
806         if (!tab)
807                 return -ENOMEM;
808
809         memcpy(&tab[slot], poke, sizeof(*poke));
810         prog->aux->size_poke_tab = size;
811         prog->aux->poke_tab = tab;
812
813         return slot;
814 }
815
816 /*
817  * BPF program pack allocator.
818  *
819  * Most BPF programs are pretty small. Allocating a hole page for each
820  * program is sometime a waste. Many small bpf program also adds pressure
821  * to instruction TLB. To solve this issue, we introduce a BPF program pack
822  * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
823  * to host BPF programs.
824  */
825 #define BPF_PROG_CHUNK_SHIFT    6
826 #define BPF_PROG_CHUNK_SIZE     (1 << BPF_PROG_CHUNK_SHIFT)
827 #define BPF_PROG_CHUNK_MASK     (~(BPF_PROG_CHUNK_SIZE - 1))
828
829 struct bpf_prog_pack {
830         struct list_head list;
831         void *ptr;
832         unsigned long bitmap[];
833 };
834
835 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size)
836 {
837         memset(area, 0, size);
838 }
839
840 #define BPF_PROG_SIZE_TO_NBITS(size)    (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
841
842 static DEFINE_MUTEX(pack_mutex);
843 static LIST_HEAD(pack_list);
844
845 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
846  * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
847  */
848 #ifdef PMD_SIZE
849 #define BPF_PROG_PACK_SIZE (PMD_SIZE * num_possible_nodes())
850 #else
851 #define BPF_PROG_PACK_SIZE PAGE_SIZE
852 #endif
853
854 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
855
856 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
857 {
858         struct bpf_prog_pack *pack;
859
860         pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
861                        GFP_KERNEL);
862         if (!pack)
863                 return NULL;
864         pack->ptr = module_alloc(BPF_PROG_PACK_SIZE);
865         if (!pack->ptr) {
866                 kfree(pack);
867                 return NULL;
868         }
869         bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
870         bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
871         list_add_tail(&pack->list, &pack_list);
872
873         set_vm_flush_reset_perms(pack->ptr);
874         set_memory_rox((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
875         return pack;
876 }
877
878 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
879 {
880         unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
881         struct bpf_prog_pack *pack;
882         unsigned long pos;
883         void *ptr = NULL;
884
885         mutex_lock(&pack_mutex);
886         if (size > BPF_PROG_PACK_SIZE) {
887                 size = round_up(size, PAGE_SIZE);
888                 ptr = module_alloc(size);
889                 if (ptr) {
890                         bpf_fill_ill_insns(ptr, size);
891                         set_vm_flush_reset_perms(ptr);
892                         set_memory_rox((unsigned long)ptr, size / PAGE_SIZE);
893                 }
894                 goto out;
895         }
896         list_for_each_entry(pack, &pack_list, list) {
897                 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
898                                                  nbits, 0);
899                 if (pos < BPF_PROG_CHUNK_COUNT)
900                         goto found_free_area;
901         }
902
903         pack = alloc_new_pack(bpf_fill_ill_insns);
904         if (!pack)
905                 goto out;
906
907         pos = 0;
908
909 found_free_area:
910         bitmap_set(pack->bitmap, pos, nbits);
911         ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
912
913 out:
914         mutex_unlock(&pack_mutex);
915         return ptr;
916 }
917
918 void bpf_prog_pack_free(struct bpf_binary_header *hdr)
919 {
920         struct bpf_prog_pack *pack = NULL, *tmp;
921         unsigned int nbits;
922         unsigned long pos;
923
924         mutex_lock(&pack_mutex);
925         if (hdr->size > BPF_PROG_PACK_SIZE) {
926                 module_memfree(hdr);
927                 goto out;
928         }
929
930         list_for_each_entry(tmp, &pack_list, list) {
931                 if ((void *)hdr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > (void *)hdr) {
932                         pack = tmp;
933                         break;
934                 }
935         }
936
937         if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
938                 goto out;
939
940         nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size);
941         pos = ((unsigned long)hdr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
942
943         WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size),
944                   "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
945
946         bitmap_clear(pack->bitmap, pos, nbits);
947         if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
948                                        BPF_PROG_CHUNK_COUNT, 0) == 0) {
949                 list_del(&pack->list);
950                 module_memfree(pack->ptr);
951                 kfree(pack);
952         }
953 out:
954         mutex_unlock(&pack_mutex);
955 }
956
957 static atomic_long_t bpf_jit_current;
958
959 /* Can be overridden by an arch's JIT compiler if it has a custom,
960  * dedicated BPF backend memory area, or if neither of the two
961  * below apply.
962  */
963 u64 __weak bpf_jit_alloc_exec_limit(void)
964 {
965 #if defined(MODULES_VADDR)
966         return MODULES_END - MODULES_VADDR;
967 #else
968         return VMALLOC_END - VMALLOC_START;
969 #endif
970 }
971
972 static int __init bpf_jit_charge_init(void)
973 {
974         /* Only used as heuristic here to derive limit. */
975         bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
976         bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1,
977                                             PAGE_SIZE), LONG_MAX);
978         return 0;
979 }
980 pure_initcall(bpf_jit_charge_init);
981
982 int bpf_jit_charge_modmem(u32 size)
983 {
984         if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
985                 if (!bpf_capable()) {
986                         atomic_long_sub(size, &bpf_jit_current);
987                         return -EPERM;
988                 }
989         }
990
991         return 0;
992 }
993
994 void bpf_jit_uncharge_modmem(u32 size)
995 {
996         atomic_long_sub(size, &bpf_jit_current);
997 }
998
999 void *__weak bpf_jit_alloc_exec(unsigned long size)
1000 {
1001         return module_alloc(size);
1002 }
1003
1004 void __weak bpf_jit_free_exec(void *addr)
1005 {
1006         module_memfree(addr);
1007 }
1008
1009 struct bpf_binary_header *
1010 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1011                      unsigned int alignment,
1012                      bpf_jit_fill_hole_t bpf_fill_ill_insns)
1013 {
1014         struct bpf_binary_header *hdr;
1015         u32 size, hole, start;
1016
1017         WARN_ON_ONCE(!is_power_of_2(alignment) ||
1018                      alignment > BPF_IMAGE_ALIGNMENT);
1019
1020         /* Most of BPF filters are really small, but if some of them
1021          * fill a page, allow at least 128 extra bytes to insert a
1022          * random section of illegal instructions.
1023          */
1024         size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1025
1026         if (bpf_jit_charge_modmem(size))
1027                 return NULL;
1028         hdr = bpf_jit_alloc_exec(size);
1029         if (!hdr) {
1030                 bpf_jit_uncharge_modmem(size);
1031                 return NULL;
1032         }
1033
1034         /* Fill space with illegal/arch-dep instructions. */
1035         bpf_fill_ill_insns(hdr, size);
1036
1037         hdr->size = size;
1038         hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1039                      PAGE_SIZE - sizeof(*hdr));
1040         start = get_random_u32_below(hole) & ~(alignment - 1);
1041
1042         /* Leave a random number of instructions before BPF code. */
1043         *image_ptr = &hdr->image[start];
1044
1045         return hdr;
1046 }
1047
1048 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1049 {
1050         u32 size = hdr->size;
1051
1052         bpf_jit_free_exec(hdr);
1053         bpf_jit_uncharge_modmem(size);
1054 }
1055
1056 /* Allocate jit binary from bpf_prog_pack allocator.
1057  * Since the allocated memory is RO+X, the JIT engine cannot write directly
1058  * to the memory. To solve this problem, a RW buffer is also allocated at
1059  * as the same time. The JIT engine should calculate offsets based on the
1060  * RO memory address, but write JITed program to the RW buffer. Once the
1061  * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1062  * the JITed program to the RO memory.
1063  */
1064 struct bpf_binary_header *
1065 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1066                           unsigned int alignment,
1067                           struct bpf_binary_header **rw_header,
1068                           u8 **rw_image,
1069                           bpf_jit_fill_hole_t bpf_fill_ill_insns)
1070 {
1071         struct bpf_binary_header *ro_header;
1072         u32 size, hole, start;
1073
1074         WARN_ON_ONCE(!is_power_of_2(alignment) ||
1075                      alignment > BPF_IMAGE_ALIGNMENT);
1076
1077         /* add 16 bytes for a random section of illegal instructions */
1078         size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1079
1080         if (bpf_jit_charge_modmem(size))
1081                 return NULL;
1082         ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1083         if (!ro_header) {
1084                 bpf_jit_uncharge_modmem(size);
1085                 return NULL;
1086         }
1087
1088         *rw_header = kvmalloc(size, GFP_KERNEL);
1089         if (!*rw_header) {
1090                 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size));
1091                 bpf_prog_pack_free(ro_header);
1092                 bpf_jit_uncharge_modmem(size);
1093                 return NULL;
1094         }
1095
1096         /* Fill space with illegal/arch-dep instructions. */
1097         bpf_fill_ill_insns(*rw_header, size);
1098         (*rw_header)->size = size;
1099
1100         hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1101                      BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1102         start = get_random_u32_below(hole) & ~(alignment - 1);
1103
1104         *image_ptr = &ro_header->image[start];
1105         *rw_image = &(*rw_header)->image[start];
1106
1107         return ro_header;
1108 }
1109
1110 /* Copy JITed text from rw_header to its final location, the ro_header. */
1111 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog,
1112                                  struct bpf_binary_header *ro_header,
1113                                  struct bpf_binary_header *rw_header)
1114 {
1115         void *ptr;
1116
1117         ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1118
1119         kvfree(rw_header);
1120
1121         if (IS_ERR(ptr)) {
1122                 bpf_prog_pack_free(ro_header);
1123                 return PTR_ERR(ptr);
1124         }
1125         return 0;
1126 }
1127
1128 /* bpf_jit_binary_pack_free is called in two different scenarios:
1129  *   1) when the program is freed after;
1130  *   2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1131  * For case 2), we need to free both the RO memory and the RW buffer.
1132  *
1133  * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1134  * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1135  * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1136  * bpf_arch_text_copy (when jit fails).
1137  */
1138 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1139                               struct bpf_binary_header *rw_header)
1140 {
1141         u32 size = ro_header->size;
1142
1143         bpf_prog_pack_free(ro_header);
1144         kvfree(rw_header);
1145         bpf_jit_uncharge_modmem(size);
1146 }
1147
1148 struct bpf_binary_header *
1149 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1150 {
1151         unsigned long real_start = (unsigned long)fp->bpf_func;
1152         unsigned long addr;
1153
1154         addr = real_start & BPF_PROG_CHUNK_MASK;
1155         return (void *)addr;
1156 }
1157
1158 static inline struct bpf_binary_header *
1159 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1160 {
1161         unsigned long real_start = (unsigned long)fp->bpf_func;
1162         unsigned long addr;
1163
1164         addr = real_start & PAGE_MASK;
1165         return (void *)addr;
1166 }
1167
1168 /* This symbol is only overridden by archs that have different
1169  * requirements than the usual eBPF JITs, f.e. when they only
1170  * implement cBPF JIT, do not set images read-only, etc.
1171  */
1172 void __weak bpf_jit_free(struct bpf_prog *fp)
1173 {
1174         if (fp->jited) {
1175                 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1176
1177                 bpf_jit_binary_free(hdr);
1178                 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1179         }
1180
1181         bpf_prog_unlock_free(fp);
1182 }
1183
1184 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1185                           const struct bpf_insn *insn, bool extra_pass,
1186                           u64 *func_addr, bool *func_addr_fixed)
1187 {
1188         s16 off = insn->off;
1189         s32 imm = insn->imm;
1190         u8 *addr;
1191         int err;
1192
1193         *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1194         if (!*func_addr_fixed) {
1195                 /* Place-holder address till the last pass has collected
1196                  * all addresses for JITed subprograms in which case we
1197                  * can pick them up from prog->aux.
1198                  */
1199                 if (!extra_pass)
1200                         addr = NULL;
1201                 else if (prog->aux->func &&
1202                          off >= 0 && off < prog->aux->func_cnt)
1203                         addr = (u8 *)prog->aux->func[off]->bpf_func;
1204                 else
1205                         return -EINVAL;
1206         } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
1207                    bpf_jit_supports_far_kfunc_call()) {
1208                 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr);
1209                 if (err)
1210                         return err;
1211         } else {
1212                 /* Address of a BPF helper call. Since part of the core
1213                  * kernel, it's always at a fixed location. __bpf_call_base
1214                  * and the helper with imm relative to it are both in core
1215                  * kernel.
1216                  */
1217                 addr = (u8 *)__bpf_call_base + imm;
1218         }
1219
1220         *func_addr = (unsigned long)addr;
1221         return 0;
1222 }
1223
1224 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1225                               const struct bpf_insn *aux,
1226                               struct bpf_insn *to_buff,
1227                               bool emit_zext)
1228 {
1229         struct bpf_insn *to = to_buff;
1230         u32 imm_rnd = get_random_u32();
1231         s16 off;
1232
1233         BUILD_BUG_ON(BPF_REG_AX  + 1 != MAX_BPF_JIT_REG);
1234         BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1235
1236         /* Constraints on AX register:
1237          *
1238          * AX register is inaccessible from user space. It is mapped in
1239          * all JITs, and used here for constant blinding rewrites. It is
1240          * typically "stateless" meaning its contents are only valid within
1241          * the executed instruction, but not across several instructions.
1242          * There are a few exceptions however which are further detailed
1243          * below.
1244          *
1245          * Constant blinding is only used by JITs, not in the interpreter.
1246          * The interpreter uses AX in some occasions as a local temporary
1247          * register e.g. in DIV or MOD instructions.
1248          *
1249          * In restricted circumstances, the verifier can also use the AX
1250          * register for rewrites as long as they do not interfere with
1251          * the above cases!
1252          */
1253         if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1254                 goto out;
1255
1256         if (from->imm == 0 &&
1257             (from->code == (BPF_ALU   | BPF_MOV | BPF_K) ||
1258              from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1259                 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1260                 goto out;
1261         }
1262
1263         switch (from->code) {
1264         case BPF_ALU | BPF_ADD | BPF_K:
1265         case BPF_ALU | BPF_SUB | BPF_K:
1266         case BPF_ALU | BPF_AND | BPF_K:
1267         case BPF_ALU | BPF_OR  | BPF_K:
1268         case BPF_ALU | BPF_XOR | BPF_K:
1269         case BPF_ALU | BPF_MUL | BPF_K:
1270         case BPF_ALU | BPF_MOV | BPF_K:
1271         case BPF_ALU | BPF_DIV | BPF_K:
1272         case BPF_ALU | BPF_MOD | BPF_K:
1273                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1274                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1275                 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1276                 break;
1277
1278         case BPF_ALU64 | BPF_ADD | BPF_K:
1279         case BPF_ALU64 | BPF_SUB | BPF_K:
1280         case BPF_ALU64 | BPF_AND | BPF_K:
1281         case BPF_ALU64 | BPF_OR  | BPF_K:
1282         case BPF_ALU64 | BPF_XOR | BPF_K:
1283         case BPF_ALU64 | BPF_MUL | BPF_K:
1284         case BPF_ALU64 | BPF_MOV | BPF_K:
1285         case BPF_ALU64 | BPF_DIV | BPF_K:
1286         case BPF_ALU64 | BPF_MOD | BPF_K:
1287                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1288                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1289                 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1290                 break;
1291
1292         case BPF_JMP | BPF_JEQ  | BPF_K:
1293         case BPF_JMP | BPF_JNE  | BPF_K:
1294         case BPF_JMP | BPF_JGT  | BPF_K:
1295         case BPF_JMP | BPF_JLT  | BPF_K:
1296         case BPF_JMP | BPF_JGE  | BPF_K:
1297         case BPF_JMP | BPF_JLE  | BPF_K:
1298         case BPF_JMP | BPF_JSGT | BPF_K:
1299         case BPF_JMP | BPF_JSLT | BPF_K:
1300         case BPF_JMP | BPF_JSGE | BPF_K:
1301         case BPF_JMP | BPF_JSLE | BPF_K:
1302         case BPF_JMP | BPF_JSET | BPF_K:
1303                 /* Accommodate for extra offset in case of a backjump. */
1304                 off = from->off;
1305                 if (off < 0)
1306                         off -= 2;
1307                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1308                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1309                 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1310                 break;
1311
1312         case BPF_JMP32 | BPF_JEQ  | BPF_K:
1313         case BPF_JMP32 | BPF_JNE  | BPF_K:
1314         case BPF_JMP32 | BPF_JGT  | BPF_K:
1315         case BPF_JMP32 | BPF_JLT  | BPF_K:
1316         case BPF_JMP32 | BPF_JGE  | BPF_K:
1317         case BPF_JMP32 | BPF_JLE  | BPF_K:
1318         case BPF_JMP32 | BPF_JSGT | BPF_K:
1319         case BPF_JMP32 | BPF_JSLT | BPF_K:
1320         case BPF_JMP32 | BPF_JSGE | BPF_K:
1321         case BPF_JMP32 | BPF_JSLE | BPF_K:
1322         case BPF_JMP32 | BPF_JSET | BPF_K:
1323                 /* Accommodate for extra offset in case of a backjump. */
1324                 off = from->off;
1325                 if (off < 0)
1326                         off -= 2;
1327                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1328                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1329                 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1330                                       off);
1331                 break;
1332
1333         case BPF_LD | BPF_IMM | BPF_DW:
1334                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1335                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1336                 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1337                 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1338                 break;
1339         case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1340                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1341                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1342                 if (emit_zext)
1343                         *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1344                 *to++ = BPF_ALU64_REG(BPF_OR,  aux[0].dst_reg, BPF_REG_AX);
1345                 break;
1346
1347         case BPF_ST | BPF_MEM | BPF_DW:
1348         case BPF_ST | BPF_MEM | BPF_W:
1349         case BPF_ST | BPF_MEM | BPF_H:
1350         case BPF_ST | BPF_MEM | BPF_B:
1351                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1352                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1353                 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1354                 break;
1355         }
1356 out:
1357         return to - to_buff;
1358 }
1359
1360 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1361                                               gfp_t gfp_extra_flags)
1362 {
1363         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1364         struct bpf_prog *fp;
1365
1366         fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1367         if (fp != NULL) {
1368                 /* aux->prog still points to the fp_other one, so
1369                  * when promoting the clone to the real program,
1370                  * this still needs to be adapted.
1371                  */
1372                 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1373         }
1374
1375         return fp;
1376 }
1377
1378 static void bpf_prog_clone_free(struct bpf_prog *fp)
1379 {
1380         /* aux was stolen by the other clone, so we cannot free
1381          * it from this path! It will be freed eventually by the
1382          * other program on release.
1383          *
1384          * At this point, we don't need a deferred release since
1385          * clone is guaranteed to not be locked.
1386          */
1387         fp->aux = NULL;
1388         fp->stats = NULL;
1389         fp->active = NULL;
1390         __bpf_prog_free(fp);
1391 }
1392
1393 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1394 {
1395         /* We have to repoint aux->prog to self, as we don't
1396          * know whether fp here is the clone or the original.
1397          */
1398         fp->aux->prog = fp;
1399         bpf_prog_clone_free(fp_other);
1400 }
1401
1402 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1403 {
1404         struct bpf_insn insn_buff[16], aux[2];
1405         struct bpf_prog *clone, *tmp;
1406         int insn_delta, insn_cnt;
1407         struct bpf_insn *insn;
1408         int i, rewritten;
1409
1410         if (!prog->blinding_requested || prog->blinded)
1411                 return prog;
1412
1413         clone = bpf_prog_clone_create(prog, GFP_USER);
1414         if (!clone)
1415                 return ERR_PTR(-ENOMEM);
1416
1417         insn_cnt = clone->len;
1418         insn = clone->insnsi;
1419
1420         for (i = 0; i < insn_cnt; i++, insn++) {
1421                 if (bpf_pseudo_func(insn)) {
1422                         /* ld_imm64 with an address of bpf subprog is not
1423                          * a user controlled constant. Don't randomize it,
1424                          * since it will conflict with jit_subprogs() logic.
1425                          */
1426                         insn++;
1427                         i++;
1428                         continue;
1429                 }
1430
1431                 /* We temporarily need to hold the original ld64 insn
1432                  * so that we can still access the first part in the
1433                  * second blinding run.
1434                  */
1435                 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1436                     insn[1].code == 0)
1437                         memcpy(aux, insn, sizeof(aux));
1438
1439                 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1440                                                 clone->aux->verifier_zext);
1441                 if (!rewritten)
1442                         continue;
1443
1444                 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1445                 if (IS_ERR(tmp)) {
1446                         /* Patching may have repointed aux->prog during
1447                          * realloc from the original one, so we need to
1448                          * fix it up here on error.
1449                          */
1450                         bpf_jit_prog_release_other(prog, clone);
1451                         return tmp;
1452                 }
1453
1454                 clone = tmp;
1455                 insn_delta = rewritten - 1;
1456
1457                 /* Walk new program and skip insns we just inserted. */
1458                 insn = clone->insnsi + i + insn_delta;
1459                 insn_cnt += insn_delta;
1460                 i        += insn_delta;
1461         }
1462
1463         clone->blinded = 1;
1464         return clone;
1465 }
1466 #endif /* CONFIG_BPF_JIT */
1467
1468 /* Base function for offset calculation. Needs to go into .text section,
1469  * therefore keeping it non-static as well; will also be used by JITs
1470  * anyway later on, so do not let the compiler omit it. This also needs
1471  * to go into kallsyms for correlation from e.g. bpftool, so naming
1472  * must not change.
1473  */
1474 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1475 {
1476         return 0;
1477 }
1478 EXPORT_SYMBOL_GPL(__bpf_call_base);
1479
1480 /* All UAPI available opcodes. */
1481 #define BPF_INSN_MAP(INSN_2, INSN_3)            \
1482         /* 32 bit ALU operations. */            \
1483         /*   Register based. */                 \
1484         INSN_3(ALU, ADD,  X),                   \
1485         INSN_3(ALU, SUB,  X),                   \
1486         INSN_3(ALU, AND,  X),                   \
1487         INSN_3(ALU, OR,   X),                   \
1488         INSN_3(ALU, LSH,  X),                   \
1489         INSN_3(ALU, RSH,  X),                   \
1490         INSN_3(ALU, XOR,  X),                   \
1491         INSN_3(ALU, MUL,  X),                   \
1492         INSN_3(ALU, MOV,  X),                   \
1493         INSN_3(ALU, ARSH, X),                   \
1494         INSN_3(ALU, DIV,  X),                   \
1495         INSN_3(ALU, MOD,  X),                   \
1496         INSN_2(ALU, NEG),                       \
1497         INSN_3(ALU, END, TO_BE),                \
1498         INSN_3(ALU, END, TO_LE),                \
1499         /*   Immediate based. */                \
1500         INSN_3(ALU, ADD,  K),                   \
1501         INSN_3(ALU, SUB,  K),                   \
1502         INSN_3(ALU, AND,  K),                   \
1503         INSN_3(ALU, OR,   K),                   \
1504         INSN_3(ALU, LSH,  K),                   \
1505         INSN_3(ALU, RSH,  K),                   \
1506         INSN_3(ALU, XOR,  K),                   \
1507         INSN_3(ALU, MUL,  K),                   \
1508         INSN_3(ALU, MOV,  K),                   \
1509         INSN_3(ALU, ARSH, K),                   \
1510         INSN_3(ALU, DIV,  K),                   \
1511         INSN_3(ALU, MOD,  K),                   \
1512         /* 64 bit ALU operations. */            \
1513         /*   Register based. */                 \
1514         INSN_3(ALU64, ADD,  X),                 \
1515         INSN_3(ALU64, SUB,  X),                 \
1516         INSN_3(ALU64, AND,  X),                 \
1517         INSN_3(ALU64, OR,   X),                 \
1518         INSN_3(ALU64, LSH,  X),                 \
1519         INSN_3(ALU64, RSH,  X),                 \
1520         INSN_3(ALU64, XOR,  X),                 \
1521         INSN_3(ALU64, MUL,  X),                 \
1522         INSN_3(ALU64, MOV,  X),                 \
1523         INSN_3(ALU64, ARSH, X),                 \
1524         INSN_3(ALU64, DIV,  X),                 \
1525         INSN_3(ALU64, MOD,  X),                 \
1526         INSN_2(ALU64, NEG),                     \
1527         /*   Immediate based. */                \
1528         INSN_3(ALU64, ADD,  K),                 \
1529         INSN_3(ALU64, SUB,  K),                 \
1530         INSN_3(ALU64, AND,  K),                 \
1531         INSN_3(ALU64, OR,   K),                 \
1532         INSN_3(ALU64, LSH,  K),                 \
1533         INSN_3(ALU64, RSH,  K),                 \
1534         INSN_3(ALU64, XOR,  K),                 \
1535         INSN_3(ALU64, MUL,  K),                 \
1536         INSN_3(ALU64, MOV,  K),                 \
1537         INSN_3(ALU64, ARSH, K),                 \
1538         INSN_3(ALU64, DIV,  K),                 \
1539         INSN_3(ALU64, MOD,  K),                 \
1540         /* Call instruction. */                 \
1541         INSN_2(JMP, CALL),                      \
1542         /* Exit instruction. */                 \
1543         INSN_2(JMP, EXIT),                      \
1544         /* 32-bit Jump instructions. */         \
1545         /*   Register based. */                 \
1546         INSN_3(JMP32, JEQ,  X),                 \
1547         INSN_3(JMP32, JNE,  X),                 \
1548         INSN_3(JMP32, JGT,  X),                 \
1549         INSN_3(JMP32, JLT,  X),                 \
1550         INSN_3(JMP32, JGE,  X),                 \
1551         INSN_3(JMP32, JLE,  X),                 \
1552         INSN_3(JMP32, JSGT, X),                 \
1553         INSN_3(JMP32, JSLT, X),                 \
1554         INSN_3(JMP32, JSGE, X),                 \
1555         INSN_3(JMP32, JSLE, X),                 \
1556         INSN_3(JMP32, JSET, X),                 \
1557         /*   Immediate based. */                \
1558         INSN_3(JMP32, JEQ,  K),                 \
1559         INSN_3(JMP32, JNE,  K),                 \
1560         INSN_3(JMP32, JGT,  K),                 \
1561         INSN_3(JMP32, JLT,  K),                 \
1562         INSN_3(JMP32, JGE,  K),                 \
1563         INSN_3(JMP32, JLE,  K),                 \
1564         INSN_3(JMP32, JSGT, K),                 \
1565         INSN_3(JMP32, JSLT, K),                 \
1566         INSN_3(JMP32, JSGE, K),                 \
1567         INSN_3(JMP32, JSLE, K),                 \
1568         INSN_3(JMP32, JSET, K),                 \
1569         /* Jump instructions. */                \
1570         /*   Register based. */                 \
1571         INSN_3(JMP, JEQ,  X),                   \
1572         INSN_3(JMP, JNE,  X),                   \
1573         INSN_3(JMP, JGT,  X),                   \
1574         INSN_3(JMP, JLT,  X),                   \
1575         INSN_3(JMP, JGE,  X),                   \
1576         INSN_3(JMP, JLE,  X),                   \
1577         INSN_3(JMP, JSGT, X),                   \
1578         INSN_3(JMP, JSLT, X),                   \
1579         INSN_3(JMP, JSGE, X),                   \
1580         INSN_3(JMP, JSLE, X),                   \
1581         INSN_3(JMP, JSET, X),                   \
1582         /*   Immediate based. */                \
1583         INSN_3(JMP, JEQ,  K),                   \
1584         INSN_3(JMP, JNE,  K),                   \
1585         INSN_3(JMP, JGT,  K),                   \
1586         INSN_3(JMP, JLT,  K),                   \
1587         INSN_3(JMP, JGE,  K),                   \
1588         INSN_3(JMP, JLE,  K),                   \
1589         INSN_3(JMP, JSGT, K),                   \
1590         INSN_3(JMP, JSLT, K),                   \
1591         INSN_3(JMP, JSGE, K),                   \
1592         INSN_3(JMP, JSLE, K),                   \
1593         INSN_3(JMP, JSET, K),                   \
1594         INSN_2(JMP, JA),                        \
1595         /* Store instructions. */               \
1596         /*   Register based. */                 \
1597         INSN_3(STX, MEM,  B),                   \
1598         INSN_3(STX, MEM,  H),                   \
1599         INSN_3(STX, MEM,  W),                   \
1600         INSN_3(STX, MEM,  DW),                  \
1601         INSN_3(STX, ATOMIC, W),                 \
1602         INSN_3(STX, ATOMIC, DW),                \
1603         /*   Immediate based. */                \
1604         INSN_3(ST, MEM, B),                     \
1605         INSN_3(ST, MEM, H),                     \
1606         INSN_3(ST, MEM, W),                     \
1607         INSN_3(ST, MEM, DW),                    \
1608         /* Load instructions. */                \
1609         /*   Register based. */                 \
1610         INSN_3(LDX, MEM, B),                    \
1611         INSN_3(LDX, MEM, H),                    \
1612         INSN_3(LDX, MEM, W),                    \
1613         INSN_3(LDX, MEM, DW),                   \
1614         INSN_3(LDX, MEMSX, B),                  \
1615         INSN_3(LDX, MEMSX, H),                  \
1616         INSN_3(LDX, MEMSX, W),                  \
1617         /*   Immediate based. */                \
1618         INSN_3(LD, IMM, DW)
1619
1620 bool bpf_opcode_in_insntable(u8 code)
1621 {
1622 #define BPF_INSN_2_TBL(x, y)    [BPF_##x | BPF_##y] = true
1623 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1624         static const bool public_insntable[256] = {
1625                 [0 ... 255] = false,
1626                 /* Now overwrite non-defaults ... */
1627                 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1628                 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1629                 [BPF_LD | BPF_ABS | BPF_B] = true,
1630                 [BPF_LD | BPF_ABS | BPF_H] = true,
1631                 [BPF_LD | BPF_ABS | BPF_W] = true,
1632                 [BPF_LD | BPF_IND | BPF_B] = true,
1633                 [BPF_LD | BPF_IND | BPF_H] = true,
1634                 [BPF_LD | BPF_IND | BPF_W] = true,
1635         };
1636 #undef BPF_INSN_3_TBL
1637 #undef BPF_INSN_2_TBL
1638         return public_insntable[code];
1639 }
1640
1641 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1642 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1643 {
1644         memset(dst, 0, size);
1645         return -EFAULT;
1646 }
1647
1648 /**
1649  *      ___bpf_prog_run - run eBPF program on a given context
1650  *      @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1651  *      @insn: is the array of eBPF instructions
1652  *
1653  * Decode and execute eBPF instructions.
1654  *
1655  * Return: whatever value is in %BPF_R0 at program exit
1656  */
1657 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1658 {
1659 #define BPF_INSN_2_LBL(x, y)    [BPF_##x | BPF_##y] = &&x##_##y
1660 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1661         static const void * const jumptable[256] __annotate_jump_table = {
1662                 [0 ... 255] = &&default_label,
1663                 /* Now overwrite non-defaults ... */
1664                 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1665                 /* Non-UAPI available opcodes. */
1666                 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1667                 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1668                 [BPF_ST  | BPF_NOSPEC] = &&ST_NOSPEC,
1669                 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1670                 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1671                 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1672                 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1673                 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B,
1674                 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H,
1675                 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W,
1676         };
1677 #undef BPF_INSN_3_LBL
1678 #undef BPF_INSN_2_LBL
1679         u32 tail_call_cnt = 0;
1680
1681 #define CONT     ({ insn++; goto select_insn; })
1682 #define CONT_JMP ({ insn++; goto select_insn; })
1683
1684 select_insn:
1685         goto *jumptable[insn->code];
1686
1687         /* Explicitly mask the register-based shift amounts with 63 or 31
1688          * to avoid undefined behavior. Normally this won't affect the
1689          * generated code, for example, in case of native 64 bit archs such
1690          * as x86-64 or arm64, the compiler is optimizing the AND away for
1691          * the interpreter. In case of JITs, each of the JIT backends compiles
1692          * the BPF shift operations to machine instructions which produce
1693          * implementation-defined results in such a case; the resulting
1694          * contents of the register may be arbitrary, but program behaviour
1695          * as a whole remains defined. In other words, in case of JIT backends,
1696          * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1697          */
1698         /* ALU (shifts) */
1699 #define SHT(OPCODE, OP)                                 \
1700         ALU64_##OPCODE##_X:                             \
1701                 DST = DST OP (SRC & 63);                \
1702                 CONT;                                   \
1703         ALU_##OPCODE##_X:                               \
1704                 DST = (u32) DST OP ((u32) SRC & 31);    \
1705                 CONT;                                   \
1706         ALU64_##OPCODE##_K:                             \
1707                 DST = DST OP IMM;                       \
1708                 CONT;                                   \
1709         ALU_##OPCODE##_K:                               \
1710                 DST = (u32) DST OP (u32) IMM;           \
1711                 CONT;
1712         /* ALU (rest) */
1713 #define ALU(OPCODE, OP)                                 \
1714         ALU64_##OPCODE##_X:                             \
1715                 DST = DST OP SRC;                       \
1716                 CONT;                                   \
1717         ALU_##OPCODE##_X:                               \
1718                 DST = (u32) DST OP (u32) SRC;           \
1719                 CONT;                                   \
1720         ALU64_##OPCODE##_K:                             \
1721                 DST = DST OP IMM;                       \
1722                 CONT;                                   \
1723         ALU_##OPCODE##_K:                               \
1724                 DST = (u32) DST OP (u32) IMM;           \
1725                 CONT;
1726         ALU(ADD,  +)
1727         ALU(SUB,  -)
1728         ALU(AND,  &)
1729         ALU(OR,   |)
1730         ALU(XOR,  ^)
1731         ALU(MUL,  *)
1732         SHT(LSH, <<)
1733         SHT(RSH, >>)
1734 #undef SHT
1735 #undef ALU
1736         ALU_NEG:
1737                 DST = (u32) -DST;
1738                 CONT;
1739         ALU64_NEG:
1740                 DST = -DST;
1741                 CONT;
1742         ALU_MOV_X:
1743                 switch (OFF) {
1744                 case 0:
1745                         DST = (u32) SRC;
1746                         break;
1747                 case 8:
1748                         DST = (u32)(s8) SRC;
1749                         break;
1750                 case 16:
1751                         DST = (u32)(s16) SRC;
1752                         break;
1753                 }
1754                 CONT;
1755         ALU_MOV_K:
1756                 DST = (u32) IMM;
1757                 CONT;
1758         ALU64_MOV_X:
1759                 switch (OFF) {
1760                 case 0:
1761                         DST = SRC;
1762                         break;
1763                 case 8:
1764                         DST = (s8) SRC;
1765                         break;
1766                 case 16:
1767                         DST = (s16) SRC;
1768                         break;
1769                 case 32:
1770                         DST = (s32) SRC;
1771                         break;
1772                 }
1773                 CONT;
1774         ALU64_MOV_K:
1775                 DST = IMM;
1776                 CONT;
1777         LD_IMM_DW:
1778                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1779                 insn++;
1780                 CONT;
1781         ALU_ARSH_X:
1782                 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1783                 CONT;
1784         ALU_ARSH_K:
1785                 DST = (u64) (u32) (((s32) DST) >> IMM);
1786                 CONT;
1787         ALU64_ARSH_X:
1788                 (*(s64 *) &DST) >>= (SRC & 63);
1789                 CONT;
1790         ALU64_ARSH_K:
1791                 (*(s64 *) &DST) >>= IMM;
1792                 CONT;
1793         ALU64_MOD_X:
1794                 div64_u64_rem(DST, SRC, &AX);
1795                 DST = AX;
1796                 CONT;
1797         ALU_MOD_X:
1798                 AX = (u32) DST;
1799                 DST = do_div(AX, (u32) SRC);
1800                 CONT;
1801         ALU64_MOD_K:
1802                 div64_u64_rem(DST, IMM, &AX);
1803                 DST = AX;
1804                 CONT;
1805         ALU_MOD_K:
1806                 AX = (u32) DST;
1807                 DST = do_div(AX, (u32) IMM);
1808                 CONT;
1809         ALU64_DIV_X:
1810                 DST = div64_u64(DST, SRC);
1811                 CONT;
1812         ALU_DIV_X:
1813                 AX = (u32) DST;
1814                 do_div(AX, (u32) SRC);
1815                 DST = (u32) AX;
1816                 CONT;
1817         ALU64_DIV_K:
1818                 DST = div64_u64(DST, IMM);
1819                 CONT;
1820         ALU_DIV_K:
1821                 AX = (u32) DST;
1822                 do_div(AX, (u32) IMM);
1823                 DST = (u32) AX;
1824                 CONT;
1825         ALU_END_TO_BE:
1826                 switch (IMM) {
1827                 case 16:
1828                         DST = (__force u16) cpu_to_be16(DST);
1829                         break;
1830                 case 32:
1831                         DST = (__force u32) cpu_to_be32(DST);
1832                         break;
1833                 case 64:
1834                         DST = (__force u64) cpu_to_be64(DST);
1835                         break;
1836                 }
1837                 CONT;
1838         ALU_END_TO_LE:
1839                 switch (IMM) {
1840                 case 16:
1841                         DST = (__force u16) cpu_to_le16(DST);
1842                         break;
1843                 case 32:
1844                         DST = (__force u32) cpu_to_le32(DST);
1845                         break;
1846                 case 64:
1847                         DST = (__force u64) cpu_to_le64(DST);
1848                         break;
1849                 }
1850                 CONT;
1851
1852         /* CALL */
1853         JMP_CALL:
1854                 /* Function call scratches BPF_R1-BPF_R5 registers,
1855                  * preserves BPF_R6-BPF_R9, and stores return value
1856                  * into BPF_R0.
1857                  */
1858                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1859                                                        BPF_R4, BPF_R5);
1860                 CONT;
1861
1862         JMP_CALL_ARGS:
1863                 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1864                                                             BPF_R3, BPF_R4,
1865                                                             BPF_R5,
1866                                                             insn + insn->off + 1);
1867                 CONT;
1868
1869         JMP_TAIL_CALL: {
1870                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1871                 struct bpf_array *array = container_of(map, struct bpf_array, map);
1872                 struct bpf_prog *prog;
1873                 u32 index = BPF_R3;
1874
1875                 if (unlikely(index >= array->map.max_entries))
1876                         goto out;
1877
1878                 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
1879                         goto out;
1880
1881                 tail_call_cnt++;
1882
1883                 prog = READ_ONCE(array->ptrs[index]);
1884                 if (!prog)
1885                         goto out;
1886
1887                 /* ARG1 at this point is guaranteed to point to CTX from
1888                  * the verifier side due to the fact that the tail call is
1889                  * handled like a helper, that is, bpf_tail_call_proto,
1890                  * where arg1_type is ARG_PTR_TO_CTX.
1891                  */
1892                 insn = prog->insnsi;
1893                 goto select_insn;
1894 out:
1895                 CONT;
1896         }
1897         JMP_JA:
1898                 insn += insn->off;
1899                 CONT;
1900         JMP_EXIT:
1901                 return BPF_R0;
1902         /* JMP */
1903 #define COND_JMP(SIGN, OPCODE, CMP_OP)                          \
1904         JMP_##OPCODE##_X:                                       \
1905                 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) {     \
1906                         insn += insn->off;                      \
1907                         CONT_JMP;                               \
1908                 }                                               \
1909                 CONT;                                           \
1910         JMP32_##OPCODE##_X:                                     \
1911                 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) {     \
1912                         insn += insn->off;                      \
1913                         CONT_JMP;                               \
1914                 }                                               \
1915                 CONT;                                           \
1916         JMP_##OPCODE##_K:                                       \
1917                 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) {     \
1918                         insn += insn->off;                      \
1919                         CONT_JMP;                               \
1920                 }                                               \
1921                 CONT;                                           \
1922         JMP32_##OPCODE##_K:                                     \
1923                 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) {     \
1924                         insn += insn->off;                      \
1925                         CONT_JMP;                               \
1926                 }                                               \
1927                 CONT;
1928         COND_JMP(u, JEQ, ==)
1929         COND_JMP(u, JNE, !=)
1930         COND_JMP(u, JGT, >)
1931         COND_JMP(u, JLT, <)
1932         COND_JMP(u, JGE, >=)
1933         COND_JMP(u, JLE, <=)
1934         COND_JMP(u, JSET, &)
1935         COND_JMP(s, JSGT, >)
1936         COND_JMP(s, JSLT, <)
1937         COND_JMP(s, JSGE, >=)
1938         COND_JMP(s, JSLE, <=)
1939 #undef COND_JMP
1940         /* ST, STX and LDX*/
1941         ST_NOSPEC:
1942                 /* Speculation barrier for mitigating Speculative Store Bypass.
1943                  * In case of arm64, we rely on the firmware mitigation as
1944                  * controlled via the ssbd kernel parameter. Whenever the
1945                  * mitigation is enabled, it works for all of the kernel code
1946                  * with no need to provide any additional instructions here.
1947                  * In case of x86, we use 'lfence' insn for mitigation. We
1948                  * reuse preexisting logic from Spectre v1 mitigation that
1949                  * happens to produce the required code on x86 for v4 as well.
1950                  */
1951                 barrier_nospec();
1952                 CONT;
1953 #define LDST(SIZEOP, SIZE)                                              \
1954         STX_MEM_##SIZEOP:                                               \
1955                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
1956                 CONT;                                                   \
1957         ST_MEM_##SIZEOP:                                                \
1958                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
1959                 CONT;                                                   \
1960         LDX_MEM_##SIZEOP:                                               \
1961                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
1962                 CONT;                                                   \
1963         LDX_PROBE_MEM_##SIZEOP:                                         \
1964                 bpf_probe_read_kernel(&DST, sizeof(SIZE),               \
1965                                       (const void *)(long) (SRC + insn->off));  \
1966                 DST = *((SIZE *)&DST);                                  \
1967                 CONT;
1968
1969         LDST(B,   u8)
1970         LDST(H,  u16)
1971         LDST(W,  u32)
1972         LDST(DW, u64)
1973 #undef LDST
1974
1975 #define LDSX(SIZEOP, SIZE)                                              \
1976         LDX_MEMSX_##SIZEOP:                                             \
1977                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
1978                 CONT;                                                   \
1979         LDX_PROBE_MEMSX_##SIZEOP:                                       \
1980                 bpf_probe_read_kernel(&DST, sizeof(SIZE),               \
1981                                       (const void *)(long) (SRC + insn->off));  \
1982                 DST = *((SIZE *)&DST);                                  \
1983                 CONT;
1984
1985         LDSX(B,   s8)
1986         LDSX(H,  s16)
1987         LDSX(W,  s32)
1988 #undef LDSX
1989
1990 #define ATOMIC_ALU_OP(BOP, KOP)                                         \
1991                 case BOP:                                               \
1992                         if (BPF_SIZE(insn->code) == BPF_W)              \
1993                                 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
1994                                              (DST + insn->off));        \
1995                         else                                            \
1996                                 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
1997                                                (DST + insn->off));      \
1998                         break;                                          \
1999                 case BOP | BPF_FETCH:                                   \
2000                         if (BPF_SIZE(insn->code) == BPF_W)              \
2001                                 SRC = (u32) atomic_fetch_##KOP(         \
2002                                         (u32) SRC,                      \
2003                                         (atomic_t *)(unsigned long) (DST + insn->off)); \
2004                         else                                            \
2005                                 SRC = (u64) atomic64_fetch_##KOP(       \
2006                                         (u64) SRC,                      \
2007                                         (atomic64_t *)(unsigned long) (DST + insn->off)); \
2008                         break;
2009
2010         STX_ATOMIC_DW:
2011         STX_ATOMIC_W:
2012                 switch (IMM) {
2013                 ATOMIC_ALU_OP(BPF_ADD, add)
2014                 ATOMIC_ALU_OP(BPF_AND, and)
2015                 ATOMIC_ALU_OP(BPF_OR, or)
2016                 ATOMIC_ALU_OP(BPF_XOR, xor)
2017 #undef ATOMIC_ALU_OP
2018
2019                 case BPF_XCHG:
2020                         if (BPF_SIZE(insn->code) == BPF_W)
2021                                 SRC = (u32) atomic_xchg(
2022                                         (atomic_t *)(unsigned long) (DST + insn->off),
2023                                         (u32) SRC);
2024                         else
2025                                 SRC = (u64) atomic64_xchg(
2026                                         (atomic64_t *)(unsigned long) (DST + insn->off),
2027                                         (u64) SRC);
2028                         break;
2029                 case BPF_CMPXCHG:
2030                         if (BPF_SIZE(insn->code) == BPF_W)
2031                                 BPF_R0 = (u32) atomic_cmpxchg(
2032                                         (atomic_t *)(unsigned long) (DST + insn->off),
2033                                         (u32) BPF_R0, (u32) SRC);
2034                         else
2035                                 BPF_R0 = (u64) atomic64_cmpxchg(
2036                                         (atomic64_t *)(unsigned long) (DST + insn->off),
2037                                         (u64) BPF_R0, (u64) SRC);
2038                         break;
2039
2040                 default:
2041                         goto default_label;
2042                 }
2043                 CONT;
2044
2045         default_label:
2046                 /* If we ever reach this, we have a bug somewhere. Die hard here
2047                  * instead of just returning 0; we could be somewhere in a subprog,
2048                  * so execution could continue otherwise which we do /not/ want.
2049                  *
2050                  * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
2051                  */
2052                 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
2053                         insn->code, insn->imm);
2054                 BUG_ON(1);
2055                 return 0;
2056 }
2057
2058 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2059 #define DEFINE_BPF_PROG_RUN(stack_size) \
2060 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2061 { \
2062         u64 stack[stack_size / sizeof(u64)]; \
2063         u64 regs[MAX_BPF_EXT_REG] = {}; \
2064 \
2065         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2066         ARG1 = (u64) (unsigned long) ctx; \
2067         return ___bpf_prog_run(regs, insn); \
2068 }
2069
2070 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2071 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2072 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2073                                       const struct bpf_insn *insn) \
2074 { \
2075         u64 stack[stack_size / sizeof(u64)]; \
2076         u64 regs[MAX_BPF_EXT_REG]; \
2077 \
2078         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2079         BPF_R1 = r1; \
2080         BPF_R2 = r2; \
2081         BPF_R3 = r3; \
2082         BPF_R4 = r4; \
2083         BPF_R5 = r5; \
2084         return ___bpf_prog_run(regs, insn); \
2085 }
2086
2087 #define EVAL1(FN, X) FN(X)
2088 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2089 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2090 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2091 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2092 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2093
2094 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2095 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2096 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2097
2098 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2099 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2100 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2101
2102 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2103
2104 static unsigned int (*interpreters[])(const void *ctx,
2105                                       const struct bpf_insn *insn) = {
2106 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2107 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2108 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2109 };
2110 #undef PROG_NAME_LIST
2111 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2112 static __maybe_unused
2113 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2114                            const struct bpf_insn *insn) = {
2115 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2116 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2117 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2118 };
2119 #undef PROG_NAME_LIST
2120
2121 #ifdef CONFIG_BPF_SYSCALL
2122 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2123 {
2124         stack_depth = max_t(u32, stack_depth, 1);
2125         insn->off = (s16) insn->imm;
2126         insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2127                 __bpf_call_base_args;
2128         insn->code = BPF_JMP | BPF_CALL_ARGS;
2129 }
2130 #endif
2131 #else
2132 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2133                                          const struct bpf_insn *insn)
2134 {
2135         /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2136          * is not working properly, so warn about it!
2137          */
2138         WARN_ON_ONCE(1);
2139         return 0;
2140 }
2141 #endif
2142
2143 bool bpf_prog_map_compatible(struct bpf_map *map,
2144                              const struct bpf_prog *fp)
2145 {
2146         enum bpf_prog_type prog_type = resolve_prog_type(fp);
2147         bool ret;
2148
2149         if (fp->kprobe_override)
2150                 return false;
2151
2152         /* XDP programs inserted into maps are not guaranteed to run on
2153          * a particular netdev (and can run outside driver context entirely
2154          * in the case of devmap and cpumap). Until device checks
2155          * are implemented, prohibit adding dev-bound programs to program maps.
2156          */
2157         if (bpf_prog_is_dev_bound(fp->aux))
2158                 return false;
2159
2160         spin_lock(&map->owner.lock);
2161         if (!map->owner.type) {
2162                 /* There's no owner yet where we could check for
2163                  * compatibility.
2164                  */
2165                 map->owner.type  = prog_type;
2166                 map->owner.jited = fp->jited;
2167                 map->owner.xdp_has_frags = fp->aux->xdp_has_frags;
2168                 ret = true;
2169         } else {
2170                 ret = map->owner.type  == prog_type &&
2171                       map->owner.jited == fp->jited &&
2172                       map->owner.xdp_has_frags == fp->aux->xdp_has_frags;
2173         }
2174         spin_unlock(&map->owner.lock);
2175
2176         return ret;
2177 }
2178
2179 static int bpf_check_tail_call(const struct bpf_prog *fp)
2180 {
2181         struct bpf_prog_aux *aux = fp->aux;
2182         int i, ret = 0;
2183
2184         mutex_lock(&aux->used_maps_mutex);
2185         for (i = 0; i < aux->used_map_cnt; i++) {
2186                 struct bpf_map *map = aux->used_maps[i];
2187
2188                 if (!map_type_contains_progs(map))
2189                         continue;
2190
2191                 if (!bpf_prog_map_compatible(map, fp)) {
2192                         ret = -EINVAL;
2193                         goto out;
2194                 }
2195         }
2196
2197 out:
2198         mutex_unlock(&aux->used_maps_mutex);
2199         return ret;
2200 }
2201
2202 static void bpf_prog_select_func(struct bpf_prog *fp)
2203 {
2204 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2205         u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2206
2207         fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2208 #else
2209         fp->bpf_func = __bpf_prog_ret0_warn;
2210 #endif
2211 }
2212
2213 /**
2214  *      bpf_prog_select_runtime - select exec runtime for BPF program
2215  *      @fp: bpf_prog populated with BPF program
2216  *      @err: pointer to error variable
2217  *
2218  * Try to JIT eBPF program, if JIT is not available, use interpreter.
2219  * The BPF program will be executed via bpf_prog_run() function.
2220  *
2221  * Return: the &fp argument along with &err set to 0 for success or
2222  * a negative errno code on failure
2223  */
2224 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2225 {
2226         /* In case of BPF to BPF calls, verifier did all the prep
2227          * work with regards to JITing, etc.
2228          */
2229         bool jit_needed = false;
2230
2231         if (fp->bpf_func)
2232                 goto finalize;
2233
2234         if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2235             bpf_prog_has_kfunc_call(fp))
2236                 jit_needed = true;
2237
2238         bpf_prog_select_func(fp);
2239
2240         /* eBPF JITs can rewrite the program in case constant
2241          * blinding is active. However, in case of error during
2242          * blinding, bpf_int_jit_compile() must always return a
2243          * valid program, which in this case would simply not
2244          * be JITed, but falls back to the interpreter.
2245          */
2246         if (!bpf_prog_is_offloaded(fp->aux)) {
2247                 *err = bpf_prog_alloc_jited_linfo(fp);
2248                 if (*err)
2249                         return fp;
2250
2251                 fp = bpf_int_jit_compile(fp);
2252                 bpf_prog_jit_attempt_done(fp);
2253                 if (!fp->jited && jit_needed) {
2254                         *err = -ENOTSUPP;
2255                         return fp;
2256                 }
2257         } else {
2258                 *err = bpf_prog_offload_compile(fp);
2259                 if (*err)
2260                         return fp;
2261         }
2262
2263 finalize:
2264         bpf_prog_lock_ro(fp);
2265
2266         /* The tail call compatibility check can only be done at
2267          * this late stage as we need to determine, if we deal
2268          * with JITed or non JITed program concatenations and not
2269          * all eBPF JITs might immediately support all features.
2270          */
2271         *err = bpf_check_tail_call(fp);
2272
2273         return fp;
2274 }
2275 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2276
2277 static unsigned int __bpf_prog_ret1(const void *ctx,
2278                                     const struct bpf_insn *insn)
2279 {
2280         return 1;
2281 }
2282
2283 static struct bpf_prog_dummy {
2284         struct bpf_prog prog;
2285 } dummy_bpf_prog = {
2286         .prog = {
2287                 .bpf_func = __bpf_prog_ret1,
2288         },
2289 };
2290
2291 struct bpf_empty_prog_array bpf_empty_prog_array = {
2292         .null_prog = NULL,
2293 };
2294 EXPORT_SYMBOL(bpf_empty_prog_array);
2295
2296 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2297 {
2298         if (prog_cnt)
2299                 return kzalloc(sizeof(struct bpf_prog_array) +
2300                                sizeof(struct bpf_prog_array_item) *
2301                                (prog_cnt + 1),
2302                                flags);
2303
2304         return &bpf_empty_prog_array.hdr;
2305 }
2306
2307 void bpf_prog_array_free(struct bpf_prog_array *progs)
2308 {
2309         if (!progs || progs == &bpf_empty_prog_array.hdr)
2310                 return;
2311         kfree_rcu(progs, rcu);
2312 }
2313
2314 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2315 {
2316         struct bpf_prog_array *progs;
2317
2318         /* If RCU Tasks Trace grace period implies RCU grace period, there is
2319          * no need to call kfree_rcu(), just call kfree() directly.
2320          */
2321         progs = container_of(rcu, struct bpf_prog_array, rcu);
2322         if (rcu_trace_implies_rcu_gp())
2323                 kfree(progs);
2324         else
2325                 kfree_rcu(progs, rcu);
2326 }
2327
2328 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2329 {
2330         if (!progs || progs == &bpf_empty_prog_array.hdr)
2331                 return;
2332         call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2333 }
2334
2335 int bpf_prog_array_length(struct bpf_prog_array *array)
2336 {
2337         struct bpf_prog_array_item *item;
2338         u32 cnt = 0;
2339
2340         for (item = array->items; item->prog; item++)
2341                 if (item->prog != &dummy_bpf_prog.prog)
2342                         cnt++;
2343         return cnt;
2344 }
2345
2346 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2347 {
2348         struct bpf_prog_array_item *item;
2349
2350         for (item = array->items; item->prog; item++)
2351                 if (item->prog != &dummy_bpf_prog.prog)
2352                         return false;
2353         return true;
2354 }
2355
2356 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2357                                      u32 *prog_ids,
2358                                      u32 request_cnt)
2359 {
2360         struct bpf_prog_array_item *item;
2361         int i = 0;
2362
2363         for (item = array->items; item->prog; item++) {
2364                 if (item->prog == &dummy_bpf_prog.prog)
2365                         continue;
2366                 prog_ids[i] = item->prog->aux->id;
2367                 if (++i == request_cnt) {
2368                         item++;
2369                         break;
2370                 }
2371         }
2372
2373         return !!(item->prog);
2374 }
2375
2376 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2377                                 __u32 __user *prog_ids, u32 cnt)
2378 {
2379         unsigned long err = 0;
2380         bool nospc;
2381         u32 *ids;
2382
2383         /* users of this function are doing:
2384          * cnt = bpf_prog_array_length();
2385          * if (cnt > 0)
2386          *     bpf_prog_array_copy_to_user(..., cnt);
2387          * so below kcalloc doesn't need extra cnt > 0 check.
2388          */
2389         ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2390         if (!ids)
2391                 return -ENOMEM;
2392         nospc = bpf_prog_array_copy_core(array, ids, cnt);
2393         err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2394         kfree(ids);
2395         if (err)
2396                 return -EFAULT;
2397         if (nospc)
2398                 return -ENOSPC;
2399         return 0;
2400 }
2401
2402 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2403                                 struct bpf_prog *old_prog)
2404 {
2405         struct bpf_prog_array_item *item;
2406
2407         for (item = array->items; item->prog; item++)
2408                 if (item->prog == old_prog) {
2409                         WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2410                         break;
2411                 }
2412 }
2413
2414 /**
2415  * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2416  *                                   index into the program array with
2417  *                                   a dummy no-op program.
2418  * @array: a bpf_prog_array
2419  * @index: the index of the program to replace
2420  *
2421  * Skips over dummy programs, by not counting them, when calculating
2422  * the position of the program to replace.
2423  *
2424  * Return:
2425  * * 0          - Success
2426  * * -EINVAL    - Invalid index value. Must be a non-negative integer.
2427  * * -ENOENT    - Index out of range
2428  */
2429 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2430 {
2431         return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2432 }
2433
2434 /**
2435  * bpf_prog_array_update_at() - Updates the program at the given index
2436  *                              into the program array.
2437  * @array: a bpf_prog_array
2438  * @index: the index of the program to update
2439  * @prog: the program to insert into the array
2440  *
2441  * Skips over dummy programs, by not counting them, when calculating
2442  * the position of the program to update.
2443  *
2444  * Return:
2445  * * 0          - Success
2446  * * -EINVAL    - Invalid index value. Must be a non-negative integer.
2447  * * -ENOENT    - Index out of range
2448  */
2449 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2450                              struct bpf_prog *prog)
2451 {
2452         struct bpf_prog_array_item *item;
2453
2454         if (unlikely(index < 0))
2455                 return -EINVAL;
2456
2457         for (item = array->items; item->prog; item++) {
2458                 if (item->prog == &dummy_bpf_prog.prog)
2459                         continue;
2460                 if (!index) {
2461                         WRITE_ONCE(item->prog, prog);
2462                         return 0;
2463                 }
2464                 index--;
2465         }
2466         return -ENOENT;
2467 }
2468
2469 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2470                         struct bpf_prog *exclude_prog,
2471                         struct bpf_prog *include_prog,
2472                         u64 bpf_cookie,
2473                         struct bpf_prog_array **new_array)
2474 {
2475         int new_prog_cnt, carry_prog_cnt = 0;
2476         struct bpf_prog_array_item *existing, *new;
2477         struct bpf_prog_array *array;
2478         bool found_exclude = false;
2479
2480         /* Figure out how many existing progs we need to carry over to
2481          * the new array.
2482          */
2483         if (old_array) {
2484                 existing = old_array->items;
2485                 for (; existing->prog; existing++) {
2486                         if (existing->prog == exclude_prog) {
2487                                 found_exclude = true;
2488                                 continue;
2489                         }
2490                         if (existing->prog != &dummy_bpf_prog.prog)
2491                                 carry_prog_cnt++;
2492                         if (existing->prog == include_prog)
2493                                 return -EEXIST;
2494                 }
2495         }
2496
2497         if (exclude_prog && !found_exclude)
2498                 return -ENOENT;
2499
2500         /* How many progs (not NULL) will be in the new array? */
2501         new_prog_cnt = carry_prog_cnt;
2502         if (include_prog)
2503                 new_prog_cnt += 1;
2504
2505         /* Do we have any prog (not NULL) in the new array? */
2506         if (!new_prog_cnt) {
2507                 *new_array = NULL;
2508                 return 0;
2509         }
2510
2511         /* +1 as the end of prog_array is marked with NULL */
2512         array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2513         if (!array)
2514                 return -ENOMEM;
2515         new = array->items;
2516
2517         /* Fill in the new prog array */
2518         if (carry_prog_cnt) {
2519                 existing = old_array->items;
2520                 for (; existing->prog; existing++) {
2521                         if (existing->prog == exclude_prog ||
2522                             existing->prog == &dummy_bpf_prog.prog)
2523                                 continue;
2524
2525                         new->prog = existing->prog;
2526                         new->bpf_cookie = existing->bpf_cookie;
2527                         new++;
2528                 }
2529         }
2530         if (include_prog) {
2531                 new->prog = include_prog;
2532                 new->bpf_cookie = bpf_cookie;
2533                 new++;
2534         }
2535         new->prog = NULL;
2536         *new_array = array;
2537         return 0;
2538 }
2539
2540 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2541                              u32 *prog_ids, u32 request_cnt,
2542                              u32 *prog_cnt)
2543 {
2544         u32 cnt = 0;
2545
2546         if (array)
2547                 cnt = bpf_prog_array_length(array);
2548
2549         *prog_cnt = cnt;
2550
2551         /* return early if user requested only program count or nothing to copy */
2552         if (!request_cnt || !cnt)
2553                 return 0;
2554
2555         /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2556         return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2557                                                                      : 0;
2558 }
2559
2560 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2561                           struct bpf_map **used_maps, u32 len)
2562 {
2563         struct bpf_map *map;
2564         u32 i;
2565
2566         for (i = 0; i < len; i++) {
2567                 map = used_maps[i];
2568                 if (map->ops->map_poke_untrack)
2569                         map->ops->map_poke_untrack(map, aux);
2570                 bpf_map_put(map);
2571         }
2572 }
2573
2574 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2575 {
2576         __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2577         kfree(aux->used_maps);
2578 }
2579
2580 void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2581                           struct btf_mod_pair *used_btfs, u32 len)
2582 {
2583 #ifdef CONFIG_BPF_SYSCALL
2584         struct btf_mod_pair *btf_mod;
2585         u32 i;
2586
2587         for (i = 0; i < len; i++) {
2588                 btf_mod = &used_btfs[i];
2589                 if (btf_mod->module)
2590                         module_put(btf_mod->module);
2591                 btf_put(btf_mod->btf);
2592         }
2593 #endif
2594 }
2595
2596 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2597 {
2598         __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2599         kfree(aux->used_btfs);
2600 }
2601
2602 static void bpf_prog_free_deferred(struct work_struct *work)
2603 {
2604         struct bpf_prog_aux *aux;
2605         int i;
2606
2607         aux = container_of(work, struct bpf_prog_aux, work);
2608 #ifdef CONFIG_BPF_SYSCALL
2609         bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2610 #endif
2611 #ifdef CONFIG_CGROUP_BPF
2612         if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2613                 bpf_cgroup_atype_put(aux->cgroup_atype);
2614 #endif
2615         bpf_free_used_maps(aux);
2616         bpf_free_used_btfs(aux);
2617         if (bpf_prog_is_dev_bound(aux))
2618                 bpf_prog_dev_bound_destroy(aux->prog);
2619 #ifdef CONFIG_PERF_EVENTS
2620         if (aux->prog->has_callchain_buf)
2621                 put_callchain_buffers();
2622 #endif
2623         if (aux->dst_trampoline)
2624                 bpf_trampoline_put(aux->dst_trampoline);
2625         for (i = 0; i < aux->func_cnt; i++) {
2626                 /* We can just unlink the subprog poke descriptor table as
2627                  * it was originally linked to the main program and is also
2628                  * released along with it.
2629                  */
2630                 aux->func[i]->aux->poke_tab = NULL;
2631                 bpf_jit_free(aux->func[i]);
2632         }
2633         if (aux->func_cnt) {
2634                 kfree(aux->func);
2635                 bpf_prog_unlock_free(aux->prog);
2636         } else {
2637                 bpf_jit_free(aux->prog);
2638         }
2639 }
2640
2641 void bpf_prog_free(struct bpf_prog *fp)
2642 {
2643         struct bpf_prog_aux *aux = fp->aux;
2644
2645         if (aux->dst_prog)
2646                 bpf_prog_put(aux->dst_prog);
2647         INIT_WORK(&aux->work, bpf_prog_free_deferred);
2648         schedule_work(&aux->work);
2649 }
2650 EXPORT_SYMBOL_GPL(bpf_prog_free);
2651
2652 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2653 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2654
2655 void bpf_user_rnd_init_once(void)
2656 {
2657         prandom_init_once(&bpf_user_rnd_state);
2658 }
2659
2660 BPF_CALL_0(bpf_user_rnd_u32)
2661 {
2662         /* Should someone ever have the rather unwise idea to use some
2663          * of the registers passed into this function, then note that
2664          * this function is called from native eBPF and classic-to-eBPF
2665          * transformations. Register assignments from both sides are
2666          * different, f.e. classic always sets fn(ctx, A, X) here.
2667          */
2668         struct rnd_state *state;
2669         u32 res;
2670
2671         state = &get_cpu_var(bpf_user_rnd_state);
2672         res = prandom_u32_state(state);
2673         put_cpu_var(bpf_user_rnd_state);
2674
2675         return res;
2676 }
2677
2678 BPF_CALL_0(bpf_get_raw_cpu_id)
2679 {
2680         return raw_smp_processor_id();
2681 }
2682
2683 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2684 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2685 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2686 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2687 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2688 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2689 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2690 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2691 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2692 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2693 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2694
2695 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2696 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2697 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2698 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2699 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2700 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2701 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak;
2702
2703 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2704 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2705 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2706 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2707 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2708 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2709 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2710 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2711 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2712 const struct bpf_func_proto bpf_set_retval_proto __weak;
2713 const struct bpf_func_proto bpf_get_retval_proto __weak;
2714
2715 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2716 {
2717         return NULL;
2718 }
2719
2720 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2721 {
2722         return NULL;
2723 }
2724
2725 u64 __weak
2726 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2727                  void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2728 {
2729         return -ENOTSUPP;
2730 }
2731 EXPORT_SYMBOL_GPL(bpf_event_output);
2732
2733 /* Always built-in helper functions. */
2734 const struct bpf_func_proto bpf_tail_call_proto = {
2735         .func           = NULL,
2736         .gpl_only       = false,
2737         .ret_type       = RET_VOID,
2738         .arg1_type      = ARG_PTR_TO_CTX,
2739         .arg2_type      = ARG_CONST_MAP_PTR,
2740         .arg3_type      = ARG_ANYTHING,
2741 };
2742
2743 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2744  * It is encouraged to implement bpf_int_jit_compile() instead, so that
2745  * eBPF and implicitly also cBPF can get JITed!
2746  */
2747 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2748 {
2749         return prog;
2750 }
2751
2752 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2753  * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2754  */
2755 void __weak bpf_jit_compile(struct bpf_prog *prog)
2756 {
2757 }
2758
2759 bool __weak bpf_helper_changes_pkt_data(void *func)
2760 {
2761         return false;
2762 }
2763
2764 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2765  * analysis code and wants explicit zero extension inserted by verifier.
2766  * Otherwise, return FALSE.
2767  *
2768  * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2769  * you don't override this. JITs that don't want these extra insns can detect
2770  * them using insn_is_zext.
2771  */
2772 bool __weak bpf_jit_needs_zext(void)
2773 {
2774         return false;
2775 }
2776
2777 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2778 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2779 {
2780         return false;
2781 }
2782
2783 bool __weak bpf_jit_supports_kfunc_call(void)
2784 {
2785         return false;
2786 }
2787
2788 bool __weak bpf_jit_supports_far_kfunc_call(void)
2789 {
2790         return false;
2791 }
2792
2793 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2794  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2795  */
2796 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2797                          int len)
2798 {
2799         return -EFAULT;
2800 }
2801
2802 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2803                               void *addr1, void *addr2)
2804 {
2805         return -ENOTSUPP;
2806 }
2807
2808 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
2809 {
2810         return ERR_PTR(-ENOTSUPP);
2811 }
2812
2813 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
2814 {
2815         return -ENOTSUPP;
2816 }
2817
2818 #ifdef CONFIG_BPF_SYSCALL
2819 static int __init bpf_global_ma_init(void)
2820 {
2821         int ret;
2822
2823         ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false);
2824         bpf_global_ma_set = !ret;
2825         return ret;
2826 }
2827 late_initcall(bpf_global_ma_init);
2828 #endif
2829
2830 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2831 EXPORT_SYMBOL(bpf_stats_enabled_key);
2832
2833 /* All definitions of tracepoints related to BPF. */
2834 #define CREATE_TRACE_POINTS
2835 #include <linux/bpf_trace.h>
2836
2837 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2838 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);