1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Linux Socket Filter - Kernel level socket filtering
5 * Based on the design of the Berkeley Packet Filter. The new
6 * internal format has been designed by PLUMgrid:
8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
12 * Jay Schulist <jschlst@samba.org>
13 * Alexei Starovoitov <ast@plumgrid.com>
14 * Daniel Borkmann <dborkman@redhat.com>
16 * Andi Kleen - Fix a few bad bugs and races.
17 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
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>
39 #include <asm/barrier.h>
40 #include <asm/unaligned.h>
43 #define BPF_R0 regs[BPF_REG_0]
44 #define BPF_R1 regs[BPF_REG_1]
45 #define BPF_R2 regs[BPF_REG_2]
46 #define BPF_R3 regs[BPF_REG_3]
47 #define BPF_R4 regs[BPF_REG_4]
48 #define BPF_R5 regs[BPF_REG_5]
49 #define BPF_R6 regs[BPF_REG_6]
50 #define BPF_R7 regs[BPF_REG_7]
51 #define BPF_R8 regs[BPF_REG_8]
52 #define BPF_R9 regs[BPF_REG_9]
53 #define BPF_R10 regs[BPF_REG_10]
56 #define DST regs[insn->dst_reg]
57 #define SRC regs[insn->src_reg]
58 #define FP regs[BPF_REG_FP]
59 #define AX regs[BPF_REG_AX]
60 #define ARG1 regs[BPF_REG_ARG1]
61 #define CTX regs[BPF_REG_CTX]
64 /* No hurry in this branch
66 * Exported for the bpf jit load helper.
68 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
72 if (k >= SKF_NET_OFF) {
73 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
74 } else if (k >= SKF_LL_OFF) {
75 if (unlikely(!skb_mac_header_was_set(skb)))
77 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
79 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
85 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
87 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
88 struct bpf_prog_aux *aux;
91 size = round_up(size, PAGE_SIZE);
92 fp = __vmalloc(size, gfp_flags);
96 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags);
101 fp->active = alloc_percpu_gfp(int, GFP_KERNEL_ACCOUNT | gfp_extra_flags);
108 fp->pages = size / PAGE_SIZE;
111 fp->jit_requested = ebpf_jit_enabled();
112 fp->blinding_requested = bpf_jit_blinding_enabled(fp);
113 #ifdef CONFIG_CGROUP_BPF
114 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
117 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
118 mutex_init(&fp->aux->used_maps_mutex);
119 mutex_init(&fp->aux->dst_mutex);
124 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
126 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
127 struct bpf_prog *prog;
130 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
134 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
136 free_percpu(prog->active);
142 for_each_possible_cpu(cpu) {
143 struct bpf_prog_stats *pstats;
145 pstats = per_cpu_ptr(prog->stats, cpu);
146 u64_stats_init(&pstats->syncp);
150 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
152 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
154 if (!prog->aux->nr_linfo || !prog->jit_requested)
157 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
158 sizeof(*prog->aux->jited_linfo),
159 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
160 if (!prog->aux->jited_linfo)
166 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
168 if (prog->aux->jited_linfo &&
169 (!prog->jited || !prog->aux->jited_linfo[0])) {
170 kvfree(prog->aux->jited_linfo);
171 prog->aux->jited_linfo = NULL;
174 kfree(prog->aux->kfunc_tab);
175 prog->aux->kfunc_tab = NULL;
178 /* The jit engine is responsible to provide an array
179 * for insn_off to the jited_off mapping (insn_to_jit_off).
181 * The idx to this array is the insn_off. Hence, the insn_off
182 * here is relative to the prog itself instead of the main prog.
183 * This array has one entry for each xlated bpf insn.
185 * jited_off is the byte off to the end of the jited insn.
189 * The first bpf insn off of the prog. The insn off
190 * here is relative to the main prog.
191 * e.g. if prog is a subprog, insn_start > 0
193 * The prog's idx to prog->aux->linfo and jited_linfo
195 * jited_linfo[linfo_idx] = prog->bpf_func
199 * jited_linfo[i] = prog->bpf_func +
200 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
202 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
203 const u32 *insn_to_jit_off)
205 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
206 const struct bpf_line_info *linfo;
209 if (!prog->aux->jited_linfo)
210 /* Userspace did not provide linfo */
213 linfo_idx = prog->aux->linfo_idx;
214 linfo = &prog->aux->linfo[linfo_idx];
215 insn_start = linfo[0].insn_off;
216 insn_end = insn_start + prog->len;
218 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
219 jited_linfo[0] = prog->bpf_func;
221 nr_linfo = prog->aux->nr_linfo - linfo_idx;
223 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
224 /* The verifier ensures that linfo[i].insn_off is
225 * strictly increasing
227 jited_linfo[i] = prog->bpf_func +
228 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
231 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
232 gfp_t gfp_extra_flags)
234 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
238 size = round_up(size, PAGE_SIZE);
239 pages = size / PAGE_SIZE;
240 if (pages <= fp_old->pages)
243 fp = __vmalloc(size, gfp_flags);
245 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
249 /* We keep fp->aux from fp_old around in the new
250 * reallocated structure.
253 fp_old->stats = NULL;
254 fp_old->active = NULL;
255 __bpf_prog_free(fp_old);
261 void __bpf_prog_free(struct bpf_prog *fp)
264 mutex_destroy(&fp->aux->used_maps_mutex);
265 mutex_destroy(&fp->aux->dst_mutex);
266 kfree(fp->aux->poke_tab);
269 free_percpu(fp->stats);
270 free_percpu(fp->active);
274 int bpf_prog_calc_tag(struct bpf_prog *fp)
276 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
277 u32 raw_size = bpf_prog_tag_scratch_size(fp);
278 u32 digest[SHA1_DIGEST_WORDS];
279 u32 ws[SHA1_WORKSPACE_WORDS];
280 u32 i, bsize, psize, blocks;
281 struct bpf_insn *dst;
287 raw = vmalloc(raw_size);
292 memset(ws, 0, sizeof(ws));
294 /* We need to take out the map fd for the digest calculation
295 * since they are unstable from user space side.
298 for (i = 0, was_ld_map = false; i < fp->len; i++) {
299 dst[i] = fp->insnsi[i];
301 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
302 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
303 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
306 } else if (was_ld_map &&
308 dst[i].dst_reg == 0 &&
309 dst[i].src_reg == 0 &&
318 psize = bpf_prog_insn_size(fp);
319 memset(&raw[psize], 0, raw_size - psize);
322 bsize = round_up(psize, SHA1_BLOCK_SIZE);
323 blocks = bsize / SHA1_BLOCK_SIZE;
325 if (bsize - psize >= sizeof(__be64)) {
326 bits = (__be64 *)(todo + bsize - sizeof(__be64));
328 bits = (__be64 *)(todo + bsize + bits_offset);
331 *bits = cpu_to_be64((psize - 1) << 3);
334 sha1_transform(digest, todo, ws);
335 todo += SHA1_BLOCK_SIZE;
338 result = (__force __be32 *)digest;
339 for (i = 0; i < SHA1_DIGEST_WORDS; i++)
340 result[i] = cpu_to_be32(digest[i]);
341 memcpy(fp->tag, result, sizeof(fp->tag));
347 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
348 s32 end_new, s32 curr, const bool probe_pass)
350 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
351 s32 delta = end_new - end_old;
354 if (curr < pos && curr + imm + 1 >= end_old)
356 else if (curr >= end_new && curr + imm + 1 < end_new)
358 if (imm < imm_min || imm > imm_max)
365 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
366 s32 end_new, s32 curr, const bool probe_pass)
368 const s32 off_min = S16_MIN, off_max = S16_MAX;
369 s32 delta = end_new - end_old;
372 if (curr < pos && curr + off + 1 >= end_old)
374 else if (curr >= end_new && curr + off + 1 < end_new)
376 if (off < off_min || off > off_max)
383 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
384 s32 end_new, const bool probe_pass)
386 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
387 struct bpf_insn *insn = prog->insnsi;
390 for (i = 0; i < insn_cnt; i++, insn++) {
393 /* In the probing pass we still operate on the original,
394 * unpatched image in order to check overflows before we
395 * do any other adjustments. Therefore skip the patchlet.
397 if (probe_pass && i == pos) {
399 insn = prog->insnsi + end_old;
401 if (bpf_pseudo_func(insn)) {
402 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
403 end_new, i, probe_pass);
409 if ((BPF_CLASS(code) != BPF_JMP &&
410 BPF_CLASS(code) != BPF_JMP32) ||
411 BPF_OP(code) == BPF_EXIT)
413 /* Adjust offset of jmps if we cross patch boundaries. */
414 if (BPF_OP(code) == BPF_CALL) {
415 if (insn->src_reg != BPF_PSEUDO_CALL)
417 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
418 end_new, i, probe_pass);
420 ret = bpf_adj_delta_to_off(insn, pos, end_old,
421 end_new, i, probe_pass);
430 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
432 struct bpf_line_info *linfo;
435 nr_linfo = prog->aux->nr_linfo;
436 if (!nr_linfo || !delta)
439 linfo = prog->aux->linfo;
441 for (i = 0; i < nr_linfo; i++)
442 if (off < linfo[i].insn_off)
445 /* Push all off < linfo[i].insn_off by delta */
446 for (; i < nr_linfo; i++)
447 linfo[i].insn_off += delta;
450 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
451 const struct bpf_insn *patch, u32 len)
453 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
454 const u32 cnt_max = S16_MAX;
455 struct bpf_prog *prog_adj;
458 /* Since our patchlet doesn't expand the image, we're done. */
459 if (insn_delta == 0) {
460 memcpy(prog->insnsi + off, patch, sizeof(*patch));
464 insn_adj_cnt = prog->len + insn_delta;
466 /* Reject anything that would potentially let the insn->off
467 * target overflow when we have excessive program expansions.
468 * We need to probe here before we do any reallocation where
469 * we afterwards may not fail anymore.
471 if (insn_adj_cnt > cnt_max &&
472 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
475 /* Several new instructions need to be inserted. Make room
476 * for them. Likely, there's no need for a new allocation as
477 * last page could have large enough tailroom.
479 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
482 return ERR_PTR(-ENOMEM);
484 prog_adj->len = insn_adj_cnt;
486 /* Patching happens in 3 steps:
488 * 1) Move over tail of insnsi from next instruction onwards,
489 * so we can patch the single target insn with one or more
490 * new ones (patching is always from 1 to n insns, n > 0).
491 * 2) Inject new instructions at the target location.
492 * 3) Adjust branch offsets if necessary.
494 insn_rest = insn_adj_cnt - off - len;
496 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
497 sizeof(*patch) * insn_rest);
498 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
500 /* We are guaranteed to not fail at this point, otherwise
501 * the ship has sailed to reverse to the original state. An
502 * overflow cannot happen at this point.
504 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
506 bpf_adj_linfo(prog_adj, off, insn_delta);
511 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
513 /* Branch offsets can't overflow when program is shrinking, no need
514 * to call bpf_adj_branches(..., true) here
516 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
517 sizeof(struct bpf_insn) * (prog->len - off - cnt));
520 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
523 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
527 for (i = 0; i < fp->aux->func_cnt; i++)
528 bpf_prog_kallsyms_del(fp->aux->func[i]);
531 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
533 bpf_prog_kallsyms_del_subprogs(fp);
534 bpf_prog_kallsyms_del(fp);
537 #ifdef CONFIG_BPF_JIT
538 /* All BPF JIT sysctl knobs here. */
539 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
540 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
541 int bpf_jit_harden __read_mostly;
542 long bpf_jit_limit __read_mostly;
543 long bpf_jit_limit_max __read_mostly;
546 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
548 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
550 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
551 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len;
555 bpf_prog_ksym_set_name(struct bpf_prog *prog)
557 char *sym = prog->aux->ksym.name;
558 const char *end = sym + KSYM_NAME_LEN;
559 const struct btf_type *type;
560 const char *func_name;
562 BUILD_BUG_ON(sizeof("bpf_prog_") +
563 sizeof(prog->tag) * 2 +
564 /* name has been null terminated.
565 * We should need +1 for the '_' preceding
566 * the name. However, the null character
567 * is double counted between the name and the
568 * sizeof("bpf_prog_") above, so we omit
571 sizeof(prog->aux->name) > KSYM_NAME_LEN);
573 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
574 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
576 /* prog->aux->name will be ignored if full btf name is available */
577 if (prog->aux->func_info_cnt) {
578 type = btf_type_by_id(prog->aux->btf,
579 prog->aux->func_info[prog->aux->func_idx].type_id);
580 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
581 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
585 if (prog->aux->name[0])
586 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
591 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
593 return container_of(n, struct bpf_ksym, tnode)->start;
596 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
597 struct latch_tree_node *b)
599 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
602 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
604 unsigned long val = (unsigned long)key;
605 const struct bpf_ksym *ksym;
607 ksym = container_of(n, struct bpf_ksym, tnode);
609 if (val < ksym->start)
611 if (val >= ksym->end)
617 static const struct latch_tree_ops bpf_tree_ops = {
618 .less = bpf_tree_less,
619 .comp = bpf_tree_comp,
622 static DEFINE_SPINLOCK(bpf_lock);
623 static LIST_HEAD(bpf_kallsyms);
624 static struct latch_tree_root bpf_tree __cacheline_aligned;
626 void bpf_ksym_add(struct bpf_ksym *ksym)
628 spin_lock_bh(&bpf_lock);
629 WARN_ON_ONCE(!list_empty(&ksym->lnode));
630 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
631 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
632 spin_unlock_bh(&bpf_lock);
635 static void __bpf_ksym_del(struct bpf_ksym *ksym)
637 if (list_empty(&ksym->lnode))
640 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
641 list_del_rcu(&ksym->lnode);
644 void bpf_ksym_del(struct bpf_ksym *ksym)
646 spin_lock_bh(&bpf_lock);
647 __bpf_ksym_del(ksym);
648 spin_unlock_bh(&bpf_lock);
651 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
653 return fp->jited && !bpf_prog_was_classic(fp);
656 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
658 if (!bpf_prog_kallsyms_candidate(fp) ||
662 bpf_prog_ksym_set_addr(fp);
663 bpf_prog_ksym_set_name(fp);
664 fp->aux->ksym.prog = true;
666 bpf_ksym_add(&fp->aux->ksym);
669 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
671 if (!bpf_prog_kallsyms_candidate(fp))
674 bpf_ksym_del(&fp->aux->ksym);
677 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
679 struct latch_tree_node *n;
681 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
682 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
685 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
686 unsigned long *off, char *sym)
688 struct bpf_ksym *ksym;
692 ksym = bpf_ksym_find(addr);
694 unsigned long symbol_start = ksym->start;
695 unsigned long symbol_end = ksym->end;
697 strncpy(sym, ksym->name, KSYM_NAME_LEN);
701 *size = symbol_end - symbol_start;
703 *off = addr - symbol_start;
710 bool is_bpf_text_address(unsigned long addr)
715 ret = bpf_ksym_find(addr) != NULL;
721 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
723 struct bpf_ksym *ksym = bpf_ksym_find(addr);
725 return ksym && ksym->prog ?
726 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
730 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
732 const struct exception_table_entry *e = NULL;
733 struct bpf_prog *prog;
736 prog = bpf_prog_ksym_find(addr);
739 if (!prog->aux->num_exentries)
742 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
748 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
751 struct bpf_ksym *ksym;
755 if (!bpf_jit_kallsyms_enabled())
759 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
763 strncpy(sym, ksym->name, KSYM_NAME_LEN);
765 *value = ksym->start;
766 *type = BPF_SYM_ELF_TYPE;
776 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
777 struct bpf_jit_poke_descriptor *poke)
779 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
780 static const u32 poke_tab_max = 1024;
781 u32 slot = prog->aux->size_poke_tab;
784 if (size > poke_tab_max)
786 if (poke->tailcall_target || poke->tailcall_target_stable ||
787 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
790 switch (poke->reason) {
791 case BPF_POKE_REASON_TAIL_CALL:
792 if (!poke->tail_call.map)
799 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
803 memcpy(&tab[slot], poke, sizeof(*poke));
804 prog->aux->size_poke_tab = size;
805 prog->aux->poke_tab = tab;
811 * BPF program pack allocator.
813 * Most BPF programs are pretty small. Allocating a hole page for each
814 * program is sometime a waste. Many small bpf program also adds pressure
815 * to instruction TLB. To solve this issue, we introduce a BPF program pack
816 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
817 * to host BPF programs.
819 #define BPF_PROG_CHUNK_SHIFT 6
820 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT)
821 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1))
823 struct bpf_prog_pack {
824 struct list_head list;
826 unsigned long bitmap[];
829 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size)
831 memset(area, 0, size);
834 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
836 static DEFINE_MUTEX(pack_mutex);
837 static LIST_HEAD(pack_list);
839 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
840 * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
843 #define BPF_PROG_PACK_SIZE (PMD_SIZE * num_possible_nodes())
845 #define BPF_PROG_PACK_SIZE PAGE_SIZE
848 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
850 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
852 struct bpf_prog_pack *pack;
854 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
858 pack->ptr = module_alloc(BPF_PROG_PACK_SIZE);
863 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
864 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
865 list_add_tail(&pack->list, &pack_list);
867 set_vm_flush_reset_perms(pack->ptr);
868 set_memory_ro((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
869 set_memory_x((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
873 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
875 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
876 struct bpf_prog_pack *pack;
880 mutex_lock(&pack_mutex);
881 if (size > BPF_PROG_PACK_SIZE) {
882 size = round_up(size, PAGE_SIZE);
883 ptr = module_alloc(size);
885 bpf_fill_ill_insns(ptr, size);
886 set_vm_flush_reset_perms(ptr);
887 set_memory_ro((unsigned long)ptr, size / PAGE_SIZE);
888 set_memory_x((unsigned long)ptr, size / PAGE_SIZE);
892 list_for_each_entry(pack, &pack_list, list) {
893 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
895 if (pos < BPF_PROG_CHUNK_COUNT)
896 goto found_free_area;
899 pack = alloc_new_pack(bpf_fill_ill_insns);
906 bitmap_set(pack->bitmap, pos, nbits);
907 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
910 mutex_unlock(&pack_mutex);
914 void bpf_prog_pack_free(struct bpf_binary_header *hdr)
916 struct bpf_prog_pack *pack = NULL, *tmp;
920 mutex_lock(&pack_mutex);
921 if (hdr->size > BPF_PROG_PACK_SIZE) {
926 list_for_each_entry(tmp, &pack_list, list) {
927 if ((void *)hdr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > (void *)hdr) {
933 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
936 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size);
937 pos = ((unsigned long)hdr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
939 WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size),
940 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
942 bitmap_clear(pack->bitmap, pos, nbits);
943 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
944 BPF_PROG_CHUNK_COUNT, 0) == 0) {
945 list_del(&pack->list);
946 module_memfree(pack->ptr);
950 mutex_unlock(&pack_mutex);
953 static atomic_long_t bpf_jit_current;
955 /* Can be overridden by an arch's JIT compiler if it has a custom,
956 * dedicated BPF backend memory area, or if neither of the two
959 u64 __weak bpf_jit_alloc_exec_limit(void)
961 #if defined(MODULES_VADDR)
962 return MODULES_END - MODULES_VADDR;
964 return VMALLOC_END - VMALLOC_START;
968 static int __init bpf_jit_charge_init(void)
970 /* Only used as heuristic here to derive limit. */
971 bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
972 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1,
973 PAGE_SIZE), LONG_MAX);
976 pure_initcall(bpf_jit_charge_init);
978 int bpf_jit_charge_modmem(u32 size)
980 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
981 if (!bpf_capable()) {
982 atomic_long_sub(size, &bpf_jit_current);
990 void bpf_jit_uncharge_modmem(u32 size)
992 atomic_long_sub(size, &bpf_jit_current);
995 void *__weak bpf_jit_alloc_exec(unsigned long size)
997 return module_alloc(size);
1000 void __weak bpf_jit_free_exec(void *addr)
1002 module_memfree(addr);
1005 struct bpf_binary_header *
1006 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1007 unsigned int alignment,
1008 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1010 struct bpf_binary_header *hdr;
1011 u32 size, hole, start;
1013 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1014 alignment > BPF_IMAGE_ALIGNMENT);
1016 /* Most of BPF filters are really small, but if some of them
1017 * fill a page, allow at least 128 extra bytes to insert a
1018 * random section of illegal instructions.
1020 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1022 if (bpf_jit_charge_modmem(size))
1024 hdr = bpf_jit_alloc_exec(size);
1026 bpf_jit_uncharge_modmem(size);
1030 /* Fill space with illegal/arch-dep instructions. */
1031 bpf_fill_ill_insns(hdr, size);
1034 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1035 PAGE_SIZE - sizeof(*hdr));
1036 start = prandom_u32_max(hole) & ~(alignment - 1);
1038 /* Leave a random number of instructions before BPF code. */
1039 *image_ptr = &hdr->image[start];
1044 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1046 u32 size = hdr->size;
1048 bpf_jit_free_exec(hdr);
1049 bpf_jit_uncharge_modmem(size);
1052 /* Allocate jit binary from bpf_prog_pack allocator.
1053 * Since the allocated memory is RO+X, the JIT engine cannot write directly
1054 * to the memory. To solve this problem, a RW buffer is also allocated at
1055 * as the same time. The JIT engine should calculate offsets based on the
1056 * RO memory address, but write JITed program to the RW buffer. Once the
1057 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1058 * the JITed program to the RO memory.
1060 struct bpf_binary_header *
1061 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1062 unsigned int alignment,
1063 struct bpf_binary_header **rw_header,
1065 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1067 struct bpf_binary_header *ro_header;
1068 u32 size, hole, start;
1070 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1071 alignment > BPF_IMAGE_ALIGNMENT);
1073 /* add 16 bytes for a random section of illegal instructions */
1074 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1076 if (bpf_jit_charge_modmem(size))
1078 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1080 bpf_jit_uncharge_modmem(size);
1084 *rw_header = kvmalloc(size, GFP_KERNEL);
1086 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size));
1087 bpf_prog_pack_free(ro_header);
1088 bpf_jit_uncharge_modmem(size);
1092 /* Fill space with illegal/arch-dep instructions. */
1093 bpf_fill_ill_insns(*rw_header, size);
1094 (*rw_header)->size = size;
1096 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1097 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1098 start = prandom_u32_max(hole) & ~(alignment - 1);
1100 *image_ptr = &ro_header->image[start];
1101 *rw_image = &(*rw_header)->image[start];
1106 /* Copy JITed text from rw_header to its final location, the ro_header. */
1107 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog,
1108 struct bpf_binary_header *ro_header,
1109 struct bpf_binary_header *rw_header)
1113 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1118 bpf_prog_pack_free(ro_header);
1119 return PTR_ERR(ptr);
1124 /* bpf_jit_binary_pack_free is called in two different scenarios:
1125 * 1) when the program is freed after;
1126 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1127 * For case 2), we need to free both the RO memory and the RW buffer.
1129 * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1130 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1131 * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1132 * bpf_arch_text_copy (when jit fails).
1134 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1135 struct bpf_binary_header *rw_header)
1137 u32 size = ro_header->size;
1139 bpf_prog_pack_free(ro_header);
1141 bpf_jit_uncharge_modmem(size);
1144 struct bpf_binary_header *
1145 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1147 unsigned long real_start = (unsigned long)fp->bpf_func;
1150 addr = real_start & BPF_PROG_CHUNK_MASK;
1151 return (void *)addr;
1154 static inline struct bpf_binary_header *
1155 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1157 unsigned long real_start = (unsigned long)fp->bpf_func;
1160 addr = real_start & PAGE_MASK;
1161 return (void *)addr;
1164 /* This symbol is only overridden by archs that have different
1165 * requirements than the usual eBPF JITs, f.e. when they only
1166 * implement cBPF JIT, do not set images read-only, etc.
1168 void __weak bpf_jit_free(struct bpf_prog *fp)
1171 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1173 bpf_jit_binary_free(hdr);
1174 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1177 bpf_prog_unlock_free(fp);
1180 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1181 const struct bpf_insn *insn, bool extra_pass,
1182 u64 *func_addr, bool *func_addr_fixed)
1184 s16 off = insn->off;
1185 s32 imm = insn->imm;
1188 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1189 if (!*func_addr_fixed) {
1190 /* Place-holder address till the last pass has collected
1191 * all addresses for JITed subprograms in which case we
1192 * can pick them up from prog->aux.
1196 else if (prog->aux->func &&
1197 off >= 0 && off < prog->aux->func_cnt)
1198 addr = (u8 *)prog->aux->func[off]->bpf_func;
1202 /* Address of a BPF helper call. Since part of the core
1203 * kernel, it's always at a fixed location. __bpf_call_base
1204 * and the helper with imm relative to it are both in core
1207 addr = (u8 *)__bpf_call_base + imm;
1210 *func_addr = (unsigned long)addr;
1214 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1215 const struct bpf_insn *aux,
1216 struct bpf_insn *to_buff,
1219 struct bpf_insn *to = to_buff;
1220 u32 imm_rnd = get_random_u32();
1223 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
1224 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1226 /* Constraints on AX register:
1228 * AX register is inaccessible from user space. It is mapped in
1229 * all JITs, and used here for constant blinding rewrites. It is
1230 * typically "stateless" meaning its contents are only valid within
1231 * the executed instruction, but not across several instructions.
1232 * There are a few exceptions however which are further detailed
1235 * Constant blinding is only used by JITs, not in the interpreter.
1236 * The interpreter uses AX in some occasions as a local temporary
1237 * register e.g. in DIV or MOD instructions.
1239 * In restricted circumstances, the verifier can also use the AX
1240 * register for rewrites as long as they do not interfere with
1243 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1246 if (from->imm == 0 &&
1247 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
1248 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1249 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1253 switch (from->code) {
1254 case BPF_ALU | BPF_ADD | BPF_K:
1255 case BPF_ALU | BPF_SUB | BPF_K:
1256 case BPF_ALU | BPF_AND | BPF_K:
1257 case BPF_ALU | BPF_OR | BPF_K:
1258 case BPF_ALU | BPF_XOR | BPF_K:
1259 case BPF_ALU | BPF_MUL | BPF_K:
1260 case BPF_ALU | BPF_MOV | BPF_K:
1261 case BPF_ALU | BPF_DIV | BPF_K:
1262 case BPF_ALU | BPF_MOD | BPF_K:
1263 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1264 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1265 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1268 case BPF_ALU64 | BPF_ADD | BPF_K:
1269 case BPF_ALU64 | BPF_SUB | BPF_K:
1270 case BPF_ALU64 | BPF_AND | BPF_K:
1271 case BPF_ALU64 | BPF_OR | BPF_K:
1272 case BPF_ALU64 | BPF_XOR | BPF_K:
1273 case BPF_ALU64 | BPF_MUL | BPF_K:
1274 case BPF_ALU64 | BPF_MOV | BPF_K:
1275 case BPF_ALU64 | BPF_DIV | BPF_K:
1276 case BPF_ALU64 | BPF_MOD | BPF_K:
1277 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1278 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1279 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1282 case BPF_JMP | BPF_JEQ | BPF_K:
1283 case BPF_JMP | BPF_JNE | BPF_K:
1284 case BPF_JMP | BPF_JGT | BPF_K:
1285 case BPF_JMP | BPF_JLT | BPF_K:
1286 case BPF_JMP | BPF_JGE | BPF_K:
1287 case BPF_JMP | BPF_JLE | BPF_K:
1288 case BPF_JMP | BPF_JSGT | BPF_K:
1289 case BPF_JMP | BPF_JSLT | BPF_K:
1290 case BPF_JMP | BPF_JSGE | BPF_K:
1291 case BPF_JMP | BPF_JSLE | BPF_K:
1292 case BPF_JMP | BPF_JSET | BPF_K:
1293 /* Accommodate for extra offset in case of a backjump. */
1297 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1298 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1299 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1302 case BPF_JMP32 | BPF_JEQ | BPF_K:
1303 case BPF_JMP32 | BPF_JNE | BPF_K:
1304 case BPF_JMP32 | BPF_JGT | BPF_K:
1305 case BPF_JMP32 | BPF_JLT | BPF_K:
1306 case BPF_JMP32 | BPF_JGE | BPF_K:
1307 case BPF_JMP32 | BPF_JLE | BPF_K:
1308 case BPF_JMP32 | BPF_JSGT | BPF_K:
1309 case BPF_JMP32 | BPF_JSLT | BPF_K:
1310 case BPF_JMP32 | BPF_JSGE | BPF_K:
1311 case BPF_JMP32 | BPF_JSLE | BPF_K:
1312 case BPF_JMP32 | BPF_JSET | BPF_K:
1313 /* Accommodate for extra offset in case of a backjump. */
1317 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1318 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1319 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1323 case BPF_LD | BPF_IMM | BPF_DW:
1324 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1325 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1326 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1327 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1329 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1330 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1331 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1333 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1334 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1337 case BPF_ST | BPF_MEM | BPF_DW:
1338 case BPF_ST | BPF_MEM | BPF_W:
1339 case BPF_ST | BPF_MEM | BPF_H:
1340 case BPF_ST | BPF_MEM | BPF_B:
1341 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1342 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1343 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1347 return to - to_buff;
1350 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1351 gfp_t gfp_extra_flags)
1353 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1354 struct bpf_prog *fp;
1356 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1358 /* aux->prog still points to the fp_other one, so
1359 * when promoting the clone to the real program,
1360 * this still needs to be adapted.
1362 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1368 static void bpf_prog_clone_free(struct bpf_prog *fp)
1370 /* aux was stolen by the other clone, so we cannot free
1371 * it from this path! It will be freed eventually by the
1372 * other program on release.
1374 * At this point, we don't need a deferred release since
1375 * clone is guaranteed to not be locked.
1380 __bpf_prog_free(fp);
1383 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1385 /* We have to repoint aux->prog to self, as we don't
1386 * know whether fp here is the clone or the original.
1389 bpf_prog_clone_free(fp_other);
1392 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1394 struct bpf_insn insn_buff[16], aux[2];
1395 struct bpf_prog *clone, *tmp;
1396 int insn_delta, insn_cnt;
1397 struct bpf_insn *insn;
1400 if (!prog->blinding_requested || prog->blinded)
1403 clone = bpf_prog_clone_create(prog, GFP_USER);
1405 return ERR_PTR(-ENOMEM);
1407 insn_cnt = clone->len;
1408 insn = clone->insnsi;
1410 for (i = 0; i < insn_cnt; i++, insn++) {
1411 if (bpf_pseudo_func(insn)) {
1412 /* ld_imm64 with an address of bpf subprog is not
1413 * a user controlled constant. Don't randomize it,
1414 * since it will conflict with jit_subprogs() logic.
1421 /* We temporarily need to hold the original ld64 insn
1422 * so that we can still access the first part in the
1423 * second blinding run.
1425 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1427 memcpy(aux, insn, sizeof(aux));
1429 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1430 clone->aux->verifier_zext);
1434 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1436 /* Patching may have repointed aux->prog during
1437 * realloc from the original one, so we need to
1438 * fix it up here on error.
1440 bpf_jit_prog_release_other(prog, clone);
1445 insn_delta = rewritten - 1;
1447 /* Walk new program and skip insns we just inserted. */
1448 insn = clone->insnsi + i + insn_delta;
1449 insn_cnt += insn_delta;
1456 #endif /* CONFIG_BPF_JIT */
1458 /* Base function for offset calculation. Needs to go into .text section,
1459 * therefore keeping it non-static as well; will also be used by JITs
1460 * anyway later on, so do not let the compiler omit it. This also needs
1461 * to go into kallsyms for correlation from e.g. bpftool, so naming
1464 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1468 EXPORT_SYMBOL_GPL(__bpf_call_base);
1470 /* All UAPI available opcodes. */
1471 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1472 /* 32 bit ALU operations. */ \
1473 /* Register based. */ \
1474 INSN_3(ALU, ADD, X), \
1475 INSN_3(ALU, SUB, X), \
1476 INSN_3(ALU, AND, X), \
1477 INSN_3(ALU, OR, X), \
1478 INSN_3(ALU, LSH, X), \
1479 INSN_3(ALU, RSH, X), \
1480 INSN_3(ALU, XOR, X), \
1481 INSN_3(ALU, MUL, X), \
1482 INSN_3(ALU, MOV, X), \
1483 INSN_3(ALU, ARSH, X), \
1484 INSN_3(ALU, DIV, X), \
1485 INSN_3(ALU, MOD, X), \
1487 INSN_3(ALU, END, TO_BE), \
1488 INSN_3(ALU, END, TO_LE), \
1489 /* Immediate based. */ \
1490 INSN_3(ALU, ADD, K), \
1491 INSN_3(ALU, SUB, K), \
1492 INSN_3(ALU, AND, K), \
1493 INSN_3(ALU, OR, K), \
1494 INSN_3(ALU, LSH, K), \
1495 INSN_3(ALU, RSH, K), \
1496 INSN_3(ALU, XOR, K), \
1497 INSN_3(ALU, MUL, K), \
1498 INSN_3(ALU, MOV, K), \
1499 INSN_3(ALU, ARSH, K), \
1500 INSN_3(ALU, DIV, K), \
1501 INSN_3(ALU, MOD, K), \
1502 /* 64 bit ALU operations. */ \
1503 /* Register based. */ \
1504 INSN_3(ALU64, ADD, X), \
1505 INSN_3(ALU64, SUB, X), \
1506 INSN_3(ALU64, AND, X), \
1507 INSN_3(ALU64, OR, X), \
1508 INSN_3(ALU64, LSH, X), \
1509 INSN_3(ALU64, RSH, X), \
1510 INSN_3(ALU64, XOR, X), \
1511 INSN_3(ALU64, MUL, X), \
1512 INSN_3(ALU64, MOV, X), \
1513 INSN_3(ALU64, ARSH, X), \
1514 INSN_3(ALU64, DIV, X), \
1515 INSN_3(ALU64, MOD, X), \
1516 INSN_2(ALU64, NEG), \
1517 /* Immediate based. */ \
1518 INSN_3(ALU64, ADD, K), \
1519 INSN_3(ALU64, SUB, K), \
1520 INSN_3(ALU64, AND, K), \
1521 INSN_3(ALU64, OR, K), \
1522 INSN_3(ALU64, LSH, K), \
1523 INSN_3(ALU64, RSH, K), \
1524 INSN_3(ALU64, XOR, K), \
1525 INSN_3(ALU64, MUL, K), \
1526 INSN_3(ALU64, MOV, K), \
1527 INSN_3(ALU64, ARSH, K), \
1528 INSN_3(ALU64, DIV, K), \
1529 INSN_3(ALU64, MOD, K), \
1530 /* Call instruction. */ \
1531 INSN_2(JMP, CALL), \
1532 /* Exit instruction. */ \
1533 INSN_2(JMP, EXIT), \
1534 /* 32-bit Jump instructions. */ \
1535 /* Register based. */ \
1536 INSN_3(JMP32, JEQ, X), \
1537 INSN_3(JMP32, JNE, X), \
1538 INSN_3(JMP32, JGT, X), \
1539 INSN_3(JMP32, JLT, X), \
1540 INSN_3(JMP32, JGE, X), \
1541 INSN_3(JMP32, JLE, X), \
1542 INSN_3(JMP32, JSGT, X), \
1543 INSN_3(JMP32, JSLT, X), \
1544 INSN_3(JMP32, JSGE, X), \
1545 INSN_3(JMP32, JSLE, X), \
1546 INSN_3(JMP32, JSET, X), \
1547 /* Immediate based. */ \
1548 INSN_3(JMP32, JEQ, K), \
1549 INSN_3(JMP32, JNE, K), \
1550 INSN_3(JMP32, JGT, K), \
1551 INSN_3(JMP32, JLT, K), \
1552 INSN_3(JMP32, JGE, K), \
1553 INSN_3(JMP32, JLE, K), \
1554 INSN_3(JMP32, JSGT, K), \
1555 INSN_3(JMP32, JSLT, K), \
1556 INSN_3(JMP32, JSGE, K), \
1557 INSN_3(JMP32, JSLE, K), \
1558 INSN_3(JMP32, JSET, K), \
1559 /* Jump instructions. */ \
1560 /* Register based. */ \
1561 INSN_3(JMP, JEQ, X), \
1562 INSN_3(JMP, JNE, X), \
1563 INSN_3(JMP, JGT, X), \
1564 INSN_3(JMP, JLT, X), \
1565 INSN_3(JMP, JGE, X), \
1566 INSN_3(JMP, JLE, X), \
1567 INSN_3(JMP, JSGT, X), \
1568 INSN_3(JMP, JSLT, X), \
1569 INSN_3(JMP, JSGE, X), \
1570 INSN_3(JMP, JSLE, X), \
1571 INSN_3(JMP, JSET, X), \
1572 /* Immediate based. */ \
1573 INSN_3(JMP, JEQ, K), \
1574 INSN_3(JMP, JNE, K), \
1575 INSN_3(JMP, JGT, K), \
1576 INSN_3(JMP, JLT, K), \
1577 INSN_3(JMP, JGE, K), \
1578 INSN_3(JMP, JLE, K), \
1579 INSN_3(JMP, JSGT, K), \
1580 INSN_3(JMP, JSLT, K), \
1581 INSN_3(JMP, JSGE, K), \
1582 INSN_3(JMP, JSLE, K), \
1583 INSN_3(JMP, JSET, K), \
1585 /* Store instructions. */ \
1586 /* Register based. */ \
1587 INSN_3(STX, MEM, B), \
1588 INSN_3(STX, MEM, H), \
1589 INSN_3(STX, MEM, W), \
1590 INSN_3(STX, MEM, DW), \
1591 INSN_3(STX, ATOMIC, W), \
1592 INSN_3(STX, ATOMIC, DW), \
1593 /* Immediate based. */ \
1594 INSN_3(ST, MEM, B), \
1595 INSN_3(ST, MEM, H), \
1596 INSN_3(ST, MEM, W), \
1597 INSN_3(ST, MEM, DW), \
1598 /* Load instructions. */ \
1599 /* Register based. */ \
1600 INSN_3(LDX, MEM, B), \
1601 INSN_3(LDX, MEM, H), \
1602 INSN_3(LDX, MEM, W), \
1603 INSN_3(LDX, MEM, DW), \
1604 /* Immediate based. */ \
1607 bool bpf_opcode_in_insntable(u8 code)
1609 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1610 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1611 static const bool public_insntable[256] = {
1612 [0 ... 255] = false,
1613 /* Now overwrite non-defaults ... */
1614 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1615 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1616 [BPF_LD | BPF_ABS | BPF_B] = true,
1617 [BPF_LD | BPF_ABS | BPF_H] = true,
1618 [BPF_LD | BPF_ABS | BPF_W] = true,
1619 [BPF_LD | BPF_IND | BPF_B] = true,
1620 [BPF_LD | BPF_IND | BPF_H] = true,
1621 [BPF_LD | BPF_IND | BPF_W] = true,
1623 #undef BPF_INSN_3_TBL
1624 #undef BPF_INSN_2_TBL
1625 return public_insntable[code];
1628 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1629 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1631 memset(dst, 0, size);
1636 * ___bpf_prog_run - run eBPF program on a given context
1637 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1638 * @insn: is the array of eBPF instructions
1640 * Decode and execute eBPF instructions.
1642 * Return: whatever value is in %BPF_R0 at program exit
1644 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1646 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1647 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1648 static const void * const jumptable[256] __annotate_jump_table = {
1649 [0 ... 255] = &&default_label,
1650 /* Now overwrite non-defaults ... */
1651 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1652 /* Non-UAPI available opcodes. */
1653 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1654 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1655 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1656 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1657 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1658 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1659 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1661 #undef BPF_INSN_3_LBL
1662 #undef BPF_INSN_2_LBL
1663 u32 tail_call_cnt = 0;
1665 #define CONT ({ insn++; goto select_insn; })
1666 #define CONT_JMP ({ insn++; goto select_insn; })
1669 goto *jumptable[insn->code];
1671 /* Explicitly mask the register-based shift amounts with 63 or 31
1672 * to avoid undefined behavior. Normally this won't affect the
1673 * generated code, for example, in case of native 64 bit archs such
1674 * as x86-64 or arm64, the compiler is optimizing the AND away for
1675 * the interpreter. In case of JITs, each of the JIT backends compiles
1676 * the BPF shift operations to machine instructions which produce
1677 * implementation-defined results in such a case; the resulting
1678 * contents of the register may be arbitrary, but program behaviour
1679 * as a whole remains defined. In other words, in case of JIT backends,
1680 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1683 #define SHT(OPCODE, OP) \
1684 ALU64_##OPCODE##_X: \
1685 DST = DST OP (SRC & 63); \
1688 DST = (u32) DST OP ((u32) SRC & 31); \
1690 ALU64_##OPCODE##_K: \
1694 DST = (u32) DST OP (u32) IMM; \
1697 #define ALU(OPCODE, OP) \
1698 ALU64_##OPCODE##_X: \
1702 DST = (u32) DST OP (u32) SRC; \
1704 ALU64_##OPCODE##_K: \
1708 DST = (u32) DST OP (u32) IMM; \
1739 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1743 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1746 DST = (u64) (u32) (((s32) DST) >> IMM);
1749 (*(s64 *) &DST) >>= (SRC & 63);
1752 (*(s64 *) &DST) >>= IMM;
1755 div64_u64_rem(DST, SRC, &AX);
1760 DST = do_div(AX, (u32) SRC);
1763 div64_u64_rem(DST, IMM, &AX);
1768 DST = do_div(AX, (u32) IMM);
1771 DST = div64_u64(DST, SRC);
1775 do_div(AX, (u32) SRC);
1779 DST = div64_u64(DST, IMM);
1783 do_div(AX, (u32) IMM);
1789 DST = (__force u16) cpu_to_be16(DST);
1792 DST = (__force u32) cpu_to_be32(DST);
1795 DST = (__force u64) cpu_to_be64(DST);
1802 DST = (__force u16) cpu_to_le16(DST);
1805 DST = (__force u32) cpu_to_le32(DST);
1808 DST = (__force u64) cpu_to_le64(DST);
1815 /* Function call scratches BPF_R1-BPF_R5 registers,
1816 * preserves BPF_R6-BPF_R9, and stores return value
1819 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1824 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1827 insn + insn->off + 1);
1831 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1832 struct bpf_array *array = container_of(map, struct bpf_array, map);
1833 struct bpf_prog *prog;
1836 if (unlikely(index >= array->map.max_entries))
1839 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
1844 prog = READ_ONCE(array->ptrs[index]);
1848 /* ARG1 at this point is guaranteed to point to CTX from
1849 * the verifier side due to the fact that the tail call is
1850 * handled like a helper, that is, bpf_tail_call_proto,
1851 * where arg1_type is ARG_PTR_TO_CTX.
1853 insn = prog->insnsi;
1864 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1866 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1867 insn += insn->off; \
1871 JMP32_##OPCODE##_X: \
1872 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1873 insn += insn->off; \
1878 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1879 insn += insn->off; \
1883 JMP32_##OPCODE##_K: \
1884 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1885 insn += insn->off; \
1889 COND_JMP(u, JEQ, ==)
1890 COND_JMP(u, JNE, !=)
1893 COND_JMP(u, JGE, >=)
1894 COND_JMP(u, JLE, <=)
1895 COND_JMP(u, JSET, &)
1896 COND_JMP(s, JSGT, >)
1897 COND_JMP(s, JSLT, <)
1898 COND_JMP(s, JSGE, >=)
1899 COND_JMP(s, JSLE, <=)
1901 /* ST, STX and LDX*/
1903 /* Speculation barrier for mitigating Speculative Store Bypass.
1904 * In case of arm64, we rely on the firmware mitigation as
1905 * controlled via the ssbd kernel parameter. Whenever the
1906 * mitigation is enabled, it works for all of the kernel code
1907 * with no need to provide any additional instructions here.
1908 * In case of x86, we use 'lfence' insn for mitigation. We
1909 * reuse preexisting logic from Spectre v1 mitigation that
1910 * happens to produce the required code on x86 for v4 as well.
1914 #define LDST(SIZEOP, SIZE) \
1916 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1919 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1922 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1924 LDX_PROBE_MEM_##SIZEOP: \
1925 bpf_probe_read_kernel(&DST, sizeof(SIZE), \
1926 (const void *)(long) (SRC + insn->off)); \
1927 DST = *((SIZE *)&DST); \
1936 #define ATOMIC_ALU_OP(BOP, KOP) \
1938 if (BPF_SIZE(insn->code) == BPF_W) \
1939 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
1940 (DST + insn->off)); \
1942 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
1943 (DST + insn->off)); \
1945 case BOP | BPF_FETCH: \
1946 if (BPF_SIZE(insn->code) == BPF_W) \
1947 SRC = (u32) atomic_fetch_##KOP( \
1949 (atomic_t *)(unsigned long) (DST + insn->off)); \
1951 SRC = (u64) atomic64_fetch_##KOP( \
1953 (atomic64_t *)(unsigned long) (DST + insn->off)); \
1959 ATOMIC_ALU_OP(BPF_ADD, add)
1960 ATOMIC_ALU_OP(BPF_AND, and)
1961 ATOMIC_ALU_OP(BPF_OR, or)
1962 ATOMIC_ALU_OP(BPF_XOR, xor)
1963 #undef ATOMIC_ALU_OP
1966 if (BPF_SIZE(insn->code) == BPF_W)
1967 SRC = (u32) atomic_xchg(
1968 (atomic_t *)(unsigned long) (DST + insn->off),
1971 SRC = (u64) atomic64_xchg(
1972 (atomic64_t *)(unsigned long) (DST + insn->off),
1976 if (BPF_SIZE(insn->code) == BPF_W)
1977 BPF_R0 = (u32) atomic_cmpxchg(
1978 (atomic_t *)(unsigned long) (DST + insn->off),
1979 (u32) BPF_R0, (u32) SRC);
1981 BPF_R0 = (u64) atomic64_cmpxchg(
1982 (atomic64_t *)(unsigned long) (DST + insn->off),
1983 (u64) BPF_R0, (u64) SRC);
1992 /* If we ever reach this, we have a bug somewhere. Die hard here
1993 * instead of just returning 0; we could be somewhere in a subprog,
1994 * so execution could continue otherwise which we do /not/ want.
1996 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
1998 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
1999 insn->code, insn->imm);
2004 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2005 #define DEFINE_BPF_PROG_RUN(stack_size) \
2006 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2008 u64 stack[stack_size / sizeof(u64)]; \
2009 u64 regs[MAX_BPF_EXT_REG] = {}; \
2011 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2012 ARG1 = (u64) (unsigned long) ctx; \
2013 return ___bpf_prog_run(regs, insn); \
2016 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2017 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2018 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2019 const struct bpf_insn *insn) \
2021 u64 stack[stack_size / sizeof(u64)]; \
2022 u64 regs[MAX_BPF_EXT_REG]; \
2024 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2030 return ___bpf_prog_run(regs, insn); \
2033 #define EVAL1(FN, X) FN(X)
2034 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2035 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2036 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2037 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2038 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2040 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2041 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2042 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2044 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2045 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2046 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2048 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2050 static unsigned int (*interpreters[])(const void *ctx,
2051 const struct bpf_insn *insn) = {
2052 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2053 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2054 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2056 #undef PROG_NAME_LIST
2057 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2058 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2059 const struct bpf_insn *insn) = {
2060 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2061 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2062 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2064 #undef PROG_NAME_LIST
2066 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2068 stack_depth = max_t(u32, stack_depth, 1);
2069 insn->off = (s16) insn->imm;
2070 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2071 __bpf_call_base_args;
2072 insn->code = BPF_JMP | BPF_CALL_ARGS;
2076 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2077 const struct bpf_insn *insn)
2079 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2080 * is not working properly, so warn about it!
2087 bool bpf_prog_map_compatible(struct bpf_map *map,
2088 const struct bpf_prog *fp)
2090 enum bpf_prog_type prog_type = resolve_prog_type(fp);
2093 if (fp->kprobe_override)
2096 spin_lock(&map->owner.lock);
2097 if (!map->owner.type) {
2098 /* There's no owner yet where we could check for
2101 map->owner.type = prog_type;
2102 map->owner.jited = fp->jited;
2103 map->owner.xdp_has_frags = fp->aux->xdp_has_frags;
2106 ret = map->owner.type == prog_type &&
2107 map->owner.jited == fp->jited &&
2108 map->owner.xdp_has_frags == fp->aux->xdp_has_frags;
2110 spin_unlock(&map->owner.lock);
2115 static int bpf_check_tail_call(const struct bpf_prog *fp)
2117 struct bpf_prog_aux *aux = fp->aux;
2120 mutex_lock(&aux->used_maps_mutex);
2121 for (i = 0; i < aux->used_map_cnt; i++) {
2122 struct bpf_map *map = aux->used_maps[i];
2124 if (!map_type_contains_progs(map))
2127 if (!bpf_prog_map_compatible(map, fp)) {
2134 mutex_unlock(&aux->used_maps_mutex);
2138 static void bpf_prog_select_func(struct bpf_prog *fp)
2140 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2141 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2143 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2145 fp->bpf_func = __bpf_prog_ret0_warn;
2150 * bpf_prog_select_runtime - select exec runtime for BPF program
2151 * @fp: bpf_prog populated with BPF program
2152 * @err: pointer to error variable
2154 * Try to JIT eBPF program, if JIT is not available, use interpreter.
2155 * The BPF program will be executed via bpf_prog_run() function.
2157 * Return: the &fp argument along with &err set to 0 for success or
2158 * a negative errno code on failure
2160 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2162 /* In case of BPF to BPF calls, verifier did all the prep
2163 * work with regards to JITing, etc.
2165 bool jit_needed = false;
2170 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2171 bpf_prog_has_kfunc_call(fp))
2174 bpf_prog_select_func(fp);
2176 /* eBPF JITs can rewrite the program in case constant
2177 * blinding is active. However, in case of error during
2178 * blinding, bpf_int_jit_compile() must always return a
2179 * valid program, which in this case would simply not
2180 * be JITed, but falls back to the interpreter.
2182 if (!bpf_prog_is_dev_bound(fp->aux)) {
2183 *err = bpf_prog_alloc_jited_linfo(fp);
2187 fp = bpf_int_jit_compile(fp);
2188 bpf_prog_jit_attempt_done(fp);
2189 if (!fp->jited && jit_needed) {
2194 *err = bpf_prog_offload_compile(fp);
2200 bpf_prog_lock_ro(fp);
2202 /* The tail call compatibility check can only be done at
2203 * this late stage as we need to determine, if we deal
2204 * with JITed or non JITed program concatenations and not
2205 * all eBPF JITs might immediately support all features.
2207 *err = bpf_check_tail_call(fp);
2211 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2213 static unsigned int __bpf_prog_ret1(const void *ctx,
2214 const struct bpf_insn *insn)
2219 static struct bpf_prog_dummy {
2220 struct bpf_prog prog;
2221 } dummy_bpf_prog = {
2223 .bpf_func = __bpf_prog_ret1,
2227 struct bpf_empty_prog_array bpf_empty_prog_array = {
2230 EXPORT_SYMBOL(bpf_empty_prog_array);
2232 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2235 return kzalloc(sizeof(struct bpf_prog_array) +
2236 sizeof(struct bpf_prog_array_item) *
2240 return &bpf_empty_prog_array.hdr;
2243 void bpf_prog_array_free(struct bpf_prog_array *progs)
2245 if (!progs || progs == &bpf_empty_prog_array.hdr)
2247 kfree_rcu(progs, rcu);
2250 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2252 struct bpf_prog_array *progs;
2254 progs = container_of(rcu, struct bpf_prog_array, rcu);
2255 kfree_rcu(progs, rcu);
2258 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2260 if (!progs || progs == &bpf_empty_prog_array.hdr)
2262 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2265 int bpf_prog_array_length(struct bpf_prog_array *array)
2267 struct bpf_prog_array_item *item;
2270 for (item = array->items; item->prog; item++)
2271 if (item->prog != &dummy_bpf_prog.prog)
2276 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2278 struct bpf_prog_array_item *item;
2280 for (item = array->items; item->prog; item++)
2281 if (item->prog != &dummy_bpf_prog.prog)
2286 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2290 struct bpf_prog_array_item *item;
2293 for (item = array->items; item->prog; item++) {
2294 if (item->prog == &dummy_bpf_prog.prog)
2296 prog_ids[i] = item->prog->aux->id;
2297 if (++i == request_cnt) {
2303 return !!(item->prog);
2306 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2307 __u32 __user *prog_ids, u32 cnt)
2309 unsigned long err = 0;
2313 /* users of this function are doing:
2314 * cnt = bpf_prog_array_length();
2316 * bpf_prog_array_copy_to_user(..., cnt);
2317 * so below kcalloc doesn't need extra cnt > 0 check.
2319 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2322 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2323 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2332 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2333 struct bpf_prog *old_prog)
2335 struct bpf_prog_array_item *item;
2337 for (item = array->items; item->prog; item++)
2338 if (item->prog == old_prog) {
2339 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2345 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2346 * index into the program array with
2347 * a dummy no-op program.
2348 * @array: a bpf_prog_array
2349 * @index: the index of the program to replace
2351 * Skips over dummy programs, by not counting them, when calculating
2352 * the position of the program to replace.
2356 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2357 * * -ENOENT - Index out of range
2359 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2361 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2365 * bpf_prog_array_update_at() - Updates the program at the given index
2366 * into the program array.
2367 * @array: a bpf_prog_array
2368 * @index: the index of the program to update
2369 * @prog: the program to insert into the array
2371 * Skips over dummy programs, by not counting them, when calculating
2372 * the position of the program to update.
2376 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2377 * * -ENOENT - Index out of range
2379 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2380 struct bpf_prog *prog)
2382 struct bpf_prog_array_item *item;
2384 if (unlikely(index < 0))
2387 for (item = array->items; item->prog; item++) {
2388 if (item->prog == &dummy_bpf_prog.prog)
2391 WRITE_ONCE(item->prog, prog);
2399 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2400 struct bpf_prog *exclude_prog,
2401 struct bpf_prog *include_prog,
2403 struct bpf_prog_array **new_array)
2405 int new_prog_cnt, carry_prog_cnt = 0;
2406 struct bpf_prog_array_item *existing, *new;
2407 struct bpf_prog_array *array;
2408 bool found_exclude = false;
2410 /* Figure out how many existing progs we need to carry over to
2414 existing = old_array->items;
2415 for (; existing->prog; existing++) {
2416 if (existing->prog == exclude_prog) {
2417 found_exclude = true;
2420 if (existing->prog != &dummy_bpf_prog.prog)
2422 if (existing->prog == include_prog)
2427 if (exclude_prog && !found_exclude)
2430 /* How many progs (not NULL) will be in the new array? */
2431 new_prog_cnt = carry_prog_cnt;
2435 /* Do we have any prog (not NULL) in the new array? */
2436 if (!new_prog_cnt) {
2441 /* +1 as the end of prog_array is marked with NULL */
2442 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2447 /* Fill in the new prog array */
2448 if (carry_prog_cnt) {
2449 existing = old_array->items;
2450 for (; existing->prog; existing++) {
2451 if (existing->prog == exclude_prog ||
2452 existing->prog == &dummy_bpf_prog.prog)
2455 new->prog = existing->prog;
2456 new->bpf_cookie = existing->bpf_cookie;
2461 new->prog = include_prog;
2462 new->bpf_cookie = bpf_cookie;
2470 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2471 u32 *prog_ids, u32 request_cnt,
2477 cnt = bpf_prog_array_length(array);
2481 /* return early if user requested only program count or nothing to copy */
2482 if (!request_cnt || !cnt)
2485 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2486 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2490 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2491 struct bpf_map **used_maps, u32 len)
2493 struct bpf_map *map;
2496 for (i = 0; i < len; i++) {
2498 if (map->ops->map_poke_untrack)
2499 map->ops->map_poke_untrack(map, aux);
2504 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2506 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2507 kfree(aux->used_maps);
2510 void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2511 struct btf_mod_pair *used_btfs, u32 len)
2513 #ifdef CONFIG_BPF_SYSCALL
2514 struct btf_mod_pair *btf_mod;
2517 for (i = 0; i < len; i++) {
2518 btf_mod = &used_btfs[i];
2519 if (btf_mod->module)
2520 module_put(btf_mod->module);
2521 btf_put(btf_mod->btf);
2526 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2528 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2529 kfree(aux->used_btfs);
2532 static void bpf_prog_free_deferred(struct work_struct *work)
2534 struct bpf_prog_aux *aux;
2537 aux = container_of(work, struct bpf_prog_aux, work);
2538 #ifdef CONFIG_BPF_SYSCALL
2539 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2541 #ifdef CONFIG_CGROUP_BPF
2542 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2543 bpf_cgroup_atype_put(aux->cgroup_atype);
2545 bpf_free_used_maps(aux);
2546 bpf_free_used_btfs(aux);
2547 if (bpf_prog_is_dev_bound(aux))
2548 bpf_prog_offload_destroy(aux->prog);
2549 #ifdef CONFIG_PERF_EVENTS
2550 if (aux->prog->has_callchain_buf)
2551 put_callchain_buffers();
2553 if (aux->dst_trampoline)
2554 bpf_trampoline_put(aux->dst_trampoline);
2555 for (i = 0; i < aux->func_cnt; i++) {
2556 /* We can just unlink the subprog poke descriptor table as
2557 * it was originally linked to the main program and is also
2558 * released along with it.
2560 aux->func[i]->aux->poke_tab = NULL;
2561 bpf_jit_free(aux->func[i]);
2563 if (aux->func_cnt) {
2565 bpf_prog_unlock_free(aux->prog);
2567 bpf_jit_free(aux->prog);
2571 void bpf_prog_free(struct bpf_prog *fp)
2573 struct bpf_prog_aux *aux = fp->aux;
2576 bpf_prog_put(aux->dst_prog);
2577 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2578 schedule_work(&aux->work);
2580 EXPORT_SYMBOL_GPL(bpf_prog_free);
2582 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2583 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2585 void bpf_user_rnd_init_once(void)
2587 prandom_init_once(&bpf_user_rnd_state);
2590 BPF_CALL_0(bpf_user_rnd_u32)
2592 /* Should someone ever have the rather unwise idea to use some
2593 * of the registers passed into this function, then note that
2594 * this function is called from native eBPF and classic-to-eBPF
2595 * transformations. Register assignments from both sides are
2596 * different, f.e. classic always sets fn(ctx, A, X) here.
2598 struct rnd_state *state;
2601 state = &get_cpu_var(bpf_user_rnd_state);
2602 res = prandom_u32_state(state);
2603 put_cpu_var(bpf_user_rnd_state);
2608 BPF_CALL_0(bpf_get_raw_cpu_id)
2610 return raw_smp_processor_id();
2613 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2614 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2615 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2616 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2617 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2618 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2619 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2620 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2621 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2622 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2623 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2625 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2626 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2627 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2628 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2629 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2630 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2631 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak;
2633 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2634 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2635 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2636 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2637 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2638 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2639 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2640 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2641 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2642 const struct bpf_func_proto bpf_set_retval_proto __weak;
2643 const struct bpf_func_proto bpf_get_retval_proto __weak;
2645 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2650 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2656 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2657 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2661 EXPORT_SYMBOL_GPL(bpf_event_output);
2663 /* Always built-in helper functions. */
2664 const struct bpf_func_proto bpf_tail_call_proto = {
2667 .ret_type = RET_VOID,
2668 .arg1_type = ARG_PTR_TO_CTX,
2669 .arg2_type = ARG_CONST_MAP_PTR,
2670 .arg3_type = ARG_ANYTHING,
2673 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2674 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2675 * eBPF and implicitly also cBPF can get JITed!
2677 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2682 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2683 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2685 void __weak bpf_jit_compile(struct bpf_prog *prog)
2689 bool __weak bpf_helper_changes_pkt_data(void *func)
2694 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2695 * analysis code and wants explicit zero extension inserted by verifier.
2696 * Otherwise, return FALSE.
2698 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2699 * you don't override this. JITs that don't want these extra insns can detect
2700 * them using insn_is_zext.
2702 bool __weak bpf_jit_needs_zext(void)
2707 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2708 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2713 bool __weak bpf_jit_supports_kfunc_call(void)
2718 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2719 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2721 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2727 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2728 void *addr1, void *addr2)
2733 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
2735 return ERR_PTR(-ENOTSUPP);
2738 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
2743 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2744 EXPORT_SYMBOL(bpf_stats_enabled_key);
2746 /* All definitions of tracepoints related to BPF. */
2747 #define CREATE_TRACE_POINTS
2748 #include <linux/bpf_trace.h>
2750 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2751 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);