2 * Linux Socket Filter - Kernel level socket filtering
4 * Based on the design of the Berkeley Packet Filter. The new
5 * internal format has been designed by PLUMgrid:
7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
11 * Jay Schulist <jschlst@samba.org>
12 * Alexei Starovoitov <ast@plumgrid.com>
13 * Daniel Borkmann <dborkman@redhat.com>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
20 * Andi Kleen - Fix a few bad bugs and races.
21 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
24 #include <uapi/linux/btf.h>
25 #include <linux/filter.h>
26 #include <linux/skbuff.h>
27 #include <linux/vmalloc.h>
28 #include <linux/random.h>
29 #include <linux/moduleloader.h>
30 #include <linux/bpf.h>
31 #include <linux/btf.h>
32 #include <linux/frame.h>
33 #include <linux/rbtree_latch.h>
34 #include <linux/kallsyms.h>
35 #include <linux/rcupdate.h>
36 #include <linux/perf_event.h>
38 #include <asm/unaligned.h>
41 #define BPF_R0 regs[BPF_REG_0]
42 #define BPF_R1 regs[BPF_REG_1]
43 #define BPF_R2 regs[BPF_REG_2]
44 #define BPF_R3 regs[BPF_REG_3]
45 #define BPF_R4 regs[BPF_REG_4]
46 #define BPF_R5 regs[BPF_REG_5]
47 #define BPF_R6 regs[BPF_REG_6]
48 #define BPF_R7 regs[BPF_REG_7]
49 #define BPF_R8 regs[BPF_REG_8]
50 #define BPF_R9 regs[BPF_REG_9]
51 #define BPF_R10 regs[BPF_REG_10]
54 #define DST regs[insn->dst_reg]
55 #define SRC regs[insn->src_reg]
56 #define FP regs[BPF_REG_FP]
57 #define AX regs[BPF_REG_AX]
58 #define ARG1 regs[BPF_REG_ARG1]
59 #define CTX regs[BPF_REG_CTX]
62 /* No hurry in this branch
64 * Exported for the bpf jit load helper.
66 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
71 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
72 else if (k >= SKF_LL_OFF)
73 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
75 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
81 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
83 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
84 struct bpf_prog_aux *aux;
87 size = round_up(size, PAGE_SIZE);
88 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
92 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
98 fp->pages = size / PAGE_SIZE;
101 fp->jit_requested = ebpf_jit_enabled();
103 INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode);
108 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
110 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
111 struct bpf_prog *prog;
114 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
118 prog->aux->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
119 if (!prog->aux->stats) {
125 for_each_possible_cpu(cpu) {
126 struct bpf_prog_stats *pstats;
128 pstats = per_cpu_ptr(prog->aux->stats, cpu);
129 u64_stats_init(&pstats->syncp);
133 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
135 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
137 if (!prog->aux->nr_linfo || !prog->jit_requested)
140 prog->aux->jited_linfo = kcalloc(prog->aux->nr_linfo,
141 sizeof(*prog->aux->jited_linfo),
142 GFP_KERNEL | __GFP_NOWARN);
143 if (!prog->aux->jited_linfo)
149 void bpf_prog_free_jited_linfo(struct bpf_prog *prog)
151 kfree(prog->aux->jited_linfo);
152 prog->aux->jited_linfo = NULL;
155 void bpf_prog_free_unused_jited_linfo(struct bpf_prog *prog)
157 if (prog->aux->jited_linfo && !prog->aux->jited_linfo[0])
158 bpf_prog_free_jited_linfo(prog);
161 /* The jit engine is responsible to provide an array
162 * for insn_off to the jited_off mapping (insn_to_jit_off).
164 * The idx to this array is the insn_off. Hence, the insn_off
165 * here is relative to the prog itself instead of the main prog.
166 * This array has one entry for each xlated bpf insn.
168 * jited_off is the byte off to the last byte of the jited insn.
172 * The first bpf insn off of the prog. The insn off
173 * here is relative to the main prog.
174 * e.g. if prog is a subprog, insn_start > 0
176 * The prog's idx to prog->aux->linfo and jited_linfo
178 * jited_linfo[linfo_idx] = prog->bpf_func
182 * jited_linfo[i] = prog->bpf_func +
183 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
185 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
186 const u32 *insn_to_jit_off)
188 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
189 const struct bpf_line_info *linfo;
192 if (!prog->aux->jited_linfo)
193 /* Userspace did not provide linfo */
196 linfo_idx = prog->aux->linfo_idx;
197 linfo = &prog->aux->linfo[linfo_idx];
198 insn_start = linfo[0].insn_off;
199 insn_end = insn_start + prog->len;
201 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
202 jited_linfo[0] = prog->bpf_func;
204 nr_linfo = prog->aux->nr_linfo - linfo_idx;
206 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
207 /* The verifier ensures that linfo[i].insn_off is
208 * strictly increasing
210 jited_linfo[i] = prog->bpf_func +
211 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
214 void bpf_prog_free_linfo(struct bpf_prog *prog)
216 bpf_prog_free_jited_linfo(prog);
217 kvfree(prog->aux->linfo);
220 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
221 gfp_t gfp_extra_flags)
223 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
228 BUG_ON(fp_old == NULL);
230 size = round_up(size, PAGE_SIZE);
231 pages = size / PAGE_SIZE;
232 if (pages <= fp_old->pages)
235 delta = pages - fp_old->pages;
236 ret = __bpf_prog_charge(fp_old->aux->user, delta);
240 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
242 __bpf_prog_uncharge(fp_old->aux->user, delta);
244 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
248 /* We keep fp->aux from fp_old around in the new
249 * reallocated structure.
252 __bpf_prog_free(fp_old);
258 void __bpf_prog_free(struct bpf_prog *fp)
261 free_percpu(fp->aux->stats);
267 int bpf_prog_calc_tag(struct bpf_prog *fp)
269 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
270 u32 raw_size = bpf_prog_tag_scratch_size(fp);
271 u32 digest[SHA_DIGEST_WORDS];
272 u32 ws[SHA_WORKSPACE_WORDS];
273 u32 i, bsize, psize, blocks;
274 struct bpf_insn *dst;
280 raw = vmalloc(raw_size);
285 memset(ws, 0, sizeof(ws));
287 /* We need to take out the map fd for the digest calculation
288 * since they are unstable from user space side.
291 for (i = 0, was_ld_map = false; i < fp->len; i++) {
292 dst[i] = fp->insnsi[i];
294 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
295 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
296 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
299 } else if (was_ld_map &&
301 dst[i].dst_reg == 0 &&
302 dst[i].src_reg == 0 &&
311 psize = bpf_prog_insn_size(fp);
312 memset(&raw[psize], 0, raw_size - psize);
315 bsize = round_up(psize, SHA_MESSAGE_BYTES);
316 blocks = bsize / SHA_MESSAGE_BYTES;
318 if (bsize - psize >= sizeof(__be64)) {
319 bits = (__be64 *)(todo + bsize - sizeof(__be64));
321 bits = (__be64 *)(todo + bsize + bits_offset);
324 *bits = cpu_to_be64((psize - 1) << 3);
327 sha_transform(digest, todo, ws);
328 todo += SHA_MESSAGE_BYTES;
331 result = (__force __be32 *)digest;
332 for (i = 0; i < SHA_DIGEST_WORDS; i++)
333 result[i] = cpu_to_be32(digest[i]);
334 memcpy(fp->tag, result, sizeof(fp->tag));
340 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
341 s32 end_new, s32 curr, const bool probe_pass)
343 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
344 s32 delta = end_new - end_old;
347 if (curr < pos && curr + imm + 1 >= end_old)
349 else if (curr >= end_new && curr + imm + 1 < end_new)
351 if (imm < imm_min || imm > imm_max)
358 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
359 s32 end_new, s32 curr, const bool probe_pass)
361 const s32 off_min = S16_MIN, off_max = S16_MAX;
362 s32 delta = end_new - end_old;
365 if (curr < pos && curr + off + 1 >= end_old)
367 else if (curr >= end_new && curr + off + 1 < end_new)
369 if (off < off_min || off > off_max)
376 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
377 s32 end_new, const bool probe_pass)
379 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
380 struct bpf_insn *insn = prog->insnsi;
383 for (i = 0; i < insn_cnt; i++, insn++) {
386 /* In the probing pass we still operate on the original,
387 * unpatched image in order to check overflows before we
388 * do any other adjustments. Therefore skip the patchlet.
390 if (probe_pass && i == pos) {
392 insn = prog->insnsi + end_old;
395 if ((BPF_CLASS(code) != BPF_JMP &&
396 BPF_CLASS(code) != BPF_JMP32) ||
397 BPF_OP(code) == BPF_EXIT)
399 /* Adjust offset of jmps if we cross patch boundaries. */
400 if (BPF_OP(code) == BPF_CALL) {
401 if (insn->src_reg != BPF_PSEUDO_CALL)
403 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
404 end_new, i, probe_pass);
406 ret = bpf_adj_delta_to_off(insn, pos, end_old,
407 end_new, i, probe_pass);
416 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
418 struct bpf_line_info *linfo;
421 nr_linfo = prog->aux->nr_linfo;
422 if (!nr_linfo || !delta)
425 linfo = prog->aux->linfo;
427 for (i = 0; i < nr_linfo; i++)
428 if (off < linfo[i].insn_off)
431 /* Push all off < linfo[i].insn_off by delta */
432 for (; i < nr_linfo; i++)
433 linfo[i].insn_off += delta;
436 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
437 const struct bpf_insn *patch, u32 len)
439 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
440 const u32 cnt_max = S16_MAX;
441 struct bpf_prog *prog_adj;
444 /* Since our patchlet doesn't expand the image, we're done. */
445 if (insn_delta == 0) {
446 memcpy(prog->insnsi + off, patch, sizeof(*patch));
450 insn_adj_cnt = prog->len + insn_delta;
452 /* Reject anything that would potentially let the insn->off
453 * target overflow when we have excessive program expansions.
454 * We need to probe here before we do any reallocation where
455 * we afterwards may not fail anymore.
457 if (insn_adj_cnt > cnt_max &&
458 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
461 /* Several new instructions need to be inserted. Make room
462 * for them. Likely, there's no need for a new allocation as
463 * last page could have large enough tailroom.
465 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
468 return ERR_PTR(-ENOMEM);
470 prog_adj->len = insn_adj_cnt;
472 /* Patching happens in 3 steps:
474 * 1) Move over tail of insnsi from next instruction onwards,
475 * so we can patch the single target insn with one or more
476 * new ones (patching is always from 1 to n insns, n > 0).
477 * 2) Inject new instructions at the target location.
478 * 3) Adjust branch offsets if necessary.
480 insn_rest = insn_adj_cnt - off - len;
482 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
483 sizeof(*patch) * insn_rest);
484 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
486 /* We are guaranteed to not fail at this point, otherwise
487 * the ship has sailed to reverse to the original state. An
488 * overflow cannot happen at this point.
490 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
492 bpf_adj_linfo(prog_adj, off, insn_delta);
497 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
499 /* Branch offsets can't overflow when program is shrinking, no need
500 * to call bpf_adj_branches(..., true) here
502 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
503 sizeof(struct bpf_insn) * (prog->len - off - cnt));
506 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
509 void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
513 for (i = 0; i < fp->aux->func_cnt; i++)
514 bpf_prog_kallsyms_del(fp->aux->func[i]);
517 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
519 bpf_prog_kallsyms_del_subprogs(fp);
520 bpf_prog_kallsyms_del(fp);
523 #ifdef CONFIG_BPF_JIT
524 /* All BPF JIT sysctl knobs here. */
525 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_ALWAYS_ON);
526 int bpf_jit_harden __read_mostly;
527 int bpf_jit_kallsyms __read_mostly;
528 long bpf_jit_limit __read_mostly;
530 static __always_inline void
531 bpf_get_prog_addr_region(const struct bpf_prog *prog,
532 unsigned long *symbol_start,
533 unsigned long *symbol_end)
535 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
536 unsigned long addr = (unsigned long)hdr;
538 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
540 *symbol_start = addr;
541 *symbol_end = addr + hdr->pages * PAGE_SIZE;
544 void bpf_get_prog_name(const struct bpf_prog *prog, char *sym)
546 const char *end = sym + KSYM_NAME_LEN;
547 const struct btf_type *type;
548 const char *func_name;
550 BUILD_BUG_ON(sizeof("bpf_prog_") +
551 sizeof(prog->tag) * 2 +
552 /* name has been null terminated.
553 * We should need +1 for the '_' preceding
554 * the name. However, the null character
555 * is double counted between the name and the
556 * sizeof("bpf_prog_") above, so we omit
559 sizeof(prog->aux->name) > KSYM_NAME_LEN);
561 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
562 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
564 /* prog->aux->name will be ignored if full btf name is available */
565 if (prog->aux->func_info_cnt) {
566 type = btf_type_by_id(prog->aux->btf,
567 prog->aux->func_info[prog->aux->func_idx].type_id);
568 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
569 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
573 if (prog->aux->name[0])
574 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
579 static __always_inline unsigned long
580 bpf_get_prog_addr_start(struct latch_tree_node *n)
582 unsigned long symbol_start, symbol_end;
583 const struct bpf_prog_aux *aux;
585 aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
586 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
591 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
592 struct latch_tree_node *b)
594 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b);
597 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
599 unsigned long val = (unsigned long)key;
600 unsigned long symbol_start, symbol_end;
601 const struct bpf_prog_aux *aux;
603 aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
604 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
606 if (val < symbol_start)
608 if (val >= symbol_end)
614 static const struct latch_tree_ops bpf_tree_ops = {
615 .less = bpf_tree_less,
616 .comp = bpf_tree_comp,
619 static DEFINE_SPINLOCK(bpf_lock);
620 static LIST_HEAD(bpf_kallsyms);
621 static struct latch_tree_root bpf_tree __cacheline_aligned;
623 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux)
625 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode));
626 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms);
627 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
630 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux)
632 if (list_empty(&aux->ksym_lnode))
635 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
636 list_del_rcu(&aux->ksym_lnode);
639 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
641 return fp->jited && !bpf_prog_was_classic(fp);
644 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
646 return list_empty(&fp->aux->ksym_lnode) ||
647 fp->aux->ksym_lnode.prev == LIST_POISON2;
650 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
652 if (!bpf_prog_kallsyms_candidate(fp) ||
653 !capable(CAP_SYS_ADMIN))
656 spin_lock_bh(&bpf_lock);
657 bpf_prog_ksym_node_add(fp->aux);
658 spin_unlock_bh(&bpf_lock);
661 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
663 if (!bpf_prog_kallsyms_candidate(fp))
666 spin_lock_bh(&bpf_lock);
667 bpf_prog_ksym_node_del(fp->aux);
668 spin_unlock_bh(&bpf_lock);
671 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr)
673 struct latch_tree_node *n;
675 if (!bpf_jit_kallsyms_enabled())
678 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
680 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog :
684 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
685 unsigned long *off, char *sym)
687 unsigned long symbol_start, symbol_end;
688 struct bpf_prog *prog;
692 prog = bpf_prog_kallsyms_find(addr);
694 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end);
695 bpf_get_prog_name(prog, sym);
699 *size = symbol_end - symbol_start;
701 *off = addr - symbol_start;
708 bool is_bpf_text_address(unsigned long addr)
713 ret = bpf_prog_kallsyms_find(addr) != NULL;
719 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
722 struct bpf_prog_aux *aux;
726 if (!bpf_jit_kallsyms_enabled())
730 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) {
734 bpf_get_prog_name(aux->prog, sym);
736 *value = (unsigned long)aux->prog->bpf_func;
737 *type = BPF_SYM_ELF_TYPE;
747 static atomic_long_t bpf_jit_current;
749 /* Can be overridden by an arch's JIT compiler if it has a custom,
750 * dedicated BPF backend memory area, or if neither of the two
753 u64 __weak bpf_jit_alloc_exec_limit(void)
755 #if defined(MODULES_VADDR)
756 return MODULES_END - MODULES_VADDR;
758 return VMALLOC_END - VMALLOC_START;
762 static int __init bpf_jit_charge_init(void)
764 /* Only used as heuristic here to derive limit. */
765 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2,
766 PAGE_SIZE), LONG_MAX);
769 pure_initcall(bpf_jit_charge_init);
771 static int bpf_jit_charge_modmem(u32 pages)
773 if (atomic_long_add_return(pages, &bpf_jit_current) >
774 (bpf_jit_limit >> PAGE_SHIFT)) {
775 if (!capable(CAP_SYS_ADMIN)) {
776 atomic_long_sub(pages, &bpf_jit_current);
784 static void bpf_jit_uncharge_modmem(u32 pages)
786 atomic_long_sub(pages, &bpf_jit_current);
789 void *__weak bpf_jit_alloc_exec(unsigned long size)
791 return module_alloc(size);
794 void __weak bpf_jit_free_exec(void *addr)
796 module_memfree(addr);
799 struct bpf_binary_header *
800 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
801 unsigned int alignment,
802 bpf_jit_fill_hole_t bpf_fill_ill_insns)
804 struct bpf_binary_header *hdr;
805 u32 size, hole, start, pages;
807 /* Most of BPF filters are really small, but if some of them
808 * fill a page, allow at least 128 extra bytes to insert a
809 * random section of illegal instructions.
811 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
812 pages = size / PAGE_SIZE;
814 if (bpf_jit_charge_modmem(pages))
816 hdr = bpf_jit_alloc_exec(size);
818 bpf_jit_uncharge_modmem(pages);
822 /* Fill space with illegal/arch-dep instructions. */
823 bpf_fill_ill_insns(hdr, size);
826 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
827 PAGE_SIZE - sizeof(*hdr));
828 start = (get_random_int() % hole) & ~(alignment - 1);
830 /* Leave a random number of instructions before BPF code. */
831 *image_ptr = &hdr->image[start];
836 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
838 u32 pages = hdr->pages;
840 bpf_jit_free_exec(hdr);
841 bpf_jit_uncharge_modmem(pages);
844 /* This symbol is only overridden by archs that have different
845 * requirements than the usual eBPF JITs, f.e. when they only
846 * implement cBPF JIT, do not set images read-only, etc.
848 void __weak bpf_jit_free(struct bpf_prog *fp)
851 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
853 bpf_jit_binary_free(hdr);
855 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
858 bpf_prog_unlock_free(fp);
861 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
862 const struct bpf_insn *insn, bool extra_pass,
863 u64 *func_addr, bool *func_addr_fixed)
869 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
870 if (!*func_addr_fixed) {
871 /* Place-holder address till the last pass has collected
872 * all addresses for JITed subprograms in which case we
873 * can pick them up from prog->aux.
877 else if (prog->aux->func &&
878 off >= 0 && off < prog->aux->func_cnt)
879 addr = (u8 *)prog->aux->func[off]->bpf_func;
883 /* Address of a BPF helper call. Since part of the core
884 * kernel, it's always at a fixed location. __bpf_call_base
885 * and the helper with imm relative to it are both in core
888 addr = (u8 *)__bpf_call_base + imm;
891 *func_addr = (unsigned long)addr;
895 static int bpf_jit_blind_insn(const struct bpf_insn *from,
896 const struct bpf_insn *aux,
897 struct bpf_insn *to_buff)
899 struct bpf_insn *to = to_buff;
900 u32 imm_rnd = get_random_int();
903 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
904 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
906 /* Constraints on AX register:
908 * AX register is inaccessible from user space. It is mapped in
909 * all JITs, and used here for constant blinding rewrites. It is
910 * typically "stateless" meaning its contents are only valid within
911 * the executed instruction, but not across several instructions.
912 * There are a few exceptions however which are further detailed
915 * Constant blinding is only used by JITs, not in the interpreter.
916 * The interpreter uses AX in some occasions as a local temporary
917 * register e.g. in DIV or MOD instructions.
919 * In restricted circumstances, the verifier can also use the AX
920 * register for rewrites as long as they do not interfere with
923 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
926 if (from->imm == 0 &&
927 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
928 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
929 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
933 switch (from->code) {
934 case BPF_ALU | BPF_ADD | BPF_K:
935 case BPF_ALU | BPF_SUB | BPF_K:
936 case BPF_ALU | BPF_AND | BPF_K:
937 case BPF_ALU | BPF_OR | BPF_K:
938 case BPF_ALU | BPF_XOR | BPF_K:
939 case BPF_ALU | BPF_MUL | BPF_K:
940 case BPF_ALU | BPF_MOV | BPF_K:
941 case BPF_ALU | BPF_DIV | BPF_K:
942 case BPF_ALU | BPF_MOD | BPF_K:
943 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
944 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
945 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
948 case BPF_ALU64 | BPF_ADD | BPF_K:
949 case BPF_ALU64 | BPF_SUB | BPF_K:
950 case BPF_ALU64 | BPF_AND | BPF_K:
951 case BPF_ALU64 | BPF_OR | BPF_K:
952 case BPF_ALU64 | BPF_XOR | BPF_K:
953 case BPF_ALU64 | BPF_MUL | BPF_K:
954 case BPF_ALU64 | BPF_MOV | BPF_K:
955 case BPF_ALU64 | BPF_DIV | BPF_K:
956 case BPF_ALU64 | BPF_MOD | BPF_K:
957 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
958 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
959 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
962 case BPF_JMP | BPF_JEQ | BPF_K:
963 case BPF_JMP | BPF_JNE | BPF_K:
964 case BPF_JMP | BPF_JGT | BPF_K:
965 case BPF_JMP | BPF_JLT | BPF_K:
966 case BPF_JMP | BPF_JGE | BPF_K:
967 case BPF_JMP | BPF_JLE | BPF_K:
968 case BPF_JMP | BPF_JSGT | BPF_K:
969 case BPF_JMP | BPF_JSLT | BPF_K:
970 case BPF_JMP | BPF_JSGE | BPF_K:
971 case BPF_JMP | BPF_JSLE | BPF_K:
972 case BPF_JMP | BPF_JSET | BPF_K:
973 /* Accommodate for extra offset in case of a backjump. */
977 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
978 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
979 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
982 case BPF_JMP32 | BPF_JEQ | BPF_K:
983 case BPF_JMP32 | BPF_JNE | BPF_K:
984 case BPF_JMP32 | BPF_JGT | BPF_K:
985 case BPF_JMP32 | BPF_JLT | BPF_K:
986 case BPF_JMP32 | BPF_JGE | BPF_K:
987 case BPF_JMP32 | BPF_JLE | BPF_K:
988 case BPF_JMP32 | BPF_JSGT | BPF_K:
989 case BPF_JMP32 | BPF_JSLT | BPF_K:
990 case BPF_JMP32 | BPF_JSGE | BPF_K:
991 case BPF_JMP32 | BPF_JSLE | BPF_K:
992 case BPF_JMP32 | BPF_JSET | BPF_K:
993 /* Accommodate for extra offset in case of a backjump. */
997 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
998 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
999 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1003 case BPF_LD | BPF_IMM | BPF_DW:
1004 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1005 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1006 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1007 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1009 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1010 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1011 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1012 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1015 case BPF_ST | BPF_MEM | BPF_DW:
1016 case BPF_ST | BPF_MEM | BPF_W:
1017 case BPF_ST | BPF_MEM | BPF_H:
1018 case BPF_ST | BPF_MEM | BPF_B:
1019 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1020 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1021 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1025 return to - to_buff;
1028 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1029 gfp_t gfp_extra_flags)
1031 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1032 struct bpf_prog *fp;
1034 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
1036 /* aux->prog still points to the fp_other one, so
1037 * when promoting the clone to the real program,
1038 * this still needs to be adapted.
1040 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1046 static void bpf_prog_clone_free(struct bpf_prog *fp)
1048 /* aux was stolen by the other clone, so we cannot free
1049 * it from this path! It will be freed eventually by the
1050 * other program on release.
1052 * At this point, we don't need a deferred release since
1053 * clone is guaranteed to not be locked.
1056 __bpf_prog_free(fp);
1059 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1061 /* We have to repoint aux->prog to self, as we don't
1062 * know whether fp here is the clone or the original.
1065 bpf_prog_clone_free(fp_other);
1068 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1070 struct bpf_insn insn_buff[16], aux[2];
1071 struct bpf_prog *clone, *tmp;
1072 int insn_delta, insn_cnt;
1073 struct bpf_insn *insn;
1076 if (!bpf_jit_blinding_enabled(prog) || prog->blinded)
1079 clone = bpf_prog_clone_create(prog, GFP_USER);
1081 return ERR_PTR(-ENOMEM);
1083 insn_cnt = clone->len;
1084 insn = clone->insnsi;
1086 for (i = 0; i < insn_cnt; i++, insn++) {
1087 /* We temporarily need to hold the original ld64 insn
1088 * so that we can still access the first part in the
1089 * second blinding run.
1091 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1093 memcpy(aux, insn, sizeof(aux));
1095 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
1099 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1101 /* Patching may have repointed aux->prog during
1102 * realloc from the original one, so we need to
1103 * fix it up here on error.
1105 bpf_jit_prog_release_other(prog, clone);
1110 insn_delta = rewritten - 1;
1112 /* Walk new program and skip insns we just inserted. */
1113 insn = clone->insnsi + i + insn_delta;
1114 insn_cnt += insn_delta;
1121 #endif /* CONFIG_BPF_JIT */
1123 /* Base function for offset calculation. Needs to go into .text section,
1124 * therefore keeping it non-static as well; will also be used by JITs
1125 * anyway later on, so do not let the compiler omit it. This also needs
1126 * to go into kallsyms for correlation from e.g. bpftool, so naming
1129 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1133 EXPORT_SYMBOL_GPL(__bpf_call_base);
1135 /* All UAPI available opcodes. */
1136 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1137 /* 32 bit ALU operations. */ \
1138 /* Register based. */ \
1139 INSN_3(ALU, ADD, X), \
1140 INSN_3(ALU, SUB, X), \
1141 INSN_3(ALU, AND, X), \
1142 INSN_3(ALU, OR, X), \
1143 INSN_3(ALU, LSH, X), \
1144 INSN_3(ALU, RSH, X), \
1145 INSN_3(ALU, XOR, X), \
1146 INSN_3(ALU, MUL, X), \
1147 INSN_3(ALU, MOV, X), \
1148 INSN_3(ALU, ARSH, X), \
1149 INSN_3(ALU, DIV, X), \
1150 INSN_3(ALU, MOD, X), \
1152 INSN_3(ALU, END, TO_BE), \
1153 INSN_3(ALU, END, TO_LE), \
1154 /* Immediate based. */ \
1155 INSN_3(ALU, ADD, K), \
1156 INSN_3(ALU, SUB, K), \
1157 INSN_3(ALU, AND, K), \
1158 INSN_3(ALU, OR, K), \
1159 INSN_3(ALU, LSH, K), \
1160 INSN_3(ALU, RSH, K), \
1161 INSN_3(ALU, XOR, K), \
1162 INSN_3(ALU, MUL, K), \
1163 INSN_3(ALU, MOV, K), \
1164 INSN_3(ALU, ARSH, K), \
1165 INSN_3(ALU, DIV, K), \
1166 INSN_3(ALU, MOD, K), \
1167 /* 64 bit ALU operations. */ \
1168 /* Register based. */ \
1169 INSN_3(ALU64, ADD, X), \
1170 INSN_3(ALU64, SUB, X), \
1171 INSN_3(ALU64, AND, X), \
1172 INSN_3(ALU64, OR, X), \
1173 INSN_3(ALU64, LSH, X), \
1174 INSN_3(ALU64, RSH, X), \
1175 INSN_3(ALU64, XOR, X), \
1176 INSN_3(ALU64, MUL, X), \
1177 INSN_3(ALU64, MOV, X), \
1178 INSN_3(ALU64, ARSH, X), \
1179 INSN_3(ALU64, DIV, X), \
1180 INSN_3(ALU64, MOD, X), \
1181 INSN_2(ALU64, NEG), \
1182 /* Immediate based. */ \
1183 INSN_3(ALU64, ADD, K), \
1184 INSN_3(ALU64, SUB, K), \
1185 INSN_3(ALU64, AND, K), \
1186 INSN_3(ALU64, OR, K), \
1187 INSN_3(ALU64, LSH, K), \
1188 INSN_3(ALU64, RSH, K), \
1189 INSN_3(ALU64, XOR, K), \
1190 INSN_3(ALU64, MUL, K), \
1191 INSN_3(ALU64, MOV, K), \
1192 INSN_3(ALU64, ARSH, K), \
1193 INSN_3(ALU64, DIV, K), \
1194 INSN_3(ALU64, MOD, K), \
1195 /* Call instruction. */ \
1196 INSN_2(JMP, CALL), \
1197 /* Exit instruction. */ \
1198 INSN_2(JMP, EXIT), \
1199 /* 32-bit Jump instructions. */ \
1200 /* Register based. */ \
1201 INSN_3(JMP32, JEQ, X), \
1202 INSN_3(JMP32, JNE, X), \
1203 INSN_3(JMP32, JGT, X), \
1204 INSN_3(JMP32, JLT, X), \
1205 INSN_3(JMP32, JGE, X), \
1206 INSN_3(JMP32, JLE, X), \
1207 INSN_3(JMP32, JSGT, X), \
1208 INSN_3(JMP32, JSLT, X), \
1209 INSN_3(JMP32, JSGE, X), \
1210 INSN_3(JMP32, JSLE, X), \
1211 INSN_3(JMP32, JSET, X), \
1212 /* Immediate based. */ \
1213 INSN_3(JMP32, JEQ, K), \
1214 INSN_3(JMP32, JNE, K), \
1215 INSN_3(JMP32, JGT, K), \
1216 INSN_3(JMP32, JLT, K), \
1217 INSN_3(JMP32, JGE, K), \
1218 INSN_3(JMP32, JLE, K), \
1219 INSN_3(JMP32, JSGT, K), \
1220 INSN_3(JMP32, JSLT, K), \
1221 INSN_3(JMP32, JSGE, K), \
1222 INSN_3(JMP32, JSLE, K), \
1223 INSN_3(JMP32, JSET, K), \
1224 /* Jump instructions. */ \
1225 /* Register based. */ \
1226 INSN_3(JMP, JEQ, X), \
1227 INSN_3(JMP, JNE, X), \
1228 INSN_3(JMP, JGT, X), \
1229 INSN_3(JMP, JLT, X), \
1230 INSN_3(JMP, JGE, X), \
1231 INSN_3(JMP, JLE, X), \
1232 INSN_3(JMP, JSGT, X), \
1233 INSN_3(JMP, JSLT, X), \
1234 INSN_3(JMP, JSGE, X), \
1235 INSN_3(JMP, JSLE, X), \
1236 INSN_3(JMP, JSET, X), \
1237 /* Immediate based. */ \
1238 INSN_3(JMP, JEQ, K), \
1239 INSN_3(JMP, JNE, K), \
1240 INSN_3(JMP, JGT, K), \
1241 INSN_3(JMP, JLT, K), \
1242 INSN_3(JMP, JGE, K), \
1243 INSN_3(JMP, JLE, K), \
1244 INSN_3(JMP, JSGT, K), \
1245 INSN_3(JMP, JSLT, K), \
1246 INSN_3(JMP, JSGE, K), \
1247 INSN_3(JMP, JSLE, K), \
1248 INSN_3(JMP, JSET, K), \
1250 /* Store instructions. */ \
1251 /* Register based. */ \
1252 INSN_3(STX, MEM, B), \
1253 INSN_3(STX, MEM, H), \
1254 INSN_3(STX, MEM, W), \
1255 INSN_3(STX, MEM, DW), \
1256 INSN_3(STX, XADD, W), \
1257 INSN_3(STX, XADD, DW), \
1258 /* Immediate based. */ \
1259 INSN_3(ST, MEM, B), \
1260 INSN_3(ST, MEM, H), \
1261 INSN_3(ST, MEM, W), \
1262 INSN_3(ST, MEM, DW), \
1263 /* Load instructions. */ \
1264 /* Register based. */ \
1265 INSN_3(LDX, MEM, B), \
1266 INSN_3(LDX, MEM, H), \
1267 INSN_3(LDX, MEM, W), \
1268 INSN_3(LDX, MEM, DW), \
1269 /* Immediate based. */ \
1272 bool bpf_opcode_in_insntable(u8 code)
1274 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1275 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1276 static const bool public_insntable[256] = {
1277 [0 ... 255] = false,
1278 /* Now overwrite non-defaults ... */
1279 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1280 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1281 [BPF_LD | BPF_ABS | BPF_B] = true,
1282 [BPF_LD | BPF_ABS | BPF_H] = true,
1283 [BPF_LD | BPF_ABS | BPF_W] = true,
1284 [BPF_LD | BPF_IND | BPF_B] = true,
1285 [BPF_LD | BPF_IND | BPF_H] = true,
1286 [BPF_LD | BPF_IND | BPF_W] = true,
1288 #undef BPF_INSN_3_TBL
1289 #undef BPF_INSN_2_TBL
1290 return public_insntable[code];
1293 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1295 * __bpf_prog_run - run eBPF program on a given context
1296 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1297 * @insn: is the array of eBPF instructions
1298 * @stack: is the eBPF storage stack
1300 * Decode and execute eBPF instructions.
1302 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack)
1304 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1305 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1306 static const void *jumptable[256] = {
1307 [0 ... 255] = &&default_label,
1308 /* Now overwrite non-defaults ... */
1309 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1310 /* Non-UAPI available opcodes. */
1311 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1312 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1314 #undef BPF_INSN_3_LBL
1315 #undef BPF_INSN_2_LBL
1316 u32 tail_call_cnt = 0;
1318 #define CONT ({ insn++; goto select_insn; })
1319 #define CONT_JMP ({ insn++; goto select_insn; })
1322 goto *jumptable[insn->code];
1325 #define ALU(OPCODE, OP) \
1326 ALU64_##OPCODE##_X: \
1330 DST = (u32) DST OP (u32) SRC; \
1332 ALU64_##OPCODE##_K: \
1336 DST = (u32) DST OP (u32) IMM; \
1367 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1371 DST = (u64) (u32) ((*(s32 *) &DST) >> SRC);
1374 DST = (u64) (u32) ((*(s32 *) &DST) >> IMM);
1377 (*(s64 *) &DST) >>= SRC;
1380 (*(s64 *) &DST) >>= IMM;
1383 div64_u64_rem(DST, SRC, &AX);
1388 DST = do_div(AX, (u32) SRC);
1391 div64_u64_rem(DST, IMM, &AX);
1396 DST = do_div(AX, (u32) IMM);
1399 DST = div64_u64(DST, SRC);
1403 do_div(AX, (u32) SRC);
1407 DST = div64_u64(DST, IMM);
1411 do_div(AX, (u32) IMM);
1417 DST = (__force u16) cpu_to_be16(DST);
1420 DST = (__force u32) cpu_to_be32(DST);
1423 DST = (__force u64) cpu_to_be64(DST);
1430 DST = (__force u16) cpu_to_le16(DST);
1433 DST = (__force u32) cpu_to_le32(DST);
1436 DST = (__force u64) cpu_to_le64(DST);
1443 /* Function call scratches BPF_R1-BPF_R5 registers,
1444 * preserves BPF_R6-BPF_R9, and stores return value
1447 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1452 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1455 insn + insn->off + 1);
1459 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1460 struct bpf_array *array = container_of(map, struct bpf_array, map);
1461 struct bpf_prog *prog;
1464 if (unlikely(index >= array->map.max_entries))
1466 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1471 prog = READ_ONCE(array->ptrs[index]);
1475 /* ARG1 at this point is guaranteed to point to CTX from
1476 * the verifier side due to the fact that the tail call is
1477 * handeled like a helper, that is, bpf_tail_call_proto,
1478 * where arg1_type is ARG_PTR_TO_CTX.
1480 insn = prog->insnsi;
1491 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1493 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1494 insn += insn->off; \
1498 JMP32_##OPCODE##_X: \
1499 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1500 insn += insn->off; \
1505 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1506 insn += insn->off; \
1510 JMP32_##OPCODE##_K: \
1511 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1512 insn += insn->off; \
1516 COND_JMP(u, JEQ, ==)
1517 COND_JMP(u, JNE, !=)
1520 COND_JMP(u, JGE, >=)
1521 COND_JMP(u, JLE, <=)
1522 COND_JMP(u, JSET, &)
1523 COND_JMP(s, JSGT, >)
1524 COND_JMP(s, JSLT, <)
1525 COND_JMP(s, JSGE, >=)
1526 COND_JMP(s, JSLE, <=)
1528 /* STX and ST and LDX*/
1529 #define LDST(SIZEOP, SIZE) \
1531 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1534 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1537 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1545 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
1546 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
1549 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
1550 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
1555 /* If we ever reach this, we have a bug somewhere. Die hard here
1556 * instead of just returning 0; we could be somewhere in a subprog,
1557 * so execution could continue otherwise which we do /not/ want.
1559 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
1561 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code);
1565 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */
1567 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
1568 #define DEFINE_BPF_PROG_RUN(stack_size) \
1569 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
1571 u64 stack[stack_size / sizeof(u64)]; \
1572 u64 regs[MAX_BPF_EXT_REG]; \
1574 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1575 ARG1 = (u64) (unsigned long) ctx; \
1576 return ___bpf_prog_run(regs, insn, stack); \
1579 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
1580 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
1581 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
1582 const struct bpf_insn *insn) \
1584 u64 stack[stack_size / sizeof(u64)]; \
1585 u64 regs[MAX_BPF_EXT_REG]; \
1587 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1593 return ___bpf_prog_run(regs, insn, stack); \
1596 #define EVAL1(FN, X) FN(X)
1597 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
1598 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
1599 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
1600 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
1601 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
1603 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
1604 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
1605 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
1607 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
1608 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
1609 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
1611 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
1613 static unsigned int (*interpreters[])(const void *ctx,
1614 const struct bpf_insn *insn) = {
1615 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1616 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1617 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1619 #undef PROG_NAME_LIST
1620 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
1621 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
1622 const struct bpf_insn *insn) = {
1623 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1624 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1625 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1627 #undef PROG_NAME_LIST
1629 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
1631 stack_depth = max_t(u32, stack_depth, 1);
1632 insn->off = (s16) insn->imm;
1633 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
1634 __bpf_call_base_args;
1635 insn->code = BPF_JMP | BPF_CALL_ARGS;
1639 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
1640 const struct bpf_insn *insn)
1642 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
1643 * is not working properly, so warn about it!
1650 bool bpf_prog_array_compatible(struct bpf_array *array,
1651 const struct bpf_prog *fp)
1653 if (fp->kprobe_override)
1656 if (!array->owner_prog_type) {
1657 /* There's no owner yet where we could check for
1660 array->owner_prog_type = fp->type;
1661 array->owner_jited = fp->jited;
1666 return array->owner_prog_type == fp->type &&
1667 array->owner_jited == fp->jited;
1670 static int bpf_check_tail_call(const struct bpf_prog *fp)
1672 struct bpf_prog_aux *aux = fp->aux;
1675 for (i = 0; i < aux->used_map_cnt; i++) {
1676 struct bpf_map *map = aux->used_maps[i];
1677 struct bpf_array *array;
1679 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1682 array = container_of(map, struct bpf_array, map);
1683 if (!bpf_prog_array_compatible(array, fp))
1690 static void bpf_prog_select_func(struct bpf_prog *fp)
1692 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1693 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
1695 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
1697 fp->bpf_func = __bpf_prog_ret0_warn;
1702 * bpf_prog_select_runtime - select exec runtime for BPF program
1703 * @fp: bpf_prog populated with internal BPF program
1704 * @err: pointer to error variable
1706 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1707 * The BPF program will be executed via BPF_PROG_RUN() macro.
1709 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1711 /* In case of BPF to BPF calls, verifier did all the prep
1712 * work with regards to JITing, etc.
1717 bpf_prog_select_func(fp);
1719 /* eBPF JITs can rewrite the program in case constant
1720 * blinding is active. However, in case of error during
1721 * blinding, bpf_int_jit_compile() must always return a
1722 * valid program, which in this case would simply not
1723 * be JITed, but falls back to the interpreter.
1725 if (!bpf_prog_is_dev_bound(fp->aux)) {
1726 *err = bpf_prog_alloc_jited_linfo(fp);
1730 fp = bpf_int_jit_compile(fp);
1732 bpf_prog_free_jited_linfo(fp);
1733 #ifdef CONFIG_BPF_JIT_ALWAYS_ON
1738 bpf_prog_free_unused_jited_linfo(fp);
1741 *err = bpf_prog_offload_compile(fp);
1747 bpf_prog_lock_ro(fp);
1749 /* The tail call compatibility check can only be done at
1750 * this late stage as we need to determine, if we deal
1751 * with JITed or non JITed program concatenations and not
1752 * all eBPF JITs might immediately support all features.
1754 *err = bpf_check_tail_call(fp);
1758 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1760 static unsigned int __bpf_prog_ret1(const void *ctx,
1761 const struct bpf_insn *insn)
1766 static struct bpf_prog_dummy {
1767 struct bpf_prog prog;
1768 } dummy_bpf_prog = {
1770 .bpf_func = __bpf_prog_ret1,
1774 /* to avoid allocating empty bpf_prog_array for cgroups that
1775 * don't have bpf program attached use one global 'empty_prog_array'
1776 * It will not be modified the caller of bpf_prog_array_alloc()
1777 * (since caller requested prog_cnt == 0)
1778 * that pointer should be 'freed' by bpf_prog_array_free()
1781 struct bpf_prog_array hdr;
1782 struct bpf_prog *null_prog;
1783 } empty_prog_array = {
1787 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
1790 return kzalloc(sizeof(struct bpf_prog_array) +
1791 sizeof(struct bpf_prog_array_item) *
1795 return &empty_prog_array.hdr;
1798 void bpf_prog_array_free(struct bpf_prog_array __rcu *progs)
1801 progs == (struct bpf_prog_array __rcu *)&empty_prog_array.hdr)
1803 kfree_rcu(progs, rcu);
1806 int bpf_prog_array_length(struct bpf_prog_array __rcu *array)
1808 struct bpf_prog_array_item *item;
1812 item = rcu_dereference(array)->items;
1813 for (; item->prog; item++)
1814 if (item->prog != &dummy_bpf_prog.prog)
1821 static bool bpf_prog_array_copy_core(struct bpf_prog_array __rcu *array,
1825 struct bpf_prog_array_item *item;
1828 item = rcu_dereference_check(array, 1)->items;
1829 for (; item->prog; item++) {
1830 if (item->prog == &dummy_bpf_prog.prog)
1832 prog_ids[i] = item->prog->aux->id;
1833 if (++i == request_cnt) {
1839 return !!(item->prog);
1842 int bpf_prog_array_copy_to_user(struct bpf_prog_array __rcu *array,
1843 __u32 __user *prog_ids, u32 cnt)
1845 unsigned long err = 0;
1849 /* users of this function are doing:
1850 * cnt = bpf_prog_array_length();
1852 * bpf_prog_array_copy_to_user(..., cnt);
1853 * so below kcalloc doesn't need extra cnt > 0 check, but
1854 * bpf_prog_array_length() releases rcu lock and
1855 * prog array could have been swapped with empty or larger array,
1856 * so always copy 'cnt' prog_ids to the user.
1857 * In a rare race the user will see zero prog_ids
1859 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
1863 nospc = bpf_prog_array_copy_core(array, ids, cnt);
1865 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
1874 void bpf_prog_array_delete_safe(struct bpf_prog_array __rcu *array,
1875 struct bpf_prog *old_prog)
1877 struct bpf_prog_array_item *item = array->items;
1879 for (; item->prog; item++)
1880 if (item->prog == old_prog) {
1881 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
1886 int bpf_prog_array_copy(struct bpf_prog_array __rcu *old_array,
1887 struct bpf_prog *exclude_prog,
1888 struct bpf_prog *include_prog,
1889 struct bpf_prog_array **new_array)
1891 int new_prog_cnt, carry_prog_cnt = 0;
1892 struct bpf_prog_array_item *existing;
1893 struct bpf_prog_array *array;
1894 bool found_exclude = false;
1895 int new_prog_idx = 0;
1897 /* Figure out how many existing progs we need to carry over to
1901 existing = old_array->items;
1902 for (; existing->prog; existing++) {
1903 if (existing->prog == exclude_prog) {
1904 found_exclude = true;
1907 if (existing->prog != &dummy_bpf_prog.prog)
1909 if (existing->prog == include_prog)
1914 if (exclude_prog && !found_exclude)
1917 /* How many progs (not NULL) will be in the new array? */
1918 new_prog_cnt = carry_prog_cnt;
1922 /* Do we have any prog (not NULL) in the new array? */
1923 if (!new_prog_cnt) {
1928 /* +1 as the end of prog_array is marked with NULL */
1929 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
1933 /* Fill in the new prog array */
1934 if (carry_prog_cnt) {
1935 existing = old_array->items;
1936 for (; existing->prog; existing++)
1937 if (existing->prog != exclude_prog &&
1938 existing->prog != &dummy_bpf_prog.prog) {
1939 array->items[new_prog_idx++].prog =
1944 array->items[new_prog_idx++].prog = include_prog;
1945 array->items[new_prog_idx].prog = NULL;
1950 int bpf_prog_array_copy_info(struct bpf_prog_array __rcu *array,
1951 u32 *prog_ids, u32 request_cnt,
1957 cnt = bpf_prog_array_length(array);
1961 /* return early if user requested only program count or nothing to copy */
1962 if (!request_cnt || !cnt)
1965 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
1966 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
1970 static void bpf_prog_free_deferred(struct work_struct *work)
1972 struct bpf_prog_aux *aux;
1975 aux = container_of(work, struct bpf_prog_aux, work);
1976 if (bpf_prog_is_dev_bound(aux))
1977 bpf_prog_offload_destroy(aux->prog);
1978 #ifdef CONFIG_PERF_EVENTS
1979 if (aux->prog->has_callchain_buf)
1980 put_callchain_buffers();
1982 for (i = 0; i < aux->func_cnt; i++)
1983 bpf_jit_free(aux->func[i]);
1984 if (aux->func_cnt) {
1986 bpf_prog_unlock_free(aux->prog);
1988 bpf_jit_free(aux->prog);
1992 /* Free internal BPF program */
1993 void bpf_prog_free(struct bpf_prog *fp)
1995 struct bpf_prog_aux *aux = fp->aux;
1997 INIT_WORK(&aux->work, bpf_prog_free_deferred);
1998 schedule_work(&aux->work);
2000 EXPORT_SYMBOL_GPL(bpf_prog_free);
2002 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2003 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2005 void bpf_user_rnd_init_once(void)
2007 prandom_init_once(&bpf_user_rnd_state);
2010 BPF_CALL_0(bpf_user_rnd_u32)
2012 /* Should someone ever have the rather unwise idea to use some
2013 * of the registers passed into this function, then note that
2014 * this function is called from native eBPF and classic-to-eBPF
2015 * transformations. Register assignments from both sides are
2016 * different, f.e. classic always sets fn(ctx, A, X) here.
2018 struct rnd_state *state;
2021 state = &get_cpu_var(bpf_user_rnd_state);
2022 res = prandom_u32_state(state);
2023 put_cpu_var(bpf_user_rnd_state);
2028 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2029 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2030 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2031 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2032 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2033 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2034 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2035 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2036 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2038 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2039 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2040 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2041 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2043 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2044 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2045 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2046 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2047 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2049 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2055 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2056 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2060 EXPORT_SYMBOL_GPL(bpf_event_output);
2062 /* Always built-in helper functions. */
2063 const struct bpf_func_proto bpf_tail_call_proto = {
2066 .ret_type = RET_VOID,
2067 .arg1_type = ARG_PTR_TO_CTX,
2068 .arg2_type = ARG_CONST_MAP_PTR,
2069 .arg3_type = ARG_ANYTHING,
2072 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2073 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2074 * eBPF and implicitly also cBPF can get JITed!
2076 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2081 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2082 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2084 void __weak bpf_jit_compile(struct bpf_prog *prog)
2088 bool __weak bpf_helper_changes_pkt_data(void *func)
2093 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2094 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2096 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2102 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2103 EXPORT_SYMBOL(bpf_stats_enabled_key);
2104 int sysctl_bpf_stats_enabled __read_mostly;
2106 /* All definitions of tracepoints related to BPF. */
2107 #define CREATE_TRACE_POINTS
2108 #include <linux/bpf_trace.h>
2110 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);