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 <linux/module.h>
25 #include <linux/types.h>
27 #include <linux/fcntl.h>
28 #include <linux/socket.h>
30 #include <linux/inet.h>
31 #include <linux/netdevice.h>
32 #include <linux/if_packet.h>
33 #include <linux/gfp.h>
35 #include <net/protocol.h>
36 #include <net/netlink.h>
37 #include <linux/skbuff.h>
39 #include <net/flow_dissector.h>
40 #include <linux/errno.h>
41 #include <linux/timer.h>
42 #include <asm/uaccess.h>
43 #include <asm/unaligned.h>
44 #include <linux/filter.h>
45 #include <linux/ratelimit.h>
46 #include <linux/seccomp.h>
47 #include <linux/if_vlan.h>
48 #include <linux/bpf.h>
49 #include <net/sch_generic.h>
50 #include <net/cls_cgroup.h>
51 #include <net/dst_metadata.h>
53 #include <net/sock_reuseport.h>
56 * sk_filter_trim_cap - run a packet through a socket filter
57 * @sk: sock associated with &sk_buff
58 * @skb: buffer to filter
59 * @cap: limit on how short the eBPF program may trim the packet
61 * Run the eBPF program and then cut skb->data to correct size returned by
62 * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
63 * than pkt_len we keep whole skb->data. This is the socket level
64 * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
65 * be accepted or -EPERM if the packet should be tossed.
68 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
71 struct sk_filter *filter;
74 * If the skb was allocated from pfmemalloc reserves, only
75 * allow SOCK_MEMALLOC sockets to use it as this socket is
78 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC))
81 err = security_sock_rcv_skb(sk, skb);
86 filter = rcu_dereference(sk->sk_filter);
88 unsigned int pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
89 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
95 EXPORT_SYMBOL(sk_filter_trim_cap);
97 static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
99 return skb_get_poff((struct sk_buff *)(unsigned long) ctx);
102 static u64 __skb_get_nlattr(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
104 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
107 if (skb_is_nonlinear(skb))
110 if (skb->len < sizeof(struct nlattr))
113 if (a > skb->len - sizeof(struct nlattr))
116 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
118 return (void *) nla - (void *) skb->data;
123 static u64 __skb_get_nlattr_nest(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
125 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
128 if (skb_is_nonlinear(skb))
131 if (skb->len < sizeof(struct nlattr))
134 if (a > skb->len - sizeof(struct nlattr))
137 nla = (struct nlattr *) &skb->data[a];
138 if (nla->nla_len > skb->len - a)
141 nla = nla_find_nested(nla, x);
143 return (void *) nla - (void *) skb->data;
148 static u64 __get_raw_cpu_id(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
150 return raw_smp_processor_id();
153 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
154 .func = __get_raw_cpu_id,
156 .ret_type = RET_INTEGER,
159 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
160 struct bpf_insn *insn_buf)
162 struct bpf_insn *insn = insn_buf;
166 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
168 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
169 offsetof(struct sk_buff, mark));
173 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
174 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
175 #ifdef __BIG_ENDIAN_BITFIELD
176 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
181 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
183 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
184 offsetof(struct sk_buff, queue_mapping));
187 case SKF_AD_VLAN_TAG:
188 case SKF_AD_VLAN_TAG_PRESENT:
189 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
190 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
192 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
193 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
194 offsetof(struct sk_buff, vlan_tci));
195 if (skb_field == SKF_AD_VLAN_TAG) {
196 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
200 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
202 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
207 return insn - insn_buf;
210 static bool convert_bpf_extensions(struct sock_filter *fp,
211 struct bpf_insn **insnp)
213 struct bpf_insn *insn = *insnp;
217 case SKF_AD_OFF + SKF_AD_PROTOCOL:
218 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
220 /* A = *(u16 *) (CTX + offsetof(protocol)) */
221 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
222 offsetof(struct sk_buff, protocol));
223 /* A = ntohs(A) [emitting a nop or swap16] */
224 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
227 case SKF_AD_OFF + SKF_AD_PKTTYPE:
228 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
232 case SKF_AD_OFF + SKF_AD_IFINDEX:
233 case SKF_AD_OFF + SKF_AD_HATYPE:
234 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
235 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
236 BUILD_BUG_ON(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)) < 0);
238 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
239 BPF_REG_TMP, BPF_REG_CTX,
240 offsetof(struct sk_buff, dev));
241 /* if (tmp != 0) goto pc + 1 */
242 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
243 *insn++ = BPF_EXIT_INSN();
244 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
245 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
246 offsetof(struct net_device, ifindex));
248 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
249 offsetof(struct net_device, type));
252 case SKF_AD_OFF + SKF_AD_MARK:
253 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
257 case SKF_AD_OFF + SKF_AD_RXHASH:
258 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
260 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
261 offsetof(struct sk_buff, hash));
264 case SKF_AD_OFF + SKF_AD_QUEUE:
265 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
269 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
270 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
271 BPF_REG_A, BPF_REG_CTX, insn);
275 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
276 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
277 BPF_REG_A, BPF_REG_CTX, insn);
281 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
282 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
284 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
285 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
286 offsetof(struct sk_buff, vlan_proto));
287 /* A = ntohs(A) [emitting a nop or swap16] */
288 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
291 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
292 case SKF_AD_OFF + SKF_AD_NLATTR:
293 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
294 case SKF_AD_OFF + SKF_AD_CPU:
295 case SKF_AD_OFF + SKF_AD_RANDOM:
297 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
299 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
301 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
302 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
304 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
305 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
307 case SKF_AD_OFF + SKF_AD_NLATTR:
308 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
310 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
311 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
313 case SKF_AD_OFF + SKF_AD_CPU:
314 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
316 case SKF_AD_OFF + SKF_AD_RANDOM:
317 *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
318 bpf_user_rnd_init_once();
323 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
325 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
329 /* This is just a dummy call to avoid letting the compiler
330 * evict __bpf_call_base() as an optimization. Placed here
331 * where no-one bothers.
333 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
342 * bpf_convert_filter - convert filter program
343 * @prog: the user passed filter program
344 * @len: the length of the user passed filter program
345 * @new_prog: buffer where converted program will be stored
346 * @new_len: pointer to store length of converted program
348 * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style.
349 * Conversion workflow:
351 * 1) First pass for calculating the new program length:
352 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
354 * 2) 2nd pass to remap in two passes: 1st pass finds new
355 * jump offsets, 2nd pass remapping:
356 * new_prog = kmalloc(sizeof(struct bpf_insn) * new_len);
357 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
359 static int bpf_convert_filter(struct sock_filter *prog, int len,
360 struct bpf_insn *new_prog, int *new_len)
362 int new_flen = 0, pass = 0, target, i;
363 struct bpf_insn *new_insn;
364 struct sock_filter *fp;
368 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
369 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
371 if (len <= 0 || len > BPF_MAXINSNS)
375 addrs = kcalloc(len, sizeof(*addrs),
376 GFP_KERNEL | __GFP_NOWARN);
385 /* Classic BPF related prologue emission. */
387 /* Classic BPF expects A and X to be reset first. These need
388 * to be guaranteed to be the first two instructions.
390 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
391 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
393 /* All programs must keep CTX in callee saved BPF_REG_CTX.
394 * In eBPF case it's done by the compiler, here we need to
395 * do this ourself. Initial CTX is present in BPF_REG_ARG1.
397 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
402 for (i = 0; i < len; fp++, i++) {
403 struct bpf_insn tmp_insns[6] = { };
404 struct bpf_insn *insn = tmp_insns;
407 addrs[i] = new_insn - new_prog;
410 /* All arithmetic insns and skb loads map as-is. */
411 case BPF_ALU | BPF_ADD | BPF_X:
412 case BPF_ALU | BPF_ADD | BPF_K:
413 case BPF_ALU | BPF_SUB | BPF_X:
414 case BPF_ALU | BPF_SUB | BPF_K:
415 case BPF_ALU | BPF_AND | BPF_X:
416 case BPF_ALU | BPF_AND | BPF_K:
417 case BPF_ALU | BPF_OR | BPF_X:
418 case BPF_ALU | BPF_OR | BPF_K:
419 case BPF_ALU | BPF_LSH | BPF_X:
420 case BPF_ALU | BPF_LSH | BPF_K:
421 case BPF_ALU | BPF_RSH | BPF_X:
422 case BPF_ALU | BPF_RSH | BPF_K:
423 case BPF_ALU | BPF_XOR | BPF_X:
424 case BPF_ALU | BPF_XOR | BPF_K:
425 case BPF_ALU | BPF_MUL | BPF_X:
426 case BPF_ALU | BPF_MUL | BPF_K:
427 case BPF_ALU | BPF_DIV | BPF_X:
428 case BPF_ALU | BPF_DIV | BPF_K:
429 case BPF_ALU | BPF_MOD | BPF_X:
430 case BPF_ALU | BPF_MOD | BPF_K:
431 case BPF_ALU | BPF_NEG:
432 case BPF_LD | BPF_ABS | BPF_W:
433 case BPF_LD | BPF_ABS | BPF_H:
434 case BPF_LD | BPF_ABS | BPF_B:
435 case BPF_LD | BPF_IND | BPF_W:
436 case BPF_LD | BPF_IND | BPF_H:
437 case BPF_LD | BPF_IND | BPF_B:
438 /* Check for overloaded BPF extension and
439 * directly convert it if found, otherwise
440 * just move on with mapping.
442 if (BPF_CLASS(fp->code) == BPF_LD &&
443 BPF_MODE(fp->code) == BPF_ABS &&
444 convert_bpf_extensions(fp, &insn))
447 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
450 /* Jump transformation cannot use BPF block macros
451 * everywhere as offset calculation and target updates
452 * require a bit more work than the rest, i.e. jump
453 * opcodes map as-is, but offsets need adjustment.
456 #define BPF_EMIT_JMP \
458 if (target >= len || target < 0) \
460 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
461 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
462 insn->off -= insn - tmp_insns; \
465 case BPF_JMP | BPF_JA:
466 target = i + fp->k + 1;
467 insn->code = fp->code;
471 case BPF_JMP | BPF_JEQ | BPF_K:
472 case BPF_JMP | BPF_JEQ | BPF_X:
473 case BPF_JMP | BPF_JSET | BPF_K:
474 case BPF_JMP | BPF_JSET | BPF_X:
475 case BPF_JMP | BPF_JGT | BPF_K:
476 case BPF_JMP | BPF_JGT | BPF_X:
477 case BPF_JMP | BPF_JGE | BPF_K:
478 case BPF_JMP | BPF_JGE | BPF_X:
479 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
480 /* BPF immediates are signed, zero extend
481 * immediate into tmp register and use it
484 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
486 insn->dst_reg = BPF_REG_A;
487 insn->src_reg = BPF_REG_TMP;
490 insn->dst_reg = BPF_REG_A;
492 bpf_src = BPF_SRC(fp->code);
493 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
496 /* Common case where 'jump_false' is next insn. */
498 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
499 target = i + fp->jt + 1;
504 /* Convert JEQ into JNE when 'jump_true' is next insn. */
505 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
506 insn->code = BPF_JMP | BPF_JNE | bpf_src;
507 target = i + fp->jf + 1;
512 /* Other jumps are mapped into two insns: Jxx and JA. */
513 target = i + fp->jt + 1;
514 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
518 insn->code = BPF_JMP | BPF_JA;
519 target = i + fp->jf + 1;
523 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
524 case BPF_LDX | BPF_MSH | BPF_B:
526 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
527 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
528 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
530 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
532 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
534 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
536 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
539 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
540 * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
542 case BPF_RET | BPF_A:
543 case BPF_RET | BPF_K:
544 if (BPF_RVAL(fp->code) == BPF_K)
545 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
547 *insn = BPF_EXIT_INSN();
550 /* Store to stack. */
553 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
554 BPF_ST ? BPF_REG_A : BPF_REG_X,
555 -(BPF_MEMWORDS - fp->k) * 4);
558 /* Load from stack. */
559 case BPF_LD | BPF_MEM:
560 case BPF_LDX | BPF_MEM:
561 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
562 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
563 -(BPF_MEMWORDS - fp->k) * 4);
567 case BPF_LD | BPF_IMM:
568 case BPF_LDX | BPF_IMM:
569 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
570 BPF_REG_A : BPF_REG_X, fp->k);
574 case BPF_MISC | BPF_TAX:
575 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
579 case BPF_MISC | BPF_TXA:
580 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
583 /* A = skb->len or X = skb->len */
584 case BPF_LD | BPF_W | BPF_LEN:
585 case BPF_LDX | BPF_W | BPF_LEN:
586 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
587 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
588 offsetof(struct sk_buff, len));
591 /* Access seccomp_data fields. */
592 case BPF_LDX | BPF_ABS | BPF_W:
593 /* A = *(u32 *) (ctx + K) */
594 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
597 /* Unknown instruction. */
604 memcpy(new_insn, tmp_insns,
605 sizeof(*insn) * (insn - tmp_insns));
606 new_insn += insn - tmp_insns;
610 /* Only calculating new length. */
611 *new_len = new_insn - new_prog;
616 if (new_flen != new_insn - new_prog) {
617 new_flen = new_insn - new_prog;
624 BUG_ON(*new_len != new_flen);
633 * As we dont want to clear mem[] array for each packet going through
634 * __bpf_prog_run(), we check that filter loaded by user never try to read
635 * a cell if not previously written, and we check all branches to be sure
636 * a malicious user doesn't try to abuse us.
638 static int check_load_and_stores(const struct sock_filter *filter, int flen)
640 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
643 BUILD_BUG_ON(BPF_MEMWORDS > 16);
645 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
649 memset(masks, 0xff, flen * sizeof(*masks));
651 for (pc = 0; pc < flen; pc++) {
652 memvalid &= masks[pc];
654 switch (filter[pc].code) {
657 memvalid |= (1 << filter[pc].k);
659 case BPF_LD | BPF_MEM:
660 case BPF_LDX | BPF_MEM:
661 if (!(memvalid & (1 << filter[pc].k))) {
666 case BPF_JMP | BPF_JA:
667 /* A jump must set masks on target */
668 masks[pc + 1 + filter[pc].k] &= memvalid;
671 case BPF_JMP | BPF_JEQ | BPF_K:
672 case BPF_JMP | BPF_JEQ | BPF_X:
673 case BPF_JMP | BPF_JGE | BPF_K:
674 case BPF_JMP | BPF_JGE | BPF_X:
675 case BPF_JMP | BPF_JGT | BPF_K:
676 case BPF_JMP | BPF_JGT | BPF_X:
677 case BPF_JMP | BPF_JSET | BPF_K:
678 case BPF_JMP | BPF_JSET | BPF_X:
679 /* A jump must set masks on targets */
680 masks[pc + 1 + filter[pc].jt] &= memvalid;
681 masks[pc + 1 + filter[pc].jf] &= memvalid;
691 static bool chk_code_allowed(u16 code_to_probe)
693 static const bool codes[] = {
694 /* 32 bit ALU operations */
695 [BPF_ALU | BPF_ADD | BPF_K] = true,
696 [BPF_ALU | BPF_ADD | BPF_X] = true,
697 [BPF_ALU | BPF_SUB | BPF_K] = true,
698 [BPF_ALU | BPF_SUB | BPF_X] = true,
699 [BPF_ALU | BPF_MUL | BPF_K] = true,
700 [BPF_ALU | BPF_MUL | BPF_X] = true,
701 [BPF_ALU | BPF_DIV | BPF_K] = true,
702 [BPF_ALU | BPF_DIV | BPF_X] = true,
703 [BPF_ALU | BPF_MOD | BPF_K] = true,
704 [BPF_ALU | BPF_MOD | BPF_X] = true,
705 [BPF_ALU | BPF_AND | BPF_K] = true,
706 [BPF_ALU | BPF_AND | BPF_X] = true,
707 [BPF_ALU | BPF_OR | BPF_K] = true,
708 [BPF_ALU | BPF_OR | BPF_X] = true,
709 [BPF_ALU | BPF_XOR | BPF_K] = true,
710 [BPF_ALU | BPF_XOR | BPF_X] = true,
711 [BPF_ALU | BPF_LSH | BPF_K] = true,
712 [BPF_ALU | BPF_LSH | BPF_X] = true,
713 [BPF_ALU | BPF_RSH | BPF_K] = true,
714 [BPF_ALU | BPF_RSH | BPF_X] = true,
715 [BPF_ALU | BPF_NEG] = true,
716 /* Load instructions */
717 [BPF_LD | BPF_W | BPF_ABS] = true,
718 [BPF_LD | BPF_H | BPF_ABS] = true,
719 [BPF_LD | BPF_B | BPF_ABS] = true,
720 [BPF_LD | BPF_W | BPF_LEN] = true,
721 [BPF_LD | BPF_W | BPF_IND] = true,
722 [BPF_LD | BPF_H | BPF_IND] = true,
723 [BPF_LD | BPF_B | BPF_IND] = true,
724 [BPF_LD | BPF_IMM] = true,
725 [BPF_LD | BPF_MEM] = true,
726 [BPF_LDX | BPF_W | BPF_LEN] = true,
727 [BPF_LDX | BPF_B | BPF_MSH] = true,
728 [BPF_LDX | BPF_IMM] = true,
729 [BPF_LDX | BPF_MEM] = true,
730 /* Store instructions */
733 /* Misc instructions */
734 [BPF_MISC | BPF_TAX] = true,
735 [BPF_MISC | BPF_TXA] = true,
736 /* Return instructions */
737 [BPF_RET | BPF_K] = true,
738 [BPF_RET | BPF_A] = true,
739 /* Jump instructions */
740 [BPF_JMP | BPF_JA] = true,
741 [BPF_JMP | BPF_JEQ | BPF_K] = true,
742 [BPF_JMP | BPF_JEQ | BPF_X] = true,
743 [BPF_JMP | BPF_JGE | BPF_K] = true,
744 [BPF_JMP | BPF_JGE | BPF_X] = true,
745 [BPF_JMP | BPF_JGT | BPF_K] = true,
746 [BPF_JMP | BPF_JGT | BPF_X] = true,
747 [BPF_JMP | BPF_JSET | BPF_K] = true,
748 [BPF_JMP | BPF_JSET | BPF_X] = true,
751 if (code_to_probe >= ARRAY_SIZE(codes))
754 return codes[code_to_probe];
757 static bool bpf_check_basics_ok(const struct sock_filter *filter,
762 if (flen == 0 || flen > BPF_MAXINSNS)
769 * bpf_check_classic - verify socket filter code
770 * @filter: filter to verify
771 * @flen: length of filter
773 * Check the user's filter code. If we let some ugly
774 * filter code slip through kaboom! The filter must contain
775 * no references or jumps that are out of range, no illegal
776 * instructions, and must end with a RET instruction.
778 * All jumps are forward as they are not signed.
780 * Returns 0 if the rule set is legal or -EINVAL if not.
782 static int bpf_check_classic(const struct sock_filter *filter,
788 /* Check the filter code now */
789 for (pc = 0; pc < flen; pc++) {
790 const struct sock_filter *ftest = &filter[pc];
792 /* May we actually operate on this code? */
793 if (!chk_code_allowed(ftest->code))
796 /* Some instructions need special checks */
797 switch (ftest->code) {
798 case BPF_ALU | BPF_DIV | BPF_K:
799 case BPF_ALU | BPF_MOD | BPF_K:
800 /* Check for division by zero */
804 case BPF_ALU | BPF_LSH | BPF_K:
805 case BPF_ALU | BPF_RSH | BPF_K:
809 case BPF_LD | BPF_MEM:
810 case BPF_LDX | BPF_MEM:
813 /* Check for invalid memory addresses */
814 if (ftest->k >= BPF_MEMWORDS)
817 case BPF_JMP | BPF_JA:
818 /* Note, the large ftest->k might cause loops.
819 * Compare this with conditional jumps below,
820 * where offsets are limited. --ANK (981016)
822 if (ftest->k >= (unsigned int)(flen - pc - 1))
825 case BPF_JMP | BPF_JEQ | BPF_K:
826 case BPF_JMP | BPF_JEQ | BPF_X:
827 case BPF_JMP | BPF_JGE | BPF_K:
828 case BPF_JMP | BPF_JGE | BPF_X:
829 case BPF_JMP | BPF_JGT | BPF_K:
830 case BPF_JMP | BPF_JGT | BPF_X:
831 case BPF_JMP | BPF_JSET | BPF_K:
832 case BPF_JMP | BPF_JSET | BPF_X:
833 /* Both conditionals must be safe */
834 if (pc + ftest->jt + 1 >= flen ||
835 pc + ftest->jf + 1 >= flen)
838 case BPF_LD | BPF_W | BPF_ABS:
839 case BPF_LD | BPF_H | BPF_ABS:
840 case BPF_LD | BPF_B | BPF_ABS:
842 if (bpf_anc_helper(ftest) & BPF_ANC)
844 /* Ancillary operation unknown or unsupported */
845 if (anc_found == false && ftest->k >= SKF_AD_OFF)
850 /* Last instruction must be a RET code */
851 switch (filter[flen - 1].code) {
852 case BPF_RET | BPF_K:
853 case BPF_RET | BPF_A:
854 return check_load_and_stores(filter, flen);
860 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
861 const struct sock_fprog *fprog)
863 unsigned int fsize = bpf_classic_proglen(fprog);
864 struct sock_fprog_kern *fkprog;
866 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
870 fkprog = fp->orig_prog;
871 fkprog->len = fprog->len;
873 fkprog->filter = kmemdup(fp->insns, fsize,
874 GFP_KERNEL | __GFP_NOWARN);
875 if (!fkprog->filter) {
876 kfree(fp->orig_prog);
883 static void bpf_release_orig_filter(struct bpf_prog *fp)
885 struct sock_fprog_kern *fprog = fp->orig_prog;
888 kfree(fprog->filter);
893 static void __bpf_prog_release(struct bpf_prog *prog)
895 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
898 bpf_release_orig_filter(prog);
903 static void __sk_filter_release(struct sk_filter *fp)
905 __bpf_prog_release(fp->prog);
910 * sk_filter_release_rcu - Release a socket filter by rcu_head
911 * @rcu: rcu_head that contains the sk_filter to free
913 static void sk_filter_release_rcu(struct rcu_head *rcu)
915 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
917 __sk_filter_release(fp);
921 * sk_filter_release - release a socket filter
922 * @fp: filter to remove
924 * Remove a filter from a socket and release its resources.
926 static void sk_filter_release(struct sk_filter *fp)
928 if (atomic_dec_and_test(&fp->refcnt))
929 call_rcu(&fp->rcu, sk_filter_release_rcu);
932 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
934 u32 filter_size = bpf_prog_size(fp->prog->len);
936 atomic_sub(filter_size, &sk->sk_omem_alloc);
937 sk_filter_release(fp);
940 /* try to charge the socket memory if there is space available
941 * return true on success
943 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
945 u32 filter_size = bpf_prog_size(fp->prog->len);
947 /* same check as in sock_kmalloc() */
948 if (filter_size <= sysctl_optmem_max &&
949 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
950 atomic_inc(&fp->refcnt);
951 atomic_add(filter_size, &sk->sk_omem_alloc);
957 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
959 struct sock_filter *old_prog;
960 struct bpf_prog *old_fp;
961 int err, new_len, old_len = fp->len;
963 /* We are free to overwrite insns et al right here as it
964 * won't be used at this point in time anymore internally
965 * after the migration to the internal BPF instruction
968 BUILD_BUG_ON(sizeof(struct sock_filter) !=
969 sizeof(struct bpf_insn));
971 /* Conversion cannot happen on overlapping memory areas,
972 * so we need to keep the user BPF around until the 2nd
973 * pass. At this time, the user BPF is stored in fp->insns.
975 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
976 GFP_KERNEL | __GFP_NOWARN);
982 /* 1st pass: calculate the new program length. */
983 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
987 /* Expand fp for appending the new filter representation. */
989 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
991 /* The old_fp is still around in case we couldn't
992 * allocate new memory, so uncharge on that one.
1001 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1002 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
1004 /* 2nd bpf_convert_filter() can fail only if it fails
1005 * to allocate memory, remapping must succeed. Note,
1006 * that at this time old_fp has already been released
1011 /* We are guaranteed to never error here with cBPF to eBPF
1012 * transitions, since there's no issue with type compatibility
1013 * checks on program arrays.
1015 fp = bpf_prog_select_runtime(fp, &err);
1023 __bpf_prog_release(fp);
1024 return ERR_PTR(err);
1027 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1028 bpf_aux_classic_check_t trans)
1032 fp->bpf_func = NULL;
1035 err = bpf_check_classic(fp->insns, fp->len);
1037 __bpf_prog_release(fp);
1038 return ERR_PTR(err);
1041 /* There might be additional checks and transformations
1042 * needed on classic filters, f.e. in case of seccomp.
1045 err = trans(fp->insns, fp->len);
1047 __bpf_prog_release(fp);
1048 return ERR_PTR(err);
1052 /* Probe if we can JIT compile the filter and if so, do
1053 * the compilation of the filter.
1055 bpf_jit_compile(fp);
1057 /* JIT compiler couldn't process this filter, so do the
1058 * internal BPF translation for the optimized interpreter.
1061 fp = bpf_migrate_filter(fp);
1067 * bpf_prog_create - create an unattached filter
1068 * @pfp: the unattached filter that is created
1069 * @fprog: the filter program
1071 * Create a filter independent of any socket. We first run some
1072 * sanity checks on it to make sure it does not explode on us later.
1073 * If an error occurs or there is insufficient memory for the filter
1074 * a negative errno code is returned. On success the return is zero.
1076 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1078 unsigned int fsize = bpf_classic_proglen(fprog);
1079 struct bpf_prog *fp;
1081 /* Make sure new filter is there and in the right amounts. */
1082 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1085 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1089 memcpy(fp->insns, fprog->filter, fsize);
1091 fp->len = fprog->len;
1092 /* Since unattached filters are not copied back to user
1093 * space through sk_get_filter(), we do not need to hold
1094 * a copy here, and can spare us the work.
1096 fp->orig_prog = NULL;
1098 /* bpf_prepare_filter() already takes care of freeing
1099 * memory in case something goes wrong.
1101 fp = bpf_prepare_filter(fp, NULL);
1108 EXPORT_SYMBOL_GPL(bpf_prog_create);
1111 * bpf_prog_create_from_user - create an unattached filter from user buffer
1112 * @pfp: the unattached filter that is created
1113 * @fprog: the filter program
1114 * @trans: post-classic verifier transformation handler
1115 * @save_orig: save classic BPF program
1117 * This function effectively does the same as bpf_prog_create(), only
1118 * that it builds up its insns buffer from user space provided buffer.
1119 * It also allows for passing a bpf_aux_classic_check_t handler.
1121 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1122 bpf_aux_classic_check_t trans, bool save_orig)
1124 unsigned int fsize = bpf_classic_proglen(fprog);
1125 struct bpf_prog *fp;
1128 /* Make sure new filter is there and in the right amounts. */
1129 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1132 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1136 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1137 __bpf_prog_free(fp);
1141 fp->len = fprog->len;
1142 fp->orig_prog = NULL;
1145 err = bpf_prog_store_orig_filter(fp, fprog);
1147 __bpf_prog_free(fp);
1152 /* bpf_prepare_filter() already takes care of freeing
1153 * memory in case something goes wrong.
1155 fp = bpf_prepare_filter(fp, trans);
1162 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1164 void bpf_prog_destroy(struct bpf_prog *fp)
1166 __bpf_prog_release(fp);
1168 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1170 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1172 struct sk_filter *fp, *old_fp;
1174 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1179 atomic_set(&fp->refcnt, 0);
1181 if (!sk_filter_charge(sk, fp)) {
1186 old_fp = rcu_dereference_protected(sk->sk_filter,
1187 lockdep_sock_is_held(sk));
1188 rcu_assign_pointer(sk->sk_filter, fp);
1191 sk_filter_uncharge(sk, old_fp);
1196 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1198 struct bpf_prog *old_prog;
1201 if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1204 if (sk_unhashed(sk) && sk->sk_reuseport) {
1205 err = reuseport_alloc(sk);
1208 } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1209 /* The socket wasn't bound with SO_REUSEPORT */
1213 old_prog = reuseport_attach_prog(sk, prog);
1215 bpf_prog_destroy(old_prog);
1221 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1223 unsigned int fsize = bpf_classic_proglen(fprog);
1224 struct bpf_prog *prog;
1227 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1228 return ERR_PTR(-EPERM);
1230 /* Make sure new filter is there and in the right amounts. */
1231 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1232 return ERR_PTR(-EINVAL);
1234 prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1236 return ERR_PTR(-ENOMEM);
1238 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1239 __bpf_prog_free(prog);
1240 return ERR_PTR(-EFAULT);
1243 prog->len = fprog->len;
1245 err = bpf_prog_store_orig_filter(prog, fprog);
1247 __bpf_prog_free(prog);
1248 return ERR_PTR(-ENOMEM);
1251 /* bpf_prepare_filter() already takes care of freeing
1252 * memory in case something goes wrong.
1254 return bpf_prepare_filter(prog, NULL);
1258 * sk_attach_filter - attach a socket filter
1259 * @fprog: the filter program
1260 * @sk: the socket to use
1262 * Attach the user's filter code. We first run some sanity checks on
1263 * it to make sure it does not explode on us later. If an error
1264 * occurs or there is insufficient memory for the filter a negative
1265 * errno code is returned. On success the return is zero.
1267 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1269 struct bpf_prog *prog = __get_filter(fprog, sk);
1273 return PTR_ERR(prog);
1275 err = __sk_attach_prog(prog, sk);
1277 __bpf_prog_release(prog);
1283 EXPORT_SYMBOL_GPL(sk_attach_filter);
1285 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1287 struct bpf_prog *prog = __get_filter(fprog, sk);
1291 return PTR_ERR(prog);
1293 err = __reuseport_attach_prog(prog, sk);
1295 __bpf_prog_release(prog);
1302 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1304 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1305 return ERR_PTR(-EPERM);
1307 return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1310 int sk_attach_bpf(u32 ufd, struct sock *sk)
1312 struct bpf_prog *prog = __get_bpf(ufd, sk);
1316 return PTR_ERR(prog);
1318 err = __sk_attach_prog(prog, sk);
1327 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1329 struct bpf_prog *prog = __get_bpf(ufd, sk);
1333 return PTR_ERR(prog);
1335 err = __reuseport_attach_prog(prog, sk);
1344 struct bpf_scratchpad {
1346 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1347 u8 buff[MAX_BPF_STACK];
1351 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1353 static inline int bpf_try_make_writable(struct sk_buff *skb,
1354 unsigned int write_len)
1358 if (!skb_cloned(skb))
1360 if (skb_clone_writable(skb, write_len))
1362 err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
1364 bpf_compute_data_end(skb);
1368 static u64 bpf_skb_store_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 flags)
1370 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1371 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1372 int offset = (int) r2;
1373 void *from = (void *) (long) r3;
1374 unsigned int len = (unsigned int) r4;
1377 if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1380 /* bpf verifier guarantees that:
1381 * 'from' pointer points to bpf program stack
1382 * 'len' bytes of it were initialized
1384 * 'skb' is a valid pointer to 'struct sk_buff'
1386 * so check for invalid 'offset' and too large 'len'
1388 if (unlikely((u32) offset > 0xffff || len > sizeof(sp->buff)))
1390 if (unlikely(bpf_try_make_writable(skb, offset + len)))
1393 ptr = skb_header_pointer(skb, offset, len, sp->buff);
1397 if (flags & BPF_F_RECOMPUTE_CSUM)
1398 skb_postpull_rcsum(skb, ptr, len);
1400 memcpy(ptr, from, len);
1402 if (ptr == sp->buff)
1403 /* skb_store_bits cannot return -EFAULT here */
1404 skb_store_bits(skb, offset, ptr, len);
1406 if (flags & BPF_F_RECOMPUTE_CSUM)
1407 skb_postpush_rcsum(skb, ptr, len);
1408 if (flags & BPF_F_INVALIDATE_HASH)
1409 skb_clear_hash(skb);
1414 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1415 .func = bpf_skb_store_bytes,
1417 .ret_type = RET_INTEGER,
1418 .arg1_type = ARG_PTR_TO_CTX,
1419 .arg2_type = ARG_ANYTHING,
1420 .arg3_type = ARG_PTR_TO_STACK,
1421 .arg4_type = ARG_CONST_STACK_SIZE,
1422 .arg5_type = ARG_ANYTHING,
1425 static u64 bpf_skb_load_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1427 const struct sk_buff *skb = (const struct sk_buff *)(unsigned long) r1;
1428 int offset = (int) r2;
1429 void *to = (void *)(unsigned long) r3;
1430 unsigned int len = (unsigned int) r4;
1433 if (unlikely((u32) offset > 0xffff))
1436 ptr = skb_header_pointer(skb, offset, len, to);
1440 memcpy(to, ptr, len);
1448 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1449 .func = bpf_skb_load_bytes,
1451 .ret_type = RET_INTEGER,
1452 .arg1_type = ARG_PTR_TO_CTX,
1453 .arg2_type = ARG_ANYTHING,
1454 .arg3_type = ARG_PTR_TO_RAW_STACK,
1455 .arg4_type = ARG_CONST_STACK_SIZE,
1458 static u64 bpf_l3_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1460 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1461 int offset = (int) r2;
1464 if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1466 if (unlikely((u32) offset > 0xffff))
1468 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(sum))))
1471 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1475 switch (flags & BPF_F_HDR_FIELD_MASK) {
1477 if (unlikely(from != 0))
1480 csum_replace_by_diff(ptr, to);
1483 csum_replace2(ptr, from, to);
1486 csum_replace4(ptr, from, to);
1493 /* skb_store_bits guaranteed to not return -EFAULT here */
1494 skb_store_bits(skb, offset, ptr, sizeof(sum));
1499 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1500 .func = bpf_l3_csum_replace,
1502 .ret_type = RET_INTEGER,
1503 .arg1_type = ARG_PTR_TO_CTX,
1504 .arg2_type = ARG_ANYTHING,
1505 .arg3_type = ARG_ANYTHING,
1506 .arg4_type = ARG_ANYTHING,
1507 .arg5_type = ARG_ANYTHING,
1510 static u64 bpf_l4_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1512 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1513 bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1514 bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1515 int offset = (int) r2;
1518 if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_PSEUDO_HDR |
1519 BPF_F_HDR_FIELD_MASK)))
1521 if (unlikely((u32) offset > 0xffff))
1523 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(sum))))
1526 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1529 if (is_mmzero && !*ptr)
1532 switch (flags & BPF_F_HDR_FIELD_MASK) {
1534 if (unlikely(from != 0))
1537 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1540 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1543 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1549 if (is_mmzero && !*ptr)
1550 *ptr = CSUM_MANGLED_0;
1552 /* skb_store_bits guaranteed to not return -EFAULT here */
1553 skb_store_bits(skb, offset, ptr, sizeof(sum));
1558 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1559 .func = bpf_l4_csum_replace,
1561 .ret_type = RET_INTEGER,
1562 .arg1_type = ARG_PTR_TO_CTX,
1563 .arg2_type = ARG_ANYTHING,
1564 .arg3_type = ARG_ANYTHING,
1565 .arg4_type = ARG_ANYTHING,
1566 .arg5_type = ARG_ANYTHING,
1569 static u64 bpf_csum_diff(u64 r1, u64 from_size, u64 r3, u64 to_size, u64 seed)
1571 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1572 u64 diff_size = from_size + to_size;
1573 __be32 *from = (__be32 *) (long) r1;
1574 __be32 *to = (__be32 *) (long) r3;
1577 /* This is quite flexible, some examples:
1579 * from_size == 0, to_size > 0, seed := csum --> pushing data
1580 * from_size > 0, to_size == 0, seed := csum --> pulling data
1581 * from_size > 0, to_size > 0, seed := 0 --> diffing data
1583 * Even for diffing, from_size and to_size don't need to be equal.
1585 if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1586 diff_size > sizeof(sp->diff)))
1589 for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1590 sp->diff[j] = ~from[i];
1591 for (i = 0; i < to_size / sizeof(__be32); i++, j++)
1592 sp->diff[j] = to[i];
1594 return csum_partial(sp->diff, diff_size, seed);
1597 static const struct bpf_func_proto bpf_csum_diff_proto = {
1598 .func = bpf_csum_diff,
1600 .ret_type = RET_INTEGER,
1601 .arg1_type = ARG_PTR_TO_STACK,
1602 .arg2_type = ARG_CONST_STACK_SIZE_OR_ZERO,
1603 .arg3_type = ARG_PTR_TO_STACK,
1604 .arg4_type = ARG_CONST_STACK_SIZE_OR_ZERO,
1605 .arg5_type = ARG_ANYTHING,
1608 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1610 if (skb_at_tc_ingress(skb))
1611 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1613 return dev_forward_skb(dev, skb);
1616 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1620 if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1621 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1628 __this_cpu_inc(xmit_recursion);
1629 ret = dev_queue_xmit(skb);
1630 __this_cpu_dec(xmit_recursion);
1635 static u64 bpf_clone_redirect(u64 r1, u64 ifindex, u64 flags, u64 r4, u64 r5)
1637 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1638 struct net_device *dev;
1640 if (unlikely(flags & ~(BPF_F_INGRESS)))
1643 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1647 skb = skb_clone(skb, GFP_ATOMIC);
1651 return flags & BPF_F_INGRESS ?
1652 __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1655 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1656 .func = bpf_clone_redirect,
1658 .ret_type = RET_INTEGER,
1659 .arg1_type = ARG_PTR_TO_CTX,
1660 .arg2_type = ARG_ANYTHING,
1661 .arg3_type = ARG_ANYTHING,
1664 struct redirect_info {
1669 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1671 static u64 bpf_redirect(u64 ifindex, u64 flags, u64 r3, u64 r4, u64 r5)
1673 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1675 if (unlikely(flags & ~(BPF_F_INGRESS)))
1678 ri->ifindex = ifindex;
1681 return TC_ACT_REDIRECT;
1684 int skb_do_redirect(struct sk_buff *skb)
1686 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1687 struct net_device *dev;
1689 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1691 if (unlikely(!dev)) {
1696 return ri->flags & BPF_F_INGRESS ?
1697 __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1700 static const struct bpf_func_proto bpf_redirect_proto = {
1701 .func = bpf_redirect,
1703 .ret_type = RET_INTEGER,
1704 .arg1_type = ARG_ANYTHING,
1705 .arg2_type = ARG_ANYTHING,
1708 static u64 bpf_get_cgroup_classid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1710 return task_get_classid((struct sk_buff *) (unsigned long) r1);
1713 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1714 .func = bpf_get_cgroup_classid,
1716 .ret_type = RET_INTEGER,
1717 .arg1_type = ARG_PTR_TO_CTX,
1720 static u64 bpf_get_route_realm(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1722 return dst_tclassid((struct sk_buff *) (unsigned long) r1);
1725 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1726 .func = bpf_get_route_realm,
1728 .ret_type = RET_INTEGER,
1729 .arg1_type = ARG_PTR_TO_CTX,
1732 static u64 bpf_get_hash_recalc(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1734 /* If skb_clear_hash() was called due to mangling, we can
1735 * trigger SW recalculation here. Later access to hash
1736 * can then use the inline skb->hash via context directly
1737 * instead of calling this helper again.
1739 return skb_get_hash((struct sk_buff *) (unsigned long) r1);
1742 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1743 .func = bpf_get_hash_recalc,
1745 .ret_type = RET_INTEGER,
1746 .arg1_type = ARG_PTR_TO_CTX,
1749 static u64 bpf_skb_vlan_push(u64 r1, u64 r2, u64 vlan_tci, u64 r4, u64 r5)
1751 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1752 __be16 vlan_proto = (__force __be16) r2;
1755 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1756 vlan_proto != htons(ETH_P_8021AD)))
1757 vlan_proto = htons(ETH_P_8021Q);
1759 ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1760 bpf_compute_data_end(skb);
1764 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1765 .func = bpf_skb_vlan_push,
1767 .ret_type = RET_INTEGER,
1768 .arg1_type = ARG_PTR_TO_CTX,
1769 .arg2_type = ARG_ANYTHING,
1770 .arg3_type = ARG_ANYTHING,
1772 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1774 static u64 bpf_skb_vlan_pop(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1776 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1779 ret = skb_vlan_pop(skb);
1780 bpf_compute_data_end(skb);
1784 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1785 .func = bpf_skb_vlan_pop,
1787 .ret_type = RET_INTEGER,
1788 .arg1_type = ARG_PTR_TO_CTX,
1790 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1792 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
1794 /* Caller already did skb_cow() with len as headroom,
1795 * so no need to do it here.
1798 memmove(skb->data, skb->data + len, off);
1799 memset(skb->data + off, 0, len);
1801 /* No skb_postpush_rcsum(skb, skb->data + off, len)
1802 * needed here as it does not change the skb->csum
1803 * result for checksum complete when summing over
1809 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
1811 /* skb_ensure_writable() is not needed here, as we're
1812 * already working on an uncloned skb.
1814 if (unlikely(!pskb_may_pull(skb, off + len)))
1817 skb_postpull_rcsum(skb, skb->data + off, len);
1818 memmove(skb->data + len, skb->data, off);
1819 __skb_pull(skb, len);
1824 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
1826 bool trans_same = skb->transport_header == skb->network_header;
1829 /* There's no need for __skb_push()/__skb_pull() pair to
1830 * get to the start of the mac header as we're guaranteed
1831 * to always start from here under eBPF.
1833 ret = bpf_skb_generic_push(skb, off, len);
1835 skb->mac_header -= len;
1836 skb->network_header -= len;
1838 skb->transport_header = skb->network_header;
1844 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
1846 bool trans_same = skb->transport_header == skb->network_header;
1849 /* Same here, __skb_push()/__skb_pull() pair not needed. */
1850 ret = bpf_skb_generic_pop(skb, off, len);
1852 skb->mac_header += len;
1853 skb->network_header += len;
1855 skb->transport_header = skb->network_header;
1861 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
1863 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
1864 u32 off = skb->network_header - skb->mac_header;
1867 ret = skb_cow(skb, len_diff);
1868 if (unlikely(ret < 0))
1871 ret = bpf_skb_net_hdr_push(skb, off, len_diff);
1872 if (unlikely(ret < 0))
1875 if (skb_is_gso(skb)) {
1876 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV4 needs to
1877 * be changed into SKB_GSO_TCPV6.
1879 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
1880 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
1881 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV6;
1884 /* Due to IPv6 header, MSS needs to be downgraded. */
1885 skb_shinfo(skb)->gso_size -= len_diff;
1886 /* Header must be checked, and gso_segs recomputed. */
1887 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1888 skb_shinfo(skb)->gso_segs = 0;
1891 skb->protocol = htons(ETH_P_IPV6);
1892 skb_clear_hash(skb);
1897 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
1899 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
1900 u32 off = skb->network_header - skb->mac_header;
1903 ret = skb_unclone(skb, GFP_ATOMIC);
1904 if (unlikely(ret < 0))
1907 ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
1908 if (unlikely(ret < 0))
1911 if (skb_is_gso(skb)) {
1912 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV6 needs to
1913 * be changed into SKB_GSO_TCPV4.
1915 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
1916 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
1917 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
1920 /* Due to IPv4 header, MSS can be upgraded. */
1921 skb_shinfo(skb)->gso_size += len_diff;
1922 /* Header must be checked, and gso_segs recomputed. */
1923 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1924 skb_shinfo(skb)->gso_segs = 0;
1927 skb->protocol = htons(ETH_P_IP);
1928 skb_clear_hash(skb);
1933 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
1935 __be16 from_proto = skb->protocol;
1937 if (from_proto == htons(ETH_P_IP) &&
1938 to_proto == htons(ETH_P_IPV6))
1939 return bpf_skb_proto_4_to_6(skb);
1941 if (from_proto == htons(ETH_P_IPV6) &&
1942 to_proto == htons(ETH_P_IP))
1943 return bpf_skb_proto_6_to_4(skb);
1948 static u64 bpf_skb_change_proto(u64 r1, u64 r2, u64 flags, u64 r4, u64 r5)
1950 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1951 __be16 proto = (__force __be16) r2;
1954 if (unlikely(flags))
1957 /* General idea is that this helper does the basic groundwork
1958 * needed for changing the protocol, and eBPF program fills the
1959 * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
1960 * and other helpers, rather than passing a raw buffer here.
1962 * The rationale is to keep this minimal and without a need to
1963 * deal with raw packet data. F.e. even if we would pass buffers
1964 * here, the program still needs to call the bpf_lX_csum_replace()
1965 * helpers anyway. Plus, this way we keep also separation of
1966 * concerns, since f.e. bpf_skb_store_bytes() should only take
1969 * Currently, additional options and extension header space are
1970 * not supported, but flags register is reserved so we can adapt
1971 * that. For offloads, we mark packet as dodgy, so that headers
1972 * need to be verified first.
1974 ret = bpf_skb_proto_xlat(skb, proto);
1975 bpf_compute_data_end(skb);
1979 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
1980 .func = bpf_skb_change_proto,
1982 .ret_type = RET_INTEGER,
1983 .arg1_type = ARG_PTR_TO_CTX,
1984 .arg2_type = ARG_ANYTHING,
1985 .arg3_type = ARG_ANYTHING,
1988 static u64 bpf_skb_change_type(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1990 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1993 /* We only allow a restricted subset to be changed for now. */
1994 if (unlikely(skb->pkt_type > PACKET_OTHERHOST ||
1995 pkt_type > PACKET_OTHERHOST))
1998 skb->pkt_type = pkt_type;
2002 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2003 .func = bpf_skb_change_type,
2005 .ret_type = RET_INTEGER,
2006 .arg1_type = ARG_PTR_TO_CTX,
2007 .arg2_type = ARG_ANYTHING,
2010 bool bpf_helper_changes_skb_data(void *func)
2012 if (func == bpf_skb_vlan_push)
2014 if (func == bpf_skb_vlan_pop)
2016 if (func == bpf_skb_store_bytes)
2018 if (func == bpf_skb_change_proto)
2020 if (func == bpf_l3_csum_replace)
2022 if (func == bpf_l4_csum_replace)
2028 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2029 unsigned long off, unsigned long len)
2031 void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2035 if (ptr != dst_buff)
2036 memcpy(dst_buff, ptr, len);
2041 static u64 bpf_skb_event_output(u64 r1, u64 r2, u64 flags, u64 r4,
2044 struct sk_buff *skb = (struct sk_buff *)(long) r1;
2045 struct bpf_map *map = (struct bpf_map *)(long) r2;
2046 u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2047 void *meta = (void *)(long) r4;
2049 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2051 if (unlikely(skb_size > skb->len))
2054 return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2058 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2059 .func = bpf_skb_event_output,
2061 .ret_type = RET_INTEGER,
2062 .arg1_type = ARG_PTR_TO_CTX,
2063 .arg2_type = ARG_CONST_MAP_PTR,
2064 .arg3_type = ARG_ANYTHING,
2065 .arg4_type = ARG_PTR_TO_STACK,
2066 .arg5_type = ARG_CONST_STACK_SIZE,
2069 static unsigned short bpf_tunnel_key_af(u64 flags)
2071 return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2074 static u64 bpf_skb_get_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
2076 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2077 struct bpf_tunnel_key *to = (struct bpf_tunnel_key *) (long) r2;
2078 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2079 u8 compat[sizeof(struct bpf_tunnel_key)];
2083 if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2087 if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2091 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2094 case offsetof(struct bpf_tunnel_key, tunnel_label):
2095 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2097 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2098 /* Fixup deprecated structure layouts here, so we have
2099 * a common path later on.
2101 if (ip_tunnel_info_af(info) != AF_INET)
2104 to = (struct bpf_tunnel_key *)compat;
2111 to->tunnel_id = be64_to_cpu(info->key.tun_id);
2112 to->tunnel_tos = info->key.tos;
2113 to->tunnel_ttl = info->key.ttl;
2115 if (flags & BPF_F_TUNINFO_IPV6) {
2116 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2117 sizeof(to->remote_ipv6));
2118 to->tunnel_label = be32_to_cpu(info->key.label);
2120 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2123 if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2124 memcpy(to_orig, to, size);
2128 memset(to_orig, 0, size);
2132 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2133 .func = bpf_skb_get_tunnel_key,
2135 .ret_type = RET_INTEGER,
2136 .arg1_type = ARG_PTR_TO_CTX,
2137 .arg2_type = ARG_PTR_TO_RAW_STACK,
2138 .arg3_type = ARG_CONST_STACK_SIZE,
2139 .arg4_type = ARG_ANYTHING,
2142 static u64 bpf_skb_get_tunnel_opt(u64 r1, u64 r2, u64 size, u64 r4, u64 r5)
2144 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2145 u8 *to = (u8 *) (long) r2;
2146 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2149 if (unlikely(!info ||
2150 !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2154 if (unlikely(size < info->options_len)) {
2159 ip_tunnel_info_opts_get(to, info);
2160 if (size > info->options_len)
2161 memset(to + info->options_len, 0, size - info->options_len);
2163 return info->options_len;
2165 memset(to, 0, size);
2169 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2170 .func = bpf_skb_get_tunnel_opt,
2172 .ret_type = RET_INTEGER,
2173 .arg1_type = ARG_PTR_TO_CTX,
2174 .arg2_type = ARG_PTR_TO_RAW_STACK,
2175 .arg3_type = ARG_CONST_STACK_SIZE,
2178 static struct metadata_dst __percpu *md_dst;
2180 static u64 bpf_skb_set_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
2182 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2183 struct bpf_tunnel_key *from = (struct bpf_tunnel_key *) (long) r2;
2184 struct metadata_dst *md = this_cpu_ptr(md_dst);
2185 u8 compat[sizeof(struct bpf_tunnel_key)];
2186 struct ip_tunnel_info *info;
2188 if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2189 BPF_F_DONT_FRAGMENT)))
2191 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2193 case offsetof(struct bpf_tunnel_key, tunnel_label):
2194 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2195 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2196 /* Fixup deprecated structure layouts here, so we have
2197 * a common path later on.
2199 memcpy(compat, from, size);
2200 memset(compat + size, 0, sizeof(compat) - size);
2201 from = (struct bpf_tunnel_key *)compat;
2207 if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
2212 dst_hold((struct dst_entry *) md);
2213 skb_dst_set(skb, (struct dst_entry *) md);
2215 info = &md->u.tun_info;
2216 info->mode = IP_TUNNEL_INFO_TX;
2218 info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
2219 if (flags & BPF_F_DONT_FRAGMENT)
2220 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
2222 info->key.tun_id = cpu_to_be64(from->tunnel_id);
2223 info->key.tos = from->tunnel_tos;
2224 info->key.ttl = from->tunnel_ttl;
2226 if (flags & BPF_F_TUNINFO_IPV6) {
2227 info->mode |= IP_TUNNEL_INFO_IPV6;
2228 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
2229 sizeof(from->remote_ipv6));
2230 info->key.label = cpu_to_be32(from->tunnel_label) &
2231 IPV6_FLOWLABEL_MASK;
2233 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
2234 if (flags & BPF_F_ZERO_CSUM_TX)
2235 info->key.tun_flags &= ~TUNNEL_CSUM;
2241 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
2242 .func = bpf_skb_set_tunnel_key,
2244 .ret_type = RET_INTEGER,
2245 .arg1_type = ARG_PTR_TO_CTX,
2246 .arg2_type = ARG_PTR_TO_STACK,
2247 .arg3_type = ARG_CONST_STACK_SIZE,
2248 .arg4_type = ARG_ANYTHING,
2251 static u64 bpf_skb_set_tunnel_opt(u64 r1, u64 r2, u64 size, u64 r4, u64 r5)
2253 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2254 u8 *from = (u8 *) (long) r2;
2255 struct ip_tunnel_info *info = skb_tunnel_info(skb);
2256 const struct metadata_dst *md = this_cpu_ptr(md_dst);
2258 if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
2260 if (unlikely(size > IP_TUNNEL_OPTS_MAX))
2263 ip_tunnel_info_opts_set(info, from, size);
2268 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
2269 .func = bpf_skb_set_tunnel_opt,
2271 .ret_type = RET_INTEGER,
2272 .arg1_type = ARG_PTR_TO_CTX,
2273 .arg2_type = ARG_PTR_TO_STACK,
2274 .arg3_type = ARG_CONST_STACK_SIZE,
2277 static const struct bpf_func_proto *
2278 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
2281 /* Race is not possible, since it's called from verifier
2282 * that is holding verifier mutex.
2284 md_dst = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
2291 case BPF_FUNC_skb_set_tunnel_key:
2292 return &bpf_skb_set_tunnel_key_proto;
2293 case BPF_FUNC_skb_set_tunnel_opt:
2294 return &bpf_skb_set_tunnel_opt_proto;
2300 #ifdef CONFIG_SOCK_CGROUP_DATA
2301 static u64 bpf_skb_in_cgroup(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
2303 struct sk_buff *skb = (struct sk_buff *)(long)r1;
2304 struct bpf_map *map = (struct bpf_map *)(long)r2;
2305 struct bpf_array *array = container_of(map, struct bpf_array, map);
2306 struct cgroup *cgrp;
2311 if (!sk || !sk_fullsock(sk))
2314 if (unlikely(i >= array->map.max_entries))
2317 cgrp = READ_ONCE(array->ptrs[i]);
2318 if (unlikely(!cgrp))
2321 return cgroup_is_descendant(sock_cgroup_ptr(&sk->sk_cgrp_data), cgrp);
2324 static const struct bpf_func_proto bpf_skb_in_cgroup_proto = {
2325 .func = bpf_skb_in_cgroup,
2327 .ret_type = RET_INTEGER,
2328 .arg1_type = ARG_PTR_TO_CTX,
2329 .arg2_type = ARG_CONST_MAP_PTR,
2330 .arg3_type = ARG_ANYTHING,
2334 static const struct bpf_func_proto *
2335 sk_filter_func_proto(enum bpf_func_id func_id)
2338 case BPF_FUNC_map_lookup_elem:
2339 return &bpf_map_lookup_elem_proto;
2340 case BPF_FUNC_map_update_elem:
2341 return &bpf_map_update_elem_proto;
2342 case BPF_FUNC_map_delete_elem:
2343 return &bpf_map_delete_elem_proto;
2344 case BPF_FUNC_get_prandom_u32:
2345 return &bpf_get_prandom_u32_proto;
2346 case BPF_FUNC_get_smp_processor_id:
2347 return &bpf_get_raw_smp_processor_id_proto;
2348 case BPF_FUNC_tail_call:
2349 return &bpf_tail_call_proto;
2350 case BPF_FUNC_ktime_get_ns:
2351 return &bpf_ktime_get_ns_proto;
2352 case BPF_FUNC_trace_printk:
2353 if (capable(CAP_SYS_ADMIN))
2354 return bpf_get_trace_printk_proto();
2360 static const struct bpf_func_proto *
2361 tc_cls_act_func_proto(enum bpf_func_id func_id)
2364 case BPF_FUNC_skb_store_bytes:
2365 return &bpf_skb_store_bytes_proto;
2366 case BPF_FUNC_skb_load_bytes:
2367 return &bpf_skb_load_bytes_proto;
2368 case BPF_FUNC_csum_diff:
2369 return &bpf_csum_diff_proto;
2370 case BPF_FUNC_l3_csum_replace:
2371 return &bpf_l3_csum_replace_proto;
2372 case BPF_FUNC_l4_csum_replace:
2373 return &bpf_l4_csum_replace_proto;
2374 case BPF_FUNC_clone_redirect:
2375 return &bpf_clone_redirect_proto;
2376 case BPF_FUNC_get_cgroup_classid:
2377 return &bpf_get_cgroup_classid_proto;
2378 case BPF_FUNC_skb_vlan_push:
2379 return &bpf_skb_vlan_push_proto;
2380 case BPF_FUNC_skb_vlan_pop:
2381 return &bpf_skb_vlan_pop_proto;
2382 case BPF_FUNC_skb_change_proto:
2383 return &bpf_skb_change_proto_proto;
2384 case BPF_FUNC_skb_change_type:
2385 return &bpf_skb_change_type_proto;
2386 case BPF_FUNC_skb_get_tunnel_key:
2387 return &bpf_skb_get_tunnel_key_proto;
2388 case BPF_FUNC_skb_set_tunnel_key:
2389 return bpf_get_skb_set_tunnel_proto(func_id);
2390 case BPF_FUNC_skb_get_tunnel_opt:
2391 return &bpf_skb_get_tunnel_opt_proto;
2392 case BPF_FUNC_skb_set_tunnel_opt:
2393 return bpf_get_skb_set_tunnel_proto(func_id);
2394 case BPF_FUNC_redirect:
2395 return &bpf_redirect_proto;
2396 case BPF_FUNC_get_route_realm:
2397 return &bpf_get_route_realm_proto;
2398 case BPF_FUNC_get_hash_recalc:
2399 return &bpf_get_hash_recalc_proto;
2400 case BPF_FUNC_perf_event_output:
2401 return &bpf_skb_event_output_proto;
2402 case BPF_FUNC_get_smp_processor_id:
2403 return &bpf_get_smp_processor_id_proto;
2404 #ifdef CONFIG_SOCK_CGROUP_DATA
2405 case BPF_FUNC_skb_in_cgroup:
2406 return &bpf_skb_in_cgroup_proto;
2409 return sk_filter_func_proto(func_id);
2413 static const struct bpf_func_proto *
2414 xdp_func_proto(enum bpf_func_id func_id)
2416 return sk_filter_func_proto(func_id);
2419 static bool __is_valid_access(int off, int size, enum bpf_access_type type)
2421 if (off < 0 || off >= sizeof(struct __sk_buff))
2423 /* The verifier guarantees that size > 0. */
2424 if (off % size != 0)
2426 if (size != sizeof(__u32))
2432 static bool sk_filter_is_valid_access(int off, int size,
2433 enum bpf_access_type type,
2434 enum bpf_reg_type *reg_type)
2437 case offsetof(struct __sk_buff, tc_classid):
2438 case offsetof(struct __sk_buff, data):
2439 case offsetof(struct __sk_buff, data_end):
2443 if (type == BPF_WRITE) {
2445 case offsetof(struct __sk_buff, cb[0]) ...
2446 offsetof(struct __sk_buff, cb[4]):
2453 return __is_valid_access(off, size, type);
2456 static bool tc_cls_act_is_valid_access(int off, int size,
2457 enum bpf_access_type type,
2458 enum bpf_reg_type *reg_type)
2460 if (type == BPF_WRITE) {
2462 case offsetof(struct __sk_buff, mark):
2463 case offsetof(struct __sk_buff, tc_index):
2464 case offsetof(struct __sk_buff, priority):
2465 case offsetof(struct __sk_buff, cb[0]) ...
2466 offsetof(struct __sk_buff, cb[4]):
2467 case offsetof(struct __sk_buff, tc_classid):
2475 case offsetof(struct __sk_buff, data):
2476 *reg_type = PTR_TO_PACKET;
2478 case offsetof(struct __sk_buff, data_end):
2479 *reg_type = PTR_TO_PACKET_END;
2483 return __is_valid_access(off, size, type);
2486 static bool __is_valid_xdp_access(int off, int size,
2487 enum bpf_access_type type)
2489 if (off < 0 || off >= sizeof(struct xdp_md))
2491 if (off % size != 0)
2499 static bool xdp_is_valid_access(int off, int size,
2500 enum bpf_access_type type,
2501 enum bpf_reg_type *reg_type)
2503 if (type == BPF_WRITE)
2507 case offsetof(struct xdp_md, data):
2508 *reg_type = PTR_TO_PACKET;
2510 case offsetof(struct xdp_md, data_end):
2511 *reg_type = PTR_TO_PACKET_END;
2515 return __is_valid_xdp_access(off, size, type);
2518 void bpf_warn_invalid_xdp_action(u32 act)
2520 WARN_ONCE(1, "Illegal XDP return value %u, expect packet loss\n", act);
2522 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
2524 static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
2525 int src_reg, int ctx_off,
2526 struct bpf_insn *insn_buf,
2527 struct bpf_prog *prog)
2529 struct bpf_insn *insn = insn_buf;
2532 case offsetof(struct __sk_buff, len):
2533 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
2535 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2536 offsetof(struct sk_buff, len));
2539 case offsetof(struct __sk_buff, protocol):
2540 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
2542 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2543 offsetof(struct sk_buff, protocol));
2546 case offsetof(struct __sk_buff, vlan_proto):
2547 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
2549 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2550 offsetof(struct sk_buff, vlan_proto));
2553 case offsetof(struct __sk_buff, priority):
2554 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
2556 if (type == BPF_WRITE)
2557 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2558 offsetof(struct sk_buff, priority));
2560 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2561 offsetof(struct sk_buff, priority));
2564 case offsetof(struct __sk_buff, ingress_ifindex):
2565 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
2567 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2568 offsetof(struct sk_buff, skb_iif));
2571 case offsetof(struct __sk_buff, ifindex):
2572 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
2574 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
2576 offsetof(struct sk_buff, dev));
2577 *insn++ = BPF_JMP_IMM(BPF_JEQ, dst_reg, 0, 1);
2578 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
2579 offsetof(struct net_device, ifindex));
2582 case offsetof(struct __sk_buff, hash):
2583 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
2585 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2586 offsetof(struct sk_buff, hash));
2589 case offsetof(struct __sk_buff, mark):
2590 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
2592 if (type == BPF_WRITE)
2593 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2594 offsetof(struct sk_buff, mark));
2596 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2597 offsetof(struct sk_buff, mark));
2600 case offsetof(struct __sk_buff, pkt_type):
2601 return convert_skb_access(SKF_AD_PKTTYPE, dst_reg, src_reg, insn);
2603 case offsetof(struct __sk_buff, queue_mapping):
2604 return convert_skb_access(SKF_AD_QUEUE, dst_reg, src_reg, insn);
2606 case offsetof(struct __sk_buff, vlan_present):
2607 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
2608 dst_reg, src_reg, insn);
2610 case offsetof(struct __sk_buff, vlan_tci):
2611 return convert_skb_access(SKF_AD_VLAN_TAG,
2612 dst_reg, src_reg, insn);
2614 case offsetof(struct __sk_buff, cb[0]) ...
2615 offsetof(struct __sk_buff, cb[4]):
2616 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
2618 prog->cb_access = 1;
2619 ctx_off -= offsetof(struct __sk_buff, cb[0]);
2620 ctx_off += offsetof(struct sk_buff, cb);
2621 ctx_off += offsetof(struct qdisc_skb_cb, data);
2622 if (type == BPF_WRITE)
2623 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2625 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2628 case offsetof(struct __sk_buff, tc_classid):
2629 ctx_off -= offsetof(struct __sk_buff, tc_classid);
2630 ctx_off += offsetof(struct sk_buff, cb);
2631 ctx_off += offsetof(struct qdisc_skb_cb, tc_classid);
2632 if (type == BPF_WRITE)
2633 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2635 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2638 case offsetof(struct __sk_buff, data):
2639 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, data)),
2641 offsetof(struct sk_buff, data));
2644 case offsetof(struct __sk_buff, data_end):
2645 ctx_off -= offsetof(struct __sk_buff, data_end);
2646 ctx_off += offsetof(struct sk_buff, cb);
2647 ctx_off += offsetof(struct bpf_skb_data_end, data_end);
2648 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(sizeof(void *)),
2649 dst_reg, src_reg, ctx_off);
2652 case offsetof(struct __sk_buff, tc_index):
2653 #ifdef CONFIG_NET_SCHED
2654 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
2656 if (type == BPF_WRITE)
2657 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg,
2658 offsetof(struct sk_buff, tc_index));
2660 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2661 offsetof(struct sk_buff, tc_index));
2664 if (type == BPF_WRITE)
2665 *insn++ = BPF_MOV64_REG(dst_reg, dst_reg);
2667 *insn++ = BPF_MOV64_IMM(dst_reg, 0);
2672 return insn - insn_buf;
2675 static u32 xdp_convert_ctx_access(enum bpf_access_type type, int dst_reg,
2676 int src_reg, int ctx_off,
2677 struct bpf_insn *insn_buf,
2678 struct bpf_prog *prog)
2680 struct bpf_insn *insn = insn_buf;
2683 case offsetof(struct xdp_md, data):
2684 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct xdp_buff, data)),
2686 offsetof(struct xdp_buff, data));
2688 case offsetof(struct xdp_md, data_end):
2689 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct xdp_buff, data_end)),
2691 offsetof(struct xdp_buff, data_end));
2695 return insn - insn_buf;
2698 static const struct bpf_verifier_ops sk_filter_ops = {
2699 .get_func_proto = sk_filter_func_proto,
2700 .is_valid_access = sk_filter_is_valid_access,
2701 .convert_ctx_access = bpf_net_convert_ctx_access,
2704 static const struct bpf_verifier_ops tc_cls_act_ops = {
2705 .get_func_proto = tc_cls_act_func_proto,
2706 .is_valid_access = tc_cls_act_is_valid_access,
2707 .convert_ctx_access = bpf_net_convert_ctx_access,
2710 static const struct bpf_verifier_ops xdp_ops = {
2711 .get_func_proto = xdp_func_proto,
2712 .is_valid_access = xdp_is_valid_access,
2713 .convert_ctx_access = xdp_convert_ctx_access,
2716 static struct bpf_prog_type_list sk_filter_type __read_mostly = {
2717 .ops = &sk_filter_ops,
2718 .type = BPF_PROG_TYPE_SOCKET_FILTER,
2721 static struct bpf_prog_type_list sched_cls_type __read_mostly = {
2722 .ops = &tc_cls_act_ops,
2723 .type = BPF_PROG_TYPE_SCHED_CLS,
2726 static struct bpf_prog_type_list sched_act_type __read_mostly = {
2727 .ops = &tc_cls_act_ops,
2728 .type = BPF_PROG_TYPE_SCHED_ACT,
2731 static struct bpf_prog_type_list xdp_type __read_mostly = {
2733 .type = BPF_PROG_TYPE_XDP,
2736 static int __init register_sk_filter_ops(void)
2738 bpf_register_prog_type(&sk_filter_type);
2739 bpf_register_prog_type(&sched_cls_type);
2740 bpf_register_prog_type(&sched_act_type);
2741 bpf_register_prog_type(&xdp_type);
2745 late_initcall(register_sk_filter_ops);
2747 int sk_detach_filter(struct sock *sk)
2750 struct sk_filter *filter;
2752 if (sock_flag(sk, SOCK_FILTER_LOCKED))
2755 filter = rcu_dereference_protected(sk->sk_filter,
2756 lockdep_sock_is_held(sk));
2758 RCU_INIT_POINTER(sk->sk_filter, NULL);
2759 sk_filter_uncharge(sk, filter);
2765 EXPORT_SYMBOL_GPL(sk_detach_filter);
2767 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
2770 struct sock_fprog_kern *fprog;
2771 struct sk_filter *filter;
2775 filter = rcu_dereference_protected(sk->sk_filter,
2776 lockdep_sock_is_held(sk));
2780 /* We're copying the filter that has been originally attached,
2781 * so no conversion/decode needed anymore. eBPF programs that
2782 * have no original program cannot be dumped through this.
2785 fprog = filter->prog->orig_prog;
2791 /* User space only enquires number of filter blocks. */
2795 if (len < fprog->len)
2799 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
2802 /* Instead of bytes, the API requests to return the number