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>
55 * sk_filter_trim_cap - run a packet through a socket filter
56 * @sk: sock associated with &sk_buff
57 * @skb: buffer to filter
58 * @cap: limit on how short the eBPF program may trim the packet
60 * Run the eBPF program and then cut skb->data to correct size returned by
61 * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
62 * than pkt_len we keep whole skb->data. This is the socket level
63 * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
64 * be accepted or -EPERM if the packet should be tossed.
67 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
70 struct sk_filter *filter;
73 * If the skb was allocated from pfmemalloc reserves, only
74 * allow SOCK_MEMALLOC sockets to use it as this socket is
77 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC))
80 err = security_sock_rcv_skb(sk, skb);
85 filter = rcu_dereference(sk->sk_filter);
87 unsigned int pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
88 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
94 EXPORT_SYMBOL(sk_filter_trim_cap);
96 static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
98 return skb_get_poff((struct sk_buff *)(unsigned long) ctx);
101 static u64 __skb_get_nlattr(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
103 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
106 if (skb_is_nonlinear(skb))
109 if (skb->len < sizeof(struct nlattr))
112 if (a > skb->len - sizeof(struct nlattr))
115 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
117 return (void *) nla - (void *) skb->data;
122 static u64 __skb_get_nlattr_nest(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
124 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
127 if (skb_is_nonlinear(skb))
130 if (skb->len < sizeof(struct nlattr))
133 if (a > skb->len - sizeof(struct nlattr))
136 nla = (struct nlattr *) &skb->data[a];
137 if (nla->nla_len > skb->len - a)
140 nla = nla_find_nested(nla, x);
142 return (void *) nla - (void *) skb->data;
147 static u64 __get_raw_cpu_id(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
149 return raw_smp_processor_id();
152 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
153 struct bpf_insn *insn_buf)
155 struct bpf_insn *insn = insn_buf;
159 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
161 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
162 offsetof(struct sk_buff, mark));
166 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
167 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
168 #ifdef __BIG_ENDIAN_BITFIELD
169 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
174 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
176 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
177 offsetof(struct sk_buff, queue_mapping));
180 case SKF_AD_VLAN_TAG:
181 case SKF_AD_VLAN_TAG_PRESENT:
182 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
183 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
185 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
186 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
187 offsetof(struct sk_buff, vlan_tci));
188 if (skb_field == SKF_AD_VLAN_TAG) {
189 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
193 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
195 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
200 return insn - insn_buf;
203 static bool convert_bpf_extensions(struct sock_filter *fp,
204 struct bpf_insn **insnp)
206 struct bpf_insn *insn = *insnp;
210 case SKF_AD_OFF + SKF_AD_PROTOCOL:
211 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
213 /* A = *(u16 *) (CTX + offsetof(protocol)) */
214 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
215 offsetof(struct sk_buff, protocol));
216 /* A = ntohs(A) [emitting a nop or swap16] */
217 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
220 case SKF_AD_OFF + SKF_AD_PKTTYPE:
221 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
225 case SKF_AD_OFF + SKF_AD_IFINDEX:
226 case SKF_AD_OFF + SKF_AD_HATYPE:
227 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
228 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
229 BUILD_BUG_ON(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)) < 0);
231 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
232 BPF_REG_TMP, BPF_REG_CTX,
233 offsetof(struct sk_buff, dev));
234 /* if (tmp != 0) goto pc + 1 */
235 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
236 *insn++ = BPF_EXIT_INSN();
237 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
238 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
239 offsetof(struct net_device, ifindex));
241 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
242 offsetof(struct net_device, type));
245 case SKF_AD_OFF + SKF_AD_MARK:
246 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
250 case SKF_AD_OFF + SKF_AD_RXHASH:
251 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
253 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
254 offsetof(struct sk_buff, hash));
257 case SKF_AD_OFF + SKF_AD_QUEUE:
258 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
262 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
263 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
264 BPF_REG_A, BPF_REG_CTX, insn);
268 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
269 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
270 BPF_REG_A, BPF_REG_CTX, insn);
274 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
275 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
277 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
278 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
279 offsetof(struct sk_buff, vlan_proto));
280 /* A = ntohs(A) [emitting a nop or swap16] */
281 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
284 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
285 case SKF_AD_OFF + SKF_AD_NLATTR:
286 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
287 case SKF_AD_OFF + SKF_AD_CPU:
288 case SKF_AD_OFF + SKF_AD_RANDOM:
290 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
292 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
294 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
295 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
297 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
298 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
300 case SKF_AD_OFF + SKF_AD_NLATTR:
301 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
303 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
304 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
306 case SKF_AD_OFF + SKF_AD_CPU:
307 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
309 case SKF_AD_OFF + SKF_AD_RANDOM:
310 *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
311 bpf_user_rnd_init_once();
316 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
318 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
322 /* This is just a dummy call to avoid letting the compiler
323 * evict __bpf_call_base() as an optimization. Placed here
324 * where no-one bothers.
326 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
335 * bpf_convert_filter - convert filter program
336 * @prog: the user passed filter program
337 * @len: the length of the user passed filter program
338 * @new_prog: buffer where converted program will be stored
339 * @new_len: pointer to store length of converted program
341 * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style.
342 * Conversion workflow:
344 * 1) First pass for calculating the new program length:
345 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
347 * 2) 2nd pass to remap in two passes: 1st pass finds new
348 * jump offsets, 2nd pass remapping:
349 * new_prog = kmalloc(sizeof(struct bpf_insn) * new_len);
350 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
352 * User BPF's register A is mapped to our BPF register 6, user BPF
353 * register X is mapped to BPF register 7; frame pointer is always
354 * register 10; Context 'void *ctx' is stored in register 1, that is,
355 * for socket filters: ctx == 'struct sk_buff *', for seccomp:
356 * ctx == 'struct seccomp_data *'.
358 static int bpf_convert_filter(struct sock_filter *prog, int len,
359 struct bpf_insn *new_prog, int *new_len)
361 int new_flen = 0, pass = 0, target, i;
362 struct bpf_insn *new_insn;
363 struct sock_filter *fp;
367 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
368 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
370 if (len <= 0 || len > BPF_MAXINSNS)
374 addrs = kcalloc(len, sizeof(*addrs),
375 GFP_KERNEL | __GFP_NOWARN);
385 *new_insn = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
388 for (i = 0; i < len; fp++, i++) {
389 struct bpf_insn tmp_insns[6] = { };
390 struct bpf_insn *insn = tmp_insns;
393 addrs[i] = new_insn - new_prog;
396 /* All arithmetic insns and skb loads map as-is. */
397 case BPF_ALU | BPF_ADD | BPF_X:
398 case BPF_ALU | BPF_ADD | BPF_K:
399 case BPF_ALU | BPF_SUB | BPF_X:
400 case BPF_ALU | BPF_SUB | BPF_K:
401 case BPF_ALU | BPF_AND | BPF_X:
402 case BPF_ALU | BPF_AND | BPF_K:
403 case BPF_ALU | BPF_OR | BPF_X:
404 case BPF_ALU | BPF_OR | BPF_K:
405 case BPF_ALU | BPF_LSH | BPF_X:
406 case BPF_ALU | BPF_LSH | BPF_K:
407 case BPF_ALU | BPF_RSH | BPF_X:
408 case BPF_ALU | BPF_RSH | BPF_K:
409 case BPF_ALU | BPF_XOR | BPF_X:
410 case BPF_ALU | BPF_XOR | BPF_K:
411 case BPF_ALU | BPF_MUL | BPF_X:
412 case BPF_ALU | BPF_MUL | BPF_K:
413 case BPF_ALU | BPF_DIV | BPF_X:
414 case BPF_ALU | BPF_DIV | BPF_K:
415 case BPF_ALU | BPF_MOD | BPF_X:
416 case BPF_ALU | BPF_MOD | BPF_K:
417 case BPF_ALU | BPF_NEG:
418 case BPF_LD | BPF_ABS | BPF_W:
419 case BPF_LD | BPF_ABS | BPF_H:
420 case BPF_LD | BPF_ABS | BPF_B:
421 case BPF_LD | BPF_IND | BPF_W:
422 case BPF_LD | BPF_IND | BPF_H:
423 case BPF_LD | BPF_IND | BPF_B:
424 /* Check for overloaded BPF extension and
425 * directly convert it if found, otherwise
426 * just move on with mapping.
428 if (BPF_CLASS(fp->code) == BPF_LD &&
429 BPF_MODE(fp->code) == BPF_ABS &&
430 convert_bpf_extensions(fp, &insn))
433 if (fp->code == (BPF_ALU | BPF_DIV | BPF_X) ||
434 fp->code == (BPF_ALU | BPF_MOD | BPF_X))
435 *insn++ = BPF_MOV32_REG(BPF_REG_X, BPF_REG_X);
437 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
440 /* Jump transformation cannot use BPF block macros
441 * everywhere as offset calculation and target updates
442 * require a bit more work than the rest, i.e. jump
443 * opcodes map as-is, but offsets need adjustment.
446 #define BPF_EMIT_JMP \
448 if (target >= len || target < 0) \
450 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
451 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
452 insn->off -= insn - tmp_insns; \
455 case BPF_JMP | BPF_JA:
456 target = i + fp->k + 1;
457 insn->code = fp->code;
461 case BPF_JMP | BPF_JEQ | BPF_K:
462 case BPF_JMP | BPF_JEQ | BPF_X:
463 case BPF_JMP | BPF_JSET | BPF_K:
464 case BPF_JMP | BPF_JSET | BPF_X:
465 case BPF_JMP | BPF_JGT | BPF_K:
466 case BPF_JMP | BPF_JGT | BPF_X:
467 case BPF_JMP | BPF_JGE | BPF_K:
468 case BPF_JMP | BPF_JGE | BPF_X:
469 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
470 /* BPF immediates are signed, zero extend
471 * immediate into tmp register and use it
474 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
476 insn->dst_reg = BPF_REG_A;
477 insn->src_reg = BPF_REG_TMP;
480 insn->dst_reg = BPF_REG_A;
482 bpf_src = BPF_SRC(fp->code);
483 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
486 /* Common case where 'jump_false' is next insn. */
488 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
489 target = i + fp->jt + 1;
494 /* Convert JEQ into JNE when 'jump_true' is next insn. */
495 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
496 insn->code = BPF_JMP | BPF_JNE | bpf_src;
497 target = i + fp->jf + 1;
502 /* Other jumps are mapped into two insns: Jxx and JA. */
503 target = i + fp->jt + 1;
504 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
508 insn->code = BPF_JMP | BPF_JA;
509 target = i + fp->jf + 1;
513 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
514 case BPF_LDX | BPF_MSH | BPF_B:
516 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
517 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
518 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
520 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
522 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
524 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
526 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
529 /* RET_K, RET_A are remaped into 2 insns. */
530 case BPF_RET | BPF_A:
531 case BPF_RET | BPF_K:
532 *insn++ = BPF_MOV32_RAW(BPF_RVAL(fp->code) == BPF_K ?
533 BPF_K : BPF_X, BPF_REG_0,
535 *insn = BPF_EXIT_INSN();
538 /* Store to stack. */
541 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
542 BPF_ST ? BPF_REG_A : BPF_REG_X,
543 -(BPF_MEMWORDS - fp->k) * 4);
546 /* Load from stack. */
547 case BPF_LD | BPF_MEM:
548 case BPF_LDX | BPF_MEM:
549 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
550 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
551 -(BPF_MEMWORDS - fp->k) * 4);
555 case BPF_LD | BPF_IMM:
556 case BPF_LDX | BPF_IMM:
557 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
558 BPF_REG_A : BPF_REG_X, fp->k);
562 case BPF_MISC | BPF_TAX:
563 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
567 case BPF_MISC | BPF_TXA:
568 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
571 /* A = skb->len or X = skb->len */
572 case BPF_LD | BPF_W | BPF_LEN:
573 case BPF_LDX | BPF_W | BPF_LEN:
574 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
575 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
576 offsetof(struct sk_buff, len));
579 /* Access seccomp_data fields. */
580 case BPF_LDX | BPF_ABS | BPF_W:
581 /* A = *(u32 *) (ctx + K) */
582 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
585 /* Unknown instruction. */
592 memcpy(new_insn, tmp_insns,
593 sizeof(*insn) * (insn - tmp_insns));
594 new_insn += insn - tmp_insns;
598 /* Only calculating new length. */
599 *new_len = new_insn - new_prog;
604 if (new_flen != new_insn - new_prog) {
605 new_flen = new_insn - new_prog;
612 BUG_ON(*new_len != new_flen);
621 * As we dont want to clear mem[] array for each packet going through
622 * __bpf_prog_run(), we check that filter loaded by user never try to read
623 * a cell if not previously written, and we check all branches to be sure
624 * a malicious user doesn't try to abuse us.
626 static int check_load_and_stores(const struct sock_filter *filter, int flen)
628 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
631 BUILD_BUG_ON(BPF_MEMWORDS > 16);
633 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
637 memset(masks, 0xff, flen * sizeof(*masks));
639 for (pc = 0; pc < flen; pc++) {
640 memvalid &= masks[pc];
642 switch (filter[pc].code) {
645 memvalid |= (1 << filter[pc].k);
647 case BPF_LD | BPF_MEM:
648 case BPF_LDX | BPF_MEM:
649 if (!(memvalid & (1 << filter[pc].k))) {
654 case BPF_JMP | BPF_JA:
655 /* A jump must set masks on target */
656 masks[pc + 1 + filter[pc].k] &= memvalid;
659 case BPF_JMP | BPF_JEQ | BPF_K:
660 case BPF_JMP | BPF_JEQ | BPF_X:
661 case BPF_JMP | BPF_JGE | BPF_K:
662 case BPF_JMP | BPF_JGE | BPF_X:
663 case BPF_JMP | BPF_JGT | BPF_K:
664 case BPF_JMP | BPF_JGT | BPF_X:
665 case BPF_JMP | BPF_JSET | BPF_K:
666 case BPF_JMP | BPF_JSET | BPF_X:
667 /* A jump must set masks on targets */
668 masks[pc + 1 + filter[pc].jt] &= memvalid;
669 masks[pc + 1 + filter[pc].jf] &= memvalid;
679 static bool chk_code_allowed(u16 code_to_probe)
681 static const bool codes[] = {
682 /* 32 bit ALU operations */
683 [BPF_ALU | BPF_ADD | BPF_K] = true,
684 [BPF_ALU | BPF_ADD | BPF_X] = true,
685 [BPF_ALU | BPF_SUB | BPF_K] = true,
686 [BPF_ALU | BPF_SUB | BPF_X] = true,
687 [BPF_ALU | BPF_MUL | BPF_K] = true,
688 [BPF_ALU | BPF_MUL | BPF_X] = true,
689 [BPF_ALU | BPF_DIV | BPF_K] = true,
690 [BPF_ALU | BPF_DIV | BPF_X] = true,
691 [BPF_ALU | BPF_MOD | BPF_K] = true,
692 [BPF_ALU | BPF_MOD | BPF_X] = true,
693 [BPF_ALU | BPF_AND | BPF_K] = true,
694 [BPF_ALU | BPF_AND | BPF_X] = true,
695 [BPF_ALU | BPF_OR | BPF_K] = true,
696 [BPF_ALU | BPF_OR | BPF_X] = true,
697 [BPF_ALU | BPF_XOR | BPF_K] = true,
698 [BPF_ALU | BPF_XOR | BPF_X] = true,
699 [BPF_ALU | BPF_LSH | BPF_K] = true,
700 [BPF_ALU | BPF_LSH | BPF_X] = true,
701 [BPF_ALU | BPF_RSH | BPF_K] = true,
702 [BPF_ALU | BPF_RSH | BPF_X] = true,
703 [BPF_ALU | BPF_NEG] = true,
704 /* Load instructions */
705 [BPF_LD | BPF_W | BPF_ABS] = true,
706 [BPF_LD | BPF_H | BPF_ABS] = true,
707 [BPF_LD | BPF_B | BPF_ABS] = true,
708 [BPF_LD | BPF_W | BPF_LEN] = true,
709 [BPF_LD | BPF_W | BPF_IND] = true,
710 [BPF_LD | BPF_H | BPF_IND] = true,
711 [BPF_LD | BPF_B | BPF_IND] = true,
712 [BPF_LD | BPF_IMM] = true,
713 [BPF_LD | BPF_MEM] = true,
714 [BPF_LDX | BPF_W | BPF_LEN] = true,
715 [BPF_LDX | BPF_B | BPF_MSH] = true,
716 [BPF_LDX | BPF_IMM] = true,
717 [BPF_LDX | BPF_MEM] = true,
718 /* Store instructions */
721 /* Misc instructions */
722 [BPF_MISC | BPF_TAX] = true,
723 [BPF_MISC | BPF_TXA] = true,
724 /* Return instructions */
725 [BPF_RET | BPF_K] = true,
726 [BPF_RET | BPF_A] = true,
727 /* Jump instructions */
728 [BPF_JMP | BPF_JA] = true,
729 [BPF_JMP | BPF_JEQ | BPF_K] = true,
730 [BPF_JMP | BPF_JEQ | BPF_X] = true,
731 [BPF_JMP | BPF_JGE | BPF_K] = true,
732 [BPF_JMP | BPF_JGE | BPF_X] = true,
733 [BPF_JMP | BPF_JGT | BPF_K] = true,
734 [BPF_JMP | BPF_JGT | BPF_X] = true,
735 [BPF_JMP | BPF_JSET | BPF_K] = true,
736 [BPF_JMP | BPF_JSET | BPF_X] = true,
739 if (code_to_probe >= ARRAY_SIZE(codes))
742 return codes[code_to_probe];
745 static bool bpf_check_basics_ok(const struct sock_filter *filter,
750 if (flen == 0 || flen > BPF_MAXINSNS)
757 * bpf_check_classic - verify socket filter code
758 * @filter: filter to verify
759 * @flen: length of filter
761 * Check the user's filter code. If we let some ugly
762 * filter code slip through kaboom! The filter must contain
763 * no references or jumps that are out of range, no illegal
764 * instructions, and must end with a RET instruction.
766 * All jumps are forward as they are not signed.
768 * Returns 0 if the rule set is legal or -EINVAL if not.
770 static int bpf_check_classic(const struct sock_filter *filter,
776 /* Check the filter code now */
777 for (pc = 0; pc < flen; pc++) {
778 const struct sock_filter *ftest = &filter[pc];
780 /* May we actually operate on this code? */
781 if (!chk_code_allowed(ftest->code))
784 /* Some instructions need special checks */
785 switch (ftest->code) {
786 case BPF_ALU | BPF_DIV | BPF_K:
787 case BPF_ALU | BPF_MOD | BPF_K:
788 /* Check for division by zero */
792 case BPF_ALU | BPF_LSH | BPF_K:
793 case BPF_ALU | BPF_RSH | BPF_K:
797 case BPF_LD | BPF_MEM:
798 case BPF_LDX | BPF_MEM:
801 /* Check for invalid memory addresses */
802 if (ftest->k >= BPF_MEMWORDS)
805 case BPF_JMP | BPF_JA:
806 /* Note, the large ftest->k might cause loops.
807 * Compare this with conditional jumps below,
808 * where offsets are limited. --ANK (981016)
810 if (ftest->k >= (unsigned int)(flen - pc - 1))
813 case BPF_JMP | BPF_JEQ | BPF_K:
814 case BPF_JMP | BPF_JEQ | BPF_X:
815 case BPF_JMP | BPF_JGE | BPF_K:
816 case BPF_JMP | BPF_JGE | BPF_X:
817 case BPF_JMP | BPF_JGT | BPF_K:
818 case BPF_JMP | BPF_JGT | BPF_X:
819 case BPF_JMP | BPF_JSET | BPF_K:
820 case BPF_JMP | BPF_JSET | BPF_X:
821 /* Both conditionals must be safe */
822 if (pc + ftest->jt + 1 >= flen ||
823 pc + ftest->jf + 1 >= flen)
826 case BPF_LD | BPF_W | BPF_ABS:
827 case BPF_LD | BPF_H | BPF_ABS:
828 case BPF_LD | BPF_B | BPF_ABS:
830 if (bpf_anc_helper(ftest) & BPF_ANC)
832 /* Ancillary operation unknown or unsupported */
833 if (anc_found == false && ftest->k >= SKF_AD_OFF)
838 /* Last instruction must be a RET code */
839 switch (filter[flen - 1].code) {
840 case BPF_RET | BPF_K:
841 case BPF_RET | BPF_A:
842 return check_load_and_stores(filter, flen);
848 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
849 const struct sock_fprog *fprog)
851 unsigned int fsize = bpf_classic_proglen(fprog);
852 struct sock_fprog_kern *fkprog;
854 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
858 fkprog = fp->orig_prog;
859 fkprog->len = fprog->len;
861 fkprog->filter = kmemdup(fp->insns, fsize,
862 GFP_KERNEL | __GFP_NOWARN);
863 if (!fkprog->filter) {
864 kfree(fp->orig_prog);
871 static void bpf_release_orig_filter(struct bpf_prog *fp)
873 struct sock_fprog_kern *fprog = fp->orig_prog;
876 kfree(fprog->filter);
881 static void __bpf_prog_release(struct bpf_prog *prog)
883 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
886 bpf_release_orig_filter(prog);
891 static void __sk_filter_release(struct sk_filter *fp)
893 __bpf_prog_release(fp->prog);
898 * sk_filter_release_rcu - Release a socket filter by rcu_head
899 * @rcu: rcu_head that contains the sk_filter to free
901 static void sk_filter_release_rcu(struct rcu_head *rcu)
903 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
905 __sk_filter_release(fp);
909 * sk_filter_release - release a socket filter
910 * @fp: filter to remove
912 * Remove a filter from a socket and release its resources.
914 static void sk_filter_release(struct sk_filter *fp)
916 if (atomic_dec_and_test(&fp->refcnt))
917 call_rcu(&fp->rcu, sk_filter_release_rcu);
920 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
922 u32 filter_size = bpf_prog_size(fp->prog->len);
924 atomic_sub(filter_size, &sk->sk_omem_alloc);
925 sk_filter_release(fp);
928 /* try to charge the socket memory if there is space available
929 * return true on success
931 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
933 u32 filter_size = bpf_prog_size(fp->prog->len);
935 /* same check as in sock_kmalloc() */
936 if (filter_size <= sysctl_optmem_max &&
937 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
938 atomic_inc(&fp->refcnt);
939 atomic_add(filter_size, &sk->sk_omem_alloc);
945 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
947 struct sock_filter *old_prog;
948 struct bpf_prog *old_fp;
949 int err, new_len, old_len = fp->len;
951 /* We are free to overwrite insns et al right here as it
952 * won't be used at this point in time anymore internally
953 * after the migration to the internal BPF instruction
956 BUILD_BUG_ON(sizeof(struct sock_filter) !=
957 sizeof(struct bpf_insn));
959 /* Conversion cannot happen on overlapping memory areas,
960 * so we need to keep the user BPF around until the 2nd
961 * pass. At this time, the user BPF is stored in fp->insns.
963 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
964 GFP_KERNEL | __GFP_NOWARN);
970 /* 1st pass: calculate the new program length. */
971 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
975 /* Expand fp for appending the new filter representation. */
977 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
979 /* The old_fp is still around in case we couldn't
980 * allocate new memory, so uncharge on that one.
989 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
990 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
992 /* 2nd bpf_convert_filter() can fail only if it fails
993 * to allocate memory, remapping must succeed. Note,
994 * that at this time old_fp has already been released
999 err = bpf_prog_select_runtime(fp);
1009 __bpf_prog_release(fp);
1010 return ERR_PTR(err);
1013 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1014 bpf_aux_classic_check_t trans)
1018 fp->bpf_func = NULL;
1021 err = bpf_check_classic(fp->insns, fp->len);
1023 __bpf_prog_release(fp);
1024 return ERR_PTR(err);
1027 /* There might be additional checks and transformations
1028 * needed on classic filters, f.e. in case of seccomp.
1031 err = trans(fp->insns, fp->len);
1033 __bpf_prog_release(fp);
1034 return ERR_PTR(err);
1038 /* Probe if we can JIT compile the filter and if so, do
1039 * the compilation of the filter.
1041 bpf_jit_compile(fp);
1043 /* JIT compiler couldn't process this filter, so do the
1044 * internal BPF translation for the optimized interpreter.
1047 fp = bpf_migrate_filter(fp);
1053 * bpf_prog_create - create an unattached filter
1054 * @pfp: the unattached filter that is created
1055 * @fprog: the filter program
1057 * Create a filter independent of any socket. We first run some
1058 * sanity checks on it to make sure it does not explode on us later.
1059 * If an error occurs or there is insufficient memory for the filter
1060 * a negative errno code is returned. On success the return is zero.
1062 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1064 unsigned int fsize = bpf_classic_proglen(fprog);
1065 struct bpf_prog *fp;
1067 /* Make sure new filter is there and in the right amounts. */
1068 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1071 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1075 memcpy(fp->insns, fprog->filter, fsize);
1077 fp->len = fprog->len;
1078 /* Since unattached filters are not copied back to user
1079 * space through sk_get_filter(), we do not need to hold
1080 * a copy here, and can spare us the work.
1082 fp->orig_prog = NULL;
1084 /* bpf_prepare_filter() already takes care of freeing
1085 * memory in case something goes wrong.
1087 fp = bpf_prepare_filter(fp, NULL);
1094 EXPORT_SYMBOL_GPL(bpf_prog_create);
1097 * bpf_prog_create_from_user - create an unattached filter from user buffer
1098 * @pfp: the unattached filter that is created
1099 * @fprog: the filter program
1100 * @trans: post-classic verifier transformation handler
1101 * @save_orig: save classic BPF program
1103 * This function effectively does the same as bpf_prog_create(), only
1104 * that it builds up its insns buffer from user space provided buffer.
1105 * It also allows for passing a bpf_aux_classic_check_t handler.
1107 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1108 bpf_aux_classic_check_t trans, bool save_orig)
1110 unsigned int fsize = bpf_classic_proglen(fprog);
1111 struct bpf_prog *fp;
1114 /* Make sure new filter is there and in the right amounts. */
1115 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1118 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1122 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1123 __bpf_prog_free(fp);
1127 fp->len = fprog->len;
1128 fp->orig_prog = NULL;
1131 err = bpf_prog_store_orig_filter(fp, fprog);
1133 __bpf_prog_free(fp);
1138 /* bpf_prepare_filter() already takes care of freeing
1139 * memory in case something goes wrong.
1141 fp = bpf_prepare_filter(fp, trans);
1148 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1150 void bpf_prog_destroy(struct bpf_prog *fp)
1152 __bpf_prog_release(fp);
1154 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1156 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk,
1159 struct sk_filter *fp, *old_fp;
1161 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1166 atomic_set(&fp->refcnt, 0);
1168 if (!sk_filter_charge(sk, fp)) {
1173 old_fp = rcu_dereference_protected(sk->sk_filter, locked);
1174 rcu_assign_pointer(sk->sk_filter, fp);
1176 sk_filter_uncharge(sk, old_fp);
1182 * sk_attach_filter - attach a socket filter
1183 * @fprog: the filter program
1184 * @sk: the socket to use
1186 * Attach the user's filter code. We first run some sanity checks on
1187 * it to make sure it does not explode on us later. If an error
1188 * occurs or there is insufficient memory for the filter a negative
1189 * errno code is returned. On success the return is zero.
1191 int __sk_attach_filter(struct sock_fprog *fprog, struct sock *sk,
1194 unsigned int fsize = bpf_classic_proglen(fprog);
1195 struct bpf_prog *prog;
1198 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1201 /* Make sure new filter is there and in the right amounts. */
1202 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1205 prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1209 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1210 __bpf_prog_free(prog);
1214 prog->len = fprog->len;
1216 err = bpf_prog_store_orig_filter(prog, fprog);
1218 __bpf_prog_free(prog);
1222 /* bpf_prepare_filter() already takes care of freeing
1223 * memory in case something goes wrong.
1225 prog = bpf_prepare_filter(prog, NULL);
1227 return PTR_ERR(prog);
1229 err = __sk_attach_prog(prog, sk, locked);
1231 __bpf_prog_release(prog);
1237 EXPORT_SYMBOL_GPL(__sk_attach_filter);
1239 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1241 return __sk_attach_filter(fprog, sk, sock_owned_by_user(sk));
1244 int sk_attach_bpf(u32 ufd, struct sock *sk)
1246 struct bpf_prog *prog;
1249 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1252 prog = bpf_prog_get(ufd);
1254 return PTR_ERR(prog);
1256 if (prog->type != BPF_PROG_TYPE_SOCKET_FILTER) {
1261 err = __sk_attach_prog(prog, sk, sock_owned_by_user(sk));
1270 #define BPF_RECOMPUTE_CSUM(flags) ((flags) & 1)
1272 static u64 bpf_skb_store_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 flags)
1274 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1275 int offset = (int) r2;
1276 void *from = (void *) (long) r3;
1277 unsigned int len = (unsigned int) r4;
1281 /* bpf verifier guarantees that:
1282 * 'from' pointer points to bpf program stack
1283 * 'len' bytes of it were initialized
1285 * 'skb' is a valid pointer to 'struct sk_buff'
1287 * so check for invalid 'offset' and too large 'len'
1289 if (unlikely((u32) offset > 0xffff || len > sizeof(buf)))
1291 if (unlikely(skb_try_make_writable(skb, offset + len)))
1294 ptr = skb_header_pointer(skb, offset, len, buf);
1298 if (BPF_RECOMPUTE_CSUM(flags))
1299 skb_postpull_rcsum(skb, ptr, len);
1301 memcpy(ptr, from, len);
1304 /* skb_store_bits cannot return -EFAULT here */
1305 skb_store_bits(skb, offset, ptr, len);
1307 if (BPF_RECOMPUTE_CSUM(flags) && skb->ip_summed == CHECKSUM_COMPLETE)
1308 skb->csum = csum_add(skb->csum, csum_partial(ptr, len, 0));
1312 const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1313 .func = bpf_skb_store_bytes,
1315 .ret_type = RET_INTEGER,
1316 .arg1_type = ARG_PTR_TO_CTX,
1317 .arg2_type = ARG_ANYTHING,
1318 .arg3_type = ARG_PTR_TO_STACK,
1319 .arg4_type = ARG_CONST_STACK_SIZE,
1320 .arg5_type = ARG_ANYTHING,
1323 #define BPF_HEADER_FIELD_SIZE(flags) ((flags) & 0x0f)
1324 #define BPF_IS_PSEUDO_HEADER(flags) ((flags) & 0x10)
1326 static u64 bpf_l3_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1328 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1329 int offset = (int) r2;
1332 if (unlikely((u32) offset > 0xffff))
1335 if (unlikely(skb_try_make_writable(skb, offset + sizeof(sum))))
1338 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1342 switch (BPF_HEADER_FIELD_SIZE(flags)) {
1344 csum_replace2(ptr, from, to);
1347 csum_replace4(ptr, from, to);
1354 /* skb_store_bits guaranteed to not return -EFAULT here */
1355 skb_store_bits(skb, offset, ptr, sizeof(sum));
1360 const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1361 .func = bpf_l3_csum_replace,
1363 .ret_type = RET_INTEGER,
1364 .arg1_type = ARG_PTR_TO_CTX,
1365 .arg2_type = ARG_ANYTHING,
1366 .arg3_type = ARG_ANYTHING,
1367 .arg4_type = ARG_ANYTHING,
1368 .arg5_type = ARG_ANYTHING,
1371 static u64 bpf_l4_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1373 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1374 bool is_pseudo = !!BPF_IS_PSEUDO_HEADER(flags);
1375 int offset = (int) r2;
1378 if (unlikely((u32) offset > 0xffff))
1380 if (unlikely(skb_try_make_writable(skb, offset + sizeof(sum))))
1383 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1387 switch (BPF_HEADER_FIELD_SIZE(flags)) {
1389 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1392 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1399 /* skb_store_bits guaranteed to not return -EFAULT here */
1400 skb_store_bits(skb, offset, ptr, sizeof(sum));
1405 const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1406 .func = bpf_l4_csum_replace,
1408 .ret_type = RET_INTEGER,
1409 .arg1_type = ARG_PTR_TO_CTX,
1410 .arg2_type = ARG_ANYTHING,
1411 .arg3_type = ARG_ANYTHING,
1412 .arg4_type = ARG_ANYTHING,
1413 .arg5_type = ARG_ANYTHING,
1416 #define BPF_IS_REDIRECT_INGRESS(flags) ((flags) & 1)
1418 static u64 bpf_clone_redirect(u64 r1, u64 ifindex, u64 flags, u64 r4, u64 r5)
1420 struct sk_buff *skb = (struct sk_buff *) (long) r1, *skb2;
1421 struct net_device *dev;
1423 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1427 skb2 = skb_clone(skb, GFP_ATOMIC);
1428 if (unlikely(!skb2))
1431 if (BPF_IS_REDIRECT_INGRESS(flags))
1432 return dev_forward_skb(dev, skb2);
1435 skb_sender_cpu_clear(skb2);
1436 return dev_queue_xmit(skb2);
1439 const struct bpf_func_proto bpf_clone_redirect_proto = {
1440 .func = bpf_clone_redirect,
1442 .ret_type = RET_INTEGER,
1443 .arg1_type = ARG_PTR_TO_CTX,
1444 .arg2_type = ARG_ANYTHING,
1445 .arg3_type = ARG_ANYTHING,
1448 struct redirect_info {
1453 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1454 static u64 bpf_redirect(u64 ifindex, u64 flags, u64 r3, u64 r4, u64 r5)
1456 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1458 ri->ifindex = ifindex;
1460 return TC_ACT_REDIRECT;
1463 int skb_do_redirect(struct sk_buff *skb)
1465 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1466 struct net_device *dev;
1468 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1470 if (unlikely(!dev)) {
1475 if (BPF_IS_REDIRECT_INGRESS(ri->flags))
1476 return dev_forward_skb(dev, skb);
1479 skb_sender_cpu_clear(skb);
1480 return dev_queue_xmit(skb);
1483 const struct bpf_func_proto bpf_redirect_proto = {
1484 .func = bpf_redirect,
1486 .ret_type = RET_INTEGER,
1487 .arg1_type = ARG_ANYTHING,
1488 .arg2_type = ARG_ANYTHING,
1491 static u64 bpf_get_cgroup_classid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1493 return task_get_classid((struct sk_buff *) (unsigned long) r1);
1496 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1497 .func = bpf_get_cgroup_classid,
1499 .ret_type = RET_INTEGER,
1500 .arg1_type = ARG_PTR_TO_CTX,
1503 static u64 bpf_get_route_realm(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1505 #ifdef CONFIG_IP_ROUTE_CLASSID
1506 const struct dst_entry *dst;
1508 dst = skb_dst((struct sk_buff *) (unsigned long) r1);
1510 return dst->tclassid;
1515 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1516 .func = bpf_get_route_realm,
1518 .ret_type = RET_INTEGER,
1519 .arg1_type = ARG_PTR_TO_CTX,
1522 static u64 bpf_skb_vlan_push(u64 r1, u64 r2, u64 vlan_tci, u64 r4, u64 r5)
1524 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1525 __be16 vlan_proto = (__force __be16) r2;
1527 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1528 vlan_proto != htons(ETH_P_8021AD)))
1529 vlan_proto = htons(ETH_P_8021Q);
1531 return skb_vlan_push(skb, vlan_proto, vlan_tci);
1534 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1535 .func = bpf_skb_vlan_push,
1537 .ret_type = RET_INTEGER,
1538 .arg1_type = ARG_PTR_TO_CTX,
1539 .arg2_type = ARG_ANYTHING,
1540 .arg3_type = ARG_ANYTHING,
1542 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1544 static u64 bpf_skb_vlan_pop(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1546 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1548 return skb_vlan_pop(skb);
1551 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1552 .func = bpf_skb_vlan_pop,
1554 .ret_type = RET_INTEGER,
1555 .arg1_type = ARG_PTR_TO_CTX,
1557 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1559 bool bpf_helper_changes_skb_data(void *func)
1561 if (func == bpf_skb_vlan_push)
1563 if (func == bpf_skb_vlan_pop)
1565 if (func == bpf_skb_store_bytes)
1567 if (func == bpf_l3_csum_replace)
1569 if (func == bpf_l4_csum_replace)
1575 static u64 bpf_skb_get_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
1577 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1578 struct bpf_tunnel_key *to = (struct bpf_tunnel_key *) (long) r2;
1579 struct ip_tunnel_info *info = skb_tunnel_info(skb);
1581 if (unlikely(size != sizeof(struct bpf_tunnel_key) || flags || !info))
1583 if (ip_tunnel_info_af(info) != AF_INET)
1586 to->tunnel_id = be64_to_cpu(info->key.tun_id);
1587 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
1592 const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
1593 .func = bpf_skb_get_tunnel_key,
1595 .ret_type = RET_INTEGER,
1596 .arg1_type = ARG_PTR_TO_CTX,
1597 .arg2_type = ARG_PTR_TO_STACK,
1598 .arg3_type = ARG_CONST_STACK_SIZE,
1599 .arg4_type = ARG_ANYTHING,
1602 static struct metadata_dst __percpu *md_dst;
1604 static u64 bpf_skb_set_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
1606 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1607 struct bpf_tunnel_key *from = (struct bpf_tunnel_key *) (long) r2;
1608 struct metadata_dst *md = this_cpu_ptr(md_dst);
1609 struct ip_tunnel_info *info;
1611 if (unlikely(size != sizeof(struct bpf_tunnel_key) || flags))
1615 dst_hold((struct dst_entry *) md);
1616 skb_dst_set(skb, (struct dst_entry *) md);
1618 info = &md->u.tun_info;
1619 info->mode = IP_TUNNEL_INFO_TX;
1620 info->key.tun_flags = TUNNEL_KEY;
1621 info->key.tun_id = cpu_to_be64(from->tunnel_id);
1622 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
1627 const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
1628 .func = bpf_skb_set_tunnel_key,
1630 .ret_type = RET_INTEGER,
1631 .arg1_type = ARG_PTR_TO_CTX,
1632 .arg2_type = ARG_PTR_TO_STACK,
1633 .arg3_type = ARG_CONST_STACK_SIZE,
1634 .arg4_type = ARG_ANYTHING,
1637 static const struct bpf_func_proto *bpf_get_skb_set_tunnel_key_proto(void)
1640 /* race is not possible, since it's called from
1641 * verifier that is holding verifier mutex
1643 md_dst = metadata_dst_alloc_percpu(0, GFP_KERNEL);
1647 return &bpf_skb_set_tunnel_key_proto;
1650 static const struct bpf_func_proto *
1651 sk_filter_func_proto(enum bpf_func_id func_id)
1654 case BPF_FUNC_map_lookup_elem:
1655 return &bpf_map_lookup_elem_proto;
1656 case BPF_FUNC_map_update_elem:
1657 return &bpf_map_update_elem_proto;
1658 case BPF_FUNC_map_delete_elem:
1659 return &bpf_map_delete_elem_proto;
1660 case BPF_FUNC_get_prandom_u32:
1661 return &bpf_get_prandom_u32_proto;
1662 case BPF_FUNC_get_smp_processor_id:
1663 return &bpf_get_smp_processor_id_proto;
1664 case BPF_FUNC_tail_call:
1665 return &bpf_tail_call_proto;
1666 case BPF_FUNC_ktime_get_ns:
1667 return &bpf_ktime_get_ns_proto;
1668 case BPF_FUNC_trace_printk:
1669 if (capable(CAP_SYS_ADMIN))
1670 return bpf_get_trace_printk_proto();
1676 static const struct bpf_func_proto *
1677 tc_cls_act_func_proto(enum bpf_func_id func_id)
1680 case BPF_FUNC_skb_store_bytes:
1681 return &bpf_skb_store_bytes_proto;
1682 case BPF_FUNC_l3_csum_replace:
1683 return &bpf_l3_csum_replace_proto;
1684 case BPF_FUNC_l4_csum_replace:
1685 return &bpf_l4_csum_replace_proto;
1686 case BPF_FUNC_clone_redirect:
1687 return &bpf_clone_redirect_proto;
1688 case BPF_FUNC_get_cgroup_classid:
1689 return &bpf_get_cgroup_classid_proto;
1690 case BPF_FUNC_skb_vlan_push:
1691 return &bpf_skb_vlan_push_proto;
1692 case BPF_FUNC_skb_vlan_pop:
1693 return &bpf_skb_vlan_pop_proto;
1694 case BPF_FUNC_skb_get_tunnel_key:
1695 return &bpf_skb_get_tunnel_key_proto;
1696 case BPF_FUNC_skb_set_tunnel_key:
1697 return bpf_get_skb_set_tunnel_key_proto();
1698 case BPF_FUNC_redirect:
1699 return &bpf_redirect_proto;
1700 case BPF_FUNC_get_route_realm:
1701 return &bpf_get_route_realm_proto;
1703 return sk_filter_func_proto(func_id);
1707 static bool __is_valid_access(int off, int size, enum bpf_access_type type)
1710 if (off < 0 || off >= sizeof(struct __sk_buff))
1713 /* disallow misaligned access */
1714 if (off % size != 0)
1717 /* all __sk_buff fields are __u32 */
1724 static bool sk_filter_is_valid_access(int off, int size,
1725 enum bpf_access_type type)
1727 if (off == offsetof(struct __sk_buff, tc_classid))
1730 if (type == BPF_WRITE) {
1732 case offsetof(struct __sk_buff, cb[0]) ...
1733 offsetof(struct __sk_buff, cb[4]):
1740 return __is_valid_access(off, size, type);
1743 static bool tc_cls_act_is_valid_access(int off, int size,
1744 enum bpf_access_type type)
1746 if (off == offsetof(struct __sk_buff, tc_classid))
1747 return type == BPF_WRITE ? true : false;
1749 if (type == BPF_WRITE) {
1751 case offsetof(struct __sk_buff, mark):
1752 case offsetof(struct __sk_buff, tc_index):
1753 case offsetof(struct __sk_buff, priority):
1754 case offsetof(struct __sk_buff, cb[0]) ...
1755 offsetof(struct __sk_buff, cb[4]):
1761 return __is_valid_access(off, size, type);
1764 static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
1765 int src_reg, int ctx_off,
1766 struct bpf_insn *insn_buf,
1767 struct bpf_prog *prog)
1769 struct bpf_insn *insn = insn_buf;
1772 case offsetof(struct __sk_buff, len):
1773 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
1775 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1776 offsetof(struct sk_buff, len));
1779 case offsetof(struct __sk_buff, protocol):
1780 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
1782 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1783 offsetof(struct sk_buff, protocol));
1786 case offsetof(struct __sk_buff, vlan_proto):
1787 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
1789 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1790 offsetof(struct sk_buff, vlan_proto));
1793 case offsetof(struct __sk_buff, priority):
1794 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
1796 if (type == BPF_WRITE)
1797 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
1798 offsetof(struct sk_buff, priority));
1800 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1801 offsetof(struct sk_buff, priority));
1804 case offsetof(struct __sk_buff, ingress_ifindex):
1805 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
1807 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1808 offsetof(struct sk_buff, skb_iif));
1811 case offsetof(struct __sk_buff, ifindex):
1812 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
1814 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
1816 offsetof(struct sk_buff, dev));
1817 *insn++ = BPF_JMP_IMM(BPF_JEQ, dst_reg, 0, 1);
1818 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
1819 offsetof(struct net_device, ifindex));
1822 case offsetof(struct __sk_buff, hash):
1823 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
1825 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1826 offsetof(struct sk_buff, hash));
1829 case offsetof(struct __sk_buff, mark):
1830 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
1832 if (type == BPF_WRITE)
1833 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
1834 offsetof(struct sk_buff, mark));
1836 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1837 offsetof(struct sk_buff, mark));
1840 case offsetof(struct __sk_buff, pkt_type):
1841 return convert_skb_access(SKF_AD_PKTTYPE, dst_reg, src_reg, insn);
1843 case offsetof(struct __sk_buff, queue_mapping):
1844 return convert_skb_access(SKF_AD_QUEUE, dst_reg, src_reg, insn);
1846 case offsetof(struct __sk_buff, vlan_present):
1847 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
1848 dst_reg, src_reg, insn);
1850 case offsetof(struct __sk_buff, vlan_tci):
1851 return convert_skb_access(SKF_AD_VLAN_TAG,
1852 dst_reg, src_reg, insn);
1854 case offsetof(struct __sk_buff, cb[0]) ...
1855 offsetof(struct __sk_buff, cb[4]):
1856 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
1858 prog->cb_access = 1;
1859 ctx_off -= offsetof(struct __sk_buff, cb[0]);
1860 ctx_off += offsetof(struct sk_buff, cb);
1861 ctx_off += offsetof(struct qdisc_skb_cb, data);
1862 if (type == BPF_WRITE)
1863 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
1865 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
1868 case offsetof(struct __sk_buff, tc_classid):
1869 ctx_off -= offsetof(struct __sk_buff, tc_classid);
1870 ctx_off += offsetof(struct sk_buff, cb);
1871 ctx_off += offsetof(struct qdisc_skb_cb, tc_classid);
1872 WARN_ON(type != BPF_WRITE);
1873 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
1876 case offsetof(struct __sk_buff, tc_index):
1877 #ifdef CONFIG_NET_SCHED
1878 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
1880 if (type == BPF_WRITE)
1881 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg,
1882 offsetof(struct sk_buff, tc_index));
1884 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1885 offsetof(struct sk_buff, tc_index));
1888 if (type == BPF_WRITE)
1889 *insn++ = BPF_MOV64_REG(dst_reg, dst_reg);
1891 *insn++ = BPF_MOV64_IMM(dst_reg, 0);
1896 return insn - insn_buf;
1899 static const struct bpf_verifier_ops sk_filter_ops = {
1900 .get_func_proto = sk_filter_func_proto,
1901 .is_valid_access = sk_filter_is_valid_access,
1902 .convert_ctx_access = bpf_net_convert_ctx_access,
1905 static const struct bpf_verifier_ops tc_cls_act_ops = {
1906 .get_func_proto = tc_cls_act_func_proto,
1907 .is_valid_access = tc_cls_act_is_valid_access,
1908 .convert_ctx_access = bpf_net_convert_ctx_access,
1911 static struct bpf_prog_type_list sk_filter_type __read_mostly = {
1912 .ops = &sk_filter_ops,
1913 .type = BPF_PROG_TYPE_SOCKET_FILTER,
1916 static struct bpf_prog_type_list sched_cls_type __read_mostly = {
1917 .ops = &tc_cls_act_ops,
1918 .type = BPF_PROG_TYPE_SCHED_CLS,
1921 static struct bpf_prog_type_list sched_act_type __read_mostly = {
1922 .ops = &tc_cls_act_ops,
1923 .type = BPF_PROG_TYPE_SCHED_ACT,
1926 static int __init register_sk_filter_ops(void)
1928 bpf_register_prog_type(&sk_filter_type);
1929 bpf_register_prog_type(&sched_cls_type);
1930 bpf_register_prog_type(&sched_act_type);
1934 late_initcall(register_sk_filter_ops);
1936 int __sk_detach_filter(struct sock *sk, bool locked)
1939 struct sk_filter *filter;
1941 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1944 filter = rcu_dereference_protected(sk->sk_filter, locked);
1946 RCU_INIT_POINTER(sk->sk_filter, NULL);
1947 sk_filter_uncharge(sk, filter);
1953 EXPORT_SYMBOL_GPL(__sk_detach_filter);
1955 int sk_detach_filter(struct sock *sk)
1957 return __sk_detach_filter(sk, sock_owned_by_user(sk));
1960 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
1963 struct sock_fprog_kern *fprog;
1964 struct sk_filter *filter;
1968 filter = rcu_dereference_protected(sk->sk_filter,
1969 sock_owned_by_user(sk));
1973 /* We're copying the filter that has been originally attached,
1974 * so no conversion/decode needed anymore. eBPF programs that
1975 * have no original program cannot be dumped through this.
1978 fprog = filter->prog->orig_prog;
1984 /* User space only enquires number of filter blocks. */
1988 if (len < fprog->len)
1992 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
1995 /* Instead of bytes, the API requests to return the number