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 - run a packet through a socket filter
57 * @sk: sock associated with &sk_buff
58 * @skb: buffer to filter
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(struct sock *sk, struct sk_buff *skb)
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);
89 err = pkt_len ? pskb_trim(skb, pkt_len) : -EPERM;
95 EXPORT_SYMBOL(sk_filter);
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 u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
154 struct bpf_insn *insn_buf)
156 struct bpf_insn *insn = insn_buf;
160 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
162 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
163 offsetof(struct sk_buff, mark));
167 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
168 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
169 #ifdef __BIG_ENDIAN_BITFIELD
170 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
175 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
177 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
178 offsetof(struct sk_buff, queue_mapping));
181 case SKF_AD_VLAN_TAG:
182 case SKF_AD_VLAN_TAG_PRESENT:
183 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
184 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
186 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
187 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
188 offsetof(struct sk_buff, vlan_tci));
189 if (skb_field == SKF_AD_VLAN_TAG) {
190 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
194 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
196 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
201 return insn - insn_buf;
204 static bool convert_bpf_extensions(struct sock_filter *fp,
205 struct bpf_insn **insnp)
207 struct bpf_insn *insn = *insnp;
211 case SKF_AD_OFF + SKF_AD_PROTOCOL:
212 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
214 /* A = *(u16 *) (CTX + offsetof(protocol)) */
215 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
216 offsetof(struct sk_buff, protocol));
217 /* A = ntohs(A) [emitting a nop or swap16] */
218 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
221 case SKF_AD_OFF + SKF_AD_PKTTYPE:
222 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
226 case SKF_AD_OFF + SKF_AD_IFINDEX:
227 case SKF_AD_OFF + SKF_AD_HATYPE:
228 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
229 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
230 BUILD_BUG_ON(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)) < 0);
232 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
233 BPF_REG_TMP, BPF_REG_CTX,
234 offsetof(struct sk_buff, dev));
235 /* if (tmp != 0) goto pc + 1 */
236 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
237 *insn++ = BPF_EXIT_INSN();
238 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
239 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
240 offsetof(struct net_device, ifindex));
242 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
243 offsetof(struct net_device, type));
246 case SKF_AD_OFF + SKF_AD_MARK:
247 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
251 case SKF_AD_OFF + SKF_AD_RXHASH:
252 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
254 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
255 offsetof(struct sk_buff, hash));
258 case SKF_AD_OFF + SKF_AD_QUEUE:
259 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
263 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
264 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
265 BPF_REG_A, BPF_REG_CTX, insn);
269 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
270 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
271 BPF_REG_A, BPF_REG_CTX, insn);
275 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
276 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
278 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
279 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
280 offsetof(struct sk_buff, vlan_proto));
281 /* A = ntohs(A) [emitting a nop or swap16] */
282 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
285 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
286 case SKF_AD_OFF + SKF_AD_NLATTR:
287 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
288 case SKF_AD_OFF + SKF_AD_CPU:
289 case SKF_AD_OFF + SKF_AD_RANDOM:
291 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
293 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
295 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
296 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
298 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
299 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
301 case SKF_AD_OFF + SKF_AD_NLATTR:
302 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
304 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
305 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
307 case SKF_AD_OFF + SKF_AD_CPU:
308 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
310 case SKF_AD_OFF + SKF_AD_RANDOM:
311 *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
312 bpf_user_rnd_init_once();
317 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
319 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
323 /* This is just a dummy call to avoid letting the compiler
324 * evict __bpf_call_base() as an optimization. Placed here
325 * where no-one bothers.
327 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
336 * bpf_convert_filter - convert filter program
337 * @prog: the user passed filter program
338 * @len: the length of the user passed filter program
339 * @new_prog: buffer where converted program will be stored
340 * @new_len: pointer to store length of converted program
342 * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style.
343 * Conversion workflow:
345 * 1) First pass for calculating the new program length:
346 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
348 * 2) 2nd pass to remap in two passes: 1st pass finds new
349 * jump offsets, 2nd pass remapping:
350 * new_prog = kmalloc(sizeof(struct bpf_insn) * new_len);
351 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
353 static int bpf_convert_filter(struct sock_filter *prog, int len,
354 struct bpf_insn *new_prog, int *new_len)
356 int new_flen = 0, pass = 0, target, i;
357 struct bpf_insn *new_insn;
358 struct sock_filter *fp;
362 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
363 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
365 if (len <= 0 || len > BPF_MAXINSNS)
369 addrs = kcalloc(len, sizeof(*addrs),
370 GFP_KERNEL | __GFP_NOWARN);
379 /* Classic BPF related prologue emission. */
381 /* Classic BPF expects A and X to be reset first. These need
382 * to be guaranteed to be the first two instructions.
384 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
385 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
387 /* All programs must keep CTX in callee saved BPF_REG_CTX.
388 * In eBPF case it's done by the compiler, here we need to
389 * do this ourself. Initial CTX is present in BPF_REG_ARG1.
391 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
396 for (i = 0; i < len; fp++, i++) {
397 struct bpf_insn tmp_insns[6] = { };
398 struct bpf_insn *insn = tmp_insns;
401 addrs[i] = new_insn - new_prog;
404 /* All arithmetic insns and skb loads map as-is. */
405 case BPF_ALU | BPF_ADD | BPF_X:
406 case BPF_ALU | BPF_ADD | BPF_K:
407 case BPF_ALU | BPF_SUB | BPF_X:
408 case BPF_ALU | BPF_SUB | BPF_K:
409 case BPF_ALU | BPF_AND | BPF_X:
410 case BPF_ALU | BPF_AND | BPF_K:
411 case BPF_ALU | BPF_OR | BPF_X:
412 case BPF_ALU | BPF_OR | BPF_K:
413 case BPF_ALU | BPF_LSH | BPF_X:
414 case BPF_ALU | BPF_LSH | BPF_K:
415 case BPF_ALU | BPF_RSH | BPF_X:
416 case BPF_ALU | BPF_RSH | BPF_K:
417 case BPF_ALU | BPF_XOR | BPF_X:
418 case BPF_ALU | BPF_XOR | BPF_K:
419 case BPF_ALU | BPF_MUL | BPF_X:
420 case BPF_ALU | BPF_MUL | BPF_K:
421 case BPF_ALU | BPF_DIV | BPF_X:
422 case BPF_ALU | BPF_DIV | BPF_K:
423 case BPF_ALU | BPF_MOD | BPF_X:
424 case BPF_ALU | BPF_MOD | BPF_K:
425 case BPF_ALU | BPF_NEG:
426 case BPF_LD | BPF_ABS | BPF_W:
427 case BPF_LD | BPF_ABS | BPF_H:
428 case BPF_LD | BPF_ABS | BPF_B:
429 case BPF_LD | BPF_IND | BPF_W:
430 case BPF_LD | BPF_IND | BPF_H:
431 case BPF_LD | BPF_IND | BPF_B:
432 /* Check for overloaded BPF extension and
433 * directly convert it if found, otherwise
434 * just move on with mapping.
436 if (BPF_CLASS(fp->code) == BPF_LD &&
437 BPF_MODE(fp->code) == BPF_ABS &&
438 convert_bpf_extensions(fp, &insn))
441 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
444 /* Jump transformation cannot use BPF block macros
445 * everywhere as offset calculation and target updates
446 * require a bit more work than the rest, i.e. jump
447 * opcodes map as-is, but offsets need adjustment.
450 #define BPF_EMIT_JMP \
452 if (target >= len || target < 0) \
454 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
455 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
456 insn->off -= insn - tmp_insns; \
459 case BPF_JMP | BPF_JA:
460 target = i + fp->k + 1;
461 insn->code = fp->code;
465 case BPF_JMP | BPF_JEQ | BPF_K:
466 case BPF_JMP | BPF_JEQ | BPF_X:
467 case BPF_JMP | BPF_JSET | BPF_K:
468 case BPF_JMP | BPF_JSET | BPF_X:
469 case BPF_JMP | BPF_JGT | BPF_K:
470 case BPF_JMP | BPF_JGT | BPF_X:
471 case BPF_JMP | BPF_JGE | BPF_K:
472 case BPF_JMP | BPF_JGE | BPF_X:
473 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
474 /* BPF immediates are signed, zero extend
475 * immediate into tmp register and use it
478 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
480 insn->dst_reg = BPF_REG_A;
481 insn->src_reg = BPF_REG_TMP;
484 insn->dst_reg = BPF_REG_A;
486 bpf_src = BPF_SRC(fp->code);
487 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
490 /* Common case where 'jump_false' is next insn. */
492 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
493 target = i + fp->jt + 1;
498 /* Convert JEQ into JNE when 'jump_true' is next insn. */
499 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
500 insn->code = BPF_JMP | BPF_JNE | bpf_src;
501 target = i + fp->jf + 1;
506 /* Other jumps are mapped into two insns: Jxx and JA. */
507 target = i + fp->jt + 1;
508 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
512 insn->code = BPF_JMP | BPF_JA;
513 target = i + fp->jf + 1;
517 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
518 case BPF_LDX | BPF_MSH | BPF_B:
520 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
521 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
522 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
524 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
526 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
528 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
530 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
533 /* RET_K, RET_A are remaped into 2 insns. */
534 case BPF_RET | BPF_A:
535 case BPF_RET | BPF_K:
536 *insn++ = BPF_MOV32_RAW(BPF_RVAL(fp->code) == BPF_K ?
537 BPF_K : BPF_X, BPF_REG_0,
539 *insn = BPF_EXIT_INSN();
542 /* Store to stack. */
545 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
546 BPF_ST ? BPF_REG_A : BPF_REG_X,
547 -(BPF_MEMWORDS - fp->k) * 4);
550 /* Load from stack. */
551 case BPF_LD | BPF_MEM:
552 case BPF_LDX | BPF_MEM:
553 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
554 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
555 -(BPF_MEMWORDS - fp->k) * 4);
559 case BPF_LD | BPF_IMM:
560 case BPF_LDX | BPF_IMM:
561 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
562 BPF_REG_A : BPF_REG_X, fp->k);
566 case BPF_MISC | BPF_TAX:
567 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
571 case BPF_MISC | BPF_TXA:
572 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
575 /* A = skb->len or X = skb->len */
576 case BPF_LD | BPF_W | BPF_LEN:
577 case BPF_LDX | BPF_W | BPF_LEN:
578 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
579 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
580 offsetof(struct sk_buff, len));
583 /* Access seccomp_data fields. */
584 case BPF_LDX | BPF_ABS | BPF_W:
585 /* A = *(u32 *) (ctx + K) */
586 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
589 /* Unknown instruction. */
596 memcpy(new_insn, tmp_insns,
597 sizeof(*insn) * (insn - tmp_insns));
598 new_insn += insn - tmp_insns;
602 /* Only calculating new length. */
603 *new_len = new_insn - new_prog;
608 if (new_flen != new_insn - new_prog) {
609 new_flen = new_insn - new_prog;
616 BUG_ON(*new_len != new_flen);
625 * As we dont want to clear mem[] array for each packet going through
626 * __bpf_prog_run(), we check that filter loaded by user never try to read
627 * a cell if not previously written, and we check all branches to be sure
628 * a malicious user doesn't try to abuse us.
630 static int check_load_and_stores(const struct sock_filter *filter, int flen)
632 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
635 BUILD_BUG_ON(BPF_MEMWORDS > 16);
637 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
641 memset(masks, 0xff, flen * sizeof(*masks));
643 for (pc = 0; pc < flen; pc++) {
644 memvalid &= masks[pc];
646 switch (filter[pc].code) {
649 memvalid |= (1 << filter[pc].k);
651 case BPF_LD | BPF_MEM:
652 case BPF_LDX | BPF_MEM:
653 if (!(memvalid & (1 << filter[pc].k))) {
658 case BPF_JMP | BPF_JA:
659 /* A jump must set masks on target */
660 masks[pc + 1 + filter[pc].k] &= memvalid;
663 case BPF_JMP | BPF_JEQ | BPF_K:
664 case BPF_JMP | BPF_JEQ | BPF_X:
665 case BPF_JMP | BPF_JGE | BPF_K:
666 case BPF_JMP | BPF_JGE | BPF_X:
667 case BPF_JMP | BPF_JGT | BPF_K:
668 case BPF_JMP | BPF_JGT | BPF_X:
669 case BPF_JMP | BPF_JSET | BPF_K:
670 case BPF_JMP | BPF_JSET | BPF_X:
671 /* A jump must set masks on targets */
672 masks[pc + 1 + filter[pc].jt] &= memvalid;
673 masks[pc + 1 + filter[pc].jf] &= memvalid;
683 static bool chk_code_allowed(u16 code_to_probe)
685 static const bool codes[] = {
686 /* 32 bit ALU operations */
687 [BPF_ALU | BPF_ADD | BPF_K] = true,
688 [BPF_ALU | BPF_ADD | BPF_X] = true,
689 [BPF_ALU | BPF_SUB | BPF_K] = true,
690 [BPF_ALU | BPF_SUB | BPF_X] = true,
691 [BPF_ALU | BPF_MUL | BPF_K] = true,
692 [BPF_ALU | BPF_MUL | BPF_X] = true,
693 [BPF_ALU | BPF_DIV | BPF_K] = true,
694 [BPF_ALU | BPF_DIV | BPF_X] = true,
695 [BPF_ALU | BPF_MOD | BPF_K] = true,
696 [BPF_ALU | BPF_MOD | BPF_X] = true,
697 [BPF_ALU | BPF_AND | BPF_K] = true,
698 [BPF_ALU | BPF_AND | BPF_X] = true,
699 [BPF_ALU | BPF_OR | BPF_K] = true,
700 [BPF_ALU | BPF_OR | BPF_X] = true,
701 [BPF_ALU | BPF_XOR | BPF_K] = true,
702 [BPF_ALU | BPF_XOR | BPF_X] = true,
703 [BPF_ALU | BPF_LSH | BPF_K] = true,
704 [BPF_ALU | BPF_LSH | BPF_X] = true,
705 [BPF_ALU | BPF_RSH | BPF_K] = true,
706 [BPF_ALU | BPF_RSH | BPF_X] = true,
707 [BPF_ALU | BPF_NEG] = true,
708 /* Load instructions */
709 [BPF_LD | BPF_W | BPF_ABS] = true,
710 [BPF_LD | BPF_H | BPF_ABS] = true,
711 [BPF_LD | BPF_B | BPF_ABS] = true,
712 [BPF_LD | BPF_W | BPF_LEN] = true,
713 [BPF_LD | BPF_W | BPF_IND] = true,
714 [BPF_LD | BPF_H | BPF_IND] = true,
715 [BPF_LD | BPF_B | BPF_IND] = true,
716 [BPF_LD | BPF_IMM] = true,
717 [BPF_LD | BPF_MEM] = true,
718 [BPF_LDX | BPF_W | BPF_LEN] = true,
719 [BPF_LDX | BPF_B | BPF_MSH] = true,
720 [BPF_LDX | BPF_IMM] = true,
721 [BPF_LDX | BPF_MEM] = true,
722 /* Store instructions */
725 /* Misc instructions */
726 [BPF_MISC | BPF_TAX] = true,
727 [BPF_MISC | BPF_TXA] = true,
728 /* Return instructions */
729 [BPF_RET | BPF_K] = true,
730 [BPF_RET | BPF_A] = true,
731 /* Jump instructions */
732 [BPF_JMP | BPF_JA] = true,
733 [BPF_JMP | BPF_JEQ | BPF_K] = true,
734 [BPF_JMP | BPF_JEQ | BPF_X] = true,
735 [BPF_JMP | BPF_JGE | BPF_K] = true,
736 [BPF_JMP | BPF_JGE | BPF_X] = true,
737 [BPF_JMP | BPF_JGT | BPF_K] = true,
738 [BPF_JMP | BPF_JGT | BPF_X] = true,
739 [BPF_JMP | BPF_JSET | BPF_K] = true,
740 [BPF_JMP | BPF_JSET | BPF_X] = true,
743 if (code_to_probe >= ARRAY_SIZE(codes))
746 return codes[code_to_probe];
750 * bpf_check_classic - verify socket filter code
751 * @filter: filter to verify
752 * @flen: length of filter
754 * Check the user's filter code. If we let some ugly
755 * filter code slip through kaboom! The filter must contain
756 * no references or jumps that are out of range, no illegal
757 * instructions, and must end with a RET instruction.
759 * All jumps are forward as they are not signed.
761 * Returns 0 if the rule set is legal or -EINVAL if not.
763 static int bpf_check_classic(const struct sock_filter *filter,
769 if (flen == 0 || flen > BPF_MAXINSNS)
772 /* Check the filter code now */
773 for (pc = 0; pc < flen; pc++) {
774 const struct sock_filter *ftest = &filter[pc];
776 /* May we actually operate on this code? */
777 if (!chk_code_allowed(ftest->code))
780 /* Some instructions need special checks */
781 switch (ftest->code) {
782 case BPF_ALU | BPF_DIV | BPF_K:
783 case BPF_ALU | BPF_MOD | BPF_K:
784 /* Check for division by zero */
788 case BPF_ALU | BPF_LSH | BPF_K:
789 case BPF_ALU | BPF_RSH | BPF_K:
793 case BPF_LD | BPF_MEM:
794 case BPF_LDX | BPF_MEM:
797 /* Check for invalid memory addresses */
798 if (ftest->k >= BPF_MEMWORDS)
801 case BPF_JMP | BPF_JA:
802 /* Note, the large ftest->k might cause loops.
803 * Compare this with conditional jumps below,
804 * where offsets are limited. --ANK (981016)
806 if (ftest->k >= (unsigned int)(flen - pc - 1))
809 case BPF_JMP | BPF_JEQ | BPF_K:
810 case BPF_JMP | BPF_JEQ | BPF_X:
811 case BPF_JMP | BPF_JGE | BPF_K:
812 case BPF_JMP | BPF_JGE | BPF_X:
813 case BPF_JMP | BPF_JGT | BPF_K:
814 case BPF_JMP | BPF_JGT | BPF_X:
815 case BPF_JMP | BPF_JSET | BPF_K:
816 case BPF_JMP | BPF_JSET | BPF_X:
817 /* Both conditionals must be safe */
818 if (pc + ftest->jt + 1 >= flen ||
819 pc + ftest->jf + 1 >= flen)
822 case BPF_LD | BPF_W | BPF_ABS:
823 case BPF_LD | BPF_H | BPF_ABS:
824 case BPF_LD | BPF_B | BPF_ABS:
826 if (bpf_anc_helper(ftest) & BPF_ANC)
828 /* Ancillary operation unknown or unsupported */
829 if (anc_found == false && ftest->k >= SKF_AD_OFF)
834 /* Last instruction must be a RET code */
835 switch (filter[flen - 1].code) {
836 case BPF_RET | BPF_K:
837 case BPF_RET | BPF_A:
838 return check_load_and_stores(filter, flen);
844 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
845 const struct sock_fprog *fprog)
847 unsigned int fsize = bpf_classic_proglen(fprog);
848 struct sock_fprog_kern *fkprog;
850 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
854 fkprog = fp->orig_prog;
855 fkprog->len = fprog->len;
857 fkprog->filter = kmemdup(fp->insns, fsize,
858 GFP_KERNEL | __GFP_NOWARN);
859 if (!fkprog->filter) {
860 kfree(fp->orig_prog);
867 static void bpf_release_orig_filter(struct bpf_prog *fp)
869 struct sock_fprog_kern *fprog = fp->orig_prog;
872 kfree(fprog->filter);
877 static void __bpf_prog_release(struct bpf_prog *prog)
879 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
882 bpf_release_orig_filter(prog);
887 static void __sk_filter_release(struct sk_filter *fp)
889 __bpf_prog_release(fp->prog);
894 * sk_filter_release_rcu - Release a socket filter by rcu_head
895 * @rcu: rcu_head that contains the sk_filter to free
897 static void sk_filter_release_rcu(struct rcu_head *rcu)
899 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
901 __sk_filter_release(fp);
905 * sk_filter_release - release a socket filter
906 * @fp: filter to remove
908 * Remove a filter from a socket and release its resources.
910 static void sk_filter_release(struct sk_filter *fp)
912 if (atomic_dec_and_test(&fp->refcnt))
913 call_rcu(&fp->rcu, sk_filter_release_rcu);
916 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
918 u32 filter_size = bpf_prog_size(fp->prog->len);
920 atomic_sub(filter_size, &sk->sk_omem_alloc);
921 sk_filter_release(fp);
924 /* try to charge the socket memory if there is space available
925 * return true on success
927 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
929 u32 filter_size = bpf_prog_size(fp->prog->len);
931 /* same check as in sock_kmalloc() */
932 if (filter_size <= sysctl_optmem_max &&
933 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
934 atomic_inc(&fp->refcnt);
935 atomic_add(filter_size, &sk->sk_omem_alloc);
941 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
943 struct sock_filter *old_prog;
944 struct bpf_prog *old_fp;
945 int err, new_len, old_len = fp->len;
947 /* We are free to overwrite insns et al right here as it
948 * won't be used at this point in time anymore internally
949 * after the migration to the internal BPF instruction
952 BUILD_BUG_ON(sizeof(struct sock_filter) !=
953 sizeof(struct bpf_insn));
955 /* Conversion cannot happen on overlapping memory areas,
956 * so we need to keep the user BPF around until the 2nd
957 * pass. At this time, the user BPF is stored in fp->insns.
959 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
960 GFP_KERNEL | __GFP_NOWARN);
966 /* 1st pass: calculate the new program length. */
967 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
971 /* Expand fp for appending the new filter representation. */
973 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
975 /* The old_fp is still around in case we couldn't
976 * allocate new memory, so uncharge on that one.
985 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
986 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
988 /* 2nd bpf_convert_filter() can fail only if it fails
989 * to allocate memory, remapping must succeed. Note,
990 * that at this time old_fp has already been released
995 bpf_prog_select_runtime(fp);
1003 __bpf_prog_release(fp);
1004 return ERR_PTR(err);
1007 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1008 bpf_aux_classic_check_t trans)
1012 fp->bpf_func = NULL;
1015 err = bpf_check_classic(fp->insns, fp->len);
1017 __bpf_prog_release(fp);
1018 return ERR_PTR(err);
1021 /* There might be additional checks and transformations
1022 * needed on classic filters, f.e. in case of seccomp.
1025 err = trans(fp->insns, fp->len);
1027 __bpf_prog_release(fp);
1028 return ERR_PTR(err);
1032 /* Probe if we can JIT compile the filter and if so, do
1033 * the compilation of the filter.
1035 bpf_jit_compile(fp);
1037 /* JIT compiler couldn't process this filter, so do the
1038 * internal BPF translation for the optimized interpreter.
1041 fp = bpf_migrate_filter(fp);
1047 * bpf_prog_create - create an unattached filter
1048 * @pfp: the unattached filter that is created
1049 * @fprog: the filter program
1051 * Create a filter independent of any socket. We first run some
1052 * sanity checks on it to make sure it does not explode on us later.
1053 * If an error occurs or there is insufficient memory for the filter
1054 * a negative errno code is returned. On success the return is zero.
1056 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1058 unsigned int fsize = bpf_classic_proglen(fprog);
1059 struct bpf_prog *fp;
1061 /* Make sure new filter is there and in the right amounts. */
1062 if (fprog->filter == NULL)
1065 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1069 memcpy(fp->insns, fprog->filter, fsize);
1071 fp->len = fprog->len;
1072 /* Since unattached filters are not copied back to user
1073 * space through sk_get_filter(), we do not need to hold
1074 * a copy here, and can spare us the work.
1076 fp->orig_prog = NULL;
1078 /* bpf_prepare_filter() already takes care of freeing
1079 * memory in case something goes wrong.
1081 fp = bpf_prepare_filter(fp, NULL);
1088 EXPORT_SYMBOL_GPL(bpf_prog_create);
1091 * bpf_prog_create_from_user - create an unattached filter from user buffer
1092 * @pfp: the unattached filter that is created
1093 * @fprog: the filter program
1094 * @trans: post-classic verifier transformation handler
1095 * @save_orig: save classic BPF program
1097 * This function effectively does the same as bpf_prog_create(), only
1098 * that it builds up its insns buffer from user space provided buffer.
1099 * It also allows for passing a bpf_aux_classic_check_t handler.
1101 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1102 bpf_aux_classic_check_t trans, bool save_orig)
1104 unsigned int fsize = bpf_classic_proglen(fprog);
1105 struct bpf_prog *fp;
1108 /* Make sure new filter is there and in the right amounts. */
1109 if (fprog->filter == NULL)
1112 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1116 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1117 __bpf_prog_free(fp);
1121 fp->len = fprog->len;
1122 fp->orig_prog = NULL;
1125 err = bpf_prog_store_orig_filter(fp, fprog);
1127 __bpf_prog_free(fp);
1132 /* bpf_prepare_filter() already takes care of freeing
1133 * memory in case something goes wrong.
1135 fp = bpf_prepare_filter(fp, trans);
1142 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1144 void bpf_prog_destroy(struct bpf_prog *fp)
1146 __bpf_prog_release(fp);
1148 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1150 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1152 struct sk_filter *fp, *old_fp;
1154 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1159 atomic_set(&fp->refcnt, 0);
1161 if (!sk_filter_charge(sk, fp)) {
1166 old_fp = rcu_dereference_protected(sk->sk_filter,
1167 sock_owned_by_user(sk));
1168 rcu_assign_pointer(sk->sk_filter, fp);
1171 sk_filter_uncharge(sk, old_fp);
1176 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1178 struct bpf_prog *old_prog;
1181 if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1184 if (sk_unhashed(sk)) {
1185 err = reuseport_alloc(sk);
1188 } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1189 /* The socket wasn't bound with SO_REUSEPORT */
1193 old_prog = reuseport_attach_prog(sk, prog);
1195 bpf_prog_destroy(old_prog);
1201 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1203 unsigned int fsize = bpf_classic_proglen(fprog);
1204 unsigned int bpf_fsize = bpf_prog_size(fprog->len);
1205 struct bpf_prog *prog;
1208 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1209 return ERR_PTR(-EPERM);
1211 /* Make sure new filter is there and in the right amounts. */
1212 if (fprog->filter == NULL)
1213 return ERR_PTR(-EINVAL);
1215 prog = bpf_prog_alloc(bpf_fsize, 0);
1217 return ERR_PTR(-ENOMEM);
1219 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1220 __bpf_prog_free(prog);
1221 return ERR_PTR(-EFAULT);
1224 prog->len = fprog->len;
1226 err = bpf_prog_store_orig_filter(prog, fprog);
1228 __bpf_prog_free(prog);
1229 return ERR_PTR(-ENOMEM);
1232 /* bpf_prepare_filter() already takes care of freeing
1233 * memory in case something goes wrong.
1235 return bpf_prepare_filter(prog, NULL);
1239 * sk_attach_filter - attach a socket filter
1240 * @fprog: the filter program
1241 * @sk: the socket to use
1243 * Attach the user's filter code. We first run some sanity checks on
1244 * it to make sure it does not explode on us later. If an error
1245 * occurs or there is insufficient memory for the filter a negative
1246 * errno code is returned. On success the return is zero.
1248 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1250 struct bpf_prog *prog = __get_filter(fprog, sk);
1254 return PTR_ERR(prog);
1256 err = __sk_attach_prog(prog, sk);
1258 __bpf_prog_release(prog);
1264 EXPORT_SYMBOL_GPL(sk_attach_filter);
1266 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1268 struct bpf_prog *prog = __get_filter(fprog, sk);
1272 return PTR_ERR(prog);
1274 err = __reuseport_attach_prog(prog, sk);
1276 __bpf_prog_release(prog);
1283 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1285 struct bpf_prog *prog;
1287 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1288 return ERR_PTR(-EPERM);
1290 prog = bpf_prog_get(ufd);
1294 if (prog->type != BPF_PROG_TYPE_SOCKET_FILTER) {
1296 return ERR_PTR(-EINVAL);
1302 int sk_attach_bpf(u32 ufd, struct sock *sk)
1304 struct bpf_prog *prog = __get_bpf(ufd, sk);
1308 return PTR_ERR(prog);
1310 err = __sk_attach_prog(prog, sk);
1319 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1321 struct bpf_prog *prog = __get_bpf(ufd, sk);
1325 return PTR_ERR(prog);
1327 err = __reuseport_attach_prog(prog, sk);
1336 #define BPF_LDST_LEN 16U
1338 static u64 bpf_skb_store_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 flags)
1340 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1341 int offset = (int) r2;
1342 void *from = (void *) (long) r3;
1343 unsigned int len = (unsigned int) r4;
1344 char buf[BPF_LDST_LEN];
1347 if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM)))
1350 /* bpf verifier guarantees that:
1351 * 'from' pointer points to bpf program stack
1352 * 'len' bytes of it were initialized
1354 * 'skb' is a valid pointer to 'struct sk_buff'
1356 * so check for invalid 'offset' and too large 'len'
1358 if (unlikely((u32) offset > 0xffff || len > sizeof(buf)))
1361 if (unlikely(skb_cloned(skb) &&
1362 !skb_clone_writable(skb, offset + len)))
1365 ptr = skb_header_pointer(skb, offset, len, buf);
1369 if (flags & BPF_F_RECOMPUTE_CSUM)
1370 skb_postpull_rcsum(skb, ptr, len);
1372 memcpy(ptr, from, len);
1375 /* skb_store_bits cannot return -EFAULT here */
1376 skb_store_bits(skb, offset, ptr, len);
1378 if (flags & BPF_F_RECOMPUTE_CSUM)
1379 skb_postpush_rcsum(skb, ptr, len);
1384 const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1385 .func = bpf_skb_store_bytes,
1387 .ret_type = RET_INTEGER,
1388 .arg1_type = ARG_PTR_TO_CTX,
1389 .arg2_type = ARG_ANYTHING,
1390 .arg3_type = ARG_PTR_TO_STACK,
1391 .arg4_type = ARG_CONST_STACK_SIZE,
1392 .arg5_type = ARG_ANYTHING,
1395 static u64 bpf_skb_load_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1397 const struct sk_buff *skb = (const struct sk_buff *)(unsigned long) r1;
1398 int offset = (int) r2;
1399 void *to = (void *)(unsigned long) r3;
1400 unsigned int len = (unsigned int) r4;
1403 if (unlikely((u32) offset > 0xffff || len > BPF_LDST_LEN))
1406 ptr = skb_header_pointer(skb, offset, len, to);
1410 memcpy(to, ptr, len);
1415 const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1416 .func = bpf_skb_load_bytes,
1418 .ret_type = RET_INTEGER,
1419 .arg1_type = ARG_PTR_TO_CTX,
1420 .arg2_type = ARG_ANYTHING,
1421 .arg3_type = ARG_PTR_TO_STACK,
1422 .arg4_type = ARG_CONST_STACK_SIZE,
1425 static u64 bpf_l3_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1427 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1428 int offset = (int) r2;
1431 if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1433 if (unlikely((u32) offset > 0xffff))
1436 if (unlikely(skb_cloned(skb) &&
1437 !skb_clone_writable(skb, offset + sizeof(sum))))
1440 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1444 switch (flags & BPF_F_HDR_FIELD_MASK) {
1446 csum_replace2(ptr, from, to);
1449 csum_replace4(ptr, from, to);
1456 /* skb_store_bits guaranteed to not return -EFAULT here */
1457 skb_store_bits(skb, offset, ptr, sizeof(sum));
1462 const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1463 .func = bpf_l3_csum_replace,
1465 .ret_type = RET_INTEGER,
1466 .arg1_type = ARG_PTR_TO_CTX,
1467 .arg2_type = ARG_ANYTHING,
1468 .arg3_type = ARG_ANYTHING,
1469 .arg4_type = ARG_ANYTHING,
1470 .arg5_type = ARG_ANYTHING,
1473 static u64 bpf_l4_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1475 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1476 bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1477 int offset = (int) r2;
1480 if (unlikely(flags & ~(BPF_F_PSEUDO_HDR | BPF_F_HDR_FIELD_MASK)))
1482 if (unlikely((u32) offset > 0xffff))
1485 if (unlikely(skb_cloned(skb) &&
1486 !skb_clone_writable(skb, offset + sizeof(sum))))
1489 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1493 switch (flags & BPF_F_HDR_FIELD_MASK) {
1495 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1498 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1505 /* skb_store_bits guaranteed to not return -EFAULT here */
1506 skb_store_bits(skb, offset, ptr, sizeof(sum));
1511 const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1512 .func = bpf_l4_csum_replace,
1514 .ret_type = RET_INTEGER,
1515 .arg1_type = ARG_PTR_TO_CTX,
1516 .arg2_type = ARG_ANYTHING,
1517 .arg3_type = ARG_ANYTHING,
1518 .arg4_type = ARG_ANYTHING,
1519 .arg5_type = ARG_ANYTHING,
1522 static u64 bpf_clone_redirect(u64 r1, u64 ifindex, u64 flags, u64 r4, u64 r5)
1524 struct sk_buff *skb = (struct sk_buff *) (long) r1, *skb2;
1525 struct net_device *dev;
1527 if (unlikely(flags & ~(BPF_F_INGRESS)))
1530 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1534 skb2 = skb_clone(skb, GFP_ATOMIC);
1535 if (unlikely(!skb2))
1538 if (flags & BPF_F_INGRESS) {
1539 if (skb_at_tc_ingress(skb2))
1540 skb_postpush_rcsum(skb2, skb_mac_header(skb2),
1542 return dev_forward_skb(dev, skb2);
1546 skb_sender_cpu_clear(skb2);
1547 return dev_queue_xmit(skb2);
1550 const struct bpf_func_proto bpf_clone_redirect_proto = {
1551 .func = bpf_clone_redirect,
1553 .ret_type = RET_INTEGER,
1554 .arg1_type = ARG_PTR_TO_CTX,
1555 .arg2_type = ARG_ANYTHING,
1556 .arg3_type = ARG_ANYTHING,
1559 struct redirect_info {
1564 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1566 static u64 bpf_redirect(u64 ifindex, u64 flags, u64 r3, u64 r4, u64 r5)
1568 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1570 if (unlikely(flags & ~(BPF_F_INGRESS)))
1573 ri->ifindex = ifindex;
1576 return TC_ACT_REDIRECT;
1579 int skb_do_redirect(struct sk_buff *skb)
1581 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1582 struct net_device *dev;
1584 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1586 if (unlikely(!dev)) {
1591 if (ri->flags & BPF_F_INGRESS) {
1592 if (skb_at_tc_ingress(skb))
1593 skb_postpush_rcsum(skb, skb_mac_header(skb),
1595 return dev_forward_skb(dev, skb);
1599 skb_sender_cpu_clear(skb);
1600 return dev_queue_xmit(skb);
1603 const struct bpf_func_proto bpf_redirect_proto = {
1604 .func = bpf_redirect,
1606 .ret_type = RET_INTEGER,
1607 .arg1_type = ARG_ANYTHING,
1608 .arg2_type = ARG_ANYTHING,
1611 static u64 bpf_get_cgroup_classid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1613 return task_get_classid((struct sk_buff *) (unsigned long) r1);
1616 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1617 .func = bpf_get_cgroup_classid,
1619 .ret_type = RET_INTEGER,
1620 .arg1_type = ARG_PTR_TO_CTX,
1623 static u64 bpf_get_route_realm(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1625 #ifdef CONFIG_IP_ROUTE_CLASSID
1626 const struct dst_entry *dst;
1628 dst = skb_dst((struct sk_buff *) (unsigned long) r1);
1630 return dst->tclassid;
1635 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1636 .func = bpf_get_route_realm,
1638 .ret_type = RET_INTEGER,
1639 .arg1_type = ARG_PTR_TO_CTX,
1642 static u64 bpf_skb_vlan_push(u64 r1, u64 r2, u64 vlan_tci, u64 r4, u64 r5)
1644 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1645 __be16 vlan_proto = (__force __be16) r2;
1647 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1648 vlan_proto != htons(ETH_P_8021AD)))
1649 vlan_proto = htons(ETH_P_8021Q);
1651 return skb_vlan_push(skb, vlan_proto, vlan_tci);
1654 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1655 .func = bpf_skb_vlan_push,
1657 .ret_type = RET_INTEGER,
1658 .arg1_type = ARG_PTR_TO_CTX,
1659 .arg2_type = ARG_ANYTHING,
1660 .arg3_type = ARG_ANYTHING,
1662 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1664 static u64 bpf_skb_vlan_pop(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1666 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1668 return skb_vlan_pop(skb);
1671 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1672 .func = bpf_skb_vlan_pop,
1674 .ret_type = RET_INTEGER,
1675 .arg1_type = ARG_PTR_TO_CTX,
1677 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1679 bool bpf_helper_changes_skb_data(void *func)
1681 if (func == bpf_skb_vlan_push)
1683 if (func == bpf_skb_vlan_pop)
1688 static unsigned short bpf_tunnel_key_af(u64 flags)
1690 return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
1693 static u64 bpf_skb_get_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
1695 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1696 struct bpf_tunnel_key *to = (struct bpf_tunnel_key *) (long) r2;
1697 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
1698 u8 compat[sizeof(struct bpf_tunnel_key)];
1700 if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6))))
1702 if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags))
1704 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
1706 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
1707 /* Fixup deprecated structure layouts here, so we have
1708 * a common path later on.
1710 if (ip_tunnel_info_af(info) != AF_INET)
1712 to = (struct bpf_tunnel_key *)compat;
1719 to->tunnel_id = be64_to_cpu(info->key.tun_id);
1720 to->tunnel_tos = info->key.tos;
1721 to->tunnel_ttl = info->key.ttl;
1723 if (flags & BPF_F_TUNINFO_IPV6)
1724 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
1725 sizeof(to->remote_ipv6));
1727 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
1729 if (unlikely(size != sizeof(struct bpf_tunnel_key)))
1730 memcpy((void *)(long) r2, to, size);
1735 const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
1736 .func = bpf_skb_get_tunnel_key,
1738 .ret_type = RET_INTEGER,
1739 .arg1_type = ARG_PTR_TO_CTX,
1740 .arg2_type = ARG_PTR_TO_STACK,
1741 .arg3_type = ARG_CONST_STACK_SIZE,
1742 .arg4_type = ARG_ANYTHING,
1745 static struct metadata_dst __percpu *md_dst;
1747 static u64 bpf_skb_set_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
1749 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1750 struct bpf_tunnel_key *from = (struct bpf_tunnel_key *) (long) r2;
1751 struct metadata_dst *md = this_cpu_ptr(md_dst);
1752 u8 compat[sizeof(struct bpf_tunnel_key)];
1753 struct ip_tunnel_info *info;
1755 if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6)))
1757 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
1759 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
1760 /* Fixup deprecated structure layouts here, so we have
1761 * a common path later on.
1763 memcpy(compat, from, size);
1764 memset(compat + size, 0, sizeof(compat) - size);
1765 from = (struct bpf_tunnel_key *)compat;
1773 dst_hold((struct dst_entry *) md);
1774 skb_dst_set(skb, (struct dst_entry *) md);
1776 info = &md->u.tun_info;
1777 info->mode = IP_TUNNEL_INFO_TX;
1779 info->key.tun_flags = TUNNEL_KEY;
1780 info->key.tun_id = cpu_to_be64(from->tunnel_id);
1781 info->key.tos = from->tunnel_tos;
1782 info->key.ttl = from->tunnel_ttl;
1784 if (flags & BPF_F_TUNINFO_IPV6) {
1785 info->mode |= IP_TUNNEL_INFO_IPV6;
1786 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
1787 sizeof(from->remote_ipv6));
1789 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
1795 const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
1796 .func = bpf_skb_set_tunnel_key,
1798 .ret_type = RET_INTEGER,
1799 .arg1_type = ARG_PTR_TO_CTX,
1800 .arg2_type = ARG_PTR_TO_STACK,
1801 .arg3_type = ARG_CONST_STACK_SIZE,
1802 .arg4_type = ARG_ANYTHING,
1805 static const struct bpf_func_proto *bpf_get_skb_set_tunnel_key_proto(void)
1808 /* race is not possible, since it's called from
1809 * verifier that is holding verifier mutex
1811 md_dst = metadata_dst_alloc_percpu(0, GFP_KERNEL);
1815 return &bpf_skb_set_tunnel_key_proto;
1818 static const struct bpf_func_proto *
1819 sk_filter_func_proto(enum bpf_func_id func_id)
1822 case BPF_FUNC_map_lookup_elem:
1823 return &bpf_map_lookup_elem_proto;
1824 case BPF_FUNC_map_update_elem:
1825 return &bpf_map_update_elem_proto;
1826 case BPF_FUNC_map_delete_elem:
1827 return &bpf_map_delete_elem_proto;
1828 case BPF_FUNC_get_prandom_u32:
1829 return &bpf_get_prandom_u32_proto;
1830 case BPF_FUNC_get_smp_processor_id:
1831 return &bpf_get_smp_processor_id_proto;
1832 case BPF_FUNC_tail_call:
1833 return &bpf_tail_call_proto;
1834 case BPF_FUNC_ktime_get_ns:
1835 return &bpf_ktime_get_ns_proto;
1836 case BPF_FUNC_trace_printk:
1837 if (capable(CAP_SYS_ADMIN))
1838 return bpf_get_trace_printk_proto();
1844 static const struct bpf_func_proto *
1845 tc_cls_act_func_proto(enum bpf_func_id func_id)
1848 case BPF_FUNC_skb_store_bytes:
1849 return &bpf_skb_store_bytes_proto;
1850 case BPF_FUNC_skb_load_bytes:
1851 return &bpf_skb_load_bytes_proto;
1852 case BPF_FUNC_l3_csum_replace:
1853 return &bpf_l3_csum_replace_proto;
1854 case BPF_FUNC_l4_csum_replace:
1855 return &bpf_l4_csum_replace_proto;
1856 case BPF_FUNC_clone_redirect:
1857 return &bpf_clone_redirect_proto;
1858 case BPF_FUNC_get_cgroup_classid:
1859 return &bpf_get_cgroup_classid_proto;
1860 case BPF_FUNC_skb_vlan_push:
1861 return &bpf_skb_vlan_push_proto;
1862 case BPF_FUNC_skb_vlan_pop:
1863 return &bpf_skb_vlan_pop_proto;
1864 case BPF_FUNC_skb_get_tunnel_key:
1865 return &bpf_skb_get_tunnel_key_proto;
1866 case BPF_FUNC_skb_set_tunnel_key:
1867 return bpf_get_skb_set_tunnel_key_proto();
1868 case BPF_FUNC_redirect:
1869 return &bpf_redirect_proto;
1870 case BPF_FUNC_get_route_realm:
1871 return &bpf_get_route_realm_proto;
1873 return sk_filter_func_proto(func_id);
1877 static bool __is_valid_access(int off, int size, enum bpf_access_type type)
1880 if (off < 0 || off >= sizeof(struct __sk_buff))
1883 /* disallow misaligned access */
1884 if (off % size != 0)
1887 /* all __sk_buff fields are __u32 */
1894 static bool sk_filter_is_valid_access(int off, int size,
1895 enum bpf_access_type type)
1897 if (off == offsetof(struct __sk_buff, tc_classid))
1900 if (type == BPF_WRITE) {
1902 case offsetof(struct __sk_buff, cb[0]) ...
1903 offsetof(struct __sk_buff, cb[4]):
1910 return __is_valid_access(off, size, type);
1913 static bool tc_cls_act_is_valid_access(int off, int size,
1914 enum bpf_access_type type)
1916 if (off == offsetof(struct __sk_buff, tc_classid))
1917 return type == BPF_WRITE ? true : false;
1919 if (type == BPF_WRITE) {
1921 case offsetof(struct __sk_buff, mark):
1922 case offsetof(struct __sk_buff, tc_index):
1923 case offsetof(struct __sk_buff, priority):
1924 case offsetof(struct __sk_buff, cb[0]) ...
1925 offsetof(struct __sk_buff, cb[4]):
1931 return __is_valid_access(off, size, type);
1934 static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
1935 int src_reg, int ctx_off,
1936 struct bpf_insn *insn_buf,
1937 struct bpf_prog *prog)
1939 struct bpf_insn *insn = insn_buf;
1942 case offsetof(struct __sk_buff, len):
1943 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
1945 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1946 offsetof(struct sk_buff, len));
1949 case offsetof(struct __sk_buff, protocol):
1950 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
1952 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1953 offsetof(struct sk_buff, protocol));
1956 case offsetof(struct __sk_buff, vlan_proto):
1957 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
1959 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1960 offsetof(struct sk_buff, vlan_proto));
1963 case offsetof(struct __sk_buff, priority):
1964 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
1966 if (type == BPF_WRITE)
1967 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
1968 offsetof(struct sk_buff, priority));
1970 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1971 offsetof(struct sk_buff, priority));
1974 case offsetof(struct __sk_buff, ingress_ifindex):
1975 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
1977 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1978 offsetof(struct sk_buff, skb_iif));
1981 case offsetof(struct __sk_buff, ifindex):
1982 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
1984 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
1986 offsetof(struct sk_buff, dev));
1987 *insn++ = BPF_JMP_IMM(BPF_JEQ, dst_reg, 0, 1);
1988 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
1989 offsetof(struct net_device, ifindex));
1992 case offsetof(struct __sk_buff, hash):
1993 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
1995 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1996 offsetof(struct sk_buff, hash));
1999 case offsetof(struct __sk_buff, mark):
2000 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
2002 if (type == BPF_WRITE)
2003 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2004 offsetof(struct sk_buff, mark));
2006 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2007 offsetof(struct sk_buff, mark));
2010 case offsetof(struct __sk_buff, pkt_type):
2011 return convert_skb_access(SKF_AD_PKTTYPE, dst_reg, src_reg, insn);
2013 case offsetof(struct __sk_buff, queue_mapping):
2014 return convert_skb_access(SKF_AD_QUEUE, dst_reg, src_reg, insn);
2016 case offsetof(struct __sk_buff, vlan_present):
2017 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
2018 dst_reg, src_reg, insn);
2020 case offsetof(struct __sk_buff, vlan_tci):
2021 return convert_skb_access(SKF_AD_VLAN_TAG,
2022 dst_reg, src_reg, insn);
2024 case offsetof(struct __sk_buff, cb[0]) ...
2025 offsetof(struct __sk_buff, cb[4]):
2026 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
2028 prog->cb_access = 1;
2029 ctx_off -= offsetof(struct __sk_buff, cb[0]);
2030 ctx_off += offsetof(struct sk_buff, cb);
2031 ctx_off += offsetof(struct qdisc_skb_cb, data);
2032 if (type == BPF_WRITE)
2033 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2035 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2038 case offsetof(struct __sk_buff, tc_classid):
2039 ctx_off -= offsetof(struct __sk_buff, tc_classid);
2040 ctx_off += offsetof(struct sk_buff, cb);
2041 ctx_off += offsetof(struct qdisc_skb_cb, tc_classid);
2042 WARN_ON(type != BPF_WRITE);
2043 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2046 case offsetof(struct __sk_buff, tc_index):
2047 #ifdef CONFIG_NET_SCHED
2048 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
2050 if (type == BPF_WRITE)
2051 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg,
2052 offsetof(struct sk_buff, tc_index));
2054 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2055 offsetof(struct sk_buff, tc_index));
2058 if (type == BPF_WRITE)
2059 *insn++ = BPF_MOV64_REG(dst_reg, dst_reg);
2061 *insn++ = BPF_MOV64_IMM(dst_reg, 0);
2066 return insn - insn_buf;
2069 static const struct bpf_verifier_ops sk_filter_ops = {
2070 .get_func_proto = sk_filter_func_proto,
2071 .is_valid_access = sk_filter_is_valid_access,
2072 .convert_ctx_access = bpf_net_convert_ctx_access,
2075 static const struct bpf_verifier_ops tc_cls_act_ops = {
2076 .get_func_proto = tc_cls_act_func_proto,
2077 .is_valid_access = tc_cls_act_is_valid_access,
2078 .convert_ctx_access = bpf_net_convert_ctx_access,
2081 static struct bpf_prog_type_list sk_filter_type __read_mostly = {
2082 .ops = &sk_filter_ops,
2083 .type = BPF_PROG_TYPE_SOCKET_FILTER,
2086 static struct bpf_prog_type_list sched_cls_type __read_mostly = {
2087 .ops = &tc_cls_act_ops,
2088 .type = BPF_PROG_TYPE_SCHED_CLS,
2091 static struct bpf_prog_type_list sched_act_type __read_mostly = {
2092 .ops = &tc_cls_act_ops,
2093 .type = BPF_PROG_TYPE_SCHED_ACT,
2096 static int __init register_sk_filter_ops(void)
2098 bpf_register_prog_type(&sk_filter_type);
2099 bpf_register_prog_type(&sched_cls_type);
2100 bpf_register_prog_type(&sched_act_type);
2104 late_initcall(register_sk_filter_ops);
2106 int sk_detach_filter(struct sock *sk)
2109 struct sk_filter *filter;
2111 if (sock_flag(sk, SOCK_FILTER_LOCKED))
2114 filter = rcu_dereference_protected(sk->sk_filter,
2115 sock_owned_by_user(sk));
2117 RCU_INIT_POINTER(sk->sk_filter, NULL);
2118 sk_filter_uncharge(sk, filter);
2124 EXPORT_SYMBOL_GPL(sk_detach_filter);
2126 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
2129 struct sock_fprog_kern *fprog;
2130 struct sk_filter *filter;
2134 filter = rcu_dereference_protected(sk->sk_filter,
2135 sock_owned_by_user(sk));
2139 /* We're copying the filter that has been originally attached,
2140 * so no conversion/decode needed anymore. eBPF programs that
2141 * have no original program cannot be dumped through this.
2144 fprog = filter->prog->orig_prog;
2150 /* User space only enquires number of filter blocks. */
2154 if (len < fprog->len)
2158 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
2161 /* Instead of bytes, the API requests to return the number