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 is remaped into 2 insns. RET_A case doesn't need an
534 * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
536 case BPF_RET | BPF_A:
537 case BPF_RET | BPF_K:
538 if (BPF_RVAL(fp->code) == BPF_K)
539 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
541 *insn = BPF_EXIT_INSN();
544 /* Store to stack. */
547 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
548 BPF_ST ? BPF_REG_A : BPF_REG_X,
549 -(BPF_MEMWORDS - fp->k) * 4);
552 /* Load from stack. */
553 case BPF_LD | BPF_MEM:
554 case BPF_LDX | BPF_MEM:
555 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
556 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
557 -(BPF_MEMWORDS - fp->k) * 4);
561 case BPF_LD | BPF_IMM:
562 case BPF_LDX | BPF_IMM:
563 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
564 BPF_REG_A : BPF_REG_X, fp->k);
568 case BPF_MISC | BPF_TAX:
569 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
573 case BPF_MISC | BPF_TXA:
574 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
577 /* A = skb->len or X = skb->len */
578 case BPF_LD | BPF_W | BPF_LEN:
579 case BPF_LDX | BPF_W | BPF_LEN:
580 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
581 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
582 offsetof(struct sk_buff, len));
585 /* Access seccomp_data fields. */
586 case BPF_LDX | BPF_ABS | BPF_W:
587 /* A = *(u32 *) (ctx + K) */
588 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
591 /* Unknown instruction. */
598 memcpy(new_insn, tmp_insns,
599 sizeof(*insn) * (insn - tmp_insns));
600 new_insn += insn - tmp_insns;
604 /* Only calculating new length. */
605 *new_len = new_insn - new_prog;
610 if (new_flen != new_insn - new_prog) {
611 new_flen = new_insn - new_prog;
618 BUG_ON(*new_len != new_flen);
627 * As we dont want to clear mem[] array for each packet going through
628 * __bpf_prog_run(), we check that filter loaded by user never try to read
629 * a cell if not previously written, and we check all branches to be sure
630 * a malicious user doesn't try to abuse us.
632 static int check_load_and_stores(const struct sock_filter *filter, int flen)
634 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
637 BUILD_BUG_ON(BPF_MEMWORDS > 16);
639 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
643 memset(masks, 0xff, flen * sizeof(*masks));
645 for (pc = 0; pc < flen; pc++) {
646 memvalid &= masks[pc];
648 switch (filter[pc].code) {
651 memvalid |= (1 << filter[pc].k);
653 case BPF_LD | BPF_MEM:
654 case BPF_LDX | BPF_MEM:
655 if (!(memvalid & (1 << filter[pc].k))) {
660 case BPF_JMP | BPF_JA:
661 /* A jump must set masks on target */
662 masks[pc + 1 + filter[pc].k] &= memvalid;
665 case BPF_JMP | BPF_JEQ | BPF_K:
666 case BPF_JMP | BPF_JEQ | BPF_X:
667 case BPF_JMP | BPF_JGE | BPF_K:
668 case BPF_JMP | BPF_JGE | BPF_X:
669 case BPF_JMP | BPF_JGT | BPF_K:
670 case BPF_JMP | BPF_JGT | BPF_X:
671 case BPF_JMP | BPF_JSET | BPF_K:
672 case BPF_JMP | BPF_JSET | BPF_X:
673 /* A jump must set masks on targets */
674 masks[pc + 1 + filter[pc].jt] &= memvalid;
675 masks[pc + 1 + filter[pc].jf] &= memvalid;
685 static bool chk_code_allowed(u16 code_to_probe)
687 static const bool codes[] = {
688 /* 32 bit ALU operations */
689 [BPF_ALU | BPF_ADD | BPF_K] = true,
690 [BPF_ALU | BPF_ADD | BPF_X] = true,
691 [BPF_ALU | BPF_SUB | BPF_K] = true,
692 [BPF_ALU | BPF_SUB | BPF_X] = true,
693 [BPF_ALU | BPF_MUL | BPF_K] = true,
694 [BPF_ALU | BPF_MUL | BPF_X] = true,
695 [BPF_ALU | BPF_DIV | BPF_K] = true,
696 [BPF_ALU | BPF_DIV | BPF_X] = true,
697 [BPF_ALU | BPF_MOD | BPF_K] = true,
698 [BPF_ALU | BPF_MOD | BPF_X] = true,
699 [BPF_ALU | BPF_AND | BPF_K] = true,
700 [BPF_ALU | BPF_AND | BPF_X] = true,
701 [BPF_ALU | BPF_OR | BPF_K] = true,
702 [BPF_ALU | BPF_OR | BPF_X] = true,
703 [BPF_ALU | BPF_XOR | BPF_K] = true,
704 [BPF_ALU | BPF_XOR | BPF_X] = true,
705 [BPF_ALU | BPF_LSH | BPF_K] = true,
706 [BPF_ALU | BPF_LSH | BPF_X] = true,
707 [BPF_ALU | BPF_RSH | BPF_K] = true,
708 [BPF_ALU | BPF_RSH | BPF_X] = true,
709 [BPF_ALU | BPF_NEG] = true,
710 /* Load instructions */
711 [BPF_LD | BPF_W | BPF_ABS] = true,
712 [BPF_LD | BPF_H | BPF_ABS] = true,
713 [BPF_LD | BPF_B | BPF_ABS] = true,
714 [BPF_LD | BPF_W | BPF_LEN] = true,
715 [BPF_LD | BPF_W | BPF_IND] = true,
716 [BPF_LD | BPF_H | BPF_IND] = true,
717 [BPF_LD | BPF_B | BPF_IND] = true,
718 [BPF_LD | BPF_IMM] = true,
719 [BPF_LD | BPF_MEM] = true,
720 [BPF_LDX | BPF_W | BPF_LEN] = true,
721 [BPF_LDX | BPF_B | BPF_MSH] = true,
722 [BPF_LDX | BPF_IMM] = true,
723 [BPF_LDX | BPF_MEM] = true,
724 /* Store instructions */
727 /* Misc instructions */
728 [BPF_MISC | BPF_TAX] = true,
729 [BPF_MISC | BPF_TXA] = true,
730 /* Return instructions */
731 [BPF_RET | BPF_K] = true,
732 [BPF_RET | BPF_A] = true,
733 /* Jump instructions */
734 [BPF_JMP | BPF_JA] = true,
735 [BPF_JMP | BPF_JEQ | BPF_K] = true,
736 [BPF_JMP | BPF_JEQ | BPF_X] = true,
737 [BPF_JMP | BPF_JGE | BPF_K] = true,
738 [BPF_JMP | BPF_JGE | BPF_X] = true,
739 [BPF_JMP | BPF_JGT | BPF_K] = true,
740 [BPF_JMP | BPF_JGT | BPF_X] = true,
741 [BPF_JMP | BPF_JSET | BPF_K] = true,
742 [BPF_JMP | BPF_JSET | BPF_X] = true,
745 if (code_to_probe >= ARRAY_SIZE(codes))
748 return codes[code_to_probe];
752 * bpf_check_classic - verify socket filter code
753 * @filter: filter to verify
754 * @flen: length of filter
756 * Check the user's filter code. If we let some ugly
757 * filter code slip through kaboom! The filter must contain
758 * no references or jumps that are out of range, no illegal
759 * instructions, and must end with a RET instruction.
761 * All jumps are forward as they are not signed.
763 * Returns 0 if the rule set is legal or -EINVAL if not.
765 static int bpf_check_classic(const struct sock_filter *filter,
771 if (flen == 0 || flen > BPF_MAXINSNS)
774 /* Check the filter code now */
775 for (pc = 0; pc < flen; pc++) {
776 const struct sock_filter *ftest = &filter[pc];
778 /* May we actually operate on this code? */
779 if (!chk_code_allowed(ftest->code))
782 /* Some instructions need special checks */
783 switch (ftest->code) {
784 case BPF_ALU | BPF_DIV | BPF_K:
785 case BPF_ALU | BPF_MOD | BPF_K:
786 /* Check for division by zero */
790 case BPF_ALU | BPF_LSH | BPF_K:
791 case BPF_ALU | BPF_RSH | BPF_K:
795 case BPF_LD | BPF_MEM:
796 case BPF_LDX | BPF_MEM:
799 /* Check for invalid memory addresses */
800 if (ftest->k >= BPF_MEMWORDS)
803 case BPF_JMP | BPF_JA:
804 /* Note, the large ftest->k might cause loops.
805 * Compare this with conditional jumps below,
806 * where offsets are limited. --ANK (981016)
808 if (ftest->k >= (unsigned int)(flen - pc - 1))
811 case BPF_JMP | BPF_JEQ | BPF_K:
812 case BPF_JMP | BPF_JEQ | BPF_X:
813 case BPF_JMP | BPF_JGE | BPF_K:
814 case BPF_JMP | BPF_JGE | BPF_X:
815 case BPF_JMP | BPF_JGT | BPF_K:
816 case BPF_JMP | BPF_JGT | BPF_X:
817 case BPF_JMP | BPF_JSET | BPF_K:
818 case BPF_JMP | BPF_JSET | BPF_X:
819 /* Both conditionals must be safe */
820 if (pc + ftest->jt + 1 >= flen ||
821 pc + ftest->jf + 1 >= flen)
824 case BPF_LD | BPF_W | BPF_ABS:
825 case BPF_LD | BPF_H | BPF_ABS:
826 case BPF_LD | BPF_B | BPF_ABS:
828 if (bpf_anc_helper(ftest) & BPF_ANC)
830 /* Ancillary operation unknown or unsupported */
831 if (anc_found == false && ftest->k >= SKF_AD_OFF)
836 /* Last instruction must be a RET code */
837 switch (filter[flen - 1].code) {
838 case BPF_RET | BPF_K:
839 case BPF_RET | BPF_A:
840 return check_load_and_stores(filter, flen);
846 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
847 const struct sock_fprog *fprog)
849 unsigned int fsize = bpf_classic_proglen(fprog);
850 struct sock_fprog_kern *fkprog;
852 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
856 fkprog = fp->orig_prog;
857 fkprog->len = fprog->len;
859 fkprog->filter = kmemdup(fp->insns, fsize,
860 GFP_KERNEL | __GFP_NOWARN);
861 if (!fkprog->filter) {
862 kfree(fp->orig_prog);
869 static void bpf_release_orig_filter(struct bpf_prog *fp)
871 struct sock_fprog_kern *fprog = fp->orig_prog;
874 kfree(fprog->filter);
879 static void __bpf_prog_release(struct bpf_prog *prog)
881 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
884 bpf_release_orig_filter(prog);
889 static void __sk_filter_release(struct sk_filter *fp)
891 __bpf_prog_release(fp->prog);
896 * sk_filter_release_rcu - Release a socket filter by rcu_head
897 * @rcu: rcu_head that contains the sk_filter to free
899 static void sk_filter_release_rcu(struct rcu_head *rcu)
901 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
903 __sk_filter_release(fp);
907 * sk_filter_release - release a socket filter
908 * @fp: filter to remove
910 * Remove a filter from a socket and release its resources.
912 static void sk_filter_release(struct sk_filter *fp)
914 if (atomic_dec_and_test(&fp->refcnt))
915 call_rcu(&fp->rcu, sk_filter_release_rcu);
918 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
920 u32 filter_size = bpf_prog_size(fp->prog->len);
922 atomic_sub(filter_size, &sk->sk_omem_alloc);
923 sk_filter_release(fp);
926 /* try to charge the socket memory if there is space available
927 * return true on success
929 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
931 u32 filter_size = bpf_prog_size(fp->prog->len);
933 /* same check as in sock_kmalloc() */
934 if (filter_size <= sysctl_optmem_max &&
935 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
936 atomic_inc(&fp->refcnt);
937 atomic_add(filter_size, &sk->sk_omem_alloc);
943 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
945 struct sock_filter *old_prog;
946 struct bpf_prog *old_fp;
947 int err, new_len, old_len = fp->len;
949 /* We are free to overwrite insns et al right here as it
950 * won't be used at this point in time anymore internally
951 * after the migration to the internal BPF instruction
954 BUILD_BUG_ON(sizeof(struct sock_filter) !=
955 sizeof(struct bpf_insn));
957 /* Conversion cannot happen on overlapping memory areas,
958 * so we need to keep the user BPF around until the 2nd
959 * pass. At this time, the user BPF is stored in fp->insns.
961 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
962 GFP_KERNEL | __GFP_NOWARN);
968 /* 1st pass: calculate the new program length. */
969 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
973 /* Expand fp for appending the new filter representation. */
975 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
977 /* The old_fp is still around in case we couldn't
978 * allocate new memory, so uncharge on that one.
987 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
988 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
990 /* 2nd bpf_convert_filter() can fail only if it fails
991 * to allocate memory, remapping must succeed. Note,
992 * that at this time old_fp has already been released
997 bpf_prog_select_runtime(fp);
1005 __bpf_prog_release(fp);
1006 return ERR_PTR(err);
1009 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1010 bpf_aux_classic_check_t trans)
1014 fp->bpf_func = NULL;
1017 err = bpf_check_classic(fp->insns, fp->len);
1019 __bpf_prog_release(fp);
1020 return ERR_PTR(err);
1023 /* There might be additional checks and transformations
1024 * needed on classic filters, f.e. in case of seccomp.
1027 err = trans(fp->insns, fp->len);
1029 __bpf_prog_release(fp);
1030 return ERR_PTR(err);
1034 /* Probe if we can JIT compile the filter and if so, do
1035 * the compilation of the filter.
1037 bpf_jit_compile(fp);
1039 /* JIT compiler couldn't process this filter, so do the
1040 * internal BPF translation for the optimized interpreter.
1043 fp = bpf_migrate_filter(fp);
1049 * bpf_prog_create - create an unattached filter
1050 * @pfp: the unattached filter that is created
1051 * @fprog: the filter program
1053 * Create a filter independent of any socket. We first run some
1054 * sanity checks on it to make sure it does not explode on us later.
1055 * If an error occurs or there is insufficient memory for the filter
1056 * a negative errno code is returned. On success the return is zero.
1058 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1060 unsigned int fsize = bpf_classic_proglen(fprog);
1061 struct bpf_prog *fp;
1063 /* Make sure new filter is there and in the right amounts. */
1064 if (fprog->filter == NULL)
1067 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1071 memcpy(fp->insns, fprog->filter, fsize);
1073 fp->len = fprog->len;
1074 /* Since unattached filters are not copied back to user
1075 * space through sk_get_filter(), we do not need to hold
1076 * a copy here, and can spare us the work.
1078 fp->orig_prog = NULL;
1080 /* bpf_prepare_filter() already takes care of freeing
1081 * memory in case something goes wrong.
1083 fp = bpf_prepare_filter(fp, NULL);
1090 EXPORT_SYMBOL_GPL(bpf_prog_create);
1093 * bpf_prog_create_from_user - create an unattached filter from user buffer
1094 * @pfp: the unattached filter that is created
1095 * @fprog: the filter program
1096 * @trans: post-classic verifier transformation handler
1097 * @save_orig: save classic BPF program
1099 * This function effectively does the same as bpf_prog_create(), only
1100 * that it builds up its insns buffer from user space provided buffer.
1101 * It also allows for passing a bpf_aux_classic_check_t handler.
1103 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1104 bpf_aux_classic_check_t trans, bool save_orig)
1106 unsigned int fsize = bpf_classic_proglen(fprog);
1107 struct bpf_prog *fp;
1110 /* Make sure new filter is there and in the right amounts. */
1111 if (fprog->filter == NULL)
1114 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1118 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1119 __bpf_prog_free(fp);
1123 fp->len = fprog->len;
1124 fp->orig_prog = NULL;
1127 err = bpf_prog_store_orig_filter(fp, fprog);
1129 __bpf_prog_free(fp);
1134 /* bpf_prepare_filter() already takes care of freeing
1135 * memory in case something goes wrong.
1137 fp = bpf_prepare_filter(fp, trans);
1144 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1146 void bpf_prog_destroy(struct bpf_prog *fp)
1148 __bpf_prog_release(fp);
1150 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1152 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1154 struct sk_filter *fp, *old_fp;
1156 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1161 atomic_set(&fp->refcnt, 0);
1163 if (!sk_filter_charge(sk, fp)) {
1168 old_fp = rcu_dereference_protected(sk->sk_filter,
1169 sock_owned_by_user(sk));
1170 rcu_assign_pointer(sk->sk_filter, fp);
1173 sk_filter_uncharge(sk, old_fp);
1178 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1180 struct bpf_prog *old_prog;
1183 if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1186 if (sk_unhashed(sk) && sk->sk_reuseport) {
1187 err = reuseport_alloc(sk);
1190 } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1191 /* The socket wasn't bound with SO_REUSEPORT */
1195 old_prog = reuseport_attach_prog(sk, prog);
1197 bpf_prog_destroy(old_prog);
1203 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1205 unsigned int fsize = bpf_classic_proglen(fprog);
1206 unsigned int bpf_fsize = bpf_prog_size(fprog->len);
1207 struct bpf_prog *prog;
1210 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1211 return ERR_PTR(-EPERM);
1213 /* Make sure new filter is there and in the right amounts. */
1214 if (fprog->filter == NULL)
1215 return ERR_PTR(-EINVAL);
1217 prog = bpf_prog_alloc(bpf_fsize, 0);
1219 return ERR_PTR(-ENOMEM);
1221 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1222 __bpf_prog_free(prog);
1223 return ERR_PTR(-EFAULT);
1226 prog->len = fprog->len;
1228 err = bpf_prog_store_orig_filter(prog, fprog);
1230 __bpf_prog_free(prog);
1231 return ERR_PTR(-ENOMEM);
1234 /* bpf_prepare_filter() already takes care of freeing
1235 * memory in case something goes wrong.
1237 return bpf_prepare_filter(prog, NULL);
1241 * sk_attach_filter - attach a socket filter
1242 * @fprog: the filter program
1243 * @sk: the socket to use
1245 * Attach the user's filter code. We first run some sanity checks on
1246 * it to make sure it does not explode on us later. If an error
1247 * occurs or there is insufficient memory for the filter a negative
1248 * errno code is returned. On success the return is zero.
1250 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1252 struct bpf_prog *prog = __get_filter(fprog, sk);
1256 return PTR_ERR(prog);
1258 err = __sk_attach_prog(prog, sk);
1260 __bpf_prog_release(prog);
1266 EXPORT_SYMBOL_GPL(sk_attach_filter);
1268 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1270 struct bpf_prog *prog = __get_filter(fprog, sk);
1274 return PTR_ERR(prog);
1276 err = __reuseport_attach_prog(prog, sk);
1278 __bpf_prog_release(prog);
1285 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1287 struct bpf_prog *prog;
1289 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1290 return ERR_PTR(-EPERM);
1292 prog = bpf_prog_get(ufd);
1296 if (prog->type != BPF_PROG_TYPE_SOCKET_FILTER) {
1298 return ERR_PTR(-EINVAL);
1304 int sk_attach_bpf(u32 ufd, struct sock *sk)
1306 struct bpf_prog *prog = __get_bpf(ufd, sk);
1310 return PTR_ERR(prog);
1312 err = __sk_attach_prog(prog, sk);
1321 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1323 struct bpf_prog *prog = __get_bpf(ufd, sk);
1327 return PTR_ERR(prog);
1329 err = __reuseport_attach_prog(prog, sk);
1338 struct bpf_scratchpad {
1340 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1341 u8 buff[MAX_BPF_STACK];
1345 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1347 static u64 bpf_skb_store_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 flags)
1349 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1350 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1351 int offset = (int) r2;
1352 void *from = (void *) (long) r3;
1353 unsigned int len = (unsigned int) r4;
1356 if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1359 /* bpf verifier guarantees that:
1360 * 'from' pointer points to bpf program stack
1361 * 'len' bytes of it were initialized
1363 * 'skb' is a valid pointer to 'struct sk_buff'
1365 * so check for invalid 'offset' and too large 'len'
1367 if (unlikely((u32) offset > 0xffff || len > sizeof(sp->buff)))
1369 if (unlikely(skb_try_make_writable(skb, offset + len)))
1372 ptr = skb_header_pointer(skb, offset, len, sp->buff);
1376 if (flags & BPF_F_RECOMPUTE_CSUM)
1377 skb_postpull_rcsum(skb, ptr, len);
1379 memcpy(ptr, from, len);
1381 if (ptr == sp->buff)
1382 /* skb_store_bits cannot return -EFAULT here */
1383 skb_store_bits(skb, offset, ptr, len);
1385 if (flags & BPF_F_RECOMPUTE_CSUM)
1386 skb_postpush_rcsum(skb, ptr, len);
1387 if (flags & BPF_F_INVALIDATE_HASH)
1388 skb_clear_hash(skb);
1393 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1394 .func = bpf_skb_store_bytes,
1396 .ret_type = RET_INTEGER,
1397 .arg1_type = ARG_PTR_TO_CTX,
1398 .arg2_type = ARG_ANYTHING,
1399 .arg3_type = ARG_PTR_TO_STACK,
1400 .arg4_type = ARG_CONST_STACK_SIZE,
1401 .arg5_type = ARG_ANYTHING,
1404 static u64 bpf_skb_load_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1406 const struct sk_buff *skb = (const struct sk_buff *)(unsigned long) r1;
1407 int offset = (int) r2;
1408 void *to = (void *)(unsigned long) r3;
1409 unsigned int len = (unsigned int) r4;
1412 if (unlikely((u32) offset > 0xffff || len > MAX_BPF_STACK))
1415 ptr = skb_header_pointer(skb, offset, len, to);
1419 memcpy(to, ptr, len);
1424 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1425 .func = bpf_skb_load_bytes,
1427 .ret_type = RET_INTEGER,
1428 .arg1_type = ARG_PTR_TO_CTX,
1429 .arg2_type = ARG_ANYTHING,
1430 .arg3_type = ARG_PTR_TO_STACK,
1431 .arg4_type = ARG_CONST_STACK_SIZE,
1434 static u64 bpf_l3_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1436 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1437 int offset = (int) r2;
1440 if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1442 if (unlikely((u32) offset > 0xffff))
1444 if (unlikely(skb_try_make_writable(skb, offset + sizeof(sum))))
1447 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1451 switch (flags & BPF_F_HDR_FIELD_MASK) {
1453 if (unlikely(from != 0))
1456 csum_replace_by_diff(ptr, to);
1459 csum_replace2(ptr, from, to);
1462 csum_replace4(ptr, from, to);
1469 /* skb_store_bits guaranteed to not return -EFAULT here */
1470 skb_store_bits(skb, offset, ptr, sizeof(sum));
1475 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1476 .func = bpf_l3_csum_replace,
1478 .ret_type = RET_INTEGER,
1479 .arg1_type = ARG_PTR_TO_CTX,
1480 .arg2_type = ARG_ANYTHING,
1481 .arg3_type = ARG_ANYTHING,
1482 .arg4_type = ARG_ANYTHING,
1483 .arg5_type = ARG_ANYTHING,
1486 static u64 bpf_l4_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1488 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1489 bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1490 bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1491 int offset = (int) r2;
1494 if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_PSEUDO_HDR |
1495 BPF_F_HDR_FIELD_MASK)))
1497 if (unlikely((u32) offset > 0xffff))
1499 if (unlikely(skb_try_make_writable(skb, offset + sizeof(sum))))
1502 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1505 if (is_mmzero && !*ptr)
1508 switch (flags & BPF_F_HDR_FIELD_MASK) {
1510 if (unlikely(from != 0))
1513 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1516 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1519 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1525 if (is_mmzero && !*ptr)
1526 *ptr = CSUM_MANGLED_0;
1528 /* skb_store_bits guaranteed to not return -EFAULT here */
1529 skb_store_bits(skb, offset, ptr, sizeof(sum));
1534 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1535 .func = bpf_l4_csum_replace,
1537 .ret_type = RET_INTEGER,
1538 .arg1_type = ARG_PTR_TO_CTX,
1539 .arg2_type = ARG_ANYTHING,
1540 .arg3_type = ARG_ANYTHING,
1541 .arg4_type = ARG_ANYTHING,
1542 .arg5_type = ARG_ANYTHING,
1545 static u64 bpf_csum_diff(u64 r1, u64 from_size, u64 r3, u64 to_size, u64 seed)
1547 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1548 u64 diff_size = from_size + to_size;
1549 __be32 *from = (__be32 *) (long) r1;
1550 __be32 *to = (__be32 *) (long) r3;
1553 /* This is quite flexible, some examples:
1555 * from_size == 0, to_size > 0, seed := csum --> pushing data
1556 * from_size > 0, to_size == 0, seed := csum --> pulling data
1557 * from_size > 0, to_size > 0, seed := 0 --> diffing data
1559 * Even for diffing, from_size and to_size don't need to be equal.
1561 if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1562 diff_size > sizeof(sp->diff)))
1565 for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1566 sp->diff[j] = ~from[i];
1567 for (i = 0; i < to_size / sizeof(__be32); i++, j++)
1568 sp->diff[j] = to[i];
1570 return csum_partial(sp->diff, diff_size, seed);
1573 static const struct bpf_func_proto bpf_csum_diff_proto = {
1574 .func = bpf_csum_diff,
1576 .ret_type = RET_INTEGER,
1577 .arg1_type = ARG_PTR_TO_STACK,
1578 .arg2_type = ARG_CONST_STACK_SIZE_OR_ZERO,
1579 .arg3_type = ARG_PTR_TO_STACK,
1580 .arg4_type = ARG_CONST_STACK_SIZE_OR_ZERO,
1581 .arg5_type = ARG_ANYTHING,
1584 static u64 bpf_clone_redirect(u64 r1, u64 ifindex, u64 flags, u64 r4, u64 r5)
1586 struct sk_buff *skb = (struct sk_buff *) (long) r1, *skb2;
1587 struct net_device *dev;
1589 if (unlikely(flags & ~(BPF_F_INGRESS)))
1592 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1596 skb2 = skb_clone(skb, GFP_ATOMIC);
1597 if (unlikely(!skb2))
1600 if (flags & BPF_F_INGRESS) {
1601 if (skb_at_tc_ingress(skb2))
1602 skb_postpush_rcsum(skb2, skb_mac_header(skb2),
1604 return dev_forward_skb(dev, skb2);
1608 return dev_queue_xmit(skb2);
1611 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1612 .func = bpf_clone_redirect,
1614 .ret_type = RET_INTEGER,
1615 .arg1_type = ARG_PTR_TO_CTX,
1616 .arg2_type = ARG_ANYTHING,
1617 .arg3_type = ARG_ANYTHING,
1620 struct redirect_info {
1625 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1627 static u64 bpf_redirect(u64 ifindex, u64 flags, u64 r3, u64 r4, u64 r5)
1629 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1631 if (unlikely(flags & ~(BPF_F_INGRESS)))
1634 ri->ifindex = ifindex;
1637 return TC_ACT_REDIRECT;
1640 int skb_do_redirect(struct sk_buff *skb)
1642 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1643 struct net_device *dev;
1645 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1647 if (unlikely(!dev)) {
1652 if (ri->flags & BPF_F_INGRESS) {
1653 if (skb_at_tc_ingress(skb))
1654 skb_postpush_rcsum(skb, skb_mac_header(skb),
1656 return dev_forward_skb(dev, skb);
1660 return dev_queue_xmit(skb);
1663 static const struct bpf_func_proto bpf_redirect_proto = {
1664 .func = bpf_redirect,
1666 .ret_type = RET_INTEGER,
1667 .arg1_type = ARG_ANYTHING,
1668 .arg2_type = ARG_ANYTHING,
1671 static u64 bpf_get_cgroup_classid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1673 return task_get_classid((struct sk_buff *) (unsigned long) r1);
1676 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1677 .func = bpf_get_cgroup_classid,
1679 .ret_type = RET_INTEGER,
1680 .arg1_type = ARG_PTR_TO_CTX,
1683 static u64 bpf_get_route_realm(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1685 #ifdef CONFIG_IP_ROUTE_CLASSID
1686 const struct dst_entry *dst;
1688 dst = skb_dst((struct sk_buff *) (unsigned long) r1);
1690 return dst->tclassid;
1695 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1696 .func = bpf_get_route_realm,
1698 .ret_type = RET_INTEGER,
1699 .arg1_type = ARG_PTR_TO_CTX,
1702 static u64 bpf_skb_vlan_push(u64 r1, u64 r2, u64 vlan_tci, u64 r4, u64 r5)
1704 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1705 __be16 vlan_proto = (__force __be16) r2;
1707 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1708 vlan_proto != htons(ETH_P_8021AD)))
1709 vlan_proto = htons(ETH_P_8021Q);
1711 return skb_vlan_push(skb, vlan_proto, vlan_tci);
1714 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1715 .func = bpf_skb_vlan_push,
1717 .ret_type = RET_INTEGER,
1718 .arg1_type = ARG_PTR_TO_CTX,
1719 .arg2_type = ARG_ANYTHING,
1720 .arg3_type = ARG_ANYTHING,
1722 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1724 static u64 bpf_skb_vlan_pop(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1726 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1728 return skb_vlan_pop(skb);
1731 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1732 .func = bpf_skb_vlan_pop,
1734 .ret_type = RET_INTEGER,
1735 .arg1_type = ARG_PTR_TO_CTX,
1737 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1739 bool bpf_helper_changes_skb_data(void *func)
1741 if (func == bpf_skb_vlan_push)
1743 if (func == bpf_skb_vlan_pop)
1745 if (func == bpf_skb_store_bytes)
1747 if (func == bpf_l3_csum_replace)
1749 if (func == bpf_l4_csum_replace)
1755 static unsigned short bpf_tunnel_key_af(u64 flags)
1757 return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
1760 static u64 bpf_skb_get_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
1762 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1763 struct bpf_tunnel_key *to = (struct bpf_tunnel_key *) (long) r2;
1764 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
1765 u8 compat[sizeof(struct bpf_tunnel_key)];
1767 if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6))))
1769 if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags))
1771 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
1773 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
1774 /* Fixup deprecated structure layouts here, so we have
1775 * a common path later on.
1777 if (ip_tunnel_info_af(info) != AF_INET)
1779 to = (struct bpf_tunnel_key *)compat;
1786 to->tunnel_id = be64_to_cpu(info->key.tun_id);
1787 to->tunnel_tos = info->key.tos;
1788 to->tunnel_ttl = info->key.ttl;
1790 if (flags & BPF_F_TUNINFO_IPV6)
1791 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
1792 sizeof(to->remote_ipv6));
1794 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
1796 if (unlikely(size != sizeof(struct bpf_tunnel_key)))
1797 memcpy((void *)(long) r2, to, size);
1802 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
1803 .func = bpf_skb_get_tunnel_key,
1805 .ret_type = RET_INTEGER,
1806 .arg1_type = ARG_PTR_TO_CTX,
1807 .arg2_type = ARG_PTR_TO_STACK,
1808 .arg3_type = ARG_CONST_STACK_SIZE,
1809 .arg4_type = ARG_ANYTHING,
1812 static struct metadata_dst __percpu *md_dst;
1814 static u64 bpf_skb_set_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
1816 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1817 struct bpf_tunnel_key *from = (struct bpf_tunnel_key *) (long) r2;
1818 struct metadata_dst *md = this_cpu_ptr(md_dst);
1819 u8 compat[sizeof(struct bpf_tunnel_key)];
1820 struct ip_tunnel_info *info;
1822 if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX)))
1824 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
1826 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
1827 /* Fixup deprecated structure layouts here, so we have
1828 * a common path later on.
1830 memcpy(compat, from, size);
1831 memset(compat + size, 0, sizeof(compat) - size);
1832 from = (struct bpf_tunnel_key *)compat;
1840 dst_hold((struct dst_entry *) md);
1841 skb_dst_set(skb, (struct dst_entry *) md);
1843 info = &md->u.tun_info;
1844 info->mode = IP_TUNNEL_INFO_TX;
1846 info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM;
1847 info->key.tun_id = cpu_to_be64(from->tunnel_id);
1848 info->key.tos = from->tunnel_tos;
1849 info->key.ttl = from->tunnel_ttl;
1851 if (flags & BPF_F_TUNINFO_IPV6) {
1852 info->mode |= IP_TUNNEL_INFO_IPV6;
1853 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
1854 sizeof(from->remote_ipv6));
1856 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
1857 if (flags & BPF_F_ZERO_CSUM_TX)
1858 info->key.tun_flags &= ~TUNNEL_CSUM;
1864 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
1865 .func = bpf_skb_set_tunnel_key,
1867 .ret_type = RET_INTEGER,
1868 .arg1_type = ARG_PTR_TO_CTX,
1869 .arg2_type = ARG_PTR_TO_STACK,
1870 .arg3_type = ARG_CONST_STACK_SIZE,
1871 .arg4_type = ARG_ANYTHING,
1874 static const struct bpf_func_proto *bpf_get_skb_set_tunnel_key_proto(void)
1877 /* race is not possible, since it's called from
1878 * verifier that is holding verifier mutex
1880 md_dst = metadata_dst_alloc_percpu(0, GFP_KERNEL);
1884 return &bpf_skb_set_tunnel_key_proto;
1887 static const struct bpf_func_proto *
1888 sk_filter_func_proto(enum bpf_func_id func_id)
1891 case BPF_FUNC_map_lookup_elem:
1892 return &bpf_map_lookup_elem_proto;
1893 case BPF_FUNC_map_update_elem:
1894 return &bpf_map_update_elem_proto;
1895 case BPF_FUNC_map_delete_elem:
1896 return &bpf_map_delete_elem_proto;
1897 case BPF_FUNC_get_prandom_u32:
1898 return &bpf_get_prandom_u32_proto;
1899 case BPF_FUNC_get_smp_processor_id:
1900 return &bpf_get_smp_processor_id_proto;
1901 case BPF_FUNC_tail_call:
1902 return &bpf_tail_call_proto;
1903 case BPF_FUNC_ktime_get_ns:
1904 return &bpf_ktime_get_ns_proto;
1905 case BPF_FUNC_trace_printk:
1906 if (capable(CAP_SYS_ADMIN))
1907 return bpf_get_trace_printk_proto();
1913 static const struct bpf_func_proto *
1914 tc_cls_act_func_proto(enum bpf_func_id func_id)
1917 case BPF_FUNC_skb_store_bytes:
1918 return &bpf_skb_store_bytes_proto;
1919 case BPF_FUNC_skb_load_bytes:
1920 return &bpf_skb_load_bytes_proto;
1921 case BPF_FUNC_csum_diff:
1922 return &bpf_csum_diff_proto;
1923 case BPF_FUNC_l3_csum_replace:
1924 return &bpf_l3_csum_replace_proto;
1925 case BPF_FUNC_l4_csum_replace:
1926 return &bpf_l4_csum_replace_proto;
1927 case BPF_FUNC_clone_redirect:
1928 return &bpf_clone_redirect_proto;
1929 case BPF_FUNC_get_cgroup_classid:
1930 return &bpf_get_cgroup_classid_proto;
1931 case BPF_FUNC_skb_vlan_push:
1932 return &bpf_skb_vlan_push_proto;
1933 case BPF_FUNC_skb_vlan_pop:
1934 return &bpf_skb_vlan_pop_proto;
1935 case BPF_FUNC_skb_get_tunnel_key:
1936 return &bpf_skb_get_tunnel_key_proto;
1937 case BPF_FUNC_skb_set_tunnel_key:
1938 return bpf_get_skb_set_tunnel_key_proto();
1939 case BPF_FUNC_redirect:
1940 return &bpf_redirect_proto;
1941 case BPF_FUNC_get_route_realm:
1942 return &bpf_get_route_realm_proto;
1944 return sk_filter_func_proto(func_id);
1948 static bool __is_valid_access(int off, int size, enum bpf_access_type type)
1951 if (off < 0 || off >= sizeof(struct __sk_buff))
1954 /* disallow misaligned access */
1955 if (off % size != 0)
1958 /* all __sk_buff fields are __u32 */
1965 static bool sk_filter_is_valid_access(int off, int size,
1966 enum bpf_access_type type)
1968 if (off == offsetof(struct __sk_buff, tc_classid))
1971 if (type == BPF_WRITE) {
1973 case offsetof(struct __sk_buff, cb[0]) ...
1974 offsetof(struct __sk_buff, cb[4]):
1981 return __is_valid_access(off, size, type);
1984 static bool tc_cls_act_is_valid_access(int off, int size,
1985 enum bpf_access_type type)
1987 if (off == offsetof(struct __sk_buff, tc_classid))
1988 return type == BPF_WRITE ? true : false;
1990 if (type == BPF_WRITE) {
1992 case offsetof(struct __sk_buff, mark):
1993 case offsetof(struct __sk_buff, tc_index):
1994 case offsetof(struct __sk_buff, priority):
1995 case offsetof(struct __sk_buff, cb[0]) ...
1996 offsetof(struct __sk_buff, cb[4]):
2002 return __is_valid_access(off, size, type);
2005 static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
2006 int src_reg, int ctx_off,
2007 struct bpf_insn *insn_buf,
2008 struct bpf_prog *prog)
2010 struct bpf_insn *insn = insn_buf;
2013 case offsetof(struct __sk_buff, len):
2014 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
2016 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2017 offsetof(struct sk_buff, len));
2020 case offsetof(struct __sk_buff, protocol):
2021 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
2023 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2024 offsetof(struct sk_buff, protocol));
2027 case offsetof(struct __sk_buff, vlan_proto):
2028 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
2030 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2031 offsetof(struct sk_buff, vlan_proto));
2034 case offsetof(struct __sk_buff, priority):
2035 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
2037 if (type == BPF_WRITE)
2038 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2039 offsetof(struct sk_buff, priority));
2041 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2042 offsetof(struct sk_buff, priority));
2045 case offsetof(struct __sk_buff, ingress_ifindex):
2046 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
2048 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2049 offsetof(struct sk_buff, skb_iif));
2052 case offsetof(struct __sk_buff, ifindex):
2053 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
2055 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
2057 offsetof(struct sk_buff, dev));
2058 *insn++ = BPF_JMP_IMM(BPF_JEQ, dst_reg, 0, 1);
2059 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
2060 offsetof(struct net_device, ifindex));
2063 case offsetof(struct __sk_buff, hash):
2064 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
2066 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2067 offsetof(struct sk_buff, hash));
2070 case offsetof(struct __sk_buff, mark):
2071 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
2073 if (type == BPF_WRITE)
2074 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2075 offsetof(struct sk_buff, mark));
2077 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2078 offsetof(struct sk_buff, mark));
2081 case offsetof(struct __sk_buff, pkt_type):
2082 return convert_skb_access(SKF_AD_PKTTYPE, dst_reg, src_reg, insn);
2084 case offsetof(struct __sk_buff, queue_mapping):
2085 return convert_skb_access(SKF_AD_QUEUE, dst_reg, src_reg, insn);
2087 case offsetof(struct __sk_buff, vlan_present):
2088 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
2089 dst_reg, src_reg, insn);
2091 case offsetof(struct __sk_buff, vlan_tci):
2092 return convert_skb_access(SKF_AD_VLAN_TAG,
2093 dst_reg, src_reg, insn);
2095 case offsetof(struct __sk_buff, cb[0]) ...
2096 offsetof(struct __sk_buff, cb[4]):
2097 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
2099 prog->cb_access = 1;
2100 ctx_off -= offsetof(struct __sk_buff, cb[0]);
2101 ctx_off += offsetof(struct sk_buff, cb);
2102 ctx_off += offsetof(struct qdisc_skb_cb, data);
2103 if (type == BPF_WRITE)
2104 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2106 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2109 case offsetof(struct __sk_buff, tc_classid):
2110 ctx_off -= offsetof(struct __sk_buff, tc_classid);
2111 ctx_off += offsetof(struct sk_buff, cb);
2112 ctx_off += offsetof(struct qdisc_skb_cb, tc_classid);
2113 WARN_ON(type != BPF_WRITE);
2114 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2117 case offsetof(struct __sk_buff, tc_index):
2118 #ifdef CONFIG_NET_SCHED
2119 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
2121 if (type == BPF_WRITE)
2122 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg,
2123 offsetof(struct sk_buff, tc_index));
2125 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2126 offsetof(struct sk_buff, tc_index));
2129 if (type == BPF_WRITE)
2130 *insn++ = BPF_MOV64_REG(dst_reg, dst_reg);
2132 *insn++ = BPF_MOV64_IMM(dst_reg, 0);
2137 return insn - insn_buf;
2140 static const struct bpf_verifier_ops sk_filter_ops = {
2141 .get_func_proto = sk_filter_func_proto,
2142 .is_valid_access = sk_filter_is_valid_access,
2143 .convert_ctx_access = bpf_net_convert_ctx_access,
2146 static const struct bpf_verifier_ops tc_cls_act_ops = {
2147 .get_func_proto = tc_cls_act_func_proto,
2148 .is_valid_access = tc_cls_act_is_valid_access,
2149 .convert_ctx_access = bpf_net_convert_ctx_access,
2152 static struct bpf_prog_type_list sk_filter_type __read_mostly = {
2153 .ops = &sk_filter_ops,
2154 .type = BPF_PROG_TYPE_SOCKET_FILTER,
2157 static struct bpf_prog_type_list sched_cls_type __read_mostly = {
2158 .ops = &tc_cls_act_ops,
2159 .type = BPF_PROG_TYPE_SCHED_CLS,
2162 static struct bpf_prog_type_list sched_act_type __read_mostly = {
2163 .ops = &tc_cls_act_ops,
2164 .type = BPF_PROG_TYPE_SCHED_ACT,
2167 static int __init register_sk_filter_ops(void)
2169 bpf_register_prog_type(&sk_filter_type);
2170 bpf_register_prog_type(&sched_cls_type);
2171 bpf_register_prog_type(&sched_act_type);
2175 late_initcall(register_sk_filter_ops);
2177 int sk_detach_filter(struct sock *sk)
2180 struct sk_filter *filter;
2182 if (sock_flag(sk, SOCK_FILTER_LOCKED))
2185 filter = rcu_dereference_protected(sk->sk_filter,
2186 sock_owned_by_user(sk));
2188 RCU_INIT_POINTER(sk->sk_filter, NULL);
2189 sk_filter_uncharge(sk, filter);
2195 EXPORT_SYMBOL_GPL(sk_detach_filter);
2197 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
2200 struct sock_fprog_kern *fprog;
2201 struct sk_filter *filter;
2205 filter = rcu_dereference_protected(sk->sk_filter,
2206 sock_owned_by_user(sk));
2210 /* We're copying the filter that has been originally attached,
2211 * so no conversion/decode needed anymore. eBPF programs that
2212 * have no original program cannot be dumped through this.
2215 fprog = filter->prog->orig_prog;
2221 /* User space only enquires number of filter blocks. */
2225 if (len < fprog->len)
2229 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
2232 /* Instead of bytes, the API requests to return the number