1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Thomas Graf <tgraf@tgraf.ch>
3 */
4
5 #include <linux/kernel.h>
6 #include <linux/module.h>
7 #include <linux/skbuff.h>
8 #include <linux/types.h>
9 #include <linux/bpf.h>
10 #include <net/lwtunnel.h>
11 #include <net/gre.h>
12 #include <net/ip6_route.h>
13 #include <net/ipv6_stubs.h>
14
15 struct bpf_lwt_prog {
16 struct bpf_prog *prog;
17 char *name;
18 };
19
20 struct bpf_lwt {
21 struct bpf_lwt_prog in;
22 struct bpf_lwt_prog out;
23 struct bpf_lwt_prog xmit;
24 int family;
25 };
26
27 #define MAX_PROG_NAME 256
28
29 static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt)
30 {
31 return (struct bpf_lwt *)lwt->data;
32 }
33
34 #define NO_REDIRECT false
35 #define CAN_REDIRECT true
36
37 static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
38 struct dst_entry *dst, bool can_redirect)
39 {
40 int ret;
41
42 /* Preempt disable is needed to protect per-cpu redirect_info between
43 * BPF prog and skb_do_redirect(). The call_rcu in bpf_prog_put() and
44 * access to maps strictly require a rcu_read_lock() for protection,
45 * mixing with BH RCU lock doesn't work.
46 */
47 preempt_disable();
48 bpf_compute_data_pointers(skb);
49 ret = bpf_prog_run_save_cb(lwt->prog, skb);
50
51 switch (ret) {
52 case BPF_OK:
53 case BPF_LWT_REROUTE:
54 break;
55
56 case BPF_REDIRECT:
57 if (unlikely(!can_redirect)) {
58 pr_warn_once("Illegal redirect return code in prog %s\n",
59 lwt->name ? : "<unknown>");
60 ret = BPF_OK;
61 } else {
62 skb_reset_mac_header(skb);
63 ret = skb_do_redirect(skb);
64 if (ret == 0)
65 ret = BPF_REDIRECT;
66 }
67 break;
68
69 case BPF_DROP:
70 kfree_skb(skb);
71 ret = -EPERM;
72 break;
73
74 default:
75 pr_warn_once("bpf-lwt: Illegal return value %u, expect packet loss\n", ret);
76 kfree_skb(skb);
77 ret = -EINVAL;
78 break;
79 }
80
81 preempt_enable();
82
83 return ret;
84 }
85
86 static int bpf_lwt_input_reroute(struct sk_buff *skb)
87 {
88 int err = -EINVAL;
89
90 if (skb->protocol == htons(ETH_P_IP)) {
91 struct net_device *dev = skb_dst(skb)->dev;
92 struct iphdr *iph = ip_hdr(skb);
93
94 dev_hold(dev);
95 skb_dst_drop(skb);
96 err = ip_route_input_noref(skb, iph->daddr, iph->saddr,
97 iph->tos, dev);
98 dev_put(dev);
99 } else if (skb->protocol == htons(ETH_P_IPV6)) {
100 skb_dst_drop(skb);
101 err = ipv6_stub->ipv6_route_input(skb);
102 } else {
103 err = -EAFNOSUPPORT;
104 }
105
106 if (err)
107 goto err;
108 return dst_input(skb);
109
110 err:
111 kfree_skb(skb);
112 return err;
113 }
114
115 static int bpf_input(struct sk_buff *skb)
116 {
117 struct dst_entry *dst = skb_dst(skb);
118 struct bpf_lwt *bpf;
119 int ret;
120
121 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
122 if (bpf->in.prog) {
123 ret = run_lwt_bpf(skb, &bpf->in, dst, NO_REDIRECT);
124 if (ret < 0)
125 return ret;
126 if (ret == BPF_LWT_REROUTE)
127 return bpf_lwt_input_reroute(skb);
128 }
129
130 if (unlikely(!dst->lwtstate->orig_input)) {
131 kfree_skb(skb);
132 return -EINVAL;
133 }
134
135 return dst->lwtstate->orig_input(skb);
136 }
137
138 static int bpf_output(struct net *net, struct sock *sk, struct sk_buff *skb)
139 {
140 struct dst_entry *dst = skb_dst(skb);
141 struct bpf_lwt *bpf;
142 int ret;
143
144 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
145 if (bpf->out.prog) {
146 ret = run_lwt_bpf(skb, &bpf->out, dst, NO_REDIRECT);
147 if (ret < 0)
148 return ret;
149 }
150
151 if (unlikely(!dst->lwtstate->orig_output)) {
152 pr_warn_once("orig_output not set on dst for prog %s\n",
153 bpf->out.name);
154 kfree_skb(skb);
155 return -EINVAL;
156 }
157
158 return dst->lwtstate->orig_output(net, sk, skb);
159 }
160
161 static int xmit_check_hhlen(struct sk_buff *skb)
162 {
163 int hh_len = skb_dst(skb)->dev->hard_header_len;
164
165 if (skb_headroom(skb) < hh_len) {
166 int nhead = HH_DATA_ALIGN(hh_len - skb_headroom(skb));
167
168 if (pskb_expand_head(skb, nhead, 0, GFP_ATOMIC))
169 return -ENOMEM;
170 }
171
172 return 0;
173 }
174
175 static int bpf_lwt_xmit_reroute(struct sk_buff *skb)
176 {
177 struct net_device *l3mdev = l3mdev_master_dev_rcu(skb_dst(skb)->dev);
178 int oif = l3mdev ? l3mdev->ifindex : 0;
179 struct dst_entry *dst = NULL;
180 int err = -EAFNOSUPPORT;
181 struct sock *sk;
182 struct net *net;
183 bool ipv4;
184
185 if (skb->protocol == htons(ETH_P_IP))
186 ipv4 = true;
187 else if (skb->protocol == htons(ETH_P_IPV6))
188 ipv4 = false;
189 else
190 goto err;
191
192 sk = sk_to_full_sk(skb->sk);
193 if (sk) {
194 if (sk->sk_bound_dev_if)
195 oif = sk->sk_bound_dev_if;
196 net = sock_net(sk);
197 } else {
198 net = dev_net(skb_dst(skb)->dev);
199 }
200
201 if (ipv4) {
202 struct iphdr *iph = ip_hdr(skb);
203 struct flowi4 fl4 = {};
204 struct rtable *rt;
205
206 fl4.flowi4_oif = oif;
207 fl4.flowi4_mark = skb->mark;
208 fl4.flowi4_uid = sock_net_uid(net, sk);
209 fl4.flowi4_tos = RT_TOS(iph->tos);
210 fl4.flowi4_flags = FLOWI_FLAG_ANYSRC;
211 fl4.flowi4_proto = iph->protocol;
212 fl4.daddr = iph->daddr;
213 fl4.saddr = iph->saddr;
214
215 rt = ip_route_output_key(net, &fl4);
216 if (IS_ERR(rt)) {
217 err = PTR_ERR(rt);
218 goto err;
219 }
220 dst = &rt->dst;
221 } else {
222 struct ipv6hdr *iph6 = ipv6_hdr(skb);
223 struct flowi6 fl6 = {};
224
225 fl6.flowi6_oif = oif;
226 fl6.flowi6_mark = skb->mark;
227 fl6.flowi6_uid = sock_net_uid(net, sk);
228 fl6.flowlabel = ip6_flowinfo(iph6);
229 fl6.flowi6_proto = iph6->nexthdr;
230 fl6.daddr = iph6->daddr;
231 fl6.saddr = iph6->saddr;
232
233 dst = ipv6_stub->ipv6_dst_lookup_flow(net, skb->sk, &fl6, NULL);
234 if (IS_ERR(dst)) {
235 err = PTR_ERR(dst);
236 goto err;
237 }
238 }
239 if (unlikely(dst->error)) {
240 err = dst->error;
241 dst_release(dst);
242 goto err;
243 }
244
245 /* Although skb header was reserved in bpf_lwt_push_ip_encap(), it
246 * was done for the previous dst, so we are doing it here again, in
247 * case the new dst needs much more space. The call below is a noop
248 * if there is enough header space in skb.
249 */
250 err = skb_cow_head(skb, LL_RESERVED_SPACE(dst->dev));
251 if (unlikely(err))
252 goto err;
253
254 skb_dst_drop(skb);
255 skb_dst_set(skb, dst);
256
257 err = dst_output(dev_net(skb_dst(skb)->dev), skb->sk, skb);
258 if (unlikely(err))
259 return err;
260
261 /* ip[6]_finish_output2 understand LWTUNNEL_XMIT_DONE */
262 return LWTUNNEL_XMIT_DONE;
263
264 err:
265 kfree_skb(skb);
266 return err;
267 }
268
269 static int bpf_xmit(struct sk_buff *skb)
270 {
271 struct dst_entry *dst = skb_dst(skb);
272 struct bpf_lwt *bpf;
273
274 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
275 if (bpf->xmit.prog) {
276 __be16 proto = skb->protocol;
277 int ret;
278
279 ret = run_lwt_bpf(skb, &bpf->xmit, dst, CAN_REDIRECT);
280 switch (ret) {
281 case BPF_OK:
282 /* If the header changed, e.g. via bpf_lwt_push_encap,
283 * BPF_LWT_REROUTE below should have been used if the
284 * protocol was also changed.
285 */
286 if (skb->protocol != proto) {
287 kfree_skb(skb);
288 return -EINVAL;
289 }
290 /* If the header was expanded, headroom might be too
291 * small for L2 header to come, expand as needed.
292 */
293 ret = xmit_check_hhlen(skb);
294 if (unlikely(ret))
295 return ret;
296
297 return LWTUNNEL_XMIT_CONTINUE;
298 case BPF_REDIRECT:
299 return LWTUNNEL_XMIT_DONE;
300 case BPF_LWT_REROUTE:
301 return bpf_lwt_xmit_reroute(skb);
302 default:
303 return ret;
304 }
305 }
306
307 return LWTUNNEL_XMIT_CONTINUE;
308 }
309
310 static void bpf_lwt_prog_destroy(struct bpf_lwt_prog *prog)
311 {
312 if (prog->prog)
313 bpf_prog_put(prog->prog);
314
315 kfree(prog->name);
316 }
317
318 static void bpf_destroy_state(struct lwtunnel_state *lwt)
319 {
320 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
321
322 bpf_lwt_prog_destroy(&bpf->in);
323 bpf_lwt_prog_destroy(&bpf->out);
324 bpf_lwt_prog_destroy(&bpf->xmit);
325 }
326
327 static const struct nla_policy bpf_prog_policy[LWT_BPF_PROG_MAX + 1] = {
328 [LWT_BPF_PROG_FD] = { .type = NLA_U32, },
329 [LWT_BPF_PROG_NAME] = { .type = NLA_NUL_STRING,
330 .len = MAX_PROG_NAME },
331 };
332
333 static int bpf_parse_prog(struct nlattr *attr, struct bpf_lwt_prog *prog,
334 enum bpf_prog_type type)
335 {
336 struct nlattr *tb[LWT_BPF_PROG_MAX + 1];
337 struct bpf_prog *p;
338 int ret;
339 u32 fd;
340
341 ret = nla_parse_nested_deprecated(tb, LWT_BPF_PROG_MAX, attr,
342 bpf_prog_policy, NULL);
343 if (ret < 0)
344 return ret;
345
346 if (!tb[LWT_BPF_PROG_FD] || !tb[LWT_BPF_PROG_NAME])
347 return -EINVAL;
348
349 prog->name = nla_memdup(tb[LWT_BPF_PROG_NAME], GFP_ATOMIC);
350 if (!prog->name)
351 return -ENOMEM;
352
353 fd = nla_get_u32(tb[LWT_BPF_PROG_FD]);
354 p = bpf_prog_get_type(fd, type);
355 if (IS_ERR(p))
356 return PTR_ERR(p);
357
358 prog->prog = p;
359
360 return 0;
361 }
362
363 static const struct nla_policy bpf_nl_policy[LWT_BPF_MAX + 1] = {
364 [LWT_BPF_IN] = { .type = NLA_NESTED, },
365 [LWT_BPF_OUT] = { .type = NLA_NESTED, },
366 [LWT_BPF_XMIT] = { .type = NLA_NESTED, },
367 [LWT_BPF_XMIT_HEADROOM] = { .type = NLA_U32 },
368 };
369
370 static int bpf_build_state(struct net *net, struct nlattr *nla,
371 unsigned int family, const void *cfg,
372 struct lwtunnel_state **ts,
373 struct netlink_ext_ack *extack)
374 {
375 struct nlattr *tb[LWT_BPF_MAX + 1];
376 struct lwtunnel_state *newts;
377 struct bpf_lwt *bpf;
378 int ret;
379
380 if (family != AF_INET && family != AF_INET6)
381 return -EAFNOSUPPORT;
382
383 ret = nla_parse_nested_deprecated(tb, LWT_BPF_MAX, nla, bpf_nl_policy,
384 extack);
385 if (ret < 0)
386 return ret;
387
388 if (!tb[LWT_BPF_IN] && !tb[LWT_BPF_OUT] && !tb[LWT_BPF_XMIT])
389 return -EINVAL;
390
391 newts = lwtunnel_state_alloc(sizeof(*bpf));
392 if (!newts)
393 return -ENOMEM;
394
395 newts->type = LWTUNNEL_ENCAP_BPF;
396 bpf = bpf_lwt_lwtunnel(newts);
397
398 if (tb[LWT_BPF_IN]) {
399 newts->flags |= LWTUNNEL_STATE_INPUT_REDIRECT;
400 ret = bpf_parse_prog(tb[LWT_BPF_IN], &bpf->in,
401 BPF_PROG_TYPE_LWT_IN);
402 if (ret < 0)
403 goto errout;
404 }
405
406 if (tb[LWT_BPF_OUT]) {
407 newts->flags |= LWTUNNEL_STATE_OUTPUT_REDIRECT;
408 ret = bpf_parse_prog(tb[LWT_BPF_OUT], &bpf->out,
409 BPF_PROG_TYPE_LWT_OUT);
410 if (ret < 0)
411 goto errout;
412 }
413
414 if (tb[LWT_BPF_XMIT]) {
415 newts->flags |= LWTUNNEL_STATE_XMIT_REDIRECT;
416 ret = bpf_parse_prog(tb[LWT_BPF_XMIT], &bpf->xmit,
417 BPF_PROG_TYPE_LWT_XMIT);
418 if (ret < 0)
419 goto errout;
420 }
421
422 if (tb[LWT_BPF_XMIT_HEADROOM]) {
423 u32 headroom = nla_get_u32(tb[LWT_BPF_XMIT_HEADROOM]);
424
425 if (headroom > LWT_BPF_MAX_HEADROOM) {
426 ret = -ERANGE;
427 goto errout;
428 }
429
430 newts->headroom = headroom;
431 }
432
433 bpf->family = family;
434 *ts = newts;
435
436 return 0;
437
438 errout:
439 bpf_destroy_state(newts);
440 kfree(newts);
441 return ret;
442 }
443
444 static int bpf_fill_lwt_prog(struct sk_buff *skb, int attr,
445 struct bpf_lwt_prog *prog)
446 {
447 struct nlattr *nest;
448
449 if (!prog->prog)
450 return 0;
451
452 nest = nla_nest_start_noflag(skb, attr);
453 if (!nest)
454 return -EMSGSIZE;
455
456 if (prog->name &&
457 nla_put_string(skb, LWT_BPF_PROG_NAME, prog->name))
458 return -EMSGSIZE;
459
460 return nla_nest_end(skb, nest);
461 }
462
463 static int bpf_fill_encap_info(struct sk_buff *skb, struct lwtunnel_state *lwt)
464 {
465 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
466
467 if (bpf_fill_lwt_prog(skb, LWT_BPF_IN, &bpf->in) < 0 ||
468 bpf_fill_lwt_prog(skb, LWT_BPF_OUT, &bpf->out) < 0 ||
469 bpf_fill_lwt_prog(skb, LWT_BPF_XMIT, &bpf->xmit) < 0)
470 return -EMSGSIZE;
471
472 return 0;
473 }
474
475 static int bpf_encap_nlsize(struct lwtunnel_state *lwtstate)
476 {
477 int nest_len = nla_total_size(sizeof(struct nlattr)) +
478 nla_total_size(MAX_PROG_NAME) + /* LWT_BPF_PROG_NAME */
479 0;
480
481 return nest_len + /* LWT_BPF_IN */
482 nest_len + /* LWT_BPF_OUT */
483 nest_len + /* LWT_BPF_XMIT */
484 0;
485 }
486
487 static int bpf_lwt_prog_cmp(struct bpf_lwt_prog *a, struct bpf_lwt_prog *b)
488 {
489 /* FIXME:
490 * The LWT state is currently rebuilt for delete requests which
491 * results in a new bpf_prog instance. Comparing names for now.
492 */
493 if (!a->name && !b->name)
494 return 0;
495
496 if (!a->name || !b->name)
497 return 1;
498
499 return strcmp(a->name, b->name);
500 }
501
502 static int bpf_encap_cmp(struct lwtunnel_state *a, struct lwtunnel_state *b)
503 {
504 struct bpf_lwt *a_bpf = bpf_lwt_lwtunnel(a);
505 struct bpf_lwt *b_bpf = bpf_lwt_lwtunnel(b);
506
507 return bpf_lwt_prog_cmp(&a_bpf->in, &b_bpf->in) ||
508 bpf_lwt_prog_cmp(&a_bpf->out, &b_bpf->out) ||
509 bpf_lwt_prog_cmp(&a_bpf->xmit, &b_bpf->xmit);
510 }
511
512 static const struct lwtunnel_encap_ops bpf_encap_ops = {
513 .build_state = bpf_build_state,
514 .destroy_state = bpf_destroy_state,
515 .input = bpf_input,
516 .output = bpf_output,
517 .xmit = bpf_xmit,
518 .fill_encap = bpf_fill_encap_info,
519 .get_encap_size = bpf_encap_nlsize,
520 .cmp_encap = bpf_encap_cmp,
521 .owner = THIS_MODULE,
522 };
523
524 static int handle_gso_type(struct sk_buff *skb, unsigned int gso_type,
525 int encap_len)
526 {
527 struct skb_shared_info *shinfo = skb_shinfo(skb);
528
529 gso_type |= SKB_GSO_DODGY;
530 shinfo->gso_type |= gso_type;
531 skb_decrease_gso_size(shinfo, encap_len);
532 shinfo->gso_segs = 0;
533 return 0;
534 }
535
536 static int handle_gso_encap(struct sk_buff *skb, bool ipv4, int encap_len)
537 {
538 int next_hdr_offset;
539 void *next_hdr;
540 __u8 protocol;
541
542 /* SCTP and UDP_L4 gso need more nuanced handling than what
543 * handle_gso_type() does above: skb_decrease_gso_size() is not enough.
544 * So at the moment only TCP GSO packets are let through.
545 */
546 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
547 return -ENOTSUPP;
548
549 if (ipv4) {
550 protocol = ip_hdr(skb)->protocol;
551 next_hdr_offset = sizeof(struct iphdr);
552 next_hdr = skb_network_header(skb) + next_hdr_offset;
553 } else {
554 protocol = ipv6_hdr(skb)->nexthdr;
555 next_hdr_offset = sizeof(struct ipv6hdr);
556 next_hdr = skb_network_header(skb) + next_hdr_offset;
557 }
558
559 switch (protocol) {
560 case IPPROTO_GRE:
561 next_hdr_offset += sizeof(struct gre_base_hdr);
562 if (next_hdr_offset > encap_len)
563 return -EINVAL;
564
565 if (((struct gre_base_hdr *)next_hdr)->flags & GRE_CSUM)
566 return handle_gso_type(skb, SKB_GSO_GRE_CSUM,
567 encap_len);
568 return handle_gso_type(skb, SKB_GSO_GRE, encap_len);
569
570 case IPPROTO_UDP:
571 next_hdr_offset += sizeof(struct udphdr);
572 if (next_hdr_offset > encap_len)
573 return -EINVAL;
574
575 if (((struct udphdr *)next_hdr)->check)
576 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL_CSUM,
577 encap_len);
578 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL, encap_len);
579
580 case IPPROTO_IP:
581 case IPPROTO_IPV6:
582 if (ipv4)
583 return handle_gso_type(skb, SKB_GSO_IPXIP4, encap_len);
584 else
585 return handle_gso_type(skb, SKB_GSO_IPXIP6, encap_len);
586
587 default:
588 return -EPROTONOSUPPORT;
589 }
590 }
591
592 int bpf_lwt_push_ip_encap(struct sk_buff *skb, void *hdr, u32 len, bool ingress)
593 {
594 struct iphdr *iph;
595 bool ipv4;
596 int err;
597
598 if (unlikely(len < sizeof(struct iphdr) || len > LWT_BPF_MAX_HEADROOM))
599 return -EINVAL;
600
601 /* validate protocol and length */
602 iph = (struct iphdr *)hdr;
603 if (iph->version == 4) {
604 ipv4 = true;
605 if (unlikely(len < iph->ihl * 4))
606 return -EINVAL;
607 } else if (iph->version == 6) {
608 ipv4 = false;
609 if (unlikely(len < sizeof(struct ipv6hdr)))
610 return -EINVAL;
611 } else {
612 return -EINVAL;
613 }
614
615 if (ingress)
616 err = skb_cow_head(skb, len + skb->mac_len);
617 else
618 err = skb_cow_head(skb,
619 len + LL_RESERVED_SPACE(skb_dst(skb)->dev));
620 if (unlikely(err))
621 return err;
622
623 /* push the encap headers and fix pointers */
624 skb_reset_inner_headers(skb);
625 skb_reset_inner_mac_header(skb); /* mac header is not yet set */
626 skb_set_inner_protocol(skb, skb->protocol);
627 skb->encapsulation = 1;
628 skb_push(skb, len);
629 if (ingress)
630 skb_postpush_rcsum(skb, iph, len);
631 skb_reset_network_header(skb);
632 memcpy(skb_network_header(skb), hdr, len);
633 bpf_compute_data_pointers(skb);
634 skb_clear_hash(skb);
635
636 if (ipv4) {
637 skb->protocol = htons(ETH_P_IP);
638 iph = ip_hdr(skb);
639
640 if (!iph->check)
641 iph->check = ip_fast_csum((unsigned char *)iph,
642 iph->ihl);
643 } else {
644 skb->protocol = htons(ETH_P_IPV6);
645 }
646
647 if (skb_is_gso(skb))
648 return handle_gso_encap(skb, ipv4, len);
649
650 return 0;
651 }
652
653 static int __init bpf_lwt_init(void)
654 {
655 return lwtunnel_encap_add_ops(&bpf_encap_ops, LWTUNNEL_ENCAP_BPF);
656 }
657
658 subsys_initcall(bpf_lwt_init)
659
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.