Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/* Copyright (c) 2016 Thomas Graf <tgraf@tgraf.ch>
3 */
4
5#include <linux/filter.h>
6#include <linux/kernel.h>
7#include <linux/module.h>
8#include <linux/skbuff.h>
9#include <linux/types.h>
10#include <linux/bpf.h>
11#include <net/lwtunnel.h>
12#include <net/gre.h>
13#include <net/ip.h>
14#include <net/ip6_route.h>
15#include <net/ipv6_stubs.h>
16#include <net/inet_dscp.h>
17
18struct bpf_lwt_prog {
19 struct bpf_prog *prog;
20 char *name;
21};
22
23struct bpf_lwt {
24 struct bpf_lwt_prog in;
25 struct bpf_lwt_prog out;
26 struct bpf_lwt_prog xmit;
27 int family;
28};
29
30#define MAX_PROG_NAME 256
31
32static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt)
33{
34 return (struct bpf_lwt *)lwt->data;
35}
36
37#define NO_REDIRECT false
38#define CAN_REDIRECT true
39
40static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
41 struct dst_entry *dst, bool can_redirect)
42{
43 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
44 int ret;
45
46 /* Disabling BH is needed to protect per-CPU bpf_redirect_info between
47 * BPF prog and skb_do_redirect().
48 */
49 local_bh_disable();
50 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
51 bpf_compute_data_pointers(skb);
52 ret = bpf_prog_run_save_cb(lwt->prog, skb);
53
54 switch (ret) {
55 case BPF_OK:
56 case BPF_LWT_REROUTE:
57 break;
58
59 case BPF_REDIRECT:
60 if (unlikely(!can_redirect)) {
61 pr_warn_once("Illegal redirect return code in prog %s\n",
62 lwt->name ? : "<unknown>");
63 ret = BPF_OK;
64 } else {
65 skb_reset_mac_header(skb);
66 skb_do_redirect(skb);
67 ret = BPF_REDIRECT;
68 }
69 break;
70
71 case BPF_DROP:
72 kfree_skb(skb);
73 ret = -EPERM;
74 break;
75
76 default:
77 pr_warn_once("bpf-lwt: Illegal return value %u, expect packet loss\n", ret);
78 kfree_skb(skb);
79 ret = -EINVAL;
80 break;
81 }
82
83 bpf_net_ctx_clear(bpf_net_ctx);
84 local_bh_enable();
85
86 return ret;
87}
88
89static int bpf_lwt_input_reroute(struct sk_buff *skb)
90{
91 enum skb_drop_reason reason;
92 int err = -EINVAL;
93
94 if (skb->protocol == htons(ETH_P_IP)) {
95 struct net_device *dev = skb_dst(skb)->dev;
96 const struct iphdr *iph = ip_hdr(skb);
97
98 dev_hold(dev);
99 skb_dst_drop(skb);
100 reason = ip_route_input_noref(skb, iph->daddr, iph->saddr,
101 ip4h_dscp(iph), dev);
102 err = reason ? -EINVAL : 0;
103 dev_put(dev);
104 } else if (skb->protocol == htons(ETH_P_IPV6)) {
105 skb_dst_drop(skb);
106 err = ipv6_stub->ipv6_route_input(skb);
107 } else {
108 err = -EAFNOSUPPORT;
109 }
110
111 if (err)
112 goto err;
113 return dst_input(skb);
114
115err:
116 kfree_skb(skb);
117 return err;
118}
119
120static int bpf_input(struct sk_buff *skb)
121{
122 struct dst_entry *dst = skb_dst(skb);
123 struct bpf_lwt *bpf;
124 int ret;
125
126 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
127 if (bpf->in.prog) {
128 ret = run_lwt_bpf(skb, &bpf->in, dst, NO_REDIRECT);
129 if (ret < 0)
130 return ret;
131 if (ret == BPF_LWT_REROUTE)
132 return bpf_lwt_input_reroute(skb);
133 }
134
135 if (unlikely(!dst->lwtstate->orig_input)) {
136 kfree_skb(skb);
137 return -EINVAL;
138 }
139
140 return dst->lwtstate->orig_input(skb);
141}
142
143static int bpf_output(struct net *net, struct sock *sk, struct sk_buff *skb)
144{
145 struct dst_entry *dst = skb_dst(skb);
146 struct bpf_lwt *bpf;
147 int ret;
148
149 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
150 if (bpf->out.prog) {
151 ret = run_lwt_bpf(skb, &bpf->out, dst, NO_REDIRECT);
152 if (ret < 0)
153 return ret;
154 }
155
156 if (unlikely(!dst->lwtstate->orig_output)) {
157 pr_warn_once("orig_output not set on dst for prog %s\n",
158 bpf->out.name);
159 kfree_skb(skb);
160 return -EINVAL;
161 }
162
163 return dst->lwtstate->orig_output(net, sk, skb);
164}
165
166static int xmit_check_hhlen(struct sk_buff *skb, int hh_len)
167{
168 if (skb_headroom(skb) < hh_len) {
169 int nhead = HH_DATA_ALIGN(hh_len - skb_headroom(skb));
170
171 if (pskb_expand_head(skb, nhead, 0, GFP_ATOMIC))
172 return -ENOMEM;
173 }
174
175 return 0;
176}
177
178static int bpf_lwt_xmit_reroute(struct sk_buff *skb)
179{
180 struct net_device *l3mdev = l3mdev_master_dev_rcu(skb_dst(skb)->dev);
181 int oif = l3mdev ? l3mdev->ifindex : 0;
182 struct dst_entry *dst = NULL;
183 int err = -EAFNOSUPPORT;
184 struct sock *sk;
185 struct net *net;
186 bool ipv4;
187
188 if (skb->protocol == htons(ETH_P_IP))
189 ipv4 = true;
190 else if (skb->protocol == htons(ETH_P_IPV6))
191 ipv4 = false;
192 else
193 goto err;
194
195 sk = sk_to_full_sk(skb->sk);
196 if (sk) {
197 if (sk->sk_bound_dev_if)
198 oif = sk->sk_bound_dev_if;
199 net = sock_net(sk);
200 } else {
201 net = dev_net(skb_dst(skb)->dev);
202 }
203
204 if (ipv4) {
205 struct iphdr *iph = ip_hdr(skb);
206 struct flowi4 fl4 = {};
207 struct rtable *rt;
208
209 fl4.flowi4_oif = oif;
210 fl4.flowi4_mark = skb->mark;
211 fl4.flowi4_uid = sock_net_uid(net, sk);
212 fl4.flowi4_tos = inet_dscp_to_dsfield(ip4h_dscp(iph));
213 fl4.flowi4_flags = FLOWI_FLAG_ANYSRC;
214 fl4.flowi4_proto = iph->protocol;
215 fl4.daddr = iph->daddr;
216 fl4.saddr = iph->saddr;
217
218 rt = ip_route_output_key(net, &fl4);
219 if (IS_ERR(rt)) {
220 err = PTR_ERR(rt);
221 goto err;
222 }
223 dst = &rt->dst;
224 } else {
225 struct ipv6hdr *iph6 = ipv6_hdr(skb);
226 struct flowi6 fl6 = {};
227
228 fl6.flowi6_oif = oif;
229 fl6.flowi6_mark = skb->mark;
230 fl6.flowi6_uid = sock_net_uid(net, sk);
231 fl6.flowlabel = ip6_flowinfo(iph6);
232 fl6.flowi6_proto = iph6->nexthdr;
233 fl6.daddr = iph6->daddr;
234 fl6.saddr = iph6->saddr;
235
236 dst = ipv6_stub->ipv6_dst_lookup_flow(net, skb->sk, &fl6, NULL);
237 if (IS_ERR(dst)) {
238 err = PTR_ERR(dst);
239 goto err;
240 }
241 }
242 if (unlikely(dst->error)) {
243 err = dst->error;
244 dst_release(dst);
245 goto err;
246 }
247
248 /* Although skb header was reserved in bpf_lwt_push_ip_encap(), it
249 * was done for the previous dst, so we are doing it here again, in
250 * case the new dst needs much more space. The call below is a noop
251 * if there is enough header space in skb.
252 */
253 err = skb_cow_head(skb, LL_RESERVED_SPACE(dst->dev));
254 if (unlikely(err))
255 goto err;
256
257 skb_dst_drop(skb);
258 skb_dst_set(skb, dst);
259
260 err = dst_output(dev_net(skb_dst(skb)->dev), skb->sk, skb);
261 if (unlikely(err))
262 return net_xmit_errno(err);
263
264 /* ip[6]_finish_output2 understand LWTUNNEL_XMIT_DONE */
265 return LWTUNNEL_XMIT_DONE;
266
267err:
268 kfree_skb(skb);
269 return err;
270}
271
272static int bpf_xmit(struct sk_buff *skb)
273{
274 struct dst_entry *dst = skb_dst(skb);
275 struct bpf_lwt *bpf;
276
277 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
278 if (bpf->xmit.prog) {
279 int hh_len = dst->dev->hard_header_len;
280 __be16 proto = skb->protocol;
281 int ret;
282
283 ret = run_lwt_bpf(skb, &bpf->xmit, dst, CAN_REDIRECT);
284 switch (ret) {
285 case BPF_OK:
286 /* If the header changed, e.g. via bpf_lwt_push_encap,
287 * BPF_LWT_REROUTE below should have been used if the
288 * protocol was also changed.
289 */
290 if (skb->protocol != proto) {
291 kfree_skb(skb);
292 return -EINVAL;
293 }
294 /* If the header was expanded, headroom might be too
295 * small for L2 header to come, expand as needed.
296 */
297 ret = xmit_check_hhlen(skb, hh_len);
298 if (unlikely(ret))
299 return ret;
300
301 return LWTUNNEL_XMIT_CONTINUE;
302 case BPF_REDIRECT:
303 return LWTUNNEL_XMIT_DONE;
304 case BPF_LWT_REROUTE:
305 return bpf_lwt_xmit_reroute(skb);
306 default:
307 return ret;
308 }
309 }
310
311 return LWTUNNEL_XMIT_CONTINUE;
312}
313
314static void bpf_lwt_prog_destroy(struct bpf_lwt_prog *prog)
315{
316 if (prog->prog)
317 bpf_prog_put(prog->prog);
318
319 kfree(prog->name);
320}
321
322static void bpf_destroy_state(struct lwtunnel_state *lwt)
323{
324 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
325
326 bpf_lwt_prog_destroy(&bpf->in);
327 bpf_lwt_prog_destroy(&bpf->out);
328 bpf_lwt_prog_destroy(&bpf->xmit);
329}
330
331static const struct nla_policy bpf_prog_policy[LWT_BPF_PROG_MAX + 1] = {
332 [LWT_BPF_PROG_FD] = { .type = NLA_U32, },
333 [LWT_BPF_PROG_NAME] = { .type = NLA_NUL_STRING,
334 .len = MAX_PROG_NAME },
335};
336
337static int bpf_parse_prog(struct nlattr *attr, struct bpf_lwt_prog *prog,
338 enum bpf_prog_type type)
339{
340 struct nlattr *tb[LWT_BPF_PROG_MAX + 1];
341 struct bpf_prog *p;
342 int ret;
343 u32 fd;
344
345 ret = nla_parse_nested_deprecated(tb, LWT_BPF_PROG_MAX, attr,
346 bpf_prog_policy, NULL);
347 if (ret < 0)
348 return ret;
349
350 if (!tb[LWT_BPF_PROG_FD] || !tb[LWT_BPF_PROG_NAME])
351 return -EINVAL;
352
353 prog->name = nla_memdup(tb[LWT_BPF_PROG_NAME], GFP_ATOMIC);
354 if (!prog->name)
355 return -ENOMEM;
356
357 fd = nla_get_u32(tb[LWT_BPF_PROG_FD]);
358 p = bpf_prog_get_type(fd, type);
359 if (IS_ERR(p))
360 return PTR_ERR(p);
361
362 prog->prog = p;
363
364 return 0;
365}
366
367static const struct nla_policy bpf_nl_policy[LWT_BPF_MAX + 1] = {
368 [LWT_BPF_IN] = { .type = NLA_NESTED, },
369 [LWT_BPF_OUT] = { .type = NLA_NESTED, },
370 [LWT_BPF_XMIT] = { .type = NLA_NESTED, },
371 [LWT_BPF_XMIT_HEADROOM] = { .type = NLA_U32 },
372};
373
374static int bpf_build_state(struct net *net, struct nlattr *nla,
375 unsigned int family, const void *cfg,
376 struct lwtunnel_state **ts,
377 struct netlink_ext_ack *extack)
378{
379 struct nlattr *tb[LWT_BPF_MAX + 1];
380 struct lwtunnel_state *newts;
381 struct bpf_lwt *bpf;
382 int ret;
383
384 if (family != AF_INET && family != AF_INET6)
385 return -EAFNOSUPPORT;
386
387 ret = nla_parse_nested_deprecated(tb, LWT_BPF_MAX, nla, bpf_nl_policy,
388 extack);
389 if (ret < 0)
390 return ret;
391
392 if (!tb[LWT_BPF_IN] && !tb[LWT_BPF_OUT] && !tb[LWT_BPF_XMIT])
393 return -EINVAL;
394
395 newts = lwtunnel_state_alloc(sizeof(*bpf));
396 if (!newts)
397 return -ENOMEM;
398
399 newts->type = LWTUNNEL_ENCAP_BPF;
400 bpf = bpf_lwt_lwtunnel(newts);
401
402 if (tb[LWT_BPF_IN]) {
403 newts->flags |= LWTUNNEL_STATE_INPUT_REDIRECT;
404 ret = bpf_parse_prog(tb[LWT_BPF_IN], &bpf->in,
405 BPF_PROG_TYPE_LWT_IN);
406 if (ret < 0)
407 goto errout;
408 }
409
410 if (tb[LWT_BPF_OUT]) {
411 newts->flags |= LWTUNNEL_STATE_OUTPUT_REDIRECT;
412 ret = bpf_parse_prog(tb[LWT_BPF_OUT], &bpf->out,
413 BPF_PROG_TYPE_LWT_OUT);
414 if (ret < 0)
415 goto errout;
416 }
417
418 if (tb[LWT_BPF_XMIT]) {
419 newts->flags |= LWTUNNEL_STATE_XMIT_REDIRECT;
420 ret = bpf_parse_prog(tb[LWT_BPF_XMIT], &bpf->xmit,
421 BPF_PROG_TYPE_LWT_XMIT);
422 if (ret < 0)
423 goto errout;
424 }
425
426 if (tb[LWT_BPF_XMIT_HEADROOM]) {
427 u32 headroom = nla_get_u32(tb[LWT_BPF_XMIT_HEADROOM]);
428
429 if (headroom > LWT_BPF_MAX_HEADROOM) {
430 ret = -ERANGE;
431 goto errout;
432 }
433
434 newts->headroom = headroom;
435 }
436
437 bpf->family = family;
438 *ts = newts;
439
440 return 0;
441
442errout:
443 bpf_destroy_state(newts);
444 kfree(newts);
445 return ret;
446}
447
448static int bpf_fill_lwt_prog(struct sk_buff *skb, int attr,
449 struct bpf_lwt_prog *prog)
450{
451 struct nlattr *nest;
452
453 if (!prog->prog)
454 return 0;
455
456 nest = nla_nest_start_noflag(skb, attr);
457 if (!nest)
458 return -EMSGSIZE;
459
460 if (prog->name &&
461 nla_put_string(skb, LWT_BPF_PROG_NAME, prog->name))
462 return -EMSGSIZE;
463
464 return nla_nest_end(skb, nest);
465}
466
467static int bpf_fill_encap_info(struct sk_buff *skb, struct lwtunnel_state *lwt)
468{
469 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
470
471 if (bpf_fill_lwt_prog(skb, LWT_BPF_IN, &bpf->in) < 0 ||
472 bpf_fill_lwt_prog(skb, LWT_BPF_OUT, &bpf->out) < 0 ||
473 bpf_fill_lwt_prog(skb, LWT_BPF_XMIT, &bpf->xmit) < 0)
474 return -EMSGSIZE;
475
476 return 0;
477}
478
479static int bpf_encap_nlsize(struct lwtunnel_state *lwtstate)
480{
481 int nest_len = nla_total_size(sizeof(struct nlattr)) +
482 nla_total_size(MAX_PROG_NAME) + /* LWT_BPF_PROG_NAME */
483 0;
484
485 return nest_len + /* LWT_BPF_IN */
486 nest_len + /* LWT_BPF_OUT */
487 nest_len + /* LWT_BPF_XMIT */
488 0;
489}
490
491static int bpf_lwt_prog_cmp(struct bpf_lwt_prog *a, struct bpf_lwt_prog *b)
492{
493 /* FIXME:
494 * The LWT state is currently rebuilt for delete requests which
495 * results in a new bpf_prog instance. Comparing names for now.
496 */
497 if (!a->name && !b->name)
498 return 0;
499
500 if (!a->name || !b->name)
501 return 1;
502
503 return strcmp(a->name, b->name);
504}
505
506static int bpf_encap_cmp(struct lwtunnel_state *a, struct lwtunnel_state *b)
507{
508 struct bpf_lwt *a_bpf = bpf_lwt_lwtunnel(a);
509 struct bpf_lwt *b_bpf = bpf_lwt_lwtunnel(b);
510
511 return bpf_lwt_prog_cmp(&a_bpf->in, &b_bpf->in) ||
512 bpf_lwt_prog_cmp(&a_bpf->out, &b_bpf->out) ||
513 bpf_lwt_prog_cmp(&a_bpf->xmit, &b_bpf->xmit);
514}
515
516static const struct lwtunnel_encap_ops bpf_encap_ops = {
517 .build_state = bpf_build_state,
518 .destroy_state = bpf_destroy_state,
519 .input = bpf_input,
520 .output = bpf_output,
521 .xmit = bpf_xmit,
522 .fill_encap = bpf_fill_encap_info,
523 .get_encap_size = bpf_encap_nlsize,
524 .cmp_encap = bpf_encap_cmp,
525 .owner = THIS_MODULE,
526};
527
528static int handle_gso_type(struct sk_buff *skb, unsigned int gso_type,
529 int encap_len)
530{
531 struct skb_shared_info *shinfo = skb_shinfo(skb);
532
533 gso_type |= SKB_GSO_DODGY;
534 shinfo->gso_type |= gso_type;
535 skb_decrease_gso_size(shinfo, encap_len);
536 shinfo->gso_segs = 0;
537 return 0;
538}
539
540static int handle_gso_encap(struct sk_buff *skb, bool ipv4, int encap_len)
541{
542 int next_hdr_offset;
543 void *next_hdr;
544 __u8 protocol;
545
546 /* SCTP and UDP_L4 gso need more nuanced handling than what
547 * handle_gso_type() does above: skb_decrease_gso_size() is not enough.
548 * So at the moment only TCP GSO packets are let through.
549 */
550 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
551 return -ENOTSUPP;
552
553 if (ipv4) {
554 protocol = ip_hdr(skb)->protocol;
555 next_hdr_offset = sizeof(struct iphdr);
556 next_hdr = skb_network_header(skb) + next_hdr_offset;
557 } else {
558 protocol = ipv6_hdr(skb)->nexthdr;
559 next_hdr_offset = sizeof(struct ipv6hdr);
560 next_hdr = skb_network_header(skb) + next_hdr_offset;
561 }
562
563 switch (protocol) {
564 case IPPROTO_GRE:
565 next_hdr_offset += sizeof(struct gre_base_hdr);
566 if (next_hdr_offset > encap_len)
567 return -EINVAL;
568
569 if (((struct gre_base_hdr *)next_hdr)->flags & GRE_CSUM)
570 return handle_gso_type(skb, SKB_GSO_GRE_CSUM,
571 encap_len);
572 return handle_gso_type(skb, SKB_GSO_GRE, encap_len);
573
574 case IPPROTO_UDP:
575 next_hdr_offset += sizeof(struct udphdr);
576 if (next_hdr_offset > encap_len)
577 return -EINVAL;
578
579 if (((struct udphdr *)next_hdr)->check)
580 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL_CSUM,
581 encap_len);
582 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL, encap_len);
583
584 case IPPROTO_IP:
585 case IPPROTO_IPV6:
586 if (ipv4)
587 return handle_gso_type(skb, SKB_GSO_IPXIP4, encap_len);
588 else
589 return handle_gso_type(skb, SKB_GSO_IPXIP6, encap_len);
590
591 default:
592 return -EPROTONOSUPPORT;
593 }
594}
595
596int bpf_lwt_push_ip_encap(struct sk_buff *skb, void *hdr, u32 len, bool ingress)
597{
598 struct iphdr *iph;
599 bool ipv4;
600 int err;
601
602 if (unlikely(len < sizeof(struct iphdr) || len > LWT_BPF_MAX_HEADROOM))
603 return -EINVAL;
604
605 /* validate protocol and length */
606 iph = (struct iphdr *)hdr;
607 if (iph->version == 4) {
608 ipv4 = true;
609 if (unlikely(len < iph->ihl * 4))
610 return -EINVAL;
611 } else if (iph->version == 6) {
612 ipv4 = false;
613 if (unlikely(len < sizeof(struct ipv6hdr)))
614 return -EINVAL;
615 } else {
616 return -EINVAL;
617 }
618
619 if (ingress)
620 err = skb_cow_head(skb, len + skb->mac_len);
621 else
622 err = skb_cow_head(skb,
623 len + LL_RESERVED_SPACE(skb_dst(skb)->dev));
624 if (unlikely(err))
625 return err;
626
627 /* push the encap headers and fix pointers */
628 skb_reset_inner_headers(skb);
629 skb_reset_inner_mac_header(skb); /* mac header is not yet set */
630 skb_set_inner_protocol(skb, skb->protocol);
631 skb->encapsulation = 1;
632 skb_push(skb, len);
633 if (ingress)
634 skb_postpush_rcsum(skb, iph, len);
635 skb_reset_network_header(skb);
636 memcpy(skb_network_header(skb), hdr, len);
637 bpf_compute_data_pointers(skb);
638 skb_clear_hash(skb);
639
640 if (ipv4) {
641 skb->protocol = htons(ETH_P_IP);
642 iph = ip_hdr(skb);
643
644 if (!iph->check)
645 iph->check = ip_fast_csum((unsigned char *)iph,
646 iph->ihl);
647 } else {
648 skb->protocol = htons(ETH_P_IPV6);
649 }
650
651 if (skb_is_gso(skb))
652 return handle_gso_encap(skb, ipv4, len);
653
654 return 0;
655}
656
657static int __init bpf_lwt_init(void)
658{
659 return lwtunnel_encap_add_ops(&bpf_encap_ops, LWTUNNEL_ENCAP_BPF);
660}
661
662subsys_initcall(bpf_lwt_init)
1// SPDX-License-Identifier: GPL-2.0-only
2/* Copyright (c) 2016 Thomas Graf <tgraf@tgraf.ch>
3 */
4
5#include <linux/filter.h>
6#include <linux/kernel.h>
7#include <linux/module.h>
8#include <linux/skbuff.h>
9#include <linux/types.h>
10#include <linux/bpf.h>
11#include <net/lwtunnel.h>
12#include <net/gre.h>
13#include <net/ip6_route.h>
14#include <net/ipv6_stubs.h>
15
16struct bpf_lwt_prog {
17 struct bpf_prog *prog;
18 char *name;
19};
20
21struct bpf_lwt {
22 struct bpf_lwt_prog in;
23 struct bpf_lwt_prog out;
24 struct bpf_lwt_prog xmit;
25 int family;
26};
27
28#define MAX_PROG_NAME 256
29
30static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt)
31{
32 return (struct bpf_lwt *)lwt->data;
33}
34
35#define NO_REDIRECT false
36#define CAN_REDIRECT true
37
38static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
39 struct dst_entry *dst, bool can_redirect)
40{
41 int ret;
42
43 /* Migration disable and BH disable are needed to protect per-cpu
44 * redirect_info between BPF prog and skb_do_redirect().
45 */
46 migrate_disable();
47 local_bh_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 local_bh_enable();
82 migrate_enable();
83
84 return ret;
85}
86
87static int bpf_lwt_input_reroute(struct sk_buff *skb)
88{
89 int err = -EINVAL;
90
91 if (skb->protocol == htons(ETH_P_IP)) {
92 struct net_device *dev = skb_dst(skb)->dev;
93 struct iphdr *iph = ip_hdr(skb);
94
95 dev_hold(dev);
96 skb_dst_drop(skb);
97 err = ip_route_input_noref(skb, iph->daddr, iph->saddr,
98 iph->tos, dev);
99 dev_put(dev);
100 } else if (skb->protocol == htons(ETH_P_IPV6)) {
101 skb_dst_drop(skb);
102 err = ipv6_stub->ipv6_route_input(skb);
103 } else {
104 err = -EAFNOSUPPORT;
105 }
106
107 if (err)
108 goto err;
109 return dst_input(skb);
110
111err:
112 kfree_skb(skb);
113 return err;
114}
115
116static int bpf_input(struct sk_buff *skb)
117{
118 struct dst_entry *dst = skb_dst(skb);
119 struct bpf_lwt *bpf;
120 int ret;
121
122 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
123 if (bpf->in.prog) {
124 ret = run_lwt_bpf(skb, &bpf->in, dst, NO_REDIRECT);
125 if (ret < 0)
126 return ret;
127 if (ret == BPF_LWT_REROUTE)
128 return bpf_lwt_input_reroute(skb);
129 }
130
131 if (unlikely(!dst->lwtstate->orig_input)) {
132 kfree_skb(skb);
133 return -EINVAL;
134 }
135
136 return dst->lwtstate->orig_input(skb);
137}
138
139static int bpf_output(struct net *net, struct sock *sk, struct sk_buff *skb)
140{
141 struct dst_entry *dst = skb_dst(skb);
142 struct bpf_lwt *bpf;
143 int ret;
144
145 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
146 if (bpf->out.prog) {
147 ret = run_lwt_bpf(skb, &bpf->out, dst, NO_REDIRECT);
148 if (ret < 0)
149 return ret;
150 }
151
152 if (unlikely(!dst->lwtstate->orig_output)) {
153 pr_warn_once("orig_output not set on dst for prog %s\n",
154 bpf->out.name);
155 kfree_skb(skb);
156 return -EINVAL;
157 }
158
159 return dst->lwtstate->orig_output(net, sk, skb);
160}
161
162static int xmit_check_hhlen(struct sk_buff *skb, int hh_len)
163{
164 if (skb_headroom(skb) < hh_len) {
165 int nhead = HH_DATA_ALIGN(hh_len - skb_headroom(skb));
166
167 if (pskb_expand_head(skb, nhead, 0, GFP_ATOMIC))
168 return -ENOMEM;
169 }
170
171 return 0;
172}
173
174static int bpf_lwt_xmit_reroute(struct sk_buff *skb)
175{
176 struct net_device *l3mdev = l3mdev_master_dev_rcu(skb_dst(skb)->dev);
177 int oif = l3mdev ? l3mdev->ifindex : 0;
178 struct dst_entry *dst = NULL;
179 int err = -EAFNOSUPPORT;
180 struct sock *sk;
181 struct net *net;
182 bool ipv4;
183
184 if (skb->protocol == htons(ETH_P_IP))
185 ipv4 = true;
186 else if (skb->protocol == htons(ETH_P_IPV6))
187 ipv4 = false;
188 else
189 goto err;
190
191 sk = sk_to_full_sk(skb->sk);
192 if (sk) {
193 if (sk->sk_bound_dev_if)
194 oif = sk->sk_bound_dev_if;
195 net = sock_net(sk);
196 } else {
197 net = dev_net(skb_dst(skb)->dev);
198 }
199
200 if (ipv4) {
201 struct iphdr *iph = ip_hdr(skb);
202 struct flowi4 fl4 = {};
203 struct rtable *rt;
204
205 fl4.flowi4_oif = oif;
206 fl4.flowi4_mark = skb->mark;
207 fl4.flowi4_uid = sock_net_uid(net, sk);
208 fl4.flowi4_tos = RT_TOS(iph->tos);
209 fl4.flowi4_flags = FLOWI_FLAG_ANYSRC;
210 fl4.flowi4_proto = iph->protocol;
211 fl4.daddr = iph->daddr;
212 fl4.saddr = iph->saddr;
213
214 rt = ip_route_output_key(net, &fl4);
215 if (IS_ERR(rt)) {
216 err = PTR_ERR(rt);
217 goto err;
218 }
219 dst = &rt->dst;
220 } else {
221 struct ipv6hdr *iph6 = ipv6_hdr(skb);
222 struct flowi6 fl6 = {};
223
224 fl6.flowi6_oif = oif;
225 fl6.flowi6_mark = skb->mark;
226 fl6.flowi6_uid = sock_net_uid(net, sk);
227 fl6.flowlabel = ip6_flowinfo(iph6);
228 fl6.flowi6_proto = iph6->nexthdr;
229 fl6.daddr = iph6->daddr;
230 fl6.saddr = iph6->saddr;
231
232 dst = ipv6_stub->ipv6_dst_lookup_flow(net, skb->sk, &fl6, NULL);
233 if (IS_ERR(dst)) {
234 err = PTR_ERR(dst);
235 goto err;
236 }
237 }
238 if (unlikely(dst->error)) {
239 err = dst->error;
240 dst_release(dst);
241 goto err;
242 }
243
244 /* Although skb header was reserved in bpf_lwt_push_ip_encap(), it
245 * was done for the previous dst, so we are doing it here again, in
246 * case the new dst needs much more space. The call below is a noop
247 * if there is enough header space in skb.
248 */
249 err = skb_cow_head(skb, LL_RESERVED_SPACE(dst->dev));
250 if (unlikely(err))
251 goto err;
252
253 skb_dst_drop(skb);
254 skb_dst_set(skb, dst);
255
256 err = dst_output(dev_net(skb_dst(skb)->dev), skb->sk, skb);
257 if (unlikely(err))
258 return err;
259
260 /* ip[6]_finish_output2 understand LWTUNNEL_XMIT_DONE */
261 return LWTUNNEL_XMIT_DONE;
262
263err:
264 kfree_skb(skb);
265 return err;
266}
267
268static int bpf_xmit(struct sk_buff *skb)
269{
270 struct dst_entry *dst = skb_dst(skb);
271 struct bpf_lwt *bpf;
272
273 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
274 if (bpf->xmit.prog) {
275 int hh_len = dst->dev->hard_header_len;
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, hh_len);
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
310static 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
318static 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
327static 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
333static 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
363static 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
370static 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
438errout:
439 bpf_destroy_state(newts);
440 kfree(newts);
441 return ret;
442}
443
444static 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
463static 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
475static 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
487static 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
502static 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
512static 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
524static 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
536static 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
592int 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
653static int __init bpf_lwt_init(void)
654{
655 return lwtunnel_encap_add_ops(&bpf_encap_ops, LWTUNNEL_ENCAP_BPF);
656}
657
658subsys_initcall(bpf_lwt_init)