1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Thomas Graf <tgraf@tgraf.ch>
5 #include <linux/kernel.h>
6 #include <linux/module.h>
7 #include <linux/skbuff.h>
8 #include <linux/types.h>
10 #include <net/lwtunnel.h>
12 #include <net/ip6_route.h>
13 #include <net/ipv6_stubs.h>
16 struct bpf_prog *prog;
21 struct bpf_lwt_prog in;
22 struct bpf_lwt_prog out;
23 struct bpf_lwt_prog xmit;
27 #define MAX_PROG_NAME 256
29 static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt)
31 return (struct bpf_lwt *)lwt->data;
34 #define NO_REDIRECT false
35 #define CAN_REDIRECT true
37 static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
38 struct dst_entry *dst, bool can_redirect)
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.
48 bpf_compute_data_pointers(skb);
49 ret = bpf_prog_run_save_cb(lwt->prog, skb);
57 if (unlikely(!can_redirect)) {
58 pr_warn_once("Illegal redirect return code in prog %s\n",
59 lwt->name ? : "<unknown>");
62 skb_reset_mac_header(skb);
63 ret = skb_do_redirect(skb);
75 pr_warn_once("bpf-lwt: Illegal return value %u, expect packet loss\n", ret);
86 static int bpf_lwt_input_reroute(struct sk_buff *skb)
90 if (skb->protocol == htons(ETH_P_IP)) {
91 struct net_device *dev = skb_dst(skb)->dev;
92 struct iphdr *iph = ip_hdr(skb);
96 err = ip_route_input_noref(skb, iph->daddr, iph->saddr,
99 } else if (skb->protocol == htons(ETH_P_IPV6)) {
101 err = ipv6_stub->ipv6_route_input(skb);
108 return dst_input(skb);
115 static int bpf_input(struct sk_buff *skb)
117 struct dst_entry *dst = skb_dst(skb);
121 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
123 ret = run_lwt_bpf(skb, &bpf->in, dst, NO_REDIRECT);
126 if (ret == BPF_LWT_REROUTE)
127 return bpf_lwt_input_reroute(skb);
130 if (unlikely(!dst->lwtstate->orig_input)) {
135 return dst->lwtstate->orig_input(skb);
138 static int bpf_output(struct net *net, struct sock *sk, struct sk_buff *skb)
140 struct dst_entry *dst = skb_dst(skb);
144 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
146 ret = run_lwt_bpf(skb, &bpf->out, dst, NO_REDIRECT);
151 if (unlikely(!dst->lwtstate->orig_output)) {
152 pr_warn_once("orig_output not set on dst for prog %s\n",
158 return dst->lwtstate->orig_output(net, sk, skb);
161 static int xmit_check_hhlen(struct sk_buff *skb)
163 int hh_len = skb_dst(skb)->dev->hard_header_len;
165 if (skb_headroom(skb) < hh_len) {
166 int nhead = HH_DATA_ALIGN(hh_len - skb_headroom(skb));
168 if (pskb_expand_head(skb, nhead, 0, GFP_ATOMIC))
175 static int bpf_lwt_xmit_reroute(struct sk_buff *skb)
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;
185 if (skb->protocol == htons(ETH_P_IP))
187 else if (skb->protocol == htons(ETH_P_IPV6))
192 sk = sk_to_full_sk(skb->sk);
194 if (sk->sk_bound_dev_if)
195 oif = sk->sk_bound_dev_if;
198 net = dev_net(skb_dst(skb)->dev);
202 struct iphdr *iph = ip_hdr(skb);
203 struct flowi4 fl4 = {};
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;
215 rt = ip_route_output_key(net, &fl4);
222 struct ipv6hdr *iph6 = ipv6_hdr(skb);
223 struct flowi6 fl6 = {};
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;
233 err = ipv6_stub->ipv6_dst_lookup(net, skb->sk, &dst, &fl6);
241 if (unlikely(dst->error)) {
247 /* Although skb header was reserved in bpf_lwt_push_ip_encap(), it
248 * was done for the previous dst, so we are doing it here again, in
249 * case the new dst needs much more space. The call below is a noop
250 * if there is enough header space in skb.
252 err = skb_cow_head(skb, LL_RESERVED_SPACE(dst->dev));
257 skb_dst_set(skb, dst);
259 err = dst_output(dev_net(skb_dst(skb)->dev), skb->sk, skb);
263 /* ip[6]_finish_output2 understand LWTUNNEL_XMIT_DONE */
264 return LWTUNNEL_XMIT_DONE;
271 static int bpf_xmit(struct sk_buff *skb)
273 struct dst_entry *dst = skb_dst(skb);
276 bpf = bpf_lwt_lwtunnel(dst->lwtstate);
277 if (bpf->xmit.prog) {
278 __be16 proto = skb->protocol;
281 ret = run_lwt_bpf(skb, &bpf->xmit, dst, CAN_REDIRECT);
284 /* If the header changed, e.g. via bpf_lwt_push_encap,
285 * BPF_LWT_REROUTE below should have been used if the
286 * protocol was also changed.
288 if (skb->protocol != proto) {
292 /* If the header was expanded, headroom might be too
293 * small for L2 header to come, expand as needed.
295 ret = xmit_check_hhlen(skb);
299 return LWTUNNEL_XMIT_CONTINUE;
301 return LWTUNNEL_XMIT_DONE;
302 case BPF_LWT_REROUTE:
303 return bpf_lwt_xmit_reroute(skb);
309 return LWTUNNEL_XMIT_CONTINUE;
312 static void bpf_lwt_prog_destroy(struct bpf_lwt_prog *prog)
315 bpf_prog_put(prog->prog);
320 static void bpf_destroy_state(struct lwtunnel_state *lwt)
322 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
324 bpf_lwt_prog_destroy(&bpf->in);
325 bpf_lwt_prog_destroy(&bpf->out);
326 bpf_lwt_prog_destroy(&bpf->xmit);
329 static const struct nla_policy bpf_prog_policy[LWT_BPF_PROG_MAX + 1] = {
330 [LWT_BPF_PROG_FD] = { .type = NLA_U32, },
331 [LWT_BPF_PROG_NAME] = { .type = NLA_NUL_STRING,
332 .len = MAX_PROG_NAME },
335 static int bpf_parse_prog(struct nlattr *attr, struct bpf_lwt_prog *prog,
336 enum bpf_prog_type type)
338 struct nlattr *tb[LWT_BPF_PROG_MAX + 1];
343 ret = nla_parse_nested_deprecated(tb, LWT_BPF_PROG_MAX, attr,
344 bpf_prog_policy, NULL);
348 if (!tb[LWT_BPF_PROG_FD] || !tb[LWT_BPF_PROG_NAME])
351 prog->name = nla_memdup(tb[LWT_BPF_PROG_NAME], GFP_ATOMIC);
355 fd = nla_get_u32(tb[LWT_BPF_PROG_FD]);
356 p = bpf_prog_get_type(fd, type);
365 static const struct nla_policy bpf_nl_policy[LWT_BPF_MAX + 1] = {
366 [LWT_BPF_IN] = { .type = NLA_NESTED, },
367 [LWT_BPF_OUT] = { .type = NLA_NESTED, },
368 [LWT_BPF_XMIT] = { .type = NLA_NESTED, },
369 [LWT_BPF_XMIT_HEADROOM] = { .type = NLA_U32 },
372 static int bpf_build_state(struct nlattr *nla,
373 unsigned int family, const void *cfg,
374 struct lwtunnel_state **ts,
375 struct netlink_ext_ack *extack)
377 struct nlattr *tb[LWT_BPF_MAX + 1];
378 struct lwtunnel_state *newts;
382 if (family != AF_INET && family != AF_INET6)
383 return -EAFNOSUPPORT;
385 ret = nla_parse_nested_deprecated(tb, LWT_BPF_MAX, nla, bpf_nl_policy,
390 if (!tb[LWT_BPF_IN] && !tb[LWT_BPF_OUT] && !tb[LWT_BPF_XMIT])
393 newts = lwtunnel_state_alloc(sizeof(*bpf));
397 newts->type = LWTUNNEL_ENCAP_BPF;
398 bpf = bpf_lwt_lwtunnel(newts);
400 if (tb[LWT_BPF_IN]) {
401 newts->flags |= LWTUNNEL_STATE_INPUT_REDIRECT;
402 ret = bpf_parse_prog(tb[LWT_BPF_IN], &bpf->in,
403 BPF_PROG_TYPE_LWT_IN);
408 if (tb[LWT_BPF_OUT]) {
409 newts->flags |= LWTUNNEL_STATE_OUTPUT_REDIRECT;
410 ret = bpf_parse_prog(tb[LWT_BPF_OUT], &bpf->out,
411 BPF_PROG_TYPE_LWT_OUT);
416 if (tb[LWT_BPF_XMIT]) {
417 newts->flags |= LWTUNNEL_STATE_XMIT_REDIRECT;
418 ret = bpf_parse_prog(tb[LWT_BPF_XMIT], &bpf->xmit,
419 BPF_PROG_TYPE_LWT_XMIT);
424 if (tb[LWT_BPF_XMIT_HEADROOM]) {
425 u32 headroom = nla_get_u32(tb[LWT_BPF_XMIT_HEADROOM]);
427 if (headroom > LWT_BPF_MAX_HEADROOM) {
432 newts->headroom = headroom;
435 bpf->family = family;
441 bpf_destroy_state(newts);
446 static int bpf_fill_lwt_prog(struct sk_buff *skb, int attr,
447 struct bpf_lwt_prog *prog)
454 nest = nla_nest_start_noflag(skb, attr);
459 nla_put_string(skb, LWT_BPF_PROG_NAME, prog->name))
462 return nla_nest_end(skb, nest);
465 static int bpf_fill_encap_info(struct sk_buff *skb, struct lwtunnel_state *lwt)
467 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
469 if (bpf_fill_lwt_prog(skb, LWT_BPF_IN, &bpf->in) < 0 ||
470 bpf_fill_lwt_prog(skb, LWT_BPF_OUT, &bpf->out) < 0 ||
471 bpf_fill_lwt_prog(skb, LWT_BPF_XMIT, &bpf->xmit) < 0)
477 static int bpf_encap_nlsize(struct lwtunnel_state *lwtstate)
479 int nest_len = nla_total_size(sizeof(struct nlattr)) +
480 nla_total_size(MAX_PROG_NAME) + /* LWT_BPF_PROG_NAME */
483 return nest_len + /* LWT_BPF_IN */
484 nest_len + /* LWT_BPF_OUT */
485 nest_len + /* LWT_BPF_XMIT */
489 static int bpf_lwt_prog_cmp(struct bpf_lwt_prog *a, struct bpf_lwt_prog *b)
492 * The LWT state is currently rebuilt for delete requests which
493 * results in a new bpf_prog instance. Comparing names for now.
495 if (!a->name && !b->name)
498 if (!a->name || !b->name)
501 return strcmp(a->name, b->name);
504 static int bpf_encap_cmp(struct lwtunnel_state *a, struct lwtunnel_state *b)
506 struct bpf_lwt *a_bpf = bpf_lwt_lwtunnel(a);
507 struct bpf_lwt *b_bpf = bpf_lwt_lwtunnel(b);
509 return bpf_lwt_prog_cmp(&a_bpf->in, &b_bpf->in) ||
510 bpf_lwt_prog_cmp(&a_bpf->out, &b_bpf->out) ||
511 bpf_lwt_prog_cmp(&a_bpf->xmit, &b_bpf->xmit);
514 static const struct lwtunnel_encap_ops bpf_encap_ops = {
515 .build_state = bpf_build_state,
516 .destroy_state = bpf_destroy_state,
518 .output = bpf_output,
520 .fill_encap = bpf_fill_encap_info,
521 .get_encap_size = bpf_encap_nlsize,
522 .cmp_encap = bpf_encap_cmp,
523 .owner = THIS_MODULE,
526 static int handle_gso_type(struct sk_buff *skb, unsigned int gso_type,
529 struct skb_shared_info *shinfo = skb_shinfo(skb);
531 gso_type |= SKB_GSO_DODGY;
532 shinfo->gso_type |= gso_type;
533 skb_decrease_gso_size(shinfo, encap_len);
534 shinfo->gso_segs = 0;
538 static int handle_gso_encap(struct sk_buff *skb, bool ipv4, int encap_len)
544 /* SCTP and UDP_L4 gso need more nuanced handling than what
545 * handle_gso_type() does above: skb_decrease_gso_size() is not enough.
546 * So at the moment only TCP GSO packets are let through.
548 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
552 protocol = ip_hdr(skb)->protocol;
553 next_hdr_offset = sizeof(struct iphdr);
554 next_hdr = skb_network_header(skb) + next_hdr_offset;
556 protocol = ipv6_hdr(skb)->nexthdr;
557 next_hdr_offset = sizeof(struct ipv6hdr);
558 next_hdr = skb_network_header(skb) + next_hdr_offset;
563 next_hdr_offset += sizeof(struct gre_base_hdr);
564 if (next_hdr_offset > encap_len)
567 if (((struct gre_base_hdr *)next_hdr)->flags & GRE_CSUM)
568 return handle_gso_type(skb, SKB_GSO_GRE_CSUM,
570 return handle_gso_type(skb, SKB_GSO_GRE, encap_len);
573 next_hdr_offset += sizeof(struct udphdr);
574 if (next_hdr_offset > encap_len)
577 if (((struct udphdr *)next_hdr)->check)
578 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL_CSUM,
580 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL, encap_len);
585 return handle_gso_type(skb, SKB_GSO_IPXIP4, encap_len);
587 return handle_gso_type(skb, SKB_GSO_IPXIP6, encap_len);
590 return -EPROTONOSUPPORT;
594 int bpf_lwt_push_ip_encap(struct sk_buff *skb, void *hdr, u32 len, bool ingress)
600 if (unlikely(len < sizeof(struct iphdr) || len > LWT_BPF_MAX_HEADROOM))
603 /* validate protocol and length */
604 iph = (struct iphdr *)hdr;
605 if (iph->version == 4) {
607 if (unlikely(len < iph->ihl * 4))
609 } else if (iph->version == 6) {
611 if (unlikely(len < sizeof(struct ipv6hdr)))
618 err = skb_cow_head(skb, len + skb->mac_len);
620 err = skb_cow_head(skb,
621 len + LL_RESERVED_SPACE(skb_dst(skb)->dev));
625 /* push the encap headers and fix pointers */
626 skb_reset_inner_headers(skb);
627 skb_reset_inner_mac_header(skb); /* mac header is not yet set */
628 skb_set_inner_protocol(skb, skb->protocol);
629 skb->encapsulation = 1;
632 skb_postpush_rcsum(skb, iph, len);
633 skb_reset_network_header(skb);
634 memcpy(skb_network_header(skb), hdr, len);
635 bpf_compute_data_pointers(skb);
639 skb->protocol = htons(ETH_P_IP);
643 iph->check = ip_fast_csum((unsigned char *)iph,
646 skb->protocol = htons(ETH_P_IPV6);
650 return handle_gso_encap(skb, ipv4, len);
655 static int __init bpf_lwt_init(void)
657 return lwtunnel_encap_add_ops(&bpf_encap_ops, LWTUNNEL_ENCAP_BPF);
660 subsys_initcall(bpf_lwt_init)