1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * IPV4 GSO/GRO offload support
4 * Linux INET implementation
9 #include <linux/skbuff.h>
12 #include <net/protocol.h>
13 #include <net/inet_common.h>
15 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
16 netdev_features_t features,
17 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
18 netdev_features_t features),
19 __be16 new_protocol, bool is_ipv6)
21 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
22 bool remcsum, need_csum, offload_csum, gso_partial;
23 struct sk_buff *segs = ERR_PTR(-EINVAL);
24 struct udphdr *uh = udp_hdr(skb);
25 u16 mac_offset = skb->mac_header;
26 __be16 protocol = skb->protocol;
27 u16 mac_len = skb->mac_len;
28 int udp_offset, outer_hlen;
32 if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
35 /* Adjust partial header checksum to negate old length.
36 * We cannot rely on the value contained in uh->len as it is
37 * possible that the actual value exceeds the boundaries of the
38 * 16 bit length field due to the header being added outside of an
39 * IP or IPv6 frame that was already limited to 64K - 1.
41 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
42 partial = (__force __wsum)uh->len;
44 partial = (__force __wsum)htonl(skb->len);
45 partial = csum_sub(csum_unfold(uh->check), partial);
47 /* setup inner skb. */
48 skb->encapsulation = 0;
49 SKB_GSO_CB(skb)->encap_level = 0;
50 __skb_pull(skb, tnl_hlen);
51 skb_reset_mac_header(skb);
52 skb_set_network_header(skb, skb_inner_network_offset(skb));
53 skb_set_transport_header(skb, skb_inner_transport_offset(skb));
54 skb->mac_len = skb_inner_network_offset(skb);
55 skb->protocol = new_protocol;
57 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
58 skb->encap_hdr_csum = need_csum;
60 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
61 skb->remcsum_offload = remcsum;
63 need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb));
64 /* Try to offload checksum if possible */
65 offload_csum = !!(need_csum &&
68 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
69 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
71 features &= skb->dev->hw_enc_features;
73 features &= ~NETIF_F_SCTP_CRC;
75 /* The only checksum offload we care about from here on out is the
76 * outer one so strip the existing checksum feature flags and
77 * instead set the flag based on our outer checksum offload value.
80 features &= ~NETIF_F_CSUM_MASK;
81 if (!need_csum || offload_csum)
82 features |= NETIF_F_HW_CSUM;
85 /* segment inner packet. */
86 segs = gso_inner_segment(skb, features);
87 if (IS_ERR_OR_NULL(segs)) {
88 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
93 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
95 outer_hlen = skb_tnl_header_len(skb);
96 udp_offset = outer_hlen - tnl_hlen;
102 skb->ip_summed = CHECKSUM_NONE;
104 /* Set up inner headers if we are offloading inner checksum */
105 if (skb->ip_summed == CHECKSUM_PARTIAL) {
106 skb_reset_inner_headers(skb);
107 skb->encapsulation = 1;
110 skb->mac_len = mac_len;
111 skb->protocol = protocol;
113 __skb_push(skb, outer_hlen);
114 skb_reset_mac_header(skb);
115 skb_set_network_header(skb, mac_len);
116 skb_set_transport_header(skb, udp_offset);
117 len = skb->len - udp_offset;
120 /* If we are only performing partial GSO the inner header
121 * will be using a length value equal to only one MSS sized
122 * segment instead of the entire frame.
124 if (gso_partial && skb_is_gso(skb)) {
125 uh->len = htons(skb_shinfo(skb)->gso_size +
126 SKB_GSO_CB(skb)->data_offset +
127 skb->head - (unsigned char *)uh);
129 uh->len = htons(len);
135 uh->check = ~csum_fold(csum_add(partial,
136 (__force __wsum)htonl(len)));
138 if (skb->encapsulation || !offload_csum) {
139 uh->check = gso_make_checksum(skb, ~uh->check);
141 uh->check = CSUM_MANGLED_0;
143 skb->ip_summed = CHECKSUM_PARTIAL;
144 skb->csum_start = skb_transport_header(skb) - skb->head;
145 skb->csum_offset = offsetof(struct udphdr, check);
147 } while ((skb = skb->next));
152 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
153 netdev_features_t features,
156 const struct net_offload __rcu **offloads;
157 __be16 protocol = skb->protocol;
158 const struct net_offload *ops;
159 struct sk_buff *segs = ERR_PTR(-EINVAL);
160 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
161 netdev_features_t features);
165 switch (skb->inner_protocol_type) {
166 case ENCAP_TYPE_ETHER:
167 protocol = skb->inner_protocol;
168 gso_inner_segment = skb_mac_gso_segment;
170 case ENCAP_TYPE_IPPROTO:
171 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
172 ops = rcu_dereference(offloads[skb->inner_ipproto]);
173 if (!ops || !ops->callbacks.gso_segment)
175 gso_inner_segment = ops->callbacks.gso_segment;
181 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
189 EXPORT_SYMBOL(skb_udp_tunnel_segment);
191 static void __udpv4_gso_segment_csum(struct sk_buff *seg,
192 __be32 *oldip, __be32 *newip,
193 __be16 *oldport, __be16 *newport)
198 if (*oldip == *newip && *oldport == *newport)
205 inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip,
207 inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport,
210 uh->check = CSUM_MANGLED_0;
214 csum_replace4(&iph->check, *oldip, *newip);
218 static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs)
221 struct udphdr *uh, *uh2;
222 struct iphdr *iph, *iph2;
228 if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) &&
229 (udp_hdr(seg)->source == udp_hdr(seg->next)->source) &&
230 (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) &&
231 (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr))
234 while ((seg = seg->next)) {
238 __udpv4_gso_segment_csum(seg,
239 &iph2->saddr, &iph->saddr,
240 &uh2->source, &uh->source);
241 __udpv4_gso_segment_csum(seg,
242 &iph2->daddr, &iph->daddr,
243 &uh2->dest, &uh->dest);
249 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
250 netdev_features_t features,
253 unsigned int mss = skb_shinfo(skb)->gso_size;
255 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
259 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
261 return is_ipv6 ? skb : __udpv4_gso_segment_list_csum(skb);
264 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
265 netdev_features_t features, bool is_ipv6)
267 struct sock *sk = gso_skb->sk;
268 unsigned int sum_truesize = 0;
269 struct sk_buff *segs, *seg;
276 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)
277 return __udp_gso_segment_list(gso_skb, features, is_ipv6);
279 mss = skb_shinfo(gso_skb)->gso_size;
280 if (gso_skb->len <= sizeof(*uh) + mss)
281 return ERR_PTR(-EINVAL);
283 skb_pull(gso_skb, sizeof(*uh));
285 /* clear destructor to avoid skb_segment assigning it to tail */
286 copy_dtor = gso_skb->destructor == sock_wfree;
288 gso_skb->destructor = NULL;
290 segs = skb_segment(gso_skb, features);
291 if (IS_ERR_OR_NULL(segs)) {
293 gso_skb->destructor = sock_wfree;
297 /* GSO partial and frag_list segmentation only requires splitting
298 * the frame into an MSS multiple and possibly a remainder, both
299 * cases return a GSO skb. So update the mss now.
301 if (skb_is_gso(segs))
302 mss *= skb_shinfo(segs)->gso_segs;
307 /* preserve TX timestamp flags and TS key for first segment */
308 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
309 skb_shinfo(seg)->tx_flags |=
310 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
312 /* compute checksum adjustment based on old length versus new */
313 newlen = htons(sizeof(*uh) + mss);
314 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
318 seg->destructor = sock_wfree;
320 sum_truesize += seg->truesize;
329 if (seg->ip_summed == CHECKSUM_PARTIAL)
330 gso_reset_checksum(seg, ~check);
332 uh->check = gso_make_checksum(seg, ~check) ? :
339 /* last packet can be partial gso_size, account for that in checksum */
340 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
342 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
347 if (seg->ip_summed == CHECKSUM_PARTIAL)
348 gso_reset_checksum(seg, ~check);
350 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
352 /* update refcount for the packet */
354 int delta = sum_truesize - gso_skb->truesize;
356 /* In some pathological cases, delta can be negative.
357 * We need to either use refcount_add() or refcount_sub_and_test()
359 if (likely(delta >= 0))
360 refcount_add(delta, &sk->sk_wmem_alloc);
362 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
366 EXPORT_SYMBOL_GPL(__udp_gso_segment);
368 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
369 netdev_features_t features)
371 struct sk_buff *segs = ERR_PTR(-EINVAL);
377 if (skb->encapsulation &&
378 (skb_shinfo(skb)->gso_type &
379 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
380 segs = skb_udp_tunnel_segment(skb, features, false);
384 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
387 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
390 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
391 return __udp_gso_segment(skb, features, false);
393 mss = skb_shinfo(skb)->gso_size;
394 if (unlikely(skb->len <= mss))
397 /* Do software UFO. Complete and fill in the UDP checksum as
398 * HW cannot do checksum of UDP packets sent as multiple
406 csum = skb_checksum(skb, 0, skb->len, 0);
407 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
409 uh->check = CSUM_MANGLED_0;
411 skb->ip_summed = CHECKSUM_UNNECESSARY;
413 /* If there is no outer header we can fake a checksum offload
414 * due to the fact that we have already done the checksum in
415 * software prior to segmenting the frame.
417 if (!skb->encap_hdr_csum)
418 features |= NETIF_F_HW_CSUM;
420 /* Fragment the skb. IP headers of the fragments are updated in
423 segs = skb_segment(skb, features);
428 static int skb_gro_receive_list(struct sk_buff *p, struct sk_buff *skb)
430 if (unlikely(p->len + skb->len >= 65536))
433 if (NAPI_GRO_CB(p)->last == p)
434 skb_shinfo(p)->frag_list = skb;
436 NAPI_GRO_CB(p)->last->next = skb;
438 skb_pull(skb, skb_gro_offset(skb));
440 NAPI_GRO_CB(p)->last = skb;
441 NAPI_GRO_CB(p)->count++;
442 p->data_len += skb->len;
444 /* sk owenrship - if any - completely transferred to the aggregated packet */
445 skb->destructor = NULL;
446 p->truesize += skb->truesize;
449 NAPI_GRO_CB(skb)->same_flow = 1;
455 #define UDP_GRO_CNT_MAX 64
456 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
459 struct udphdr *uh = udp_gro_udphdr(skb);
460 struct sk_buff *pp = NULL;
466 /* requires non zero csum, for symmetry with GSO */
468 NAPI_GRO_CB(skb)->flush = 1;
472 /* Do not deal with padded or malicious packets, sorry ! */
473 ulen = ntohs(uh->len);
474 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
475 NAPI_GRO_CB(skb)->flush = 1;
478 /* pull encapsulating udp header */
479 skb_gro_pull(skb, sizeof(struct udphdr));
481 list_for_each_entry(p, head, list) {
482 if (!NAPI_GRO_CB(p)->same_flow)
487 /* Match ports only, as csum is always non zero */
488 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
489 NAPI_GRO_CB(p)->same_flow = 0;
493 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
494 NAPI_GRO_CB(skb)->flush = 1;
498 /* Terminate the flow on len mismatch or if it grow "too much".
499 * Under small packet flood GRO count could elsewhere grow a lot
500 * leading to excessive truesize values.
501 * On len mismatch merge the first packet shorter than gso_size,
502 * otherwise complete the GRO packet.
504 if (ulen > ntohs(uh2->len)) {
507 if (NAPI_GRO_CB(skb)->is_flist) {
508 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
509 NAPI_GRO_CB(skb)->flush = 1;
512 if ((skb->ip_summed != p->ip_summed) ||
513 (skb->csum_level != p->csum_level)) {
514 NAPI_GRO_CB(skb)->flush = 1;
517 ret = skb_gro_receive_list(p, skb);
519 skb_gro_postpull_rcsum(skb, uh,
520 sizeof(struct udphdr));
522 ret = skb_gro_receive(p, skb);
526 if (ret || ulen != ntohs(uh2->len) ||
527 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
533 /* mismatch, but we never need to flush */
537 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
538 struct udphdr *uh, struct sock *sk)
540 struct sk_buff *pp = NULL;
543 unsigned int off = skb_gro_offset(skb);
546 /* we can do L4 aggregation only if the packet can't land in a tunnel
547 * otherwise we could corrupt the inner stream
549 NAPI_GRO_CB(skb)->is_flist = 0;
550 if (!sk || !udp_sk(sk)->gro_receive) {
551 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
552 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled : 1;
554 if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) ||
555 (sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist)
556 return call_gro_receive(udp_gro_receive_segment, head, skb);
558 /* no GRO, be sure flush the current packet */
562 if (NAPI_GRO_CB(skb)->encap_mark ||
563 (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
564 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
565 !NAPI_GRO_CB(skb)->csum_valid))
568 /* mark that this skb passed once through the tunnel gro layer */
569 NAPI_GRO_CB(skb)->encap_mark = 1;
573 list_for_each_entry(p, head, list) {
574 if (!NAPI_GRO_CB(p)->same_flow)
577 uh2 = (struct udphdr *)(p->data + off);
579 /* Match ports and either checksums are either both zero
582 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
583 (!uh->check ^ !uh2->check)) {
584 NAPI_GRO_CB(p)->same_flow = 0;
589 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
590 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
591 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
594 skb_gro_flush_final(skb, pp, flush);
597 EXPORT_SYMBOL(udp_gro_receive);
599 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
602 const struct iphdr *iph = skb_gro_network_header(skb);
604 return __udp4_lib_lookup(dev_net(skb->dev), iph->saddr, sport,
605 iph->daddr, dport, inet_iif(skb),
606 inet_sdif(skb), &udp_table, NULL);
609 INDIRECT_CALLABLE_SCOPE
610 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
612 struct udphdr *uh = udp_gro_udphdr(skb);
613 struct sock *sk = NULL;
619 /* Don't bother verifying checksum if we're going to flush anyway. */
620 if (NAPI_GRO_CB(skb)->flush)
623 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
624 inet_gro_compute_pseudo))
627 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
628 inet_gro_compute_pseudo);
630 NAPI_GRO_CB(skb)->is_ipv6 = 0;
632 if (static_branch_unlikely(&udp_encap_needed_key))
633 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
635 pp = udp_gro_receive(head, skb, uh, sk);
639 NAPI_GRO_CB(skb)->flush = 1;
643 static int udp_gro_complete_segment(struct sk_buff *skb)
645 struct udphdr *uh = udp_hdr(skb);
647 skb->csum_start = (unsigned char *)uh - skb->head;
648 skb->csum_offset = offsetof(struct udphdr, check);
649 skb->ip_summed = CHECKSUM_PARTIAL;
651 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
652 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
654 if (skb->encapsulation)
655 skb->inner_transport_header = skb->transport_header;
660 int udp_gro_complete(struct sk_buff *skb, int nhoff,
663 __be16 newlen = htons(skb->len - nhoff);
664 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
670 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
671 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
672 if (sk && udp_sk(sk)->gro_complete) {
673 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
674 : SKB_GSO_UDP_TUNNEL;
676 /* clear the encap mark, so that inner frag_list gro_complete
679 NAPI_GRO_CB(skb)->encap_mark = 0;
681 /* Set encapsulation before calling into inner gro_complete()
682 * functions to make them set up the inner offsets.
684 skb->encapsulation = 1;
685 err = udp_sk(sk)->gro_complete(sk, skb,
686 nhoff + sizeof(struct udphdr));
688 err = udp_gro_complete_segment(skb);
691 if (skb->remcsum_offload)
692 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
696 EXPORT_SYMBOL(udp_gro_complete);
698 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
700 const struct iphdr *iph = ip_hdr(skb);
701 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
703 /* do fraglist only if there is no outer UDP encap (or we already processed it) */
704 if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) {
705 uh->len = htons(skb->len - nhoff);
707 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
708 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
710 if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
711 if (skb->csum_level < SKB_MAX_CSUM_LEVEL)
714 skb->ip_summed = CHECKSUM_UNNECESSARY;
722 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
725 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
728 static const struct net_offload udpv4_offload = {
730 .gso_segment = udp4_ufo_fragment,
731 .gro_receive = udp4_gro_receive,
732 .gro_complete = udp4_gro_complete,
736 int __init udpv4_offload_init(void)
738 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);