1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * IPV4 GSO/GRO offload support
4 * Linux INET implementation
9 #include <linux/skbuff.h>
11 #include <net/protocol.h>
12 #include <net/inet_common.h>
14 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
15 netdev_features_t features,
16 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
17 netdev_features_t features),
18 __be16 new_protocol, bool is_ipv6)
20 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
21 bool remcsum, need_csum, offload_csum, gso_partial;
22 struct sk_buff *segs = ERR_PTR(-EINVAL);
23 struct udphdr *uh = udp_hdr(skb);
24 u16 mac_offset = skb->mac_header;
25 __be16 protocol = skb->protocol;
26 u16 mac_len = skb->mac_len;
27 int udp_offset, outer_hlen;
31 if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
34 /* Adjust partial header checksum to negate old length.
35 * We cannot rely on the value contained in uh->len as it is
36 * possible that the actual value exceeds the boundaries of the
37 * 16 bit length field due to the header being added outside of an
38 * IP or IPv6 frame that was already limited to 64K - 1.
40 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
41 partial = (__force __wsum)uh->len;
43 partial = (__force __wsum)htonl(skb->len);
44 partial = csum_sub(csum_unfold(uh->check), partial);
46 /* setup inner skb. */
47 skb->encapsulation = 0;
48 SKB_GSO_CB(skb)->encap_level = 0;
49 __skb_pull(skb, tnl_hlen);
50 skb_reset_mac_header(skb);
51 skb_set_network_header(skb, skb_inner_network_offset(skb));
52 skb->mac_len = skb_inner_network_offset(skb);
53 skb->protocol = new_protocol;
55 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
56 skb->encap_hdr_csum = need_csum;
58 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
59 skb->remcsum_offload = remcsum;
61 need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb));
62 /* Try to offload checksum if possible */
63 offload_csum = !!(need_csum &&
66 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
67 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
69 features &= skb->dev->hw_enc_features;
71 /* The only checksum offload we care about from here on out is the
72 * outer one so strip the existing checksum feature flags and
73 * instead set the flag based on our outer checksum offload value.
76 features &= ~NETIF_F_CSUM_MASK;
77 if (!need_csum || offload_csum)
78 features |= NETIF_F_HW_CSUM;
81 /* segment inner packet. */
82 segs = gso_inner_segment(skb, features);
83 if (IS_ERR_OR_NULL(segs)) {
84 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
89 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
91 outer_hlen = skb_tnl_header_len(skb);
92 udp_offset = outer_hlen - tnl_hlen;
98 skb->ip_summed = CHECKSUM_NONE;
100 /* Set up inner headers if we are offloading inner checksum */
101 if (skb->ip_summed == CHECKSUM_PARTIAL) {
102 skb_reset_inner_headers(skb);
103 skb->encapsulation = 1;
106 skb->mac_len = mac_len;
107 skb->protocol = protocol;
109 __skb_push(skb, outer_hlen);
110 skb_reset_mac_header(skb);
111 skb_set_network_header(skb, mac_len);
112 skb_set_transport_header(skb, udp_offset);
113 len = skb->len - udp_offset;
116 /* If we are only performing partial GSO the inner header
117 * will be using a length value equal to only one MSS sized
118 * segment instead of the entire frame.
120 if (gso_partial && skb_is_gso(skb)) {
121 uh->len = htons(skb_shinfo(skb)->gso_size +
122 SKB_GSO_CB(skb)->data_offset +
123 skb->head - (unsigned char *)uh);
125 uh->len = htons(len);
131 uh->check = ~csum_fold(csum_add(partial,
132 (__force __wsum)htonl(len)));
134 if (skb->encapsulation || !offload_csum) {
135 uh->check = gso_make_checksum(skb, ~uh->check);
137 uh->check = CSUM_MANGLED_0;
139 skb->ip_summed = CHECKSUM_PARTIAL;
140 skb->csum_start = skb_transport_header(skb) - skb->head;
141 skb->csum_offset = offsetof(struct udphdr, check);
143 } while ((skb = skb->next));
148 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
149 netdev_features_t features,
152 __be16 protocol = skb->protocol;
153 const struct net_offload **offloads;
154 const struct net_offload *ops;
155 struct sk_buff *segs = ERR_PTR(-EINVAL);
156 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
157 netdev_features_t features);
161 switch (skb->inner_protocol_type) {
162 case ENCAP_TYPE_ETHER:
163 protocol = skb->inner_protocol;
164 gso_inner_segment = skb_mac_gso_segment;
166 case ENCAP_TYPE_IPPROTO:
167 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
168 ops = rcu_dereference(offloads[skb->inner_ipproto]);
169 if (!ops || !ops->callbacks.gso_segment)
171 gso_inner_segment = ops->callbacks.gso_segment;
177 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
185 EXPORT_SYMBOL(skb_udp_tunnel_segment);
187 static void __udpv4_gso_segment_csum(struct sk_buff *seg,
188 __be32 *oldip, __be32 *newip,
189 __be16 *oldport, __be16 *newport)
194 if (*oldip == *newip && *oldport == *newport)
201 inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip,
203 inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport,
206 uh->check = CSUM_MANGLED_0;
210 csum_replace4(&iph->check, *oldip, *newip);
214 static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs)
217 struct udphdr *uh, *uh2;
218 struct iphdr *iph, *iph2;
224 if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) &&
225 (udp_hdr(seg)->source == udp_hdr(seg->next)->source) &&
226 (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) &&
227 (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr))
230 while ((seg = seg->next)) {
234 __udpv4_gso_segment_csum(seg,
235 &iph2->saddr, &iph->saddr,
236 &uh2->source, &uh->source);
237 __udpv4_gso_segment_csum(seg,
238 &iph2->daddr, &iph->daddr,
239 &uh2->dest, &uh->dest);
245 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
246 netdev_features_t features,
249 unsigned int mss = skb_shinfo(skb)->gso_size;
251 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
255 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
257 return is_ipv6 ? skb : __udpv4_gso_segment_list_csum(skb);
260 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
261 netdev_features_t features, bool is_ipv6)
263 struct sock *sk = gso_skb->sk;
264 unsigned int sum_truesize = 0;
265 struct sk_buff *segs, *seg;
272 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)
273 return __udp_gso_segment_list(gso_skb, features, is_ipv6);
275 mss = skb_shinfo(gso_skb)->gso_size;
276 if (gso_skb->len <= sizeof(*uh) + mss)
277 return ERR_PTR(-EINVAL);
279 skb_pull(gso_skb, sizeof(*uh));
281 /* clear destructor to avoid skb_segment assigning it to tail */
282 copy_dtor = gso_skb->destructor == sock_wfree;
284 gso_skb->destructor = NULL;
286 segs = skb_segment(gso_skb, features);
287 if (IS_ERR_OR_NULL(segs)) {
289 gso_skb->destructor = sock_wfree;
293 /* GSO partial and frag_list segmentation only requires splitting
294 * the frame into an MSS multiple and possibly a remainder, both
295 * cases return a GSO skb. So update the mss now.
297 if (skb_is_gso(segs))
298 mss *= skb_shinfo(segs)->gso_segs;
303 /* preserve TX timestamp flags and TS key for first segment */
304 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
305 skb_shinfo(seg)->tx_flags |=
306 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
308 /* compute checksum adjustment based on old length versus new */
309 newlen = htons(sizeof(*uh) + mss);
310 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
314 seg->destructor = sock_wfree;
316 sum_truesize += seg->truesize;
325 if (seg->ip_summed == CHECKSUM_PARTIAL)
326 gso_reset_checksum(seg, ~check);
328 uh->check = gso_make_checksum(seg, ~check) ? :
335 /* last packet can be partial gso_size, account for that in checksum */
336 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
338 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
343 if (seg->ip_summed == CHECKSUM_PARTIAL)
344 gso_reset_checksum(seg, ~check);
346 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
348 /* update refcount for the packet */
350 int delta = sum_truesize - gso_skb->truesize;
352 /* In some pathological cases, delta can be negative.
353 * We need to either use refcount_add() or refcount_sub_and_test()
355 if (likely(delta >= 0))
356 refcount_add(delta, &sk->sk_wmem_alloc);
358 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
362 EXPORT_SYMBOL_GPL(__udp_gso_segment);
364 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
365 netdev_features_t features)
367 struct sk_buff *segs = ERR_PTR(-EINVAL);
373 if (skb->encapsulation &&
374 (skb_shinfo(skb)->gso_type &
375 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
376 segs = skb_udp_tunnel_segment(skb, features, false);
380 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
383 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
386 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
387 return __udp_gso_segment(skb, features, false);
389 mss = skb_shinfo(skb)->gso_size;
390 if (unlikely(skb->len <= mss))
393 /* Do software UFO. Complete and fill in the UDP checksum as
394 * HW cannot do checksum of UDP packets sent as multiple
402 csum = skb_checksum(skb, 0, skb->len, 0);
403 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
405 uh->check = CSUM_MANGLED_0;
407 skb->ip_summed = CHECKSUM_UNNECESSARY;
409 /* If there is no outer header we can fake a checksum offload
410 * due to the fact that we have already done the checksum in
411 * software prior to segmenting the frame.
413 if (!skb->encap_hdr_csum)
414 features |= NETIF_F_HW_CSUM;
416 /* Fragment the skb. IP headers of the fragments are updated in
419 segs = skb_segment(skb, features);
424 #define UDP_GRO_CNT_MAX 64
425 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
428 struct udphdr *uh = udp_gro_udphdr(skb);
429 struct sk_buff *pp = NULL;
435 /* requires non zero csum, for symmetry with GSO */
437 NAPI_GRO_CB(skb)->flush = 1;
441 /* Do not deal with padded or malicious packets, sorry ! */
442 ulen = ntohs(uh->len);
443 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
444 NAPI_GRO_CB(skb)->flush = 1;
447 /* pull encapsulating udp header */
448 skb_gro_pull(skb, sizeof(struct udphdr));
450 list_for_each_entry(p, head, list) {
451 if (!NAPI_GRO_CB(p)->same_flow)
456 /* Match ports only, as csum is always non zero */
457 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
458 NAPI_GRO_CB(p)->same_flow = 0;
462 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
463 NAPI_GRO_CB(skb)->flush = 1;
467 /* Terminate the flow on len mismatch or if it grow "too much".
468 * Under small packet flood GRO count could elsewhere grow a lot
469 * leading to excessive truesize values.
470 * On len mismatch merge the first packet shorter than gso_size,
471 * otherwise complete the GRO packet.
473 if (ulen > ntohs(uh2->len)) {
476 if (NAPI_GRO_CB(skb)->is_flist) {
477 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
478 NAPI_GRO_CB(skb)->flush = 1;
481 if ((skb->ip_summed != p->ip_summed) ||
482 (skb->csum_level != p->csum_level)) {
483 NAPI_GRO_CB(skb)->flush = 1;
486 ret = skb_gro_receive_list(p, skb);
488 skb_gro_postpull_rcsum(skb, uh,
489 sizeof(struct udphdr));
491 ret = skb_gro_receive(p, skb);
495 if (ret || ulen != ntohs(uh2->len) ||
496 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
502 /* mismatch, but we never need to flush */
506 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
507 struct udphdr *uh, struct sock *sk)
509 struct sk_buff *pp = NULL;
512 unsigned int off = skb_gro_offset(skb);
515 NAPI_GRO_CB(skb)->is_flist = 0;
516 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
517 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled: 1;
519 if ((sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist) {
520 pp = call_gro_receive(udp_gro_receive_segment, head, skb);
524 if (!sk || NAPI_GRO_CB(skb)->encap_mark ||
525 (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
526 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
527 !NAPI_GRO_CB(skb)->csum_valid) ||
528 !udp_sk(sk)->gro_receive)
531 /* mark that this skb passed once through the tunnel gro layer */
532 NAPI_GRO_CB(skb)->encap_mark = 1;
536 list_for_each_entry(p, head, list) {
537 if (!NAPI_GRO_CB(p)->same_flow)
540 uh2 = (struct udphdr *)(p->data + off);
542 /* Match ports and either checksums are either both zero
545 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
546 (!uh->check ^ !uh2->check)) {
547 NAPI_GRO_CB(p)->same_flow = 0;
552 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
553 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
554 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
557 skb_gro_flush_final(skb, pp, flush);
560 EXPORT_SYMBOL(udp_gro_receive);
562 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
565 const struct iphdr *iph = skb_gro_network_header(skb);
567 return __udp4_lib_lookup(dev_net(skb->dev), iph->saddr, sport,
568 iph->daddr, dport, inet_iif(skb),
569 inet_sdif(skb), &udp_table, NULL);
572 INDIRECT_CALLABLE_SCOPE
573 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
575 struct udphdr *uh = udp_gro_udphdr(skb);
576 struct sock *sk = NULL;
582 /* Don't bother verifying checksum if we're going to flush anyway. */
583 if (NAPI_GRO_CB(skb)->flush)
586 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
587 inet_gro_compute_pseudo))
590 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
591 inet_gro_compute_pseudo);
593 NAPI_GRO_CB(skb)->is_ipv6 = 0;
596 if (static_branch_unlikely(&udp_encap_needed_key))
597 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
599 pp = udp_gro_receive(head, skb, uh, sk);
604 NAPI_GRO_CB(skb)->flush = 1;
608 static int udp_gro_complete_segment(struct sk_buff *skb)
610 struct udphdr *uh = udp_hdr(skb);
612 skb->csum_start = (unsigned char *)uh - skb->head;
613 skb->csum_offset = offsetof(struct udphdr, check);
614 skb->ip_summed = CHECKSUM_PARTIAL;
616 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
617 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
619 if (skb->encapsulation)
620 skb->inner_transport_header = skb->transport_header;
625 int udp_gro_complete(struct sk_buff *skb, int nhoff,
628 __be16 newlen = htons(skb->len - nhoff);
629 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
636 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
637 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
638 if (sk && udp_sk(sk)->gro_complete) {
639 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
640 : SKB_GSO_UDP_TUNNEL;
642 /* Set encapsulation before calling into inner gro_complete()
643 * functions to make them set up the inner offsets.
645 skb->encapsulation = 1;
646 err = udp_sk(sk)->gro_complete(sk, skb,
647 nhoff + sizeof(struct udphdr));
649 err = udp_gro_complete_segment(skb);
653 if (skb->remcsum_offload)
654 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
658 EXPORT_SYMBOL(udp_gro_complete);
660 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
662 const struct iphdr *iph = ip_hdr(skb);
663 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
665 if (NAPI_GRO_CB(skb)->is_flist) {
666 uh->len = htons(skb->len - nhoff);
668 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
669 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
671 if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
672 if (skb->csum_level < SKB_MAX_CSUM_LEVEL)
675 skb->ip_summed = CHECKSUM_UNNECESSARY;
683 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
686 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
689 static const struct net_offload udpv4_offload = {
691 .gso_segment = udp4_ufo_fragment,
692 .gro_receive = udp4_gro_receive,
693 .gro_complete = udp4_gro_complete,
697 int __init udpv4_offload_init(void)
699 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);