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_set_transport_header(skb, skb_inner_transport_offset(skb));
53 skb->mac_len = skb_inner_network_offset(skb);
54 skb->protocol = new_protocol;
56 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
57 skb->encap_hdr_csum = need_csum;
59 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
60 skb->remcsum_offload = remcsum;
62 need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb));
63 /* Try to offload checksum if possible */
64 offload_csum = !!(need_csum &&
67 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
68 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
70 features &= skb->dev->hw_enc_features;
72 features &= ~NETIF_F_SCTP_CRC;
74 /* The only checksum offload we care about from here on out is the
75 * outer one so strip the existing checksum feature flags and
76 * instead set the flag based on our outer checksum offload value.
79 features &= ~NETIF_F_CSUM_MASK;
80 if (!need_csum || offload_csum)
81 features |= NETIF_F_HW_CSUM;
84 /* segment inner packet. */
85 segs = gso_inner_segment(skb, features);
86 if (IS_ERR_OR_NULL(segs)) {
87 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
92 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
94 outer_hlen = skb_tnl_header_len(skb);
95 udp_offset = outer_hlen - tnl_hlen;
101 skb->ip_summed = CHECKSUM_NONE;
103 /* Set up inner headers if we are offloading inner checksum */
104 if (skb->ip_summed == CHECKSUM_PARTIAL) {
105 skb_reset_inner_headers(skb);
106 skb->encapsulation = 1;
109 skb->mac_len = mac_len;
110 skb->protocol = protocol;
112 __skb_push(skb, outer_hlen);
113 skb_reset_mac_header(skb);
114 skb_set_network_header(skb, mac_len);
115 skb_set_transport_header(skb, udp_offset);
116 len = skb->len - udp_offset;
119 /* If we are only performing partial GSO the inner header
120 * will be using a length value equal to only one MSS sized
121 * segment instead of the entire frame.
123 if (gso_partial && skb_is_gso(skb)) {
124 uh->len = htons(skb_shinfo(skb)->gso_size +
125 SKB_GSO_CB(skb)->data_offset +
126 skb->head - (unsigned char *)uh);
128 uh->len = htons(len);
134 uh->check = ~csum_fold(csum_add(partial,
135 (__force __wsum)htonl(len)));
137 if (skb->encapsulation || !offload_csum) {
138 uh->check = gso_make_checksum(skb, ~uh->check);
140 uh->check = CSUM_MANGLED_0;
142 skb->ip_summed = CHECKSUM_PARTIAL;
143 skb->csum_start = skb_transport_header(skb) - skb->head;
144 skb->csum_offset = offsetof(struct udphdr, check);
146 } while ((skb = skb->next));
151 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
152 netdev_features_t features,
155 const struct net_offload __rcu **offloads;
156 __be16 protocol = skb->protocol;
157 const struct net_offload *ops;
158 struct sk_buff *segs = ERR_PTR(-EINVAL);
159 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
160 netdev_features_t features);
164 switch (skb->inner_protocol_type) {
165 case ENCAP_TYPE_ETHER:
166 protocol = skb->inner_protocol;
167 gso_inner_segment = skb_mac_gso_segment;
169 case ENCAP_TYPE_IPPROTO:
170 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
171 ops = rcu_dereference(offloads[skb->inner_ipproto]);
172 if (!ops || !ops->callbacks.gso_segment)
174 gso_inner_segment = ops->callbacks.gso_segment;
180 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
188 EXPORT_SYMBOL(skb_udp_tunnel_segment);
190 static void __udpv4_gso_segment_csum(struct sk_buff *seg,
191 __be32 *oldip, __be32 *newip,
192 __be16 *oldport, __be16 *newport)
197 if (*oldip == *newip && *oldport == *newport)
204 inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip,
206 inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport,
209 uh->check = CSUM_MANGLED_0;
213 csum_replace4(&iph->check, *oldip, *newip);
217 static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs)
220 struct udphdr *uh, *uh2;
221 struct iphdr *iph, *iph2;
227 if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) &&
228 (udp_hdr(seg)->source == udp_hdr(seg->next)->source) &&
229 (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) &&
230 (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr))
233 while ((seg = seg->next)) {
237 __udpv4_gso_segment_csum(seg,
238 &iph2->saddr, &iph->saddr,
239 &uh2->source, &uh->source);
240 __udpv4_gso_segment_csum(seg,
241 &iph2->daddr, &iph->daddr,
242 &uh2->dest, &uh->dest);
248 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
249 netdev_features_t features,
252 unsigned int mss = skb_shinfo(skb)->gso_size;
254 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
258 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
260 return is_ipv6 ? skb : __udpv4_gso_segment_list_csum(skb);
263 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
264 netdev_features_t features, bool is_ipv6)
266 struct sock *sk = gso_skb->sk;
267 unsigned int sum_truesize = 0;
268 struct sk_buff *segs, *seg;
275 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)
276 return __udp_gso_segment_list(gso_skb, features, is_ipv6);
278 mss = skb_shinfo(gso_skb)->gso_size;
279 if (gso_skb->len <= sizeof(*uh) + mss)
280 return ERR_PTR(-EINVAL);
282 skb_pull(gso_skb, sizeof(*uh));
284 /* clear destructor to avoid skb_segment assigning it to tail */
285 copy_dtor = gso_skb->destructor == sock_wfree;
287 gso_skb->destructor = NULL;
289 segs = skb_segment(gso_skb, features);
290 if (IS_ERR_OR_NULL(segs)) {
292 gso_skb->destructor = sock_wfree;
296 /* GSO partial and frag_list segmentation only requires splitting
297 * the frame into an MSS multiple and possibly a remainder, both
298 * cases return a GSO skb. So update the mss now.
300 if (skb_is_gso(segs))
301 mss *= skb_shinfo(segs)->gso_segs;
306 /* preserve TX timestamp flags and TS key for first segment */
307 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
308 skb_shinfo(seg)->tx_flags |=
309 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
311 /* compute checksum adjustment based on old length versus new */
312 newlen = htons(sizeof(*uh) + mss);
313 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
317 seg->destructor = sock_wfree;
319 sum_truesize += seg->truesize;
328 if (seg->ip_summed == CHECKSUM_PARTIAL)
329 gso_reset_checksum(seg, ~check);
331 uh->check = gso_make_checksum(seg, ~check) ? :
338 /* last packet can be partial gso_size, account for that in checksum */
339 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
341 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
346 if (seg->ip_summed == CHECKSUM_PARTIAL)
347 gso_reset_checksum(seg, ~check);
349 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
351 /* update refcount for the packet */
353 int delta = sum_truesize - gso_skb->truesize;
355 /* In some pathological cases, delta can be negative.
356 * We need to either use refcount_add() or refcount_sub_and_test()
358 if (likely(delta >= 0))
359 refcount_add(delta, &sk->sk_wmem_alloc);
361 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
365 EXPORT_SYMBOL_GPL(__udp_gso_segment);
367 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
368 netdev_features_t features)
370 struct sk_buff *segs = ERR_PTR(-EINVAL);
376 if (skb->encapsulation &&
377 (skb_shinfo(skb)->gso_type &
378 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
379 segs = skb_udp_tunnel_segment(skb, features, false);
383 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
386 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
389 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
390 return __udp_gso_segment(skb, features, false);
392 mss = skb_shinfo(skb)->gso_size;
393 if (unlikely(skb->len <= mss))
396 /* Do software UFO. Complete and fill in the UDP checksum as
397 * HW cannot do checksum of UDP packets sent as multiple
405 csum = skb_checksum(skb, 0, skb->len, 0);
406 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
408 uh->check = CSUM_MANGLED_0;
410 skb->ip_summed = CHECKSUM_UNNECESSARY;
412 /* If there is no outer header we can fake a checksum offload
413 * due to the fact that we have already done the checksum in
414 * software prior to segmenting the frame.
416 if (!skb->encap_hdr_csum)
417 features |= NETIF_F_HW_CSUM;
419 /* Fragment the skb. IP headers of the fragments are updated in
422 segs = skb_segment(skb, features);
427 #define UDP_GRO_CNT_MAX 64
428 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
431 struct udphdr *uh = udp_gro_udphdr(skb);
432 struct sk_buff *pp = NULL;
438 /* requires non zero csum, for symmetry with GSO */
440 NAPI_GRO_CB(skb)->flush = 1;
444 /* Do not deal with padded or malicious packets, sorry ! */
445 ulen = ntohs(uh->len);
446 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
447 NAPI_GRO_CB(skb)->flush = 1;
450 /* pull encapsulating udp header */
451 skb_gro_pull(skb, sizeof(struct udphdr));
453 list_for_each_entry(p, head, list) {
454 if (!NAPI_GRO_CB(p)->same_flow)
459 /* Match ports only, as csum is always non zero */
460 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
461 NAPI_GRO_CB(p)->same_flow = 0;
465 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
466 NAPI_GRO_CB(skb)->flush = 1;
470 /* Terminate the flow on len mismatch or if it grow "too much".
471 * Under small packet flood GRO count could elsewhere grow a lot
472 * leading to excessive truesize values.
473 * On len mismatch merge the first packet shorter than gso_size,
474 * otherwise complete the GRO packet.
476 if (ulen > ntohs(uh2->len)) {
479 if (NAPI_GRO_CB(skb)->is_flist) {
480 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
481 NAPI_GRO_CB(skb)->flush = 1;
484 if ((skb->ip_summed != p->ip_summed) ||
485 (skb->csum_level != p->csum_level)) {
486 NAPI_GRO_CB(skb)->flush = 1;
489 ret = skb_gro_receive_list(p, skb);
491 skb_gro_postpull_rcsum(skb, uh,
492 sizeof(struct udphdr));
494 ret = skb_gro_receive(p, skb);
498 if (ret || ulen != ntohs(uh2->len) ||
499 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
505 /* mismatch, but we never need to flush */
509 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
510 struct udphdr *uh, struct sock *sk)
512 struct sk_buff *pp = NULL;
515 unsigned int off = skb_gro_offset(skb);
518 /* we can do L4 aggregation only if the packet can't land in a tunnel
519 * otherwise we could corrupt the inner stream
521 NAPI_GRO_CB(skb)->is_flist = 0;
522 if (!sk || !udp_sk(sk)->gro_receive) {
523 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
524 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled : 1;
526 if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) ||
527 (sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist)
528 return call_gro_receive(udp_gro_receive_segment, head, skb);
530 /* no GRO, be sure flush the current packet */
534 if (NAPI_GRO_CB(skb)->encap_mark ||
535 (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
536 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
537 !NAPI_GRO_CB(skb)->csum_valid))
540 /* mark that this skb passed once through the tunnel gro layer */
541 NAPI_GRO_CB(skb)->encap_mark = 1;
545 list_for_each_entry(p, head, list) {
546 if (!NAPI_GRO_CB(p)->same_flow)
549 uh2 = (struct udphdr *)(p->data + off);
551 /* Match ports and either checksums are either both zero
554 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
555 (!uh->check ^ !uh2->check)) {
556 NAPI_GRO_CB(p)->same_flow = 0;
561 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
562 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
563 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
566 skb_gro_flush_final(skb, pp, flush);
569 EXPORT_SYMBOL(udp_gro_receive);
571 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
574 const struct iphdr *iph = skb_gro_network_header(skb);
576 return __udp4_lib_lookup(dev_net(skb->dev), iph->saddr, sport,
577 iph->daddr, dport, inet_iif(skb),
578 inet_sdif(skb), &udp_table, NULL);
581 INDIRECT_CALLABLE_SCOPE
582 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
584 struct udphdr *uh = udp_gro_udphdr(skb);
585 struct sock *sk = NULL;
591 /* Don't bother verifying checksum if we're going to flush anyway. */
592 if (NAPI_GRO_CB(skb)->flush)
595 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
596 inet_gro_compute_pseudo))
599 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
600 inet_gro_compute_pseudo);
602 NAPI_GRO_CB(skb)->is_ipv6 = 0;
605 if (static_branch_unlikely(&udp_encap_needed_key))
606 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
608 pp = udp_gro_receive(head, skb, uh, sk);
613 NAPI_GRO_CB(skb)->flush = 1;
617 static int udp_gro_complete_segment(struct sk_buff *skb)
619 struct udphdr *uh = udp_hdr(skb);
621 skb->csum_start = (unsigned char *)uh - skb->head;
622 skb->csum_offset = offsetof(struct udphdr, check);
623 skb->ip_summed = CHECKSUM_PARTIAL;
625 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
626 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
628 if (skb->encapsulation)
629 skb->inner_transport_header = skb->transport_header;
634 int udp_gro_complete(struct sk_buff *skb, int nhoff,
637 __be16 newlen = htons(skb->len - nhoff);
638 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
645 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
646 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
647 if (sk && udp_sk(sk)->gro_complete) {
648 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
649 : SKB_GSO_UDP_TUNNEL;
651 /* clear the encap mark, so that inner frag_list gro_complete
654 NAPI_GRO_CB(skb)->encap_mark = 0;
656 /* Set encapsulation before calling into inner gro_complete()
657 * functions to make them set up the inner offsets.
659 skb->encapsulation = 1;
660 err = udp_sk(sk)->gro_complete(sk, skb,
661 nhoff + sizeof(struct udphdr));
663 err = udp_gro_complete_segment(skb);
667 if (skb->remcsum_offload)
668 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
672 EXPORT_SYMBOL(udp_gro_complete);
674 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
676 const struct iphdr *iph = ip_hdr(skb);
677 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
679 /* do fraglist only if there is no outer UDP encap (or we already processed it) */
680 if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) {
681 uh->len = htons(skb->len - nhoff);
683 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
684 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
686 if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
687 if (skb->csum_level < SKB_MAX_CSUM_LEVEL)
690 skb->ip_summed = CHECKSUM_UNNECESSARY;
698 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
701 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
704 static const struct net_offload udpv4_offload = {
706 .gso_segment = udp4_ufo_fragment,
707 .gro_receive = udp4_gro_receive,
708 .gro_complete = udp4_gro_complete,
712 int __init udpv4_offload_init(void)
714 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);