1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * IPV4 GSO/GRO offload support
4 * Linux INET implementation
9 #include <linux/skbuff.h>
13 #include <net/protocol.h>
14 #include <net/inet_common.h>
16 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
17 netdev_features_t features,
18 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
19 netdev_features_t features),
20 __be16 new_protocol, bool is_ipv6)
22 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
23 bool remcsum, need_csum, offload_csum, gso_partial;
24 struct sk_buff *segs = ERR_PTR(-EINVAL);
25 struct udphdr *uh = udp_hdr(skb);
26 u16 mac_offset = skb->mac_header;
27 __be16 protocol = skb->protocol;
28 u16 mac_len = skb->mac_len;
29 int udp_offset, outer_hlen;
33 if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
36 /* Adjust partial header checksum to negate old length.
37 * We cannot rely on the value contained in uh->len as it is
38 * possible that the actual value exceeds the boundaries of the
39 * 16 bit length field due to the header being added outside of an
40 * IP or IPv6 frame that was already limited to 64K - 1.
42 if (skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)
43 partial = (__force __wsum)uh->len;
45 partial = (__force __wsum)htonl(skb->len);
46 partial = csum_sub(csum_unfold(uh->check), partial);
48 /* setup inner skb. */
49 skb->encapsulation = 0;
50 SKB_GSO_CB(skb)->encap_level = 0;
51 __skb_pull(skb, tnl_hlen);
52 skb_reset_mac_header(skb);
53 skb_set_network_header(skb, skb_inner_network_offset(skb));
54 skb_set_transport_header(skb, skb_inner_transport_offset(skb));
55 skb->mac_len = skb_inner_network_offset(skb);
56 skb->protocol = new_protocol;
58 need_csum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM);
59 skb->encap_hdr_csum = need_csum;
61 remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
62 skb->remcsum_offload = remcsum;
64 need_ipsec = skb_dst(skb) && dst_xfrm(skb_dst(skb));
65 /* Try to offload checksum if possible */
66 offload_csum = !!(need_csum &&
69 (is_ipv6 ? (NETIF_F_HW_CSUM | NETIF_F_IPV6_CSUM) :
70 (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM))));
72 features &= skb->dev->hw_enc_features;
74 features &= ~NETIF_F_SCTP_CRC;
76 /* The only checksum offload we care about from here on out is the
77 * outer one so strip the existing checksum feature flags and
78 * instead set the flag based on our outer checksum offload value.
81 features &= ~NETIF_F_CSUM_MASK;
82 if (!need_csum || offload_csum)
83 features |= NETIF_F_HW_CSUM;
86 /* segment inner packet. */
87 segs = gso_inner_segment(skb, features);
88 if (IS_ERR_OR_NULL(segs)) {
89 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
94 gso_partial = !!(skb_shinfo(segs)->gso_type & SKB_GSO_PARTIAL);
96 outer_hlen = skb_tnl_header_len(skb);
97 udp_offset = outer_hlen - tnl_hlen;
103 skb->ip_summed = CHECKSUM_NONE;
105 /* Set up inner headers if we are offloading inner checksum */
106 if (skb->ip_summed == CHECKSUM_PARTIAL) {
107 skb_reset_inner_headers(skb);
108 skb->encapsulation = 1;
111 skb->mac_len = mac_len;
112 skb->protocol = protocol;
114 __skb_push(skb, outer_hlen);
115 skb_reset_mac_header(skb);
116 skb_set_network_header(skb, mac_len);
117 skb_set_transport_header(skb, udp_offset);
118 len = skb->len - udp_offset;
121 /* If we are only performing partial GSO the inner header
122 * will be using a length value equal to only one MSS sized
123 * segment instead of the entire frame.
125 if (gso_partial && skb_is_gso(skb)) {
126 uh->len = htons(skb_shinfo(skb)->gso_size +
127 SKB_GSO_CB(skb)->data_offset +
128 skb->head - (unsigned char *)uh);
130 uh->len = htons(len);
136 uh->check = ~csum_fold(csum_add(partial,
137 (__force __wsum)htonl(len)));
139 if (skb->encapsulation || !offload_csum) {
140 uh->check = gso_make_checksum(skb, ~uh->check);
142 uh->check = CSUM_MANGLED_0;
144 skb->ip_summed = CHECKSUM_PARTIAL;
145 skb->csum_start = skb_transport_header(skb) - skb->head;
146 skb->csum_offset = offsetof(struct udphdr, check);
148 } while ((skb = skb->next));
153 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
154 netdev_features_t features,
157 const struct net_offload __rcu **offloads;
158 __be16 protocol = skb->protocol;
159 const struct net_offload *ops;
160 struct sk_buff *segs = ERR_PTR(-EINVAL);
161 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
162 netdev_features_t features);
166 switch (skb->inner_protocol_type) {
167 case ENCAP_TYPE_ETHER:
168 protocol = skb->inner_protocol;
169 gso_inner_segment = skb_mac_gso_segment;
171 case ENCAP_TYPE_IPPROTO:
172 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
173 ops = rcu_dereference(offloads[skb->inner_ipproto]);
174 if (!ops || !ops->callbacks.gso_segment)
176 gso_inner_segment = ops->callbacks.gso_segment;
182 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
190 EXPORT_SYMBOL(skb_udp_tunnel_segment);
192 static void __udpv4_gso_segment_csum(struct sk_buff *seg,
193 __be32 *oldip, __be32 *newip,
194 __be16 *oldport, __be16 *newport)
199 if (*oldip == *newip && *oldport == *newport)
206 inet_proto_csum_replace4(&uh->check, seg, *oldip, *newip,
208 inet_proto_csum_replace2(&uh->check, seg, *oldport, *newport,
211 uh->check = CSUM_MANGLED_0;
215 csum_replace4(&iph->check, *oldip, *newip);
219 static struct sk_buff *__udpv4_gso_segment_list_csum(struct sk_buff *segs)
222 struct udphdr *uh, *uh2;
223 struct iphdr *iph, *iph2;
229 if ((udp_hdr(seg)->dest == udp_hdr(seg->next)->dest) &&
230 (udp_hdr(seg)->source == udp_hdr(seg->next)->source) &&
231 (ip_hdr(seg)->daddr == ip_hdr(seg->next)->daddr) &&
232 (ip_hdr(seg)->saddr == ip_hdr(seg->next)->saddr))
235 while ((seg = seg->next)) {
239 __udpv4_gso_segment_csum(seg,
240 &iph2->saddr, &iph->saddr,
241 &uh2->source, &uh->source);
242 __udpv4_gso_segment_csum(seg,
243 &iph2->daddr, &iph->daddr,
244 &uh2->dest, &uh->dest);
250 static struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
251 netdev_features_t features,
254 unsigned int mss = skb_shinfo(skb)->gso_size;
256 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
260 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
262 return is_ipv6 ? skb : __udpv4_gso_segment_list_csum(skb);
265 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
266 netdev_features_t features, bool is_ipv6)
268 struct sock *sk = gso_skb->sk;
269 unsigned int sum_truesize = 0;
270 struct sk_buff *segs, *seg;
277 mss = skb_shinfo(gso_skb)->gso_size;
278 if (gso_skb->len <= sizeof(*uh) + mss)
279 return ERR_PTR(-EINVAL);
281 if (skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) {
282 /* Packet is from an untrusted source, reset gso_segs. */
283 skb_shinfo(gso_skb)->gso_segs = DIV_ROUND_UP(gso_skb->len - sizeof(*uh),
288 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)
289 return __udp_gso_segment_list(gso_skb, features, is_ipv6);
291 skb_pull(gso_skb, sizeof(*uh));
293 /* clear destructor to avoid skb_segment assigning it to tail */
294 copy_dtor = gso_skb->destructor == sock_wfree;
296 gso_skb->destructor = NULL;
298 segs = skb_segment(gso_skb, features);
299 if (IS_ERR_OR_NULL(segs)) {
301 gso_skb->destructor = sock_wfree;
305 /* GSO partial and frag_list segmentation only requires splitting
306 * the frame into an MSS multiple and possibly a remainder, both
307 * cases return a GSO skb. So update the mss now.
309 if (skb_is_gso(segs))
310 mss *= skb_shinfo(segs)->gso_segs;
315 /* preserve TX timestamp flags and TS key for first segment */
316 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
317 skb_shinfo(seg)->tx_flags |=
318 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
320 /* compute checksum adjustment based on old length versus new */
321 newlen = htons(sizeof(*uh) + mss);
322 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
326 seg->destructor = sock_wfree;
328 sum_truesize += seg->truesize;
337 if (seg->ip_summed == CHECKSUM_PARTIAL)
338 gso_reset_checksum(seg, ~check);
340 uh->check = gso_make_checksum(seg, ~check) ? :
347 /* last packet can be partial gso_size, account for that in checksum */
348 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
350 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
355 if (seg->ip_summed == CHECKSUM_PARTIAL)
356 gso_reset_checksum(seg, ~check);
358 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
360 /* update refcount for the packet */
362 int delta = sum_truesize - gso_skb->truesize;
364 /* In some pathological cases, delta can be negative.
365 * We need to either use refcount_add() or refcount_sub_and_test()
367 if (likely(delta >= 0))
368 refcount_add(delta, &sk->sk_wmem_alloc);
370 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
374 EXPORT_SYMBOL_GPL(__udp_gso_segment);
376 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
377 netdev_features_t features)
379 struct sk_buff *segs = ERR_PTR(-EINVAL);
385 if (skb->encapsulation &&
386 (skb_shinfo(skb)->gso_type &
387 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
388 segs = skb_udp_tunnel_segment(skb, features, false);
392 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
395 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
398 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
399 return __udp_gso_segment(skb, features, false);
401 mss = skb_shinfo(skb)->gso_size;
402 if (unlikely(skb->len <= mss))
405 /* Do software UFO. Complete and fill in the UDP checksum as
406 * HW cannot do checksum of UDP packets sent as multiple
414 csum = skb_checksum(skb, 0, skb->len, 0);
415 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
417 uh->check = CSUM_MANGLED_0;
419 skb->ip_summed = CHECKSUM_UNNECESSARY;
421 /* If there is no outer header we can fake a checksum offload
422 * due to the fact that we have already done the checksum in
423 * software prior to segmenting the frame.
425 if (!skb->encap_hdr_csum)
426 features |= NETIF_F_HW_CSUM;
428 /* Fragment the skb. IP headers of the fragments are updated in
431 segs = skb_segment(skb, features);
436 static int skb_gro_receive_list(struct sk_buff *p, struct sk_buff *skb)
438 if (unlikely(p->len + skb->len >= 65536))
441 if (NAPI_GRO_CB(p)->last == p)
442 skb_shinfo(p)->frag_list = skb;
444 NAPI_GRO_CB(p)->last->next = skb;
446 skb_pull(skb, skb_gro_offset(skb));
448 NAPI_GRO_CB(p)->last = skb;
449 NAPI_GRO_CB(p)->count++;
450 p->data_len += skb->len;
452 /* sk owenrship - if any - completely transferred to the aggregated packet */
453 skb->destructor = NULL;
454 p->truesize += skb->truesize;
457 NAPI_GRO_CB(skb)->same_flow = 1;
463 #define UDP_GRO_CNT_MAX 64
464 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
467 struct udphdr *uh = udp_gro_udphdr(skb);
468 struct sk_buff *pp = NULL;
474 /* requires non zero csum, for symmetry with GSO */
476 NAPI_GRO_CB(skb)->flush = 1;
480 /* Do not deal with padded or malicious packets, sorry ! */
481 ulen = ntohs(uh->len);
482 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
483 NAPI_GRO_CB(skb)->flush = 1;
486 /* pull encapsulating udp header */
487 skb_gro_pull(skb, sizeof(struct udphdr));
489 list_for_each_entry(p, head, list) {
490 if (!NAPI_GRO_CB(p)->same_flow)
495 /* Match ports only, as csum is always non zero */
496 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
497 NAPI_GRO_CB(p)->same_flow = 0;
501 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
502 NAPI_GRO_CB(skb)->flush = 1;
506 /* Terminate the flow on len mismatch or if it grow "too much".
507 * Under small packet flood GRO count could elsewhere grow a lot
508 * leading to excessive truesize values.
509 * On len mismatch merge the first packet shorter than gso_size,
510 * otherwise complete the GRO packet.
512 if (ulen > ntohs(uh2->len)) {
515 if (NAPI_GRO_CB(skb)->is_flist) {
516 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
517 NAPI_GRO_CB(skb)->flush = 1;
520 if ((skb->ip_summed != p->ip_summed) ||
521 (skb->csum_level != p->csum_level)) {
522 NAPI_GRO_CB(skb)->flush = 1;
525 ret = skb_gro_receive_list(p, skb);
527 skb_gro_postpull_rcsum(skb, uh,
528 sizeof(struct udphdr));
530 ret = skb_gro_receive(p, skb);
534 if (ret || ulen != ntohs(uh2->len) ||
535 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
541 /* mismatch, but we never need to flush */
545 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
546 struct udphdr *uh, struct sock *sk)
548 struct sk_buff *pp = NULL;
551 unsigned int off = skb_gro_offset(skb);
554 /* we can do L4 aggregation only if the packet can't land in a tunnel
555 * otherwise we could corrupt the inner stream
557 NAPI_GRO_CB(skb)->is_flist = 0;
558 if (!sk || !udp_sk(sk)->gro_receive) {
559 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
560 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled : 1;
562 if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) ||
563 (sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist)
564 return call_gro_receive(udp_gro_receive_segment, head, skb);
566 /* no GRO, be sure flush the current packet */
570 if (NAPI_GRO_CB(skb)->encap_mark ||
571 (uh->check && skb->ip_summed != CHECKSUM_PARTIAL &&
572 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
573 !NAPI_GRO_CB(skb)->csum_valid))
576 /* mark that this skb passed once through the tunnel gro layer */
577 NAPI_GRO_CB(skb)->encap_mark = 1;
581 list_for_each_entry(p, head, list) {
582 if (!NAPI_GRO_CB(p)->same_flow)
585 uh2 = (struct udphdr *)(p->data + off);
587 /* Match ports and either checksums are either both zero
590 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
591 (!uh->check ^ !uh2->check)) {
592 NAPI_GRO_CB(p)->same_flow = 0;
597 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
598 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
599 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
602 skb_gro_flush_final(skb, pp, flush);
605 EXPORT_SYMBOL(udp_gro_receive);
607 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
610 const struct iphdr *iph = skb_gro_network_header(skb);
611 struct net *net = dev_net(skb->dev);
614 inet_get_iif_sdif(skb, &iif, &sdif);
616 return __udp4_lib_lookup(net, iph->saddr, sport,
617 iph->daddr, dport, iif,
618 sdif, net->ipv4.udp_table, NULL);
621 INDIRECT_CALLABLE_SCOPE
622 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
624 struct udphdr *uh = udp_gro_udphdr(skb);
625 struct sock *sk = NULL;
631 /* Don't bother verifying checksum if we're going to flush anyway. */
632 if (NAPI_GRO_CB(skb)->flush)
635 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
636 inet_gro_compute_pseudo))
639 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
640 inet_gro_compute_pseudo);
642 NAPI_GRO_CB(skb)->is_ipv6 = 0;
644 if (static_branch_unlikely(&udp_encap_needed_key))
645 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
647 pp = udp_gro_receive(head, skb, uh, sk);
651 NAPI_GRO_CB(skb)->flush = 1;
655 static int udp_gro_complete_segment(struct sk_buff *skb)
657 struct udphdr *uh = udp_hdr(skb);
659 skb->csum_start = (unsigned char *)uh - skb->head;
660 skb->csum_offset = offsetof(struct udphdr, check);
661 skb->ip_summed = CHECKSUM_PARTIAL;
663 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
664 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
666 if (skb->encapsulation)
667 skb->inner_transport_header = skb->transport_header;
672 int udp_gro_complete(struct sk_buff *skb, int nhoff,
675 __be16 newlen = htons(skb->len - nhoff);
676 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
682 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
683 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
684 if (sk && udp_sk(sk)->gro_complete) {
685 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
686 : SKB_GSO_UDP_TUNNEL;
688 /* clear the encap mark, so that inner frag_list gro_complete
691 NAPI_GRO_CB(skb)->encap_mark = 0;
693 /* Set encapsulation before calling into inner gro_complete()
694 * functions to make them set up the inner offsets.
696 skb->encapsulation = 1;
697 err = udp_sk(sk)->gro_complete(sk, skb,
698 nhoff + sizeof(struct udphdr));
700 err = udp_gro_complete_segment(skb);
703 if (skb->remcsum_offload)
704 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
708 EXPORT_SYMBOL(udp_gro_complete);
710 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
712 const struct iphdr *iph = ip_hdr(skb);
713 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
715 /* do fraglist only if there is no outer UDP encap (or we already processed it) */
716 if (NAPI_GRO_CB(skb)->is_flist && !NAPI_GRO_CB(skb)->encap_mark) {
717 uh->len = htons(skb->len - nhoff);
719 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
720 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
722 if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
723 if (skb->csum_level < SKB_MAX_CSUM_LEVEL)
726 skb->ip_summed = CHECKSUM_UNNECESSARY;
734 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
737 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
740 static const struct net_offload udpv4_offload = {
742 .gso_segment = udp4_ufo_fragment,
743 .gro_receive = udp4_gro_receive,
744 .gro_complete = udp4_gro_complete,
748 int __init udpv4_offload_init(void)
750 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);