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;
71 /* CRC checksum can't be handled by HW when it's a UDP tunneling packet. */
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 __be16 protocol = skb->protocol;
156 const struct net_offload **offloads;
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 struct sk_buff *__udp_gso_segment_list(struct sk_buff *skb,
191 netdev_features_t features)
193 unsigned int mss = skb_shinfo(skb)->gso_size;
195 skb = skb_segment_list(skb, features, skb_mac_header_len(skb));
199 udp_hdr(skb)->len = htons(sizeof(struct udphdr) + mss);
204 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
205 netdev_features_t features)
207 struct sock *sk = gso_skb->sk;
208 unsigned int sum_truesize = 0;
209 struct sk_buff *segs, *seg;
216 if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)
217 return __udp_gso_segment_list(gso_skb, features);
219 mss = skb_shinfo(gso_skb)->gso_size;
220 if (gso_skb->len <= sizeof(*uh) + mss)
221 return ERR_PTR(-EINVAL);
223 skb_pull(gso_skb, sizeof(*uh));
225 /* clear destructor to avoid skb_segment assigning it to tail */
226 copy_dtor = gso_skb->destructor == sock_wfree;
228 gso_skb->destructor = NULL;
230 segs = skb_segment(gso_skb, features);
231 if (IS_ERR_OR_NULL(segs)) {
233 gso_skb->destructor = sock_wfree;
237 /* GSO partial and frag_list segmentation only requires splitting
238 * the frame into an MSS multiple and possibly a remainder, both
239 * cases return a GSO skb. So update the mss now.
241 if (skb_is_gso(segs))
242 mss *= skb_shinfo(segs)->gso_segs;
247 /* preserve TX timestamp flags and TS key for first segment */
248 skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
249 skb_shinfo(seg)->tx_flags |=
250 (skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
252 /* compute checksum adjustment based on old length versus new */
253 newlen = htons(sizeof(*uh) + mss);
254 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
258 seg->destructor = sock_wfree;
260 sum_truesize += seg->truesize;
269 if (seg->ip_summed == CHECKSUM_PARTIAL)
270 gso_reset_checksum(seg, ~check);
272 uh->check = gso_make_checksum(seg, ~check) ? :
279 /* last packet can be partial gso_size, account for that in checksum */
280 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
282 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
287 if (seg->ip_summed == CHECKSUM_PARTIAL)
288 gso_reset_checksum(seg, ~check);
290 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
292 /* update refcount for the packet */
294 int delta = sum_truesize - gso_skb->truesize;
296 /* In some pathological cases, delta can be negative.
297 * We need to either use refcount_add() or refcount_sub_and_test()
299 if (likely(delta >= 0))
300 refcount_add(delta, &sk->sk_wmem_alloc);
302 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
306 EXPORT_SYMBOL_GPL(__udp_gso_segment);
308 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
309 netdev_features_t features)
311 struct sk_buff *segs = ERR_PTR(-EINVAL);
317 if (skb->encapsulation &&
318 (skb_shinfo(skb)->gso_type &
319 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
320 segs = skb_udp_tunnel_segment(skb, features, false);
324 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
327 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
330 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
331 return __udp_gso_segment(skb, features);
333 mss = skb_shinfo(skb)->gso_size;
334 if (unlikely(skb->len <= mss))
337 /* Do software UFO. Complete and fill in the UDP checksum as
338 * HW cannot do checksum of UDP packets sent as multiple
346 csum = skb_checksum(skb, 0, skb->len, 0);
347 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
349 uh->check = CSUM_MANGLED_0;
351 skb->ip_summed = CHECKSUM_UNNECESSARY;
353 /* If there is no outer header we can fake a checksum offload
354 * due to the fact that we have already done the checksum in
355 * software prior to segmenting the frame.
357 if (!skb->encap_hdr_csum)
358 features |= NETIF_F_HW_CSUM;
360 /* Fragment the skb. IP headers of the fragments are updated in
363 segs = skb_segment(skb, features);
368 #define UDP_GRO_CNT_MAX 64
369 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
372 struct udphdr *uh = udp_gro_udphdr(skb);
373 struct sk_buff *pp = NULL;
379 /* requires non zero csum, for symmetry with GSO */
381 NAPI_GRO_CB(skb)->flush = 1;
385 /* Do not deal with padded or malicious packets, sorry ! */
386 ulen = ntohs(uh->len);
387 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
388 NAPI_GRO_CB(skb)->flush = 1;
391 /* pull encapsulating udp header */
392 skb_gro_pull(skb, sizeof(struct udphdr));
394 list_for_each_entry(p, head, list) {
395 if (!NAPI_GRO_CB(p)->same_flow)
400 /* Match ports only, as csum is always non zero */
401 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
402 NAPI_GRO_CB(p)->same_flow = 0;
406 if (NAPI_GRO_CB(skb)->is_flist != NAPI_GRO_CB(p)->is_flist) {
407 NAPI_GRO_CB(skb)->flush = 1;
411 /* Terminate the flow on len mismatch or if it grow "too much".
412 * Under small packet flood GRO count could elsewhere grow a lot
413 * leading to excessive truesize values.
414 * On len mismatch merge the first packet shorter than gso_size,
415 * otherwise complete the GRO packet.
417 if (ulen > ntohs(uh2->len)) {
420 if (NAPI_GRO_CB(skb)->is_flist) {
421 if (!pskb_may_pull(skb, skb_gro_offset(skb))) {
422 NAPI_GRO_CB(skb)->flush = 1;
425 if ((skb->ip_summed != p->ip_summed) ||
426 (skb->csum_level != p->csum_level)) {
427 NAPI_GRO_CB(skb)->flush = 1;
430 ret = skb_gro_receive_list(p, skb);
432 skb_gro_postpull_rcsum(skb, uh,
433 sizeof(struct udphdr));
435 ret = skb_gro_receive(p, skb);
439 if (ret || ulen != ntohs(uh2->len) ||
440 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
446 /* mismatch, but we never need to flush */
450 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
451 struct udphdr *uh, struct sock *sk)
453 struct sk_buff *pp = NULL;
456 unsigned int off = skb_gro_offset(skb);
459 NAPI_GRO_CB(skb)->is_flist = 0;
460 if (skb->dev->features & NETIF_F_GRO_FRAGLIST)
461 NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled: 1;
463 if ((sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist) {
464 pp = call_gro_receive(udp_gro_receive_segment, head, skb);
468 if (!sk || NAPI_GRO_CB(skb)->encap_mark ||
469 (skb->ip_summed != CHECKSUM_PARTIAL &&
470 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
471 !NAPI_GRO_CB(skb)->csum_valid) ||
472 !udp_sk(sk)->gro_receive)
475 /* mark that this skb passed once through the tunnel gro layer */
476 NAPI_GRO_CB(skb)->encap_mark = 1;
480 list_for_each_entry(p, head, list) {
481 if (!NAPI_GRO_CB(p)->same_flow)
484 uh2 = (struct udphdr *)(p->data + off);
486 /* Match ports and either checksums are either both zero
489 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
490 (!uh->check ^ !uh2->check)) {
491 NAPI_GRO_CB(p)->same_flow = 0;
496 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
497 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
498 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
501 skb_gro_flush_final(skb, pp, flush);
504 EXPORT_SYMBOL(udp_gro_receive);
506 static struct sock *udp4_gro_lookup_skb(struct sk_buff *skb, __be16 sport,
509 const struct iphdr *iph = skb_gro_network_header(skb);
511 return __udp4_lib_lookup(dev_net(skb->dev), iph->saddr, sport,
512 iph->daddr, dport, inet_iif(skb),
513 inet_sdif(skb), &udp_table, NULL);
516 INDIRECT_CALLABLE_SCOPE
517 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
519 struct udphdr *uh = udp_gro_udphdr(skb);
520 struct sock *sk = NULL;
526 /* Don't bother verifying checksum if we're going to flush anyway. */
527 if (NAPI_GRO_CB(skb)->flush)
530 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
531 inet_gro_compute_pseudo))
534 skb_gro_checksum_try_convert(skb, IPPROTO_UDP,
535 inet_gro_compute_pseudo);
537 NAPI_GRO_CB(skb)->is_ipv6 = 0;
540 if (static_branch_unlikely(&udp_encap_needed_key))
541 sk = udp4_gro_lookup_skb(skb, uh->source, uh->dest);
543 pp = udp_gro_receive(head, skb, uh, sk);
548 NAPI_GRO_CB(skb)->flush = 1;
552 static int udp_gro_complete_segment(struct sk_buff *skb)
554 struct udphdr *uh = udp_hdr(skb);
556 skb->csum_start = (unsigned char *)uh - skb->head;
557 skb->csum_offset = offsetof(struct udphdr, check);
558 skb->ip_summed = CHECKSUM_PARTIAL;
560 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
561 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
565 int udp_gro_complete(struct sk_buff *skb, int nhoff,
568 __be16 newlen = htons(skb->len - nhoff);
569 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
576 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
577 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
578 if (sk && udp_sk(sk)->gro_complete) {
579 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
580 : SKB_GSO_UDP_TUNNEL;
582 /* Set encapsulation before calling into inner gro_complete()
583 * functions to make them set up the inner offsets.
585 skb->encapsulation = 1;
586 err = udp_sk(sk)->gro_complete(sk, skb,
587 nhoff + sizeof(struct udphdr));
589 err = udp_gro_complete_segment(skb);
593 if (skb->remcsum_offload)
594 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
598 EXPORT_SYMBOL(udp_gro_complete);
600 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
602 const struct iphdr *iph = ip_hdr(skb);
603 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
605 if (NAPI_GRO_CB(skb)->is_flist) {
606 uh->len = htons(skb->len - nhoff);
608 skb_shinfo(skb)->gso_type |= (SKB_GSO_FRAGLIST|SKB_GSO_UDP_L4);
609 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
611 if (skb->ip_summed == CHECKSUM_UNNECESSARY) {
612 if (skb->csum_level < SKB_MAX_CSUM_LEVEL)
615 skb->ip_summed = CHECKSUM_UNNECESSARY;
623 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
626 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
629 static const struct net_offload udpv4_offload = {
631 .gso_segment = udp4_ufo_fragment,
632 .gro_receive = udp4_gro_receive,
633 .gro_complete = udp4_gro_complete,
637 int __init udpv4_offload_init(void)
639 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);