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 struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
188 netdev_features_t features)
190 struct sock *sk = gso_skb->sk;
191 unsigned int sum_truesize = 0;
192 struct sk_buff *segs, *seg;
199 mss = skb_shinfo(gso_skb)->gso_size;
200 if (gso_skb->len <= sizeof(*uh) + mss)
201 return ERR_PTR(-EINVAL);
203 skb_pull(gso_skb, sizeof(*uh));
205 /* clear destructor to avoid skb_segment assigning it to tail */
206 copy_dtor = gso_skb->destructor == sock_wfree;
208 gso_skb->destructor = NULL;
210 segs = skb_segment(gso_skb, features);
211 if (unlikely(IS_ERR_OR_NULL(segs))) {
213 gso_skb->destructor = sock_wfree;
217 /* GSO partial and frag_list segmentation only requires splitting
218 * the frame into an MSS multiple and possibly a remainder, both
219 * cases return a GSO skb. So update the mss now.
221 if (skb_is_gso(segs))
222 mss *= skb_shinfo(segs)->gso_segs;
227 /* compute checksum adjustment based on old length versus new */
228 newlen = htons(sizeof(*uh) + mss);
229 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
233 seg->destructor = sock_wfree;
235 sum_truesize += seg->truesize;
244 if (seg->ip_summed == CHECKSUM_PARTIAL)
245 gso_reset_checksum(seg, ~check);
247 uh->check = gso_make_checksum(seg, ~check) ? :
254 /* last packet can be partial gso_size, account for that in checksum */
255 newlen = htons(skb_tail_pointer(seg) - skb_transport_header(seg) +
257 check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
262 if (seg->ip_summed == CHECKSUM_PARTIAL)
263 gso_reset_checksum(seg, ~check);
265 uh->check = gso_make_checksum(seg, ~check) ? : CSUM_MANGLED_0;
267 /* update refcount for the packet */
269 int delta = sum_truesize - gso_skb->truesize;
271 /* In some pathological cases, delta can be negative.
272 * We need to either use refcount_add() or refcount_sub_and_test()
274 if (likely(delta >= 0))
275 refcount_add(delta, &sk->sk_wmem_alloc);
277 WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));
281 EXPORT_SYMBOL_GPL(__udp_gso_segment);
283 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
284 netdev_features_t features)
286 struct sk_buff *segs = ERR_PTR(-EINVAL);
292 if (skb->encapsulation &&
293 (skb_shinfo(skb)->gso_type &
294 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
295 segs = skb_udp_tunnel_segment(skb, features, false);
299 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_UDP | SKB_GSO_UDP_L4)))
302 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
305 if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
306 return __udp_gso_segment(skb, features);
308 mss = skb_shinfo(skb)->gso_size;
309 if (unlikely(skb->len <= mss))
312 /* Do software UFO. Complete and fill in the UDP checksum as
313 * HW cannot do checksum of UDP packets sent as multiple
321 csum = skb_checksum(skb, 0, skb->len, 0);
322 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
324 uh->check = CSUM_MANGLED_0;
326 skb->ip_summed = CHECKSUM_UNNECESSARY;
328 /* If there is no outer header we can fake a checksum offload
329 * due to the fact that we have already done the checksum in
330 * software prior to segmenting the frame.
332 if (!skb->encap_hdr_csum)
333 features |= NETIF_F_HW_CSUM;
335 /* Fragment the skb. IP headers of the fragments are updated in
338 segs = skb_segment(skb, features);
343 #define UDP_GRO_CNT_MAX 64
344 static struct sk_buff *udp_gro_receive_segment(struct list_head *head,
347 struct udphdr *uh = udp_hdr(skb);
348 struct sk_buff *pp = NULL;
353 /* requires non zero csum, for symmetry with GSO */
355 NAPI_GRO_CB(skb)->flush = 1;
359 /* Do not deal with padded or malicious packets, sorry ! */
360 ulen = ntohs(uh->len);
361 if (ulen <= sizeof(*uh) || ulen != skb_gro_len(skb)) {
362 NAPI_GRO_CB(skb)->flush = 1;
365 /* pull encapsulating udp header */
366 skb_gro_pull(skb, sizeof(struct udphdr));
367 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
369 list_for_each_entry(p, head, list) {
370 if (!NAPI_GRO_CB(p)->same_flow)
375 /* Match ports only, as csum is always non zero */
376 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source)) {
377 NAPI_GRO_CB(p)->same_flow = 0;
381 /* Terminate the flow on len mismatch or if it grow "too much".
382 * Under small packet flood GRO count could elsewhere grow a lot
383 * leading to excessive truesize values.
384 * On len mismatch merge the first packet shorter than gso_size,
385 * otherwise complete the GRO packet.
387 if (ulen > ntohs(uh2->len) || skb_gro_receive(p, skb) ||
388 ulen != ntohs(uh2->len) ||
389 NAPI_GRO_CB(p)->count >= UDP_GRO_CNT_MAX)
395 /* mismatch, but we never need to flush */
399 INDIRECT_CALLABLE_DECLARE(struct sock *udp6_lib_lookup_skb(struct sk_buff *skb,
400 __be16 sport, __be16 dport));
401 struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
402 struct udphdr *uh, udp_lookup_t lookup)
404 struct sk_buff *pp = NULL;
407 unsigned int off = skb_gro_offset(skb);
412 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
413 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
417 if (udp_sk(sk)->gro_enabled) {
418 pp = call_gro_receive(udp_gro_receive_segment, head, skb);
423 if (NAPI_GRO_CB(skb)->encap_mark ||
424 (skb->ip_summed != CHECKSUM_PARTIAL &&
425 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
426 !NAPI_GRO_CB(skb)->csum_valid) ||
427 !udp_sk(sk)->gro_receive)
430 /* mark that this skb passed once through the tunnel gro layer */
431 NAPI_GRO_CB(skb)->encap_mark = 1;
435 list_for_each_entry(p, head, list) {
436 if (!NAPI_GRO_CB(p)->same_flow)
439 uh2 = (struct udphdr *)(p->data + off);
441 /* Match ports and either checksums are either both zero
444 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
445 (!uh->check ^ !uh2->check)) {
446 NAPI_GRO_CB(p)->same_flow = 0;
451 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
452 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
453 pp = call_gro_receive_sk(udp_sk(sk)->gro_receive, sk, head, skb);
457 skb_gro_flush_final(skb, pp, flush);
460 EXPORT_SYMBOL(udp_gro_receive);
462 INDIRECT_CALLABLE_SCOPE
463 struct sk_buff *udp4_gro_receive(struct list_head *head, struct sk_buff *skb)
465 struct udphdr *uh = udp_gro_udphdr(skb);
467 if (unlikely(!uh) || !static_branch_unlikely(&udp_encap_needed_key))
470 /* Don't bother verifying checksum if we're going to flush anyway. */
471 if (NAPI_GRO_CB(skb)->flush)
474 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
475 inet_gro_compute_pseudo))
478 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
479 inet_gro_compute_pseudo);
481 NAPI_GRO_CB(skb)->is_ipv6 = 0;
482 return udp_gro_receive(head, skb, uh, udp4_lib_lookup_skb);
485 NAPI_GRO_CB(skb)->flush = 1;
489 static int udp_gro_complete_segment(struct sk_buff *skb)
491 struct udphdr *uh = udp_hdr(skb);
493 skb->csum_start = (unsigned char *)uh - skb->head;
494 skb->csum_offset = offsetof(struct udphdr, check);
495 skb->ip_summed = CHECKSUM_PARTIAL;
497 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
498 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_L4;
502 int udp_gro_complete(struct sk_buff *skb, int nhoff,
505 __be16 newlen = htons(skb->len - nhoff);
506 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
513 sk = INDIRECT_CALL_INET(lookup, udp6_lib_lookup_skb,
514 udp4_lib_lookup_skb, skb, uh->source, uh->dest);
515 if (sk && udp_sk(sk)->gro_enabled) {
516 err = udp_gro_complete_segment(skb);
517 } else if (sk && udp_sk(sk)->gro_complete) {
518 skb_shinfo(skb)->gso_type = uh->check ? SKB_GSO_UDP_TUNNEL_CSUM
519 : SKB_GSO_UDP_TUNNEL;
521 /* Set encapsulation before calling into inner gro_complete()
522 * functions to make them set up the inner offsets.
524 skb->encapsulation = 1;
525 err = udp_sk(sk)->gro_complete(sk, skb,
526 nhoff + sizeof(struct udphdr));
530 if (skb->remcsum_offload)
531 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
535 EXPORT_SYMBOL(udp_gro_complete);
537 INDIRECT_CALLABLE_SCOPE int udp4_gro_complete(struct sk_buff *skb, int nhoff)
539 const struct iphdr *iph = ip_hdr(skb);
540 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
543 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
546 return udp_gro_complete(skb, nhoff, udp4_lib_lookup_skb);
549 static const struct net_offload udpv4_offload = {
551 .gso_segment = udp4_ufo_fragment,
552 .gro_receive = udp4_gro_receive,
553 .gro_complete = udp4_gro_complete,
557 int __init udpv4_offload_init(void)
559 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);