2 * IPV4 GSO/GRO offload support
3 * Linux INET implementation
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License
7 * as published by the Free Software Foundation; either version
8 * 2 of the License, or (at your option) any later version.
13 #include <linux/skbuff.h>
15 #include <net/protocol.h>
17 static DEFINE_SPINLOCK(udp_offload_lock);
18 static struct udp_offload_priv __rcu *udp_offload_base __read_mostly;
20 #define udp_deref_protected(X) rcu_dereference_protected(X, lockdep_is_held(&udp_offload_lock))
22 struct udp_offload_priv {
23 struct udp_offload *offload;
25 struct udp_offload_priv __rcu *next;
28 static struct sk_buff *__skb_udp_tunnel_segment(struct sk_buff *skb,
29 netdev_features_t features,
30 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
31 netdev_features_t features),
32 __be16 new_protocol, bool is_ipv6)
34 struct sk_buff *segs = ERR_PTR(-EINVAL);
35 u16 mac_offset = skb->mac_header;
36 int mac_len = skb->mac_len;
37 int tnl_hlen = skb_inner_mac_header(skb) - skb_transport_header(skb);
38 __be16 protocol = skb->protocol;
39 netdev_features_t enc_features;
40 int udp_offset, outer_hlen;
42 bool need_csum = !!(skb_shinfo(skb)->gso_type &
43 SKB_GSO_UDP_TUNNEL_CSUM);
44 bool remcsum = !!(skb_shinfo(skb)->gso_type & SKB_GSO_TUNNEL_REMCSUM);
45 bool offload_csum = false, dont_encap = (need_csum || remcsum);
47 oldlen = (u16)~skb->len;
49 if (unlikely(!pskb_may_pull(skb, tnl_hlen)))
52 skb->encapsulation = 0;
53 __skb_pull(skb, tnl_hlen);
54 skb_reset_mac_header(skb);
55 skb_set_network_header(skb, skb_inner_network_offset(skb));
56 skb->mac_len = skb_inner_network_offset(skb);
57 skb->protocol = new_protocol;
58 skb->encap_hdr_csum = need_csum;
59 skb->remcsum_offload = remcsum;
61 /* Try to offload checksum if possible */
62 offload_csum = !!(need_csum &&
63 ((skb->dev->features & NETIF_F_HW_CSUM) ||
64 (skb->dev->features & (is_ipv6 ?
65 NETIF_F_IPV6_CSUM : NETIF_F_IP_CSUM))));
67 /* segment inner packet. */
68 enc_features = skb->dev->hw_enc_features & features;
69 segs = gso_inner_segment(skb, enc_features);
70 if (IS_ERR_OR_NULL(segs)) {
71 skb_gso_error_unwind(skb, protocol, tnl_hlen, mac_offset,
76 outer_hlen = skb_tnl_header_len(skb);
77 udp_offset = outer_hlen - tnl_hlen;
85 skb->encapsulation = 0;
86 skb->ip_summed = CHECKSUM_NONE;
88 /* Only set up inner headers if we might be offloading
91 skb_reset_inner_headers(skb);
92 skb->encapsulation = 1;
95 skb->mac_len = mac_len;
96 skb->protocol = protocol;
98 skb_push(skb, outer_hlen);
99 skb_reset_mac_header(skb);
100 skb_set_network_header(skb, mac_len);
101 skb_set_transport_header(skb, udp_offset);
102 len = skb->len - udp_offset;
104 uh->len = htons(len);
109 delta = htonl(oldlen + len);
111 uh->check = ~csum_fold((__force __wsum)
112 ((__force u32)uh->check +
113 (__force u32)delta));
115 skb->ip_summed = CHECKSUM_PARTIAL;
116 skb->csum_start = skb_transport_header(skb) - skb->head;
117 skb->csum_offset = offsetof(struct udphdr, check);
118 } else if (remcsum) {
119 /* Need to calculate checksum from scratch,
120 * inner checksums are never when doing
121 * remote_checksum_offload.
124 skb->csum = skb_checksum(skb, udp_offset,
125 skb->len - udp_offset,
127 uh->check = csum_fold(skb->csum);
129 uh->check = CSUM_MANGLED_0;
131 uh->check = gso_make_checksum(skb, ~uh->check);
134 uh->check = CSUM_MANGLED_0;
136 } while ((skb = skb->next));
141 struct sk_buff *skb_udp_tunnel_segment(struct sk_buff *skb,
142 netdev_features_t features,
145 __be16 protocol = skb->protocol;
146 const struct net_offload **offloads;
147 const struct net_offload *ops;
148 struct sk_buff *segs = ERR_PTR(-EINVAL);
149 struct sk_buff *(*gso_inner_segment)(struct sk_buff *skb,
150 netdev_features_t features);
154 switch (skb->inner_protocol_type) {
155 case ENCAP_TYPE_ETHER:
156 protocol = skb->inner_protocol;
157 gso_inner_segment = skb_mac_gso_segment;
159 case ENCAP_TYPE_IPPROTO:
160 offloads = is_ipv6 ? inet6_offloads : inet_offloads;
161 ops = rcu_dereference(offloads[skb->inner_ipproto]);
162 if (!ops || !ops->callbacks.gso_segment)
164 gso_inner_segment = ops->callbacks.gso_segment;
170 segs = __skb_udp_tunnel_segment(skb, features, gso_inner_segment,
179 static struct sk_buff *udp4_ufo_fragment(struct sk_buff *skb,
180 netdev_features_t features)
182 struct sk_buff *segs = ERR_PTR(-EINVAL);
188 if (skb->encapsulation &&
189 (skb_shinfo(skb)->gso_type &
190 (SKB_GSO_UDP_TUNNEL|SKB_GSO_UDP_TUNNEL_CSUM))) {
191 segs = skb_udp_tunnel_segment(skb, features, false);
195 if (!pskb_may_pull(skb, sizeof(struct udphdr)))
198 mss = skb_shinfo(skb)->gso_size;
199 if (unlikely(skb->len <= mss))
202 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
203 /* Packet is from an untrusted source, reset gso_segs. */
204 int type = skb_shinfo(skb)->gso_type;
206 if (unlikely(type & ~(SKB_GSO_UDP | SKB_GSO_DODGY |
208 SKB_GSO_UDP_TUNNEL_CSUM |
209 SKB_GSO_TUNNEL_REMCSUM |
211 SKB_GSO_GRE | SKB_GSO_GRE_CSUM) ||
212 !(type & (SKB_GSO_UDP))))
215 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
221 /* Do software UFO. Complete and fill in the UDP checksum as
222 * HW cannot do checksum of UDP packets sent as multiple
230 csum = skb_checksum(skb, 0, skb->len, 0);
231 uh->check = udp_v4_check(skb->len, iph->saddr, iph->daddr, csum);
233 uh->check = CSUM_MANGLED_0;
235 skb->ip_summed = CHECKSUM_NONE;
237 /* Fragment the skb. IP headers of the fragments are updated in
240 segs = skb_segment(skb, features);
245 int udp_add_offload(struct udp_offload *uo)
247 struct udp_offload_priv *new_offload = kzalloc(sizeof(*new_offload), GFP_ATOMIC);
252 new_offload->offload = uo;
254 spin_lock(&udp_offload_lock);
255 new_offload->next = udp_offload_base;
256 rcu_assign_pointer(udp_offload_base, new_offload);
257 spin_unlock(&udp_offload_lock);
261 EXPORT_SYMBOL(udp_add_offload);
263 static void udp_offload_free_routine(struct rcu_head *head)
265 struct udp_offload_priv *ou_priv = container_of(head, struct udp_offload_priv, rcu);
269 void udp_del_offload(struct udp_offload *uo)
271 struct udp_offload_priv __rcu **head = &udp_offload_base;
272 struct udp_offload_priv *uo_priv;
274 spin_lock(&udp_offload_lock);
276 uo_priv = udp_deref_protected(*head);
277 for (; uo_priv != NULL;
278 uo_priv = udp_deref_protected(*head)) {
279 if (uo_priv->offload == uo) {
280 rcu_assign_pointer(*head,
281 udp_deref_protected(uo_priv->next));
284 head = &uo_priv->next;
286 pr_warn("udp_del_offload: didn't find offload for port %d\n", ntohs(uo->port));
288 spin_unlock(&udp_offload_lock);
290 call_rcu(&uo_priv->rcu, udp_offload_free_routine);
292 EXPORT_SYMBOL(udp_del_offload);
294 struct sk_buff **udp_gro_receive(struct sk_buff **head, struct sk_buff *skb,
297 struct udp_offload_priv *uo_priv;
298 struct sk_buff *p, **pp = NULL;
300 unsigned int off = skb_gro_offset(skb);
303 if (NAPI_GRO_CB(skb)->udp_mark ||
304 (skb->ip_summed != CHECKSUM_PARTIAL &&
305 NAPI_GRO_CB(skb)->csum_cnt == 0 &&
306 !NAPI_GRO_CB(skb)->csum_valid))
309 /* mark that this skb passed once through the udp gro layer */
310 NAPI_GRO_CB(skb)->udp_mark = 1;
313 uo_priv = rcu_dereference(udp_offload_base);
314 for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
315 if (uo_priv->offload->port == uh->dest &&
316 uo_priv->offload->callbacks.gro_receive)
324 for (p = *head; p; p = p->next) {
325 if (!NAPI_GRO_CB(p)->same_flow)
328 uh2 = (struct udphdr *)(p->data + off);
330 /* Match ports and either checksums are either both zero
333 if ((*(u32 *)&uh->source != *(u32 *)&uh2->source) ||
334 (!uh->check ^ !uh2->check)) {
335 NAPI_GRO_CB(p)->same_flow = 0;
340 skb_gro_pull(skb, sizeof(struct udphdr)); /* pull encapsulating udp header */
341 skb_gro_postpull_rcsum(skb, uh, sizeof(struct udphdr));
342 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
343 pp = uo_priv->offload->callbacks.gro_receive(head, skb,
349 NAPI_GRO_CB(skb)->flush |= flush;
353 static struct sk_buff **udp4_gro_receive(struct sk_buff **head,
356 struct udphdr *uh = udp_gro_udphdr(skb);
361 /* Don't bother verifying checksum if we're going to flush anyway. */
362 if (NAPI_GRO_CB(skb)->flush)
365 if (skb_gro_checksum_validate_zero_check(skb, IPPROTO_UDP, uh->check,
366 inet_gro_compute_pseudo))
369 skb_gro_checksum_try_convert(skb, IPPROTO_UDP, uh->check,
370 inet_gro_compute_pseudo);
372 NAPI_GRO_CB(skb)->is_ipv6 = 0;
373 return udp_gro_receive(head, skb, uh);
376 NAPI_GRO_CB(skb)->flush = 1;
380 int udp_gro_complete(struct sk_buff *skb, int nhoff)
382 struct udp_offload_priv *uo_priv;
383 __be16 newlen = htons(skb->len - nhoff);
384 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
391 uo_priv = rcu_dereference(udp_offload_base);
392 for (; uo_priv != NULL; uo_priv = rcu_dereference(uo_priv->next)) {
393 if (uo_priv->offload->port == uh->dest &&
394 uo_priv->offload->callbacks.gro_complete)
399 NAPI_GRO_CB(skb)->proto = uo_priv->offload->ipproto;
400 err = uo_priv->offload->callbacks.gro_complete(skb,
401 nhoff + sizeof(struct udphdr),
407 if (skb->remcsum_offload)
408 skb_shinfo(skb)->gso_type |= SKB_GSO_TUNNEL_REMCSUM;
410 skb->encapsulation = 1;
411 skb_set_inner_mac_header(skb, nhoff + sizeof(struct udphdr));
416 static int udp4_gro_complete(struct sk_buff *skb, int nhoff)
418 const struct iphdr *iph = ip_hdr(skb);
419 struct udphdr *uh = (struct udphdr *)(skb->data + nhoff);
422 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL_CSUM;
423 uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
426 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
429 return udp_gro_complete(skb, nhoff);
432 static const struct net_offload udpv4_offload = {
434 .gso_segment = udp4_ufo_fragment,
435 .gro_receive = udp4_gro_receive,
436 .gro_complete = udp4_gro_complete,
440 int __init udpv4_offload_init(void)
442 return inet_add_offload(&udpv4_offload, IPPROTO_UDP);