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.
10 * TCPv4 GSO/GRO support
13 #include <linux/skbuff.h>
15 #include <net/protocol.h>
17 struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
18 netdev_features_t features)
20 struct sk_buff *segs = ERR_PTR(-EINVAL);
27 struct sk_buff *gso_skb = skb;
29 bool ooo_okay, copy_destructor;
31 if (!pskb_may_pull(skb, sizeof(*th)))
36 if (thlen < sizeof(*th))
39 if (!pskb_may_pull(skb, thlen))
42 oldlen = (u16)~skb->len;
43 __skb_pull(skb, thlen);
45 mss = tcp_skb_mss(skb);
46 if (unlikely(skb->len <= mss))
49 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
50 /* Packet is from an untrusted source, reset gso_segs. */
51 int type = skb_shinfo(skb)->gso_type;
63 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
66 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
72 copy_destructor = gso_skb->destructor == tcp_wfree;
73 ooo_okay = gso_skb->ooo_okay;
74 /* All segments but the first should have ooo_okay cleared */
77 segs = skb_segment(skb, features);
81 /* Only first segment might have ooo_okay set */
82 segs->ooo_okay = ooo_okay;
84 delta = htonl(oldlen + (thlen + mss));
90 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
94 th->fin = th->psh = 0;
97 if (skb->ip_summed != CHECKSUM_PARTIAL)
99 csum_fold(csum_partial(skb_transport_header(skb),
103 if (copy_destructor) {
104 skb->destructor = gso_skb->destructor;
105 skb->sk = gso_skb->sk;
106 /* {tcp|sock}_wfree() use exact truesize accounting :
107 * sum(skb->truesize) MUST be exactly be gso_skb->truesize
108 * So we account mss bytes of 'true size' for each segment.
109 * The last segment will contain the remaining.
112 gso_skb->truesize -= mss;
117 th->seq = htonl(seq);
121 /* Following permits TCP Small Queues to work well with GSO :
122 * The callback to TCP stack will be called at the time last frag
123 * is freed at TX completion, and not right now when gso_skb
124 * is freed by GSO engine
126 if (copy_destructor) {
127 swap(gso_skb->sk, skb->sk);
128 swap(gso_skb->destructor, skb->destructor);
129 swap(gso_skb->truesize, skb->truesize);
132 delta = htonl(oldlen + (skb_tail_pointer(skb) -
133 skb_transport_header(skb)) +
135 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
136 (__force u32)delta));
137 if (skb->ip_summed != CHECKSUM_PARTIAL)
138 th->check = csum_fold(csum_partial(skb_transport_header(skb),
143 EXPORT_SYMBOL(tcp_gso_segment);
145 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
147 struct sk_buff **pp = NULL;
154 unsigned int mss = 1;
160 off = skb_gro_offset(skb);
161 hlen = off + sizeof(*th);
162 th = skb_gro_header_fast(skb, off);
163 if (skb_gro_header_hard(skb, hlen)) {
164 th = skb_gro_header_slow(skb, hlen, off);
169 thlen = th->doff * 4;
170 if (thlen < sizeof(*th))
174 if (skb_gro_header_hard(skb, hlen)) {
175 th = skb_gro_header_slow(skb, hlen, off);
180 skb_gro_pull(skb, thlen);
182 len = skb_gro_len(skb);
183 flags = tcp_flag_word(th);
185 for (; (p = *head); head = &p->next) {
186 if (!NAPI_GRO_CB(p)->same_flow)
191 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
192 NAPI_GRO_CB(p)->same_flow = 0;
199 goto out_check_final;
202 flush = NAPI_GRO_CB(p)->flush;
203 flush |= (__force int)(flags & TCP_FLAG_CWR);
204 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
205 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
206 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
207 for (i = sizeof(*th); i < thlen; i += 4)
208 flush |= *(u32 *)((u8 *)th + i) ^
209 *(u32 *)((u8 *)th2 + i);
211 mss = tcp_skb_mss(p);
213 flush |= (len - 1) >= mss;
214 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
216 if (flush || skb_gro_receive(head, skb)) {
218 goto out_check_final;
223 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
227 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
228 TCP_FLAG_RST | TCP_FLAG_SYN |
231 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
235 NAPI_GRO_CB(skb)->flush |= flush;
239 EXPORT_SYMBOL(tcp_gro_receive);
241 int tcp_gro_complete(struct sk_buff *skb)
243 struct tcphdr *th = tcp_hdr(skb);
245 skb->csum_start = skb_transport_header(skb) - skb->head;
246 skb->csum_offset = offsetof(struct tcphdr, check);
247 skb->ip_summed = CHECKSUM_PARTIAL;
249 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
252 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
256 EXPORT_SYMBOL(tcp_gro_complete);
258 static int tcp_v4_gso_send_check(struct sk_buff *skb)
260 const struct iphdr *iph;
263 if (!pskb_may_pull(skb, sizeof(*th)))
270 skb->ip_summed = CHECKSUM_PARTIAL;
271 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
275 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
277 const struct iphdr *iph = skb_gro_network_header(skb);
281 switch (skb->ip_summed) {
282 case CHECKSUM_COMPLETE:
283 if (!tcp_v4_check(skb_gro_len(skb), iph->saddr, iph->daddr,
285 skb->ip_summed = CHECKSUM_UNNECESSARY;
289 NAPI_GRO_CB(skb)->flush = 1;
293 wsum = csum_tcpudp_nofold(iph->saddr, iph->daddr,
294 skb_gro_len(skb), IPPROTO_TCP, 0);
295 sum = csum_fold(skb_checksum(skb,
302 skb->ip_summed = CHECKSUM_UNNECESSARY;
306 return tcp_gro_receive(head, skb);
309 static int tcp4_gro_complete(struct sk_buff *skb)
311 const struct iphdr *iph = ip_hdr(skb);
312 struct tcphdr *th = tcp_hdr(skb);
314 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
315 iph->saddr, iph->daddr, 0);
316 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
318 return tcp_gro_complete(skb);
321 static const struct net_offload tcpv4_offload = {
323 .gso_send_check = tcp_v4_gso_send_check,
324 .gso_segment = tcp_gso_segment,
325 .gro_receive = tcp4_gro_receive,
326 .gro_complete = tcp4_gro_complete,
330 int __init tcpv4_offload_init(void)
332 return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);