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;
64 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
67 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
73 copy_destructor = gso_skb->destructor == tcp_wfree;
74 ooo_okay = gso_skb->ooo_okay;
75 /* All segments but the first should have ooo_okay cleared */
78 segs = skb_segment(skb, features);
82 /* Only first segment might have ooo_okay set */
83 segs->ooo_okay = ooo_okay;
85 delta = htonl(oldlen + (thlen + mss));
91 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
95 th->fin = th->psh = 0;
98 if (skb->ip_summed != CHECKSUM_PARTIAL)
100 csum_fold(csum_partial(skb_transport_header(skb),
104 if (copy_destructor) {
105 skb->destructor = gso_skb->destructor;
106 skb->sk = gso_skb->sk;
107 /* {tcp|sock}_wfree() use exact truesize accounting :
108 * sum(skb->truesize) MUST be exactly be gso_skb->truesize
109 * So we account mss bytes of 'true size' for each segment.
110 * The last segment will contain the remaining.
113 gso_skb->truesize -= mss;
118 th->seq = htonl(seq);
122 /* Following permits TCP Small Queues to work well with GSO :
123 * The callback to TCP stack will be called at the time last frag
124 * is freed at TX completion, and not right now when gso_skb
125 * is freed by GSO engine
127 if (copy_destructor) {
128 swap(gso_skb->sk, skb->sk);
129 swap(gso_skb->destructor, skb->destructor);
130 swap(gso_skb->truesize, skb->truesize);
133 delta = htonl(oldlen + (skb_tail_pointer(skb) -
134 skb_transport_header(skb)) +
136 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
137 (__force u32)delta));
138 if (skb->ip_summed != CHECKSUM_PARTIAL)
139 th->check = csum_fold(csum_partial(skb_transport_header(skb),
144 EXPORT_SYMBOL(tcp_gso_segment);
146 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
148 struct sk_buff **pp = NULL;
155 unsigned int mss = 1;
161 off = skb_gro_offset(skb);
162 hlen = off + sizeof(*th);
163 th = skb_gro_header_fast(skb, off);
164 if (skb_gro_header_hard(skb, hlen)) {
165 th = skb_gro_header_slow(skb, hlen, off);
170 thlen = th->doff * 4;
171 if (thlen < sizeof(*th))
175 if (skb_gro_header_hard(skb, hlen)) {
176 th = skb_gro_header_slow(skb, hlen, off);
181 skb_gro_pull(skb, thlen);
183 len = skb_gro_len(skb);
184 flags = tcp_flag_word(th);
186 for (; (p = *head); head = &p->next) {
187 if (!NAPI_GRO_CB(p)->same_flow)
192 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
193 NAPI_GRO_CB(p)->same_flow = 0;
200 goto out_check_final;
203 flush = NAPI_GRO_CB(p)->flush;
204 flush |= (__force int)(flags & TCP_FLAG_CWR);
205 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
206 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
207 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
208 for (i = sizeof(*th); i < thlen; i += 4)
209 flush |= *(u32 *)((u8 *)th + i) ^
210 *(u32 *)((u8 *)th2 + i);
212 mss = tcp_skb_mss(p);
214 flush |= (len - 1) >= mss;
215 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
217 if (flush || skb_gro_receive(head, skb)) {
219 goto out_check_final;
224 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
228 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
229 TCP_FLAG_RST | TCP_FLAG_SYN |
232 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
236 NAPI_GRO_CB(skb)->flush |= flush;
240 EXPORT_SYMBOL(tcp_gro_receive);
242 int tcp_gro_complete(struct sk_buff *skb)
244 struct tcphdr *th = tcp_hdr(skb);
246 skb->csum_start = skb_transport_header(skb) - skb->head;
247 skb->csum_offset = offsetof(struct tcphdr, check);
248 skb->ip_summed = CHECKSUM_PARTIAL;
250 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
253 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
257 EXPORT_SYMBOL(tcp_gro_complete);
259 static int tcp_v4_gso_send_check(struct sk_buff *skb)
261 const struct iphdr *iph;
264 if (!pskb_may_pull(skb, sizeof(*th)))
271 skb->ip_summed = CHECKSUM_PARTIAL;
272 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
276 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
278 const struct iphdr *iph = skb_gro_network_header(skb);
282 switch (skb->ip_summed) {
283 case CHECKSUM_COMPLETE:
284 if (!tcp_v4_check(skb_gro_len(skb), iph->saddr, iph->daddr,
286 skb->ip_summed = CHECKSUM_UNNECESSARY;
290 NAPI_GRO_CB(skb)->flush = 1;
294 wsum = csum_tcpudp_nofold(iph->saddr, iph->daddr,
295 skb_gro_len(skb), IPPROTO_TCP, 0);
296 sum = csum_fold(skb_checksum(skb,
303 skb->ip_summed = CHECKSUM_UNNECESSARY;
307 return tcp_gro_receive(head, skb);
310 static int tcp4_gro_complete(struct sk_buff *skb)
312 const struct iphdr *iph = ip_hdr(skb);
313 struct tcphdr *th = tcp_hdr(skb);
315 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
316 iph->saddr, iph->daddr, 0);
317 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
319 return tcp_gro_complete(skb);
322 static const struct net_offload tcpv4_offload = {
324 .gso_send_check = tcp_v4_gso_send_check,
325 .gso_segment = tcp_gso_segment,
326 .gro_receive = tcp4_gro_receive,
327 .gro_complete = tcp4_gro_complete,
331 int __init tcpv4_offload_init(void)
333 return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);