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_tso_segment(struct sk_buff *skb,
18 netdev_features_t features)
20 struct sk_buff *segs = ERR_PTR(-EINVAL);
21 unsigned int sum_truesize = 0;
28 struct sk_buff *gso_skb = skb;
30 bool ooo_okay, copy_destructor;
32 if (!pskb_may_pull(skb, sizeof(*th)))
37 if (thlen < sizeof(*th))
40 if (!pskb_may_pull(skb, thlen))
43 oldlen = (u16)~skb->len;
44 __skb_pull(skb, thlen);
46 mss = tcp_skb_mss(skb);
47 if (unlikely(skb->len <= mss))
50 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
51 /* Packet is from an untrusted source, reset gso_segs. */
52 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 sum_truesize += skb->truesize;
111 th->seq = htonl(seq);
115 /* Following permits TCP Small Queues to work well with GSO :
116 * The callback to TCP stack will be called at the time last frag
117 * is freed at TX completion, and not right now when gso_skb
118 * is freed by GSO engine
120 if (copy_destructor) {
121 swap(gso_skb->sk, skb->sk);
122 swap(gso_skb->destructor, skb->destructor);
123 sum_truesize += skb->truesize;
124 atomic_add(sum_truesize - gso_skb->truesize,
125 &skb->sk->sk_wmem_alloc);
128 delta = htonl(oldlen + (skb_tail_pointer(skb) -
129 skb_transport_header(skb)) +
131 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
132 (__force u32)delta));
133 if (skb->ip_summed != CHECKSUM_PARTIAL)
134 th->check = csum_fold(csum_partial(skb_transport_header(skb),
139 EXPORT_SYMBOL(tcp_tso_segment);
141 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
143 struct sk_buff **pp = NULL;
150 unsigned int mss = 1;
156 off = skb_gro_offset(skb);
157 hlen = off + sizeof(*th);
158 th = skb_gro_header_fast(skb, off);
159 if (skb_gro_header_hard(skb, hlen)) {
160 th = skb_gro_header_slow(skb, hlen, off);
165 thlen = th->doff * 4;
166 if (thlen < sizeof(*th))
170 if (skb_gro_header_hard(skb, hlen)) {
171 th = skb_gro_header_slow(skb, hlen, off);
176 skb_gro_pull(skb, thlen);
178 len = skb_gro_len(skb);
179 flags = tcp_flag_word(th);
181 for (; (p = *head); head = &p->next) {
182 if (!NAPI_GRO_CB(p)->same_flow)
187 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
188 NAPI_GRO_CB(p)->same_flow = 0;
195 goto out_check_final;
198 flush = NAPI_GRO_CB(p)->flush;
199 flush |= (__force int)(flags & TCP_FLAG_CWR);
200 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
201 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
202 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
203 for (i = sizeof(*th); i < thlen; i += 4)
204 flush |= *(u32 *)((u8 *)th + i) ^
205 *(u32 *)((u8 *)th2 + i);
207 mss = tcp_skb_mss(p);
209 flush |= (len - 1) >= mss;
210 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
212 if (flush || skb_gro_receive(head, skb)) {
214 goto out_check_final;
219 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
223 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
224 TCP_FLAG_RST | TCP_FLAG_SYN |
227 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
231 NAPI_GRO_CB(skb)->flush |= flush;
235 EXPORT_SYMBOL(tcp_gro_receive);
237 int tcp_gro_complete(struct sk_buff *skb)
239 struct tcphdr *th = tcp_hdr(skb);
241 skb->csum_start = skb_transport_header(skb) - skb->head;
242 skb->csum_offset = offsetof(struct tcphdr, check);
243 skb->ip_summed = CHECKSUM_PARTIAL;
245 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
248 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
252 EXPORT_SYMBOL(tcp_gro_complete);
254 static int tcp_v4_gso_send_check(struct sk_buff *skb)
256 const struct iphdr *iph;
259 if (!pskb_may_pull(skb, sizeof(*th)))
266 skb->ip_summed = CHECKSUM_PARTIAL;
267 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
271 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
273 const struct iphdr *iph = skb_gro_network_header(skb);
277 switch (skb->ip_summed) {
278 case CHECKSUM_COMPLETE:
279 if (!tcp_v4_check(skb_gro_len(skb), iph->saddr, iph->daddr,
281 skb->ip_summed = CHECKSUM_UNNECESSARY;
285 NAPI_GRO_CB(skb)->flush = 1;
289 wsum = csum_tcpudp_nofold(iph->saddr, iph->daddr,
290 skb_gro_len(skb), IPPROTO_TCP, 0);
291 sum = csum_fold(skb_checksum(skb,
298 skb->ip_summed = CHECKSUM_UNNECESSARY;
302 return tcp_gro_receive(head, skb);
305 static int tcp4_gro_complete(struct sk_buff *skb)
307 const struct iphdr *iph = ip_hdr(skb);
308 struct tcphdr *th = tcp_hdr(skb);
310 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
311 iph->saddr, iph->daddr, 0);
312 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
314 return tcp_gro_complete(skb);
317 static const struct net_offload tcpv4_offload = {
319 .gso_send_check = tcp_v4_gso_send_check,
320 .gso_segment = tcp_tso_segment,
321 .gro_receive = tcp4_gro_receive,
322 .gro_complete = tcp4_gro_complete,
326 int __init tcpv4_offload_init(void)
328 return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);