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);
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;
62 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
65 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
71 copy_destructor = gso_skb->destructor == tcp_wfree;
72 ooo_okay = gso_skb->ooo_okay;
73 /* All segments but the first should have ooo_okay cleared */
76 segs = skb_segment(skb, features);
80 /* Only first segment might have ooo_okay set */
81 segs->ooo_okay = ooo_okay;
83 delta = htonl(oldlen + (thlen + mss));
89 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
93 th->fin = th->psh = 0;
96 if (skb->ip_summed != CHECKSUM_PARTIAL)
98 csum_fold(csum_partial(skb_transport_header(skb),
102 if (copy_destructor) {
103 skb->destructor = gso_skb->destructor;
104 skb->sk = gso_skb->sk;
105 /* {tcp|sock}_wfree() use exact truesize accounting :
106 * sum(skb->truesize) MUST be exactly be gso_skb->truesize
107 * So we account mss bytes of 'true size' for each segment.
108 * The last segment will contain the remaining.
111 gso_skb->truesize -= mss;
116 th->seq = htonl(seq);
120 /* Following permits TCP Small Queues to work well with GSO :
121 * The callback to TCP stack will be called at the time last frag
122 * is freed at TX completion, and not right now when gso_skb
123 * is freed by GSO engine
125 if (copy_destructor) {
126 swap(gso_skb->sk, skb->sk);
127 swap(gso_skb->destructor, skb->destructor);
128 swap(gso_skb->truesize, skb->truesize);
131 delta = htonl(oldlen + (skb_tail_pointer(skb) -
132 skb_transport_header(skb)) +
134 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
135 (__force u32)delta));
136 if (skb->ip_summed != CHECKSUM_PARTIAL)
137 th->check = csum_fold(csum_partial(skb_transport_header(skb),
142 EXPORT_SYMBOL(tcp_tso_segment);
144 struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
146 struct sk_buff **pp = NULL;
153 unsigned int mss = 1;
159 off = skb_gro_offset(skb);
160 hlen = off + sizeof(*th);
161 th = skb_gro_header_fast(skb, off);
162 if (skb_gro_header_hard(skb, hlen)) {
163 th = skb_gro_header_slow(skb, hlen, off);
168 thlen = th->doff * 4;
169 if (thlen < sizeof(*th))
173 if (skb_gro_header_hard(skb, hlen)) {
174 th = skb_gro_header_slow(skb, hlen, off);
179 skb_gro_pull(skb, thlen);
181 len = skb_gro_len(skb);
182 flags = tcp_flag_word(th);
184 for (; (p = *head); head = &p->next) {
185 if (!NAPI_GRO_CB(p)->same_flow)
190 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
191 NAPI_GRO_CB(p)->same_flow = 0;
198 goto out_check_final;
201 flush = NAPI_GRO_CB(p)->flush;
202 flush |= (__force int)(flags & TCP_FLAG_CWR);
203 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
204 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
205 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
206 for (i = sizeof(*th); i < thlen; i += 4)
207 flush |= *(u32 *)((u8 *)th + i) ^
208 *(u32 *)((u8 *)th2 + i);
210 mss = tcp_skb_mss(p);
212 flush |= (len - 1) >= mss;
213 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
215 if (flush || skb_gro_receive(head, skb)) {
217 goto out_check_final;
222 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
226 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
227 TCP_FLAG_RST | TCP_FLAG_SYN |
230 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
234 NAPI_GRO_CB(skb)->flush |= flush;
238 EXPORT_SYMBOL(tcp_gro_receive);
240 int tcp_gro_complete(struct sk_buff *skb)
242 struct tcphdr *th = tcp_hdr(skb);
244 skb->csum_start = skb_transport_header(skb) - skb->head;
245 skb->csum_offset = offsetof(struct tcphdr, check);
246 skb->ip_summed = CHECKSUM_PARTIAL;
248 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
251 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
255 EXPORT_SYMBOL(tcp_gro_complete);
257 static int tcp_v4_gso_send_check(struct sk_buff *skb)
259 const struct iphdr *iph;
262 if (!pskb_may_pull(skb, sizeof(*th)))
269 skb->ip_summed = CHECKSUM_PARTIAL;
270 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
274 static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
276 const struct iphdr *iph = skb_gro_network_header(skb);
280 switch (skb->ip_summed) {
281 case CHECKSUM_COMPLETE:
282 if (!tcp_v4_check(skb_gro_len(skb), iph->saddr, iph->daddr,
284 skb->ip_summed = CHECKSUM_UNNECESSARY;
288 NAPI_GRO_CB(skb)->flush = 1;
292 wsum = csum_tcpudp_nofold(iph->saddr, iph->daddr,
293 skb_gro_len(skb), IPPROTO_TCP, 0);
294 sum = csum_fold(skb_checksum(skb,
301 skb->ip_summed = CHECKSUM_UNNECESSARY;
305 return tcp_gro_receive(head, skb);
308 static int tcp4_gro_complete(struct sk_buff *skb)
310 const struct iphdr *iph = ip_hdr(skb);
311 struct tcphdr *th = tcp_hdr(skb);
313 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
314 iph->saddr, iph->daddr, 0);
315 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
317 return tcp_gro_complete(skb);
320 static const struct net_offload tcpv4_offload = {
322 .gso_send_check = tcp_v4_gso_send_check,
323 .gso_segment = tcp_tso_segment,
324 .gro_receive = tcp4_gro_receive,
325 .gro_complete = tcp4_gro_complete,
329 int __init tcpv4_offload_init(void)
331 return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);