blob: 55aeec93014023ba08b6126230b8a73e1aebf032 [file] [log] [blame]
Daniel Borkmann28850dc2013-06-07 05:11:46 +00001/*
2 * IPV4 GSO/GRO offload support
3 * Linux INET implementation
4 *
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.
9 *
10 * TCPv4 GSO/GRO support
11 */
12
13#include <linux/skbuff.h>
14#include <net/tcp.h>
15#include <net/protocol.h>
16
Eric Dumazet28be6e02013-10-18 10:36:17 -070017struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
Daniel Borkmann28850dc2013-06-07 05:11:46 +000018 netdev_features_t features)
19{
20 struct sk_buff *segs = ERR_PTR(-EINVAL);
Eric Dumazet0d08c422013-10-25 17:26:17 -070021 unsigned int sum_truesize = 0;
Daniel Borkmann28850dc2013-06-07 05:11:46 +000022 struct tcphdr *th;
23 unsigned int thlen;
24 unsigned int seq;
25 __be32 delta;
26 unsigned int oldlen;
27 unsigned int mss;
28 struct sk_buff *gso_skb = skb;
29 __sum16 newcheck;
30 bool ooo_okay, copy_destructor;
31
32 if (!pskb_may_pull(skb, sizeof(*th)))
33 goto out;
34
35 th = tcp_hdr(skb);
36 thlen = th->doff * 4;
37 if (thlen < sizeof(*th))
38 goto out;
39
40 if (!pskb_may_pull(skb, thlen))
41 goto out;
42
43 oldlen = (u16)~skb->len;
44 __skb_pull(skb, thlen);
45
46 mss = tcp_skb_mss(skb);
47 if (unlikely(skb->len <= mss))
48 goto out;
49
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;
53
54 if (unlikely(type &
55 ~(SKB_GSO_TCPV4 |
56 SKB_GSO_DODGY |
57 SKB_GSO_TCP_ECN |
58 SKB_GSO_TCPV6 |
59 SKB_GSO_GRE |
Eric Dumazetcb32f512013-10-19 11:42:57 -070060 SKB_GSO_IPIP |
Eric Dumazet61c1db72013-10-20 20:47:30 -070061 SKB_GSO_SIT |
Daniel Borkmann28850dc2013-06-07 05:11:46 +000062 SKB_GSO_MPLS |
63 SKB_GSO_UDP_TUNNEL |
64 0) ||
65 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
66 goto out;
67
68 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
69
70 segs = NULL;
71 goto out;
72 }
73
74 copy_destructor = gso_skb->destructor == tcp_wfree;
75 ooo_okay = gso_skb->ooo_okay;
76 /* All segments but the first should have ooo_okay cleared */
77 skb->ooo_okay = 0;
78
79 segs = skb_segment(skb, features);
80 if (IS_ERR(segs))
81 goto out;
82
83 /* Only first segment might have ooo_okay set */
84 segs->ooo_okay = ooo_okay;
85
86 delta = htonl(oldlen + (thlen + mss));
87
88 skb = segs;
89 th = tcp_hdr(skb);
90 seq = ntohl(th->seq);
91
92 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
93 (__force u32)delta));
94
95 do {
96 th->fin = th->psh = 0;
97 th->check = newcheck;
98
99 if (skb->ip_summed != CHECKSUM_PARTIAL)
100 th->check =
101 csum_fold(csum_partial(skb_transport_header(skb),
102 thlen, skb->csum));
103
104 seq += mss;
105 if (copy_destructor) {
106 skb->destructor = gso_skb->destructor;
107 skb->sk = gso_skb->sk;
Eric Dumazet0d08c422013-10-25 17:26:17 -0700108 sum_truesize += skb->truesize;
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000109 }
110 skb = skb->next;
111 th = tcp_hdr(skb);
112
113 th->seq = htonl(seq);
114 th->cwr = 0;
115 } while (skb->next);
116
117 /* Following permits TCP Small Queues to work well with GSO :
118 * The callback to TCP stack will be called at the time last frag
119 * is freed at TX completion, and not right now when gso_skb
120 * is freed by GSO engine
121 */
122 if (copy_destructor) {
123 swap(gso_skb->sk, skb->sk);
124 swap(gso_skb->destructor, skb->destructor);
Eric Dumazet0d08c422013-10-25 17:26:17 -0700125 sum_truesize += skb->truesize;
126 atomic_add(sum_truesize - gso_skb->truesize,
127 &skb->sk->sk_wmem_alloc);
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000128 }
129
130 delta = htonl(oldlen + (skb_tail_pointer(skb) -
131 skb_transport_header(skb)) +
132 skb->data_len);
133 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
134 (__force u32)delta));
135 if (skb->ip_summed != CHECKSUM_PARTIAL)
136 th->check = csum_fold(csum_partial(skb_transport_header(skb),
137 thlen, skb->csum));
138out:
139 return segs;
140}
Eric Dumazet28be6e02013-10-18 10:36:17 -0700141EXPORT_SYMBOL(tcp_gso_segment);
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000142
143struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
144{
145 struct sk_buff **pp = NULL;
146 struct sk_buff *p;
147 struct tcphdr *th;
148 struct tcphdr *th2;
149 unsigned int len;
150 unsigned int thlen;
151 __be32 flags;
152 unsigned int mss = 1;
153 unsigned int hlen;
154 unsigned int off;
155 int flush = 1;
156 int i;
157
158 off = skb_gro_offset(skb);
159 hlen = off + sizeof(*th);
160 th = skb_gro_header_fast(skb, off);
161 if (skb_gro_header_hard(skb, hlen)) {
162 th = skb_gro_header_slow(skb, hlen, off);
163 if (unlikely(!th))
164 goto out;
165 }
166
167 thlen = th->doff * 4;
168 if (thlen < sizeof(*th))
169 goto out;
170
171 hlen = off + thlen;
172 if (skb_gro_header_hard(skb, hlen)) {
173 th = skb_gro_header_slow(skb, hlen, off);
174 if (unlikely(!th))
175 goto out;
176 }
177
178 skb_gro_pull(skb, thlen);
179
180 len = skb_gro_len(skb);
181 flags = tcp_flag_word(th);
182
183 for (; (p = *head); head = &p->next) {
184 if (!NAPI_GRO_CB(p)->same_flow)
185 continue;
186
187 th2 = tcp_hdr(p);
188
189 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
190 NAPI_GRO_CB(p)->same_flow = 0;
191 continue;
192 }
193
194 goto found;
195 }
196
197 goto out_check_final;
198
199found:
200 flush = NAPI_GRO_CB(p)->flush;
201 flush |= (__force int)(flags & TCP_FLAG_CWR);
202 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
203 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
204 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
205 for (i = sizeof(*th); i < thlen; i += 4)
206 flush |= *(u32 *)((u8 *)th + i) ^
207 *(u32 *)((u8 *)th2 + i);
208
209 mss = tcp_skb_mss(p);
210
211 flush |= (len - 1) >= mss;
212 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
213
214 if (flush || skb_gro_receive(head, skb)) {
215 mss = 1;
216 goto out_check_final;
217 }
218
219 p = *head;
220 th2 = tcp_hdr(p);
221 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
222
223out_check_final:
224 flush = len < mss;
225 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
226 TCP_FLAG_RST | TCP_FLAG_SYN |
227 TCP_FLAG_FIN));
228
229 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
230 pp = head;
231
232out:
233 NAPI_GRO_CB(skb)->flush |= flush;
234
235 return pp;
236}
237EXPORT_SYMBOL(tcp_gro_receive);
238
239int tcp_gro_complete(struct sk_buff *skb)
240{
241 struct tcphdr *th = tcp_hdr(skb);
242
243 skb->csum_start = skb_transport_header(skb) - skb->head;
244 skb->csum_offset = offsetof(struct tcphdr, check);
245 skb->ip_summed = CHECKSUM_PARTIAL;
246
247 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
248
249 if (th->cwr)
250 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
251
252 return 0;
253}
254EXPORT_SYMBOL(tcp_gro_complete);
255
256static int tcp_v4_gso_send_check(struct sk_buff *skb)
257{
258 const struct iphdr *iph;
259 struct tcphdr *th;
260
261 if (!pskb_may_pull(skb, sizeof(*th)))
262 return -EINVAL;
263
264 iph = ip_hdr(skb);
265 th = tcp_hdr(skb);
266
267 th->check = 0;
268 skb->ip_summed = CHECKSUM_PARTIAL;
269 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
270 return 0;
271}
272
273static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
274{
275 const struct iphdr *iph = skb_gro_network_header(skb);
276 __wsum wsum;
277 __sum16 sum;
278
Herbert Xucc5c00b2013-11-22 10:31:29 +0800279 /* Don't bother verifying checksum if we're going to flush anyway. */
280 if (NAPI_GRO_CB(skb)->flush)
281 goto skip_csum;
282
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000283 switch (skb->ip_summed) {
284 case CHECKSUM_COMPLETE:
285 if (!tcp_v4_check(skb_gro_len(skb), iph->saddr, iph->daddr,
286 skb->csum)) {
287 skb->ip_summed = CHECKSUM_UNNECESSARY;
288 break;
289 }
290flush:
291 NAPI_GRO_CB(skb)->flush = 1;
292 return NULL;
293
294 case CHECKSUM_NONE:
295 wsum = csum_tcpudp_nofold(iph->saddr, iph->daddr,
296 skb_gro_len(skb), IPPROTO_TCP, 0);
297 sum = csum_fold(skb_checksum(skb,
298 skb_gro_offset(skb),
299 skb_gro_len(skb),
300 wsum));
301 if (sum)
302 goto flush;
303
304 skb->ip_summed = CHECKSUM_UNNECESSARY;
305 break;
306 }
307
Herbert Xucc5c00b2013-11-22 10:31:29 +0800308skip_csum:
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000309 return tcp_gro_receive(head, skb);
310}
311
312static int tcp4_gro_complete(struct sk_buff *skb)
313{
314 const struct iphdr *iph = ip_hdr(skb);
315 struct tcphdr *th = tcp_hdr(skb);
316
317 th->check = ~tcp_v4_check(skb->len - skb_transport_offset(skb),
318 iph->saddr, iph->daddr, 0);
319 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
320
321 return tcp_gro_complete(skb);
322}
323
324static const struct net_offload tcpv4_offload = {
325 .callbacks = {
326 .gso_send_check = tcp_v4_gso_send_check,
Eric Dumazet28be6e02013-10-18 10:36:17 -0700327 .gso_segment = tcp_gso_segment,
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000328 .gro_receive = tcp4_gro_receive,
329 .gro_complete = tcp4_gro_complete,
330 },
331};
332
333int __init tcpv4_offload_init(void)
334{
335 return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
336}