blob: 9d7930ba8e0f4c51c8d5434c5ea186b585ff3a35 [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
Willem de Bruijnf066e2b2014-08-06 15:09:44 -040017static void tcp_gso_tstamp(struct sk_buff *skb, unsigned int ts_seq,
18 unsigned int seq, unsigned int mss)
Willem de Bruijn4ed2d762014-08-04 22:11:49 -040019{
20 while (skb) {
Willem de Bruijnf066e2b2014-08-06 15:09:44 -040021 if (before(ts_seq, seq + mss)) {
22 skb_shinfo(skb)->tx_flags |= SKBTX_SW_TSTAMP;
Willem de Bruijn4ed2d762014-08-04 22:11:49 -040023 skb_shinfo(skb)->tskey = ts_seq;
24 return;
25 }
26
27 skb = skb->next;
28 seq += mss;
29 }
30}
31
Tom Herbertd020f8f2014-09-20 14:52:28 -070032struct sk_buff *tcp4_gso_segment(struct sk_buff *skb,
33 netdev_features_t features)
34{
35 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
36 return ERR_PTR(-EINVAL);
37
38 if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
39 const struct iphdr *iph = ip_hdr(skb);
40 struct tcphdr *th = tcp_hdr(skb);
41
42 /* Set up checksum pseudo header, usually expect stack to
43 * have done this already.
44 */
45
46 th->check = 0;
47 skb->ip_summed = CHECKSUM_PARTIAL;
48 __tcp_v4_send_check(skb, iph->saddr, iph->daddr);
49 }
50
51 return tcp_gso_segment(skb, features);
52}
53
Eric Dumazet28be6e02013-10-18 10:36:17 -070054struct sk_buff *tcp_gso_segment(struct sk_buff *skb,
Daniel Borkmann28850dc2013-06-07 05:11:46 +000055 netdev_features_t features)
56{
57 struct sk_buff *segs = ERR_PTR(-EINVAL);
Eric Dumazet0d08c422013-10-25 17:26:17 -070058 unsigned int sum_truesize = 0;
Daniel Borkmann28850dc2013-06-07 05:11:46 +000059 struct tcphdr *th;
60 unsigned int thlen;
61 unsigned int seq;
62 __be32 delta;
63 unsigned int oldlen;
64 unsigned int mss;
65 struct sk_buff *gso_skb = skb;
66 __sum16 newcheck;
67 bool ooo_okay, copy_destructor;
68
Daniel Borkmann28850dc2013-06-07 05:11:46 +000069 th = tcp_hdr(skb);
70 thlen = th->doff * 4;
71 if (thlen < sizeof(*th))
72 goto out;
73
74 if (!pskb_may_pull(skb, thlen))
75 goto out;
76
77 oldlen = (u16)~skb->len;
78 __skb_pull(skb, thlen);
79
80 mss = tcp_skb_mss(skb);
81 if (unlikely(skb->len <= mss))
82 goto out;
83
84 if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
85 /* Packet is from an untrusted source, reset gso_segs. */
86 int type = skb_shinfo(skb)->gso_type;
87
88 if (unlikely(type &
89 ~(SKB_GSO_TCPV4 |
90 SKB_GSO_DODGY |
91 SKB_GSO_TCP_ECN |
92 SKB_GSO_TCPV6 |
93 SKB_GSO_GRE |
Tom Herbert4749c092014-06-04 17:20:23 -070094 SKB_GSO_GRE_CSUM |
Eric Dumazetcb32f512013-10-19 11:42:57 -070095 SKB_GSO_IPIP |
Eric Dumazet61c1db72013-10-20 20:47:30 -070096 SKB_GSO_SIT |
Daniel Borkmann28850dc2013-06-07 05:11:46 +000097 SKB_GSO_UDP_TUNNEL |
Tom Herbert0f4f4ff2014-06-04 17:20:16 -070098 SKB_GSO_UDP_TUNNEL_CSUM |
Tom Herberte585f232014-11-04 09:06:54 -080099 SKB_GSO_TUNNEL_REMCSUM |
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000100 0) ||
101 !(type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))))
102 goto out;
103
104 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
105
106 segs = NULL;
107 goto out;
108 }
109
110 copy_destructor = gso_skb->destructor == tcp_wfree;
111 ooo_okay = gso_skb->ooo_okay;
112 /* All segments but the first should have ooo_okay cleared */
113 skb->ooo_okay = 0;
114
115 segs = skb_segment(skb, features);
116 if (IS_ERR(segs))
117 goto out;
118
119 /* Only first segment might have ooo_okay set */
120 segs->ooo_okay = ooo_okay;
121
122 delta = htonl(oldlen + (thlen + mss));
123
124 skb = segs;
125 th = tcp_hdr(skb);
126 seq = ntohl(th->seq);
127
Willem de Bruijn4ed2d762014-08-04 22:11:49 -0400128 if (unlikely(skb_shinfo(gso_skb)->tx_flags & SKBTX_SW_TSTAMP))
129 tcp_gso_tstamp(segs, skb_shinfo(gso_skb)->tskey, seq, mss);
130
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000131 newcheck = ~csum_fold((__force __wsum)((__force u32)th->check +
132 (__force u32)delta));
133
134 do {
135 th->fin = th->psh = 0;
136 th->check = newcheck;
137
138 if (skb->ip_summed != CHECKSUM_PARTIAL)
Tom Herberte9c3a242014-06-04 17:20:09 -0700139 th->check = gso_make_checksum(skb, ~th->check);
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000140
141 seq += mss;
142 if (copy_destructor) {
143 skb->destructor = gso_skb->destructor;
144 skb->sk = gso_skb->sk;
Eric Dumazet0d08c422013-10-25 17:26:17 -0700145 sum_truesize += skb->truesize;
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000146 }
147 skb = skb->next;
148 th = tcp_hdr(skb);
149
150 th->seq = htonl(seq);
151 th->cwr = 0;
152 } while (skb->next);
153
154 /* Following permits TCP Small Queues to work well with GSO :
155 * The callback to TCP stack will be called at the time last frag
156 * is freed at TX completion, and not right now when gso_skb
157 * is freed by GSO engine
158 */
159 if (copy_destructor) {
160 swap(gso_skb->sk, skb->sk);
161 swap(gso_skb->destructor, skb->destructor);
Eric Dumazet0d08c422013-10-25 17:26:17 -0700162 sum_truesize += skb->truesize;
163 atomic_add(sum_truesize - gso_skb->truesize,
164 &skb->sk->sk_wmem_alloc);
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000165 }
166
167 delta = htonl(oldlen + (skb_tail_pointer(skb) -
168 skb_transport_header(skb)) +
169 skb->data_len);
170 th->check = ~csum_fold((__force __wsum)((__force u32)th->check +
171 (__force u32)delta));
172 if (skb->ip_summed != CHECKSUM_PARTIAL)
Tom Herberte9c3a242014-06-04 17:20:09 -0700173 th->check = gso_make_checksum(skb, ~th->check);
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000174out:
175 return segs;
176}
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000177
178struct sk_buff **tcp_gro_receive(struct sk_buff **head, struct sk_buff *skb)
179{
180 struct sk_buff **pp = NULL;
181 struct sk_buff *p;
182 struct tcphdr *th;
183 struct tcphdr *th2;
184 unsigned int len;
185 unsigned int thlen;
186 __be32 flags;
187 unsigned int mss = 1;
188 unsigned int hlen;
189 unsigned int off;
190 int flush = 1;
191 int i;
192
193 off = skb_gro_offset(skb);
194 hlen = off + sizeof(*th);
195 th = skb_gro_header_fast(skb, off);
196 if (skb_gro_header_hard(skb, hlen)) {
197 th = skb_gro_header_slow(skb, hlen, off);
198 if (unlikely(!th))
199 goto out;
200 }
201
202 thlen = th->doff * 4;
203 if (thlen < sizeof(*th))
204 goto out;
205
206 hlen = off + thlen;
207 if (skb_gro_header_hard(skb, hlen)) {
208 th = skb_gro_header_slow(skb, hlen, off);
209 if (unlikely(!th))
210 goto out;
211 }
212
213 skb_gro_pull(skb, thlen);
214
215 len = skb_gro_len(skb);
216 flags = tcp_flag_word(th);
217
218 for (; (p = *head); head = &p->next) {
219 if (!NAPI_GRO_CB(p)->same_flow)
220 continue;
221
222 th2 = tcp_hdr(p);
223
224 if (*(u32 *)&th->source ^ *(u32 *)&th2->source) {
225 NAPI_GRO_CB(p)->same_flow = 0;
226 continue;
227 }
228
229 goto found;
230 }
231
232 goto out_check_final;
233
234found:
Jerry Chubf5a7552014-01-07 10:23:19 -0800235 /* Include the IP ID check below from the inner most IP hdr */
236 flush = NAPI_GRO_CB(p)->flush | NAPI_GRO_CB(p)->flush_id;
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000237 flush |= (__force int)(flags & TCP_FLAG_CWR);
238 flush |= (__force int)((flags ^ tcp_flag_word(th2)) &
239 ~(TCP_FLAG_CWR | TCP_FLAG_FIN | TCP_FLAG_PSH));
240 flush |= (__force int)(th->ack_seq ^ th2->ack_seq);
241 for (i = sizeof(*th); i < thlen; i += 4)
242 flush |= *(u32 *)((u8 *)th + i) ^
243 *(u32 *)((u8 *)th2 + i);
244
245 mss = tcp_skb_mss(p);
246
247 flush |= (len - 1) >= mss;
248 flush |= (ntohl(th2->seq) + skb_gro_len(p)) ^ ntohl(th->seq);
249
250 if (flush || skb_gro_receive(head, skb)) {
251 mss = 1;
252 goto out_check_final;
253 }
254
255 p = *head;
256 th2 = tcp_hdr(p);
257 tcp_flag_word(th2) |= flags & (TCP_FLAG_FIN | TCP_FLAG_PSH);
258
259out_check_final:
260 flush = len < mss;
261 flush |= (__force int)(flags & (TCP_FLAG_URG | TCP_FLAG_PSH |
262 TCP_FLAG_RST | TCP_FLAG_SYN |
263 TCP_FLAG_FIN));
264
265 if (p && (!NAPI_GRO_CB(skb)->same_flow || flush))
266 pp = head;
267
268out:
Jerry Chubf5a7552014-01-07 10:23:19 -0800269 NAPI_GRO_CB(skb)->flush |= (flush != 0);
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000270
271 return pp;
272}
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000273
274int tcp_gro_complete(struct sk_buff *skb)
275{
276 struct tcphdr *th = tcp_hdr(skb);
277
Jerry Chu299603e82013-12-11 20:53:45 -0800278 skb->csum_start = (unsigned char *)th - skb->head;
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000279 skb->csum_offset = offsetof(struct tcphdr, check);
280 skb->ip_summed = CHECKSUM_PARTIAL;
281
282 skb_shinfo(skb)->gso_segs = NAPI_GRO_CB(skb)->count;
283
284 if (th->cwr)
285 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
286
287 return 0;
288}
289EXPORT_SYMBOL(tcp_gro_complete);
290
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000291static struct sk_buff **tcp4_gro_receive(struct sk_buff **head, struct sk_buff *skb)
292{
Herbert Xucc5c00b2013-11-22 10:31:29 +0800293 /* Don't bother verifying checksum if we're going to flush anyway. */
Tom Herbert149d0772014-08-22 13:34:30 -0700294 if (!NAPI_GRO_CB(skb)->flush &&
295 skb_gro_checksum_validate(skb, IPPROTO_TCP,
296 inet_gro_compute_pseudo)) {
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000297 NAPI_GRO_CB(skb)->flush = 1;
298 return NULL;
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000299 }
300
301 return tcp_gro_receive(head, skb);
302}
303
Jerry Chu299603e82013-12-11 20:53:45 -0800304static int tcp4_gro_complete(struct sk_buff *skb, int thoff)
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000305{
306 const struct iphdr *iph = ip_hdr(skb);
307 struct tcphdr *th = tcp_hdr(skb);
308
Jerry Chu299603e82013-12-11 20:53:45 -0800309 th->check = ~tcp_v4_check(skb->len - thoff, iph->saddr,
310 iph->daddr, 0);
Jerry Chuc3caf112014-07-14 15:54:46 -0700311 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000312
313 return tcp_gro_complete(skb);
314}
315
316static const struct net_offload tcpv4_offload = {
317 .callbacks = {
Tom Herbertd020f8f2014-09-20 14:52:28 -0700318 .gso_segment = tcp4_gso_segment,
Daniel Borkmann28850dc2013-06-07 05:11:46 +0000319 .gro_receive = tcp4_gro_receive,
320 .gro_complete = tcp4_gro_complete,
321 },
322};
323
324int __init tcpv4_offload_init(void)
325{
326 return inet_add_offload(&tcpv4_offload, IPPROTO_TCP);
327}