blob: a3679174e78a624c217797392edcab95fd9aded0 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
9 *
Jesper Juhl02c30a82005-05-05 16:16:16 -070010 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
21 */
22
23/*
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
27 * : AF independence
28 *
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
36 *
37 */
38
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
42#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070045int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Rick Jones15d99e02006-03-20 22:40:29 -080047/* People can turn this on to work with those rare, broken TCPs that
48 * interpret the window field as a signed quantity.
49 */
Brian Haleyab32ea52006-09-22 14:15:41 -070050int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080051
Linus Torvalds1da177e2005-04-16 15:20:36 -070052/* This limits the percentage of the congestion window which we
53 * will allow a single TSO frame to consume. Building TSO frames
54 * which are too large can cause TCP streams to be bursty.
55 */
Brian Haleyab32ea52006-09-22 14:15:41 -070056int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070057
Brian Haleyab32ea52006-09-22 14:15:41 -070058int sysctl_tcp_mtu_probing __read_mostly = 0;
59int sysctl_tcp_base_mss __read_mostly = 512;
John Heffner5d424d52006-03-20 17:53:41 -080060
David S. Miller35089bb2006-06-13 22:33:04 -070061/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070062int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070063
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070064static void update_send_head(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -070065{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070066 struct tcp_sock *tp = tcp_sk(sk);
67
David S. Millerfe067e82007-03-07 12:12:44 -080068 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070070 tcp_packets_out_inc(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070071}
72
73/* SND.NXT, if window was not shrunk.
74 * If window has been shrunk, what should we make? It is not clear at all.
75 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
76 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
77 * invalid. OK, let's make this for now:
78 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070079static inline __u32 tcp_acceptable_seq(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070080{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070081 struct tcp_sock *tp = tcp_sk(sk);
82
Linus Torvalds1da177e2005-04-16 15:20:36 -070083 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
84 return tp->snd_nxt;
85 else
86 return tp->snd_una+tp->snd_wnd;
87}
88
89/* Calculate mss to advertise in SYN segment.
90 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
91 *
92 * 1. It is independent of path mtu.
93 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95 * attached devices, because some buggy hosts are confused by
96 * large MSS.
97 * 4. We do not make 3, we advertise MSS, calculated from first
98 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
99 * This may be overridden via information stored in routing table.
100 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101 * probably even Jumbo".
102 */
103static __u16 tcp_advertise_mss(struct sock *sk)
104{
105 struct tcp_sock *tp = tcp_sk(sk);
106 struct dst_entry *dst = __sk_dst_get(sk);
107 int mss = tp->advmss;
108
109 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
110 mss = dst_metric(dst, RTAX_ADVMSS);
111 tp->advmss = mss;
112 }
113
114 return (__u16)mss;
115}
116
117/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700119static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700121 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 s32 delta = tcp_time_stamp - tp->lsndtime;
123 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
124 u32 cwnd = tp->snd_cwnd;
125
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300126 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300128 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 restart_cwnd = min(restart_cwnd, cwnd);
130
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700131 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 cwnd >>= 1;
133 tp->snd_cwnd = max(cwnd, restart_cwnd);
134 tp->snd_cwnd_stamp = tcp_time_stamp;
135 tp->snd_cwnd_used = 0;
136}
137
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800138static void tcp_event_data_sent(struct tcp_sock *tp,
139 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700141 struct inet_connection_sock *icsk = inet_csk(sk);
142 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143
David S. Miller35089bb2006-06-13 22:33:04 -0700144 if (sysctl_tcp_slow_start_after_idle &&
145 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700146 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
148 tp->lsndtime = now;
149
150 /* If it is a reply for ato after last received
151 * packet, enter pingpong mode.
152 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700153 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
154 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155}
156
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800157static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700159 tcp_dec_quickack_mode(sk, pkts);
160 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161}
162
163/* Determine a window scaling and initial window to offer.
164 * Based on the assumption that the given amount of space
165 * will be offered. Store the results in the tp structure.
166 * NOTE: for smooth operation initial space offering should
167 * be a multiple of mss if possible. We assume here that mss >= 1.
168 * This MUST be enforced by all callers.
169 */
170void tcp_select_initial_window(int __space, __u32 mss,
171 __u32 *rcv_wnd, __u32 *window_clamp,
172 int wscale_ok, __u8 *rcv_wscale)
173{
174 unsigned int space = (__space < 0 ? 0 : __space);
175
176 /* If no clamp set the clamp to the max possible scaled window */
177 if (*window_clamp == 0)
178 (*window_clamp) = (65535 << 14);
179 space = min(*window_clamp, space);
180
181 /* Quantize space offering to a multiple of mss if possible. */
182 if (space > mss)
183 space = (space / mss) * mss;
184
185 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800186 * will break some buggy TCP stacks. If the admin tells us
187 * it is likely we could be speaking with such a buggy stack
188 * we will truncate our initial window offering to 32K-1
189 * unless the remote has sent us a window scaling option,
190 * which we interpret as a sign the remote TCP is not
191 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 */
Rick Jones15d99e02006-03-20 22:40:29 -0800193 if (sysctl_tcp_workaround_signed_windows)
194 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
195 else
196 (*rcv_wnd) = space;
197
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 (*rcv_wscale) = 0;
199 if (wscale_ok) {
200 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900201 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 */
203 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700204 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 while (space > 65535 && (*rcv_wscale) < 14) {
206 space >>= 1;
207 (*rcv_wscale)++;
208 }
209 }
210
211 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700212 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 * will be satisfied with 2.
214 */
215 if (mss > (1<<*rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700216 int init_cwnd = 4;
217 if (mss > 1460*3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700219 else if (mss > 1460)
220 init_cwnd = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 if (*rcv_wnd > init_cwnd*mss)
222 *rcv_wnd = init_cwnd*mss;
223 }
224
225 /* Set the clamp no higher than max representable value */
226 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
227}
228
229/* Chose a new window to advertise, update state in tcp_sock for the
230 * socket, and return result with RFC1323 scaling applied. The return
231 * value can be stuffed directly into th->window for an outgoing
232 * frame.
233 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800234static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235{
236 struct tcp_sock *tp = tcp_sk(sk);
237 u32 cur_win = tcp_receive_window(tp);
238 u32 new_win = __tcp_select_window(sk);
239
240 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800241 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 /* Danger Will Robinson!
243 * Don't update rcv_wup/rcv_wnd here or else
244 * we will not be able to advertise a zero
245 * window in time. --DaveM
246 *
247 * Relax Will Robinson.
248 */
249 new_win = cur_win;
250 }
251 tp->rcv_wnd = new_win;
252 tp->rcv_wup = tp->rcv_nxt;
253
254 /* Make sure we do not exceed the maximum possible
255 * scaled window.
256 */
Rick Jones15d99e02006-03-20 22:40:29 -0800257 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 new_win = min(new_win, MAX_TCP_WINDOW);
259 else
260 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
261
262 /* RFC1323 scaling applied */
263 new_win >>= tp->rx_opt.rcv_wscale;
264
265 /* If we advertise zero window, disable fast path. */
266 if (new_win == 0)
267 tp->pred_flags = 0;
268
269 return new_win;
270}
271
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700272static inline void TCP_ECN_send_synack(struct tcp_sock *tp,
273 struct sk_buff *skb)
274{
275 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
276 if (!(tp->ecn_flags&TCP_ECN_OK))
277 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
278}
279
280static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
281{
282 struct tcp_sock *tp = tcp_sk(sk);
283
284 tp->ecn_flags = 0;
285 if (sysctl_tcp_ecn) {
286 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR;
287 tp->ecn_flags = TCP_ECN_OK;
288 }
289}
290
291static __inline__ void
292TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
293{
294 if (inet_rsk(req)->ecn_ok)
295 th->ece = 1;
296}
297
298static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
299 int tcp_header_len)
300{
301 struct tcp_sock *tp = tcp_sk(sk);
302
303 if (tp->ecn_flags & TCP_ECN_OK) {
304 /* Not-retransmitted data segment: set ECT and inject CWR. */
305 if (skb->len != tcp_header_len &&
306 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
307 INET_ECN_xmit(sk);
308 if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) {
309 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
310 tcp_hdr(skb)->cwr = 1;
311 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
312 }
313 } else {
314 /* ACK or retransmitted segment: clear ECT|CE */
315 INET_ECN_dontxmit(sk);
316 }
317 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
318 tcp_hdr(skb)->ece = 1;
319 }
320}
321
Al Virodf7a3b02006-09-27 18:38:52 -0700322static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800323 __u32 tstamp, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800324{
325 if (tp->rx_opt.tstamp_ok) {
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700326 *ptr++ = htonl((TCPOPT_NOP << 24) |
327 (TCPOPT_NOP << 16) |
328 (TCPOPT_TIMESTAMP << 8) |
329 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800330 *ptr++ = htonl(tstamp);
331 *ptr++ = htonl(tp->rx_opt.ts_recent);
332 }
333 if (tp->rx_opt.eff_sacks) {
334 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
335 int this_sack;
336
337 *ptr++ = htonl((TCPOPT_NOP << 24) |
338 (TCPOPT_NOP << 16) |
339 (TCPOPT_SACK << 8) |
340 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
341 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800342
343 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800344 *ptr++ = htonl(sp[this_sack].start_seq);
345 *ptr++ = htonl(sp[this_sack].end_seq);
346 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800347
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800348 if (tp->rx_opt.dsack) {
349 tp->rx_opt.dsack = 0;
350 tp->rx_opt.eff_sacks--;
351 }
352 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800353#ifdef CONFIG_TCP_MD5SIG
354 if (md5_hash) {
355 *ptr++ = htonl((TCPOPT_NOP << 24) |
356 (TCPOPT_NOP << 16) |
357 (TCPOPT_MD5SIG << 8) |
358 TCPOLEN_MD5SIG);
359 *md5_hash = (__u8 *)ptr;
360 }
361#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800362}
363
364/* Construct a tcp options header for a SYN or SYN_ACK packet.
365 * If this is every changed make sure to change the definition of
366 * MAX_SYN_SIZE to match the new maximum number of options that you
367 * can generate.
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800368 *
369 * Note - that with the RFC2385 TCP option, we make room for the
370 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
371 * location to be filled is passed back up.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800372 */
Al Virodf7a3b02006-09-27 18:38:52 -0700373static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800374 int offer_wscale, int wscale, __u32 tstamp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800375 __u32 ts_recent, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800376{
377 /* We always get an MSS option.
378 * The option bytes which will be seen in normal data
379 * packets should timestamps be used, must be in the MSS
380 * advertised. But we subtract them from tp->mss_cache so
381 * that calculations in tcp_sendmsg are simpler etc.
382 * So account for this fact here if necessary. If we
383 * don't do this correctly, as a receiver we won't
384 * recognize data packets as being full sized when we
385 * should, and thus we won't abide by the delayed ACK
386 * rules correctly.
387 * SACKs don't matter, we never delay an ACK when we
388 * have any of those going out.
389 */
390 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
391 if (ts) {
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800392 if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700393 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
394 (TCPOLEN_SACK_PERM << 16) |
395 (TCPOPT_TIMESTAMP << 8) |
396 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800397 else
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700398 *ptr++ = htonl((TCPOPT_NOP << 24) |
399 (TCPOPT_NOP << 16) |
400 (TCPOPT_TIMESTAMP << 8) |
401 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800402 *ptr++ = htonl(tstamp); /* TSVAL */
403 *ptr++ = htonl(ts_recent); /* TSECR */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800404 } else if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700405 *ptr++ = htonl((TCPOPT_NOP << 24) |
406 (TCPOPT_NOP << 16) |
407 (TCPOPT_SACK_PERM << 8) |
408 TCPOLEN_SACK_PERM);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800409 if (offer_wscale)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700410 *ptr++ = htonl((TCPOPT_NOP << 24) |
411 (TCPOPT_WINDOW << 16) |
412 (TCPOLEN_WINDOW << 8) |
413 (wscale));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800414#ifdef CONFIG_TCP_MD5SIG
415 /*
416 * If MD5 is enabled, then we set the option, and include the size
417 * (always 18). The actual MD5 hash is added just before the
418 * packet is sent.
419 */
420 if (md5_hash) {
421 *ptr++ = htonl((TCPOPT_NOP << 24) |
422 (TCPOPT_NOP << 16) |
423 (TCPOPT_MD5SIG << 8) |
424 TCPOLEN_MD5SIG);
425 *md5_hash = (__u8 *) ptr;
426 }
427#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800428}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
430/* This routine actually transmits TCP packets queued in by
431 * tcp_do_sendmsg(). This is used by both the initial
432 * transmission and possible later retransmissions.
433 * All SKB's seen here are completely headerless. It is our
434 * job to build the TCP header, and pass the packet down to
435 * IP so it can do the same plus pass the packet off to the
436 * device.
437 *
438 * We are working here with either a clone of the original
439 * SKB, or a fresh unique copy made by the retransmit engine.
440 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800441static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800443 const struct inet_connection_sock *icsk = inet_csk(sk);
444 struct inet_sock *inet;
445 struct tcp_sock *tp;
446 struct tcp_skb_cb *tcb;
447 int tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800448#ifdef CONFIG_TCP_MD5SIG
449 struct tcp_md5sig_key *md5;
450 __u8 *md5_hash_location;
451#endif
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800452 struct tcphdr *th;
453 int sysctl_flags;
454 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800456 BUG_ON(!skb || !tcp_skb_pcount(skb));
457
458 /* If congestion control is doing timestamping, we must
459 * take such a timestamp before we potentially clone/copy.
460 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700461 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800462 __net_timestamp(skb);
463
464 if (likely(clone_it)) {
465 if (unlikely(skb_cloned(skb)))
466 skb = pskb_copy(skb, gfp_mask);
467 else
468 skb = skb_clone(skb, gfp_mask);
469 if (unlikely(!skb))
470 return -ENOBUFS;
471 }
472
473 inet = inet_sk(sk);
474 tp = tcp_sk(sk);
475 tcb = TCP_SKB_CB(skb);
476 tcp_header_size = tp->tcp_header_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477
478#define SYSCTL_FLAG_TSTAMPS 0x1
479#define SYSCTL_FLAG_WSCALE 0x2
480#define SYSCTL_FLAG_SACK 0x4
481
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800482 sysctl_flags = 0;
483 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
484 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800485 if (sysctl_tcp_timestamps) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800486 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
487 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800489 if (sysctl_tcp_window_scaling) {
490 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
491 sysctl_flags |= SYSCTL_FLAG_WSCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800493 if (sysctl_tcp_sack) {
494 sysctl_flags |= SYSCTL_FLAG_SACK;
495 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
496 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800498 } else if (unlikely(tp->rx_opt.eff_sacks)) {
499 /* A SACK is 2 pad bytes, a 2 byte header, plus
500 * 2 32-bit sequence numbers for each SACK block.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800502 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
503 (tp->rx_opt.eff_sacks *
504 TCPOLEN_SACK_PERBLOCK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900506
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800507 if (tcp_packets_in_flight(tp) == 0)
508 tcp_ca_event(sk, CA_EVENT_TX_START);
509
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800510#ifdef CONFIG_TCP_MD5SIG
511 /*
512 * Are we doing MD5 on this segment? If so - make
513 * room for it.
514 */
515 md5 = tp->af_specific->md5_lookup(sk, sk);
516 if (md5)
517 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
518#endif
519
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700520 skb_push(skb, tcp_header_size);
521 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800522 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800523
524 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700525 th = tcp_hdr(skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800526 th->source = inet->sport;
527 th->dest = inet->dport;
528 th->seq = htonl(tcb->seq);
529 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700530 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800531 tcb->flags);
532
533 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
534 /* RFC1323: The window in SYN & SYN/ACK segments
535 * is never scaled.
536 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800537 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800538 } else {
539 th->window = htons(tcp_select_window(sk));
540 }
541 th->check = 0;
542 th->urg_ptr = 0;
543
544 if (unlikely(tp->urg_mode &&
545 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
546 th->urg_ptr = htons(tp->snd_up-tcb->seq);
547 th->urg = 1;
548 }
549
550 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
Al Virodf7a3b02006-09-27 18:38:52 -0700551 tcp_syn_build_options((__be32 *)(th + 1),
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800552 tcp_advertise_mss(sk),
553 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
554 (sysctl_flags & SYSCTL_FLAG_SACK),
555 (sysctl_flags & SYSCTL_FLAG_WSCALE),
556 tp->rx_opt.rcv_wscale,
557 tcb->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800558 tp->rx_opt.ts_recent,
559
560#ifdef CONFIG_TCP_MD5SIG
561 md5 ? &md5_hash_location :
562#endif
563 NULL);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800564 } else {
Al Virodf7a3b02006-09-27 18:38:52 -0700565 tcp_build_and_update_options((__be32 *)(th + 1),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800566 tp, tcb->when,
567#ifdef CONFIG_TCP_MD5SIG
568 md5 ? &md5_hash_location :
569#endif
570 NULL);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700571 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800572 }
573
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800574#ifdef CONFIG_TCP_MD5SIG
575 /* Calculate the MD5 hash, as we have all we need now */
576 if (md5) {
577 tp->af_specific->calc_md5_hash(md5_hash_location,
578 md5,
579 sk, NULL, NULL,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700580 tcp_hdr(skb),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800581 sk->sk_protocol,
582 skb->len);
583 }
584#endif
585
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800586 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800587
588 if (likely(tcb->flags & TCPCB_FLAG_ACK))
589 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
590
591 if (skb->len != tcp_header_size)
592 tcp_event_data_sent(tp, skb, sk);
593
Wei Yongjunbd37a082006-08-07 21:04:15 -0700594 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
595 TCP_INC_STATS(TCP_MIB_OUTSEGS);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800596
David S. Millere89862f2007-01-26 01:04:55 -0800597 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700598 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800599 return err;
600
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800601 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800602
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200603 return net_xmit_eval(err);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800604
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605#undef SYSCTL_FLAG_TSTAMPS
606#undef SYSCTL_FLAG_WSCALE
607#undef SYSCTL_FLAG_SACK
608}
609
610
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900611/* This routine just queue's the buffer
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 *
613 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
614 * otherwise socket can stall.
615 */
616static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
617{
618 struct tcp_sock *tp = tcp_sk(sk);
619
620 /* Advance write_seq and place onto the write_queue. */
621 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
622 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800623 tcp_add_write_queue_tail(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 sk_charge_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625}
626
David S. Miller846998a2005-08-04 19:52:01 -0700627static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700628{
Herbert Xubcd76112006-06-30 13:36:35 -0700629 if (skb->len <= mss_now || !sk_can_gso(sk)) {
David S. Millerf6302d12005-07-05 15:18:03 -0700630 /* Avoid the costly divide in the normal
631 * non-TSO case.
632 */
Herbert Xu79671682006-06-22 02:40:14 -0700633 skb_shinfo(skb)->gso_segs = 1;
634 skb_shinfo(skb)->gso_size = 0;
635 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700636 } else {
637 unsigned int factor;
638
David S. Miller846998a2005-08-04 19:52:01 -0700639 factor = skb->len + (mss_now - 1);
640 factor /= mss_now;
Herbert Xu79671682006-06-22 02:40:14 -0700641 skb_shinfo(skb)->gso_segs = factor;
642 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700643 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 }
645}
646
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647/* Function to create two new TCP segments. Shrinks the given segment
648 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900649 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 * Remember, these are still headerless SKBs at this point.
651 */
David S. Miller6475be12005-09-01 22:47:01 -0700652int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653{
654 struct tcp_sock *tp = tcp_sk(sk);
655 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700656 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700657 int nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 u16 flags;
659
Herbert Xub2cc99f2005-10-20 17:13:13 -0200660 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800661
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900662 clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 nsize = skb_headlen(skb) - len;
664 if (nsize < 0)
665 nsize = 0;
666
667 if (skb_cloned(skb) &&
668 skb_is_nonlinear(skb) &&
669 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
670 return -ENOMEM;
671
672 /* Get a new skb... force flag on. */
673 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
674 if (buff == NULL)
675 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -0700676
Herbert Xub60b49e2006-04-19 21:35:00 -0700677 sk_charge_skb(sk, buff);
678 nlen = skb->len - len - nsize;
679 buff->truesize += nlen;
680 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681
682 /* Correct the sequence numbers. */
683 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
684 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
685 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
686
687 /* PSH and FIN should only be set in the second packet. */
688 flags = TCP_SKB_CB(skb)->flags;
689 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
690 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700691 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
693
Patrick McHardy84fa7932006-08-29 16:44:56 -0700694 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 /* Copy and checksum data tail into the new buffer. */
696 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
697 nsize, 0);
698
699 skb_trim(skb, len);
700
701 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
702 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -0700703 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 skb_split(skb, buff, len);
705 }
706
707 buff->ip_summed = skb->ip_summed;
708
709 /* Looks stupid, but our code really uses when of
710 * skbs, which it never sent before. --ANK
711 */
712 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700713 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714
David S. Miller6475be12005-09-01 22:47:01 -0700715 old_factor = tcp_skb_pcount(skb);
716
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700718 tcp_set_skb_tso_segs(sk, skb, mss_now);
719 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720
David S. Miller6475be12005-09-01 22:47:01 -0700721 /* If this packet has been sent out already, we must
722 * adjust the various packet counters.
723 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700724 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700725 int diff = old_factor - tcp_skb_pcount(skb) -
726 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727
David S. Miller6475be12005-09-01 22:47:01 -0700728 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700729
730 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
731 tp->sacked_out -= diff;
732 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
733 tp->retrans_out -= diff;
734
Ilpo Järvinenb5860bb2007-08-09 14:33:18 +0300735 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
David S. Miller6475be12005-09-01 22:47:01 -0700736 tp->lost_out -= diff;
Herbert Xu83ca28b2005-09-22 23:32:56 -0700737
David S. Miller6475be12005-09-01 22:47:01 -0700738 if (diff > 0) {
Herbert Xu83ca28b2005-09-22 23:32:56 -0700739 /* Adjust Reno SACK estimate. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300740 if (tcp_is_reno(tp)) {
Ilpo Järvinenaf610b42007-06-15 12:58:38 +0300741 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
Ilpo Järvinen005903b2007-08-09 14:44:16 +0300742 tcp_verify_left_out(tp);
Herbert Xu83ca28b2005-09-22 23:32:56 -0700743 }
744
Ilpo Järvinenaf610b42007-06-15 12:58:38 +0300745 tcp_dec_pcount_approx_int(&tp->fackets_out, diff);
Ilpo Järvinend0410052007-07-02 22:07:22 -0700746 /* SACK fastpath might overwrite it unless dealt with */
747 if (tp->fastpath_skb_hint != NULL &&
748 after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq,
749 TCP_SKB_CB(skb)->seq)) {
Ilpo Järvinenaf610b42007-06-15 12:58:38 +0300750 tcp_dec_pcount_approx_int(&tp->fastpath_cnt_hint, diff);
Ilpo Järvinend0410052007-07-02 22:07:22 -0700751 }
David S. Miller6475be12005-09-01 22:47:01 -0700752 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 }
754
755 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700756 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -0800757 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758
759 return 0;
760}
761
762/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
763 * eventually). The difference is that pulled data not copied, but
764 * immediately discarded.
765 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700766static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767{
768 int i, k, eat;
769
770 eat = len;
771 k = 0;
772 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
773 if (skb_shinfo(skb)->frags[i].size <= eat) {
774 put_page(skb_shinfo(skb)->frags[i].page);
775 eat -= skb_shinfo(skb)->frags[i].size;
776 } else {
777 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
778 if (eat) {
779 skb_shinfo(skb)->frags[k].page_offset += eat;
780 skb_shinfo(skb)->frags[k].size -= eat;
781 eat = 0;
782 }
783 k++;
784 }
785 }
786 skb_shinfo(skb)->nr_frags = k;
787
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -0700788 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 skb->data_len -= len;
790 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791}
792
793int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
794{
795 if (skb_cloned(skb) &&
796 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
797 return -ENOMEM;
798
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700799 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
800 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700802 else
803 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804
805 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -0700806 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807
808 skb->truesize -= len;
809 sk->sk_wmem_queued -= len;
810 sk->sk_forward_alloc += len;
811 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
812
813 /* Any change of skb->len requires recalculation of tso
814 * factor and mss.
815 */
816 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700817 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818
819 return 0;
820}
821
John Heffner5d424d52006-03-20 17:53:41 -0800822/* Not accounting for SACKs here. */
823int tcp_mtu_to_mss(struct sock *sk, int pmtu)
824{
825 struct tcp_sock *tp = tcp_sk(sk);
826 struct inet_connection_sock *icsk = inet_csk(sk);
827 int mss_now;
828
829 /* Calculate base mss without TCP options:
830 It is MMS_S - sizeof(tcphdr) of rfc1122
831 */
832 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
833
834 /* Clamp it (mss_clamp does not include tcp options) */
835 if (mss_now > tp->rx_opt.mss_clamp)
836 mss_now = tp->rx_opt.mss_clamp;
837
838 /* Now subtract optional transport overhead */
839 mss_now -= icsk->icsk_ext_hdr_len;
840
841 /* Then reserve room for full set of TCP options and 8 bytes of data */
842 if (mss_now < 48)
843 mss_now = 48;
844
845 /* Now subtract TCP options size, not including SACKs */
846 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
847
848 return mss_now;
849}
850
851/* Inverse of above */
852int tcp_mss_to_mtu(struct sock *sk, int mss)
853{
854 struct tcp_sock *tp = tcp_sk(sk);
855 struct inet_connection_sock *icsk = inet_csk(sk);
856 int mtu;
857
858 mtu = mss +
859 tp->tcp_header_len +
860 icsk->icsk_ext_hdr_len +
861 icsk->icsk_af_ops->net_header_len;
862
863 return mtu;
864}
865
866void tcp_mtup_init(struct sock *sk)
867{
868 struct tcp_sock *tp = tcp_sk(sk);
869 struct inet_connection_sock *icsk = inet_csk(sk);
870
871 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
872 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900873 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -0800874 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
875 icsk->icsk_mtup.probe_size = 0;
876}
877
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878/* This function synchronize snd mss to current pmtu/exthdr set.
879
880 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
881 for TCP options, but includes only bare TCP header.
882
883 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800884 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 It also does not include TCP options.
886
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800887 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888
889 tp->mss_cache is current effective sending mss, including
890 all tcp options except for SACKs. It is evaluated,
891 taking into account current pmtu, but never exceeds
892 tp->rx_opt.mss_clamp.
893
894 NOTE1. rfc1122 clearly states that advertised MSS
895 DOES NOT include either tcp or ip options.
896
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800897 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
898 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 */
900
901unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
902{
903 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800904 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -0800905 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906
John Heffner5d424d52006-03-20 17:53:41 -0800907 if (icsk->icsk_mtup.search_high > pmtu)
908 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909
John Heffner5d424d52006-03-20 17:53:41 -0800910 mss_now = tcp_mtu_to_mss(sk, pmtu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911
912 /* Bound mss with half of window */
913 if (tp->max_window && mss_now > (tp->max_window>>1))
914 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
915
916 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800917 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -0800918 if (icsk->icsk_mtup.enabled)
919 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700920 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
922 return mss_now;
923}
924
925/* Compute the current effective MSS, taking SACKs and IP options,
926 * and even PMTU discovery events into account.
927 *
928 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
929 * cannot be large. However, taking into account rare use of URG, this
930 * is not a big flaw.
931 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700932unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933{
934 struct tcp_sock *tp = tcp_sk(sk);
935 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700936 u32 mss_now;
937 u16 xmit_size_goal;
938 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700940 mss_now = tp->mss_cache;
941
Herbert Xubcd76112006-06-30 13:36:35 -0700942 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700943 doing_tso = 1;
944
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 if (dst) {
946 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800947 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 mss_now = tcp_sync_mss(sk, mtu);
949 }
950
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 if (tp->rx_opt.eff_sacks)
952 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
953 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700954
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800955#ifdef CONFIG_TCP_MD5SIG
956 if (tp->af_specific->md5_lookup(sk, sk))
957 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
958#endif
959
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700960 xmit_size_goal = mss_now;
961
962 if (doing_tso) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800963 xmit_size_goal = (65535 -
964 inet_csk(sk)->icsk_af_ops->net_header_len -
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800965 inet_csk(sk)->icsk_ext_hdr_len -
966 tp->tcp_header_len);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700967
968 if (tp->max_window &&
969 (xmit_size_goal > (tp->max_window >> 1)))
970 xmit_size_goal = max((tp->max_window >> 1),
971 68U - tp->tcp_header_len);
972
973 xmit_size_goal -= (xmit_size_goal % mss_now);
974 }
975 tp->xmit_size_goal = xmit_size_goal;
976
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 return mss_now;
978}
979
David S. Millera762a982005-07-05 15:18:51 -0700980/* Congestion window validation. (RFC2861) */
981
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700982static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -0700983{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700984 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -0700985 __u32 packets_out = tp->packets_out;
986
987 if (packets_out >= tp->snd_cwnd) {
988 /* Network is feed fully. */
989 tp->snd_cwnd_used = 0;
990 tp->snd_cwnd_stamp = tcp_time_stamp;
991 } else {
992 /* Network starves. */
993 if (tp->packets_out > tp->snd_cwnd_used)
994 tp->snd_cwnd_used = tp->packets_out;
995
David S. Miller15d33c02007-04-09 13:23:14 -0700996 if (sysctl_tcp_slow_start_after_idle &&
997 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -0700998 tcp_cwnd_application_limited(sk);
999 }
1000}
1001
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001002static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
1003{
1004 u32 window, cwnd_len;
1005
1006 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
1007 cwnd_len = mss_now * cwnd;
1008 return min(window, cwnd_len);
1009}
1010
1011/* Can at least one segment of SKB be sent right now, according to the
1012 * congestion window rules? If so, return how many segments are allowed.
1013 */
1014static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
1015{
1016 u32 in_flight, cwnd;
1017
1018 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -08001019 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1020 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001021 return 1;
1022
1023 in_flight = tcp_packets_in_flight(tp);
1024 cwnd = tp->snd_cwnd;
1025 if (in_flight < cwnd)
1026 return (cwnd - in_flight);
1027
1028 return 0;
1029}
1030
1031/* This must be invoked the first time we consider transmitting
1032 * SKB onto the wire.
1033 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08001034static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001035{
1036 int tso_segs = tcp_skb_pcount(skb);
1037
David S. Miller846998a2005-08-04 19:52:01 -07001038 if (!tso_segs ||
1039 (tso_segs > 1 &&
Herbert Xu79671682006-06-22 02:40:14 -07001040 tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001041 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001042 tso_segs = tcp_skb_pcount(skb);
1043 }
1044 return tso_segs;
1045}
1046
1047static inline int tcp_minshall_check(const struct tcp_sock *tp)
1048{
1049 return after(tp->snd_sml,tp->snd_una) &&
1050 !after(tp->snd_sml, tp->snd_nxt);
1051}
1052
1053/* Return 0, if packet can be sent now without violation Nagle's rules:
1054 * 1. It is full sized.
1055 * 2. Or it contains FIN. (already checked by caller)
1056 * 3. Or TCP_NODELAY was set.
1057 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1058 * With Minshall's modification: all sent small packets are ACKed.
1059 */
1060
1061static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001062 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001063 unsigned mss_now, int nonagle)
1064{
1065 return (skb->len < mss_now &&
1066 ((nonagle&TCP_NAGLE_CORK) ||
1067 (!nonagle &&
1068 tp->packets_out &&
1069 tcp_minshall_check(tp))));
1070}
1071
1072/* Return non-zero if the Nagle test allows this packet to be
1073 * sent now.
1074 */
1075static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1076 unsigned int cur_mss, int nonagle)
1077{
1078 /* Nagle rule does not apply to frames, which sit in the middle of the
1079 * write_queue (they have no chances to get new data).
1080 *
1081 * This is implemented in the callers, where they modify the 'nonagle'
1082 * argument based upon the location of SKB in the send queue.
1083 */
1084 if (nonagle & TCP_NAGLE_PUSH)
1085 return 1;
1086
Ilpo Järvinend551e452007-04-30 00:42:20 -07001087 /* Don't use the nagle rule for urgent data (or for the final FIN).
1088 * Nagle can be ignored during F-RTO too (see RFC4138).
1089 */
1090 if (tp->urg_mode || (tp->frto_counter == 2) ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001091 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1092 return 1;
1093
1094 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1095 return 1;
1096
1097 return 0;
1098}
1099
1100/* Does at least the first segment of SKB fit into the send window? */
1101static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
1102{
1103 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1104
1105 if (skb->len > cur_mss)
1106 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1107
1108 return !after(end_seq, tp->snd_una + tp->snd_wnd);
1109}
1110
David S. Millerfe067e82007-03-07 12:12:44 -08001111/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001112 * should be put on the wire right now. If so, it returns the number of
1113 * packets allowed by the congestion window.
1114 */
1115static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1116 unsigned int cur_mss, int nonagle)
1117{
1118 struct tcp_sock *tp = tcp_sk(sk);
1119 unsigned int cwnd_quota;
1120
David S. Miller846998a2005-08-04 19:52:01 -07001121 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001122
1123 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1124 return 0;
1125
1126 cwnd_quota = tcp_cwnd_test(tp, skb);
1127 if (cwnd_quota &&
1128 !tcp_snd_wnd_test(tp, skb, cur_mss))
1129 cwnd_quota = 0;
1130
1131 return cwnd_quota;
1132}
1133
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001134int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001135{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001136 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001137 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001138
1139 return (skb &&
1140 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1141 (tcp_skb_is_last(sk, skb) ?
1142 TCP_NAGLE_PUSH :
1143 tp->nonagle)));
1144}
1145
1146/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1147 * which is put after SKB on the list. It is very much like
1148 * tcp_fragment() except that it may make several kinds of assumptions
1149 * in order to speed up the splitting operation. In particular, we
1150 * know that all the data is in scatter-gather pages, and that the
1151 * packet has never been sent out before (and thus is not cloned).
1152 */
David S. Miller846998a2005-08-04 19:52:01 -07001153static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001154{
1155 struct sk_buff *buff;
1156 int nlen = skb->len - len;
1157 u16 flags;
1158
1159 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001160 if (skb->len != skb->data_len)
1161 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001162
1163 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1164 if (unlikely(buff == NULL))
1165 return -ENOMEM;
1166
Herbert Xub60b49e2006-04-19 21:35:00 -07001167 sk_charge_skb(sk, buff);
1168 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001169 skb->truesize -= nlen;
1170
1171 /* Correct the sequence numbers. */
1172 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1173 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1174 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1175
1176 /* PSH and FIN should only be set in the second packet. */
1177 flags = TCP_SKB_CB(skb)->flags;
1178 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1179 TCP_SKB_CB(buff)->flags = flags;
1180
1181 /* This packet was never sent out yet, so no SACK bits. */
1182 TCP_SKB_CB(buff)->sacked = 0;
1183
Patrick McHardy84fa7932006-08-29 16:44:56 -07001184 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001185 skb_split(skb, buff, len);
1186
1187 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001188 tcp_set_skb_tso_segs(sk, skb, mss_now);
1189 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001190
1191 /* Link BUFF into the send queue. */
1192 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001193 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001194
1195 return 0;
1196}
1197
1198/* Try to defer sending, if possible, in order to minimize the amount
1199 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1200 *
1201 * This algorithm is from John Heffner.
1202 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001203static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001204{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001205 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001206 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001207 u32 send_win, cong_win, limit, in_flight;
1208
1209 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001210 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001211
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001212 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001213 goto send_now;
1214
1215 /* Defer for less than two clock ticks. */
1216 if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1)
1217 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001218
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001219 in_flight = tcp_packets_in_flight(tp);
1220
1221 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1222 (tp->snd_cwnd <= in_flight));
1223
1224 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1225
1226 /* From in_flight test above, we know that cwnd > in_flight. */
1227 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1228
1229 limit = min(send_win, cong_win);
1230
David S. Millerba244fe2006-03-11 18:51:49 -08001231 /* If a full-sized TSO skb can be sent, do it. */
1232 if (limit >= 65536)
John Heffnerae8064a2006-10-18 20:36:48 -07001233 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001234
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001235 if (sysctl_tcp_tso_win_divisor) {
1236 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1237
1238 /* If at least some fraction of a window is available,
1239 * just use it.
1240 */
1241 chunk /= sysctl_tcp_tso_win_divisor;
1242 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001243 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001244 } else {
1245 /* Different approach, try not to defer past a single
1246 * ACK. Receiver should ACK every other full sized
1247 * frame, so if we have space for more than 3 frames
1248 * then send now.
1249 */
1250 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001251 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001252 }
1253
1254 /* Ok, it looks like it is advisable to defer. */
John Heffnerae8064a2006-10-18 20:36:48 -07001255 tp->tso_deferred = 1 | (jiffies<<1);
1256
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001257 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001258
1259send_now:
1260 tp->tso_deferred = 0;
1261 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001262}
1263
John Heffner5d424d52006-03-20 17:53:41 -08001264/* Create a new MTU probe if we are ready.
1265 * Returns 0 if we should wait to probe (no cwnd available),
1266 * 1 if a probe was sent,
1267 * -1 otherwise */
1268static int tcp_mtu_probe(struct sock *sk)
1269{
1270 struct tcp_sock *tp = tcp_sk(sk);
1271 struct inet_connection_sock *icsk = inet_csk(sk);
1272 struct sk_buff *skb, *nskb, *next;
1273 int len;
1274 int probe_size;
1275 unsigned int pif;
1276 int copy;
1277 int mss_now;
1278
1279 /* Not currently probing/verifying,
1280 * not in recovery,
1281 * have enough cwnd, and
1282 * not SACKing (the variable headers throw things off) */
1283 if (!icsk->icsk_mtup.enabled ||
1284 icsk->icsk_mtup.probe_size ||
1285 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1286 tp->snd_cwnd < 11 ||
1287 tp->rx_opt.eff_sacks)
1288 return -1;
1289
1290 /* Very simple search strategy: just double the MSS. */
1291 mss_now = tcp_current_mss(sk, 0);
1292 probe_size = 2*tp->mss_cache;
1293 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1294 /* TODO: set timer for probe_converge_event */
1295 return -1;
1296 }
1297
1298 /* Have enough data in the send queue to probe? */
1299 len = 0;
David S. Millerfe067e82007-03-07 12:12:44 -08001300 if ((skb = tcp_send_head(sk)) == NULL)
John Heffner5d424d52006-03-20 17:53:41 -08001301 return -1;
1302 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
David S. Millerfe067e82007-03-07 12:12:44 -08001303 skb = tcp_write_queue_next(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001304 if (len < probe_size)
1305 return -1;
1306
1307 /* Receive window check. */
1308 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1309 if (tp->snd_wnd < probe_size)
1310 return -1;
1311 else
1312 return 0;
1313 }
1314
1315 /* Do we need to wait to drain cwnd? */
1316 pif = tcp_packets_in_flight(tp);
1317 if (pif + 2 > tp->snd_cwnd) {
1318 /* With no packets in flight, don't stall. */
1319 if (pif == 0)
1320 return -1;
1321 else
1322 return 0;
1323 }
1324
1325 /* We're allowed to probe. Build it now. */
1326 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1327 return -1;
1328 sk_charge_skb(sk, nskb);
1329
David S. Millerfe067e82007-03-07 12:12:44 -08001330 skb = tcp_send_head(sk);
1331 tcp_insert_write_queue_before(nskb, skb, sk);
1332 tcp_advance_send_head(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001333
1334 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1335 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1336 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1337 TCP_SKB_CB(nskb)->sacked = 0;
1338 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001339 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001340
1341 len = 0;
1342 while (len < probe_size) {
David S. Millerfe067e82007-03-07 12:12:44 -08001343 next = tcp_write_queue_next(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001344
1345 copy = min_t(int, skb->len, probe_size - len);
1346 if (nskb->ip_summed)
1347 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1348 else
1349 nskb->csum = skb_copy_and_csum_bits(skb, 0,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001350 skb_put(nskb, copy), copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001351
1352 if (skb->len <= copy) {
1353 /* We've eaten all the data from this skb.
1354 * Throw it away. */
1355 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001356 tcp_unlink_write_queue(skb, sk);
John Heffner5d424d52006-03-20 17:53:41 -08001357 sk_stream_free_skb(sk, skb);
1358 } else {
1359 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001360 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001361 if (!skb_shinfo(skb)->nr_frags) {
1362 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001363 if (skb->ip_summed != CHECKSUM_PARTIAL)
John Heffner5d424d52006-03-20 17:53:41 -08001364 skb->csum = csum_partial(skb->data, skb->len, 0);
1365 } else {
1366 __pskb_trim_head(skb, copy);
1367 tcp_set_skb_tso_segs(sk, skb, mss_now);
1368 }
1369 TCP_SKB_CB(skb)->seq += copy;
1370 }
1371
1372 len += copy;
1373 skb = next;
1374 }
1375 tcp_init_tso_segs(sk, nskb, nskb->len);
1376
1377 /* We're ready to send. If this fails, the probe will
1378 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1379 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1380 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1381 /* Decrement cwnd here because we are sending
1382 * effectively two packets. */
1383 tp->snd_cwnd--;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001384 update_send_head(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001385
1386 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001387 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1388 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001389
1390 return 1;
1391 }
1392
1393 return -1;
1394}
1395
1396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397/* This routine writes packets to the network. It advances the
1398 * send_head. This happens as incoming acks open up the remote
1399 * window for us.
1400 *
1401 * Returns 1, if no segments are in flight and we have queued segments, but
1402 * cannot send anything now because of SWS or another problem.
1403 */
David S. Millera2e2a592005-07-05 15:19:23 -07001404static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405{
1406 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001407 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001408 unsigned int tso_segs, sent_pkts;
1409 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001410 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411
1412 /* If we are closed, the bytes will have to remain here.
1413 * In time closedown will finish, we empty the write queue and all
1414 * will be happy.
1415 */
David S. Miller92df7b52005-07-05 15:19:06 -07001416 if (unlikely(sk->sk_state == TCP_CLOSE))
1417 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
David S. Miller92df7b52005-07-05 15:19:06 -07001419 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001420
1421 /* Do MTU probing. */
1422 if ((result = tcp_mtu_probe(sk)) == 0) {
1423 return 0;
1424 } else if (result > 0) {
1425 sent_pkts = 1;
1426 }
1427
David S. Millerfe067e82007-03-07 12:12:44 -08001428 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001429 unsigned int limit;
1430
Herbert Xub68e9f82005-08-04 19:52:02 -07001431 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001432 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001433
Herbert Xub68e9f82005-08-04 19:52:02 -07001434 cwnd_quota = tcp_cwnd_test(tp, skb);
1435 if (!cwnd_quota)
1436 break;
1437
1438 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1439 break;
1440
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001441 if (tso_segs == 1) {
1442 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1443 (tcp_skb_is_last(sk, skb) ?
1444 nonagle : TCP_NAGLE_PUSH))))
1445 break;
1446 } else {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001447 if (tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001448 break;
1449 }
David S. Milleraa934662005-07-05 15:20:09 -07001450
Herbert Xuc8ac3772005-08-16 20:43:40 -07001451 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001452 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001453 limit = tcp_window_allows(tp, skb,
1454 mss_now, cwnd_quota);
David S. Milleraa934662005-07-05 15:20:09 -07001455
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001456 if (skb->len < limit) {
1457 unsigned int trim = skb->len % mss_now;
1458
1459 if (trim)
1460 limit = skb->len - trim;
1461 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 }
1463
Herbert Xuc8ac3772005-08-16 20:43:40 -07001464 if (skb->len > limit &&
1465 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1466 break;
1467
David S. Miller92df7b52005-07-05 15:19:06 -07001468 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001469
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001470 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
David S. Miller92df7b52005-07-05 15:19:06 -07001471 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
David S. Miller92df7b52005-07-05 15:19:06 -07001473 /* Advance the send_head. This one is sent out.
1474 * This call will increment packets_out.
1475 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001476 update_send_head(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001477
1478 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001479 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 }
David S. Miller92df7b52005-07-05 15:19:06 -07001481
David S. Milleraa934662005-07-05 15:20:09 -07001482 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001483 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001484 return 0;
1485 }
David S. Millerfe067e82007-03-07 12:12:44 -08001486 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487}
1488
David S. Millera762a982005-07-05 15:18:51 -07001489/* Push out any pending frames which were held back due to
1490 * TCP_CORK or attempt at coalescing tiny packets.
1491 * The socket must be locked by the caller.
1492 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001493void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1494 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001495{
David S. Millerfe067e82007-03-07 12:12:44 -08001496 struct sk_buff *skb = tcp_send_head(sk);
David S. Millera762a982005-07-05 15:18:51 -07001497
1498 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001499 if (tcp_write_xmit(sk, cur_mss, nonagle))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001500 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001501 }
1502}
1503
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001504/* Send _single_ skb sitting at the send head. This function requires
1505 * true push pending frames to setup probe timer etc.
1506 */
1507void tcp_push_one(struct sock *sk, unsigned int mss_now)
1508{
1509 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001510 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001511 unsigned int tso_segs, cwnd_quota;
1512
1513 BUG_ON(!skb || skb->len < mss_now);
1514
David S. Miller846998a2005-08-04 19:52:01 -07001515 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001516 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1517
1518 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001519 unsigned int limit;
1520
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001521 BUG_ON(!tso_segs);
1522
Herbert Xuc8ac3772005-08-16 20:43:40 -07001523 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001524 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001525 limit = tcp_window_allows(tp, skb,
1526 mss_now, cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001527
1528 if (skb->len < limit) {
1529 unsigned int trim = skb->len % mss_now;
1530
1531 if (trim)
1532 limit = skb->len - trim;
1533 }
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001534 }
1535
Herbert Xuc8ac3772005-08-16 20:43:40 -07001536 if (skb->len > limit &&
1537 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1538 return;
1539
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001540 /* Send it out now. */
1541 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1542
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001543 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001544 update_send_head(sk, skb);
1545 tcp_cwnd_validate(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001546 return;
1547 }
1548 }
1549}
1550
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551/* This function returns the amount that we can raise the
1552 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001553 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 * 1. The window can never be shrunk once it is offered (RFC 793)
1555 * 2. We limit memory per socket
1556 *
1557 * RFC 1122:
1558 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1559 * RECV.NEXT + RCV.WIN fixed until:
1560 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1561 *
1562 * i.e. don't raise the right edge of the window until you can raise
1563 * it at least MSS bytes.
1564 *
1565 * Unfortunately, the recommended algorithm breaks header prediction,
1566 * since header prediction assumes th->window stays fixed.
1567 *
1568 * Strictly speaking, keeping th->window fixed violates the receiver
1569 * side SWS prevention criteria. The problem is that under this rule
1570 * a stream of single byte packets will cause the right side of the
1571 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001572 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 * Of course, if the sender implements sender side SWS prevention
1574 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001575 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001577 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 * If the free space is less than the 1/4 of the maximum
1579 * space available and the free space is less than 1/2 mss,
1580 * then set the window to 0.
1581 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1582 * Otherwise, just prevent the window from shrinking
1583 * and from being larger than the largest representable value.
1584 *
1585 * This prevents incremental opening of the window in the regime
1586 * where TCP is limited by the speed of the reader side taking
1587 * data out of the TCP receive queue. It does nothing about
1588 * those cases where the window is constrained on the sender side
1589 * because the pipeline is full.
1590 *
1591 * BSD also seems to "accidentally" limit itself to windows that are a
1592 * multiple of MSS, at least until the free space gets quite small.
1593 * This would appear to be a side effect of the mbuf implementation.
1594 * Combining these two algorithms results in the observed behavior
1595 * of having a fixed window size at almost all times.
1596 *
1597 * Below we obtain similar behavior by forcing the offered window to
1598 * a multiple of the mss when it is feasible to do so.
1599 *
1600 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1601 * Regular options like TIMESTAMP are taken into account.
1602 */
1603u32 __tcp_select_window(struct sock *sk)
1604{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001605 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001607 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 * here. I don't know if the value based on our guesses
1609 * of peer's MSS is better for the performance. It's more correct
1610 * but may be worse for the performance because of rcv_mss
1611 * fluctuations. --SAW 1998/11/1
1612 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001613 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 int free_space = tcp_space(sk);
1615 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1616 int window;
1617
1618 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001619 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620
1621 if (free_space < full_space/2) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001622 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
1624 if (tcp_memory_pressure)
1625 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1626
1627 if (free_space < mss)
1628 return 0;
1629 }
1630
1631 if (free_space > tp->rcv_ssthresh)
1632 free_space = tp->rcv_ssthresh;
1633
1634 /* Don't do rounding if we are using window scaling, since the
1635 * scaled window will not line up with the MSS boundary anyway.
1636 */
1637 window = tp->rcv_wnd;
1638 if (tp->rx_opt.rcv_wscale) {
1639 window = free_space;
1640
1641 /* Advertise enough space so that it won't get scaled away.
1642 * Import case: prevent zero window announcement if
1643 * 1<<rcv_wscale > mss.
1644 */
1645 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1646 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1647 << tp->rx_opt.rcv_wscale);
1648 } else {
1649 /* Get the largest window that is a nice multiple of mss.
1650 * Window clamp already applied above.
1651 * If our current window offering is within 1 mss of the
1652 * free space we just keep it. This prevents the divide
1653 * and multiply from happening most of the time.
1654 * We also don't do any window rounding when the free space
1655 * is too small.
1656 */
1657 if (window <= free_space - mss || window > free_space)
1658 window = (free_space/mss)*mss;
John Heffner84565072007-04-02 13:56:32 -07001659 else if (mss == full_space &&
YOSHIFUJI Hideaki9c681b42007-07-19 10:43:47 +09001660 free_space > window + full_space/2)
John Heffner84565072007-04-02 13:56:32 -07001661 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 }
1663
1664 return window;
1665}
1666
1667/* Attempt to collapse two adjacent SKB's during retransmission. */
1668static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1669{
1670 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001671 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672
1673 /* The first test we must make is that neither of these two
1674 * SKB's are still referenced by someone else.
1675 */
1676 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1677 int skb_size = skb->len, next_skb_size = next_skb->len;
1678 u16 flags = TCP_SKB_CB(skb)->flags;
1679
1680 /* Also punt if next skb has been SACK'd. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001681 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 return;
1683
1684 /* Next skb is out of window. */
1685 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1686 return;
1687
1688 /* Punt if not enough space exists in the first SKB for
1689 * the data in the second, or the total combined payload
1690 * would exceed the MSS.
1691 */
1692 if ((next_skb_size > skb_tailroom(skb)) ||
1693 ((skb_size + next_skb_size) > mss_now))
1694 return;
1695
1696 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1697 tcp_skb_pcount(next_skb) != 1);
1698
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001699 /* changing transmit queue under us so clear hints */
1700 clear_all_retrans_hints(tp);
1701
1702 /* Ok. We will be able to collapse the packet. */
David S. Millerfe067e82007-03-07 12:12:44 -08001703 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704
Arnaldo Carvalho de Melo1a4e2d02007-03-31 11:55:45 -03001705 skb_copy_from_linear_data(next_skb,
1706 skb_put(skb, next_skb_size),
1707 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708
Jarek Poplawski52d570a2007-01-23 22:07:12 -08001709 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1710 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
Patrick McHardy84fa7932006-08-29 16:44:56 -07001712 if (skb->ip_summed != CHECKSUM_PARTIAL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1714
1715 /* Update sequence range on original skb. */
1716 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1717
1718 /* Merge over control information. */
1719 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1720 TCP_SKB_CB(skb)->flags = flags;
1721
1722 /* All done, get rid of second SKB and account for it so
1723 * packet counting does not break.
1724 */
1725 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1726 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1727 tp->retrans_out -= tcp_skb_pcount(next_skb);
Ilpo Järvinenb5860bb2007-08-09 14:33:18 +03001728 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 tp->lost_out -= tcp_skb_pcount(next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 /* Reno case is special. Sigh... */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001731 if (tcp_is_reno(tp) && tp->sacked_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733
1734 /* Not quite right: it can be > snd.fack, but
1735 * it is better to underestimate fackets.
1736 */
1737 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1738 tcp_packets_out_dec(tp, next_skb);
1739 sk_stream_free_skb(sk, next_skb);
1740 }
1741}
1742
1743/* Do a simple retransmit without using the backoff mechanisms in
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001744 * tcp_timer. This is used for path mtu discovery.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 * The socket is already locked here.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001746 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747void tcp_simple_retransmit(struct sock *sk)
1748{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001749 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 struct tcp_sock *tp = tcp_sk(sk);
1751 struct sk_buff *skb;
1752 unsigned int mss = tcp_current_mss(sk, 0);
1753 int lost = 0;
1754
David S. Millerfe067e82007-03-07 12:12:44 -08001755 tcp_for_write_queue(skb, sk) {
1756 if (skb == tcp_send_head(sk))
1757 break;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001758 if (skb->len > mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1760 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1761 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1762 tp->retrans_out -= tcp_skb_pcount(skb);
1763 }
1764 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1765 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1766 tp->lost_out += tcp_skb_pcount(skb);
1767 lost = 1;
1768 }
1769 }
1770 }
1771
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001772 clear_all_retrans_hints(tp);
1773
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 if (!lost)
1775 return;
1776
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001777 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001779 /* Don't muck with the congestion window here.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 * Reason is that we do not increase amount of _data_
1781 * in network, but units changed and effective
1782 * cwnd/ssthresh really reduced now.
1783 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001784 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001786 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 tp->prior_ssthresh = 0;
1788 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001789 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 }
1791 tcp_xmit_retransmit_queue(sk);
1792}
1793
1794/* This retransmits one SKB. Policy decisions and retransmit queue
1795 * state updates are done by the caller. Returns non-zero if an
1796 * error occurred which prevented the send.
1797 */
1798int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1799{
1800 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001801 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001802 unsigned int cur_mss = tcp_current_mss(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 int err;
1804
John Heffner5d424d52006-03-20 17:53:41 -08001805 /* Inconslusive MTU probe */
1806 if (icsk->icsk_mtup.probe_size) {
1807 icsk->icsk_mtup.probe_size = 0;
1808 }
1809
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001811 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 */
1813 if (atomic_read(&sk->sk_wmem_alloc) >
1814 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1815 return -EAGAIN;
1816
1817 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1818 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1819 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1821 return -ENOMEM;
1822 }
1823
1824 /* If receiver has shrunk his window, and skb is out of
1825 * new window, do not retransmit it. The exception is the
1826 * case, when window is shrunk to zero. In this case
1827 * our retransmit serves as a zero window probe.
1828 */
1829 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1830 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1831 return -EAGAIN;
1832
1833 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001834 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 }
1837
1838 /* Collapse two adjacent packets if worthwhile and we can. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001839 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1840 (skb->len < (cur_mss >> 1)) &&
1841 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1842 (!tcp_skb_is_last(sk, skb)) &&
1843 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1844 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1845 (sysctl_tcp_retrans_collapse != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 tcp_retrans_try_collapse(sk, skb, cur_mss);
1847
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001848 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 return -EHOSTUNREACH; /* Routing failure or similar. */
1850
1851 /* Some Solaris stacks overoptimize and ignore the FIN on a
1852 * retransmit when old data is attached. So strip it off
1853 * since it is cheap to do so and saves bytes on the network.
1854 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001855 if (skb->len > 0 &&
1856 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1857 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 if (!pskb_trim(skb, 0)) {
1859 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
Herbert Xu79671682006-06-22 02:40:14 -07001860 skb_shinfo(skb)->gso_segs = 1;
1861 skb_shinfo(skb)->gso_size = 0;
1862 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 skb->ip_summed = CHECKSUM_NONE;
1864 skb->csum = 0;
1865 }
1866 }
1867
1868 /* Make a copy, if the first transmission SKB clone we made
1869 * is still in somebody's hands, else make a clone.
1870 */
1871 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001873 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874
1875 if (err == 0) {
1876 /* Update global TCP statistics. */
1877 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1878
1879 tp->total_retrans++;
1880
1881#if FASTRETRANS_DEBUG > 0
1882 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1883 if (net_ratelimit())
1884 printk(KERN_DEBUG "retrans_out leaked.\n");
1885 }
1886#endif
1887 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1888 tp->retrans_out += tcp_skb_pcount(skb);
1889
1890 /* Save stamp of the first retransmit. */
1891 if (!tp->retrans_stamp)
1892 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1893
1894 tp->undo_retrans++;
1895
1896 /* snd_nxt is stored to detect loss of retransmitted segment,
1897 * see tcp_input.c tcp_sacktag_write_queue().
1898 */
1899 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1900 }
1901 return err;
1902}
1903
1904/* This gets called after a retransmit timeout, and the initially
1905 * retransmitted data is acknowledged. It tries to continue
1906 * resending the rest of the retransmit queue, until either
1907 * we've sent it all or the congestion window limit is reached.
1908 * If doing SACK, the first ACK which comes back for a timeout
1909 * based retransmit packet might feed us FACK information again.
1910 * If so, we use it to avoid unnecessarily retransmissions.
1911 */
1912void tcp_xmit_retransmit_queue(struct sock *sk)
1913{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001914 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 struct tcp_sock *tp = tcp_sk(sk);
1916 struct sk_buff *skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001917 int packet_cnt;
1918
1919 if (tp->retransmit_skb_hint) {
1920 skb = tp->retransmit_skb_hint;
1921 packet_cnt = tp->retransmit_cnt_hint;
1922 }else{
David S. Millerfe067e82007-03-07 12:12:44 -08001923 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001924 packet_cnt = 0;
1925 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926
1927 /* First pass: retransmit lost packets. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001928 if (tp->lost_out) {
David S. Millerfe067e82007-03-07 12:12:44 -08001929 tcp_for_write_queue_from(skb, sk) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1931
David S. Millerfe067e82007-03-07 12:12:44 -08001932 if (skb == tcp_send_head(sk))
1933 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001934 /* we could do better than to assign each time */
1935 tp->retransmit_skb_hint = skb;
1936 tp->retransmit_cnt_hint = packet_cnt;
1937
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 /* Assume this retransmit will generate
1939 * only one packet for congestion window
1940 * calculation purposes. This works because
1941 * tcp_retransmit_skb() will chop up the
1942 * packet to be MSS sized and all the
1943 * packet counting works out.
1944 */
1945 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1946 return;
1947
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001948 if (sacked & TCPCB_LOST) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001950 if (tcp_retransmit_skb(sk, skb)) {
1951 tp->retransmit_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001953 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001954 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1956 else
1957 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1958
David S. Millerfe067e82007-03-07 12:12:44 -08001959 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001960 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001961 inet_csk(sk)->icsk_rto,
1962 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 }
1964
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001965 packet_cnt += tcp_skb_pcount(skb);
1966 if (packet_cnt >= tp->lost_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967 break;
1968 }
1969 }
1970 }
1971
1972 /* OK, demanded retransmission is finished. */
1973
1974 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001975 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 return;
1977
1978 /* No forward retransmissions in Reno are possible. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001979 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 return;
1981
1982 /* Yeah, we have to make difficult choice between forward transmission
1983 * and retransmission... Both ways have their merits...
1984 *
1985 * For now we do not retransmit anything, while we have some new
Ilpo Järvinen539d2432007-05-27 02:03:20 -07001986 * segments to send. In the other cases, follow rule 3 for
1987 * NextSeg() specified in RFC3517.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 */
1989
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001990 if (tcp_may_send_now(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 return;
1992
Ilpo Järvinen539d2432007-05-27 02:03:20 -07001993 /* If nothing is SACKed, highest_sack in the loop won't be valid */
1994 if (!tp->sacked_out)
1995 return;
1996
1997 if (tp->forward_skb_hint)
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001998 skb = tp->forward_skb_hint;
Ilpo Järvinen539d2432007-05-27 02:03:20 -07001999 else
David S. Millerfe067e82007-03-07 12:12:44 -08002000 skb = tcp_write_queue_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001
David S. Millerfe067e82007-03-07 12:12:44 -08002002 tcp_for_write_queue_from(skb, sk) {
2003 if (skb == tcp_send_head(sk))
2004 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002005 tp->forward_skb_hint = skb;
2006
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002007 if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 break;
2009
2010 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2011 break;
2012
2013 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2014 continue;
2015
2016 /* Ok, retransmit it. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002017 if (tcp_retransmit_skb(sk, skb)) {
2018 tp->forward_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002020 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021
David S. Millerfe067e82007-03-07 12:12:44 -08002022 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002023 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2024 inet_csk(sk)->icsk_rto,
2025 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026
2027 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2028 }
2029}
2030
2031
2032/* Send a fin. The caller locks the socket for us. This cannot be
2033 * allowed to fail queueing a FIN frame under any circumstances.
2034 */
2035void tcp_send_fin(struct sock *sk)
2036{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002037 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002038 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002040
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 /* Optimization, tack on the FIN if we have a queue of
2042 * unsent frames. But be careful about outgoing SACKS
2043 * and IP options.
2044 */
2045 mss_now = tcp_current_mss(sk, 1);
2046
David S. Millerfe067e82007-03-07 12:12:44 -08002047 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2049 TCP_SKB_CB(skb)->end_seq++;
2050 tp->write_seq++;
2051 } else {
2052 /* Socket is locked, keep trying until memory is available. */
2053 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07002054 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 if (skb)
2056 break;
2057 yield();
2058 }
2059
2060 /* Reserve space for headers and prepare control bits. */
2061 skb_reserve(skb, MAX_TCP_HEADER);
2062 skb->csum = 0;
2063 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2064 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002065 skb_shinfo(skb)->gso_segs = 1;
2066 skb_shinfo(skb)->gso_size = 0;
2067 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068
2069 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2070 TCP_SKB_CB(skb)->seq = tp->write_seq;
2071 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2072 tcp_queue_skb(sk, skb);
2073 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002074 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075}
2076
2077/* We get here when a process closes a file descriptor (either due to
2078 * an explicit close() or as a byproduct of exit()'ing) and there
2079 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002080 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 */
Al Virodd0fc662005-10-07 07:46:04 +01002082void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 struct sk_buff *skb;
2085
2086 /* NOTE: No TCP options attached and we never retransmit this. */
2087 skb = alloc_skb(MAX_TCP_HEADER, priority);
2088 if (!skb) {
2089 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2090 return;
2091 }
2092
2093 /* Reserve space for headers and prepare control bits. */
2094 skb_reserve(skb, MAX_TCP_HEADER);
2095 skb->csum = 0;
2096 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2097 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002098 skb_shinfo(skb)->gso_segs = 1;
2099 skb_shinfo(skb)->gso_size = 0;
2100 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101
2102 /* Send it off. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002103 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2105 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002106 if (tcp_transmit_skb(sk, skb, 0, priority))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2108}
2109
2110/* WARNING: This routine must only be called when we have already sent
2111 * a SYN packet that crossed the incoming SYN that caused this routine
2112 * to get called. If this assumption fails then the initial rcv_wnd
2113 * and rcv_wscale values will not be correct.
2114 */
2115int tcp_send_synack(struct sock *sk)
2116{
2117 struct sk_buff* skb;
2118
David S. Millerfe067e82007-03-07 12:12:44 -08002119 skb = tcp_write_queue_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
2121 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2122 return -EFAULT;
2123 }
2124 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
2125 if (skb_cloned(skb)) {
2126 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2127 if (nskb == NULL)
2128 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002129 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002131 __tcp_add_write_queue_head(sk, nskb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 sk_stream_free_skb(sk, skb);
2133 sk_charge_skb(sk, nskb);
2134 skb = nskb;
2135 }
2136
2137 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2138 TCP_ECN_send_synack(tcp_sk(sk), skb);
2139 }
2140 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002141 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142}
2143
2144/*
2145 * Prepare a SYN-ACK.
2146 */
2147struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
Arnaldo Carvalho de Melo60236fd2005-06-18 22:47:21 -07002148 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002150 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 struct tcp_sock *tp = tcp_sk(sk);
2152 struct tcphdr *th;
2153 int tcp_header_size;
2154 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002155#ifdef CONFIG_TCP_MD5SIG
2156 struct tcp_md5sig_key *md5;
2157 __u8 *md5_hash_location;
2158#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159
2160 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2161 if (skb == NULL)
2162 return NULL;
2163
2164 /* Reserve space for headers. */
2165 skb_reserve(skb, MAX_TCP_HEADER);
2166
2167 skb->dst = dst_clone(dst);
2168
2169 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002170 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2171 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002173 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002174
2175#ifdef CONFIG_TCP_MD5SIG
2176 /* Are we doing MD5 on this segment? If so - make room for it */
2177 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2178 if (md5)
2179 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2180#endif
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002181 skb_push(skb, tcp_header_size);
2182 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002184 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 memset(th, 0, sizeof(struct tcphdr));
2186 th->syn = 1;
2187 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 TCP_ECN_make_synack(req, th);
2189 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002190 th->dest = ireq->rmt_port;
2191 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2193 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002194 skb_shinfo(skb)->gso_segs = 1;
2195 skb_shinfo(skb)->gso_size = 0;
2196 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002198 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002200 __u8 rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 /* Set this up on the first call only */
2202 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2203 /* tcp_full_space because it is guaranteed to be the first packet */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002204 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002205 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 &req->rcv_wnd,
2207 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002208 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 &rcv_wscale);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002210 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 }
2212
2213 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002214 th->window = htons(min(req->rcv_wnd, 65535U));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215
2216 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Al Virodf7a3b02006-09-27 18:38:52 -07002217 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002218 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 TCP_SKB_CB(skb)->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002220 req->ts_recent,
2221 (
2222#ifdef CONFIG_TCP_MD5SIG
2223 md5 ? &md5_hash_location :
2224#endif
2225 NULL)
2226 );
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227
2228 skb->csum = 0;
2229 th->doff = (tcp_header_size >> 2);
2230 TCP_INC_STATS(TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002231
2232#ifdef CONFIG_TCP_MD5SIG
2233 /* Okay, we have all we need - do the md5 hash if needed */
2234 if (md5) {
2235 tp->af_specific->calc_md5_hash(md5_hash_location,
2236 md5,
2237 NULL, dst, req,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002238 tcp_hdr(skb), sk->sk_protocol,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002239 skb->len);
2240 }
2241#endif
2242
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 return skb;
2244}
2245
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002246/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247 * Do all connect socket setups that can be done AF independent.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002248 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002249static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250{
2251 struct dst_entry *dst = __sk_dst_get(sk);
2252 struct tcp_sock *tp = tcp_sk(sk);
2253 __u8 rcv_wscale;
2254
2255 /* We'll fix this up when we get a response from the other end.
2256 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2257 */
2258 tp->tcp_header_len = sizeof(struct tcphdr) +
2259 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2260
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002261#ifdef CONFIG_TCP_MD5SIG
2262 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2263 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2264#endif
2265
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 /* If user gave his TCP_MAXSEG, record it to clamp */
2267 if (tp->rx_opt.user_mss)
2268 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2269 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002270 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 tcp_sync_mss(sk, dst_mtu(dst));
2272
2273 if (!tp->window_clamp)
2274 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2275 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2276 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277
2278 tcp_select_initial_window(tcp_full_space(sk),
2279 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2280 &tp->rcv_wnd,
2281 &tp->window_clamp,
2282 sysctl_tcp_window_scaling,
2283 &rcv_wscale);
2284
2285 tp->rx_opt.rcv_wscale = rcv_wscale;
2286 tp->rcv_ssthresh = tp->rcv_wnd;
2287
2288 sk->sk_err = 0;
2289 sock_reset_flag(sk, SOCK_DONE);
2290 tp->snd_wnd = 0;
2291 tcp_init_wl(tp, tp->write_seq, 0);
2292 tp->snd_una = tp->write_seq;
2293 tp->snd_sml = tp->write_seq;
2294 tp->rcv_nxt = 0;
2295 tp->rcv_wup = 0;
2296 tp->copied_seq = 0;
2297
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002298 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2299 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 tcp_clear_retrans(tp);
2301}
2302
2303/*
2304 * Build a SYN and send it off.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002305 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306int tcp_connect(struct sock *sk)
2307{
2308 struct tcp_sock *tp = tcp_sk(sk);
2309 struct sk_buff *buff;
2310
2311 tcp_connect_init(sk);
2312
David S. Millerd179cd12005-08-17 14:57:30 -07002313 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 if (unlikely(buff == NULL))
2315 return -ENOBUFS;
2316
2317 /* Reserve space for headers. */
2318 skb_reserve(buff, MAX_TCP_HEADER);
2319
2320 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002321 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002323 skb_shinfo(buff)->gso_segs = 1;
2324 skb_shinfo(buff)->gso_size = 0;
2325 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 buff->csum = 0;
Wei Yongjunbd37a082006-08-07 21:04:15 -07002327 tp->snd_nxt = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2329 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330
2331 /* Send it off. */
2332 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2333 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2334 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002335 __tcp_add_write_queue_tail(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 sk_charge_skb(sk, buff);
2337 tp->packets_out += tcp_skb_pcount(buff);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002338 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002339
2340 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2341 * in order to make this packet get counted in tcpOutSegs.
2342 */
2343 tp->snd_nxt = tp->write_seq;
2344 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2346
2347 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002348 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2349 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 return 0;
2351}
2352
2353/* Send out a delayed ack, the caller does the policy checking
2354 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2355 * for details.
2356 */
2357void tcp_send_delayed_ack(struct sock *sk)
2358{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002359 struct inet_connection_sock *icsk = inet_csk(sk);
2360 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 unsigned long timeout;
2362
2363 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002364 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 int max_ato = HZ/2;
2366
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002367 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 max_ato = TCP_DELACK_MAX;
2369
2370 /* Slow path, intersegment interval is "high". */
2371
2372 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002373 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 * directly.
2375 */
2376 if (tp->srtt) {
2377 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2378
2379 if (rtt < max_ato)
2380 max_ato = rtt;
2381 }
2382
2383 ato = min(ato, max_ato);
2384 }
2385
2386 /* Stay within the limit we were given */
2387 timeout = jiffies + ato;
2388
2389 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002390 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 /* If delack timer was blocked or is about to expire,
2392 * send ACK now.
2393 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002394 if (icsk->icsk_ack.blocked ||
2395 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 tcp_send_ack(sk);
2397 return;
2398 }
2399
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002400 if (!time_before(timeout, icsk->icsk_ack.timeout))
2401 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002403 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2404 icsk->icsk_ack.timeout = timeout;
2405 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406}
2407
2408/* This routine sends an ack and also updates the window. */
2409void tcp_send_ack(struct sock *sk)
2410{
2411 /* If we have been reset, we may not send again. */
2412 if (sk->sk_state != TCP_CLOSE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 struct sk_buff *buff;
2414
2415 /* We are not putting this on the write queue, so
2416 * tcp_transmit_skb() will set the ownership to this
2417 * sock.
2418 */
2419 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2420 if (buff == NULL) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002421 inet_csk_schedule_ack(sk);
2422 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002423 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2424 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 return;
2426 }
2427
2428 /* Reserve space for headers and prepare control bits. */
2429 skb_reserve(buff, MAX_TCP_HEADER);
2430 buff->csum = 0;
2431 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2432 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002433 skb_shinfo(buff)->gso_segs = 1;
2434 skb_shinfo(buff)->gso_size = 0;
2435 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436
2437 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002438 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 TCP_SKB_CB(buff)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002440 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 }
2442}
2443
2444/* This routine sends a packet with an out of date sequence
2445 * number. It assumes the other end will try to ack it.
2446 *
2447 * Question: what should we make while urgent mode?
2448 * 4.4BSD forces sending single byte of data. We cannot send
2449 * out of window data, because we have SND.NXT==SND.MAX...
2450 *
2451 * Current solution: to send TWO zero-length segments in urgent mode:
2452 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2453 * out-of-date with SND.UNA-1 to probe window.
2454 */
2455static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2456{
2457 struct tcp_sock *tp = tcp_sk(sk);
2458 struct sk_buff *skb;
2459
2460 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2461 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002462 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 return -1;
2464
2465 /* Reserve space for headers and set control bits. */
2466 skb_reserve(skb, MAX_TCP_HEADER);
2467 skb->csum = 0;
2468 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2469 TCP_SKB_CB(skb)->sacked = urgent;
Herbert Xu79671682006-06-22 02:40:14 -07002470 skb_shinfo(skb)->gso_segs = 1;
2471 skb_shinfo(skb)->gso_size = 0;
2472 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473
2474 /* Use a previous sequence. This should cause the other
2475 * end to send an ack. Don't queue or clone SKB, just
2476 * send it.
2477 */
2478 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2479 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2480 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002481 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482}
2483
2484int tcp_write_wakeup(struct sock *sk)
2485{
2486 if (sk->sk_state != TCP_CLOSE) {
2487 struct tcp_sock *tp = tcp_sk(sk);
2488 struct sk_buff *skb;
2489
David S. Millerfe067e82007-03-07 12:12:44 -08002490 if ((skb = tcp_send_head(sk)) != NULL &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2492 int err;
2493 unsigned int mss = tcp_current_mss(sk, 0);
2494 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2495
2496 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2497 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2498
2499 /* We are probing the opening of a window
2500 * but the window size is != 0
2501 * must have been a result SWS avoidance ( sender )
2502 */
2503 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2504 skb->len > mss) {
2505 seg_size = min(seg_size, mss);
2506 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
David S. Miller846998a2005-08-04 19:52:01 -07002507 if (tcp_fragment(sk, skb, seg_size, mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 } else if (!tcp_skb_pcount(skb))
David S. Miller846998a2005-08-04 19:52:01 -07002510 tcp_set_skb_tso_segs(sk, skb, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511
2512 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2513 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002514 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 if (!err) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002516 update_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 }
2518 return err;
2519 } else {
2520 if (tp->urg_mode &&
2521 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2522 tcp_xmit_probe_skb(sk, TCPCB_URG);
2523 return tcp_xmit_probe_skb(sk, 0);
2524 }
2525 }
2526 return -1;
2527}
2528
2529/* A window probe timeout has occurred. If window is not closed send
2530 * a partial packet else a zero probe.
2531 */
2532void tcp_send_probe0(struct sock *sk)
2533{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002534 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 struct tcp_sock *tp = tcp_sk(sk);
2536 int err;
2537
2538 err = tcp_write_wakeup(sk);
2539
David S. Millerfe067e82007-03-07 12:12:44 -08002540 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002542 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002543 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 return;
2545 }
2546
2547 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002548 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2549 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002550 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002551 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002552 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2553 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 } else {
2555 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002556 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 * Let local senders to fight for local resources.
2558 *
2559 * Use accumulated backoff yet.
2560 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002561 if (!icsk->icsk_probes_out)
2562 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002563 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002564 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002565 TCP_RESOURCE_PROBE_INTERVAL),
2566 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 }
2568}
2569
2570EXPORT_SYMBOL(tcp_connect);
2571EXPORT_SYMBOL(tcp_make_synack);
2572EXPORT_SYMBOL(tcp_simple_retransmit);
2573EXPORT_SYMBOL(tcp_sync_mss);
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -08002574EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
John Heffner5d424d52006-03-20 17:53:41 -08002575EXPORT_SYMBOL(tcp_mtup_init);