blob: b5fa3c19afeecf81ccdee0b3c9a65be7dace0cce [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>
43#include <linux/smp_lock.h>
44
45/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070046int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Rick Jones15d99e02006-03-20 22:40:29 -080048/* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
50 */
Brian Haleyab32ea52006-09-22 14:15:41 -070051int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080052
Linus Torvalds1da177e2005-04-16 15:20:36 -070053/* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
56 */
Brian Haleyab32ea52006-09-22 14:15:41 -070057int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
Brian Haleyab32ea52006-09-22 14:15:41 -070059int sysctl_tcp_mtu_probing __read_mostly = 0;
60int sysctl_tcp_base_mss __read_mostly = 512;
John Heffner5d424d52006-03-20 17:53:41 -080061
David S. Miller35089bb2006-06-13 22:33:04 -070062/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070063int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070064
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070065static void update_send_head(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -070066{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070067 struct tcp_sock *tp = tcp_sk(sk);
68
David S. Millerfe067e82007-03-07 12:12:44 -080069 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070071 tcp_packets_out_inc(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070072}
73
74/* SND.NXT, if window was not shrunk.
75 * If window has been shrunk, what should we make? It is not clear at all.
76 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
77 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
78 * invalid. OK, let's make this for now:
79 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070080static inline __u32 tcp_acceptable_seq(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070081{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070082 struct tcp_sock *tp = tcp_sk(sk);
83
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
85 return tp->snd_nxt;
86 else
87 return tp->snd_una+tp->snd_wnd;
88}
89
90/* Calculate mss to advertise in SYN segment.
91 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92 *
93 * 1. It is independent of path mtu.
94 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
95 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
96 * attached devices, because some buggy hosts are confused by
97 * large MSS.
98 * 4. We do not make 3, we advertise MSS, calculated from first
99 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
100 * This may be overridden via information stored in routing table.
101 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
102 * probably even Jumbo".
103 */
104static __u16 tcp_advertise_mss(struct sock *sk)
105{
106 struct tcp_sock *tp = tcp_sk(sk);
107 struct dst_entry *dst = __sk_dst_get(sk);
108 int mss = tp->advmss;
109
110 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
111 mss = dst_metric(dst, RTAX_ADVMSS);
112 tp->advmss = mss;
113 }
114
115 return (__u16)mss;
116}
117
118/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
119 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700120static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700122 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 s32 delta = tcp_time_stamp - tp->lsndtime;
124 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
125 u32 cwnd = tp->snd_cwnd;
126
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300127 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300129 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 restart_cwnd = min(restart_cwnd, cwnd);
131
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700132 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 cwnd >>= 1;
134 tp->snd_cwnd = max(cwnd, restart_cwnd);
135 tp->snd_cwnd_stamp = tcp_time_stamp;
136 tp->snd_cwnd_used = 0;
137}
138
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800139static void tcp_event_data_sent(struct tcp_sock *tp,
140 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700142 struct inet_connection_sock *icsk = inet_csk(sk);
143 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
David S. Miller35089bb2006-06-13 22:33:04 -0700145 if (sysctl_tcp_slow_start_after_idle &&
146 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700147 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148
149 tp->lsndtime = now;
150
151 /* If it is a reply for ato after last received
152 * packet, enter pingpong mode.
153 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700154 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
155 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156}
157
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800158static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700160 tcp_dec_quickack_mode(sk, pkts);
161 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162}
163
164/* Determine a window scaling and initial window to offer.
165 * Based on the assumption that the given amount of space
166 * will be offered. Store the results in the tp structure.
167 * NOTE: for smooth operation initial space offering should
168 * be a multiple of mss if possible. We assume here that mss >= 1.
169 * This MUST be enforced by all callers.
170 */
171void tcp_select_initial_window(int __space, __u32 mss,
172 __u32 *rcv_wnd, __u32 *window_clamp,
173 int wscale_ok, __u8 *rcv_wscale)
174{
175 unsigned int space = (__space < 0 ? 0 : __space);
176
177 /* If no clamp set the clamp to the max possible scaled window */
178 if (*window_clamp == 0)
179 (*window_clamp) = (65535 << 14);
180 space = min(*window_clamp, space);
181
182 /* Quantize space offering to a multiple of mss if possible. */
183 if (space > mss)
184 space = (space / mss) * mss;
185
186 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800187 * will break some buggy TCP stacks. If the admin tells us
188 * it is likely we could be speaking with such a buggy stack
189 * we will truncate our initial window offering to 32K-1
190 * unless the remote has sent us a window scaling option,
191 * which we interpret as a sign the remote TCP is not
192 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 */
Rick Jones15d99e02006-03-20 22:40:29 -0800194 if (sysctl_tcp_workaround_signed_windows)
195 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
196 else
197 (*rcv_wnd) = space;
198
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199 (*rcv_wscale) = 0;
200 if (wscale_ok) {
201 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900202 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 */
204 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700205 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 while (space > 65535 && (*rcv_wscale) < 14) {
207 space >>= 1;
208 (*rcv_wscale)++;
209 }
210 }
211
212 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700213 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 * will be satisfied with 2.
215 */
216 if (mss > (1<<*rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700217 int init_cwnd = 4;
218 if (mss > 1460*3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700220 else if (mss > 1460)
221 init_cwnd = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 if (*rcv_wnd > init_cwnd*mss)
223 *rcv_wnd = init_cwnd*mss;
224 }
225
226 /* Set the clamp no higher than max representable value */
227 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
228}
229
230/* Chose a new window to advertise, update state in tcp_sock for the
231 * socket, and return result with RFC1323 scaling applied. The return
232 * value can be stuffed directly into th->window for an outgoing
233 * frame.
234 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800235static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
237 struct tcp_sock *tp = tcp_sk(sk);
238 u32 cur_win = tcp_receive_window(tp);
239 u32 new_win = __tcp_select_window(sk);
240
241 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800242 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 /* Danger Will Robinson!
244 * Don't update rcv_wup/rcv_wnd here or else
245 * we will not be able to advertise a zero
246 * window in time. --DaveM
247 *
248 * Relax Will Robinson.
249 */
250 new_win = cur_win;
251 }
252 tp->rcv_wnd = new_win;
253 tp->rcv_wup = tp->rcv_nxt;
254
255 /* Make sure we do not exceed the maximum possible
256 * scaled window.
257 */
Rick Jones15d99e02006-03-20 22:40:29 -0800258 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 new_win = min(new_win, MAX_TCP_WINDOW);
260 else
261 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
262
263 /* RFC1323 scaling applied */
264 new_win >>= tp->rx_opt.rcv_wscale;
265
266 /* If we advertise zero window, disable fast path. */
267 if (new_win == 0)
268 tp->pred_flags = 0;
269
270 return new_win;
271}
272
Al Virodf7a3b02006-09-27 18:38:52 -0700273static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800274 __u32 tstamp, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800275{
276 if (tp->rx_opt.tstamp_ok) {
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700277 *ptr++ = htonl((TCPOPT_NOP << 24) |
278 (TCPOPT_NOP << 16) |
279 (TCPOPT_TIMESTAMP << 8) |
280 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800281 *ptr++ = htonl(tstamp);
282 *ptr++ = htonl(tp->rx_opt.ts_recent);
283 }
284 if (tp->rx_opt.eff_sacks) {
285 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
286 int this_sack;
287
288 *ptr++ = htonl((TCPOPT_NOP << 24) |
289 (TCPOPT_NOP << 16) |
290 (TCPOPT_SACK << 8) |
291 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
292 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800293
294 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800295 *ptr++ = htonl(sp[this_sack].start_seq);
296 *ptr++ = htonl(sp[this_sack].end_seq);
297 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800298
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800299 if (tp->rx_opt.dsack) {
300 tp->rx_opt.dsack = 0;
301 tp->rx_opt.eff_sacks--;
302 }
303 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800304#ifdef CONFIG_TCP_MD5SIG
305 if (md5_hash) {
306 *ptr++ = htonl((TCPOPT_NOP << 24) |
307 (TCPOPT_NOP << 16) |
308 (TCPOPT_MD5SIG << 8) |
309 TCPOLEN_MD5SIG);
310 *md5_hash = (__u8 *)ptr;
311 }
312#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800313}
314
315/* Construct a tcp options header for a SYN or SYN_ACK packet.
316 * If this is every changed make sure to change the definition of
317 * MAX_SYN_SIZE to match the new maximum number of options that you
318 * can generate.
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800319 *
320 * Note - that with the RFC2385 TCP option, we make room for the
321 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
322 * location to be filled is passed back up.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800323 */
Al Virodf7a3b02006-09-27 18:38:52 -0700324static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800325 int offer_wscale, int wscale, __u32 tstamp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800326 __u32 ts_recent, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800327{
328 /* We always get an MSS option.
329 * The option bytes which will be seen in normal data
330 * packets should timestamps be used, must be in the MSS
331 * advertised. But we subtract them from tp->mss_cache so
332 * that calculations in tcp_sendmsg are simpler etc.
333 * So account for this fact here if necessary. If we
334 * don't do this correctly, as a receiver we won't
335 * recognize data packets as being full sized when we
336 * should, and thus we won't abide by the delayed ACK
337 * rules correctly.
338 * SACKs don't matter, we never delay an ACK when we
339 * have any of those going out.
340 */
341 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
342 if (ts) {
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800343 if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700344 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
345 (TCPOLEN_SACK_PERM << 16) |
346 (TCPOPT_TIMESTAMP << 8) |
347 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800348 else
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700349 *ptr++ = htonl((TCPOPT_NOP << 24) |
350 (TCPOPT_NOP << 16) |
351 (TCPOPT_TIMESTAMP << 8) |
352 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800353 *ptr++ = htonl(tstamp); /* TSVAL */
354 *ptr++ = htonl(ts_recent); /* TSECR */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800355 } else if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700356 *ptr++ = htonl((TCPOPT_NOP << 24) |
357 (TCPOPT_NOP << 16) |
358 (TCPOPT_SACK_PERM << 8) |
359 TCPOLEN_SACK_PERM);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800360 if (offer_wscale)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700361 *ptr++ = htonl((TCPOPT_NOP << 24) |
362 (TCPOPT_WINDOW << 16) |
363 (TCPOLEN_WINDOW << 8) |
364 (wscale));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800365#ifdef CONFIG_TCP_MD5SIG
366 /*
367 * If MD5 is enabled, then we set the option, and include the size
368 * (always 18). The actual MD5 hash is added just before the
369 * packet is sent.
370 */
371 if (md5_hash) {
372 *ptr++ = htonl((TCPOPT_NOP << 24) |
373 (TCPOPT_NOP << 16) |
374 (TCPOPT_MD5SIG << 8) |
375 TCPOLEN_MD5SIG);
376 *md5_hash = (__u8 *) ptr;
377 }
378#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800379}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380
381/* This routine actually transmits TCP packets queued in by
382 * tcp_do_sendmsg(). This is used by both the initial
383 * transmission and possible later retransmissions.
384 * All SKB's seen here are completely headerless. It is our
385 * job to build the TCP header, and pass the packet down to
386 * IP so it can do the same plus pass the packet off to the
387 * device.
388 *
389 * We are working here with either a clone of the original
390 * SKB, or a fresh unique copy made by the retransmit engine.
391 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800392static 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 -0700393{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800394 const struct inet_connection_sock *icsk = inet_csk(sk);
395 struct inet_sock *inet;
396 struct tcp_sock *tp;
397 struct tcp_skb_cb *tcb;
398 int tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800399#ifdef CONFIG_TCP_MD5SIG
400 struct tcp_md5sig_key *md5;
401 __u8 *md5_hash_location;
402#endif
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800403 struct tcphdr *th;
404 int sysctl_flags;
405 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800407 BUG_ON(!skb || !tcp_skb_pcount(skb));
408
409 /* If congestion control is doing timestamping, we must
410 * take such a timestamp before we potentially clone/copy.
411 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700412 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800413 __net_timestamp(skb);
414
415 if (likely(clone_it)) {
416 if (unlikely(skb_cloned(skb)))
417 skb = pskb_copy(skb, gfp_mask);
418 else
419 skb = skb_clone(skb, gfp_mask);
420 if (unlikely(!skb))
421 return -ENOBUFS;
422 }
423
424 inet = inet_sk(sk);
425 tp = tcp_sk(sk);
426 tcb = TCP_SKB_CB(skb);
427 tcp_header_size = tp->tcp_header_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
429#define SYSCTL_FLAG_TSTAMPS 0x1
430#define SYSCTL_FLAG_WSCALE 0x2
431#define SYSCTL_FLAG_SACK 0x4
432
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800433 sysctl_flags = 0;
434 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
435 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800436 if (sysctl_tcp_timestamps) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800437 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
438 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800440 if (sysctl_tcp_window_scaling) {
441 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
442 sysctl_flags |= SYSCTL_FLAG_WSCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800444 if (sysctl_tcp_sack) {
445 sysctl_flags |= SYSCTL_FLAG_SACK;
446 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
447 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800449 } else if (unlikely(tp->rx_opt.eff_sacks)) {
450 /* A SACK is 2 pad bytes, a 2 byte header, plus
451 * 2 32-bit sequence numbers for each SACK block.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800453 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
454 (tp->rx_opt.eff_sacks *
455 TCPOLEN_SACK_PERBLOCK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900457
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800458 if (tcp_packets_in_flight(tp) == 0)
459 tcp_ca_event(sk, CA_EVENT_TX_START);
460
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800461#ifdef CONFIG_TCP_MD5SIG
462 /*
463 * Are we doing MD5 on this segment? If so - make
464 * room for it.
465 */
466 md5 = tp->af_specific->md5_lookup(sk, sk);
467 if (md5)
468 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
469#endif
470
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700471 skb_push(skb, tcp_header_size);
472 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800473 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800474
475 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700476 th = tcp_hdr(skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800477 th->source = inet->sport;
478 th->dest = inet->dport;
479 th->seq = htonl(tcb->seq);
480 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700481 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800482 tcb->flags);
483
484 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
485 /* RFC1323: The window in SYN & SYN/ACK segments
486 * is never scaled.
487 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800488 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800489 } else {
490 th->window = htons(tcp_select_window(sk));
491 }
492 th->check = 0;
493 th->urg_ptr = 0;
494
495 if (unlikely(tp->urg_mode &&
496 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
497 th->urg_ptr = htons(tp->snd_up-tcb->seq);
498 th->urg = 1;
499 }
500
501 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
Al Virodf7a3b02006-09-27 18:38:52 -0700502 tcp_syn_build_options((__be32 *)(th + 1),
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800503 tcp_advertise_mss(sk),
504 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
505 (sysctl_flags & SYSCTL_FLAG_SACK),
506 (sysctl_flags & SYSCTL_FLAG_WSCALE),
507 tp->rx_opt.rcv_wscale,
508 tcb->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800509 tp->rx_opt.ts_recent,
510
511#ifdef CONFIG_TCP_MD5SIG
512 md5 ? &md5_hash_location :
513#endif
514 NULL);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800515 } else {
Al Virodf7a3b02006-09-27 18:38:52 -0700516 tcp_build_and_update_options((__be32 *)(th + 1),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800517 tp, tcb->when,
518#ifdef CONFIG_TCP_MD5SIG
519 md5 ? &md5_hash_location :
520#endif
521 NULL);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700522 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800523 }
524
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800525#ifdef CONFIG_TCP_MD5SIG
526 /* Calculate the MD5 hash, as we have all we need now */
527 if (md5) {
528 tp->af_specific->calc_md5_hash(md5_hash_location,
529 md5,
530 sk, NULL, NULL,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700531 tcp_hdr(skb),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800532 sk->sk_protocol,
533 skb->len);
534 }
535#endif
536
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800537 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800538
539 if (likely(tcb->flags & TCPCB_FLAG_ACK))
540 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
541
542 if (skb->len != tcp_header_size)
543 tcp_event_data_sent(tp, skb, sk);
544
Wei Yongjunbd37a082006-08-07 21:04:15 -0700545 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
546 TCP_INC_STATS(TCP_MIB_OUTSEGS);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800547
David S. Millere89862f2007-01-26 01:04:55 -0800548 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700549 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800550 return err;
551
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800552 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800553
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200554 return net_xmit_eval(err);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800555
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556#undef SYSCTL_FLAG_TSTAMPS
557#undef SYSCTL_FLAG_WSCALE
558#undef SYSCTL_FLAG_SACK
559}
560
561
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900562/* This routine just queue's the buffer
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 *
564 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
565 * otherwise socket can stall.
566 */
567static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
568{
569 struct tcp_sock *tp = tcp_sk(sk);
570
571 /* Advance write_seq and place onto the write_queue. */
572 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
573 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800574 tcp_add_write_queue_tail(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 sk_charge_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576}
577
David S. Miller846998a2005-08-04 19:52:01 -0700578static 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 -0700579{
Herbert Xubcd76112006-06-30 13:36:35 -0700580 if (skb->len <= mss_now || !sk_can_gso(sk)) {
David S. Millerf6302d12005-07-05 15:18:03 -0700581 /* Avoid the costly divide in the normal
582 * non-TSO case.
583 */
Herbert Xu79671682006-06-22 02:40:14 -0700584 skb_shinfo(skb)->gso_segs = 1;
585 skb_shinfo(skb)->gso_size = 0;
586 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700587 } else {
588 unsigned int factor;
589
David S. Miller846998a2005-08-04 19:52:01 -0700590 factor = skb->len + (mss_now - 1);
591 factor /= mss_now;
Herbert Xu79671682006-06-22 02:40:14 -0700592 skb_shinfo(skb)->gso_segs = factor;
593 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700594 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 }
596}
597
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598/* Function to create two new TCP segments. Shrinks the given segment
599 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900600 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 * Remember, these are still headerless SKBs at this point.
602 */
David S. Miller6475be12005-09-01 22:47:01 -0700603int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604{
605 struct tcp_sock *tp = tcp_sk(sk);
606 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700607 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700608 int nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 u16 flags;
610
Herbert Xub2cc99f2005-10-20 17:13:13 -0200611 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800612
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900613 clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 nsize = skb_headlen(skb) - len;
615 if (nsize < 0)
616 nsize = 0;
617
618 if (skb_cloned(skb) &&
619 skb_is_nonlinear(skb) &&
620 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
621 return -ENOMEM;
622
623 /* Get a new skb... force flag on. */
624 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
625 if (buff == NULL)
626 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -0700627
Herbert Xub60b49e2006-04-19 21:35:00 -0700628 sk_charge_skb(sk, buff);
629 nlen = skb->len - len - nsize;
630 buff->truesize += nlen;
631 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
633 /* Correct the sequence numbers. */
634 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
635 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
636 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
637
638 /* PSH and FIN should only be set in the second packet. */
639 flags = TCP_SKB_CB(skb)->flags;
640 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
641 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700642 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
644
Patrick McHardy84fa7932006-08-29 16:44:56 -0700645 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 /* Copy and checksum data tail into the new buffer. */
647 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
648 nsize, 0);
649
650 skb_trim(skb, len);
651
652 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
653 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -0700654 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 skb_split(skb, buff, len);
656 }
657
658 buff->ip_summed = skb->ip_summed;
659
660 /* Looks stupid, but our code really uses when of
661 * skbs, which it never sent before. --ANK
662 */
663 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700664 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665
David S. Miller6475be12005-09-01 22:47:01 -0700666 old_factor = tcp_skb_pcount(skb);
667
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700669 tcp_set_skb_tso_segs(sk, skb, mss_now);
670 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
David S. Miller6475be12005-09-01 22:47:01 -0700672 /* If this packet has been sent out already, we must
673 * adjust the various packet counters.
674 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700675 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700676 int diff = old_factor - tcp_skb_pcount(skb) -
677 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
David S. Miller6475be12005-09-01 22:47:01 -0700679 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700680
681 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
682 tp->sacked_out -= diff;
683 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
684 tp->retrans_out -= diff;
685
David S. Miller6475be12005-09-01 22:47:01 -0700686 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
687 tp->lost_out -= diff;
688 tp->left_out -= diff;
689 }
Herbert Xu83ca28b2005-09-22 23:32:56 -0700690
David S. Miller6475be12005-09-01 22:47:01 -0700691 if (diff > 0) {
Herbert Xu83ca28b2005-09-22 23:32:56 -0700692 /* Adjust Reno SACK estimate. */
693 if (!tp->rx_opt.sack_ok) {
694 tp->sacked_out -= diff;
695 if ((int)tp->sacked_out < 0)
696 tp->sacked_out = 0;
697 tcp_sync_left_out(tp);
698 }
699
David S. Miller6475be12005-09-01 22:47:01 -0700700 tp->fackets_out -= diff;
701 if ((int)tp->fackets_out < 0)
702 tp->fackets_out = 0;
703 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 }
705
706 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700707 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -0800708 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
710 return 0;
711}
712
713/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
714 * eventually). The difference is that pulled data not copied, but
715 * immediately discarded.
716 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700717static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718{
719 int i, k, eat;
720
721 eat = len;
722 k = 0;
723 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
724 if (skb_shinfo(skb)->frags[i].size <= eat) {
725 put_page(skb_shinfo(skb)->frags[i].page);
726 eat -= skb_shinfo(skb)->frags[i].size;
727 } else {
728 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
729 if (eat) {
730 skb_shinfo(skb)->frags[k].page_offset += eat;
731 skb_shinfo(skb)->frags[k].size -= eat;
732 eat = 0;
733 }
734 k++;
735 }
736 }
737 skb_shinfo(skb)->nr_frags = k;
738
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -0700739 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 skb->data_len -= len;
741 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742}
743
744int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
745{
746 if (skb_cloned(skb) &&
747 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
748 return -ENOMEM;
749
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700750 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
751 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700753 else
754 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755
756 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -0700757 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758
759 skb->truesize -= len;
760 sk->sk_wmem_queued -= len;
761 sk->sk_forward_alloc += len;
762 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
763
764 /* Any change of skb->len requires recalculation of tso
765 * factor and mss.
766 */
767 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700768 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769
770 return 0;
771}
772
John Heffner5d424d52006-03-20 17:53:41 -0800773/* Not accounting for SACKs here. */
774int tcp_mtu_to_mss(struct sock *sk, int pmtu)
775{
776 struct tcp_sock *tp = tcp_sk(sk);
777 struct inet_connection_sock *icsk = inet_csk(sk);
778 int mss_now;
779
780 /* Calculate base mss without TCP options:
781 It is MMS_S - sizeof(tcphdr) of rfc1122
782 */
783 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
784
785 /* Clamp it (mss_clamp does not include tcp options) */
786 if (mss_now > tp->rx_opt.mss_clamp)
787 mss_now = tp->rx_opt.mss_clamp;
788
789 /* Now subtract optional transport overhead */
790 mss_now -= icsk->icsk_ext_hdr_len;
791
792 /* Then reserve room for full set of TCP options and 8 bytes of data */
793 if (mss_now < 48)
794 mss_now = 48;
795
796 /* Now subtract TCP options size, not including SACKs */
797 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
798
799 return mss_now;
800}
801
802/* Inverse of above */
803int tcp_mss_to_mtu(struct sock *sk, int mss)
804{
805 struct tcp_sock *tp = tcp_sk(sk);
806 struct inet_connection_sock *icsk = inet_csk(sk);
807 int mtu;
808
809 mtu = mss +
810 tp->tcp_header_len +
811 icsk->icsk_ext_hdr_len +
812 icsk->icsk_af_ops->net_header_len;
813
814 return mtu;
815}
816
817void tcp_mtup_init(struct sock *sk)
818{
819 struct tcp_sock *tp = tcp_sk(sk);
820 struct inet_connection_sock *icsk = inet_csk(sk);
821
822 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
823 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900824 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -0800825 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
826 icsk->icsk_mtup.probe_size = 0;
827}
828
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829/* This function synchronize snd mss to current pmtu/exthdr set.
830
831 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
832 for TCP options, but includes only bare TCP header.
833
834 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800835 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 It also does not include TCP options.
837
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800838 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839
840 tp->mss_cache is current effective sending mss, including
841 all tcp options except for SACKs. It is evaluated,
842 taking into account current pmtu, but never exceeds
843 tp->rx_opt.mss_clamp.
844
845 NOTE1. rfc1122 clearly states that advertised MSS
846 DOES NOT include either tcp or ip options.
847
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800848 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
849 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 */
851
852unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
853{
854 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800855 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -0800856 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857
John Heffner5d424d52006-03-20 17:53:41 -0800858 if (icsk->icsk_mtup.search_high > pmtu)
859 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860
John Heffner5d424d52006-03-20 17:53:41 -0800861 mss_now = tcp_mtu_to_mss(sk, pmtu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862
863 /* Bound mss with half of window */
864 if (tp->max_window && mss_now > (tp->max_window>>1))
865 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
866
867 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800868 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -0800869 if (icsk->icsk_mtup.enabled)
870 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700871 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872
873 return mss_now;
874}
875
876/* Compute the current effective MSS, taking SACKs and IP options,
877 * and even PMTU discovery events into account.
878 *
879 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
880 * cannot be large. However, taking into account rare use of URG, this
881 * is not a big flaw.
882 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700883unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884{
885 struct tcp_sock *tp = tcp_sk(sk);
886 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700887 u32 mss_now;
888 u16 xmit_size_goal;
889 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700891 mss_now = tp->mss_cache;
892
Herbert Xubcd76112006-06-30 13:36:35 -0700893 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700894 doing_tso = 1;
895
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 if (dst) {
897 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800898 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 mss_now = tcp_sync_mss(sk, mtu);
900 }
901
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 if (tp->rx_opt.eff_sacks)
903 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
904 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700905
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800906#ifdef CONFIG_TCP_MD5SIG
907 if (tp->af_specific->md5_lookup(sk, sk))
908 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
909#endif
910
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700911 xmit_size_goal = mss_now;
912
913 if (doing_tso) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800914 xmit_size_goal = (65535 -
915 inet_csk(sk)->icsk_af_ops->net_header_len -
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800916 inet_csk(sk)->icsk_ext_hdr_len -
917 tp->tcp_header_len);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700918
919 if (tp->max_window &&
920 (xmit_size_goal > (tp->max_window >> 1)))
921 xmit_size_goal = max((tp->max_window >> 1),
922 68U - tp->tcp_header_len);
923
924 xmit_size_goal -= (xmit_size_goal % mss_now);
925 }
926 tp->xmit_size_goal = xmit_size_goal;
927
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 return mss_now;
929}
930
David S. Millera762a982005-07-05 15:18:51 -0700931/* Congestion window validation. (RFC2861) */
932
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700933static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -0700934{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700935 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -0700936 __u32 packets_out = tp->packets_out;
937
938 if (packets_out >= tp->snd_cwnd) {
939 /* Network is feed fully. */
940 tp->snd_cwnd_used = 0;
941 tp->snd_cwnd_stamp = tcp_time_stamp;
942 } else {
943 /* Network starves. */
944 if (tp->packets_out > tp->snd_cwnd_used)
945 tp->snd_cwnd_used = tp->packets_out;
946
David S. Miller15d33c02007-04-09 13:23:14 -0700947 if (sysctl_tcp_slow_start_after_idle &&
948 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -0700949 tcp_cwnd_application_limited(sk);
950 }
951}
952
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700953static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
954{
955 u32 window, cwnd_len;
956
957 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
958 cwnd_len = mss_now * cwnd;
959 return min(window, cwnd_len);
960}
961
962/* Can at least one segment of SKB be sent right now, according to the
963 * congestion window rules? If so, return how many segments are allowed.
964 */
965static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
966{
967 u32 in_flight, cwnd;
968
969 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -0800970 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
971 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700972 return 1;
973
974 in_flight = tcp_packets_in_flight(tp);
975 cwnd = tp->snd_cwnd;
976 if (in_flight < cwnd)
977 return (cwnd - in_flight);
978
979 return 0;
980}
981
982/* This must be invoked the first time we consider transmitting
983 * SKB onto the wire.
984 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800985static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700986{
987 int tso_segs = tcp_skb_pcount(skb);
988
David S. Miller846998a2005-08-04 19:52:01 -0700989 if (!tso_segs ||
990 (tso_segs > 1 &&
Herbert Xu79671682006-06-22 02:40:14 -0700991 tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -0700992 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700993 tso_segs = tcp_skb_pcount(skb);
994 }
995 return tso_segs;
996}
997
998static inline int tcp_minshall_check(const struct tcp_sock *tp)
999{
1000 return after(tp->snd_sml,tp->snd_una) &&
1001 !after(tp->snd_sml, tp->snd_nxt);
1002}
1003
1004/* Return 0, if packet can be sent now without violation Nagle's rules:
1005 * 1. It is full sized.
1006 * 2. Or it contains FIN. (already checked by caller)
1007 * 3. Or TCP_NODELAY was set.
1008 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1009 * With Minshall's modification: all sent small packets are ACKed.
1010 */
1011
1012static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001013 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001014 unsigned mss_now, int nonagle)
1015{
1016 return (skb->len < mss_now &&
1017 ((nonagle&TCP_NAGLE_CORK) ||
1018 (!nonagle &&
1019 tp->packets_out &&
1020 tcp_minshall_check(tp))));
1021}
1022
1023/* Return non-zero if the Nagle test allows this packet to be
1024 * sent now.
1025 */
1026static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1027 unsigned int cur_mss, int nonagle)
1028{
1029 /* Nagle rule does not apply to frames, which sit in the middle of the
1030 * write_queue (they have no chances to get new data).
1031 *
1032 * This is implemented in the callers, where they modify the 'nonagle'
1033 * argument based upon the location of SKB in the send queue.
1034 */
1035 if (nonagle & TCP_NAGLE_PUSH)
1036 return 1;
1037
1038 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1039 if (tp->urg_mode ||
1040 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1041 return 1;
1042
1043 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1044 return 1;
1045
1046 return 0;
1047}
1048
1049/* Does at least the first segment of SKB fit into the send window? */
1050static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
1051{
1052 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1053
1054 if (skb->len > cur_mss)
1055 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1056
1057 return !after(end_seq, tp->snd_una + tp->snd_wnd);
1058}
1059
David S. Millerfe067e82007-03-07 12:12:44 -08001060/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001061 * should be put on the wire right now. If so, it returns the number of
1062 * packets allowed by the congestion window.
1063 */
1064static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1065 unsigned int cur_mss, int nonagle)
1066{
1067 struct tcp_sock *tp = tcp_sk(sk);
1068 unsigned int cwnd_quota;
1069
David S. Miller846998a2005-08-04 19:52:01 -07001070 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001071
1072 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1073 return 0;
1074
1075 cwnd_quota = tcp_cwnd_test(tp, skb);
1076 if (cwnd_quota &&
1077 !tcp_snd_wnd_test(tp, skb, cur_mss))
1078 cwnd_quota = 0;
1079
1080 return cwnd_quota;
1081}
1082
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001083int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001084{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001085 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001086 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001087
1088 return (skb &&
1089 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1090 (tcp_skb_is_last(sk, skb) ?
1091 TCP_NAGLE_PUSH :
1092 tp->nonagle)));
1093}
1094
1095/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1096 * which is put after SKB on the list. It is very much like
1097 * tcp_fragment() except that it may make several kinds of assumptions
1098 * in order to speed up the splitting operation. In particular, we
1099 * know that all the data is in scatter-gather pages, and that the
1100 * packet has never been sent out before (and thus is not cloned).
1101 */
David S. Miller846998a2005-08-04 19:52:01 -07001102static 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 -07001103{
1104 struct sk_buff *buff;
1105 int nlen = skb->len - len;
1106 u16 flags;
1107
1108 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001109 if (skb->len != skb->data_len)
1110 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001111
1112 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1113 if (unlikely(buff == NULL))
1114 return -ENOMEM;
1115
Herbert Xub60b49e2006-04-19 21:35:00 -07001116 sk_charge_skb(sk, buff);
1117 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001118 skb->truesize -= nlen;
1119
1120 /* Correct the sequence numbers. */
1121 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1122 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1123 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1124
1125 /* PSH and FIN should only be set in the second packet. */
1126 flags = TCP_SKB_CB(skb)->flags;
1127 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1128 TCP_SKB_CB(buff)->flags = flags;
1129
1130 /* This packet was never sent out yet, so no SACK bits. */
1131 TCP_SKB_CB(buff)->sacked = 0;
1132
Patrick McHardy84fa7932006-08-29 16:44:56 -07001133 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001134 skb_split(skb, buff, len);
1135
1136 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001137 tcp_set_skb_tso_segs(sk, skb, mss_now);
1138 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001139
1140 /* Link BUFF into the send queue. */
1141 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001142 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001143
1144 return 0;
1145}
1146
1147/* Try to defer sending, if possible, in order to minimize the amount
1148 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1149 *
1150 * This algorithm is from John Heffner.
1151 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001152static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001153{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001154 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001155 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001156 u32 send_win, cong_win, limit, in_flight;
1157
1158 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001159 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001160
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001161 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001162 goto send_now;
1163
1164 /* Defer for less than two clock ticks. */
1165 if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1)
1166 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001167
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001168 in_flight = tcp_packets_in_flight(tp);
1169
1170 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1171 (tp->snd_cwnd <= in_flight));
1172
1173 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1174
1175 /* From in_flight test above, we know that cwnd > in_flight. */
1176 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1177
1178 limit = min(send_win, cong_win);
1179
David S. Millerba244fe2006-03-11 18:51:49 -08001180 /* If a full-sized TSO skb can be sent, do it. */
1181 if (limit >= 65536)
John Heffnerae8064a2006-10-18 20:36:48 -07001182 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001183
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001184 if (sysctl_tcp_tso_win_divisor) {
1185 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1186
1187 /* If at least some fraction of a window is available,
1188 * just use it.
1189 */
1190 chunk /= sysctl_tcp_tso_win_divisor;
1191 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001192 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001193 } else {
1194 /* Different approach, try not to defer past a single
1195 * ACK. Receiver should ACK every other full sized
1196 * frame, so if we have space for more than 3 frames
1197 * then send now.
1198 */
1199 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001200 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001201 }
1202
1203 /* Ok, it looks like it is advisable to defer. */
John Heffnerae8064a2006-10-18 20:36:48 -07001204 tp->tso_deferred = 1 | (jiffies<<1);
1205
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001206 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001207
1208send_now:
1209 tp->tso_deferred = 0;
1210 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001211}
1212
John Heffner5d424d52006-03-20 17:53:41 -08001213/* Create a new MTU probe if we are ready.
1214 * Returns 0 if we should wait to probe (no cwnd available),
1215 * 1 if a probe was sent,
1216 * -1 otherwise */
1217static int tcp_mtu_probe(struct sock *sk)
1218{
1219 struct tcp_sock *tp = tcp_sk(sk);
1220 struct inet_connection_sock *icsk = inet_csk(sk);
1221 struct sk_buff *skb, *nskb, *next;
1222 int len;
1223 int probe_size;
1224 unsigned int pif;
1225 int copy;
1226 int mss_now;
1227
1228 /* Not currently probing/verifying,
1229 * not in recovery,
1230 * have enough cwnd, and
1231 * not SACKing (the variable headers throw things off) */
1232 if (!icsk->icsk_mtup.enabled ||
1233 icsk->icsk_mtup.probe_size ||
1234 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1235 tp->snd_cwnd < 11 ||
1236 tp->rx_opt.eff_sacks)
1237 return -1;
1238
1239 /* Very simple search strategy: just double the MSS. */
1240 mss_now = tcp_current_mss(sk, 0);
1241 probe_size = 2*tp->mss_cache;
1242 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1243 /* TODO: set timer for probe_converge_event */
1244 return -1;
1245 }
1246
1247 /* Have enough data in the send queue to probe? */
1248 len = 0;
David S. Millerfe067e82007-03-07 12:12:44 -08001249 if ((skb = tcp_send_head(sk)) == NULL)
John Heffner5d424d52006-03-20 17:53:41 -08001250 return -1;
1251 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
David S. Millerfe067e82007-03-07 12:12:44 -08001252 skb = tcp_write_queue_next(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001253 if (len < probe_size)
1254 return -1;
1255
1256 /* Receive window check. */
1257 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1258 if (tp->snd_wnd < probe_size)
1259 return -1;
1260 else
1261 return 0;
1262 }
1263
1264 /* Do we need to wait to drain cwnd? */
1265 pif = tcp_packets_in_flight(tp);
1266 if (pif + 2 > tp->snd_cwnd) {
1267 /* With no packets in flight, don't stall. */
1268 if (pif == 0)
1269 return -1;
1270 else
1271 return 0;
1272 }
1273
1274 /* We're allowed to probe. Build it now. */
1275 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1276 return -1;
1277 sk_charge_skb(sk, nskb);
1278
David S. Millerfe067e82007-03-07 12:12:44 -08001279 skb = tcp_send_head(sk);
1280 tcp_insert_write_queue_before(nskb, skb, sk);
1281 tcp_advance_send_head(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001282
1283 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1284 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1285 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1286 TCP_SKB_CB(nskb)->sacked = 0;
1287 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001288 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001289
1290 len = 0;
1291 while (len < probe_size) {
David S. Millerfe067e82007-03-07 12:12:44 -08001292 next = tcp_write_queue_next(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001293
1294 copy = min_t(int, skb->len, probe_size - len);
1295 if (nskb->ip_summed)
1296 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1297 else
1298 nskb->csum = skb_copy_and_csum_bits(skb, 0,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001299 skb_put(nskb, copy), copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001300
1301 if (skb->len <= copy) {
1302 /* We've eaten all the data from this skb.
1303 * Throw it away. */
1304 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001305 tcp_unlink_write_queue(skb, sk);
John Heffner5d424d52006-03-20 17:53:41 -08001306 sk_stream_free_skb(sk, skb);
1307 } else {
1308 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001309 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001310 if (!skb_shinfo(skb)->nr_frags) {
1311 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001312 if (skb->ip_summed != CHECKSUM_PARTIAL)
John Heffner5d424d52006-03-20 17:53:41 -08001313 skb->csum = csum_partial(skb->data, skb->len, 0);
1314 } else {
1315 __pskb_trim_head(skb, copy);
1316 tcp_set_skb_tso_segs(sk, skb, mss_now);
1317 }
1318 TCP_SKB_CB(skb)->seq += copy;
1319 }
1320
1321 len += copy;
1322 skb = next;
1323 }
1324 tcp_init_tso_segs(sk, nskb, nskb->len);
1325
1326 /* We're ready to send. If this fails, the probe will
1327 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1328 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1329 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1330 /* Decrement cwnd here because we are sending
1331 * effectively two packets. */
1332 tp->snd_cwnd--;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001333 update_send_head(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001334
1335 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001336 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1337 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001338
1339 return 1;
1340 }
1341
1342 return -1;
1343}
1344
1345
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346/* This routine writes packets to the network. It advances the
1347 * send_head. This happens as incoming acks open up the remote
1348 * window for us.
1349 *
1350 * Returns 1, if no segments are in flight and we have queued segments, but
1351 * cannot send anything now because of SWS or another problem.
1352 */
David S. Millera2e2a592005-07-05 15:19:23 -07001353static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354{
1355 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001356 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001357 unsigned int tso_segs, sent_pkts;
1358 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001359 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360
1361 /* If we are closed, the bytes will have to remain here.
1362 * In time closedown will finish, we empty the write queue and all
1363 * will be happy.
1364 */
David S. Miller92df7b52005-07-05 15:19:06 -07001365 if (unlikely(sk->sk_state == TCP_CLOSE))
1366 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367
David S. Miller92df7b52005-07-05 15:19:06 -07001368 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001369
1370 /* Do MTU probing. */
1371 if ((result = tcp_mtu_probe(sk)) == 0) {
1372 return 0;
1373 } else if (result > 0) {
1374 sent_pkts = 1;
1375 }
1376
David S. Millerfe067e82007-03-07 12:12:44 -08001377 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001378 unsigned int limit;
1379
Herbert Xub68e9f82005-08-04 19:52:02 -07001380 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001381 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001382
Herbert Xub68e9f82005-08-04 19:52:02 -07001383 cwnd_quota = tcp_cwnd_test(tp, skb);
1384 if (!cwnd_quota)
1385 break;
1386
1387 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1388 break;
1389
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001390 if (tso_segs == 1) {
1391 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1392 (tcp_skb_is_last(sk, skb) ?
1393 nonagle : TCP_NAGLE_PUSH))))
1394 break;
1395 } else {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001396 if (tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001397 break;
1398 }
David S. Milleraa934662005-07-05 15:20:09 -07001399
Herbert Xuc8ac3772005-08-16 20:43:40 -07001400 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001401 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001402 limit = tcp_window_allows(tp, skb,
1403 mss_now, cwnd_quota);
David S. Milleraa934662005-07-05 15:20:09 -07001404
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001405 if (skb->len < limit) {
1406 unsigned int trim = skb->len % mss_now;
1407
1408 if (trim)
1409 limit = skb->len - trim;
1410 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 }
1412
Herbert Xuc8ac3772005-08-16 20:43:40 -07001413 if (skb->len > limit &&
1414 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1415 break;
1416
David S. Miller92df7b52005-07-05 15:19:06 -07001417 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001418
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001419 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
David S. Miller92df7b52005-07-05 15:19:06 -07001420 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421
David S. Miller92df7b52005-07-05 15:19:06 -07001422 /* Advance the send_head. This one is sent out.
1423 * This call will increment packets_out.
1424 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001425 update_send_head(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001426
1427 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001428 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429 }
David S. Miller92df7b52005-07-05 15:19:06 -07001430
David S. Milleraa934662005-07-05 15:20:09 -07001431 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001432 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001433 return 0;
1434 }
David S. Millerfe067e82007-03-07 12:12:44 -08001435 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436}
1437
David S. Millera762a982005-07-05 15:18:51 -07001438/* Push out any pending frames which were held back due to
1439 * TCP_CORK or attempt at coalescing tiny packets.
1440 * The socket must be locked by the caller.
1441 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001442void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1443 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001444{
David S. Millerfe067e82007-03-07 12:12:44 -08001445 struct sk_buff *skb = tcp_send_head(sk);
David S. Millera762a982005-07-05 15:18:51 -07001446
1447 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001448 if (tcp_write_xmit(sk, cur_mss, nonagle))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001449 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001450 }
1451}
1452
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001453/* Send _single_ skb sitting at the send head. This function requires
1454 * true push pending frames to setup probe timer etc.
1455 */
1456void tcp_push_one(struct sock *sk, unsigned int mss_now)
1457{
1458 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001459 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001460 unsigned int tso_segs, cwnd_quota;
1461
1462 BUG_ON(!skb || skb->len < mss_now);
1463
David S. Miller846998a2005-08-04 19:52:01 -07001464 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001465 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1466
1467 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001468 unsigned int limit;
1469
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001470 BUG_ON(!tso_segs);
1471
Herbert Xuc8ac3772005-08-16 20:43:40 -07001472 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001473 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001474 limit = tcp_window_allows(tp, skb,
1475 mss_now, cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001476
1477 if (skb->len < limit) {
1478 unsigned int trim = skb->len % mss_now;
1479
1480 if (trim)
1481 limit = skb->len - trim;
1482 }
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001483 }
1484
Herbert Xuc8ac3772005-08-16 20:43:40 -07001485 if (skb->len > limit &&
1486 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1487 return;
1488
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001489 /* Send it out now. */
1490 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1491
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001492 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001493 update_send_head(sk, skb);
1494 tcp_cwnd_validate(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001495 return;
1496 }
1497 }
1498}
1499
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500/* This function returns the amount that we can raise the
1501 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001502 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 * 1. The window can never be shrunk once it is offered (RFC 793)
1504 * 2. We limit memory per socket
1505 *
1506 * RFC 1122:
1507 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1508 * RECV.NEXT + RCV.WIN fixed until:
1509 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1510 *
1511 * i.e. don't raise the right edge of the window until you can raise
1512 * it at least MSS bytes.
1513 *
1514 * Unfortunately, the recommended algorithm breaks header prediction,
1515 * since header prediction assumes th->window stays fixed.
1516 *
1517 * Strictly speaking, keeping th->window fixed violates the receiver
1518 * side SWS prevention criteria. The problem is that under this rule
1519 * a stream of single byte packets will cause the right side of the
1520 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001521 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 * Of course, if the sender implements sender side SWS prevention
1523 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001524 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001526 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 * If the free space is less than the 1/4 of the maximum
1528 * space available and the free space is less than 1/2 mss,
1529 * then set the window to 0.
1530 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1531 * Otherwise, just prevent the window from shrinking
1532 * and from being larger than the largest representable value.
1533 *
1534 * This prevents incremental opening of the window in the regime
1535 * where TCP is limited by the speed of the reader side taking
1536 * data out of the TCP receive queue. It does nothing about
1537 * those cases where the window is constrained on the sender side
1538 * because the pipeline is full.
1539 *
1540 * BSD also seems to "accidentally" limit itself to windows that are a
1541 * multiple of MSS, at least until the free space gets quite small.
1542 * This would appear to be a side effect of the mbuf implementation.
1543 * Combining these two algorithms results in the observed behavior
1544 * of having a fixed window size at almost all times.
1545 *
1546 * Below we obtain similar behavior by forcing the offered window to
1547 * a multiple of the mss when it is feasible to do so.
1548 *
1549 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1550 * Regular options like TIMESTAMP are taken into account.
1551 */
1552u32 __tcp_select_window(struct sock *sk)
1553{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001554 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001556 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557 * here. I don't know if the value based on our guesses
1558 * of peer's MSS is better for the performance. It's more correct
1559 * but may be worse for the performance because of rcv_mss
1560 * fluctuations. --SAW 1998/11/1
1561 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001562 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 int free_space = tcp_space(sk);
1564 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1565 int window;
1566
1567 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001568 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569
1570 if (free_space < full_space/2) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001571 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572
1573 if (tcp_memory_pressure)
1574 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1575
1576 if (free_space < mss)
1577 return 0;
1578 }
1579
1580 if (free_space > tp->rcv_ssthresh)
1581 free_space = tp->rcv_ssthresh;
1582
1583 /* Don't do rounding if we are using window scaling, since the
1584 * scaled window will not line up with the MSS boundary anyway.
1585 */
1586 window = tp->rcv_wnd;
1587 if (tp->rx_opt.rcv_wscale) {
1588 window = free_space;
1589
1590 /* Advertise enough space so that it won't get scaled away.
1591 * Import case: prevent zero window announcement if
1592 * 1<<rcv_wscale > mss.
1593 */
1594 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1595 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1596 << tp->rx_opt.rcv_wscale);
1597 } else {
1598 /* Get the largest window that is a nice multiple of mss.
1599 * Window clamp already applied above.
1600 * If our current window offering is within 1 mss of the
1601 * free space we just keep it. This prevents the divide
1602 * and multiply from happening most of the time.
1603 * We also don't do any window rounding when the free space
1604 * is too small.
1605 */
1606 if (window <= free_space - mss || window > free_space)
1607 window = (free_space/mss)*mss;
John Heffner84565072007-04-02 13:56:32 -07001608 else if (mss == full_space &&
1609 free_space > window + full_space/2)
1610 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 }
1612
1613 return window;
1614}
1615
1616/* Attempt to collapse two adjacent SKB's during retransmission. */
1617static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1618{
1619 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001620 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
1622 /* The first test we must make is that neither of these two
1623 * SKB's are still referenced by someone else.
1624 */
1625 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1626 int skb_size = skb->len, next_skb_size = next_skb->len;
1627 u16 flags = TCP_SKB_CB(skb)->flags;
1628
1629 /* Also punt if next skb has been SACK'd. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001630 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 return;
1632
1633 /* Next skb is out of window. */
1634 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1635 return;
1636
1637 /* Punt if not enough space exists in the first SKB for
1638 * the data in the second, or the total combined payload
1639 * would exceed the MSS.
1640 */
1641 if ((next_skb_size > skb_tailroom(skb)) ||
1642 ((skb_size + next_skb_size) > mss_now))
1643 return;
1644
1645 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1646 tcp_skb_pcount(next_skb) != 1);
1647
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001648 /* changing transmit queue under us so clear hints */
1649 clear_all_retrans_hints(tp);
1650
1651 /* Ok. We will be able to collapse the packet. */
David S. Millerfe067e82007-03-07 12:12:44 -08001652 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653
Arnaldo Carvalho de Melo1a4e2d02007-03-31 11:55:45 -03001654 skb_copy_from_linear_data(next_skb,
1655 skb_put(skb, next_skb_size),
1656 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657
Jarek Poplawski52d570a2007-01-23 22:07:12 -08001658 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1659 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660
Patrick McHardy84fa7932006-08-29 16:44:56 -07001661 if (skb->ip_summed != CHECKSUM_PARTIAL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1663
1664 /* Update sequence range on original skb. */
1665 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1666
1667 /* Merge over control information. */
1668 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1669 TCP_SKB_CB(skb)->flags = flags;
1670
1671 /* All done, get rid of second SKB and account for it so
1672 * packet counting does not break.
1673 */
1674 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1675 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1676 tp->retrans_out -= tcp_skb_pcount(next_skb);
1677 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1678 tp->lost_out -= tcp_skb_pcount(next_skb);
1679 tp->left_out -= tcp_skb_pcount(next_skb);
1680 }
1681 /* Reno case is special. Sigh... */
1682 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1683 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1684 tp->left_out -= tcp_skb_pcount(next_skb);
1685 }
1686
1687 /* Not quite right: it can be > snd.fack, but
1688 * it is better to underestimate fackets.
1689 */
1690 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1691 tcp_packets_out_dec(tp, next_skb);
1692 sk_stream_free_skb(sk, next_skb);
1693 }
1694}
1695
1696/* Do a simple retransmit without using the backoff mechanisms in
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001697 * tcp_timer. This is used for path mtu discovery.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 * The socket is already locked here.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001699 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700void tcp_simple_retransmit(struct sock *sk)
1701{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001702 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 struct tcp_sock *tp = tcp_sk(sk);
1704 struct sk_buff *skb;
1705 unsigned int mss = tcp_current_mss(sk, 0);
1706 int lost = 0;
1707
David S. Millerfe067e82007-03-07 12:12:44 -08001708 tcp_for_write_queue(skb, sk) {
1709 if (skb == tcp_send_head(sk))
1710 break;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001711 if (skb->len > mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1713 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1714 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1715 tp->retrans_out -= tcp_skb_pcount(skb);
1716 }
1717 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1718 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1719 tp->lost_out += tcp_skb_pcount(skb);
1720 lost = 1;
1721 }
1722 }
1723 }
1724
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001725 clear_all_retrans_hints(tp);
1726
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 if (!lost)
1728 return;
1729
1730 tcp_sync_left_out(tp);
1731
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001732 /* Don't muck with the congestion window here.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733 * Reason is that we do not increase amount of _data_
1734 * in network, but units changed and effective
1735 * cwnd/ssthresh really reduced now.
1736 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001737 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001739 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740 tp->prior_ssthresh = 0;
1741 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001742 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 }
1744 tcp_xmit_retransmit_queue(sk);
1745}
1746
1747/* This retransmits one SKB. Policy decisions and retransmit queue
1748 * state updates are done by the caller. Returns non-zero if an
1749 * error occurred which prevented the send.
1750 */
1751int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1752{
1753 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001754 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001755 unsigned int cur_mss = tcp_current_mss(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 int err;
1757
John Heffner5d424d52006-03-20 17:53:41 -08001758 /* Inconslusive MTU probe */
1759 if (icsk->icsk_mtup.probe_size) {
1760 icsk->icsk_mtup.probe_size = 0;
1761 }
1762
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001764 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 */
1766 if (atomic_read(&sk->sk_wmem_alloc) >
1767 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1768 return -EAGAIN;
1769
1770 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1771 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1772 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1774 return -ENOMEM;
1775 }
1776
1777 /* If receiver has shrunk his window, and skb is out of
1778 * new window, do not retransmit it. The exception is the
1779 * case, when window is shrunk to zero. In this case
1780 * our retransmit serves as a zero window probe.
1781 */
1782 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1783 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1784 return -EAGAIN;
1785
1786 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001787 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 }
1790
1791 /* Collapse two adjacent packets if worthwhile and we can. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001792 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1793 (skb->len < (cur_mss >> 1)) &&
1794 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1795 (!tcp_skb_is_last(sk, skb)) &&
1796 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1797 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1798 (sysctl_tcp_retrans_collapse != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 tcp_retrans_try_collapse(sk, skb, cur_mss);
1800
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001801 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 return -EHOSTUNREACH; /* Routing failure or similar. */
1803
1804 /* Some Solaris stacks overoptimize and ignore the FIN on a
1805 * retransmit when old data is attached. So strip it off
1806 * since it is cheap to do so and saves bytes on the network.
1807 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001808 if (skb->len > 0 &&
1809 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1810 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 if (!pskb_trim(skb, 0)) {
1812 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
Herbert Xu79671682006-06-22 02:40:14 -07001813 skb_shinfo(skb)->gso_segs = 1;
1814 skb_shinfo(skb)->gso_size = 0;
1815 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 skb->ip_summed = CHECKSUM_NONE;
1817 skb->csum = 0;
1818 }
1819 }
1820
1821 /* Make a copy, if the first transmission SKB clone we made
1822 * is still in somebody's hands, else make a clone.
1823 */
1824 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001826 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827
1828 if (err == 0) {
1829 /* Update global TCP statistics. */
1830 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1831
1832 tp->total_retrans++;
1833
1834#if FASTRETRANS_DEBUG > 0
1835 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1836 if (net_ratelimit())
1837 printk(KERN_DEBUG "retrans_out leaked.\n");
1838 }
1839#endif
1840 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1841 tp->retrans_out += tcp_skb_pcount(skb);
1842
1843 /* Save stamp of the first retransmit. */
1844 if (!tp->retrans_stamp)
1845 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1846
1847 tp->undo_retrans++;
1848
1849 /* snd_nxt is stored to detect loss of retransmitted segment,
1850 * see tcp_input.c tcp_sacktag_write_queue().
1851 */
1852 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1853 }
1854 return err;
1855}
1856
1857/* This gets called after a retransmit timeout, and the initially
1858 * retransmitted data is acknowledged. It tries to continue
1859 * resending the rest of the retransmit queue, until either
1860 * we've sent it all or the congestion window limit is reached.
1861 * If doing SACK, the first ACK which comes back for a timeout
1862 * based retransmit packet might feed us FACK information again.
1863 * If so, we use it to avoid unnecessarily retransmissions.
1864 */
1865void tcp_xmit_retransmit_queue(struct sock *sk)
1866{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001867 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 struct tcp_sock *tp = tcp_sk(sk);
1869 struct sk_buff *skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001870 int packet_cnt;
1871
1872 if (tp->retransmit_skb_hint) {
1873 skb = tp->retransmit_skb_hint;
1874 packet_cnt = tp->retransmit_cnt_hint;
1875 }else{
David S. Millerfe067e82007-03-07 12:12:44 -08001876 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001877 packet_cnt = 0;
1878 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879
1880 /* First pass: retransmit lost packets. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001881 if (tp->lost_out) {
David S. Millerfe067e82007-03-07 12:12:44 -08001882 tcp_for_write_queue_from(skb, sk) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1884
David S. Millerfe067e82007-03-07 12:12:44 -08001885 if (skb == tcp_send_head(sk))
1886 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001887 /* we could do better than to assign each time */
1888 tp->retransmit_skb_hint = skb;
1889 tp->retransmit_cnt_hint = packet_cnt;
1890
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 /* Assume this retransmit will generate
1892 * only one packet for congestion window
1893 * calculation purposes. This works because
1894 * tcp_retransmit_skb() will chop up the
1895 * packet to be MSS sized and all the
1896 * packet counting works out.
1897 */
1898 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1899 return;
1900
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001901 if (sacked & TCPCB_LOST) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001903 if (tcp_retransmit_skb(sk, skb)) {
1904 tp->retransmit_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001906 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001907 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1909 else
1910 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1911
David S. Millerfe067e82007-03-07 12:12:44 -08001912 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001913 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001914 inet_csk(sk)->icsk_rto,
1915 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 }
1917
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001918 packet_cnt += tcp_skb_pcount(skb);
1919 if (packet_cnt >= tp->lost_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 break;
1921 }
1922 }
1923 }
1924
1925 /* OK, demanded retransmission is finished. */
1926
1927 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001928 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 return;
1930
1931 /* No forward retransmissions in Reno are possible. */
1932 if (!tp->rx_opt.sack_ok)
1933 return;
1934
1935 /* Yeah, we have to make difficult choice between forward transmission
1936 * and retransmission... Both ways have their merits...
1937 *
1938 * For now we do not retransmit anything, while we have some new
1939 * segments to send.
1940 */
1941
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001942 if (tcp_may_send_now(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943 return;
1944
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001945 if (tp->forward_skb_hint) {
1946 skb = tp->forward_skb_hint;
1947 packet_cnt = tp->forward_cnt_hint;
1948 } else{
David S. Millerfe067e82007-03-07 12:12:44 -08001949 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001950 packet_cnt = 0;
1951 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952
David S. Millerfe067e82007-03-07 12:12:44 -08001953 tcp_for_write_queue_from(skb, sk) {
1954 if (skb == tcp_send_head(sk))
1955 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001956 tp->forward_cnt_hint = packet_cnt;
1957 tp->forward_skb_hint = skb;
1958
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 /* Similar to the retransmit loop above we
1960 * can pretend that the retransmitted SKB
1961 * we send out here will be composed of one
1962 * real MSS sized packet because tcp_retransmit_skb()
1963 * will fragment it if necessary.
1964 */
1965 if (++packet_cnt > tp->fackets_out)
1966 break;
1967
1968 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1969 break;
1970
1971 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1972 continue;
1973
1974 /* Ok, retransmit it. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001975 if (tcp_retransmit_skb(sk, skb)) {
1976 tp->forward_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001978 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979
David S. Millerfe067e82007-03-07 12:12:44 -08001980 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001981 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1982 inet_csk(sk)->icsk_rto,
1983 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984
1985 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1986 }
1987}
1988
1989
1990/* Send a fin. The caller locks the socket for us. This cannot be
1991 * allowed to fail queueing a FIN frame under any circumstances.
1992 */
1993void tcp_send_fin(struct sock *sk)
1994{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001995 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001996 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001998
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 /* Optimization, tack on the FIN if we have a queue of
2000 * unsent frames. But be careful about outgoing SACKS
2001 * and IP options.
2002 */
2003 mss_now = tcp_current_mss(sk, 1);
2004
David S. Millerfe067e82007-03-07 12:12:44 -08002005 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2007 TCP_SKB_CB(skb)->end_seq++;
2008 tp->write_seq++;
2009 } else {
2010 /* Socket is locked, keep trying until memory is available. */
2011 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07002012 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 if (skb)
2014 break;
2015 yield();
2016 }
2017
2018 /* Reserve space for headers and prepare control bits. */
2019 skb_reserve(skb, MAX_TCP_HEADER);
2020 skb->csum = 0;
2021 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2022 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002023 skb_shinfo(skb)->gso_segs = 1;
2024 skb_shinfo(skb)->gso_size = 0;
2025 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026
2027 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2028 TCP_SKB_CB(skb)->seq = tp->write_seq;
2029 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2030 tcp_queue_skb(sk, skb);
2031 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002032 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033}
2034
2035/* We get here when a process closes a file descriptor (either due to
2036 * an explicit close() or as a byproduct of exit()'ing) and there
2037 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002038 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 */
Al Virodd0fc662005-10-07 07:46:04 +01002040void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 struct sk_buff *skb;
2043
2044 /* NOTE: No TCP options attached and we never retransmit this. */
2045 skb = alloc_skb(MAX_TCP_HEADER, priority);
2046 if (!skb) {
2047 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2048 return;
2049 }
2050
2051 /* Reserve space for headers and prepare control bits. */
2052 skb_reserve(skb, MAX_TCP_HEADER);
2053 skb->csum = 0;
2054 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2055 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002056 skb_shinfo(skb)->gso_segs = 1;
2057 skb_shinfo(skb)->gso_size = 0;
2058 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059
2060 /* Send it off. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002061 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2063 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002064 if (tcp_transmit_skb(sk, skb, 0, priority))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2066}
2067
2068/* WARNING: This routine must only be called when we have already sent
2069 * a SYN packet that crossed the incoming SYN that caused this routine
2070 * to get called. If this assumption fails then the initial rcv_wnd
2071 * and rcv_wscale values will not be correct.
2072 */
2073int tcp_send_synack(struct sock *sk)
2074{
2075 struct sk_buff* skb;
2076
David S. Millerfe067e82007-03-07 12:12:44 -08002077 skb = tcp_write_queue_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
2079 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2080 return -EFAULT;
2081 }
2082 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
2083 if (skb_cloned(skb)) {
2084 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2085 if (nskb == NULL)
2086 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002087 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002089 __tcp_add_write_queue_head(sk, nskb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 sk_stream_free_skb(sk, skb);
2091 sk_charge_skb(sk, nskb);
2092 skb = nskb;
2093 }
2094
2095 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2096 TCP_ECN_send_synack(tcp_sk(sk), skb);
2097 }
2098 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002099 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100}
2101
2102/*
2103 * Prepare a SYN-ACK.
2104 */
2105struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
Arnaldo Carvalho de Melo60236fd2005-06-18 22:47:21 -07002106 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002108 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 struct tcp_sock *tp = tcp_sk(sk);
2110 struct tcphdr *th;
2111 int tcp_header_size;
2112 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002113#ifdef CONFIG_TCP_MD5SIG
2114 struct tcp_md5sig_key *md5;
2115 __u8 *md5_hash_location;
2116#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
2118 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2119 if (skb == NULL)
2120 return NULL;
2121
2122 /* Reserve space for headers. */
2123 skb_reserve(skb, MAX_TCP_HEADER);
2124
2125 skb->dst = dst_clone(dst);
2126
2127 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002128 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2129 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002131 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002132
2133#ifdef CONFIG_TCP_MD5SIG
2134 /* Are we doing MD5 on this segment? If so - make room for it */
2135 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2136 if (md5)
2137 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2138#endif
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002139 skb_push(skb, tcp_header_size);
2140 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002142 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 memset(th, 0, sizeof(struct tcphdr));
2144 th->syn = 1;
2145 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 TCP_ECN_make_synack(req, th);
2147 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002148 th->dest = ireq->rmt_port;
2149 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2151 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002152 skb_shinfo(skb)->gso_segs = 1;
2153 skb_shinfo(skb)->gso_size = 0;
2154 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002156 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002158 __u8 rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 /* Set this up on the first call only */
2160 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2161 /* tcp_full_space because it is guaranteed to be the first packet */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002162 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002163 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 &req->rcv_wnd,
2165 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002166 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 &rcv_wscale);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002168 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169 }
2170
2171 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002172 th->window = htons(min(req->rcv_wnd, 65535U));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173
2174 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Al Virodf7a3b02006-09-27 18:38:52 -07002175 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002176 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 TCP_SKB_CB(skb)->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002178 req->ts_recent,
2179 (
2180#ifdef CONFIG_TCP_MD5SIG
2181 md5 ? &md5_hash_location :
2182#endif
2183 NULL)
2184 );
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185
2186 skb->csum = 0;
2187 th->doff = (tcp_header_size >> 2);
2188 TCP_INC_STATS(TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002189
2190#ifdef CONFIG_TCP_MD5SIG
2191 /* Okay, we have all we need - do the md5 hash if needed */
2192 if (md5) {
2193 tp->af_specific->calc_md5_hash(md5_hash_location,
2194 md5,
2195 NULL, dst, req,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002196 tcp_hdr(skb), sk->sk_protocol,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002197 skb->len);
2198 }
2199#endif
2200
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 return skb;
2202}
2203
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002204/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 * Do all connect socket setups that can be done AF independent.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002206 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002207static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208{
2209 struct dst_entry *dst = __sk_dst_get(sk);
2210 struct tcp_sock *tp = tcp_sk(sk);
2211 __u8 rcv_wscale;
2212
2213 /* We'll fix this up when we get a response from the other end.
2214 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2215 */
2216 tp->tcp_header_len = sizeof(struct tcphdr) +
2217 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2218
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002219#ifdef CONFIG_TCP_MD5SIG
2220 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2221 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2222#endif
2223
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 /* If user gave his TCP_MAXSEG, record it to clamp */
2225 if (tp->rx_opt.user_mss)
2226 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2227 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002228 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 tcp_sync_mss(sk, dst_mtu(dst));
2230
2231 if (!tp->window_clamp)
2232 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2233 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2234 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235
2236 tcp_select_initial_window(tcp_full_space(sk),
2237 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2238 &tp->rcv_wnd,
2239 &tp->window_clamp,
2240 sysctl_tcp_window_scaling,
2241 &rcv_wscale);
2242
2243 tp->rx_opt.rcv_wscale = rcv_wscale;
2244 tp->rcv_ssthresh = tp->rcv_wnd;
2245
2246 sk->sk_err = 0;
2247 sock_reset_flag(sk, SOCK_DONE);
2248 tp->snd_wnd = 0;
2249 tcp_init_wl(tp, tp->write_seq, 0);
2250 tp->snd_una = tp->write_seq;
2251 tp->snd_sml = tp->write_seq;
2252 tp->rcv_nxt = 0;
2253 tp->rcv_wup = 0;
2254 tp->copied_seq = 0;
2255
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002256 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2257 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 tcp_clear_retrans(tp);
2259}
2260
2261/*
2262 * Build a SYN and send it off.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002263 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264int tcp_connect(struct sock *sk)
2265{
2266 struct tcp_sock *tp = tcp_sk(sk);
2267 struct sk_buff *buff;
2268
2269 tcp_connect_init(sk);
2270
David S. Millerd179cd12005-08-17 14:57:30 -07002271 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 if (unlikely(buff == NULL))
2273 return -ENOBUFS;
2274
2275 /* Reserve space for headers. */
2276 skb_reserve(buff, MAX_TCP_HEADER);
2277
2278 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002279 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002281 skb_shinfo(buff)->gso_segs = 1;
2282 skb_shinfo(buff)->gso_size = 0;
2283 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 buff->csum = 0;
Wei Yongjunbd37a082006-08-07 21:04:15 -07002285 tp->snd_nxt = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2287 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288
2289 /* Send it off. */
2290 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2291 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2292 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002293 __tcp_add_write_queue_tail(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 sk_charge_skb(sk, buff);
2295 tp->packets_out += tcp_skb_pcount(buff);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002296 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002297
2298 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2299 * in order to make this packet get counted in tcpOutSegs.
2300 */
2301 tp->snd_nxt = tp->write_seq;
2302 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2304
2305 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002306 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2307 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 return 0;
2309}
2310
2311/* Send out a delayed ack, the caller does the policy checking
2312 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2313 * for details.
2314 */
2315void tcp_send_delayed_ack(struct sock *sk)
2316{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002317 struct inet_connection_sock *icsk = inet_csk(sk);
2318 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 unsigned long timeout;
2320
2321 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002322 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 int max_ato = HZ/2;
2324
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002325 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 max_ato = TCP_DELACK_MAX;
2327
2328 /* Slow path, intersegment interval is "high". */
2329
2330 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002331 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 * directly.
2333 */
2334 if (tp->srtt) {
2335 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2336
2337 if (rtt < max_ato)
2338 max_ato = rtt;
2339 }
2340
2341 ato = min(ato, max_ato);
2342 }
2343
2344 /* Stay within the limit we were given */
2345 timeout = jiffies + ato;
2346
2347 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002348 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 /* If delack timer was blocked or is about to expire,
2350 * send ACK now.
2351 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002352 if (icsk->icsk_ack.blocked ||
2353 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 tcp_send_ack(sk);
2355 return;
2356 }
2357
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002358 if (!time_before(timeout, icsk->icsk_ack.timeout))
2359 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002361 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2362 icsk->icsk_ack.timeout = timeout;
2363 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364}
2365
2366/* This routine sends an ack and also updates the window. */
2367void tcp_send_ack(struct sock *sk)
2368{
2369 /* If we have been reset, we may not send again. */
2370 if (sk->sk_state != TCP_CLOSE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 struct sk_buff *buff;
2372
2373 /* We are not putting this on the write queue, so
2374 * tcp_transmit_skb() will set the ownership to this
2375 * sock.
2376 */
2377 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2378 if (buff == NULL) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002379 inet_csk_schedule_ack(sk);
2380 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002381 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2382 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 return;
2384 }
2385
2386 /* Reserve space for headers and prepare control bits. */
2387 skb_reserve(buff, MAX_TCP_HEADER);
2388 buff->csum = 0;
2389 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2390 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002391 skb_shinfo(buff)->gso_segs = 1;
2392 skb_shinfo(buff)->gso_size = 0;
2393 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394
2395 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002396 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 TCP_SKB_CB(buff)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002398 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 }
2400}
2401
2402/* This routine sends a packet with an out of date sequence
2403 * number. It assumes the other end will try to ack it.
2404 *
2405 * Question: what should we make while urgent mode?
2406 * 4.4BSD forces sending single byte of data. We cannot send
2407 * out of window data, because we have SND.NXT==SND.MAX...
2408 *
2409 * Current solution: to send TWO zero-length segments in urgent mode:
2410 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2411 * out-of-date with SND.UNA-1 to probe window.
2412 */
2413static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2414{
2415 struct tcp_sock *tp = tcp_sk(sk);
2416 struct sk_buff *skb;
2417
2418 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2419 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002420 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 return -1;
2422
2423 /* Reserve space for headers and set control bits. */
2424 skb_reserve(skb, MAX_TCP_HEADER);
2425 skb->csum = 0;
2426 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2427 TCP_SKB_CB(skb)->sacked = urgent;
Herbert Xu79671682006-06-22 02:40:14 -07002428 skb_shinfo(skb)->gso_segs = 1;
2429 skb_shinfo(skb)->gso_size = 0;
2430 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431
2432 /* Use a previous sequence. This should cause the other
2433 * end to send an ack. Don't queue or clone SKB, just
2434 * send it.
2435 */
2436 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2437 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2438 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002439 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440}
2441
2442int tcp_write_wakeup(struct sock *sk)
2443{
2444 if (sk->sk_state != TCP_CLOSE) {
2445 struct tcp_sock *tp = tcp_sk(sk);
2446 struct sk_buff *skb;
2447
David S. Millerfe067e82007-03-07 12:12:44 -08002448 if ((skb = tcp_send_head(sk)) != NULL &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2450 int err;
2451 unsigned int mss = tcp_current_mss(sk, 0);
2452 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2453
2454 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2455 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2456
2457 /* We are probing the opening of a window
2458 * but the window size is != 0
2459 * must have been a result SWS avoidance ( sender )
2460 */
2461 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2462 skb->len > mss) {
2463 seg_size = min(seg_size, mss);
2464 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
David S. Miller846998a2005-08-04 19:52:01 -07002465 if (tcp_fragment(sk, skb, seg_size, mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 } else if (!tcp_skb_pcount(skb))
David S. Miller846998a2005-08-04 19:52:01 -07002468 tcp_set_skb_tso_segs(sk, skb, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469
2470 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2471 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002472 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 if (!err) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002474 update_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475 }
2476 return err;
2477 } else {
2478 if (tp->urg_mode &&
2479 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2480 tcp_xmit_probe_skb(sk, TCPCB_URG);
2481 return tcp_xmit_probe_skb(sk, 0);
2482 }
2483 }
2484 return -1;
2485}
2486
2487/* A window probe timeout has occurred. If window is not closed send
2488 * a partial packet else a zero probe.
2489 */
2490void tcp_send_probe0(struct sock *sk)
2491{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002492 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 struct tcp_sock *tp = tcp_sk(sk);
2494 int err;
2495
2496 err = tcp_write_wakeup(sk);
2497
David S. Millerfe067e82007-03-07 12:12:44 -08002498 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002500 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002501 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 return;
2503 }
2504
2505 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002506 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2507 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002508 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002509 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002510 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2511 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 } else {
2513 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002514 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 * Let local senders to fight for local resources.
2516 *
2517 * Use accumulated backoff yet.
2518 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002519 if (!icsk->icsk_probes_out)
2520 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002521 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002522 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002523 TCP_RESOURCE_PROBE_INTERVAL),
2524 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 }
2526}
2527
2528EXPORT_SYMBOL(tcp_connect);
2529EXPORT_SYMBOL(tcp_make_synack);
2530EXPORT_SYMBOL(tcp_simple_retransmit);
2531EXPORT_SYMBOL(tcp_sync_mss);
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -08002532EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
John Heffner5d424d52006-03-20 17:53:41 -08002533EXPORT_SYMBOL(tcp_mtup_init);