blob: f4c1eef89af0b3ac1e63ed91ecf3725a761d6949 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
9 *
Jesper Juhl02c30a82005-05-05 16:16:16 -070010 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
21 */
22
23/*
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
27 * : AF independence
28 *
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
36 *
37 */
38
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
42#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070045int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Rick Jones15d99e02006-03-20 22:40:29 -080047/* People can turn this on to work with those rare, broken TCPs that
48 * interpret the window field as a signed quantity.
49 */
Brian Haleyab32ea52006-09-22 14:15:41 -070050int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080051
Linus Torvalds1da177e2005-04-16 15:20:36 -070052/* This limits the percentage of the congestion window which we
53 * will allow a single TSO frame to consume. Building TSO frames
54 * which are too large can cause TCP streams to be bursty.
55 */
Brian Haleyab32ea52006-09-22 14:15:41 -070056int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070057
Brian Haleyab32ea52006-09-22 14:15:41 -070058int sysctl_tcp_mtu_probing __read_mostly = 0;
59int sysctl_tcp_base_mss __read_mostly = 512;
John Heffner5d424d52006-03-20 17:53:41 -080060
David S. Miller35089bb2006-06-13 22:33:04 -070061/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070062int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070063
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070064static inline void tcp_packets_out_inc(struct sock *sk,
65 const struct sk_buff *skb)
66{
67 struct tcp_sock *tp = tcp_sk(sk);
68 int orig = tp->packets_out;
69
70 tp->packets_out += tcp_skb_pcount(skb);
71 if (!orig)
72 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
73 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
74}
75
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070076static void update_send_head(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -070077{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070078 struct tcp_sock *tp = tcp_sk(sk);
79
David S. Millerfe067e82007-03-07 12:12:44 -080080 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070081 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070082 tcp_packets_out_inc(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070083}
84
85/* SND.NXT, if window was not shrunk.
86 * If window has been shrunk, what should we make? It is not clear at all.
87 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
88 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
89 * invalid. OK, let's make this for now:
90 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070091static inline __u32 tcp_acceptable_seq(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070092{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070093 struct tcp_sock *tp = tcp_sk(sk);
94
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
96 return tp->snd_nxt;
97 else
98 return tp->snd_una+tp->snd_wnd;
99}
100
101/* Calculate mss to advertise in SYN segment.
102 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
103 *
104 * 1. It is independent of path mtu.
105 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
106 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
107 * attached devices, because some buggy hosts are confused by
108 * large MSS.
109 * 4. We do not make 3, we advertise MSS, calculated from first
110 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
111 * This may be overridden via information stored in routing table.
112 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
113 * probably even Jumbo".
114 */
115static __u16 tcp_advertise_mss(struct sock *sk)
116{
117 struct tcp_sock *tp = tcp_sk(sk);
118 struct dst_entry *dst = __sk_dst_get(sk);
119 int mss = tp->advmss;
120
121 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
122 mss = dst_metric(dst, RTAX_ADVMSS);
123 tp->advmss = mss;
124 }
125
126 return (__u16)mss;
127}
128
129/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
130 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700131static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700133 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 s32 delta = tcp_time_stamp - tp->lsndtime;
135 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
136 u32 cwnd = tp->snd_cwnd;
137
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300138 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300140 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 restart_cwnd = min(restart_cwnd, cwnd);
142
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700143 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 cwnd >>= 1;
145 tp->snd_cwnd = max(cwnd, restart_cwnd);
146 tp->snd_cwnd_stamp = tcp_time_stamp;
147 tp->snd_cwnd_used = 0;
148}
149
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800150static void tcp_event_data_sent(struct tcp_sock *tp,
151 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700153 struct inet_connection_sock *icsk = inet_csk(sk);
154 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
David S. Miller35089bb2006-06-13 22:33:04 -0700156 if (sysctl_tcp_slow_start_after_idle &&
157 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700158 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
160 tp->lsndtime = now;
161
162 /* If it is a reply for ato after last received
163 * packet, enter pingpong mode.
164 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700165 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
166 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167}
168
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800169static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700171 tcp_dec_quickack_mode(sk, pkts);
172 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173}
174
175/* Determine a window scaling and initial window to offer.
176 * Based on the assumption that the given amount of space
177 * will be offered. Store the results in the tp structure.
178 * NOTE: for smooth operation initial space offering should
179 * be a multiple of mss if possible. We assume here that mss >= 1.
180 * This MUST be enforced by all callers.
181 */
182void tcp_select_initial_window(int __space, __u32 mss,
183 __u32 *rcv_wnd, __u32 *window_clamp,
184 int wscale_ok, __u8 *rcv_wscale)
185{
186 unsigned int space = (__space < 0 ? 0 : __space);
187
188 /* If no clamp set the clamp to the max possible scaled window */
189 if (*window_clamp == 0)
190 (*window_clamp) = (65535 << 14);
191 space = min(*window_clamp, space);
192
193 /* Quantize space offering to a multiple of mss if possible. */
194 if (space > mss)
195 space = (space / mss) * mss;
196
197 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800198 * will break some buggy TCP stacks. If the admin tells us
199 * it is likely we could be speaking with such a buggy stack
200 * we will truncate our initial window offering to 32K-1
201 * unless the remote has sent us a window scaling option,
202 * which we interpret as a sign the remote TCP is not
203 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 */
Rick Jones15d99e02006-03-20 22:40:29 -0800205 if (sysctl_tcp_workaround_signed_windows)
206 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
207 else
208 (*rcv_wnd) = space;
209
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210 (*rcv_wscale) = 0;
211 if (wscale_ok) {
212 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900213 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 */
215 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700216 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 while (space > 65535 && (*rcv_wscale) < 14) {
218 space >>= 1;
219 (*rcv_wscale)++;
220 }
221 }
222
223 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700224 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 * will be satisfied with 2.
226 */
227 if (mss > (1<<*rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700228 int init_cwnd = 4;
229 if (mss > 1460*3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700231 else if (mss > 1460)
232 init_cwnd = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 if (*rcv_wnd > init_cwnd*mss)
234 *rcv_wnd = init_cwnd*mss;
235 }
236
237 /* Set the clamp no higher than max representable value */
238 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
239}
240
241/* Chose a new window to advertise, update state in tcp_sock for the
242 * socket, and return result with RFC1323 scaling applied. The return
243 * value can be stuffed directly into th->window for an outgoing
244 * frame.
245 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800246static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247{
248 struct tcp_sock *tp = tcp_sk(sk);
249 u32 cur_win = tcp_receive_window(tp);
250 u32 new_win = __tcp_select_window(sk);
251
252 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800253 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 /* Danger Will Robinson!
255 * Don't update rcv_wup/rcv_wnd here or else
256 * we will not be able to advertise a zero
257 * window in time. --DaveM
258 *
259 * Relax Will Robinson.
260 */
261 new_win = cur_win;
262 }
263 tp->rcv_wnd = new_win;
264 tp->rcv_wup = tp->rcv_nxt;
265
266 /* Make sure we do not exceed the maximum possible
267 * scaled window.
268 */
Rick Jones15d99e02006-03-20 22:40:29 -0800269 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 new_win = min(new_win, MAX_TCP_WINDOW);
271 else
272 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
273
274 /* RFC1323 scaling applied */
275 new_win >>= tp->rx_opt.rcv_wscale;
276
277 /* If we advertise zero window, disable fast path. */
278 if (new_win == 0)
279 tp->pred_flags = 0;
280
281 return new_win;
282}
283
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700284static inline void TCP_ECN_send_synack(struct tcp_sock *tp,
285 struct sk_buff *skb)
286{
287 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
288 if (!(tp->ecn_flags&TCP_ECN_OK))
289 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
290}
291
292static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
293{
294 struct tcp_sock *tp = tcp_sk(sk);
295
296 tp->ecn_flags = 0;
297 if (sysctl_tcp_ecn) {
298 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR;
299 tp->ecn_flags = TCP_ECN_OK;
300 }
301}
302
303static __inline__ void
304TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
305{
306 if (inet_rsk(req)->ecn_ok)
307 th->ece = 1;
308}
309
310static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
311 int tcp_header_len)
312{
313 struct tcp_sock *tp = tcp_sk(sk);
314
315 if (tp->ecn_flags & TCP_ECN_OK) {
316 /* Not-retransmitted data segment: set ECT and inject CWR. */
317 if (skb->len != tcp_header_len &&
318 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
319 INET_ECN_xmit(sk);
320 if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) {
321 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
322 tcp_hdr(skb)->cwr = 1;
323 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
324 }
325 } else {
326 /* ACK or retransmitted segment: clear ECT|CE */
327 INET_ECN_dontxmit(sk);
328 }
329 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
330 tcp_hdr(skb)->ece = 1;
331 }
332}
333
Al Virodf7a3b02006-09-27 18:38:52 -0700334static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800335 __u32 tstamp, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800336{
337 if (tp->rx_opt.tstamp_ok) {
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700338 *ptr++ = htonl((TCPOPT_NOP << 24) |
339 (TCPOPT_NOP << 16) |
340 (TCPOPT_TIMESTAMP << 8) |
341 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800342 *ptr++ = htonl(tstamp);
343 *ptr++ = htonl(tp->rx_opt.ts_recent);
344 }
345 if (tp->rx_opt.eff_sacks) {
346 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
347 int this_sack;
348
349 *ptr++ = htonl((TCPOPT_NOP << 24) |
350 (TCPOPT_NOP << 16) |
351 (TCPOPT_SACK << 8) |
352 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
353 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800354
355 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800356 *ptr++ = htonl(sp[this_sack].start_seq);
357 *ptr++ = htonl(sp[this_sack].end_seq);
358 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800359
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800360 if (tp->rx_opt.dsack) {
361 tp->rx_opt.dsack = 0;
362 tp->rx_opt.eff_sacks--;
363 }
364 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800365#ifdef CONFIG_TCP_MD5SIG
366 if (md5_hash) {
367 *ptr++ = htonl((TCPOPT_NOP << 24) |
368 (TCPOPT_NOP << 16) |
369 (TCPOPT_MD5SIG << 8) |
370 TCPOLEN_MD5SIG);
371 *md5_hash = (__u8 *)ptr;
372 }
373#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800374}
375
376/* Construct a tcp options header for a SYN or SYN_ACK packet.
377 * If this is every changed make sure to change the definition of
378 * MAX_SYN_SIZE to match the new maximum number of options that you
379 * can generate.
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800380 *
381 * Note - that with the RFC2385 TCP option, we make room for the
382 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
383 * location to be filled is passed back up.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800384 */
Al Virodf7a3b02006-09-27 18:38:52 -0700385static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800386 int offer_wscale, int wscale, __u32 tstamp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800387 __u32 ts_recent, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800388{
389 /* We always get an MSS option.
390 * The option bytes which will be seen in normal data
391 * packets should timestamps be used, must be in the MSS
392 * advertised. But we subtract them from tp->mss_cache so
393 * that calculations in tcp_sendmsg are simpler etc.
394 * So account for this fact here if necessary. If we
395 * don't do this correctly, as a receiver we won't
396 * recognize data packets as being full sized when we
397 * should, and thus we won't abide by the delayed ACK
398 * rules correctly.
399 * SACKs don't matter, we never delay an ACK when we
400 * have any of those going out.
401 */
402 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
403 if (ts) {
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800404 if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700405 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
406 (TCPOLEN_SACK_PERM << 16) |
407 (TCPOPT_TIMESTAMP << 8) |
408 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800409 else
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700410 *ptr++ = htonl((TCPOPT_NOP << 24) |
411 (TCPOPT_NOP << 16) |
412 (TCPOPT_TIMESTAMP << 8) |
413 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800414 *ptr++ = htonl(tstamp); /* TSVAL */
415 *ptr++ = htonl(ts_recent); /* TSECR */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800416 } else if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700417 *ptr++ = htonl((TCPOPT_NOP << 24) |
418 (TCPOPT_NOP << 16) |
419 (TCPOPT_SACK_PERM << 8) |
420 TCPOLEN_SACK_PERM);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800421 if (offer_wscale)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700422 *ptr++ = htonl((TCPOPT_NOP << 24) |
423 (TCPOPT_WINDOW << 16) |
424 (TCPOLEN_WINDOW << 8) |
425 (wscale));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800426#ifdef CONFIG_TCP_MD5SIG
427 /*
428 * If MD5 is enabled, then we set the option, and include the size
429 * (always 18). The actual MD5 hash is added just before the
430 * packet is sent.
431 */
432 if (md5_hash) {
433 *ptr++ = htonl((TCPOPT_NOP << 24) |
434 (TCPOPT_NOP << 16) |
435 (TCPOPT_MD5SIG << 8) |
436 TCPOLEN_MD5SIG);
437 *md5_hash = (__u8 *) ptr;
438 }
439#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800440}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
442/* This routine actually transmits TCP packets queued in by
443 * tcp_do_sendmsg(). This is used by both the initial
444 * transmission and possible later retransmissions.
445 * All SKB's seen here are completely headerless. It is our
446 * job to build the TCP header, and pass the packet down to
447 * IP so it can do the same plus pass the packet off to the
448 * device.
449 *
450 * We are working here with either a clone of the original
451 * SKB, or a fresh unique copy made by the retransmit engine.
452 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800453static 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 -0700454{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800455 const struct inet_connection_sock *icsk = inet_csk(sk);
456 struct inet_sock *inet;
457 struct tcp_sock *tp;
458 struct tcp_skb_cb *tcb;
459 int tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800460#ifdef CONFIG_TCP_MD5SIG
461 struct tcp_md5sig_key *md5;
462 __u8 *md5_hash_location;
463#endif
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800464 struct tcphdr *th;
465 int sysctl_flags;
466 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800468 BUG_ON(!skb || !tcp_skb_pcount(skb));
469
470 /* If congestion control is doing timestamping, we must
471 * take such a timestamp before we potentially clone/copy.
472 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700473 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800474 __net_timestamp(skb);
475
476 if (likely(clone_it)) {
477 if (unlikely(skb_cloned(skb)))
478 skb = pskb_copy(skb, gfp_mask);
479 else
480 skb = skb_clone(skb, gfp_mask);
481 if (unlikely(!skb))
482 return -ENOBUFS;
483 }
484
485 inet = inet_sk(sk);
486 tp = tcp_sk(sk);
487 tcb = TCP_SKB_CB(skb);
488 tcp_header_size = tp->tcp_header_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
490#define SYSCTL_FLAG_TSTAMPS 0x1
491#define SYSCTL_FLAG_WSCALE 0x2
492#define SYSCTL_FLAG_SACK 0x4
493
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800494 sysctl_flags = 0;
495 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
496 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800497 if (sysctl_tcp_timestamps) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800498 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
499 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800501 if (sysctl_tcp_window_scaling) {
502 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
503 sysctl_flags |= SYSCTL_FLAG_WSCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800505 if (sysctl_tcp_sack) {
506 sysctl_flags |= SYSCTL_FLAG_SACK;
507 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
508 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800510 } else if (unlikely(tp->rx_opt.eff_sacks)) {
511 /* A SACK is 2 pad bytes, a 2 byte header, plus
512 * 2 32-bit sequence numbers for each SACK block.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800514 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
515 (tp->rx_opt.eff_sacks *
516 TCPOLEN_SACK_PERBLOCK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900518
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800519 if (tcp_packets_in_flight(tp) == 0)
520 tcp_ca_event(sk, CA_EVENT_TX_START);
521
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800522#ifdef CONFIG_TCP_MD5SIG
523 /*
524 * Are we doing MD5 on this segment? If so - make
525 * room for it.
526 */
527 md5 = tp->af_specific->md5_lookup(sk, sk);
528 if (md5)
529 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
530#endif
531
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700532 skb_push(skb, tcp_header_size);
533 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800534 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800535
536 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700537 th = tcp_hdr(skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800538 th->source = inet->sport;
539 th->dest = inet->dport;
540 th->seq = htonl(tcb->seq);
541 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700542 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800543 tcb->flags);
544
545 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
546 /* RFC1323: The window in SYN & SYN/ACK segments
547 * is never scaled.
548 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800549 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800550 } else {
551 th->window = htons(tcp_select_window(sk));
552 }
553 th->check = 0;
554 th->urg_ptr = 0;
555
556 if (unlikely(tp->urg_mode &&
557 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
558 th->urg_ptr = htons(tp->snd_up-tcb->seq);
559 th->urg = 1;
560 }
561
562 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
Al Virodf7a3b02006-09-27 18:38:52 -0700563 tcp_syn_build_options((__be32 *)(th + 1),
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800564 tcp_advertise_mss(sk),
565 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
566 (sysctl_flags & SYSCTL_FLAG_SACK),
567 (sysctl_flags & SYSCTL_FLAG_WSCALE),
568 tp->rx_opt.rcv_wscale,
569 tcb->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800570 tp->rx_opt.ts_recent,
571
572#ifdef CONFIG_TCP_MD5SIG
573 md5 ? &md5_hash_location :
574#endif
575 NULL);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800576 } else {
Al Virodf7a3b02006-09-27 18:38:52 -0700577 tcp_build_and_update_options((__be32 *)(th + 1),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800578 tp, tcb->when,
579#ifdef CONFIG_TCP_MD5SIG
580 md5 ? &md5_hash_location :
581#endif
582 NULL);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700583 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800584 }
585
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800586#ifdef CONFIG_TCP_MD5SIG
587 /* Calculate the MD5 hash, as we have all we need now */
588 if (md5) {
589 tp->af_specific->calc_md5_hash(md5_hash_location,
590 md5,
591 sk, NULL, NULL,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700592 tcp_hdr(skb),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800593 sk->sk_protocol,
594 skb->len);
595 }
596#endif
597
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800598 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800599
600 if (likely(tcb->flags & TCPCB_FLAG_ACK))
601 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
602
603 if (skb->len != tcp_header_size)
604 tcp_event_data_sent(tp, skb, sk);
605
Wei Yongjunbd37a082006-08-07 21:04:15 -0700606 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
607 TCP_INC_STATS(TCP_MIB_OUTSEGS);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800608
David S. Millere89862f2007-01-26 01:04:55 -0800609 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700610 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800611 return err;
612
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800613 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800614
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200615 return net_xmit_eval(err);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800616
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617#undef SYSCTL_FLAG_TSTAMPS
618#undef SYSCTL_FLAG_WSCALE
619#undef SYSCTL_FLAG_SACK
620}
621
622
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900623/* This routine just queue's the buffer
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 *
625 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
626 * otherwise socket can stall.
627 */
628static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
629{
630 struct tcp_sock *tp = tcp_sk(sk);
631
632 /* Advance write_seq and place onto the write_queue. */
633 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
634 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800635 tcp_add_write_queue_tail(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 sk_charge_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637}
638
David S. Miller846998a2005-08-04 19:52:01 -0700639static 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 -0700640{
Herbert Xubcd76112006-06-30 13:36:35 -0700641 if (skb->len <= mss_now || !sk_can_gso(sk)) {
David S. Millerf6302d12005-07-05 15:18:03 -0700642 /* Avoid the costly divide in the normal
643 * non-TSO case.
644 */
Herbert Xu79671682006-06-22 02:40:14 -0700645 skb_shinfo(skb)->gso_segs = 1;
646 skb_shinfo(skb)->gso_size = 0;
647 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700648 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700649 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700650 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700651 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 }
653}
654
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700655/* When a modification to fackets out becomes necessary, we need to check
656 * skb is counted to fackets_out or not. Another important thing is to
657 * tweak SACK fastpath hint too as it would overwrite all changes unless
658 * hint is also changed.
659 */
660static void tcp_adjust_fackets_out(struct tcp_sock *tp, struct sk_buff *skb,
661 int decr)
662{
Ilpo Järvinendc869672007-10-01 15:27:19 -0700663 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700664 return;
665
666 if (!before(tp->highest_sack, TCP_SKB_CB(skb)->seq))
667 tp->fackets_out -= decr;
668
669 /* cnt_hint is "off-by-one" compared with fackets_out (see sacktag) */
670 if (tp->fastpath_skb_hint != NULL &&
671 after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
672 tp->fastpath_cnt_hint -= decr;
673}
674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675/* Function to create two new TCP segments. Shrinks the given segment
676 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900677 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 * Remember, these are still headerless SKBs at this point.
679 */
David S. Miller6475be12005-09-01 22:47:01 -0700680int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681{
682 struct tcp_sock *tp = tcp_sk(sk);
683 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700684 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700685 int nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 u16 flags;
687
Herbert Xub2cc99f02005-10-20 17:13:13 -0200688 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800689
Ilpo Järvinenb7689202007-09-20 11:37:19 -0700690 tcp_clear_retrans_hints_partial(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 nsize = skb_headlen(skb) - len;
692 if (nsize < 0)
693 nsize = 0;
694
695 if (skb_cloned(skb) &&
696 skb_is_nonlinear(skb) &&
697 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
698 return -ENOMEM;
699
700 /* Get a new skb... force flag on. */
701 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
702 if (buff == NULL)
703 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -0700704
Herbert Xub60b49e2006-04-19 21:35:00 -0700705 sk_charge_skb(sk, buff);
706 nlen = skb->len - len - nsize;
707 buff->truesize += nlen;
708 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
710 /* Correct the sequence numbers. */
711 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
712 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
713 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
714
Ilpo Järvinendc869672007-10-01 15:27:19 -0700715 if (tcp_is_sack(tp) && tp->sacked_out &&
716 (TCP_SKB_CB(skb)->seq == tp->highest_sack))
Ilpo Järvinen0dde7b52007-09-20 11:28:05 -0700717 tp->highest_sack = TCP_SKB_CB(buff)->seq;
718
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 /* PSH and FIN should only be set in the second packet. */
720 flags = TCP_SKB_CB(skb)->flags;
721 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
722 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700723 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
725
Patrick McHardy84fa7932006-08-29 16:44:56 -0700726 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 /* Copy and checksum data tail into the new buffer. */
728 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
729 nsize, 0);
730
731 skb_trim(skb, len);
732
733 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
734 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -0700735 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 skb_split(skb, buff, len);
737 }
738
739 buff->ip_summed = skb->ip_summed;
740
741 /* Looks stupid, but our code really uses when of
742 * skbs, which it never sent before. --ANK
743 */
744 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700745 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746
David S. Miller6475be12005-09-01 22:47:01 -0700747 old_factor = tcp_skb_pcount(skb);
748
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700750 tcp_set_skb_tso_segs(sk, skb, mss_now);
751 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752
David S. Miller6475be12005-09-01 22:47:01 -0700753 /* If this packet has been sent out already, we must
754 * adjust the various packet counters.
755 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700756 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700757 int diff = old_factor - tcp_skb_pcount(skb) -
758 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
David S. Miller6475be12005-09-01 22:47:01 -0700760 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700761
762 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
763 tp->sacked_out -= diff;
764 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
765 tp->retrans_out -= diff;
766
Ilpo Järvinenb5860bb2007-08-09 14:33:18 +0300767 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
David S. Miller6475be12005-09-01 22:47:01 -0700768 tp->lost_out -= diff;
Herbert Xu83ca28b2005-09-22 23:32:56 -0700769
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700770 /* Adjust Reno SACK estimate. */
771 if (tcp_is_reno(tp) && diff > 0) {
772 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
773 tcp_verify_left_out(tp);
David S. Miller6475be12005-09-01 22:47:01 -0700774 }
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700775 tcp_adjust_fackets_out(tp, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 }
777
778 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700779 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -0800780 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781
782 return 0;
783}
784
785/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
786 * eventually). The difference is that pulled data not copied, but
787 * immediately discarded.
788 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700789static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790{
791 int i, k, eat;
792
793 eat = len;
794 k = 0;
795 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
796 if (skb_shinfo(skb)->frags[i].size <= eat) {
797 put_page(skb_shinfo(skb)->frags[i].page);
798 eat -= skb_shinfo(skb)->frags[i].size;
799 } else {
800 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
801 if (eat) {
802 skb_shinfo(skb)->frags[k].page_offset += eat;
803 skb_shinfo(skb)->frags[k].size -= eat;
804 eat = 0;
805 }
806 k++;
807 }
808 }
809 skb_shinfo(skb)->nr_frags = k;
810
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -0700811 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 skb->data_len -= len;
813 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814}
815
816int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
817{
818 if (skb_cloned(skb) &&
819 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
820 return -ENOMEM;
821
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700822 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
823 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700825 else
826 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827
828 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -0700829 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
831 skb->truesize -= len;
832 sk->sk_wmem_queued -= len;
833 sk->sk_forward_alloc += len;
834 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
835
836 /* Any change of skb->len requires recalculation of tso
837 * factor and mss.
838 */
839 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700840 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841
842 return 0;
843}
844
John Heffner5d424d52006-03-20 17:53:41 -0800845/* Not accounting for SACKs here. */
846int tcp_mtu_to_mss(struct sock *sk, int pmtu)
847{
848 struct tcp_sock *tp = tcp_sk(sk);
849 struct inet_connection_sock *icsk = inet_csk(sk);
850 int mss_now;
851
852 /* Calculate base mss without TCP options:
853 It is MMS_S - sizeof(tcphdr) of rfc1122
854 */
855 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
856
857 /* Clamp it (mss_clamp does not include tcp options) */
858 if (mss_now > tp->rx_opt.mss_clamp)
859 mss_now = tp->rx_opt.mss_clamp;
860
861 /* Now subtract optional transport overhead */
862 mss_now -= icsk->icsk_ext_hdr_len;
863
864 /* Then reserve room for full set of TCP options and 8 bytes of data */
865 if (mss_now < 48)
866 mss_now = 48;
867
868 /* Now subtract TCP options size, not including SACKs */
869 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
870
871 return mss_now;
872}
873
874/* Inverse of above */
875int tcp_mss_to_mtu(struct sock *sk, int mss)
876{
877 struct tcp_sock *tp = tcp_sk(sk);
878 struct inet_connection_sock *icsk = inet_csk(sk);
879 int mtu;
880
881 mtu = mss +
882 tp->tcp_header_len +
883 icsk->icsk_ext_hdr_len +
884 icsk->icsk_af_ops->net_header_len;
885
886 return mtu;
887}
888
889void tcp_mtup_init(struct sock *sk)
890{
891 struct tcp_sock *tp = tcp_sk(sk);
892 struct inet_connection_sock *icsk = inet_csk(sk);
893
894 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
895 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900896 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -0800897 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
898 icsk->icsk_mtup.probe_size = 0;
899}
900
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901/* This function synchronize snd mss to current pmtu/exthdr set.
902
903 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
904 for TCP options, but includes only bare TCP header.
905
906 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800907 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 It also does not include TCP options.
909
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800910 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911
912 tp->mss_cache is current effective sending mss, including
913 all tcp options except for SACKs. It is evaluated,
914 taking into account current pmtu, but never exceeds
915 tp->rx_opt.mss_clamp.
916
917 NOTE1. rfc1122 clearly states that advertised MSS
918 DOES NOT include either tcp or ip options.
919
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800920 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
921 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 */
923
924unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
925{
926 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800927 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -0800928 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
John Heffner5d424d52006-03-20 17:53:41 -0800930 if (icsk->icsk_mtup.search_high > pmtu)
931 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
John Heffner5d424d52006-03-20 17:53:41 -0800933 mss_now = tcp_mtu_to_mss(sk, pmtu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934
935 /* Bound mss with half of window */
936 if (tp->max_window && mss_now > (tp->max_window>>1))
937 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
938
939 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800940 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -0800941 if (icsk->icsk_mtup.enabled)
942 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700943 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944
945 return mss_now;
946}
947
948/* Compute the current effective MSS, taking SACKs and IP options,
949 * and even PMTU discovery events into account.
950 *
951 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
952 * cannot be large. However, taking into account rare use of URG, this
953 * is not a big flaw.
954 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700955unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956{
957 struct tcp_sock *tp = tcp_sk(sk);
958 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700959 u32 mss_now;
960 u16 xmit_size_goal;
961 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700963 mss_now = tp->mss_cache;
964
Herbert Xubcd76112006-06-30 13:36:35 -0700965 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700966 doing_tso = 1;
967
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 if (dst) {
969 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800970 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 mss_now = tcp_sync_mss(sk, mtu);
972 }
973
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 if (tp->rx_opt.eff_sacks)
975 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
976 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700977
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800978#ifdef CONFIG_TCP_MD5SIG
979 if (tp->af_specific->md5_lookup(sk, sk))
980 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
981#endif
982
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700983 xmit_size_goal = mss_now;
984
985 if (doing_tso) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800986 xmit_size_goal = (65535 -
987 inet_csk(sk)->icsk_af_ops->net_header_len -
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800988 inet_csk(sk)->icsk_ext_hdr_len -
989 tp->tcp_header_len);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700990
991 if (tp->max_window &&
992 (xmit_size_goal > (tp->max_window >> 1)))
993 xmit_size_goal = max((tp->max_window >> 1),
994 68U - tp->tcp_header_len);
995
996 xmit_size_goal -= (xmit_size_goal % mss_now);
997 }
998 tp->xmit_size_goal = xmit_size_goal;
999
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 return mss_now;
1001}
1002
David S. Millera762a982005-07-05 15:18:51 -07001003/* Congestion window validation. (RFC2861) */
1004
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001005static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001006{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001007 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001008 __u32 packets_out = tp->packets_out;
1009
1010 if (packets_out >= tp->snd_cwnd) {
1011 /* Network is feed fully. */
1012 tp->snd_cwnd_used = 0;
1013 tp->snd_cwnd_stamp = tcp_time_stamp;
1014 } else {
1015 /* Network starves. */
1016 if (tp->packets_out > tp->snd_cwnd_used)
1017 tp->snd_cwnd_used = tp->packets_out;
1018
David S. Miller15d33c02007-04-09 13:23:14 -07001019 if (sysctl_tcp_slow_start_after_idle &&
1020 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001021 tcp_cwnd_application_limited(sk);
1022 }
1023}
1024
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001025static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
1026{
1027 u32 window, cwnd_len;
1028
1029 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
1030 cwnd_len = mss_now * cwnd;
1031 return min(window, cwnd_len);
1032}
1033
1034/* Can at least one segment of SKB be sent right now, according to the
1035 * congestion window rules? If so, return how many segments are allowed.
1036 */
1037static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
1038{
1039 u32 in_flight, cwnd;
1040
1041 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -08001042 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1043 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001044 return 1;
1045
1046 in_flight = tcp_packets_in_flight(tp);
1047 cwnd = tp->snd_cwnd;
1048 if (in_flight < cwnd)
1049 return (cwnd - in_flight);
1050
1051 return 0;
1052}
1053
1054/* This must be invoked the first time we consider transmitting
1055 * SKB onto the wire.
1056 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08001057static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001058{
1059 int tso_segs = tcp_skb_pcount(skb);
1060
David S. Miller846998a2005-08-04 19:52:01 -07001061 if (!tso_segs ||
1062 (tso_segs > 1 &&
Herbert Xu79671682006-06-22 02:40:14 -07001063 tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001064 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001065 tso_segs = tcp_skb_pcount(skb);
1066 }
1067 return tso_segs;
1068}
1069
1070static inline int tcp_minshall_check(const struct tcp_sock *tp)
1071{
1072 return after(tp->snd_sml,tp->snd_una) &&
1073 !after(tp->snd_sml, tp->snd_nxt);
1074}
1075
1076/* Return 0, if packet can be sent now without violation Nagle's rules:
1077 * 1. It is full sized.
1078 * 2. Or it contains FIN. (already checked by caller)
1079 * 3. Or TCP_NODELAY was set.
1080 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1081 * With Minshall's modification: all sent small packets are ACKed.
1082 */
1083
1084static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001085 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001086 unsigned mss_now, int nonagle)
1087{
1088 return (skb->len < mss_now &&
1089 ((nonagle&TCP_NAGLE_CORK) ||
1090 (!nonagle &&
1091 tp->packets_out &&
1092 tcp_minshall_check(tp))));
1093}
1094
1095/* Return non-zero if the Nagle test allows this packet to be
1096 * sent now.
1097 */
1098static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1099 unsigned int cur_mss, int nonagle)
1100{
1101 /* Nagle rule does not apply to frames, which sit in the middle of the
1102 * write_queue (they have no chances to get new data).
1103 *
1104 * This is implemented in the callers, where they modify the 'nonagle'
1105 * argument based upon the location of SKB in the send queue.
1106 */
1107 if (nonagle & TCP_NAGLE_PUSH)
1108 return 1;
1109
Ilpo Järvinend551e452007-04-30 00:42:20 -07001110 /* Don't use the nagle rule for urgent data (or for the final FIN).
1111 * Nagle can be ignored during F-RTO too (see RFC4138).
1112 */
1113 if (tp->urg_mode || (tp->frto_counter == 2) ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001114 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1115 return 1;
1116
1117 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1118 return 1;
1119
1120 return 0;
1121}
1122
1123/* Does at least the first segment of SKB fit into the send window? */
1124static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
1125{
1126 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1127
1128 if (skb->len > cur_mss)
1129 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1130
1131 return !after(end_seq, tp->snd_una + tp->snd_wnd);
1132}
1133
David S. Millerfe067e82007-03-07 12:12:44 -08001134/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001135 * should be put on the wire right now. If so, it returns the number of
1136 * packets allowed by the congestion window.
1137 */
1138static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1139 unsigned int cur_mss, int nonagle)
1140{
1141 struct tcp_sock *tp = tcp_sk(sk);
1142 unsigned int cwnd_quota;
1143
David S. Miller846998a2005-08-04 19:52:01 -07001144 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001145
1146 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1147 return 0;
1148
1149 cwnd_quota = tcp_cwnd_test(tp, skb);
1150 if (cwnd_quota &&
1151 !tcp_snd_wnd_test(tp, skb, cur_mss))
1152 cwnd_quota = 0;
1153
1154 return cwnd_quota;
1155}
1156
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001157int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001158{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001159 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001160 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001161
1162 return (skb &&
1163 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1164 (tcp_skb_is_last(sk, skb) ?
Ilpo Järvinen4e67d872007-12-05 02:25:32 -08001165 tp->nonagle : TCP_NAGLE_PUSH)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001166}
1167
1168/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1169 * which is put after SKB on the list. It is very much like
1170 * tcp_fragment() except that it may make several kinds of assumptions
1171 * in order to speed up the splitting operation. In particular, we
1172 * know that all the data is in scatter-gather pages, and that the
1173 * packet has never been sent out before (and thus is not cloned).
1174 */
David S. Miller846998a2005-08-04 19:52:01 -07001175static 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 -07001176{
1177 struct sk_buff *buff;
1178 int nlen = skb->len - len;
1179 u16 flags;
1180
1181 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001182 if (skb->len != skb->data_len)
1183 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001184
1185 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1186 if (unlikely(buff == NULL))
1187 return -ENOMEM;
1188
Herbert Xub60b49e2006-04-19 21:35:00 -07001189 sk_charge_skb(sk, buff);
1190 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001191 skb->truesize -= nlen;
1192
1193 /* Correct the sequence numbers. */
1194 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1195 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1196 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1197
1198 /* PSH and FIN should only be set in the second packet. */
1199 flags = TCP_SKB_CB(skb)->flags;
1200 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1201 TCP_SKB_CB(buff)->flags = flags;
1202
1203 /* This packet was never sent out yet, so no SACK bits. */
1204 TCP_SKB_CB(buff)->sacked = 0;
1205
Patrick McHardy84fa7932006-08-29 16:44:56 -07001206 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001207 skb_split(skb, buff, len);
1208
1209 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001210 tcp_set_skb_tso_segs(sk, skb, mss_now);
1211 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001212
1213 /* Link BUFF into the send queue. */
1214 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001215 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001216
1217 return 0;
1218}
1219
1220/* Try to defer sending, if possible, in order to minimize the amount
1221 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1222 *
1223 * This algorithm is from John Heffner.
1224 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001225static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001226{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001227 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001228 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001229 u32 send_win, cong_win, limit, in_flight;
1230
1231 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001232 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001233
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001234 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001235 goto send_now;
1236
1237 /* Defer for less than two clock ticks. */
1238 if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1)
1239 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001240
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001241 in_flight = tcp_packets_in_flight(tp);
1242
1243 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1244 (tp->snd_cwnd <= in_flight));
1245
1246 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1247
1248 /* From in_flight test above, we know that cwnd > in_flight. */
1249 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1250
1251 limit = min(send_win, cong_win);
1252
David S. Millerba244fe2006-03-11 18:51:49 -08001253 /* If a full-sized TSO skb can be sent, do it. */
1254 if (limit >= 65536)
John Heffnerae8064a2006-10-18 20:36:48 -07001255 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001256
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001257 if (sysctl_tcp_tso_win_divisor) {
1258 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1259
1260 /* If at least some fraction of a window is available,
1261 * just use it.
1262 */
1263 chunk /= sysctl_tcp_tso_win_divisor;
1264 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001265 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001266 } else {
1267 /* Different approach, try not to defer past a single
1268 * ACK. Receiver should ACK every other full sized
1269 * frame, so if we have space for more than 3 frames
1270 * then send now.
1271 */
1272 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001273 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001274 }
1275
1276 /* Ok, it looks like it is advisable to defer. */
John Heffnerae8064a2006-10-18 20:36:48 -07001277 tp->tso_deferred = 1 | (jiffies<<1);
1278
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001279 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001280
1281send_now:
1282 tp->tso_deferred = 0;
1283 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001284}
1285
John Heffner5d424d52006-03-20 17:53:41 -08001286/* Create a new MTU probe if we are ready.
1287 * Returns 0 if we should wait to probe (no cwnd available),
1288 * 1 if a probe was sent,
1289 * -1 otherwise */
1290static int tcp_mtu_probe(struct sock *sk)
1291{
1292 struct tcp_sock *tp = tcp_sk(sk);
1293 struct inet_connection_sock *icsk = inet_csk(sk);
1294 struct sk_buff *skb, *nskb, *next;
1295 int len;
1296 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001297 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001298 unsigned int pif;
1299 int copy;
1300 int mss_now;
1301
1302 /* Not currently probing/verifying,
1303 * not in recovery,
1304 * have enough cwnd, and
1305 * not SACKing (the variable headers throw things off) */
1306 if (!icsk->icsk_mtup.enabled ||
1307 icsk->icsk_mtup.probe_size ||
1308 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1309 tp->snd_cwnd < 11 ||
1310 tp->rx_opt.eff_sacks)
1311 return -1;
1312
1313 /* Very simple search strategy: just double the MSS. */
1314 mss_now = tcp_current_mss(sk, 0);
1315 probe_size = 2*tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001316 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001317 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1318 /* TODO: set timer for probe_converge_event */
1319 return -1;
1320 }
1321
1322 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001323 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001324 return -1;
1325
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001326 if (tp->snd_wnd < size_needed)
1327 return -1;
1328 if (after(tp->snd_nxt + size_needed, tp->snd_una + tp->snd_wnd))
1329 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001330
1331 /* Do we need to wait to drain cwnd? */
1332 pif = tcp_packets_in_flight(tp);
1333 if (pif + 2 > tp->snd_cwnd) {
1334 /* With no packets in flight, don't stall. */
1335 if (pif == 0)
1336 return -1;
1337 else
1338 return 0;
1339 }
1340
1341 /* We're allowed to probe. Build it now. */
1342 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1343 return -1;
1344 sk_charge_skb(sk, nskb);
1345
David S. Millerfe067e82007-03-07 12:12:44 -08001346 skb = tcp_send_head(sk);
1347 tcp_insert_write_queue_before(nskb, skb, sk);
John Heffner5d424d52006-03-20 17:53:41 -08001348
1349 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1350 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1351 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1352 TCP_SKB_CB(nskb)->sacked = 0;
1353 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001354 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001355
1356 len = 0;
1357 while (len < probe_size) {
David S. Millerfe067e82007-03-07 12:12:44 -08001358 next = tcp_write_queue_next(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001359
1360 copy = min_t(int, skb->len, probe_size - len);
1361 if (nskb->ip_summed)
1362 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1363 else
1364 nskb->csum = skb_copy_and_csum_bits(skb, 0,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001365 skb_put(nskb, copy), copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001366
1367 if (skb->len <= copy) {
1368 /* We've eaten all the data from this skb.
1369 * Throw it away. */
1370 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001371 tcp_unlink_write_queue(skb, sk);
John Heffner5d424d52006-03-20 17:53:41 -08001372 sk_stream_free_skb(sk, skb);
1373 } else {
1374 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001375 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001376 if (!skb_shinfo(skb)->nr_frags) {
1377 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001378 if (skb->ip_summed != CHECKSUM_PARTIAL)
John Heffner5d424d52006-03-20 17:53:41 -08001379 skb->csum = csum_partial(skb->data, skb->len, 0);
1380 } else {
1381 __pskb_trim_head(skb, copy);
1382 tcp_set_skb_tso_segs(sk, skb, mss_now);
1383 }
1384 TCP_SKB_CB(skb)->seq += copy;
1385 }
1386
1387 len += copy;
1388 skb = next;
1389 }
1390 tcp_init_tso_segs(sk, nskb, nskb->len);
1391
1392 /* We're ready to send. If this fails, the probe will
1393 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1394 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1395 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1396 /* Decrement cwnd here because we are sending
1397 * effectively two packets. */
1398 tp->snd_cwnd--;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001399 update_send_head(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001400
1401 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001402 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1403 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001404
1405 return 1;
1406 }
1407
1408 return -1;
1409}
1410
1411
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412/* This routine writes packets to the network. It advances the
1413 * send_head. This happens as incoming acks open up the remote
1414 * window for us.
1415 *
1416 * Returns 1, if no segments are in flight and we have queued segments, but
1417 * cannot send anything now because of SWS or another problem.
1418 */
David S. Millera2e2a592005-07-05 15:19:23 -07001419static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420{
1421 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001422 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001423 unsigned int tso_segs, sent_pkts;
1424 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001425 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426
1427 /* If we are closed, the bytes will have to remain here.
1428 * In time closedown will finish, we empty the write queue and all
1429 * will be happy.
1430 */
David S. Miller92df7b52005-07-05 15:19:06 -07001431 if (unlikely(sk->sk_state == TCP_CLOSE))
1432 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433
David S. Miller92df7b52005-07-05 15:19:06 -07001434 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001435
1436 /* Do MTU probing. */
1437 if ((result = tcp_mtu_probe(sk)) == 0) {
1438 return 0;
1439 } else if (result > 0) {
1440 sent_pkts = 1;
1441 }
1442
David S. Millerfe067e82007-03-07 12:12:44 -08001443 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001444 unsigned int limit;
1445
Herbert Xub68e9f82005-08-04 19:52:02 -07001446 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001447 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001448
Herbert Xub68e9f82005-08-04 19:52:02 -07001449 cwnd_quota = tcp_cwnd_test(tp, skb);
1450 if (!cwnd_quota)
1451 break;
1452
1453 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1454 break;
1455
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001456 if (tso_segs == 1) {
1457 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1458 (tcp_skb_is_last(sk, skb) ?
1459 nonagle : TCP_NAGLE_PUSH))))
1460 break;
1461 } else {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001462 if (tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001463 break;
1464 }
David S. Milleraa934662005-07-05 15:20:09 -07001465
Herbert Xuc8ac3772005-08-16 20:43:40 -07001466 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001467 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001468 limit = tcp_window_allows(tp, skb,
1469 mss_now, cwnd_quota);
David S. Milleraa934662005-07-05 15:20:09 -07001470
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001471 if (skb->len < limit) {
1472 unsigned int trim = skb->len % mss_now;
1473
1474 if (trim)
1475 limit = skb->len - trim;
1476 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 }
1478
Herbert Xuc8ac3772005-08-16 20:43:40 -07001479 if (skb->len > limit &&
1480 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1481 break;
1482
David S. Miller92df7b52005-07-05 15:19:06 -07001483 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001484
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001485 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
David S. Miller92df7b52005-07-05 15:19:06 -07001486 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487
David S. Miller92df7b52005-07-05 15:19:06 -07001488 /* Advance the send_head. This one is sent out.
1489 * This call will increment packets_out.
1490 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001491 update_send_head(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001492
1493 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001494 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 }
David S. Miller92df7b52005-07-05 15:19:06 -07001496
David S. Milleraa934662005-07-05 15:20:09 -07001497 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001498 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001499 return 0;
1500 }
David S. Millerfe067e82007-03-07 12:12:44 -08001501 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502}
1503
David S. Millera762a982005-07-05 15:18:51 -07001504/* Push out any pending frames which were held back due to
1505 * TCP_CORK or attempt at coalescing tiny packets.
1506 * The socket must be locked by the caller.
1507 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001508void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1509 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001510{
David S. Millerfe067e82007-03-07 12:12:44 -08001511 struct sk_buff *skb = tcp_send_head(sk);
David S. Millera762a982005-07-05 15:18:51 -07001512
1513 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001514 if (tcp_write_xmit(sk, cur_mss, nonagle))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001515 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001516 }
1517}
1518
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001519/* Send _single_ skb sitting at the send head. This function requires
1520 * true push pending frames to setup probe timer etc.
1521 */
1522void tcp_push_one(struct sock *sk, unsigned int mss_now)
1523{
1524 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001525 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001526 unsigned int tso_segs, cwnd_quota;
1527
1528 BUG_ON(!skb || skb->len < mss_now);
1529
David S. Miller846998a2005-08-04 19:52:01 -07001530 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001531 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1532
1533 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001534 unsigned int limit;
1535
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001536 BUG_ON(!tso_segs);
1537
Herbert Xuc8ac3772005-08-16 20:43:40 -07001538 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001539 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001540 limit = tcp_window_allows(tp, skb,
1541 mss_now, cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001542
1543 if (skb->len < limit) {
1544 unsigned int trim = skb->len % mss_now;
1545
1546 if (trim)
1547 limit = skb->len - trim;
1548 }
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001549 }
1550
Herbert Xuc8ac3772005-08-16 20:43:40 -07001551 if (skb->len > limit &&
1552 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1553 return;
1554
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001555 /* Send it out now. */
1556 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1557
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001558 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001559 update_send_head(sk, skb);
1560 tcp_cwnd_validate(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001561 return;
1562 }
1563 }
1564}
1565
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566/* This function returns the amount that we can raise the
1567 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001568 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 * 1. The window can never be shrunk once it is offered (RFC 793)
1570 * 2. We limit memory per socket
1571 *
1572 * RFC 1122:
1573 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1574 * RECV.NEXT + RCV.WIN fixed until:
1575 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1576 *
1577 * i.e. don't raise the right edge of the window until you can raise
1578 * it at least MSS bytes.
1579 *
1580 * Unfortunately, the recommended algorithm breaks header prediction,
1581 * since header prediction assumes th->window stays fixed.
1582 *
1583 * Strictly speaking, keeping th->window fixed violates the receiver
1584 * side SWS prevention criteria. The problem is that under this rule
1585 * a stream of single byte packets will cause the right side of the
1586 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001587 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 * Of course, if the sender implements sender side SWS prevention
1589 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001590 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001592 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 * If the free space is less than the 1/4 of the maximum
1594 * space available and the free space is less than 1/2 mss,
1595 * then set the window to 0.
1596 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1597 * Otherwise, just prevent the window from shrinking
1598 * and from being larger than the largest representable value.
1599 *
1600 * This prevents incremental opening of the window in the regime
1601 * where TCP is limited by the speed of the reader side taking
1602 * data out of the TCP receive queue. It does nothing about
1603 * those cases where the window is constrained on the sender side
1604 * because the pipeline is full.
1605 *
1606 * BSD also seems to "accidentally" limit itself to windows that are a
1607 * multiple of MSS, at least until the free space gets quite small.
1608 * This would appear to be a side effect of the mbuf implementation.
1609 * Combining these two algorithms results in the observed behavior
1610 * of having a fixed window size at almost all times.
1611 *
1612 * Below we obtain similar behavior by forcing the offered window to
1613 * a multiple of the mss when it is feasible to do so.
1614 *
1615 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1616 * Regular options like TIMESTAMP are taken into account.
1617 */
1618u32 __tcp_select_window(struct sock *sk)
1619{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001620 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001622 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 * here. I don't know if the value based on our guesses
1624 * of peer's MSS is better for the performance. It's more correct
1625 * but may be worse for the performance because of rcv_mss
1626 * fluctuations. --SAW 1998/11/1
1627 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001628 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 int free_space = tcp_space(sk);
1630 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1631 int window;
1632
1633 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001634 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635
1636 if (free_space < full_space/2) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001637 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638
1639 if (tcp_memory_pressure)
1640 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1641
1642 if (free_space < mss)
1643 return 0;
1644 }
1645
1646 if (free_space > tp->rcv_ssthresh)
1647 free_space = tp->rcv_ssthresh;
1648
1649 /* Don't do rounding if we are using window scaling, since the
1650 * scaled window will not line up with the MSS boundary anyway.
1651 */
1652 window = tp->rcv_wnd;
1653 if (tp->rx_opt.rcv_wscale) {
1654 window = free_space;
1655
1656 /* Advertise enough space so that it won't get scaled away.
1657 * Import case: prevent zero window announcement if
1658 * 1<<rcv_wscale > mss.
1659 */
1660 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1661 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1662 << tp->rx_opt.rcv_wscale);
1663 } else {
1664 /* Get the largest window that is a nice multiple of mss.
1665 * Window clamp already applied above.
1666 * If our current window offering is within 1 mss of the
1667 * free space we just keep it. This prevents the divide
1668 * and multiply from happening most of the time.
1669 * We also don't do any window rounding when the free space
1670 * is too small.
1671 */
1672 if (window <= free_space - mss || window > free_space)
1673 window = (free_space/mss)*mss;
John Heffner84565072007-04-02 13:56:32 -07001674 else if (mss == full_space &&
YOSHIFUJI Hideaki9c681b42007-07-19 10:43:47 +09001675 free_space > window + full_space/2)
John Heffner84565072007-04-02 13:56:32 -07001676 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 }
1678
1679 return window;
1680}
1681
1682/* Attempt to collapse two adjacent SKB's during retransmission. */
1683static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1684{
1685 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001686 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687
1688 /* The first test we must make is that neither of these two
1689 * SKB's are still referenced by someone else.
1690 */
1691 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1692 int skb_size = skb->len, next_skb_size = next_skb->len;
1693 u16 flags = TCP_SKB_CB(skb)->flags;
1694
1695 /* Also punt if next skb has been SACK'd. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001696 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 return;
1698
1699 /* Next skb is out of window. */
1700 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1701 return;
1702
1703 /* Punt if not enough space exists in the first SKB for
1704 * the data in the second, or the total combined payload
1705 * would exceed the MSS.
1706 */
1707 if ((next_skb_size > skb_tailroom(skb)) ||
1708 ((skb_size + next_skb_size) > mss_now))
1709 return;
1710
1711 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1712 tcp_skb_pcount(next_skb) != 1);
1713
Ilpo Järvinendc869672007-10-01 15:27:19 -07001714 if (WARN_ON(tcp_is_sack(tp) && tp->sacked_out &&
Ilpo Järvinena6963a62007-09-25 22:44:14 -07001715 (TCP_SKB_CB(next_skb)->seq == tp->highest_sack)))
1716 return;
1717
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001718 /* Ok. We will be able to collapse the packet. */
David S. Millerfe067e82007-03-07 12:12:44 -08001719 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720
Arnaldo Carvalho de Melo1a4e2d02007-03-31 11:55:45 -03001721 skb_copy_from_linear_data(next_skb,
1722 skb_put(skb, next_skb_size),
1723 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724
Jarek Poplawski52d570a2007-01-23 22:07:12 -08001725 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1726 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727
Patrick McHardy84fa7932006-08-29 16:44:56 -07001728 if (skb->ip_summed != CHECKSUM_PARTIAL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1730
1731 /* Update sequence range on original skb. */
1732 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1733
1734 /* Merge over control information. */
1735 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1736 TCP_SKB_CB(skb)->flags = flags;
1737
1738 /* All done, get rid of second SKB and account for it so
1739 * packet counting does not break.
1740 */
1741 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1742 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1743 tp->retrans_out -= tcp_skb_pcount(next_skb);
Ilpo Järvinenb5860bb2007-08-09 14:33:18 +03001744 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 tp->lost_out -= tcp_skb_pcount(next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 /* Reno case is special. Sigh... */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001747 if (tcp_is_reno(tp) && tp->sacked_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749
Ilpo Järvinen29d0a302007-10-07 23:36:41 -07001750 tcp_adjust_fackets_out(tp, next_skb, tcp_skb_pcount(next_skb));
Ilpo Järvinene9144bd2007-08-24 22:43:14 -07001751 tp->packets_out -= tcp_skb_pcount(next_skb);
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001752
1753 /* changed transmit queue under us so clear hints */
1754 tcp_clear_retrans_hints_partial(tp);
1755 /* manually tune sacktag skb hint */
Ilpo Järvinen29d0a302007-10-07 23:36:41 -07001756 if (tp->fastpath_skb_hint == next_skb) {
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001757 tp->fastpath_skb_hint = skb;
Ilpo Järvinen29d0a302007-10-07 23:36:41 -07001758 tp->fastpath_cnt_hint -= tcp_skb_pcount(skb);
1759 }
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001760
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 sk_stream_free_skb(sk, next_skb);
1762 }
1763}
1764
1765/* Do a simple retransmit without using the backoff mechanisms in
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001766 * tcp_timer. This is used for path mtu discovery.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767 * The socket is already locked here.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001768 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769void tcp_simple_retransmit(struct sock *sk)
1770{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001771 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 struct tcp_sock *tp = tcp_sk(sk);
1773 struct sk_buff *skb;
1774 unsigned int mss = tcp_current_mss(sk, 0);
1775 int lost = 0;
1776
David S. Millerfe067e82007-03-07 12:12:44 -08001777 tcp_for_write_queue(skb, sk) {
1778 if (skb == tcp_send_head(sk))
1779 break;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001780 if (skb->len > mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1782 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1783 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1784 tp->retrans_out -= tcp_skb_pcount(skb);
1785 }
1786 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1787 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1788 tp->lost_out += tcp_skb_pcount(skb);
1789 lost = 1;
1790 }
1791 }
1792 }
1793
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07001794 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001795
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 if (!lost)
1797 return;
1798
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001799 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001801 /* Don't muck with the congestion window here.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 * Reason is that we do not increase amount of _data_
1803 * in network, but units changed and effective
1804 * cwnd/ssthresh really reduced now.
1805 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001806 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001808 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809 tp->prior_ssthresh = 0;
1810 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001811 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 }
1813 tcp_xmit_retransmit_queue(sk);
1814}
1815
1816/* This retransmits one SKB. Policy decisions and retransmit queue
1817 * state updates are done by the caller. Returns non-zero if an
1818 * error occurred which prevented the send.
1819 */
1820int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1821{
1822 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001823 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001824 unsigned int cur_mss = tcp_current_mss(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 int err;
1826
John Heffner5d424d52006-03-20 17:53:41 -08001827 /* Inconslusive MTU probe */
1828 if (icsk->icsk_mtup.probe_size) {
1829 icsk->icsk_mtup.probe_size = 0;
1830 }
1831
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001833 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 */
1835 if (atomic_read(&sk->sk_wmem_alloc) >
1836 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1837 return -EAGAIN;
1838
1839 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1840 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1841 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1843 return -ENOMEM;
1844 }
1845
1846 /* If receiver has shrunk his window, and skb is out of
1847 * new window, do not retransmit it. The exception is the
1848 * case, when window is shrunk to zero. In this case
1849 * our retransmit serves as a zero window probe.
1850 */
1851 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1852 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1853 return -EAGAIN;
1854
1855 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001856 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 }
1859
1860 /* Collapse two adjacent packets if worthwhile and we can. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001861 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1862 (skb->len < (cur_mss >> 1)) &&
1863 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1864 (!tcp_skb_is_last(sk, skb)) &&
1865 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1866 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1867 (sysctl_tcp_retrans_collapse != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 tcp_retrans_try_collapse(sk, skb, cur_mss);
1869
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001870 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 return -EHOSTUNREACH; /* Routing failure or similar. */
1872
1873 /* Some Solaris stacks overoptimize and ignore the FIN on a
1874 * retransmit when old data is attached. So strip it off
1875 * since it is cheap to do so and saves bytes on the network.
1876 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001877 if (skb->len > 0 &&
1878 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1879 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 if (!pskb_trim(skb, 0)) {
1881 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
Herbert Xu79671682006-06-22 02:40:14 -07001882 skb_shinfo(skb)->gso_segs = 1;
1883 skb_shinfo(skb)->gso_size = 0;
1884 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 skb->ip_summed = CHECKSUM_NONE;
1886 skb->csum = 0;
1887 }
1888 }
1889
1890 /* Make a copy, if the first transmission SKB clone we made
1891 * is still in somebody's hands, else make a clone.
1892 */
1893 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001895 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896
1897 if (err == 0) {
1898 /* Update global TCP statistics. */
1899 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1900
1901 tp->total_retrans++;
1902
1903#if FASTRETRANS_DEBUG > 0
1904 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1905 if (net_ratelimit())
1906 printk(KERN_DEBUG "retrans_out leaked.\n");
1907 }
1908#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001909 if (!tp->retrans_out)
1910 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1912 tp->retrans_out += tcp_skb_pcount(skb);
1913
1914 /* Save stamp of the first retransmit. */
1915 if (!tp->retrans_stamp)
1916 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1917
1918 tp->undo_retrans++;
1919
1920 /* snd_nxt is stored to detect loss of retransmitted segment,
1921 * see tcp_input.c tcp_sacktag_write_queue().
1922 */
1923 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1924 }
1925 return err;
1926}
1927
1928/* This gets called after a retransmit timeout, and the initially
1929 * retransmitted data is acknowledged. It tries to continue
1930 * resending the rest of the retransmit queue, until either
1931 * we've sent it all or the congestion window limit is reached.
1932 * If doing SACK, the first ACK which comes back for a timeout
1933 * based retransmit packet might feed us FACK information again.
1934 * If so, we use it to avoid unnecessarily retransmissions.
1935 */
1936void tcp_xmit_retransmit_queue(struct sock *sk)
1937{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001938 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 struct tcp_sock *tp = tcp_sk(sk);
1940 struct sk_buff *skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001941 int packet_cnt;
1942
1943 if (tp->retransmit_skb_hint) {
1944 skb = tp->retransmit_skb_hint;
1945 packet_cnt = tp->retransmit_cnt_hint;
1946 }else{
David S. Millerfe067e82007-03-07 12:12:44 -08001947 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001948 packet_cnt = 0;
1949 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950
1951 /* First pass: retransmit lost packets. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001952 if (tp->lost_out) {
David S. Millerfe067e82007-03-07 12:12:44 -08001953 tcp_for_write_queue_from(skb, sk) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1955
David S. Millerfe067e82007-03-07 12:12:44 -08001956 if (skb == tcp_send_head(sk))
1957 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001958 /* we could do better than to assign each time */
1959 tp->retransmit_skb_hint = skb;
1960 tp->retransmit_cnt_hint = packet_cnt;
1961
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 /* Assume this retransmit will generate
1963 * only one packet for congestion window
1964 * calculation purposes. This works because
1965 * tcp_retransmit_skb() will chop up the
1966 * packet to be MSS sized and all the
1967 * packet counting works out.
1968 */
1969 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1970 return;
1971
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001972 if (sacked & TCPCB_LOST) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001974 if (tcp_retransmit_skb(sk, skb)) {
1975 tp->retransmit_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001977 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001978 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1980 else
1981 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1982
David S. Millerfe067e82007-03-07 12:12:44 -08001983 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001984 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001985 inet_csk(sk)->icsk_rto,
1986 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987 }
1988
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001989 packet_cnt += tcp_skb_pcount(skb);
1990 if (packet_cnt >= tp->lost_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 break;
1992 }
1993 }
1994 }
1995
1996 /* OK, demanded retransmission is finished. */
1997
1998 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001999 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000 return;
2001
2002 /* No forward retransmissions in Reno are possible. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002003 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 return;
2005
2006 /* Yeah, we have to make difficult choice between forward transmission
2007 * and retransmission... Both ways have their merits...
2008 *
2009 * For now we do not retransmit anything, while we have some new
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002010 * segments to send. In the other cases, follow rule 3 for
2011 * NextSeg() specified in RFC3517.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 */
2013
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002014 if (tcp_may_send_now(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 return;
2016
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002017 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2018 if (!tp->sacked_out)
2019 return;
2020
2021 if (tp->forward_skb_hint)
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002022 skb = tp->forward_skb_hint;
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002023 else
David S. Millerfe067e82007-03-07 12:12:44 -08002024 skb = tcp_write_queue_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025
David S. Millerfe067e82007-03-07 12:12:44 -08002026 tcp_for_write_queue_from(skb, sk) {
2027 if (skb == tcp_send_head(sk))
2028 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002029 tp->forward_skb_hint = skb;
2030
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002031 if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 break;
2033
2034 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2035 break;
2036
2037 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2038 continue;
2039
2040 /* Ok, retransmit it. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002041 if (tcp_retransmit_skb(sk, skb)) {
2042 tp->forward_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002044 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045
David S. Millerfe067e82007-03-07 12:12:44 -08002046 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002047 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2048 inet_csk(sk)->icsk_rto,
2049 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050
2051 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2052 }
2053}
2054
2055
2056/* Send a fin. The caller locks the socket for us. This cannot be
2057 * allowed to fail queueing a FIN frame under any circumstances.
2058 */
2059void tcp_send_fin(struct sock *sk)
2060{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002061 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002062 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002064
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 /* Optimization, tack on the FIN if we have a queue of
2066 * unsent frames. But be careful about outgoing SACKS
2067 * and IP options.
2068 */
2069 mss_now = tcp_current_mss(sk, 1);
2070
David S. Millerfe067e82007-03-07 12:12:44 -08002071 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2073 TCP_SKB_CB(skb)->end_seq++;
2074 tp->write_seq++;
2075 } else {
2076 /* Socket is locked, keep trying until memory is available. */
2077 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07002078 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 if (skb)
2080 break;
2081 yield();
2082 }
2083
2084 /* Reserve space for headers and prepare control bits. */
2085 skb_reserve(skb, MAX_TCP_HEADER);
2086 skb->csum = 0;
2087 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2088 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002089 skb_shinfo(skb)->gso_segs = 1;
2090 skb_shinfo(skb)->gso_size = 0;
2091 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092
2093 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2094 TCP_SKB_CB(skb)->seq = tp->write_seq;
2095 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2096 tcp_queue_skb(sk, skb);
2097 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002098 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099}
2100
2101/* We get here when a process closes a file descriptor (either due to
2102 * an explicit close() or as a byproduct of exit()'ing) and there
2103 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002104 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 */
Al Virodd0fc662005-10-07 07:46:04 +01002106void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 struct sk_buff *skb;
2109
2110 /* NOTE: No TCP options attached and we never retransmit this. */
2111 skb = alloc_skb(MAX_TCP_HEADER, priority);
2112 if (!skb) {
2113 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2114 return;
2115 }
2116
2117 /* Reserve space for headers and prepare control bits. */
2118 skb_reserve(skb, MAX_TCP_HEADER);
2119 skb->csum = 0;
2120 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2121 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002122 skb_shinfo(skb)->gso_segs = 1;
2123 skb_shinfo(skb)->gso_size = 0;
2124 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125
2126 /* Send it off. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002127 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2129 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002130 if (tcp_transmit_skb(sk, skb, 0, priority))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2132}
2133
2134/* WARNING: This routine must only be called when we have already sent
2135 * a SYN packet that crossed the incoming SYN that caused this routine
2136 * to get called. If this assumption fails then the initial rcv_wnd
2137 * and rcv_wscale values will not be correct.
2138 */
2139int tcp_send_synack(struct sock *sk)
2140{
2141 struct sk_buff* skb;
2142
David S. Millerfe067e82007-03-07 12:12:44 -08002143 skb = tcp_write_queue_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
2145 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2146 return -EFAULT;
2147 }
2148 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
2149 if (skb_cloned(skb)) {
2150 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2151 if (nskb == NULL)
2152 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002153 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002155 __tcp_add_write_queue_head(sk, nskb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 sk_stream_free_skb(sk, skb);
2157 sk_charge_skb(sk, nskb);
2158 skb = nskb;
2159 }
2160
2161 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2162 TCP_ECN_send_synack(tcp_sk(sk), skb);
2163 }
2164 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002165 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166}
2167
2168/*
2169 * Prepare a SYN-ACK.
2170 */
2171struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
Arnaldo Carvalho de Melo60236fd2005-06-18 22:47:21 -07002172 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002174 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 struct tcp_sock *tp = tcp_sk(sk);
2176 struct tcphdr *th;
2177 int tcp_header_size;
2178 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002179#ifdef CONFIG_TCP_MD5SIG
2180 struct tcp_md5sig_key *md5;
2181 __u8 *md5_hash_location;
2182#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183
2184 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2185 if (skb == NULL)
2186 return NULL;
2187
2188 /* Reserve space for headers. */
2189 skb_reserve(skb, MAX_TCP_HEADER);
2190
2191 skb->dst = dst_clone(dst);
2192
2193 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002194 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2195 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002197 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002198
2199#ifdef CONFIG_TCP_MD5SIG
2200 /* Are we doing MD5 on this segment? If so - make room for it */
2201 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2202 if (md5)
2203 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2204#endif
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002205 skb_push(skb, tcp_header_size);
2206 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002208 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 memset(th, 0, sizeof(struct tcphdr));
2210 th->syn = 1;
2211 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 TCP_ECN_make_synack(req, th);
2213 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002214 th->dest = ireq->rmt_port;
2215 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2217 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002218 skb_shinfo(skb)->gso_segs = 1;
2219 skb_shinfo(skb)->gso_size = 0;
2220 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002222 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002224 __u8 rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 /* Set this up on the first call only */
2226 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2227 /* tcp_full_space because it is guaranteed to be the first packet */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002228 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002229 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 &req->rcv_wnd,
2231 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002232 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 &rcv_wscale);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002234 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 }
2236
2237 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002238 th->window = htons(min(req->rcv_wnd, 65535U));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239
2240 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Al Virodf7a3b02006-09-27 18:38:52 -07002241 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002242 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 TCP_SKB_CB(skb)->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002244 req->ts_recent,
2245 (
2246#ifdef CONFIG_TCP_MD5SIG
2247 md5 ? &md5_hash_location :
2248#endif
2249 NULL)
2250 );
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251
2252 skb->csum = 0;
2253 th->doff = (tcp_header_size >> 2);
2254 TCP_INC_STATS(TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002255
2256#ifdef CONFIG_TCP_MD5SIG
2257 /* Okay, we have all we need - do the md5 hash if needed */
2258 if (md5) {
2259 tp->af_specific->calc_md5_hash(md5_hash_location,
2260 md5,
2261 NULL, dst, req,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002262 tcp_hdr(skb), sk->sk_protocol,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002263 skb->len);
2264 }
2265#endif
2266
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 return skb;
2268}
2269
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002270/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 * Do all connect socket setups that can be done AF independent.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002272 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002273static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274{
2275 struct dst_entry *dst = __sk_dst_get(sk);
2276 struct tcp_sock *tp = tcp_sk(sk);
2277 __u8 rcv_wscale;
2278
2279 /* We'll fix this up when we get a response from the other end.
2280 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2281 */
2282 tp->tcp_header_len = sizeof(struct tcphdr) +
2283 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2284
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002285#ifdef CONFIG_TCP_MD5SIG
2286 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2287 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2288#endif
2289
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 /* If user gave his TCP_MAXSEG, record it to clamp */
2291 if (tp->rx_opt.user_mss)
2292 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2293 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002294 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 tcp_sync_mss(sk, dst_mtu(dst));
2296
2297 if (!tp->window_clamp)
2298 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2299 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2300 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301
2302 tcp_select_initial_window(tcp_full_space(sk),
2303 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2304 &tp->rcv_wnd,
2305 &tp->window_clamp,
2306 sysctl_tcp_window_scaling,
2307 &rcv_wscale);
2308
2309 tp->rx_opt.rcv_wscale = rcv_wscale;
2310 tp->rcv_ssthresh = tp->rcv_wnd;
2311
2312 sk->sk_err = 0;
2313 sock_reset_flag(sk, SOCK_DONE);
2314 tp->snd_wnd = 0;
2315 tcp_init_wl(tp, tp->write_seq, 0);
2316 tp->snd_una = tp->write_seq;
2317 tp->snd_sml = tp->write_seq;
2318 tp->rcv_nxt = 0;
2319 tp->rcv_wup = 0;
2320 tp->copied_seq = 0;
2321
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002322 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2323 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 tcp_clear_retrans(tp);
2325}
2326
2327/*
2328 * Build a SYN and send it off.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002329 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330int tcp_connect(struct sock *sk)
2331{
2332 struct tcp_sock *tp = tcp_sk(sk);
2333 struct sk_buff *buff;
2334
2335 tcp_connect_init(sk);
2336
David S. Millerd179cd12005-08-17 14:57:30 -07002337 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 if (unlikely(buff == NULL))
2339 return -ENOBUFS;
2340
2341 /* Reserve space for headers. */
2342 skb_reserve(buff, MAX_TCP_HEADER);
2343
2344 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002345 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002347 skb_shinfo(buff)->gso_segs = 1;
2348 skb_shinfo(buff)->gso_size = 0;
2349 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 buff->csum = 0;
Wei Yongjunbd37a082006-08-07 21:04:15 -07002351 tp->snd_nxt = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2353 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354
2355 /* Send it off. */
2356 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2357 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2358 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002359 __tcp_add_write_queue_tail(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 sk_charge_skb(sk, buff);
2361 tp->packets_out += tcp_skb_pcount(buff);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002362 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002363
2364 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2365 * in order to make this packet get counted in tcpOutSegs.
2366 */
2367 tp->snd_nxt = tp->write_seq;
2368 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2370
2371 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002372 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2373 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 return 0;
2375}
2376
2377/* Send out a delayed ack, the caller does the policy checking
2378 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2379 * for details.
2380 */
2381void tcp_send_delayed_ack(struct sock *sk)
2382{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002383 struct inet_connection_sock *icsk = inet_csk(sk);
2384 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 unsigned long timeout;
2386
2387 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002388 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 int max_ato = HZ/2;
2390
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002391 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 max_ato = TCP_DELACK_MAX;
2393
2394 /* Slow path, intersegment interval is "high". */
2395
2396 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002397 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 * directly.
2399 */
2400 if (tp->srtt) {
2401 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2402
2403 if (rtt < max_ato)
2404 max_ato = rtt;
2405 }
2406
2407 ato = min(ato, max_ato);
2408 }
2409
2410 /* Stay within the limit we were given */
2411 timeout = jiffies + ato;
2412
2413 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002414 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 /* If delack timer was blocked or is about to expire,
2416 * send ACK now.
2417 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002418 if (icsk->icsk_ack.blocked ||
2419 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 tcp_send_ack(sk);
2421 return;
2422 }
2423
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002424 if (!time_before(timeout, icsk->icsk_ack.timeout))
2425 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002427 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2428 icsk->icsk_ack.timeout = timeout;
2429 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430}
2431
2432/* This routine sends an ack and also updates the window. */
2433void tcp_send_ack(struct sock *sk)
2434{
2435 /* If we have been reset, we may not send again. */
2436 if (sk->sk_state != TCP_CLOSE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 struct sk_buff *buff;
2438
2439 /* We are not putting this on the write queue, so
2440 * tcp_transmit_skb() will set the ownership to this
2441 * sock.
2442 */
2443 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2444 if (buff == NULL) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002445 inet_csk_schedule_ack(sk);
2446 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002447 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2448 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 return;
2450 }
2451
2452 /* Reserve space for headers and prepare control bits. */
2453 skb_reserve(buff, MAX_TCP_HEADER);
2454 buff->csum = 0;
2455 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2456 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002457 skb_shinfo(buff)->gso_segs = 1;
2458 skb_shinfo(buff)->gso_size = 0;
2459 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460
2461 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002462 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 TCP_SKB_CB(buff)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002464 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 }
2466}
2467
2468/* This routine sends a packet with an out of date sequence
2469 * number. It assumes the other end will try to ack it.
2470 *
2471 * Question: what should we make while urgent mode?
2472 * 4.4BSD forces sending single byte of data. We cannot send
2473 * out of window data, because we have SND.NXT==SND.MAX...
2474 *
2475 * Current solution: to send TWO zero-length segments in urgent mode:
2476 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2477 * out-of-date with SND.UNA-1 to probe window.
2478 */
2479static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2480{
2481 struct tcp_sock *tp = tcp_sk(sk);
2482 struct sk_buff *skb;
2483
2484 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2485 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002486 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 return -1;
2488
2489 /* Reserve space for headers and set control bits. */
2490 skb_reserve(skb, MAX_TCP_HEADER);
2491 skb->csum = 0;
2492 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2493 TCP_SKB_CB(skb)->sacked = urgent;
Herbert Xu79671682006-06-22 02:40:14 -07002494 skb_shinfo(skb)->gso_segs = 1;
2495 skb_shinfo(skb)->gso_size = 0;
2496 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497
2498 /* Use a previous sequence. This should cause the other
2499 * end to send an ack. Don't queue or clone SKB, just
2500 * send it.
2501 */
2502 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2503 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2504 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002505 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506}
2507
2508int tcp_write_wakeup(struct sock *sk)
2509{
2510 if (sk->sk_state != TCP_CLOSE) {
2511 struct tcp_sock *tp = tcp_sk(sk);
2512 struct sk_buff *skb;
2513
David S. Millerfe067e82007-03-07 12:12:44 -08002514 if ((skb = tcp_send_head(sk)) != NULL &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2516 int err;
2517 unsigned int mss = tcp_current_mss(sk, 0);
2518 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2519
2520 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2521 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2522
2523 /* We are probing the opening of a window
2524 * but the window size is != 0
2525 * must have been a result SWS avoidance ( sender )
2526 */
2527 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2528 skb->len > mss) {
2529 seg_size = min(seg_size, mss);
2530 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
David S. Miller846998a2005-08-04 19:52:01 -07002531 if (tcp_fragment(sk, skb, seg_size, mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 } else if (!tcp_skb_pcount(skb))
David S. Miller846998a2005-08-04 19:52:01 -07002534 tcp_set_skb_tso_segs(sk, skb, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535
2536 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2537 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002538 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 if (!err) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002540 update_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 }
2542 return err;
2543 } else {
2544 if (tp->urg_mode &&
2545 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2546 tcp_xmit_probe_skb(sk, TCPCB_URG);
2547 return tcp_xmit_probe_skb(sk, 0);
2548 }
2549 }
2550 return -1;
2551}
2552
2553/* A window probe timeout has occurred. If window is not closed send
2554 * a partial packet else a zero probe.
2555 */
2556void tcp_send_probe0(struct sock *sk)
2557{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002558 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 struct tcp_sock *tp = tcp_sk(sk);
2560 int err;
2561
2562 err = tcp_write_wakeup(sk);
2563
David S. Millerfe067e82007-03-07 12:12:44 -08002564 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002566 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002567 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568 return;
2569 }
2570
2571 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002572 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2573 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002574 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002575 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002576 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2577 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 } else {
2579 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002580 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581 * Let local senders to fight for local resources.
2582 *
2583 * Use accumulated backoff yet.
2584 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002585 if (!icsk->icsk_probes_out)
2586 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002587 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002588 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002589 TCP_RESOURCE_PROBE_INTERVAL),
2590 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 }
2592}
2593
2594EXPORT_SYMBOL(tcp_connect);
2595EXPORT_SYMBOL(tcp_make_synack);
2596EXPORT_SYMBOL(tcp_simple_retransmit);
2597EXPORT_SYMBOL(tcp_sync_mss);
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -08002598EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
John Heffner5d424d52006-03-20 17:53:41 -08002599EXPORT_SYMBOL(tcp_mtup_init);