blob: bb7e80a284e7e046ac5fe52d939d93e417718443 [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 Järvinen66f5fe62007-12-31 04:43:57 -080064static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070065{
66 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080067 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070068
David S. Millerfe067e82007-03-07 12:12:44 -080069 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080071
72 /* Don't override Nagle indefinately with F-RTO */
73 if (tp->frto_counter == 2)
74 tp->frto_counter = 3;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080075
76 tp->packets_out += tcp_skb_pcount(skb);
77 if (!prior_packets)
78 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
79 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -070080}
81
82/* SND.NXT, if window was not shrunk.
83 * If window has been shrunk, what should we make? It is not clear at all.
84 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
85 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
86 * invalid. OK, let's make this for now:
87 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070088static inline __u32 tcp_acceptable_seq(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070089{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070090 struct tcp_sock *tp = tcp_sk(sk);
91
Ilpo Järvinen90840de2007-12-31 04:48:41 -080092 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -070093 return tp->snd_nxt;
94 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -080095 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -070096}
97
98/* Calculate mss to advertise in SYN segment.
99 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
100 *
101 * 1. It is independent of path mtu.
102 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
103 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
104 * attached devices, because some buggy hosts are confused by
105 * large MSS.
106 * 4. We do not make 3, we advertise MSS, calculated from first
107 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
108 * This may be overridden via information stored in routing table.
109 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
110 * probably even Jumbo".
111 */
112static __u16 tcp_advertise_mss(struct sock *sk)
113{
114 struct tcp_sock *tp = tcp_sk(sk);
115 struct dst_entry *dst = __sk_dst_get(sk);
116 int mss = tp->advmss;
117
118 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
119 mss = dst_metric(dst, RTAX_ADVMSS);
120 tp->advmss = mss;
121 }
122
123 return (__u16)mss;
124}
125
126/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
127 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700128static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700130 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 s32 delta = tcp_time_stamp - tp->lsndtime;
132 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
133 u32 cwnd = tp->snd_cwnd;
134
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300135 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300137 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 restart_cwnd = min(restart_cwnd, cwnd);
139
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700140 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 cwnd >>= 1;
142 tp->snd_cwnd = max(cwnd, restart_cwnd);
143 tp->snd_cwnd_stamp = tcp_time_stamp;
144 tp->snd_cwnd_used = 0;
145}
146
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800147static void tcp_event_data_sent(struct tcp_sock *tp,
148 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700150 struct inet_connection_sock *icsk = inet_csk(sk);
151 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
David S. Miller35089bb2006-06-13 22:33:04 -0700153 if (sysctl_tcp_slow_start_after_idle &&
154 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700155 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
157 tp->lsndtime = now;
158
159 /* If it is a reply for ato after last received
160 * packet, enter pingpong mode.
161 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700162 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
163 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164}
165
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800166static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700168 tcp_dec_quickack_mode(sk, pkts);
169 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170}
171
172/* Determine a window scaling and initial window to offer.
173 * Based on the assumption that the given amount of space
174 * will be offered. Store the results in the tp structure.
175 * NOTE: for smooth operation initial space offering should
176 * be a multiple of mss if possible. We assume here that mss >= 1.
177 * This MUST be enforced by all callers.
178 */
179void tcp_select_initial_window(int __space, __u32 mss,
180 __u32 *rcv_wnd, __u32 *window_clamp,
181 int wscale_ok, __u8 *rcv_wscale)
182{
183 unsigned int space = (__space < 0 ? 0 : __space);
184
185 /* If no clamp set the clamp to the max possible scaled window */
186 if (*window_clamp == 0)
187 (*window_clamp) = (65535 << 14);
188 space = min(*window_clamp, space);
189
190 /* Quantize space offering to a multiple of mss if possible. */
191 if (space > mss)
192 space = (space / mss) * mss;
193
194 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800195 * will break some buggy TCP stacks. If the admin tells us
196 * it is likely we could be speaking with such a buggy stack
197 * we will truncate our initial window offering to 32K-1
198 * unless the remote has sent us a window scaling option,
199 * which we interpret as a sign the remote TCP is not
200 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201 */
Rick Jones15d99e02006-03-20 22:40:29 -0800202 if (sysctl_tcp_workaround_signed_windows)
203 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
204 else
205 (*rcv_wnd) = space;
206
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207 (*rcv_wscale) = 0;
208 if (wscale_ok) {
209 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900210 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 */
212 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700213 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 while (space > 65535 && (*rcv_wscale) < 14) {
215 space >>= 1;
216 (*rcv_wscale)++;
217 }
218 }
219
220 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700221 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 * will be satisfied with 2.
223 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800224 if (mss > (1 << *rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700225 int init_cwnd = 4;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800226 if (mss > 1460 * 3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700228 else if (mss > 1460)
229 init_cwnd = 3;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800230 if (*rcv_wnd > init_cwnd * mss)
231 *rcv_wnd = init_cwnd * mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 }
233
234 /* Set the clamp no higher than max representable value */
235 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
236}
237
238/* Chose a new window to advertise, update state in tcp_sock for the
239 * socket, and return result with RFC1323 scaling applied. The return
240 * value can be stuffed directly into th->window for an outgoing
241 * frame.
242 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800243static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244{
245 struct tcp_sock *tp = tcp_sk(sk);
246 u32 cur_win = tcp_receive_window(tp);
247 u32 new_win = __tcp_select_window(sk);
248
249 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800250 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 /* Danger Will Robinson!
252 * Don't update rcv_wup/rcv_wnd here or else
253 * we will not be able to advertise a zero
254 * window in time. --DaveM
255 *
256 * Relax Will Robinson.
257 */
258 new_win = cur_win;
259 }
260 tp->rcv_wnd = new_win;
261 tp->rcv_wup = tp->rcv_nxt;
262
263 /* Make sure we do not exceed the maximum possible
264 * scaled window.
265 */
Rick Jones15d99e02006-03-20 22:40:29 -0800266 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 new_win = min(new_win, MAX_TCP_WINDOW);
268 else
269 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
270
271 /* RFC1323 scaling applied */
272 new_win >>= tp->rx_opt.rcv_wscale;
273
274 /* If we advertise zero window, disable fast path. */
275 if (new_win == 0)
276 tp->pred_flags = 0;
277
278 return new_win;
279}
280
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800281static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700282{
283 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800284 if (!(tp->ecn_flags & TCP_ECN_OK))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700285 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
286}
287
288static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
289{
290 struct tcp_sock *tp = tcp_sk(sk);
291
292 tp->ecn_flags = 0;
293 if (sysctl_tcp_ecn) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800294 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700295 tp->ecn_flags = TCP_ECN_OK;
296 }
297}
298
299static __inline__ void
300TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
301{
302 if (inet_rsk(req)->ecn_ok)
303 th->ece = 1;
304}
305
306static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
307 int tcp_header_len)
308{
309 struct tcp_sock *tp = tcp_sk(sk);
310
311 if (tp->ecn_flags & TCP_ECN_OK) {
312 /* Not-retransmitted data segment: set ECT and inject CWR. */
313 if (skb->len != tcp_header_len &&
314 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
315 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800316 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700317 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
318 tcp_hdr(skb)->cwr = 1;
319 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
320 }
321 } else {
322 /* ACK or retransmitted segment: clear ECT|CE */
323 INET_ECN_dontxmit(sk);
324 }
325 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
326 tcp_hdr(skb)->ece = 1;
327 }
328}
329
Al Virodf7a3b02006-09-27 18:38:52 -0700330static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800331 __u32 tstamp, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800332{
333 if (tp->rx_opt.tstamp_ok) {
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700334 *ptr++ = htonl((TCPOPT_NOP << 24) |
335 (TCPOPT_NOP << 16) |
336 (TCPOPT_TIMESTAMP << 8) |
337 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800338 *ptr++ = htonl(tstamp);
339 *ptr++ = htonl(tp->rx_opt.ts_recent);
340 }
341 if (tp->rx_opt.eff_sacks) {
342 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
343 int this_sack;
344
345 *ptr++ = htonl((TCPOPT_NOP << 24) |
346 (TCPOPT_NOP << 16) |
347 (TCPOPT_SACK << 8) |
348 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
349 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800350
351 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800352 *ptr++ = htonl(sp[this_sack].start_seq);
353 *ptr++ = htonl(sp[this_sack].end_seq);
354 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800355
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800356 if (tp->rx_opt.dsack) {
357 tp->rx_opt.dsack = 0;
358 tp->rx_opt.eff_sacks--;
359 }
360 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800361#ifdef CONFIG_TCP_MD5SIG
362 if (md5_hash) {
363 *ptr++ = htonl((TCPOPT_NOP << 24) |
364 (TCPOPT_NOP << 16) |
365 (TCPOPT_MD5SIG << 8) |
366 TCPOLEN_MD5SIG);
367 *md5_hash = (__u8 *)ptr;
368 }
369#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800370}
371
372/* Construct a tcp options header for a SYN or SYN_ACK packet.
373 * If this is every changed make sure to change the definition of
374 * MAX_SYN_SIZE to match the new maximum number of options that you
375 * can generate.
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800376 *
377 * Note - that with the RFC2385 TCP option, we make room for the
378 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
379 * location to be filled is passed back up.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800380 */
Al Virodf7a3b02006-09-27 18:38:52 -0700381static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800382 int offer_wscale, int wscale, __u32 tstamp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800383 __u32 ts_recent, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800384{
385 /* We always get an MSS option.
386 * The option bytes which will be seen in normal data
387 * packets should timestamps be used, must be in the MSS
388 * advertised. But we subtract them from tp->mss_cache so
389 * that calculations in tcp_sendmsg are simpler etc.
390 * So account for this fact here if necessary. If we
391 * don't do this correctly, as a receiver we won't
392 * recognize data packets as being full sized when we
393 * should, and thus we won't abide by the delayed ACK
394 * rules correctly.
395 * SACKs don't matter, we never delay an ACK when we
396 * have any of those going out.
397 */
398 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
399 if (ts) {
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800400 if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700401 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
402 (TCPOLEN_SACK_PERM << 16) |
403 (TCPOPT_TIMESTAMP << 8) |
404 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800405 else
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700406 *ptr++ = htonl((TCPOPT_NOP << 24) |
407 (TCPOPT_NOP << 16) |
408 (TCPOPT_TIMESTAMP << 8) |
409 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800410 *ptr++ = htonl(tstamp); /* TSVAL */
411 *ptr++ = htonl(ts_recent); /* TSECR */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800412 } else if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700413 *ptr++ = htonl((TCPOPT_NOP << 24) |
414 (TCPOPT_NOP << 16) |
415 (TCPOPT_SACK_PERM << 8) |
416 TCPOLEN_SACK_PERM);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800417 if (offer_wscale)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700418 *ptr++ = htonl((TCPOPT_NOP << 24) |
419 (TCPOPT_WINDOW << 16) |
420 (TCPOLEN_WINDOW << 8) |
421 (wscale));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800422#ifdef CONFIG_TCP_MD5SIG
423 /*
424 * If MD5 is enabled, then we set the option, and include the size
425 * (always 18). The actual MD5 hash is added just before the
426 * packet is sent.
427 */
428 if (md5_hash) {
429 *ptr++ = htonl((TCPOPT_NOP << 24) |
430 (TCPOPT_NOP << 16) |
431 (TCPOPT_MD5SIG << 8) |
432 TCPOLEN_MD5SIG);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800433 *md5_hash = (__u8 *)ptr;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800434 }
435#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800436}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437
438/* This routine actually transmits TCP packets queued in by
439 * tcp_do_sendmsg(). This is used by both the initial
440 * transmission and possible later retransmissions.
441 * All SKB's seen here are completely headerless. It is our
442 * job to build the TCP header, and pass the packet down to
443 * IP so it can do the same plus pass the packet off to the
444 * device.
445 *
446 * We are working here with either a clone of the original
447 * SKB, or a fresh unique copy made by the retransmit engine.
448 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800449static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
450 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800452 const struct inet_connection_sock *icsk = inet_csk(sk);
453 struct inet_sock *inet;
454 struct tcp_sock *tp;
455 struct tcp_skb_cb *tcb;
456 int tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800457#ifdef CONFIG_TCP_MD5SIG
458 struct tcp_md5sig_key *md5;
459 __u8 *md5_hash_location;
460#endif
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800461 struct tcphdr *th;
462 int sysctl_flags;
463 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800465 BUG_ON(!skb || !tcp_skb_pcount(skb));
466
467 /* If congestion control is doing timestamping, we must
468 * take such a timestamp before we potentially clone/copy.
469 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700470 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800471 __net_timestamp(skb);
472
473 if (likely(clone_it)) {
474 if (unlikely(skb_cloned(skb)))
475 skb = pskb_copy(skb, gfp_mask);
476 else
477 skb = skb_clone(skb, gfp_mask);
478 if (unlikely(!skb))
479 return -ENOBUFS;
480 }
481
482 inet = inet_sk(sk);
483 tp = tcp_sk(sk);
484 tcb = TCP_SKB_CB(skb);
485 tcp_header_size = tp->tcp_header_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
487#define SYSCTL_FLAG_TSTAMPS 0x1
488#define SYSCTL_FLAG_WSCALE 0x2
489#define SYSCTL_FLAG_SACK 0x4
490
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800491 sysctl_flags = 0;
492 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
493 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800494 if (sysctl_tcp_timestamps) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800495 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
496 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800498 if (sysctl_tcp_window_scaling) {
499 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
500 sysctl_flags |= SYSCTL_FLAG_WSCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800502 if (sysctl_tcp_sack) {
503 sysctl_flags |= SYSCTL_FLAG_SACK;
504 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
505 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800507 } else if (unlikely(tp->rx_opt.eff_sacks)) {
508 /* A SACK is 2 pad bytes, a 2 byte header, plus
509 * 2 32-bit sequence numbers for each SACK block.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800511 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
512 (tp->rx_opt.eff_sacks *
513 TCPOLEN_SACK_PERBLOCK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900515
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800516 if (tcp_packets_in_flight(tp) == 0)
517 tcp_ca_event(sk, CA_EVENT_TX_START);
518
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800519#ifdef CONFIG_TCP_MD5SIG
520 /*
521 * Are we doing MD5 on this segment? If so - make
522 * room for it.
523 */
524 md5 = tp->af_specific->md5_lookup(sk, sk);
525 if (md5)
526 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
527#endif
528
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700529 skb_push(skb, tcp_header_size);
530 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800531 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800532
533 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700534 th = tcp_hdr(skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800535 th->source = inet->sport;
536 th->dest = inet->dport;
537 th->seq = htonl(tcb->seq);
538 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700539 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800540 tcb->flags);
541
542 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
543 /* RFC1323: The window in SYN & SYN/ACK segments
544 * is never scaled.
545 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800546 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800547 } else {
548 th->window = htons(tcp_select_window(sk));
549 }
550 th->check = 0;
551 th->urg_ptr = 0;
552
553 if (unlikely(tp->urg_mode &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800554 between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
555 th->urg_ptr = htons(tp->snd_up - tcb->seq);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800556 th->urg = 1;
557 }
558
559 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
Al Virodf7a3b02006-09-27 18:38:52 -0700560 tcp_syn_build_options((__be32 *)(th + 1),
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800561 tcp_advertise_mss(sk),
562 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
563 (sysctl_flags & SYSCTL_FLAG_SACK),
564 (sysctl_flags & SYSCTL_FLAG_WSCALE),
565 tp->rx_opt.rcv_wscale,
566 tcb->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800567 tp->rx_opt.ts_recent,
568
569#ifdef CONFIG_TCP_MD5SIG
570 md5 ? &md5_hash_location :
571#endif
572 NULL);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800573 } else {
Al Virodf7a3b02006-09-27 18:38:52 -0700574 tcp_build_and_update_options((__be32 *)(th + 1),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800575 tp, tcb->when,
576#ifdef CONFIG_TCP_MD5SIG
577 md5 ? &md5_hash_location :
578#endif
579 NULL);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700580 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800581 }
582
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800583#ifdef CONFIG_TCP_MD5SIG
584 /* Calculate the MD5 hash, as we have all we need now */
585 if (md5) {
586 tp->af_specific->calc_md5_hash(md5_hash_location,
587 md5,
588 sk, NULL, NULL,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700589 tcp_hdr(skb),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800590 sk->sk_protocol,
591 skb->len);
592 }
593#endif
594
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800595 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800596
597 if (likely(tcb->flags & TCPCB_FLAG_ACK))
598 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
599
600 if (skb->len != tcp_header_size)
601 tcp_event_data_sent(tp, skb, sk);
602
Wei Yongjunbd37a082006-08-07 21:04:15 -0700603 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
604 TCP_INC_STATS(TCP_MIB_OUTSEGS);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800605
David S. Millere89862f2007-01-26 01:04:55 -0800606 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700607 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800608 return err;
609
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800610 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800611
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200612 return net_xmit_eval(err);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800613
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614#undef SYSCTL_FLAG_TSTAMPS
615#undef SYSCTL_FLAG_WSCALE
616#undef SYSCTL_FLAG_SACK
617}
618
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900619/* This routine just queue's the buffer
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 *
621 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
622 * otherwise socket can stall.
623 */
624static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
625{
626 struct tcp_sock *tp = tcp_sk(sk);
627
628 /* Advance write_seq and place onto the write_queue. */
629 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
630 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800631 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800632 sk->sk_wmem_queued += skb->truesize;
633 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634}
635
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800636static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
637 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700638{
Herbert Xubcd76112006-06-30 13:36:35 -0700639 if (skb->len <= mss_now || !sk_can_gso(sk)) {
David S. Millerf6302d12005-07-05 15:18:03 -0700640 /* Avoid the costly divide in the normal
641 * non-TSO case.
642 */
Herbert Xu79671682006-06-22 02:40:14 -0700643 skb_shinfo(skb)->gso_segs = 1;
644 skb_shinfo(skb)->gso_size = 0;
645 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700646 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700647 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700648 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700649 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 }
651}
652
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700653/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -0800654 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700655 */
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800656static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700657 int decr)
658{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800659 struct tcp_sock *tp = tcp_sk(sk);
660
Ilpo Järvinendc869672007-10-01 15:27:19 -0700661 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700662 return;
663
Ilpo Järvinen6859d492007-12-02 00:48:06 +0200664 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700665 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700666}
667
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668/* Function to create two new TCP segments. Shrinks the given segment
669 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900670 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 * Remember, these are still headerless SKBs at this point.
672 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800673int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
674 unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675{
676 struct tcp_sock *tp = tcp_sk(sk);
677 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700678 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700679 int nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 u16 flags;
681
Herbert Xub2cc99f02005-10-20 17:13:13 -0200682 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800683
Ilpo Järvinenb7689202007-09-20 11:37:19 -0700684 tcp_clear_retrans_hints_partial(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 nsize = skb_headlen(skb) - len;
686 if (nsize < 0)
687 nsize = 0;
688
689 if (skb_cloned(skb) &&
690 skb_is_nonlinear(skb) &&
691 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
692 return -ENOMEM;
693
694 /* Get a new skb... force flag on. */
695 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
696 if (buff == NULL)
697 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -0700698
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800699 sk->sk_wmem_queued += buff->truesize;
700 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -0700701 nlen = skb->len - len - nsize;
702 buff->truesize += nlen;
703 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704
705 /* Correct the sequence numbers. */
706 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
707 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
708 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
709
710 /* PSH and FIN should only be set in the second packet. */
711 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800712 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700714 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
Patrick McHardy84fa7932006-08-29 16:44:56 -0700716 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800718 buff->csum = csum_partial_copy_nocheck(skb->data + len,
719 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 nsize, 0);
721
722 skb_trim(skb, len);
723
724 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
725 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -0700726 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 skb_split(skb, buff, len);
728 }
729
730 buff->ip_summed = skb->ip_summed;
731
732 /* Looks stupid, but our code really uses when of
733 * skbs, which it never sent before. --ANK
734 */
735 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700736 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737
David S. Miller6475be12005-09-01 22:47:01 -0700738 old_factor = tcp_skb_pcount(skb);
739
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700741 tcp_set_skb_tso_segs(sk, skb, mss_now);
742 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743
David S. Miller6475be12005-09-01 22:47:01 -0700744 /* If this packet has been sent out already, we must
745 * adjust the various packet counters.
746 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700747 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700748 int diff = old_factor - tcp_skb_pcount(skb) -
749 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750
David S. Miller6475be12005-09-01 22:47:01 -0700751 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700752
753 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
754 tp->sacked_out -= diff;
755 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
756 tp->retrans_out -= diff;
757
Ilpo Järvinenb5860bb2007-08-09 14:33:18 +0300758 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
David S. Miller6475be12005-09-01 22:47:01 -0700759 tp->lost_out -= diff;
Herbert Xu83ca28b2005-09-22 23:32:56 -0700760
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700761 /* Adjust Reno SACK estimate. */
762 if (tcp_is_reno(tp) && diff > 0) {
763 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
764 tcp_verify_left_out(tp);
David S. Miller6475be12005-09-01 22:47:01 -0700765 }
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800766 tcp_adjust_fackets_out(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 }
768
769 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700770 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -0800771 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772
773 return 0;
774}
775
776/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
777 * eventually). The difference is that pulled data not copied, but
778 * immediately discarded.
779 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700780static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781{
782 int i, k, eat;
783
784 eat = len;
785 k = 0;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800786 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 if (skb_shinfo(skb)->frags[i].size <= eat) {
788 put_page(skb_shinfo(skb)->frags[i].page);
789 eat -= skb_shinfo(skb)->frags[i].size;
790 } else {
791 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
792 if (eat) {
793 skb_shinfo(skb)->frags[k].page_offset += eat;
794 skb_shinfo(skb)->frags[k].size -= eat;
795 eat = 0;
796 }
797 k++;
798 }
799 }
800 skb_shinfo(skb)->nr_frags = k;
801
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -0700802 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 skb->data_len -= len;
804 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805}
806
807int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
808{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800809 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 return -ENOMEM;
811
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700812 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
813 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700815 else
816 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
818 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -0700819 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820
821 skb->truesize -= len;
822 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800823 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
825
826 /* Any change of skb->len requires recalculation of tso
827 * factor and mss.
828 */
829 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700830 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
832 return 0;
833}
834
John Heffner5d424d52006-03-20 17:53:41 -0800835/* Not accounting for SACKs here. */
836int tcp_mtu_to_mss(struct sock *sk, int pmtu)
837{
838 struct tcp_sock *tp = tcp_sk(sk);
839 struct inet_connection_sock *icsk = inet_csk(sk);
840 int mss_now;
841
842 /* Calculate base mss without TCP options:
843 It is MMS_S - sizeof(tcphdr) of rfc1122
844 */
845 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
846
847 /* Clamp it (mss_clamp does not include tcp options) */
848 if (mss_now > tp->rx_opt.mss_clamp)
849 mss_now = tp->rx_opt.mss_clamp;
850
851 /* Now subtract optional transport overhead */
852 mss_now -= icsk->icsk_ext_hdr_len;
853
854 /* Then reserve room for full set of TCP options and 8 bytes of data */
855 if (mss_now < 48)
856 mss_now = 48;
857
858 /* Now subtract TCP options size, not including SACKs */
859 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
860
861 return mss_now;
862}
863
864/* Inverse of above */
865int tcp_mss_to_mtu(struct sock *sk, int mss)
866{
867 struct tcp_sock *tp = tcp_sk(sk);
868 struct inet_connection_sock *icsk = inet_csk(sk);
869 int mtu;
870
871 mtu = mss +
872 tp->tcp_header_len +
873 icsk->icsk_ext_hdr_len +
874 icsk->icsk_af_ops->net_header_len;
875
876 return mtu;
877}
878
879void tcp_mtup_init(struct sock *sk)
880{
881 struct tcp_sock *tp = tcp_sk(sk);
882 struct inet_connection_sock *icsk = inet_csk(sk);
883
884 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
885 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900886 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -0800887 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
888 icsk->icsk_mtup.probe_size = 0;
889}
890
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891/* This function synchronize snd mss to current pmtu/exthdr set.
892
893 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
894 for TCP options, but includes only bare TCP header.
895
896 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800897 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 It also does not include TCP options.
899
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800900 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901
902 tp->mss_cache is current effective sending mss, including
903 all tcp options except for SACKs. It is evaluated,
904 taking into account current pmtu, but never exceeds
905 tp->rx_opt.mss_clamp.
906
907 NOTE1. rfc1122 clearly states that advertised MSS
908 DOES NOT include either tcp or ip options.
909
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800910 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
911 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
914{
915 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800916 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -0800917 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
John Heffner5d424d52006-03-20 17:53:41 -0800919 if (icsk->icsk_mtup.search_high > pmtu)
920 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
John Heffner5d424d52006-03-20 17:53:41 -0800922 mss_now = tcp_mtu_to_mss(sk, pmtu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923
924 /* Bound mss with half of window */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800925 if (tp->max_window && mss_now > (tp->max_window >> 1))
926 mss_now = max((tp->max_window >> 1), 68U - tp->tcp_header_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927
928 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800929 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -0800930 if (icsk->icsk_mtup.enabled)
931 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700932 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933
934 return mss_now;
935}
936
937/* Compute the current effective MSS, taking SACKs and IP options,
938 * and even PMTU discovery events into account.
939 *
940 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
941 * cannot be large. However, taking into account rare use of URG, this
942 * is not a big flaw.
943 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700944unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945{
946 struct tcp_sock *tp = tcp_sk(sk);
947 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700948 u32 mss_now;
949 u16 xmit_size_goal;
950 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700952 mss_now = tp->mss_cache;
953
Herbert Xubcd76112006-06-30 13:36:35 -0700954 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700955 doing_tso = 1;
956
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 if (dst) {
958 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800959 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 mss_now = tcp_sync_mss(sk, mtu);
961 }
962
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 if (tp->rx_opt.eff_sacks)
964 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
965 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700966
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800967#ifdef CONFIG_TCP_MD5SIG
968 if (tp->af_specific->md5_lookup(sk, sk))
969 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
970#endif
971
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700972 xmit_size_goal = mss_now;
973
974 if (doing_tso) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800975 xmit_size_goal = (65535 -
976 inet_csk(sk)->icsk_af_ops->net_header_len -
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800977 inet_csk(sk)->icsk_ext_hdr_len -
978 tp->tcp_header_len);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700979
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800980 if (tp->max_window && (xmit_size_goal > (tp->max_window >> 1)))
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700981 xmit_size_goal = max((tp->max_window >> 1),
982 68U - tp->tcp_header_len);
983
984 xmit_size_goal -= (xmit_size_goal % mss_now);
985 }
986 tp->xmit_size_goal = xmit_size_goal;
987
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 return mss_now;
989}
990
David S. Millera762a982005-07-05 15:18:51 -0700991/* Congestion window validation. (RFC2861) */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700992static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -0700993{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700994 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -0700995 __u32 packets_out = tp->packets_out;
996
997 if (packets_out >= tp->snd_cwnd) {
998 /* Network is feed fully. */
999 tp->snd_cwnd_used = 0;
1000 tp->snd_cwnd_stamp = tcp_time_stamp;
1001 } else {
1002 /* Network starves. */
1003 if (tp->packets_out > tp->snd_cwnd_used)
1004 tp->snd_cwnd_used = tp->packets_out;
1005
David S. Miller15d33c02007-04-09 13:23:14 -07001006 if (sysctl_tcp_slow_start_after_idle &&
1007 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001008 tcp_cwnd_application_limited(sk);
1009 }
1010}
1011
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001012/* Returns the portion of skb which can be sent right away without
1013 * introducing MSS oddities to segment boundaries. In rare cases where
1014 * mss_now != mss_cache, we will request caller to create a small skb
1015 * per input skb which could be mostly avoided here (if desired).
1016 */
1017static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001018 unsigned int mss_now, unsigned int cwnd)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001019{
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001020 struct tcp_sock *tp = tcp_sk(sk);
1021 u32 needed, window, cwnd_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001022
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001023 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001024 cwnd_len = mss_now * cwnd;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001025
1026 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1027 return cwnd_len;
1028
1029 if (skb == tcp_write_queue_tail(sk) && cwnd_len <= skb->len)
1030 return cwnd_len;
1031
1032 needed = min(skb->len, window);
1033 return needed - needed % mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001034}
1035
1036/* Can at least one segment of SKB be sent right now, according to the
1037 * congestion window rules? If so, return how many segments are allowed.
1038 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001039static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1040 struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001041{
1042 u32 in_flight, cwnd;
1043
1044 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -08001045 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1046 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001047 return 1;
1048
1049 in_flight = tcp_packets_in_flight(tp);
1050 cwnd = tp->snd_cwnd;
1051 if (in_flight < cwnd)
1052 return (cwnd - in_flight);
1053
1054 return 0;
1055}
1056
1057/* This must be invoked the first time we consider transmitting
1058 * SKB onto the wire.
1059 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001060static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1061 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001062{
1063 int tso_segs = tcp_skb_pcount(skb);
1064
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001065 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001066 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001067 tso_segs = tcp_skb_pcount(skb);
1068 }
1069 return tso_segs;
1070}
1071
1072static inline int tcp_minshall_check(const struct tcp_sock *tp)
1073{
1074 return after(tp->snd_sml,tp->snd_una) &&
1075 !after(tp->snd_sml, tp->snd_nxt);
1076}
1077
1078/* Return 0, if packet can be sent now without violation Nagle's rules:
1079 * 1. It is full sized.
1080 * 2. Or it contains FIN. (already checked by caller)
1081 * 3. Or TCP_NODELAY was set.
1082 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1083 * With Minshall's modification: all sent small packets are ACKed.
1084 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001085static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001086 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001087 unsigned mss_now, int nonagle)
1088{
1089 return (skb->len < mss_now &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001090 ((nonagle & TCP_NAGLE_CORK) ||
1091 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001092}
1093
1094/* Return non-zero if the Nagle test allows this packet to be
1095 * sent now.
1096 */
1097static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1098 unsigned int cur_mss, int nonagle)
1099{
1100 /* Nagle rule does not apply to frames, which sit in the middle of the
1101 * write_queue (they have no chances to get new data).
1102 *
1103 * This is implemented in the callers, where they modify the 'nonagle'
1104 * argument based upon the location of SKB in the send queue.
1105 */
1106 if (nonagle & TCP_NAGLE_PUSH)
1107 return 1;
1108
Ilpo Järvinend551e452007-04-30 00:42:20 -07001109 /* Don't use the nagle rule for urgent data (or for the final FIN).
1110 * Nagle can be ignored during F-RTO too (see RFC4138).
1111 */
1112 if (tp->urg_mode || (tp->frto_counter == 2) ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001113 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1114 return 1;
1115
1116 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1117 return 1;
1118
1119 return 0;
1120}
1121
1122/* Does at least the first segment of SKB fit into the send window? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001123static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1124 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001125{
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
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001131 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001132}
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);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001150 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001151 cwnd_quota = 0;
1152
1153 return cwnd_quota;
1154}
1155
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001156int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001157{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001158 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001159 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001160
1161 return (skb &&
1162 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1163 (tcp_skb_is_last(sk, skb) ?
Ilpo Järvinen4e67d872007-12-05 02:25:32 -08001164 tp->nonagle : TCP_NAGLE_PUSH)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001165}
1166
1167/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1168 * which is put after SKB on the list. It is very much like
1169 * tcp_fragment() except that it may make several kinds of assumptions
1170 * in order to speed up the splitting operation. In particular, we
1171 * know that all the data is in scatter-gather pages, and that the
1172 * packet has never been sent out before (and thus is not cloned).
1173 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001174static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1175 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
Pavel Emelyanovdf97c702007-11-29 21:22:33 +11001185 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001186 if (unlikely(buff == NULL))
1187 return -ENOMEM;
1188
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001189 sk->sk_wmem_queued += buff->truesize;
1190 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001191 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001192 skb->truesize -= nlen;
1193
1194 /* Correct the sequence numbers. */
1195 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1196 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1197 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1198
1199 /* PSH and FIN should only be set in the second packet. */
1200 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001201 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001202 TCP_SKB_CB(buff)->flags = flags;
1203
1204 /* This packet was never sent out yet, so no SACK bits. */
1205 TCP_SKB_CB(buff)->sacked = 0;
1206
Patrick McHardy84fa7932006-08-29 16:44:56 -07001207 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001208 skb_split(skb, buff, len);
1209
1210 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001211 tcp_set_skb_tso_segs(sk, skb, mss_now);
1212 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001213
1214 /* Link BUFF into the send queue. */
1215 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001216 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001217
1218 return 0;
1219}
1220
1221/* Try to defer sending, if possible, in order to minimize the amount
1222 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1223 *
1224 * This algorithm is from John Heffner.
1225 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001226static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001227{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001228 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001229 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001230 u32 send_win, cong_win, limit, in_flight;
1231
1232 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001233 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001234
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001235 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001236 goto send_now;
1237
1238 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001239 if (tp->tso_deferred &&
1240 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001241 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001242
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001243 in_flight = tcp_packets_in_flight(tp);
1244
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001245 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001246
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001247 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001248
1249 /* From in_flight test above, we know that cwnd > in_flight. */
1250 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1251
1252 limit = min(send_win, cong_win);
1253
David S. Millerba244fe2006-03-11 18:51:49 -08001254 /* If a full-sized TSO skb can be sent, do it. */
1255 if (limit >= 65536)
John Heffnerae8064a2006-10-18 20:36:48 -07001256 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001257
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001258 if (sysctl_tcp_tso_win_divisor) {
1259 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1260
1261 /* If at least some fraction of a window is available,
1262 * just use it.
1263 */
1264 chunk /= sysctl_tcp_tso_win_divisor;
1265 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001266 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001267 } else {
1268 /* Different approach, try not to defer past a single
1269 * ACK. Receiver should ACK every other full sized
1270 * frame, so if we have space for more than 3 frames
1271 * then send now.
1272 */
1273 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001274 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001275 }
1276
1277 /* Ok, it looks like it is advisable to defer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001278 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001279
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001280 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001281
1282send_now:
1283 tp->tso_deferred = 0;
1284 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001285}
1286
John Heffner5d424d52006-03-20 17:53:41 -08001287/* Create a new MTU probe if we are ready.
1288 * Returns 0 if we should wait to probe (no cwnd available),
1289 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001290 * -1 otherwise
1291 */
John Heffner5d424d52006-03-20 17:53:41 -08001292static int tcp_mtu_probe(struct sock *sk)
1293{
1294 struct tcp_sock *tp = tcp_sk(sk);
1295 struct inet_connection_sock *icsk = inet_csk(sk);
1296 struct sk_buff *skb, *nskb, *next;
1297 int len;
1298 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001299 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001300 int copy;
1301 int mss_now;
1302
1303 /* Not currently probing/verifying,
1304 * not in recovery,
1305 * have enough cwnd, and
1306 * not SACKing (the variable headers throw things off) */
1307 if (!icsk->icsk_mtup.enabled ||
1308 icsk->icsk_mtup.probe_size ||
1309 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1310 tp->snd_cwnd < 11 ||
1311 tp->rx_opt.eff_sacks)
1312 return -1;
1313
1314 /* Very simple search strategy: just double the MSS. */
1315 mss_now = tcp_current_mss(sk, 0);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001316 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001317 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001318 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1319 /* TODO: set timer for probe_converge_event */
1320 return -1;
1321 }
1322
1323 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001324 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001325 return -1;
1326
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001327 if (tp->snd_wnd < size_needed)
1328 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001329 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001330 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001331
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001332 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1333 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1334 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001335 return -1;
1336 else
1337 return 0;
1338 }
1339
1340 /* We're allowed to probe. Build it now. */
1341 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1342 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001343 sk->sk_wmem_queued += nskb->truesize;
1344 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001345
David S. Millerfe067e82007-03-07 12:12:44 -08001346 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001347
1348 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1349 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1350 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1351 TCP_SKB_CB(nskb)->sacked = 0;
1352 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001353 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001354
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001355 tcp_insert_write_queue_before(nskb, skb, sk);
1356
John Heffner5d424d52006-03-20 17:53:41 -08001357 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001358 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001359 copy = min_t(int, skb->len, probe_size - len);
1360 if (nskb->ip_summed)
1361 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1362 else
1363 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001364 skb_put(nskb, copy),
1365 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);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001372 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001373 } 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)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001379 skb->csum = csum_partial(skb->data,
1380 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001381 } else {
1382 __pskb_trim_head(skb, copy);
1383 tcp_set_skb_tso_segs(sk, skb, mss_now);
1384 }
1385 TCP_SKB_CB(skb)->seq += copy;
1386 }
1387
1388 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001389
1390 if (len >= probe_size)
1391 break;
John Heffner5d424d52006-03-20 17:53:41 -08001392 }
1393 tcp_init_tso_segs(sk, nskb, nskb->len);
1394
1395 /* We're ready to send. If this fails, the probe will
1396 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1397 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1398 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1399 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001400 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001401 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001402 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001403
1404 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001405 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1406 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001407
1408 return 1;
1409 }
1410
1411 return -1;
1412}
1413
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414/* This routine writes packets to the network. It advances the
1415 * send_head. This happens as incoming acks open up the remote
1416 * window for us.
1417 *
1418 * Returns 1, if no segments are in flight and we have queued segments, but
1419 * cannot send anything now because of SWS or another problem.
1420 */
David S. Millera2e2a592005-07-05 15:19:23 -07001421static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422{
1423 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001424 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001425 unsigned int tso_segs, sent_pkts;
1426 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001427 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428
1429 /* If we are closed, the bytes will have to remain here.
1430 * In time closedown will finish, we empty the write queue and all
1431 * will be happy.
1432 */
David S. Miller92df7b52005-07-05 15:19:06 -07001433 if (unlikely(sk->sk_state == TCP_CLOSE))
1434 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435
David S. Miller92df7b52005-07-05 15:19:06 -07001436 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001437
1438 /* Do MTU probing. */
1439 if ((result = tcp_mtu_probe(sk)) == 0) {
1440 return 0;
1441 } else if (result > 0) {
1442 sent_pkts = 1;
1443 }
1444
David S. Millerfe067e82007-03-07 12:12:44 -08001445 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001446 unsigned int limit;
1447
Herbert Xub68e9f82005-08-04 19:52:02 -07001448 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001449 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001450
Herbert Xub68e9f82005-08-04 19:52:02 -07001451 cwnd_quota = tcp_cwnd_test(tp, skb);
1452 if (!cwnd_quota)
1453 break;
1454
1455 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1456 break;
1457
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001458 if (tso_segs == 1) {
1459 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1460 (tcp_skb_is_last(sk, skb) ?
1461 nonagle : TCP_NAGLE_PUSH))))
1462 break;
1463 } else {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001464 if (tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001465 break;
1466 }
David S. Milleraa934662005-07-05 15:20:09 -07001467
Herbert Xuc8ac3772005-08-16 20:43:40 -07001468 limit = mss_now;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001469 if (tso_segs > 1)
1470 limit = tcp_mss_split_point(sk, skb, mss_now,
1471 cwnd_quota);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
Herbert Xuc8ac3772005-08-16 20:43:40 -07001473 if (skb->len > limit &&
1474 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1475 break;
1476
David S. Miller92df7b52005-07-05 15:19:06 -07001477 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001478
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001479 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
David S. Miller92df7b52005-07-05 15:19:06 -07001480 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
David S. Miller92df7b52005-07-05 15:19:06 -07001482 /* Advance the send_head. This one is sent out.
1483 * This call will increment packets_out.
1484 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001485 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001486
1487 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001488 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 }
David S. Miller92df7b52005-07-05 15:19:06 -07001490
David S. Milleraa934662005-07-05 15:20:09 -07001491 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001492 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001493 return 0;
1494 }
David S. Millerfe067e82007-03-07 12:12:44 -08001495 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496}
1497
David S. Millera762a982005-07-05 15:18:51 -07001498/* Push out any pending frames which were held back due to
1499 * TCP_CORK or attempt at coalescing tiny packets.
1500 * The socket must be locked by the caller.
1501 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001502void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1503 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001504{
David S. Millerfe067e82007-03-07 12:12:44 -08001505 struct sk_buff *skb = tcp_send_head(sk);
David S. Millera762a982005-07-05 15:18:51 -07001506
1507 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001508 if (tcp_write_xmit(sk, cur_mss, nonagle))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001509 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001510 }
1511}
1512
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001513/* Send _single_ skb sitting at the send head. This function requires
1514 * true push pending frames to setup probe timer etc.
1515 */
1516void tcp_push_one(struct sock *sk, unsigned int mss_now)
1517{
David S. Millerfe067e82007-03-07 12:12:44 -08001518 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001519 unsigned int tso_segs, cwnd_quota;
1520
1521 BUG_ON(!skb || skb->len < mss_now);
1522
David S. Miller846998a2005-08-04 19:52:01 -07001523 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001524 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1525
1526 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001527 unsigned int limit;
1528
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001529 BUG_ON(!tso_segs);
1530
Herbert Xuc8ac3772005-08-16 20:43:40 -07001531 limit = mss_now;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001532 if (tso_segs > 1)
1533 limit = tcp_mss_split_point(sk, skb, mss_now,
1534 cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001535
Herbert Xuc8ac3772005-08-16 20:43:40 -07001536 if (skb->len > limit &&
1537 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1538 return;
1539
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001540 /* Send it out now. */
1541 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1542
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001543 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001544 tcp_event_new_data_sent(sk, skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001545 tcp_cwnd_validate(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001546 return;
1547 }
1548 }
1549}
1550
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551/* This function returns the amount that we can raise the
1552 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001553 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 * 1. The window can never be shrunk once it is offered (RFC 793)
1555 * 2. We limit memory per socket
1556 *
1557 * RFC 1122:
1558 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1559 * RECV.NEXT + RCV.WIN fixed until:
1560 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1561 *
1562 * i.e. don't raise the right edge of the window until you can raise
1563 * it at least MSS bytes.
1564 *
1565 * Unfortunately, the recommended algorithm breaks header prediction,
1566 * since header prediction assumes th->window stays fixed.
1567 *
1568 * Strictly speaking, keeping th->window fixed violates the receiver
1569 * side SWS prevention criteria. The problem is that under this rule
1570 * a stream of single byte packets will cause the right side of the
1571 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001572 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 * Of course, if the sender implements sender side SWS prevention
1574 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001575 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001577 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 * If the free space is less than the 1/4 of the maximum
1579 * space available and the free space is less than 1/2 mss,
1580 * then set the window to 0.
1581 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1582 * Otherwise, just prevent the window from shrinking
1583 * and from being larger than the largest representable value.
1584 *
1585 * This prevents incremental opening of the window in the regime
1586 * where TCP is limited by the speed of the reader side taking
1587 * data out of the TCP receive queue. It does nothing about
1588 * those cases where the window is constrained on the sender side
1589 * because the pipeline is full.
1590 *
1591 * BSD also seems to "accidentally" limit itself to windows that are a
1592 * multiple of MSS, at least until the free space gets quite small.
1593 * This would appear to be a side effect of the mbuf implementation.
1594 * Combining these two algorithms results in the observed behavior
1595 * of having a fixed window size at almost all times.
1596 *
1597 * Below we obtain similar behavior by forcing the offered window to
1598 * a multiple of the mss when it is feasible to do so.
1599 *
1600 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1601 * Regular options like TIMESTAMP are taken into account.
1602 */
1603u32 __tcp_select_window(struct sock *sk)
1604{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001605 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001607 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 * here. I don't know if the value based on our guesses
1609 * of peer's MSS is better for the performance. It's more correct
1610 * but may be worse for the performance because of rcv_mss
1611 * fluctuations. --SAW 1998/11/1
1612 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001613 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 int free_space = tcp_space(sk);
1615 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1616 int window;
1617
1618 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001619 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001621 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001622 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
1624 if (tcp_memory_pressure)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001625 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1626 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
1628 if (free_space < mss)
1629 return 0;
1630 }
1631
1632 if (free_space > tp->rcv_ssthresh)
1633 free_space = tp->rcv_ssthresh;
1634
1635 /* Don't do rounding if we are using window scaling, since the
1636 * scaled window will not line up with the MSS boundary anyway.
1637 */
1638 window = tp->rcv_wnd;
1639 if (tp->rx_opt.rcv_wscale) {
1640 window = free_space;
1641
1642 /* Advertise enough space so that it won't get scaled away.
1643 * Import case: prevent zero window announcement if
1644 * 1<<rcv_wscale > mss.
1645 */
1646 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1647 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1648 << tp->rx_opt.rcv_wscale);
1649 } else {
1650 /* Get the largest window that is a nice multiple of mss.
1651 * Window clamp already applied above.
1652 * If our current window offering is within 1 mss of the
1653 * free space we just keep it. This prevents the divide
1654 * and multiply from happening most of the time.
1655 * We also don't do any window rounding when the free space
1656 * is too small.
1657 */
1658 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001659 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07001660 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001661 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07001662 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 }
1664
1665 return window;
1666}
1667
1668/* Attempt to collapse two adjacent SKB's during retransmission. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001669static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1670 int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671{
1672 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001673 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001674 int skb_size, next_skb_size;
1675 u16 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676
1677 /* The first test we must make is that neither of these two
1678 * SKB's are still referenced by someone else.
1679 */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001680 if (skb_cloned(skb) || skb_cloned(next_skb))
1681 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001683 skb_size = skb->len;
1684 next_skb_size = next_skb->len;
1685 flags = TCP_SKB_CB(skb)->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001687 /* Also punt if next skb has been SACK'd. */
1688 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1689 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001691 /* Next skb is out of window. */
1692 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1693 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001695 /* Punt if not enough space exists in the first SKB for
1696 * the data in the second, or the total combined payload
1697 * would exceed the MSS.
1698 */
1699 if ((next_skb_size > skb_tailroom(skb)) ||
1700 ((skb_size + next_skb_size) > mss_now))
1701 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001703 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07001704
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001705 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001707 /* Ok. We will be able to collapse the packet. */
1708 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001710 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1711 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001713 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1714 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001716 if (skb->ip_summed != CHECKSUM_PARTIAL)
1717 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001719 /* Update sequence range on original skb. */
1720 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001722 /* Merge over control information. */
1723 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1724 TCP_SKB_CB(skb)->flags = flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001726 /* All done, get rid of second SKB and account for it so
1727 * packet counting does not break.
1728 */
1729 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1730 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1731 tp->retrans_out -= tcp_skb_pcount(next_skb);
1732 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1733 tp->lost_out -= tcp_skb_pcount(next_skb);
1734 /* Reno case is special. Sigh... */
1735 if (tcp_is_reno(tp) && tp->sacked_out)
1736 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001737
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001738 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1739 tp->packets_out -= tcp_skb_pcount(next_skb);
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001740
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001741 /* changed transmit queue under us so clear hints */
1742 tcp_clear_retrans_hints_partial(tp);
1743
1744 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745}
1746
1747/* Do a simple retransmit without using the backoff mechanisms in
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001748 * tcp_timer. This is used for path mtu discovery.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 * The socket is already locked here.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001750 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751void tcp_simple_retransmit(struct sock *sk)
1752{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001753 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 struct tcp_sock *tp = tcp_sk(sk);
1755 struct sk_buff *skb;
1756 unsigned int mss = tcp_current_mss(sk, 0);
1757 int lost = 0;
1758
David S. Millerfe067e82007-03-07 12:12:44 -08001759 tcp_for_write_queue(skb, sk) {
1760 if (skb == tcp_send_head(sk))
1761 break;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001762 if (skb->len > mss &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001763 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1764 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1766 tp->retrans_out -= tcp_skb_pcount(skb);
1767 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001768 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1770 tp->lost_out += tcp_skb_pcount(skb);
1771 lost = 1;
1772 }
1773 }
1774 }
1775
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07001776 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001777
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 if (!lost)
1779 return;
1780
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001781 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001783 /* Don't muck with the congestion window here.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 * Reason is that we do not increase amount of _data_
1785 * in network, but units changed and effective
1786 * cwnd/ssthresh really reduced now.
1787 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001788 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001790 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 tp->prior_ssthresh = 0;
1792 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001793 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 }
1795 tcp_xmit_retransmit_queue(sk);
1796}
1797
1798/* This retransmits one SKB. Policy decisions and retransmit queue
1799 * state updates are done by the caller. Returns non-zero if an
1800 * error occurred which prevented the send.
1801 */
1802int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1803{
1804 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001805 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001806 unsigned int cur_mss = tcp_current_mss(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 int err;
1808
John Heffner5d424d52006-03-20 17:53:41 -08001809 /* Inconslusive MTU probe */
1810 if (icsk->icsk_mtup.probe_size) {
1811 icsk->icsk_mtup.probe_size = 0;
1812 }
1813
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001815 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 */
1817 if (atomic_read(&sk->sk_wmem_alloc) >
1818 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1819 return -EAGAIN;
1820
1821 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1822 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1823 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1825 return -ENOMEM;
1826 }
1827
1828 /* If receiver has shrunk his window, and skb is out of
1829 * new window, do not retransmit it. The exception is the
1830 * case, when window is shrunk to zero. In this case
1831 * our retransmit serves as a zero window probe.
1832 */
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001833 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1835 return -EAGAIN;
1836
1837 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001838 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 }
1841
1842 /* Collapse two adjacent packets if worthwhile and we can. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001843 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1844 (skb->len < (cur_mss >> 1)) &&
1845 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1846 (!tcp_skb_is_last(sk, skb)) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001847 (skb_shinfo(skb)->nr_frags == 0 &&
1848 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1849 (tcp_skb_pcount(skb) == 1 &&
1850 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001851 (sysctl_tcp_retrans_collapse != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 tcp_retrans_try_collapse(sk, skb, cur_mss);
1853
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001854 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 return -EHOSTUNREACH; /* Routing failure or similar. */
1856
1857 /* Some Solaris stacks overoptimize and ignore the FIN on a
1858 * retransmit when old data is attached. So strip it off
1859 * since it is cheap to do so and saves bytes on the network.
1860 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001861 if (skb->len > 0 &&
1862 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1863 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 if (!pskb_trim(skb, 0)) {
1865 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
Herbert Xu79671682006-06-22 02:40:14 -07001866 skb_shinfo(skb)->gso_segs = 1;
1867 skb_shinfo(skb)->gso_size = 0;
1868 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 skb->ip_summed = CHECKSUM_NONE;
1870 skb->csum = 0;
1871 }
1872 }
1873
1874 /* Make a copy, if the first transmission SKB clone we made
1875 * is still in somebody's hands, else make a clone.
1876 */
1877 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001879 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880
1881 if (err == 0) {
1882 /* Update global TCP statistics. */
1883 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1884
1885 tp->total_retrans++;
1886
1887#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001888 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 if (net_ratelimit())
1890 printk(KERN_DEBUG "retrans_out leaked.\n");
1891 }
1892#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001893 if (!tp->retrans_out)
1894 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1896 tp->retrans_out += tcp_skb_pcount(skb);
1897
1898 /* Save stamp of the first retransmit. */
1899 if (!tp->retrans_stamp)
1900 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1901
1902 tp->undo_retrans++;
1903
1904 /* snd_nxt is stored to detect loss of retransmitted segment,
1905 * see tcp_input.c tcp_sacktag_write_queue().
1906 */
1907 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1908 }
1909 return err;
1910}
1911
1912/* This gets called after a retransmit timeout, and the initially
1913 * retransmitted data is acknowledged. It tries to continue
1914 * resending the rest of the retransmit queue, until either
1915 * we've sent it all or the congestion window limit is reached.
1916 * If doing SACK, the first ACK which comes back for a timeout
1917 * based retransmit packet might feed us FACK information again.
1918 * If so, we use it to avoid unnecessarily retransmissions.
1919 */
1920void tcp_xmit_retransmit_queue(struct sock *sk)
1921{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001922 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 struct tcp_sock *tp = tcp_sk(sk);
1924 struct sk_buff *skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001925 int packet_cnt;
1926
1927 if (tp->retransmit_skb_hint) {
1928 skb = tp->retransmit_skb_hint;
1929 packet_cnt = tp->retransmit_cnt_hint;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001930 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08001931 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001932 packet_cnt = 0;
1933 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934
1935 /* First pass: retransmit lost packets. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001936 if (tp->lost_out) {
David S. Millerfe067e82007-03-07 12:12:44 -08001937 tcp_for_write_queue_from(skb, sk) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1939
David S. Millerfe067e82007-03-07 12:12:44 -08001940 if (skb == tcp_send_head(sk))
1941 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001942 /* we could do better than to assign each time */
1943 tp->retransmit_skb_hint = skb;
1944 tp->retransmit_cnt_hint = packet_cnt;
1945
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 /* Assume this retransmit will generate
1947 * only one packet for congestion window
1948 * calculation purposes. This works because
1949 * tcp_retransmit_skb() will chop up the
1950 * packet to be MSS sized and all the
1951 * packet counting works out.
1952 */
1953 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1954 return;
1955
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001956 if (sacked & TCPCB_LOST) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001957 if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001958 if (tcp_retransmit_skb(sk, skb)) {
1959 tp->retransmit_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001961 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001962 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1964 else
1965 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1966
David S. Millerfe067e82007-03-07 12:12:44 -08001967 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001968 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001969 inet_csk(sk)->icsk_rto,
1970 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 }
1972
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001973 packet_cnt += tcp_skb_pcount(skb);
1974 if (packet_cnt >= tp->lost_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 break;
1976 }
1977 }
1978 }
1979
1980 /* OK, demanded retransmission is finished. */
1981
1982 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001983 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 return;
1985
1986 /* No forward retransmissions in Reno are possible. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001987 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 return;
1989
1990 /* Yeah, we have to make difficult choice between forward transmission
1991 * and retransmission... Both ways have their merits...
1992 *
1993 * For now we do not retransmit anything, while we have some new
Ilpo Järvinen539d2432007-05-27 02:03:20 -07001994 * segments to send. In the other cases, follow rule 3 for
1995 * NextSeg() specified in RFC3517.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 */
1997
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001998 if (tcp_may_send_now(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 return;
2000
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002001 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2002 if (!tp->sacked_out)
2003 return;
2004
2005 if (tp->forward_skb_hint)
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002006 skb = tp->forward_skb_hint;
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002007 else
David S. Millerfe067e82007-03-07 12:12:44 -08002008 skb = tcp_write_queue_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009
David S. Millerfe067e82007-03-07 12:12:44 -08002010 tcp_for_write_queue_from(skb, sk) {
2011 if (skb == tcp_send_head(sk))
2012 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002013 tp->forward_skb_hint = skb;
2014
Ilpo Järvinen6859d492007-12-02 00:48:06 +02002015 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016 break;
2017
2018 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2019 break;
2020
2021 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2022 continue;
2023
2024 /* Ok, retransmit it. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002025 if (tcp_retransmit_skb(sk, skb)) {
2026 tp->forward_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002028 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029
David S. Millerfe067e82007-03-07 12:12:44 -08002030 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002031 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2032 inet_csk(sk)->icsk_rto,
2033 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034
2035 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2036 }
2037}
2038
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039/* Send a fin. The caller locks the socket for us. This cannot be
2040 * allowed to fail queueing a FIN frame under any circumstances.
2041 */
2042void tcp_send_fin(struct sock *sk)
2043{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002044 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002045 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002047
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 /* Optimization, tack on the FIN if we have a queue of
2049 * unsent frames. But be careful about outgoing SACKS
2050 * and IP options.
2051 */
2052 mss_now = tcp_current_mss(sk, 1);
2053
David S. Millerfe067e82007-03-07 12:12:44 -08002054 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2056 TCP_SKB_CB(skb)->end_seq++;
2057 tp->write_seq++;
2058 } else {
2059 /* Socket is locked, keep trying until memory is available. */
2060 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07002061 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 if (skb)
2063 break;
2064 yield();
2065 }
2066
2067 /* Reserve space for headers and prepare control bits. */
2068 skb_reserve(skb, MAX_TCP_HEADER);
2069 skb->csum = 0;
2070 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2071 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002072 skb_shinfo(skb)->gso_segs = 1;
2073 skb_shinfo(skb)->gso_size = 0;
2074 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075
2076 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2077 TCP_SKB_CB(skb)->seq = tp->write_seq;
2078 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2079 tcp_queue_skb(sk, skb);
2080 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002081 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082}
2083
2084/* We get here when a process closes a file descriptor (either due to
2085 * an explicit close() or as a byproduct of exit()'ing) and there
2086 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002087 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 */
Al Virodd0fc662005-10-07 07:46:04 +01002089void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 struct sk_buff *skb;
2092
2093 /* NOTE: No TCP options attached and we never retransmit this. */
2094 skb = alloc_skb(MAX_TCP_HEADER, priority);
2095 if (!skb) {
2096 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2097 return;
2098 }
2099
2100 /* Reserve space for headers and prepare control bits. */
2101 skb_reserve(skb, MAX_TCP_HEADER);
2102 skb->csum = 0;
2103 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2104 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002105 skb_shinfo(skb)->gso_segs = 1;
2106 skb_shinfo(skb)->gso_size = 0;
2107 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108
2109 /* Send it off. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002110 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2112 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002113 if (tcp_transmit_skb(sk, skb, 0, priority))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2115}
2116
2117/* WARNING: This routine must only be called when we have already sent
2118 * a SYN packet that crossed the incoming SYN that caused this routine
2119 * to get called. If this assumption fails then the initial rcv_wnd
2120 * and rcv_wscale values will not be correct.
2121 */
2122int tcp_send_synack(struct sock *sk)
2123{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002124 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125
David S. Millerfe067e82007-03-07 12:12:44 -08002126 skb = tcp_write_queue_head(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002127 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2129 return -EFAULT;
2130 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002131 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 if (skb_cloned(skb)) {
2133 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2134 if (nskb == NULL)
2135 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002136 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002138 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002139 sk_wmem_free_skb(sk, skb);
2140 sk->sk_wmem_queued += nskb->truesize;
2141 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142 skb = nskb;
2143 }
2144
2145 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2146 TCP_ECN_send_synack(tcp_sk(sk), skb);
2147 }
2148 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002149 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150}
2151
2152/*
2153 * Prepare a SYN-ACK.
2154 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002155struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2156 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002158 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 struct tcp_sock *tp = tcp_sk(sk);
2160 struct tcphdr *th;
2161 int tcp_header_size;
2162 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002163#ifdef CONFIG_TCP_MD5SIG
2164 struct tcp_md5sig_key *md5;
2165 __u8 *md5_hash_location;
2166#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167
2168 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2169 if (skb == NULL)
2170 return NULL;
2171
2172 /* Reserve space for headers. */
2173 skb_reserve(skb, MAX_TCP_HEADER);
2174
2175 skb->dst = dst_clone(dst);
2176
2177 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002178 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2179 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002181 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002182
2183#ifdef CONFIG_TCP_MD5SIG
2184 /* Are we doing MD5 on this segment? If so - make room for it */
2185 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2186 if (md5)
2187 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2188#endif
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002189 skb_push(skb, tcp_header_size);
2190 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002192 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 memset(th, 0, sizeof(struct tcphdr));
2194 th->syn = 1;
2195 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 TCP_ECN_make_synack(req, th);
2197 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002198 th->dest = ireq->rmt_port;
2199 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2201 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002202 skb_shinfo(skb)->gso_segs = 1;
2203 skb_shinfo(skb)->gso_size = 0;
2204 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002206 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002208 __u8 rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 /* Set this up on the first call only */
2210 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2211 /* tcp_full_space because it is guaranteed to be the first packet */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002212 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002213 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 &req->rcv_wnd,
2215 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002216 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 &rcv_wscale);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002218 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 }
2220
2221 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002222 th->window = htons(min(req->rcv_wnd, 65535U));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223
2224 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Al Virodf7a3b02006-09-27 18:38:52 -07002225 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002226 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 TCP_SKB_CB(skb)->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002228 req->ts_recent,
2229 (
2230#ifdef CONFIG_TCP_MD5SIG
2231 md5 ? &md5_hash_location :
2232#endif
2233 NULL)
2234 );
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235
2236 skb->csum = 0;
2237 th->doff = (tcp_header_size >> 2);
2238 TCP_INC_STATS(TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002239
2240#ifdef CONFIG_TCP_MD5SIG
2241 /* Okay, we have all we need - do the md5 hash if needed */
2242 if (md5) {
2243 tp->af_specific->calc_md5_hash(md5_hash_location,
2244 md5,
2245 NULL, dst, req,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002246 tcp_hdr(skb), sk->sk_protocol,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002247 skb->len);
2248 }
2249#endif
2250
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 return skb;
2252}
2253
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002254/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 * Do all connect socket setups that can be done AF independent.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002256 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002257static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258{
2259 struct dst_entry *dst = __sk_dst_get(sk);
2260 struct tcp_sock *tp = tcp_sk(sk);
2261 __u8 rcv_wscale;
2262
2263 /* We'll fix this up when we get a response from the other end.
2264 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2265 */
2266 tp->tcp_header_len = sizeof(struct tcphdr) +
2267 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2268
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002269#ifdef CONFIG_TCP_MD5SIG
2270 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2271 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2272#endif
2273
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 /* If user gave his TCP_MAXSEG, record it to clamp */
2275 if (tp->rx_opt.user_mss)
2276 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2277 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002278 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 tcp_sync_mss(sk, dst_mtu(dst));
2280
2281 if (!tp->window_clamp)
2282 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2283 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2284 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
2286 tcp_select_initial_window(tcp_full_space(sk),
2287 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2288 &tp->rcv_wnd,
2289 &tp->window_clamp,
2290 sysctl_tcp_window_scaling,
2291 &rcv_wscale);
2292
2293 tp->rx_opt.rcv_wscale = rcv_wscale;
2294 tp->rcv_ssthresh = tp->rcv_wnd;
2295
2296 sk->sk_err = 0;
2297 sock_reset_flag(sk, SOCK_DONE);
2298 tp->snd_wnd = 0;
2299 tcp_init_wl(tp, tp->write_seq, 0);
2300 tp->snd_una = tp->write_seq;
2301 tp->snd_sml = tp->write_seq;
2302 tp->rcv_nxt = 0;
2303 tp->rcv_wup = 0;
2304 tp->copied_seq = 0;
2305
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002306 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2307 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 tcp_clear_retrans(tp);
2309}
2310
2311/*
2312 * Build a SYN and send it off.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002313 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314int tcp_connect(struct sock *sk)
2315{
2316 struct tcp_sock *tp = tcp_sk(sk);
2317 struct sk_buff *buff;
2318
2319 tcp_connect_init(sk);
2320
David S. Millerd179cd12005-08-17 14:57:30 -07002321 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 if (unlikely(buff == NULL))
2323 return -ENOBUFS;
2324
2325 /* Reserve space for headers. */
2326 skb_reserve(buff, MAX_TCP_HEADER);
2327
2328 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002329 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002331 skb_shinfo(buff)->gso_segs = 1;
2332 skb_shinfo(buff)->gso_size = 0;
2333 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 buff->csum = 0;
Wei Yongjunbd37a082006-08-07 21:04:15 -07002335 tp->snd_nxt = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2337 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338
2339 /* Send it off. */
2340 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2341 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2342 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002343 __tcp_add_write_queue_tail(sk, buff);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002344 sk->sk_wmem_queued += buff->truesize;
2345 sk_mem_charge(sk, buff->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 tp->packets_out += tcp_skb_pcount(buff);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002347 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002348
2349 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2350 * in order to make this packet get counted in tcpOutSegs.
2351 */
2352 tp->snd_nxt = tp->write_seq;
2353 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2355
2356 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002357 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2358 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 return 0;
2360}
2361
2362/* Send out a delayed ack, the caller does the policy checking
2363 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2364 * for details.
2365 */
2366void tcp_send_delayed_ack(struct sock *sk)
2367{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002368 struct inet_connection_sock *icsk = inet_csk(sk);
2369 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 unsigned long timeout;
2371
2372 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002373 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002374 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002376 if (icsk->icsk_ack.pingpong ||
2377 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 max_ato = TCP_DELACK_MAX;
2379
2380 /* Slow path, intersegment interval is "high". */
2381
2382 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002383 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 * directly.
2385 */
2386 if (tp->srtt) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002387 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
2389 if (rtt < max_ato)
2390 max_ato = rtt;
2391 }
2392
2393 ato = min(ato, max_ato);
2394 }
2395
2396 /* Stay within the limit we were given */
2397 timeout = jiffies + ato;
2398
2399 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002400 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 /* If delack timer was blocked or is about to expire,
2402 * send ACK now.
2403 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002404 if (icsk->icsk_ack.blocked ||
2405 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406 tcp_send_ack(sk);
2407 return;
2408 }
2409
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002410 if (!time_before(timeout, icsk->icsk_ack.timeout))
2411 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002413 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2414 icsk->icsk_ack.timeout = timeout;
2415 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416}
2417
2418/* This routine sends an ack and also updates the window. */
2419void tcp_send_ack(struct sock *sk)
2420{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002421 struct sk_buff *buff;
2422
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002424 if (sk->sk_state == TCP_CLOSE)
2425 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002427 /* We are not putting this on the write queue, so
2428 * tcp_transmit_skb() will set the ownership to this
2429 * sock.
2430 */
2431 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2432 if (buff == NULL) {
2433 inet_csk_schedule_ack(sk);
2434 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2435 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2436 TCP_DELACK_MAX, TCP_RTO_MAX);
2437 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002439
2440 /* Reserve space for headers and prepare control bits. */
2441 skb_reserve(buff, MAX_TCP_HEADER);
2442 buff->csum = 0;
2443 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2444 TCP_SKB_CB(buff)->sacked = 0;
2445 skb_shinfo(buff)->gso_segs = 1;
2446 skb_shinfo(buff)->gso_size = 0;
2447 skb_shinfo(buff)->gso_type = 0;
2448
2449 /* Send it off, this clears delayed acks for us. */
2450 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
2451 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2452 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453}
2454
2455/* This routine sends a packet with an out of date sequence
2456 * number. It assumes the other end will try to ack it.
2457 *
2458 * Question: what should we make while urgent mode?
2459 * 4.4BSD forces sending single byte of data. We cannot send
2460 * out of window data, because we have SND.NXT==SND.MAX...
2461 *
2462 * Current solution: to send TWO zero-length segments in urgent mode:
2463 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2464 * out-of-date with SND.UNA-1 to probe window.
2465 */
2466static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2467{
2468 struct tcp_sock *tp = tcp_sk(sk);
2469 struct sk_buff *skb;
2470
2471 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2472 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002473 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 return -1;
2475
2476 /* Reserve space for headers and set control bits. */
2477 skb_reserve(skb, MAX_TCP_HEADER);
2478 skb->csum = 0;
2479 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
Ilpo Järvinen4828e7f2007-12-31 04:50:19 -08002480 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002481 skb_shinfo(skb)->gso_segs = 1;
2482 skb_shinfo(skb)->gso_size = 0;
2483 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484
2485 /* Use a previous sequence. This should cause the other
2486 * end to send an ack. Don't queue or clone SKB, just
2487 * send it.
2488 */
2489 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2490 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2491 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002492 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493}
2494
2495int tcp_write_wakeup(struct sock *sk)
2496{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002497 struct tcp_sock *tp = tcp_sk(sk);
2498 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002500 if (sk->sk_state == TCP_CLOSE)
2501 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002503 if ((skb = tcp_send_head(sk)) != NULL &&
2504 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2505 int err;
2506 unsigned int mss = tcp_current_mss(sk, 0);
2507 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002509 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2510 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002512 /* We are probing the opening of a window
2513 * but the window size is != 0
2514 * must have been a result SWS avoidance ( sender )
2515 */
2516 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2517 skb->len > mss) {
2518 seg_size = min(seg_size, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002520 if (tcp_fragment(sk, skb, seg_size, mss))
2521 return -1;
2522 } else if (!tcp_skb_pcount(skb))
2523 tcp_set_skb_tso_segs(sk, skb, mss);
2524
2525 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2526 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2527 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2528 if (!err)
2529 tcp_event_new_data_sent(sk, skb);
2530 return err;
2531 } else {
2532 if (tp->urg_mode &&
2533 between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2534 tcp_xmit_probe_skb(sk, 1);
2535 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537}
2538
2539/* A window probe timeout has occurred. If window is not closed send
2540 * a partial packet else a zero probe.
2541 */
2542void tcp_send_probe0(struct sock *sk)
2543{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002544 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 struct tcp_sock *tp = tcp_sk(sk);
2546 int err;
2547
2548 err = tcp_write_wakeup(sk);
2549
David S. Millerfe067e82007-03-07 12:12:44 -08002550 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002552 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002553 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 return;
2555 }
2556
2557 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002558 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2559 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002560 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002561 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002562 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2563 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 } else {
2565 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002566 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 * Let local senders to fight for local resources.
2568 *
2569 * Use accumulated backoff yet.
2570 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002571 if (!icsk->icsk_probes_out)
2572 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002573 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002574 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002575 TCP_RESOURCE_PROBE_INTERVAL),
2576 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 }
2578}
2579
2580EXPORT_SYMBOL(tcp_connect);
2581EXPORT_SYMBOL(tcp_make_synack);
2582EXPORT_SYMBOL(tcp_simple_retransmit);
2583EXPORT_SYMBOL(tcp_sync_mss);
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -08002584EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
John Heffner5d424d52006-03-20 17:53:41 -08002585EXPORT_SYMBOL(tcp_mtup_init);