blob: debf235816067cc8a268b015c9adbae851a538a1 [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 */
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700258 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 }
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
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800330/* Constructs common control bits of non-data skb. If SYN/FIN is present,
331 * auto increment end seqno.
332 */
333static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
334{
335 skb->csum = 0;
336
337 TCP_SKB_CB(skb)->flags = flags;
338 TCP_SKB_CB(skb)->sacked = 0;
339
340 skb_shinfo(skb)->gso_segs = 1;
341 skb_shinfo(skb)->gso_size = 0;
342 skb_shinfo(skb)->gso_type = 0;
343
344 TCP_SKB_CB(skb)->seq = seq;
345 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
346 seq++;
347 TCP_SKB_CB(skb)->end_seq = seq;
348}
349
Al Virodf7a3b02006-09-27 18:38:52 -0700350static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800351 __u32 tstamp, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800352{
353 if (tp->rx_opt.tstamp_ok) {
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700354 *ptr++ = htonl((TCPOPT_NOP << 24) |
355 (TCPOPT_NOP << 16) |
356 (TCPOPT_TIMESTAMP << 8) |
357 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800358 *ptr++ = htonl(tstamp);
359 *ptr++ = htonl(tp->rx_opt.ts_recent);
360 }
361 if (tp->rx_opt.eff_sacks) {
362 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
363 int this_sack;
364
365 *ptr++ = htonl((TCPOPT_NOP << 24) |
366 (TCPOPT_NOP << 16) |
367 (TCPOPT_SACK << 8) |
368 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
369 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800370
371 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800372 *ptr++ = htonl(sp[this_sack].start_seq);
373 *ptr++ = htonl(sp[this_sack].end_seq);
374 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800375
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800376 if (tp->rx_opt.dsack) {
377 tp->rx_opt.dsack = 0;
378 tp->rx_opt.eff_sacks--;
379 }
380 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800381#ifdef CONFIG_TCP_MD5SIG
382 if (md5_hash) {
383 *ptr++ = htonl((TCPOPT_NOP << 24) |
384 (TCPOPT_NOP << 16) |
385 (TCPOPT_MD5SIG << 8) |
386 TCPOLEN_MD5SIG);
387 *md5_hash = (__u8 *)ptr;
388 }
389#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800390}
391
392/* Construct a tcp options header for a SYN or SYN_ACK packet.
393 * If this is every changed make sure to change the definition of
394 * MAX_SYN_SIZE to match the new maximum number of options that you
395 * can generate.
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800396 *
397 * Note - that with the RFC2385 TCP option, we make room for the
398 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
399 * location to be filled is passed back up.
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800400 */
Al Virodf7a3b02006-09-27 18:38:52 -0700401static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800402 int offer_wscale, int wscale, __u32 tstamp,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800403 __u32 ts_recent, __u8 **md5_hash)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800404{
405 /* We always get an MSS option.
406 * The option bytes which will be seen in normal data
407 * packets should timestamps be used, must be in the MSS
408 * advertised. But we subtract them from tp->mss_cache so
409 * that calculations in tcp_sendmsg are simpler etc.
410 * So account for this fact here if necessary. If we
411 * don't do this correctly, as a receiver we won't
412 * recognize data packets as being full sized when we
413 * should, and thus we won't abide by the delayed ACK
414 * rules correctly.
415 * SACKs don't matter, we never delay an ACK when we
416 * have any of those going out.
417 */
418 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
419 if (ts) {
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800420 if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700421 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
422 (TCPOLEN_SACK_PERM << 16) |
423 (TCPOPT_TIMESTAMP << 8) |
424 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800425 else
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700426 *ptr++ = htonl((TCPOPT_NOP << 24) |
427 (TCPOPT_NOP << 16) |
428 (TCPOPT_TIMESTAMP << 8) |
429 TCPOLEN_TIMESTAMP);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800430 *ptr++ = htonl(tstamp); /* TSVAL */
431 *ptr++ = htonl(ts_recent); /* TSECR */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800432 } else if (sack)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700433 *ptr++ = htonl((TCPOPT_NOP << 24) |
434 (TCPOPT_NOP << 16) |
435 (TCPOPT_SACK_PERM << 8) |
436 TCPOLEN_SACK_PERM);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800437 if (offer_wscale)
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700438 *ptr++ = htonl((TCPOPT_NOP << 24) |
439 (TCPOPT_WINDOW << 16) |
440 (TCPOLEN_WINDOW << 8) |
441 (wscale));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800442#ifdef CONFIG_TCP_MD5SIG
443 /*
444 * If MD5 is enabled, then we set the option, and include the size
445 * (always 18). The actual MD5 hash is added just before the
446 * packet is sent.
447 */
448 if (md5_hash) {
449 *ptr++ = htonl((TCPOPT_NOP << 24) |
450 (TCPOPT_NOP << 16) |
451 (TCPOPT_MD5SIG << 8) |
452 TCPOLEN_MD5SIG);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800453 *md5_hash = (__u8 *)ptr;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800454 }
455#endif
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800456}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457
458/* This routine actually transmits TCP packets queued in by
459 * tcp_do_sendmsg(). This is used by both the initial
460 * transmission and possible later retransmissions.
461 * All SKB's seen here are completely headerless. It is our
462 * job to build the TCP header, and pass the packet down to
463 * IP so it can do the same plus pass the packet off to the
464 * device.
465 *
466 * We are working here with either a clone of the original
467 * SKB, or a fresh unique copy made by the retransmit engine.
468 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800469static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
470 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800472 const struct inet_connection_sock *icsk = inet_csk(sk);
473 struct inet_sock *inet;
474 struct tcp_sock *tp;
475 struct tcp_skb_cb *tcb;
476 int tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800477#ifdef CONFIG_TCP_MD5SIG
478 struct tcp_md5sig_key *md5;
479 __u8 *md5_hash_location;
480#endif
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800481 struct tcphdr *th;
482 int sysctl_flags;
483 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800485 BUG_ON(!skb || !tcp_skb_pcount(skb));
486
487 /* If congestion control is doing timestamping, we must
488 * take such a timestamp before we potentially clone/copy.
489 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700490 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800491 __net_timestamp(skb);
492
493 if (likely(clone_it)) {
494 if (unlikely(skb_cloned(skb)))
495 skb = pskb_copy(skb, gfp_mask);
496 else
497 skb = skb_clone(skb, gfp_mask);
498 if (unlikely(!skb))
499 return -ENOBUFS;
500 }
501
502 inet = inet_sk(sk);
503 tp = tcp_sk(sk);
504 tcb = TCP_SKB_CB(skb);
505 tcp_header_size = tp->tcp_header_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506
507#define SYSCTL_FLAG_TSTAMPS 0x1
508#define SYSCTL_FLAG_WSCALE 0x2
509#define SYSCTL_FLAG_SACK 0x4
510
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800511 sysctl_flags = 0;
512 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
513 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800514 if (sysctl_tcp_timestamps) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800515 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
516 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800518 if (sysctl_tcp_window_scaling) {
519 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
520 sysctl_flags |= SYSCTL_FLAG_WSCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800522 if (sysctl_tcp_sack) {
523 sysctl_flags |= SYSCTL_FLAG_SACK;
524 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
525 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800527 } else if (unlikely(tp->rx_opt.eff_sacks)) {
528 /* A SACK is 2 pad bytes, a 2 byte header, plus
529 * 2 32-bit sequence numbers for each SACK block.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800531 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
532 (tp->rx_opt.eff_sacks *
533 TCPOLEN_SACK_PERBLOCK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900535
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800536 if (tcp_packets_in_flight(tp) == 0)
537 tcp_ca_event(sk, CA_EVENT_TX_START);
538
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800539#ifdef CONFIG_TCP_MD5SIG
540 /*
541 * Are we doing MD5 on this segment? If so - make
542 * room for it.
543 */
544 md5 = tp->af_specific->md5_lookup(sk, sk);
545 if (md5)
546 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
547#endif
548
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700549 skb_push(skb, tcp_header_size);
550 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800551 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800552
553 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700554 th = tcp_hdr(skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800555 th->source = inet->sport;
556 th->dest = inet->dport;
557 th->seq = htonl(tcb->seq);
558 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700559 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800560 tcb->flags);
561
562 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
563 /* RFC1323: The window in SYN & SYN/ACK segments
564 * is never scaled.
565 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800566 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800567 } else {
568 th->window = htons(tcp_select_window(sk));
569 }
570 th->check = 0;
571 th->urg_ptr = 0;
572
573 if (unlikely(tp->urg_mode &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800574 between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
575 th->urg_ptr = htons(tp->snd_up - tcb->seq);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800576 th->urg = 1;
577 }
578
579 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
Al Virodf7a3b02006-09-27 18:38:52 -0700580 tcp_syn_build_options((__be32 *)(th + 1),
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800581 tcp_advertise_mss(sk),
582 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
583 (sysctl_flags & SYSCTL_FLAG_SACK),
584 (sysctl_flags & SYSCTL_FLAG_WSCALE),
585 tp->rx_opt.rcv_wscale,
586 tcb->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800587 tp->rx_opt.ts_recent,
588
589#ifdef CONFIG_TCP_MD5SIG
590 md5 ? &md5_hash_location :
591#endif
592 NULL);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800593 } else {
Al Virodf7a3b02006-09-27 18:38:52 -0700594 tcp_build_and_update_options((__be32 *)(th + 1),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800595 tp, tcb->when,
596#ifdef CONFIG_TCP_MD5SIG
597 md5 ? &md5_hash_location :
598#endif
599 NULL);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700600 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800601 }
602
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800603#ifdef CONFIG_TCP_MD5SIG
604 /* Calculate the MD5 hash, as we have all we need now */
605 if (md5) {
606 tp->af_specific->calc_md5_hash(md5_hash_location,
607 md5,
608 sk, NULL, NULL,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700609 tcp_hdr(skb),
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800610 sk->sk_protocol,
611 skb->len);
612 }
613#endif
614
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800615 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800616
617 if (likely(tcb->flags & TCPCB_FLAG_ACK))
618 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
619
620 if (skb->len != tcp_header_size)
621 tcp_event_data_sent(tp, skb, sk);
622
Wei Yongjunbd37a082006-08-07 21:04:15 -0700623 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
624 TCP_INC_STATS(TCP_MIB_OUTSEGS);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800625
David S. Millere89862f2007-01-26 01:04:55 -0800626 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700627 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800628 return err;
629
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800630 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800631
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200632 return net_xmit_eval(err);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800633
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634#undef SYSCTL_FLAG_TSTAMPS
635#undef SYSCTL_FLAG_WSCALE
636#undef SYSCTL_FLAG_SACK
637}
638
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900639/* This routine just queue's the buffer
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 *
641 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
642 * otherwise socket can stall.
643 */
644static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
645{
646 struct tcp_sock *tp = tcp_sk(sk);
647
648 /* Advance write_seq and place onto the write_queue. */
649 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
650 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800651 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800652 sk->sk_wmem_queued += skb->truesize;
653 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654}
655
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800656static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
657 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700658{
Herbert Xubcd76112006-06-30 13:36:35 -0700659 if (skb->len <= mss_now || !sk_can_gso(sk)) {
David S. Millerf6302d12005-07-05 15:18:03 -0700660 /* Avoid the costly divide in the normal
661 * non-TSO case.
662 */
Herbert Xu79671682006-06-22 02:40:14 -0700663 skb_shinfo(skb)->gso_segs = 1;
664 skb_shinfo(skb)->gso_size = 0;
665 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700666 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700667 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700668 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700669 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 }
671}
672
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700673/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -0800674 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700675 */
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800676static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700677 int decr)
678{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800679 struct tcp_sock *tp = tcp_sk(sk);
680
Ilpo Järvinendc869672007-10-01 15:27:19 -0700681 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700682 return;
683
Ilpo Järvinen6859d492007-12-02 00:48:06 +0200684 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700685 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700686}
687
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688/* Function to create two new TCP segments. Shrinks the given segment
689 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900690 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 * Remember, these are still headerless SKBs at this point.
692 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800693int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
694 unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695{
696 struct tcp_sock *tp = tcp_sk(sk);
697 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700698 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700699 int nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 u16 flags;
701
Herbert Xub2cc99f02005-10-20 17:13:13 -0200702 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800703
Ilpo Järvinenb7689202007-09-20 11:37:19 -0700704 tcp_clear_retrans_hints_partial(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 nsize = skb_headlen(skb) - len;
706 if (nsize < 0)
707 nsize = 0;
708
709 if (skb_cloned(skb) &&
710 skb_is_nonlinear(skb) &&
711 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
712 return -ENOMEM;
713
714 /* Get a new skb... force flag on. */
715 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
716 if (buff == NULL)
717 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -0700718
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800719 sk->sk_wmem_queued += buff->truesize;
720 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -0700721 nlen = skb->len - len - nsize;
722 buff->truesize += nlen;
723 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724
725 /* Correct the sequence numbers. */
726 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
727 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
728 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
729
730 /* PSH and FIN should only be set in the second packet. */
731 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800732 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700734 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
Patrick McHardy84fa7932006-08-29 16:44:56 -0700736 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800738 buff->csum = csum_partial_copy_nocheck(skb->data + len,
739 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 nsize, 0);
741
742 skb_trim(skb, len);
743
744 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
745 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -0700746 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 skb_split(skb, buff, len);
748 }
749
750 buff->ip_summed = skb->ip_summed;
751
752 /* Looks stupid, but our code really uses when of
753 * skbs, which it never sent before. --ANK
754 */
755 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700756 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757
David S. Miller6475be12005-09-01 22:47:01 -0700758 old_factor = tcp_skb_pcount(skb);
759
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700761 tcp_set_skb_tso_segs(sk, skb, mss_now);
762 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763
David S. Miller6475be12005-09-01 22:47:01 -0700764 /* If this packet has been sent out already, we must
765 * adjust the various packet counters.
766 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700767 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700768 int diff = old_factor - tcp_skb_pcount(skb) -
769 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770
David S. Miller6475be12005-09-01 22:47:01 -0700771 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700772
773 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
774 tp->sacked_out -= diff;
775 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
776 tp->retrans_out -= diff;
777
Ilpo Järvinenb5860bb2007-08-09 14:33:18 +0300778 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
David S. Miller6475be12005-09-01 22:47:01 -0700779 tp->lost_out -= diff;
Herbert Xu83ca28b2005-09-22 23:32:56 -0700780
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700781 /* Adjust Reno SACK estimate. */
782 if (tcp_is_reno(tp) && diff > 0) {
783 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
784 tcp_verify_left_out(tp);
David S. Miller6475be12005-09-01 22:47:01 -0700785 }
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800786 tcp_adjust_fackets_out(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787 }
788
789 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700790 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -0800791 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792
793 return 0;
794}
795
796/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
797 * eventually). The difference is that pulled data not copied, but
798 * immediately discarded.
799 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700800static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801{
802 int i, k, eat;
803
804 eat = len;
805 k = 0;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800806 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 if (skb_shinfo(skb)->frags[i].size <= eat) {
808 put_page(skb_shinfo(skb)->frags[i].page);
809 eat -= skb_shinfo(skb)->frags[i].size;
810 } else {
811 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
812 if (eat) {
813 skb_shinfo(skb)->frags[k].page_offset += eat;
814 skb_shinfo(skb)->frags[k].size -= eat;
815 eat = 0;
816 }
817 k++;
818 }
819 }
820 skb_shinfo(skb)->nr_frags = k;
821
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -0700822 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 skb->data_len -= len;
824 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825}
826
827int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
828{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800829 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 return -ENOMEM;
831
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700832 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
833 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700835 else
836 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837
838 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -0700839 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840
841 skb->truesize -= len;
842 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800843 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
845
846 /* Any change of skb->len requires recalculation of tso
847 * factor and mss.
848 */
849 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700850 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851
852 return 0;
853}
854
John Heffner5d424d52006-03-20 17:53:41 -0800855/* Not accounting for SACKs here. */
856int tcp_mtu_to_mss(struct sock *sk, int pmtu)
857{
858 struct tcp_sock *tp = tcp_sk(sk);
859 struct inet_connection_sock *icsk = inet_csk(sk);
860 int mss_now;
861
862 /* Calculate base mss without TCP options:
863 It is MMS_S - sizeof(tcphdr) of rfc1122
864 */
865 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
866
867 /* Clamp it (mss_clamp does not include tcp options) */
868 if (mss_now > tp->rx_opt.mss_clamp)
869 mss_now = tp->rx_opt.mss_clamp;
870
871 /* Now subtract optional transport overhead */
872 mss_now -= icsk->icsk_ext_hdr_len;
873
874 /* Then reserve room for full set of TCP options and 8 bytes of data */
875 if (mss_now < 48)
876 mss_now = 48;
877
878 /* Now subtract TCP options size, not including SACKs */
879 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
880
881 return mss_now;
882}
883
884/* Inverse of above */
885int tcp_mss_to_mtu(struct sock *sk, int mss)
886{
887 struct tcp_sock *tp = tcp_sk(sk);
888 struct inet_connection_sock *icsk = inet_csk(sk);
889 int mtu;
890
891 mtu = mss +
892 tp->tcp_header_len +
893 icsk->icsk_ext_hdr_len +
894 icsk->icsk_af_ops->net_header_len;
895
896 return mtu;
897}
898
899void tcp_mtup_init(struct sock *sk)
900{
901 struct tcp_sock *tp = tcp_sk(sk);
902 struct inet_connection_sock *icsk = inet_csk(sk);
903
904 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
905 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900906 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -0800907 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
908 icsk->icsk_mtup.probe_size = 0;
909}
910
Ilpo Järvinen409d22b2007-12-31 14:57:40 -0800911/* Bound MSS / TSO packet size with the half of the window */
912static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
913{
914 if (tp->max_window && pktsize > (tp->max_window >> 1))
915 return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
916 else
917 return pktsize;
918}
919
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920/* This function synchronize snd mss to current pmtu/exthdr set.
921
922 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
923 for TCP options, but includes only bare TCP header.
924
925 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800926 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 It also does not include TCP options.
928
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800929 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930
931 tp->mss_cache is current effective sending mss, including
932 all tcp options except for SACKs. It is evaluated,
933 taking into account current pmtu, but never exceeds
934 tp->rx_opt.mss_clamp.
935
936 NOTE1. rfc1122 clearly states that advertised MSS
937 DOES NOT include either tcp or ip options.
938
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800939 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
940 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
943{
944 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800945 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -0800946 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
John Heffner5d424d52006-03-20 17:53:41 -0800948 if (icsk->icsk_mtup.search_high > pmtu)
949 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950
John Heffner5d424d52006-03-20 17:53:41 -0800951 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -0800952 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953
954 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800955 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -0800956 if (icsk->icsk_mtup.enabled)
957 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700958 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959
960 return mss_now;
961}
962
963/* Compute the current effective MSS, taking SACKs and IP options,
964 * and even PMTU discovery events into account.
965 *
966 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
967 * cannot be large. However, taking into account rare use of URG, this
968 * is not a big flaw.
969 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700970unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971{
972 struct tcp_sock *tp = tcp_sk(sk);
973 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700974 u32 mss_now;
975 u16 xmit_size_goal;
976 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700978 mss_now = tp->mss_cache;
979
Herbert Xubcd76112006-06-30 13:36:35 -0700980 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700981 doing_tso = 1;
982
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 if (dst) {
984 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800985 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 mss_now = tcp_sync_mss(sk, mtu);
987 }
988
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 if (tp->rx_opt.eff_sacks)
990 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
991 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700992
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800993#ifdef CONFIG_TCP_MD5SIG
994 if (tp->af_specific->md5_lookup(sk, sk))
995 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
996#endif
997
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700998 xmit_size_goal = mss_now;
999
1000 if (doing_tso) {
Peter P Waskiewicz Jr82cc1a72008-03-21 03:43:19 -07001001 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001002 inet_csk(sk)->icsk_af_ops->net_header_len -
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001003 inet_csk(sk)->icsk_ext_hdr_len -
1004 tp->tcp_header_len);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001005
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001006 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001007 xmit_size_goal -= (xmit_size_goal % mss_now);
1008 }
1009 tp->xmit_size_goal = xmit_size_goal;
1010
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 return mss_now;
1012}
1013
David S. Millera762a982005-07-05 15:18:51 -07001014/* Congestion window validation. (RFC2861) */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001015static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001016{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001017 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001018
Ilpo Järvinend436d682007-12-31 14:58:00 -08001019 if (tp->packets_out >= tp->snd_cwnd) {
David S. Millera762a982005-07-05 15:18:51 -07001020 /* Network is feed fully. */
1021 tp->snd_cwnd_used = 0;
1022 tp->snd_cwnd_stamp = tcp_time_stamp;
1023 } else {
1024 /* Network starves. */
1025 if (tp->packets_out > tp->snd_cwnd_used)
1026 tp->snd_cwnd_used = tp->packets_out;
1027
David S. Miller15d33c02007-04-09 13:23:14 -07001028 if (sysctl_tcp_slow_start_after_idle &&
1029 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001030 tcp_cwnd_application_limited(sk);
1031 }
1032}
1033
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001034/* Returns the portion of skb which can be sent right away without
1035 * introducing MSS oddities to segment boundaries. In rare cases where
1036 * mss_now != mss_cache, we will request caller to create a small skb
1037 * per input skb which could be mostly avoided here (if desired).
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001038 *
1039 * We explicitly want to create a request for splitting write queue tail
1040 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1041 * thus all the complexity (cwnd_len is always MSS multiple which we
1042 * return whenever allowed by the other factors). Basically we need the
1043 * modulo only when the receiver window alone is the limiting factor or
1044 * when we would be allowed to send the split-due-to-Nagle skb fully.
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001045 */
1046static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001047 unsigned int mss_now, unsigned int cwnd)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001048{
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001049 struct tcp_sock *tp = tcp_sk(sk);
1050 u32 needed, window, cwnd_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001051
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001052 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001053 cwnd_len = mss_now * cwnd;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001054
1055 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1056 return cwnd_len;
1057
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001058 needed = min(skb->len, window);
1059
Ilpo Järvinen17515402008-04-15 20:36:55 -07001060 if (cwnd_len <= needed)
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001061 return cwnd_len;
1062
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001063 return needed - needed % mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001064}
1065
1066/* Can at least one segment of SKB be sent right now, according to the
1067 * congestion window rules? If so, return how many segments are allowed.
1068 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001069static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1070 struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001071{
1072 u32 in_flight, cwnd;
1073
1074 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -08001075 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1076 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001077 return 1;
1078
1079 in_flight = tcp_packets_in_flight(tp);
1080 cwnd = tp->snd_cwnd;
1081 if (in_flight < cwnd)
1082 return (cwnd - in_flight);
1083
1084 return 0;
1085}
1086
1087/* This must be invoked the first time we consider transmitting
1088 * SKB onto the wire.
1089 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001090static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1091 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001092{
1093 int tso_segs = tcp_skb_pcount(skb);
1094
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001095 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001096 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001097 tso_segs = tcp_skb_pcount(skb);
1098 }
1099 return tso_segs;
1100}
1101
1102static inline int tcp_minshall_check(const struct tcp_sock *tp)
1103{
1104 return after(tp->snd_sml,tp->snd_una) &&
1105 !after(tp->snd_sml, tp->snd_nxt);
1106}
1107
1108/* Return 0, if packet can be sent now without violation Nagle's rules:
1109 * 1. It is full sized.
1110 * 2. Or it contains FIN. (already checked by caller)
1111 * 3. Or TCP_NODELAY was set.
1112 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1113 * With Minshall's modification: all sent small packets are ACKed.
1114 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001115static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001116 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001117 unsigned mss_now, int nonagle)
1118{
1119 return (skb->len < mss_now &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001120 ((nonagle & TCP_NAGLE_CORK) ||
1121 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001122}
1123
1124/* Return non-zero if the Nagle test allows this packet to be
1125 * sent now.
1126 */
1127static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1128 unsigned int cur_mss, int nonagle)
1129{
1130 /* Nagle rule does not apply to frames, which sit in the middle of the
1131 * write_queue (they have no chances to get new data).
1132 *
1133 * This is implemented in the callers, where they modify the 'nonagle'
1134 * argument based upon the location of SKB in the send queue.
1135 */
1136 if (nonagle & TCP_NAGLE_PUSH)
1137 return 1;
1138
Ilpo Järvinend551e452007-04-30 00:42:20 -07001139 /* Don't use the nagle rule for urgent data (or for the final FIN).
1140 * Nagle can be ignored during F-RTO too (see RFC4138).
1141 */
1142 if (tp->urg_mode || (tp->frto_counter == 2) ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001143 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1144 return 1;
1145
1146 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1147 return 1;
1148
1149 return 0;
1150}
1151
1152/* Does at least the first segment of SKB fit into the send window? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001153static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1154 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001155{
1156 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1157
1158 if (skb->len > cur_mss)
1159 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1160
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001161 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001162}
1163
David S. Millerfe067e82007-03-07 12:12:44 -08001164/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001165 * should be put on the wire right now. If so, it returns the number of
1166 * packets allowed by the congestion window.
1167 */
1168static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1169 unsigned int cur_mss, int nonagle)
1170{
1171 struct tcp_sock *tp = tcp_sk(sk);
1172 unsigned int cwnd_quota;
1173
David S. Miller846998a2005-08-04 19:52:01 -07001174 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001175
1176 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1177 return 0;
1178
1179 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001180 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001181 cwnd_quota = 0;
1182
1183 return cwnd_quota;
1184}
1185
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001186int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001187{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001188 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001189 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001190
1191 return (skb &&
1192 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1193 (tcp_skb_is_last(sk, skb) ?
Ilpo Järvinen4e67d872007-12-05 02:25:32 -08001194 tp->nonagle : TCP_NAGLE_PUSH)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001195}
1196
1197/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1198 * which is put after SKB on the list. It is very much like
1199 * tcp_fragment() except that it may make several kinds of assumptions
1200 * in order to speed up the splitting operation. In particular, we
1201 * know that all the data is in scatter-gather pages, and that the
1202 * packet has never been sent out before (and thus is not cloned).
1203 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001204static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1205 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001206{
1207 struct sk_buff *buff;
1208 int nlen = skb->len - len;
1209 u16 flags;
1210
1211 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001212 if (skb->len != skb->data_len)
1213 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001214
Pavel Emelyanovdf97c702007-11-29 21:22:33 +11001215 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001216 if (unlikely(buff == NULL))
1217 return -ENOMEM;
1218
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001219 sk->sk_wmem_queued += buff->truesize;
1220 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001221 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001222 skb->truesize -= nlen;
1223
1224 /* Correct the sequence numbers. */
1225 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1226 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1227 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1228
1229 /* PSH and FIN should only be set in the second packet. */
1230 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001231 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001232 TCP_SKB_CB(buff)->flags = flags;
1233
1234 /* This packet was never sent out yet, so no SACK bits. */
1235 TCP_SKB_CB(buff)->sacked = 0;
1236
Patrick McHardy84fa7932006-08-29 16:44:56 -07001237 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001238 skb_split(skb, buff, len);
1239
1240 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001241 tcp_set_skb_tso_segs(sk, skb, mss_now);
1242 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001243
1244 /* Link BUFF into the send queue. */
1245 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001246 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001247
1248 return 0;
1249}
1250
1251/* Try to defer sending, if possible, in order to minimize the amount
1252 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1253 *
1254 * This algorithm is from John Heffner.
1255 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001256static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001257{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001258 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001259 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001260 u32 send_win, cong_win, limit, in_flight;
1261
1262 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001263 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001264
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001265 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001266 goto send_now;
1267
1268 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001269 if (tp->tso_deferred &&
1270 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001271 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001272
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001273 in_flight = tcp_packets_in_flight(tp);
1274
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001275 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001276
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001277 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001278
1279 /* From in_flight test above, we know that cwnd > in_flight. */
1280 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1281
1282 limit = min(send_win, cong_win);
1283
David S. Millerba244fe2006-03-11 18:51:49 -08001284 /* If a full-sized TSO skb can be sent, do it. */
Peter P Waskiewicz Jr82cc1a72008-03-21 03:43:19 -07001285 if (limit >= sk->sk_gso_max_size)
John Heffnerae8064a2006-10-18 20:36:48 -07001286 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001287
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001288 if (sysctl_tcp_tso_win_divisor) {
1289 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1290
1291 /* If at least some fraction of a window is available,
1292 * just use it.
1293 */
1294 chunk /= sysctl_tcp_tso_win_divisor;
1295 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001296 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001297 } else {
1298 /* Different approach, try not to defer past a single
1299 * ACK. Receiver should ACK every other full sized
1300 * frame, so if we have space for more than 3 frames
1301 * then send now.
1302 */
1303 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001304 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001305 }
1306
1307 /* Ok, it looks like it is advisable to defer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001308 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001309
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001310 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001311
1312send_now:
1313 tp->tso_deferred = 0;
1314 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001315}
1316
John Heffner5d424d52006-03-20 17:53:41 -08001317/* Create a new MTU probe if we are ready.
1318 * Returns 0 if we should wait to probe (no cwnd available),
1319 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001320 * -1 otherwise
1321 */
John Heffner5d424d52006-03-20 17:53:41 -08001322static int tcp_mtu_probe(struct sock *sk)
1323{
1324 struct tcp_sock *tp = tcp_sk(sk);
1325 struct inet_connection_sock *icsk = inet_csk(sk);
1326 struct sk_buff *skb, *nskb, *next;
1327 int len;
1328 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001329 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001330 int copy;
1331 int mss_now;
1332
1333 /* Not currently probing/verifying,
1334 * not in recovery,
1335 * have enough cwnd, and
1336 * not SACKing (the variable headers throw things off) */
1337 if (!icsk->icsk_mtup.enabled ||
1338 icsk->icsk_mtup.probe_size ||
1339 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1340 tp->snd_cwnd < 11 ||
1341 tp->rx_opt.eff_sacks)
1342 return -1;
1343
1344 /* Very simple search strategy: just double the MSS. */
1345 mss_now = tcp_current_mss(sk, 0);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001346 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001347 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001348 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1349 /* TODO: set timer for probe_converge_event */
1350 return -1;
1351 }
1352
1353 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001354 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001355 return -1;
1356
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001357 if (tp->snd_wnd < size_needed)
1358 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001359 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001360 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001361
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001362 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1363 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1364 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001365 return -1;
1366 else
1367 return 0;
1368 }
1369
1370 /* We're allowed to probe. Build it now. */
1371 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1372 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001373 sk->sk_wmem_queued += nskb->truesize;
1374 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001375
David S. Millerfe067e82007-03-07 12:12:44 -08001376 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001377
1378 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1379 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1380 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1381 TCP_SKB_CB(nskb)->sacked = 0;
1382 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001383 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001384
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001385 tcp_insert_write_queue_before(nskb, skb, sk);
1386
John Heffner5d424d52006-03-20 17:53:41 -08001387 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001388 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001389 copy = min_t(int, skb->len, probe_size - len);
1390 if (nskb->ip_summed)
1391 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1392 else
1393 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001394 skb_put(nskb, copy),
1395 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001396
1397 if (skb->len <= copy) {
1398 /* We've eaten all the data from this skb.
1399 * Throw it away. */
1400 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001401 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001402 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001403 } else {
1404 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001405 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001406 if (!skb_shinfo(skb)->nr_frags) {
1407 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001408 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001409 skb->csum = csum_partial(skb->data,
1410 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001411 } else {
1412 __pskb_trim_head(skb, copy);
1413 tcp_set_skb_tso_segs(sk, skb, mss_now);
1414 }
1415 TCP_SKB_CB(skb)->seq += copy;
1416 }
1417
1418 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001419
1420 if (len >= probe_size)
1421 break;
John Heffner5d424d52006-03-20 17:53:41 -08001422 }
1423 tcp_init_tso_segs(sk, nskb, nskb->len);
1424
1425 /* We're ready to send. If this fails, the probe will
1426 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1427 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1428 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1429 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001430 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001431 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001432 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001433
1434 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001435 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1436 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001437
1438 return 1;
1439 }
1440
1441 return -1;
1442}
1443
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444/* This routine writes packets to the network. It advances the
1445 * send_head. This happens as incoming acks open up the remote
1446 * window for us.
1447 *
1448 * Returns 1, if no segments are in flight and we have queued segments, but
1449 * cannot send anything now because of SWS or another problem.
1450 */
David S. Millera2e2a592005-07-05 15:19:23 -07001451static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452{
1453 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001454 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001455 unsigned int tso_segs, sent_pkts;
1456 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001457 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458
1459 /* If we are closed, the bytes will have to remain here.
1460 * In time closedown will finish, we empty the write queue and all
1461 * will be happy.
1462 */
David S. Miller92df7b52005-07-05 15:19:06 -07001463 if (unlikely(sk->sk_state == TCP_CLOSE))
1464 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
David S. Miller92df7b52005-07-05 15:19:06 -07001466 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001467
1468 /* Do MTU probing. */
1469 if ((result = tcp_mtu_probe(sk)) == 0) {
1470 return 0;
1471 } else if (result > 0) {
1472 sent_pkts = 1;
1473 }
1474
David S. Millerfe067e82007-03-07 12:12:44 -08001475 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001476 unsigned int limit;
1477
Herbert Xub68e9f82005-08-04 19:52:02 -07001478 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001479 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001480
Herbert Xub68e9f82005-08-04 19:52:02 -07001481 cwnd_quota = tcp_cwnd_test(tp, skb);
1482 if (!cwnd_quota)
1483 break;
1484
1485 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1486 break;
1487
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001488 if (tso_segs == 1) {
1489 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1490 (tcp_skb_is_last(sk, skb) ?
1491 nonagle : TCP_NAGLE_PUSH))))
1492 break;
1493 } else {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001494 if (tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001495 break;
1496 }
David S. Milleraa934662005-07-05 15:20:09 -07001497
Herbert Xuc8ac3772005-08-16 20:43:40 -07001498 limit = mss_now;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001499 if (tso_segs > 1)
1500 limit = tcp_mss_split_point(sk, skb, mss_now,
1501 cwnd_quota);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502
Herbert Xuc8ac3772005-08-16 20:43:40 -07001503 if (skb->len > limit &&
1504 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1505 break;
1506
David S. Miller92df7b52005-07-05 15:19:06 -07001507 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001508
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001509 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
David S. Miller92df7b52005-07-05 15:19:06 -07001510 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
David S. Miller92df7b52005-07-05 15:19:06 -07001512 /* Advance the send_head. This one is sent out.
1513 * This call will increment packets_out.
1514 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001515 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001516
1517 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001518 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519 }
David S. Miller92df7b52005-07-05 15:19:06 -07001520
David S. Milleraa934662005-07-05 15:20:09 -07001521 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001522 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001523 return 0;
1524 }
David S. Millerfe067e82007-03-07 12:12:44 -08001525 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526}
1527
David S. Millera762a982005-07-05 15:18:51 -07001528/* Push out any pending frames which were held back due to
1529 * TCP_CORK or attempt at coalescing tiny packets.
1530 * The socket must be locked by the caller.
1531 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001532void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1533 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001534{
David S. Millerfe067e82007-03-07 12:12:44 -08001535 struct sk_buff *skb = tcp_send_head(sk);
David S. Millera762a982005-07-05 15:18:51 -07001536
1537 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001538 if (tcp_write_xmit(sk, cur_mss, nonagle))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001539 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001540 }
1541}
1542
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001543/* Send _single_ skb sitting at the send head. This function requires
1544 * true push pending frames to setup probe timer etc.
1545 */
1546void tcp_push_one(struct sock *sk, unsigned int mss_now)
1547{
David S. Millerfe067e82007-03-07 12:12:44 -08001548 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001549 unsigned int tso_segs, cwnd_quota;
1550
1551 BUG_ON(!skb || skb->len < mss_now);
1552
David S. Miller846998a2005-08-04 19:52:01 -07001553 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001554 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1555
1556 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001557 unsigned int limit;
1558
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001559 BUG_ON(!tso_segs);
1560
Herbert Xuc8ac3772005-08-16 20:43:40 -07001561 limit = mss_now;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001562 if (tso_segs > 1)
1563 limit = tcp_mss_split_point(sk, skb, mss_now,
1564 cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001565
Herbert Xuc8ac3772005-08-16 20:43:40 -07001566 if (skb->len > limit &&
1567 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1568 return;
1569
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001570 /* Send it out now. */
1571 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1572
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001573 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001574 tcp_event_new_data_sent(sk, skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001575 tcp_cwnd_validate(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001576 return;
1577 }
1578 }
1579}
1580
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581/* This function returns the amount that we can raise the
1582 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001583 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 * 1. The window can never be shrunk once it is offered (RFC 793)
1585 * 2. We limit memory per socket
1586 *
1587 * RFC 1122:
1588 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1589 * RECV.NEXT + RCV.WIN fixed until:
1590 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1591 *
1592 * i.e. don't raise the right edge of the window until you can raise
1593 * it at least MSS bytes.
1594 *
1595 * Unfortunately, the recommended algorithm breaks header prediction,
1596 * since header prediction assumes th->window stays fixed.
1597 *
1598 * Strictly speaking, keeping th->window fixed violates the receiver
1599 * side SWS prevention criteria. The problem is that under this rule
1600 * a stream of single byte packets will cause the right side of the
1601 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001602 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 * Of course, if the sender implements sender side SWS prevention
1604 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001605 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001607 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 * If the free space is less than the 1/4 of the maximum
1609 * space available and the free space is less than 1/2 mss,
1610 * then set the window to 0.
1611 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1612 * Otherwise, just prevent the window from shrinking
1613 * and from being larger than the largest representable value.
1614 *
1615 * This prevents incremental opening of the window in the regime
1616 * where TCP is limited by the speed of the reader side taking
1617 * data out of the TCP receive queue. It does nothing about
1618 * those cases where the window is constrained on the sender side
1619 * because the pipeline is full.
1620 *
1621 * BSD also seems to "accidentally" limit itself to windows that are a
1622 * multiple of MSS, at least until the free space gets quite small.
1623 * This would appear to be a side effect of the mbuf implementation.
1624 * Combining these two algorithms results in the observed behavior
1625 * of having a fixed window size at almost all times.
1626 *
1627 * Below we obtain similar behavior by forcing the offered window to
1628 * a multiple of the mss when it is feasible to do so.
1629 *
1630 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1631 * Regular options like TIMESTAMP are taken into account.
1632 */
1633u32 __tcp_select_window(struct sock *sk)
1634{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001635 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001637 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638 * here. I don't know if the value based on our guesses
1639 * of peer's MSS is better for the performance. It's more correct
1640 * but may be worse for the performance because of rcv_mss
1641 * fluctuations. --SAW 1998/11/1
1642 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001643 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 int free_space = tcp_space(sk);
1645 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1646 int window;
1647
1648 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001649 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001651 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001652 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653
1654 if (tcp_memory_pressure)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001655 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1656 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657
1658 if (free_space < mss)
1659 return 0;
1660 }
1661
1662 if (free_space > tp->rcv_ssthresh)
1663 free_space = tp->rcv_ssthresh;
1664
1665 /* Don't do rounding if we are using window scaling, since the
1666 * scaled window will not line up with the MSS boundary anyway.
1667 */
1668 window = tp->rcv_wnd;
1669 if (tp->rx_opt.rcv_wscale) {
1670 window = free_space;
1671
1672 /* Advertise enough space so that it won't get scaled away.
1673 * Import case: prevent zero window announcement if
1674 * 1<<rcv_wscale > mss.
1675 */
1676 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1677 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1678 << tp->rx_opt.rcv_wscale);
1679 } else {
1680 /* Get the largest window that is a nice multiple of mss.
1681 * Window clamp already applied above.
1682 * If our current window offering is within 1 mss of the
1683 * free space we just keep it. This prevents the divide
1684 * and multiply from happening most of the time.
1685 * We also don't do any window rounding when the free space
1686 * is too small.
1687 */
1688 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001689 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07001690 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001691 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07001692 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 }
1694
1695 return window;
1696}
1697
1698/* Attempt to collapse two adjacent SKB's during retransmission. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001699static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1700 int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701{
1702 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001703 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001704 int skb_size, next_skb_size;
1705 u16 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706
1707 /* The first test we must make is that neither of these two
1708 * SKB's are still referenced by someone else.
1709 */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001710 if (skb_cloned(skb) || skb_cloned(next_skb))
1711 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001713 skb_size = skb->len;
1714 next_skb_size = next_skb->len;
1715 flags = TCP_SKB_CB(skb)->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001717 /* Also punt if next skb has been SACK'd. */
1718 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1719 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001721 /* Next skb is out of window. */
1722 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1723 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001725 /* Punt if not enough space exists in the first SKB for
1726 * the data in the second, or the total combined payload
1727 * would exceed the MSS.
1728 */
1729 if ((next_skb_size > skb_tailroom(skb)) ||
1730 ((skb_size + next_skb_size) > mss_now))
1731 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001733 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07001734
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001735 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001737 /* Ok. We will be able to collapse the packet. */
1738 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001740 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1741 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001743 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1744 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001746 if (skb->ip_summed != CHECKSUM_PARTIAL)
1747 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001749 /* Update sequence range on original skb. */
1750 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001752 /* Merge over control information. */
1753 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1754 TCP_SKB_CB(skb)->flags = flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001756 /* All done, get rid of second SKB and account for it so
1757 * packet counting does not break.
1758 */
1759 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1760 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1761 tp->retrans_out -= tcp_skb_pcount(next_skb);
1762 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1763 tp->lost_out -= tcp_skb_pcount(next_skb);
1764 /* Reno case is special. Sigh... */
1765 if (tcp_is_reno(tp) && tp->sacked_out)
1766 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001767
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001768 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1769 tp->packets_out -= tcp_skb_pcount(next_skb);
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001770
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001771 /* changed transmit queue under us so clear hints */
1772 tcp_clear_retrans_hints_partial(tp);
1773
1774 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775}
1776
1777/* Do a simple retransmit without using the backoff mechanisms in
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001778 * tcp_timer. This is used for path mtu discovery.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 * The socket is already locked here.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001780 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781void tcp_simple_retransmit(struct sock *sk)
1782{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001783 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 struct tcp_sock *tp = tcp_sk(sk);
1785 struct sk_buff *skb;
1786 unsigned int mss = tcp_current_mss(sk, 0);
1787 int lost = 0;
1788
David S. Millerfe067e82007-03-07 12:12:44 -08001789 tcp_for_write_queue(skb, sk) {
1790 if (skb == tcp_send_head(sk))
1791 break;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001792 if (skb->len > mss &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001793 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1794 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1796 tp->retrans_out -= tcp_skb_pcount(skb);
1797 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001798 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1800 tp->lost_out += tcp_skb_pcount(skb);
1801 lost = 1;
1802 }
1803 }
1804 }
1805
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07001806 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001807
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 if (!lost)
1809 return;
1810
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001811 if (tcp_is_reno(tp))
1812 tcp_limit_reno_sacked(tp);
1813
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001814 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001816 /* Don't muck with the congestion window here.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 * Reason is that we do not increase amount of _data_
1818 * in network, but units changed and effective
1819 * cwnd/ssthresh really reduced now.
1820 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001821 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001823 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 tp->prior_ssthresh = 0;
1825 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001826 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 }
1828 tcp_xmit_retransmit_queue(sk);
1829}
1830
1831/* This retransmits one SKB. Policy decisions and retransmit queue
1832 * state updates are done by the caller. Returns non-zero if an
1833 * error occurred which prevented the send.
1834 */
1835int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1836{
1837 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001838 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001839 unsigned int cur_mss = tcp_current_mss(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 int err;
1841
John Heffner5d424d52006-03-20 17:53:41 -08001842 /* Inconslusive MTU probe */
1843 if (icsk->icsk_mtup.probe_size) {
1844 icsk->icsk_mtup.probe_size = 0;
1845 }
1846
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001848 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 */
1850 if (atomic_read(&sk->sk_wmem_alloc) >
1851 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1852 return -EAGAIN;
1853
1854 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1855 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1856 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1858 return -ENOMEM;
1859 }
1860
1861 /* If receiver has shrunk his window, and skb is out of
1862 * new window, do not retransmit it. The exception is the
1863 * case, when window is shrunk to zero. In this case
1864 * our retransmit serves as a zero window probe.
1865 */
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001866 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1868 return -EAGAIN;
1869
1870 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001871 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873 }
1874
1875 /* Collapse two adjacent packets if worthwhile and we can. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001876 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1877 (skb->len < (cur_mss >> 1)) &&
1878 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1879 (!tcp_skb_is_last(sk, skb)) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001880 (skb_shinfo(skb)->nr_frags == 0 &&
1881 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1882 (tcp_skb_pcount(skb) == 1 &&
1883 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001884 (sysctl_tcp_retrans_collapse != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 tcp_retrans_try_collapse(sk, skb, cur_mss);
1886
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001887 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 return -EHOSTUNREACH; /* Routing failure or similar. */
1889
1890 /* Some Solaris stacks overoptimize and ignore the FIN on a
1891 * retransmit when old data is attached. So strip it off
1892 * since it is cheap to do so and saves bytes on the network.
1893 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001894 if (skb->len > 0 &&
1895 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1896 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 if (!pskb_trim(skb, 0)) {
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08001898 /* Reuse, even though it does some unnecessary work */
1899 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1900 TCP_SKB_CB(skb)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 skb->ip_summed = CHECKSUM_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 }
1903 }
1904
1905 /* Make a copy, if the first transmission SKB clone we made
1906 * is still in somebody's hands, else make a clone.
1907 */
1908 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001910 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
1912 if (err == 0) {
1913 /* Update global TCP statistics. */
1914 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1915
1916 tp->total_retrans++;
1917
1918#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001919 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 if (net_ratelimit())
1921 printk(KERN_DEBUG "retrans_out leaked.\n");
1922 }
1923#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001924 if (!tp->retrans_out)
1925 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1927 tp->retrans_out += tcp_skb_pcount(skb);
1928
1929 /* Save stamp of the first retransmit. */
1930 if (!tp->retrans_stamp)
1931 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1932
1933 tp->undo_retrans++;
1934
1935 /* snd_nxt is stored to detect loss of retransmitted segment,
1936 * see tcp_input.c tcp_sacktag_write_queue().
1937 */
1938 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1939 }
1940 return err;
1941}
1942
1943/* This gets called after a retransmit timeout, and the initially
1944 * retransmitted data is acknowledged. It tries to continue
1945 * resending the rest of the retransmit queue, until either
1946 * we've sent it all or the congestion window limit is reached.
1947 * If doing SACK, the first ACK which comes back for a timeout
1948 * based retransmit packet might feed us FACK information again.
1949 * If so, we use it to avoid unnecessarily retransmissions.
1950 */
1951void tcp_xmit_retransmit_queue(struct sock *sk)
1952{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001953 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 struct tcp_sock *tp = tcp_sk(sk);
1955 struct sk_buff *skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001956 int packet_cnt;
1957
1958 if (tp->retransmit_skb_hint) {
1959 skb = tp->retransmit_skb_hint;
1960 packet_cnt = tp->retransmit_cnt_hint;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001961 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08001962 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001963 packet_cnt = 0;
1964 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965
1966 /* First pass: retransmit lost packets. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001967 if (tp->lost_out) {
David S. Millerfe067e82007-03-07 12:12:44 -08001968 tcp_for_write_queue_from(skb, sk) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1970
David S. Millerfe067e82007-03-07 12:12:44 -08001971 if (skb == tcp_send_head(sk))
1972 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001973 /* we could do better than to assign each time */
1974 tp->retransmit_skb_hint = skb;
1975 tp->retransmit_cnt_hint = packet_cnt;
1976
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 /* Assume this retransmit will generate
1978 * only one packet for congestion window
1979 * calculation purposes. This works because
1980 * tcp_retransmit_skb() will chop up the
1981 * packet to be MSS sized and all the
1982 * packet counting works out.
1983 */
1984 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1985 return;
1986
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001987 if (sacked & TCPCB_LOST) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001988 if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001989 if (tcp_retransmit_skb(sk, skb)) {
1990 tp->retransmit_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001992 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001993 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1995 else
1996 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1997
David S. Millerfe067e82007-03-07 12:12:44 -08001998 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001999 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002000 inet_csk(sk)->icsk_rto,
2001 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 }
2003
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002004 packet_cnt += tcp_skb_pcount(skb);
2005 if (packet_cnt >= tp->lost_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 break;
2007 }
2008 }
2009 }
2010
2011 /* OK, demanded retransmission is finished. */
2012
2013 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002014 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 return;
2016
2017 /* No forward retransmissions in Reno are possible. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002018 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 return;
2020
2021 /* Yeah, we have to make difficult choice between forward transmission
2022 * and retransmission... Both ways have their merits...
2023 *
2024 * For now we do not retransmit anything, while we have some new
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002025 * segments to send. In the other cases, follow rule 3 for
2026 * NextSeg() specified in RFC3517.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 */
2028
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002029 if (tcp_may_send_now(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 return;
2031
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002032 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2033 if (!tp->sacked_out)
2034 return;
2035
2036 if (tp->forward_skb_hint)
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002037 skb = tp->forward_skb_hint;
Ilpo Järvinen539d2432007-05-27 02:03:20 -07002038 else
David S. Millerfe067e82007-03-07 12:12:44 -08002039 skb = tcp_write_queue_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040
David S. Millerfe067e82007-03-07 12:12:44 -08002041 tcp_for_write_queue_from(skb, sk) {
2042 if (skb == tcp_send_head(sk))
2043 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002044 tp->forward_skb_hint = skb;
2045
Ilpo Järvinen6859d492007-12-02 00:48:06 +02002046 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 break;
2048
2049 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2050 break;
2051
2052 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2053 continue;
2054
2055 /* Ok, retransmit it. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002056 if (tcp_retransmit_skb(sk, skb)) {
2057 tp->forward_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002059 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060
David S. Millerfe067e82007-03-07 12:12:44 -08002061 if (skb == tcp_write_queue_head(sk))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002062 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2063 inet_csk(sk)->icsk_rto,
2064 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065
2066 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2067 }
2068}
2069
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070/* Send a fin. The caller locks the socket for us. This cannot be
2071 * allowed to fail queueing a FIN frame under any circumstances.
2072 */
2073void tcp_send_fin(struct sock *sk)
2074{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002075 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002076 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002078
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 /* Optimization, tack on the FIN if we have a queue of
2080 * unsent frames. But be careful about outgoing SACKS
2081 * and IP options.
2082 */
2083 mss_now = tcp_current_mss(sk, 1);
2084
David S. Millerfe067e82007-03-07 12:12:44 -08002085 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2087 TCP_SKB_CB(skb)->end_seq++;
2088 tp->write_seq++;
2089 } else {
2090 /* Socket is locked, keep trying until memory is available. */
2091 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07002092 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093 if (skb)
2094 break;
2095 yield();
2096 }
2097
2098 /* Reserve space for headers and prepare control bits. */
2099 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002101 tcp_init_nondata_skb(skb, tp->write_seq,
2102 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 tcp_queue_skb(sk, skb);
2104 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002105 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106}
2107
2108/* We get here when a process closes a file descriptor (either due to
2109 * an explicit close() or as a byproduct of exit()'ing) and there
2110 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002111 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 */
Al Virodd0fc662005-10-07 07:46:04 +01002113void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 struct sk_buff *skb;
2116
2117 /* NOTE: No TCP options attached and we never retransmit this. */
2118 skb = alloc_skb(MAX_TCP_HEADER, priority);
2119 if (!skb) {
2120 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2121 return;
2122 }
2123
2124 /* Reserve space for headers and prepare control bits. */
2125 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002126 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2127 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 /* Send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002130 if (tcp_transmit_skb(sk, skb, 0, priority))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2132}
2133
2134/* WARNING: This routine must only be called when we have already sent
2135 * a SYN packet that crossed the incoming SYN that caused this routine
2136 * to get called. If this assumption fails then the initial rcv_wnd
2137 * and rcv_wscale values will not be correct.
2138 */
2139int tcp_send_synack(struct sock *sk)
2140{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002141 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142
David S. Millerfe067e82007-03-07 12:12:44 -08002143 skb = tcp_write_queue_head(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002144 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2146 return -EFAULT;
2147 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002148 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 if (skb_cloned(skb)) {
2150 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2151 if (nskb == NULL)
2152 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002153 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002155 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002156 sk_wmem_free_skb(sk, skb);
2157 sk->sk_wmem_queued += nskb->truesize;
2158 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 skb = nskb;
2160 }
2161
2162 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2163 TCP_ECN_send_synack(tcp_sk(sk), skb);
2164 }
2165 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002166 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167}
2168
2169/*
2170 * Prepare a SYN-ACK.
2171 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002172struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2173 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002175 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 struct tcp_sock *tp = tcp_sk(sk);
2177 struct tcphdr *th;
2178 int tcp_header_size;
2179 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002180#ifdef CONFIG_TCP_MD5SIG
2181 struct tcp_md5sig_key *md5;
2182 __u8 *md5_hash_location;
2183#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184
2185 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2186 if (skb == NULL)
2187 return NULL;
2188
2189 /* Reserve space for headers. */
2190 skb_reserve(skb, MAX_TCP_HEADER);
2191
2192 skb->dst = dst_clone(dst);
2193
2194 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002195 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2196 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002198 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002199
2200#ifdef CONFIG_TCP_MD5SIG
2201 /* Are we doing MD5 on this segment? If so - make room for it */
2202 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2203 if (md5)
2204 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2205#endif
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002206 skb_push(skb, tcp_header_size);
2207 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002209 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 memset(th, 0, sizeof(struct tcphdr));
2211 th->syn = 1;
2212 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 TCP_ECN_make_synack(req, th);
2214 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002215 th->dest = ireq->rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002216 /* Setting of flags are superfluous here for callers (and ECE is
2217 * not even correctly set)
2218 */
2219 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2220 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002222 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002224 __u8 rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 /* Set this up on the first call only */
2226 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2227 /* tcp_full_space because it is guaranteed to be the first packet */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002228 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002229 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 &req->rcv_wnd,
2231 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002232 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 &rcv_wscale);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002234 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 }
2236
2237 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002238 th->window = htons(min(req->rcv_wnd, 65535U));
Florian Westphal4dfc2812008-04-10 03:12:40 -07002239#ifdef CONFIG_SYN_COOKIES
2240 if (unlikely(req->cookie_ts))
2241 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2242 else
2243#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Al Virodf7a3b02006-09-27 18:38:52 -07002245 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002246 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247 TCP_SKB_CB(skb)->when,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002248 req->ts_recent,
2249 (
2250#ifdef CONFIG_TCP_MD5SIG
2251 md5 ? &md5_hash_location :
2252#endif
2253 NULL)
2254 );
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 th->doff = (tcp_header_size >> 2);
2257 TCP_INC_STATS(TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002258
2259#ifdef CONFIG_TCP_MD5SIG
2260 /* Okay, we have all we need - do the md5 hash if needed */
2261 if (md5) {
2262 tp->af_specific->calc_md5_hash(md5_hash_location,
2263 md5,
2264 NULL, dst, req,
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002265 tcp_hdr(skb), sk->sk_protocol,
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002266 skb->len);
2267 }
2268#endif
2269
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 return skb;
2271}
2272
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002273/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 * Do all connect socket setups that can be done AF independent.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002275 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002276static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277{
2278 struct dst_entry *dst = __sk_dst_get(sk);
2279 struct tcp_sock *tp = tcp_sk(sk);
2280 __u8 rcv_wscale;
2281
2282 /* We'll fix this up when we get a response from the other end.
2283 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2284 */
2285 tp->tcp_header_len = sizeof(struct tcphdr) +
2286 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2287
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002288#ifdef CONFIG_TCP_MD5SIG
2289 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2290 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2291#endif
2292
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 /* If user gave his TCP_MAXSEG, record it to clamp */
2294 if (tp->rx_opt.user_mss)
2295 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2296 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002297 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 tcp_sync_mss(sk, dst_mtu(dst));
2299
2300 if (!tp->window_clamp)
2301 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2302 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2303 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304
2305 tcp_select_initial_window(tcp_full_space(sk),
2306 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2307 &tp->rcv_wnd,
2308 &tp->window_clamp,
2309 sysctl_tcp_window_scaling,
2310 &rcv_wscale);
2311
2312 tp->rx_opt.rcv_wscale = rcv_wscale;
2313 tp->rcv_ssthresh = tp->rcv_wnd;
2314
2315 sk->sk_err = 0;
2316 sock_reset_flag(sk, SOCK_DONE);
2317 tp->snd_wnd = 0;
2318 tcp_init_wl(tp, tp->write_seq, 0);
2319 tp->snd_una = tp->write_seq;
2320 tp->snd_sml = tp->write_seq;
2321 tp->rcv_nxt = 0;
2322 tp->rcv_wup = 0;
2323 tp->copied_seq = 0;
2324
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002325 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2326 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 tcp_clear_retrans(tp);
2328}
2329
2330/*
2331 * Build a SYN and send it off.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002332 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333int tcp_connect(struct sock *sk)
2334{
2335 struct tcp_sock *tp = tcp_sk(sk);
2336 struct sk_buff *buff;
2337
2338 tcp_connect_init(sk);
2339
David S. Millerd179cd12005-08-17 14:57:30 -07002340 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 if (unlikely(buff == NULL))
2342 return -ENOBUFS;
2343
2344 /* Reserve space for headers. */
2345 skb_reserve(buff, MAX_TCP_HEADER);
2346
Wei Yongjunbd37a082006-08-07 21:04:15 -07002347 tp->snd_nxt = tp->write_seq;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002348 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2349 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350
2351 /* Send it off. */
2352 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2353 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2354 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002355 __tcp_add_write_queue_tail(sk, buff);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002356 sk->sk_wmem_queued += buff->truesize;
2357 sk_mem_charge(sk, buff->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 tp->packets_out += tcp_skb_pcount(buff);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002359 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002360
2361 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2362 * in order to make this packet get counted in tcpOutSegs.
2363 */
2364 tp->snd_nxt = tp->write_seq;
2365 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2367
2368 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002369 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2370 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 return 0;
2372}
2373
2374/* Send out a delayed ack, the caller does the policy checking
2375 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2376 * for details.
2377 */
2378void tcp_send_delayed_ack(struct sock *sk)
2379{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002380 struct inet_connection_sock *icsk = inet_csk(sk);
2381 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 unsigned long timeout;
2383
2384 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002385 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002386 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002388 if (icsk->icsk_ack.pingpong ||
2389 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 max_ato = TCP_DELACK_MAX;
2391
2392 /* Slow path, intersegment interval is "high". */
2393
2394 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002395 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 * directly.
2397 */
2398 if (tp->srtt) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002399 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400
2401 if (rtt < max_ato)
2402 max_ato = rtt;
2403 }
2404
2405 ato = min(ato, max_ato);
2406 }
2407
2408 /* Stay within the limit we were given */
2409 timeout = jiffies + ato;
2410
2411 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002412 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 /* If delack timer was blocked or is about to expire,
2414 * send ACK now.
2415 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002416 if (icsk->icsk_ack.blocked ||
2417 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 tcp_send_ack(sk);
2419 return;
2420 }
2421
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002422 if (!time_before(timeout, icsk->icsk_ack.timeout))
2423 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002425 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2426 icsk->icsk_ack.timeout = timeout;
2427 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428}
2429
2430/* This routine sends an ack and also updates the window. */
2431void tcp_send_ack(struct sock *sk)
2432{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002433 struct sk_buff *buff;
2434
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002436 if (sk->sk_state == TCP_CLOSE)
2437 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002439 /* We are not putting this on the write queue, so
2440 * tcp_transmit_skb() will set the ownership to this
2441 * sock.
2442 */
2443 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2444 if (buff == NULL) {
2445 inet_csk_schedule_ack(sk);
2446 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2447 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2448 TCP_DELACK_MAX, TCP_RTO_MAX);
2449 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002451
2452 /* Reserve space for headers and prepare control bits. */
2453 skb_reserve(buff, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002454 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002455
2456 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002457 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2458 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459}
2460
2461/* This routine sends a packet with an out of date sequence
2462 * number. It assumes the other end will try to ack it.
2463 *
2464 * Question: what should we make while urgent mode?
2465 * 4.4BSD forces sending single byte of data. We cannot send
2466 * out of window data, because we have SND.NXT==SND.MAX...
2467 *
2468 * Current solution: to send TWO zero-length segments in urgent mode:
2469 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2470 * out-of-date with SND.UNA-1 to probe window.
2471 */
2472static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2473{
2474 struct tcp_sock *tp = tcp_sk(sk);
2475 struct sk_buff *skb;
2476
2477 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2478 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002479 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 return -1;
2481
2482 /* Reserve space for headers and set control bits. */
2483 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 /* Use a previous sequence. This should cause the other
2485 * end to send an ack. Don't queue or clone SKB, just
2486 * send it.
2487 */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002488 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002490 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491}
2492
2493int tcp_write_wakeup(struct sock *sk)
2494{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002495 struct tcp_sock *tp = tcp_sk(sk);
2496 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002498 if (sk->sk_state == TCP_CLOSE)
2499 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002501 if ((skb = tcp_send_head(sk)) != NULL &&
2502 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2503 int err;
2504 unsigned int mss = tcp_current_mss(sk, 0);
2505 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002507 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2508 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002510 /* We are probing the opening of a window
2511 * but the window size is != 0
2512 * must have been a result SWS avoidance ( sender )
2513 */
2514 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2515 skb->len > mss) {
2516 seg_size = min(seg_size, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002518 if (tcp_fragment(sk, skb, seg_size, mss))
2519 return -1;
2520 } else if (!tcp_skb_pcount(skb))
2521 tcp_set_skb_tso_segs(sk, skb, mss);
2522
2523 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2524 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2525 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2526 if (!err)
2527 tcp_event_new_data_sent(sk, skb);
2528 return err;
2529 } else {
2530 if (tp->urg_mode &&
2531 between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2532 tcp_xmit_probe_skb(sk, 1);
2533 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535}
2536
2537/* A window probe timeout has occurred. If window is not closed send
2538 * a partial packet else a zero probe.
2539 */
2540void tcp_send_probe0(struct sock *sk)
2541{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002542 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543 struct tcp_sock *tp = tcp_sk(sk);
2544 int err;
2545
2546 err = tcp_write_wakeup(sk);
2547
David S. Millerfe067e82007-03-07 12:12:44 -08002548 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002550 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002551 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 return;
2553 }
2554
2555 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002556 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2557 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002558 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002559 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002560 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2561 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 } else {
2563 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002564 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 * Let local senders to fight for local resources.
2566 *
2567 * Use accumulated backoff yet.
2568 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002569 if (!icsk->icsk_probes_out)
2570 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002571 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002572 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002573 TCP_RESOURCE_PROBE_INTERVAL),
2574 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575 }
2576}
2577
Glenn Griffinc6aefaf2008-02-07 21:49:26 -08002578EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579EXPORT_SYMBOL(tcp_connect);
2580EXPORT_SYMBOL(tcp_make_synack);
2581EXPORT_SYMBOL(tcp_simple_retransmit);
2582EXPORT_SYMBOL(tcp_sync_mss);
John Heffner5d424d52006-03-20 17:53:41 -08002583EXPORT_SYMBOL(tcp_mtup_init);