blob: d6e3d269e90611250fabaea9005d0dc3abc3dea3 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
9 *
Jesper Juhl02c30a82005-05-05 16:16:16 -070010 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
21 */
22
23/*
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
27 * : AF independence
28 *
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
36 *
37 */
38
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
42#include <linux/module.h>
43#include <linux/smp_lock.h>
44
45/* People can turn this off for buggy TCP's found in printers etc. */
46int sysctl_tcp_retrans_collapse = 1;
47
48/* This limits the percentage of the congestion window which we
49 * will allow a single TSO frame to consume. Building TSO frames
50 * which are too large can cause TCP streams to be bursty.
51 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -070052int sysctl_tcp_tso_win_divisor = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
54static inline void update_send_head(struct sock *sk, struct tcp_sock *tp,
55 struct sk_buff *skb)
56{
57 sk->sk_send_head = skb->next;
58 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
59 sk->sk_send_head = NULL;
60 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
61 tcp_packets_out_inc(sk, tp, skb);
62}
63
64/* SND.NXT, if window was not shrunk.
65 * If window has been shrunk, what should we make? It is not clear at all.
66 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
67 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
68 * invalid. OK, let's make this for now:
69 */
70static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
71{
72 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
73 return tp->snd_nxt;
74 else
75 return tp->snd_una+tp->snd_wnd;
76}
77
78/* Calculate mss to advertise in SYN segment.
79 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
80 *
81 * 1. It is independent of path mtu.
82 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
83 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
84 * attached devices, because some buggy hosts are confused by
85 * large MSS.
86 * 4. We do not make 3, we advertise MSS, calculated from first
87 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
88 * This may be overridden via information stored in routing table.
89 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
90 * probably even Jumbo".
91 */
92static __u16 tcp_advertise_mss(struct sock *sk)
93{
94 struct tcp_sock *tp = tcp_sk(sk);
95 struct dst_entry *dst = __sk_dst_get(sk);
96 int mss = tp->advmss;
97
98 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
99 mss = dst_metric(dst, RTAX_ADVMSS);
100 tp->advmss = mss;
101 }
102
103 return (__u16)mss;
104}
105
106/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
107 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700108static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700110 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 s32 delta = tcp_time_stamp - tp->lsndtime;
112 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
113 u32 cwnd = tp->snd_cwnd;
114
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300115 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300117 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118 restart_cwnd = min(restart_cwnd, cwnd);
119
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700120 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121 cwnd >>= 1;
122 tp->snd_cwnd = max(cwnd, restart_cwnd);
123 tp->snd_cwnd_stamp = tcp_time_stamp;
124 tp->snd_cwnd_used = 0;
125}
126
127static inline void tcp_event_data_sent(struct tcp_sock *tp,
128 struct sk_buff *skb, struct sock *sk)
129{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700130 struct inet_connection_sock *icsk = inet_csk(sk);
131 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700133 if (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)
134 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135
136 tp->lsndtime = now;
137
138 /* If it is a reply for ato after last received
139 * packet, enter pingpong mode.
140 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700141 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
142 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143}
144
David S. Millerfc6415b2005-07-05 15:17:45 -0700145static __inline__ void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700147 tcp_dec_quickack_mode(sk, pkts);
148 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149}
150
151/* Determine a window scaling and initial window to offer.
152 * Based on the assumption that the given amount of space
153 * will be offered. Store the results in the tp structure.
154 * NOTE: for smooth operation initial space offering should
155 * be a multiple of mss if possible. We assume here that mss >= 1.
156 * This MUST be enforced by all callers.
157 */
158void tcp_select_initial_window(int __space, __u32 mss,
159 __u32 *rcv_wnd, __u32 *window_clamp,
160 int wscale_ok, __u8 *rcv_wscale)
161{
162 unsigned int space = (__space < 0 ? 0 : __space);
163
164 /* If no clamp set the clamp to the max possible scaled window */
165 if (*window_clamp == 0)
166 (*window_clamp) = (65535 << 14);
167 space = min(*window_clamp, space);
168
169 /* Quantize space offering to a multiple of mss if possible. */
170 if (space > mss)
171 space = (space / mss) * mss;
172
173 /* NOTE: offering an initial window larger than 32767
174 * will break some buggy TCP stacks. We try to be nice.
175 * If we are not window scaling, then this truncates
176 * our initial window offering to 32k. There should also
177 * be a sysctl option to stop being nice.
178 */
179 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
180 (*rcv_wscale) = 0;
181 if (wscale_ok) {
182 /* Set window scaling on max possible window
183 * See RFC1323 for an explanation of the limit to 14
184 */
185 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
186 while (space > 65535 && (*rcv_wscale) < 14) {
187 space >>= 1;
188 (*rcv_wscale)++;
189 }
190 }
191
192 /* Set initial window to value enough for senders,
193 * following RFC1414. Senders, not following this RFC,
194 * will be satisfied with 2.
195 */
196 if (mss > (1<<*rcv_wscale)) {
197 int init_cwnd = 4;
198 if (mss > 1460*3)
199 init_cwnd = 2;
200 else if (mss > 1460)
201 init_cwnd = 3;
202 if (*rcv_wnd > init_cwnd*mss)
203 *rcv_wnd = init_cwnd*mss;
204 }
205
206 /* Set the clamp no higher than max representable value */
207 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
208}
209
210/* Chose a new window to advertise, update state in tcp_sock for the
211 * socket, and return result with RFC1323 scaling applied. The return
212 * value can be stuffed directly into th->window for an outgoing
213 * frame.
214 */
215static __inline__ u16 tcp_select_window(struct sock *sk)
216{
217 struct tcp_sock *tp = tcp_sk(sk);
218 u32 cur_win = tcp_receive_window(tp);
219 u32 new_win = __tcp_select_window(sk);
220
221 /* Never shrink the offered window */
222 if(new_win < cur_win) {
223 /* Danger Will Robinson!
224 * Don't update rcv_wup/rcv_wnd here or else
225 * we will not be able to advertise a zero
226 * window in time. --DaveM
227 *
228 * Relax Will Robinson.
229 */
230 new_win = cur_win;
231 }
232 tp->rcv_wnd = new_win;
233 tp->rcv_wup = tp->rcv_nxt;
234
235 /* Make sure we do not exceed the maximum possible
236 * scaled window.
237 */
238 if (!tp->rx_opt.rcv_wscale)
239 new_win = min(new_win, MAX_TCP_WINDOW);
240 else
241 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
242
243 /* RFC1323 scaling applied */
244 new_win >>= tp->rx_opt.rcv_wscale;
245
246 /* If we advertise zero window, disable fast path. */
247 if (new_win == 0)
248 tp->pred_flags = 0;
249
250 return new_win;
251}
252
253
254/* This routine actually transmits TCP packets queued in by
255 * tcp_do_sendmsg(). This is used by both the initial
256 * transmission and possible later retransmissions.
257 * All SKB's seen here are completely headerless. It is our
258 * job to build the TCP header, and pass the packet down to
259 * IP so it can do the same plus pass the packet off to the
260 * device.
261 *
262 * We are working here with either a clone of the original
263 * SKB, or a fresh unique copy made by the retransmit engine.
264 */
265static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb)
266{
267 if (skb != NULL) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300268 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 struct inet_sock *inet = inet_sk(sk);
270 struct tcp_sock *tp = tcp_sk(sk);
271 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
272 int tcp_header_size = tp->tcp_header_len;
273 struct tcphdr *th;
274 int sysctl_flags;
275 int err;
276
277 BUG_ON(!tcp_skb_pcount(skb));
278
279#define SYSCTL_FLAG_TSTAMPS 0x1
280#define SYSCTL_FLAG_WSCALE 0x2
281#define SYSCTL_FLAG_SACK 0x4
282
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700283 /* If congestion control is doing timestamping */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300284 if (icsk->icsk_ca_ops->rtt_sample)
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700285 __net_timestamp(skb);
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700286
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 sysctl_flags = 0;
288 if (tcb->flags & TCPCB_FLAG_SYN) {
289 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
290 if(sysctl_tcp_timestamps) {
291 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
292 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
293 }
294 if(sysctl_tcp_window_scaling) {
295 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
296 sysctl_flags |= SYSCTL_FLAG_WSCALE;
297 }
298 if(sysctl_tcp_sack) {
299 sysctl_flags |= SYSCTL_FLAG_SACK;
300 if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
301 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
302 }
303 } else if (tp->rx_opt.eff_sacks) {
304 /* A SACK is 2 pad bytes, a 2 byte header, plus
305 * 2 32-bit sequence numbers for each SACK block.
306 */
307 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
308 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
309 }
310
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700311 if (tcp_packets_in_flight(tp) == 0)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300312 tcp_ca_event(sk, CA_EVENT_TX_START);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
314 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
315 skb->h.th = th;
316 skb_set_owner_w(skb, sk);
317
318 /* Build TCP header and checksum it. */
319 th->source = inet->sport;
320 th->dest = inet->dport;
321 th->seq = htonl(tcb->seq);
322 th->ack_seq = htonl(tp->rcv_nxt);
323 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags);
324 if (tcb->flags & TCPCB_FLAG_SYN) {
325 /* RFC1323: The window in SYN & SYN/ACK segments
326 * is never scaled.
327 */
328 th->window = htons(tp->rcv_wnd);
329 } else {
330 th->window = htons(tcp_select_window(sk));
331 }
332 th->check = 0;
333 th->urg_ptr = 0;
334
335 if (tp->urg_mode &&
336 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) {
337 th->urg_ptr = htons(tp->snd_up-tcb->seq);
338 th->urg = 1;
339 }
340
341 if (tcb->flags & TCPCB_FLAG_SYN) {
342 tcp_syn_build_options((__u32 *)(th + 1),
343 tcp_advertise_mss(sk),
344 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
345 (sysctl_flags & SYSCTL_FLAG_SACK),
346 (sysctl_flags & SYSCTL_FLAG_WSCALE),
347 tp->rx_opt.rcv_wscale,
348 tcb->when,
349 tp->rx_opt.ts_recent);
350 } else {
351 tcp_build_and_update_options((__u32 *)(th + 1),
352 tp, tcb->when);
353
354 TCP_ECN_send(sk, tp, skb, tcp_header_size);
355 }
356 tp->af_specific->send_check(sk, th, skb->len, skb);
357
358 if (tcb->flags & TCPCB_FLAG_ACK)
David S. Millerfc6415b2005-07-05 15:17:45 -0700359 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360
361 if (skb->len != tcp_header_size)
362 tcp_event_data_sent(tp, skb, sk);
363
364 TCP_INC_STATS(TCP_MIB_OUTSEGS);
365
366 err = tp->af_specific->queue_xmit(skb, 0);
367 if (err <= 0)
368 return err;
369
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300370 tcp_enter_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371
372 /* NET_XMIT_CN is special. It does not guarantee,
373 * that this packet is lost. It tells that device
374 * is about to start to drop packets or already
375 * drops some packets of the same priority and
376 * invokes us to send less aggressively.
377 */
378 return err == NET_XMIT_CN ? 0 : err;
379 }
380 return -ENOBUFS;
381#undef SYSCTL_FLAG_TSTAMPS
382#undef SYSCTL_FLAG_WSCALE
383#undef SYSCTL_FLAG_SACK
384}
385
386
387/* This routine just queue's the buffer
388 *
389 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
390 * otherwise socket can stall.
391 */
392static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
393{
394 struct tcp_sock *tp = tcp_sk(sk);
395
396 /* Advance write_seq and place onto the write_queue. */
397 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
398 skb_header_release(skb);
399 __skb_queue_tail(&sk->sk_write_queue, skb);
400 sk_charge_skb(sk, skb);
401
402 /* Queue it, remembering where we must start sending. */
403 if (sk->sk_send_head == NULL)
404 sk->sk_send_head = skb;
405}
406
David S. Miller846998a2005-08-04 19:52:01 -0700407static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700408{
David S. Miller846998a2005-08-04 19:52:01 -0700409 if (skb->len <= mss_now ||
David S. Millerf6302d12005-07-05 15:18:03 -0700410 !(sk->sk_route_caps & NETIF_F_TSO)) {
411 /* Avoid the costly divide in the normal
412 * non-TSO case.
413 */
414 skb_shinfo(skb)->tso_segs = 1;
415 skb_shinfo(skb)->tso_size = 0;
416 } else {
417 unsigned int factor;
418
David S. Miller846998a2005-08-04 19:52:01 -0700419 factor = skb->len + (mss_now - 1);
420 factor /= mss_now;
David S. Millerf6302d12005-07-05 15:18:03 -0700421 skb_shinfo(skb)->tso_segs = factor;
David S. Miller846998a2005-08-04 19:52:01 -0700422 skb_shinfo(skb)->tso_size = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 }
424}
425
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426/* Function to create two new TCP segments. Shrinks the given segment
427 * to the specified size and appends a new segment with the rest of the
428 * packet to the list. This won't be called frequently, I hope.
429 * Remember, these are still headerless SKBs at this point.
430 */
David S. Miller6475be12005-09-01 22:47:01 -0700431int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
433 struct tcp_sock *tp = tcp_sk(sk);
434 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700435 int nsize, old_factor;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 u16 flags;
437
Herbert Xu3c05d922005-09-14 20:50:35 -0700438 BUG_ON(len >= skb->len);
439
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 nsize = skb_headlen(skb) - len;
441 if (nsize < 0)
442 nsize = 0;
443
444 if (skb_cloned(skb) &&
445 skb_is_nonlinear(skb) &&
446 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
447 return -ENOMEM;
448
449 /* Get a new skb... force flag on. */
450 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
451 if (buff == NULL)
452 return -ENOMEM; /* We'll just try again later. */
453 sk_charge_skb(sk, buff);
454
455 /* Correct the sequence numbers. */
456 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
457 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
458 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
459
460 /* PSH and FIN should only be set in the second packet. */
461 flags = TCP_SKB_CB(skb)->flags;
462 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
463 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700464 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
466
467 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
468 /* Copy and checksum data tail into the new buffer. */
469 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
470 nsize, 0);
471
472 skb_trim(skb, len);
473
474 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
475 } else {
476 skb->ip_summed = CHECKSUM_HW;
477 skb_split(skb, buff, len);
478 }
479
480 buff->ip_summed = skb->ip_summed;
481
482 /* Looks stupid, but our code really uses when of
483 * skbs, which it never sent before. --ANK
484 */
485 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700486 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487
David S. Miller6475be12005-09-01 22:47:01 -0700488 old_factor = tcp_skb_pcount(skb);
489
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700491 tcp_set_skb_tso_segs(sk, skb, mss_now);
492 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493
David S. Miller6475be12005-09-01 22:47:01 -0700494 /* If this packet has been sent out already, we must
495 * adjust the various packet counters.
496 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700497 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700498 int diff = old_factor - tcp_skb_pcount(skb) -
499 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500
David S. Miller6475be12005-09-01 22:47:01 -0700501 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700502
503 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
504 tp->sacked_out -= diff;
505 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
506 tp->retrans_out -= diff;
507
David S. Miller6475be12005-09-01 22:47:01 -0700508 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
509 tp->lost_out -= diff;
510 tp->left_out -= diff;
511 }
Herbert Xu83ca28b2005-09-22 23:32:56 -0700512
David S. Miller6475be12005-09-01 22:47:01 -0700513 if (diff > 0) {
Herbert Xu83ca28b2005-09-22 23:32:56 -0700514 /* Adjust Reno SACK estimate. */
515 if (!tp->rx_opt.sack_ok) {
516 tp->sacked_out -= diff;
517 if ((int)tp->sacked_out < 0)
518 tp->sacked_out = 0;
519 tcp_sync_left_out(tp);
520 }
521
David S. Miller6475be12005-09-01 22:47:01 -0700522 tp->fackets_out -= diff;
523 if ((int)tp->fackets_out < 0)
524 tp->fackets_out = 0;
525 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 }
527
528 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700529 skb_header_release(buff);
David S. Miller8728b832005-08-09 19:25:21 -0700530 __skb_append(skb, buff, &sk->sk_write_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531
532 return 0;
533}
534
535/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
536 * eventually). The difference is that pulled data not copied, but
537 * immediately discarded.
538 */
539static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
540{
541 int i, k, eat;
542
543 eat = len;
544 k = 0;
545 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
546 if (skb_shinfo(skb)->frags[i].size <= eat) {
547 put_page(skb_shinfo(skb)->frags[i].page);
548 eat -= skb_shinfo(skb)->frags[i].size;
549 } else {
550 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
551 if (eat) {
552 skb_shinfo(skb)->frags[k].page_offset += eat;
553 skb_shinfo(skb)->frags[k].size -= eat;
554 eat = 0;
555 }
556 k++;
557 }
558 }
559 skb_shinfo(skb)->nr_frags = k;
560
561 skb->tail = skb->data;
562 skb->data_len -= len;
563 skb->len = skb->data_len;
564 return skb->tail;
565}
566
567int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
568{
569 if (skb_cloned(skb) &&
570 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
571 return -ENOMEM;
572
573 if (len <= skb_headlen(skb)) {
574 __skb_pull(skb, len);
575 } else {
576 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
577 return -ENOMEM;
578 }
579
580 TCP_SKB_CB(skb)->seq += len;
581 skb->ip_summed = CHECKSUM_HW;
582
583 skb->truesize -= len;
584 sk->sk_wmem_queued -= len;
585 sk->sk_forward_alloc += len;
586 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
587
588 /* Any change of skb->len requires recalculation of tso
589 * factor and mss.
590 */
591 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700592 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
594 return 0;
595}
596
597/* This function synchronize snd mss to current pmtu/exthdr set.
598
599 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
600 for TCP options, but includes only bare TCP header.
601
602 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
603 It is minumum of user_mss and mss received with SYN.
604 It also does not include TCP options.
605
606 tp->pmtu_cookie is last pmtu, seen by this function.
607
608 tp->mss_cache is current effective sending mss, including
609 all tcp options except for SACKs. It is evaluated,
610 taking into account current pmtu, but never exceeds
611 tp->rx_opt.mss_clamp.
612
613 NOTE1. rfc1122 clearly states that advertised MSS
614 DOES NOT include either tcp or ip options.
615
616 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
617 this function. --ANK (980731)
618 */
619
620unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
621{
622 struct tcp_sock *tp = tcp_sk(sk);
623 int mss_now;
624
625 /* Calculate base mss without TCP options:
626 It is MMS_S - sizeof(tcphdr) of rfc1122
627 */
628 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
629
630 /* Clamp it (mss_clamp does not include tcp options) */
631 if (mss_now > tp->rx_opt.mss_clamp)
632 mss_now = tp->rx_opt.mss_clamp;
633
634 /* Now subtract optional transport overhead */
635 mss_now -= tp->ext_header_len;
636
637 /* Then reserve room for full set of TCP options and 8 bytes of data */
638 if (mss_now < 48)
639 mss_now = 48;
640
641 /* Now subtract TCP options size, not including SACKs */
642 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
643
644 /* Bound mss with half of window */
645 if (tp->max_window && mss_now > (tp->max_window>>1))
646 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
647
648 /* And store cached results */
649 tp->pmtu_cookie = pmtu;
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700650 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
652 return mss_now;
653}
654
655/* Compute the current effective MSS, taking SACKs and IP options,
656 * and even PMTU discovery events into account.
657 *
658 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
659 * cannot be large. However, taking into account rare use of URG, this
660 * is not a big flaw.
661 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700662unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663{
664 struct tcp_sock *tp = tcp_sk(sk);
665 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700666 u32 mss_now;
667 u16 xmit_size_goal;
668 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700670 mss_now = tp->mss_cache;
671
672 if (large_allowed &&
673 (sk->sk_route_caps & NETIF_F_TSO) &&
674 !tp->urg_mode)
675 doing_tso = 1;
676
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 if (dst) {
678 u32 mtu = dst_mtu(dst);
679 if (mtu != tp->pmtu_cookie)
680 mss_now = tcp_sync_mss(sk, mtu);
681 }
682
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 if (tp->rx_opt.eff_sacks)
684 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
685 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700686
687 xmit_size_goal = mss_now;
688
689 if (doing_tso) {
690 xmit_size_goal = 65535 -
691 tp->af_specific->net_header_len -
692 tp->ext_header_len - tp->tcp_header_len;
693
694 if (tp->max_window &&
695 (xmit_size_goal > (tp->max_window >> 1)))
696 xmit_size_goal = max((tp->max_window >> 1),
697 68U - tp->tcp_header_len);
698
699 xmit_size_goal -= (xmit_size_goal % mss_now);
700 }
701 tp->xmit_size_goal = xmit_size_goal;
702
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 return mss_now;
704}
705
David S. Millera762a982005-07-05 15:18:51 -0700706/* Congestion window validation. (RFC2861) */
707
708static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
709{
710 __u32 packets_out = tp->packets_out;
711
712 if (packets_out >= tp->snd_cwnd) {
713 /* Network is feed fully. */
714 tp->snd_cwnd_used = 0;
715 tp->snd_cwnd_stamp = tcp_time_stamp;
716 } else {
717 /* Network starves. */
718 if (tp->packets_out > tp->snd_cwnd_used)
719 tp->snd_cwnd_used = tp->packets_out;
720
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700721 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -0700722 tcp_cwnd_application_limited(sk);
723 }
724}
725
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700726static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
727{
728 u32 window, cwnd_len;
729
730 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
731 cwnd_len = mss_now * cwnd;
732 return min(window, cwnd_len);
733}
734
735/* Can at least one segment of SKB be sent right now, according to the
736 * congestion window rules? If so, return how many segments are allowed.
737 */
738static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
739{
740 u32 in_flight, cwnd;
741
742 /* Don't be strict about the congestion window for the final FIN. */
743 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
744 return 1;
745
746 in_flight = tcp_packets_in_flight(tp);
747 cwnd = tp->snd_cwnd;
748 if (in_flight < cwnd)
749 return (cwnd - in_flight);
750
751 return 0;
752}
753
754/* This must be invoked the first time we consider transmitting
755 * SKB onto the wire.
756 */
David S. Miller846998a2005-08-04 19:52:01 -0700757static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700758{
759 int tso_segs = tcp_skb_pcount(skb);
760
David S. Miller846998a2005-08-04 19:52:01 -0700761 if (!tso_segs ||
762 (tso_segs > 1 &&
763 skb_shinfo(skb)->tso_size != mss_now)) {
764 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700765 tso_segs = tcp_skb_pcount(skb);
766 }
767 return tso_segs;
768}
769
770static inline int tcp_minshall_check(const struct tcp_sock *tp)
771{
772 return after(tp->snd_sml,tp->snd_una) &&
773 !after(tp->snd_sml, tp->snd_nxt);
774}
775
776/* Return 0, if packet can be sent now without violation Nagle's rules:
777 * 1. It is full sized.
778 * 2. Or it contains FIN. (already checked by caller)
779 * 3. Or TCP_NODELAY was set.
780 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
781 * With Minshall's modification: all sent small packets are ACKed.
782 */
783
784static inline int tcp_nagle_check(const struct tcp_sock *tp,
785 const struct sk_buff *skb,
786 unsigned mss_now, int nonagle)
787{
788 return (skb->len < mss_now &&
789 ((nonagle&TCP_NAGLE_CORK) ||
790 (!nonagle &&
791 tp->packets_out &&
792 tcp_minshall_check(tp))));
793}
794
795/* Return non-zero if the Nagle test allows this packet to be
796 * sent now.
797 */
798static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
799 unsigned int cur_mss, int nonagle)
800{
801 /* Nagle rule does not apply to frames, which sit in the middle of the
802 * write_queue (they have no chances to get new data).
803 *
804 * This is implemented in the callers, where they modify the 'nonagle'
805 * argument based upon the location of SKB in the send queue.
806 */
807 if (nonagle & TCP_NAGLE_PUSH)
808 return 1;
809
810 /* Don't use the nagle rule for urgent data (or for the final FIN). */
811 if (tp->urg_mode ||
812 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
813 return 1;
814
815 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
816 return 1;
817
818 return 0;
819}
820
821/* Does at least the first segment of SKB fit into the send window? */
822static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
823{
824 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
825
826 if (skb->len > cur_mss)
827 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
828
829 return !after(end_seq, tp->snd_una + tp->snd_wnd);
830}
831
832/* This checks if the data bearing packet SKB (usually sk->sk_send_head)
833 * should be put on the wire right now. If so, it returns the number of
834 * packets allowed by the congestion window.
835 */
836static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
837 unsigned int cur_mss, int nonagle)
838{
839 struct tcp_sock *tp = tcp_sk(sk);
840 unsigned int cwnd_quota;
841
David S. Miller846998a2005-08-04 19:52:01 -0700842 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700843
844 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
845 return 0;
846
847 cwnd_quota = tcp_cwnd_test(tp, skb);
848 if (cwnd_quota &&
849 !tcp_snd_wnd_test(tp, skb, cur_mss))
850 cwnd_quota = 0;
851
852 return cwnd_quota;
853}
854
855static inline int tcp_skb_is_last(const struct sock *sk,
856 const struct sk_buff *skb)
857{
858 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
859}
860
861int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
862{
863 struct sk_buff *skb = sk->sk_send_head;
864
865 return (skb &&
866 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
867 (tcp_skb_is_last(sk, skb) ?
868 TCP_NAGLE_PUSH :
869 tp->nonagle)));
870}
871
872/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
873 * which is put after SKB on the list. It is very much like
874 * tcp_fragment() except that it may make several kinds of assumptions
875 * in order to speed up the splitting operation. In particular, we
876 * know that all the data is in scatter-gather pages, and that the
877 * packet has never been sent out before (and thus is not cloned).
878 */
David S. Miller846998a2005-08-04 19:52:01 -0700879static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700880{
881 struct sk_buff *buff;
882 int nlen = skb->len - len;
883 u16 flags;
884
885 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -0700886 if (skb->len != skb->data_len)
887 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700888
889 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
890 if (unlikely(buff == NULL))
891 return -ENOMEM;
892
893 buff->truesize = nlen;
894 skb->truesize -= nlen;
895
896 /* Correct the sequence numbers. */
897 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
898 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
899 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
900
901 /* PSH and FIN should only be set in the second packet. */
902 flags = TCP_SKB_CB(skb)->flags;
903 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
904 TCP_SKB_CB(buff)->flags = flags;
905
906 /* This packet was never sent out yet, so no SACK bits. */
907 TCP_SKB_CB(buff)->sacked = 0;
908
909 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
910 skb_split(skb, buff, len);
911
912 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700913 tcp_set_skb_tso_segs(sk, skb, mss_now);
914 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700915
916 /* Link BUFF into the send queue. */
917 skb_header_release(buff);
David S. Miller8728b832005-08-09 19:25:21 -0700918 __skb_append(skb, buff, &sk->sk_write_queue);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700919
920 return 0;
921}
922
923/* Try to defer sending, if possible, in order to minimize the amount
924 * of TSO splitting we do. View it as a kind of TSO Nagle test.
925 *
926 * This algorithm is from John Heffner.
927 */
928static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
929{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300930 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700931 u32 send_win, cong_win, limit, in_flight;
932
933 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
934 return 0;
935
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300936 if (icsk->icsk_ca_state != TCP_CA_Open)
David S. Miller908a75c2005-07-05 15:43:58 -0700937 return 0;
938
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700939 in_flight = tcp_packets_in_flight(tp);
940
941 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
942 (tp->snd_cwnd <= in_flight));
943
944 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
945
946 /* From in_flight test above, we know that cwnd > in_flight. */
947 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
948
949 limit = min(send_win, cong_win);
950
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700951 if (sysctl_tcp_tso_win_divisor) {
952 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
953
954 /* If at least some fraction of a window is available,
955 * just use it.
956 */
957 chunk /= sysctl_tcp_tso_win_divisor;
958 if (limit >= chunk)
959 return 0;
960 } else {
961 /* Different approach, try not to defer past a single
962 * ACK. Receiver should ACK every other full sized
963 * frame, so if we have space for more than 3 frames
964 * then send now.
965 */
966 if (limit > tcp_max_burst(tp) * tp->mss_cache)
967 return 0;
968 }
969
970 /* Ok, it looks like it is advisable to defer. */
971 return 1;
972}
973
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974/* This routine writes packets to the network. It advances the
975 * send_head. This happens as incoming acks open up the remote
976 * window for us.
977 *
978 * Returns 1, if no segments are in flight and we have queued segments, but
979 * cannot send anything now because of SWS or another problem.
980 */
David S. Millera2e2a592005-07-05 15:19:23 -0700981static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982{
983 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -0700984 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700985 unsigned int tso_segs, sent_pkts;
986 int cwnd_quota;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987
988 /* If we are closed, the bytes will have to remain here.
989 * In time closedown will finish, we empty the write queue and all
990 * will be happy.
991 */
David S. Miller92df7b52005-07-05 15:19:06 -0700992 if (unlikely(sk->sk_state == TCP_CLOSE))
993 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994
David S. Miller92df7b52005-07-05 15:19:06 -0700995 sent_pkts = 0;
Herbert Xub68e9f82005-08-04 19:52:02 -0700996 while ((skb = sk->sk_send_head)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -0700997 unsigned int limit;
998
Herbert Xub68e9f82005-08-04 19:52:02 -0700999 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001000 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001001
Herbert Xub68e9f82005-08-04 19:52:02 -07001002 cwnd_quota = tcp_cwnd_test(tp, skb);
1003 if (!cwnd_quota)
1004 break;
1005
1006 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1007 break;
1008
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001009 if (tso_segs == 1) {
1010 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1011 (tcp_skb_is_last(sk, skb) ?
1012 nonagle : TCP_NAGLE_PUSH))))
1013 break;
1014 } else {
1015 if (tcp_tso_should_defer(sk, tp, skb))
1016 break;
1017 }
David S. Milleraa934662005-07-05 15:20:09 -07001018
Herbert Xuc8ac3772005-08-16 20:43:40 -07001019 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001020 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001021 limit = tcp_window_allows(tp, skb,
1022 mss_now, cwnd_quota);
David S. Milleraa934662005-07-05 15:20:09 -07001023
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001024 if (skb->len < limit) {
1025 unsigned int trim = skb->len % mss_now;
1026
1027 if (trim)
1028 limit = skb->len - trim;
1029 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 }
1031
Herbert Xuc8ac3772005-08-16 20:43:40 -07001032 if (skb->len > limit &&
1033 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1034 break;
1035
David S. Miller92df7b52005-07-05 15:19:06 -07001036 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001037
David S. Milleraa934662005-07-05 15:20:09 -07001038 if (unlikely(tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))))
David S. Miller92df7b52005-07-05 15:19:06 -07001039 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040
David S. Miller92df7b52005-07-05 15:19:06 -07001041 /* Advance the send_head. This one is sent out.
1042 * This call will increment packets_out.
1043 */
1044 update_send_head(sk, tp, skb);
1045
1046 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001047 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 }
David S. Miller92df7b52005-07-05 15:19:06 -07001049
David S. Milleraa934662005-07-05 15:20:09 -07001050 if (likely(sent_pkts)) {
David S. Miller92df7b52005-07-05 15:19:06 -07001051 tcp_cwnd_validate(sk, tp);
1052 return 0;
1053 }
David S. Miller92df7b52005-07-05 15:19:06 -07001054 return !tp->packets_out && sk->sk_send_head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055}
1056
David S. Millera762a982005-07-05 15:18:51 -07001057/* Push out any pending frames which were held back due to
1058 * TCP_CORK or attempt at coalescing tiny packets.
1059 * The socket must be locked by the caller.
1060 */
1061void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
David S. Millera2e2a592005-07-05 15:19:23 -07001062 unsigned int cur_mss, int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001063{
1064 struct sk_buff *skb = sk->sk_send_head;
1065
1066 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001067 if (tcp_write_xmit(sk, cur_mss, nonagle))
David S. Millera762a982005-07-05 15:18:51 -07001068 tcp_check_probe_timer(sk, tp);
1069 }
1070}
1071
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001072/* Send _single_ skb sitting at the send head. This function requires
1073 * true push pending frames to setup probe timer etc.
1074 */
1075void tcp_push_one(struct sock *sk, unsigned int mss_now)
1076{
1077 struct tcp_sock *tp = tcp_sk(sk);
1078 struct sk_buff *skb = sk->sk_send_head;
1079 unsigned int tso_segs, cwnd_quota;
1080
1081 BUG_ON(!skb || skb->len < mss_now);
1082
David S. Miller846998a2005-08-04 19:52:01 -07001083 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001084 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1085
1086 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001087 unsigned int limit;
1088
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001089 BUG_ON(!tso_segs);
1090
Herbert Xuc8ac3772005-08-16 20:43:40 -07001091 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001092 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001093 limit = tcp_window_allows(tp, skb,
1094 mss_now, cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001095
1096 if (skb->len < limit) {
1097 unsigned int trim = skb->len % mss_now;
1098
1099 if (trim)
1100 limit = skb->len - trim;
1101 }
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001102 }
1103
Herbert Xuc8ac3772005-08-16 20:43:40 -07001104 if (skb->len > limit &&
1105 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1106 return;
1107
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001108 /* Send it out now. */
1109 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1110
1111 if (likely(!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation)))) {
1112 update_send_head(sk, tp, skb);
1113 tcp_cwnd_validate(sk, tp);
1114 return;
1115 }
1116 }
1117}
1118
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119/* This function returns the amount that we can raise the
1120 * usable window based on the following constraints
1121 *
1122 * 1. The window can never be shrunk once it is offered (RFC 793)
1123 * 2. We limit memory per socket
1124 *
1125 * RFC 1122:
1126 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1127 * RECV.NEXT + RCV.WIN fixed until:
1128 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1129 *
1130 * i.e. don't raise the right edge of the window until you can raise
1131 * it at least MSS bytes.
1132 *
1133 * Unfortunately, the recommended algorithm breaks header prediction,
1134 * since header prediction assumes th->window stays fixed.
1135 *
1136 * Strictly speaking, keeping th->window fixed violates the receiver
1137 * side SWS prevention criteria. The problem is that under this rule
1138 * a stream of single byte packets will cause the right side of the
1139 * window to always advance by a single byte.
1140 *
1141 * Of course, if the sender implements sender side SWS prevention
1142 * then this will not be a problem.
1143 *
1144 * BSD seems to make the following compromise:
1145 *
1146 * If the free space is less than the 1/4 of the maximum
1147 * space available and the free space is less than 1/2 mss,
1148 * then set the window to 0.
1149 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1150 * Otherwise, just prevent the window from shrinking
1151 * and from being larger than the largest representable value.
1152 *
1153 * This prevents incremental opening of the window in the regime
1154 * where TCP is limited by the speed of the reader side taking
1155 * data out of the TCP receive queue. It does nothing about
1156 * those cases where the window is constrained on the sender side
1157 * because the pipeline is full.
1158 *
1159 * BSD also seems to "accidentally" limit itself to windows that are a
1160 * multiple of MSS, at least until the free space gets quite small.
1161 * This would appear to be a side effect of the mbuf implementation.
1162 * Combining these two algorithms results in the observed behavior
1163 * of having a fixed window size at almost all times.
1164 *
1165 * Below we obtain similar behavior by forcing the offered window to
1166 * a multiple of the mss when it is feasible to do so.
1167 *
1168 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1169 * Regular options like TIMESTAMP are taken into account.
1170 */
1171u32 __tcp_select_window(struct sock *sk)
1172{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001173 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 struct tcp_sock *tp = tcp_sk(sk);
1175 /* MSS for the peer's data. Previous verions used mss_clamp
1176 * here. I don't know if the value based on our guesses
1177 * of peer's MSS is better for the performance. It's more correct
1178 * but may be worse for the performance because of rcv_mss
1179 * fluctuations. --SAW 1998/11/1
1180 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001181 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 int free_space = tcp_space(sk);
1183 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1184 int window;
1185
1186 if (mss > full_space)
1187 mss = full_space;
1188
1189 if (free_space < full_space/2) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001190 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191
1192 if (tcp_memory_pressure)
1193 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1194
1195 if (free_space < mss)
1196 return 0;
1197 }
1198
1199 if (free_space > tp->rcv_ssthresh)
1200 free_space = tp->rcv_ssthresh;
1201
1202 /* Don't do rounding if we are using window scaling, since the
1203 * scaled window will not line up with the MSS boundary anyway.
1204 */
1205 window = tp->rcv_wnd;
1206 if (tp->rx_opt.rcv_wscale) {
1207 window = free_space;
1208
1209 /* Advertise enough space so that it won't get scaled away.
1210 * Import case: prevent zero window announcement if
1211 * 1<<rcv_wscale > mss.
1212 */
1213 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1214 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1215 << tp->rx_opt.rcv_wscale);
1216 } else {
1217 /* Get the largest window that is a nice multiple of mss.
1218 * Window clamp already applied above.
1219 * If our current window offering is within 1 mss of the
1220 * free space we just keep it. This prevents the divide
1221 * and multiply from happening most of the time.
1222 * We also don't do any window rounding when the free space
1223 * is too small.
1224 */
1225 if (window <= free_space - mss || window > free_space)
1226 window = (free_space/mss)*mss;
1227 }
1228
1229 return window;
1230}
1231
1232/* Attempt to collapse two adjacent SKB's during retransmission. */
1233static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1234{
1235 struct tcp_sock *tp = tcp_sk(sk);
1236 struct sk_buff *next_skb = skb->next;
1237
1238 /* The first test we must make is that neither of these two
1239 * SKB's are still referenced by someone else.
1240 */
1241 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1242 int skb_size = skb->len, next_skb_size = next_skb->len;
1243 u16 flags = TCP_SKB_CB(skb)->flags;
1244
1245 /* Also punt if next skb has been SACK'd. */
1246 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1247 return;
1248
1249 /* Next skb is out of window. */
1250 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1251 return;
1252
1253 /* Punt if not enough space exists in the first SKB for
1254 * the data in the second, or the total combined payload
1255 * would exceed the MSS.
1256 */
1257 if ((next_skb_size > skb_tailroom(skb)) ||
1258 ((skb_size + next_skb_size) > mss_now))
1259 return;
1260
1261 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1262 tcp_skb_pcount(next_skb) != 1);
1263
1264 /* Ok. We will be able to collapse the packet. */
David S. Miller8728b832005-08-09 19:25:21 -07001265 __skb_unlink(next_skb, &sk->sk_write_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266
1267 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1268
1269 if (next_skb->ip_summed == CHECKSUM_HW)
1270 skb->ip_summed = CHECKSUM_HW;
1271
1272 if (skb->ip_summed != CHECKSUM_HW)
1273 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1274
1275 /* Update sequence range on original skb. */
1276 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1277
1278 /* Merge over control information. */
1279 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1280 TCP_SKB_CB(skb)->flags = flags;
1281
1282 /* All done, get rid of second SKB and account for it so
1283 * packet counting does not break.
1284 */
1285 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1286 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1287 tp->retrans_out -= tcp_skb_pcount(next_skb);
1288 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1289 tp->lost_out -= tcp_skb_pcount(next_skb);
1290 tp->left_out -= tcp_skb_pcount(next_skb);
1291 }
1292 /* Reno case is special. Sigh... */
1293 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1294 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1295 tp->left_out -= tcp_skb_pcount(next_skb);
1296 }
1297
1298 /* Not quite right: it can be > snd.fack, but
1299 * it is better to underestimate fackets.
1300 */
1301 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1302 tcp_packets_out_dec(tp, next_skb);
1303 sk_stream_free_skb(sk, next_skb);
1304 }
1305}
1306
1307/* Do a simple retransmit without using the backoff mechanisms in
1308 * tcp_timer. This is used for path mtu discovery.
1309 * The socket is already locked here.
1310 */
1311void tcp_simple_retransmit(struct sock *sk)
1312{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001313 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 struct tcp_sock *tp = tcp_sk(sk);
1315 struct sk_buff *skb;
1316 unsigned int mss = tcp_current_mss(sk, 0);
1317 int lost = 0;
1318
1319 sk_stream_for_retrans_queue(skb, sk) {
1320 if (skb->len > mss &&
1321 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1322 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1323 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1324 tp->retrans_out -= tcp_skb_pcount(skb);
1325 }
1326 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1327 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1328 tp->lost_out += tcp_skb_pcount(skb);
1329 lost = 1;
1330 }
1331 }
1332 }
1333
1334 if (!lost)
1335 return;
1336
1337 tcp_sync_left_out(tp);
1338
1339 /* Don't muck with the congestion window here.
1340 * Reason is that we do not increase amount of _data_
1341 * in network, but units changed and effective
1342 * cwnd/ssthresh really reduced now.
1343 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001344 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001346 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 tp->prior_ssthresh = 0;
1348 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001349 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 }
1351 tcp_xmit_retransmit_queue(sk);
1352}
1353
1354/* This retransmits one SKB. Policy decisions and retransmit queue
1355 * state updates are done by the caller. Returns non-zero if an
1356 * error occurred which prevented the send.
1357 */
1358int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1359{
1360 struct tcp_sock *tp = tcp_sk(sk);
1361 unsigned int cur_mss = tcp_current_mss(sk, 0);
1362 int err;
1363
1364 /* Do not sent more than we queued. 1/4 is reserved for possible
1365 * copying overhead: frgagmentation, tunneling, mangling etc.
1366 */
1367 if (atomic_read(&sk->sk_wmem_alloc) >
1368 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1369 return -EAGAIN;
1370
1371 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1372 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1373 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1375 return -ENOMEM;
1376 }
1377
1378 /* If receiver has shrunk his window, and skb is out of
1379 * new window, do not retransmit it. The exception is the
1380 * case, when window is shrunk to zero. In this case
1381 * our retransmit serves as a zero window probe.
1382 */
1383 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1384 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1385 return -EAGAIN;
1386
1387 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001388 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 }
1391
1392 /* Collapse two adjacent packets if worthwhile and we can. */
1393 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1394 (skb->len < (cur_mss >> 1)) &&
1395 (skb->next != sk->sk_send_head) &&
1396 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1397 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1398 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1399 (sysctl_tcp_retrans_collapse != 0))
1400 tcp_retrans_try_collapse(sk, skb, cur_mss);
1401
1402 if(tp->af_specific->rebuild_header(sk))
1403 return -EHOSTUNREACH; /* Routing failure or similar. */
1404
1405 /* Some Solaris stacks overoptimize and ignore the FIN on a
1406 * retransmit when old data is attached. So strip it off
1407 * since it is cheap to do so and saves bytes on the network.
1408 */
1409 if(skb->len > 0 &&
1410 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1411 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1412 if (!pskb_trim(skb, 0)) {
1413 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1414 skb_shinfo(skb)->tso_segs = 1;
1415 skb_shinfo(skb)->tso_size = 0;
1416 skb->ip_summed = CHECKSUM_NONE;
1417 skb->csum = 0;
1418 }
1419 }
1420
1421 /* Make a copy, if the first transmission SKB clone we made
1422 * is still in somebody's hands, else make a clone.
1423 */
1424 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
1426 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1427 pskb_copy(skb, GFP_ATOMIC):
1428 skb_clone(skb, GFP_ATOMIC)));
1429
1430 if (err == 0) {
1431 /* Update global TCP statistics. */
1432 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1433
1434 tp->total_retrans++;
1435
1436#if FASTRETRANS_DEBUG > 0
1437 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1438 if (net_ratelimit())
1439 printk(KERN_DEBUG "retrans_out leaked.\n");
1440 }
1441#endif
1442 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1443 tp->retrans_out += tcp_skb_pcount(skb);
1444
1445 /* Save stamp of the first retransmit. */
1446 if (!tp->retrans_stamp)
1447 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1448
1449 tp->undo_retrans++;
1450
1451 /* snd_nxt is stored to detect loss of retransmitted segment,
1452 * see tcp_input.c tcp_sacktag_write_queue().
1453 */
1454 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1455 }
1456 return err;
1457}
1458
1459/* This gets called after a retransmit timeout, and the initially
1460 * retransmitted data is acknowledged. It tries to continue
1461 * resending the rest of the retransmit queue, until either
1462 * we've sent it all or the congestion window limit is reached.
1463 * If doing SACK, the first ACK which comes back for a timeout
1464 * based retransmit packet might feed us FACK information again.
1465 * If so, we use it to avoid unnecessarily retransmissions.
1466 */
1467void tcp_xmit_retransmit_queue(struct sock *sk)
1468{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001469 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 struct tcp_sock *tp = tcp_sk(sk);
1471 struct sk_buff *skb;
1472 int packet_cnt = tp->lost_out;
1473
1474 /* First pass: retransmit lost packets. */
1475 if (packet_cnt) {
1476 sk_stream_for_retrans_queue(skb, sk) {
1477 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1478
1479 /* Assume this retransmit will generate
1480 * only one packet for congestion window
1481 * calculation purposes. This works because
1482 * tcp_retransmit_skb() will chop up the
1483 * packet to be MSS sized and all the
1484 * packet counting works out.
1485 */
1486 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1487 return;
1488
1489 if (sacked&TCPCB_LOST) {
1490 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1491 if (tcp_retransmit_skb(sk, skb))
1492 return;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001493 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1495 else
1496 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1497
1498 if (skb ==
1499 skb_peek(&sk->sk_write_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001500 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001501 inet_csk(sk)->icsk_rto,
1502 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 }
1504
1505 packet_cnt -= tcp_skb_pcount(skb);
1506 if (packet_cnt <= 0)
1507 break;
1508 }
1509 }
1510 }
1511
1512 /* OK, demanded retransmission is finished. */
1513
1514 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001515 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 return;
1517
1518 /* No forward retransmissions in Reno are possible. */
1519 if (!tp->rx_opt.sack_ok)
1520 return;
1521
1522 /* Yeah, we have to make difficult choice between forward transmission
1523 * and retransmission... Both ways have their merits...
1524 *
1525 * For now we do not retransmit anything, while we have some new
1526 * segments to send.
1527 */
1528
1529 if (tcp_may_send_now(sk, tp))
1530 return;
1531
1532 packet_cnt = 0;
1533
1534 sk_stream_for_retrans_queue(skb, sk) {
1535 /* Similar to the retransmit loop above we
1536 * can pretend that the retransmitted SKB
1537 * we send out here will be composed of one
1538 * real MSS sized packet because tcp_retransmit_skb()
1539 * will fragment it if necessary.
1540 */
1541 if (++packet_cnt > tp->fackets_out)
1542 break;
1543
1544 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1545 break;
1546
1547 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1548 continue;
1549
1550 /* Ok, retransmit it. */
1551 if (tcp_retransmit_skb(sk, skb))
1552 break;
1553
1554 if (skb == skb_peek(&sk->sk_write_queue))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001555 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1556 inet_csk(sk)->icsk_rto,
1557 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558
1559 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1560 }
1561}
1562
1563
1564/* Send a fin. The caller locks the socket for us. This cannot be
1565 * allowed to fail queueing a FIN frame under any circumstances.
1566 */
1567void tcp_send_fin(struct sock *sk)
1568{
1569 struct tcp_sock *tp = tcp_sk(sk);
1570 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1571 int mss_now;
1572
1573 /* Optimization, tack on the FIN if we have a queue of
1574 * unsent frames. But be careful about outgoing SACKS
1575 * and IP options.
1576 */
1577 mss_now = tcp_current_mss(sk, 1);
1578
1579 if (sk->sk_send_head != NULL) {
1580 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1581 TCP_SKB_CB(skb)->end_seq++;
1582 tp->write_seq++;
1583 } else {
1584 /* Socket is locked, keep trying until memory is available. */
1585 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07001586 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 if (skb)
1588 break;
1589 yield();
1590 }
1591
1592 /* Reserve space for headers and prepare control bits. */
1593 skb_reserve(skb, MAX_TCP_HEADER);
1594 skb->csum = 0;
1595 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1596 TCP_SKB_CB(skb)->sacked = 0;
1597 skb_shinfo(skb)->tso_segs = 1;
1598 skb_shinfo(skb)->tso_size = 0;
1599
1600 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1601 TCP_SKB_CB(skb)->seq = tp->write_seq;
1602 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1603 tcp_queue_skb(sk, skb);
1604 }
1605 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1606}
1607
1608/* We get here when a process closes a file descriptor (either due to
1609 * an explicit close() or as a byproduct of exit()'ing) and there
1610 * was unread data in the receive queue. This behavior is recommended
1611 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1612 */
Victor Fusco86a76ca2005-07-08 14:57:47 -07001613void tcp_send_active_reset(struct sock *sk, unsigned int __nocast priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614{
1615 struct tcp_sock *tp = tcp_sk(sk);
1616 struct sk_buff *skb;
1617
1618 /* NOTE: No TCP options attached and we never retransmit this. */
1619 skb = alloc_skb(MAX_TCP_HEADER, priority);
1620 if (!skb) {
1621 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1622 return;
1623 }
1624
1625 /* Reserve space for headers and prepare control bits. */
1626 skb_reserve(skb, MAX_TCP_HEADER);
1627 skb->csum = 0;
1628 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1629 TCP_SKB_CB(skb)->sacked = 0;
1630 skb_shinfo(skb)->tso_segs = 1;
1631 skb_shinfo(skb)->tso_size = 0;
1632
1633 /* Send it off. */
1634 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1635 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1636 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1637 if (tcp_transmit_skb(sk, skb))
1638 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1639}
1640
1641/* WARNING: This routine must only be called when we have already sent
1642 * a SYN packet that crossed the incoming SYN that caused this routine
1643 * to get called. If this assumption fails then the initial rcv_wnd
1644 * and rcv_wscale values will not be correct.
1645 */
1646int tcp_send_synack(struct sock *sk)
1647{
1648 struct sk_buff* skb;
1649
1650 skb = skb_peek(&sk->sk_write_queue);
1651 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1652 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1653 return -EFAULT;
1654 }
1655 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1656 if (skb_cloned(skb)) {
1657 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1658 if (nskb == NULL)
1659 return -ENOMEM;
1660 __skb_unlink(skb, &sk->sk_write_queue);
1661 skb_header_release(nskb);
1662 __skb_queue_head(&sk->sk_write_queue, nskb);
1663 sk_stream_free_skb(sk, skb);
1664 sk_charge_skb(sk, nskb);
1665 skb = nskb;
1666 }
1667
1668 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1669 TCP_ECN_send_synack(tcp_sk(sk), skb);
1670 }
1671 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1672 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1673}
1674
1675/*
1676 * Prepare a SYN-ACK.
1677 */
1678struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
Arnaldo Carvalho de Melo60236fd2005-06-18 22:47:21 -07001679 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001681 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 struct tcp_sock *tp = tcp_sk(sk);
1683 struct tcphdr *th;
1684 int tcp_header_size;
1685 struct sk_buff *skb;
1686
1687 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1688 if (skb == NULL)
1689 return NULL;
1690
1691 /* Reserve space for headers. */
1692 skb_reserve(skb, MAX_TCP_HEADER);
1693
1694 skb->dst = dst_clone(dst);
1695
1696 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001697 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1698 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001700 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1702
1703 memset(th, 0, sizeof(struct tcphdr));
1704 th->syn = 1;
1705 th->ack = 1;
1706 if (dst->dev->features&NETIF_F_TSO)
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001707 ireq->ecn_ok = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 TCP_ECN_make_synack(req, th);
1709 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001710 th->dest = ireq->rmt_port;
1711 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1713 TCP_SKB_CB(skb)->sacked = 0;
1714 skb_shinfo(skb)->tso_segs = 1;
1715 skb_shinfo(skb)->tso_size = 0;
1716 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001717 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1719 __u8 rcv_wscale;
1720 /* Set this up on the first call only */
1721 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1722 /* tcp_full_space because it is guaranteed to be the first packet */
1723 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001724 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 &req->rcv_wnd,
1726 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001727 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 &rcv_wscale);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001729 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 }
1731
1732 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1733 th->window = htons(req->rcv_wnd);
1734
1735 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001736 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
1737 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 TCP_SKB_CB(skb)->when,
1739 req->ts_recent);
1740
1741 skb->csum = 0;
1742 th->doff = (tcp_header_size >> 2);
1743 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1744 return skb;
1745}
1746
1747/*
1748 * Do all connect socket setups that can be done AF independent.
1749 */
1750static inline void tcp_connect_init(struct sock *sk)
1751{
1752 struct dst_entry *dst = __sk_dst_get(sk);
1753 struct tcp_sock *tp = tcp_sk(sk);
1754 __u8 rcv_wscale;
1755
1756 /* We'll fix this up when we get a response from the other end.
1757 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1758 */
1759 tp->tcp_header_len = sizeof(struct tcphdr) +
1760 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1761
1762 /* If user gave his TCP_MAXSEG, record it to clamp */
1763 if (tp->rx_opt.user_mss)
1764 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1765 tp->max_window = 0;
1766 tcp_sync_mss(sk, dst_mtu(dst));
1767
1768 if (!tp->window_clamp)
1769 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1770 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1771 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
1773 tcp_select_initial_window(tcp_full_space(sk),
1774 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1775 &tp->rcv_wnd,
1776 &tp->window_clamp,
1777 sysctl_tcp_window_scaling,
1778 &rcv_wscale);
1779
1780 tp->rx_opt.rcv_wscale = rcv_wscale;
1781 tp->rcv_ssthresh = tp->rcv_wnd;
1782
1783 sk->sk_err = 0;
1784 sock_reset_flag(sk, SOCK_DONE);
1785 tp->snd_wnd = 0;
1786 tcp_init_wl(tp, tp->write_seq, 0);
1787 tp->snd_una = tp->write_seq;
1788 tp->snd_sml = tp->write_seq;
1789 tp->rcv_nxt = 0;
1790 tp->rcv_wup = 0;
1791 tp->copied_seq = 0;
1792
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001793 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
1794 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 tcp_clear_retrans(tp);
1796}
1797
1798/*
1799 * Build a SYN and send it off.
1800 */
1801int tcp_connect(struct sock *sk)
1802{
1803 struct tcp_sock *tp = tcp_sk(sk);
1804 struct sk_buff *buff;
1805
1806 tcp_connect_init(sk);
1807
David S. Millerd179cd12005-08-17 14:57:30 -07001808 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809 if (unlikely(buff == NULL))
1810 return -ENOBUFS;
1811
1812 /* Reserve space for headers. */
1813 skb_reserve(buff, MAX_TCP_HEADER);
1814
1815 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1816 TCP_ECN_send_syn(sk, tp, buff);
1817 TCP_SKB_CB(buff)->sacked = 0;
1818 skb_shinfo(buff)->tso_segs = 1;
1819 skb_shinfo(buff)->tso_size = 0;
1820 buff->csum = 0;
1821 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1822 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1823 tp->snd_nxt = tp->write_seq;
1824 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
1826 /* Send it off. */
1827 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1828 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1829 skb_header_release(buff);
1830 __skb_queue_tail(&sk->sk_write_queue, buff);
1831 sk_charge_skb(sk, buff);
1832 tp->packets_out += tcp_skb_pcount(buff);
1833 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1834 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1835
1836 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001837 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1838 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 return 0;
1840}
1841
1842/* Send out a delayed ack, the caller does the policy checking
1843 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1844 * for details.
1845 */
1846void tcp_send_delayed_ack(struct sock *sk)
1847{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001848 struct inet_connection_sock *icsk = inet_csk(sk);
1849 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 unsigned long timeout;
1851
1852 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001853 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 int max_ato = HZ/2;
1855
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001856 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 max_ato = TCP_DELACK_MAX;
1858
1859 /* Slow path, intersegment interval is "high". */
1860
1861 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001862 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 * directly.
1864 */
1865 if (tp->srtt) {
1866 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1867
1868 if (rtt < max_ato)
1869 max_ato = rtt;
1870 }
1871
1872 ato = min(ato, max_ato);
1873 }
1874
1875 /* Stay within the limit we were given */
1876 timeout = jiffies + ato;
1877
1878 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001879 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 /* If delack timer was blocked or is about to expire,
1881 * send ACK now.
1882 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001883 if (icsk->icsk_ack.blocked ||
1884 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 tcp_send_ack(sk);
1886 return;
1887 }
1888
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001889 if (!time_before(timeout, icsk->icsk_ack.timeout))
1890 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001892 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
1893 icsk->icsk_ack.timeout = timeout;
1894 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895}
1896
1897/* This routine sends an ack and also updates the window. */
1898void tcp_send_ack(struct sock *sk)
1899{
1900 /* If we have been reset, we may not send again. */
1901 if (sk->sk_state != TCP_CLOSE) {
1902 struct tcp_sock *tp = tcp_sk(sk);
1903 struct sk_buff *buff;
1904
1905 /* We are not putting this on the write queue, so
1906 * tcp_transmit_skb() will set the ownership to this
1907 * sock.
1908 */
1909 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1910 if (buff == NULL) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001911 inet_csk_schedule_ack(sk);
1912 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001913 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
1914 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 return;
1916 }
1917
1918 /* Reserve space for headers and prepare control bits. */
1919 skb_reserve(buff, MAX_TCP_HEADER);
1920 buff->csum = 0;
1921 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1922 TCP_SKB_CB(buff)->sacked = 0;
1923 skb_shinfo(buff)->tso_segs = 1;
1924 skb_shinfo(buff)->tso_size = 0;
1925
1926 /* Send it off, this clears delayed acks for us. */
1927 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1928 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1929 tcp_transmit_skb(sk, buff);
1930 }
1931}
1932
1933/* This routine sends a packet with an out of date sequence
1934 * number. It assumes the other end will try to ack it.
1935 *
1936 * Question: what should we make while urgent mode?
1937 * 4.4BSD forces sending single byte of data. We cannot send
1938 * out of window data, because we have SND.NXT==SND.MAX...
1939 *
1940 * Current solution: to send TWO zero-length segments in urgent mode:
1941 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1942 * out-of-date with SND.UNA-1 to probe window.
1943 */
1944static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1945{
1946 struct tcp_sock *tp = tcp_sk(sk);
1947 struct sk_buff *skb;
1948
1949 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1950 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1951 if (skb == NULL)
1952 return -1;
1953
1954 /* Reserve space for headers and set control bits. */
1955 skb_reserve(skb, MAX_TCP_HEADER);
1956 skb->csum = 0;
1957 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1958 TCP_SKB_CB(skb)->sacked = urgent;
1959 skb_shinfo(skb)->tso_segs = 1;
1960 skb_shinfo(skb)->tso_size = 0;
1961
1962 /* Use a previous sequence. This should cause the other
1963 * end to send an ack. Don't queue or clone SKB, just
1964 * send it.
1965 */
1966 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1967 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1968 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1969 return tcp_transmit_skb(sk, skb);
1970}
1971
1972int tcp_write_wakeup(struct sock *sk)
1973{
1974 if (sk->sk_state != TCP_CLOSE) {
1975 struct tcp_sock *tp = tcp_sk(sk);
1976 struct sk_buff *skb;
1977
1978 if ((skb = sk->sk_send_head) != NULL &&
1979 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1980 int err;
1981 unsigned int mss = tcp_current_mss(sk, 0);
1982 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1983
1984 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1985 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1986
1987 /* We are probing the opening of a window
1988 * but the window size is != 0
1989 * must have been a result SWS avoidance ( sender )
1990 */
1991 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1992 skb->len > mss) {
1993 seg_size = min(seg_size, mss);
1994 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
David S. Miller846998a2005-08-04 19:52:01 -07001995 if (tcp_fragment(sk, skb, seg_size, mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 } else if (!tcp_skb_pcount(skb))
David S. Miller846998a2005-08-04 19:52:01 -07001998 tcp_set_skb_tso_segs(sk, skb, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999
2000 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2001 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
2003 if (!err) {
2004 update_send_head(sk, tp, skb);
2005 }
2006 return err;
2007 } else {
2008 if (tp->urg_mode &&
2009 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2010 tcp_xmit_probe_skb(sk, TCPCB_URG);
2011 return tcp_xmit_probe_skb(sk, 0);
2012 }
2013 }
2014 return -1;
2015}
2016
2017/* A window probe timeout has occurred. If window is not closed send
2018 * a partial packet else a zero probe.
2019 */
2020void tcp_send_probe0(struct sock *sk)
2021{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002022 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 struct tcp_sock *tp = tcp_sk(sk);
2024 int err;
2025
2026 err = tcp_write_wakeup(sk);
2027
2028 if (tp->packets_out || !sk->sk_send_head) {
2029 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002030 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002031 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 return;
2033 }
2034
2035 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002036 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2037 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002038 icsk->icsk_probes_out++;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002039 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002040 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2041 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 } else {
2043 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002044 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 * Let local senders to fight for local resources.
2046 *
2047 * Use accumulated backoff yet.
2048 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002049 if (!icsk->icsk_probes_out)
2050 icsk->icsk_probes_out = 1;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002051 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2052 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002053 TCP_RESOURCE_PROBE_INTERVAL),
2054 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 }
2056}
2057
2058EXPORT_SYMBOL(tcp_connect);
2059EXPORT_SYMBOL(tcp_make_synack);
2060EXPORT_SYMBOL(tcp_simple_retransmit);
2061EXPORT_SYMBOL(tcp_sync_mss);