blob: 8d92ab562aedb079bd8f7c3226862d79d6452cc8 [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. Miller846998a2005-08-04 19:52:01 -0700431static int 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;
435 int nsize;
436 u16 flags;
437
438 nsize = skb_headlen(skb) - len;
439 if (nsize < 0)
440 nsize = 0;
441
442 if (skb_cloned(skb) &&
443 skb_is_nonlinear(skb) &&
444 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
445 return -ENOMEM;
446
447 /* Get a new skb... force flag on. */
448 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
449 if (buff == NULL)
450 return -ENOMEM; /* We'll just try again later. */
451 sk_charge_skb(sk, buff);
452
453 /* Correct the sequence numbers. */
454 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
455 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
456 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
457
458 /* PSH and FIN should only be set in the second packet. */
459 flags = TCP_SKB_CB(skb)->flags;
460 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
461 TCP_SKB_CB(buff)->flags = flags;
462 TCP_SKB_CB(buff)->sacked =
463 (TCP_SKB_CB(skb)->sacked &
464 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL));
465 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
488 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
489 tp->lost_out -= tcp_skb_pcount(skb);
490 tp->left_out -= tcp_skb_pcount(skb);
491 }
492
493 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700494 tcp_set_skb_tso_segs(sk, skb, mss_now);
495 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496
497 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
498 tp->lost_out += tcp_skb_pcount(skb);
499 tp->left_out += tcp_skb_pcount(skb);
500 }
501
502 if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) {
503 tp->lost_out += tcp_skb_pcount(buff);
504 tp->left_out += tcp_skb_pcount(buff);
505 }
506
507 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700508 skb_header_release(buff);
David S. Miller8728b832005-08-09 19:25:21 -0700509 __skb_append(skb, buff, &sk->sk_write_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510
511 return 0;
512}
513
514/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
515 * eventually). The difference is that pulled data not copied, but
516 * immediately discarded.
517 */
518static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
519{
520 int i, k, eat;
521
522 eat = len;
523 k = 0;
524 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
525 if (skb_shinfo(skb)->frags[i].size <= eat) {
526 put_page(skb_shinfo(skb)->frags[i].page);
527 eat -= skb_shinfo(skb)->frags[i].size;
528 } else {
529 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
530 if (eat) {
531 skb_shinfo(skb)->frags[k].page_offset += eat;
532 skb_shinfo(skb)->frags[k].size -= eat;
533 eat = 0;
534 }
535 k++;
536 }
537 }
538 skb_shinfo(skb)->nr_frags = k;
539
540 skb->tail = skb->data;
541 skb->data_len -= len;
542 skb->len = skb->data_len;
543 return skb->tail;
544}
545
546int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
547{
548 if (skb_cloned(skb) &&
549 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
550 return -ENOMEM;
551
552 if (len <= skb_headlen(skb)) {
553 __skb_pull(skb, len);
554 } else {
555 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
556 return -ENOMEM;
557 }
558
559 TCP_SKB_CB(skb)->seq += len;
560 skb->ip_summed = CHECKSUM_HW;
561
562 skb->truesize -= len;
563 sk->sk_wmem_queued -= len;
564 sk->sk_forward_alloc += len;
565 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
566
567 /* Any change of skb->len requires recalculation of tso
568 * factor and mss.
569 */
570 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700571 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572
573 return 0;
574}
575
576/* This function synchronize snd mss to current pmtu/exthdr set.
577
578 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
579 for TCP options, but includes only bare TCP header.
580
581 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
582 It is minumum of user_mss and mss received with SYN.
583 It also does not include TCP options.
584
585 tp->pmtu_cookie is last pmtu, seen by this function.
586
587 tp->mss_cache is current effective sending mss, including
588 all tcp options except for SACKs. It is evaluated,
589 taking into account current pmtu, but never exceeds
590 tp->rx_opt.mss_clamp.
591
592 NOTE1. rfc1122 clearly states that advertised MSS
593 DOES NOT include either tcp or ip options.
594
595 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
596 this function. --ANK (980731)
597 */
598
599unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
600{
601 struct tcp_sock *tp = tcp_sk(sk);
602 int mss_now;
603
604 /* Calculate base mss without TCP options:
605 It is MMS_S - sizeof(tcphdr) of rfc1122
606 */
607 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
608
609 /* Clamp it (mss_clamp does not include tcp options) */
610 if (mss_now > tp->rx_opt.mss_clamp)
611 mss_now = tp->rx_opt.mss_clamp;
612
613 /* Now subtract optional transport overhead */
614 mss_now -= tp->ext_header_len;
615
616 /* Then reserve room for full set of TCP options and 8 bytes of data */
617 if (mss_now < 48)
618 mss_now = 48;
619
620 /* Now subtract TCP options size, not including SACKs */
621 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
622
623 /* Bound mss with half of window */
624 if (tp->max_window && mss_now > (tp->max_window>>1))
625 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
626
627 /* And store cached results */
628 tp->pmtu_cookie = pmtu;
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700629 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630
631 return mss_now;
632}
633
634/* Compute the current effective MSS, taking SACKs and IP options,
635 * and even PMTU discovery events into account.
636 *
637 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
638 * cannot be large. However, taking into account rare use of URG, this
639 * is not a big flaw.
640 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700641unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642{
643 struct tcp_sock *tp = tcp_sk(sk);
644 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700645 u32 mss_now;
646 u16 xmit_size_goal;
647 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700649 mss_now = tp->mss_cache;
650
651 if (large_allowed &&
652 (sk->sk_route_caps & NETIF_F_TSO) &&
653 !tp->urg_mode)
654 doing_tso = 1;
655
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 if (dst) {
657 u32 mtu = dst_mtu(dst);
658 if (mtu != tp->pmtu_cookie)
659 mss_now = tcp_sync_mss(sk, mtu);
660 }
661
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 if (tp->rx_opt.eff_sacks)
663 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
664 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700665
666 xmit_size_goal = mss_now;
667
668 if (doing_tso) {
669 xmit_size_goal = 65535 -
670 tp->af_specific->net_header_len -
671 tp->ext_header_len - tp->tcp_header_len;
672
673 if (tp->max_window &&
674 (xmit_size_goal > (tp->max_window >> 1)))
675 xmit_size_goal = max((tp->max_window >> 1),
676 68U - tp->tcp_header_len);
677
678 xmit_size_goal -= (xmit_size_goal % mss_now);
679 }
680 tp->xmit_size_goal = xmit_size_goal;
681
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 return mss_now;
683}
684
David S. Millera762a982005-07-05 15:18:51 -0700685/* Congestion window validation. (RFC2861) */
686
687static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
688{
689 __u32 packets_out = tp->packets_out;
690
691 if (packets_out >= tp->snd_cwnd) {
692 /* Network is feed fully. */
693 tp->snd_cwnd_used = 0;
694 tp->snd_cwnd_stamp = tcp_time_stamp;
695 } else {
696 /* Network starves. */
697 if (tp->packets_out > tp->snd_cwnd_used)
698 tp->snd_cwnd_used = tp->packets_out;
699
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700700 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -0700701 tcp_cwnd_application_limited(sk);
702 }
703}
704
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700705static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
706{
707 u32 window, cwnd_len;
708
709 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
710 cwnd_len = mss_now * cwnd;
711 return min(window, cwnd_len);
712}
713
714/* Can at least one segment of SKB be sent right now, according to the
715 * congestion window rules? If so, return how many segments are allowed.
716 */
717static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
718{
719 u32 in_flight, cwnd;
720
721 /* Don't be strict about the congestion window for the final FIN. */
722 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
723 return 1;
724
725 in_flight = tcp_packets_in_flight(tp);
726 cwnd = tp->snd_cwnd;
727 if (in_flight < cwnd)
728 return (cwnd - in_flight);
729
730 return 0;
731}
732
733/* This must be invoked the first time we consider transmitting
734 * SKB onto the wire.
735 */
David S. Miller846998a2005-08-04 19:52:01 -0700736static 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 -0700737{
738 int tso_segs = tcp_skb_pcount(skb);
739
David S. Miller846998a2005-08-04 19:52:01 -0700740 if (!tso_segs ||
741 (tso_segs > 1 &&
742 skb_shinfo(skb)->tso_size != mss_now)) {
743 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700744 tso_segs = tcp_skb_pcount(skb);
745 }
746 return tso_segs;
747}
748
749static inline int tcp_minshall_check(const struct tcp_sock *tp)
750{
751 return after(tp->snd_sml,tp->snd_una) &&
752 !after(tp->snd_sml, tp->snd_nxt);
753}
754
755/* Return 0, if packet can be sent now without violation Nagle's rules:
756 * 1. It is full sized.
757 * 2. Or it contains FIN. (already checked by caller)
758 * 3. Or TCP_NODELAY was set.
759 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
760 * With Minshall's modification: all sent small packets are ACKed.
761 */
762
763static inline int tcp_nagle_check(const struct tcp_sock *tp,
764 const struct sk_buff *skb,
765 unsigned mss_now, int nonagle)
766{
767 return (skb->len < mss_now &&
768 ((nonagle&TCP_NAGLE_CORK) ||
769 (!nonagle &&
770 tp->packets_out &&
771 tcp_minshall_check(tp))));
772}
773
774/* Return non-zero if the Nagle test allows this packet to be
775 * sent now.
776 */
777static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
778 unsigned int cur_mss, int nonagle)
779{
780 /* Nagle rule does not apply to frames, which sit in the middle of the
781 * write_queue (they have no chances to get new data).
782 *
783 * This is implemented in the callers, where they modify the 'nonagle'
784 * argument based upon the location of SKB in the send queue.
785 */
786 if (nonagle & TCP_NAGLE_PUSH)
787 return 1;
788
789 /* Don't use the nagle rule for urgent data (or for the final FIN). */
790 if (tp->urg_mode ||
791 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
792 return 1;
793
794 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
795 return 1;
796
797 return 0;
798}
799
800/* Does at least the first segment of SKB fit into the send window? */
801static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
802{
803 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
804
805 if (skb->len > cur_mss)
806 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
807
808 return !after(end_seq, tp->snd_una + tp->snd_wnd);
809}
810
811/* This checks if the data bearing packet SKB (usually sk->sk_send_head)
812 * should be put on the wire right now. If so, it returns the number of
813 * packets allowed by the congestion window.
814 */
815static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
816 unsigned int cur_mss, int nonagle)
817{
818 struct tcp_sock *tp = tcp_sk(sk);
819 unsigned int cwnd_quota;
820
David S. Miller846998a2005-08-04 19:52:01 -0700821 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700822
823 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
824 return 0;
825
826 cwnd_quota = tcp_cwnd_test(tp, skb);
827 if (cwnd_quota &&
828 !tcp_snd_wnd_test(tp, skb, cur_mss))
829 cwnd_quota = 0;
830
831 return cwnd_quota;
832}
833
834static inline int tcp_skb_is_last(const struct sock *sk,
835 const struct sk_buff *skb)
836{
837 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
838}
839
840int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
841{
842 struct sk_buff *skb = sk->sk_send_head;
843
844 return (skb &&
845 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
846 (tcp_skb_is_last(sk, skb) ?
847 TCP_NAGLE_PUSH :
848 tp->nonagle)));
849}
850
851/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
852 * which is put after SKB on the list. It is very much like
853 * tcp_fragment() except that it may make several kinds of assumptions
854 * in order to speed up the splitting operation. In particular, we
855 * know that all the data is in scatter-gather pages, and that the
856 * packet has never been sent out before (and thus is not cloned).
857 */
David S. Miller846998a2005-08-04 19:52:01 -0700858static 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 -0700859{
860 struct sk_buff *buff;
861 int nlen = skb->len - len;
862 u16 flags;
863
864 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -0700865 if (skb->len != skb->data_len)
866 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700867
868 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
869 if (unlikely(buff == NULL))
870 return -ENOMEM;
871
872 buff->truesize = nlen;
873 skb->truesize -= nlen;
874
875 /* Correct the sequence numbers. */
876 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
877 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
878 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
879
880 /* PSH and FIN should only be set in the second packet. */
881 flags = TCP_SKB_CB(skb)->flags;
882 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
883 TCP_SKB_CB(buff)->flags = flags;
884
885 /* This packet was never sent out yet, so no SACK bits. */
886 TCP_SKB_CB(buff)->sacked = 0;
887
888 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
889 skb_split(skb, buff, len);
890
891 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700892 tcp_set_skb_tso_segs(sk, skb, mss_now);
893 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700894
895 /* Link BUFF into the send queue. */
896 skb_header_release(buff);
David S. Miller8728b832005-08-09 19:25:21 -0700897 __skb_append(skb, buff, &sk->sk_write_queue);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700898
899 return 0;
900}
901
902/* Try to defer sending, if possible, in order to minimize the amount
903 * of TSO splitting we do. View it as a kind of TSO Nagle test.
904 *
905 * This algorithm is from John Heffner.
906 */
907static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
908{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300909 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700910 u32 send_win, cong_win, limit, in_flight;
911
912 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
913 return 0;
914
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300915 if (icsk->icsk_ca_state != TCP_CA_Open)
David S. Miller908a75c2005-07-05 15:43:58 -0700916 return 0;
917
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700918 in_flight = tcp_packets_in_flight(tp);
919
920 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
921 (tp->snd_cwnd <= in_flight));
922
923 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
924
925 /* From in_flight test above, we know that cwnd > in_flight. */
926 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
927
928 limit = min(send_win, cong_win);
929
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700930 if (sysctl_tcp_tso_win_divisor) {
931 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
932
933 /* If at least some fraction of a window is available,
934 * just use it.
935 */
936 chunk /= sysctl_tcp_tso_win_divisor;
937 if (limit >= chunk)
938 return 0;
939 } else {
940 /* Different approach, try not to defer past a single
941 * ACK. Receiver should ACK every other full sized
942 * frame, so if we have space for more than 3 frames
943 * then send now.
944 */
945 if (limit > tcp_max_burst(tp) * tp->mss_cache)
946 return 0;
947 }
948
949 /* Ok, it looks like it is advisable to defer. */
950 return 1;
951}
952
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953/* This routine writes packets to the network. It advances the
954 * send_head. This happens as incoming acks open up the remote
955 * window for us.
956 *
957 * Returns 1, if no segments are in flight and we have queued segments, but
958 * cannot send anything now because of SWS or another problem.
959 */
David S. Millera2e2a592005-07-05 15:19:23 -0700960static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961{
962 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -0700963 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700964 unsigned int tso_segs, sent_pkts;
965 int cwnd_quota;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966
967 /* If we are closed, the bytes will have to remain here.
968 * In time closedown will finish, we empty the write queue and all
969 * will be happy.
970 */
David S. Miller92df7b52005-07-05 15:19:06 -0700971 if (unlikely(sk->sk_state == TCP_CLOSE))
972 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973
David S. Miller92df7b52005-07-05 15:19:06 -0700974 sent_pkts = 0;
Herbert Xub68e9f82005-08-04 19:52:02 -0700975 while ((skb = sk->sk_send_head)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -0700976 unsigned int limit;
977
Herbert Xub68e9f82005-08-04 19:52:02 -0700978 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700979 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -0700980
Herbert Xub68e9f82005-08-04 19:52:02 -0700981 cwnd_quota = tcp_cwnd_test(tp, skb);
982 if (!cwnd_quota)
983 break;
984
985 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
986 break;
987
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700988 if (tso_segs == 1) {
989 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
990 (tcp_skb_is_last(sk, skb) ?
991 nonagle : TCP_NAGLE_PUSH))))
992 break;
993 } else {
994 if (tcp_tso_should_defer(sk, tp, skb))
995 break;
996 }
David S. Milleraa934662005-07-05 15:20:09 -0700997
Herbert Xuc8ac3772005-08-16 20:43:40 -0700998 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700999 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001000 limit = tcp_window_allows(tp, skb,
1001 mss_now, cwnd_quota);
David S. Milleraa934662005-07-05 15:20:09 -07001002
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001003 if (skb->len < limit) {
1004 unsigned int trim = skb->len % mss_now;
1005
1006 if (trim)
1007 limit = skb->len - trim;
1008 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 }
1010
Herbert Xuc8ac3772005-08-16 20:43:40 -07001011 if (skb->len > limit &&
1012 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1013 break;
1014
David S. Miller92df7b52005-07-05 15:19:06 -07001015 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001016
David S. Milleraa934662005-07-05 15:20:09 -07001017 if (unlikely(tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))))
David S. Miller92df7b52005-07-05 15:19:06 -07001018 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019
David S. Miller92df7b52005-07-05 15:19:06 -07001020 /* Advance the send_head. This one is sent out.
1021 * This call will increment packets_out.
1022 */
1023 update_send_head(sk, tp, skb);
1024
1025 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001026 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 }
David S. Miller92df7b52005-07-05 15:19:06 -07001028
David S. Milleraa934662005-07-05 15:20:09 -07001029 if (likely(sent_pkts)) {
David S. Miller92df7b52005-07-05 15:19:06 -07001030 tcp_cwnd_validate(sk, tp);
1031 return 0;
1032 }
David S. Miller92df7b52005-07-05 15:19:06 -07001033 return !tp->packets_out && sk->sk_send_head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034}
1035
David S. Millera762a982005-07-05 15:18:51 -07001036/* Push out any pending frames which were held back due to
1037 * TCP_CORK or attempt at coalescing tiny packets.
1038 * The socket must be locked by the caller.
1039 */
1040void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
David S. Millera2e2a592005-07-05 15:19:23 -07001041 unsigned int cur_mss, int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001042{
1043 struct sk_buff *skb = sk->sk_send_head;
1044
1045 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001046 if (tcp_write_xmit(sk, cur_mss, nonagle))
David S. Millera762a982005-07-05 15:18:51 -07001047 tcp_check_probe_timer(sk, tp);
1048 }
1049}
1050
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001051/* Send _single_ skb sitting at the send head. This function requires
1052 * true push pending frames to setup probe timer etc.
1053 */
1054void tcp_push_one(struct sock *sk, unsigned int mss_now)
1055{
1056 struct tcp_sock *tp = tcp_sk(sk);
1057 struct sk_buff *skb = sk->sk_send_head;
1058 unsigned int tso_segs, cwnd_quota;
1059
1060 BUG_ON(!skb || skb->len < mss_now);
1061
David S. Miller846998a2005-08-04 19:52:01 -07001062 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001063 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1064
1065 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001066 unsigned int limit;
1067
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001068 BUG_ON(!tso_segs);
1069
Herbert Xuc8ac3772005-08-16 20:43:40 -07001070 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001071 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001072 limit = tcp_window_allows(tp, skb,
1073 mss_now, cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001074
1075 if (skb->len < limit) {
1076 unsigned int trim = skb->len % mss_now;
1077
1078 if (trim)
1079 limit = skb->len - trim;
1080 }
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001081 }
1082
Herbert Xuc8ac3772005-08-16 20:43:40 -07001083 if (skb->len > limit &&
1084 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1085 return;
1086
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001087 /* Send it out now. */
1088 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1089
1090 if (likely(!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation)))) {
1091 update_send_head(sk, tp, skb);
1092 tcp_cwnd_validate(sk, tp);
1093 return;
1094 }
1095 }
1096}
1097
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098/* This function returns the amount that we can raise the
1099 * usable window based on the following constraints
1100 *
1101 * 1. The window can never be shrunk once it is offered (RFC 793)
1102 * 2. We limit memory per socket
1103 *
1104 * RFC 1122:
1105 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1106 * RECV.NEXT + RCV.WIN fixed until:
1107 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1108 *
1109 * i.e. don't raise the right edge of the window until you can raise
1110 * it at least MSS bytes.
1111 *
1112 * Unfortunately, the recommended algorithm breaks header prediction,
1113 * since header prediction assumes th->window stays fixed.
1114 *
1115 * Strictly speaking, keeping th->window fixed violates the receiver
1116 * side SWS prevention criteria. The problem is that under this rule
1117 * a stream of single byte packets will cause the right side of the
1118 * window to always advance by a single byte.
1119 *
1120 * Of course, if the sender implements sender side SWS prevention
1121 * then this will not be a problem.
1122 *
1123 * BSD seems to make the following compromise:
1124 *
1125 * If the free space is less than the 1/4 of the maximum
1126 * space available and the free space is less than 1/2 mss,
1127 * then set the window to 0.
1128 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1129 * Otherwise, just prevent the window from shrinking
1130 * and from being larger than the largest representable value.
1131 *
1132 * This prevents incremental opening of the window in the regime
1133 * where TCP is limited by the speed of the reader side taking
1134 * data out of the TCP receive queue. It does nothing about
1135 * those cases where the window is constrained on the sender side
1136 * because the pipeline is full.
1137 *
1138 * BSD also seems to "accidentally" limit itself to windows that are a
1139 * multiple of MSS, at least until the free space gets quite small.
1140 * This would appear to be a side effect of the mbuf implementation.
1141 * Combining these two algorithms results in the observed behavior
1142 * of having a fixed window size at almost all times.
1143 *
1144 * Below we obtain similar behavior by forcing the offered window to
1145 * a multiple of the mss when it is feasible to do so.
1146 *
1147 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1148 * Regular options like TIMESTAMP are taken into account.
1149 */
1150u32 __tcp_select_window(struct sock *sk)
1151{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001152 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 struct tcp_sock *tp = tcp_sk(sk);
1154 /* MSS for the peer's data. Previous verions used mss_clamp
1155 * here. I don't know if the value based on our guesses
1156 * of peer's MSS is better for the performance. It's more correct
1157 * but may be worse for the performance because of rcv_mss
1158 * fluctuations. --SAW 1998/11/1
1159 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001160 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 int free_space = tcp_space(sk);
1162 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1163 int window;
1164
1165 if (mss > full_space)
1166 mss = full_space;
1167
1168 if (free_space < full_space/2) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001169 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170
1171 if (tcp_memory_pressure)
1172 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1173
1174 if (free_space < mss)
1175 return 0;
1176 }
1177
1178 if (free_space > tp->rcv_ssthresh)
1179 free_space = tp->rcv_ssthresh;
1180
1181 /* Don't do rounding if we are using window scaling, since the
1182 * scaled window will not line up with the MSS boundary anyway.
1183 */
1184 window = tp->rcv_wnd;
1185 if (tp->rx_opt.rcv_wscale) {
1186 window = free_space;
1187
1188 /* Advertise enough space so that it won't get scaled away.
1189 * Import case: prevent zero window announcement if
1190 * 1<<rcv_wscale > mss.
1191 */
1192 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1193 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1194 << tp->rx_opt.rcv_wscale);
1195 } else {
1196 /* Get the largest window that is a nice multiple of mss.
1197 * Window clamp already applied above.
1198 * If our current window offering is within 1 mss of the
1199 * free space we just keep it. This prevents the divide
1200 * and multiply from happening most of the time.
1201 * We also don't do any window rounding when the free space
1202 * is too small.
1203 */
1204 if (window <= free_space - mss || window > free_space)
1205 window = (free_space/mss)*mss;
1206 }
1207
1208 return window;
1209}
1210
1211/* Attempt to collapse two adjacent SKB's during retransmission. */
1212static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1213{
1214 struct tcp_sock *tp = tcp_sk(sk);
1215 struct sk_buff *next_skb = skb->next;
1216
1217 /* The first test we must make is that neither of these two
1218 * SKB's are still referenced by someone else.
1219 */
1220 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1221 int skb_size = skb->len, next_skb_size = next_skb->len;
1222 u16 flags = TCP_SKB_CB(skb)->flags;
1223
1224 /* Also punt if next skb has been SACK'd. */
1225 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1226 return;
1227
1228 /* Next skb is out of window. */
1229 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1230 return;
1231
1232 /* Punt if not enough space exists in the first SKB for
1233 * the data in the second, or the total combined payload
1234 * would exceed the MSS.
1235 */
1236 if ((next_skb_size > skb_tailroom(skb)) ||
1237 ((skb_size + next_skb_size) > mss_now))
1238 return;
1239
1240 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1241 tcp_skb_pcount(next_skb) != 1);
1242
1243 /* Ok. We will be able to collapse the packet. */
David S. Miller8728b832005-08-09 19:25:21 -07001244 __skb_unlink(next_skb, &sk->sk_write_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245
1246 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1247
1248 if (next_skb->ip_summed == CHECKSUM_HW)
1249 skb->ip_summed = CHECKSUM_HW;
1250
1251 if (skb->ip_summed != CHECKSUM_HW)
1252 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1253
1254 /* Update sequence range on original skb. */
1255 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1256
1257 /* Merge over control information. */
1258 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1259 TCP_SKB_CB(skb)->flags = flags;
1260
1261 /* All done, get rid of second SKB and account for it so
1262 * packet counting does not break.
1263 */
1264 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1265 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1266 tp->retrans_out -= tcp_skb_pcount(next_skb);
1267 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1268 tp->lost_out -= tcp_skb_pcount(next_skb);
1269 tp->left_out -= tcp_skb_pcount(next_skb);
1270 }
1271 /* Reno case is special. Sigh... */
1272 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1273 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1274 tp->left_out -= tcp_skb_pcount(next_skb);
1275 }
1276
1277 /* Not quite right: it can be > snd.fack, but
1278 * it is better to underestimate fackets.
1279 */
1280 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1281 tcp_packets_out_dec(tp, next_skb);
1282 sk_stream_free_skb(sk, next_skb);
1283 }
1284}
1285
1286/* Do a simple retransmit without using the backoff mechanisms in
1287 * tcp_timer. This is used for path mtu discovery.
1288 * The socket is already locked here.
1289 */
1290void tcp_simple_retransmit(struct sock *sk)
1291{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001292 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 struct tcp_sock *tp = tcp_sk(sk);
1294 struct sk_buff *skb;
1295 unsigned int mss = tcp_current_mss(sk, 0);
1296 int lost = 0;
1297
1298 sk_stream_for_retrans_queue(skb, sk) {
1299 if (skb->len > mss &&
1300 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1301 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1302 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1303 tp->retrans_out -= tcp_skb_pcount(skb);
1304 }
1305 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1306 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1307 tp->lost_out += tcp_skb_pcount(skb);
1308 lost = 1;
1309 }
1310 }
1311 }
1312
1313 if (!lost)
1314 return;
1315
1316 tcp_sync_left_out(tp);
1317
1318 /* Don't muck with the congestion window here.
1319 * Reason is that we do not increase amount of _data_
1320 * in network, but units changed and effective
1321 * cwnd/ssthresh really reduced now.
1322 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001323 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001325 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 tp->prior_ssthresh = 0;
1327 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001328 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329 }
1330 tcp_xmit_retransmit_queue(sk);
1331}
1332
1333/* This retransmits one SKB. Policy decisions and retransmit queue
1334 * state updates are done by the caller. Returns non-zero if an
1335 * error occurred which prevented the send.
1336 */
1337int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1338{
1339 struct tcp_sock *tp = tcp_sk(sk);
1340 unsigned int cur_mss = tcp_current_mss(sk, 0);
1341 int err;
1342
1343 /* Do not sent more than we queued. 1/4 is reserved for possible
1344 * copying overhead: frgagmentation, tunneling, mangling etc.
1345 */
1346 if (atomic_read(&sk->sk_wmem_alloc) >
1347 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1348 return -EAGAIN;
1349
1350 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1351 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1352 BUG();
1353
1354 if (sk->sk_route_caps & NETIF_F_TSO) {
1355 sk->sk_route_caps &= ~NETIF_F_TSO;
1356 sock_set_flag(sk, SOCK_NO_LARGESEND);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 }
1358
1359 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1360 return -ENOMEM;
1361 }
1362
1363 /* If receiver has shrunk his window, and skb is out of
1364 * new window, do not retransmit it. The exception is the
1365 * case, when window is shrunk to zero. In this case
1366 * our retransmit serves as a zero window probe.
1367 */
1368 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1369 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1370 return -EAGAIN;
1371
1372 if (skb->len > cur_mss) {
1373 int old_factor = tcp_skb_pcount(skb);
Herbert Xub5da6232005-08-10 18:32:36 -07001374 int diff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
David S. Miller846998a2005-08-04 19:52:01 -07001376 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 return -ENOMEM; /* We'll try again later. */
1378
1379 /* New SKB created, account for it. */
Herbert Xub5da6232005-08-10 18:32:36 -07001380 diff = old_factor - tcp_skb_pcount(skb) -
1381 tcp_skb_pcount(skb->next);
1382 tp->packets_out -= diff;
1383
1384 if (diff > 0) {
1385 tp->fackets_out -= diff;
1386 if ((int)tp->fackets_out < 0)
1387 tp->fackets_out = 0;
1388 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 }
1390
1391 /* Collapse two adjacent packets if worthwhile and we can. */
1392 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1393 (skb->len < (cur_mss >> 1)) &&
1394 (skb->next != sk->sk_send_head) &&
1395 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1396 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1397 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1398 (sysctl_tcp_retrans_collapse != 0))
1399 tcp_retrans_try_collapse(sk, skb, cur_mss);
1400
1401 if(tp->af_specific->rebuild_header(sk))
1402 return -EHOSTUNREACH; /* Routing failure or similar. */
1403
1404 /* Some Solaris stacks overoptimize and ignore the FIN on a
1405 * retransmit when old data is attached. So strip it off
1406 * since it is cheap to do so and saves bytes on the network.
1407 */
1408 if(skb->len > 0 &&
1409 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1410 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1411 if (!pskb_trim(skb, 0)) {
1412 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1413 skb_shinfo(skb)->tso_segs = 1;
1414 skb_shinfo(skb)->tso_size = 0;
1415 skb->ip_summed = CHECKSUM_NONE;
1416 skb->csum = 0;
1417 }
1418 }
1419
1420 /* Make a copy, if the first transmission SKB clone we made
1421 * is still in somebody's hands, else make a clone.
1422 */
1423 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424
1425 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1426 pskb_copy(skb, GFP_ATOMIC):
1427 skb_clone(skb, GFP_ATOMIC)));
1428
1429 if (err == 0) {
1430 /* Update global TCP statistics. */
1431 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1432
1433 tp->total_retrans++;
1434
1435#if FASTRETRANS_DEBUG > 0
1436 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1437 if (net_ratelimit())
1438 printk(KERN_DEBUG "retrans_out leaked.\n");
1439 }
1440#endif
1441 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1442 tp->retrans_out += tcp_skb_pcount(skb);
1443
1444 /* Save stamp of the first retransmit. */
1445 if (!tp->retrans_stamp)
1446 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1447
1448 tp->undo_retrans++;
1449
1450 /* snd_nxt is stored to detect loss of retransmitted segment,
1451 * see tcp_input.c tcp_sacktag_write_queue().
1452 */
1453 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1454 }
1455 return err;
1456}
1457
1458/* This gets called after a retransmit timeout, and the initially
1459 * retransmitted data is acknowledged. It tries to continue
1460 * resending the rest of the retransmit queue, until either
1461 * we've sent it all or the congestion window limit is reached.
1462 * If doing SACK, the first ACK which comes back for a timeout
1463 * based retransmit packet might feed us FACK information again.
1464 * If so, we use it to avoid unnecessarily retransmissions.
1465 */
1466void tcp_xmit_retransmit_queue(struct sock *sk)
1467{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001468 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 struct tcp_sock *tp = tcp_sk(sk);
1470 struct sk_buff *skb;
1471 int packet_cnt = tp->lost_out;
1472
1473 /* First pass: retransmit lost packets. */
1474 if (packet_cnt) {
1475 sk_stream_for_retrans_queue(skb, sk) {
1476 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1477
1478 /* Assume this retransmit will generate
1479 * only one packet for congestion window
1480 * calculation purposes. This works because
1481 * tcp_retransmit_skb() will chop up the
1482 * packet to be MSS sized and all the
1483 * packet counting works out.
1484 */
1485 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1486 return;
1487
1488 if (sacked&TCPCB_LOST) {
1489 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1490 if (tcp_retransmit_skb(sk, skb))
1491 return;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001492 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1494 else
1495 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1496
1497 if (skb ==
1498 skb_peek(&sk->sk_write_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001499 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001500 inet_csk(sk)->icsk_rto,
1501 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 }
1503
1504 packet_cnt -= tcp_skb_pcount(skb);
1505 if (packet_cnt <= 0)
1506 break;
1507 }
1508 }
1509 }
1510
1511 /* OK, demanded retransmission is finished. */
1512
1513 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001514 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 return;
1516
1517 /* No forward retransmissions in Reno are possible. */
1518 if (!tp->rx_opt.sack_ok)
1519 return;
1520
1521 /* Yeah, we have to make difficult choice between forward transmission
1522 * and retransmission... Both ways have their merits...
1523 *
1524 * For now we do not retransmit anything, while we have some new
1525 * segments to send.
1526 */
1527
1528 if (tcp_may_send_now(sk, tp))
1529 return;
1530
1531 packet_cnt = 0;
1532
1533 sk_stream_for_retrans_queue(skb, sk) {
1534 /* Similar to the retransmit loop above we
1535 * can pretend that the retransmitted SKB
1536 * we send out here will be composed of one
1537 * real MSS sized packet because tcp_retransmit_skb()
1538 * will fragment it if necessary.
1539 */
1540 if (++packet_cnt > tp->fackets_out)
1541 break;
1542
1543 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1544 break;
1545
1546 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1547 continue;
1548
1549 /* Ok, retransmit it. */
1550 if (tcp_retransmit_skb(sk, skb))
1551 break;
1552
1553 if (skb == skb_peek(&sk->sk_write_queue))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001554 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1555 inet_csk(sk)->icsk_rto,
1556 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557
1558 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1559 }
1560}
1561
1562
1563/* Send a fin. The caller locks the socket for us. This cannot be
1564 * allowed to fail queueing a FIN frame under any circumstances.
1565 */
1566void tcp_send_fin(struct sock *sk)
1567{
1568 struct tcp_sock *tp = tcp_sk(sk);
1569 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1570 int mss_now;
1571
1572 /* Optimization, tack on the FIN if we have a queue of
1573 * unsent frames. But be careful about outgoing SACKS
1574 * and IP options.
1575 */
1576 mss_now = tcp_current_mss(sk, 1);
1577
1578 if (sk->sk_send_head != NULL) {
1579 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1580 TCP_SKB_CB(skb)->end_seq++;
1581 tp->write_seq++;
1582 } else {
1583 /* Socket is locked, keep trying until memory is available. */
1584 for (;;) {
1585 skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL);
1586 if (skb)
1587 break;
1588 yield();
1589 }
1590
1591 /* Reserve space for headers and prepare control bits. */
1592 skb_reserve(skb, MAX_TCP_HEADER);
1593 skb->csum = 0;
1594 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1595 TCP_SKB_CB(skb)->sacked = 0;
1596 skb_shinfo(skb)->tso_segs = 1;
1597 skb_shinfo(skb)->tso_size = 0;
1598
1599 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1600 TCP_SKB_CB(skb)->seq = tp->write_seq;
1601 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1602 tcp_queue_skb(sk, skb);
1603 }
1604 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1605}
1606
1607/* We get here when a process closes a file descriptor (either due to
1608 * an explicit close() or as a byproduct of exit()'ing) and there
1609 * was unread data in the receive queue. This behavior is recommended
1610 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1611 */
Victor Fusco86a76ca2005-07-08 14:57:47 -07001612void tcp_send_active_reset(struct sock *sk, unsigned int __nocast priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613{
1614 struct tcp_sock *tp = tcp_sk(sk);
1615 struct sk_buff *skb;
1616
1617 /* NOTE: No TCP options attached and we never retransmit this. */
1618 skb = alloc_skb(MAX_TCP_HEADER, priority);
1619 if (!skb) {
1620 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1621 return;
1622 }
1623
1624 /* Reserve space for headers and prepare control bits. */
1625 skb_reserve(skb, MAX_TCP_HEADER);
1626 skb->csum = 0;
1627 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1628 TCP_SKB_CB(skb)->sacked = 0;
1629 skb_shinfo(skb)->tso_segs = 1;
1630 skb_shinfo(skb)->tso_size = 0;
1631
1632 /* Send it off. */
1633 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1634 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1635 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1636 if (tcp_transmit_skb(sk, skb))
1637 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1638}
1639
1640/* WARNING: This routine must only be called when we have already sent
1641 * a SYN packet that crossed the incoming SYN that caused this routine
1642 * to get called. If this assumption fails then the initial rcv_wnd
1643 * and rcv_wscale values will not be correct.
1644 */
1645int tcp_send_synack(struct sock *sk)
1646{
1647 struct sk_buff* skb;
1648
1649 skb = skb_peek(&sk->sk_write_queue);
1650 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1651 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1652 return -EFAULT;
1653 }
1654 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1655 if (skb_cloned(skb)) {
1656 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1657 if (nskb == NULL)
1658 return -ENOMEM;
1659 __skb_unlink(skb, &sk->sk_write_queue);
1660 skb_header_release(nskb);
1661 __skb_queue_head(&sk->sk_write_queue, nskb);
1662 sk_stream_free_skb(sk, skb);
1663 sk_charge_skb(sk, nskb);
1664 skb = nskb;
1665 }
1666
1667 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1668 TCP_ECN_send_synack(tcp_sk(sk), skb);
1669 }
1670 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1671 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1672}
1673
1674/*
1675 * Prepare a SYN-ACK.
1676 */
1677struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
Arnaldo Carvalho de Melo60236fd2005-06-18 22:47:21 -07001678 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001680 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 struct tcp_sock *tp = tcp_sk(sk);
1682 struct tcphdr *th;
1683 int tcp_header_size;
1684 struct sk_buff *skb;
1685
1686 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1687 if (skb == NULL)
1688 return NULL;
1689
1690 /* Reserve space for headers. */
1691 skb_reserve(skb, MAX_TCP_HEADER);
1692
1693 skb->dst = dst_clone(dst);
1694
1695 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001696 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1697 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001699 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1701
1702 memset(th, 0, sizeof(struct tcphdr));
1703 th->syn = 1;
1704 th->ack = 1;
1705 if (dst->dev->features&NETIF_F_TSO)
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001706 ireq->ecn_ok = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 TCP_ECN_make_synack(req, th);
1708 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001709 th->dest = ireq->rmt_port;
1710 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1712 TCP_SKB_CB(skb)->sacked = 0;
1713 skb_shinfo(skb)->tso_segs = 1;
1714 skb_shinfo(skb)->tso_size = 0;
1715 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001716 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1718 __u8 rcv_wscale;
1719 /* Set this up on the first call only */
1720 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1721 /* tcp_full_space because it is guaranteed to be the first packet */
1722 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001723 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 &req->rcv_wnd,
1725 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001726 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 &rcv_wscale);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001728 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 }
1730
1731 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1732 th->window = htons(req->rcv_wnd);
1733
1734 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07001735 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
1736 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 TCP_SKB_CB(skb)->when,
1738 req->ts_recent);
1739
1740 skb->csum = 0;
1741 th->doff = (tcp_header_size >> 2);
1742 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1743 return skb;
1744}
1745
1746/*
1747 * Do all connect socket setups that can be done AF independent.
1748 */
1749static inline void tcp_connect_init(struct sock *sk)
1750{
1751 struct dst_entry *dst = __sk_dst_get(sk);
1752 struct tcp_sock *tp = tcp_sk(sk);
1753 __u8 rcv_wscale;
1754
1755 /* We'll fix this up when we get a response from the other end.
1756 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1757 */
1758 tp->tcp_header_len = sizeof(struct tcphdr) +
1759 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1760
1761 /* If user gave his TCP_MAXSEG, record it to clamp */
1762 if (tp->rx_opt.user_mss)
1763 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1764 tp->max_window = 0;
1765 tcp_sync_mss(sk, dst_mtu(dst));
1766
1767 if (!tp->window_clamp)
1768 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1769 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1770 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771
1772 tcp_select_initial_window(tcp_full_space(sk),
1773 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1774 &tp->rcv_wnd,
1775 &tp->window_clamp,
1776 sysctl_tcp_window_scaling,
1777 &rcv_wscale);
1778
1779 tp->rx_opt.rcv_wscale = rcv_wscale;
1780 tp->rcv_ssthresh = tp->rcv_wnd;
1781
1782 sk->sk_err = 0;
1783 sock_reset_flag(sk, SOCK_DONE);
1784 tp->snd_wnd = 0;
1785 tcp_init_wl(tp, tp->write_seq, 0);
1786 tp->snd_una = tp->write_seq;
1787 tp->snd_sml = tp->write_seq;
1788 tp->rcv_nxt = 0;
1789 tp->rcv_wup = 0;
1790 tp->copied_seq = 0;
1791
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001792 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
1793 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 tcp_clear_retrans(tp);
1795}
1796
1797/*
1798 * Build a SYN and send it off.
1799 */
1800int tcp_connect(struct sock *sk)
1801{
1802 struct tcp_sock *tp = tcp_sk(sk);
1803 struct sk_buff *buff;
1804
1805 tcp_connect_init(sk);
1806
1807 buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation);
1808 if (unlikely(buff == NULL))
1809 return -ENOBUFS;
1810
1811 /* Reserve space for headers. */
1812 skb_reserve(buff, MAX_TCP_HEADER);
1813
1814 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1815 TCP_ECN_send_syn(sk, tp, buff);
1816 TCP_SKB_CB(buff)->sacked = 0;
1817 skb_shinfo(buff)->tso_segs = 1;
1818 skb_shinfo(buff)->tso_size = 0;
1819 buff->csum = 0;
1820 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1821 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1822 tp->snd_nxt = tp->write_seq;
1823 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824
1825 /* Send it off. */
1826 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1827 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1828 skb_header_release(buff);
1829 __skb_queue_tail(&sk->sk_write_queue, buff);
1830 sk_charge_skb(sk, buff);
1831 tp->packets_out += tcp_skb_pcount(buff);
1832 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1833 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1834
1835 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001836 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1837 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 return 0;
1839}
1840
1841/* Send out a delayed ack, the caller does the policy checking
1842 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1843 * for details.
1844 */
1845void tcp_send_delayed_ack(struct sock *sk)
1846{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001847 struct inet_connection_sock *icsk = inet_csk(sk);
1848 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 unsigned long timeout;
1850
1851 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001852 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 int max_ato = HZ/2;
1854
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001855 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 max_ato = TCP_DELACK_MAX;
1857
1858 /* Slow path, intersegment interval is "high". */
1859
1860 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001861 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 * directly.
1863 */
1864 if (tp->srtt) {
1865 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1866
1867 if (rtt < max_ato)
1868 max_ato = rtt;
1869 }
1870
1871 ato = min(ato, max_ato);
1872 }
1873
1874 /* Stay within the limit we were given */
1875 timeout = jiffies + ato;
1876
1877 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001878 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 /* If delack timer was blocked or is about to expire,
1880 * send ACK now.
1881 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001882 if (icsk->icsk_ack.blocked ||
1883 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884 tcp_send_ack(sk);
1885 return;
1886 }
1887
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001888 if (!time_before(timeout, icsk->icsk_ack.timeout))
1889 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001891 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
1892 icsk->icsk_ack.timeout = timeout;
1893 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894}
1895
1896/* This routine sends an ack and also updates the window. */
1897void tcp_send_ack(struct sock *sk)
1898{
1899 /* If we have been reset, we may not send again. */
1900 if (sk->sk_state != TCP_CLOSE) {
1901 struct tcp_sock *tp = tcp_sk(sk);
1902 struct sk_buff *buff;
1903
1904 /* We are not putting this on the write queue, so
1905 * tcp_transmit_skb() will set the ownership to this
1906 * sock.
1907 */
1908 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1909 if (buff == NULL) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001910 inet_csk_schedule_ack(sk);
1911 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001912 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
1913 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 return;
1915 }
1916
1917 /* Reserve space for headers and prepare control bits. */
1918 skb_reserve(buff, MAX_TCP_HEADER);
1919 buff->csum = 0;
1920 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1921 TCP_SKB_CB(buff)->sacked = 0;
1922 skb_shinfo(buff)->tso_segs = 1;
1923 skb_shinfo(buff)->tso_size = 0;
1924
1925 /* Send it off, this clears delayed acks for us. */
1926 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1927 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1928 tcp_transmit_skb(sk, buff);
1929 }
1930}
1931
1932/* This routine sends a packet with an out of date sequence
1933 * number. It assumes the other end will try to ack it.
1934 *
1935 * Question: what should we make while urgent mode?
1936 * 4.4BSD forces sending single byte of data. We cannot send
1937 * out of window data, because we have SND.NXT==SND.MAX...
1938 *
1939 * Current solution: to send TWO zero-length segments in urgent mode:
1940 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1941 * out-of-date with SND.UNA-1 to probe window.
1942 */
1943static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1944{
1945 struct tcp_sock *tp = tcp_sk(sk);
1946 struct sk_buff *skb;
1947
1948 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1949 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1950 if (skb == NULL)
1951 return -1;
1952
1953 /* Reserve space for headers and set control bits. */
1954 skb_reserve(skb, MAX_TCP_HEADER);
1955 skb->csum = 0;
1956 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1957 TCP_SKB_CB(skb)->sacked = urgent;
1958 skb_shinfo(skb)->tso_segs = 1;
1959 skb_shinfo(skb)->tso_size = 0;
1960
1961 /* Use a previous sequence. This should cause the other
1962 * end to send an ack. Don't queue or clone SKB, just
1963 * send it.
1964 */
1965 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1966 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1967 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1968 return tcp_transmit_skb(sk, skb);
1969}
1970
1971int tcp_write_wakeup(struct sock *sk)
1972{
1973 if (sk->sk_state != TCP_CLOSE) {
1974 struct tcp_sock *tp = tcp_sk(sk);
1975 struct sk_buff *skb;
1976
1977 if ((skb = sk->sk_send_head) != NULL &&
1978 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1979 int err;
1980 unsigned int mss = tcp_current_mss(sk, 0);
1981 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1982
1983 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1984 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1985
1986 /* We are probing the opening of a window
1987 * but the window size is != 0
1988 * must have been a result SWS avoidance ( sender )
1989 */
1990 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1991 skb->len > mss) {
1992 seg_size = min(seg_size, mss);
1993 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
David S. Miller846998a2005-08-04 19:52:01 -07001994 if (tcp_fragment(sk, skb, seg_size, mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 return -1;
1996 /* SWS override triggered forced fragmentation.
1997 * Disable TSO, the connection is too sick. */
1998 if (sk->sk_route_caps & NETIF_F_TSO) {
1999 sock_set_flag(sk, SOCK_NO_LARGESEND);
2000 sk->sk_route_caps &= ~NETIF_F_TSO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 }
2002 } else if (!tcp_skb_pcount(skb))
David S. Miller846998a2005-08-04 19:52:01 -07002003 tcp_set_skb_tso_segs(sk, skb, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004
2005 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2006 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
2008 if (!err) {
2009 update_send_head(sk, tp, skb);
2010 }
2011 return err;
2012 } else {
2013 if (tp->urg_mode &&
2014 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2015 tcp_xmit_probe_skb(sk, TCPCB_URG);
2016 return tcp_xmit_probe_skb(sk, 0);
2017 }
2018 }
2019 return -1;
2020}
2021
2022/* A window probe timeout has occurred. If window is not closed send
2023 * a partial packet else a zero probe.
2024 */
2025void tcp_send_probe0(struct sock *sk)
2026{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002027 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 struct tcp_sock *tp = tcp_sk(sk);
2029 int err;
2030
2031 err = tcp_write_wakeup(sk);
2032
2033 if (tp->packets_out || !sk->sk_send_head) {
2034 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002035 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002036 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 return;
2038 }
2039
2040 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002041 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2042 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002043 icsk->icsk_probes_out++;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002044 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002045 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2046 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 } else {
2048 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002049 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 * Let local senders to fight for local resources.
2051 *
2052 * Use accumulated backoff yet.
2053 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002054 if (!icsk->icsk_probes_out)
2055 icsk->icsk_probes_out = 1;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002056 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2057 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002058 TCP_RESOURCE_PROBE_INTERVAL),
2059 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 }
2061}
2062
2063EXPORT_SYMBOL(tcp_connect);
2064EXPORT_SYMBOL(tcp_make_synack);
2065EXPORT_SYMBOL(tcp_simple_retransmit);
2066EXPORT_SYMBOL(tcp_sync_mss);