blob: bdd71db8bf908e3d20dad25ee43ec777970c139c [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
Rick Jones15d99e02006-03-20 22:40:29 -080048/* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
50 */
51int sysctl_tcp_workaround_signed_windows = 0;
52
Linus Torvalds1da177e2005-04-16 15:20:36 -070053/* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
56 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -070057int sysctl_tcp_tso_win_divisor = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
John Heffner5d424d52006-03-20 17:53:41 -080059int sysctl_tcp_mtu_probing = 0;
60int sysctl_tcp_base_mss = 512;
61
David S. Miller35089bb2006-06-13 22:33:04 -070062/* By default, RFC2861 behavior. */
63int sysctl_tcp_slow_start_after_idle = 1;
64
Stephen Hemminger40efc6f2006-01-03 16:03:49 -080065static void update_send_head(struct sock *sk, struct tcp_sock *tp,
66 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -070067{
68 sk->sk_send_head = skb->next;
69 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
70 sk->sk_send_head = NULL;
71 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
72 tcp_packets_out_inc(sk, tp, skb);
73}
74
75/* SND.NXT, if window was not shrunk.
76 * If window has been shrunk, what should we make? It is not clear at all.
77 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
78 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
79 * invalid. OK, let's make this for now:
80 */
81static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
82{
83 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
84 return tp->snd_nxt;
85 else
86 return tp->snd_una+tp->snd_wnd;
87}
88
89/* Calculate mss to advertise in SYN segment.
90 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
91 *
92 * 1. It is independent of path mtu.
93 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95 * attached devices, because some buggy hosts are confused by
96 * large MSS.
97 * 4. We do not make 3, we advertise MSS, calculated from first
98 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
99 * This may be overridden via information stored in routing table.
100 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101 * probably even Jumbo".
102 */
103static __u16 tcp_advertise_mss(struct sock *sk)
104{
105 struct tcp_sock *tp = tcp_sk(sk);
106 struct dst_entry *dst = __sk_dst_get(sk);
107 int mss = tp->advmss;
108
109 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
110 mss = dst_metric(dst, RTAX_ADVMSS);
111 tp->advmss = mss;
112 }
113
114 return (__u16)mss;
115}
116
117/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700119static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700121 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 s32 delta = tcp_time_stamp - tp->lsndtime;
123 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
124 u32 cwnd = tp->snd_cwnd;
125
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300126 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300128 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 restart_cwnd = min(restart_cwnd, cwnd);
130
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700131 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 cwnd >>= 1;
133 tp->snd_cwnd = max(cwnd, restart_cwnd);
134 tp->snd_cwnd_stamp = tcp_time_stamp;
135 tp->snd_cwnd_used = 0;
136}
137
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800138static void tcp_event_data_sent(struct tcp_sock *tp,
139 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700141 struct inet_connection_sock *icsk = inet_csk(sk);
142 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143
David S. Miller35089bb2006-06-13 22:33:04 -0700144 if (sysctl_tcp_slow_start_after_idle &&
145 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700146 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
148 tp->lsndtime = now;
149
150 /* If it is a reply for ato after last received
151 * packet, enter pingpong mode.
152 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700153 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
154 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155}
156
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800157static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700159 tcp_dec_quickack_mode(sk, pkts);
160 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161}
162
163/* Determine a window scaling and initial window to offer.
164 * Based on the assumption that the given amount of space
165 * will be offered. Store the results in the tp structure.
166 * NOTE: for smooth operation initial space offering should
167 * be a multiple of mss if possible. We assume here that mss >= 1.
168 * This MUST be enforced by all callers.
169 */
170void tcp_select_initial_window(int __space, __u32 mss,
171 __u32 *rcv_wnd, __u32 *window_clamp,
172 int wscale_ok, __u8 *rcv_wscale)
173{
174 unsigned int space = (__space < 0 ? 0 : __space);
175
176 /* If no clamp set the clamp to the max possible scaled window */
177 if (*window_clamp == 0)
178 (*window_clamp) = (65535 << 14);
179 space = min(*window_clamp, space);
180
181 /* Quantize space offering to a multiple of mss if possible. */
182 if (space > mss)
183 space = (space / mss) * mss;
184
185 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800186 * will break some buggy TCP stacks. If the admin tells us
187 * it is likely we could be speaking with such a buggy stack
188 * we will truncate our initial window offering to 32K-1
189 * unless the remote has sent us a window scaling option,
190 * which we interpret as a sign the remote TCP is not
191 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 */
Rick Jones15d99e02006-03-20 22:40:29 -0800193 if (sysctl_tcp_workaround_signed_windows)
194 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
195 else
196 (*rcv_wnd) = space;
197
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198 (*rcv_wscale) = 0;
199 if (wscale_ok) {
200 /* Set window scaling on max possible window
201 * See RFC1323 for an explanation of the limit to 14
202 */
203 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
204 while (space > 65535 && (*rcv_wscale) < 14) {
205 space >>= 1;
206 (*rcv_wscale)++;
207 }
208 }
209
210 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700211 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 * will be satisfied with 2.
213 */
214 if (mss > (1<<*rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700215 int init_cwnd = 4;
216 if (mss > 1460*3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700218 else if (mss > 1460)
219 init_cwnd = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 if (*rcv_wnd > init_cwnd*mss)
221 *rcv_wnd = init_cwnd*mss;
222 }
223
224 /* Set the clamp no higher than max representable value */
225 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
226}
227
228/* Chose a new window to advertise, update state in tcp_sock for the
229 * socket, and return result with RFC1323 scaling applied. The return
230 * value can be stuffed directly into th->window for an outgoing
231 * frame.
232 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800233static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234{
235 struct tcp_sock *tp = tcp_sk(sk);
236 u32 cur_win = tcp_receive_window(tp);
237 u32 new_win = __tcp_select_window(sk);
238
239 /* Never shrink the offered window */
240 if(new_win < cur_win) {
241 /* Danger Will Robinson!
242 * Don't update rcv_wup/rcv_wnd here or else
243 * we will not be able to advertise a zero
244 * window in time. --DaveM
245 *
246 * Relax Will Robinson.
247 */
248 new_win = cur_win;
249 }
250 tp->rcv_wnd = new_win;
251 tp->rcv_wup = tp->rcv_nxt;
252
253 /* Make sure we do not exceed the maximum possible
254 * scaled window.
255 */
Rick Jones15d99e02006-03-20 22:40:29 -0800256 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 new_win = min(new_win, MAX_TCP_WINDOW);
258 else
259 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
260
261 /* RFC1323 scaling applied */
262 new_win >>= tp->rx_opt.rcv_wscale;
263
264 /* If we advertise zero window, disable fast path. */
265 if (new_win == 0)
266 tp->pred_flags = 0;
267
268 return new_win;
269}
270
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800271static void tcp_build_and_update_options(__u32 *ptr, struct tcp_sock *tp,
272 __u32 tstamp)
273{
274 if (tp->rx_opt.tstamp_ok) {
275 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) |
276 (TCPOPT_NOP << 16) |
277 (TCPOPT_TIMESTAMP << 8) |
278 TCPOLEN_TIMESTAMP);
279 *ptr++ = htonl(tstamp);
280 *ptr++ = htonl(tp->rx_opt.ts_recent);
281 }
282 if (tp->rx_opt.eff_sacks) {
283 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
284 int this_sack;
285
286 *ptr++ = htonl((TCPOPT_NOP << 24) |
287 (TCPOPT_NOP << 16) |
288 (TCPOPT_SACK << 8) |
289 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
290 TCPOLEN_SACK_PERBLOCK)));
291 for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
292 *ptr++ = htonl(sp[this_sack].start_seq);
293 *ptr++ = htonl(sp[this_sack].end_seq);
294 }
295 if (tp->rx_opt.dsack) {
296 tp->rx_opt.dsack = 0;
297 tp->rx_opt.eff_sacks--;
298 }
299 }
300}
301
302/* Construct a tcp options header for a SYN or SYN_ACK packet.
303 * If this is every changed make sure to change the definition of
304 * MAX_SYN_SIZE to match the new maximum number of options that you
305 * can generate.
306 */
307static void tcp_syn_build_options(__u32 *ptr, int mss, int ts, int sack,
308 int offer_wscale, int wscale, __u32 tstamp,
309 __u32 ts_recent)
310{
311 /* We always get an MSS option.
312 * The option bytes which will be seen in normal data
313 * packets should timestamps be used, must be in the MSS
314 * advertised. But we subtract them from tp->mss_cache so
315 * that calculations in tcp_sendmsg are simpler etc.
316 * So account for this fact here if necessary. If we
317 * don't do this correctly, as a receiver we won't
318 * recognize data packets as being full sized when we
319 * should, and thus we won't abide by the delayed ACK
320 * rules correctly.
321 * SACKs don't matter, we never delay an ACK when we
322 * have any of those going out.
323 */
324 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
325 if (ts) {
326 if(sack)
327 *ptr++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) |
328 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
329 else
330 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
331 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
332 *ptr++ = htonl(tstamp); /* TSVAL */
333 *ptr++ = htonl(ts_recent); /* TSECR */
334 } else if(sack)
335 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
336 (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM);
337 if (offer_wscale)
338 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale));
339}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
341/* This routine actually transmits TCP packets queued in by
342 * tcp_do_sendmsg(). This is used by both the initial
343 * transmission and possible later retransmissions.
344 * All SKB's seen here are completely headerless. It is our
345 * job to build the TCP header, and pass the packet down to
346 * IP so it can do the same plus pass the packet off to the
347 * device.
348 *
349 * We are working here with either a clone of the original
350 * SKB, or a fresh unique copy made by the retransmit engine.
351 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800352static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800354 const struct inet_connection_sock *icsk = inet_csk(sk);
355 struct inet_sock *inet;
356 struct tcp_sock *tp;
357 struct tcp_skb_cb *tcb;
358 int tcp_header_size;
359 struct tcphdr *th;
360 int sysctl_flags;
361 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800363 BUG_ON(!skb || !tcp_skb_pcount(skb));
364
365 /* If congestion control is doing timestamping, we must
366 * take such a timestamp before we potentially clone/copy.
367 */
368 if (icsk->icsk_ca_ops->rtt_sample)
369 __net_timestamp(skb);
370
371 if (likely(clone_it)) {
372 if (unlikely(skb_cloned(skb)))
373 skb = pskb_copy(skb, gfp_mask);
374 else
375 skb = skb_clone(skb, gfp_mask);
376 if (unlikely(!skb))
377 return -ENOBUFS;
378 }
379
380 inet = inet_sk(sk);
381 tp = tcp_sk(sk);
382 tcb = TCP_SKB_CB(skb);
383 tcp_header_size = tp->tcp_header_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
385#define SYSCTL_FLAG_TSTAMPS 0x1
386#define SYSCTL_FLAG_WSCALE 0x2
387#define SYSCTL_FLAG_SACK 0x4
388
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800389 sysctl_flags = 0;
390 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
391 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
392 if(sysctl_tcp_timestamps) {
393 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
394 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800396 if (sysctl_tcp_window_scaling) {
397 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
398 sysctl_flags |= SYSCTL_FLAG_WSCALE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800400 if (sysctl_tcp_sack) {
401 sysctl_flags |= SYSCTL_FLAG_SACK;
402 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
403 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800405 } else if (unlikely(tp->rx_opt.eff_sacks)) {
406 /* A SACK is 2 pad bytes, a 2 byte header, plus
407 * 2 32-bit sequence numbers for each SACK block.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 */
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800409 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
410 (tp->rx_opt.eff_sacks *
411 TCPOLEN_SACK_PERBLOCK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800413
414 if (tcp_packets_in_flight(tp) == 0)
415 tcp_ca_event(sk, CA_EVENT_TX_START);
416
417 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
418 skb->h.th = th;
419 skb_set_owner_w(skb, sk);
420
421 /* Build TCP header and checksum it. */
422 th->source = inet->sport;
423 th->dest = inet->dport;
424 th->seq = htonl(tcb->seq);
425 th->ack_seq = htonl(tp->rcv_nxt);
426 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
427 tcb->flags);
428
429 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
430 /* RFC1323: The window in SYN & SYN/ACK segments
431 * is never scaled.
432 */
433 th->window = htons(tp->rcv_wnd);
434 } else {
435 th->window = htons(tcp_select_window(sk));
436 }
437 th->check = 0;
438 th->urg_ptr = 0;
439
440 if (unlikely(tp->urg_mode &&
441 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
442 th->urg_ptr = htons(tp->snd_up-tcb->seq);
443 th->urg = 1;
444 }
445
446 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
447 tcp_syn_build_options((__u32 *)(th + 1),
448 tcp_advertise_mss(sk),
449 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
450 (sysctl_flags & SYSCTL_FLAG_SACK),
451 (sysctl_flags & SYSCTL_FLAG_WSCALE),
452 tp->rx_opt.rcv_wscale,
453 tcb->when,
454 tp->rx_opt.ts_recent);
455 } else {
456 tcp_build_and_update_options((__u32 *)(th + 1),
457 tp, tcb->when);
458 TCP_ECN_send(sk, tp, skb, tcp_header_size);
459 }
460
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800461 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800462
463 if (likely(tcb->flags & TCPCB_FLAG_ACK))
464 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
465
466 if (skb->len != tcp_header_size)
467 tcp_event_data_sent(tp, skb, sk);
468
469 TCP_INC_STATS(TCP_MIB_OUTSEGS);
470
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800471 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700472 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800473 return err;
474
475 tcp_enter_cwr(sk);
476
477 /* NET_XMIT_CN is special. It does not guarantee,
478 * that this packet is lost. It tells that device
479 * is about to start to drop packets or already
480 * drops some packets of the same priority and
481 * invokes us to send less aggressively.
482 */
483 return err == NET_XMIT_CN ? 0 : err;
484
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485#undef SYSCTL_FLAG_TSTAMPS
486#undef SYSCTL_FLAG_WSCALE
487#undef SYSCTL_FLAG_SACK
488}
489
490
491/* This routine just queue's the buffer
492 *
493 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
494 * otherwise socket can stall.
495 */
496static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
497{
498 struct tcp_sock *tp = tcp_sk(sk);
499
500 /* Advance write_seq and place onto the write_queue. */
501 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
502 skb_header_release(skb);
503 __skb_queue_tail(&sk->sk_write_queue, skb);
504 sk_charge_skb(sk, skb);
505
506 /* Queue it, remembering where we must start sending. */
507 if (sk->sk_send_head == NULL)
508 sk->sk_send_head = skb;
509}
510
David S. Miller846998a2005-08-04 19:52:01 -0700511static 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 -0700512{
David S. Miller846998a2005-08-04 19:52:01 -0700513 if (skb->len <= mss_now ||
David S. Millerf6302d12005-07-05 15:18:03 -0700514 !(sk->sk_route_caps & NETIF_F_TSO)) {
515 /* Avoid the costly divide in the normal
516 * non-TSO case.
517 */
Herbert Xu79671682006-06-22 02:40:14 -0700518 skb_shinfo(skb)->gso_segs = 1;
519 skb_shinfo(skb)->gso_size = 0;
520 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700521 } else {
522 unsigned int factor;
523
David S. Miller846998a2005-08-04 19:52:01 -0700524 factor = skb->len + (mss_now - 1);
525 factor /= mss_now;
Herbert Xu79671682006-06-22 02:40:14 -0700526 skb_shinfo(skb)->gso_segs = factor;
527 skb_shinfo(skb)->gso_size = mss_now;
528 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 }
530}
531
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532/* Function to create two new TCP segments. Shrinks the given segment
533 * to the specified size and appends a new segment with the rest of the
534 * packet to the list. This won't be called frequently, I hope.
535 * Remember, these are still headerless SKBs at this point.
536 */
David S. Miller6475be12005-09-01 22:47:01 -0700537int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538{
539 struct tcp_sock *tp = tcp_sk(sk);
540 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700541 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700542 int nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 u16 flags;
544
Herbert Xub2cc99f2005-10-20 17:13:13 -0200545 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800546
547 clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 nsize = skb_headlen(skb) - len;
549 if (nsize < 0)
550 nsize = 0;
551
552 if (skb_cloned(skb) &&
553 skb_is_nonlinear(skb) &&
554 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
555 return -ENOMEM;
556
557 /* Get a new skb... force flag on. */
558 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
559 if (buff == NULL)
560 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -0700561
Herbert Xub60b49e2006-04-19 21:35:00 -0700562 sk_charge_skb(sk, buff);
563 nlen = skb->len - len - nsize;
564 buff->truesize += nlen;
565 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566
567 /* Correct the sequence numbers. */
568 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
569 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
570 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
571
572 /* PSH and FIN should only be set in the second packet. */
573 flags = TCP_SKB_CB(skb)->flags;
574 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
575 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700576 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
578
579 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
580 /* Copy and checksum data tail into the new buffer. */
581 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
582 nsize, 0);
583
584 skb_trim(skb, len);
585
586 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
587 } else {
588 skb->ip_summed = CHECKSUM_HW;
589 skb_split(skb, buff, len);
590 }
591
592 buff->ip_summed = skb->ip_summed;
593
594 /* Looks stupid, but our code really uses when of
595 * skbs, which it never sent before. --ANK
596 */
597 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700598 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
David S. Miller6475be12005-09-01 22:47:01 -0700600 old_factor = tcp_skb_pcount(skb);
601
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700603 tcp_set_skb_tso_segs(sk, skb, mss_now);
604 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
David S. Miller6475be12005-09-01 22:47:01 -0700606 /* If this packet has been sent out already, we must
607 * adjust the various packet counters.
608 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700609 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700610 int diff = old_factor - tcp_skb_pcount(skb) -
611 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612
David S. Miller6475be12005-09-01 22:47:01 -0700613 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700614
615 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
616 tp->sacked_out -= diff;
617 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
618 tp->retrans_out -= diff;
619
David S. Miller6475be12005-09-01 22:47:01 -0700620 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
621 tp->lost_out -= diff;
622 tp->left_out -= diff;
623 }
Herbert Xu83ca28b2005-09-22 23:32:56 -0700624
David S. Miller6475be12005-09-01 22:47:01 -0700625 if (diff > 0) {
Herbert Xu83ca28b2005-09-22 23:32:56 -0700626 /* Adjust Reno SACK estimate. */
627 if (!tp->rx_opt.sack_ok) {
628 tp->sacked_out -= diff;
629 if ((int)tp->sacked_out < 0)
630 tp->sacked_out = 0;
631 tcp_sync_left_out(tp);
632 }
633
David S. Miller6475be12005-09-01 22:47:01 -0700634 tp->fackets_out -= diff;
635 if ((int)tp->fackets_out < 0)
636 tp->fackets_out = 0;
637 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 }
639
640 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700641 skb_header_release(buff);
David S. Miller8728b832005-08-09 19:25:21 -0700642 __skb_append(skb, buff, &sk->sk_write_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
644 return 0;
645}
646
647/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
648 * eventually). The difference is that pulled data not copied, but
649 * immediately discarded.
650 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700651static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652{
653 int i, k, eat;
654
655 eat = len;
656 k = 0;
657 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
658 if (skb_shinfo(skb)->frags[i].size <= eat) {
659 put_page(skb_shinfo(skb)->frags[i].page);
660 eat -= skb_shinfo(skb)->frags[i].size;
661 } else {
662 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
663 if (eat) {
664 skb_shinfo(skb)->frags[k].page_offset += eat;
665 skb_shinfo(skb)->frags[k].size -= eat;
666 eat = 0;
667 }
668 k++;
669 }
670 }
671 skb_shinfo(skb)->nr_frags = k;
672
673 skb->tail = skb->data;
674 skb->data_len -= len;
675 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
678int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
679{
680 if (skb_cloned(skb) &&
681 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
682 return -ENOMEM;
683
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700684 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
685 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700687 else
688 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
690 TCP_SKB_CB(skb)->seq += len;
691 skb->ip_summed = CHECKSUM_HW;
692
693 skb->truesize -= len;
694 sk->sk_wmem_queued -= len;
695 sk->sk_forward_alloc += len;
696 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
697
698 /* Any change of skb->len requires recalculation of tso
699 * factor and mss.
700 */
701 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700702 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
704 return 0;
705}
706
John Heffner5d424d52006-03-20 17:53:41 -0800707/* Not accounting for SACKs here. */
708int tcp_mtu_to_mss(struct sock *sk, int pmtu)
709{
710 struct tcp_sock *tp = tcp_sk(sk);
711 struct inet_connection_sock *icsk = inet_csk(sk);
712 int mss_now;
713
714 /* Calculate base mss without TCP options:
715 It is MMS_S - sizeof(tcphdr) of rfc1122
716 */
717 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
718
719 /* Clamp it (mss_clamp does not include tcp options) */
720 if (mss_now > tp->rx_opt.mss_clamp)
721 mss_now = tp->rx_opt.mss_clamp;
722
723 /* Now subtract optional transport overhead */
724 mss_now -= icsk->icsk_ext_hdr_len;
725
726 /* Then reserve room for full set of TCP options and 8 bytes of data */
727 if (mss_now < 48)
728 mss_now = 48;
729
730 /* Now subtract TCP options size, not including SACKs */
731 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
732
733 return mss_now;
734}
735
736/* Inverse of above */
737int tcp_mss_to_mtu(struct sock *sk, int mss)
738{
739 struct tcp_sock *tp = tcp_sk(sk);
740 struct inet_connection_sock *icsk = inet_csk(sk);
741 int mtu;
742
743 mtu = mss +
744 tp->tcp_header_len +
745 icsk->icsk_ext_hdr_len +
746 icsk->icsk_af_ops->net_header_len;
747
748 return mtu;
749}
750
751void tcp_mtup_init(struct sock *sk)
752{
753 struct tcp_sock *tp = tcp_sk(sk);
754 struct inet_connection_sock *icsk = inet_csk(sk);
755
756 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
757 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
758 icsk->icsk_af_ops->net_header_len;
759 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
760 icsk->icsk_mtup.probe_size = 0;
761}
762
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763/* This function synchronize snd mss to current pmtu/exthdr set.
764
765 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
766 for TCP options, but includes only bare TCP header.
767
768 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800769 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 It also does not include TCP options.
771
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800772 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773
774 tp->mss_cache is current effective sending mss, including
775 all tcp options except for SACKs. It is evaluated,
776 taking into account current pmtu, but never exceeds
777 tp->rx_opt.mss_clamp.
778
779 NOTE1. rfc1122 clearly states that advertised MSS
780 DOES NOT include either tcp or ip options.
781
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800782 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
783 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 */
785
786unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
787{
788 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800789 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -0800790 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791
John Heffner5d424d52006-03-20 17:53:41 -0800792 if (icsk->icsk_mtup.search_high > pmtu)
793 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794
John Heffner5d424d52006-03-20 17:53:41 -0800795 mss_now = tcp_mtu_to_mss(sk, pmtu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
797 /* Bound mss with half of window */
798 if (tp->max_window && mss_now > (tp->max_window>>1))
799 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
800
801 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800802 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -0800803 if (icsk->icsk_mtup.enabled)
804 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700805 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
807 return mss_now;
808}
809
810/* Compute the current effective MSS, taking SACKs and IP options,
811 * and even PMTU discovery events into account.
812 *
813 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
814 * cannot be large. However, taking into account rare use of URG, this
815 * is not a big flaw.
816 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700817unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818{
819 struct tcp_sock *tp = tcp_sk(sk);
820 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700821 u32 mss_now;
822 u16 xmit_size_goal;
823 int doing_tso = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700825 mss_now = tp->mss_cache;
826
827 if (large_allowed &&
828 (sk->sk_route_caps & NETIF_F_TSO) &&
829 !tp->urg_mode)
830 doing_tso = 1;
831
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 if (dst) {
833 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800834 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 mss_now = tcp_sync_mss(sk, mtu);
836 }
837
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 if (tp->rx_opt.eff_sacks)
839 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
840 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700841
842 xmit_size_goal = mss_now;
843
844 if (doing_tso) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800845 xmit_size_goal = (65535 -
846 inet_csk(sk)->icsk_af_ops->net_header_len -
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800847 inet_csk(sk)->icsk_ext_hdr_len -
848 tp->tcp_header_len);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700849
850 if (tp->max_window &&
851 (xmit_size_goal > (tp->max_window >> 1)))
852 xmit_size_goal = max((tp->max_window >> 1),
853 68U - tp->tcp_header_len);
854
855 xmit_size_goal -= (xmit_size_goal % mss_now);
856 }
857 tp->xmit_size_goal = xmit_size_goal;
858
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 return mss_now;
860}
861
David S. Millera762a982005-07-05 15:18:51 -0700862/* Congestion window validation. (RFC2861) */
863
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800864static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
David S. Millera762a982005-07-05 15:18:51 -0700865{
866 __u32 packets_out = tp->packets_out;
867
868 if (packets_out >= tp->snd_cwnd) {
869 /* Network is feed fully. */
870 tp->snd_cwnd_used = 0;
871 tp->snd_cwnd_stamp = tcp_time_stamp;
872 } else {
873 /* Network starves. */
874 if (tp->packets_out > tp->snd_cwnd_used)
875 tp->snd_cwnd_used = tp->packets_out;
876
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700877 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -0700878 tcp_cwnd_application_limited(sk);
879 }
880}
881
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700882static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
883{
884 u32 window, cwnd_len;
885
886 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
887 cwnd_len = mss_now * cwnd;
888 return min(window, cwnd_len);
889}
890
891/* Can at least one segment of SKB be sent right now, according to the
892 * congestion window rules? If so, return how many segments are allowed.
893 */
894static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
895{
896 u32 in_flight, cwnd;
897
898 /* Don't be strict about the congestion window for the final FIN. */
899 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
900 return 1;
901
902 in_flight = tcp_packets_in_flight(tp);
903 cwnd = tp->snd_cwnd;
904 if (in_flight < cwnd)
905 return (cwnd - in_flight);
906
907 return 0;
908}
909
910/* This must be invoked the first time we consider transmitting
911 * SKB onto the wire.
912 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800913static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700914{
915 int tso_segs = tcp_skb_pcount(skb);
916
David S. Miller846998a2005-08-04 19:52:01 -0700917 if (!tso_segs ||
918 (tso_segs > 1 &&
Herbert Xu79671682006-06-22 02:40:14 -0700919 tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -0700920 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700921 tso_segs = tcp_skb_pcount(skb);
922 }
923 return tso_segs;
924}
925
926static inline int tcp_minshall_check(const struct tcp_sock *tp)
927{
928 return after(tp->snd_sml,tp->snd_una) &&
929 !after(tp->snd_sml, tp->snd_nxt);
930}
931
932/* Return 0, if packet can be sent now without violation Nagle's rules:
933 * 1. It is full sized.
934 * 2. Or it contains FIN. (already checked by caller)
935 * 3. Or TCP_NODELAY was set.
936 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
937 * With Minshall's modification: all sent small packets are ACKed.
938 */
939
940static inline int tcp_nagle_check(const struct tcp_sock *tp,
941 const struct sk_buff *skb,
942 unsigned mss_now, int nonagle)
943{
944 return (skb->len < mss_now &&
945 ((nonagle&TCP_NAGLE_CORK) ||
946 (!nonagle &&
947 tp->packets_out &&
948 tcp_minshall_check(tp))));
949}
950
951/* Return non-zero if the Nagle test allows this packet to be
952 * sent now.
953 */
954static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
955 unsigned int cur_mss, int nonagle)
956{
957 /* Nagle rule does not apply to frames, which sit in the middle of the
958 * write_queue (they have no chances to get new data).
959 *
960 * This is implemented in the callers, where they modify the 'nonagle'
961 * argument based upon the location of SKB in the send queue.
962 */
963 if (nonagle & TCP_NAGLE_PUSH)
964 return 1;
965
966 /* Don't use the nagle rule for urgent data (or for the final FIN). */
967 if (tp->urg_mode ||
968 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
969 return 1;
970
971 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
972 return 1;
973
974 return 0;
975}
976
977/* Does at least the first segment of SKB fit into the send window? */
978static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
979{
980 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
981
982 if (skb->len > cur_mss)
983 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
984
985 return !after(end_seq, tp->snd_una + tp->snd_wnd);
986}
987
988/* This checks if the data bearing packet SKB (usually sk->sk_send_head)
989 * should be put on the wire right now. If so, it returns the number of
990 * packets allowed by the congestion window.
991 */
992static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
993 unsigned int cur_mss, int nonagle)
994{
995 struct tcp_sock *tp = tcp_sk(sk);
996 unsigned int cwnd_quota;
997
David S. Miller846998a2005-08-04 19:52:01 -0700998 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700999
1000 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1001 return 0;
1002
1003 cwnd_quota = tcp_cwnd_test(tp, skb);
1004 if (cwnd_quota &&
1005 !tcp_snd_wnd_test(tp, skb, cur_mss))
1006 cwnd_quota = 0;
1007
1008 return cwnd_quota;
1009}
1010
1011static inline int tcp_skb_is_last(const struct sock *sk,
1012 const struct sk_buff *skb)
1013{
1014 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
1015}
1016
1017int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
1018{
1019 struct sk_buff *skb = sk->sk_send_head;
1020
1021 return (skb &&
1022 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1023 (tcp_skb_is_last(sk, skb) ?
1024 TCP_NAGLE_PUSH :
1025 tp->nonagle)));
1026}
1027
1028/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1029 * which is put after SKB on the list. It is very much like
1030 * tcp_fragment() except that it may make several kinds of assumptions
1031 * in order to speed up the splitting operation. In particular, we
1032 * know that all the data is in scatter-gather pages, and that the
1033 * packet has never been sent out before (and thus is not cloned).
1034 */
David S. Miller846998a2005-08-04 19:52:01 -07001035static 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 -07001036{
1037 struct sk_buff *buff;
1038 int nlen = skb->len - len;
1039 u16 flags;
1040
1041 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001042 if (skb->len != skb->data_len)
1043 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001044
1045 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1046 if (unlikely(buff == NULL))
1047 return -ENOMEM;
1048
Herbert Xub60b49e2006-04-19 21:35:00 -07001049 sk_charge_skb(sk, buff);
1050 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001051 skb->truesize -= nlen;
1052
1053 /* Correct the sequence numbers. */
1054 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1055 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1056 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1057
1058 /* PSH and FIN should only be set in the second packet. */
1059 flags = TCP_SKB_CB(skb)->flags;
1060 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1061 TCP_SKB_CB(buff)->flags = flags;
1062
1063 /* This packet was never sent out yet, so no SACK bits. */
1064 TCP_SKB_CB(buff)->sacked = 0;
1065
1066 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
1067 skb_split(skb, buff, len);
1068
1069 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001070 tcp_set_skb_tso_segs(sk, skb, mss_now);
1071 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001072
1073 /* Link BUFF into the send queue. */
1074 skb_header_release(buff);
David S. Miller8728b832005-08-09 19:25:21 -07001075 __skb_append(skb, buff, &sk->sk_write_queue);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001076
1077 return 0;
1078}
1079
1080/* Try to defer sending, if possible, in order to minimize the amount
1081 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1082 *
1083 * This algorithm is from John Heffner.
1084 */
1085static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1086{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001087 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001088 u32 send_win, cong_win, limit, in_flight;
1089
1090 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1091 return 0;
1092
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001093 if (icsk->icsk_ca_state != TCP_CA_Open)
David S. Miller908a75c2005-07-05 15:43:58 -07001094 return 0;
1095
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001096 in_flight = tcp_packets_in_flight(tp);
1097
1098 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1099 (tp->snd_cwnd <= in_flight));
1100
1101 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1102
1103 /* From in_flight test above, we know that cwnd > in_flight. */
1104 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1105
1106 limit = min(send_win, cong_win);
1107
David S. Millerba244fe2006-03-11 18:51:49 -08001108 /* If a full-sized TSO skb can be sent, do it. */
1109 if (limit >= 65536)
1110 return 0;
1111
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001112 if (sysctl_tcp_tso_win_divisor) {
1113 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1114
1115 /* If at least some fraction of a window is available,
1116 * just use it.
1117 */
1118 chunk /= sysctl_tcp_tso_win_divisor;
1119 if (limit >= chunk)
1120 return 0;
1121 } else {
1122 /* Different approach, try not to defer past a single
1123 * ACK. Receiver should ACK every other full sized
1124 * frame, so if we have space for more than 3 frames
1125 * then send now.
1126 */
1127 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1128 return 0;
1129 }
1130
1131 /* Ok, it looks like it is advisable to defer. */
1132 return 1;
1133}
1134
John Heffner5d424d52006-03-20 17:53:41 -08001135/* Create a new MTU probe if we are ready.
1136 * Returns 0 if we should wait to probe (no cwnd available),
1137 * 1 if a probe was sent,
1138 * -1 otherwise */
1139static int tcp_mtu_probe(struct sock *sk)
1140{
1141 struct tcp_sock *tp = tcp_sk(sk);
1142 struct inet_connection_sock *icsk = inet_csk(sk);
1143 struct sk_buff *skb, *nskb, *next;
1144 int len;
1145 int probe_size;
1146 unsigned int pif;
1147 int copy;
1148 int mss_now;
1149
1150 /* Not currently probing/verifying,
1151 * not in recovery,
1152 * have enough cwnd, and
1153 * not SACKing (the variable headers throw things off) */
1154 if (!icsk->icsk_mtup.enabled ||
1155 icsk->icsk_mtup.probe_size ||
1156 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1157 tp->snd_cwnd < 11 ||
1158 tp->rx_opt.eff_sacks)
1159 return -1;
1160
1161 /* Very simple search strategy: just double the MSS. */
1162 mss_now = tcp_current_mss(sk, 0);
1163 probe_size = 2*tp->mss_cache;
1164 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1165 /* TODO: set timer for probe_converge_event */
1166 return -1;
1167 }
1168
1169 /* Have enough data in the send queue to probe? */
1170 len = 0;
1171 if ((skb = sk->sk_send_head) == NULL)
1172 return -1;
1173 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1174 skb = skb->next;
1175 if (len < probe_size)
1176 return -1;
1177
1178 /* Receive window check. */
1179 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1180 if (tp->snd_wnd < probe_size)
1181 return -1;
1182 else
1183 return 0;
1184 }
1185
1186 /* Do we need to wait to drain cwnd? */
1187 pif = tcp_packets_in_flight(tp);
1188 if (pif + 2 > tp->snd_cwnd) {
1189 /* With no packets in flight, don't stall. */
1190 if (pif == 0)
1191 return -1;
1192 else
1193 return 0;
1194 }
1195
1196 /* We're allowed to probe. Build it now. */
1197 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1198 return -1;
1199 sk_charge_skb(sk, nskb);
1200
1201 skb = sk->sk_send_head;
1202 __skb_insert(nskb, skb->prev, skb, &sk->sk_write_queue);
1203 sk->sk_send_head = nskb;
1204
1205 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1206 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1207 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1208 TCP_SKB_CB(nskb)->sacked = 0;
1209 nskb->csum = 0;
1210 if (skb->ip_summed == CHECKSUM_HW)
1211 nskb->ip_summed = CHECKSUM_HW;
1212
1213 len = 0;
1214 while (len < probe_size) {
1215 next = skb->next;
1216
1217 copy = min_t(int, skb->len, probe_size - len);
1218 if (nskb->ip_summed)
1219 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1220 else
1221 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1222 skb_put(nskb, copy), copy, nskb->csum);
1223
1224 if (skb->len <= copy) {
1225 /* We've eaten all the data from this skb.
1226 * Throw it away. */
1227 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1228 __skb_unlink(skb, &sk->sk_write_queue);
1229 sk_stream_free_skb(sk, skb);
1230 } else {
1231 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1232 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1233 if (!skb_shinfo(skb)->nr_frags) {
1234 skb_pull(skb, copy);
1235 if (skb->ip_summed != CHECKSUM_HW)
1236 skb->csum = csum_partial(skb->data, skb->len, 0);
1237 } else {
1238 __pskb_trim_head(skb, copy);
1239 tcp_set_skb_tso_segs(sk, skb, mss_now);
1240 }
1241 TCP_SKB_CB(skb)->seq += copy;
1242 }
1243
1244 len += copy;
1245 skb = next;
1246 }
1247 tcp_init_tso_segs(sk, nskb, nskb->len);
1248
1249 /* We're ready to send. If this fails, the probe will
1250 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1251 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1252 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1253 /* Decrement cwnd here because we are sending
1254 * effectively two packets. */
1255 tp->snd_cwnd--;
1256 update_send_head(sk, tp, nskb);
1257
1258 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001259 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1260 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001261
1262 return 1;
1263 }
1264
1265 return -1;
1266}
1267
1268
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269/* This routine writes packets to the network. It advances the
1270 * send_head. This happens as incoming acks open up the remote
1271 * window for us.
1272 *
1273 * Returns 1, if no segments are in flight and we have queued segments, but
1274 * cannot send anything now because of SWS or another problem.
1275 */
David S. Millera2e2a592005-07-05 15:19:23 -07001276static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277{
1278 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001279 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001280 unsigned int tso_segs, sent_pkts;
1281 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001282 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283
1284 /* If we are closed, the bytes will have to remain here.
1285 * In time closedown will finish, we empty the write queue and all
1286 * will be happy.
1287 */
David S. Miller92df7b52005-07-05 15:19:06 -07001288 if (unlikely(sk->sk_state == TCP_CLOSE))
1289 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
David S. Miller92df7b52005-07-05 15:19:06 -07001291 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001292
1293 /* Do MTU probing. */
1294 if ((result = tcp_mtu_probe(sk)) == 0) {
1295 return 0;
1296 } else if (result > 0) {
1297 sent_pkts = 1;
1298 }
1299
Herbert Xub68e9f82005-08-04 19:52:02 -07001300 while ((skb = sk->sk_send_head)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001301 unsigned int limit;
1302
Herbert Xub68e9f82005-08-04 19:52:02 -07001303 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001304 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001305
Herbert Xub68e9f82005-08-04 19:52:02 -07001306 cwnd_quota = tcp_cwnd_test(tp, skb);
1307 if (!cwnd_quota)
1308 break;
1309
1310 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1311 break;
1312
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001313 if (tso_segs == 1) {
1314 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1315 (tcp_skb_is_last(sk, skb) ?
1316 nonagle : TCP_NAGLE_PUSH))))
1317 break;
1318 } else {
1319 if (tcp_tso_should_defer(sk, tp, skb))
1320 break;
1321 }
David S. Milleraa934662005-07-05 15:20:09 -07001322
Herbert Xuc8ac3772005-08-16 20:43:40 -07001323 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001324 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001325 limit = tcp_window_allows(tp, skb,
1326 mss_now, cwnd_quota);
David S. Milleraa934662005-07-05 15:20:09 -07001327
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001328 if (skb->len < limit) {
1329 unsigned int trim = skb->len % mss_now;
1330
1331 if (trim)
1332 limit = skb->len - trim;
1333 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 }
1335
Herbert Xuc8ac3772005-08-16 20:43:40 -07001336 if (skb->len > limit &&
1337 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1338 break;
1339
David S. Miller92df7b52005-07-05 15:19:06 -07001340 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001341
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001342 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
David S. Miller92df7b52005-07-05 15:19:06 -07001343 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344
David S. Miller92df7b52005-07-05 15:19:06 -07001345 /* Advance the send_head. This one is sent out.
1346 * This call will increment packets_out.
1347 */
1348 update_send_head(sk, tp, skb);
1349
1350 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001351 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 }
David S. Miller92df7b52005-07-05 15:19:06 -07001353
David S. Milleraa934662005-07-05 15:20:09 -07001354 if (likely(sent_pkts)) {
David S. Miller92df7b52005-07-05 15:19:06 -07001355 tcp_cwnd_validate(sk, tp);
1356 return 0;
1357 }
David S. Miller92df7b52005-07-05 15:19:06 -07001358 return !tp->packets_out && sk->sk_send_head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359}
1360
David S. Millera762a982005-07-05 15:18:51 -07001361/* Push out any pending frames which were held back due to
1362 * TCP_CORK or attempt at coalescing tiny packets.
1363 * The socket must be locked by the caller.
1364 */
1365void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
David S. Millera2e2a592005-07-05 15:19:23 -07001366 unsigned int cur_mss, int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001367{
1368 struct sk_buff *skb = sk->sk_send_head;
1369
1370 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001371 if (tcp_write_xmit(sk, cur_mss, nonagle))
David S. Millera762a982005-07-05 15:18:51 -07001372 tcp_check_probe_timer(sk, tp);
1373 }
1374}
1375
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001376/* Send _single_ skb sitting at the send head. This function requires
1377 * true push pending frames to setup probe timer etc.
1378 */
1379void tcp_push_one(struct sock *sk, unsigned int mss_now)
1380{
1381 struct tcp_sock *tp = tcp_sk(sk);
1382 struct sk_buff *skb = sk->sk_send_head;
1383 unsigned int tso_segs, cwnd_quota;
1384
1385 BUG_ON(!skb || skb->len < mss_now);
1386
David S. Miller846998a2005-08-04 19:52:01 -07001387 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001388 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1389
1390 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001391 unsigned int limit;
1392
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001393 BUG_ON(!tso_segs);
1394
Herbert Xuc8ac3772005-08-16 20:43:40 -07001395 limit = mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001396 if (tso_segs > 1) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001397 limit = tcp_window_allows(tp, skb,
1398 mss_now, cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001399
1400 if (skb->len < limit) {
1401 unsigned int trim = skb->len % mss_now;
1402
1403 if (trim)
1404 limit = skb->len - trim;
1405 }
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001406 }
1407
Herbert Xuc8ac3772005-08-16 20:43:40 -07001408 if (skb->len > limit &&
1409 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1410 return;
1411
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001412 /* Send it out now. */
1413 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1414
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001415 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001416 update_send_head(sk, tp, skb);
1417 tcp_cwnd_validate(sk, tp);
1418 return;
1419 }
1420 }
1421}
1422
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423/* This function returns the amount that we can raise the
1424 * usable window based on the following constraints
1425 *
1426 * 1. The window can never be shrunk once it is offered (RFC 793)
1427 * 2. We limit memory per socket
1428 *
1429 * RFC 1122:
1430 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1431 * RECV.NEXT + RCV.WIN fixed until:
1432 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1433 *
1434 * i.e. don't raise the right edge of the window until you can raise
1435 * it at least MSS bytes.
1436 *
1437 * Unfortunately, the recommended algorithm breaks header prediction,
1438 * since header prediction assumes th->window stays fixed.
1439 *
1440 * Strictly speaking, keeping th->window fixed violates the receiver
1441 * side SWS prevention criteria. The problem is that under this rule
1442 * a stream of single byte packets will cause the right side of the
1443 * window to always advance by a single byte.
1444 *
1445 * Of course, if the sender implements sender side SWS prevention
1446 * then this will not be a problem.
1447 *
1448 * BSD seems to make the following compromise:
1449 *
1450 * If the free space is less than the 1/4 of the maximum
1451 * space available and the free space is less than 1/2 mss,
1452 * then set the window to 0.
1453 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1454 * Otherwise, just prevent the window from shrinking
1455 * and from being larger than the largest representable value.
1456 *
1457 * This prevents incremental opening of the window in the regime
1458 * where TCP is limited by the speed of the reader side taking
1459 * data out of the TCP receive queue. It does nothing about
1460 * those cases where the window is constrained on the sender side
1461 * because the pipeline is full.
1462 *
1463 * BSD also seems to "accidentally" limit itself to windows that are a
1464 * multiple of MSS, at least until the free space gets quite small.
1465 * This would appear to be a side effect of the mbuf implementation.
1466 * Combining these two algorithms results in the observed behavior
1467 * of having a fixed window size at almost all times.
1468 *
1469 * Below we obtain similar behavior by forcing the offered window to
1470 * a multiple of the mss when it is feasible to do so.
1471 *
1472 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1473 * Regular options like TIMESTAMP are taken into account.
1474 */
1475u32 __tcp_select_window(struct sock *sk)
1476{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001477 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001479 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 * here. I don't know if the value based on our guesses
1481 * of peer's MSS is better for the performance. It's more correct
1482 * but may be worse for the performance because of rcv_mss
1483 * fluctuations. --SAW 1998/11/1
1484 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001485 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 int free_space = tcp_space(sk);
1487 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1488 int window;
1489
1490 if (mss > full_space)
1491 mss = full_space;
1492
1493 if (free_space < full_space/2) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001494 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495
1496 if (tcp_memory_pressure)
1497 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1498
1499 if (free_space < mss)
1500 return 0;
1501 }
1502
1503 if (free_space > tp->rcv_ssthresh)
1504 free_space = tp->rcv_ssthresh;
1505
1506 /* Don't do rounding if we are using window scaling, since the
1507 * scaled window will not line up with the MSS boundary anyway.
1508 */
1509 window = tp->rcv_wnd;
1510 if (tp->rx_opt.rcv_wscale) {
1511 window = free_space;
1512
1513 /* Advertise enough space so that it won't get scaled away.
1514 * Import case: prevent zero window announcement if
1515 * 1<<rcv_wscale > mss.
1516 */
1517 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1518 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1519 << tp->rx_opt.rcv_wscale);
1520 } else {
1521 /* Get the largest window that is a nice multiple of mss.
1522 * Window clamp already applied above.
1523 * If our current window offering is within 1 mss of the
1524 * free space we just keep it. This prevents the divide
1525 * and multiply from happening most of the time.
1526 * We also don't do any window rounding when the free space
1527 * is too small.
1528 */
1529 if (window <= free_space - mss || window > free_space)
1530 window = (free_space/mss)*mss;
1531 }
1532
1533 return window;
1534}
1535
1536/* Attempt to collapse two adjacent SKB's during retransmission. */
1537static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1538{
1539 struct tcp_sock *tp = tcp_sk(sk);
1540 struct sk_buff *next_skb = skb->next;
1541
1542 /* The first test we must make is that neither of these two
1543 * SKB's are still referenced by someone else.
1544 */
1545 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1546 int skb_size = skb->len, next_skb_size = next_skb->len;
1547 u16 flags = TCP_SKB_CB(skb)->flags;
1548
1549 /* Also punt if next skb has been SACK'd. */
1550 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1551 return;
1552
1553 /* Next skb is out of window. */
1554 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1555 return;
1556
1557 /* Punt if not enough space exists in the first SKB for
1558 * the data in the second, or the total combined payload
1559 * would exceed the MSS.
1560 */
1561 if ((next_skb_size > skb_tailroom(skb)) ||
1562 ((skb_size + next_skb_size) > mss_now))
1563 return;
1564
1565 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1566 tcp_skb_pcount(next_skb) != 1);
1567
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001568 /* changing transmit queue under us so clear hints */
1569 clear_all_retrans_hints(tp);
1570
1571 /* Ok. We will be able to collapse the packet. */
David S. Miller8728b832005-08-09 19:25:21 -07001572 __skb_unlink(next_skb, &sk->sk_write_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573
1574 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1575
1576 if (next_skb->ip_summed == CHECKSUM_HW)
1577 skb->ip_summed = CHECKSUM_HW;
1578
1579 if (skb->ip_summed != CHECKSUM_HW)
1580 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1581
1582 /* Update sequence range on original skb. */
1583 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1584
1585 /* Merge over control information. */
1586 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1587 TCP_SKB_CB(skb)->flags = flags;
1588
1589 /* All done, get rid of second SKB and account for it so
1590 * packet counting does not break.
1591 */
1592 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1593 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1594 tp->retrans_out -= tcp_skb_pcount(next_skb);
1595 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1596 tp->lost_out -= tcp_skb_pcount(next_skb);
1597 tp->left_out -= tcp_skb_pcount(next_skb);
1598 }
1599 /* Reno case is special. Sigh... */
1600 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1601 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1602 tp->left_out -= tcp_skb_pcount(next_skb);
1603 }
1604
1605 /* Not quite right: it can be > snd.fack, but
1606 * it is better to underestimate fackets.
1607 */
1608 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1609 tcp_packets_out_dec(tp, next_skb);
1610 sk_stream_free_skb(sk, next_skb);
1611 }
1612}
1613
1614/* Do a simple retransmit without using the backoff mechanisms in
1615 * tcp_timer. This is used for path mtu discovery.
1616 * The socket is already locked here.
1617 */
1618void tcp_simple_retransmit(struct sock *sk)
1619{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001620 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 struct tcp_sock *tp = tcp_sk(sk);
1622 struct sk_buff *skb;
1623 unsigned int mss = tcp_current_mss(sk, 0);
1624 int lost = 0;
1625
1626 sk_stream_for_retrans_queue(skb, sk) {
1627 if (skb->len > mss &&
1628 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1629 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1630 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1631 tp->retrans_out -= tcp_skb_pcount(skb);
1632 }
1633 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1634 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1635 tp->lost_out += tcp_skb_pcount(skb);
1636 lost = 1;
1637 }
1638 }
1639 }
1640
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001641 clear_all_retrans_hints(tp);
1642
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 if (!lost)
1644 return;
1645
1646 tcp_sync_left_out(tp);
1647
1648 /* Don't muck with the congestion window here.
1649 * Reason is that we do not increase amount of _data_
1650 * in network, but units changed and effective
1651 * cwnd/ssthresh really reduced now.
1652 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001653 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001655 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 tp->prior_ssthresh = 0;
1657 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001658 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 }
1660 tcp_xmit_retransmit_queue(sk);
1661}
1662
1663/* This retransmits one SKB. Policy decisions and retransmit queue
1664 * state updates are done by the caller. Returns non-zero if an
1665 * error occurred which prevented the send.
1666 */
1667int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1668{
1669 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001670 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 unsigned int cur_mss = tcp_current_mss(sk, 0);
1672 int err;
1673
John Heffner5d424d52006-03-20 17:53:41 -08001674 /* Inconslusive MTU probe */
1675 if (icsk->icsk_mtup.probe_size) {
1676 icsk->icsk_mtup.probe_size = 0;
1677 }
1678
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001680 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 */
1682 if (atomic_read(&sk->sk_wmem_alloc) >
1683 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1684 return -EAGAIN;
1685
1686 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1687 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1688 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1690 return -ENOMEM;
1691 }
1692
1693 /* If receiver has shrunk his window, and skb is out of
1694 * new window, do not retransmit it. The exception is the
1695 * case, when window is shrunk to zero. In this case
1696 * our retransmit serves as a zero window probe.
1697 */
1698 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1699 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1700 return -EAGAIN;
1701
1702 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001703 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 }
1706
1707 /* Collapse two adjacent packets if worthwhile and we can. */
1708 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1709 (skb->len < (cur_mss >> 1)) &&
1710 (skb->next != sk->sk_send_head) &&
1711 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1712 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1713 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1714 (sysctl_tcp_retrans_collapse != 0))
1715 tcp_retrans_try_collapse(sk, skb, cur_mss);
1716
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001717 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 return -EHOSTUNREACH; /* Routing failure or similar. */
1719
1720 /* Some Solaris stacks overoptimize and ignore the FIN on a
1721 * retransmit when old data is attached. So strip it off
1722 * since it is cheap to do so and saves bytes on the network.
1723 */
1724 if(skb->len > 0 &&
1725 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1726 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1727 if (!pskb_trim(skb, 0)) {
1728 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
Herbert Xu79671682006-06-22 02:40:14 -07001729 skb_shinfo(skb)->gso_segs = 1;
1730 skb_shinfo(skb)->gso_size = 0;
1731 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 skb->ip_summed = CHECKSUM_NONE;
1733 skb->csum = 0;
1734 }
1735 }
1736
1737 /* Make a copy, if the first transmission SKB clone we made
1738 * is still in somebody's hands, else make a clone.
1739 */
1740 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001742 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
1744 if (err == 0) {
1745 /* Update global TCP statistics. */
1746 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1747
1748 tp->total_retrans++;
1749
1750#if FASTRETRANS_DEBUG > 0
1751 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1752 if (net_ratelimit())
1753 printk(KERN_DEBUG "retrans_out leaked.\n");
1754 }
1755#endif
1756 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1757 tp->retrans_out += tcp_skb_pcount(skb);
1758
1759 /* Save stamp of the first retransmit. */
1760 if (!tp->retrans_stamp)
1761 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1762
1763 tp->undo_retrans++;
1764
1765 /* snd_nxt is stored to detect loss of retransmitted segment,
1766 * see tcp_input.c tcp_sacktag_write_queue().
1767 */
1768 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1769 }
1770 return err;
1771}
1772
1773/* This gets called after a retransmit timeout, and the initially
1774 * retransmitted data is acknowledged. It tries to continue
1775 * resending the rest of the retransmit queue, until either
1776 * we've sent it all or the congestion window limit is reached.
1777 * If doing SACK, the first ACK which comes back for a timeout
1778 * based retransmit packet might feed us FACK information again.
1779 * If so, we use it to avoid unnecessarily retransmissions.
1780 */
1781void tcp_xmit_retransmit_queue(struct sock *sk)
1782{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001783 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 struct tcp_sock *tp = tcp_sk(sk);
1785 struct sk_buff *skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001786 int packet_cnt;
1787
1788 if (tp->retransmit_skb_hint) {
1789 skb = tp->retransmit_skb_hint;
1790 packet_cnt = tp->retransmit_cnt_hint;
1791 }else{
1792 skb = sk->sk_write_queue.next;
1793 packet_cnt = 0;
1794 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795
1796 /* First pass: retransmit lost packets. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001797 if (tp->lost_out) {
1798 sk_stream_for_retrans_queue_from(skb, sk) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1800
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001801 /* we could do better than to assign each time */
1802 tp->retransmit_skb_hint = skb;
1803 tp->retransmit_cnt_hint = packet_cnt;
1804
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 /* Assume this retransmit will generate
1806 * only one packet for congestion window
1807 * calculation purposes. This works because
1808 * tcp_retransmit_skb() will chop up the
1809 * packet to be MSS sized and all the
1810 * packet counting works out.
1811 */
1812 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1813 return;
1814
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001815 if (sacked & TCPCB_LOST) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001817 if (tcp_retransmit_skb(sk, skb)) {
1818 tp->retransmit_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001820 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001821 if (icsk->icsk_ca_state != TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1823 else
1824 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1825
1826 if (skb ==
1827 skb_peek(&sk->sk_write_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001828 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001829 inet_csk(sk)->icsk_rto,
1830 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 }
1832
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001833 packet_cnt += tcp_skb_pcount(skb);
1834 if (packet_cnt >= tp->lost_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 break;
1836 }
1837 }
1838 }
1839
1840 /* OK, demanded retransmission is finished. */
1841
1842 /* Forward retransmissions are possible only during Recovery. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001843 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 return;
1845
1846 /* No forward retransmissions in Reno are possible. */
1847 if (!tp->rx_opt.sack_ok)
1848 return;
1849
1850 /* Yeah, we have to make difficult choice between forward transmission
1851 * and retransmission... Both ways have their merits...
1852 *
1853 * For now we do not retransmit anything, while we have some new
1854 * segments to send.
1855 */
1856
1857 if (tcp_may_send_now(sk, tp))
1858 return;
1859
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001860 if (tp->forward_skb_hint) {
1861 skb = tp->forward_skb_hint;
1862 packet_cnt = tp->forward_cnt_hint;
1863 } else{
1864 skb = sk->sk_write_queue.next;
1865 packet_cnt = 0;
1866 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001868 sk_stream_for_retrans_queue_from(skb, sk) {
1869 tp->forward_cnt_hint = packet_cnt;
1870 tp->forward_skb_hint = skb;
1871
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 /* Similar to the retransmit loop above we
1873 * can pretend that the retransmitted SKB
1874 * we send out here will be composed of one
1875 * real MSS sized packet because tcp_retransmit_skb()
1876 * will fragment it if necessary.
1877 */
1878 if (++packet_cnt > tp->fackets_out)
1879 break;
1880
1881 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1882 break;
1883
1884 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1885 continue;
1886
1887 /* Ok, retransmit it. */
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001888 if (tcp_retransmit_skb(sk, skb)) {
1889 tp->forward_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001891 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892
1893 if (skb == skb_peek(&sk->sk_write_queue))
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07001894 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1895 inet_csk(sk)->icsk_rto,
1896 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897
1898 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1899 }
1900}
1901
1902
1903/* Send a fin. The caller locks the socket for us. This cannot be
1904 * allowed to fail queueing a FIN frame under any circumstances.
1905 */
1906void tcp_send_fin(struct sock *sk)
1907{
1908 struct tcp_sock *tp = tcp_sk(sk);
1909 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1910 int mss_now;
1911
1912 /* Optimization, tack on the FIN if we have a queue of
1913 * unsent frames. But be careful about outgoing SACKS
1914 * and IP options.
1915 */
1916 mss_now = tcp_current_mss(sk, 1);
1917
1918 if (sk->sk_send_head != NULL) {
1919 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1920 TCP_SKB_CB(skb)->end_seq++;
1921 tp->write_seq++;
1922 } else {
1923 /* Socket is locked, keep trying until memory is available. */
1924 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07001925 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 if (skb)
1927 break;
1928 yield();
1929 }
1930
1931 /* Reserve space for headers and prepare control bits. */
1932 skb_reserve(skb, MAX_TCP_HEADER);
1933 skb->csum = 0;
1934 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1935 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07001936 skb_shinfo(skb)->gso_segs = 1;
1937 skb_shinfo(skb)->gso_size = 0;
1938 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939
1940 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1941 TCP_SKB_CB(skb)->seq = tp->write_seq;
1942 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1943 tcp_queue_skb(sk, skb);
1944 }
1945 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1946}
1947
1948/* We get here when a process closes a file descriptor (either due to
1949 * an explicit close() or as a byproduct of exit()'ing) and there
1950 * was unread data in the receive queue. This behavior is recommended
1951 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1952 */
Al Virodd0fc662005-10-07 07:46:04 +01001953void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954{
1955 struct tcp_sock *tp = tcp_sk(sk);
1956 struct sk_buff *skb;
1957
1958 /* NOTE: No TCP options attached and we never retransmit this. */
1959 skb = alloc_skb(MAX_TCP_HEADER, priority);
1960 if (!skb) {
1961 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1962 return;
1963 }
1964
1965 /* Reserve space for headers and prepare control bits. */
1966 skb_reserve(skb, MAX_TCP_HEADER);
1967 skb->csum = 0;
1968 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1969 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07001970 skb_shinfo(skb)->gso_segs = 1;
1971 skb_shinfo(skb)->gso_size = 0;
1972 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973
1974 /* Send it off. */
1975 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1976 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1977 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001978 if (tcp_transmit_skb(sk, skb, 0, priority))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1980}
1981
1982/* WARNING: This routine must only be called when we have already sent
1983 * a SYN packet that crossed the incoming SYN that caused this routine
1984 * to get called. If this assumption fails then the initial rcv_wnd
1985 * and rcv_wscale values will not be correct.
1986 */
1987int tcp_send_synack(struct sock *sk)
1988{
1989 struct sk_buff* skb;
1990
1991 skb = skb_peek(&sk->sk_write_queue);
1992 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1993 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1994 return -EFAULT;
1995 }
1996 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1997 if (skb_cloned(skb)) {
1998 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1999 if (nskb == NULL)
2000 return -ENOMEM;
2001 __skb_unlink(skb, &sk->sk_write_queue);
2002 skb_header_release(nskb);
2003 __skb_queue_head(&sk->sk_write_queue, nskb);
2004 sk_stream_free_skb(sk, skb);
2005 sk_charge_skb(sk, nskb);
2006 skb = nskb;
2007 }
2008
2009 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2010 TCP_ECN_send_synack(tcp_sk(sk), skb);
2011 }
2012 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002013 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014}
2015
2016/*
2017 * Prepare a SYN-ACK.
2018 */
2019struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
Arnaldo Carvalho de Melo60236fd2005-06-18 22:47:21 -07002020 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002022 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 struct tcp_sock *tp = tcp_sk(sk);
2024 struct tcphdr *th;
2025 int tcp_header_size;
2026 struct sk_buff *skb;
2027
2028 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2029 if (skb == NULL)
2030 return NULL;
2031
2032 /* Reserve space for headers. */
2033 skb_reserve(skb, MAX_TCP_HEADER);
2034
2035 skb->dst = dst_clone(dst);
2036
2037 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002038 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2039 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 /* SACK_PERM is in the place of NOP NOP of TS */
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002041 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
2043
2044 memset(th, 0, sizeof(struct tcphdr));
2045 th->syn = 1;
2046 th->ack = 1;
2047 if (dst->dev->features&NETIF_F_TSO)
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002048 ireq->ecn_ok = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 TCP_ECN_make_synack(req, th);
2050 th->source = inet_sk(sk)->sport;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002051 th->dest = ireq->rmt_port;
2052 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2054 TCP_SKB_CB(skb)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002055 skb_shinfo(skb)->gso_segs = 1;
2056 skb_shinfo(skb)->gso_size = 0;
2057 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002059 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2061 __u8 rcv_wscale;
2062 /* Set this up on the first call only */
2063 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2064 /* tcp_full_space because it is guaranteed to be the first packet */
2065 tcp_select_initial_window(tcp_full_space(sk),
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002066 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 &req->rcv_wnd,
2068 &req->window_clamp,
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002069 ireq->wscale_ok,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 &rcv_wscale);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002071 ireq->rcv_wscale = rcv_wscale;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 }
2073
2074 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2075 th->window = htons(req->rcv_wnd);
2076
2077 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002078 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2079 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 TCP_SKB_CB(skb)->when,
2081 req->ts_recent);
2082
2083 skb->csum = 0;
2084 th->doff = (tcp_header_size >> 2);
2085 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2086 return skb;
2087}
2088
2089/*
2090 * Do all connect socket setups that can be done AF independent.
2091 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002092static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093{
2094 struct dst_entry *dst = __sk_dst_get(sk);
2095 struct tcp_sock *tp = tcp_sk(sk);
2096 __u8 rcv_wscale;
2097
2098 /* We'll fix this up when we get a response from the other end.
2099 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2100 */
2101 tp->tcp_header_len = sizeof(struct tcphdr) +
2102 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2103
2104 /* If user gave his TCP_MAXSEG, record it to clamp */
2105 if (tp->rx_opt.user_mss)
2106 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2107 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002108 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 tcp_sync_mss(sk, dst_mtu(dst));
2110
2111 if (!tp->window_clamp)
2112 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2113 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2114 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115
2116 tcp_select_initial_window(tcp_full_space(sk),
2117 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2118 &tp->rcv_wnd,
2119 &tp->window_clamp,
2120 sysctl_tcp_window_scaling,
2121 &rcv_wscale);
2122
2123 tp->rx_opt.rcv_wscale = rcv_wscale;
2124 tp->rcv_ssthresh = tp->rcv_wnd;
2125
2126 sk->sk_err = 0;
2127 sock_reset_flag(sk, SOCK_DONE);
2128 tp->snd_wnd = 0;
2129 tcp_init_wl(tp, tp->write_seq, 0);
2130 tp->snd_una = tp->write_seq;
2131 tp->snd_sml = tp->write_seq;
2132 tp->rcv_nxt = 0;
2133 tp->rcv_wup = 0;
2134 tp->copied_seq = 0;
2135
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002136 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2137 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 tcp_clear_retrans(tp);
2139}
2140
2141/*
2142 * Build a SYN and send it off.
2143 */
2144int tcp_connect(struct sock *sk)
2145{
2146 struct tcp_sock *tp = tcp_sk(sk);
2147 struct sk_buff *buff;
2148
2149 tcp_connect_init(sk);
2150
David S. Millerd179cd12005-08-17 14:57:30 -07002151 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 if (unlikely(buff == NULL))
2153 return -ENOBUFS;
2154
2155 /* Reserve space for headers. */
2156 skb_reserve(buff, MAX_TCP_HEADER);
2157
2158 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2159 TCP_ECN_send_syn(sk, tp, buff);
2160 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002161 skb_shinfo(buff)->gso_segs = 1;
2162 skb_shinfo(buff)->gso_size = 0;
2163 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 buff->csum = 0;
2165 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2166 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2167 tp->snd_nxt = tp->write_seq;
2168 tp->pushed_seq = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169
2170 /* Send it off. */
2171 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2172 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2173 skb_header_release(buff);
2174 __skb_queue_tail(&sk->sk_write_queue, buff);
2175 sk_charge_skb(sk, buff);
2176 tp->packets_out += tcp_skb_pcount(buff);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002177 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2179
2180 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002181 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2182 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 return 0;
2184}
2185
2186/* Send out a delayed ack, the caller does the policy checking
2187 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2188 * for details.
2189 */
2190void tcp_send_delayed_ack(struct sock *sk)
2191{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002192 struct inet_connection_sock *icsk = inet_csk(sk);
2193 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 unsigned long timeout;
2195
2196 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002197 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 int max_ato = HZ/2;
2199
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002200 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 max_ato = TCP_DELACK_MAX;
2202
2203 /* Slow path, intersegment interval is "high". */
2204
2205 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002206 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 * directly.
2208 */
2209 if (tp->srtt) {
2210 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2211
2212 if (rtt < max_ato)
2213 max_ato = rtt;
2214 }
2215
2216 ato = min(ato, max_ato);
2217 }
2218
2219 /* Stay within the limit we were given */
2220 timeout = jiffies + ato;
2221
2222 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002223 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 /* If delack timer was blocked or is about to expire,
2225 * send ACK now.
2226 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002227 if (icsk->icsk_ack.blocked ||
2228 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 tcp_send_ack(sk);
2230 return;
2231 }
2232
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002233 if (!time_before(timeout, icsk->icsk_ack.timeout))
2234 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002236 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2237 icsk->icsk_ack.timeout = timeout;
2238 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239}
2240
2241/* This routine sends an ack and also updates the window. */
2242void tcp_send_ack(struct sock *sk)
2243{
2244 /* If we have been reset, we may not send again. */
2245 if (sk->sk_state != TCP_CLOSE) {
2246 struct tcp_sock *tp = tcp_sk(sk);
2247 struct sk_buff *buff;
2248
2249 /* We are not putting this on the write queue, so
2250 * tcp_transmit_skb() will set the ownership to this
2251 * sock.
2252 */
2253 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2254 if (buff == NULL) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002255 inet_csk_schedule_ack(sk);
2256 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002257 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2258 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 return;
2260 }
2261
2262 /* Reserve space for headers and prepare control bits. */
2263 skb_reserve(buff, MAX_TCP_HEADER);
2264 buff->csum = 0;
2265 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2266 TCP_SKB_CB(buff)->sacked = 0;
Herbert Xu79671682006-06-22 02:40:14 -07002267 skb_shinfo(buff)->gso_segs = 1;
2268 skb_shinfo(buff)->gso_size = 0;
2269 skb_shinfo(buff)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270
2271 /* Send it off, this clears delayed acks for us. */
2272 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2273 TCP_SKB_CB(buff)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002274 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 }
2276}
2277
2278/* This routine sends a packet with an out of date sequence
2279 * number. It assumes the other end will try to ack it.
2280 *
2281 * Question: what should we make while urgent mode?
2282 * 4.4BSD forces sending single byte of data. We cannot send
2283 * out of window data, because we have SND.NXT==SND.MAX...
2284 *
2285 * Current solution: to send TWO zero-length segments in urgent mode:
2286 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2287 * out-of-date with SND.UNA-1 to probe window.
2288 */
2289static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2290{
2291 struct tcp_sock *tp = tcp_sk(sk);
2292 struct sk_buff *skb;
2293
2294 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2295 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2296 if (skb == NULL)
2297 return -1;
2298
2299 /* Reserve space for headers and set control bits. */
2300 skb_reserve(skb, MAX_TCP_HEADER);
2301 skb->csum = 0;
2302 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2303 TCP_SKB_CB(skb)->sacked = urgent;
Herbert Xu79671682006-06-22 02:40:14 -07002304 skb_shinfo(skb)->gso_segs = 1;
2305 skb_shinfo(skb)->gso_size = 0;
2306 skb_shinfo(skb)->gso_type = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307
2308 /* Use a previous sequence. This should cause the other
2309 * end to send an ack. Don't queue or clone SKB, just
2310 * send it.
2311 */
2312 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2313 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2314 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002315 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316}
2317
2318int tcp_write_wakeup(struct sock *sk)
2319{
2320 if (sk->sk_state != TCP_CLOSE) {
2321 struct tcp_sock *tp = tcp_sk(sk);
2322 struct sk_buff *skb;
2323
2324 if ((skb = sk->sk_send_head) != NULL &&
2325 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2326 int err;
2327 unsigned int mss = tcp_current_mss(sk, 0);
2328 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2329
2330 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2331 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2332
2333 /* We are probing the opening of a window
2334 * but the window size is != 0
2335 * must have been a result SWS avoidance ( sender )
2336 */
2337 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2338 skb->len > mss) {
2339 seg_size = min(seg_size, mss);
2340 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
David S. Miller846998a2005-08-04 19:52:01 -07002341 if (tcp_fragment(sk, skb, seg_size, mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 } else if (!tcp_skb_pcount(skb))
David S. Miller846998a2005-08-04 19:52:01 -07002344 tcp_set_skb_tso_segs(sk, skb, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345
2346 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2347 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002348 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 if (!err) {
2350 update_send_head(sk, tp, skb);
2351 }
2352 return err;
2353 } else {
2354 if (tp->urg_mode &&
2355 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2356 tcp_xmit_probe_skb(sk, TCPCB_URG);
2357 return tcp_xmit_probe_skb(sk, 0);
2358 }
2359 }
2360 return -1;
2361}
2362
2363/* A window probe timeout has occurred. If window is not closed send
2364 * a partial packet else a zero probe.
2365 */
2366void tcp_send_probe0(struct sock *sk)
2367{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002368 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 struct tcp_sock *tp = tcp_sk(sk);
2370 int err;
2371
2372 err = tcp_write_wakeup(sk);
2373
2374 if (tp->packets_out || !sk->sk_send_head) {
2375 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002376 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002377 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 return;
2379 }
2380
2381 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002382 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2383 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002384 icsk->icsk_probes_out++;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002385 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002386 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2387 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 } else {
2389 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002390 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 * Let local senders to fight for local resources.
2392 *
2393 * Use accumulated backoff yet.
2394 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002395 if (!icsk->icsk_probes_out)
2396 icsk->icsk_probes_out = 1;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002397 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2398 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002399 TCP_RESOURCE_PROBE_INTERVAL),
2400 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 }
2402}
2403
2404EXPORT_SYMBOL(tcp_connect);
2405EXPORT_SYMBOL(tcp_make_synack);
2406EXPORT_SYMBOL(tcp_simple_retransmit);
2407EXPORT_SYMBOL(tcp_sync_mss);
Stephen Hemmingerf4805ed2005-11-10 16:53:30 -08002408EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
John Heffner5d424d52006-03-20 17:53:41 -08002409EXPORT_SYMBOL(tcp_mtup_init);