blob: 53a8a5399f1e6e6de3b13a05d691612c533f433c [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_input.c,v 1.243 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:
25 * Pedro Roque : Fast Retransmit/Recovery.
26 * Two receive queues.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
30 * Header prediction.
31 * Variable renaming.
32 *
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presnce of
46 * timestamps.
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
49 * data segments.
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
57 * fast path.
58 * J Hadi Salim: ECN support
59 * Andrei Gurtov,
60 * Pasi Sarolahti,
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 */
65
66#include <linux/config.h>
67#include <linux/mm.h>
68#include <linux/module.h>
69#include <linux/sysctl.h>
70#include <net/tcp.h>
71#include <net/inet_common.h>
72#include <linux/ipsec.h>
73#include <asm/unaligned.h>
74
75int sysctl_tcp_timestamps = 1;
76int sysctl_tcp_window_scaling = 1;
77int sysctl_tcp_sack = 1;
78int sysctl_tcp_fack = 1;
79int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
80int sysctl_tcp_ecn;
81int sysctl_tcp_dsack = 1;
82int sysctl_tcp_app_win = 31;
83int sysctl_tcp_adv_win_scale = 2;
84
85int sysctl_tcp_stdurg;
86int sysctl_tcp_rfc1337;
87int sysctl_tcp_max_orphans = NR_FILE;
88int sysctl_tcp_frto;
89int sysctl_tcp_nometrics_save;
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
91int sysctl_tcp_moderate_rcvbuf = 1;
92
Linus Torvalds1da177e2005-04-16 15:20:36 -070093#define FLAG_DATA 0x01 /* Incoming frame contained data. */
94#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
95#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
96#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
97#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
98#define FLAG_DATA_SACKED 0x20 /* New SACK. */
99#define FLAG_ECE 0x40 /* ECE in this ACK */
100#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
101#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
102
103#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
106#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
107
108#define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
109#define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
110#define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
111
112#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
113
114/* Adapt the MSS value used to make delayed ack decision to the
115 * real world.
116 */
117static inline void tcp_measure_rcv_mss(struct tcp_sock *tp,
118 struct sk_buff *skb)
119{
120 unsigned int len, lss;
121
122 lss = tp->ack.last_seg_size;
123 tp->ack.last_seg_size = 0;
124
125 /* skb->len may jitter because of SACKs, even if peer
126 * sends good full-sized frames.
127 */
128 len = skb->len;
129 if (len >= tp->ack.rcv_mss) {
130 tp->ack.rcv_mss = len;
131 } else {
132 /* Otherwise, we make more careful check taking into account,
133 * that SACKs block is variable.
134 *
135 * "len" is invariant segment length, including TCP header.
136 */
137 len += skb->data - skb->h.raw;
138 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
139 /* If PSH is not set, packet should be
140 * full sized, provided peer TCP is not badly broken.
141 * This observation (if it is correct 8)) allows
142 * to handle super-low mtu links fairly.
143 */
144 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
145 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
146 /* Subtract also invariant (if peer is RFC compliant),
147 * tcp header plus fixed timestamp option length.
148 * Resulting "len" is MSS free of SACK jitter.
149 */
150 len -= tp->tcp_header_len;
151 tp->ack.last_seg_size = len;
152 if (len == lss) {
153 tp->ack.rcv_mss = len;
154 return;
155 }
156 }
157 tp->ack.pending |= TCP_ACK_PUSHED;
158 }
159}
160
161static void tcp_incr_quickack(struct tcp_sock *tp)
162{
163 unsigned quickacks = tp->rcv_wnd/(2*tp->ack.rcv_mss);
164
165 if (quickacks==0)
166 quickacks=2;
167 if (quickacks > tp->ack.quick)
168 tp->ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
169}
170
171void tcp_enter_quickack_mode(struct tcp_sock *tp)
172{
173 tcp_incr_quickack(tp);
174 tp->ack.pingpong = 0;
175 tp->ack.ato = TCP_ATO_MIN;
176}
177
178/* Send ACKs quickly, if "quick" count is not exhausted
179 * and the session is not interactive.
180 */
181
182static __inline__ int tcp_in_quickack_mode(struct tcp_sock *tp)
183{
184 return (tp->ack.quick && !tp->ack.pingpong);
185}
186
187/* Buffer size and advertised window tuning.
188 *
189 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
190 */
191
192static void tcp_fixup_sndbuf(struct sock *sk)
193{
194 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
195 sizeof(struct sk_buff);
196
197 if (sk->sk_sndbuf < 3 * sndmem)
198 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
199}
200
201/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
202 *
203 * All tcp_full_space() is split to two parts: "network" buffer, allocated
204 * forward and advertised in receiver window (tp->rcv_wnd) and
205 * "application buffer", required to isolate scheduling/application
206 * latencies from network.
207 * window_clamp is maximal advertised window. It can be less than
208 * tcp_full_space(), in this case tcp_full_space() - window_clamp
209 * is reserved for "application" buffer. The less window_clamp is
210 * the smoother our behaviour from viewpoint of network, but the lower
211 * throughput and the higher sensitivity of the connection to losses. 8)
212 *
213 * rcv_ssthresh is more strict window_clamp used at "slow start"
214 * phase to predict further behaviour of this connection.
215 * It is used for two goals:
216 * - to enforce header prediction at sender, even when application
217 * requires some significant "application buffer". It is check #1.
218 * - to prevent pruning of receive queue because of misprediction
219 * of receiver window. Check #2.
220 *
221 * The scheme does not work when sender sends good segments opening
222 * window and then starts to feed us spagetti. But it should work
223 * in common situations. Otherwise, we have to rely on queue collapsing.
224 */
225
226/* Slow part of check#2. */
227static int __tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
228 struct sk_buff *skb)
229{
230 /* Optimize this! */
231 int truesize = tcp_win_from_space(skb->truesize)/2;
232 int window = tcp_full_space(sk)/2;
233
234 while (tp->rcv_ssthresh <= window) {
235 if (truesize <= skb->len)
236 return 2*tp->ack.rcv_mss;
237
238 truesize >>= 1;
239 window >>= 1;
240 }
241 return 0;
242}
243
244static inline void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
245 struct sk_buff *skb)
246{
247 /* Check #1 */
248 if (tp->rcv_ssthresh < tp->window_clamp &&
249 (int)tp->rcv_ssthresh < tcp_space(sk) &&
250 !tcp_memory_pressure) {
251 int incr;
252
253 /* Check #2. Increase window, if skb with such overhead
254 * will fit to rcvbuf in future.
255 */
256 if (tcp_win_from_space(skb->truesize) <= skb->len)
257 incr = 2*tp->advmss;
258 else
259 incr = __tcp_grow_window(sk, tp, skb);
260
261 if (incr) {
262 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
263 tp->ack.quick |= 1;
264 }
265 }
266}
267
268/* 3. Tuning rcvbuf, when connection enters established state. */
269
270static void tcp_fixup_rcvbuf(struct sock *sk)
271{
272 struct tcp_sock *tp = tcp_sk(sk);
273 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
274
275 /* Try to select rcvbuf so that 4 mss-sized segments
276 * will fit to window and correspoding skbs will fit to our rcvbuf.
277 * (was 3; 4 is minimum to allow fast retransmit to work.)
278 */
279 while (tcp_win_from_space(rcvmem) < tp->advmss)
280 rcvmem += 128;
281 if (sk->sk_rcvbuf < 4 * rcvmem)
282 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
283}
284
285/* 4. Try to fixup all. It is made iimediately after connection enters
286 * established state.
287 */
288static void tcp_init_buffer_space(struct sock *sk)
289{
290 struct tcp_sock *tp = tcp_sk(sk);
291 int maxwin;
292
293 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
294 tcp_fixup_rcvbuf(sk);
295 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
296 tcp_fixup_sndbuf(sk);
297
298 tp->rcvq_space.space = tp->rcv_wnd;
299
300 maxwin = tcp_full_space(sk);
301
302 if (tp->window_clamp >= maxwin) {
303 tp->window_clamp = maxwin;
304
305 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
306 tp->window_clamp = max(maxwin -
307 (maxwin >> sysctl_tcp_app_win),
308 4 * tp->advmss);
309 }
310
311 /* Force reservation of one segment. */
312 if (sysctl_tcp_app_win &&
313 tp->window_clamp > 2 * tp->advmss &&
314 tp->window_clamp + tp->advmss > maxwin)
315 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
316
317 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
318 tp->snd_cwnd_stamp = tcp_time_stamp;
319}
320
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321/* 5. Recalculate window clamp after socket hit its memory bounds. */
322static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
323{
324 struct sk_buff *skb;
325 unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
326 int ofo_win = 0;
327
328 tp->ack.quick = 0;
329
330 skb_queue_walk(&tp->out_of_order_queue, skb) {
331 ofo_win += skb->len;
332 }
333
334 /* If overcommit is due to out of order segments,
335 * do not clamp window. Try to expand rcvbuf instead.
336 */
337 if (ofo_win) {
338 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
339 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
340 !tcp_memory_pressure &&
341 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
342 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
343 sysctl_tcp_rmem[2]);
344 }
345 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
346 app_win += ofo_win;
347 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
348 app_win >>= 1;
349 if (app_win > tp->ack.rcv_mss)
350 app_win -= tp->ack.rcv_mss;
351 app_win = max(app_win, 2U*tp->advmss);
352
353 if (!ofo_win)
354 tp->window_clamp = min(tp->window_clamp, app_win);
355 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
356 }
357}
358
359/* Receiver "autotuning" code.
360 *
361 * The algorithm for RTT estimation w/o timestamps is based on
362 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
363 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
364 *
365 * More detail on this code can be found at
366 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
367 * though this reference is out of date. A new paper
368 * is pending.
369 */
370static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
371{
372 u32 new_sample = tp->rcv_rtt_est.rtt;
373 long m = sample;
374
375 if (m == 0)
376 m = 1;
377
378 if (new_sample != 0) {
379 /* If we sample in larger samples in the non-timestamp
380 * case, we could grossly overestimate the RTT especially
381 * with chatty applications or bulk transfer apps which
382 * are stalled on filesystem I/O.
383 *
384 * Also, since we are only going for a minimum in the
385 * non-timestamp case, we do not smoothe things out
386 * else with timestamps disabled convergance takes too
387 * long.
388 */
389 if (!win_dep) {
390 m -= (new_sample >> 3);
391 new_sample += m;
392 } else if (m < new_sample)
393 new_sample = m << 3;
394 } else {
395 /* No previous mesaure. */
396 new_sample = m << 3;
397 }
398
399 if (tp->rcv_rtt_est.rtt != new_sample)
400 tp->rcv_rtt_est.rtt = new_sample;
401}
402
403static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
404{
405 if (tp->rcv_rtt_est.time == 0)
406 goto new_measure;
407 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
408 return;
409 tcp_rcv_rtt_update(tp,
410 jiffies - tp->rcv_rtt_est.time,
411 1);
412
413new_measure:
414 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
415 tp->rcv_rtt_est.time = tcp_time_stamp;
416}
417
418static inline void tcp_rcv_rtt_measure_ts(struct tcp_sock *tp, struct sk_buff *skb)
419{
420 if (tp->rx_opt.rcv_tsecr &&
421 (TCP_SKB_CB(skb)->end_seq -
422 TCP_SKB_CB(skb)->seq >= tp->ack.rcv_mss))
423 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
424}
425
426/*
427 * This function should be called every time data is copied to user space.
428 * It calculates the appropriate TCP receive buffer space.
429 */
430void tcp_rcv_space_adjust(struct sock *sk)
431{
432 struct tcp_sock *tp = tcp_sk(sk);
433 int time;
434 int space;
435
436 if (tp->rcvq_space.time == 0)
437 goto new_measure;
438
439 time = tcp_time_stamp - tp->rcvq_space.time;
440 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
441 tp->rcv_rtt_est.rtt == 0)
442 return;
443
444 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
445
446 space = max(tp->rcvq_space.space, space);
447
448 if (tp->rcvq_space.space != space) {
449 int rcvmem;
450
451 tp->rcvq_space.space = space;
452
453 if (sysctl_tcp_moderate_rcvbuf) {
454 int new_clamp = space;
455
456 /* Receive space grows, normalize in order to
457 * take into account packet headers and sk_buff
458 * structure overhead.
459 */
460 space /= tp->advmss;
461 if (!space)
462 space = 1;
463 rcvmem = (tp->advmss + MAX_TCP_HEADER +
464 16 + sizeof(struct sk_buff));
465 while (tcp_win_from_space(rcvmem) < tp->advmss)
466 rcvmem += 128;
467 space *= rcvmem;
468 space = min(space, sysctl_tcp_rmem[2]);
469 if (space > sk->sk_rcvbuf) {
470 sk->sk_rcvbuf = space;
471
472 /* Make the window clamp follow along. */
473 tp->window_clamp = new_clamp;
474 }
475 }
476 }
477
478new_measure:
479 tp->rcvq_space.seq = tp->copied_seq;
480 tp->rcvq_space.time = tcp_time_stamp;
481}
482
483/* There is something which you must keep in mind when you analyze the
484 * behavior of the tp->ato delayed ack timeout interval. When a
485 * connection starts up, we want to ack as quickly as possible. The
486 * problem is that "good" TCP's do slow start at the beginning of data
487 * transmission. The means that until we send the first few ACK's the
488 * sender will sit on his end and only queue most of his data, because
489 * he can only send snd_cwnd unacked packets at any given time. For
490 * each ACK we send, he increments snd_cwnd and transmits more of his
491 * queue. -DaveM
492 */
493static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
494{
495 u32 now;
496
497 tcp_schedule_ack(tp);
498
499 tcp_measure_rcv_mss(tp, skb);
500
501 tcp_rcv_rtt_measure(tp);
502
503 now = tcp_time_stamp;
504
505 if (!tp->ack.ato) {
506 /* The _first_ data packet received, initialize
507 * delayed ACK engine.
508 */
509 tcp_incr_quickack(tp);
510 tp->ack.ato = TCP_ATO_MIN;
511 } else {
512 int m = now - tp->ack.lrcvtime;
513
514 if (m <= TCP_ATO_MIN/2) {
515 /* The fastest case is the first. */
516 tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2;
517 } else if (m < tp->ack.ato) {
518 tp->ack.ato = (tp->ack.ato>>1) + m;
519 if (tp->ack.ato > tp->rto)
520 tp->ack.ato = tp->rto;
521 } else if (m > tp->rto) {
522 /* Too long gap. Apparently sender falled to
523 * restart window, so that we send ACKs quickly.
524 */
525 tcp_incr_quickack(tp);
526 sk_stream_mem_reclaim(sk);
527 }
528 }
529 tp->ack.lrcvtime = now;
530
531 TCP_ECN_check_ce(tp, skb);
532
533 if (skb->len >= 128)
534 tcp_grow_window(sk, tp, skb);
535}
536
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537/* Called to compute a smoothed rtt estimate. The data fed to this
538 * routine either comes from timestamps, or from segments that were
539 * known _not_ to have been retransmitted [see Karn/Partridge
540 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
541 * piece by Van Jacobson.
542 * NOTE: the next three routines used to be one big routine.
543 * To save cycles in the RFC 1323 implementation it was better to break
544 * it up into three procedures. -- erics
545 */
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700546static void tcp_rtt_estimator(struct tcp_sock *tp, __u32 mrtt, u32 *usrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547{
548 long m = mrtt; /* RTT */
549
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 /* The following amusing code comes from Jacobson's
551 * article in SIGCOMM '88. Note that rtt and mdev
552 * are scaled versions of rtt and mean deviation.
553 * This is designed to be as fast as possible
554 * m stands for "measurement".
555 *
556 * On a 1990 paper the rto value is changed to:
557 * RTO = rtt + 4 * mdev
558 *
559 * Funny. This algorithm seems to be very broken.
560 * These formulae increase RTO, when it should be decreased, increase
561 * too slowly, when it should be incresed fastly, decrease too fastly
562 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
563 * does not matter how to _calculate_ it. Seems, it was trap
564 * that VJ failed to avoid. 8)
565 */
566 if(m == 0)
567 m = 1;
568 if (tp->srtt != 0) {
569 m -= (tp->srtt >> 3); /* m is now error in rtt est */
570 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
571 if (m < 0) {
572 m = -m; /* m is now abs(error) */
573 m -= (tp->mdev >> 2); /* similar update on mdev */
574 /* This is similar to one of Eifel findings.
575 * Eifel blocks mdev updates when rtt decreases.
576 * This solution is a bit different: we use finer gain
577 * for mdev in this case (alpha*beta).
578 * Like Eifel it also prevents growth of rto,
579 * but also it limits too fast rto decreases,
580 * happening in pure Eifel.
581 */
582 if (m > 0)
583 m >>= 3;
584 } else {
585 m -= (tp->mdev >> 2); /* similar update on mdev */
586 }
587 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
588 if (tp->mdev > tp->mdev_max) {
589 tp->mdev_max = tp->mdev;
590 if (tp->mdev_max > tp->rttvar)
591 tp->rttvar = tp->mdev_max;
592 }
593 if (after(tp->snd_una, tp->rtt_seq)) {
594 if (tp->mdev_max < tp->rttvar)
595 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
596 tp->rtt_seq = tp->snd_nxt;
597 tp->mdev_max = TCP_RTO_MIN;
598 }
599 } else {
600 /* no previous measure. */
601 tp->srtt = m<<3; /* take the measured time to be rtt */
602 tp->mdev = m<<1; /* make sure rto = 3*rtt */
603 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
604 tp->rtt_seq = tp->snd_nxt;
605 }
606
Stephen Hemminger317a76f2005-06-23 12:19:55 -0700607 if (tp->ca_ops->rtt_sample)
608 tp->ca_ops->rtt_sample(tp, *usrtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609}
610
611/* Calculate rto without backoff. This is the second half of Van Jacobson's
612 * routine referred to above.
613 */
614static inline void tcp_set_rto(struct tcp_sock *tp)
615{
616 /* Old crap is replaced with new one. 8)
617 *
618 * More seriously:
619 * 1. If rtt variance happened to be less 50msec, it is hallucination.
620 * It cannot be less due to utterly erratic ACK generation made
621 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
622 * to do with delayed acks, because at cwnd>2 true delack timeout
623 * is invisible. Actually, Linux-2.4 also generates erratic
624 * ACKs in some curcumstances.
625 */
626 tp->rto = (tp->srtt >> 3) + tp->rttvar;
627
628 /* 2. Fixups made earlier cannot be right.
629 * If we do not estimate RTO correctly without them,
630 * all the algo is pure shit and should be replaced
631 * with correct one. It is exaclty, which we pretend to do.
632 */
633}
634
635/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
636 * guarantees that rto is higher.
637 */
638static inline void tcp_bound_rto(struct tcp_sock *tp)
639{
640 if (tp->rto > TCP_RTO_MAX)
641 tp->rto = TCP_RTO_MAX;
642}
643
644/* Save metrics learned by this TCP session.
645 This function is called only, when TCP finishes successfully
646 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
647 */
648void tcp_update_metrics(struct sock *sk)
649{
650 struct tcp_sock *tp = tcp_sk(sk);
651 struct dst_entry *dst = __sk_dst_get(sk);
652
653 if (sysctl_tcp_nometrics_save)
654 return;
655
656 dst_confirm(dst);
657
658 if (dst && (dst->flags&DST_HOST)) {
659 int m;
660
661 if (tp->backoff || !tp->srtt) {
662 /* This session failed to estimate rtt. Why?
663 * Probably, no packets returned in time.
664 * Reset our results.
665 */
666 if (!(dst_metric_locked(dst, RTAX_RTT)))
667 dst->metrics[RTAX_RTT-1] = 0;
668 return;
669 }
670
671 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
672
673 /* If newly calculated rtt larger than stored one,
674 * store new one. Otherwise, use EWMA. Remember,
675 * rtt overestimation is always better than underestimation.
676 */
677 if (!(dst_metric_locked(dst, RTAX_RTT))) {
678 if (m <= 0)
679 dst->metrics[RTAX_RTT-1] = tp->srtt;
680 else
681 dst->metrics[RTAX_RTT-1] -= (m>>3);
682 }
683
684 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
685 if (m < 0)
686 m = -m;
687
688 /* Scale deviation to rttvar fixed point */
689 m >>= 1;
690 if (m < tp->mdev)
691 m = tp->mdev;
692
693 if (m >= dst_metric(dst, RTAX_RTTVAR))
694 dst->metrics[RTAX_RTTVAR-1] = m;
695 else
696 dst->metrics[RTAX_RTTVAR-1] -=
697 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
698 }
699
700 if (tp->snd_ssthresh >= 0xFFFF) {
701 /* Slow start still did not finish. */
702 if (dst_metric(dst, RTAX_SSTHRESH) &&
703 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
704 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
705 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
706 if (!dst_metric_locked(dst, RTAX_CWND) &&
707 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
708 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
709 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
710 tp->ca_state == TCP_CA_Open) {
711 /* Cong. avoidance phase, cwnd is reliable. */
712 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
713 dst->metrics[RTAX_SSTHRESH-1] =
714 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
715 if (!dst_metric_locked(dst, RTAX_CWND))
716 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
717 } else {
718 /* Else slow start did not finish, cwnd is non-sense,
719 ssthresh may be also invalid.
720 */
721 if (!dst_metric_locked(dst, RTAX_CWND))
722 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
723 if (dst->metrics[RTAX_SSTHRESH-1] &&
724 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
725 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
726 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
727 }
728
729 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
730 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
731 tp->reordering != sysctl_tcp_reordering)
732 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
733 }
734 }
735}
736
737/* Numbers are taken from RFC2414. */
738__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
739{
740 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
741
742 if (!cwnd) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700743 if (tp->mss_cache > 1460)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 cwnd = 2;
745 else
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700746 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 }
748 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
749}
750
751/* Initialize metrics on socket. */
752
753static void tcp_init_metrics(struct sock *sk)
754{
755 struct tcp_sock *tp = tcp_sk(sk);
756 struct dst_entry *dst = __sk_dst_get(sk);
757
758 if (dst == NULL)
759 goto reset;
760
761 dst_confirm(dst);
762
763 if (dst_metric_locked(dst, RTAX_CWND))
764 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
765 if (dst_metric(dst, RTAX_SSTHRESH)) {
766 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
767 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
768 tp->snd_ssthresh = tp->snd_cwnd_clamp;
769 }
770 if (dst_metric(dst, RTAX_REORDERING) &&
771 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
772 tp->rx_opt.sack_ok &= ~2;
773 tp->reordering = dst_metric(dst, RTAX_REORDERING);
774 }
775
776 if (dst_metric(dst, RTAX_RTT) == 0)
777 goto reset;
778
779 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
780 goto reset;
781
782 /* Initial rtt is determined from SYN,SYN-ACK.
783 * The segment is small and rtt may appear much
784 * less than real one. Use per-dst memory
785 * to make it more realistic.
786 *
787 * A bit of theory. RTT is time passed after "normal" sized packet
788 * is sent until it is ACKed. In normal curcumstances sending small
789 * packets force peer to delay ACKs and calculation is correct too.
790 * The algorithm is adaptive and, provided we follow specs, it
791 * NEVER underestimate RTT. BUT! If peer tries to make some clever
792 * tricks sort of "quick acks" for time long enough to decrease RTT
793 * to low value, and then abruptly stops to do it and starts to delay
794 * ACKs, wait for troubles.
795 */
796 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
797 tp->srtt = dst_metric(dst, RTAX_RTT);
798 tp->rtt_seq = tp->snd_nxt;
799 }
800 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
801 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
802 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
803 }
804 tcp_set_rto(tp);
805 tcp_bound_rto(tp);
806 if (tp->rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
807 goto reset;
808 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
809 tp->snd_cwnd_stamp = tcp_time_stamp;
810 return;
811
812reset:
813 /* Play conservative. If timestamps are not
814 * supported, TCP will fail to recalculate correct
815 * rtt, if initial rto is too small. FORGET ALL AND RESET!
816 */
817 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
818 tp->srtt = 0;
819 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
820 tp->rto = TCP_TIMEOUT_INIT;
821 }
822}
823
824static void tcp_update_reordering(struct tcp_sock *tp, int metric, int ts)
825{
826 if (metric > tp->reordering) {
827 tp->reordering = min(TCP_MAX_REORDERING, metric);
828
829 /* This exciting event is worth to be remembered. 8) */
830 if (ts)
831 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
832 else if (IsReno(tp))
833 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
834 else if (IsFack(tp))
835 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
836 else
837 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
838#if FASTRETRANS_DEBUG > 1
839 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
840 tp->rx_opt.sack_ok, tp->ca_state,
841 tp->reordering,
842 tp->fackets_out,
843 tp->sacked_out,
844 tp->undo_marker ? tp->undo_retrans : 0);
845#endif
846 /* Disable FACK yet. */
847 tp->rx_opt.sack_ok &= ~2;
848 }
849}
850
851/* This procedure tags the retransmission queue when SACKs arrive.
852 *
853 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
854 * Packets in queue with these bits set are counted in variables
855 * sacked_out, retrans_out and lost_out, correspondingly.
856 *
857 * Valid combinations are:
858 * Tag InFlight Description
859 * 0 1 - orig segment is in flight.
860 * S 0 - nothing flies, orig reached receiver.
861 * L 0 - nothing flies, orig lost by net.
862 * R 2 - both orig and retransmit are in flight.
863 * L|R 1 - orig is lost, retransmit is in flight.
864 * S|R 1 - orig reached receiver, retrans is still in flight.
865 * (L|S|R is logically valid, it could occur when L|R is sacked,
866 * but it is equivalent to plain S and code short-curcuits it to S.
867 * L|S is logically invalid, it would mean -1 packet in flight 8))
868 *
869 * These 6 states form finite state machine, controlled by the following events:
870 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
871 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
872 * 3. Loss detection event of one of three flavors:
873 * A. Scoreboard estimator decided the packet is lost.
874 * A'. Reno "three dupacks" marks head of queue lost.
875 * A''. Its FACK modfication, head until snd.fack is lost.
876 * B. SACK arrives sacking data transmitted after never retransmitted
877 * hole was sent out.
878 * C. SACK arrives sacking SND.NXT at the moment, when the
879 * segment was retransmitted.
880 * 4. D-SACK added new rule: D-SACK changes any tag to S.
881 *
882 * It is pleasant to note, that state diagram turns out to be commutative,
883 * so that we are allowed not to be bothered by order of our actions,
884 * when multiple events arrive simultaneously. (see the function below).
885 *
886 * Reordering detection.
887 * --------------------
888 * Reordering metric is maximal distance, which a packet can be displaced
889 * in packet stream. With SACKs we can estimate it:
890 *
891 * 1. SACK fills old hole and the corresponding segment was not
892 * ever retransmitted -> reordering. Alas, we cannot use it
893 * when segment was retransmitted.
894 * 2. The last flaw is solved with D-SACK. D-SACK arrives
895 * for retransmitted and already SACKed segment -> reordering..
896 * Both of these heuristics are not used in Loss state, when we cannot
897 * account for retransmits accurately.
898 */
899static int
900tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
901{
902 struct tcp_sock *tp = tcp_sk(sk);
903 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
904 struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
905 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
906 int reord = tp->packets_out;
907 int prior_fackets;
908 u32 lost_retrans = 0;
909 int flag = 0;
910 int i;
911
912 /* So, SACKs for already sent large segments will be lost.
913 * Not good, but alternative is to resegment the queue. */
914 if (sk->sk_route_caps & NETIF_F_TSO) {
915 sk->sk_route_caps &= ~NETIF_F_TSO;
916 sock_set_flag(sk, SOCK_NO_LARGESEND);
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700917 tp->mss_cache = tp->mss_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 }
919
920 if (!tp->sacked_out)
921 tp->fackets_out = 0;
922 prior_fackets = tp->fackets_out;
923
924 for (i=0; i<num_sacks; i++, sp++) {
925 struct sk_buff *skb;
926 __u32 start_seq = ntohl(sp->start_seq);
927 __u32 end_seq = ntohl(sp->end_seq);
928 int fack_count = 0;
929 int dup_sack = 0;
930
931 /* Check for D-SACK. */
932 if (i == 0) {
933 u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
934
935 if (before(start_seq, ack)) {
936 dup_sack = 1;
937 tp->rx_opt.sack_ok |= 4;
938 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
939 } else if (num_sacks > 1 &&
940 !after(end_seq, ntohl(sp[1].end_seq)) &&
941 !before(start_seq, ntohl(sp[1].start_seq))) {
942 dup_sack = 1;
943 tp->rx_opt.sack_ok |= 4;
944 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
945 }
946
947 /* D-SACK for already forgotten data...
948 * Do dumb counting. */
949 if (dup_sack &&
950 !after(end_seq, prior_snd_una) &&
951 after(end_seq, tp->undo_marker))
952 tp->undo_retrans--;
953
954 /* Eliminate too old ACKs, but take into
955 * account more or less fresh ones, they can
956 * contain valid SACK info.
957 */
958 if (before(ack, prior_snd_una - tp->max_window))
959 return 0;
960 }
961
962 /* Event "B" in the comment above. */
963 if (after(end_seq, tp->high_seq))
964 flag |= FLAG_DATA_LOST;
965
966 sk_stream_for_retrans_queue(skb, sk) {
967 u8 sacked = TCP_SKB_CB(skb)->sacked;
968 int in_sack;
969
970 /* The retransmission queue is always in order, so
971 * we can short-circuit the walk early.
972 */
973 if(!before(TCP_SKB_CB(skb)->seq, end_seq))
974 break;
975
976 fack_count += tcp_skb_pcount(skb);
977
978 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
979 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
980
981 /* Account D-SACK for retransmitted packet. */
982 if ((dup_sack && in_sack) &&
983 (sacked & TCPCB_RETRANS) &&
984 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
985 tp->undo_retrans--;
986
987 /* The frame is ACKed. */
988 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
989 if (sacked&TCPCB_RETRANS) {
990 if ((dup_sack && in_sack) &&
991 (sacked&TCPCB_SACKED_ACKED))
992 reord = min(fack_count, reord);
993 } else {
994 /* If it was in a hole, we detected reordering. */
995 if (fack_count < prior_fackets &&
996 !(sacked&TCPCB_SACKED_ACKED))
997 reord = min(fack_count, reord);
998 }
999
1000 /* Nothing to do; acked frame is about to be dropped. */
1001 continue;
1002 }
1003
1004 if ((sacked&TCPCB_SACKED_RETRANS) &&
1005 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1006 (!lost_retrans || after(end_seq, lost_retrans)))
1007 lost_retrans = end_seq;
1008
1009 if (!in_sack)
1010 continue;
1011
1012 if (!(sacked&TCPCB_SACKED_ACKED)) {
1013 if (sacked & TCPCB_SACKED_RETRANS) {
1014 /* If the segment is not tagged as lost,
1015 * we do not clear RETRANS, believing
1016 * that retransmission is still in flight.
1017 */
1018 if (sacked & TCPCB_LOST) {
1019 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1020 tp->lost_out -= tcp_skb_pcount(skb);
1021 tp->retrans_out -= tcp_skb_pcount(skb);
1022 }
1023 } else {
1024 /* New sack for not retransmitted frame,
1025 * which was in hole. It is reordering.
1026 */
1027 if (!(sacked & TCPCB_RETRANS) &&
1028 fack_count < prior_fackets)
1029 reord = min(fack_count, reord);
1030
1031 if (sacked & TCPCB_LOST) {
1032 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1033 tp->lost_out -= tcp_skb_pcount(skb);
1034 }
1035 }
1036
1037 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1038 flag |= FLAG_DATA_SACKED;
1039 tp->sacked_out += tcp_skb_pcount(skb);
1040
1041 if (fack_count > tp->fackets_out)
1042 tp->fackets_out = fack_count;
1043 } else {
1044 if (dup_sack && (sacked&TCPCB_RETRANS))
1045 reord = min(fack_count, reord);
1046 }
1047
1048 /* D-SACK. We can detect redundant retransmission
1049 * in S|R and plain R frames and clear it.
1050 * undo_retrans is decreased above, L|R frames
1051 * are accounted above as well.
1052 */
1053 if (dup_sack &&
1054 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1055 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1056 tp->retrans_out -= tcp_skb_pcount(skb);
1057 }
1058 }
1059 }
1060
1061 /* Check for lost retransmit. This superb idea is
1062 * borrowed from "ratehalving". Event "C".
1063 * Later note: FACK people cheated me again 8),
1064 * we have to account for reordering! Ugly,
1065 * but should help.
1066 */
1067 if (lost_retrans && tp->ca_state == TCP_CA_Recovery) {
1068 struct sk_buff *skb;
1069
1070 sk_stream_for_retrans_queue(skb, sk) {
1071 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1072 break;
1073 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1074 continue;
1075 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1076 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1077 (IsFack(tp) ||
1078 !before(lost_retrans,
1079 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001080 tp->mss_cache))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1082 tp->retrans_out -= tcp_skb_pcount(skb);
1083
1084 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1085 tp->lost_out += tcp_skb_pcount(skb);
1086 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1087 flag |= FLAG_DATA_SACKED;
1088 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1089 }
1090 }
1091 }
1092 }
1093
1094 tp->left_out = tp->sacked_out + tp->lost_out;
1095
1096 if ((reord < tp->fackets_out) && tp->ca_state != TCP_CA_Loss)
1097 tcp_update_reordering(tp, ((tp->fackets_out + 1) - reord), 0);
1098
1099#if FASTRETRANS_DEBUG > 0
1100 BUG_TRAP((int)tp->sacked_out >= 0);
1101 BUG_TRAP((int)tp->lost_out >= 0);
1102 BUG_TRAP((int)tp->retrans_out >= 0);
1103 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1104#endif
1105 return flag;
1106}
1107
1108/* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1109 * segments to see from the next ACKs whether any data was really missing.
1110 * If the RTO was spurious, new ACKs should arrive.
1111 */
1112void tcp_enter_frto(struct sock *sk)
1113{
1114 struct tcp_sock *tp = tcp_sk(sk);
1115 struct sk_buff *skb;
1116
1117 tp->frto_counter = 1;
1118
1119 if (tp->ca_state <= TCP_CA_Disorder ||
1120 tp->snd_una == tp->high_seq ||
1121 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1122 tp->prior_ssthresh = tcp_current_ssthresh(tp);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001123 tp->snd_ssthresh = tp->ca_ops->ssthresh(tp);
1124 tcp_ca_event(tp, CA_EVENT_FRTO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 }
1126
1127 /* Have to clear retransmission markers here to keep the bookkeeping
1128 * in shape, even though we are not yet in Loss state.
1129 * If something was really lost, it is eventually caught up
1130 * in tcp_enter_frto_loss.
1131 */
1132 tp->retrans_out = 0;
1133 tp->undo_marker = tp->snd_una;
1134 tp->undo_retrans = 0;
1135
1136 sk_stream_for_retrans_queue(skb, sk) {
1137 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1138 }
1139 tcp_sync_left_out(tp);
1140
1141 tcp_set_ca_state(tp, TCP_CA_Open);
1142 tp->frto_highmark = tp->snd_nxt;
1143}
1144
1145/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1146 * which indicates that we should follow the traditional RTO recovery,
1147 * i.e. mark everything lost and do go-back-N retransmission.
1148 */
1149static void tcp_enter_frto_loss(struct sock *sk)
1150{
1151 struct tcp_sock *tp = tcp_sk(sk);
1152 struct sk_buff *skb;
1153 int cnt = 0;
1154
1155 tp->sacked_out = 0;
1156 tp->lost_out = 0;
1157 tp->fackets_out = 0;
1158
1159 sk_stream_for_retrans_queue(skb, sk) {
1160 cnt += tcp_skb_pcount(skb);
1161 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1162 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1163
1164 /* Do not mark those segments lost that were
1165 * forward transmitted after RTO
1166 */
1167 if (!after(TCP_SKB_CB(skb)->end_seq,
1168 tp->frto_highmark)) {
1169 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1170 tp->lost_out += tcp_skb_pcount(skb);
1171 }
1172 } else {
1173 tp->sacked_out += tcp_skb_pcount(skb);
1174 tp->fackets_out = cnt;
1175 }
1176 }
1177 tcp_sync_left_out(tp);
1178
1179 tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1180 tp->snd_cwnd_cnt = 0;
1181 tp->snd_cwnd_stamp = tcp_time_stamp;
1182 tp->undo_marker = 0;
1183 tp->frto_counter = 0;
1184
1185 tp->reordering = min_t(unsigned int, tp->reordering,
1186 sysctl_tcp_reordering);
1187 tcp_set_ca_state(tp, TCP_CA_Loss);
1188 tp->high_seq = tp->frto_highmark;
1189 TCP_ECN_queue_cwr(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190}
1191
1192void tcp_clear_retrans(struct tcp_sock *tp)
1193{
1194 tp->left_out = 0;
1195 tp->retrans_out = 0;
1196
1197 tp->fackets_out = 0;
1198 tp->sacked_out = 0;
1199 tp->lost_out = 0;
1200
1201 tp->undo_marker = 0;
1202 tp->undo_retrans = 0;
1203}
1204
1205/* Enter Loss state. If "how" is not zero, forget all SACK information
1206 * and reset tags completely, otherwise preserve SACKs. If receiver
1207 * dropped its ofo queue, we will know this due to reneging detection.
1208 */
1209void tcp_enter_loss(struct sock *sk, int how)
1210{
1211 struct tcp_sock *tp = tcp_sk(sk);
1212 struct sk_buff *skb;
1213 int cnt = 0;
1214
1215 /* Reduce ssthresh if it has not yet been made inside this window. */
1216 if (tp->ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1217 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1218 tp->prior_ssthresh = tcp_current_ssthresh(tp);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001219 tp->snd_ssthresh = tp->ca_ops->ssthresh(tp);
1220 tcp_ca_event(tp, CA_EVENT_LOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 }
1222 tp->snd_cwnd = 1;
1223 tp->snd_cwnd_cnt = 0;
1224 tp->snd_cwnd_stamp = tcp_time_stamp;
1225
1226 tcp_clear_retrans(tp);
1227
1228 /* Push undo marker, if it was plain RTO and nothing
1229 * was retransmitted. */
1230 if (!how)
1231 tp->undo_marker = tp->snd_una;
1232
1233 sk_stream_for_retrans_queue(skb, sk) {
1234 cnt += tcp_skb_pcount(skb);
1235 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1236 tp->undo_marker = 0;
1237 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1238 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1239 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1240 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1241 tp->lost_out += tcp_skb_pcount(skb);
1242 } else {
1243 tp->sacked_out += tcp_skb_pcount(skb);
1244 tp->fackets_out = cnt;
1245 }
1246 }
1247 tcp_sync_left_out(tp);
1248
1249 tp->reordering = min_t(unsigned int, tp->reordering,
1250 sysctl_tcp_reordering);
1251 tcp_set_ca_state(tp, TCP_CA_Loss);
1252 tp->high_seq = tp->snd_nxt;
1253 TCP_ECN_queue_cwr(tp);
1254}
1255
1256static int tcp_check_sack_reneging(struct sock *sk, struct tcp_sock *tp)
1257{
1258 struct sk_buff *skb;
1259
1260 /* If ACK arrived pointing to a remembered SACK,
1261 * it means that our remembered SACKs do not reflect
1262 * real state of receiver i.e.
1263 * receiver _host_ is heavily congested (or buggy).
1264 * Do processing similar to RTO timeout.
1265 */
1266 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1267 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1268 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1269
1270 tcp_enter_loss(sk, 1);
1271 tp->retransmits++;
1272 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1273 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1274 return 1;
1275 }
1276 return 0;
1277}
1278
1279static inline int tcp_fackets_out(struct tcp_sock *tp)
1280{
1281 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1282}
1283
1284static inline int tcp_skb_timedout(struct tcp_sock *tp, struct sk_buff *skb)
1285{
1286 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto);
1287}
1288
1289static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1290{
1291 return tp->packets_out &&
1292 tcp_skb_timedout(tp, skb_peek(&sk->sk_write_queue));
1293}
1294
1295/* Linux NewReno/SACK/FACK/ECN state machine.
1296 * --------------------------------------
1297 *
1298 * "Open" Normal state, no dubious events, fast path.
1299 * "Disorder" In all the respects it is "Open",
1300 * but requires a bit more attention. It is entered when
1301 * we see some SACKs or dupacks. It is split of "Open"
1302 * mainly to move some processing from fast path to slow one.
1303 * "CWR" CWND was reduced due to some Congestion Notification event.
1304 * It can be ECN, ICMP source quench, local device congestion.
1305 * "Recovery" CWND was reduced, we are fast-retransmitting.
1306 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1307 *
1308 * tcp_fastretrans_alert() is entered:
1309 * - each incoming ACK, if state is not "Open"
1310 * - when arrived ACK is unusual, namely:
1311 * * SACK
1312 * * Duplicate ACK.
1313 * * ECN ECE.
1314 *
1315 * Counting packets in flight is pretty simple.
1316 *
1317 * in_flight = packets_out - left_out + retrans_out
1318 *
1319 * packets_out is SND.NXT-SND.UNA counted in packets.
1320 *
1321 * retrans_out is number of retransmitted segments.
1322 *
1323 * left_out is number of segments left network, but not ACKed yet.
1324 *
1325 * left_out = sacked_out + lost_out
1326 *
1327 * sacked_out: Packets, which arrived to receiver out of order
1328 * and hence not ACKed. With SACKs this number is simply
1329 * amount of SACKed data. Even without SACKs
1330 * it is easy to give pretty reliable estimate of this number,
1331 * counting duplicate ACKs.
1332 *
1333 * lost_out: Packets lost by network. TCP has no explicit
1334 * "loss notification" feedback from network (for now).
1335 * It means that this number can be only _guessed_.
1336 * Actually, it is the heuristics to predict lossage that
1337 * distinguishes different algorithms.
1338 *
1339 * F.e. after RTO, when all the queue is considered as lost,
1340 * lost_out = packets_out and in_flight = retrans_out.
1341 *
1342 * Essentially, we have now two algorithms counting
1343 * lost packets.
1344 *
1345 * FACK: It is the simplest heuristics. As soon as we decided
1346 * that something is lost, we decide that _all_ not SACKed
1347 * packets until the most forward SACK are lost. I.e.
1348 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1349 * It is absolutely correct estimate, if network does not reorder
1350 * packets. And it loses any connection to reality when reordering
1351 * takes place. We use FACK by default until reordering
1352 * is suspected on the path to this destination.
1353 *
1354 * NewReno: when Recovery is entered, we assume that one segment
1355 * is lost (classic Reno). While we are in Recovery and
1356 * a partial ACK arrives, we assume that one more packet
1357 * is lost (NewReno). This heuristics are the same in NewReno
1358 * and SACK.
1359 *
1360 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1361 * deflation etc. CWND is real congestion window, never inflated, changes
1362 * only according to classic VJ rules.
1363 *
1364 * Really tricky (and requiring careful tuning) part of algorithm
1365 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1366 * The first determines the moment _when_ we should reduce CWND and,
1367 * hence, slow down forward transmission. In fact, it determines the moment
1368 * when we decide that hole is caused by loss, rather than by a reorder.
1369 *
1370 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1371 * holes, caused by lost packets.
1372 *
1373 * And the most logically complicated part of algorithm is undo
1374 * heuristics. We detect false retransmits due to both too early
1375 * fast retransmit (reordering) and underestimated RTO, analyzing
1376 * timestamps and D-SACKs. When we detect that some segments were
1377 * retransmitted by mistake and CWND reduction was wrong, we undo
1378 * window reduction and abort recovery phase. This logic is hidden
1379 * inside several functions named tcp_try_undo_<something>.
1380 */
1381
1382/* This function decides, when we should leave Disordered state
1383 * and enter Recovery phase, reducing congestion window.
1384 *
1385 * Main question: may we further continue forward transmission
1386 * with the same cwnd?
1387 */
1388static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1389{
1390 __u32 packets_out;
1391
1392 /* Trick#1: The loss is proven. */
1393 if (tp->lost_out)
1394 return 1;
1395
1396 /* Not-A-Trick#2 : Classic rule... */
1397 if (tcp_fackets_out(tp) > tp->reordering)
1398 return 1;
1399
1400 /* Trick#3 : when we use RFC2988 timer restart, fast
1401 * retransmit can be triggered by timeout of queue head.
1402 */
1403 if (tcp_head_timedout(sk, tp))
1404 return 1;
1405
1406 /* Trick#4: It is still not OK... But will it be useful to delay
1407 * recovery more?
1408 */
1409 packets_out = tp->packets_out;
1410 if (packets_out <= tp->reordering &&
1411 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1412 !tcp_may_send_now(sk, tp)) {
1413 /* We have nothing to send. This connection is limited
1414 * either by receiver window or by application.
1415 */
1416 return 1;
1417 }
1418
1419 return 0;
1420}
1421
1422/* If we receive more dupacks than we expected counting segments
1423 * in assumption of absent reordering, interpret this as reordering.
1424 * The only another reason could be bug in receiver TCP.
1425 */
1426static void tcp_check_reno_reordering(struct tcp_sock *tp, int addend)
1427{
1428 u32 holes;
1429
1430 holes = max(tp->lost_out, 1U);
1431 holes = min(holes, tp->packets_out);
1432
1433 if ((tp->sacked_out + holes) > tp->packets_out) {
1434 tp->sacked_out = tp->packets_out - holes;
1435 tcp_update_reordering(tp, tp->packets_out+addend, 0);
1436 }
1437}
1438
1439/* Emulate SACKs for SACKless connection: account for a new dupack. */
1440
1441static void tcp_add_reno_sack(struct tcp_sock *tp)
1442{
1443 tp->sacked_out++;
1444 tcp_check_reno_reordering(tp, 0);
1445 tcp_sync_left_out(tp);
1446}
1447
1448/* Account for ACK, ACKing some data in Reno Recovery phase. */
1449
1450static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1451{
1452 if (acked > 0) {
1453 /* One ACK acked hole. The rest eat duplicate ACKs. */
1454 if (acked-1 >= tp->sacked_out)
1455 tp->sacked_out = 0;
1456 else
1457 tp->sacked_out -= acked-1;
1458 }
1459 tcp_check_reno_reordering(tp, acked);
1460 tcp_sync_left_out(tp);
1461}
1462
1463static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1464{
1465 tp->sacked_out = 0;
1466 tp->left_out = tp->lost_out;
1467}
1468
1469/* Mark head of queue up as lost. */
1470static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1471 int packets, u32 high_seq)
1472{
1473 struct sk_buff *skb;
1474 int cnt = packets;
1475
1476 BUG_TRAP(cnt <= tp->packets_out);
1477
1478 sk_stream_for_retrans_queue(skb, sk) {
1479 cnt -= tcp_skb_pcount(skb);
1480 if (cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1481 break;
1482 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1483 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1484 tp->lost_out += tcp_skb_pcount(skb);
1485 }
1486 }
1487 tcp_sync_left_out(tp);
1488}
1489
1490/* Account newly detected lost packet(s) */
1491
1492static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1493{
1494 if (IsFack(tp)) {
1495 int lost = tp->fackets_out - tp->reordering;
1496 if (lost <= 0)
1497 lost = 1;
1498 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1499 } else {
1500 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1501 }
1502
1503 /* New heuristics: it is possible only after we switched
1504 * to restart timer each time when something is ACKed.
1505 * Hence, we can detect timed out packets during fast
1506 * retransmit without falling to slow start.
1507 */
1508 if (tcp_head_timedout(sk, tp)) {
1509 struct sk_buff *skb;
1510
1511 sk_stream_for_retrans_queue(skb, sk) {
1512 if (tcp_skb_timedout(tp, skb) &&
1513 !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1514 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1515 tp->lost_out += tcp_skb_pcount(skb);
1516 }
1517 }
1518 tcp_sync_left_out(tp);
1519 }
1520}
1521
1522/* CWND moderation, preventing bursts due to too big ACKs
1523 * in dubious situations.
1524 */
1525static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1526{
1527 tp->snd_cwnd = min(tp->snd_cwnd,
1528 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1529 tp->snd_cwnd_stamp = tcp_time_stamp;
1530}
1531
1532/* Decrease cwnd each second ack. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533static void tcp_cwnd_down(struct tcp_sock *tp)
1534{
1535 int decr = tp->snd_cwnd_cnt + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
1537 tp->snd_cwnd_cnt = decr&1;
1538 decr >>= 1;
1539
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001540 if (decr && tp->snd_cwnd > tp->ca_ops->min_cwnd(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 tp->snd_cwnd -= decr;
1542
1543 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1544 tp->snd_cwnd_stamp = tcp_time_stamp;
1545}
1546
1547/* Nothing was retransmitted or returned timestamp is less
1548 * than timestamp of the first retransmission.
1549 */
1550static inline int tcp_packet_delayed(struct tcp_sock *tp)
1551{
1552 return !tp->retrans_stamp ||
1553 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1554 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1555}
1556
1557/* Undo procedures. */
1558
1559#if FASTRETRANS_DEBUG > 1
1560static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1561{
1562 struct inet_sock *inet = inet_sk(sk);
1563 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1564 msg,
1565 NIPQUAD(inet->daddr), ntohs(inet->dport),
1566 tp->snd_cwnd, tp->left_out,
1567 tp->snd_ssthresh, tp->prior_ssthresh,
1568 tp->packets_out);
1569}
1570#else
1571#define DBGUNDO(x...) do { } while (0)
1572#endif
1573
1574static void tcp_undo_cwr(struct tcp_sock *tp, int undo)
1575{
1576 if (tp->prior_ssthresh) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001577 if (tp->ca_ops->undo_cwnd)
1578 tp->snd_cwnd = tp->ca_ops->undo_cwnd(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 else
1580 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1581
1582 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1583 tp->snd_ssthresh = tp->prior_ssthresh;
1584 TCP_ECN_withdraw_cwr(tp);
1585 }
1586 } else {
1587 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1588 }
1589 tcp_moderate_cwnd(tp);
1590 tp->snd_cwnd_stamp = tcp_time_stamp;
1591}
1592
1593static inline int tcp_may_undo(struct tcp_sock *tp)
1594{
1595 return tp->undo_marker &&
1596 (!tp->undo_retrans || tcp_packet_delayed(tp));
1597}
1598
1599/* People celebrate: "We love our President!" */
1600static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1601{
1602 if (tcp_may_undo(tp)) {
1603 /* Happy end! We did not retransmit anything
1604 * or our original transmission succeeded.
1605 */
1606 DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans");
1607 tcp_undo_cwr(tp, 1);
1608 if (tp->ca_state == TCP_CA_Loss)
1609 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1610 else
1611 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1612 tp->undo_marker = 0;
1613 }
1614 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1615 /* Hold old state until something *above* high_seq
1616 * is ACKed. For Reno it is MUST to prevent false
1617 * fast retransmits (RFC2582). SACK TCP is safe. */
1618 tcp_moderate_cwnd(tp);
1619 return 1;
1620 }
1621 tcp_set_ca_state(tp, TCP_CA_Open);
1622 return 0;
1623}
1624
1625/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1626static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1627{
1628 if (tp->undo_marker && !tp->undo_retrans) {
1629 DBGUNDO(sk, tp, "D-SACK");
1630 tcp_undo_cwr(tp, 1);
1631 tp->undo_marker = 0;
1632 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1633 }
1634}
1635
1636/* Undo during fast recovery after partial ACK. */
1637
1638static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1639 int acked)
1640{
1641 /* Partial ACK arrived. Force Hoe's retransmit. */
1642 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1643
1644 if (tcp_may_undo(tp)) {
1645 /* Plain luck! Hole if filled with delayed
1646 * packet, rather than with a retransmit.
1647 */
1648 if (tp->retrans_out == 0)
1649 tp->retrans_stamp = 0;
1650
1651 tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1);
1652
1653 DBGUNDO(sk, tp, "Hoe");
1654 tcp_undo_cwr(tp, 0);
1655 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1656
1657 /* So... Do not make Hoe's retransmit yet.
1658 * If the first packet was delayed, the rest
1659 * ones are most probably delayed as well.
1660 */
1661 failed = 0;
1662 }
1663 return failed;
1664}
1665
1666/* Undo during loss recovery after partial ACK. */
1667static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1668{
1669 if (tcp_may_undo(tp)) {
1670 struct sk_buff *skb;
1671 sk_stream_for_retrans_queue(skb, sk) {
1672 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1673 }
1674 DBGUNDO(sk, tp, "partial loss");
1675 tp->lost_out = 0;
1676 tp->left_out = tp->sacked_out;
1677 tcp_undo_cwr(tp, 1);
1678 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1679 tp->retransmits = 0;
1680 tp->undo_marker = 0;
1681 if (!IsReno(tp))
1682 tcp_set_ca_state(tp, TCP_CA_Open);
1683 return 1;
1684 }
1685 return 0;
1686}
1687
1688static inline void tcp_complete_cwr(struct tcp_sock *tp)
1689{
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001690 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 tp->snd_cwnd_stamp = tcp_time_stamp;
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001692 tcp_ca_event(tp, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693}
1694
1695static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1696{
1697 tp->left_out = tp->sacked_out;
1698
1699 if (tp->retrans_out == 0)
1700 tp->retrans_stamp = 0;
1701
1702 if (flag&FLAG_ECE)
1703 tcp_enter_cwr(tp);
1704
1705 if (tp->ca_state != TCP_CA_CWR) {
1706 int state = TCP_CA_Open;
1707
1708 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1709 state = TCP_CA_Disorder;
1710
1711 if (tp->ca_state != state) {
1712 tcp_set_ca_state(tp, state);
1713 tp->high_seq = tp->snd_nxt;
1714 }
1715 tcp_moderate_cwnd(tp);
1716 } else {
1717 tcp_cwnd_down(tp);
1718 }
1719}
1720
1721/* Process an event, which can update packets-in-flight not trivially.
1722 * Main goal of this function is to calculate new estimate for left_out,
1723 * taking into account both packets sitting in receiver's buffer and
1724 * packets lost by network.
1725 *
1726 * Besides that it does CWND reduction, when packet loss is detected
1727 * and changes state of machine.
1728 *
1729 * It does _not_ decide what to send, it is made in function
1730 * tcp_xmit_retransmit_queue().
1731 */
1732static void
1733tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1734 int prior_packets, int flag)
1735{
1736 struct tcp_sock *tp = tcp_sk(sk);
1737 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1738
1739 /* Some technical things:
1740 * 1. Reno does not count dupacks (sacked_out) automatically. */
1741 if (!tp->packets_out)
1742 tp->sacked_out = 0;
1743 /* 2. SACK counts snd_fack in packets inaccurately. */
1744 if (tp->sacked_out == 0)
1745 tp->fackets_out = 0;
1746
1747 /* Now state machine starts.
1748 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1749 if (flag&FLAG_ECE)
1750 tp->prior_ssthresh = 0;
1751
1752 /* B. In all the states check for reneging SACKs. */
1753 if (tp->sacked_out && tcp_check_sack_reneging(sk, tp))
1754 return;
1755
1756 /* C. Process data loss notification, provided it is valid. */
1757 if ((flag&FLAG_DATA_LOST) &&
1758 before(tp->snd_una, tp->high_seq) &&
1759 tp->ca_state != TCP_CA_Open &&
1760 tp->fackets_out > tp->reordering) {
1761 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1762 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1763 }
1764
1765 /* D. Synchronize left_out to current state. */
1766 tcp_sync_left_out(tp);
1767
1768 /* E. Check state exit conditions. State can be terminated
1769 * when high_seq is ACKed. */
1770 if (tp->ca_state == TCP_CA_Open) {
1771 if (!sysctl_tcp_frto)
1772 BUG_TRAP(tp->retrans_out == 0);
1773 tp->retrans_stamp = 0;
1774 } else if (!before(tp->snd_una, tp->high_seq)) {
1775 switch (tp->ca_state) {
1776 case TCP_CA_Loss:
1777 tp->retransmits = 0;
1778 if (tcp_try_undo_recovery(sk, tp))
1779 return;
1780 break;
1781
1782 case TCP_CA_CWR:
1783 /* CWR is to be held something *above* high_seq
1784 * is ACKed for CWR bit to reach receiver. */
1785 if (tp->snd_una != tp->high_seq) {
1786 tcp_complete_cwr(tp);
1787 tcp_set_ca_state(tp, TCP_CA_Open);
1788 }
1789 break;
1790
1791 case TCP_CA_Disorder:
1792 tcp_try_undo_dsack(sk, tp);
1793 if (!tp->undo_marker ||
1794 /* For SACK case do not Open to allow to undo
1795 * catching for all duplicate ACKs. */
1796 IsReno(tp) || tp->snd_una != tp->high_seq) {
1797 tp->undo_marker = 0;
1798 tcp_set_ca_state(tp, TCP_CA_Open);
1799 }
1800 break;
1801
1802 case TCP_CA_Recovery:
1803 if (IsReno(tp))
1804 tcp_reset_reno_sack(tp);
1805 if (tcp_try_undo_recovery(sk, tp))
1806 return;
1807 tcp_complete_cwr(tp);
1808 break;
1809 }
1810 }
1811
1812 /* F. Process state. */
1813 switch (tp->ca_state) {
1814 case TCP_CA_Recovery:
1815 if (prior_snd_una == tp->snd_una) {
1816 if (IsReno(tp) && is_dupack)
1817 tcp_add_reno_sack(tp);
1818 } else {
1819 int acked = prior_packets - tp->packets_out;
1820 if (IsReno(tp))
1821 tcp_remove_reno_sacks(sk, tp, acked);
1822 is_dupack = tcp_try_undo_partial(sk, tp, acked);
1823 }
1824 break;
1825 case TCP_CA_Loss:
1826 if (flag&FLAG_DATA_ACKED)
1827 tp->retransmits = 0;
1828 if (!tcp_try_undo_loss(sk, tp)) {
1829 tcp_moderate_cwnd(tp);
1830 tcp_xmit_retransmit_queue(sk);
1831 return;
1832 }
1833 if (tp->ca_state != TCP_CA_Open)
1834 return;
1835 /* Loss is undone; fall through to processing in Open state. */
1836 default:
1837 if (IsReno(tp)) {
1838 if (tp->snd_una != prior_snd_una)
1839 tcp_reset_reno_sack(tp);
1840 if (is_dupack)
1841 tcp_add_reno_sack(tp);
1842 }
1843
1844 if (tp->ca_state == TCP_CA_Disorder)
1845 tcp_try_undo_dsack(sk, tp);
1846
1847 if (!tcp_time_to_recover(sk, tp)) {
1848 tcp_try_to_open(sk, tp, flag);
1849 return;
1850 }
1851
1852 /* Otherwise enter Recovery state */
1853
1854 if (IsReno(tp))
1855 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
1856 else
1857 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
1858
1859 tp->high_seq = tp->snd_nxt;
1860 tp->prior_ssthresh = 0;
1861 tp->undo_marker = tp->snd_una;
1862 tp->undo_retrans = tp->retrans_out;
1863
1864 if (tp->ca_state < TCP_CA_CWR) {
1865 if (!(flag&FLAG_ECE))
1866 tp->prior_ssthresh = tcp_current_ssthresh(tp);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001867 tp->snd_ssthresh = tp->ca_ops->ssthresh(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 TCP_ECN_queue_cwr(tp);
1869 }
1870
1871 tp->snd_cwnd_cnt = 0;
1872 tcp_set_ca_state(tp, TCP_CA_Recovery);
1873 }
1874
1875 if (is_dupack || tcp_head_timedout(sk, tp))
1876 tcp_update_scoreboard(sk, tp);
1877 tcp_cwnd_down(tp);
1878 tcp_xmit_retransmit_queue(sk);
1879}
1880
1881/* Read draft-ietf-tcplw-high-performance before mucking
1882 * with this code. (Superceeds RFC1323)
1883 */
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001884static void tcp_ack_saw_tstamp(struct tcp_sock *tp, u32 *usrtt, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885{
1886 __u32 seq_rtt;
1887
1888 /* RTTM Rule: A TSecr value received in a segment is used to
1889 * update the averaged RTT measurement only if the segment
1890 * acknowledges some new data, i.e., only if it advances the
1891 * left edge of the send window.
1892 *
1893 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1894 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1895 *
1896 * Changed: reset backoff as soon as we see the first valid sample.
1897 * If we do not, we get strongly overstimated rto. With timestamps
1898 * samples are accepted even from very old segments: f.e., when rtt=1
1899 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1900 * answer arrives rto becomes 120 seconds! If at least one of segments
1901 * in window is lost... Voila. --ANK (010210)
1902 */
1903 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001904 tcp_rtt_estimator(tp, seq_rtt, usrtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 tcp_set_rto(tp);
1906 tp->backoff = 0;
1907 tcp_bound_rto(tp);
1908}
1909
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001910static void tcp_ack_no_tstamp(struct tcp_sock *tp, u32 seq_rtt, u32 *usrtt, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911{
1912 /* We don't have a timestamp. Can only use
1913 * packets that are not retransmitted to determine
1914 * rtt estimates. Also, we must not reset the
1915 * backoff for rto until we get a non-retransmitted
1916 * packet. This allows us to deal with a situation
1917 * where the network delay has increased suddenly.
1918 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1919 */
1920
1921 if (flag & FLAG_RETRANS_DATA_ACKED)
1922 return;
1923
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001924 tcp_rtt_estimator(tp, seq_rtt, usrtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 tcp_set_rto(tp);
1926 tp->backoff = 0;
1927 tcp_bound_rto(tp);
1928}
1929
1930static inline void tcp_ack_update_rtt(struct tcp_sock *tp,
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001931 int flag, s32 seq_rtt, u32 *usrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932{
1933 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1934 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001935 tcp_ack_saw_tstamp(tp, usrtt, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 else if (seq_rtt >= 0)
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001937 tcp_ack_no_tstamp(tp, seq_rtt, usrtt, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938}
1939
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001940static inline void tcp_cong_avoid(struct tcp_sock *tp, u32 ack, u32 rtt,
1941 u32 in_flight, int good)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942{
Stephen Hemminger317a76f2005-06-23 12:19:55 -07001943 tp->ca_ops->cong_avoid(tp, ack, rtt, in_flight, good);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 tp->snd_cwnd_stamp = tcp_time_stamp;
1945}
1946
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947/* Restart timer after forward progress on connection.
1948 * RFC2988 recommends to restart timer to now+rto.
1949 */
1950
1951static inline void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
1952{
1953 if (!tp->packets_out) {
1954 tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS);
1955 } else {
1956 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1957 }
1958}
1959
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
1961 __u32 now, __s32 *seq_rtt)
1962{
1963 struct tcp_sock *tp = tcp_sk(sk);
1964 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
1965 __u32 seq = tp->snd_una;
1966 __u32 packets_acked;
1967 int acked = 0;
1968
1969 /* If we get here, the whole TSO packet has not been
1970 * acked.
1971 */
1972 BUG_ON(!after(scb->end_seq, seq));
1973
1974 packets_acked = tcp_skb_pcount(skb);
1975 if (tcp_trim_head(sk, skb, seq - scb->seq))
1976 return 0;
1977 packets_acked -= tcp_skb_pcount(skb);
1978
1979 if (packets_acked) {
1980 __u8 sacked = scb->sacked;
1981
1982 acked |= FLAG_DATA_ACKED;
1983 if (sacked) {
1984 if (sacked & TCPCB_RETRANS) {
1985 if (sacked & TCPCB_SACKED_RETRANS)
1986 tp->retrans_out -= packets_acked;
1987 acked |= FLAG_RETRANS_DATA_ACKED;
1988 *seq_rtt = -1;
1989 } else if (*seq_rtt < 0)
1990 *seq_rtt = now - scb->when;
1991 if (sacked & TCPCB_SACKED_ACKED)
1992 tp->sacked_out -= packets_acked;
1993 if (sacked & TCPCB_LOST)
1994 tp->lost_out -= packets_acked;
1995 if (sacked & TCPCB_URG) {
1996 if (tp->urg_mode &&
1997 !before(seq, tp->snd_up))
1998 tp->urg_mode = 0;
1999 }
2000 } else if (*seq_rtt < 0)
2001 *seq_rtt = now - scb->when;
2002
2003 if (tp->fackets_out) {
2004 __u32 dval = min(tp->fackets_out, packets_acked);
2005 tp->fackets_out -= dval;
2006 }
2007 tp->packets_out -= packets_acked;
2008
2009 BUG_ON(tcp_skb_pcount(skb) == 0);
2010 BUG_ON(!before(scb->seq, scb->end_seq));
2011 }
2012
2013 return acked;
2014}
2015
2016
2017/* Remove acknowledged frames from the retransmission queue. */
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002018static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p, s32 *seq_usrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019{
2020 struct tcp_sock *tp = tcp_sk(sk);
2021 struct sk_buff *skb;
2022 __u32 now = tcp_time_stamp;
2023 int acked = 0;
2024 __s32 seq_rtt = -1;
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002025 struct timeval usnow;
2026 u32 pkts_acked = 0;
2027
2028 if (seq_usrtt)
2029 do_gettimeofday(&usnow);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030
2031 while ((skb = skb_peek(&sk->sk_write_queue)) &&
2032 skb != sk->sk_send_head) {
2033 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2034 __u8 sacked = scb->sacked;
2035
2036 /* If our packet is before the ack sequence we can
2037 * discard it as it's confirmed to have arrived at
2038 * the other end.
2039 */
2040 if (after(scb->end_seq, tp->snd_una)) {
David S. Millercb831992005-07-05 15:20:55 -07002041 if (tcp_skb_pcount(skb) > 1 &&
2042 after(tp->snd_una, scb->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 acked |= tcp_tso_acked(sk, skb,
2044 now, &seq_rtt);
2045 break;
2046 }
2047
2048 /* Initial outgoing SYN's get put onto the write_queue
2049 * just like anything else we transmit. It is not
2050 * true data, and if we misinform our callers that
2051 * this ACK acks real data, we will erroneously exit
2052 * connection startup slow start one packet too
2053 * quickly. This is severely frowned upon behavior.
2054 */
2055 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2056 acked |= FLAG_DATA_ACKED;
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002057 ++pkts_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 } else {
2059 acked |= FLAG_SYN_ACKED;
2060 tp->retrans_stamp = 0;
2061 }
2062
2063 if (sacked) {
2064 if (sacked & TCPCB_RETRANS) {
2065 if(sacked & TCPCB_SACKED_RETRANS)
2066 tp->retrans_out -= tcp_skb_pcount(skb);
2067 acked |= FLAG_RETRANS_DATA_ACKED;
2068 seq_rtt = -1;
2069 } else if (seq_rtt < 0)
2070 seq_rtt = now - scb->when;
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002071 if (seq_usrtt)
2072 *seq_usrtt = (usnow.tv_sec - skb->stamp.tv_sec) * 1000000
2073 + (usnow.tv_usec - skb->stamp.tv_usec);
2074
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 if (sacked & TCPCB_SACKED_ACKED)
2076 tp->sacked_out -= tcp_skb_pcount(skb);
2077 if (sacked & TCPCB_LOST)
2078 tp->lost_out -= tcp_skb_pcount(skb);
2079 if (sacked & TCPCB_URG) {
2080 if (tp->urg_mode &&
2081 !before(scb->end_seq, tp->snd_up))
2082 tp->urg_mode = 0;
2083 }
2084 } else if (seq_rtt < 0)
2085 seq_rtt = now - scb->when;
2086 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2087 tcp_packets_out_dec(tp, skb);
2088 __skb_unlink(skb, skb->list);
2089 sk_stream_free_skb(sk, skb);
2090 }
2091
2092 if (acked&FLAG_ACKED) {
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002093 tcp_ack_update_rtt(tp, acked, seq_rtt, seq_usrtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 tcp_ack_packets_out(sk, tp);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002095
2096 if (tp->ca_ops->pkts_acked)
2097 tp->ca_ops->pkts_acked(tp, pkts_acked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 }
2099
2100#if FASTRETRANS_DEBUG > 0
2101 BUG_TRAP((int)tp->sacked_out >= 0);
2102 BUG_TRAP((int)tp->lost_out >= 0);
2103 BUG_TRAP((int)tp->retrans_out >= 0);
2104 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2105 if (tp->lost_out) {
2106 printk(KERN_DEBUG "Leak l=%u %d\n",
2107 tp->lost_out, tp->ca_state);
2108 tp->lost_out = 0;
2109 }
2110 if (tp->sacked_out) {
2111 printk(KERN_DEBUG "Leak s=%u %d\n",
2112 tp->sacked_out, tp->ca_state);
2113 tp->sacked_out = 0;
2114 }
2115 if (tp->retrans_out) {
2116 printk(KERN_DEBUG "Leak r=%u %d\n",
2117 tp->retrans_out, tp->ca_state);
2118 tp->retrans_out = 0;
2119 }
2120 }
2121#endif
2122 *seq_rtt_p = seq_rtt;
2123 return acked;
2124}
2125
2126static void tcp_ack_probe(struct sock *sk)
2127{
2128 struct tcp_sock *tp = tcp_sk(sk);
2129
2130 /* Was it a usable window open? */
2131
2132 if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2133 tp->snd_una + tp->snd_wnd)) {
2134 tp->backoff = 0;
2135 tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0);
2136 /* Socket must be waked up by subsequent tcp_data_snd_check().
2137 * This function is not for random using!
2138 */
2139 } else {
2140 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0,
2141 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2142 }
2143}
2144
2145static inline int tcp_ack_is_dubious(struct tcp_sock *tp, int flag)
2146{
2147 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2148 tp->ca_state != TCP_CA_Open);
2149}
2150
2151static inline int tcp_may_raise_cwnd(struct tcp_sock *tp, int flag)
2152{
2153 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2154 !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR));
2155}
2156
2157/* Check that window update is acceptable.
2158 * The function assumes that snd_una<=ack<=snd_next.
2159 */
2160static inline int tcp_may_update_window(struct tcp_sock *tp, u32 ack,
2161 u32 ack_seq, u32 nwin)
2162{
2163 return (after(ack, tp->snd_una) ||
2164 after(ack_seq, tp->snd_wl1) ||
2165 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2166}
2167
2168/* Update our send window.
2169 *
2170 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2171 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2172 */
2173static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2174 struct sk_buff *skb, u32 ack, u32 ack_seq)
2175{
2176 int flag = 0;
2177 u32 nwin = ntohs(skb->h.th->window);
2178
2179 if (likely(!skb->h.th->syn))
2180 nwin <<= tp->rx_opt.snd_wscale;
2181
2182 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2183 flag |= FLAG_WIN_UPDATE;
2184 tcp_update_wl(tp, ack, ack_seq);
2185
2186 if (tp->snd_wnd != nwin) {
2187 tp->snd_wnd = nwin;
2188
2189 /* Note, it is the only place, where
2190 * fast path is recovered for sending TCP.
2191 */
2192 tcp_fast_path_check(sk, tp);
2193
2194 if (nwin > tp->max_window) {
2195 tp->max_window = nwin;
2196 tcp_sync_mss(sk, tp->pmtu_cookie);
2197 }
2198 }
2199 }
2200
2201 tp->snd_una = ack;
2202
2203 return flag;
2204}
2205
2206static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2207{
2208 struct tcp_sock *tp = tcp_sk(sk);
2209
2210 tcp_sync_left_out(tp);
2211
2212 if (tp->snd_una == prior_snd_una ||
2213 !before(tp->snd_una, tp->frto_highmark)) {
2214 /* RTO was caused by loss, start retransmitting in
2215 * go-back-N slow start
2216 */
2217 tcp_enter_frto_loss(sk);
2218 return;
2219 }
2220
2221 if (tp->frto_counter == 1) {
2222 /* First ACK after RTO advances the window: allow two new
2223 * segments out.
2224 */
2225 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2226 } else {
2227 /* Also the second ACK after RTO advances the window.
2228 * The RTO was likely spurious. Reduce cwnd and continue
2229 * in congestion avoidance
2230 */
2231 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2232 tcp_moderate_cwnd(tp);
2233 }
2234
2235 /* F-RTO affects on two new ACKs following RTO.
2236 * At latest on third ACK the TCP behavor is back to normal.
2237 */
2238 tp->frto_counter = (tp->frto_counter + 1) % 3;
2239}
2240
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241/* This routine deals with incoming acks, but not outgoing ones. */
2242static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2243{
2244 struct tcp_sock *tp = tcp_sk(sk);
2245 u32 prior_snd_una = tp->snd_una;
2246 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2247 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2248 u32 prior_in_flight;
2249 s32 seq_rtt;
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002250 s32 seq_usrtt = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 int prior_packets;
2252
2253 /* If the ack is newer than sent or older than previous acks
2254 * then we can probably ignore it.
2255 */
2256 if (after(ack, tp->snd_nxt))
2257 goto uninteresting_ack;
2258
2259 if (before(ack, prior_snd_una))
2260 goto old_ack;
2261
2262 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2263 /* Window is constant, pure forward advance.
2264 * No more checks are required.
2265 * Note, we use the fact that SND.UNA>=SND.WL2.
2266 */
2267 tcp_update_wl(tp, ack, ack_seq);
2268 tp->snd_una = ack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 flag |= FLAG_WIN_UPDATE;
2270
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002271 tcp_ca_event(tp, CA_EVENT_FAST_ACK);
2272
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2274 } else {
2275 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2276 flag |= FLAG_DATA;
2277 else
2278 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2279
2280 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2281
2282 if (TCP_SKB_CB(skb)->sacked)
2283 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2284
2285 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2286 flag |= FLAG_ECE;
2287
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002288 tcp_ca_event(tp, CA_EVENT_SLOW_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 }
2290
2291 /* We passed data and got it acked, remove any soft error
2292 * log. Something worked...
2293 */
2294 sk->sk_err_soft = 0;
2295 tp->rcv_tstamp = tcp_time_stamp;
2296 prior_packets = tp->packets_out;
2297 if (!prior_packets)
2298 goto no_queue;
2299
2300 prior_in_flight = tcp_packets_in_flight(tp);
2301
2302 /* See if we can take anything off of the retransmit queue. */
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002303 flag |= tcp_clean_rtx_queue(sk, &seq_rtt,
2304 tp->ca_ops->rtt_sample ? &seq_usrtt : NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305
2306 if (tp->frto_counter)
2307 tcp_process_frto(sk, prior_snd_una);
2308
2309 if (tcp_ack_is_dubious(tp, flag)) {
2310 /* Advanve CWND, if state allows this. */
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002311 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(tp, flag))
2312 tcp_cong_avoid(tp, ack, seq_rtt, prior_in_flight, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2314 } else {
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002315 if ((flag & FLAG_DATA_ACKED))
2316 tcp_cong_avoid(tp, ack, seq_rtt, prior_in_flight, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 }
2318
2319 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2320 dst_confirm(sk->sk_dst_cache);
2321
2322 return 1;
2323
2324no_queue:
2325 tp->probes_out = 0;
2326
2327 /* If this ack opens up a zero window, clear backoff. It was
2328 * being used to time the probes, and is probably far higher than
2329 * it needs to be for normal retransmission.
2330 */
2331 if (sk->sk_send_head)
2332 tcp_ack_probe(sk);
2333 return 1;
2334
2335old_ack:
2336 if (TCP_SKB_CB(skb)->sacked)
2337 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2338
2339uninteresting_ack:
2340 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2341 return 0;
2342}
2343
2344
2345/* Look for tcp options. Normally only called on SYN and SYNACK packets.
2346 * But, this can also be called on packets in the established flow when
2347 * the fast version below fails.
2348 */
2349void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2350{
2351 unsigned char *ptr;
2352 struct tcphdr *th = skb->h.th;
2353 int length=(th->doff*4)-sizeof(struct tcphdr);
2354
2355 ptr = (unsigned char *)(th + 1);
2356 opt_rx->saw_tstamp = 0;
2357
2358 while(length>0) {
2359 int opcode=*ptr++;
2360 int opsize;
2361
2362 switch (opcode) {
2363 case TCPOPT_EOL:
2364 return;
2365 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2366 length--;
2367 continue;
2368 default:
2369 opsize=*ptr++;
2370 if (opsize < 2) /* "silly options" */
2371 return;
2372 if (opsize > length)
2373 return; /* don't parse partial options */
2374 switch(opcode) {
2375 case TCPOPT_MSS:
2376 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2377 u16 in_mss = ntohs(get_unaligned((__u16 *)ptr));
2378 if (in_mss) {
2379 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2380 in_mss = opt_rx->user_mss;
2381 opt_rx->mss_clamp = in_mss;
2382 }
2383 }
2384 break;
2385 case TCPOPT_WINDOW:
2386 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2387 if (sysctl_tcp_window_scaling) {
2388 __u8 snd_wscale = *(__u8 *) ptr;
2389 opt_rx->wscale_ok = 1;
2390 if (snd_wscale > 14) {
2391 if(net_ratelimit())
2392 printk(KERN_INFO "tcp_parse_options: Illegal window "
2393 "scaling value %d >14 received.\n",
2394 snd_wscale);
2395 snd_wscale = 14;
2396 }
2397 opt_rx->snd_wscale = snd_wscale;
2398 }
2399 break;
2400 case TCPOPT_TIMESTAMP:
2401 if(opsize==TCPOLEN_TIMESTAMP) {
2402 if ((estab && opt_rx->tstamp_ok) ||
2403 (!estab && sysctl_tcp_timestamps)) {
2404 opt_rx->saw_tstamp = 1;
2405 opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr));
2406 opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4)));
2407 }
2408 }
2409 break;
2410 case TCPOPT_SACK_PERM:
2411 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2412 if (sysctl_tcp_sack) {
2413 opt_rx->sack_ok = 1;
2414 tcp_sack_reset(opt_rx);
2415 }
2416 }
2417 break;
2418
2419 case TCPOPT_SACK:
2420 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2421 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2422 opt_rx->sack_ok) {
2423 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2424 }
2425 };
2426 ptr+=opsize-2;
2427 length-=opsize;
2428 };
2429 }
2430}
2431
2432/* Fast parse options. This hopes to only see timestamps.
2433 * If it is wrong it falls back on tcp_parse_options().
2434 */
2435static inline int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2436 struct tcp_sock *tp)
2437{
2438 if (th->doff == sizeof(struct tcphdr)>>2) {
2439 tp->rx_opt.saw_tstamp = 0;
2440 return 0;
2441 } else if (tp->rx_opt.tstamp_ok &&
2442 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2443 __u32 *ptr = (__u32 *)(th + 1);
2444 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2445 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2446 tp->rx_opt.saw_tstamp = 1;
2447 ++ptr;
2448 tp->rx_opt.rcv_tsval = ntohl(*ptr);
2449 ++ptr;
2450 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2451 return 1;
2452 }
2453 }
2454 tcp_parse_options(skb, &tp->rx_opt, 1);
2455 return 1;
2456}
2457
2458static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2459{
2460 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2461 tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2462}
2463
2464static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2465{
2466 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2467 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2468 * extra check below makes sure this can only happen
2469 * for pure ACK frames. -DaveM
2470 *
2471 * Not only, also it occurs for expired timestamps.
2472 */
2473
2474 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2475 xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2476 tcp_store_ts_recent(tp);
2477 }
2478}
2479
2480/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2481 *
2482 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2483 * it can pass through stack. So, the following predicate verifies that
2484 * this segment is not used for anything but congestion avoidance or
2485 * fast retransmit. Moreover, we even are able to eliminate most of such
2486 * second order effects, if we apply some small "replay" window (~RTO)
2487 * to timestamp space.
2488 *
2489 * All these measures still do not guarantee that we reject wrapped ACKs
2490 * on networks with high bandwidth, when sequence space is recycled fastly,
2491 * but it guarantees that such events will be very rare and do not affect
2492 * connection seriously. This doesn't look nice, but alas, PAWS is really
2493 * buggy extension.
2494 *
2495 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2496 * states that events when retransmit arrives after original data are rare.
2497 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2498 * the biggest problem on large power networks even with minor reordering.
2499 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2500 * up to bandwidth of 18Gigabit/sec. 8) ]
2501 */
2502
2503static int tcp_disordered_ack(struct tcp_sock *tp, struct sk_buff *skb)
2504{
2505 struct tcphdr *th = skb->h.th;
2506 u32 seq = TCP_SKB_CB(skb)->seq;
2507 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2508
2509 return (/* 1. Pure ACK with correct sequence number. */
2510 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2511
2512 /* 2. ... and duplicate ACK. */
2513 ack == tp->snd_una &&
2514
2515 /* 3. ... and does not update window. */
2516 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2517
2518 /* 4. ... and sits in replay window. */
2519 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (tp->rto*1024)/HZ);
2520}
2521
2522static inline int tcp_paws_discard(struct tcp_sock *tp, struct sk_buff *skb)
2523{
2524 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2525 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2526 !tcp_disordered_ack(tp, skb));
2527}
2528
2529/* Check segment sequence number for validity.
2530 *
2531 * Segment controls are considered valid, if the segment
2532 * fits to the window after truncation to the window. Acceptability
2533 * of data (and SYN, FIN, of course) is checked separately.
2534 * See tcp_data_queue(), for example.
2535 *
2536 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2537 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2538 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2539 * (borrowed from freebsd)
2540 */
2541
2542static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2543{
2544 return !before(end_seq, tp->rcv_wup) &&
2545 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2546}
2547
2548/* When we get a reset we do this. */
2549static void tcp_reset(struct sock *sk)
2550{
2551 /* We want the right error as BSD sees it (and indeed as we do). */
2552 switch (sk->sk_state) {
2553 case TCP_SYN_SENT:
2554 sk->sk_err = ECONNREFUSED;
2555 break;
2556 case TCP_CLOSE_WAIT:
2557 sk->sk_err = EPIPE;
2558 break;
2559 case TCP_CLOSE:
2560 return;
2561 default:
2562 sk->sk_err = ECONNRESET;
2563 }
2564
2565 if (!sock_flag(sk, SOCK_DEAD))
2566 sk->sk_error_report(sk);
2567
2568 tcp_done(sk);
2569}
2570
2571/*
2572 * Process the FIN bit. This now behaves as it is supposed to work
2573 * and the FIN takes effect when it is validly part of sequence
2574 * space. Not before when we get holes.
2575 *
2576 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2577 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2578 * TIME-WAIT)
2579 *
2580 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2581 * close and we go into CLOSING (and later onto TIME-WAIT)
2582 *
2583 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2584 */
2585static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2586{
2587 struct tcp_sock *tp = tcp_sk(sk);
2588
2589 tcp_schedule_ack(tp);
2590
2591 sk->sk_shutdown |= RCV_SHUTDOWN;
2592 sock_set_flag(sk, SOCK_DONE);
2593
2594 switch (sk->sk_state) {
2595 case TCP_SYN_RECV:
2596 case TCP_ESTABLISHED:
2597 /* Move to CLOSE_WAIT */
2598 tcp_set_state(sk, TCP_CLOSE_WAIT);
2599 tp->ack.pingpong = 1;
2600 break;
2601
2602 case TCP_CLOSE_WAIT:
2603 case TCP_CLOSING:
2604 /* Received a retransmission of the FIN, do
2605 * nothing.
2606 */
2607 break;
2608 case TCP_LAST_ACK:
2609 /* RFC793: Remain in the LAST-ACK state. */
2610 break;
2611
2612 case TCP_FIN_WAIT1:
2613 /* This case occurs when a simultaneous close
2614 * happens, we must ack the received FIN and
2615 * enter the CLOSING state.
2616 */
2617 tcp_send_ack(sk);
2618 tcp_set_state(sk, TCP_CLOSING);
2619 break;
2620 case TCP_FIN_WAIT2:
2621 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2622 tcp_send_ack(sk);
2623 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2624 break;
2625 default:
2626 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2627 * cases we should never reach this piece of code.
2628 */
2629 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2630 __FUNCTION__, sk->sk_state);
2631 break;
2632 };
2633
2634 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2635 * Probably, we should reset in this case. For now drop them.
2636 */
2637 __skb_queue_purge(&tp->out_of_order_queue);
2638 if (tp->rx_opt.sack_ok)
2639 tcp_sack_reset(&tp->rx_opt);
2640 sk_stream_mem_reclaim(sk);
2641
2642 if (!sock_flag(sk, SOCK_DEAD)) {
2643 sk->sk_state_change(sk);
2644
2645 /* Do not send POLL_HUP for half duplex close. */
2646 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2647 sk->sk_state == TCP_CLOSE)
2648 sk_wake_async(sk, 1, POLL_HUP);
2649 else
2650 sk_wake_async(sk, 1, POLL_IN);
2651 }
2652}
2653
2654static __inline__ int
2655tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2656{
2657 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2658 if (before(seq, sp->start_seq))
2659 sp->start_seq = seq;
2660 if (after(end_seq, sp->end_seq))
2661 sp->end_seq = end_seq;
2662 return 1;
2663 }
2664 return 0;
2665}
2666
2667static inline void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
2668{
2669 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2670 if (before(seq, tp->rcv_nxt))
2671 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
2672 else
2673 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
2674
2675 tp->rx_opt.dsack = 1;
2676 tp->duplicate_sack[0].start_seq = seq;
2677 tp->duplicate_sack[0].end_seq = end_seq;
2678 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
2679 }
2680}
2681
2682static inline void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
2683{
2684 if (!tp->rx_opt.dsack)
2685 tcp_dsack_set(tp, seq, end_seq);
2686 else
2687 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2688}
2689
2690static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2691{
2692 struct tcp_sock *tp = tcp_sk(sk);
2693
2694 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2695 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2696 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2697 tcp_enter_quickack_mode(tp);
2698
2699 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2700 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2701
2702 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2703 end_seq = tp->rcv_nxt;
2704 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2705 }
2706 }
2707
2708 tcp_send_ack(sk);
2709}
2710
2711/* These routines update the SACK block as out-of-order packets arrive or
2712 * in-order packets close up the sequence space.
2713 */
2714static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
2715{
2716 int this_sack;
2717 struct tcp_sack_block *sp = &tp->selective_acks[0];
2718 struct tcp_sack_block *swalk = sp+1;
2719
2720 /* See if the recent change to the first SACK eats into
2721 * or hits the sequence space of other SACK blocks, if so coalesce.
2722 */
2723 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
2724 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2725 int i;
2726
2727 /* Zap SWALK, by moving every further SACK up by one slot.
2728 * Decrease num_sacks.
2729 */
2730 tp->rx_opt.num_sacks--;
2731 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2732 for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
2733 sp[i] = sp[i+1];
2734 continue;
2735 }
2736 this_sack++, swalk++;
2737 }
2738}
2739
2740static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2741{
2742 __u32 tmp;
2743
2744 tmp = sack1->start_seq;
2745 sack1->start_seq = sack2->start_seq;
2746 sack2->start_seq = tmp;
2747
2748 tmp = sack1->end_seq;
2749 sack1->end_seq = sack2->end_seq;
2750 sack2->end_seq = tmp;
2751}
2752
2753static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
2754{
2755 struct tcp_sock *tp = tcp_sk(sk);
2756 struct tcp_sack_block *sp = &tp->selective_acks[0];
2757 int cur_sacks = tp->rx_opt.num_sacks;
2758 int this_sack;
2759
2760 if (!cur_sacks)
2761 goto new_sack;
2762
2763 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
2764 if (tcp_sack_extend(sp, seq, end_seq)) {
2765 /* Rotate this_sack to the first one. */
2766 for (; this_sack>0; this_sack--, sp--)
2767 tcp_sack_swap(sp, sp-1);
2768 if (cur_sacks > 1)
2769 tcp_sack_maybe_coalesce(tp);
2770 return;
2771 }
2772 }
2773
2774 /* Could not find an adjacent existing SACK, build a new one,
2775 * put it at the front, and shift everyone else down. We
2776 * always know there is at least one SACK present already here.
2777 *
2778 * If the sack array is full, forget about the last one.
2779 */
2780 if (this_sack >= 4) {
2781 this_sack--;
2782 tp->rx_opt.num_sacks--;
2783 sp--;
2784 }
2785 for(; this_sack > 0; this_sack--, sp--)
2786 *sp = *(sp-1);
2787
2788new_sack:
2789 /* Build the new head SACK, and we're done. */
2790 sp->start_seq = seq;
2791 sp->end_seq = end_seq;
2792 tp->rx_opt.num_sacks++;
2793 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2794}
2795
2796/* RCV.NXT advances, some SACKs should be eaten. */
2797
2798static void tcp_sack_remove(struct tcp_sock *tp)
2799{
2800 struct tcp_sack_block *sp = &tp->selective_acks[0];
2801 int num_sacks = tp->rx_opt.num_sacks;
2802 int this_sack;
2803
2804 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07002805 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 tp->rx_opt.num_sacks = 0;
2807 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
2808 return;
2809 }
2810
2811 for(this_sack = 0; this_sack < num_sacks; ) {
2812 /* Check if the start of the sack is covered by RCV.NXT. */
2813 if (!before(tp->rcv_nxt, sp->start_seq)) {
2814 int i;
2815
2816 /* RCV.NXT must cover all the block! */
2817 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
2818
2819 /* Zap this SACK, by moving forward any other SACKS. */
2820 for (i=this_sack+1; i < num_sacks; i++)
2821 tp->selective_acks[i-1] = tp->selective_acks[i];
2822 num_sacks--;
2823 continue;
2824 }
2825 this_sack++;
2826 sp++;
2827 }
2828 if (num_sacks != tp->rx_opt.num_sacks) {
2829 tp->rx_opt.num_sacks = num_sacks;
2830 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2831 }
2832}
2833
2834/* This one checks to see if we can put data from the
2835 * out_of_order queue into the receive_queue.
2836 */
2837static void tcp_ofo_queue(struct sock *sk)
2838{
2839 struct tcp_sock *tp = tcp_sk(sk);
2840 __u32 dsack_high = tp->rcv_nxt;
2841 struct sk_buff *skb;
2842
2843 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
2844 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
2845 break;
2846
2847 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
2848 __u32 dsack = dsack_high;
2849 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
2850 dsack_high = TCP_SKB_CB(skb)->end_seq;
2851 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
2852 }
2853
2854 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
2855 SOCK_DEBUG(sk, "ofo packet was already received \n");
2856 __skb_unlink(skb, skb->list);
2857 __kfree_skb(skb);
2858 continue;
2859 }
2860 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
2861 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
2862 TCP_SKB_CB(skb)->end_seq);
2863
2864 __skb_unlink(skb, skb->list);
2865 __skb_queue_tail(&sk->sk_receive_queue, skb);
2866 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2867 if(skb->h.th->fin)
2868 tcp_fin(skb, sk, skb->h.th);
2869 }
2870}
2871
2872static int tcp_prune_queue(struct sock *sk);
2873
2874static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
2875{
2876 struct tcphdr *th = skb->h.th;
2877 struct tcp_sock *tp = tcp_sk(sk);
2878 int eaten = -1;
2879
2880 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
2881 goto drop;
2882
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 __skb_pull(skb, th->doff*4);
2884
2885 TCP_ECN_accept_cwr(tp, skb);
2886
2887 if (tp->rx_opt.dsack) {
2888 tp->rx_opt.dsack = 0;
2889 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
2890 4 - tp->rx_opt.tstamp_ok);
2891 }
2892
2893 /* Queue data for delivery to the user.
2894 * Packets in sequence go to the receive queue.
2895 * Out of sequence packets to the out_of_order_queue.
2896 */
2897 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
2898 if (tcp_receive_window(tp) == 0)
2899 goto out_of_window;
2900
2901 /* Ok. In sequence. In window. */
2902 if (tp->ucopy.task == current &&
2903 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
2904 sock_owned_by_user(sk) && !tp->urg_data) {
2905 int chunk = min_t(unsigned int, skb->len,
2906 tp->ucopy.len);
2907
2908 __set_current_state(TASK_RUNNING);
2909
2910 local_bh_enable();
2911 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
2912 tp->ucopy.len -= chunk;
2913 tp->copied_seq += chunk;
2914 eaten = (chunk == skb->len && !th->fin);
2915 tcp_rcv_space_adjust(sk);
2916 }
2917 local_bh_disable();
2918 }
2919
2920 if (eaten <= 0) {
2921queue_and_out:
2922 if (eaten < 0 &&
2923 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
2924 !sk_stream_rmem_schedule(sk, skb))) {
2925 if (tcp_prune_queue(sk) < 0 ||
2926 !sk_stream_rmem_schedule(sk, skb))
2927 goto drop;
2928 }
2929 sk_stream_set_owner_r(skb, sk);
2930 __skb_queue_tail(&sk->sk_receive_queue, skb);
2931 }
2932 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2933 if(skb->len)
2934 tcp_event_data_recv(sk, tp, skb);
2935 if(th->fin)
2936 tcp_fin(skb, sk, th);
2937
David S. Millerb03efcf2005-07-08 14:57:23 -07002938 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 tcp_ofo_queue(sk);
2940
2941 /* RFC2581. 4.2. SHOULD send immediate ACK, when
2942 * gap in queue is filled.
2943 */
David S. Millerb03efcf2005-07-08 14:57:23 -07002944 if (skb_queue_empty(&tp->out_of_order_queue))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 tp->ack.pingpong = 0;
2946 }
2947
2948 if (tp->rx_opt.num_sacks)
2949 tcp_sack_remove(tp);
2950
2951 tcp_fast_path_check(sk, tp);
2952
2953 if (eaten > 0)
2954 __kfree_skb(skb);
2955 else if (!sock_flag(sk, SOCK_DEAD))
2956 sk->sk_data_ready(sk, 0);
2957 return;
2958 }
2959
2960 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
2961 /* A retransmit, 2nd most common case. Force an immediate ack. */
2962 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2963 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
2964
2965out_of_window:
2966 tcp_enter_quickack_mode(tp);
2967 tcp_schedule_ack(tp);
2968drop:
2969 __kfree_skb(skb);
2970 return;
2971 }
2972
2973 /* Out of window. F.e. zero window probe. */
2974 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
2975 goto out_of_window;
2976
2977 tcp_enter_quickack_mode(tp);
2978
2979 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2980 /* Partial packet, seq < rcv_next < end_seq */
2981 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
2982 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
2983 TCP_SKB_CB(skb)->end_seq);
2984
2985 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
2986
2987 /* If window is closed, drop tail of packet. But after
2988 * remembering D-SACK for its head made in previous line.
2989 */
2990 if (!tcp_receive_window(tp))
2991 goto out_of_window;
2992 goto queue_and_out;
2993 }
2994
2995 TCP_ECN_check_ce(tp, skb);
2996
2997 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
2998 !sk_stream_rmem_schedule(sk, skb)) {
2999 if (tcp_prune_queue(sk) < 0 ||
3000 !sk_stream_rmem_schedule(sk, skb))
3001 goto drop;
3002 }
3003
3004 /* Disable header prediction. */
3005 tp->pred_flags = 0;
3006 tcp_schedule_ack(tp);
3007
3008 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3009 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3010
3011 sk_stream_set_owner_r(skb, sk);
3012
3013 if (!skb_peek(&tp->out_of_order_queue)) {
3014 /* Initial out of order segment, build 1 SACK. */
3015 if (tp->rx_opt.sack_ok) {
3016 tp->rx_opt.num_sacks = 1;
3017 tp->rx_opt.dsack = 0;
3018 tp->rx_opt.eff_sacks = 1;
3019 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3020 tp->selective_acks[0].end_seq =
3021 TCP_SKB_CB(skb)->end_seq;
3022 }
3023 __skb_queue_head(&tp->out_of_order_queue,skb);
3024 } else {
3025 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3026 u32 seq = TCP_SKB_CB(skb)->seq;
3027 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3028
3029 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3030 __skb_append(skb1, skb);
3031
3032 if (!tp->rx_opt.num_sacks ||
3033 tp->selective_acks[0].end_seq != seq)
3034 goto add_sack;
3035
3036 /* Common case: data arrive in order after hole. */
3037 tp->selective_acks[0].end_seq = end_seq;
3038 return;
3039 }
3040
3041 /* Find place to insert this segment. */
3042 do {
3043 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3044 break;
3045 } while ((skb1 = skb1->prev) !=
3046 (struct sk_buff*)&tp->out_of_order_queue);
3047
3048 /* Do skb overlap to previous one? */
3049 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3050 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3051 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3052 /* All the bits are present. Drop. */
3053 __kfree_skb(skb);
3054 tcp_dsack_set(tp, seq, end_seq);
3055 goto add_sack;
3056 }
3057 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3058 /* Partial overlap. */
3059 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3060 } else {
3061 skb1 = skb1->prev;
3062 }
3063 }
3064 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3065
3066 /* And clean segments covered by new one as whole. */
3067 while ((skb1 = skb->next) !=
3068 (struct sk_buff*)&tp->out_of_order_queue &&
3069 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3070 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3071 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3072 break;
3073 }
3074 __skb_unlink(skb1, skb1->list);
3075 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3076 __kfree_skb(skb1);
3077 }
3078
3079add_sack:
3080 if (tp->rx_opt.sack_ok)
3081 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3082 }
3083}
3084
3085/* Collapse contiguous sequence of skbs head..tail with
3086 * sequence numbers start..end.
3087 * Segments with FIN/SYN are not collapsed (only because this
3088 * simplifies code)
3089 */
3090static void
3091tcp_collapse(struct sock *sk, struct sk_buff *head,
3092 struct sk_buff *tail, u32 start, u32 end)
3093{
3094 struct sk_buff *skb;
3095
3096 /* First, check that queue is collapsable and find
3097 * the point where collapsing can be useful. */
3098 for (skb = head; skb != tail; ) {
3099 /* No new bits? It is possible on ofo queue. */
3100 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3101 struct sk_buff *next = skb->next;
3102 __skb_unlink(skb, skb->list);
3103 __kfree_skb(skb);
3104 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3105 skb = next;
3106 continue;
3107 }
3108
3109 /* The first skb to collapse is:
3110 * - not SYN/FIN and
3111 * - bloated or contains data before "start" or
3112 * overlaps to the next one.
3113 */
3114 if (!skb->h.th->syn && !skb->h.th->fin &&
3115 (tcp_win_from_space(skb->truesize) > skb->len ||
3116 before(TCP_SKB_CB(skb)->seq, start) ||
3117 (skb->next != tail &&
3118 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3119 break;
3120
3121 /* Decided to skip this, advance start seq. */
3122 start = TCP_SKB_CB(skb)->end_seq;
3123 skb = skb->next;
3124 }
3125 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3126 return;
3127
3128 while (before(start, end)) {
3129 struct sk_buff *nskb;
3130 int header = skb_headroom(skb);
3131 int copy = SKB_MAX_ORDER(header, 0);
3132
3133 /* Too big header? This can happen with IPv6. */
3134 if (copy < 0)
3135 return;
3136 if (end-start < copy)
3137 copy = end-start;
3138 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3139 if (!nskb)
3140 return;
3141 skb_reserve(nskb, header);
3142 memcpy(nskb->head, skb->head, header);
3143 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3144 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3145 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3146 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3147 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3148 __skb_insert(nskb, skb->prev, skb, skb->list);
3149 sk_stream_set_owner_r(nskb, sk);
3150
3151 /* Copy data, releasing collapsed skbs. */
3152 while (copy > 0) {
3153 int offset = start - TCP_SKB_CB(skb)->seq;
3154 int size = TCP_SKB_CB(skb)->end_seq - start;
3155
3156 if (offset < 0) BUG();
3157 if (size > 0) {
3158 size = min(copy, size);
3159 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3160 BUG();
3161 TCP_SKB_CB(nskb)->end_seq += size;
3162 copy -= size;
3163 start += size;
3164 }
3165 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3166 struct sk_buff *next = skb->next;
3167 __skb_unlink(skb, skb->list);
3168 __kfree_skb(skb);
3169 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3170 skb = next;
3171 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3172 return;
3173 }
3174 }
3175 }
3176}
3177
3178/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3179 * and tcp_collapse() them until all the queue is collapsed.
3180 */
3181static void tcp_collapse_ofo_queue(struct sock *sk)
3182{
3183 struct tcp_sock *tp = tcp_sk(sk);
3184 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3185 struct sk_buff *head;
3186 u32 start, end;
3187
3188 if (skb == NULL)
3189 return;
3190
3191 start = TCP_SKB_CB(skb)->seq;
3192 end = TCP_SKB_CB(skb)->end_seq;
3193 head = skb;
3194
3195 for (;;) {
3196 skb = skb->next;
3197
3198 /* Segment is terminated when we see gap or when
3199 * we are at the end of all the queue. */
3200 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3201 after(TCP_SKB_CB(skb)->seq, end) ||
3202 before(TCP_SKB_CB(skb)->end_seq, start)) {
3203 tcp_collapse(sk, head, skb, start, end);
3204 head = skb;
3205 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3206 break;
3207 /* Start new segment */
3208 start = TCP_SKB_CB(skb)->seq;
3209 end = TCP_SKB_CB(skb)->end_seq;
3210 } else {
3211 if (before(TCP_SKB_CB(skb)->seq, start))
3212 start = TCP_SKB_CB(skb)->seq;
3213 if (after(TCP_SKB_CB(skb)->end_seq, end))
3214 end = TCP_SKB_CB(skb)->end_seq;
3215 }
3216 }
3217}
3218
3219/* Reduce allocated memory if we can, trying to get
3220 * the socket within its memory limits again.
3221 *
3222 * Return less than zero if we should start dropping frames
3223 * until the socket owning process reads some of the data
3224 * to stabilize the situation.
3225 */
3226static int tcp_prune_queue(struct sock *sk)
3227{
3228 struct tcp_sock *tp = tcp_sk(sk);
3229
3230 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3231
3232 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3233
3234 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3235 tcp_clamp_window(sk, tp);
3236 else if (tcp_memory_pressure)
3237 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3238
3239 tcp_collapse_ofo_queue(sk);
3240 tcp_collapse(sk, sk->sk_receive_queue.next,
3241 (struct sk_buff*)&sk->sk_receive_queue,
3242 tp->copied_seq, tp->rcv_nxt);
3243 sk_stream_mem_reclaim(sk);
3244
3245 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3246 return 0;
3247
3248 /* Collapsing did not help, destructive actions follow.
3249 * This must not ever occur. */
3250
3251 /* First, purge the out_of_order queue. */
David S. Millerb03efcf2005-07-08 14:57:23 -07003252 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3253 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 __skb_queue_purge(&tp->out_of_order_queue);
3255
3256 /* Reset SACK state. A conforming SACK implementation will
3257 * do the same at a timeout based retransmit. When a connection
3258 * is in a sad state like this, we care only about integrity
3259 * of the connection not performance.
3260 */
3261 if (tp->rx_opt.sack_ok)
3262 tcp_sack_reset(&tp->rx_opt);
3263 sk_stream_mem_reclaim(sk);
3264 }
3265
3266 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3267 return 0;
3268
3269 /* If we are really being abused, tell the caller to silently
3270 * drop receive data on the floor. It will get retransmitted
3271 * and hopefully then we'll have sufficient space.
3272 */
3273 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3274
3275 /* Massive buffer overcommit. */
3276 tp->pred_flags = 0;
3277 return -1;
3278}
3279
3280
3281/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3282 * As additional protections, we do not touch cwnd in retransmission phases,
3283 * and if application hit its sndbuf limit recently.
3284 */
3285void tcp_cwnd_application_limited(struct sock *sk)
3286{
3287 struct tcp_sock *tp = tcp_sk(sk);
3288
3289 if (tp->ca_state == TCP_CA_Open &&
3290 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3291 /* Limited by application or receiver window. */
3292 u32 win_used = max(tp->snd_cwnd_used, 2U);
3293 if (win_used < tp->snd_cwnd) {
3294 tp->snd_ssthresh = tcp_current_ssthresh(tp);
3295 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3296 }
3297 tp->snd_cwnd_used = 0;
3298 }
3299 tp->snd_cwnd_stamp = tcp_time_stamp;
3300}
3301
David S. Miller0d9901d2005-07-05 15:21:10 -07003302static inline int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3303{
3304 /* If the user specified a specific send buffer setting, do
3305 * not modify it.
3306 */
3307 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3308 return 0;
3309
3310 /* If we are under global TCP memory pressure, do not expand. */
3311 if (tcp_memory_pressure)
3312 return 0;
3313
3314 /* If we are under soft global TCP memory pressure, do not expand. */
3315 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3316 return 0;
3317
3318 /* If we filled the congestion window, do not expand. */
3319 if (tp->packets_out >= tp->snd_cwnd)
3320 return 0;
3321
3322 return 1;
3323}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324
3325/* When incoming ACK allowed to free some skb from write_queue,
3326 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3327 * on the exit from tcp input handler.
3328 *
3329 * PROBLEM: sndbuf expansion does not work well with largesend.
3330 */
3331static void tcp_new_space(struct sock *sk)
3332{
3333 struct tcp_sock *tp = tcp_sk(sk);
3334
David S. Miller0d9901d2005-07-05 15:21:10 -07003335 if (tcp_should_expand_sndbuf(sk, tp)) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07003336 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3338 demanded = max_t(unsigned int, tp->snd_cwnd,
3339 tp->reordering + 1);
3340 sndmem *= 2*demanded;
3341 if (sndmem > sk->sk_sndbuf)
3342 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3343 tp->snd_cwnd_stamp = tcp_time_stamp;
3344 }
3345
3346 sk->sk_write_space(sk);
3347}
3348
3349static inline void tcp_check_space(struct sock *sk)
3350{
3351 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3352 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3353 if (sk->sk_socket &&
3354 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3355 tcp_new_space(sk);
3356 }
3357}
3358
David S. Miller55c97f32005-07-05 15:19:38 -07003359static __inline__ void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360{
David S. Miller55c97f32005-07-05 15:19:38 -07003361 tcp_push_pending_frames(sk, tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362 tcp_check_space(sk);
3363}
3364
3365/*
3366 * Check if sending an ack is needed.
3367 */
3368static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3369{
3370 struct tcp_sock *tp = tcp_sk(sk);
3371
3372 /* More than one full frame received... */
3373 if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss
3374 /* ... and right edge of window advances far enough.
3375 * (tcp_recvmsg() will send ACK otherwise). Or...
3376 */
3377 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3378 /* We ACK each frame or... */
3379 tcp_in_quickack_mode(tp) ||
3380 /* We have out of order data. */
3381 (ofo_possible &&
3382 skb_peek(&tp->out_of_order_queue))) {
3383 /* Then ack it now */
3384 tcp_send_ack(sk);
3385 } else {
3386 /* Else, send delayed ack. */
3387 tcp_send_delayed_ack(sk);
3388 }
3389}
3390
3391static __inline__ void tcp_ack_snd_check(struct sock *sk)
3392{
3393 struct tcp_sock *tp = tcp_sk(sk);
3394 if (!tcp_ack_scheduled(tp)) {
3395 /* We sent a data segment already. */
3396 return;
3397 }
3398 __tcp_ack_snd_check(sk, 1);
3399}
3400
3401/*
3402 * This routine is only called when we have urgent data
3403 * signalled. Its the 'slow' part of tcp_urg. It could be
3404 * moved inline now as tcp_urg is only called from one
3405 * place. We handle URGent data wrong. We have to - as
3406 * BSD still doesn't use the correction from RFC961.
3407 * For 1003.1g we should support a new option TCP_STDURG to permit
3408 * either form (or just set the sysctl tcp_stdurg).
3409 */
3410
3411static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3412{
3413 struct tcp_sock *tp = tcp_sk(sk);
3414 u32 ptr = ntohs(th->urg_ptr);
3415
3416 if (ptr && !sysctl_tcp_stdurg)
3417 ptr--;
3418 ptr += ntohl(th->seq);
3419
3420 /* Ignore urgent data that we've already seen and read. */
3421 if (after(tp->copied_seq, ptr))
3422 return;
3423
3424 /* Do not replay urg ptr.
3425 *
3426 * NOTE: interesting situation not covered by specs.
3427 * Misbehaving sender may send urg ptr, pointing to segment,
3428 * which we already have in ofo queue. We are not able to fetch
3429 * such data and will stay in TCP_URG_NOTYET until will be eaten
3430 * by recvmsg(). Seems, we are not obliged to handle such wicked
3431 * situations. But it is worth to think about possibility of some
3432 * DoSes using some hypothetical application level deadlock.
3433 */
3434 if (before(ptr, tp->rcv_nxt))
3435 return;
3436
3437 /* Do we already have a newer (or duplicate) urgent pointer? */
3438 if (tp->urg_data && !after(ptr, tp->urg_seq))
3439 return;
3440
3441 /* Tell the world about our new urgent pointer. */
3442 sk_send_sigurg(sk);
3443
3444 /* We may be adding urgent data when the last byte read was
3445 * urgent. To do this requires some care. We cannot just ignore
3446 * tp->copied_seq since we would read the last urgent byte again
3447 * as data, nor can we alter copied_seq until this data arrives
3448 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3449 *
3450 * NOTE. Double Dutch. Rendering to plain English: author of comment
3451 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3452 * and expect that both A and B disappear from stream. This is _wrong_.
3453 * Though this happens in BSD with high probability, this is occasional.
3454 * Any application relying on this is buggy. Note also, that fix "works"
3455 * only in this artificial test. Insert some normal data between A and B and we will
3456 * decline of BSD again. Verdict: it is better to remove to trap
3457 * buggy users.
3458 */
3459 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3460 !sock_flag(sk, SOCK_URGINLINE) &&
3461 tp->copied_seq != tp->rcv_nxt) {
3462 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3463 tp->copied_seq++;
3464 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3465 __skb_unlink(skb, skb->list);
3466 __kfree_skb(skb);
3467 }
3468 }
3469
3470 tp->urg_data = TCP_URG_NOTYET;
3471 tp->urg_seq = ptr;
3472
3473 /* Disable header prediction. */
3474 tp->pred_flags = 0;
3475}
3476
3477/* This is the 'fast' part of urgent handling. */
3478static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3479{
3480 struct tcp_sock *tp = tcp_sk(sk);
3481
3482 /* Check if we get a new urgent pointer - normally not. */
3483 if (th->urg)
3484 tcp_check_urg(sk,th);
3485
3486 /* Do we wait for any urgent data? - normally not... */
3487 if (tp->urg_data == TCP_URG_NOTYET) {
3488 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3489 th->syn;
3490
3491 /* Is the urgent pointer pointing into this packet? */
3492 if (ptr < skb->len) {
3493 u8 tmp;
3494 if (skb_copy_bits(skb, ptr, &tmp, 1))
3495 BUG();
3496 tp->urg_data = TCP_URG_VALID | tmp;
3497 if (!sock_flag(sk, SOCK_DEAD))
3498 sk->sk_data_ready(sk, 0);
3499 }
3500 }
3501}
3502
3503static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3504{
3505 struct tcp_sock *tp = tcp_sk(sk);
3506 int chunk = skb->len - hlen;
3507 int err;
3508
3509 local_bh_enable();
3510 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3511 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3512 else
3513 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3514 tp->ucopy.iov);
3515
3516 if (!err) {
3517 tp->ucopy.len -= chunk;
3518 tp->copied_seq += chunk;
3519 tcp_rcv_space_adjust(sk);
3520 }
3521
3522 local_bh_disable();
3523 return err;
3524}
3525
3526static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3527{
3528 int result;
3529
3530 if (sock_owned_by_user(sk)) {
3531 local_bh_enable();
3532 result = __tcp_checksum_complete(skb);
3533 local_bh_disable();
3534 } else {
3535 result = __tcp_checksum_complete(skb);
3536 }
3537 return result;
3538}
3539
3540static __inline__ int
3541tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3542{
3543 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3544 __tcp_checksum_complete_user(sk, skb);
3545}
3546
3547/*
3548 * TCP receive function for the ESTABLISHED state.
3549 *
3550 * It is split into a fast path and a slow path. The fast path is
3551 * disabled when:
3552 * - A zero window was announced from us - zero window probing
3553 * is only handled properly in the slow path.
3554 * - Out of order segments arrived.
3555 * - Urgent data is expected.
3556 * - There is no buffer space left
3557 * - Unexpected TCP flags/window values/header lengths are received
3558 * (detected by checking the TCP header against pred_flags)
3559 * - Data is sent in both directions. Fast path only supports pure senders
3560 * or pure receivers (this means either the sequence number or the ack
3561 * value must stay constant)
3562 * - Unexpected TCP option.
3563 *
3564 * When these conditions are not satisfied it drops into a standard
3565 * receive procedure patterned after RFC793 to handle all cases.
3566 * The first three cases are guaranteed by proper pred_flags setting,
3567 * the rest is checked inline. Fast processing is turned on in
3568 * tcp_data_queue when everything is OK.
3569 */
3570int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3571 struct tcphdr *th, unsigned len)
3572{
3573 struct tcp_sock *tp = tcp_sk(sk);
3574
3575 /*
3576 * Header prediction.
3577 * The code loosely follows the one in the famous
3578 * "30 instruction TCP receive" Van Jacobson mail.
3579 *
3580 * Van's trick is to deposit buffers into socket queue
3581 * on a device interrupt, to call tcp_recv function
3582 * on the receive process context and checksum and copy
3583 * the buffer to user space. smart...
3584 *
3585 * Our current scheme is not silly either but we take the
3586 * extra cost of the net_bh soft interrupt processing...
3587 * We do checksum and copy also but from device to kernel.
3588 */
3589
3590 tp->rx_opt.saw_tstamp = 0;
3591
3592 /* pred_flags is 0xS?10 << 16 + snd_wnd
3593 * if header_predition is to be made
3594 * 'S' will always be tp->tcp_header_len >> 2
3595 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3596 * turn it off (when there are holes in the receive
3597 * space for instance)
3598 * PSH flag is ignored.
3599 */
3600
3601 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3602 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3603 int tcp_header_len = tp->tcp_header_len;
3604
3605 /* Timestamp header prediction: tcp_header_len
3606 * is automatically equal to th->doff*4 due to pred_flags
3607 * match.
3608 */
3609
3610 /* Check timestamp */
3611 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3612 __u32 *ptr = (__u32 *)(th + 1);
3613
3614 /* No? Slow path! */
3615 if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3616 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3617 goto slow_path;
3618
3619 tp->rx_opt.saw_tstamp = 1;
3620 ++ptr;
3621 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3622 ++ptr;
3623 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3624
3625 /* If PAWS failed, check it more carefully in slow path */
3626 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
3627 goto slow_path;
3628
3629 /* DO NOT update ts_recent here, if checksum fails
3630 * and timestamp was corrupted part, it will result
3631 * in a hung connection since we will drop all
3632 * future packets due to the PAWS test.
3633 */
3634 }
3635
3636 if (len <= tcp_header_len) {
3637 /* Bulk data transfer: sender */
3638 if (len == tcp_header_len) {
3639 /* Predicted packet is in window by definition.
3640 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3641 * Hence, check seq<=rcv_wup reduces to:
3642 */
3643 if (tcp_header_len ==
3644 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3645 tp->rcv_nxt == tp->rcv_wup)
3646 tcp_store_ts_recent(tp);
3647
3648 tcp_rcv_rtt_measure_ts(tp, skb);
3649
3650 /* We know that such packets are checksummed
3651 * on entry.
3652 */
3653 tcp_ack(sk, skb, 0);
3654 __kfree_skb(skb);
David S. Miller55c97f32005-07-05 15:19:38 -07003655 tcp_data_snd_check(sk, tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656 return 0;
3657 } else { /* Header too small */
3658 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3659 goto discard;
3660 }
3661 } else {
3662 int eaten = 0;
3663
3664 if (tp->ucopy.task == current &&
3665 tp->copied_seq == tp->rcv_nxt &&
3666 len - tcp_header_len <= tp->ucopy.len &&
3667 sock_owned_by_user(sk)) {
3668 __set_current_state(TASK_RUNNING);
3669
3670 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
3671 /* Predicted packet is in window by definition.
3672 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3673 * Hence, check seq<=rcv_wup reduces to:
3674 */
3675 if (tcp_header_len ==
3676 (sizeof(struct tcphdr) +
3677 TCPOLEN_TSTAMP_ALIGNED) &&
3678 tp->rcv_nxt == tp->rcv_wup)
3679 tcp_store_ts_recent(tp);
3680
3681 tcp_rcv_rtt_measure_ts(tp, skb);
3682
3683 __skb_pull(skb, tcp_header_len);
3684 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3685 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
3686 eaten = 1;
3687 }
3688 }
3689 if (!eaten) {
3690 if (tcp_checksum_complete_user(sk, skb))
3691 goto csum_error;
3692
3693 /* Predicted packet is in window by definition.
3694 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3695 * Hence, check seq<=rcv_wup reduces to:
3696 */
3697 if (tcp_header_len ==
3698 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3699 tp->rcv_nxt == tp->rcv_wup)
3700 tcp_store_ts_recent(tp);
3701
3702 tcp_rcv_rtt_measure_ts(tp, skb);
3703
3704 if ((int)skb->truesize > sk->sk_forward_alloc)
3705 goto step5;
3706
3707 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
3708
3709 /* Bulk data transfer: receiver */
3710 __skb_pull(skb,tcp_header_len);
3711 __skb_queue_tail(&sk->sk_receive_queue, skb);
3712 sk_stream_set_owner_r(skb, sk);
3713 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3714 }
3715
3716 tcp_event_data_recv(sk, tp, skb);
3717
3718 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
3719 /* Well, only one small jumplet in fast path... */
3720 tcp_ack(sk, skb, FLAG_DATA);
David S. Miller55c97f32005-07-05 15:19:38 -07003721 tcp_data_snd_check(sk, tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722 if (!tcp_ack_scheduled(tp))
3723 goto no_ack;
3724 }
3725
David S. Miller31432412005-05-23 12:03:06 -07003726 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727no_ack:
3728 if (eaten)
3729 __kfree_skb(skb);
3730 else
3731 sk->sk_data_ready(sk, 0);
3732 return 0;
3733 }
3734 }
3735
3736slow_path:
3737 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
3738 goto csum_error;
3739
3740 /*
3741 * RFC1323: H1. Apply PAWS check first.
3742 */
3743 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
3744 tcp_paws_discard(tp, skb)) {
3745 if (!th->rst) {
3746 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
3747 tcp_send_dupack(sk, skb);
3748 goto discard;
3749 }
3750 /* Resets are accepted even if PAWS failed.
3751
3752 ts_recent update must be made after we are sure
3753 that the packet is in window.
3754 */
3755 }
3756
3757 /*
3758 * Standard slow path.
3759 */
3760
3761 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3762 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3763 * (RST) segments are validated by checking their SEQ-fields."
3764 * And page 69: "If an incoming segment is not acceptable,
3765 * an acknowledgment should be sent in reply (unless the RST bit
3766 * is set, if so drop the segment and return)".
3767 */
3768 if (!th->rst)
3769 tcp_send_dupack(sk, skb);
3770 goto discard;
3771 }
3772
3773 if(th->rst) {
3774 tcp_reset(sk);
3775 goto discard;
3776 }
3777
3778 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3779
3780 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3781 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3782 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
3783 tcp_reset(sk);
3784 return 1;
3785 }
3786
3787step5:
3788 if(th->ack)
3789 tcp_ack(sk, skb, FLAG_SLOWPATH);
3790
3791 tcp_rcv_rtt_measure_ts(tp, skb);
3792
3793 /* Process urgent data. */
3794 tcp_urg(sk, skb, th);
3795
3796 /* step 7: process the segment text */
3797 tcp_data_queue(sk, skb);
3798
David S. Miller55c97f32005-07-05 15:19:38 -07003799 tcp_data_snd_check(sk, tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800 tcp_ack_snd_check(sk);
3801 return 0;
3802
3803csum_error:
3804 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3805
3806discard:
3807 __kfree_skb(skb);
3808 return 0;
3809}
3810
3811static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
3812 struct tcphdr *th, unsigned len)
3813{
3814 struct tcp_sock *tp = tcp_sk(sk);
3815 int saved_clamp = tp->rx_opt.mss_clamp;
3816
3817 tcp_parse_options(skb, &tp->rx_opt, 0);
3818
3819 if (th->ack) {
3820 /* rfc793:
3821 * "If the state is SYN-SENT then
3822 * first check the ACK bit
3823 * If the ACK bit is set
3824 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3825 * a reset (unless the RST bit is set, if so drop
3826 * the segment and return)"
3827 *
3828 * We do not send data with SYN, so that RFC-correct
3829 * test reduces to:
3830 */
3831 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
3832 goto reset_and_undo;
3833
3834 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3835 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
3836 tcp_time_stamp)) {
3837 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
3838 goto reset_and_undo;
3839 }
3840
3841 /* Now ACK is acceptable.
3842 *
3843 * "If the RST bit is set
3844 * If the ACK was acceptable then signal the user "error:
3845 * connection reset", drop the segment, enter CLOSED state,
3846 * delete TCB, and return."
3847 */
3848
3849 if (th->rst) {
3850 tcp_reset(sk);
3851 goto discard;
3852 }
3853
3854 /* rfc793:
3855 * "fifth, if neither of the SYN or RST bits is set then
3856 * drop the segment and return."
3857 *
3858 * See note below!
3859 * --ANK(990513)
3860 */
3861 if (!th->syn)
3862 goto discard_and_undo;
3863
3864 /* rfc793:
3865 * "If the SYN bit is on ...
3866 * are acceptable then ...
3867 * (our SYN has been ACKed), change the connection
3868 * state to ESTABLISHED..."
3869 */
3870
3871 TCP_ECN_rcv_synack(tp, th);
3872 if (tp->ecn_flags&TCP_ECN_OK)
3873 sock_set_flag(sk, SOCK_NO_LARGESEND);
3874
3875 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
3876 tcp_ack(sk, skb, FLAG_SLOWPATH);
3877
3878 /* Ok.. it's good. Set up sequence numbers and
3879 * move to established.
3880 */
3881 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
3882 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
3883
3884 /* RFC1323: The window in SYN & SYN/ACK segments is
3885 * never scaled.
3886 */
3887 tp->snd_wnd = ntohs(th->window);
3888 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
3889
3890 if (!tp->rx_opt.wscale_ok) {
3891 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
3892 tp->window_clamp = min(tp->window_clamp, 65535U);
3893 }
3894
3895 if (tp->rx_opt.saw_tstamp) {
3896 tp->rx_opt.tstamp_ok = 1;
3897 tp->tcp_header_len =
3898 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
3899 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
3900 tcp_store_ts_recent(tp);
3901 } else {
3902 tp->tcp_header_len = sizeof(struct tcphdr);
3903 }
3904
3905 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
3906 tp->rx_opt.sack_ok |= 2;
3907
3908 tcp_sync_mss(sk, tp->pmtu_cookie);
3909 tcp_initialize_rcv_mss(sk);
3910
3911 /* Remember, tcp_poll() does not lock socket!
3912 * Change state from SYN-SENT only after copied_seq
3913 * is initialized. */
3914 tp->copied_seq = tp->rcv_nxt;
3915 mb();
3916 tcp_set_state(sk, TCP_ESTABLISHED);
3917
3918 /* Make sure socket is routed, for correct metrics. */
3919 tp->af_specific->rebuild_header(sk);
3920
3921 tcp_init_metrics(sk);
3922
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003923 tcp_init_congestion_control(tp);
3924
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925 /* Prevent spurious tcp_cwnd_restart() on first data
3926 * packet.
3927 */
3928 tp->lsndtime = tcp_time_stamp;
3929
3930 tcp_init_buffer_space(sk);
3931
3932 if (sock_flag(sk, SOCK_KEEPOPEN))
3933 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
3934
3935 if (!tp->rx_opt.snd_wscale)
3936 __tcp_fast_path_on(tp, tp->snd_wnd);
3937 else
3938 tp->pred_flags = 0;
3939
3940 if (!sock_flag(sk, SOCK_DEAD)) {
3941 sk->sk_state_change(sk);
3942 sk_wake_async(sk, 0, POLL_OUT);
3943 }
3944
3945 if (sk->sk_write_pending || tp->defer_accept || tp->ack.pingpong) {
3946 /* Save one ACK. Data will be ready after
3947 * several ticks, if write_pending is set.
3948 *
3949 * It may be deleted, but with this feature tcpdumps
3950 * look so _wonderfully_ clever, that I was not able
3951 * to stand against the temptation 8) --ANK
3952 */
3953 tcp_schedule_ack(tp);
3954 tp->ack.lrcvtime = tcp_time_stamp;
3955 tp->ack.ato = TCP_ATO_MIN;
3956 tcp_incr_quickack(tp);
3957 tcp_enter_quickack_mode(tp);
3958 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
3959
3960discard:
3961 __kfree_skb(skb);
3962 return 0;
3963 } else {
3964 tcp_send_ack(sk);
3965 }
3966 return -1;
3967 }
3968
3969 /* No ACK in the segment */
3970
3971 if (th->rst) {
3972 /* rfc793:
3973 * "If the RST bit is set
3974 *
3975 * Otherwise (no ACK) drop the segment and return."
3976 */
3977
3978 goto discard_and_undo;
3979 }
3980
3981 /* PAWS check. */
3982 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
3983 goto discard_and_undo;
3984
3985 if (th->syn) {
3986 /* We see SYN without ACK. It is attempt of
3987 * simultaneous connect with crossed SYNs.
3988 * Particularly, it can be connect to self.
3989 */
3990 tcp_set_state(sk, TCP_SYN_RECV);
3991
3992 if (tp->rx_opt.saw_tstamp) {
3993 tp->rx_opt.tstamp_ok = 1;
3994 tcp_store_ts_recent(tp);
3995 tp->tcp_header_len =
3996 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
3997 } else {
3998 tp->tcp_header_len = sizeof(struct tcphdr);
3999 }
4000
4001 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4002 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4003
4004 /* RFC1323: The window in SYN & SYN/ACK segments is
4005 * never scaled.
4006 */
4007 tp->snd_wnd = ntohs(th->window);
4008 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4009 tp->max_window = tp->snd_wnd;
4010
4011 TCP_ECN_rcv_syn(tp, th);
4012 if (tp->ecn_flags&TCP_ECN_OK)
4013 sock_set_flag(sk, SOCK_NO_LARGESEND);
4014
4015 tcp_sync_mss(sk, tp->pmtu_cookie);
4016 tcp_initialize_rcv_mss(sk);
4017
4018
4019 tcp_send_synack(sk);
4020#if 0
4021 /* Note, we could accept data and URG from this segment.
4022 * There are no obstacles to make this.
4023 *
4024 * However, if we ignore data in ACKless segments sometimes,
4025 * we have no reasons to accept it sometimes.
4026 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4027 * is not flawless. So, discard packet for sanity.
4028 * Uncomment this return to process the data.
4029 */
4030 return -1;
4031#else
4032 goto discard;
4033#endif
4034 }
4035 /* "fifth, if neither of the SYN or RST bits is set then
4036 * drop the segment and return."
4037 */
4038
4039discard_and_undo:
4040 tcp_clear_options(&tp->rx_opt);
4041 tp->rx_opt.mss_clamp = saved_clamp;
4042 goto discard;
4043
4044reset_and_undo:
4045 tcp_clear_options(&tp->rx_opt);
4046 tp->rx_opt.mss_clamp = saved_clamp;
4047 return 1;
4048}
4049
4050
4051/*
4052 * This function implements the receiving procedure of RFC 793 for
4053 * all states except ESTABLISHED and TIME_WAIT.
4054 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4055 * address independent.
4056 */
4057
4058int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4059 struct tcphdr *th, unsigned len)
4060{
4061 struct tcp_sock *tp = tcp_sk(sk);
4062 int queued = 0;
4063
4064 tp->rx_opt.saw_tstamp = 0;
4065
4066 switch (sk->sk_state) {
4067 case TCP_CLOSE:
4068 goto discard;
4069
4070 case TCP_LISTEN:
4071 if(th->ack)
4072 return 1;
4073
4074 if(th->rst)
4075 goto discard;
4076
4077 if(th->syn) {
4078 if(tp->af_specific->conn_request(sk, skb) < 0)
4079 return 1;
4080
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 /* Now we have several options: In theory there is
4082 * nothing else in the frame. KA9Q has an option to
4083 * send data with the syn, BSD accepts data with the
4084 * syn up to the [to be] advertised window and
4085 * Solaris 2.1 gives you a protocol error. For now
4086 * we just ignore it, that fits the spec precisely
4087 * and avoids incompatibilities. It would be nice in
4088 * future to drop through and process the data.
4089 *
4090 * Now that TTCP is starting to be used we ought to
4091 * queue this data.
4092 * But, this leaves one open to an easy denial of
4093 * service attack, and SYN cookies can't defend
4094 * against this problem. So, we drop the data
4095 * in the interest of security over speed.
4096 */
4097 goto discard;
4098 }
4099 goto discard;
4100
4101 case TCP_SYN_SENT:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4103 if (queued >= 0)
4104 return queued;
4105
4106 /* Do step6 onward by hand. */
4107 tcp_urg(sk, skb, th);
4108 __kfree_skb(skb);
David S. Miller55c97f32005-07-05 15:19:38 -07004109 tcp_data_snd_check(sk, tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110 return 0;
4111 }
4112
4113 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4114 tcp_paws_discard(tp, skb)) {
4115 if (!th->rst) {
4116 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4117 tcp_send_dupack(sk, skb);
4118 goto discard;
4119 }
4120 /* Reset is accepted even if it did not pass PAWS. */
4121 }
4122
4123 /* step 1: check sequence number */
4124 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4125 if (!th->rst)
4126 tcp_send_dupack(sk, skb);
4127 goto discard;
4128 }
4129
4130 /* step 2: check RST bit */
4131 if(th->rst) {
4132 tcp_reset(sk);
4133 goto discard;
4134 }
4135
4136 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4137
4138 /* step 3: check security and precedence [ignored] */
4139
4140 /* step 4:
4141 *
4142 * Check for a SYN in window.
4143 */
4144 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4145 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4146 tcp_reset(sk);
4147 return 1;
4148 }
4149
4150 /* step 5: check the ACK field */
4151 if (th->ack) {
4152 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4153
4154 switch(sk->sk_state) {
4155 case TCP_SYN_RECV:
4156 if (acceptable) {
4157 tp->copied_seq = tp->rcv_nxt;
4158 mb();
4159 tcp_set_state(sk, TCP_ESTABLISHED);
4160 sk->sk_state_change(sk);
4161
4162 /* Note, that this wakeup is only for marginal
4163 * crossed SYN case. Passively open sockets
4164 * are not waked up, because sk->sk_sleep ==
4165 * NULL and sk->sk_socket == NULL.
4166 */
4167 if (sk->sk_socket) {
4168 sk_wake_async(sk,0,POLL_OUT);
4169 }
4170
4171 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4172 tp->snd_wnd = ntohs(th->window) <<
4173 tp->rx_opt.snd_wscale;
4174 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4175 TCP_SKB_CB(skb)->seq);
4176
4177 /* tcp_ack considers this ACK as duplicate
4178 * and does not calculate rtt.
4179 * Fix it at least with timestamps.
4180 */
4181 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4182 !tp->srtt)
Stephen Hemminger317a76f2005-06-23 12:19:55 -07004183 tcp_ack_saw_tstamp(tp, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184
4185 if (tp->rx_opt.tstamp_ok)
4186 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4187
4188 /* Make sure socket is routed, for
4189 * correct metrics.
4190 */
4191 tp->af_specific->rebuild_header(sk);
4192
4193 tcp_init_metrics(sk);
4194
Stephen Hemminger317a76f2005-06-23 12:19:55 -07004195 tcp_init_congestion_control(tp);
4196
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197 /* Prevent spurious tcp_cwnd_restart() on
4198 * first data packet.
4199 */
4200 tp->lsndtime = tcp_time_stamp;
4201
4202 tcp_initialize_rcv_mss(sk);
4203 tcp_init_buffer_space(sk);
4204 tcp_fast_path_on(tp);
4205 } else {
4206 return 1;
4207 }
4208 break;
4209
4210 case TCP_FIN_WAIT1:
4211 if (tp->snd_una == tp->write_seq) {
4212 tcp_set_state(sk, TCP_FIN_WAIT2);
4213 sk->sk_shutdown |= SEND_SHUTDOWN;
4214 dst_confirm(sk->sk_dst_cache);
4215
4216 if (!sock_flag(sk, SOCK_DEAD))
4217 /* Wake up lingering close() */
4218 sk->sk_state_change(sk);
4219 else {
4220 int tmo;
4221
4222 if (tp->linger2 < 0 ||
4223 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4224 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4225 tcp_done(sk);
4226 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4227 return 1;
4228 }
4229
4230 tmo = tcp_fin_time(tp);
4231 if (tmo > TCP_TIMEWAIT_LEN) {
4232 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4233 } else if (th->fin || sock_owned_by_user(sk)) {
4234 /* Bad case. We could lose such FIN otherwise.
4235 * It is not a big problem, but it looks confusing
4236 * and not so rare event. We still can lose it now,
4237 * if it spins in bh_lock_sock(), but it is really
4238 * marginal case.
4239 */
4240 tcp_reset_keepalive_timer(sk, tmo);
4241 } else {
4242 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4243 goto discard;
4244 }
4245 }
4246 }
4247 break;
4248
4249 case TCP_CLOSING:
4250 if (tp->snd_una == tp->write_seq) {
4251 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4252 goto discard;
4253 }
4254 break;
4255
4256 case TCP_LAST_ACK:
4257 if (tp->snd_una == tp->write_seq) {
4258 tcp_update_metrics(sk);
4259 tcp_done(sk);
4260 goto discard;
4261 }
4262 break;
4263 }
4264 } else
4265 goto discard;
4266
4267 /* step 6: check the URG bit */
4268 tcp_urg(sk, skb, th);
4269
4270 /* step 7: process the segment text */
4271 switch (sk->sk_state) {
4272 case TCP_CLOSE_WAIT:
4273 case TCP_CLOSING:
4274 case TCP_LAST_ACK:
4275 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4276 break;
4277 case TCP_FIN_WAIT1:
4278 case TCP_FIN_WAIT2:
4279 /* RFC 793 says to queue data in these states,
4280 * RFC 1122 says we MUST send a reset.
4281 * BSD 4.4 also does reset.
4282 */
4283 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4284 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4285 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4286 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4287 tcp_reset(sk);
4288 return 1;
4289 }
4290 }
4291 /* Fall through */
4292 case TCP_ESTABLISHED:
4293 tcp_data_queue(sk, skb);
4294 queued = 1;
4295 break;
4296 }
4297
4298 /* tcp_data could move socket to TIME-WAIT */
4299 if (sk->sk_state != TCP_CLOSE) {
David S. Miller55c97f32005-07-05 15:19:38 -07004300 tcp_data_snd_check(sk, tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301 tcp_ack_snd_check(sk);
4302 }
4303
4304 if (!queued) {
4305discard:
4306 __kfree_skb(skb);
4307 }
4308 return 0;
4309}
4310
4311EXPORT_SYMBOL(sysctl_tcp_ecn);
4312EXPORT_SYMBOL(sysctl_tcp_reordering);
4313EXPORT_SYMBOL(tcp_parse_options);
4314EXPORT_SYMBOL(tcp_rcv_established);
4315EXPORT_SYMBOL(tcp_rcv_state_process);