File: /usr/src/linux/net/ipv4/tcp_input.c
1 /*
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 1.236 2001/09/18 22:29:09 davem Exp $
9 *
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
11 * 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 */
64
65 #include <linux/config.h>
66 #include <linux/mm.h>
67 #include <linux/sysctl.h>
68 #include <net/tcp.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71
72
73 /* These are on by default so the code paths get tested.
74 * For the final 2.2 this may be undone at our discretion. -DaveM
75 */
76 int sysctl_tcp_timestamps = 1;
77 int sysctl_tcp_window_scaling = 1;
78 int sysctl_tcp_sack = 1;
79 int sysctl_tcp_fack = 1;
80 int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
81 #ifdef CONFIG_INET_ECN
82 int sysctl_tcp_ecn = 1;
83 #else
84 int sysctl_tcp_ecn = 0;
85 #endif
86 int sysctl_tcp_dsack = 1;
87 int sysctl_tcp_app_win = 31;
88 int sysctl_tcp_adv_win_scale = 2;
89
90 int sysctl_tcp_stdurg = 0;
91 int sysctl_tcp_rfc1337 = 0;
92 int sysctl_tcp_max_orphans = NR_FILE;
93
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108
109 #define IsReno(tp) ((tp)->sack_ok == 0)
110 #define IsFack(tp) ((tp)->sack_ok & 2)
111 #define IsDSack(tp) ((tp)->sack_ok & 4)
112
113 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
114
115 /* Adapt the MSS value used to make delayed ack decision to the
116 * real world.
117 */
118 static __inline__ void tcp_measure_rcv_mss(struct tcp_opt *tp, 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 /* Dubious? Rather, it is final cut. 8) */
132 if (tcp_flag_word(skb->h.th)&TCP_REMNANT)
133 tp->ack.pending |= TCP_ACK_PUSHED;
134 } else {
135 /* Otherwise, we make more careful check taking into account,
136 * that SACKs block is variable.
137 *
138 * "len" is invariant segment length, including TCP header.
139 */
140 len += skb->data - skb->h.raw;
141 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
142 /* If PSH is not set, packet should be
143 * full sized, provided peer TCP is not badly broken.
144 * This observation (if it is correct 8)) allows
145 * to handle super-low mtu links fairly.
146 */
147 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
148 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
149 /* Subtract also invariant (if peer is RFC compliant),
150 * tcp header plus fixed timestamp option length.
151 * Resulting "len" is MSS free of SACK jitter.
152 */
153 len -= tp->tcp_header_len;
154 tp->ack.last_seg_size = len;
155 if (len == lss) {
156 tp->ack.rcv_mss = len;
157 return;
158 }
159 }
160 tp->ack.pending |= TCP_ACK_PUSHED;
161 }
162 }
163
164 static void tcp_incr_quickack(struct tcp_opt *tp)
165 {
166 unsigned quickacks = tp->rcv_wnd/(2*tp->ack.rcv_mss);
167
168 if (quickacks==0)
169 quickacks=2;
170 if (quickacks > tp->ack.quick)
171 tp->ack.quick = min_t(unsigned int, quickacks, TCP_MAX_QUICKACKS);
172 }
173
174 void tcp_enter_quickack_mode(struct tcp_opt *tp)
175 {
176 tcp_incr_quickack(tp);
177 tp->ack.pingpong = 0;
178 tp->ack.ato = TCP_ATO_MIN;
179 }
180
181 /* Send ACKs quickly, if "quick" count is not exhausted
182 * and the session is not interactive.
183 */
184
185 static __inline__ int tcp_in_quickack_mode(struct tcp_opt *tp)
186 {
187 return (tp->ack.quick && !tp->ack.pingpong);
188 }
189
190 /* Buffer size and advertised window tuning.
191 *
192 * 1. Tuning sk->sndbuf, when connection enters established state.
193 */
194
195 static void tcp_fixup_sndbuf(struct sock *sk)
196 {
197 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
198 int sndmem = tp->mss_clamp+MAX_TCP_HEADER+16+sizeof(struct sk_buff);
199
200 if (sk->sndbuf < 3*sndmem)
201 sk->sndbuf = min_t(int, 3*sndmem, sysctl_tcp_wmem[2]);
202 }
203
204 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
205 *
206 * All tcp_full_space() is split to two parts: "network" buffer, allocated
207 * forward and advertised in receiver window (tp->rcv_wnd) and
208 * "application buffer", required to isolate scheduling/application
209 * latencies from network.
210 * window_clamp is maximal advertised window. It can be less than
211 * tcp_full_space(), in this case tcp_full_space() - window_clamp
212 * is reserved for "application" buffer. The less window_clamp is
213 * the smoother our behaviour from viewpoint of network, but the lower
214 * throughput and the higher sensitivity of the connection to losses. 8)
215 *
216 * rcv_ssthresh is more strict window_clamp used at "slow start"
217 * phase to predict further behaviour of this connection.
218 * It is used for two goals:
219 * - to enforce header prediction at sender, even when application
220 * requires some significant "application buffer". It is check #1.
221 * - to prevent pruning of receive queue because of misprediction
222 * of receiver window. Check #2.
223 *
224 * The scheme does not work when sender sends good segments opening
225 * window and then starts to feed us spagetti. But it should work
226 * in common situations. Otherwise, we have to rely on queue collapsing.
227 */
228
229 /* Slow part of check#2. */
230 static int
231 __tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
232 {
233 /* Optimize this! */
234 int truesize = tcp_win_from_space(skb->truesize)/2;
235 int window = tcp_full_space(sk)/2;
236
237 while (tp->rcv_ssthresh <= window) {
238 if (truesize <= skb->len)
239 return 2*tp->ack.rcv_mss;
240
241 truesize >>= 1;
242 window >>= 1;
243 }
244 return 0;
245 }
246
247 static __inline__ void
248 tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
249 {
250 /* Check #1 */
251 if (tp->rcv_ssthresh < tp->window_clamp &&
252 (int)tp->rcv_ssthresh < tcp_space(sk) &&
253 !tcp_memory_pressure) {
254 int incr;
255
256 /* Check #2. Increase window, if skb with such overhead
257 * will fit to rcvbuf in future.
258 */
259 if (tcp_win_from_space(skb->truesize) <= skb->len)
260 incr = 2*tp->advmss;
261 else
262 incr = __tcp_grow_window(sk, tp, skb);
263
264 if (incr) {
265 tp->rcv_ssthresh = min_t(u32, tp->rcv_ssthresh + incr, tp->window_clamp);
266 tp->ack.quick |= 1;
267 }
268 }
269 }
270
271 /* 3. Tuning rcvbuf, when connection enters established state. */
272
273 static void tcp_fixup_rcvbuf(struct sock *sk)
274 {
275 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
276 int rcvmem = tp->advmss+MAX_TCP_HEADER+16+sizeof(struct sk_buff);
277
278 /* Try to select rcvbuf so that 4 mss-sized segments
279 * will fit to window and correspoding skbs will fit to our rcvbuf.
280 * (was 3; 4 is minimum to allow fast retransmit to work.)
281 */
282 while (tcp_win_from_space(rcvmem) < tp->advmss)
283 rcvmem += 128;
284 if (sk->rcvbuf < 4*rcvmem)
285 sk->rcvbuf = min_t(int, 4*rcvmem, sysctl_tcp_rmem[2]);
286 }
287
288 /* 4. Try to fixup all. It is made iimediately after connection enters
289 * established state.
290 */
291 static void tcp_init_buffer_space(struct sock *sk)
292 {
293 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
294 int maxwin;
295
296 if (!(sk->userlocks&SOCK_RCVBUF_LOCK))
297 tcp_fixup_rcvbuf(sk);
298 if (!(sk->userlocks&SOCK_SNDBUF_LOCK))
299 tcp_fixup_sndbuf(sk);
300
301 maxwin = tcp_full_space(sk);
302
303 if (tp->window_clamp >= maxwin) {
304 tp->window_clamp = maxwin;
305
306 if (sysctl_tcp_app_win && maxwin>4*tp->advmss)
307 tp->window_clamp = max_t(u32, maxwin-(maxwin>>sysctl_tcp_app_win), 4*tp->advmss);
308 }
309
310 /* Force reservation of one segment. */
311 if (sysctl_tcp_app_win &&
312 tp->window_clamp > 2*tp->advmss &&
313 tp->window_clamp + tp->advmss > maxwin)
314 tp->window_clamp = max_t(u32, 2*tp->advmss, maxwin-tp->advmss);
315
316 tp->rcv_ssthresh = min_t(u32, tp->rcv_ssthresh, tp->window_clamp);
317 tp->snd_cwnd_stamp = tcp_time_stamp;
318 }
319
320 /* 5. Recalculate window clamp after socket hit its memory bounds. */
321 static void tcp_clamp_window(struct sock *sk, struct tcp_opt *tp)
322 {
323 struct sk_buff *skb;
324 int app_win = tp->rcv_nxt - tp->copied_seq;
325 int ofo_win = 0;
326
327 tp->ack.quick = 0;
328
329 skb_queue_walk(&tp->out_of_order_queue, skb) {
330 ofo_win += skb->len;
331 }
332
333 /* If overcommit is due to out of order segments,
334 * do not clamp window. Try to expand rcvbuf instead.
335 */
336 if (ofo_win) {
337 if (sk->rcvbuf < sysctl_tcp_rmem[2] &&
338 !(sk->userlocks&SOCK_RCVBUF_LOCK) &&
339 !tcp_memory_pressure &&
340 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
341 sk->rcvbuf = min_t(int, atomic_read(&sk->rmem_alloc), sysctl_tcp_rmem[2]);
342 }
343 if (atomic_read(&sk->rmem_alloc) > sk->rcvbuf) {
344 app_win += ofo_win;
345 if (atomic_read(&sk->rmem_alloc) >= 2*sk->rcvbuf)
346 app_win >>= 1;
347 if (app_win > tp->ack.rcv_mss)
348 app_win -= tp->ack.rcv_mss;
349 app_win = max_t(unsigned int, app_win, 2*tp->advmss);
350
351 if (!ofo_win)
352 tp->window_clamp = min_t(u32, tp->window_clamp, app_win);
353 tp->rcv_ssthresh = min_t(u32, tp->window_clamp, 2*tp->advmss);
354 }
355 }
356
357 /* There is something which you must keep in mind when you analyze the
358 * behavior of the tp->ato delayed ack timeout interval. When a
359 * connection starts up, we want to ack as quickly as possible. The
360 * problem is that "good" TCP's do slow start at the beginning of data
361 * transmission. The means that until we send the first few ACK's the
362 * sender will sit on his end and only queue most of his data, because
363 * he can only send snd_cwnd unacked packets at any given time. For
364 * each ACK we send, he increments snd_cwnd and transmits more of his
365 * queue. -DaveM
366 */
367 static void tcp_event_data_recv(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
368 {
369 u32 now;
370
371 tcp_schedule_ack(tp);
372
373 tcp_measure_rcv_mss(tp, skb);
374
375 now = tcp_time_stamp;
376
377 if (!tp->ack.ato) {
378 /* The _first_ data packet received, initialize
379 * delayed ACK engine.
380 */
381 tcp_incr_quickack(tp);
382 tp->ack.ato = TCP_ATO_MIN;
383 } else {
384 int m = now - tp->ack.lrcvtime;
385
386 if (m <= TCP_ATO_MIN/2) {
387 /* The fastest case is the first. */
388 tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2;
389 } else if (m < tp->ack.ato) {
390 tp->ack.ato = (tp->ack.ato>>1) + m;
391 if (tp->ack.ato > tp->rto)
392 tp->ack.ato = tp->rto;
393 } else if (m > tp->rto) {
394 /* Too long gap. Apparently sender falled to
395 * restart window, so that we send ACKs quickly.
396 */
397 tcp_incr_quickack(tp);
398 tcp_mem_reclaim(sk);
399 }
400 }
401 tp->ack.lrcvtime = now;
402
403 TCP_ECN_check_ce(tp, skb);
404
405 if (skb->len >= 128)
406 tcp_grow_window(sk, tp, skb);
407 }
408
409 /* Called to compute a smoothed rtt estimate. The data fed to this
410 * routine either comes from timestamps, or from segments that were
411 * known _not_ to have been retransmitted [see Karn/Partridge
412 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
413 * piece by Van Jacobson.
414 * NOTE: the next three routines used to be one big routine.
415 * To save cycles in the RFC 1323 implementation it was better to break
416 * it up into three procedures. -- erics
417 */
418 static __inline__ void tcp_rtt_estimator(struct tcp_opt *tp, __u32 mrtt)
419 {
420 long m = mrtt; /* RTT */
421
422 /* The following amusing code comes from Jacobson's
423 * article in SIGCOMM '88. Note that rtt and mdev
424 * are scaled versions of rtt and mean deviation.
425 * This is designed to be as fast as possible
426 * m stands for "measurement".
427 *
428 * On a 1990 paper the rto value is changed to:
429 * RTO = rtt + 4 * mdev
430 *
431 * Funny. This algorithm seems to be very broken.
432 * These formulae increase RTO, when it should be decreased, increase
433 * too slowly, when it should be incresed fastly, decrease too fastly
434 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
435 * does not matter how to _calculate_ it. Seems, it was trap
436 * that VJ failed to avoid. 8)
437 */
438 if(m == 0)
439 m = 1;
440 if (tp->srtt != 0) {
441 m -= (tp->srtt >> 3); /* m is now error in rtt est */
442 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
443 if (m < 0) {
444 m = -m; /* m is now abs(error) */
445 m -= (tp->mdev >> 2); /* similar update on mdev */
446 /* This is similar to one of Eifel findings.
447 * Eifel blocks mdev updates when rtt decreases.
448 * This solution is a bit different: we use finer gain
449 * for mdev in this case (alpha*beta).
450 * Like Eifel it also prevents growth of rto,
451 * but also it limits too fast rto decreases,
452 * happening in pure Eifel.
453 */
454 if (m > 0)
455 m >>= 3;
456 } else {
457 m -= (tp->mdev >> 2); /* similar update on mdev */
458 }
459 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
460 if (tp->mdev > tp->mdev_max) {
461 tp->mdev_max = tp->mdev;
462 if (tp->mdev_max > tp->rttvar)
463 tp->rttvar = tp->mdev_max;
464 }
465 if (after(tp->snd_una, tp->rtt_seq)) {
466 if (tp->mdev_max < tp->rttvar)
467 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
468 tp->rtt_seq = tp->snd_una;
469 tp->mdev_max = TCP_RTO_MIN;
470 }
471 } else {
472 /* no previous measure. */
473 tp->srtt = m<<3; /* take the measured time to be rtt */
474 tp->mdev = m<<2; /* make sure rto = 3*rtt */
475 tp->mdev_max = tp->rttvar = max_t(u32, tp->mdev, TCP_RTO_MIN);
476 tp->rtt_seq = tp->snd_nxt;
477 }
478 }
479
480 /* Calculate rto without backoff. This is the second half of Van Jacobson's
481 * routine referred to above.
482 */
483 static __inline__ void tcp_set_rto(struct tcp_opt *tp)
484 {
485 /* Old crap is replaced with new one. 8)
486 *
487 * More seriously:
488 * 1. If rtt variance happened to be less 50msec, it is hallucination.
489 * It cannot be less due to utterly erratic ACK generation made
490 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
491 * to do with delayed acks, because at cwnd>2 true delack timeout
492 * is invisible. Actually, Linux-2.4 also generates erratic
493 * ACKs in some curcumstances.
494 */
495 tp->rto = (tp->srtt >> 3) + tp->rttvar;
496
497 /* 2. Fixups made earlier cannot be right.
498 * If we do not estimate RTO correctly without them,
499 * all the algo is pure shit and should be replaced
500 * with correct one. It is exaclty, which we pretend to do.
501 */
502 }
503
504 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
505 * guarantees that rto is higher.
506 */
507 static __inline__ void tcp_bound_rto(struct tcp_opt *tp)
508 {
509 if (tp->rto > TCP_RTO_MAX)
510 tp->rto = TCP_RTO_MAX;
511 }
512
513 /* Save metrics learned by this TCP session.
514 This function is called only, when TCP finishes successfully
515 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
516 */
517 void tcp_update_metrics(struct sock *sk)
518 {
519 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
520 struct dst_entry *dst = __sk_dst_get(sk);
521
522 dst_confirm(dst);
523
524 if (dst && (dst->flags&DST_HOST)) {
525 int m;
526
527 if (tp->backoff || !tp->srtt) {
528 /* This session failed to estimate rtt. Why?
529 * Probably, no packets returned in time.
530 * Reset our results.
531 */
532 if (!(dst->mxlock&(1<<RTAX_RTT)))
533 dst->rtt = 0;
534 return;
535 }
536
537 m = dst->rtt - tp->srtt;
538
539 /* If newly calculated rtt larger than stored one,
540 * store new one. Otherwise, use EWMA. Remember,
541 * rtt overestimation is always better than underestimation.
542 */
543 if (!(dst->mxlock&(1<<RTAX_RTT))) {
544 if (m <= 0)
545 dst->rtt = tp->srtt;
546 else
547 dst->rtt -= (m>>3);
548 }
549
550 if (!(dst->mxlock&(1<<RTAX_RTTVAR))) {
551 if (m < 0)
552 m = -m;
553
554 /* Scale deviation to rttvar fixed point */
555 m >>= 1;
556 if (m < tp->mdev)
557 m = tp->mdev;
558
559 if (m >= dst->rttvar)
560 dst->rttvar = m;
561 else
562 dst->rttvar -= (dst->rttvar - m)>>2;
563 }
564
565 if (tp->snd_ssthresh >= 0xFFFF) {
566 /* Slow start still did not finish. */
567 if (dst->ssthresh &&
568 !(dst->mxlock&(1<<RTAX_SSTHRESH)) &&
569 (tp->snd_cwnd>>1) > dst->ssthresh)
570 dst->ssthresh = (tp->snd_cwnd>>1);
571 if (!(dst->mxlock&(1<<RTAX_CWND)) &&
572 tp->snd_cwnd > dst->cwnd)
573 dst->cwnd = tp->snd_cwnd;
574 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
575 tp->ca_state == TCP_CA_Open) {
576 /* Cong. avoidance phase, cwnd is reliable. */
577 if (!(dst->mxlock&(1<<RTAX_SSTHRESH)))
578 dst->ssthresh = max_t(u32, tp->snd_cwnd>>1, tp->snd_ssthresh);
579 if (!(dst->mxlock&(1<<RTAX_CWND)))
580 dst->cwnd = (dst->cwnd + tp->snd_cwnd)>>1;
581 } else {
582 /* Else slow start did not finish, cwnd is non-sense,
583 ssthresh may be also invalid.
584 */
585 if (!(dst->mxlock&(1<<RTAX_CWND)))
586 dst->cwnd = (dst->cwnd + tp->snd_ssthresh)>>1;
587 if (dst->ssthresh &&
588 !(dst->mxlock&(1<<RTAX_SSTHRESH)) &&
589 tp->snd_ssthresh > dst->ssthresh)
590 dst->ssthresh = tp->snd_ssthresh;
591 }
592
593 if (!(dst->mxlock&(1<<RTAX_REORDERING))) {
594 if (dst->reordering < tp->reordering &&
595 tp->reordering != sysctl_tcp_reordering)
596 dst->reordering = tp->reordering;
597 }
598 }
599 }
600
601 /* Increase initial CWND conservatively: if estimated
602 * RTT is low enough (<20msec) or if we have some preset ssthresh.
603 *
604 * Numbers are taken from RFC2414.
605 */
606 __u32 tcp_init_cwnd(struct tcp_opt *tp)
607 {
608 __u32 cwnd;
609
610 if (tp->mss_cache > 1460)
611 return 2;
612
613 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
614
615 if (!tp->srtt || (tp->snd_ssthresh >= 0xFFFF && tp->srtt > ((HZ/50)<<3)))
616 cwnd = 2;
617 else if (cwnd > tp->snd_ssthresh)
618 cwnd = tp->snd_ssthresh;
619
620 return min_t(u32, cwnd, tp->snd_cwnd_clamp);
621 }
622
623 /* Initialize metrics on socket. */
624
625 static void tcp_init_metrics(struct sock *sk)
626 {
627 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
628 struct dst_entry *dst = __sk_dst_get(sk);
629
630 if (dst == NULL)
631 goto reset;
632
633 dst_confirm(dst);
634
635 if (dst->mxlock&(1<<RTAX_CWND))
636 tp->snd_cwnd_clamp = dst->cwnd;
637 if (dst->ssthresh) {
638 tp->snd_ssthresh = dst->ssthresh;
639 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
640 tp->snd_ssthresh = tp->snd_cwnd_clamp;
641 }
642 if (dst->reordering && tp->reordering != dst->reordering) {
643 tp->sack_ok &= ~2;
644 tp->reordering = dst->reordering;
645 }
646
647 if (dst->rtt == 0)
648 goto reset;
649
650 if (!tp->srtt && dst->rtt < (TCP_TIMEOUT_INIT<<3))
651 goto reset;
652
653 /* Initial rtt is determined from SYN,SYN-ACK.
654 * The segment is small and rtt may appear much
655 * less than real one. Use per-dst memory
656 * to make it more realistic.
657 *
658 * A bit of theory. RTT is time passed after "normal" sized packet
659 * is sent until it is ACKed. In normal curcumstances sending small
660 * packets force peer to delay ACKs and calculation is correct too.
661 * The algorithm is adaptive and, provided we follow specs, it
662 * NEVER underestimate RTT. BUT! If peer tries to make some clever
663 * tricks sort of "quick acks" for time long enough to decrease RTT
664 * to low value, and then abruptly stops to do it and starts to delay
665 * ACKs, wait for troubles.
666 */
667 if (dst->rtt > tp->srtt)
668 tp->srtt = dst->rtt;
669 if (dst->rttvar > tp->mdev) {
670 tp->mdev = dst->rttvar;
671 tp->mdev_max = tp->rttvar = max_t(u32, tp->mdev, TCP_RTO_MIN);
672 }
673 tcp_set_rto(tp);
674 tcp_bound_rto(tp);
675 if (tp->rto < TCP_TIMEOUT_INIT && !tp->saw_tstamp)
676 goto reset;
677 tp->snd_cwnd = tcp_init_cwnd(tp);
678 tp->snd_cwnd_stamp = tcp_time_stamp;
679 return;
680
681 reset:
682 /* Play conservative. If timestamps are not
683 * supported, TCP will fail to recalculate correct
684 * rtt, if initial rto is too small. FORGET ALL AND RESET!
685 */
686 if (!tp->saw_tstamp && tp->srtt) {
687 tp->srtt = 0;
688 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
689 tp->rto = TCP_TIMEOUT_INIT;
690 }
691 }
692
693 static void tcp_update_reordering(struct tcp_opt *tp, int metric, int ts)
694 {
695 if (metric > tp->reordering) {
696 tp->reordering = min_t(unsigned int, TCP_MAX_REORDERING, metric);
697
698 /* This exciting event is worth to be remembered. 8) */
699 if (ts)
700 NET_INC_STATS_BH(TCPTSReorder);
701 else if (IsReno(tp))
702 NET_INC_STATS_BH(TCPRenoReorder);
703 else if (IsFack(tp))
704 NET_INC_STATS_BH(TCPFACKReorder);
705 else
706 NET_INC_STATS_BH(TCPSACKReorder);
707 #if FASTRETRANS_DEBUG > 1
708 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
709 tp->sack_ok, tp->ca_state,
710 tp->reordering, tp->fackets_out, tp->sacked_out,
711 tp->undo_marker ? tp->undo_retrans : 0);
712 #endif
713 /* Disable FACK yet. */
714 tp->sack_ok &= ~2;
715 }
716 }
717
718 /* This procedure tags the retransmission queue when SACKs arrive.
719 *
720 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
721 * Packets in queue with these bits set are counted in variables
722 * sacked_out, retrans_out and lost_out, correspondingly.
723 *
724 * Valid combinations are:
725 * Tag InFlight Description
726 * 0 1 - orig segment is in flight.
727 * S 0 - nothing flies, orig reached receiver.
728 * L 0 - nothing flies, orig lost by net.
729 * R 2 - both orig and retransmit are in flight.
730 * L|R 1 - orig is lost, retransmit is in flight.
731 * S|R 1 - orig reached receiver, retrans is still in flight.
732 * (L|S|R is logically valid, it could occur when L|R is sacked,
733 * but it is equivalent to plain S and code short-curcuits it to S.
734 * L|S is logically invalid, it would mean -1 packet in flight 8))
735 *
736 * These 6 states form finite state machine, controlled by the following events:
737 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
738 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
739 * 3. Loss detection event of one of three flavors:
740 * A. Scoreboard estimator decided the packet is lost.
741 * A'. Reno "three dupacks" marks head of queue lost.
742 * A''. Its FACK modfication, head until snd.fack is lost.
743 * B. SACK arrives sacking data transmitted after never retransmitted
744 * hole was sent out.
745 * C. SACK arrives sacking SND.NXT at the moment, when the
746 * segment was retransmitted.
747 * 4. D-SACK added new rule: D-SACK changes any tag to S.
748 *
749 * It is pleasant to note, that state diagram turns out to be commutative,
750 * so that we are allowed not to be bothered by order of our actions,
751 * when multiple events arrive simultaneously. (see the function below).
752 *
753 * Reordering detection.
754 * --------------------
755 * Reordering metric is maximal distance, which a packet can be displaced
756 * in packet stream. With SACKs we can estimate it:
757 *
758 * 1. SACK fills old hole and the corresponding segment was not
759 * ever retransmitted -> reordering. Alas, we cannot use it
760 * when segment was retransmitted.
761 * 2. The last flaw is solved with D-SACK. D-SACK arrives
762 * for retransmitted and already SACKed segment -> reordering..
763 * Both of these heuristics are not used in Loss state, when we cannot
764 * account for retransmits accurately.
765 */
766 static int
767 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
768 {
769 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
770 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
771 struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
772 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
773 int reord = tp->packets_out;
774 int prior_fackets;
775 u32 lost_retrans = 0;
776 int flag = 0;
777 int i;
778
779 if (!tp->sacked_out)
780 tp->fackets_out = 0;
781 prior_fackets = tp->fackets_out;
782
783 for (i=0; i<num_sacks; i++, sp++) {
784 struct sk_buff *skb;
785 __u32 start_seq = ntohl(sp->start_seq);
786 __u32 end_seq = ntohl(sp->end_seq);
787 int fack_count = 0;
788 int dup_sack = 0;
789
790 /* Check for D-SACK. */
791 if (i == 0) {
792 u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
793
794 if (before(start_seq, ack)) {
795 dup_sack = 1;
796 tp->sack_ok |= 4;
797 NET_INC_STATS_BH(TCPDSACKRecv);
798 } else if (num_sacks > 1 &&
799 !after(end_seq, ntohl(sp[1].end_seq)) &&
800 !before(start_seq, ntohl(sp[1].start_seq))) {
801 dup_sack = 1;
802 tp->sack_ok |= 4;
803 NET_INC_STATS_BH(TCPDSACKOfoRecv);
804 }
805
806 /* D-SACK for already forgotten data...
807 * Do dumb counting. */
808 if (dup_sack &&
809 !after(end_seq, prior_snd_una) &&
810 after(end_seq, tp->undo_marker))
811 tp->undo_retrans--;
812
813 /* Eliminate too old ACKs, but take into
814 * account more or less fresh ones, they can
815 * contain valid SACK info.
816 */
817 if (before(ack, prior_snd_una-tp->max_window))
818 return 0;
819 }
820
821 /* Event "B" in the comment above. */
822 if (after(end_seq, tp->high_seq))
823 flag |= FLAG_DATA_LOST;
824
825 for_retrans_queue(skb, sk, tp) {
826 u8 sacked = TCP_SKB_CB(skb)->sacked;
827 int in_sack;
828
829 /* The retransmission queue is always in order, so
830 * we can short-circuit the walk early.
831 */
832 if(!before(TCP_SKB_CB(skb)->seq, end_seq))
833 break;
834
835 fack_count++;
836
837 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
838 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
839
840 /* Account D-SACK for retransmitted packet. */
841 if ((dup_sack && in_sack) &&
842 (sacked & TCPCB_RETRANS) &&
843 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
844 tp->undo_retrans--;
845
846 /* The frame is ACKed. */
847 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
848 if (sacked&TCPCB_RETRANS) {
849 if ((dup_sack && in_sack) &&
850 (sacked&TCPCB_SACKED_ACKED))
851 reord = min_t(int, fack_count, reord);
852 } else {
853 /* If it was in a hole, we detected reordering. */
854 if (fack_count < prior_fackets &&
855 !(sacked&TCPCB_SACKED_ACKED))
856 reord = min_t(int, fack_count, reord);
857 }
858
859 /* Nothing to do; acked frame is about to be dropped. */
860 continue;
861 }
862
863 if ((sacked&TCPCB_SACKED_RETRANS) &&
864 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
865 (!lost_retrans || after(end_seq, lost_retrans)))
866 lost_retrans = end_seq;
867
868 if (!in_sack)
869 continue;
870
871 if (!(sacked&TCPCB_SACKED_ACKED)) {
872 if (sacked & TCPCB_SACKED_RETRANS) {
873 /* If the segment is not tagged as lost,
874 * we do not clear RETRANS, believing
875 * that retransmission is still in flight.
876 */
877 if (sacked & TCPCB_LOST) {
878 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
879 tp->lost_out--;
880 tp->retrans_out--;
881 }
882 } else {
883 /* New sack for not retransmitted frame,
884 * which was in hole. It is reordering.
885 */
886 if (!(sacked & TCPCB_RETRANS) &&
887 fack_count < prior_fackets)
888 reord = min_t(int, fack_count, reord);
889
890 if (sacked & TCPCB_LOST) {
891 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
892 tp->lost_out--;
893 }
894 }
895
896 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
897 flag |= FLAG_DATA_SACKED;
898 tp->sacked_out++;
899
900 if (fack_count > tp->fackets_out)
901 tp->fackets_out = fack_count;
902 } else {
903 if (dup_sack && (sacked&TCPCB_RETRANS))
904 reord = min_t(int, fack_count, reord);
905 }
906
907 /* D-SACK. We can detect redundant retransmission
908 * in S|R and plain R frames and clear it.
909 * undo_retrans is decreased above, L|R frames
910 * are accounted above as well.
911 */
912 if (dup_sack &&
913 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
914 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
915 tp->retrans_out--;
916 }
917 }
918 }
919
920 /* Check for lost retransmit. This superb idea is
921 * borrowed from "ratehalving". Event "C".
922 * Later note: FACK people cheated me again 8),
923 * we have to account for reordering! Ugly,
924 * but should help.
925 */
926 if (lost_retrans && tp->ca_state == TCP_CA_Recovery) {
927 struct sk_buff *skb;
928
929 for_retrans_queue(skb, sk, tp) {
930 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
931 break;
932 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
933 continue;
934 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
935 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
936 (IsFack(tp) ||
937 !before(lost_retrans, TCP_SKB_CB(skb)->ack_seq+tp->reordering*tp->mss_cache))) {
938 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
939 tp->retrans_out--;
940
941 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
942 tp->lost_out++;
943 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
944 flag |= FLAG_DATA_SACKED;
945 NET_INC_STATS_BH(TCPLostRetransmit);
946 }
947 }
948 }
949 }
950
951 tp->left_out = tp->sacked_out + tp->lost_out;
952
953 if (reord < tp->fackets_out && tp->ca_state != TCP_CA_Loss)
954 tcp_update_reordering(tp, (tp->fackets_out+1)-reord, 0);
955
956 #if FASTRETRANS_DEBUG > 0
957 BUG_TRAP((int)tp->sacked_out >= 0);
958 BUG_TRAP((int)tp->lost_out >= 0);
959 BUG_TRAP((int)tp->retrans_out >= 0);
960 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
961 #endif
962 return flag;
963 }
964
965 void tcp_clear_retrans(struct tcp_opt *tp)
966 {
967 tp->left_out = 0;
968 tp->retrans_out = 0;
969
970 tp->fackets_out = 0;
971 tp->sacked_out = 0;
972 tp->lost_out = 0;
973
974 tp->undo_marker = 0;
975 tp->undo_retrans = 0;
976 }
977
978 /* Enter Loss state. If "how" is not zero, forget all SACK information
979 * and reset tags completely, otherwise preserve SACKs. If receiver
980 * dropped its ofo queue, we will know this due to reneging detection.
981 */
982 void tcp_enter_loss(struct sock *sk, int how)
983 {
984 struct tcp_opt *tp = &sk->tp_pinfo.af_tcp;
985 struct sk_buff *skb;
986 int cnt = 0;
987
988 /* Reduce ssthresh if it has not yet been made inside this window. */
989 if (tp->ca_state <= TCP_CA_Disorder ||
990 tp->snd_una == tp->high_seq ||
991 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
992 tp->prior_ssthresh = tcp_current_ssthresh(tp);
993 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
994 }
995 tp->snd_cwnd = 1;
996 tp->snd_cwnd_cnt = 0;
997 tp->snd_cwnd_stamp = tcp_time_stamp;
998
999 tcp_clear_retrans(tp);
1000
1001 /* Push undo marker, if it was plain RTO and nothing
1002 * was retransmitted. */
1003 if (!how)
1004 tp->undo_marker = tp->snd_una;
1005
1006 for_retrans_queue(skb, sk, tp) {
1007 cnt++;
1008 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1009 tp->undo_marker = 0;
1010 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1011 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1012 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1013 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1014 tp->lost_out++;
1015 } else {
1016 tp->sacked_out++;
1017 tp->fackets_out = cnt;
1018 }
1019 }
1020 tcp_sync_left_out(tp);
1021
1022 tp->reordering = min_t(unsigned int, tp->reordering, sysctl_tcp_reordering);
1023 tp->ca_state = TCP_CA_Loss;
1024 tp->high_seq = tp->snd_nxt;
1025 TCP_ECN_queue_cwr(tp);
1026 }
1027
1028 static int tcp_check_sack_reneging(struct sock *sk, struct tcp_opt *tp)
1029 {
1030 struct sk_buff *skb;
1031
1032 /* If ACK arrived pointing to a remembered SACK,
1033 * it means that our remembered SACKs do not reflect
1034 * real state of receiver i.e.
1035 * receiver _host_ is heavily congested (or buggy).
1036 * Do processing similar to RTO timeout.
1037 */
1038 if ((skb = skb_peek(&sk->write_queue)) != NULL &&
1039 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1040 NET_INC_STATS_BH(TCPSACKReneging);
1041
1042 tcp_enter_loss(sk, 1);
1043 tp->retransmits++;
1044 tcp_retransmit_skb(sk, skb_peek(&sk->write_queue));
1045 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1046 return 1;
1047 }
1048 return 0;
1049 }
1050
1051 static inline int tcp_fackets_out(struct tcp_opt *tp)
1052 {
1053 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1054 }
1055
1056 static inline int tcp_skb_timedout(struct tcp_opt *tp, struct sk_buff *skb)
1057 {
1058 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto);
1059 }
1060
1061 static inline int tcp_head_timedout(struct sock *sk, struct tcp_opt *tp)
1062 {
1063 return tp->packets_out && tcp_skb_timedout(tp, skb_peek(&sk->write_queue));
1064 }
1065
1066 /* Linux NewReno/SACK/FACK/ECN state machine.
1067 * --------------------------------------
1068 *
1069 * "Open" Normal state, no dubious events, fast path.
1070 * "Disorder" In all the respects it is "Open",
1071 * but requires a bit more attention. It is entered when
1072 * we see some SACKs or dupacks. It is split of "Open"
1073 * mainly to move some processing from fast path to slow one.
1074 * "CWR" CWND was reduced due to some Congestion Notification event.
1075 * It can be ECN, ICMP source quench, local device congestion.
1076 * "Recovery" CWND was reduced, we are fast-retransmitting.
1077 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1078 *
1079 * tcp_fastretrans_alert() is entered:
1080 * - each incoming ACK, if state is not "Open"
1081 * - when arrived ACK is unusual, namely:
1082 * * SACK
1083 * * Duplicate ACK.
1084 * * ECN ECE.
1085 *
1086 * Counting packets in flight is pretty simple.
1087 *
1088 * in_flight = packets_out - left_out + retrans_out
1089 *
1090 * packets_out is SND.NXT-SND.UNA counted in packets.
1091 *
1092 * retrans_out is number of retransmitted segments.
1093 *
1094 * left_out is number of segments left network, but not ACKed yet.
1095 *
1096 * left_out = sacked_out + lost_out
1097 *
1098 * sacked_out: Packets, which arrived to receiver out of order
1099 * and hence not ACKed. With SACKs this number is simply
1100 * amount of SACKed data. Even without SACKs
1101 * it is easy to give pretty reliable estimate of this number,
1102 * counting duplicate ACKs.
1103 *
1104 * lost_out: Packets lost by network. TCP has no explicit
1105 * "loss notification" feedback from network (for now).
1106 * It means that this number can be only _guessed_.
1107 * Actually, it is the heuristics to predict lossage that
1108 * distinguishes different algorithms.
1109 *
1110 * F.e. after RTO, when all the queue is considered as lost,
1111 * lost_out = packets_out and in_flight = retrans_out.
1112 *
1113 * Essentially, we have now two algorithms counting
1114 * lost packets.
1115 *
1116 * FACK: It is the simplest heuristics. As soon as we decided
1117 * that something is lost, we decide that _all_ not SACKed
1118 * packets until the most forward SACK are lost. I.e.
1119 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1120 * It is absolutely correct estimate, if network does not reorder
1121 * packets. And it loses any connection to reality when reordering
1122 * takes place. We use FACK by default until reordering
1123 * is suspected on the path to this destination.
1124 *
1125 * NewReno: when Recovery is entered, we assume that one segment
1126 * is lost (classic Reno). While we are in Recovery and
1127 * a partial ACK arrives, we assume that one more packet
1128 * is lost (NewReno). This heuristics are the same in NewReno
1129 * and SACK.
1130 *
1131 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1132 * deflation etc. CWND is real congestion window, never inflated, changes
1133 * only according to classic VJ rules.
1134 *
1135 * Really tricky (and requiring careful tuning) part of algorithm
1136 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1137 * The first determines the moment _when_ we should reduce CWND and,
1138 * hence, slow down forward transmission. In fact, it determines the moment
1139 * when we decide that hole is caused by loss, rather than by a reorder.
1140 *
1141 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1142 * holes, caused by lost packets.
1143 *
1144 * And the most logically complicated part of algorithm is undo
1145 * heuristics. We detect false retransmits due to both too early
1146 * fast retransmit (reordering) and underestimated RTO, analyzing
1147 * timestamps and D-SACKs. When we detect that some segments were
1148 * retransmitted by mistake and CWND reduction was wrong, we undo
1149 * window reduction and abort recovery phase. This logic is hidden
1150 * inside several functions named tcp_try_undo_<something>.
1151 */
1152
1153 /* This function decides, when we should leave Disordered state
1154 * and enter Recovery phase, reducing congestion window.
1155 *
1156 * Main question: may we further continue forward transmission
1157 * with the same cwnd?
1158 */
1159 static int
1160 tcp_time_to_recover(struct sock *sk, struct tcp_opt *tp)
1161 {
1162 /* Trick#1: The loss is proven. */
1163 if (tp->lost_out)
1164 return 1;
1165
1166 /* Not-A-Trick#2 : Classic rule... */
1167 if (tcp_fackets_out(tp) > tp->reordering)
1168 return 1;
1169
1170 /* Trick#3 : when we use RFC2988 timer restart, fast
1171 * retransmit can be triggered by timeout of queue head.
1172 */
1173 if (tcp_head_timedout(sk, tp))
1174 return 1;
1175
1176 /* Trick#4: It is still not OK... But will it be useful to delay
1177 * recovery more?
1178 */
1179 if (tp->packets_out <= tp->reordering &&
1180 tp->sacked_out >= max_t(u32, tp->packets_out/2, sysctl_tcp_reordering) &&
1181 !tcp_may_send_now(sk, tp)) {
1182 /* We have nothing to send. This connection is limited
1183 * either by receiver window or by application.
1184 */
1185 return 1;
1186 }
1187
1188 return 0;
1189 }
1190
1191 /* If we receive more dupacks than we expected counting segments
1192 * in assumption of absent reordering, interpret this as reordering.
1193 * The only another reason could be bug in receiver TCP.
1194 */
1195 static void tcp_check_reno_reordering(struct tcp_opt *tp, int addend)
1196 {
1197 u32 holes = min_t(unsigned int,
1198 max_t(unsigned int, tp->lost_out, 1),
1199 tp->packets_out);
1200
1201 if (tp->sacked_out + holes > tp->packets_out) {
1202 tp->sacked_out = tp->packets_out - holes;
1203 tcp_update_reordering(tp, tp->packets_out+addend, 0);
1204 }
1205 }
1206
1207 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1208
1209 static void tcp_add_reno_sack(struct tcp_opt *tp)
1210 {
1211 ++tp->sacked_out;
1212 tcp_check_reno_reordering(tp, 0);
1213 tcp_sync_left_out(tp);
1214 }
1215
1216 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1217
1218 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_opt *tp, int acked)
1219 {
1220 if (acked > 0) {
1221 /* One ACK acked hole. The rest eat duplicate ACKs. */
1222 if (acked-1 >= tp->sacked_out)
1223 tp->sacked_out = 0;
1224 else
1225 tp->sacked_out -= acked-1;
1226 }
1227 tcp_check_reno_reordering(tp, acked);
1228 tcp_sync_left_out(tp);
1229 }
1230
1231 static inline void tcp_reset_reno_sack(struct tcp_opt *tp)
1232 {
1233 tp->sacked_out = 0;
1234 tp->left_out = tp->lost_out;
1235 }
1236
1237 /* Mark head of queue up as lost. */
1238 static void
1239 tcp_mark_head_lost(struct sock *sk, struct tcp_opt *tp, int packets, u32 high_seq)
1240 {
1241 struct sk_buff *skb;
1242 int cnt = packets;
1243
1244 BUG_TRAP(cnt <= tp->packets_out);
1245
1246 for_retrans_queue(skb, sk, tp) {
1247 if (--cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1248 break;
1249 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1250 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1251 tp->lost_out++;
1252 }
1253 }
1254 tcp_sync_left_out(tp);
1255 }
1256
1257 /* Account newly detected lost packet(s) */
1258
1259 static void tcp_update_scoreboard(struct sock *sk, struct tcp_opt *tp)
1260 {
1261 if (IsFack(tp)) {
1262 int lost = tp->fackets_out - tp->reordering;
1263 if (lost <= 0)
1264 lost = 1;
1265 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1266 } else {
1267 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1268 }
1269
1270 /* New heuristics: it is possible only after we switched
1271 * to restart timer each time when something is ACKed.
1272 * Hence, we can detect timed out packets during fast
1273 * retransmit without falling to slow start.
1274 */
1275 if (tcp_head_timedout(sk, tp)) {
1276 struct sk_buff *skb;
1277
1278 for_retrans_queue(skb, sk, tp) {
1279 if (tcp_skb_timedout(tp, skb) &&
1280 !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1281 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1282 tp->lost_out++;
1283 }
1284 }
1285 tcp_sync_left_out(tp);
1286 }
1287 }
1288
1289 /* CWND moderation, preventing bursts due to too big ACKs
1290 * in dubious situations.
1291 */
1292 static __inline__ void tcp_moderate_cwnd(struct tcp_opt *tp)
1293 {
1294 tp->snd_cwnd = min_t(u32, tp->snd_cwnd,
1295 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1296 tp->snd_cwnd_stamp = tcp_time_stamp;
1297 }
1298
1299 /* Decrease cwnd each second ack. */
1300
1301 static void tcp_cwnd_down(struct tcp_opt *tp)
1302 {
1303 int decr = tp->snd_cwnd_cnt + 1;
1304
1305 tp->snd_cwnd_cnt = decr&1;
1306 decr >>= 1;
1307
1308 if (decr && tp->snd_cwnd > tp->snd_ssthresh/2)
1309 tp->snd_cwnd -= decr;
1310
1311 tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1312 tp->snd_cwnd_stamp = tcp_time_stamp;
1313 }
1314
1315 /* Nothing was retransmitted or returned timestamp is less
1316 * than timestamp of the first retransmission.
1317 */
1318 static __inline__ int tcp_packet_delayed(struct tcp_opt *tp)
1319 {
1320 return !tp->retrans_stamp ||
1321 (tp->saw_tstamp && tp->rcv_tsecr &&
1322 (__s32)(tp->rcv_tsecr - tp->retrans_stamp) < 0);
1323 }
1324
1325 /* Undo procedures. */
1326
1327 #if FASTRETRANS_DEBUG > 1
1328 static void DBGUNDO(struct sock *sk, struct tcp_opt *tp, const char *msg)
1329 {
1330 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1331 msg,
1332 NIPQUAD(sk->daddr), ntohs(sk->dport),
1333 tp->snd_cwnd, tp->left_out,
1334 tp->snd_ssthresh, tp->prior_ssthresh, tp->packets_out);
1335 }
1336 #else
1337 #define DBGUNDO(x...) do { } while (0)
1338 #endif
1339
1340 static void tcp_undo_cwr(struct tcp_opt *tp, int undo)
1341 {
1342 if (tp->prior_ssthresh) {
1343 tp->snd_cwnd = max_t(unsigned int,
1344 tp->snd_cwnd, tp->snd_ssthresh<<1);
1345
1346 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1347 tp->snd_ssthresh = tp->prior_ssthresh;
1348 TCP_ECN_withdraw_cwr(tp);
1349 }
1350 } else {
1351 tp->snd_cwnd = max_t(unsigned int, tp->snd_cwnd, tp->snd_ssthresh);
1352 }
1353 tcp_moderate_cwnd(tp);
1354 tp->snd_cwnd_stamp = tcp_time_stamp;
1355 }
1356
1357 static inline int tcp_may_undo(struct tcp_opt *tp)
1358 {
1359 return tp->undo_marker &&
1360 (!tp->undo_retrans || tcp_packet_delayed(tp));
1361 }
1362
1363 /* People celebrate: "We love our President!" */
1364 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_opt *tp)
1365 {
1366 if (tcp_may_undo(tp)) {
1367 /* Happy end! We did not retransmit anything
1368 * or our original transmission succeeded.
1369 */
1370 DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans");
1371 tcp_undo_cwr(tp, 1);
1372 if (tp->ca_state == TCP_CA_Loss)
1373 NET_INC_STATS_BH(TCPLossUndo);
1374 else
1375 NET_INC_STATS_BH(TCPFullUndo);
1376 tp->undo_marker = 0;
1377 }
1378 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1379 /* Hold old state until something *above* high_seq
1380 * is ACKed. For Reno it is MUST to prevent false
1381 * fast retransmits (RFC2582). SACK TCP is safe. */
1382 tcp_moderate_cwnd(tp);
1383 return 1;
1384 }
1385 tp->ca_state = TCP_CA_Open;
1386 return 0;
1387 }
1388
1389 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1390 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_opt *tp)
1391 {
1392 if (tp->undo_marker && !tp->undo_retrans) {
1393 DBGUNDO(sk, tp, "D-SACK");
1394 tcp_undo_cwr(tp, 1);
1395 tp->undo_marker = 0;
1396 NET_INC_STATS_BH(TCPDSACKUndo);
1397 }
1398 }
1399
1400 /* Undo during fast recovery after partial ACK. */
1401
1402 static int tcp_try_undo_partial(struct sock *sk, struct tcp_opt *tp, int acked)
1403 {
1404 /* Partial ACK arrived. Force Hoe's retransmit. */
1405 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1406
1407 if (tcp_may_undo(tp)) {
1408 /* Plain luck! Hole if filled with delayed
1409 * packet, rather than with a retransmit.
1410 */
1411 if (tp->retrans_out == 0)
1412 tp->retrans_stamp = 0;
1413
1414 tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1);
1415
1416 DBGUNDO(sk, tp, "Hoe");
1417 tcp_undo_cwr(tp, 0);
1418 NET_INC_STATS_BH(TCPPartialUndo);
1419
1420 /* So... Do not make Hoe's retransmit yet.
1421 * If the first packet was delayed, the rest
1422 * ones are most probably delayed as well.
1423 */
1424 failed = 0;
1425 }
1426 return failed;
1427 }
1428
1429 /* Undo during loss recovery after partial ACK. */
1430 static int tcp_try_undo_loss(struct sock *sk, struct tcp_opt *tp)
1431 {
1432 if (tcp_may_undo(tp)) {
1433 struct sk_buff *skb;
1434 for_retrans_queue(skb, sk, tp) {
1435 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1436 }
1437 DBGUNDO(sk, tp, "partial loss");
1438 tp->lost_out = 0;
1439 tp->left_out = tp->sacked_out;
1440 tcp_undo_cwr(tp, 1);
1441 NET_INC_STATS_BH(TCPLossUndo);
1442 tp->retransmits = 0;
1443 tp->undo_marker = 0;
1444 if (!IsReno(tp))
1445 tp->ca_state = TCP_CA_Open;
1446 return 1;
1447 }
1448 return 0;
1449 }
1450
1451 static __inline__ void tcp_complete_cwr(struct tcp_opt *tp)
1452 {
1453 tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_ssthresh);
1454 tp->snd_cwnd_stamp = tcp_time_stamp;
1455 }
1456
1457 static void tcp_try_to_open(struct sock *sk, struct tcp_opt *tp, int flag)
1458 {
1459 tp->left_out = tp->sacked_out;
1460
1461 if (tp->retrans_out == 0)
1462 tp->retrans_stamp = 0;
1463
1464 if (flag&FLAG_ECE)
1465 tcp_enter_cwr(tp);
1466
1467 if (tp->ca_state != TCP_CA_CWR) {
1468 int state = TCP_CA_Open;
1469
1470 if (tp->left_out ||
1471 tp->retrans_out ||
1472 tp->undo_marker)
1473 state = TCP_CA_Disorder;
1474
1475 if (tp->ca_state != state) {
1476 tp->ca_state = state;
1477 tp->high_seq = tp->snd_nxt;
1478 }
1479 tcp_moderate_cwnd(tp);
1480 } else {
1481 tcp_cwnd_down(tp);
1482 }
1483 }
1484
1485 /* Process an event, which can update packets-in-flight not trivially.
1486 * Main goal of this function is to calculate new estimate for left_out,
1487 * taking into account both packets sitting in receiver's buffer and
1488 * packets lost by network.
1489 *
1490 * Besides that it does CWND reduction, when packet loss is detected
1491 * and changes state of machine.
1492 *
1493 * It does _not_ decide what to send, it is made in function
1494 * tcp_xmit_retransmit_queue().
1495 */
1496 static void
1497 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1498 int prior_packets, int flag)
1499 {
1500 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
1501 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1502
1503 /* Some technical things:
1504 * 1. Reno does not count dupacks (sacked_out) automatically. */
1505 if (!tp->packets_out)
1506 tp->sacked_out = 0;
1507 /* 2. SACK counts snd_fack in packets inaccurately. */
1508 if (tp->sacked_out == 0)
1509 tp->fackets_out = 0;
1510
1511 /* Now state machine starts.
1512 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1513 if (flag&FLAG_ECE)
1514 tp->prior_ssthresh = 0;
1515
1516 /* B. In all the states check for reneging SACKs. */
1517 if (tp->sacked_out && tcp_check_sack_reneging(sk, tp))
1518 return;
1519
1520 /* C. Process data loss notification, provided it is valid. */
1521 if ((flag&FLAG_DATA_LOST) &&
1522 before(tp->snd_una, tp->high_seq) &&
1523 tp->ca_state != TCP_CA_Open &&
1524 tp->fackets_out > tp->reordering) {
1525 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1526 NET_INC_STATS_BH(TCPLoss);
1527 }
1528
1529 /* D. Synchronize left_out to current state. */
1530 tcp_sync_left_out(tp);
1531
1532 /* E. Check state exit conditions. State can be terminated
1533 * when high_seq is ACKed. */
1534 if (tp->ca_state == TCP_CA_Open) {
1535 BUG_TRAP(tp->retrans_out == 0);
1536 tp->retrans_stamp = 0;
1537 } else if (!before(tp->snd_una, tp->high_seq)) {
1538 switch (tp->ca_state) {
1539 case TCP_CA_Loss:
1540 tp->retransmits = 0;
1541 if (tcp_try_undo_recovery(sk, tp))
1542 return;
1543 break;
1544
1545 case TCP_CA_CWR:
1546 /* CWR is to be held something *above* high_seq
1547 * is ACKed for CWR bit to reach receiver. */
1548 if (tp->snd_una != tp->high_seq) {
1549 tcp_complete_cwr(tp);
1550 tp->ca_state = TCP_CA_Open;
1551 }
1552 break;
1553
1554 case TCP_CA_Disorder:
1555 tcp_try_undo_dsack(sk, tp);
1556 if (!tp->undo_marker ||
1557 /* For SACK case do not Open to allow to undo
1558 * catching for all duplicate ACKs. */
1559 IsReno(tp) || tp->snd_una != tp->high_seq) {
1560 tp->undo_marker = 0;
1561 tp->ca_state = TCP_CA_Open;
1562 }
1563 break;
1564
1565 case TCP_CA_Recovery:
1566 if (IsReno(tp))
1567 tcp_reset_reno_sack(tp);
1568 if (tcp_try_undo_recovery(sk, tp))
1569 return;
1570 tcp_complete_cwr(tp);
1571 break;
1572 }
1573 }
1574
1575 /* F. Process state. */
1576 switch (tp->ca_state) {
1577 case TCP_CA_Recovery:
1578 if (prior_snd_una == tp->snd_una) {
1579 if (IsReno(tp) && is_dupack)
1580 tcp_add_reno_sack(tp);
1581 } else {
1582 int acked = prior_packets - tp->packets_out;
1583 if (IsReno(tp))
1584 tcp_remove_reno_sacks(sk, tp, acked);
1585 is_dupack = tcp_try_undo_partial(sk, tp, acked);
1586 }
1587 break;
1588 case TCP_CA_Loss:
1589 if (flag&FLAG_DATA_ACKED)
1590 tp->retransmits = 0;
1591 if (!tcp_try_undo_loss(sk, tp)) {
1592 tcp_moderate_cwnd(tp);
1593 tcp_xmit_retransmit_queue(sk);
1594 return;
1595 }
1596 if (tp->ca_state != TCP_CA_Open)
1597 return;
1598 /* Loss is undone; fall through to processing in Open state. */
1599 default:
1600 if (IsReno(tp)) {
1601 if (tp->snd_una != prior_snd_una)
1602 tcp_reset_reno_sack(tp);
1603 if (is_dupack)
1604 tcp_add_reno_sack(tp);
1605 }
1606
1607 if (tp->ca_state == TCP_CA_Disorder)
1608 tcp_try_undo_dsack(sk, tp);
1609
1610 if (!tcp_time_to_recover(sk, tp)) {
1611 tcp_try_to_open(sk, tp, flag);
1612 return;
1613 }
1614
1615 /* Otherwise enter Recovery state */
1616
1617 if (IsReno(tp))
1618 NET_INC_STATS_BH(TCPRenoRecovery);
1619 else
1620 NET_INC_STATS_BH(TCPSackRecovery);
1621
1622 tp->high_seq = tp->snd_nxt;
1623 tp->prior_ssthresh = 0;
1624 tp->undo_marker = tp->snd_una;
1625 tp->undo_retrans = tp->retrans_out;
1626
1627 if (tp->ca_state < TCP_CA_CWR) {
1628 if (!(flag&FLAG_ECE))
1629 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1630 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1631 TCP_ECN_queue_cwr(tp);
1632 }
1633
1634 tp->snd_cwnd_cnt = 0;
1635 tp->ca_state = TCP_CA_Recovery;
1636 }
1637
1638 if (is_dupack || tcp_head_timedout(sk, tp))
1639 tcp_update_scoreboard(sk, tp);
1640 tcp_cwnd_down(tp);
1641 tcp_xmit_retransmit_queue(sk);
1642 }
1643
1644 /* Read draft-ietf-tcplw-high-performance before mucking
1645 * with this code. (Superceeds RFC1323)
1646 */
1647 static void tcp_ack_saw_tstamp(struct tcp_opt *tp, int flag)
1648 {
1649 __u32 seq_rtt;
1650
1651 /* RTTM Rule: A TSecr value received in a segment is used to
1652 * update the averaged RTT measurement only if the segment
1653 * acknowledges some new data, i.e., only if it advances the
1654 * left edge of the send window.
1655 *
1656 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1657 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1658 *
1659 * Changed: reset backoff as soon as we see the first valid sample.
1660 * If we do not, we get strongly overstimated rto. With timestamps
1661 * samples are accepted even from very old segments: f.e., when rtt=1
1662 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1663 * answer arrives rto becomes 120 seconds! If at least one of segments
1664 * in window is lost... Voila. --ANK (010210)
1665 */
1666 seq_rtt = tcp_time_stamp - tp->rcv_tsecr;
1667 tcp_rtt_estimator(tp, seq_rtt);
1668 tcp_set_rto(tp);
1669 tp->backoff = 0;
1670 tcp_bound_rto(tp);
1671 }
1672
1673 static void tcp_ack_no_tstamp(struct tcp_opt *tp, u32 seq_rtt, int flag)
1674 {
1675 /* We don't have a timestamp. Can only use
1676 * packets that are not retransmitted to determine
1677 * rtt estimates. Also, we must not reset the
1678 * backoff for rto until we get a non-retransmitted
1679 * packet. This allows us to deal with a situation
1680 * where the network delay has increased suddenly.
1681 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1682 */
1683
1684 if (flag & FLAG_RETRANS_DATA_ACKED)
1685 return;
1686
1687 tcp_rtt_estimator(tp, seq_rtt);
1688 tcp_set_rto(tp);
1689 tp->backoff = 0;
1690 tcp_bound_rto(tp);
1691 }
1692
1693 static __inline__ void
1694 tcp_ack_update_rtt(struct tcp_opt *tp, int flag, s32 seq_rtt)
1695 {
1696 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1697 if (tp->saw_tstamp && tp->rcv_tsecr)
1698 tcp_ack_saw_tstamp(tp, flag);
1699 else if (seq_rtt >= 0)
1700 tcp_ack_no_tstamp(tp, seq_rtt, flag);
1701 }
1702
1703 /* This is Jacobson's slow start and congestion avoidance.
1704 * SIGCOMM '88, p. 328.
1705 */
1706 static __inline__ void tcp_cong_avoid(struct tcp_opt *tp)
1707 {
1708 if (tp->snd_cwnd <= tp->snd_ssthresh) {
1709 /* In "safe" area, increase. */
1710 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
1711 tp->snd_cwnd++;
1712 } else {
1713 /* In dangerous area, increase slowly.
1714 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
1715 */
1716 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
1717 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
1718 tp->snd_cwnd++;
1719 tp->snd_cwnd_cnt=0;
1720 } else
1721 tp->snd_cwnd_cnt++;
1722 }
1723 tp->snd_cwnd_stamp = tcp_time_stamp;
1724 }
1725
1726 /* Restart timer after forward progress on connection.
1727 * RFC2988 recommends to restart timer to now+rto.
1728 */
1729
1730 static __inline__ void tcp_ack_packets_out(struct sock *sk, struct tcp_opt *tp)
1731 {
1732 if (tp->packets_out==0) {
1733 tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS);
1734 } else {
1735 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1736 }
1737 }
1738
1739 /* Remove acknowledged frames from the retransmission queue. */
1740 static int tcp_clean_rtx_queue(struct sock *sk)
1741 {
1742 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
1743 struct sk_buff *skb;
1744 __u32 now = tcp_time_stamp;
1745 int acked = 0;
1746 __s32 seq_rtt = -1;
1747
1748 while((skb=skb_peek(&sk->write_queue)) && (skb != tp->send_head)) {
1749 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
1750 __u8 sacked = scb->sacked;
1751
1752 /* If our packet is before the ack sequence we can
1753 * discard it as it's confirmed to have arrived at
1754 * the other end.
1755 */
1756 if (after(scb->end_seq, tp->snd_una))
1757 break;
1758
1759 /* Initial outgoing SYN's get put onto the write_queue
1760 * just like anything else we transmit. It is not
1761 * true data, and if we misinform our callers that
1762 * this ACK acks real data, we will erroneously exit
1763 * connection startup slow start one packet too
1764 * quickly. This is severely frowned upon behavior.
1765 */
1766 if(!(scb->flags & TCPCB_FLAG_SYN)) {
1767 acked |= FLAG_DATA_ACKED;
1768 } else {
1769 acked |= FLAG_SYN_ACKED;
1770 }
1771
1772 if (sacked) {
1773 if(sacked & TCPCB_RETRANS) {
1774 if(sacked & TCPCB_SACKED_RETRANS)
1775 tp->retrans_out--;
1776 acked |= FLAG_RETRANS_DATA_ACKED;
1777 seq_rtt = -1;
1778 } else if (seq_rtt < 0)
1779 seq_rtt = now - scb->when;
1780 if(sacked & TCPCB_SACKED_ACKED)
1781 tp->sacked_out--;
1782 if(sacked & TCPCB_LOST)
1783 tp->lost_out--;
1784 if(sacked & TCPCB_URG) {
1785 if (tp->urg_mode &&
1786 !before(scb->end_seq, tp->snd_up))
1787 tp->urg_mode = 0;
1788 }
1789 } else if (seq_rtt < 0)
1790 seq_rtt = now - scb->when;
1791 if(tp->fackets_out)
1792 tp->fackets_out--;
1793 tp->packets_out--;
1794 __skb_unlink(skb, skb->list);
1795 tcp_free_skb(sk, skb);
1796 }
1797
1798 if (acked&FLAG_ACKED) {
1799 tcp_ack_update_rtt(tp, acked, seq_rtt);
1800 tcp_ack_packets_out(sk, tp);
1801 }
1802
1803 #if FASTRETRANS_DEBUG > 0
1804 BUG_TRAP((int)tp->sacked_out >= 0);
1805 BUG_TRAP((int)tp->lost_out >= 0);
1806 BUG_TRAP((int)tp->retrans_out >= 0);
1807 if (tp->packets_out==0 && tp->sack_ok) {
1808 if (tp->lost_out) {
1809 printk(KERN_DEBUG "Leak l=%u %d\n", tp->lost_out, tp->ca_state);
1810 tp->lost_out = 0;
1811 }
1812 if (tp->sacked_out) {
1813 printk(KERN_DEBUG "Leak s=%u %d\n", tp->sacked_out, tp->ca_state);
1814 tp->sacked_out = 0;
1815 }
1816 if (tp->retrans_out) {
1817 printk(KERN_DEBUG "Leak r=%u %d\n", tp->retrans_out, tp->ca_state);
1818 tp->retrans_out = 0;
1819 }
1820 }
1821 #endif
1822 return acked;
1823 }
1824
1825 static void tcp_ack_probe(struct sock *sk)
1826 {
1827 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
1828
1829 /* Was it a usable window open? */
1830
1831 if (!after(TCP_SKB_CB(tp->send_head)->end_seq, tp->snd_una + tp->snd_wnd)) {
1832 tp->backoff = 0;
1833 tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0);
1834 /* Socket must be waked up by subsequent tcp_data_snd_check().
1835 * This function is not for random using!
1836 */
1837 } else {
1838 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0,
1839 min_t(u32, tp->rto << tp->backoff, TCP_RTO_MAX));
1840 }
1841 }
1842
1843 static __inline__ int tcp_ack_is_dubious(struct tcp_opt *tp, int flag)
1844 {
1845 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
1846 tp->ca_state != TCP_CA_Open);
1847 }
1848
1849 static __inline__ int tcp_may_raise_cwnd(struct tcp_opt *tp, int flag)
1850 {
1851 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
1852 !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR));
1853 }
1854
1855 /* Check that window update is acceptable.
1856 * The function assumes that snd_una<=ack<=snd_next.
1857 */
1858 static __inline__ int
1859 tcp_may_update_window(struct tcp_opt *tp, u32 ack, u32 ack_seq, u32 nwin)
1860 {
1861 return (after(ack, tp->snd_una) ||
1862 after(ack_seq, tp->snd_wl1) ||
1863 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
1864 }
1865
1866 /* Update our send window.
1867 *
1868 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
1869 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
1870 */
1871 static int tcp_ack_update_window(struct sock *sk, struct tcp_opt *tp,
1872 struct sk_buff *skb, u32 ack, u32 ack_seq)
1873 {
1874 int flag = 0;
1875 u32 nwin = ntohs(skb->h.th->window) << tp->snd_wscale;
1876
1877 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
1878 flag |= FLAG_WIN_UPDATE;
1879 tcp_update_wl(tp, ack, ack_seq);
1880
1881 if (tp->snd_wnd != nwin) {
1882 tp->snd_wnd = nwin;
1883
1884 /* Note, it is the only place, where
1885 * fast path is recovered for sending TCP.
1886 */
1887 tcp_fast_path_check(sk, tp);
1888
1889 if (nwin > tp->max_window) {
1890 tp->max_window = nwin;
1891 tcp_sync_mss(sk, tp->pmtu_cookie);
1892 }
1893 }
1894 }
1895
1896 tp->snd_una = ack;
1897
1898 return flag;
1899 }
1900
1901 /* This routine deals with incoming acks, but not outgoing ones. */
1902 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
1903 {
1904 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
1905 u32 prior_snd_una = tp->snd_una;
1906 u32 ack_seq = TCP_SKB_CB(skb)->seq;
1907 u32 ack = TCP_SKB_CB(skb)->ack_seq;
1908 u32 prior_in_flight;
1909 int prior_packets;
1910
1911 /* If the ack is newer than sent or older than previous acks
1912 * then we can probably ignore it.
1913 */
1914 if (after(ack, tp->snd_nxt))
1915 goto uninteresting_ack;
1916
1917 if (before(ack, prior_snd_una))
1918 goto old_ack;
1919
1920 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
1921 /* Window is constant, pure forward advance.
1922 * No more checks are required.
1923 * Note, we use the fact that SND.UNA>=SND.WL2.
1924 */
1925 tcp_update_wl(tp, ack, ack_seq);
1926 tp->snd_una = ack;
1927 flag |= FLAG_WIN_UPDATE;
1928
1929 NET_INC_STATS_BH(TCPHPAcks);
1930 } else {
1931 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
1932 flag |= FLAG_DATA;
1933 else
1934 NET_INC_STATS_BH(TCPPureAcks);
1935
1936 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
1937
1938 if (TCP_SKB_CB(skb)->sacked)
1939 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
1940
1941 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
1942 flag |= FLAG_ECE;
1943 }
1944
1945 /* We passed data and got it acked, remove any soft error
1946 * log. Something worked...
1947 */
1948 sk->err_soft = 0;
1949 tp->rcv_tstamp = tcp_time_stamp;
1950 if ((prior_packets = tp->packets_out) == 0)
1951 goto no_queue;
1952
1953 prior_in_flight = tcp_packets_in_flight(tp);
1954
1955 /* See if we can take anything off of the retransmit queue. */
1956 flag |= tcp_clean_rtx_queue(sk);
1957
1958 if (tcp_ack_is_dubious(tp, flag)) {
1959 /* Advanve CWND, if state allows this. */
1960 if ((flag&FLAG_DATA_ACKED) && prior_in_flight >= tp->snd_cwnd &&
1961 tcp_may_raise_cwnd(tp, flag))
1962 tcp_cong_avoid(tp);
1963 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
1964 } else {
1965 if ((flag&FLAG_DATA_ACKED) && prior_in_flight >= tp->snd_cwnd)
1966 tcp_cong_avoid(tp);
1967 }
1968
1969 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
1970 dst_confirm(sk->dst_cache);
1971
1972 return 1;
1973
1974 no_queue:
1975 tp->probes_out = 0;
1976
1977 /* If this ack opens up a zero window, clear backoff. It was
1978 * being used to time the probes, and is probably far higher than
1979 * it needs to be for normal retransmission.
1980 */
1981 if (tp->send_head)
1982 tcp_ack_probe(sk);
1983 return 1;
1984
1985 old_ack:
1986 if (TCP_SKB_CB(skb)->sacked)
1987 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
1988
1989 uninteresting_ack:
1990 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
1991 return 0;
1992 }
1993
1994
1995 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
1996 * But, this can also be called on packets in the established flow when
1997 * the fast version below fails.
1998 */
1999 void tcp_parse_options(struct sk_buff *skb, struct tcp_opt *tp, int estab)
2000 {
2001 unsigned char *ptr;
2002 struct tcphdr *th = skb->h.th;
2003 int length=(th->doff*4)-sizeof(struct tcphdr);
2004
2005 ptr = (unsigned char *)(th + 1);
2006 tp->saw_tstamp = 0;
2007
2008 while(length>0) {
2009 int opcode=*ptr++;
2010 int opsize;
2011
2012 switch (opcode) {
2013 case TCPOPT_EOL:
2014 return;
2015 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2016 length--;
2017 continue;
2018 default:
2019 opsize=*ptr++;
2020 if (opsize < 2) /* "silly options" */
2021 return;
2022 if (opsize > length)
2023 return; /* don't parse partial options */
2024 switch(opcode) {
2025 case TCPOPT_MSS:
2026 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2027 u16 in_mss = ntohs(*(__u16 *)ptr);
2028 if (in_mss) {
2029 if (tp->user_mss && tp->user_mss < in_mss)
2030 in_mss = tp->user_mss;
2031 tp->mss_clamp = in_mss;
2032 }
2033 }
2034 break;
2035 case TCPOPT_WINDOW:
2036 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2037 if (sysctl_tcp_window_scaling) {
2038 tp->wscale_ok = 1;
2039 tp->snd_wscale = *(__u8 *)ptr;
2040 if(tp->snd_wscale > 14) {
2041 if(net_ratelimit())
2042 printk("tcp_parse_options: Illegal window "
2043 "scaling value %d >14 received.",
2044 tp->snd_wscale);
2045 tp->snd_wscale = 14;
2046 }
2047 }
2048 break;
2049 case TCPOPT_TIMESTAMP:
2050 if(opsize==TCPOLEN_TIMESTAMP) {
2051 if ((estab && tp->tstamp_ok) ||
2052 (!estab && sysctl_tcp_timestamps)) {
2053 tp->saw_tstamp = 1;
2054 tp->rcv_tsval = ntohl(*(__u32 *)ptr);
2055 tp->rcv_tsecr = ntohl(*(__u32 *)(ptr+4));
2056 }
2057 }
2058 break;
2059 case TCPOPT_SACK_PERM:
2060 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2061 if (sysctl_tcp_sack) {
2062 tp->sack_ok = 1;
2063 tcp_sack_reset(tp);
2064 }
2065 }
2066 break;
2067
2068 case TCPOPT_SACK:
2069 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2070 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2071 tp->sack_ok) {
2072 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2073 }
2074 };
2075 ptr+=opsize-2;
2076 length-=opsize;
2077 };
2078 }
2079 }
2080
2081 /* Fast parse options. This hopes to only see timestamps.
2082 * If it is wrong it falls back on tcp_parse_options().
2083 */
2084 static __inline__ int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, struct tcp_opt *tp)
2085 {
2086 if (th->doff == sizeof(struct tcphdr)>>2) {
2087 tp->saw_tstamp = 0;
2088 return 0;
2089 } else if (tp->tstamp_ok &&
2090 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2091 __u32 *ptr = (__u32 *)(th + 1);
2092 if (*ptr == __constant_ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2093 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2094 tp->saw_tstamp = 1;
2095 ++ptr;
2096 tp->rcv_tsval = ntohl(*ptr);
2097 ++ptr;
2098 tp->rcv_tsecr = ntohl(*ptr);
2099 return 1;
2100 }
2101 }
2102 tcp_parse_options(skb, tp, 1);
2103 return 1;
2104 }
2105
2106 extern __inline__ void
2107 tcp_store_ts_recent(struct tcp_opt *tp)
2108 {
2109 tp->ts_recent = tp->rcv_tsval;
2110 tp->ts_recent_stamp = xtime.tv_sec;
2111 }
2112
2113 extern __inline__ void
2114 tcp_replace_ts_recent(struct tcp_opt *tp, u32 seq)
2115 {
2116 if (tp->saw_tstamp && !after(seq, tp->rcv_wup)) {
2117 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2118 * extra check below makes sure this can only happen
2119 * for pure ACK frames. -DaveM
2120 *
2121 * Not only, also it occurs for expired timestamps.
2122 */
2123
2124 if((s32)(tp->rcv_tsval - tp->ts_recent) >= 0 ||
2125 xtime.tv_sec >= tp->ts_recent_stamp + TCP_PAWS_24DAYS)
2126 tcp_store_ts_recent(tp);
2127 }
2128 }
2129
2130 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2131 *
2132 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2133 * it can pass through stack. So, the following predicate verifies that
2134 * this segment is not used for anything but congestion avoidance or
2135 * fast retransmit. Moreover, we even are able to eliminate most of such
2136 * second order effects, if we apply some small "replay" window (~RTO)
2137 * to timestamp space.
2138 *
2139 * All these measures still do not guarantee that we reject wrapped ACKs
2140 * on networks with high bandwidth, when sequence space is recycled fastly,
2141 * but it guarantees that such events will be very rare and do not affect
2142 * connection seriously. This doesn't look nice, but alas, PAWS is really
2143 * buggy extension.
2144 *
2145 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2146 * states that events when retransmit arrives after original data are rare.
2147 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2148 * the biggest problem on large power networks even with minor reordering.
2149 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2150 * up to bandwidth of 18Gigabit/sec. 8) ]
2151 */
2152
2153 static int tcp_disordered_ack(struct tcp_opt *tp, struct sk_buff *skb)
2154 {
2155 struct tcphdr *th = skb->h.th;
2156 u32 seq = TCP_SKB_CB(skb)->seq;
2157 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2158
2159 return (/* 1. Pure ACK with correct sequence number. */
2160 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2161
2162 /* 2. ... and duplicate ACK. */
2163 ack == tp->snd_una &&
2164
2165 /* 3. ... and does not update window. */
2166 !tcp_may_update_window(tp, ack, seq, ntohs(th->window)<<tp->snd_wscale) &&
2167
2168 /* 4. ... and sits in replay window. */
2169 (s32)(tp->ts_recent - tp->rcv_tsval) <= (tp->rto*1024)/HZ);
2170 }
2171
2172 extern __inline__ int tcp_paws_discard(struct tcp_opt *tp, struct sk_buff *skb)
2173 {
2174 return ((s32)(tp->ts_recent - tp->rcv_tsval) > TCP_PAWS_WINDOW &&
2175 xtime.tv_sec < tp->ts_recent_stamp + TCP_PAWS_24DAYS &&
2176 !tcp_disordered_ack(tp, skb));
2177 }
2178
2179 /* Check segment sequence number for validity.
2180 *
2181 * Segment controls are considered valid, if the segment
2182 * fits to the window after truncation to the window. Acceptability
2183 * of data (and SYN, FIN, of course) is checked separately.
2184 * See tcp_data_queue(), for example.
2185 *
2186 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2187 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2188 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2189 * (borrowed from freebsd)
2190 */
2191
2192 static inline int tcp_sequence(struct tcp_opt *tp, u32 seq, u32 end_seq)
2193 {
2194 return !before(end_seq, tp->rcv_wup) &&
2195 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2196 }
2197
2198 /* When we get a reset we do this. */
2199 static void tcp_reset(struct sock *sk)
2200 {
2201 /* We want the right error as BSD sees it (and indeed as we do). */
2202 switch (sk->state) {
2203 case TCP_SYN_SENT:
2204 sk->err = ECONNREFUSED;
2205 break;
2206 case TCP_CLOSE_WAIT:
2207 sk->err = EPIPE;
2208 break;
2209 case TCP_CLOSE:
2210 return;
2211 default:
2212 sk->err = ECONNRESET;
2213 }
2214
2215 if (!sk->dead)
2216 sk->error_report(sk);
2217
2218 tcp_done(sk);
2219 }
2220
2221 /*
2222 * Process the FIN bit. This now behaves as it is supposed to work
2223 * and the FIN takes effect when it is validly part of sequence
2224 * space. Not before when we get holes.
2225 *
2226 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2227 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2228 * TIME-WAIT)
2229 *
2230 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2231 * close and we go into CLOSING (and later onto TIME-WAIT)
2232 *
2233 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2234 */
2235 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2236 {
2237 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2238
2239 tcp_schedule_ack(tp);
2240
2241 sk->shutdown |= RCV_SHUTDOWN;
2242 sk->done = 1;
2243
2244 switch(sk->state) {
2245 case TCP_SYN_RECV:
2246 case TCP_ESTABLISHED:
2247 /* Move to CLOSE_WAIT */
2248 tcp_set_state(sk, TCP_CLOSE_WAIT);
2249 tp->ack.pingpong = 1;
2250 break;
2251
2252 case TCP_CLOSE_WAIT:
2253 case TCP_CLOSING:
2254 /* Received a retransmission of the FIN, do
2255 * nothing.
2256 */
2257 break;
2258 case TCP_LAST_ACK:
2259 /* RFC793: Remain in the LAST-ACK state. */
2260 break;
2261
2262 case TCP_FIN_WAIT1:
2263 /* This case occurs when a simultaneous close
2264 * happens, we must ack the received FIN and
2265 * enter the CLOSING state.
2266 */
2267 tcp_send_ack(sk);
2268 tcp_set_state(sk, TCP_CLOSING);
2269 break;
2270 case TCP_FIN_WAIT2:
2271 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2272 tcp_send_ack(sk);
2273 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2274 break;
2275 default:
2276 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2277 * cases we should never reach this piece of code.
2278 */
2279 printk("tcp_fin: Impossible, sk->state=%d\n", sk->state);
2280 break;
2281 };
2282
2283 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2284 * Probably, we should reset in this case. For now drop them.
2285 */
2286 __skb_queue_purge(&tp->out_of_order_queue);
2287 if (tp->sack_ok)
2288 tcp_sack_reset(tp);
2289 tcp_mem_reclaim(sk);
2290
2291 if (!sk->dead) {
2292 sk->state_change(sk);
2293
2294 /* Do not send POLL_HUP for half duplex close. */
2295 if (sk->shutdown == SHUTDOWN_MASK || sk->state == TCP_CLOSE)
2296 sk_wake_async(sk, 1, POLL_HUP);
2297 else
2298 sk_wake_async(sk, 1, POLL_IN);
2299 }
2300 }
2301
2302 static __inline__ int
2303 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2304 {
2305 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2306 if (before(seq, sp->start_seq))
2307 sp->start_seq = seq;
2308 if (after(end_seq, sp->end_seq))
2309 sp->end_seq = end_seq;
2310 return 1;
2311 }
2312 return 0;
2313 }
2314
2315 static __inline__ void tcp_dsack_set(struct tcp_opt *tp, u32 seq, u32 end_seq)
2316 {
2317 if (tp->sack_ok && sysctl_tcp_dsack) {
2318 if (before(seq, tp->rcv_nxt))
2319 NET_INC_STATS_BH(TCPDSACKOldSent);
2320 else
2321 NET_INC_STATS_BH(TCPDSACKOfoSent);
2322
2323 tp->dsack = 1;
2324 tp->duplicate_sack[0].start_seq = seq;
2325 tp->duplicate_sack[0].end_seq = end_seq;
2326 tp->eff_sacks = min_t(unsigned int, tp->num_sacks+1, 4-tp->tstamp_ok);
2327 }
2328 }
2329
2330 static __inline__ void tcp_dsack_extend(struct tcp_opt *tp, u32 seq, u32 end_seq)
2331 {
2332 if (!tp->dsack)
2333 tcp_dsack_set(tp, seq, end_seq);
2334 else
2335 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2336 }
2337
2338 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2339 {
2340 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2341
2342 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2343 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2344 NET_INC_STATS_BH(DelayedACKLost);
2345 tcp_enter_quickack_mode(tp);
2346
2347 if (tp->sack_ok && sysctl_tcp_dsack) {
2348 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2349
2350 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2351 end_seq = tp->rcv_nxt;
2352 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2353 }
2354 }
2355
2356 tcp_send_ack(sk);
2357 }
2358
2359 /* These routines update the SACK block as out-of-order packets arrive or
2360 * in-order packets close up the sequence space.
2361 */
2362 static void tcp_sack_maybe_coalesce(struct tcp_opt *tp)
2363 {
2364 int this_sack;
2365 struct tcp_sack_block *sp = &tp->selective_acks[0];
2366 struct tcp_sack_block *swalk = sp+1;
2367
2368 /* See if the recent change to the first SACK eats into
2369 * or hits the sequence space of other SACK blocks, if so coalesce.
2370 */
2371 for (this_sack = 1; this_sack < tp->num_sacks; ) {
2372 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2373 int i;
2374
2375 /* Zap SWALK, by moving every further SACK up by one slot.
2376 * Decrease num_sacks.
2377 */
2378 tp->num_sacks--;
2379 tp->eff_sacks = min_t(unsigned int, tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
2380 for(i=this_sack; i < tp->num_sacks; i++)
2381 sp[i] = sp[i+1];
2382 continue;
2383 }
2384 this_sack++, swalk++;
2385 }
2386 }
2387
2388 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2389 {
2390 __u32 tmp;
2391
2392 tmp = sack1->start_seq;
2393 sack1->start_seq = sack2->start_seq;
2394 sack2->start_seq = tmp;
2395
2396 tmp = sack1->end_seq;
2397 sack1->end_seq = sack2->end_seq;
2398 sack2->end_seq = tmp;
2399 }
2400
2401 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
2402 {
2403 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2404 struct tcp_sack_block *sp = &tp->selective_acks[0];
2405 int cur_sacks = tp->num_sacks;
2406 int this_sack;
2407
2408 if (!cur_sacks)
2409 goto new_sack;
2410
2411 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
2412 if (tcp_sack_extend(sp, seq, end_seq)) {
2413 /* Rotate this_sack to the first one. */
2414 for (; this_sack>0; this_sack--, sp--)
2415 tcp_sack_swap(sp, sp-1);
2416 if (cur_sacks > 1)
2417 tcp_sack_maybe_coalesce(tp);
2418 return;
2419 }
2420 }
2421
2422 /* Could not find an adjacent existing SACK, build a new one,
2423 * put it at the front, and shift everyone else down. We
2424 * always know there is at least one SACK present already here.
2425 *
2426 * If the sack array is full, forget about the last one.
2427 */
2428 if (this_sack >= 4) {
2429 this_sack--;
2430 tp->num_sacks--;
2431 sp--;
2432 }
2433 for(; this_sack > 0; this_sack--, sp--)
2434 *sp = *(sp-1);
2435
2436 new_sack:
2437 /* Build the new head SACK, and we're done. */
2438 sp->start_seq = seq;
2439 sp->end_seq = end_seq;
2440 tp->num_sacks++;
2441 tp->eff_sacks = min_t(unsigned int, tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
2442 }
2443
2444 /* RCV.NXT advances, some SACKs should be eaten. */
2445
2446 static void tcp_sack_remove(struct tcp_opt *tp)
2447 {
2448 struct tcp_sack_block *sp = &tp->selective_acks[0];
2449 int num_sacks = tp->num_sacks;
2450 int this_sack;
2451
2452 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2453 if (skb_queue_len(&tp->out_of_order_queue) == 0) {
2454 tp->num_sacks = 0;
2455 tp->eff_sacks = tp->dsack;
2456 return;
2457 }
2458
2459 for(this_sack = 0; this_sack < num_sacks; ) {
2460 /* Check if the start of the sack is covered by RCV.NXT. */
2461 if (!before(tp->rcv_nxt, sp->start_seq)) {
2462 int i;
2463
2464 /* RCV.NXT must cover all the block! */
2465 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
2466
2467 /* Zap this SACK, by moving forward any other SACKS. */
2468 for (i=this_sack+1; i < num_sacks; i++)
2469 sp[i-1] = sp[i];
2470 num_sacks--;
2471 continue;
2472 }
2473 this_sack++;
2474 sp++;
2475 }
2476 if (num_sacks != tp->num_sacks) {
2477 tp->num_sacks = num_sacks;
2478 tp->eff_sacks = min_t(unsigned int, tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
2479 }
2480 }
2481
2482 /* This one checks to see if we can put data from the
2483 * out_of_order queue into the receive_queue.
2484 */
2485 static void tcp_ofo_queue(struct sock *sk)
2486 {
2487 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2488 __u32 dsack_high = tp->rcv_nxt;
2489 struct sk_buff *skb;
2490
2491 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
2492 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
2493 break;
2494
2495 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
2496 __u32 dsack = dsack_high;
2497 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
2498 dsack_high = TCP_SKB_CB(skb)->end_seq;
2499 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
2500 }
2501
2502 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
2503 SOCK_DEBUG(sk, "ofo packet was already received \n");
2504 __skb_unlink(skb, skb->list);
2505 __kfree_skb(skb);
2506 continue;
2507 }
2508 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
2509 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
2510 TCP_SKB_CB(skb)->end_seq);
2511
2512 __skb_unlink(skb, skb->list);
2513 __skb_queue_tail(&sk->receive_queue, skb);
2514 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2515 if(skb->h.th->fin)
2516 tcp_fin(skb, sk, skb->h.th);
2517 }
2518 }
2519
2520 static inline int tcp_rmem_schedule(struct sock *sk, struct sk_buff *skb)
2521 {
2522 return (int)skb->truesize <= sk->forward_alloc ||
2523 tcp_mem_schedule(sk, skb->truesize, 1);
2524 }
2525
2526 static int tcp_prune_queue(struct sock *sk);
2527
2528 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
2529 {
2530 struct tcphdr *th = skb->h.th;
2531 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2532 int eaten = -1;
2533
2534 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
2535 goto drop;
2536
2537 th = skb->h.th;
2538 __skb_pull(skb, th->doff*4);
2539
2540 TCP_ECN_accept_cwr(tp, skb);
2541
2542 if (tp->dsack) {
2543 tp->dsack = 0;
2544 tp->eff_sacks = min_t(unsigned int, tp->num_sacks, 4-tp->tstamp_ok);
2545 }
2546
2547 /* Queue data for delivery to the user.
2548 * Packets in sequence go to the receive queue.
2549 * Out of sequence packets to the out_of_order_queue.
2550 */
2551 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
2552 if (tcp_receive_window(tp) == 0)
2553 goto out_of_window;
2554
2555 /* Ok. In sequence. In window. */
2556 if (tp->ucopy.task == current &&
2557 tp->copied_seq == tp->rcv_nxt &&
2558 tp->ucopy.len &&
2559 sk->lock.users &&
2560 !tp->urg_data) {
2561 int chunk = min_t(unsigned int, skb->len, tp->ucopy.len);
2562
2563 __set_current_state(TASK_RUNNING);
2564
2565 local_bh_enable();
2566 if (skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
2567 sk->err = EFAULT;
2568 sk->error_report(sk);
2569 }
2570 local_bh_disable();
2571 tp->ucopy.len -= chunk;
2572 tp->copied_seq += chunk;
2573 eaten = (chunk == skb->len && !th->fin);
2574 }
2575
2576 if (eaten <= 0) {
2577 queue_and_out:
2578 if (eaten < 0 &&
2579 (atomic_read(&sk->rmem_alloc) > sk->rcvbuf ||
2580 !tcp_rmem_schedule(sk, skb))) {
2581 if (tcp_prune_queue(sk) < 0 || !tcp_rmem_schedule(sk, skb))
2582 goto drop;
2583 }
2584 tcp_set_owner_r(skb, sk);
2585 __skb_queue_tail(&sk->receive_queue, skb);
2586 }
2587 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2588 if(skb->len)
2589 tcp_event_data_recv(sk, tp, skb);
2590 if(th->fin)
2591 tcp_fin(skb, sk, th);
2592
2593 if (skb_queue_len(&tp->out_of_order_queue)) {
2594 tcp_ofo_queue(sk);
2595
2596 /* RFC2581. 4.2. SHOULD send immediate ACK, when
2597 * gap in queue is filled.
2598 */
2599 if (skb_queue_len(&tp->out_of_order_queue) == 0)
2600 tp->ack.pingpong = 0;
2601 }
2602
2603 if(tp->num_sacks)
2604 tcp_sack_remove(tp);
2605
2606 tcp_fast_path_check(sk, tp);
2607
2608 if (eaten > 0) {
2609 __kfree_skb(skb);
2610 } else if (!sk->dead)
2611 sk->data_ready(sk, 0);
2612 return;
2613 }
2614
2615 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
2616 /* A retransmit, 2nd most common case. Force an immediate ack. */
2617 NET_INC_STATS_BH(DelayedACKLost);
2618 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
2619
2620 out_of_window:
2621 tcp_enter_quickack_mode(tp);
2622 tcp_schedule_ack(tp);
2623 drop:
2624 __kfree_skb(skb);
2625 return;
2626 }
2627
2628 /* Out of window. F.e. zero window probe. */
2629 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt+tcp_receive_window(tp)))
2630 goto out_of_window;
2631
2632 tcp_enter_quickack_mode(tp);
2633
2634 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2635 /* Partial packet, seq < rcv_next < end_seq */
2636 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
2637 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
2638 TCP_SKB_CB(skb)->end_seq);
2639
2640 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
2641
2642 /* If window is closed, drop tail of packet. But after
2643 * remembering D-SACK for its head made in previous line.
2644 */
2645 if (!tcp_receive_window(tp))
2646 goto out_of_window;
2647 goto queue_and_out;
2648 }
2649
2650 TCP_ECN_check_ce(tp, skb);
2651
2652 if (atomic_read(&sk->rmem_alloc) > sk->rcvbuf ||
2653 !tcp_rmem_schedule(sk, skb)) {
2654 if (tcp_prune_queue(sk) < 0 || !tcp_rmem_schedule(sk, skb))
2655 goto drop;
2656 }
2657
2658 /* Disable header prediction. */
2659 tp->pred_flags = 0;
2660 tcp_schedule_ack(tp);
2661
2662 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
2663 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
2664
2665 tcp_set_owner_r(skb, sk);
2666
2667 if (skb_peek(&tp->out_of_order_queue) == NULL) {
2668 /* Initial out of order segment, build 1 SACK. */
2669 if(tp->sack_ok) {
2670 tp->num_sacks = 1;
2671 tp->dsack = 0;
2672 tp->eff_sacks = 1;
2673 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
2674 tp->selective_acks[0].end_seq = TCP_SKB_CB(skb)->end_seq;
2675 }
2676 __skb_queue_head(&tp->out_of_order_queue,skb);
2677 } else {
2678 struct sk_buff *skb1=tp->out_of_order_queue.prev;
2679 u32 seq = TCP_SKB_CB(skb)->seq;
2680 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2681
2682 if (seq == TCP_SKB_CB(skb1)->end_seq) {
2683 __skb_append(skb1, skb);
2684
2685 if (tp->num_sacks == 0 ||
2686 tp->selective_acks[0].end_seq != seq)
2687 goto add_sack;
2688
2689 /* Common case: data arrive in order after hole. */
2690 tp->selective_acks[0].end_seq = end_seq;
2691 return;
2692 }
2693
2694 /* Find place to insert this segment. */
2695 do {
2696 if (!after(TCP_SKB_CB(skb1)->seq, seq))
2697 break;
2698 } while ((skb1=skb1->prev) != (struct sk_buff*)&tp->out_of_order_queue);
2699
2700 /* Do skb overlap to previous one? */
2701 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
2702 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
2703 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
2704 /* All the bits are present. Drop. */
2705 __kfree_skb(skb);
2706 tcp_dsack_set(tp, seq, end_seq);
2707 goto add_sack;
2708 }
2709 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
2710 /* Partial overlap. */
2711 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
2712 } else {
2713 skb1 = skb1->prev;
2714 }
2715 }
2716 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
2717
2718 /* And clean segments covered by new one as whole. */
2719 while ((skb1 = skb->next) != (struct sk_buff*)&tp->out_of_order_queue &&
2720 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
2721 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
2722 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
2723 break;
2724 }
2725 __skb_unlink(skb1, skb1->list);
2726 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
2727 __kfree_skb(skb1);
2728 }
2729
2730 add_sack:
2731 if (tp->sack_ok)
2732 tcp_sack_new_ofo_skb(sk, seq, end_seq);
2733 }
2734 }
2735
2736 /* Collapse contiguous sequence of skbs head..tail with
2737 * sequence numbers start..end.
2738 * Segments with FIN/SYN are not collapsed (only because this
2739 * simplifies code)
2740 */
2741 static void
2742 tcp_collapse(struct sock *sk, struct sk_buff *head,
2743 struct sk_buff *tail, u32 start, u32 end)
2744 {
2745 struct sk_buff *skb;
2746
2747 /* First, check that queue is collapsable and find
2748 * the point where collapsing can be useful. */
2749 for (skb = head; skb != tail; ) {
2750 /* No new bits? It is possible on ofo queue. */
2751 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
2752 struct sk_buff *next = skb->next;
2753 __skb_unlink(skb, skb->list);
2754 __kfree_skb(skb);
2755 NET_INC_STATS_BH(TCPRcvCollapsed);
2756 skb = next;
2757 continue;
2758 }
2759
2760 /* The first skb to collapse is:
2761 * - not SYN/FIN and
2762 * - bloated or contains data before "start" or
2763 * overlaps to the next one.
2764 */
2765 if (!skb->h.th->syn && !skb->h.th->fin &&
2766 (tcp_win_from_space(skb->truesize) > skb->len ||
2767 before(TCP_SKB_CB(skb)->seq, start) ||
2768 (skb->next != tail &&
2769 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
2770 break;
2771
2772 /* Decided to skip this, advance start seq. */
2773 start = TCP_SKB_CB(skb)->end_seq;
2774 skb = skb->next;
2775 }
2776 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
2777 return;
2778
2779 while (before(start, end)) {
2780 struct sk_buff *nskb;
2781 int header = skb_headroom(skb);
2782 int copy = (PAGE_SIZE - sizeof(struct sk_buff) -
2783 sizeof(struct skb_shared_info) - header - 31)&~15;
2784
2785 /* Too big header? This can happen with IPv6. */
2786 if (copy < 0)
2787 return;
2788 if (end-start < copy)
2789 copy = end-start;
2790 nskb = alloc_skb(copy+header, GFP_ATOMIC);
2791 if (!nskb)
2792 return;
2793 skb_reserve(nskb, header);
2794 memcpy(nskb->head, skb->head, header);
2795 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
2796 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
2797 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
2798 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
2799 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
2800 __skb_insert(nskb, skb->prev, skb, skb->list);
2801 tcp_set_owner_r(nskb, sk);
2802
2803 /* Copy data, releasing collapsed skbs. */
2804 while (copy > 0) {
2805 int offset = start - TCP_SKB_CB(skb)->seq;
2806 int size = TCP_SKB_CB(skb)->end_seq - start;
2807
2808 if (offset < 0) BUG();
2809 if (size > 0) {
2810 size = min_t(int, copy, size);
2811 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
2812 BUG();
2813 TCP_SKB_CB(nskb)->end_seq += size;
2814 copy -= size;
2815 start += size;
2816 }
2817 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
2818 struct sk_buff *next = skb->next;
2819 __skb_unlink(skb, skb->list);
2820 __kfree_skb(skb);
2821 NET_INC_STATS_BH(TCPRcvCollapsed);
2822 skb = next;
2823 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
2824 return;
2825 }
2826 }
2827 }
2828 }
2829
2830 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
2831 * and tcp_collapse() them until all the queue is collapsed.
2832 */
2833 static void tcp_collapse_ofo_queue(struct sock *sk)
2834 {
2835 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2836 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
2837 struct sk_buff *head;
2838 u32 start, end;
2839
2840 if (skb == NULL)
2841 return;
2842
2843 start = TCP_SKB_CB(skb)->seq;
2844 end = TCP_SKB_CB(skb)->end_seq;
2845 head = skb;
2846
2847 for (;;) {
2848 skb = skb->next;
2849
2850 /* Segment is terminated when we see gap or when
2851 * we are at the end of all the queue. */
2852 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
2853 after(TCP_SKB_CB(skb)->seq, end) ||
2854 before(TCP_SKB_CB(skb)->end_seq, start)) {
2855 tcp_collapse(sk, head, skb, start, end);
2856 head = skb;
2857 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
2858 break;
2859 /* Start new segment */
2860 start = TCP_SKB_CB(skb)->seq;
2861 end = TCP_SKB_CB(skb)->end_seq;
2862 } else {
2863 if (before(TCP_SKB_CB(skb)->seq, start))
2864 start = TCP_SKB_CB(skb)->seq;
2865 if (after(TCP_SKB_CB(skb)->end_seq, end))
2866 end = TCP_SKB_CB(skb)->end_seq;
2867 }
2868 }
2869 }
2870
2871 /* Reduce allocated memory if we can, trying to get
2872 * the socket within its memory limits again.
2873 *
2874 * Return less than zero if we should start dropping frames
2875 * until the socket owning process reads some of the data
2876 * to stabilize the situation.
2877 */
2878 static int tcp_prune_queue(struct sock *sk)
2879 {
2880 struct tcp_opt *tp = &sk->tp_pinfo.af_tcp;
2881
2882 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
2883
2884 NET_INC_STATS_BH(PruneCalled);
2885
2886 if (atomic_read(&sk->rmem_alloc) >= sk->rcvbuf)
2887 tcp_clamp_window(sk, tp);
2888 else if (tcp_memory_pressure)
2889 tp->rcv_ssthresh = min_t(u32, tp->rcv_ssthresh, 4*tp->advmss);
2890
2891 tcp_collapse_ofo_queue(sk);
2892 tcp_collapse(sk, sk->receive_queue.next,
2893 (struct sk_buff*)&sk->receive_queue,
2894 tp->copied_seq, tp->rcv_nxt);
2895 tcp_mem_reclaim(sk);
2896
2897 if (atomic_read(&sk->rmem_alloc) <= sk->rcvbuf)
2898 return 0;
2899
2900 /* Collapsing did not help, destructive actions follow.
2901 * This must not ever occur. */
2902
2903 /* First, purge the out_of_order queue. */
2904 if (skb_queue_len(&tp->out_of_order_queue)) {
2905 net_statistics[smp_processor_id()*2].OfoPruned += skb_queue_len(&tp->out_of_order_queue);
2906 __skb_queue_purge(&tp->out_of_order_queue);
2907
2908 /* Reset SACK state. A conforming SACK implementation will
2909 * do the same at a timeout based retransmit. When a connection
2910 * is in a sad state like this, we care only about integrity
2911 * of the connection not performance.
2912 */
2913 if(tp->sack_ok)
2914 tcp_sack_reset(tp);
2915 tcp_mem_reclaim(sk);
2916 }
2917
2918 if(atomic_read(&sk->rmem_alloc) <= sk->rcvbuf)
2919 return 0;
2920
2921 /* If we are really being abused, tell the caller to silently
2922 * drop receive data on the floor. It will get retransmitted
2923 * and hopefully then we'll have sufficient space.
2924 */
2925 NET_INC_STATS_BH(RcvPruned);
2926
2927 /* Massive buffer overcommit. */
2928 tp->pred_flags = 0;
2929 return -1;
2930 }
2931
2932
2933 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
2934 * As additional protections, we do not touch cwnd in retransmission phases,
2935 * and if application hit its sndbuf limit recently.
2936 */
2937 void tcp_cwnd_application_limited(struct sock *sk)
2938 {
2939 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2940
2941 if (tp->ca_state == TCP_CA_Open &&
2942 sk->socket && !test_bit(SOCK_NOSPACE, &sk->socket->flags)) {
2943 /* Limited by application or receiver window. */
2944 u32 win_used = max_t(u32, tp->snd_cwnd_used, 2);
2945 if (win_used < tp->snd_cwnd) {
2946 tp->snd_ssthresh = tcp_current_ssthresh(tp);
2947 tp->snd_cwnd = (tp->snd_cwnd+win_used)>>1;
2948 }
2949 tp->snd_cwnd_used = 0;
2950 }
2951 tp->snd_cwnd_stamp = tcp_time_stamp;
2952 }
2953
2954
2955 /* When incoming ACK allowed to free some skb from write_queue,
2956 * we remember this event in flag tp->queue_shrunk and wake up socket
2957 * on the exit from tcp input handler.
2958 */
2959 static void tcp_new_space(struct sock *sk)
2960 {
2961 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2962
2963 if (tp->packets_out < tp->snd_cwnd &&
2964 !(sk->userlocks&SOCK_SNDBUF_LOCK) &&
2965 !tcp_memory_pressure &&
2966 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
2967 int sndmem, demanded;
2968
2969 sndmem = tp->mss_clamp+MAX_TCP_HEADER+16+sizeof(struct sk_buff);
2970 demanded = max_t(unsigned int, tp->snd_cwnd, tp->reordering+1);
2971 sndmem *= 2*demanded;
2972 if (sndmem > sk->sndbuf)
2973 sk->sndbuf = min_t(int, sndmem, sysctl_tcp_wmem[2]);
2974 tp->snd_cwnd_stamp = tcp_time_stamp;
2975 }
2976
2977 sk->write_space(sk);
2978 }
2979
2980 static inline void tcp_check_space(struct sock *sk)
2981 {
2982 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2983
2984 if (tp->queue_shrunk) {
2985 tp->queue_shrunk = 0;
2986 if (sk->socket && test_bit(SOCK_NOSPACE, &sk->socket->flags))
2987 tcp_new_space(sk);
2988 }
2989 }
2990
2991 static void __tcp_data_snd_check(struct sock *sk, struct sk_buff *skb)
2992 {
2993 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
2994
2995 if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd) ||
2996 tcp_packets_in_flight(tp) >= tp->snd_cwnd ||
2997 tcp_write_xmit(sk, tp->nonagle))
2998 tcp_check_probe_timer(sk, tp);
2999 }
3000
3001 static __inline__ void tcp_data_snd_check(struct sock *sk)
3002 {
3003 struct sk_buff *skb = sk->tp_pinfo.af_tcp.send_head;
3004
3005 if (skb != NULL)
3006 __tcp_data_snd_check(sk, skb);
3007 tcp_check_space(sk);
3008 }
3009
3010 /*
3011 * Check if sending an ack is needed.
3012 */
3013 static __inline__ void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3014 {
3015 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3016
3017 /* More than one full frame received... */
3018 if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss
3019 /* ... and right edge of window advances far enough.
3020 * (tcp_recvmsg() will send ACK otherwise). Or...
3021 */
3022 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3023 /* We ACK each frame or... */
3024 tcp_in_quickack_mode(tp) ||
3025 /* We have out of order data. */
3026 (ofo_possible &&
3027 skb_peek(&tp->out_of_order_queue) != NULL)) {
3028 /* Then ack it now */
3029 tcp_send_ack(sk);
3030 } else {
3031 /* Else, send delayed ack. */
3032 tcp_send_delayed_ack(sk);
3033 }
3034 }
3035
3036 static __inline__ void tcp_ack_snd_check(struct sock *sk)
3037 {
3038 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3039 if (!tcp_ack_scheduled(tp)) {
3040 /* We sent a data segment already. */
3041 return;
3042 }
3043 __tcp_ack_snd_check(sk, 1);
3044 }
3045
3046 /*
3047 * This routine is only called when we have urgent data
3048 * signalled. Its the 'slow' part of tcp_urg. It could be
3049 * moved inline now as tcp_urg is only called from one
3050 * place. We handle URGent data wrong. We have to - as
3051 * BSD still doesn't use the correction from RFC961.
3052 * For 1003.1g we should support a new option TCP_STDURG to permit
3053 * either form (or just set the sysctl tcp_stdurg).
3054 */
3055
3056 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3057 {
3058 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3059 u32 ptr = ntohs(th->urg_ptr);
3060
3061 if (ptr && !sysctl_tcp_stdurg)
3062 ptr--;
3063 ptr += ntohl(th->seq);
3064
3065 /* Ignore urgent data that we've already seen and read. */
3066 if (after(tp->copied_seq, ptr))
3067 return;
3068
3069 /* Do not replay urg ptr.
3070 *
3071 * NOTE: interesting situation not covered by specs.
3072 * Misbehaving sender may send urg ptr, pointing to segment,
3073 * which we already have in ofo queue. We are not able to fetch
3074 * such data and will stay in TCP_URG_NOTYET until will be eaten
3075 * by recvmsg(). Seems, we are not obliged to handle such wicked
3076 * situations. But it is worth to think about possibility of some
3077 * DoSes using some hypothetical application level deadlock.
3078 */
3079 if (before(ptr, tp->rcv_nxt))
3080 return;
3081
3082 /* Do we already have a newer (or duplicate) urgent pointer? */
3083 if (tp->urg_data && !after(ptr, tp->urg_seq))
3084 return;
3085
3086 /* Tell the world about our new urgent pointer. */
3087 if (sk->proc != 0) {
3088 if (sk->proc > 0)
3089 kill_proc(sk->proc, SIGURG, 1);
3090 else
3091 kill_pg(-sk->proc, SIGURG, 1);
3092 sk_wake_async(sk, 3, POLL_PRI);
3093 }
3094
3095 /* We may be adding urgent data when the last byte read was
3096 * urgent. To do this requires some care. We cannot just ignore
3097 * tp->copied_seq since we would read the last urgent byte again
3098 * as data, nor can we alter copied_seq until this data arrives
3099 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3100 *
3101 * NOTE. Double Dutch. Rendering to plain English: author of comment
3102 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3103 * and expect that both A and B disappear from stream. This is _wrong_.
3104 * Though this happens in BSD with high probability, this is occasional.
3105 * Any application relying on this is buggy. Note also, that fix "works"
3106 * only in this artificial test. Insert some normal data between A and B and we will
3107 * decline of BSD again. Verdict: it is better to remove to trap
3108 * buggy users.
3109 */
3110 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3111 !sk->urginline &&
3112 tp->copied_seq != tp->rcv_nxt) {
3113 struct sk_buff *skb = skb_peek(&sk->receive_queue);
3114 tp->copied_seq++;
3115 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3116 __skb_unlink(skb, skb->list);
3117 __kfree_skb(skb);
3118 }
3119 }
3120
3121 tp->urg_data = TCP_URG_NOTYET;
3122 tp->urg_seq = ptr;
3123
3124 /* Disable header prediction. */
3125 tp->pred_flags = 0;
3126 }
3127
3128 /* This is the 'fast' part of urgent handling. */
3129 static inline void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3130 {
3131 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3132
3133 /* Check if we get a new urgent pointer - normally not. */
3134 if (th->urg)
3135 tcp_check_urg(sk,th);
3136
3137 /* Do we wait for any urgent data? - normally not... */
3138 if (tp->urg_data == TCP_URG_NOTYET) {
3139 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff*4) - th->syn;
3140
3141 /* Is the urgent pointer pointing into this packet? */
3142 if (ptr < skb->len) {
3143 u8 tmp;
3144 if (skb_copy_bits(skb, ptr, &tmp, 1))
3145 BUG();
3146 tp->urg_data = TCP_URG_VALID | tmp;
3147 if (!sk->dead)
3148 sk->data_ready(sk,0);
3149 }
3150 }
3151 }
3152
3153 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3154 {
3155 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3156 int chunk = skb->len - hlen;
3157 int err;
3158
3159 local_bh_enable();
3160 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3161 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3162 else
3163 err = skb_copy_and_csum_datagram_iovec(skb, hlen, tp->ucopy.iov);
3164
3165 if (!err) {
3166 update:
3167 tp->ucopy.len -= chunk;
3168 tp->copied_seq += chunk;
3169 local_bh_disable();
3170 return 0;
3171 }
3172
3173 if (err == -EFAULT) {
3174 sk->err = EFAULT;
3175 sk->error_report(sk);
3176 goto update;
3177 }
3178
3179 local_bh_disable();
3180 return err;
3181 }
3182
3183 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3184 {
3185 int result;
3186
3187 if (sk->lock.users) {
3188 local_bh_enable();
3189 result = __tcp_checksum_complete(skb);
3190 local_bh_disable();
3191 } else {
3192 result = __tcp_checksum_complete(skb);
3193 }
3194 return result;
3195 }
3196
3197 static __inline__ int
3198 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3199 {
3200 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3201 __tcp_checksum_complete_user(sk, skb);
3202 }
3203
3204 /*
3205 * TCP receive function for the ESTABLISHED state.
3206 *
3207 * It is split into a fast path and a slow path. The fast path is
3208 * disabled when:
3209 * - A zero window was announced from us - zero window probing
3210 * is only handled properly in the slow path.
3211 * - Out of order segments arrived.
3212 * - Urgent data is expected.
3213 * - There is no buffer space left
3214 * - Unexpected TCP flags/window values/header lengths are received
3215 * (detected by checking the TCP header against pred_flags)
3216 * - Data is sent in both directions. Fast path only supports pure senders
3217 * or pure receivers (this means either the sequence number or the ack
3218 * value must stay constant)
3219 * - Unexpected TCP option.
3220 *
3221 * When these conditions are not satisfied it drops into a standard
3222 * receive procedure patterned after RFC793 to handle all cases.
3223 * The first three cases are guaranteed by proper pred_flags setting,
3224 * the rest is checked inline. Fast processing is turned on in
3225 * tcp_data_queue when everything is OK.
3226 */
3227 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3228 struct tcphdr *th, unsigned len)
3229 {
3230 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3231
3232 /*
3233 * Header prediction.
3234 * The code losely follows the one in the famous
3235 * "30 instruction TCP receive" Van Jacobson mail.
3236 *
3237 * Van's trick is to deposit buffers into socket queue
3238 * on a device interrupt, to call tcp_recv function
3239 * on the receive process context and checksum and copy
3240 * the buffer to user space. smart...
3241 *
3242 * Our current scheme is not silly either but we take the
3243 * extra cost of the net_bh soft interrupt processing...
3244 * We do checksum and copy also but from device to kernel.
3245 */
3246
3247 tp->saw_tstamp = 0;
3248
3249 /* pred_flags is 0xS?10 << 16 + snd_wnd
3250 * if header_predition is to be made
3251 * 'S' will always be tp->tcp_header_len >> 2
3252 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3253 * turn it off (when there are holes in the receive
3254 * space for instance)
3255 * PSH flag is ignored.
3256 */
3257
3258 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3259 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3260 int tcp_header_len = tp->tcp_header_len;
3261
3262 /* Timestamp header prediction: tcp_header_len
3263 * is automatically equal to th->doff*4 due to pred_flags
3264 * match.
3265 */
3266
3267 /* Check timestamp */
3268 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3269 __u32 *ptr = (__u32 *)(th + 1);
3270
3271 /* No? Slow path! */
3272 if (*ptr != __constant_ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3273 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3274 goto slow_path;
3275
3276 tp->saw_tstamp = 1;
3277 ++ptr;
3278 tp->rcv_tsval = ntohl(*ptr);
3279 ++ptr;
3280 tp->rcv_tsecr = ntohl(*ptr);
3281
3282 /* If PAWS failed, check it more carefully in slow path */
3283 if ((s32)(tp->rcv_tsval - tp->ts_recent) < 0)
3284 goto slow_path;
3285
3286 /* Predicted packet is in window by definition.
3287 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3288 * Hence, check seq<=rcv_wup reduces to:
3289 */
3290 if (tp->rcv_nxt == tp->rcv_wup)
3291 tcp_store_ts_recent(tp);
3292 }
3293
3294 if (len <= tcp_header_len) {
3295 /* Bulk data transfer: sender */
3296 if (len == tcp_header_len) {
3297 /* We know that such packets are checksummed
3298 * on entry.
3299 */
3300 tcp_ack(sk, skb, 0);
3301 __kfree_skb(skb);
3302 tcp_data_snd_check(sk);
3303 return 0;
3304 } else { /* Header too small */
3305 TCP_INC_STATS_BH(TcpInErrs);
3306 goto discard;
3307 }
3308 } else {
3309 int eaten = 0;
3310
3311 if (tp->ucopy.task == current &&
3312 tp->copied_seq == tp->rcv_nxt &&
3313 len - tcp_header_len <= tp->ucopy.len &&
3314 sk->lock.users) {
3315 eaten = 1;
3316
3317 NET_INC_STATS_BH(TCPHPHitsToUser);
3318
3319 __set_current_state(TASK_RUNNING);
3320
3321 if (tcp_copy_to_iovec(sk, skb, tcp_header_len))
3322 goto csum_error;
3323
3324 __skb_pull(skb,tcp_header_len);
3325
3326 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3327 } else {
3328 if (tcp_checksum_complete_user(sk, skb))
3329 goto csum_error;
3330
3331 if ((int)skb->truesize > sk->forward_alloc)
3332 goto step5;
3333
3334 NET_INC_STATS_BH(TCPHPHits);
3335
3336 /* Bulk data transfer: receiver */
3337 __skb_pull(skb,tcp_header_len);
3338 __skb_queue_tail(&sk->receive_queue, skb);
3339 tcp_set_owner_r(skb, sk);
3340 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3341 }
3342
3343 tcp_event_data_recv(sk, tp, skb);
3344
3345 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
3346 /* Well, only one small jumplet in fast path... */
3347 tcp_ack(sk, skb, FLAG_DATA);
3348 tcp_data_snd_check(sk);
3349 if (!tcp_ack_scheduled(tp))
3350 goto no_ack;
3351 }
3352
3353 if (eaten) {
3354 if (tcp_in_quickack_mode(tp)) {
3355 tcp_send_ack(sk);
3356 } else {
3357 tcp_send_delayed_ack(sk);
3358 }
3359 } else {
3360 __tcp_ack_snd_check(sk, 0);
3361 }
3362
3363 no_ack:
3364 if (eaten)
3365 __kfree_skb(skb);
3366 else
3367 sk->data_ready(sk, 0);
3368 return 0;
3369 }
3370 }
3371
3372 slow_path:
3373 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
3374 goto csum_error;
3375
3376 /*
3377 * RFC1323: H1. Apply PAWS check first.
3378 */
3379 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
3380 tcp_paws_discard(tp, skb)) {
3381 if (!th->rst) {
3382 NET_INC_STATS_BH(PAWSEstabRejected);
3383 tcp_send_dupack(sk, skb);
3384 goto discard;
3385 }
3386 /* Resets are accepted even if PAWS failed.
3387
3388 ts_recent update must be made after we are sure
3389 that the packet is in window.
3390 */
3391 }
3392
3393 /*
3394 * Standard slow path.
3395 */
3396
3397 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3398 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3399 * (RST) segments are validated by checking their SEQ-fields."
3400 * And page 69: "If an incoming segment is not acceptable,
3401 * an acknowledgment should be sent in reply (unless the RST bit
3402 * is set, if so drop the segment and return)".
3403 */
3404 if (!th->rst)
3405 tcp_send_dupack(sk, skb);
3406 goto discard;
3407 }
3408
3409 if(th->rst) {
3410 tcp_reset(sk);
3411 goto discard;
3412 }
3413
3414 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3415
3416 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3417 TCP_INC_STATS_BH(TcpInErrs);
3418 NET_INC_STATS_BH(TCPAbortOnSyn);
3419 tcp_reset(sk);
3420 return 1;
3421 }
3422
3423 step5:
3424 if(th->ack)
3425 tcp_ack(sk, skb, FLAG_SLOWPATH);
3426
3427 /* Process urgent data. */
3428 tcp_urg(sk, skb, th);
3429
3430 /* step 7: process the segment text */
3431 tcp_data_queue(sk, skb);
3432
3433 tcp_data_snd_check(sk);
3434 tcp_ack_snd_check(sk);
3435 return 0;
3436
3437 csum_error:
3438 TCP_INC_STATS_BH(TcpInErrs);
3439
3440 discard:
3441 __kfree_skb(skb);
3442 return 0;
3443 }
3444
3445 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
3446 struct tcphdr *th, unsigned len)
3447 {
3448 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3449 int saved_clamp = tp->mss_clamp;
3450
3451 tcp_parse_options(skb, tp, 0);
3452
3453 if (th->ack) {
3454 /* rfc793:
3455 * "If the state is SYN-SENT then
3456 * first check the ACK bit
3457 * If the ACK bit is set
3458 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3459 * a reset (unless the RST bit is set, if so drop
3460 * the segment and return)"
3461 *
3462 * We do not send data with SYN, so that RFC-correct
3463 * test reduces to:
3464 */
3465 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
3466 goto reset_and_undo;
3467
3468 if (tp->saw_tstamp && tp->rcv_tsecr &&
3469 !between(tp->rcv_tsecr, tp->retrans_stamp, tcp_time_stamp)) {
3470 NET_INC_STATS_BH(PAWSActiveRejected);
3471 goto reset_and_undo;
3472 }
3473
3474 /* Now ACK is acceptable.
3475 *
3476 * "If the RST bit is set
3477 * If the ACK was acceptable then signal the user "error:
3478 * connection reset", drop the segment, enter CLOSED state,
3479 * delete TCB, and return."
3480 */
3481
3482 if (th->rst) {
3483 tcp_reset(sk);
3484 goto discard;
3485 }
3486
3487 /* rfc793:
3488 * "fifth, if neither of the SYN or RST bits is set then
3489 * drop the segment and return."
3490 *
3491 * See note below!
3492 * --ANK(990513)
3493 */
3494 if (!th->syn)
3495 goto discard_and_undo;
3496
3497 /* rfc793:
3498 * "If the SYN bit is on ...
3499 * are acceptable then ...
3500 * (our SYN has been ACKed), change the connection
3501 * state to ESTABLISHED..."
3502 */
3503
3504 TCP_ECN_rcv_synack(tp, th);
3505
3506 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
3507 tcp_ack(sk, skb, FLAG_SLOWPATH);
3508
3509 /* Ok.. it's good. Set up sequence numbers and
3510 * move to established.
3511 */
3512 tp->rcv_nxt = TCP_SKB_CB(skb)->seq+1;
3513 tp->rcv_wup = TCP_SKB_CB(skb)->seq+1;
3514
3515 /* RFC1323: The window in SYN & SYN/ACK segments is
3516 * never scaled.
3517 */
3518 tp->snd_wnd = ntohs(th->window);
3519 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
3520
3521 if (tp->wscale_ok == 0) {
3522 tp->snd_wscale = tp->rcv_wscale = 0;
3523 tp->window_clamp = min_t(u32, tp->window_clamp, 65535);
3524 }
3525
3526 if (tp->saw_tstamp) {
3527 tp->tstamp_ok = 1;
3528 tp->tcp_header_len =
3529 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
3530 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
3531 tcp_store_ts_recent(tp);
3532 } else {
3533 tp->tcp_header_len = sizeof(struct tcphdr);
3534 }
3535
3536 if (tp->sack_ok && sysctl_tcp_fack)
3537 tp->sack_ok |= 2;
3538
3539 tcp_sync_mss(sk, tp->pmtu_cookie);
3540 tcp_initialize_rcv_mss(sk);
3541 tcp_init_metrics(sk);
3542 tcp_init_buffer_space(sk);
3543
3544 if (sk->keepopen)
3545 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
3546
3547 if (tp->snd_wscale == 0)
3548 __tcp_fast_path_on(tp, tp->snd_wnd);
3549 else
3550 tp->pred_flags = 0;
3551
3552 /* Remember, tcp_poll() does not lock socket!
3553 * Change state from SYN-SENT only after copied_seq
3554 * is initialized. */
3555 tp->copied_seq = tp->rcv_nxt;
3556 mb();
3557 tcp_set_state(sk, TCP_ESTABLISHED);
3558
3559 if(!sk->dead) {
3560 sk->state_change(sk);
3561 sk_wake_async(sk, 0, POLL_OUT);
3562 }
3563
3564 if (tp->write_pending || tp->defer_accept || tp->ack.pingpong) {
3565 /* Save one ACK. Data will be ready after
3566 * several ticks, if write_pending is set.
3567 *
3568 * It may be deleted, but with this feature tcpdumps
3569 * look so _wonderfully_ clever, that I was not able
3570 * to stand against the temptation 8) --ANK
3571 */
3572 tcp_schedule_ack(tp);
3573 tp->ack.lrcvtime = tcp_time_stamp;
3574 tp->ack.ato = TCP_ATO_MIN;
3575 tcp_incr_quickack(tp);
3576 tcp_enter_quickack_mode(tp);
3577 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
3578
3579 discard:
3580 __kfree_skb(skb);
3581 return 0;
3582 } else {
3583 tcp_send_ack(sk);
3584 }
3585 return -1;
3586 }
3587
3588 /* No ACK in the segment */
3589
3590 if (th->rst) {
3591 /* rfc793:
3592 * "If the RST bit is set
3593 *
3594 * Otherwise (no ACK) drop the segment and return."
3595 */
3596
3597 goto discard_and_undo;
3598 }
3599
3600 /* PAWS check. */
3601 if (tp->ts_recent_stamp && tp->saw_tstamp && tcp_paws_check(tp, 0))
3602 goto discard_and_undo;
3603
3604 if (th->syn) {
3605 /* We see SYN without ACK. It is attempt of
3606 * simultaneous connect with crossed SYNs.
3607 * Particularly, it can be connect to self.
3608 */
3609 tcp_set_state(sk, TCP_SYN_RECV);
3610
3611 if (tp->saw_tstamp) {
3612 tp->tstamp_ok = 1;
3613 tcp_store_ts_recent(tp);
3614 tp->tcp_header_len =
3615 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
3616 } else {
3617 tp->tcp_header_len = sizeof(struct tcphdr);
3618 }
3619
3620 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
3621 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
3622
3623 /* RFC1323: The window in SYN & SYN/ACK segments is
3624 * never scaled.
3625 */
3626 tp->snd_wnd = ntohs(th->window);
3627 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
3628 tp->max_window = tp->snd_wnd;
3629
3630 tcp_sync_mss(sk, tp->pmtu_cookie);
3631 tcp_initialize_rcv_mss(sk);
3632
3633 TCP_ECN_rcv_syn(tp, th);
3634
3635 tcp_send_synack(sk);
3636 #if 0
3637 /* Note, we could accept data and URG from this segment.
3638 * There are no obstacles to make this.
3639 *
3640 * However, if we ignore data in ACKless segments sometimes,
3641 * we have no reasons to accept it sometimes.
3642 * Also, seems the code doing it in step6 of tcp_rcv_state_process
3643 * is not flawless. So, discard packet for sanity.
3644 * Uncomment this return to process the data.
3645 */
3646 return -1;
3647 #else
3648 goto discard;
3649 #endif
3650 }
3651 /* "fifth, if neither of the SYN or RST bits is set then
3652 * drop the segment and return."
3653 */
3654
3655 discard_and_undo:
3656 tcp_clear_options(tp);
3657 tp->mss_clamp = saved_clamp;
3658 goto discard;
3659
3660 reset_and_undo:
3661 tcp_clear_options(tp);
3662 tp->mss_clamp = saved_clamp;
3663 return 1;
3664 }
3665
3666
3667 /*
3668 * This function implements the receiving procedure of RFC 793 for
3669 * all states except ESTABLISHED and TIME_WAIT.
3670 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
3671 * address independent.
3672 */
3673
3674 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
3675 struct tcphdr *th, unsigned len)
3676 {
3677 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
3678 int queued = 0;
3679
3680 tp->saw_tstamp = 0;
3681
3682 switch (sk->state) {
3683 case TCP_CLOSE:
3684 goto discard;
3685
3686 case TCP_LISTEN:
3687 if(th->ack)
3688 return 1;
3689
3690 if(th->syn) {
3691 if(tp->af_specific->conn_request(sk, skb) < 0)
3692 return 1;
3693
3694 /* Now we have several options: In theory there is
3695 * nothing else in the frame. KA9Q has an option to
3696 * send data with the syn, BSD accepts data with the
3697 * syn up to the [to be] advertised window and
3698 * Solaris 2.1 gives you a protocol error. For now
3699 * we just ignore it, that fits the spec precisely
3700 * and avoids incompatibilities. It would be nice in
3701 * future to drop through and process the data.
3702 *
3703 * Now that TTCP is starting to be used we ought to
3704 * queue this data.
3705 * But, this leaves one open to an easy denial of
3706 * service attack, and SYN cookies can't defend
3707 * against this problem. So, we drop the data
3708 * in the interest of security over speed.
3709 */
3710 goto discard;
3711 }
3712 goto discard;
3713
3714 case TCP_SYN_SENT:
3715 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
3716 if (queued >= 0)
3717 return queued;
3718
3719 /* Do step6 onward by hand. */
3720 tcp_urg(sk, skb, th);
3721 __kfree_skb(skb);
3722 tcp_data_snd_check(sk);
3723 return 0;
3724 }
3725
3726 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
3727 tcp_paws_discard(tp, skb)) {
3728 if (!th->rst) {
3729 NET_INC_STATS_BH(PAWSEstabRejected);
3730 tcp_send_dupack(sk, skb);
3731 goto discard;
3732 }
3733 /* Reset is accepted even if it did not pass PAWS. */
3734 }
3735
3736 /* step 1: check sequence number */
3737 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3738 if (!th->rst)
3739 tcp_send_dupack(sk, skb);
3740 goto discard;
3741 }
3742
3743 /* step 2: check RST bit */
3744 if(th->rst) {
3745 tcp_reset(sk);
3746 goto discard;
3747 }
3748
3749 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3750
3751 /* step 3: check security and precedence [ignored] */
3752
3753 /* step 4:
3754 *
3755 * Check for a SYN in window.
3756 */
3757 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3758 NET_INC_STATS_BH(TCPAbortOnSyn);
3759 tcp_reset(sk);
3760 return 1;
3761 }
3762
3763 /* step 5: check the ACK field */
3764 if (th->ack) {
3765 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
3766
3767 switch(sk->state) {
3768 case TCP_SYN_RECV:
3769 if (acceptable) {
3770 tp->copied_seq = tp->rcv_nxt;
3771 mb();
3772 tcp_set_state(sk, TCP_ESTABLISHED);
3773 sk->state_change(sk);
3774
3775 /* Note, that this wakeup is only for marginal
3776 * crossed SYN case. Passively open sockets
3777 * are not waked up, because sk->sleep == NULL
3778 * and sk->socket == NULL.
3779 */
3780 if (sk->socket) {
3781 sk_wake_async(sk,0,POLL_OUT);
3782 }
3783
3784 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
3785 tp->snd_wnd = ntohs(th->window) << tp->snd_wscale;
3786 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
3787
3788 /* tcp_ack considers this ACK as duplicate
3789 * and does not calculate rtt.
3790 * Fix it at least with timestamps.
3791 */
3792 if (tp->saw_tstamp && tp->rcv_tsecr && !tp->srtt)
3793 tcp_ack_saw_tstamp(tp, 0);
3794
3795 if (tp->tstamp_ok)
3796 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
3797
3798 tcp_init_metrics(sk);
3799 tcp_initialize_rcv_mss(sk);
3800 tcp_init_buffer_space(sk);
3801 tcp_fast_path_on(tp);
3802 } else {
3803 return 1;
3804 }
3805 break;
3806
3807 case TCP_FIN_WAIT1:
3808 if (tp->snd_una == tp->write_seq) {
3809 tcp_set_state(sk, TCP_FIN_WAIT2);
3810 sk->shutdown |= SEND_SHUTDOWN;
3811 dst_confirm(sk->dst_cache);
3812
3813 if (!sk->dead) {
3814 /* Wake up lingering close() */
3815 sk->state_change(sk);
3816 } else {
3817 int tmo;
3818
3819 if (tp->linger2 < 0 ||
3820 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3821 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
3822 tcp_done(sk);
3823 NET_INC_STATS_BH(TCPAbortOnData);
3824 return 1;
3825 }
3826
3827 tmo = tcp_fin_time(tp);
3828 if (tmo > TCP_TIMEWAIT_LEN) {
3829 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
3830 } else if (th->fin || sk->lock.users) {
3831 /* Bad case. We could lose such FIN otherwise.
3832 * It is not a big problem, but it looks confusing
3833 * and not so rare event. We still can lose it now,
3834 * if it spins in bh_lock_sock(), but it is really
3835 * marginal case.
3836 */
3837 tcp_reset_keepalive_timer(sk, tmo);
3838 } else {
3839 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
3840 goto discard;
3841 }
3842 }
3843 }
3844 break;
3845
3846 case TCP_CLOSING:
3847 if (tp->snd_una == tp->write_seq) {
3848 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3849 goto discard;
3850 }
3851 break;
3852
3853 case TCP_LAST_ACK:
3854 if (tp->snd_una == tp->write_seq) {
3855 tcp_update_metrics(sk);
3856 tcp_done(sk);
3857 goto discard;
3858 }
3859 break;
3860 }
3861 } else
3862 goto discard;
3863
3864 /* step 6: check the URG bit */
3865 tcp_urg(sk, skb, th);
3866
3867 /* step 7: process the segment text */
3868 switch (sk->state) {
3869 case TCP_CLOSE_WAIT:
3870 case TCP_CLOSING:
3871 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3872 break;
3873 case TCP_FIN_WAIT1:
3874 case TCP_FIN_WAIT2:
3875 /* RFC 793 says to queue data in these states,
3876 * RFC 1122 says we MUST send a reset.
3877 * BSD 4.4 also does reset.
3878 */
3879 if (sk->shutdown & RCV_SHUTDOWN) {
3880 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3881 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
3882 NET_INC_STATS_BH(TCPAbortOnData);
3883 tcp_reset(sk);
3884 return 1;
3885 }
3886 }
3887 /* Fall through */
3888 case TCP_ESTABLISHED:
3889 tcp_data_queue(sk, skb);
3890 queued = 1;
3891 break;
3892 }
3893
3894 /* tcp_data could move socket to TIME-WAIT */
3895 if (sk->state != TCP_CLOSE) {
3896 tcp_data_snd_check(sk);
3897 tcp_ack_snd_check(sk);
3898 }
3899
3900 if (!queued) {
3901 discard:
3902 __kfree_skb(skb);
3903 }
3904 return 0;
3905 }
3906