94fe5b1f9dcbeb1b819da503b8d5a09d0c35569c
[linux-2.6.git] / 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 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
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 presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make 
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks. 
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/config.h>
67 #include <linux/mm.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <net/tcp.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
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 int sysctl_tcp_ecn;
82 int sysctl_tcp_dsack = 1;
83 int sysctl_tcp_app_win = 31;
84 int sysctl_tcp_adv_win_scale = 2;
85
86 int sysctl_tcp_stdurg;
87 int sysctl_tcp_rfc1337;
88 int sysctl_tcp_max_orphans = NR_FILE;
89 int sysctl_tcp_frto;
90 int sysctl_tcp_nometrics_save;
91
92 int sysctl_tcp_moderate_rcvbuf = 1;
93 int sysctl_tcp_abc = 1;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104
105 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
106 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
107 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
108 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
109
110 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
111 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
112 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
113
114 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
115
116 /* Adapt the MSS value used to make delayed ack decision to the 
117  * real world.
118  */ 
119 static void tcp_measure_rcv_mss(struct sock *sk,
120                                 const struct sk_buff *skb)
121 {
122         struct inet_connection_sock *icsk = inet_csk(sk);
123         const unsigned int lss = icsk->icsk_ack.last_seg_size; 
124         unsigned int len;
125
126         icsk->icsk_ack.last_seg_size = 0; 
127
128         /* skb->len may jitter because of SACKs, even if peer
129          * sends good full-sized frames.
130          */
131         len = skb->len;
132         if (len >= icsk->icsk_ack.rcv_mss) {
133                 icsk->icsk_ack.rcv_mss = len;
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 -= tcp_sk(sk)->tcp_header_len;
154                         icsk->icsk_ack.last_seg_size = len;
155                         if (len == lss) {
156                                 icsk->icsk_ack.rcv_mss = len;
157                                 return;
158                         }
159                 }
160                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
161         }
162 }
163
164 static void tcp_incr_quickack(struct sock *sk)
165 {
166         struct inet_connection_sock *icsk = inet_csk(sk);
167         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
168
169         if (quickacks==0)
170                 quickacks=2;
171         if (quickacks > icsk->icsk_ack.quick)
172                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
173 }
174
175 void tcp_enter_quickack_mode(struct sock *sk)
176 {
177         struct inet_connection_sock *icsk = inet_csk(sk);
178         tcp_incr_quickack(sk);
179         icsk->icsk_ack.pingpong = 0;
180         icsk->icsk_ack.ato = TCP_ATO_MIN;
181 }
182
183 /* Send ACKs quickly, if "quick" count is not exhausted
184  * and the session is not interactive.
185  */
186
187 static inline int tcp_in_quickack_mode(const struct sock *sk)
188 {
189         const struct inet_connection_sock *icsk = inet_csk(sk);
190         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
191 }
192
193 /* Buffer size and advertised window tuning.
194  *
195  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
196  */
197
198 static void tcp_fixup_sndbuf(struct sock *sk)
199 {
200         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
201                      sizeof(struct sk_buff);
202
203         if (sk->sk_sndbuf < 3 * sndmem)
204                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
205 }
206
207 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
208  *
209  * All tcp_full_space() is split to two parts: "network" buffer, allocated
210  * forward and advertised in receiver window (tp->rcv_wnd) and
211  * "application buffer", required to isolate scheduling/application
212  * latencies from network.
213  * window_clamp is maximal advertised window. It can be less than
214  * tcp_full_space(), in this case tcp_full_space() - window_clamp
215  * is reserved for "application" buffer. The less window_clamp is
216  * the smoother our behaviour from viewpoint of network, but the lower
217  * throughput and the higher sensitivity of the connection to losses. 8)
218  *
219  * rcv_ssthresh is more strict window_clamp used at "slow start"
220  * phase to predict further behaviour of this connection.
221  * It is used for two goals:
222  * - to enforce header prediction at sender, even when application
223  *   requires some significant "application buffer". It is check #1.
224  * - to prevent pruning of receive queue because of misprediction
225  *   of receiver window. Check #2.
226  *
227  * The scheme does not work when sender sends good segments opening
228  * window and then starts to feed us spaghetti. But it should work
229  * in common situations. Otherwise, we have to rely on queue collapsing.
230  */
231
232 /* Slow part of check#2. */
233 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
234                              const struct sk_buff *skb)
235 {
236         /* Optimize this! */
237         int truesize = tcp_win_from_space(skb->truesize)/2;
238         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
239
240         while (tp->rcv_ssthresh <= window) {
241                 if (truesize <= skb->len)
242                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
243
244                 truesize >>= 1;
245                 window >>= 1;
246         }
247         return 0;
248 }
249
250 static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
251                             struct sk_buff *skb)
252 {
253         /* Check #1 */
254         if (tp->rcv_ssthresh < tp->window_clamp &&
255             (int)tp->rcv_ssthresh < tcp_space(sk) &&
256             !tcp_memory_pressure) {
257                 int incr;
258
259                 /* Check #2. Increase window, if skb with such overhead
260                  * will fit to rcvbuf in future.
261                  */
262                 if (tcp_win_from_space(skb->truesize) <= skb->len)
263                         incr = 2*tp->advmss;
264                 else
265                         incr = __tcp_grow_window(sk, tp, skb);
266
267                 if (incr) {
268                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
269                         inet_csk(sk)->icsk_ack.quick |= 1;
270                 }
271         }
272 }
273
274 /* 3. Tuning rcvbuf, when connection enters established state. */
275
276 static void tcp_fixup_rcvbuf(struct sock *sk)
277 {
278         struct tcp_sock *tp = tcp_sk(sk);
279         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
280
281         /* Try to select rcvbuf so that 4 mss-sized segments
282          * will fit to window and corresponding skbs will fit to our rcvbuf.
283          * (was 3; 4 is minimum to allow fast retransmit to work.)
284          */
285         while (tcp_win_from_space(rcvmem) < tp->advmss)
286                 rcvmem += 128;
287         if (sk->sk_rcvbuf < 4 * rcvmem)
288                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
289 }
290
291 /* 4. Try to fixup all. It is made immediately after connection enters
292  *    established state.
293  */
294 static void tcp_init_buffer_space(struct sock *sk)
295 {
296         struct tcp_sock *tp = tcp_sk(sk);
297         int maxwin;
298
299         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
300                 tcp_fixup_rcvbuf(sk);
301         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
302                 tcp_fixup_sndbuf(sk);
303
304         tp->rcvq_space.space = tp->rcv_wnd;
305
306         maxwin = tcp_full_space(sk);
307
308         if (tp->window_clamp >= maxwin) {
309                 tp->window_clamp = maxwin;
310
311                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
312                         tp->window_clamp = max(maxwin -
313                                                (maxwin >> sysctl_tcp_app_win),
314                                                4 * tp->advmss);
315         }
316
317         /* Force reservation of one segment. */
318         if (sysctl_tcp_app_win &&
319             tp->window_clamp > 2 * tp->advmss &&
320             tp->window_clamp + tp->advmss > maxwin)
321                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
322
323         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
324         tp->snd_cwnd_stamp = tcp_time_stamp;
325 }
326
327 /* 5. Recalculate window clamp after socket hit its memory bounds. */
328 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
329 {
330         struct inet_connection_sock *icsk = inet_csk(sk);
331
332         icsk->icsk_ack.quick = 0;
333
334         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
335             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
336             !tcp_memory_pressure &&
337             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
338                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
339                                     sysctl_tcp_rmem[2]);
340         }
341         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
342                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
343 }
344
345
346 /* Initialize RCV_MSS value.
347  * RCV_MSS is an our guess about MSS used by the peer.
348  * We haven't any direct information about the MSS.
349  * It's better to underestimate the RCV_MSS rather than overestimate.
350  * Overestimations make us ACKing less frequently than needed.
351  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
352  */
353 void tcp_initialize_rcv_mss(struct sock *sk)
354 {
355         struct tcp_sock *tp = tcp_sk(sk);
356         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
357
358         hint = min(hint, tp->rcv_wnd/2);
359         hint = min(hint, TCP_MIN_RCVMSS);
360         hint = max(hint, TCP_MIN_MSS);
361
362         inet_csk(sk)->icsk_ack.rcv_mss = hint;
363 }
364
365 /* Receiver "autotuning" code.
366  *
367  * The algorithm for RTT estimation w/o timestamps is based on
368  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
369  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
370  *
371  * More detail on this code can be found at
372  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
373  * though this reference is out of date.  A new paper
374  * is pending.
375  */
376 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
377 {
378         u32 new_sample = tp->rcv_rtt_est.rtt;
379         long m = sample;
380
381         if (m == 0)
382                 m = 1;
383
384         if (new_sample != 0) {
385                 /* If we sample in larger samples in the non-timestamp
386                  * case, we could grossly overestimate the RTT especially
387                  * with chatty applications or bulk transfer apps which
388                  * are stalled on filesystem I/O.
389                  *
390                  * Also, since we are only going for a minimum in the
391                  * non-timestamp case, we do not smooth things out
392                  * else with timestamps disabled convergence takes too
393                  * long.
394                  */
395                 if (!win_dep) {
396                         m -= (new_sample >> 3);
397                         new_sample += m;
398                 } else if (m < new_sample)
399                         new_sample = m << 3;
400         } else {
401                 /* No previous measure. */
402                 new_sample = m << 3;
403         }
404
405         if (tp->rcv_rtt_est.rtt != new_sample)
406                 tp->rcv_rtt_est.rtt = new_sample;
407 }
408
409 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
410 {
411         if (tp->rcv_rtt_est.time == 0)
412                 goto new_measure;
413         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
414                 return;
415         tcp_rcv_rtt_update(tp,
416                            jiffies - tp->rcv_rtt_est.time,
417                            1);
418
419 new_measure:
420         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
421         tp->rcv_rtt_est.time = tcp_time_stamp;
422 }
423
424 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
425 {
426         struct tcp_sock *tp = tcp_sk(sk);
427         if (tp->rx_opt.rcv_tsecr &&
428             (TCP_SKB_CB(skb)->end_seq -
429              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
430                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
431 }
432
433 /*
434  * This function should be called every time data is copied to user space.
435  * It calculates the appropriate TCP receive buffer space.
436  */
437 void tcp_rcv_space_adjust(struct sock *sk)
438 {
439         struct tcp_sock *tp = tcp_sk(sk);
440         int time;
441         int space;
442         
443         if (tp->rcvq_space.time == 0)
444                 goto new_measure;
445         
446         time = tcp_time_stamp - tp->rcvq_space.time;
447         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
448             tp->rcv_rtt_est.rtt == 0)
449                 return;
450         
451         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
452
453         space = max(tp->rcvq_space.space, space);
454
455         if (tp->rcvq_space.space != space) {
456                 int rcvmem;
457
458                 tp->rcvq_space.space = space;
459
460                 if (sysctl_tcp_moderate_rcvbuf &&
461                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
462                         int new_clamp = space;
463
464                         /* Receive space grows, normalize in order to
465                          * take into account packet headers and sk_buff
466                          * structure overhead.
467                          */
468                         space /= tp->advmss;
469                         if (!space)
470                                 space = 1;
471                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
472                                   16 + sizeof(struct sk_buff));
473                         while (tcp_win_from_space(rcvmem) < tp->advmss)
474                                 rcvmem += 128;
475                         space *= rcvmem;
476                         space = min(space, sysctl_tcp_rmem[2]);
477                         if (space > sk->sk_rcvbuf) {
478                                 sk->sk_rcvbuf = space;
479
480                                 /* Make the window clamp follow along.  */
481                                 tp->window_clamp = new_clamp;
482                         }
483                 }
484         }
485         
486 new_measure:
487         tp->rcvq_space.seq = tp->copied_seq;
488         tp->rcvq_space.time = tcp_time_stamp;
489 }
490
491 /* There is something which you must keep in mind when you analyze the
492  * behavior of the tp->ato delayed ack timeout interval.  When a
493  * connection starts up, we want to ack as quickly as possible.  The
494  * problem is that "good" TCP's do slow start at the beginning of data
495  * transmission.  The means that until we send the first few ACK's the
496  * sender will sit on his end and only queue most of his data, because
497  * he can only send snd_cwnd unacked packets at any given time.  For
498  * each ACK we send, he increments snd_cwnd and transmits more of his
499  * queue.  -DaveM
500  */
501 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
502 {
503         struct inet_connection_sock *icsk = inet_csk(sk);
504         u32 now;
505
506         inet_csk_schedule_ack(sk);
507
508         tcp_measure_rcv_mss(sk, skb);
509
510         tcp_rcv_rtt_measure(tp);
511         
512         now = tcp_time_stamp;
513
514         if (!icsk->icsk_ack.ato) {
515                 /* The _first_ data packet received, initialize
516                  * delayed ACK engine.
517                  */
518                 tcp_incr_quickack(sk);
519                 icsk->icsk_ack.ato = TCP_ATO_MIN;
520         } else {
521                 int m = now - icsk->icsk_ack.lrcvtime;
522
523                 if (m <= TCP_ATO_MIN/2) {
524                         /* The fastest case is the first. */
525                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
526                 } else if (m < icsk->icsk_ack.ato) {
527                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
528                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
529                                 icsk->icsk_ack.ato = icsk->icsk_rto;
530                 } else if (m > icsk->icsk_rto) {
531                         /* Too long gap. Apparently sender failed to
532                          * restart window, so that we send ACKs quickly.
533                          */
534                         tcp_incr_quickack(sk);
535                         sk_stream_mem_reclaim(sk);
536                 }
537         }
538         icsk->icsk_ack.lrcvtime = now;
539
540         TCP_ECN_check_ce(tp, skb);
541
542         if (skb->len >= 128)
543                 tcp_grow_window(sk, tp, skb);
544 }
545
546 /* Called to compute a smoothed rtt estimate. The data fed to this
547  * routine either comes from timestamps, or from segments that were
548  * known _not_ to have been retransmitted [see Karn/Partridge
549  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
550  * piece by Van Jacobson.
551  * NOTE: the next three routines used to be one big routine.
552  * To save cycles in the RFC 1323 implementation it was better to break
553  * it up into three procedures. -- erics
554  */
555 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
556 {
557         struct tcp_sock *tp = tcp_sk(sk);
558         long m = mrtt; /* RTT */
559
560         /*      The following amusing code comes from Jacobson's
561          *      article in SIGCOMM '88.  Note that rtt and mdev
562          *      are scaled versions of rtt and mean deviation.
563          *      This is designed to be as fast as possible 
564          *      m stands for "measurement".
565          *
566          *      On a 1990 paper the rto value is changed to:
567          *      RTO = rtt + 4 * mdev
568          *
569          * Funny. This algorithm seems to be very broken.
570          * These formulae increase RTO, when it should be decreased, increase
571          * too slowly, when it should be increased quickly, decrease too quickly
572          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
573          * does not matter how to _calculate_ it. Seems, it was trap
574          * that VJ failed to avoid. 8)
575          */
576         if(m == 0)
577                 m = 1;
578         if (tp->srtt != 0) {
579                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
580                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
581                 if (m < 0) {
582                         m = -m;         /* m is now abs(error) */
583                         m -= (tp->mdev >> 2);   /* similar update on mdev */
584                         /* This is similar to one of Eifel findings.
585                          * Eifel blocks mdev updates when rtt decreases.
586                          * This solution is a bit different: we use finer gain
587                          * for mdev in this case (alpha*beta).
588                          * Like Eifel it also prevents growth of rto,
589                          * but also it limits too fast rto decreases,
590                          * happening in pure Eifel.
591                          */
592                         if (m > 0)
593                                 m >>= 3;
594                 } else {
595                         m -= (tp->mdev >> 2);   /* similar update on mdev */
596                 }
597                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
598                 if (tp->mdev > tp->mdev_max) {
599                         tp->mdev_max = tp->mdev;
600                         if (tp->mdev_max > tp->rttvar)
601                                 tp->rttvar = tp->mdev_max;
602                 }
603                 if (after(tp->snd_una, tp->rtt_seq)) {
604                         if (tp->mdev_max < tp->rttvar)
605                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
606                         tp->rtt_seq = tp->snd_nxt;
607                         tp->mdev_max = TCP_RTO_MIN;
608                 }
609         } else {
610                 /* no previous measure. */
611                 tp->srtt = m<<3;        /* take the measured time to be rtt */
612                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
613                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
614                 tp->rtt_seq = tp->snd_nxt;
615         }
616 }
617
618 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
619  * routine referred to above.
620  */
621 static inline void tcp_set_rto(struct sock *sk)
622 {
623         const struct tcp_sock *tp = tcp_sk(sk);
624         /* Old crap is replaced with new one. 8)
625          *
626          * More seriously:
627          * 1. If rtt variance happened to be less 50msec, it is hallucination.
628          *    It cannot be less due to utterly erratic ACK generation made
629          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
630          *    to do with delayed acks, because at cwnd>2 true delack timeout
631          *    is invisible. Actually, Linux-2.4 also generates erratic
632          *    ACKs in some circumstances.
633          */
634         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
635
636         /* 2. Fixups made earlier cannot be right.
637          *    If we do not estimate RTO correctly without them,
638          *    all the algo is pure shit and should be replaced
639          *    with correct one. It is exactly, which we pretend to do.
640          */
641 }
642
643 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
644  * guarantees that rto is higher.
645  */
646 static inline void tcp_bound_rto(struct sock *sk)
647 {
648         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
649                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
650 }
651
652 /* Save metrics learned by this TCP session.
653    This function is called only, when TCP finishes successfully
654    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
655  */
656 void tcp_update_metrics(struct sock *sk)
657 {
658         struct tcp_sock *tp = tcp_sk(sk);
659         struct dst_entry *dst = __sk_dst_get(sk);
660
661         if (sysctl_tcp_nometrics_save)
662                 return;
663
664         dst_confirm(dst);
665
666         if (dst && (dst->flags&DST_HOST)) {
667                 const struct inet_connection_sock *icsk = inet_csk(sk);
668                 int m;
669
670                 if (icsk->icsk_backoff || !tp->srtt) {
671                         /* This session failed to estimate rtt. Why?
672                          * Probably, no packets returned in time.
673                          * Reset our results.
674                          */
675                         if (!(dst_metric_locked(dst, RTAX_RTT)))
676                                 dst->metrics[RTAX_RTT-1] = 0;
677                         return;
678                 }
679
680                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
681
682                 /* If newly calculated rtt larger than stored one,
683                  * store new one. Otherwise, use EWMA. Remember,
684                  * rtt overestimation is always better than underestimation.
685                  */
686                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
687                         if (m <= 0)
688                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
689                         else
690                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
691                 }
692
693                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
694                         if (m < 0)
695                                 m = -m;
696
697                         /* Scale deviation to rttvar fixed point */
698                         m >>= 1;
699                         if (m < tp->mdev)
700                                 m = tp->mdev;
701
702                         if (m >= dst_metric(dst, RTAX_RTTVAR))
703                                 dst->metrics[RTAX_RTTVAR-1] = m;
704                         else
705                                 dst->metrics[RTAX_RTTVAR-1] -=
706                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
707                 }
708
709                 if (tp->snd_ssthresh >= 0xFFFF) {
710                         /* Slow start still did not finish. */
711                         if (dst_metric(dst, RTAX_SSTHRESH) &&
712                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
713                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
714                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
715                         if (!dst_metric_locked(dst, RTAX_CWND) &&
716                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
717                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
718                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
719                            icsk->icsk_ca_state == TCP_CA_Open) {
720                         /* Cong. avoidance phase, cwnd is reliable. */
721                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
722                                 dst->metrics[RTAX_SSTHRESH-1] =
723                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
724                         if (!dst_metric_locked(dst, RTAX_CWND))
725                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
726                 } else {
727                         /* Else slow start did not finish, cwnd is non-sense,
728                            ssthresh may be also invalid.
729                          */
730                         if (!dst_metric_locked(dst, RTAX_CWND))
731                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
732                         if (dst->metrics[RTAX_SSTHRESH-1] &&
733                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
734                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
735                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
736                 }
737
738                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
739                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
740                             tp->reordering != sysctl_tcp_reordering)
741                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
742                 }
743         }
744 }
745
746 /* Numbers are taken from RFC2414.  */
747 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
748 {
749         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
750
751         if (!cwnd) {
752                 if (tp->mss_cache > 1460)
753                         cwnd = 2;
754                 else
755                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
756         }
757         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
758 }
759
760 /* Set slow start threshold and cwnd not falling to slow start */
761 void tcp_enter_cwr(struct sock *sk)
762 {
763         struct tcp_sock *tp = tcp_sk(sk);
764
765         tp->prior_ssthresh = 0;
766         tp->bytes_acked = 0;
767         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
768                 tp->undo_marker = 0;
769                 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
770                 tp->snd_cwnd = min(tp->snd_cwnd,
771                                    tcp_packets_in_flight(tp) + 1U);
772                 tp->snd_cwnd_cnt = 0;
773                 tp->high_seq = tp->snd_nxt;
774                 tp->snd_cwnd_stamp = tcp_time_stamp;
775                 TCP_ECN_queue_cwr(tp);
776
777                 tcp_set_ca_state(sk, TCP_CA_CWR);
778         }
779 }
780
781 /* Initialize metrics on socket. */
782
783 static void tcp_init_metrics(struct sock *sk)
784 {
785         struct tcp_sock *tp = tcp_sk(sk);
786         struct dst_entry *dst = __sk_dst_get(sk);
787
788         if (dst == NULL)
789                 goto reset;
790
791         dst_confirm(dst);
792
793         if (dst_metric_locked(dst, RTAX_CWND))
794                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
795         if (dst_metric(dst, RTAX_SSTHRESH)) {
796                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
797                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
798                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
799         }
800         if (dst_metric(dst, RTAX_REORDERING) &&
801             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
802                 tp->rx_opt.sack_ok &= ~2;
803                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
804         }
805
806         if (dst_metric(dst, RTAX_RTT) == 0)
807                 goto reset;
808
809         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
810                 goto reset;
811
812         /* Initial rtt is determined from SYN,SYN-ACK.
813          * The segment is small and rtt may appear much
814          * less than real one. Use per-dst memory
815          * to make it more realistic.
816          *
817          * A bit of theory. RTT is time passed after "normal" sized packet
818          * is sent until it is ACKed. In normal circumstances sending small
819          * packets force peer to delay ACKs and calculation is correct too.
820          * The algorithm is adaptive and, provided we follow specs, it
821          * NEVER underestimate RTT. BUT! If peer tries to make some clever
822          * tricks sort of "quick acks" for time long enough to decrease RTT
823          * to low value, and then abruptly stops to do it and starts to delay
824          * ACKs, wait for troubles.
825          */
826         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
827                 tp->srtt = dst_metric(dst, RTAX_RTT);
828                 tp->rtt_seq = tp->snd_nxt;
829         }
830         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
831                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
832                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
833         }
834         tcp_set_rto(sk);
835         tcp_bound_rto(sk);
836         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
837                 goto reset;
838         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
839         tp->snd_cwnd_stamp = tcp_time_stamp;
840         return;
841
842 reset:
843         /* Play conservative. If timestamps are not
844          * supported, TCP will fail to recalculate correct
845          * rtt, if initial rto is too small. FORGET ALL AND RESET!
846          */
847         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
848                 tp->srtt = 0;
849                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
850                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
851         }
852 }
853
854 static void tcp_update_reordering(struct sock *sk, const int metric,
855                                   const int ts)
856 {
857         struct tcp_sock *tp = tcp_sk(sk);
858         if (metric > tp->reordering) {
859                 tp->reordering = min(TCP_MAX_REORDERING, metric);
860
861                 /* This exciting event is worth to be remembered. 8) */
862                 if (ts)
863                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
864                 else if (IsReno(tp))
865                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
866                 else if (IsFack(tp))
867                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
868                 else
869                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
870 #if FASTRETRANS_DEBUG > 1
871                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
872                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
873                        tp->reordering,
874                        tp->fackets_out,
875                        tp->sacked_out,
876                        tp->undo_marker ? tp->undo_retrans : 0);
877 #endif
878                 /* Disable FACK yet. */
879                 tp->rx_opt.sack_ok &= ~2;
880         }
881 }
882
883 /* This procedure tags the retransmission queue when SACKs arrive.
884  *
885  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
886  * Packets in queue with these bits set are counted in variables
887  * sacked_out, retrans_out and lost_out, correspondingly.
888  *
889  * Valid combinations are:
890  * Tag  InFlight        Description
891  * 0    1               - orig segment is in flight.
892  * S    0               - nothing flies, orig reached receiver.
893  * L    0               - nothing flies, orig lost by net.
894  * R    2               - both orig and retransmit are in flight.
895  * L|R  1               - orig is lost, retransmit is in flight.
896  * S|R  1               - orig reached receiver, retrans is still in flight.
897  * (L|S|R is logically valid, it could occur when L|R is sacked,
898  *  but it is equivalent to plain S and code short-curcuits it to S.
899  *  L|S is logically invalid, it would mean -1 packet in flight 8))
900  *
901  * These 6 states form finite state machine, controlled by the following events:
902  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
903  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
904  * 3. Loss detection event of one of three flavors:
905  *      A. Scoreboard estimator decided the packet is lost.
906  *         A'. Reno "three dupacks" marks head of queue lost.
907  *         A''. Its FACK modfication, head until snd.fack is lost.
908  *      B. SACK arrives sacking data transmitted after never retransmitted
909  *         hole was sent out.
910  *      C. SACK arrives sacking SND.NXT at the moment, when the
911  *         segment was retransmitted.
912  * 4. D-SACK added new rule: D-SACK changes any tag to S.
913  *
914  * It is pleasant to note, that state diagram turns out to be commutative,
915  * so that we are allowed not to be bothered by order of our actions,
916  * when multiple events arrive simultaneously. (see the function below).
917  *
918  * Reordering detection.
919  * --------------------
920  * Reordering metric is maximal distance, which a packet can be displaced
921  * in packet stream. With SACKs we can estimate it:
922  *
923  * 1. SACK fills old hole and the corresponding segment was not
924  *    ever retransmitted -> reordering. Alas, we cannot use it
925  *    when segment was retransmitted.
926  * 2. The last flaw is solved with D-SACK. D-SACK arrives
927  *    for retransmitted and already SACKed segment -> reordering..
928  * Both of these heuristics are not used in Loss state, when we cannot
929  * account for retransmits accurately.
930  */
931 static int
932 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
933 {
934         const struct inet_connection_sock *icsk = inet_csk(sk);
935         struct tcp_sock *tp = tcp_sk(sk);
936         unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
937         struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
938         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
939         int reord = tp->packets_out;
940         int prior_fackets;
941         u32 lost_retrans = 0;
942         int flag = 0;
943         int dup_sack = 0;
944         int i;
945
946         if (!tp->sacked_out)
947                 tp->fackets_out = 0;
948         prior_fackets = tp->fackets_out;
949
950         /* SACK fastpath:
951          * if the only SACK change is the increase of the end_seq of
952          * the first block then only apply that SACK block
953          * and use retrans queue hinting otherwise slowpath */
954         flag = 1;
955         for (i = 0; i< num_sacks; i++) {
956                 __u32 start_seq = ntohl(sp[i].start_seq);
957                 __u32 end_seq =  ntohl(sp[i].end_seq);
958
959                 if (i == 0){
960                         if (tp->recv_sack_cache[i].start_seq != start_seq)
961                                 flag = 0;
962                 } else {
963                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
964                             (tp->recv_sack_cache[i].end_seq != end_seq))
965                                 flag = 0;
966                 }
967                 tp->recv_sack_cache[i].start_seq = start_seq;
968                 tp->recv_sack_cache[i].end_seq = end_seq;
969
970                 /* Check for D-SACK. */
971                 if (i == 0) {
972                         u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
973
974                         if (before(start_seq, ack)) {
975                                 dup_sack = 1;
976                                 tp->rx_opt.sack_ok |= 4;
977                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
978                         } else if (num_sacks > 1 &&
979                                    !after(end_seq, ntohl(sp[1].end_seq)) &&
980                                    !before(start_seq, ntohl(sp[1].start_seq))) {
981                                 dup_sack = 1;
982                                 tp->rx_opt.sack_ok |= 4;
983                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
984                         }
985
986                         /* D-SACK for already forgotten data...
987                          * Do dumb counting. */
988                         if (dup_sack &&
989                             !after(end_seq, prior_snd_una) &&
990                             after(end_seq, tp->undo_marker))
991                                 tp->undo_retrans--;
992
993                         /* Eliminate too old ACKs, but take into
994                          * account more or less fresh ones, they can
995                          * contain valid SACK info.
996                          */
997                         if (before(ack, prior_snd_una - tp->max_window))
998                                 return 0;
999                 }
1000         }
1001
1002         if (flag)
1003                 num_sacks = 1;
1004         else {
1005                 int j;
1006                 tp->fastpath_skb_hint = NULL;
1007
1008                 /* order SACK blocks to allow in order walk of the retrans queue */
1009                 for (i = num_sacks-1; i > 0; i--) {
1010                         for (j = 0; j < i; j++){
1011                                 if (after(ntohl(sp[j].start_seq),
1012                                           ntohl(sp[j+1].start_seq))){
1013                                         sp[j].start_seq = htonl(tp->recv_sack_cache[j+1].start_seq);
1014                                         sp[j].end_seq = htonl(tp->recv_sack_cache[j+1].end_seq);
1015                                         sp[j+1].start_seq = htonl(tp->recv_sack_cache[j].start_seq);
1016                                         sp[j+1].end_seq = htonl(tp->recv_sack_cache[j].end_seq);
1017                                 }
1018
1019                         }
1020                 }
1021         }
1022
1023         /* clear flag as used for different purpose in following code */
1024         flag = 0;
1025
1026         for (i=0; i<num_sacks; i++, sp++) {
1027                 struct sk_buff *skb;
1028                 __u32 start_seq = ntohl(sp->start_seq);
1029                 __u32 end_seq = ntohl(sp->end_seq);
1030                 int fack_count;
1031
1032                 /* Use SACK fastpath hint if valid */
1033                 if (tp->fastpath_skb_hint) {
1034                         skb = tp->fastpath_skb_hint;
1035                         fack_count = tp->fastpath_cnt_hint;
1036                 } else {
1037                         skb = sk->sk_write_queue.next;
1038                         fack_count = 0;
1039                 }
1040
1041                 /* Event "B" in the comment above. */
1042                 if (after(end_seq, tp->high_seq))
1043                         flag |= FLAG_DATA_LOST;
1044
1045                 sk_stream_for_retrans_queue_from(skb, sk) {
1046                         int in_sack, pcount;
1047                         u8 sacked;
1048
1049                         tp->fastpath_skb_hint = skb;
1050                         tp->fastpath_cnt_hint = fack_count;
1051
1052                         /* The retransmission queue is always in order, so
1053                          * we can short-circuit the walk early.
1054                          */
1055                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1056                                 break;
1057
1058                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1059                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1060
1061                         pcount = tcp_skb_pcount(skb);
1062
1063                         if (pcount > 1 && !in_sack &&
1064                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1065                                 unsigned int pkt_len;
1066
1067                                 in_sack = !after(start_seq,
1068                                                  TCP_SKB_CB(skb)->seq);
1069
1070                                 if (!in_sack)
1071                                         pkt_len = (start_seq -
1072                                                    TCP_SKB_CB(skb)->seq);
1073                                 else
1074                                         pkt_len = (end_seq -
1075                                                    TCP_SKB_CB(skb)->seq);
1076                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1077                                         break;
1078                                 pcount = tcp_skb_pcount(skb);
1079                         }
1080
1081                         fack_count += pcount;
1082
1083                         sacked = TCP_SKB_CB(skb)->sacked;
1084
1085                         /* Account D-SACK for retransmitted packet. */
1086                         if ((dup_sack && in_sack) &&
1087                             (sacked & TCPCB_RETRANS) &&
1088                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1089                                 tp->undo_retrans--;
1090
1091                         /* The frame is ACKed. */
1092                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1093                                 if (sacked&TCPCB_RETRANS) {
1094                                         if ((dup_sack && in_sack) &&
1095                                             (sacked&TCPCB_SACKED_ACKED))
1096                                                 reord = min(fack_count, reord);
1097                                 } else {
1098                                         /* If it was in a hole, we detected reordering. */
1099                                         if (fack_count < prior_fackets &&
1100                                             !(sacked&TCPCB_SACKED_ACKED))
1101                                                 reord = min(fack_count, reord);
1102                                 }
1103
1104                                 /* Nothing to do; acked frame is about to be dropped. */
1105                                 continue;
1106                         }
1107
1108                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1109                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1110                             (!lost_retrans || after(end_seq, lost_retrans)))
1111                                 lost_retrans = end_seq;
1112
1113                         if (!in_sack)
1114                                 continue;
1115
1116                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1117                                 if (sacked & TCPCB_SACKED_RETRANS) {
1118                                         /* If the segment is not tagged as lost,
1119                                          * we do not clear RETRANS, believing
1120                                          * that retransmission is still in flight.
1121                                          */
1122                                         if (sacked & TCPCB_LOST) {
1123                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1124                                                 tp->lost_out -= tcp_skb_pcount(skb);
1125                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1126
1127                                                 /* clear lost hint */
1128                                                 tp->retransmit_skb_hint = NULL;
1129                                         }
1130                                 } else {
1131                                         /* New sack for not retransmitted frame,
1132                                          * which was in hole. It is reordering.
1133                                          */
1134                                         if (!(sacked & TCPCB_RETRANS) &&
1135                                             fack_count < prior_fackets)
1136                                                 reord = min(fack_count, reord);
1137
1138                                         if (sacked & TCPCB_LOST) {
1139                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1140                                                 tp->lost_out -= tcp_skb_pcount(skb);
1141
1142                                                 /* clear lost hint */
1143                                                 tp->retransmit_skb_hint = NULL;
1144                                         }
1145                                 }
1146
1147                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1148                                 flag |= FLAG_DATA_SACKED;
1149                                 tp->sacked_out += tcp_skb_pcount(skb);
1150
1151                                 if (fack_count > tp->fackets_out)
1152                                         tp->fackets_out = fack_count;
1153                         } else {
1154                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1155                                         reord = min(fack_count, reord);
1156                         }
1157
1158                         /* D-SACK. We can detect redundant retransmission
1159                          * in S|R and plain R frames and clear it.
1160                          * undo_retrans is decreased above, L|R frames
1161                          * are accounted above as well.
1162                          */
1163                         if (dup_sack &&
1164                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1165                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1166                                 tp->retrans_out -= tcp_skb_pcount(skb);
1167                                 tp->retransmit_skb_hint = NULL;
1168                         }
1169                 }
1170         }
1171
1172         /* Check for lost retransmit. This superb idea is
1173          * borrowed from "ratehalving". Event "C".
1174          * Later note: FACK people cheated me again 8),
1175          * we have to account for reordering! Ugly,
1176          * but should help.
1177          */
1178         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1179                 struct sk_buff *skb;
1180
1181                 sk_stream_for_retrans_queue(skb, sk) {
1182                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1183                                 break;
1184                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1185                                 continue;
1186                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1187                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1188                             (IsFack(tp) ||
1189                              !before(lost_retrans,
1190                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1191                                      tp->mss_cache))) {
1192                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1193                                 tp->retrans_out -= tcp_skb_pcount(skb);
1194
1195                                 /* clear lost hint */
1196                                 tp->retransmit_skb_hint = NULL;
1197
1198                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1199                                         tp->lost_out += tcp_skb_pcount(skb);
1200                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1201                                         flag |= FLAG_DATA_SACKED;
1202                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1203                                 }
1204                         }
1205                 }
1206         }
1207
1208         tp->left_out = tp->sacked_out + tp->lost_out;
1209
1210         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
1211                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1212
1213 #if FASTRETRANS_DEBUG > 0
1214         BUG_TRAP((int)tp->sacked_out >= 0);
1215         BUG_TRAP((int)tp->lost_out >= 0);
1216         BUG_TRAP((int)tp->retrans_out >= 0);
1217         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1218 #endif
1219         return flag;
1220 }
1221
1222 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1223  * segments to see from the next ACKs whether any data was really missing.
1224  * If the RTO was spurious, new ACKs should arrive.
1225  */
1226 void tcp_enter_frto(struct sock *sk)
1227 {
1228         const struct inet_connection_sock *icsk = inet_csk(sk);
1229         struct tcp_sock *tp = tcp_sk(sk);
1230         struct sk_buff *skb;
1231
1232         tp->frto_counter = 1;
1233
1234         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1235             tp->snd_una == tp->high_seq ||
1236             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1237                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1238                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1239                 tcp_ca_event(sk, CA_EVENT_FRTO);
1240         }
1241
1242         /* Have to clear retransmission markers here to keep the bookkeeping
1243          * in shape, even though we are not yet in Loss state.
1244          * If something was really lost, it is eventually caught up
1245          * in tcp_enter_frto_loss.
1246          */
1247         tp->retrans_out = 0;
1248         tp->undo_marker = tp->snd_una;
1249         tp->undo_retrans = 0;
1250
1251         sk_stream_for_retrans_queue(skb, sk) {
1252                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1253         }
1254         tcp_sync_left_out(tp);
1255
1256         tcp_set_ca_state(sk, TCP_CA_Open);
1257         tp->frto_highmark = tp->snd_nxt;
1258 }
1259
1260 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1261  * which indicates that we should follow the traditional RTO recovery,
1262  * i.e. mark everything lost and do go-back-N retransmission.
1263  */
1264 static void tcp_enter_frto_loss(struct sock *sk)
1265 {
1266         struct tcp_sock *tp = tcp_sk(sk);
1267         struct sk_buff *skb;
1268         int cnt = 0;
1269
1270         tp->sacked_out = 0;
1271         tp->lost_out = 0;
1272         tp->fackets_out = 0;
1273
1274         sk_stream_for_retrans_queue(skb, sk) {
1275                 cnt += tcp_skb_pcount(skb);
1276                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1277                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1278
1279                         /* Do not mark those segments lost that were
1280                          * forward transmitted after RTO
1281                          */
1282                         if (!after(TCP_SKB_CB(skb)->end_seq,
1283                                    tp->frto_highmark)) {
1284                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1285                                 tp->lost_out += tcp_skb_pcount(skb);
1286                         }
1287                 } else {
1288                         tp->sacked_out += tcp_skb_pcount(skb);
1289                         tp->fackets_out = cnt;
1290                 }
1291         }
1292         tcp_sync_left_out(tp);
1293
1294         tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1295         tp->snd_cwnd_cnt = 0;
1296         tp->snd_cwnd_stamp = tcp_time_stamp;
1297         tp->undo_marker = 0;
1298         tp->frto_counter = 0;
1299
1300         tp->reordering = min_t(unsigned int, tp->reordering,
1301                                              sysctl_tcp_reordering);
1302         tcp_set_ca_state(sk, TCP_CA_Loss);
1303         tp->high_seq = tp->frto_highmark;
1304         TCP_ECN_queue_cwr(tp);
1305
1306         clear_all_retrans_hints(tp);
1307 }
1308
1309 void tcp_clear_retrans(struct tcp_sock *tp)
1310 {
1311         tp->left_out = 0;
1312         tp->retrans_out = 0;
1313
1314         tp->fackets_out = 0;
1315         tp->sacked_out = 0;
1316         tp->lost_out = 0;
1317
1318         tp->undo_marker = 0;
1319         tp->undo_retrans = 0;
1320 }
1321
1322 /* Enter Loss state. If "how" is not zero, forget all SACK information
1323  * and reset tags completely, otherwise preserve SACKs. If receiver
1324  * dropped its ofo queue, we will know this due to reneging detection.
1325  */
1326 void tcp_enter_loss(struct sock *sk, int how)
1327 {
1328         const struct inet_connection_sock *icsk = inet_csk(sk);
1329         struct tcp_sock *tp = tcp_sk(sk);
1330         struct sk_buff *skb;
1331         int cnt = 0;
1332
1333         /* Reduce ssthresh if it has not yet been made inside this window. */
1334         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1335             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1336                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1337                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1338                 tcp_ca_event(sk, CA_EVENT_LOSS);
1339         }
1340         tp->snd_cwnd       = 1;
1341         tp->snd_cwnd_cnt   = 0;
1342         tp->snd_cwnd_stamp = tcp_time_stamp;
1343
1344         tp->bytes_acked = 0;
1345         tcp_clear_retrans(tp);
1346
1347         /* Push undo marker, if it was plain RTO and nothing
1348          * was retransmitted. */
1349         if (!how)
1350                 tp->undo_marker = tp->snd_una;
1351
1352         sk_stream_for_retrans_queue(skb, sk) {
1353                 cnt += tcp_skb_pcount(skb);
1354                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1355                         tp->undo_marker = 0;
1356                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1357                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1358                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1359                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1360                         tp->lost_out += tcp_skb_pcount(skb);
1361                 } else {
1362                         tp->sacked_out += tcp_skb_pcount(skb);
1363                         tp->fackets_out = cnt;
1364                 }
1365         }
1366         tcp_sync_left_out(tp);
1367
1368         tp->reordering = min_t(unsigned int, tp->reordering,
1369                                              sysctl_tcp_reordering);
1370         tcp_set_ca_state(sk, TCP_CA_Loss);
1371         tp->high_seq = tp->snd_nxt;
1372         TCP_ECN_queue_cwr(tp);
1373
1374         clear_all_retrans_hints(tp);
1375 }
1376
1377 static int tcp_check_sack_reneging(struct sock *sk)
1378 {
1379         struct sk_buff *skb;
1380
1381         /* If ACK arrived pointing to a remembered SACK,
1382          * it means that our remembered SACKs do not reflect
1383          * real state of receiver i.e.
1384          * receiver _host_ is heavily congested (or buggy).
1385          * Do processing similar to RTO timeout.
1386          */
1387         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1388             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1389                 struct inet_connection_sock *icsk = inet_csk(sk);
1390                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1391
1392                 tcp_enter_loss(sk, 1);
1393                 icsk->icsk_retransmits++;
1394                 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1395                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1396                                           icsk->icsk_rto, TCP_RTO_MAX);
1397                 return 1;
1398         }
1399         return 0;
1400 }
1401
1402 static inline int tcp_fackets_out(struct tcp_sock *tp)
1403 {
1404         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1405 }
1406
1407 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1408 {
1409         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1410 }
1411
1412 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1413 {
1414         return tp->packets_out &&
1415                tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
1416 }
1417
1418 /* Linux NewReno/SACK/FACK/ECN state machine.
1419  * --------------------------------------
1420  *
1421  * "Open"       Normal state, no dubious events, fast path.
1422  * "Disorder"   In all the respects it is "Open",
1423  *              but requires a bit more attention. It is entered when
1424  *              we see some SACKs or dupacks. It is split of "Open"
1425  *              mainly to move some processing from fast path to slow one.
1426  * "CWR"        CWND was reduced due to some Congestion Notification event.
1427  *              It can be ECN, ICMP source quench, local device congestion.
1428  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1429  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1430  *
1431  * tcp_fastretrans_alert() is entered:
1432  * - each incoming ACK, if state is not "Open"
1433  * - when arrived ACK is unusual, namely:
1434  *      * SACK
1435  *      * Duplicate ACK.
1436  *      * ECN ECE.
1437  *
1438  * Counting packets in flight is pretty simple.
1439  *
1440  *      in_flight = packets_out - left_out + retrans_out
1441  *
1442  *      packets_out is SND.NXT-SND.UNA counted in packets.
1443  *
1444  *      retrans_out is number of retransmitted segments.
1445  *
1446  *      left_out is number of segments left network, but not ACKed yet.
1447  *
1448  *              left_out = sacked_out + lost_out
1449  *
1450  *     sacked_out: Packets, which arrived to receiver out of order
1451  *                 and hence not ACKed. With SACKs this number is simply
1452  *                 amount of SACKed data. Even without SACKs
1453  *                 it is easy to give pretty reliable estimate of this number,
1454  *                 counting duplicate ACKs.
1455  *
1456  *       lost_out: Packets lost by network. TCP has no explicit
1457  *                 "loss notification" feedback from network (for now).
1458  *                 It means that this number can be only _guessed_.
1459  *                 Actually, it is the heuristics to predict lossage that
1460  *                 distinguishes different algorithms.
1461  *
1462  *      F.e. after RTO, when all the queue is considered as lost,
1463  *      lost_out = packets_out and in_flight = retrans_out.
1464  *
1465  *              Essentially, we have now two algorithms counting
1466  *              lost packets.
1467  *
1468  *              FACK: It is the simplest heuristics. As soon as we decided
1469  *              that something is lost, we decide that _all_ not SACKed
1470  *              packets until the most forward SACK are lost. I.e.
1471  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1472  *              It is absolutely correct estimate, if network does not reorder
1473  *              packets. And it loses any connection to reality when reordering
1474  *              takes place. We use FACK by default until reordering
1475  *              is suspected on the path to this destination.
1476  *
1477  *              NewReno: when Recovery is entered, we assume that one segment
1478  *              is lost (classic Reno). While we are in Recovery and
1479  *              a partial ACK arrives, we assume that one more packet
1480  *              is lost (NewReno). This heuristics are the same in NewReno
1481  *              and SACK.
1482  *
1483  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1484  *  deflation etc. CWND is real congestion window, never inflated, changes
1485  *  only according to classic VJ rules.
1486  *
1487  * Really tricky (and requiring careful tuning) part of algorithm
1488  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1489  * The first determines the moment _when_ we should reduce CWND and,
1490  * hence, slow down forward transmission. In fact, it determines the moment
1491  * when we decide that hole is caused by loss, rather than by a reorder.
1492  *
1493  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1494  * holes, caused by lost packets.
1495  *
1496  * And the most logically complicated part of algorithm is undo
1497  * heuristics. We detect false retransmits due to both too early
1498  * fast retransmit (reordering) and underestimated RTO, analyzing
1499  * timestamps and D-SACKs. When we detect that some segments were
1500  * retransmitted by mistake and CWND reduction was wrong, we undo
1501  * window reduction and abort recovery phase. This logic is hidden
1502  * inside several functions named tcp_try_undo_<something>.
1503  */
1504
1505 /* This function decides, when we should leave Disordered state
1506  * and enter Recovery phase, reducing congestion window.
1507  *
1508  * Main question: may we further continue forward transmission
1509  * with the same cwnd?
1510  */
1511 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1512 {
1513         __u32 packets_out;
1514
1515         /* Trick#1: The loss is proven. */
1516         if (tp->lost_out)
1517                 return 1;
1518
1519         /* Not-A-Trick#2 : Classic rule... */
1520         if (tcp_fackets_out(tp) > tp->reordering)
1521                 return 1;
1522
1523         /* Trick#3 : when we use RFC2988 timer restart, fast
1524          * retransmit can be triggered by timeout of queue head.
1525          */
1526         if (tcp_head_timedout(sk, tp))
1527                 return 1;
1528
1529         /* Trick#4: It is still not OK... But will it be useful to delay
1530          * recovery more?
1531          */
1532         packets_out = tp->packets_out;
1533         if (packets_out <= tp->reordering &&
1534             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1535             !tcp_may_send_now(sk, tp)) {
1536                 /* We have nothing to send. This connection is limited
1537                  * either by receiver window or by application.
1538                  */
1539                 return 1;
1540         }
1541
1542         return 0;
1543 }
1544
1545 /* If we receive more dupacks than we expected counting segments
1546  * in assumption of absent reordering, interpret this as reordering.
1547  * The only another reason could be bug in receiver TCP.
1548  */
1549 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1550 {
1551         struct tcp_sock *tp = tcp_sk(sk);
1552         u32 holes;
1553
1554         holes = max(tp->lost_out, 1U);
1555         holes = min(holes, tp->packets_out);
1556
1557         if ((tp->sacked_out + holes) > tp->packets_out) {
1558                 tp->sacked_out = tp->packets_out - holes;
1559                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1560         }
1561 }
1562
1563 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1564
1565 static void tcp_add_reno_sack(struct sock *sk)
1566 {
1567         struct tcp_sock *tp = tcp_sk(sk);
1568         tp->sacked_out++;
1569         tcp_check_reno_reordering(sk, 0);
1570         tcp_sync_left_out(tp);
1571 }
1572
1573 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1574
1575 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1576 {
1577         if (acked > 0) {
1578                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1579                 if (acked-1 >= tp->sacked_out)
1580                         tp->sacked_out = 0;
1581                 else
1582                         tp->sacked_out -= acked-1;
1583         }
1584         tcp_check_reno_reordering(sk, acked);
1585         tcp_sync_left_out(tp);
1586 }
1587
1588 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1589 {
1590         tp->sacked_out = 0;
1591         tp->left_out = tp->lost_out;
1592 }
1593
1594 /* Mark head of queue up as lost. */
1595 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1596                                int packets, u32 high_seq)
1597 {
1598         struct sk_buff *skb;
1599         int cnt;
1600
1601         BUG_TRAP(packets <= tp->packets_out);
1602         if (tp->lost_skb_hint) {
1603                 skb = tp->lost_skb_hint;
1604                 cnt = tp->lost_cnt_hint;
1605         } else {
1606                 skb = sk->sk_write_queue.next;
1607                 cnt = 0;
1608         }
1609
1610         sk_stream_for_retrans_queue_from(skb, sk) {
1611                 /* TODO: do this better */
1612                 /* this is not the most efficient way to do this... */
1613                 tp->lost_skb_hint = skb;
1614                 tp->lost_cnt_hint = cnt;
1615                 cnt += tcp_skb_pcount(skb);
1616                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1617                         break;
1618                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1619                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1620                         tp->lost_out += tcp_skb_pcount(skb);
1621
1622                         /* clear xmit_retransmit_queue hints
1623                          *  if this is beyond hint */
1624                         if(tp->retransmit_skb_hint != NULL &&
1625                            before(TCP_SKB_CB(skb)->seq,
1626                                   TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) {
1627
1628                                 tp->retransmit_skb_hint = NULL;
1629                         }
1630                 }
1631         }
1632         tcp_sync_left_out(tp);
1633 }
1634
1635 /* Account newly detected lost packet(s) */
1636
1637 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1638 {
1639         if (IsFack(tp)) {
1640                 int lost = tp->fackets_out - tp->reordering;
1641                 if (lost <= 0)
1642                         lost = 1;
1643                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1644         } else {
1645                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1646         }
1647
1648         /* New heuristics: it is possible only after we switched
1649          * to restart timer each time when something is ACKed.
1650          * Hence, we can detect timed out packets during fast
1651          * retransmit without falling to slow start.
1652          */
1653         if (!IsReno(tp) && tcp_head_timedout(sk, tp)) {
1654                 struct sk_buff *skb;
1655
1656                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1657                         : sk->sk_write_queue.next;
1658
1659                 sk_stream_for_retrans_queue_from(skb, sk) {
1660                         if (!tcp_skb_timedout(sk, skb))
1661                                 break;
1662
1663                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1664                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1665                                 tp->lost_out += tcp_skb_pcount(skb);
1666
1667                                 /* clear xmit_retrans hint */
1668                                 if (tp->retransmit_skb_hint &&
1669                                     before(TCP_SKB_CB(skb)->seq,
1670                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1671
1672                                         tp->retransmit_skb_hint = NULL;
1673                         }
1674                 }
1675
1676                 tp->scoreboard_skb_hint = skb;
1677
1678                 tcp_sync_left_out(tp);
1679         }
1680 }
1681
1682 /* CWND moderation, preventing bursts due to too big ACKs
1683  * in dubious situations.
1684  */
1685 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1686 {
1687         tp->snd_cwnd = min(tp->snd_cwnd,
1688                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1689         tp->snd_cwnd_stamp = tcp_time_stamp;
1690 }
1691
1692 /* Lower bound on congestion window is slow start threshold
1693  * unless congestion avoidance choice decides to overide it.
1694  */
1695 static inline u32 tcp_cwnd_min(const struct sock *sk)
1696 {
1697         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1698
1699         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1700 }
1701
1702 /* Decrease cwnd each second ack. */
1703 static void tcp_cwnd_down(struct sock *sk)
1704 {
1705         struct tcp_sock *tp = tcp_sk(sk);
1706         int decr = tp->snd_cwnd_cnt + 1;
1707
1708         tp->snd_cwnd_cnt = decr&1;
1709         decr >>= 1;
1710
1711         if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1712                 tp->snd_cwnd -= decr;
1713
1714         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1715         tp->snd_cwnd_stamp = tcp_time_stamp;
1716 }
1717
1718 /* Nothing was retransmitted or returned timestamp is less
1719  * than timestamp of the first retransmission.
1720  */
1721 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1722 {
1723         return !tp->retrans_stamp ||
1724                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1725                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1726 }
1727
1728 /* Undo procedures. */
1729
1730 #if FASTRETRANS_DEBUG > 1
1731 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1732 {
1733         struct inet_sock *inet = inet_sk(sk);
1734         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1735                msg,
1736                NIPQUAD(inet->daddr), ntohs(inet->dport),
1737                tp->snd_cwnd, tp->left_out,
1738                tp->snd_ssthresh, tp->prior_ssthresh,
1739                tp->packets_out);
1740 }
1741 #else
1742 #define DBGUNDO(x...) do { } while (0)
1743 #endif
1744
1745 static void tcp_undo_cwr(struct sock *sk, const int undo)
1746 {
1747         struct tcp_sock *tp = tcp_sk(sk);
1748
1749         if (tp->prior_ssthresh) {
1750                 const struct inet_connection_sock *icsk = inet_csk(sk);
1751
1752                 if (icsk->icsk_ca_ops->undo_cwnd)
1753                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1754                 else
1755                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1756
1757                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1758                         tp->snd_ssthresh = tp->prior_ssthresh;
1759                         TCP_ECN_withdraw_cwr(tp);
1760                 }
1761         } else {
1762                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1763         }
1764         tcp_moderate_cwnd(tp);
1765         tp->snd_cwnd_stamp = tcp_time_stamp;
1766
1767         /* There is something screwy going on with the retrans hints after
1768            an undo */
1769         clear_all_retrans_hints(tp);
1770 }
1771
1772 static inline int tcp_may_undo(struct tcp_sock *tp)
1773 {
1774         return tp->undo_marker &&
1775                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1776 }
1777
1778 /* People celebrate: "We love our President!" */
1779 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1780 {
1781         if (tcp_may_undo(tp)) {
1782                 /* Happy end! We did not retransmit anything
1783                  * or our original transmission succeeded.
1784                  */
1785                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1786                 tcp_undo_cwr(sk, 1);
1787                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1788                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1789                 else
1790                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1791                 tp->undo_marker = 0;
1792         }
1793         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1794                 /* Hold old state until something *above* high_seq
1795                  * is ACKed. For Reno it is MUST to prevent false
1796                  * fast retransmits (RFC2582). SACK TCP is safe. */
1797                 tcp_moderate_cwnd(tp);
1798                 return 1;
1799         }
1800         tcp_set_ca_state(sk, TCP_CA_Open);
1801         return 0;
1802 }
1803
1804 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1805 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1806 {
1807         if (tp->undo_marker && !tp->undo_retrans) {
1808                 DBGUNDO(sk, tp, "D-SACK");
1809                 tcp_undo_cwr(sk, 1);
1810                 tp->undo_marker = 0;
1811                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1812         }
1813 }
1814
1815 /* Undo during fast recovery after partial ACK. */
1816
1817 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1818                                 int acked)
1819 {
1820         /* Partial ACK arrived. Force Hoe's retransmit. */
1821         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1822
1823         if (tcp_may_undo(tp)) {
1824                 /* Plain luck! Hole if filled with delayed
1825                  * packet, rather than with a retransmit.
1826                  */
1827                 if (tp->retrans_out == 0)
1828                         tp->retrans_stamp = 0;
1829
1830                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1831
1832                 DBGUNDO(sk, tp, "Hoe");
1833                 tcp_undo_cwr(sk, 0);
1834                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1835
1836                 /* So... Do not make Hoe's retransmit yet.
1837                  * If the first packet was delayed, the rest
1838                  * ones are most probably delayed as well.
1839                  */
1840                 failed = 0;
1841         }
1842         return failed;
1843 }
1844
1845 /* Undo during loss recovery after partial ACK. */
1846 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1847 {
1848         if (tcp_may_undo(tp)) {
1849                 struct sk_buff *skb;
1850                 sk_stream_for_retrans_queue(skb, sk) {
1851                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1852                 }
1853
1854                 clear_all_retrans_hints(tp);
1855
1856                 DBGUNDO(sk, tp, "partial loss");
1857                 tp->lost_out = 0;
1858                 tp->left_out = tp->sacked_out;
1859                 tcp_undo_cwr(sk, 1);
1860                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1861                 inet_csk(sk)->icsk_retransmits = 0;
1862                 tp->undo_marker = 0;
1863                 if (!IsReno(tp))
1864                         tcp_set_ca_state(sk, TCP_CA_Open);
1865                 return 1;
1866         }
1867         return 0;
1868 }
1869
1870 static inline void tcp_complete_cwr(struct sock *sk)
1871 {
1872         struct tcp_sock *tp = tcp_sk(sk);
1873         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1874         tp->snd_cwnd_stamp = tcp_time_stamp;
1875         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
1876 }
1877
1878 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1879 {
1880         tp->left_out = tp->sacked_out;
1881
1882         if (tp->retrans_out == 0)
1883                 tp->retrans_stamp = 0;
1884
1885         if (flag&FLAG_ECE)
1886                 tcp_enter_cwr(sk);
1887
1888         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
1889                 int state = TCP_CA_Open;
1890
1891                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1892                         state = TCP_CA_Disorder;
1893
1894                 if (inet_csk(sk)->icsk_ca_state != state) {
1895                         tcp_set_ca_state(sk, state);
1896                         tp->high_seq = tp->snd_nxt;
1897                 }
1898                 tcp_moderate_cwnd(tp);
1899         } else {
1900                 tcp_cwnd_down(sk);
1901         }
1902 }
1903
1904 static void tcp_mtup_probe_failed(struct sock *sk)
1905 {
1906         struct inet_connection_sock *icsk = inet_csk(sk);
1907
1908         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
1909         icsk->icsk_mtup.probe_size = 0;
1910 }
1911
1912 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
1913 {
1914         struct tcp_sock *tp = tcp_sk(sk);
1915         struct inet_connection_sock *icsk = inet_csk(sk);
1916
1917         /* FIXME: breaks with very large cwnd */
1918         tp->prior_ssthresh = tcp_current_ssthresh(sk);
1919         tp->snd_cwnd = tp->snd_cwnd *
1920                        tcp_mss_to_mtu(sk, tp->mss_cache) /
1921                        icsk->icsk_mtup.probe_size;
1922         tp->snd_cwnd_cnt = 0;
1923         tp->snd_cwnd_stamp = tcp_time_stamp;
1924         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
1925
1926         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
1927         icsk->icsk_mtup.probe_size = 0;
1928         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
1929 }
1930
1931
1932 /* Process an event, which can update packets-in-flight not trivially.
1933  * Main goal of this function is to calculate new estimate for left_out,
1934  * taking into account both packets sitting in receiver's buffer and
1935  * packets lost by network.
1936  *
1937  * Besides that it does CWND reduction, when packet loss is detected
1938  * and changes state of machine.
1939  *
1940  * It does _not_ decide what to send, it is made in function
1941  * tcp_xmit_retransmit_queue().
1942  */
1943 static void
1944 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1945                       int prior_packets, int flag)
1946 {
1947         struct inet_connection_sock *icsk = inet_csk(sk);
1948         struct tcp_sock *tp = tcp_sk(sk);
1949         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1950
1951         /* Some technical things:
1952          * 1. Reno does not count dupacks (sacked_out) automatically. */
1953         if (!tp->packets_out)
1954                 tp->sacked_out = 0;
1955         /* 2. SACK counts snd_fack in packets inaccurately. */
1956         if (tp->sacked_out == 0)
1957                 tp->fackets_out = 0;
1958
1959         /* Now state machine starts.
1960          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1961         if (flag&FLAG_ECE)
1962                 tp->prior_ssthresh = 0;
1963
1964         /* B. In all the states check for reneging SACKs. */
1965         if (tp->sacked_out && tcp_check_sack_reneging(sk))
1966                 return;
1967
1968         /* C. Process data loss notification, provided it is valid. */
1969         if ((flag&FLAG_DATA_LOST) &&
1970             before(tp->snd_una, tp->high_seq) &&
1971             icsk->icsk_ca_state != TCP_CA_Open &&
1972             tp->fackets_out > tp->reordering) {
1973                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1974                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1975         }
1976
1977         /* D. Synchronize left_out to current state. */
1978         tcp_sync_left_out(tp);
1979
1980         /* E. Check state exit conditions. State can be terminated
1981          *    when high_seq is ACKed. */
1982         if (icsk->icsk_ca_state == TCP_CA_Open) {
1983                 if (!sysctl_tcp_frto)
1984                         BUG_TRAP(tp->retrans_out == 0);
1985                 tp->retrans_stamp = 0;
1986         } else if (!before(tp->snd_una, tp->high_seq)) {
1987                 switch (icsk->icsk_ca_state) {
1988                 case TCP_CA_Loss:
1989                         icsk->icsk_retransmits = 0;
1990                         if (tcp_try_undo_recovery(sk, tp))
1991                                 return;
1992                         break;
1993
1994                 case TCP_CA_CWR:
1995                         /* CWR is to be held something *above* high_seq
1996                          * is ACKed for CWR bit to reach receiver. */
1997                         if (tp->snd_una != tp->high_seq) {
1998                                 tcp_complete_cwr(sk);
1999                                 tcp_set_ca_state(sk, TCP_CA_Open);
2000                         }
2001                         break;
2002
2003                 case TCP_CA_Disorder:
2004                         tcp_try_undo_dsack(sk, tp);
2005                         if (!tp->undo_marker ||
2006                             /* For SACK case do not Open to allow to undo
2007                              * catching for all duplicate ACKs. */
2008                             IsReno(tp) || tp->snd_una != tp->high_seq) {
2009                                 tp->undo_marker = 0;
2010                                 tcp_set_ca_state(sk, TCP_CA_Open);
2011                         }
2012                         break;
2013
2014                 case TCP_CA_Recovery:
2015                         if (IsReno(tp))
2016                                 tcp_reset_reno_sack(tp);
2017                         if (tcp_try_undo_recovery(sk, tp))
2018                                 return;
2019                         tcp_complete_cwr(sk);
2020                         break;
2021                 }
2022         }
2023
2024         /* F. Process state. */
2025         switch (icsk->icsk_ca_state) {
2026         case TCP_CA_Recovery:
2027                 if (prior_snd_una == tp->snd_una) {
2028                         if (IsReno(tp) && is_dupack)
2029                                 tcp_add_reno_sack(sk);
2030                 } else {
2031                         int acked = prior_packets - tp->packets_out;
2032                         if (IsReno(tp))
2033                                 tcp_remove_reno_sacks(sk, tp, acked);
2034                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
2035                 }
2036                 break;
2037         case TCP_CA_Loss:
2038                 if (flag&FLAG_DATA_ACKED)
2039                         icsk->icsk_retransmits = 0;
2040                 if (!tcp_try_undo_loss(sk, tp)) {
2041                         tcp_moderate_cwnd(tp);
2042                         tcp_xmit_retransmit_queue(sk);
2043                         return;
2044                 }
2045                 if (icsk->icsk_ca_state != TCP_CA_Open)
2046                         return;
2047                 /* Loss is undone; fall through to processing in Open state. */
2048         default:
2049                 if (IsReno(tp)) {
2050                         if (tp->snd_una != prior_snd_una)
2051                                 tcp_reset_reno_sack(tp);
2052                         if (is_dupack)
2053                                 tcp_add_reno_sack(sk);
2054                 }
2055
2056                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2057                         tcp_try_undo_dsack(sk, tp);
2058
2059                 if (!tcp_time_to_recover(sk, tp)) {
2060                         tcp_try_to_open(sk, tp, flag);
2061                         return;
2062                 }
2063
2064                 /* MTU probe failure: don't reduce cwnd */
2065                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2066                     icsk->icsk_mtup.probe_size &&
2067                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2068                         tcp_mtup_probe_failed(sk);
2069                         /* Restores the reduction we did in tcp_mtup_probe() */
2070                         tp->snd_cwnd++;
2071                         tcp_simple_retransmit(sk);
2072                         return;
2073                 }
2074
2075                 /* Otherwise enter Recovery state */
2076
2077                 if (IsReno(tp))
2078                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2079                 else
2080                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2081
2082                 tp->high_seq = tp->snd_nxt;
2083                 tp->prior_ssthresh = 0;
2084                 tp->undo_marker = tp->snd_una;
2085                 tp->undo_retrans = tp->retrans_out;
2086
2087                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2088                         if (!(flag&FLAG_ECE))
2089                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2090                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2091                         TCP_ECN_queue_cwr(tp);
2092                 }
2093
2094                 tp->bytes_acked = 0;
2095                 tp->snd_cwnd_cnt = 0;
2096                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2097         }
2098
2099         if (is_dupack || tcp_head_timedout(sk, tp))
2100                 tcp_update_scoreboard(sk, tp);
2101         tcp_cwnd_down(sk);
2102         tcp_xmit_retransmit_queue(sk);
2103 }
2104
2105 /* Read draft-ietf-tcplw-high-performance before mucking
2106  * with this code. (Supersedes RFC1323)
2107  */
2108 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2109 {
2110         /* RTTM Rule: A TSecr value received in a segment is used to
2111          * update the averaged RTT measurement only if the segment
2112          * acknowledges some new data, i.e., only if it advances the
2113          * left edge of the send window.
2114          *
2115          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2116          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2117          *
2118          * Changed: reset backoff as soon as we see the first valid sample.
2119          * If we do not, we get strongly overestimated rto. With timestamps
2120          * samples are accepted even from very old segments: f.e., when rtt=1
2121          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2122          * answer arrives rto becomes 120 seconds! If at least one of segments
2123          * in window is lost... Voila.                          --ANK (010210)
2124          */
2125         struct tcp_sock *tp = tcp_sk(sk);
2126         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2127         tcp_rtt_estimator(sk, seq_rtt);
2128         tcp_set_rto(sk);
2129         inet_csk(sk)->icsk_backoff = 0;
2130         tcp_bound_rto(sk);
2131 }
2132
2133 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2134 {
2135         /* We don't have a timestamp. Can only use
2136          * packets that are not retransmitted to determine
2137          * rtt estimates. Also, we must not reset the
2138          * backoff for rto until we get a non-retransmitted
2139          * packet. This allows us to deal with a situation
2140          * where the network delay has increased suddenly.
2141          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2142          */
2143
2144         if (flag & FLAG_RETRANS_DATA_ACKED)
2145                 return;
2146
2147         tcp_rtt_estimator(sk, seq_rtt);
2148         tcp_set_rto(sk);
2149         inet_csk(sk)->icsk_backoff = 0;
2150         tcp_bound_rto(sk);
2151 }
2152
2153 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2154                                       const s32 seq_rtt)
2155 {
2156         const struct tcp_sock *tp = tcp_sk(sk);
2157         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2158         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2159                 tcp_ack_saw_tstamp(sk, flag);
2160         else if (seq_rtt >= 0)
2161                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2162 }
2163
2164 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2165                            u32 in_flight, int good)
2166 {
2167         const struct inet_connection_sock *icsk = inet_csk(sk);
2168         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2169         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2170 }
2171
2172 /* Restart timer after forward progress on connection.
2173  * RFC2988 recommends to restart timer to now+rto.
2174  */
2175
2176 static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2177 {
2178         if (!tp->packets_out) {
2179                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2180         } else {
2181                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2182         }
2183 }
2184
2185 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2186                          __u32 now, __s32 *seq_rtt)
2187 {
2188         struct tcp_sock *tp = tcp_sk(sk);
2189         struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2190         __u32 seq = tp->snd_una;
2191         __u32 packets_acked;
2192         int acked = 0;
2193
2194         /* If we get here, the whole TSO packet has not been
2195          * acked.
2196          */
2197         BUG_ON(!after(scb->end_seq, seq));
2198
2199         packets_acked = tcp_skb_pcount(skb);
2200         if (tcp_trim_head(sk, skb, seq - scb->seq))
2201                 return 0;
2202         packets_acked -= tcp_skb_pcount(skb);
2203
2204         if (packets_acked) {
2205                 __u8 sacked = scb->sacked;
2206
2207                 acked |= FLAG_DATA_ACKED;
2208                 if (sacked) {
2209                         if (sacked & TCPCB_RETRANS) {
2210                                 if (sacked & TCPCB_SACKED_RETRANS)
2211                                         tp->retrans_out -= packets_acked;
2212                                 acked |= FLAG_RETRANS_DATA_ACKED;
2213                                 *seq_rtt = -1;
2214                         } else if (*seq_rtt < 0)
2215                                 *seq_rtt = now - scb->when;
2216                         if (sacked & TCPCB_SACKED_ACKED)
2217                                 tp->sacked_out -= packets_acked;
2218                         if (sacked & TCPCB_LOST)
2219                                 tp->lost_out -= packets_acked;
2220                         if (sacked & TCPCB_URG) {
2221                                 if (tp->urg_mode &&
2222                                     !before(seq, tp->snd_up))
2223                                         tp->urg_mode = 0;
2224                         }
2225                 } else if (*seq_rtt < 0)
2226                         *seq_rtt = now - scb->when;
2227
2228                 if (tp->fackets_out) {
2229                         __u32 dval = min(tp->fackets_out, packets_acked);
2230                         tp->fackets_out -= dval;
2231                 }
2232                 tp->packets_out -= packets_acked;
2233
2234                 BUG_ON(tcp_skb_pcount(skb) == 0);
2235                 BUG_ON(!before(scb->seq, scb->end_seq));
2236         }
2237
2238         return acked;
2239 }
2240
2241 static u32 tcp_usrtt(const struct sk_buff *skb)
2242 {
2243         struct timeval tv, now;
2244
2245         do_gettimeofday(&now);
2246         skb_get_timestamp(skb, &tv);
2247         return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec);
2248 }
2249
2250 /* Remove acknowledged frames from the retransmission queue. */
2251 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2252 {
2253         struct tcp_sock *tp = tcp_sk(sk);
2254         const struct inet_connection_sock *icsk = inet_csk(sk);
2255         struct sk_buff *skb;
2256         __u32 now = tcp_time_stamp;
2257         int acked = 0;
2258         __s32 seq_rtt = -1;
2259         u32 pkts_acked = 0;
2260         void (*rtt_sample)(struct sock *sk, u32 usrtt)
2261                 = icsk->icsk_ca_ops->rtt_sample;
2262
2263         while ((skb = skb_peek(&sk->sk_write_queue)) &&
2264                skb != sk->sk_send_head) {
2265                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2266                 __u8 sacked = scb->sacked;
2267
2268                 /* If our packet is before the ack sequence we can
2269                  * discard it as it's confirmed to have arrived at
2270                  * the other end.
2271                  */
2272                 if (after(scb->end_seq, tp->snd_una)) {
2273                         if (tcp_skb_pcount(skb) > 1 &&
2274                             after(tp->snd_una, scb->seq))
2275                                 acked |= tcp_tso_acked(sk, skb,
2276                                                        now, &seq_rtt);
2277                         break;
2278                 }
2279
2280                 /* Initial outgoing SYN's get put onto the write_queue
2281                  * just like anything else we transmit.  It is not
2282                  * true data, and if we misinform our callers that
2283                  * this ACK acks real data, we will erroneously exit
2284                  * connection startup slow start one packet too
2285                  * quickly.  This is severely frowned upon behavior.
2286                  */
2287                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2288                         acked |= FLAG_DATA_ACKED;
2289                         ++pkts_acked;
2290                 } else {
2291                         acked |= FLAG_SYN_ACKED;
2292                         tp->retrans_stamp = 0;
2293                 }
2294
2295                 /* MTU probing checks */
2296                 if (icsk->icsk_mtup.probe_size) {
2297                         if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2298                                 tcp_mtup_probe_success(sk, skb);
2299                         }
2300                 }
2301
2302                 if (sacked) {
2303                         if (sacked & TCPCB_RETRANS) {
2304                                 if(sacked & TCPCB_SACKED_RETRANS)
2305                                         tp->retrans_out -= tcp_skb_pcount(skb);
2306                                 acked |= FLAG_RETRANS_DATA_ACKED;
2307                                 seq_rtt = -1;
2308                         } else if (seq_rtt < 0) {
2309                                 seq_rtt = now - scb->when;
2310                                 if (rtt_sample)
2311                                         (*rtt_sample)(sk, tcp_usrtt(skb));
2312                         }
2313                         if (sacked & TCPCB_SACKED_ACKED)
2314                                 tp->sacked_out -= tcp_skb_pcount(skb);
2315                         if (sacked & TCPCB_LOST)
2316                                 tp->lost_out -= tcp_skb_pcount(skb);
2317                         if (sacked & TCPCB_URG) {
2318                                 if (tp->urg_mode &&
2319                                     !before(scb->end_seq, tp->snd_up))
2320                                         tp->urg_mode = 0;
2321                         }
2322                 } else if (seq_rtt < 0) {
2323                         seq_rtt = now - scb->when;
2324                         if (rtt_sample)
2325                                 (*rtt_sample)(sk, tcp_usrtt(skb));
2326                 }
2327                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2328                 tcp_packets_out_dec(tp, skb);
2329                 __skb_unlink(skb, &sk->sk_write_queue);
2330                 sk_stream_free_skb(sk, skb);
2331                 clear_all_retrans_hints(tp);
2332         }
2333
2334         if (acked&FLAG_ACKED) {
2335                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2336                 tcp_ack_packets_out(sk, tp);
2337
2338                 if (icsk->icsk_ca_ops->pkts_acked)
2339                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2340         }
2341
2342 #if FASTRETRANS_DEBUG > 0
2343         BUG_TRAP((int)tp->sacked_out >= 0);
2344         BUG_TRAP((int)tp->lost_out >= 0);
2345         BUG_TRAP((int)tp->retrans_out >= 0);
2346         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2347                 const struct inet_connection_sock *icsk = inet_csk(sk);
2348                 if (tp->lost_out) {
2349                         printk(KERN_DEBUG "Leak l=%u %d\n",
2350                                tp->lost_out, icsk->icsk_ca_state);
2351                         tp->lost_out = 0;
2352                 }
2353                 if (tp->sacked_out) {
2354                         printk(KERN_DEBUG "Leak s=%u %d\n",
2355                                tp->sacked_out, icsk->icsk_ca_state);
2356                         tp->sacked_out = 0;
2357                 }
2358                 if (tp->retrans_out) {
2359                         printk(KERN_DEBUG "Leak r=%u %d\n",
2360                                tp->retrans_out, icsk->icsk_ca_state);
2361                         tp->retrans_out = 0;
2362                 }
2363         }
2364 #endif
2365         *seq_rtt_p = seq_rtt;
2366         return acked;
2367 }
2368
2369 static void tcp_ack_probe(struct sock *sk)
2370 {
2371         const struct tcp_sock *tp = tcp_sk(sk);
2372         struct inet_connection_sock *icsk = inet_csk(sk);
2373
2374         /* Was it a usable window open? */
2375
2376         if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2377                    tp->snd_una + tp->snd_wnd)) {
2378                 icsk->icsk_backoff = 0;
2379                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2380                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2381                  * This function is not for random using!
2382                  */
2383         } else {
2384                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2385                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2386                                           TCP_RTO_MAX);
2387         }
2388 }
2389
2390 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2391 {
2392         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2393                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2394 }
2395
2396 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2397 {
2398         const struct tcp_sock *tp = tcp_sk(sk);
2399         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2400                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2401 }
2402
2403 /* Check that window update is acceptable.
2404  * The function assumes that snd_una<=ack<=snd_next.
2405  */
2406 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2407                                         const u32 ack_seq, const u32 nwin)
2408 {
2409         return (after(ack, tp->snd_una) ||
2410                 after(ack_seq, tp->snd_wl1) ||
2411                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2412 }
2413
2414 /* Update our send window.
2415  *
2416  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2417  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2418  */
2419 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2420                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2421 {
2422         int flag = 0;
2423         u32 nwin = ntohs(skb->h.th->window);
2424
2425         if (likely(!skb->h.th->syn))
2426                 nwin <<= tp->rx_opt.snd_wscale;
2427
2428         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2429                 flag |= FLAG_WIN_UPDATE;
2430                 tcp_update_wl(tp, ack, ack_seq);
2431
2432                 if (tp->snd_wnd != nwin) {
2433                         tp->snd_wnd = nwin;
2434
2435                         /* Note, it is the only place, where
2436                          * fast path is recovered for sending TCP.
2437                          */
2438                         tp->pred_flags = 0;
2439                         tcp_fast_path_check(sk, tp);
2440
2441                         if (nwin > tp->max_window) {
2442                                 tp->max_window = nwin;
2443                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2444                         }
2445                 }
2446         }
2447
2448         tp->snd_una = ack;
2449
2450         return flag;
2451 }
2452
2453 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2454 {
2455         struct tcp_sock *tp = tcp_sk(sk);
2456         
2457         tcp_sync_left_out(tp);
2458         
2459         if (tp->snd_una == prior_snd_una ||
2460             !before(tp->snd_una, tp->frto_highmark)) {
2461                 /* RTO was caused by loss, start retransmitting in
2462                  * go-back-N slow start
2463                  */
2464                 tcp_enter_frto_loss(sk);
2465                 return;
2466         }
2467
2468         if (tp->frto_counter == 1) {
2469                 /* First ACK after RTO advances the window: allow two new
2470                  * segments out.
2471                  */
2472                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2473         } else {
2474                 /* Also the second ACK after RTO advances the window.
2475                  * The RTO was likely spurious. Reduce cwnd and continue
2476                  * in congestion avoidance
2477                  */
2478                 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2479                 tcp_moderate_cwnd(tp);
2480         }
2481
2482         /* F-RTO affects on two new ACKs following RTO.
2483          * At latest on third ACK the TCP behavior is back to normal.
2484          */
2485         tp->frto_counter = (tp->frto_counter + 1) % 3;
2486 }
2487
2488 /* This routine deals with incoming acks, but not outgoing ones. */
2489 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2490 {
2491         struct inet_connection_sock *icsk = inet_csk(sk);
2492         struct tcp_sock *tp = tcp_sk(sk);
2493         u32 prior_snd_una = tp->snd_una;
2494         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2495         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2496         u32 prior_in_flight;
2497         s32 seq_rtt;
2498         int prior_packets;
2499
2500         /* If the ack is newer than sent or older than previous acks
2501          * then we can probably ignore it.
2502          */
2503         if (after(ack, tp->snd_nxt))
2504                 goto uninteresting_ack;
2505
2506         if (before(ack, prior_snd_una))
2507                 goto old_ack;
2508
2509         if (sysctl_tcp_abc && icsk->icsk_ca_state < TCP_CA_CWR)
2510                 tp->bytes_acked += ack - prior_snd_una;
2511
2512         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2513                 /* Window is constant, pure forward advance.
2514                  * No more checks are required.
2515                  * Note, we use the fact that SND.UNA>=SND.WL2.
2516                  */
2517                 tcp_update_wl(tp, ack, ack_seq);
2518                 tp->snd_una = ack;
2519                 flag |= FLAG_WIN_UPDATE;
2520
2521                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2522
2523                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2524         } else {
2525                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2526                         flag |= FLAG_DATA;
2527                 else
2528                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2529
2530                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2531
2532                 if (TCP_SKB_CB(skb)->sacked)
2533                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2534
2535                 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2536                         flag |= FLAG_ECE;
2537
2538                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2539         }
2540
2541         /* We passed data and got it acked, remove any soft error
2542          * log. Something worked...
2543          */
2544         sk->sk_err_soft = 0;
2545         tp->rcv_tstamp = tcp_time_stamp;
2546         prior_packets = tp->packets_out;
2547         if (!prior_packets)
2548                 goto no_queue;
2549
2550         prior_in_flight = tcp_packets_in_flight(tp);
2551
2552         /* See if we can take anything off of the retransmit queue. */
2553         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2554
2555         if (tp->frto_counter)
2556                 tcp_process_frto(sk, prior_snd_una);
2557
2558         if (tcp_ack_is_dubious(sk, flag)) {
2559                 /* Advance CWND, if state allows this. */
2560                 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
2561                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2562                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2563         } else {
2564                 if ((flag & FLAG_DATA_ACKED))
2565                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2566         }
2567
2568         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2569                 dst_confirm(sk->sk_dst_cache);
2570
2571         return 1;
2572
2573 no_queue:
2574         icsk->icsk_probes_out = 0;
2575
2576         /* If this ack opens up a zero window, clear backoff.  It was
2577          * being used to time the probes, and is probably far higher than
2578          * it needs to be for normal retransmission.
2579          */
2580         if (sk->sk_send_head)
2581                 tcp_ack_probe(sk);
2582         return 1;
2583
2584 old_ack:
2585         if (TCP_SKB_CB(skb)->sacked)
2586                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2587
2588 uninteresting_ack:
2589         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2590         return 0;
2591 }
2592
2593
2594 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2595  * But, this can also be called on packets in the established flow when
2596  * the fast version below fails.
2597  */
2598 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2599 {
2600         unsigned char *ptr;
2601         struct tcphdr *th = skb->h.th;
2602         int length=(th->doff*4)-sizeof(struct tcphdr);
2603
2604         ptr = (unsigned char *)(th + 1);
2605         opt_rx->saw_tstamp = 0;
2606
2607         while(length>0) {
2608                 int opcode=*ptr++;
2609                 int opsize;
2610
2611                 switch (opcode) {
2612                         case TCPOPT_EOL:
2613                                 return;
2614                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2615                                 length--;
2616                                 continue;
2617                         default:
2618                                 opsize=*ptr++;
2619                                 if (opsize < 2) /* "silly options" */
2620                                         return;
2621                                 if (opsize > length)
2622                                         return; /* don't parse partial options */
2623                                 switch(opcode) {
2624                                 case TCPOPT_MSS:
2625                                         if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2626                                                 u16 in_mss = ntohs(get_unaligned((__u16 *)ptr));
2627                                                 if (in_mss) {
2628                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2629                                                                 in_mss = opt_rx->user_mss;
2630                                                         opt_rx->mss_clamp = in_mss;
2631                                                 }
2632                                         }
2633                                         break;
2634                                 case TCPOPT_WINDOW:
2635                                         if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2636                                                 if (sysctl_tcp_window_scaling) {
2637                                                         __u8 snd_wscale = *(__u8 *) ptr;
2638                                                         opt_rx->wscale_ok = 1;
2639                                                         if (snd_wscale > 14) {
2640                                                                 if(net_ratelimit())
2641                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2642                                                                                "scaling value %d >14 received.\n",
2643                                                                                snd_wscale);
2644                                                                 snd_wscale = 14;
2645                                                         }
2646                                                         opt_rx->snd_wscale = snd_wscale;
2647                                                 }
2648                                         break;
2649                                 case TCPOPT_TIMESTAMP:
2650                                         if(opsize==TCPOLEN_TIMESTAMP) {
2651                                                 if ((estab && opt_rx->tstamp_ok) ||
2652                                                     (!estab && sysctl_tcp_timestamps)) {
2653                                                         opt_rx->saw_tstamp = 1;
2654                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr));
2655                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4)));
2656                                                 }
2657                                         }
2658                                         break;
2659                                 case TCPOPT_SACK_PERM:
2660                                         if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2661                                                 if (sysctl_tcp_sack) {
2662                                                         opt_rx->sack_ok = 1;
2663                                                         tcp_sack_reset(opt_rx);
2664                                                 }
2665                                         }
2666                                         break;
2667
2668                                 case TCPOPT_SACK:
2669                                         if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2670                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2671                                            opt_rx->sack_ok) {
2672                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2673                                         }
2674                                 };
2675                                 ptr+=opsize-2;
2676                                 length-=opsize;
2677                 };
2678         }
2679 }
2680
2681 /* Fast parse options. This hopes to only see timestamps.
2682  * If it is wrong it falls back on tcp_parse_options().
2683  */
2684 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2685                                   struct tcp_sock *tp)
2686 {
2687         if (th->doff == sizeof(struct tcphdr)>>2) {
2688                 tp->rx_opt.saw_tstamp = 0;
2689                 return 0;
2690         } else if (tp->rx_opt.tstamp_ok &&
2691                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2692                 __u32 *ptr = (__u32 *)(th + 1);
2693                 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2694                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2695                         tp->rx_opt.saw_tstamp = 1;
2696                         ++ptr;
2697                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2698                         ++ptr;
2699                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2700                         return 1;
2701                 }
2702         }
2703         tcp_parse_options(skb, &tp->rx_opt, 1);
2704         return 1;
2705 }
2706
2707 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2708 {
2709         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2710         tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2711 }
2712
2713 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2714 {
2715         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2716                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2717                  * extra check below makes sure this can only happen
2718                  * for pure ACK frames.  -DaveM
2719                  *
2720                  * Not only, also it occurs for expired timestamps.
2721                  */
2722
2723                 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2724                    xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2725                         tcp_store_ts_recent(tp);
2726         }
2727 }
2728
2729 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2730  *
2731  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2732  * it can pass through stack. So, the following predicate verifies that
2733  * this segment is not used for anything but congestion avoidance or
2734  * fast retransmit. Moreover, we even are able to eliminate most of such
2735  * second order effects, if we apply some small "replay" window (~RTO)
2736  * to timestamp space.
2737  *
2738  * All these measures still do not guarantee that we reject wrapped ACKs
2739  * on networks with high bandwidth, when sequence space is recycled fastly,
2740  * but it guarantees that such events will be very rare and do not affect
2741  * connection seriously. This doesn't look nice, but alas, PAWS is really
2742  * buggy extension.
2743  *
2744  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2745  * states that events when retransmit arrives after original data are rare.
2746  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2747  * the biggest problem on large power networks even with minor reordering.
2748  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2749  * up to bandwidth of 18Gigabit/sec. 8) ]
2750  */
2751
2752 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2753 {
2754         struct tcp_sock *tp = tcp_sk(sk);
2755         struct tcphdr *th = skb->h.th;
2756         u32 seq = TCP_SKB_CB(skb)->seq;
2757         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2758
2759         return (/* 1. Pure ACK with correct sequence number. */
2760                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2761
2762                 /* 2. ... and duplicate ACK. */
2763                 ack == tp->snd_una &&
2764
2765                 /* 3. ... and does not update window. */
2766                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2767
2768                 /* 4. ... and sits in replay window. */
2769                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
2770 }
2771
2772 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
2773 {
2774         const struct tcp_sock *tp = tcp_sk(sk);
2775         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2776                 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2777                 !tcp_disordered_ack(sk, skb));
2778 }
2779
2780 /* Check segment sequence number for validity.
2781  *
2782  * Segment controls are considered valid, if the segment
2783  * fits to the window after truncation to the window. Acceptability
2784  * of data (and SYN, FIN, of course) is checked separately.
2785  * See tcp_data_queue(), for example.
2786  *
2787  * Also, controls (RST is main one) are accepted using RCV.WUP instead
2788  * of RCV.NXT. Peer still did not advance his SND.UNA when we
2789  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2790  * (borrowed from freebsd)
2791  */
2792
2793 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2794 {
2795         return  !before(end_seq, tp->rcv_wup) &&
2796                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2797 }
2798
2799 /* When we get a reset we do this. */
2800 static void tcp_reset(struct sock *sk)
2801 {
2802         /* We want the right error as BSD sees it (and indeed as we do). */
2803         switch (sk->sk_state) {
2804                 case TCP_SYN_SENT:
2805                         sk->sk_err = ECONNREFUSED;
2806                         break;
2807                 case TCP_CLOSE_WAIT:
2808                         sk->sk_err = EPIPE;
2809                         break;
2810                 case TCP_CLOSE:
2811                         return;
2812                 default:
2813                         sk->sk_err = ECONNRESET;
2814         }
2815
2816         if (!sock_flag(sk, SOCK_DEAD))
2817                 sk->sk_error_report(sk);
2818
2819         tcp_done(sk);
2820 }
2821
2822 /*
2823  *      Process the FIN bit. This now behaves as it is supposed to work
2824  *      and the FIN takes effect when it is validly part of sequence
2825  *      space. Not before when we get holes.
2826  *
2827  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2828  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
2829  *      TIME-WAIT)
2830  *
2831  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
2832  *      close and we go into CLOSING (and later onto TIME-WAIT)
2833  *
2834  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2835  */
2836 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2837 {
2838         struct tcp_sock *tp = tcp_sk(sk);
2839
2840         inet_csk_schedule_ack(sk);
2841
2842         sk->sk_shutdown |= RCV_SHUTDOWN;
2843         sock_set_flag(sk, SOCK_DONE);
2844
2845         switch (sk->sk_state) {
2846                 case TCP_SYN_RECV:
2847                 case TCP_ESTABLISHED:
2848                         /* Move to CLOSE_WAIT */
2849                         tcp_set_state(sk, TCP_CLOSE_WAIT);
2850                         inet_csk(sk)->icsk_ack.pingpong = 1;
2851                         break;
2852
2853                 case TCP_CLOSE_WAIT:
2854                 case TCP_CLOSING:
2855                         /* Received a retransmission of the FIN, do
2856                          * nothing.
2857                          */
2858                         break;
2859                 case TCP_LAST_ACK:
2860                         /* RFC793: Remain in the LAST-ACK state. */
2861                         break;
2862
2863                 case TCP_FIN_WAIT1:
2864                         /* This case occurs when a simultaneous close
2865                          * happens, we must ack the received FIN and
2866                          * enter the CLOSING state.
2867                          */
2868                         tcp_send_ack(sk);
2869                         tcp_set_state(sk, TCP_CLOSING);
2870                         break;
2871                 case TCP_FIN_WAIT2:
2872                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
2873                         tcp_send_ack(sk);
2874                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2875                         break;
2876                 default:
2877                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2878                          * cases we should never reach this piece of code.
2879                          */
2880                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2881                                __FUNCTION__, sk->sk_state);
2882                         break;
2883         };
2884
2885         /* It _is_ possible, that we have something out-of-order _after_ FIN.
2886          * Probably, we should reset in this case. For now drop them.
2887          */
2888         __skb_queue_purge(&tp->out_of_order_queue);
2889         if (tp->rx_opt.sack_ok)
2890                 tcp_sack_reset(&tp->rx_opt);
2891         sk_stream_mem_reclaim(sk);
2892
2893         if (!sock_flag(sk, SOCK_DEAD)) {
2894                 sk->sk_state_change(sk);
2895
2896                 /* Do not send POLL_HUP for half duplex close. */
2897                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2898                     sk->sk_state == TCP_CLOSE)
2899                         sk_wake_async(sk, 1, POLL_HUP);
2900                 else
2901                         sk_wake_async(sk, 1, POLL_IN);
2902         }
2903 }
2904
2905 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2906 {
2907         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2908                 if (before(seq, sp->start_seq))
2909                         sp->start_seq = seq;
2910                 if (after(end_seq, sp->end_seq))
2911                         sp->end_seq = end_seq;
2912                 return 1;
2913         }
2914         return 0;
2915 }
2916
2917 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
2918 {
2919         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2920                 if (before(seq, tp->rcv_nxt))
2921                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
2922                 else
2923                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
2924
2925                 tp->rx_opt.dsack = 1;
2926                 tp->duplicate_sack[0].start_seq = seq;
2927                 tp->duplicate_sack[0].end_seq = end_seq;
2928                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
2929         }
2930 }
2931
2932 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
2933 {
2934         if (!tp->rx_opt.dsack)
2935                 tcp_dsack_set(tp, seq, end_seq);
2936         else
2937                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2938 }
2939
2940 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2941 {
2942         struct tcp_sock *tp = tcp_sk(sk);
2943
2944         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2945             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2946                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2947                 tcp_enter_quickack_mode(sk);
2948
2949                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2950                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2951
2952                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2953                                 end_seq = tp->rcv_nxt;
2954                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2955                 }
2956         }
2957
2958         tcp_send_ack(sk);
2959 }
2960
2961 /* These routines update the SACK block as out-of-order packets arrive or
2962  * in-order packets close up the sequence space.
2963  */
2964 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
2965 {
2966         int this_sack;
2967         struct tcp_sack_block *sp = &tp->selective_acks[0];
2968         struct tcp_sack_block *swalk = sp+1;
2969
2970         /* See if the recent change to the first SACK eats into
2971          * or hits the sequence space of other SACK blocks, if so coalesce.
2972          */
2973         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
2974                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2975                         int i;
2976
2977                         /* Zap SWALK, by moving every further SACK up by one slot.
2978                          * Decrease num_sacks.
2979                          */
2980                         tp->rx_opt.num_sacks--;
2981                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2982                         for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
2983                                 sp[i] = sp[i+1];
2984                         continue;
2985                 }
2986                 this_sack++, swalk++;
2987         }
2988 }
2989
2990 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2991 {
2992         __u32 tmp;
2993
2994         tmp = sack1->start_seq;
2995         sack1->start_seq = sack2->start_seq;
2996         sack2->start_seq = tmp;
2997
2998         tmp = sack1->end_seq;
2999         sack1->end_seq = sack2->end_seq;
3000         sack2->end_seq = tmp;
3001 }
3002
3003 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3004 {
3005         struct tcp_sock *tp = tcp_sk(sk);
3006         struct tcp_sack_block *sp = &tp->selective_acks[0];
3007         int cur_sacks = tp->rx_opt.num_sacks;
3008         int this_sack;
3009
3010         if (!cur_sacks)
3011                 goto new_sack;
3012
3013         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3014                 if (tcp_sack_extend(sp, seq, end_seq)) {
3015                         /* Rotate this_sack to the first one. */
3016                         for (; this_sack>0; this_sack--, sp--)
3017                                 tcp_sack_swap(sp, sp-1);
3018                         if (cur_sacks > 1)
3019                                 tcp_sack_maybe_coalesce(tp);
3020                         return;
3021                 }
3022         }
3023
3024         /* Could not find an adjacent existing SACK, build a new one,
3025          * put it at the front, and shift everyone else down.  We
3026          * always know there is at least one SACK present already here.
3027          *
3028          * If the sack array is full, forget about the last one.
3029          */
3030         if (this_sack >= 4) {
3031                 this_sack--;
3032                 tp->rx_opt.num_sacks--;
3033                 sp--;
3034         }
3035         for(; this_sack > 0; this_sack--, sp--)
3036                 *sp = *(sp-1);
3037
3038 new_sack:
3039         /* Build the new head SACK, and we're done. */
3040         sp->start_seq = seq;
3041         sp->end_seq = end_seq;
3042         tp->rx_opt.num_sacks++;
3043         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3044 }
3045
3046 /* RCV.NXT advances, some SACKs should be eaten. */
3047
3048 static void tcp_sack_remove(struct tcp_sock *tp)
3049 {
3050         struct tcp_sack_block *sp = &tp->selective_acks[0];
3051         int num_sacks = tp->rx_opt.num_sacks;
3052         int this_sack;
3053
3054         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3055         if (skb_queue_empty(&tp->out_of_order_queue)) {
3056                 tp->rx_opt.num_sacks = 0;
3057                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3058                 return;
3059         }
3060
3061         for(this_sack = 0; this_sack < num_sacks; ) {
3062                 /* Check if the start of the sack is covered by RCV.NXT. */
3063                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3064                         int i;
3065
3066                         /* RCV.NXT must cover all the block! */
3067                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3068
3069                         /* Zap this SACK, by moving forward any other SACKS. */
3070                         for (i=this_sack+1; i < num_sacks; i++)
3071                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3072                         num_sacks--;
3073                         continue;
3074                 }
3075                 this_sack++;
3076                 sp++;
3077         }
3078         if (num_sacks != tp->rx_opt.num_sacks) {
3079                 tp->rx_opt.num_sacks = num_sacks;
3080                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3081         }
3082 }
3083
3084 /* This one checks to see if we can put data from the
3085  * out_of_order queue into the receive_queue.
3086  */
3087 static void tcp_ofo_queue(struct sock *sk)
3088 {
3089         struct tcp_sock *tp = tcp_sk(sk);
3090         __u32 dsack_high = tp->rcv_nxt;
3091         struct sk_buff *skb;
3092
3093         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3094                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3095                         break;
3096
3097                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3098                         __u32 dsack = dsack_high;
3099                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3100                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3101                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3102                 }
3103
3104                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3105                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3106                         __skb_unlink(skb, &tp->out_of_order_queue);
3107                         __kfree_skb(skb);
3108                         continue;
3109                 }
3110                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3111                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3112                            TCP_SKB_CB(skb)->end_seq);
3113
3114                 __skb_unlink(skb, &tp->out_of_order_queue);
3115                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3116                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3117                 if(skb->h.th->fin)
3118                         tcp_fin(skb, sk, skb->h.th);
3119         }
3120 }
3121
3122 static int tcp_prune_queue(struct sock *sk);
3123
3124 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3125 {
3126         struct tcphdr *th = skb->h.th;
3127         struct tcp_sock *tp = tcp_sk(sk);
3128         int eaten = -1;
3129
3130         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3131                 goto drop;
3132
3133         __skb_pull(skb, th->doff*4);
3134
3135         TCP_ECN_accept_cwr(tp, skb);
3136
3137         if (tp->rx_opt.dsack) {
3138                 tp->rx_opt.dsack = 0;
3139                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3140                                                     4 - tp->rx_opt.tstamp_ok);
3141         }
3142
3143         /*  Queue data for delivery to the user.
3144          *  Packets in sequence go to the receive queue.
3145          *  Out of sequence packets to the out_of_order_queue.
3146          */
3147         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3148                 if (tcp_receive_window(tp) == 0)
3149                         goto out_of_window;
3150
3151                 /* Ok. In sequence. In window. */
3152                 if (tp->ucopy.task == current &&
3153                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3154                     sock_owned_by_user(sk) && !tp->urg_data) {
3155                         int chunk = min_t(unsigned int, skb->len,
3156                                                         tp->ucopy.len);
3157
3158                         __set_current_state(TASK_RUNNING);
3159
3160                         local_bh_enable();
3161                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3162                                 tp->ucopy.len -= chunk;
3163                                 tp->copied_seq += chunk;
3164                                 eaten = (chunk == skb->len && !th->fin);
3165                                 tcp_rcv_space_adjust(sk);
3166                         }
3167                         local_bh_disable();
3168                 }
3169
3170                 if (eaten <= 0) {
3171 queue_and_out:
3172                         if (eaten < 0 &&
3173                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3174                              !sk_stream_rmem_schedule(sk, skb))) {
3175                                 if (tcp_prune_queue(sk) < 0 ||
3176                                     !sk_stream_rmem_schedule(sk, skb))
3177                                         goto drop;
3178                         }
3179                         sk_stream_set_owner_r(skb, sk);
3180                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3181                 }
3182                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3183                 if(skb->len)
3184                         tcp_event_data_recv(sk, tp, skb);
3185                 if(th->fin)
3186                         tcp_fin(skb, sk, th);
3187
3188                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3189                         tcp_ofo_queue(sk);
3190
3191                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3192                          * gap in queue is filled.
3193                          */
3194                         if (skb_queue_empty(&tp->out_of_order_queue))
3195                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3196                 }
3197
3198                 if (tp->rx_opt.num_sacks)
3199                         tcp_sack_remove(tp);
3200
3201                 tcp_fast_path_check(sk, tp);
3202
3203                 if (eaten > 0)
3204                         __kfree_skb(skb);
3205                 else if (!sock_flag(sk, SOCK_DEAD))
3206                         sk->sk_data_ready(sk, 0);
3207                 return;
3208         }
3209
3210         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3211                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3212                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3213                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3214
3215 out_of_window:
3216                 tcp_enter_quickack_mode(sk);
3217                 inet_csk_schedule_ack(sk);
3218 drop:
3219                 __kfree_skb(skb);
3220                 return;
3221         }
3222
3223         /* Out of window. F.e. zero window probe. */
3224         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3225                 goto out_of_window;
3226
3227         tcp_enter_quickack_mode(sk);
3228
3229         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3230                 /* Partial packet, seq < rcv_next < end_seq */
3231                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3232                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3233                            TCP_SKB_CB(skb)->end_seq);
3234
3235                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3236                 
3237                 /* If window is closed, drop tail of packet. But after
3238                  * remembering D-SACK for its head made in previous line.
3239                  */
3240                 if (!tcp_receive_window(tp))
3241                         goto out_of_window;
3242                 goto queue_and_out;
3243         }
3244
3245         TCP_ECN_check_ce(tp, skb);
3246
3247         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3248             !sk_stream_rmem_schedule(sk, skb)) {
3249                 if (tcp_prune_queue(sk) < 0 ||
3250                     !sk_stream_rmem_schedule(sk, skb))
3251                         goto drop;
3252         }
3253
3254         /* Disable header prediction. */
3255         tp->pred_flags = 0;
3256         inet_csk_schedule_ack(sk);
3257
3258         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3259                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3260
3261         sk_stream_set_owner_r(skb, sk);
3262
3263         if (!skb_peek(&tp->out_of_order_queue)) {
3264                 /* Initial out of order segment, build 1 SACK. */
3265                 if (tp->rx_opt.sack_ok) {
3266                         tp->rx_opt.num_sacks = 1;
3267                         tp->rx_opt.dsack     = 0;
3268                         tp->rx_opt.eff_sacks = 1;
3269                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3270                         tp->selective_acks[0].end_seq =
3271                                                 TCP_SKB_CB(skb)->end_seq;
3272                 }
3273                 __skb_queue_head(&tp->out_of_order_queue,skb);
3274         } else {
3275                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3276                 u32 seq = TCP_SKB_CB(skb)->seq;
3277                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3278
3279                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3280                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3281
3282                         if (!tp->rx_opt.num_sacks ||
3283                             tp->selective_acks[0].end_seq != seq)
3284                                 goto add_sack;
3285
3286                         /* Common case: data arrive in order after hole. */
3287                         tp->selective_acks[0].end_seq = end_seq;
3288                         return;
3289                 }
3290
3291                 /* Find place to insert this segment. */
3292                 do {
3293                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3294                                 break;
3295                 } while ((skb1 = skb1->prev) !=
3296                          (struct sk_buff*)&tp->out_of_order_queue);
3297
3298                 /* Do skb overlap to previous one? */
3299                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3300                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3301                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3302                                 /* All the bits are present. Drop. */
3303                                 __kfree_skb(skb);
3304                                 tcp_dsack_set(tp, seq, end_seq);
3305                                 goto add_sack;
3306                         }
3307                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3308                                 /* Partial overlap. */
3309                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3310                         } else {
3311                                 skb1 = skb1->prev;
3312                         }
3313                 }
3314                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3315                 
3316                 /* And clean segments covered by new one as whole. */
3317                 while ((skb1 = skb->next) !=
3318                        (struct sk_buff*)&tp->out_of_order_queue &&
3319                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3320                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3321                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3322                                break;
3323                        }
3324                        __skb_unlink(skb1, &tp->out_of_order_queue);
3325                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3326                        __kfree_skb(skb1);
3327                 }
3328
3329 add_sack:
3330                 if (tp->rx_opt.sack_ok)
3331                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3332         }
3333 }
3334
3335 /* Collapse contiguous sequence of skbs head..tail with
3336  * sequence numbers start..end.
3337  * Segments with FIN/SYN are not collapsed (only because this
3338  * simplifies code)
3339  */
3340 static void
3341 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3342              struct sk_buff *head, struct sk_buff *tail,
3343              u32 start, u32 end)
3344 {
3345         struct sk_buff *skb;
3346
3347         /* First, check that queue is collapsible and find
3348          * the point where collapsing can be useful. */
3349         for (skb = head; skb != tail; ) {
3350                 /* No new bits? It is possible on ofo queue. */
3351                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3352                         struct sk_buff *next = skb->next;
3353                         __skb_unlink(skb, list);
3354                         __kfree_skb(skb);
3355                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3356                         skb = next;
3357                         continue;
3358                 }
3359
3360                 /* The first skb to collapse is:
3361                  * - not SYN/FIN and
3362                  * - bloated or contains data before "start" or
3363                  *   overlaps to the next one.
3364                  */
3365                 if (!skb->h.th->syn && !skb->h.th->fin &&
3366                     (tcp_win_from_space(skb->truesize) > skb->len ||
3367                      before(TCP_SKB_CB(skb)->seq, start) ||
3368                      (skb->next != tail &&
3369                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3370                         break;
3371
3372                 /* Decided to skip this, advance start seq. */
3373                 start = TCP_SKB_CB(skb)->end_seq;
3374                 skb = skb->next;
3375         }
3376         if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3377                 return;
3378
3379         while (before(start, end)) {
3380                 struct sk_buff *nskb;
3381                 int header = skb_headroom(skb);
3382                 int copy = SKB_MAX_ORDER(header, 0);
3383
3384                 /* Too big header? This can happen with IPv6. */
3385                 if (copy < 0)
3386                         return;
3387                 if (end-start < copy)
3388                         copy = end-start;
3389                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3390                 if (!nskb)
3391                         return;
3392                 skb_reserve(nskb, header);
3393                 memcpy(nskb->head, skb->head, header);
3394                 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3395                 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3396                 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3397                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3398                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3399                 __skb_insert(nskb, skb->prev, skb, list);
3400                 sk_stream_set_owner_r(nskb, sk);
3401
3402                 /* Copy data, releasing collapsed skbs. */
3403                 while (copy > 0) {
3404                         int offset = start - TCP_SKB_CB(skb)->seq;
3405                         int size = TCP_SKB_CB(skb)->end_seq - start;
3406
3407                         BUG_ON(offset < 0);
3408                         if (size > 0) {
3409                                 size = min(copy, size);
3410                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3411                                         BUG();
3412                                 TCP_SKB_CB(nskb)->end_seq += size;
3413                                 copy -= size;
3414                                 start += size;
3415                         }
3416                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3417                                 struct sk_buff *next = skb->next;
3418                                 __skb_unlink(skb, list);
3419                                 __kfree_skb(skb);
3420                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3421                                 skb = next;
3422                                 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3423                                         return;
3424                         }
3425                 }
3426         }
3427 }
3428
3429 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3430  * and tcp_collapse() them until all the queue is collapsed.
3431  */
3432 static void tcp_collapse_ofo_queue(struct sock *sk)
3433 {
3434         struct tcp_sock *tp = tcp_sk(sk);
3435         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3436         struct sk_buff *head;
3437         u32 start, end;
3438
3439         if (skb == NULL)
3440                 return;
3441
3442         start = TCP_SKB_CB(skb)->seq;
3443         end = TCP_SKB_CB(skb)->end_seq;
3444         head = skb;
3445
3446         for (;;) {
3447                 skb = skb->next;
3448
3449                 /* Segment is terminated when we see gap or when
3450                  * we are at the end of all the queue. */
3451                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3452                     after(TCP_SKB_CB(skb)->seq, end) ||
3453                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3454                         tcp_collapse(sk, &tp->out_of_order_queue,
3455                                      head, skb, start, end);
3456                         head = skb;
3457                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3458                                 break;
3459                         /* Start new segment */
3460                         start = TCP_SKB_CB(skb)->seq;
3461                         end = TCP_SKB_CB(skb)->end_seq;
3462                 } else {
3463                         if (before(TCP_SKB_CB(skb)->seq, start))
3464                                 start = TCP_SKB_CB(skb)->seq;
3465                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3466                                 end = TCP_SKB_CB(skb)->end_seq;
3467                 }
3468         }
3469 }
3470
3471 /* Reduce allocated memory if we can, trying to get
3472  * the socket within its memory limits again.
3473  *
3474  * Return less than zero if we should start dropping frames
3475  * until the socket owning process reads some of the data
3476  * to stabilize the situation.
3477  */
3478 static int tcp_prune_queue(struct sock *sk)
3479 {
3480         struct tcp_sock *tp = tcp_sk(sk); 
3481
3482         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3483
3484         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3485
3486         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3487                 tcp_clamp_window(sk, tp);
3488         else if (tcp_memory_pressure)
3489                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3490
3491         tcp_collapse_ofo_queue(sk);
3492         tcp_collapse(sk, &sk->sk_receive_queue,
3493                      sk->sk_receive_queue.next,
3494                      (struct sk_buff*)&sk->sk_receive_queue,
3495                      tp->copied_seq, tp->rcv_nxt);
3496         sk_stream_mem_reclaim(sk);
3497
3498         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3499                 return 0;
3500
3501         /* Collapsing did not help, destructive actions follow.
3502          * This must not ever occur. */
3503
3504         /* First, purge the out_of_order queue. */
3505         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3506                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3507                 __skb_queue_purge(&tp->out_of_order_queue);
3508
3509                 /* Reset SACK state.  A conforming SACK implementation will
3510                  * do the same at a timeout based retransmit.  When a connection
3511                  * is in a sad state like this, we care only about integrity
3512                  * of the connection not performance.
3513                  */
3514                 if (tp->rx_opt.sack_ok)
3515                         tcp_sack_reset(&tp->rx_opt);
3516                 sk_stream_mem_reclaim(sk);
3517         }
3518
3519         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3520                 return 0;
3521
3522         /* If we are really being abused, tell the caller to silently
3523          * drop receive data on the floor.  It will get retransmitted
3524          * and hopefully then we'll have sufficient space.
3525          */
3526         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3527
3528         /* Massive buffer overcommit. */
3529         tp->pred_flags = 0;
3530         return -1;
3531 }
3532
3533
3534 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3535  * As additional protections, we do not touch cwnd in retransmission phases,
3536  * and if application hit its sndbuf limit recently.
3537  */
3538 void tcp_cwnd_application_limited(struct sock *sk)
3539 {
3540         struct tcp_sock *tp = tcp_sk(sk);
3541
3542         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3543             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3544                 /* Limited by application or receiver window. */
3545                 u32 win_used = max(tp->snd_cwnd_used, 2U);
3546                 if (win_used < tp->snd_cwnd) {
3547                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3548                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3549                 }
3550                 tp->snd_cwnd_used = 0;
3551         }
3552         tp->snd_cwnd_stamp = tcp_time_stamp;
3553 }
3554
3555 static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3556 {
3557         /* If the user specified a specific send buffer setting, do
3558          * not modify it.
3559          */
3560         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3561                 return 0;
3562
3563         /* If we are under global TCP memory pressure, do not expand.  */
3564         if (tcp_memory_pressure)
3565                 return 0;
3566
3567         /* If we are under soft global TCP memory pressure, do not expand.  */
3568         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3569                 return 0;
3570
3571         /* If we filled the congestion window, do not expand.  */
3572         if (tp->packets_out >= tp->snd_cwnd)
3573                 return 0;
3574
3575         return 1;
3576 }
3577
3578 /* When incoming ACK allowed to free some skb from write_queue,
3579  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3580  * on the exit from tcp input handler.
3581  *
3582  * PROBLEM: sndbuf expansion does not work well with largesend.
3583  */
3584 static void tcp_new_space(struct sock *sk)
3585 {
3586         struct tcp_sock *tp = tcp_sk(sk);
3587
3588         if (tcp_should_expand_sndbuf(sk, tp)) {
3589                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3590                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3591                     demanded = max_t(unsigned int, tp->snd_cwnd,
3592                                                    tp->reordering + 1);
3593                 sndmem *= 2*demanded;
3594                 if (sndmem > sk->sk_sndbuf)
3595                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3596                 tp->snd_cwnd_stamp = tcp_time_stamp;
3597         }
3598
3599         sk->sk_write_space(sk);
3600 }
3601
3602 static void tcp_check_space(struct sock *sk)
3603 {
3604         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3605                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3606                 if (sk->sk_socket &&
3607                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3608                         tcp_new_space(sk);
3609         }
3610 }
3611
3612 static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3613 {
3614         tcp_push_pending_frames(sk, tp);
3615         tcp_check_space(sk);
3616 }
3617
3618 /*
3619  * Check if sending an ack is needed.
3620  */
3621 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3622 {
3623         struct tcp_sock *tp = tcp_sk(sk);
3624
3625             /* More than one full frame received... */
3626         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3627              /* ... and right edge of window advances far enough.
3628               * (tcp_recvmsg() will send ACK otherwise). Or...
3629               */
3630              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3631             /* We ACK each frame or... */
3632             tcp_in_quickack_mode(sk) ||
3633             /* We have out of order data. */
3634             (ofo_possible &&
3635              skb_peek(&tp->out_of_order_queue))) {
3636                 /* Then ack it now */
3637                 tcp_send_ack(sk);
3638         } else {
3639                 /* Else, send delayed ack. */
3640                 tcp_send_delayed_ack(sk);
3641         }
3642 }
3643
3644 static inline void tcp_ack_snd_check(struct sock *sk)
3645 {
3646         if (!inet_csk_ack_scheduled(sk)) {
3647                 /* We sent a data segment already. */
3648                 return;
3649         }
3650         __tcp_ack_snd_check(sk, 1);
3651 }
3652
3653 /*
3654  *      This routine is only called when we have urgent data
3655  *      signaled. Its the 'slow' part of tcp_urg. It could be
3656  *      moved inline now as tcp_urg is only called from one
3657  *      place. We handle URGent data wrong. We have to - as
3658  *      BSD still doesn't use the correction from RFC961.
3659  *      For 1003.1g we should support a new option TCP_STDURG to permit
3660  *      either form (or just set the sysctl tcp_stdurg).
3661  */
3662  
3663 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3664 {
3665         struct tcp_sock *tp = tcp_sk(sk);
3666         u32 ptr = ntohs(th->urg_ptr);
3667
3668         if (ptr && !sysctl_tcp_stdurg)
3669                 ptr--;
3670         ptr += ntohl(th->seq);
3671
3672         /* Ignore urgent data that we've already seen and read. */
3673         if (after(tp->copied_seq, ptr))
3674                 return;
3675
3676         /* Do not replay urg ptr.
3677          *
3678          * NOTE: interesting situation not covered by specs.
3679          * Misbehaving sender may send urg ptr, pointing to segment,
3680          * which we already have in ofo queue. We are not able to fetch
3681          * such data and will stay in TCP_URG_NOTYET until will be eaten
3682          * by recvmsg(). Seems, we are not obliged to handle such wicked
3683          * situations. But it is worth to think about possibility of some
3684          * DoSes using some hypothetical application level deadlock.
3685          */
3686         if (before(ptr, tp->rcv_nxt))
3687                 return;
3688
3689         /* Do we already have a newer (or duplicate) urgent pointer? */
3690         if (tp->urg_data && !after(ptr, tp->urg_seq))
3691                 return;
3692
3693         /* Tell the world about our new urgent pointer. */
3694         sk_send_sigurg(sk);
3695
3696         /* We may be adding urgent data when the last byte read was
3697          * urgent. To do this requires some care. We cannot just ignore
3698          * tp->copied_seq since we would read the last urgent byte again
3699          * as data, nor can we alter copied_seq until this data arrives
3700          * or we break the semantics of SIOCATMARK (and thus sockatmark())
3701          *
3702          * NOTE. Double Dutch. Rendering to plain English: author of comment
3703          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3704          * and expect that both A and B disappear from stream. This is _wrong_.
3705          * Though this happens in BSD with high probability, this is occasional.
3706          * Any application relying on this is buggy. Note also, that fix "works"
3707          * only in this artificial test. Insert some normal data between A and B and we will
3708          * decline of BSD again. Verdict: it is better to remove to trap
3709          * buggy users.
3710          */
3711         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3712             !sock_flag(sk, SOCK_URGINLINE) &&
3713             tp->copied_seq != tp->rcv_nxt) {
3714                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3715                 tp->copied_seq++;
3716                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3717                         __skb_unlink(skb, &sk->sk_receive_queue);
3718                         __kfree_skb(skb);
3719                 }
3720         }
3721
3722         tp->urg_data   = TCP_URG_NOTYET;
3723         tp->urg_seq    = ptr;
3724
3725         /* Disable header prediction. */
3726         tp->pred_flags = 0;
3727 }
3728
3729 /* This is the 'fast' part of urgent handling. */
3730 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3731 {
3732         struct tcp_sock *tp = tcp_sk(sk);
3733
3734         /* Check if we get a new urgent pointer - normally not. */
3735         if (th->urg)
3736                 tcp_check_urg(sk,th);
3737
3738         /* Do we wait for any urgent data? - normally not... */
3739         if (tp->urg_data == TCP_URG_NOTYET) {
3740                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3741                           th->syn;
3742
3743                 /* Is the urgent pointer pointing into this packet? */   
3744                 if (ptr < skb->len) {
3745                         u8 tmp;
3746                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3747                                 BUG();
3748                         tp->urg_data = TCP_URG_VALID | tmp;
3749                         if (!sock_flag(sk, SOCK_DEAD))
3750                                 sk->sk_data_ready(sk, 0);
3751                 }
3752         }
3753 }
3754
3755 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3756 {
3757         struct tcp_sock *tp = tcp_sk(sk);
3758         int chunk = skb->len - hlen;
3759         int err;
3760
3761         local_bh_enable();
3762         if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3763                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3764         else
3765                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3766                                                        tp->ucopy.iov);
3767
3768         if (!err) {
3769                 tp->ucopy.len -= chunk;
3770                 tp->copied_seq += chunk;
3771                 tcp_rcv_space_adjust(sk);
3772         }
3773
3774         local_bh_disable();
3775         return err;
3776 }
3777
3778 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3779 {
3780         int result;
3781
3782         if (sock_owned_by_user(sk)) {
3783                 local_bh_enable();
3784                 result = __tcp_checksum_complete(skb);
3785                 local_bh_disable();
3786         } else {
3787                 result = __tcp_checksum_complete(skb);
3788         }
3789         return result;
3790 }
3791
3792 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3793 {
3794         return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3795                 __tcp_checksum_complete_user(sk, skb);
3796 }
3797
3798 #ifdef CONFIG_NET_DMA
3799 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
3800 {
3801         struct tcp_sock *tp = tcp_sk(sk);
3802         int chunk = skb->len - hlen;
3803         int dma_cookie;
3804         int copied_early = 0;
3805
3806         if (tp->ucopy.wakeup)
3807                 return 0;
3808
3809         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
3810                 tp->ucopy.dma_chan = get_softnet_dma();
3811
3812         if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) {
3813
3814                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
3815                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
3816
3817                 if (dma_cookie < 0)
3818                         goto out;
3819
3820                 tp->ucopy.dma_cookie = dma_cookie;
3821                 copied_early = 1;
3822
3823                 tp->ucopy.len -= chunk;
3824                 tp->copied_seq += chunk;
3825                 tcp_rcv_space_adjust(sk);
3826
3827                 if ((tp->ucopy.len == 0) ||
3828                     (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) ||
3829                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
3830                         tp->ucopy.wakeup = 1;
3831                         sk->sk_data_ready(sk, 0);
3832                 }
3833         } else if (chunk > 0) {
3834                 tp->ucopy.wakeup = 1;
3835                 sk->sk_data_ready(sk, 0);
3836         }
3837 out:
3838         return copied_early;
3839 }
3840 #endif /* CONFIG_NET_DMA */
3841
3842 /*
3843  *      TCP receive function for the ESTABLISHED state. 
3844  *
3845  *      It is split into a fast path and a slow path. The fast path is 
3846  *      disabled when:
3847  *      - A zero window was announced from us - zero window probing
3848  *        is only handled properly in the slow path. 
3849  *      - Out of order segments arrived.
3850  *      - Urgent data is expected.
3851  *      - There is no buffer space left
3852  *      - Unexpected TCP flags/window values/header lengths are received
3853  *        (detected by checking the TCP header against pred_flags) 
3854  *      - Data is sent in both directions. Fast path only supports pure senders
3855  *        or pure receivers (this means either the sequence number or the ack
3856  *        value must stay constant)
3857  *      - Unexpected TCP option.
3858  *
3859  *      When these conditions are not satisfied it drops into a standard 
3860  *      receive procedure patterned after RFC793 to handle all cases.
3861  *      The first three cases are guaranteed by proper pred_flags setting,
3862  *      the rest is checked inline. Fast processing is turned on in 
3863  *      tcp_data_queue when everything is OK.
3864  */
3865 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3866                         struct tcphdr *th, unsigned len)
3867 {
3868         struct tcp_sock *tp = tcp_sk(sk);
3869
3870         /*
3871          *      Header prediction.
3872          *      The code loosely follows the one in the famous 
3873          *      "30 instruction TCP receive" Van Jacobson mail.
3874          *      
3875          *      Van's trick is to deposit buffers into socket queue 
3876          *      on a device interrupt, to call tcp_recv function
3877          *      on the receive process context and checksum and copy
3878          *      the buffer to user space. smart...
3879          *
3880          *      Our current scheme is not silly either but we take the 
3881          *      extra cost of the net_bh soft interrupt processing...
3882          *      We do checksum and copy also but from device to kernel.
3883          */
3884
3885         tp->rx_opt.saw_tstamp = 0;
3886
3887         /*      pred_flags is 0xS?10 << 16 + snd_wnd
3888          *      if header_prediction is to be made
3889          *      'S' will always be tp->tcp_header_len >> 2
3890          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3891          *  turn it off (when there are holes in the receive 
3892          *       space for instance)
3893          *      PSH flag is ignored.
3894          */
3895
3896         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3897                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3898                 int tcp_header_len = tp->tcp_header_len;
3899
3900                 /* Timestamp header prediction: tcp_header_len
3901                  * is automatically equal to th->doff*4 due to pred_flags
3902                  * match.
3903                  */
3904
3905                 /* Check timestamp */
3906                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3907                         __u32 *ptr = (__u32 *)(th + 1);
3908
3909                         /* No? Slow path! */
3910                         if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3911                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3912                                 goto slow_path;
3913
3914                         tp->rx_opt.saw_tstamp = 1;
3915                         ++ptr; 
3916                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3917                         ++ptr;
3918                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3919
3920                         /* If PAWS failed, check it more carefully in slow path */
3921                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
3922                                 goto slow_path;
3923
3924                         /* DO NOT update ts_recent here, if checksum fails
3925                          * and timestamp was corrupted part, it will result
3926                          * in a hung connection since we will drop all
3927                          * future packets due to the PAWS test.
3928                          */
3929                 }
3930
3931                 if (len <= tcp_header_len) {
3932                         /* Bulk data transfer: sender */
3933                         if (len == tcp_header_len) {
3934                                 /* Predicted packet is in window by definition.
3935                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3936                                  * Hence, check seq<=rcv_wup reduces to:
3937                                  */
3938                                 if (tcp_header_len ==
3939                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3940                                     tp->rcv_nxt == tp->rcv_wup)
3941                                         tcp_store_ts_recent(tp);
3942
3943                                 /* We know that such packets are checksummed
3944                                  * on entry.
3945                                  */
3946                                 tcp_ack(sk, skb, 0);
3947                                 __kfree_skb(skb); 
3948                                 tcp_data_snd_check(sk, tp);
3949                                 return 0;
3950                         } else { /* Header too small */
3951                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3952                                 goto discard;
3953                         }
3954                 } else {
3955                         int eaten = 0;
3956                         int copied_early = 0;
3957
3958                         if (tp->copied_seq == tp->rcv_nxt &&
3959                             len - tcp_header_len <= tp->ucopy.len) {
3960 #ifdef CONFIG_NET_DMA
3961                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
3962                                         copied_early = 1;
3963                                         eaten = 1;
3964                                 }
3965 #endif
3966                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
3967                                         __set_current_state(TASK_RUNNING);
3968
3969                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
3970                                                 eaten = 1;
3971                                 }
3972                                 if (eaten) {
3973                                         /* Predicted packet is in window by definition.
3974                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3975                                          * Hence, check seq<=rcv_wup reduces to:
3976                                          */
3977                                         if (tcp_header_len ==
3978                                             (sizeof(struct tcphdr) +
3979                                              TCPOLEN_TSTAMP_ALIGNED) &&
3980                                             tp->rcv_nxt == tp->rcv_wup)
3981                                                 tcp_store_ts_recent(tp);
3982
3983                                         tcp_rcv_rtt_measure_ts(sk, skb);
3984
3985                                         __skb_pull(skb, tcp_header_len);
3986                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3987                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
3988                                 }
3989                                 if (copied_early)
3990                                         tcp_cleanup_rbuf(sk, skb->len);
3991                         }
3992                         if (!eaten) {
3993                                 if (tcp_checksum_complete_user(sk, skb))
3994                                         goto csum_error;
3995
3996                                 /* Predicted packet is in window by definition.
3997                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3998                                  * Hence, check seq<=rcv_wup reduces to:
3999                                  */
4000                                 if (tcp_header_len ==
4001                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4002                                     tp->rcv_nxt == tp->rcv_wup)
4003                                         tcp_store_ts_recent(tp);
4004
4005                                 tcp_rcv_rtt_measure_ts(sk, skb);
4006
4007                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4008                                         goto step5;
4009
4010                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4011
4012                                 /* Bulk data transfer: receiver */
4013                                 __skb_pull(skb,tcp_header_len);
4014                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4015                                 sk_stream_set_owner_r(skb, sk);
4016                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4017                         }
4018
4019                         tcp_event_data_recv(sk, tp, skb);
4020
4021                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4022                                 /* Well, only one small jumplet in fast path... */
4023                                 tcp_ack(sk, skb, FLAG_DATA);
4024                                 tcp_data_snd_check(sk, tp);
4025                                 if (!inet_csk_ack_scheduled(sk))
4026                                         goto no_ack;
4027                         }
4028
4029                         __tcp_ack_snd_check(sk, 0);
4030 no_ack:
4031 #ifdef CONFIG_NET_DMA
4032                         if (copied_early)
4033                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4034                         else
4035 #endif
4036                         if (eaten)
4037                                 __kfree_skb(skb);
4038                         else
4039                                 sk->sk_data_ready(sk, 0);
4040                         return 0;
4041                 }
4042         }
4043
4044 slow_path:
4045         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4046                 goto csum_error;
4047
4048         /*
4049          * RFC1323: H1. Apply PAWS check first.
4050          */
4051         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4052             tcp_paws_discard(sk, skb)) {
4053                 if (!th->rst) {
4054                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4055                         tcp_send_dupack(sk, skb);
4056                         goto discard;
4057                 }
4058                 /* Resets are accepted even if PAWS failed.
4059
4060                    ts_recent update must be made after we are sure
4061                    that the packet is in window.
4062                  */
4063         }
4064
4065         /*
4066          *      Standard slow path.
4067          */
4068
4069         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4070                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4071                  * (RST) segments are validated by checking their SEQ-fields."
4072                  * And page 69: "If an incoming segment is not acceptable,
4073                  * an acknowledgment should be sent in reply (unless the RST bit
4074                  * is set, if so drop the segment and return)".
4075                  */
4076                 if (!th->rst)
4077                         tcp_send_dupack(sk, skb);
4078                 goto discard;
4079         }
4080
4081         if(th->rst) {
4082                 tcp_reset(sk);
4083                 goto discard;
4084         }
4085
4086         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4087
4088         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4089                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4090                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4091                 tcp_reset(sk);
4092                 return 1;
4093         }
4094
4095 step5:
4096         if(th->ack)
4097                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4098
4099         tcp_rcv_rtt_measure_ts(sk, skb);
4100
4101         /* Process urgent data. */
4102         tcp_urg(sk, skb, th);
4103
4104         /* step 7: process the segment text */
4105         tcp_data_queue(sk, skb);
4106
4107         tcp_data_snd_check(sk, tp);
4108         tcp_ack_snd_check(sk);
4109         return 0;
4110
4111 csum_error:
4112         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4113
4114 discard:
4115         __kfree_skb(skb);
4116         return 0;
4117 }
4118
4119 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4120                                          struct tcphdr *th, unsigned len)
4121 {
4122         struct tcp_sock *tp = tcp_sk(sk);
4123         struct inet_connection_sock *icsk = inet_csk(sk);
4124         int saved_clamp = tp->rx_opt.mss_clamp;
4125
4126         tcp_parse_options(skb, &tp->rx_opt, 0);
4127
4128         if (th->ack) {
4129                 /* rfc793:
4130                  * "If the state is SYN-SENT then
4131                  *    first check the ACK bit
4132                  *      If the ACK bit is set
4133                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4134                  *        a reset (unless the RST bit is set, if so drop
4135                  *        the segment and return)"
4136                  *
4137                  *  We do not send data with SYN, so that RFC-correct
4138                  *  test reduces to:
4139                  */
4140                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4141                         goto reset_and_undo;
4142
4143                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4144                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4145                              tcp_time_stamp)) {
4146                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4147                         goto reset_and_undo;
4148                 }
4149
4150                 /* Now ACK is acceptable.
4151                  *
4152                  * "If the RST bit is set
4153                  *    If the ACK was acceptable then signal the user "error:
4154                  *    connection reset", drop the segment, enter CLOSED state,
4155                  *    delete TCB, and return."
4156                  */
4157
4158                 if (th->rst) {
4159                         tcp_reset(sk);
4160                         goto discard;
4161                 }
4162
4163                 /* rfc793:
4164                  *   "fifth, if neither of the SYN or RST bits is set then
4165                  *    drop the segment and return."
4166                  *
4167                  *    See note below!
4168                  *                                        --ANK(990513)
4169                  */
4170                 if (!th->syn)
4171                         goto discard_and_undo;
4172
4173                 /* rfc793:
4174                  *   "If the SYN bit is on ...
4175                  *    are acceptable then ...
4176                  *    (our SYN has been ACKed), change the connection
4177                  *    state to ESTABLISHED..."
4178                  */
4179
4180                 TCP_ECN_rcv_synack(tp, th);
4181                 if (tp->ecn_flags&TCP_ECN_OK)
4182                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4183
4184                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4185                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4186
4187                 /* Ok.. it's good. Set up sequence numbers and
4188                  * move to established.
4189                  */
4190                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4191                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4192
4193                 /* RFC1323: The window in SYN & SYN/ACK segments is
4194                  * never scaled.
4195                  */
4196                 tp->snd_wnd = ntohs(th->window);
4197                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4198
4199                 if (!tp->rx_opt.wscale_ok) {
4200                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4201                         tp->window_clamp = min(tp->window_clamp, 65535U);
4202                 }
4203
4204                 if (tp->rx_opt.saw_tstamp) {
4205                         tp->rx_opt.tstamp_ok       = 1;
4206                         tp->tcp_header_len =
4207                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4208                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4209                         tcp_store_ts_recent(tp);
4210                 } else {
4211                         tp->tcp_header_len = sizeof(struct tcphdr);
4212                 }
4213
4214                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4215                         tp->rx_opt.sack_ok |= 2;
4216
4217                 tcp_mtup_init(sk);
4218                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4219                 tcp_initialize_rcv_mss(sk);
4220
4221                 /* Remember, tcp_poll() does not lock socket!
4222                  * Change state from SYN-SENT only after copied_seq
4223                  * is initialized. */
4224                 tp->copied_seq = tp->rcv_nxt;
4225                 mb();
4226                 tcp_set_state(sk, TCP_ESTABLISHED);
4227
4228                 /* Make sure socket is routed, for correct metrics.  */
4229                 icsk->icsk_af_ops->rebuild_header(sk);
4230
4231                 tcp_init_metrics(sk);
4232
4233                 tcp_init_congestion_control(sk);
4234
4235                 /* Prevent spurious tcp_cwnd_restart() on first data
4236                  * packet.
4237                  */
4238                 tp->lsndtime = tcp_time_stamp;
4239
4240                 tcp_init_buffer_space(sk);
4241
4242                 if (sock_flag(sk, SOCK_KEEPOPEN))
4243                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4244
4245                 if (!tp->rx_opt.snd_wscale)
4246                         __tcp_fast_path_on(tp, tp->snd_wnd);
4247                 else
4248                         tp->pred_flags = 0;
4249
4250                 if (!sock_flag(sk, SOCK_DEAD)) {
4251                         sk->sk_state_change(sk);
4252                         sk_wake_async(sk, 0, POLL_OUT);
4253                 }
4254
4255                 if (sk->sk_write_pending ||
4256                     icsk->icsk_accept_queue.rskq_defer_accept ||
4257                     icsk->icsk_ack.pingpong) {
4258                         /* Save one ACK. Data will be ready after
4259                          * several ticks, if write_pending is set.
4260                          *
4261                          * It may be deleted, but with this feature tcpdumps
4262                          * look so _wonderfully_ clever, that I was not able
4263                          * to stand against the temptation 8)     --ANK
4264                          */
4265                         inet_csk_schedule_ack(sk);
4266                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4267                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4268                         tcp_incr_quickack(sk);
4269                         tcp_enter_quickack_mode(sk);
4270                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4271                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4272
4273 discard:
4274                         __kfree_skb(skb);
4275                         return 0;
4276                 } else {
4277                         tcp_send_ack(sk);
4278                 }
4279                 return -1;
4280         }
4281
4282         /* No ACK in the segment */
4283
4284         if (th->rst) {
4285                 /* rfc793:
4286                  * "If the RST bit is set
4287                  *
4288                  *      Otherwise (no ACK) drop the segment and return."
4289                  */
4290
4291                 goto discard_and_undo;
4292         }
4293
4294         /* PAWS check. */
4295         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4296                 goto discard_and_undo;
4297
4298         if (th->syn) {
4299                 /* We see SYN without ACK. It is attempt of
4300                  * simultaneous connect with crossed SYNs.
4301                  * Particularly, it can be connect to self.
4302                  */
4303                 tcp_set_state(sk, TCP_SYN_RECV);
4304
4305                 if (tp->rx_opt.saw_tstamp) {
4306                         tp->rx_opt.tstamp_ok = 1;
4307                         tcp_store_ts_recent(tp);
4308                         tp->tcp_header_len =
4309                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4310                 } else {
4311                         tp->tcp_header_len = sizeof(struct tcphdr);
4312                 }
4313
4314                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4315                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4316
4317                 /* RFC1323: The window in SYN & SYN/ACK segments is
4318                  * never scaled.
4319                  */
4320                 tp->snd_wnd    = ntohs(th->window);
4321                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4322                 tp->max_window = tp->snd_wnd;
4323
4324                 TCP_ECN_rcv_syn(tp, th);
4325                 if (tp->ecn_flags&TCP_ECN_OK)
4326                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4327
4328                 tcp_mtup_init(sk);
4329                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4330                 tcp_initialize_rcv_mss(sk);
4331
4332
4333                 tcp_send_synack(sk);
4334 #if 0
4335                 /* Note, we could accept data and URG from this segment.
4336                  * There are no obstacles to make this.
4337                  *
4338                  * However, if we ignore data in ACKless segments sometimes,
4339                  * we have no reasons to accept it sometimes.
4340                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4341                  * is not flawless. So, discard packet for sanity.
4342                  * Uncomment this return to process the data.
4343                  */
4344                 return -1;
4345 #else
4346                 goto discard;
4347 #endif
4348         }
4349         /* "fifth, if neither of the SYN or RST bits is set then
4350          * drop the segment and return."
4351          */
4352
4353 discard_and_undo:
4354         tcp_clear_options(&tp->rx_opt);
4355         tp->rx_opt.mss_clamp = saved_clamp;
4356         goto discard;
4357
4358 reset_and_undo:
4359         tcp_clear_options(&tp->rx_opt);
4360         tp->rx_opt.mss_clamp = saved_clamp;
4361         return 1;
4362 }
4363
4364
4365 /*
4366  *      This function implements the receiving procedure of RFC 793 for
4367  *      all states except ESTABLISHED and TIME_WAIT. 
4368  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4369  *      address independent.
4370  */
4371         
4372 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4373                           struct tcphdr *th, unsigned len)
4374 {
4375         struct tcp_sock *tp = tcp_sk(sk);
4376         struct inet_connection_sock *icsk = inet_csk(sk);
4377         int queued = 0;
4378
4379         tp->rx_opt.saw_tstamp = 0;
4380
4381         switch (sk->sk_state) {
4382         case TCP_CLOSE:
4383                 goto discard;
4384
4385         case TCP_LISTEN:
4386                 if(th->ack)
4387                         return 1;
4388
4389                 if(th->rst)
4390                         goto discard;
4391
4392                 if(th->syn) {
4393                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4394                                 return 1;
4395
4396                         /* Now we have several options: In theory there is 
4397                          * nothing else in the frame. KA9Q has an option to 
4398                          * send data with the syn, BSD accepts data with the
4399                          * syn up to the [to be] advertised window and 
4400                          * Solaris 2.1 gives you a protocol error. For now 
4401                          * we just ignore it, that fits the spec precisely 
4402                          * and avoids incompatibilities. It would be nice in
4403                          * future to drop through and process the data.
4404                          *
4405                          * Now that TTCP is starting to be used we ought to 
4406                          * queue this data.
4407                          * But, this leaves one open to an easy denial of
4408                          * service attack, and SYN cookies can't defend
4409                          * against this problem. So, we drop the data
4410                          * in the interest of security over speed.
4411                          */
4412                         goto discard;
4413                 }
4414                 goto discard;
4415
4416         case TCP_SYN_SENT:
4417                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4418                 if (queued >= 0)
4419                         return queued;
4420
4421                 /* Do step6 onward by hand. */
4422                 tcp_urg(sk, skb, th);
4423                 __kfree_skb(skb);
4424                 tcp_data_snd_check(sk, tp);
4425                 return 0;
4426         }
4427
4428         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4429             tcp_paws_discard(sk, skb)) {
4430                 if (!th->rst) {
4431                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4432                         tcp_send_dupack(sk, skb);
4433                         goto discard;
4434                 }
4435                 /* Reset is accepted even if it did not pass PAWS. */
4436         }
4437
4438         /* step 1: check sequence number */
4439         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4440                 if (!th->rst)
4441                         tcp_send_dupack(sk, skb);
4442                 goto discard;
4443         }
4444
4445         /* step 2: check RST bit */
4446         if(th->rst) {
4447                 tcp_reset(sk);
4448                 goto discard;
4449         }
4450
4451         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4452
4453         /* step 3: check security and precedence [ignored] */
4454
4455         /*      step 4:
4456          *
4457          *      Check for a SYN in window.
4458          */
4459         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4460                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4461                 tcp_reset(sk);
4462                 return 1;
4463         }
4464
4465         /* step 5: check the ACK field */
4466         if (th->ack) {
4467                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4468
4469                 switch(sk->sk_state) {
4470                 case TCP_SYN_RECV:
4471                         if (acceptable) {
4472                                 tp->copied_seq = tp->rcv_nxt;
4473                                 mb();
4474                                 tcp_set_state(sk, TCP_ESTABLISHED);
4475                                 sk->sk_state_change(sk);
4476
4477                                 /* Note, that this wakeup is only for marginal
4478                                  * crossed SYN case. Passively open sockets
4479                                  * are not waked up, because sk->sk_sleep ==
4480                                  * NULL and sk->sk_socket == NULL.
4481                                  */
4482                                 if (sk->sk_socket) {
4483                                         sk_wake_async(sk,0,POLL_OUT);
4484                                 }
4485
4486                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4487                                 tp->snd_wnd = ntohs(th->window) <<
4488                                               tp->rx_opt.snd_wscale;
4489                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4490                                             TCP_SKB_CB(skb)->seq);
4491
4492                                 /* tcp_ack considers this ACK as duplicate
4493                                  * and does not calculate rtt.
4494                                  * Fix it at least with timestamps.
4495                                  */
4496                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4497                                     !tp->srtt)
4498                                         tcp_ack_saw_tstamp(sk, 0);
4499
4500                                 if (tp->rx_opt.tstamp_ok)
4501                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4502
4503                                 /* Make sure socket is routed, for
4504                                  * correct metrics.
4505                                  */
4506                                 icsk->icsk_af_ops->rebuild_header(sk);
4507
4508                                 tcp_init_metrics(sk);
4509
4510                                 tcp_init_congestion_control(sk);
4511
4512                                 /* Prevent spurious tcp_cwnd_restart() on
4513                                  * first data packet.
4514                                  */
4515                                 tp->lsndtime = tcp_time_stamp;
4516
4517                                 tcp_mtup_init(sk);
4518                                 tcp_initialize_rcv_mss(sk);
4519                                 tcp_init_buffer_space(sk);
4520                                 tcp_fast_path_on(tp);
4521                         } else {
4522                                 return 1;
4523                         }
4524                         break;
4525
4526                 case TCP_FIN_WAIT1:
4527                         if (tp->snd_una == tp->write_seq) {
4528                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4529                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4530                                 dst_confirm(sk->sk_dst_cache);
4531
4532                                 if (!sock_flag(sk, SOCK_DEAD))
4533                                         /* Wake up lingering close() */
4534                                         sk->sk_state_change(sk);
4535                                 else {
4536                                         int tmo;
4537
4538                                         if (tp->linger2 < 0 ||
4539                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4540                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4541                                                 tcp_done(sk);
4542                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4543                                                 return 1;
4544                                         }
4545
4546                                         tmo = tcp_fin_time(sk);
4547                                         if (tmo > TCP_TIMEWAIT_LEN) {
4548                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4549                                         } else if (th->fin || sock_owned_by_user(sk)) {
4550                                                 /* Bad case. We could lose such FIN otherwise.
4551                                                  * It is not a big problem, but it looks confusing
4552                                                  * and not so rare event. We still can lose it now,
4553                                                  * if it spins in bh_lock_sock(), but it is really
4554                                                  * marginal case.
4555                                                  */
4556                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4557                                         } else {
4558                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4559                                                 goto discard;
4560                                         }
4561                                 }
4562                         }
4563                         break;
4564
4565                 case TCP_CLOSING:
4566                         if (tp->snd_una == tp->write_seq) {
4567                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4568                                 goto discard;
4569                         }
4570                         break;
4571
4572                 case TCP_LAST_ACK:
4573                         if (tp->snd_una == tp->write_seq) {
4574                                 tcp_update_metrics(sk);
4575                                 tcp_done(sk);
4576                                 goto discard;
4577                         }
4578                         break;
4579                 }
4580         } else
4581                 goto discard;
4582
4583         /* step 6: check the URG bit */
4584         tcp_urg(sk, skb, th);
4585
4586         /* step 7: process the segment text */
4587         switch (sk->sk_state) {
4588         case TCP_CLOSE_WAIT:
4589         case TCP_CLOSING:
4590         case TCP_LAST_ACK:
4591                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4592                         break;
4593         case TCP_FIN_WAIT1:
4594         case TCP_FIN_WAIT2:
4595                 /* RFC 793 says to queue data in these states,
4596                  * RFC 1122 says we MUST send a reset. 
4597                  * BSD 4.4 also does reset.
4598                  */
4599                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4600                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4601                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4602                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4603                                 tcp_reset(sk);
4604                                 return 1;
4605                         }
4606                 }
4607                 /* Fall through */
4608         case TCP_ESTABLISHED: 
4609                 tcp_data_queue(sk, skb);
4610                 queued = 1;
4611                 break;
4612         }
4613
4614         /* tcp_data could move socket to TIME-WAIT */
4615         if (sk->sk_state != TCP_CLOSE) {
4616                 tcp_data_snd_check(sk, tp);
4617                 tcp_ack_snd_check(sk);
4618         }
4619
4620         if (!queued) { 
4621 discard:
4622                 __kfree_skb(skb);
4623         }
4624         return 0;
4625 }
4626
4627 EXPORT_SYMBOL(sysctl_tcp_ecn);
4628 EXPORT_SYMBOL(sysctl_tcp_reordering);
4629 EXPORT_SYMBOL(tcp_parse_options);
4630 EXPORT_SYMBOL(tcp_rcv_established);
4631 EXPORT_SYMBOL(tcp_rcv_state_process);
4632 EXPORT_SYMBOL(tcp_initialize_rcv_mss);