Pull mca-cleanup into release branch
[linux-2.6] / 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
75 int sysctl_tcp_timestamps = 1;
76 int sysctl_tcp_window_scaling = 1;
77 int sysctl_tcp_sack = 1;
78 int sysctl_tcp_fack = 1;
79 int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn;
81 int sysctl_tcp_dsack = 1;
82 int sysctl_tcp_app_win = 31;
83 int sysctl_tcp_adv_win_scale = 2;
84
85 int sysctl_tcp_stdurg;
86 int sysctl_tcp_rfc1337;
87 int sysctl_tcp_max_orphans = NR_FILE;
88 int sysctl_tcp_frto;
89 int sysctl_tcp_nometrics_save;
90
91 int sysctl_tcp_moderate_rcvbuf = 1;
92 int sysctl_tcp_abc = 1;
93
94 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
95 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
96 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
98 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
99 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
100 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
101 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
102 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
103
104 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
107 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
108
109 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
110 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
111 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
112
113 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
114
115 /* Adapt the MSS value used to make delayed ack decision to the 
116  * real world.
117  */ 
118 static void tcp_measure_rcv_mss(struct sock *sk,
119                                 const struct sk_buff *skb)
120 {
121         struct inet_connection_sock *icsk = inet_csk(sk);
122         const unsigned int lss = icsk->icsk_ack.last_seg_size; 
123         unsigned int len;
124
125         icsk->icsk_ack.last_seg_size = 0; 
126
127         /* skb->len may jitter because of SACKs, even if peer
128          * sends good full-sized frames.
129          */
130         len = skb->len;
131         if (len >= icsk->icsk_ack.rcv_mss) {
132                 icsk->icsk_ack.rcv_mss = len;
133         } else {
134                 /* Otherwise, we make more careful check taking into account,
135                  * that SACKs block is variable.
136                  *
137                  * "len" is invariant segment length, including TCP header.
138                  */
139                 len += skb->data - skb->h.raw;
140                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
141                     /* If PSH is not set, packet should be
142                      * full sized, provided peer TCP is not badly broken.
143                      * This observation (if it is correct 8)) allows
144                      * to handle super-low mtu links fairly.
145                      */
146                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
147                      !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
148                         /* Subtract also invariant (if peer is RFC compliant),
149                          * tcp header plus fixed timestamp option length.
150                          * Resulting "len" is MSS free of SACK jitter.
151                          */
152                         len -= tcp_sk(sk)->tcp_header_len;
153                         icsk->icsk_ack.last_seg_size = len;
154                         if (len == lss) {
155                                 icsk->icsk_ack.rcv_mss = len;
156                                 return;
157                         }
158                 }
159                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
160         }
161 }
162
163 static void tcp_incr_quickack(struct sock *sk)
164 {
165         struct inet_connection_sock *icsk = inet_csk(sk);
166         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
167
168         if (quickacks==0)
169                 quickacks=2;
170         if (quickacks > icsk->icsk_ack.quick)
171                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
172 }
173
174 void tcp_enter_quickack_mode(struct sock *sk)
175 {
176         struct inet_connection_sock *icsk = inet_csk(sk);
177         tcp_incr_quickack(sk);
178         icsk->icsk_ack.pingpong = 0;
179         icsk->icsk_ack.ato = TCP_ATO_MIN;
180 }
181
182 /* Send ACKs quickly, if "quick" count is not exhausted
183  * and the session is not interactive.
184  */
185
186 static inline int tcp_in_quickack_mode(const struct sock *sk)
187 {
188         const struct inet_connection_sock *icsk = inet_csk(sk);
189         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
190 }
191
192 /* Buffer size and advertised window tuning.
193  *
194  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
195  */
196
197 static void tcp_fixup_sndbuf(struct sock *sk)
198 {
199         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
200                      sizeof(struct sk_buff);
201
202         if (sk->sk_sndbuf < 3 * sndmem)
203                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
204 }
205
206 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
207  *
208  * All tcp_full_space() is split to two parts: "network" buffer, allocated
209  * forward and advertised in receiver window (tp->rcv_wnd) and
210  * "application buffer", required to isolate scheduling/application
211  * latencies from network.
212  * window_clamp is maximal advertised window. It can be less than
213  * tcp_full_space(), in this case tcp_full_space() - window_clamp
214  * is reserved for "application" buffer. The less window_clamp is
215  * the smoother our behaviour from viewpoint of network, but the lower
216  * throughput and the higher sensitivity of the connection to losses. 8)
217  *
218  * rcv_ssthresh is more strict window_clamp used at "slow start"
219  * phase to predict further behaviour of this connection.
220  * It is used for two goals:
221  * - to enforce header prediction at sender, even when application
222  *   requires some significant "application buffer". It is check #1.
223  * - to prevent pruning of receive queue because of misprediction
224  *   of receiver window. Check #2.
225  *
226  * The scheme does not work when sender sends good segments opening
227  * window and then starts to feed us spaghetti. But it should work
228  * in common situations. Otherwise, we have to rely on queue collapsing.
229  */
230
231 /* Slow part of check#2. */
232 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
233                              const struct sk_buff *skb)
234 {
235         /* Optimize this! */
236         int truesize = tcp_win_from_space(skb->truesize)/2;
237         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
238
239         while (tp->rcv_ssthresh <= window) {
240                 if (truesize <= skb->len)
241                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
242
243                 truesize >>= 1;
244                 window >>= 1;
245         }
246         return 0;
247 }
248
249 static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
250                             struct sk_buff *skb)
251 {
252         /* Check #1 */
253         if (tp->rcv_ssthresh < tp->window_clamp &&
254             (int)tp->rcv_ssthresh < tcp_space(sk) &&
255             !tcp_memory_pressure) {
256                 int incr;
257
258                 /* Check #2. Increase window, if skb with such overhead
259                  * will fit to rcvbuf in future.
260                  */
261                 if (tcp_win_from_space(skb->truesize) <= skb->len)
262                         incr = 2*tp->advmss;
263                 else
264                         incr = __tcp_grow_window(sk, tp, skb);
265
266                 if (incr) {
267                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
268                         inet_csk(sk)->icsk_ack.quick |= 1;
269                 }
270         }
271 }
272
273 /* 3. Tuning rcvbuf, when connection enters established state. */
274
275 static void tcp_fixup_rcvbuf(struct sock *sk)
276 {
277         struct tcp_sock *tp = tcp_sk(sk);
278         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
279
280         /* Try to select rcvbuf so that 4 mss-sized segments
281          * will fit to window and corresponding skbs will fit to our rcvbuf.
282          * (was 3; 4 is minimum to allow fast retransmit to work.)
283          */
284         while (tcp_win_from_space(rcvmem) < tp->advmss)
285                 rcvmem += 128;
286         if (sk->sk_rcvbuf < 4 * rcvmem)
287                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
288 }
289
290 /* 4. Try to fixup all. It is made immediately after connection enters
291  *    established state.
292  */
293 static void tcp_init_buffer_space(struct sock *sk)
294 {
295         struct tcp_sock *tp = tcp_sk(sk);
296         int maxwin;
297
298         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
299                 tcp_fixup_rcvbuf(sk);
300         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
301                 tcp_fixup_sndbuf(sk);
302
303         tp->rcvq_space.space = tp->rcv_wnd;
304
305         maxwin = tcp_full_space(sk);
306
307         if (tp->window_clamp >= maxwin) {
308                 tp->window_clamp = maxwin;
309
310                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
311                         tp->window_clamp = max(maxwin -
312                                                (maxwin >> sysctl_tcp_app_win),
313                                                4 * tp->advmss);
314         }
315
316         /* Force reservation of one segment. */
317         if (sysctl_tcp_app_win &&
318             tp->window_clamp > 2 * tp->advmss &&
319             tp->window_clamp + tp->advmss > maxwin)
320                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
321
322         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
323         tp->snd_cwnd_stamp = tcp_time_stamp;
324 }
325
326 /* 5. Recalculate window clamp after socket hit its memory bounds. */
327 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
328 {
329         struct inet_connection_sock *icsk = inet_csk(sk);
330
331         icsk->icsk_ack.quick = 0;
332
333         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
334             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
335             !tcp_memory_pressure &&
336             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
337                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
338                                     sysctl_tcp_rmem[2]);
339         }
340         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
341                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
342 }
343
344
345 /* Initialize RCV_MSS value.
346  * RCV_MSS is an our guess about MSS used by the peer.
347  * We haven't any direct information about the MSS.
348  * It's better to underestimate the RCV_MSS rather than overestimate.
349  * Overestimations make us ACKing less frequently than needed.
350  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
351  */
352 void tcp_initialize_rcv_mss(struct sock *sk)
353 {
354         struct tcp_sock *tp = tcp_sk(sk);
355         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
356
357         hint = min(hint, tp->rcv_wnd/2);
358         hint = min(hint, TCP_MIN_RCVMSS);
359         hint = max(hint, TCP_MIN_MSS);
360
361         inet_csk(sk)->icsk_ack.rcv_mss = hint;
362 }
363
364 /* Receiver "autotuning" code.
365  *
366  * The algorithm for RTT estimation w/o timestamps is based on
367  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
368  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
369  *
370  * More detail on this code can be found at
371  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
372  * though this reference is out of date.  A new paper
373  * is pending.
374  */
375 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
376 {
377         u32 new_sample = tp->rcv_rtt_est.rtt;
378         long m = sample;
379
380         if (m == 0)
381                 m = 1;
382
383         if (new_sample != 0) {
384                 /* If we sample in larger samples in the non-timestamp
385                  * case, we could grossly overestimate the RTT especially
386                  * with chatty applications or bulk transfer apps which
387                  * are stalled on filesystem I/O.
388                  *
389                  * Also, since we are only going for a minimum in the
390                  * non-timestamp case, we do not smooth things out
391                  * else with timestamps disabled convergence takes too
392                  * long.
393                  */
394                 if (!win_dep) {
395                         m -= (new_sample >> 3);
396                         new_sample += m;
397                 } else if (m < new_sample)
398                         new_sample = m << 3;
399         } else {
400                 /* No previous measure. */
401                 new_sample = m << 3;
402         }
403
404         if (tp->rcv_rtt_est.rtt != new_sample)
405                 tp->rcv_rtt_est.rtt = new_sample;
406 }
407
408 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
409 {
410         if (tp->rcv_rtt_est.time == 0)
411                 goto new_measure;
412         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
413                 return;
414         tcp_rcv_rtt_update(tp,
415                            jiffies - tp->rcv_rtt_est.time,
416                            1);
417
418 new_measure:
419         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
420         tp->rcv_rtt_est.time = tcp_time_stamp;
421 }
422
423 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
424 {
425         struct tcp_sock *tp = tcp_sk(sk);
426         if (tp->rx_opt.rcv_tsecr &&
427             (TCP_SKB_CB(skb)->end_seq -
428              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
429                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
430 }
431
432 /*
433  * This function should be called every time data is copied to user space.
434  * It calculates the appropriate TCP receive buffer space.
435  */
436 void tcp_rcv_space_adjust(struct sock *sk)
437 {
438         struct tcp_sock *tp = tcp_sk(sk);
439         int time;
440         int space;
441         
442         if (tp->rcvq_space.time == 0)
443                 goto new_measure;
444         
445         time = tcp_time_stamp - tp->rcvq_space.time;
446         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
447             tp->rcv_rtt_est.rtt == 0)
448                 return;
449         
450         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
451
452         space = max(tp->rcvq_space.space, space);
453
454         if (tp->rcvq_space.space != space) {
455                 int rcvmem;
456
457                 tp->rcvq_space.space = space;
458
459                 if (sysctl_tcp_moderate_rcvbuf &&
460                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
461                         int new_clamp = space;
462
463                         /* Receive space grows, normalize in order to
464                          * take into account packet headers and sk_buff
465                          * structure overhead.
466                          */
467                         space /= tp->advmss;
468                         if (!space)
469                                 space = 1;
470                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
471                                   16 + sizeof(struct sk_buff));
472                         while (tcp_win_from_space(rcvmem) < tp->advmss)
473                                 rcvmem += 128;
474                         space *= rcvmem;
475                         space = min(space, sysctl_tcp_rmem[2]);
476                         if (space > sk->sk_rcvbuf) {
477                                 sk->sk_rcvbuf = space;
478
479                                 /* Make the window clamp follow along.  */
480                                 tp->window_clamp = new_clamp;
481                         }
482                 }
483         }
484         
485 new_measure:
486         tp->rcvq_space.seq = tp->copied_seq;
487         tp->rcvq_space.time = tcp_time_stamp;
488 }
489
490 /* There is something which you must keep in mind when you analyze the
491  * behavior of the tp->ato delayed ack timeout interval.  When a
492  * connection starts up, we want to ack as quickly as possible.  The
493  * problem is that "good" TCP's do slow start at the beginning of data
494  * transmission.  The means that until we send the first few ACK's the
495  * sender will sit on his end and only queue most of his data, because
496  * he can only send snd_cwnd unacked packets at any given time.  For
497  * each ACK we send, he increments snd_cwnd and transmits more of his
498  * queue.  -DaveM
499  */
500 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
501 {
502         struct inet_connection_sock *icsk = inet_csk(sk);
503         u32 now;
504
505         inet_csk_schedule_ack(sk);
506
507         tcp_measure_rcv_mss(sk, skb);
508
509         tcp_rcv_rtt_measure(tp);
510         
511         now = tcp_time_stamp;
512
513         if (!icsk->icsk_ack.ato) {
514                 /* The _first_ data packet received, initialize
515                  * delayed ACK engine.
516                  */
517                 tcp_incr_quickack(sk);
518                 icsk->icsk_ack.ato = TCP_ATO_MIN;
519         } else {
520                 int m = now - icsk->icsk_ack.lrcvtime;
521
522                 if (m <= TCP_ATO_MIN/2) {
523                         /* The fastest case is the first. */
524                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
525                 } else if (m < icsk->icsk_ack.ato) {
526                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
527                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
528                                 icsk->icsk_ack.ato = icsk->icsk_rto;
529                 } else if (m > icsk->icsk_rto) {
530                         /* Too long gap. Apparently sender failed to
531                          * restart window, so that we send ACKs quickly.
532                          */
533                         tcp_incr_quickack(sk);
534                         sk_stream_mem_reclaim(sk);
535                 }
536         }
537         icsk->icsk_ack.lrcvtime = now;
538
539         TCP_ECN_check_ce(tp, skb);
540
541         if (skb->len >= 128)
542                 tcp_grow_window(sk, tp, skb);
543 }
544
545 /* Called to compute a smoothed rtt estimate. The data fed to this
546  * routine either comes from timestamps, or from segments that were
547  * known _not_ to have been retransmitted [see Karn/Partridge
548  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
549  * piece by Van Jacobson.
550  * NOTE: the next three routines used to be one big routine.
551  * To save cycles in the RFC 1323 implementation it was better to break
552  * it up into three procedures. -- erics
553  */
554 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
555 {
556         struct tcp_sock *tp = tcp_sk(sk);
557         long m = mrtt; /* RTT */
558
559         /*      The following amusing code comes from Jacobson's
560          *      article in SIGCOMM '88.  Note that rtt and mdev
561          *      are scaled versions of rtt and mean deviation.
562          *      This is designed to be as fast as possible 
563          *      m stands for "measurement".
564          *
565          *      On a 1990 paper the rto value is changed to:
566          *      RTO = rtt + 4 * mdev
567          *
568          * Funny. This algorithm seems to be very broken.
569          * These formulae increase RTO, when it should be decreased, increase
570          * too slowly, when it should be increased quickly, decrease too quickly
571          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
572          * does not matter how to _calculate_ it. Seems, it was trap
573          * that VJ failed to avoid. 8)
574          */
575         if(m == 0)
576                 m = 1;
577         if (tp->srtt != 0) {
578                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
579                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
580                 if (m < 0) {
581                         m = -m;         /* m is now abs(error) */
582                         m -= (tp->mdev >> 2);   /* similar update on mdev */
583                         /* This is similar to one of Eifel findings.
584                          * Eifel blocks mdev updates when rtt decreases.
585                          * This solution is a bit different: we use finer gain
586                          * for mdev in this case (alpha*beta).
587                          * Like Eifel it also prevents growth of rto,
588                          * but also it limits too fast rto decreases,
589                          * happening in pure Eifel.
590                          */
591                         if (m > 0)
592                                 m >>= 3;
593                 } else {
594                         m -= (tp->mdev >> 2);   /* similar update on mdev */
595                 }
596                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
597                 if (tp->mdev > tp->mdev_max) {
598                         tp->mdev_max = tp->mdev;
599                         if (tp->mdev_max > tp->rttvar)
600                                 tp->rttvar = tp->mdev_max;
601                 }
602                 if (after(tp->snd_una, tp->rtt_seq)) {
603                         if (tp->mdev_max < tp->rttvar)
604                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
605                         tp->rtt_seq = tp->snd_nxt;
606                         tp->mdev_max = TCP_RTO_MIN;
607                 }
608         } else {
609                 /* no previous measure. */
610                 tp->srtt = m<<3;        /* take the measured time to be rtt */
611                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
612                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
613                 tp->rtt_seq = tp->snd_nxt;
614         }
615 }
616
617 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
618  * routine referred to above.
619  */
620 static inline void tcp_set_rto(struct sock *sk)
621 {
622         const struct tcp_sock *tp = tcp_sk(sk);
623         /* Old crap is replaced with new one. 8)
624          *
625          * More seriously:
626          * 1. If rtt variance happened to be less 50msec, it is hallucination.
627          *    It cannot be less due to utterly erratic ACK generation made
628          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
629          *    to do with delayed acks, because at cwnd>2 true delack timeout
630          *    is invisible. Actually, Linux-2.4 also generates erratic
631          *    ACKs in some circumstances.
632          */
633         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
634
635         /* 2. Fixups made earlier cannot be right.
636          *    If we do not estimate RTO correctly without them,
637          *    all the algo is pure shit and should be replaced
638          *    with correct one. It is exactly, which we pretend to do.
639          */
640 }
641
642 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
643  * guarantees that rto is higher.
644  */
645 static inline void tcp_bound_rto(struct sock *sk)
646 {
647         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
648                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
649 }
650
651 /* Save metrics learned by this TCP session.
652    This function is called only, when TCP finishes successfully
653    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
654  */
655 void tcp_update_metrics(struct sock *sk)
656 {
657         struct tcp_sock *tp = tcp_sk(sk);
658         struct dst_entry *dst = __sk_dst_get(sk);
659
660         if (sysctl_tcp_nometrics_save)
661                 return;
662
663         dst_confirm(dst);
664
665         if (dst && (dst->flags&DST_HOST)) {
666                 const struct inet_connection_sock *icsk = inet_csk(sk);
667                 int m;
668
669                 if (icsk->icsk_backoff || !tp->srtt) {
670                         /* This session failed to estimate rtt. Why?
671                          * Probably, no packets returned in time.
672                          * Reset our results.
673                          */
674                         if (!(dst_metric_locked(dst, RTAX_RTT)))
675                                 dst->metrics[RTAX_RTT-1] = 0;
676                         return;
677                 }
678
679                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
680
681                 /* If newly calculated rtt larger than stored one,
682                  * store new one. Otherwise, use EWMA. Remember,
683                  * rtt overestimation is always better than underestimation.
684                  */
685                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
686                         if (m <= 0)
687                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
688                         else
689                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
690                 }
691
692                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
693                         if (m < 0)
694                                 m = -m;
695
696                         /* Scale deviation to rttvar fixed point */
697                         m >>= 1;
698                         if (m < tp->mdev)
699                                 m = tp->mdev;
700
701                         if (m >= dst_metric(dst, RTAX_RTTVAR))
702                                 dst->metrics[RTAX_RTTVAR-1] = m;
703                         else
704                                 dst->metrics[RTAX_RTTVAR-1] -=
705                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
706                 }
707
708                 if (tp->snd_ssthresh >= 0xFFFF) {
709                         /* Slow start still did not finish. */
710                         if (dst_metric(dst, RTAX_SSTHRESH) &&
711                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
712                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
713                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
714                         if (!dst_metric_locked(dst, RTAX_CWND) &&
715                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
716                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
717                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
718                            icsk->icsk_ca_state == TCP_CA_Open) {
719                         /* Cong. avoidance phase, cwnd is reliable. */
720                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
721                                 dst->metrics[RTAX_SSTHRESH-1] =
722                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
723                         if (!dst_metric_locked(dst, RTAX_CWND))
724                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
725                 } else {
726                         /* Else slow start did not finish, cwnd is non-sense,
727                            ssthresh may be also invalid.
728                          */
729                         if (!dst_metric_locked(dst, RTAX_CWND))
730                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
731                         if (dst->metrics[RTAX_SSTHRESH-1] &&
732                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
733                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
734                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
735                 }
736
737                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
738                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
739                             tp->reordering != sysctl_tcp_reordering)
740                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
741                 }
742         }
743 }
744
745 /* Numbers are taken from RFC2414.  */
746 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
747 {
748         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
749
750         if (!cwnd) {
751                 if (tp->mss_cache > 1460)
752                         cwnd = 2;
753                 else
754                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
755         }
756         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
757 }
758
759 /* Set slow start threshold and cwnd not falling to slow start */
760 void tcp_enter_cwr(struct sock *sk)
761 {
762         struct tcp_sock *tp = tcp_sk(sk);
763
764         tp->prior_ssthresh = 0;
765         tp->bytes_acked = 0;
766         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
767                 tp->undo_marker = 0;
768                 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
769                 tp->snd_cwnd = min(tp->snd_cwnd,
770                                    tcp_packets_in_flight(tp) + 1U);
771                 tp->snd_cwnd_cnt = 0;
772                 tp->high_seq = tp->snd_nxt;
773                 tp->snd_cwnd_stamp = tcp_time_stamp;
774                 TCP_ECN_queue_cwr(tp);
775
776                 tcp_set_ca_state(sk, TCP_CA_CWR);
777         }
778 }
779
780 /* Initialize metrics on socket. */
781
782 static void tcp_init_metrics(struct sock *sk)
783 {
784         struct tcp_sock *tp = tcp_sk(sk);
785         struct dst_entry *dst = __sk_dst_get(sk);
786
787         if (dst == NULL)
788                 goto reset;
789
790         dst_confirm(dst);
791
792         if (dst_metric_locked(dst, RTAX_CWND))
793                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
794         if (dst_metric(dst, RTAX_SSTHRESH)) {
795                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
796                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
797                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
798         }
799         if (dst_metric(dst, RTAX_REORDERING) &&
800             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
801                 tp->rx_opt.sack_ok &= ~2;
802                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
803         }
804
805         if (dst_metric(dst, RTAX_RTT) == 0)
806                 goto reset;
807
808         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
809                 goto reset;
810
811         /* Initial rtt is determined from SYN,SYN-ACK.
812          * The segment is small and rtt may appear much
813          * less than real one. Use per-dst memory
814          * to make it more realistic.
815          *
816          * A bit of theory. RTT is time passed after "normal" sized packet
817          * is sent until it is ACKed. In normal circumstances sending small
818          * packets force peer to delay ACKs and calculation is correct too.
819          * The algorithm is adaptive and, provided we follow specs, it
820          * NEVER underestimate RTT. BUT! If peer tries to make some clever
821          * tricks sort of "quick acks" for time long enough to decrease RTT
822          * to low value, and then abruptly stops to do it and starts to delay
823          * ACKs, wait for troubles.
824          */
825         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
826                 tp->srtt = dst_metric(dst, RTAX_RTT);
827                 tp->rtt_seq = tp->snd_nxt;
828         }
829         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
830                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
831                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
832         }
833         tcp_set_rto(sk);
834         tcp_bound_rto(sk);
835         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
836                 goto reset;
837         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
838         tp->snd_cwnd_stamp = tcp_time_stamp;
839         return;
840
841 reset:
842         /* Play conservative. If timestamps are not
843          * supported, TCP will fail to recalculate correct
844          * rtt, if initial rto is too small. FORGET ALL AND RESET!
845          */
846         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
847                 tp->srtt = 0;
848                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
849                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
850         }
851 }
852
853 static void tcp_update_reordering(struct sock *sk, const int metric,
854                                   const int ts)
855 {
856         struct tcp_sock *tp = tcp_sk(sk);
857         if (metric > tp->reordering) {
858                 tp->reordering = min(TCP_MAX_REORDERING, metric);
859
860                 /* This exciting event is worth to be remembered. 8) */
861                 if (ts)
862                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
863                 else if (IsReno(tp))
864                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
865                 else if (IsFack(tp))
866                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
867                 else
868                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
869 #if FASTRETRANS_DEBUG > 1
870                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
871                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
872                        tp->reordering,
873                        tp->fackets_out,
874                        tp->sacked_out,
875                        tp->undo_marker ? tp->undo_retrans : 0);
876 #endif
877                 /* Disable FACK yet. */
878                 tp->rx_opt.sack_ok &= ~2;
879         }
880 }
881
882 /* This procedure tags the retransmission queue when SACKs arrive.
883  *
884  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
885  * Packets in queue with these bits set are counted in variables
886  * sacked_out, retrans_out and lost_out, correspondingly.
887  *
888  * Valid combinations are:
889  * Tag  InFlight        Description
890  * 0    1               - orig segment is in flight.
891  * S    0               - nothing flies, orig reached receiver.
892  * L    0               - nothing flies, orig lost by net.
893  * R    2               - both orig and retransmit are in flight.
894  * L|R  1               - orig is lost, retransmit is in flight.
895  * S|R  1               - orig reached receiver, retrans is still in flight.
896  * (L|S|R is logically valid, it could occur when L|R is sacked,
897  *  but it is equivalent to plain S and code short-curcuits it to S.
898  *  L|S is logically invalid, it would mean -1 packet in flight 8))
899  *
900  * These 6 states form finite state machine, controlled by the following events:
901  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
902  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
903  * 3. Loss detection event of one of three flavors:
904  *      A. Scoreboard estimator decided the packet is lost.
905  *         A'. Reno "three dupacks" marks head of queue lost.
906  *         A''. Its FACK modfication, head until snd.fack is lost.
907  *      B. SACK arrives sacking data transmitted after never retransmitted
908  *         hole was sent out.
909  *      C. SACK arrives sacking SND.NXT at the moment, when the
910  *         segment was retransmitted.
911  * 4. D-SACK added new rule: D-SACK changes any tag to S.
912  *
913  * It is pleasant to note, that state diagram turns out to be commutative,
914  * so that we are allowed not to be bothered by order of our actions,
915  * when multiple events arrive simultaneously. (see the function below).
916  *
917  * Reordering detection.
918  * --------------------
919  * Reordering metric is maximal distance, which a packet can be displaced
920  * in packet stream. With SACKs we can estimate it:
921  *
922  * 1. SACK fills old hole and the corresponding segment was not
923  *    ever retransmitted -> reordering. Alas, we cannot use it
924  *    when segment was retransmitted.
925  * 2. The last flaw is solved with D-SACK. D-SACK arrives
926  *    for retransmitted and already SACKed segment -> reordering..
927  * Both of these heuristics are not used in Loss state, when we cannot
928  * account for retransmits accurately.
929  */
930 static int
931 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
932 {
933         const struct inet_connection_sock *icsk = inet_csk(sk);
934         struct tcp_sock *tp = tcp_sk(sk);
935         unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
936         struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
937         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
938         int reord = tp->packets_out;
939         int prior_fackets;
940         u32 lost_retrans = 0;
941         int flag = 0;
942         int dup_sack = 0;
943         int i;
944
945         if (!tp->sacked_out)
946                 tp->fackets_out = 0;
947         prior_fackets = tp->fackets_out;
948
949         /* SACK fastpath:
950          * if the only SACK change is the increase of the end_seq of
951          * the first block then only apply that SACK block
952          * and use retrans queue hinting otherwise slowpath */
953         flag = 1;
954         for (i = 0; i< num_sacks; i++) {
955                 __u32 start_seq = ntohl(sp[i].start_seq);
956                 __u32 end_seq =  ntohl(sp[i].end_seq);
957
958                 if (i == 0){
959                         if (tp->recv_sack_cache[i].start_seq != start_seq)
960                                 flag = 0;
961                 } else {
962                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
963                             (tp->recv_sack_cache[i].end_seq != end_seq))
964                                 flag = 0;
965                 }
966                 tp->recv_sack_cache[i].start_seq = start_seq;
967                 tp->recv_sack_cache[i].end_seq = end_seq;
968
969                 /* Check for D-SACK. */
970                 if (i == 0) {
971                         u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
972
973                         if (before(start_seq, ack)) {
974                                 dup_sack = 1;
975                                 tp->rx_opt.sack_ok |= 4;
976                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
977                         } else if (num_sacks > 1 &&
978                                    !after(end_seq, ntohl(sp[1].end_seq)) &&
979                                    !before(start_seq, ntohl(sp[1].start_seq))) {
980                                 dup_sack = 1;
981                                 tp->rx_opt.sack_ok |= 4;
982                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
983                         }
984
985                         /* D-SACK for already forgotten data...
986                          * Do dumb counting. */
987                         if (dup_sack &&
988                             !after(end_seq, prior_snd_una) &&
989                             after(end_seq, tp->undo_marker))
990                                 tp->undo_retrans--;
991
992                         /* Eliminate too old ACKs, but take into
993                          * account more or less fresh ones, they can
994                          * contain valid SACK info.
995                          */
996                         if (before(ack, prior_snd_una - tp->max_window))
997                                 return 0;
998                 }
999         }
1000
1001         if (flag)
1002                 num_sacks = 1;
1003         else {
1004                 int j;
1005                 tp->fastpath_skb_hint = NULL;
1006
1007                 /* order SACK blocks to allow in order walk of the retrans queue */
1008                 for (i = num_sacks-1; i > 0; i--) {
1009                         for (j = 0; j < i; j++){
1010                                 if (after(ntohl(sp[j].start_seq),
1011                                           ntohl(sp[j+1].start_seq))){
1012                                         sp[j].start_seq = htonl(tp->recv_sack_cache[j+1].start_seq);
1013                                         sp[j].end_seq = htonl(tp->recv_sack_cache[j+1].end_seq);
1014                                         sp[j+1].start_seq = htonl(tp->recv_sack_cache[j].start_seq);
1015                                         sp[j+1].end_seq = htonl(tp->recv_sack_cache[j].end_seq);
1016                                 }
1017
1018                         }
1019                 }
1020         }
1021
1022         /* clear flag as used for different purpose in following code */
1023         flag = 0;
1024
1025         for (i=0; i<num_sacks; i++, sp++) {
1026                 struct sk_buff *skb;
1027                 __u32 start_seq = ntohl(sp->start_seq);
1028                 __u32 end_seq = ntohl(sp->end_seq);
1029                 int fack_count;
1030
1031                 /* Use SACK fastpath hint if valid */
1032                 if (tp->fastpath_skb_hint) {
1033                         skb = tp->fastpath_skb_hint;
1034                         fack_count = tp->fastpath_cnt_hint;
1035                 } else {
1036                         skb = sk->sk_write_queue.next;
1037                         fack_count = 0;
1038                 }
1039
1040                 /* Event "B" in the comment above. */
1041                 if (after(end_seq, tp->high_seq))
1042                         flag |= FLAG_DATA_LOST;
1043
1044                 sk_stream_for_retrans_queue_from(skb, sk) {
1045                         int in_sack, pcount;
1046                         u8 sacked;
1047
1048                         tp->fastpath_skb_hint = skb;
1049                         tp->fastpath_cnt_hint = fack_count;
1050
1051                         /* The retransmission queue is always in order, so
1052                          * we can short-circuit the walk early.
1053                          */
1054                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1055                                 break;
1056
1057                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1058                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1059
1060                         pcount = tcp_skb_pcount(skb);
1061
1062                         if (pcount > 1 && !in_sack &&
1063                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1064                                 unsigned int pkt_len;
1065
1066                                 in_sack = !after(start_seq,
1067                                                  TCP_SKB_CB(skb)->seq);
1068
1069                                 if (!in_sack)
1070                                         pkt_len = (start_seq -
1071                                                    TCP_SKB_CB(skb)->seq);
1072                                 else
1073                                         pkt_len = (end_seq -
1074                                                    TCP_SKB_CB(skb)->seq);
1075                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->tso_size))
1076                                         break;
1077                                 pcount = tcp_skb_pcount(skb);
1078                         }
1079
1080                         fack_count += pcount;
1081
1082                         sacked = TCP_SKB_CB(skb)->sacked;
1083
1084                         /* Account D-SACK for retransmitted packet. */
1085                         if ((dup_sack && in_sack) &&
1086                             (sacked & TCPCB_RETRANS) &&
1087                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1088                                 tp->undo_retrans--;
1089
1090                         /* The frame is ACKed. */
1091                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1092                                 if (sacked&TCPCB_RETRANS) {
1093                                         if ((dup_sack && in_sack) &&
1094                                             (sacked&TCPCB_SACKED_ACKED))
1095                                                 reord = min(fack_count, reord);
1096                                 } else {
1097                                         /* If it was in a hole, we detected reordering. */
1098                                         if (fack_count < prior_fackets &&
1099                                             !(sacked&TCPCB_SACKED_ACKED))
1100                                                 reord = min(fack_count, reord);
1101                                 }
1102
1103                                 /* Nothing to do; acked frame is about to be dropped. */
1104                                 continue;
1105                         }
1106
1107                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1108                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1109                             (!lost_retrans || after(end_seq, lost_retrans)))
1110                                 lost_retrans = end_seq;
1111
1112                         if (!in_sack)
1113                                 continue;
1114
1115                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1116                                 if (sacked & TCPCB_SACKED_RETRANS) {
1117                                         /* If the segment is not tagged as lost,
1118                                          * we do not clear RETRANS, believing
1119                                          * that retransmission is still in flight.
1120                                          */
1121                                         if (sacked & TCPCB_LOST) {
1122                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1123                                                 tp->lost_out -= tcp_skb_pcount(skb);
1124                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1125
1126                                                 /* clear lost hint */
1127                                                 tp->retransmit_skb_hint = NULL;
1128                                         }
1129                                 } else {
1130                                         /* New sack for not retransmitted frame,
1131                                          * which was in hole. It is reordering.
1132                                          */
1133                                         if (!(sacked & TCPCB_RETRANS) &&
1134                                             fack_count < prior_fackets)
1135                                                 reord = min(fack_count, reord);
1136
1137                                         if (sacked & TCPCB_LOST) {
1138                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1139                                                 tp->lost_out -= tcp_skb_pcount(skb);
1140
1141                                                 /* clear lost hint */
1142                                                 tp->retransmit_skb_hint = NULL;
1143                                         }
1144                                 }
1145
1146                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1147                                 flag |= FLAG_DATA_SACKED;
1148                                 tp->sacked_out += tcp_skb_pcount(skb);
1149
1150                                 if (fack_count > tp->fackets_out)
1151                                         tp->fackets_out = fack_count;
1152                         } else {
1153                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1154                                         reord = min(fack_count, reord);
1155                         }
1156
1157                         /* D-SACK. We can detect redundant retransmission
1158                          * in S|R and plain R frames and clear it.
1159                          * undo_retrans is decreased above, L|R frames
1160                          * are accounted above as well.
1161                          */
1162                         if (dup_sack &&
1163                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1164                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1165                                 tp->retrans_out -= tcp_skb_pcount(skb);
1166                                 tp->retransmit_skb_hint = NULL;
1167                         }
1168                 }
1169         }
1170
1171         /* Check for lost retransmit. This superb idea is
1172          * borrowed from "ratehalving". Event "C".
1173          * Later note: FACK people cheated me again 8),
1174          * we have to account for reordering! Ugly,
1175          * but should help.
1176          */
1177         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1178                 struct sk_buff *skb;
1179
1180                 sk_stream_for_retrans_queue(skb, sk) {
1181                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1182                                 break;
1183                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1184                                 continue;
1185                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1186                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1187                             (IsFack(tp) ||
1188                              !before(lost_retrans,
1189                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1190                                      tp->mss_cache))) {
1191                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1192                                 tp->retrans_out -= tcp_skb_pcount(skb);
1193
1194                                 /* clear lost hint */
1195                                 tp->retransmit_skb_hint = NULL;
1196
1197                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1198                                         tp->lost_out += tcp_skb_pcount(skb);
1199                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1200                                         flag |= FLAG_DATA_SACKED;
1201                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1202                                 }
1203                         }
1204                 }
1205         }
1206
1207         tp->left_out = tp->sacked_out + tp->lost_out;
1208
1209         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
1210                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1211
1212 #if FASTRETRANS_DEBUG > 0
1213         BUG_TRAP((int)tp->sacked_out >= 0);
1214         BUG_TRAP((int)tp->lost_out >= 0);
1215         BUG_TRAP((int)tp->retrans_out >= 0);
1216         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1217 #endif
1218         return flag;
1219 }
1220
1221 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1222  * segments to see from the next ACKs whether any data was really missing.
1223  * If the RTO was spurious, new ACKs should arrive.
1224  */
1225 void tcp_enter_frto(struct sock *sk)
1226 {
1227         const struct inet_connection_sock *icsk = inet_csk(sk);
1228         struct tcp_sock *tp = tcp_sk(sk);
1229         struct sk_buff *skb;
1230
1231         tp->frto_counter = 1;
1232
1233         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1234             tp->snd_una == tp->high_seq ||
1235             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1236                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1237                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1238                 tcp_ca_event(sk, CA_EVENT_FRTO);
1239         }
1240
1241         /* Have to clear retransmission markers here to keep the bookkeeping
1242          * in shape, even though we are not yet in Loss state.
1243          * If something was really lost, it is eventually caught up
1244          * in tcp_enter_frto_loss.
1245          */
1246         tp->retrans_out = 0;
1247         tp->undo_marker = tp->snd_una;
1248         tp->undo_retrans = 0;
1249
1250         sk_stream_for_retrans_queue(skb, sk) {
1251                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1252         }
1253         tcp_sync_left_out(tp);
1254
1255         tcp_set_ca_state(sk, TCP_CA_Open);
1256         tp->frto_highmark = tp->snd_nxt;
1257 }
1258
1259 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1260  * which indicates that we should follow the traditional RTO recovery,
1261  * i.e. mark everything lost and do go-back-N retransmission.
1262  */
1263 static void tcp_enter_frto_loss(struct sock *sk)
1264 {
1265         struct tcp_sock *tp = tcp_sk(sk);
1266         struct sk_buff *skb;
1267         int cnt = 0;
1268
1269         tp->sacked_out = 0;
1270         tp->lost_out = 0;
1271         tp->fackets_out = 0;
1272
1273         sk_stream_for_retrans_queue(skb, sk) {
1274                 cnt += tcp_skb_pcount(skb);
1275                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1276                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1277
1278                         /* Do not mark those segments lost that were
1279                          * forward transmitted after RTO
1280                          */
1281                         if (!after(TCP_SKB_CB(skb)->end_seq,
1282                                    tp->frto_highmark)) {
1283                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1284                                 tp->lost_out += tcp_skb_pcount(skb);
1285                         }
1286                 } else {
1287                         tp->sacked_out += tcp_skb_pcount(skb);
1288                         tp->fackets_out = cnt;
1289                 }
1290         }
1291         tcp_sync_left_out(tp);
1292
1293         tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1294         tp->snd_cwnd_cnt = 0;
1295         tp->snd_cwnd_stamp = tcp_time_stamp;
1296         tp->undo_marker = 0;
1297         tp->frto_counter = 0;
1298
1299         tp->reordering = min_t(unsigned int, tp->reordering,
1300                                              sysctl_tcp_reordering);
1301         tcp_set_ca_state(sk, TCP_CA_Loss);
1302         tp->high_seq = tp->frto_highmark;
1303         TCP_ECN_queue_cwr(tp);
1304
1305         clear_all_retrans_hints(tp);
1306 }
1307
1308 void tcp_clear_retrans(struct tcp_sock *tp)
1309 {
1310         tp->left_out = 0;
1311         tp->retrans_out = 0;
1312
1313         tp->fackets_out = 0;
1314         tp->sacked_out = 0;
1315         tp->lost_out = 0;
1316
1317         tp->undo_marker = 0;
1318         tp->undo_retrans = 0;
1319 }
1320
1321 /* Enter Loss state. If "how" is not zero, forget all SACK information
1322  * and reset tags completely, otherwise preserve SACKs. If receiver
1323  * dropped its ofo queue, we will know this due to reneging detection.
1324  */
1325 void tcp_enter_loss(struct sock *sk, int how)
1326 {
1327         const struct inet_connection_sock *icsk = inet_csk(sk);
1328         struct tcp_sock *tp = tcp_sk(sk);
1329         struct sk_buff *skb;
1330         int cnt = 0;
1331
1332         /* Reduce ssthresh if it has not yet been made inside this window. */
1333         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1334             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1335                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1336                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1337                 tcp_ca_event(sk, CA_EVENT_LOSS);
1338         }
1339         tp->snd_cwnd       = 1;
1340         tp->snd_cwnd_cnt   = 0;
1341         tp->snd_cwnd_stamp = tcp_time_stamp;
1342
1343         tp->bytes_acked = 0;
1344         tcp_clear_retrans(tp);
1345
1346         /* Push undo marker, if it was plain RTO and nothing
1347          * was retransmitted. */
1348         if (!how)
1349                 tp->undo_marker = tp->snd_una;
1350
1351         sk_stream_for_retrans_queue(skb, sk) {
1352                 cnt += tcp_skb_pcount(skb);
1353                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1354                         tp->undo_marker = 0;
1355                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1356                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1357                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1358                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1359                         tp->lost_out += tcp_skb_pcount(skb);
1360                 } else {
1361                         tp->sacked_out += tcp_skb_pcount(skb);
1362                         tp->fackets_out = cnt;
1363                 }
1364         }
1365         tcp_sync_left_out(tp);
1366
1367         tp->reordering = min_t(unsigned int, tp->reordering,
1368                                              sysctl_tcp_reordering);
1369         tcp_set_ca_state(sk, TCP_CA_Loss);
1370         tp->high_seq = tp->snd_nxt;
1371         TCP_ECN_queue_cwr(tp);
1372
1373         clear_all_retrans_hints(tp);
1374 }
1375
1376 static int tcp_check_sack_reneging(struct sock *sk)
1377 {
1378         struct sk_buff *skb;
1379
1380         /* If ACK arrived pointing to a remembered SACK,
1381          * it means that our remembered SACKs do not reflect
1382          * real state of receiver i.e.
1383          * receiver _host_ is heavily congested (or buggy).
1384          * Do processing similar to RTO timeout.
1385          */
1386         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1387             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1388                 struct inet_connection_sock *icsk = inet_csk(sk);
1389                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1390
1391                 tcp_enter_loss(sk, 1);
1392                 icsk->icsk_retransmits++;
1393                 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1394                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1395                                           icsk->icsk_rto, TCP_RTO_MAX);
1396                 return 1;
1397         }
1398         return 0;
1399 }
1400
1401 static inline int tcp_fackets_out(struct tcp_sock *tp)
1402 {
1403         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1404 }
1405
1406 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1407 {
1408         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1409 }
1410
1411 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1412 {
1413         return tp->packets_out &&
1414                tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
1415 }
1416
1417 /* Linux NewReno/SACK/FACK/ECN state machine.
1418  * --------------------------------------
1419  *
1420  * "Open"       Normal state, no dubious events, fast path.
1421  * "Disorder"   In all the respects it is "Open",
1422  *              but requires a bit more attention. It is entered when
1423  *              we see some SACKs or dupacks. It is split of "Open"
1424  *              mainly to move some processing from fast path to slow one.
1425  * "CWR"        CWND was reduced due to some Congestion Notification event.
1426  *              It can be ECN, ICMP source quench, local device congestion.
1427  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1428  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1429  *
1430  * tcp_fastretrans_alert() is entered:
1431  * - each incoming ACK, if state is not "Open"
1432  * - when arrived ACK is unusual, namely:
1433  *      * SACK
1434  *      * Duplicate ACK.
1435  *      * ECN ECE.
1436  *
1437  * Counting packets in flight is pretty simple.
1438  *
1439  *      in_flight = packets_out - left_out + retrans_out
1440  *
1441  *      packets_out is SND.NXT-SND.UNA counted in packets.
1442  *
1443  *      retrans_out is number of retransmitted segments.
1444  *
1445  *      left_out is number of segments left network, but not ACKed yet.
1446  *
1447  *              left_out = sacked_out + lost_out
1448  *
1449  *     sacked_out: Packets, which arrived to receiver out of order
1450  *                 and hence not ACKed. With SACKs this number is simply
1451  *                 amount of SACKed data. Even without SACKs
1452  *                 it is easy to give pretty reliable estimate of this number,
1453  *                 counting duplicate ACKs.
1454  *
1455  *       lost_out: Packets lost by network. TCP has no explicit
1456  *                 "loss notification" feedback from network (for now).
1457  *                 It means that this number can be only _guessed_.
1458  *                 Actually, it is the heuristics to predict lossage that
1459  *                 distinguishes different algorithms.
1460  *
1461  *      F.e. after RTO, when all the queue is considered as lost,
1462  *      lost_out = packets_out and in_flight = retrans_out.
1463  *
1464  *              Essentially, we have now two algorithms counting
1465  *              lost packets.
1466  *
1467  *              FACK: It is the simplest heuristics. As soon as we decided
1468  *              that something is lost, we decide that _all_ not SACKed
1469  *              packets until the most forward SACK are lost. I.e.
1470  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1471  *              It is absolutely correct estimate, if network does not reorder
1472  *              packets. And it loses any connection to reality when reordering
1473  *              takes place. We use FACK by default until reordering
1474  *              is suspected on the path to this destination.
1475  *
1476  *              NewReno: when Recovery is entered, we assume that one segment
1477  *              is lost (classic Reno). While we are in Recovery and
1478  *              a partial ACK arrives, we assume that one more packet
1479  *              is lost (NewReno). This heuristics are the same in NewReno
1480  *              and SACK.
1481  *
1482  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1483  *  deflation etc. CWND is real congestion window, never inflated, changes
1484  *  only according to classic VJ rules.
1485  *
1486  * Really tricky (and requiring careful tuning) part of algorithm
1487  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1488  * The first determines the moment _when_ we should reduce CWND and,
1489  * hence, slow down forward transmission. In fact, it determines the moment
1490  * when we decide that hole is caused by loss, rather than by a reorder.
1491  *
1492  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1493  * holes, caused by lost packets.
1494  *
1495  * And the most logically complicated part of algorithm is undo
1496  * heuristics. We detect false retransmits due to both too early
1497  * fast retransmit (reordering) and underestimated RTO, analyzing
1498  * timestamps and D-SACKs. When we detect that some segments were
1499  * retransmitted by mistake and CWND reduction was wrong, we undo
1500  * window reduction and abort recovery phase. This logic is hidden
1501  * inside several functions named tcp_try_undo_<something>.
1502  */
1503
1504 /* This function decides, when we should leave Disordered state
1505  * and enter Recovery phase, reducing congestion window.
1506  *
1507  * Main question: may we further continue forward transmission
1508  * with the same cwnd?
1509  */
1510 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1511 {
1512         __u32 packets_out;
1513
1514         /* Trick#1: The loss is proven. */
1515         if (tp->lost_out)
1516                 return 1;
1517
1518         /* Not-A-Trick#2 : Classic rule... */
1519         if (tcp_fackets_out(tp) > tp->reordering)
1520                 return 1;
1521
1522         /* Trick#3 : when we use RFC2988 timer restart, fast
1523          * retransmit can be triggered by timeout of queue head.
1524          */
1525         if (tcp_head_timedout(sk, tp))
1526                 return 1;
1527
1528         /* Trick#4: It is still not OK... But will it be useful to delay
1529          * recovery more?
1530          */
1531         packets_out = tp->packets_out;
1532         if (packets_out <= tp->reordering &&
1533             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1534             !tcp_may_send_now(sk, tp)) {
1535                 /* We have nothing to send. This connection is limited
1536                  * either by receiver window or by application.
1537                  */
1538                 return 1;
1539         }
1540
1541         return 0;
1542 }
1543
1544 /* If we receive more dupacks than we expected counting segments
1545  * in assumption of absent reordering, interpret this as reordering.
1546  * The only another reason could be bug in receiver TCP.
1547  */
1548 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1549 {
1550         struct tcp_sock *tp = tcp_sk(sk);
1551         u32 holes;
1552
1553         holes = max(tp->lost_out, 1U);
1554         holes = min(holes, tp->packets_out);
1555
1556         if ((tp->sacked_out + holes) > tp->packets_out) {
1557                 tp->sacked_out = tp->packets_out - holes;
1558                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1559         }
1560 }
1561
1562 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1563
1564 static void tcp_add_reno_sack(struct sock *sk)
1565 {
1566         struct tcp_sock *tp = tcp_sk(sk);
1567         tp->sacked_out++;
1568         tcp_check_reno_reordering(sk, 0);
1569         tcp_sync_left_out(tp);
1570 }
1571
1572 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1573
1574 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1575 {
1576         if (acked > 0) {
1577                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1578                 if (acked-1 >= tp->sacked_out)
1579                         tp->sacked_out = 0;
1580                 else
1581                         tp->sacked_out -= acked-1;
1582         }
1583         tcp_check_reno_reordering(sk, acked);
1584         tcp_sync_left_out(tp);
1585 }
1586
1587 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1588 {
1589         tp->sacked_out = 0;
1590         tp->left_out = tp->lost_out;
1591 }
1592
1593 /* Mark head of queue up as lost. */
1594 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1595                                int packets, u32 high_seq)
1596 {
1597         struct sk_buff *skb;
1598         int cnt;
1599
1600         BUG_TRAP(packets <= tp->packets_out);
1601         if (tp->lost_skb_hint) {
1602                 skb = tp->lost_skb_hint;
1603                 cnt = tp->lost_cnt_hint;
1604         } else {
1605                 skb = sk->sk_write_queue.next;
1606                 cnt = 0;
1607         }
1608
1609         sk_stream_for_retrans_queue_from(skb, sk) {
1610                 /* TODO: do this better */
1611                 /* this is not the most efficient way to do this... */
1612                 tp->lost_skb_hint = skb;
1613                 tp->lost_cnt_hint = cnt;
1614                 cnt += tcp_skb_pcount(skb);
1615                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1616                         break;
1617                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1618                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1619                         tp->lost_out += tcp_skb_pcount(skb);
1620
1621                         /* clear xmit_retransmit_queue hints
1622                          *  if this is beyond hint */
1623                         if(tp->retransmit_skb_hint != NULL &&
1624                            before(TCP_SKB_CB(skb)->seq,
1625                                   TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) {
1626
1627                                 tp->retransmit_skb_hint = NULL;
1628                         }
1629                 }
1630         }
1631         tcp_sync_left_out(tp);
1632 }
1633
1634 /* Account newly detected lost packet(s) */
1635
1636 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1637 {
1638         if (IsFack(tp)) {
1639                 int lost = tp->fackets_out - tp->reordering;
1640                 if (lost <= 0)
1641                         lost = 1;
1642                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1643         } else {
1644                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1645         }
1646
1647         /* New heuristics: it is possible only after we switched
1648          * to restart timer each time when something is ACKed.
1649          * Hence, we can detect timed out packets during fast
1650          * retransmit without falling to slow start.
1651          */
1652         if (tcp_head_timedout(sk, tp)) {
1653                 struct sk_buff *skb;
1654
1655                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1656                         : sk->sk_write_queue.next;
1657
1658                 sk_stream_for_retrans_queue_from(skb, sk) {
1659                         if (!tcp_skb_timedout(sk, skb))
1660                                 break;
1661
1662                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1663                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1664                                 tp->lost_out += tcp_skb_pcount(skb);
1665
1666                                 /* clear xmit_retrans hint */
1667                                 if (tp->retransmit_skb_hint &&
1668                                     before(TCP_SKB_CB(skb)->seq,
1669                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1670
1671                                         tp->retransmit_skb_hint = NULL;
1672                         }
1673                 }
1674
1675                 tp->scoreboard_skb_hint = skb;
1676
1677                 tcp_sync_left_out(tp);
1678         }
1679 }
1680
1681 /* CWND moderation, preventing bursts due to too big ACKs
1682  * in dubious situations.
1683  */
1684 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1685 {
1686         tp->snd_cwnd = min(tp->snd_cwnd,
1687                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1688         tp->snd_cwnd_stamp = tcp_time_stamp;
1689 }
1690
1691 /* Decrease cwnd each second ack. */
1692 static void tcp_cwnd_down(struct sock *sk)
1693 {
1694         const struct inet_connection_sock *icsk = inet_csk(sk);
1695         struct tcp_sock *tp = tcp_sk(sk);
1696         int decr = tp->snd_cwnd_cnt + 1;
1697
1698         tp->snd_cwnd_cnt = decr&1;
1699         decr >>= 1;
1700
1701         if (decr && tp->snd_cwnd > icsk->icsk_ca_ops->min_cwnd(sk))
1702                 tp->snd_cwnd -= decr;
1703
1704         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1705         tp->snd_cwnd_stamp = tcp_time_stamp;
1706 }
1707
1708 /* Nothing was retransmitted or returned timestamp is less
1709  * than timestamp of the first retransmission.
1710  */
1711 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1712 {
1713         return !tp->retrans_stamp ||
1714                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1715                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1716 }
1717
1718 /* Undo procedures. */
1719
1720 #if FASTRETRANS_DEBUG > 1
1721 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1722 {
1723         struct inet_sock *inet = inet_sk(sk);
1724         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1725                msg,
1726                NIPQUAD(inet->daddr), ntohs(inet->dport),
1727                tp->snd_cwnd, tp->left_out,
1728                tp->snd_ssthresh, tp->prior_ssthresh,
1729                tp->packets_out);
1730 }
1731 #else
1732 #define DBGUNDO(x...) do { } while (0)
1733 #endif
1734
1735 static void tcp_undo_cwr(struct sock *sk, const int undo)
1736 {
1737         struct tcp_sock *tp = tcp_sk(sk);
1738
1739         if (tp->prior_ssthresh) {
1740                 const struct inet_connection_sock *icsk = inet_csk(sk);
1741
1742                 if (icsk->icsk_ca_ops->undo_cwnd)
1743                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1744                 else
1745                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1746
1747                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1748                         tp->snd_ssthresh = tp->prior_ssthresh;
1749                         TCP_ECN_withdraw_cwr(tp);
1750                 }
1751         } else {
1752                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1753         }
1754         tcp_moderate_cwnd(tp);
1755         tp->snd_cwnd_stamp = tcp_time_stamp;
1756
1757         /* There is something screwy going on with the retrans hints after
1758            an undo */
1759         clear_all_retrans_hints(tp);
1760 }
1761
1762 static inline int tcp_may_undo(struct tcp_sock *tp)
1763 {
1764         return tp->undo_marker &&
1765                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1766 }
1767
1768 /* People celebrate: "We love our President!" */
1769 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1770 {
1771         if (tcp_may_undo(tp)) {
1772                 /* Happy end! We did not retransmit anything
1773                  * or our original transmission succeeded.
1774                  */
1775                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1776                 tcp_undo_cwr(sk, 1);
1777                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1778                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1779                 else
1780                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1781                 tp->undo_marker = 0;
1782         }
1783         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1784                 /* Hold old state until something *above* high_seq
1785                  * is ACKed. For Reno it is MUST to prevent false
1786                  * fast retransmits (RFC2582). SACK TCP is safe. */
1787                 tcp_moderate_cwnd(tp);
1788                 return 1;
1789         }
1790         tcp_set_ca_state(sk, TCP_CA_Open);
1791         return 0;
1792 }
1793
1794 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1795 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1796 {
1797         if (tp->undo_marker && !tp->undo_retrans) {
1798                 DBGUNDO(sk, tp, "D-SACK");
1799                 tcp_undo_cwr(sk, 1);
1800                 tp->undo_marker = 0;
1801                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1802         }
1803 }
1804
1805 /* Undo during fast recovery after partial ACK. */
1806
1807 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1808                                 int acked)
1809 {
1810         /* Partial ACK arrived. Force Hoe's retransmit. */
1811         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1812
1813         if (tcp_may_undo(tp)) {
1814                 /* Plain luck! Hole if filled with delayed
1815                  * packet, rather than with a retransmit.
1816                  */
1817                 if (tp->retrans_out == 0)
1818                         tp->retrans_stamp = 0;
1819
1820                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1821
1822                 DBGUNDO(sk, tp, "Hoe");
1823                 tcp_undo_cwr(sk, 0);
1824                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1825
1826                 /* So... Do not make Hoe's retransmit yet.
1827                  * If the first packet was delayed, the rest
1828                  * ones are most probably delayed as well.
1829                  */
1830                 failed = 0;
1831         }
1832         return failed;
1833 }
1834
1835 /* Undo during loss recovery after partial ACK. */
1836 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1837 {
1838         if (tcp_may_undo(tp)) {
1839                 struct sk_buff *skb;
1840                 sk_stream_for_retrans_queue(skb, sk) {
1841                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1842                 }
1843
1844                 clear_all_retrans_hints(tp);
1845
1846                 DBGUNDO(sk, tp, "partial loss");
1847                 tp->lost_out = 0;
1848                 tp->left_out = tp->sacked_out;
1849                 tcp_undo_cwr(sk, 1);
1850                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1851                 inet_csk(sk)->icsk_retransmits = 0;
1852                 tp->undo_marker = 0;
1853                 if (!IsReno(tp))
1854                         tcp_set_ca_state(sk, TCP_CA_Open);
1855                 return 1;
1856         }
1857         return 0;
1858 }
1859
1860 static inline void tcp_complete_cwr(struct sock *sk)
1861 {
1862         struct tcp_sock *tp = tcp_sk(sk);
1863         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1864         tp->snd_cwnd_stamp = tcp_time_stamp;
1865         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
1866 }
1867
1868 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1869 {
1870         tp->left_out = tp->sacked_out;
1871
1872         if (tp->retrans_out == 0)
1873                 tp->retrans_stamp = 0;
1874
1875         if (flag&FLAG_ECE)
1876                 tcp_enter_cwr(sk);
1877
1878         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
1879                 int state = TCP_CA_Open;
1880
1881                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1882                         state = TCP_CA_Disorder;
1883
1884                 if (inet_csk(sk)->icsk_ca_state != state) {
1885                         tcp_set_ca_state(sk, state);
1886                         tp->high_seq = tp->snd_nxt;
1887                 }
1888                 tcp_moderate_cwnd(tp);
1889         } else {
1890                 tcp_cwnd_down(sk);
1891         }
1892 }
1893
1894 /* Process an event, which can update packets-in-flight not trivially.
1895  * Main goal of this function is to calculate new estimate for left_out,
1896  * taking into account both packets sitting in receiver's buffer and
1897  * packets lost by network.
1898  *
1899  * Besides that it does CWND reduction, when packet loss is detected
1900  * and changes state of machine.
1901  *
1902  * It does _not_ decide what to send, it is made in function
1903  * tcp_xmit_retransmit_queue().
1904  */
1905 static void
1906 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1907                       int prior_packets, int flag)
1908 {
1909         struct inet_connection_sock *icsk = inet_csk(sk);
1910         struct tcp_sock *tp = tcp_sk(sk);
1911         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1912
1913         /* Some technical things:
1914          * 1. Reno does not count dupacks (sacked_out) automatically. */
1915         if (!tp->packets_out)
1916                 tp->sacked_out = 0;
1917         /* 2. SACK counts snd_fack in packets inaccurately. */
1918         if (tp->sacked_out == 0)
1919                 tp->fackets_out = 0;
1920
1921         /* Now state machine starts.
1922          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1923         if (flag&FLAG_ECE)
1924                 tp->prior_ssthresh = 0;
1925
1926         /* B. In all the states check for reneging SACKs. */
1927         if (tp->sacked_out && tcp_check_sack_reneging(sk))
1928                 return;
1929
1930         /* C. Process data loss notification, provided it is valid. */
1931         if ((flag&FLAG_DATA_LOST) &&
1932             before(tp->snd_una, tp->high_seq) &&
1933             icsk->icsk_ca_state != TCP_CA_Open &&
1934             tp->fackets_out > tp->reordering) {
1935                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1936                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1937         }
1938
1939         /* D. Synchronize left_out to current state. */
1940         tcp_sync_left_out(tp);
1941
1942         /* E. Check state exit conditions. State can be terminated
1943          *    when high_seq is ACKed. */
1944         if (icsk->icsk_ca_state == TCP_CA_Open) {
1945                 if (!sysctl_tcp_frto)
1946                         BUG_TRAP(tp->retrans_out == 0);
1947                 tp->retrans_stamp = 0;
1948         } else if (!before(tp->snd_una, tp->high_seq)) {
1949                 switch (icsk->icsk_ca_state) {
1950                 case TCP_CA_Loss:
1951                         icsk->icsk_retransmits = 0;
1952                         if (tcp_try_undo_recovery(sk, tp))
1953                                 return;
1954                         break;
1955
1956                 case TCP_CA_CWR:
1957                         /* CWR is to be held something *above* high_seq
1958                          * is ACKed for CWR bit to reach receiver. */
1959                         if (tp->snd_una != tp->high_seq) {
1960                                 tcp_complete_cwr(sk);
1961                                 tcp_set_ca_state(sk, TCP_CA_Open);
1962                         }
1963                         break;
1964
1965                 case TCP_CA_Disorder:
1966                         tcp_try_undo_dsack(sk, tp);
1967                         if (!tp->undo_marker ||
1968                             /* For SACK case do not Open to allow to undo
1969                              * catching for all duplicate ACKs. */
1970                             IsReno(tp) || tp->snd_una != tp->high_seq) {
1971                                 tp->undo_marker = 0;
1972                                 tcp_set_ca_state(sk, TCP_CA_Open);
1973                         }
1974                         break;
1975
1976                 case TCP_CA_Recovery:
1977                         if (IsReno(tp))
1978                                 tcp_reset_reno_sack(tp);
1979                         if (tcp_try_undo_recovery(sk, tp))
1980                                 return;
1981                         tcp_complete_cwr(sk);
1982                         break;
1983                 }
1984         }
1985
1986         /* F. Process state. */
1987         switch (icsk->icsk_ca_state) {
1988         case TCP_CA_Recovery:
1989                 if (prior_snd_una == tp->snd_una) {
1990                         if (IsReno(tp) && is_dupack)
1991                                 tcp_add_reno_sack(sk);
1992                 } else {
1993                         int acked = prior_packets - tp->packets_out;
1994                         if (IsReno(tp))
1995                                 tcp_remove_reno_sacks(sk, tp, acked);
1996                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
1997                 }
1998                 break;
1999         case TCP_CA_Loss:
2000                 if (flag&FLAG_DATA_ACKED)
2001                         icsk->icsk_retransmits = 0;
2002                 if (!tcp_try_undo_loss(sk, tp)) {
2003                         tcp_moderate_cwnd(tp);
2004                         tcp_xmit_retransmit_queue(sk);
2005                         return;
2006                 }
2007                 if (icsk->icsk_ca_state != TCP_CA_Open)
2008                         return;
2009                 /* Loss is undone; fall through to processing in Open state. */
2010         default:
2011                 if (IsReno(tp)) {
2012                         if (tp->snd_una != prior_snd_una)
2013                                 tcp_reset_reno_sack(tp);
2014                         if (is_dupack)
2015                                 tcp_add_reno_sack(sk);
2016                 }
2017
2018                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2019                         tcp_try_undo_dsack(sk, tp);
2020
2021                 if (!tcp_time_to_recover(sk, tp)) {
2022                         tcp_try_to_open(sk, tp, flag);
2023                         return;
2024                 }
2025
2026                 /* Otherwise enter Recovery state */
2027
2028                 if (IsReno(tp))
2029                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2030                 else
2031                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2032
2033                 tp->high_seq = tp->snd_nxt;
2034                 tp->prior_ssthresh = 0;
2035                 tp->undo_marker = tp->snd_una;
2036                 tp->undo_retrans = tp->retrans_out;
2037
2038                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2039                         if (!(flag&FLAG_ECE))
2040                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2041                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2042                         TCP_ECN_queue_cwr(tp);
2043                 }
2044
2045                 tp->bytes_acked = 0;
2046                 tp->snd_cwnd_cnt = 0;
2047                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2048         }
2049
2050         if (is_dupack || tcp_head_timedout(sk, tp))
2051                 tcp_update_scoreboard(sk, tp);
2052         tcp_cwnd_down(sk);
2053         tcp_xmit_retransmit_queue(sk);
2054 }
2055
2056 /* Read draft-ietf-tcplw-high-performance before mucking
2057  * with this code. (Supersedes RFC1323)
2058  */
2059 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2060 {
2061         /* RTTM Rule: A TSecr value received in a segment is used to
2062          * update the averaged RTT measurement only if the segment
2063          * acknowledges some new data, i.e., only if it advances the
2064          * left edge of the send window.
2065          *
2066          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2067          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2068          *
2069          * Changed: reset backoff as soon as we see the first valid sample.
2070          * If we do not, we get strongly overestimated rto. With timestamps
2071          * samples are accepted even from very old segments: f.e., when rtt=1
2072          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2073          * answer arrives rto becomes 120 seconds! If at least one of segments
2074          * in window is lost... Voila.                          --ANK (010210)
2075          */
2076         struct tcp_sock *tp = tcp_sk(sk);
2077         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2078         tcp_rtt_estimator(sk, seq_rtt);
2079         tcp_set_rto(sk);
2080         inet_csk(sk)->icsk_backoff = 0;
2081         tcp_bound_rto(sk);
2082 }
2083
2084 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2085 {
2086         /* We don't have a timestamp. Can only use
2087          * packets that are not retransmitted to determine
2088          * rtt estimates. Also, we must not reset the
2089          * backoff for rto until we get a non-retransmitted
2090          * packet. This allows us to deal with a situation
2091          * where the network delay has increased suddenly.
2092          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2093          */
2094
2095         if (flag & FLAG_RETRANS_DATA_ACKED)
2096                 return;
2097
2098         tcp_rtt_estimator(sk, seq_rtt);
2099         tcp_set_rto(sk);
2100         inet_csk(sk)->icsk_backoff = 0;
2101         tcp_bound_rto(sk);
2102 }
2103
2104 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2105                                       const s32 seq_rtt)
2106 {
2107         const struct tcp_sock *tp = tcp_sk(sk);
2108         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2109         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2110                 tcp_ack_saw_tstamp(sk, flag);
2111         else if (seq_rtt >= 0)
2112                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2113 }
2114
2115 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2116                            u32 in_flight, int good)
2117 {
2118         const struct inet_connection_sock *icsk = inet_csk(sk);
2119         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2120         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2121 }
2122
2123 /* Restart timer after forward progress on connection.
2124  * RFC2988 recommends to restart timer to now+rto.
2125  */
2126
2127 static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2128 {
2129         if (!tp->packets_out) {
2130                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2131         } else {
2132                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2133         }
2134 }
2135
2136 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2137                          __u32 now, __s32 *seq_rtt)
2138 {
2139         struct tcp_sock *tp = tcp_sk(sk);
2140         struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2141         __u32 seq = tp->snd_una;
2142         __u32 packets_acked;
2143         int acked = 0;
2144
2145         /* If we get here, the whole TSO packet has not been
2146          * acked.
2147          */
2148         BUG_ON(!after(scb->end_seq, seq));
2149
2150         packets_acked = tcp_skb_pcount(skb);
2151         if (tcp_trim_head(sk, skb, seq - scb->seq))
2152                 return 0;
2153         packets_acked -= tcp_skb_pcount(skb);
2154
2155         if (packets_acked) {
2156                 __u8 sacked = scb->sacked;
2157
2158                 acked |= FLAG_DATA_ACKED;
2159                 if (sacked) {
2160                         if (sacked & TCPCB_RETRANS) {
2161                                 if (sacked & TCPCB_SACKED_RETRANS)
2162                                         tp->retrans_out -= packets_acked;
2163                                 acked |= FLAG_RETRANS_DATA_ACKED;
2164                                 *seq_rtt = -1;
2165                         } else if (*seq_rtt < 0)
2166                                 *seq_rtt = now - scb->when;
2167                         if (sacked & TCPCB_SACKED_ACKED)
2168                                 tp->sacked_out -= packets_acked;
2169                         if (sacked & TCPCB_LOST)
2170                                 tp->lost_out -= packets_acked;
2171                         if (sacked & TCPCB_URG) {
2172                                 if (tp->urg_mode &&
2173                                     !before(seq, tp->snd_up))
2174                                         tp->urg_mode = 0;
2175                         }
2176                 } else if (*seq_rtt < 0)
2177                         *seq_rtt = now - scb->when;
2178
2179                 if (tp->fackets_out) {
2180                         __u32 dval = min(tp->fackets_out, packets_acked);
2181                         tp->fackets_out -= dval;
2182                 }
2183                 tp->packets_out -= packets_acked;
2184
2185                 BUG_ON(tcp_skb_pcount(skb) == 0);
2186                 BUG_ON(!before(scb->seq, scb->end_seq));
2187         }
2188
2189         return acked;
2190 }
2191
2192 static u32 tcp_usrtt(const struct sk_buff *skb)
2193 {
2194         struct timeval tv, now;
2195
2196         do_gettimeofday(&now);
2197         skb_get_timestamp(skb, &tv);
2198         return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec);
2199 }
2200
2201 /* Remove acknowledged frames from the retransmission queue. */
2202 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2203 {
2204         struct tcp_sock *tp = tcp_sk(sk);
2205         const struct inet_connection_sock *icsk = inet_csk(sk);
2206         struct sk_buff *skb;
2207         __u32 now = tcp_time_stamp;
2208         int acked = 0;
2209         __s32 seq_rtt = -1;
2210         u32 pkts_acked = 0;
2211         void (*rtt_sample)(struct sock *sk, u32 usrtt)
2212                 = icsk->icsk_ca_ops->rtt_sample;
2213
2214         while ((skb = skb_peek(&sk->sk_write_queue)) &&
2215                skb != sk->sk_send_head) {
2216                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2217                 __u8 sacked = scb->sacked;
2218
2219                 /* If our packet is before the ack sequence we can
2220                  * discard it as it's confirmed to have arrived at
2221                  * the other end.
2222                  */
2223                 if (after(scb->end_seq, tp->snd_una)) {
2224                         if (tcp_skb_pcount(skb) > 1 &&
2225                             after(tp->snd_una, scb->seq))
2226                                 acked |= tcp_tso_acked(sk, skb,
2227                                                        now, &seq_rtt);
2228                         break;
2229                 }
2230
2231                 /* Initial outgoing SYN's get put onto the write_queue
2232                  * just like anything else we transmit.  It is not
2233                  * true data, and if we misinform our callers that
2234                  * this ACK acks real data, we will erroneously exit
2235                  * connection startup slow start one packet too
2236                  * quickly.  This is severely frowned upon behavior.
2237                  */
2238                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2239                         acked |= FLAG_DATA_ACKED;
2240                         ++pkts_acked;
2241                 } else {
2242                         acked |= FLAG_SYN_ACKED;
2243                         tp->retrans_stamp = 0;
2244                 }
2245
2246                 if (sacked) {
2247                         if (sacked & TCPCB_RETRANS) {
2248                                 if(sacked & TCPCB_SACKED_RETRANS)
2249                                         tp->retrans_out -= tcp_skb_pcount(skb);
2250                                 acked |= FLAG_RETRANS_DATA_ACKED;
2251                                 seq_rtt = -1;
2252                         } else if (seq_rtt < 0) {
2253                                 seq_rtt = now - scb->when;
2254                                 if (rtt_sample)
2255                                         (*rtt_sample)(sk, tcp_usrtt(skb));
2256                         }
2257                         if (sacked & TCPCB_SACKED_ACKED)
2258                                 tp->sacked_out -= tcp_skb_pcount(skb);
2259                         if (sacked & TCPCB_LOST)
2260                                 tp->lost_out -= tcp_skb_pcount(skb);
2261                         if (sacked & TCPCB_URG) {
2262                                 if (tp->urg_mode &&
2263                                     !before(scb->end_seq, tp->snd_up))
2264                                         tp->urg_mode = 0;
2265                         }
2266                 } else if (seq_rtt < 0) {
2267                         seq_rtt = now - scb->when;
2268                         if (rtt_sample)
2269                                 (*rtt_sample)(sk, tcp_usrtt(skb));
2270                 }
2271                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2272                 tcp_packets_out_dec(tp, skb);
2273                 __skb_unlink(skb, &sk->sk_write_queue);
2274                 sk_stream_free_skb(sk, skb);
2275                 clear_all_retrans_hints(tp);
2276         }
2277
2278         if (acked&FLAG_ACKED) {
2279                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2280                 tcp_ack_packets_out(sk, tp);
2281
2282                 if (icsk->icsk_ca_ops->pkts_acked)
2283                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2284         }
2285
2286 #if FASTRETRANS_DEBUG > 0
2287         BUG_TRAP((int)tp->sacked_out >= 0);
2288         BUG_TRAP((int)tp->lost_out >= 0);
2289         BUG_TRAP((int)tp->retrans_out >= 0);
2290         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2291                 const struct inet_connection_sock *icsk = inet_csk(sk);
2292                 if (tp->lost_out) {
2293                         printk(KERN_DEBUG "Leak l=%u %d\n",
2294                                tp->lost_out, icsk->icsk_ca_state);
2295                         tp->lost_out = 0;
2296                 }
2297                 if (tp->sacked_out) {
2298                         printk(KERN_DEBUG "Leak s=%u %d\n",
2299                                tp->sacked_out, icsk->icsk_ca_state);
2300                         tp->sacked_out = 0;
2301                 }
2302                 if (tp->retrans_out) {
2303                         printk(KERN_DEBUG "Leak r=%u %d\n",
2304                                tp->retrans_out, icsk->icsk_ca_state);
2305                         tp->retrans_out = 0;
2306                 }
2307         }
2308 #endif
2309         *seq_rtt_p = seq_rtt;
2310         return acked;
2311 }
2312
2313 static void tcp_ack_probe(struct sock *sk)
2314 {
2315         const struct tcp_sock *tp = tcp_sk(sk);
2316         struct inet_connection_sock *icsk = inet_csk(sk);
2317
2318         /* Was it a usable window open? */
2319
2320         if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2321                    tp->snd_una + tp->snd_wnd)) {
2322                 icsk->icsk_backoff = 0;
2323                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2324                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2325                  * This function is not for random using!
2326                  */
2327         } else {
2328                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2329                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2330                                           TCP_RTO_MAX);
2331         }
2332 }
2333
2334 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2335 {
2336         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2337                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2338 }
2339
2340 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2341 {
2342         const struct tcp_sock *tp = tcp_sk(sk);
2343         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2344                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2345 }
2346
2347 /* Check that window update is acceptable.
2348  * The function assumes that snd_una<=ack<=snd_next.
2349  */
2350 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2351                                         const u32 ack_seq, const u32 nwin)
2352 {
2353         return (after(ack, tp->snd_una) ||
2354                 after(ack_seq, tp->snd_wl1) ||
2355                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2356 }
2357
2358 /* Update our send window.
2359  *
2360  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2361  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2362  */
2363 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2364                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2365 {
2366         int flag = 0;
2367         u32 nwin = ntohs(skb->h.th->window);
2368
2369         if (likely(!skb->h.th->syn))
2370                 nwin <<= tp->rx_opt.snd_wscale;
2371
2372         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2373                 flag |= FLAG_WIN_UPDATE;
2374                 tcp_update_wl(tp, ack, ack_seq);
2375
2376                 if (tp->snd_wnd != nwin) {
2377                         tp->snd_wnd = nwin;
2378
2379                         /* Note, it is the only place, where
2380                          * fast path is recovered for sending TCP.
2381                          */
2382                         tp->pred_flags = 0;
2383                         tcp_fast_path_check(sk, tp);
2384
2385                         if (nwin > tp->max_window) {
2386                                 tp->max_window = nwin;
2387                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2388                         }
2389                 }
2390         }
2391
2392         tp->snd_una = ack;
2393
2394         return flag;
2395 }
2396
2397 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2398 {
2399         struct tcp_sock *tp = tcp_sk(sk);
2400         
2401         tcp_sync_left_out(tp);
2402         
2403         if (tp->snd_una == prior_snd_una ||
2404             !before(tp->snd_una, tp->frto_highmark)) {
2405                 /* RTO was caused by loss, start retransmitting in
2406                  * go-back-N slow start
2407                  */
2408                 tcp_enter_frto_loss(sk);
2409                 return;
2410         }
2411
2412         if (tp->frto_counter == 1) {
2413                 /* First ACK after RTO advances the window: allow two new
2414                  * segments out.
2415                  */
2416                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2417         } else {
2418                 /* Also the second ACK after RTO advances the window.
2419                  * The RTO was likely spurious. Reduce cwnd and continue
2420                  * in congestion avoidance
2421                  */
2422                 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2423                 tcp_moderate_cwnd(tp);
2424         }
2425
2426         /* F-RTO affects on two new ACKs following RTO.
2427          * At latest on third ACK the TCP behavior is back to normal.
2428          */
2429         tp->frto_counter = (tp->frto_counter + 1) % 3;
2430 }
2431
2432 /* This routine deals with incoming acks, but not outgoing ones. */
2433 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2434 {
2435         struct inet_connection_sock *icsk = inet_csk(sk);
2436         struct tcp_sock *tp = tcp_sk(sk);
2437         u32 prior_snd_una = tp->snd_una;
2438         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2439         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2440         u32 prior_in_flight;
2441         s32 seq_rtt;
2442         int prior_packets;
2443
2444         /* If the ack is newer than sent or older than previous acks
2445          * then we can probably ignore it.
2446          */
2447         if (after(ack, tp->snd_nxt))
2448                 goto uninteresting_ack;
2449
2450         if (before(ack, prior_snd_una))
2451                 goto old_ack;
2452
2453         if (sysctl_tcp_abc && icsk->icsk_ca_state < TCP_CA_CWR)
2454                 tp->bytes_acked += ack - prior_snd_una;
2455
2456         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2457                 /* Window is constant, pure forward advance.
2458                  * No more checks are required.
2459                  * Note, we use the fact that SND.UNA>=SND.WL2.
2460                  */
2461                 tcp_update_wl(tp, ack, ack_seq);
2462                 tp->snd_una = ack;
2463                 flag |= FLAG_WIN_UPDATE;
2464
2465                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2466
2467                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2468         } else {
2469                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2470                         flag |= FLAG_DATA;
2471                 else
2472                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2473
2474                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2475
2476                 if (TCP_SKB_CB(skb)->sacked)
2477                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2478
2479                 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2480                         flag |= FLAG_ECE;
2481
2482                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2483         }
2484
2485         /* We passed data and got it acked, remove any soft error
2486          * log. Something worked...
2487          */
2488         sk->sk_err_soft = 0;
2489         tp->rcv_tstamp = tcp_time_stamp;
2490         prior_packets = tp->packets_out;
2491         if (!prior_packets)
2492                 goto no_queue;
2493
2494         prior_in_flight = tcp_packets_in_flight(tp);
2495
2496         /* See if we can take anything off of the retransmit queue. */
2497         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2498
2499         if (tp->frto_counter)
2500                 tcp_process_frto(sk, prior_snd_una);
2501
2502         if (tcp_ack_is_dubious(sk, flag)) {
2503                 /* Advance CWND, if state allows this. */
2504                 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
2505                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2506                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2507         } else {
2508                 if ((flag & FLAG_DATA_ACKED))
2509                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2510         }
2511
2512         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2513                 dst_confirm(sk->sk_dst_cache);
2514
2515         return 1;
2516
2517 no_queue:
2518         icsk->icsk_probes_out = 0;
2519
2520         /* If this ack opens up a zero window, clear backoff.  It was
2521          * being used to time the probes, and is probably far higher than
2522          * it needs to be for normal retransmission.
2523          */
2524         if (sk->sk_send_head)
2525                 tcp_ack_probe(sk);
2526         return 1;
2527
2528 old_ack:
2529         if (TCP_SKB_CB(skb)->sacked)
2530                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2531
2532 uninteresting_ack:
2533         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2534         return 0;
2535 }
2536
2537
2538 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2539  * But, this can also be called on packets in the established flow when
2540  * the fast version below fails.
2541  */
2542 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2543 {
2544         unsigned char *ptr;
2545         struct tcphdr *th = skb->h.th;
2546         int length=(th->doff*4)-sizeof(struct tcphdr);
2547
2548         ptr = (unsigned char *)(th + 1);
2549         opt_rx->saw_tstamp = 0;
2550
2551         while(length>0) {
2552                 int opcode=*ptr++;
2553                 int opsize;
2554
2555                 switch (opcode) {
2556                         case TCPOPT_EOL:
2557                                 return;
2558                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2559                                 length--;
2560                                 continue;
2561                         default:
2562                                 opsize=*ptr++;
2563                                 if (opsize < 2) /* "silly options" */
2564                                         return;
2565                                 if (opsize > length)
2566                                         return; /* don't parse partial options */
2567                                 switch(opcode) {
2568                                 case TCPOPT_MSS:
2569                                         if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2570                                                 u16 in_mss = ntohs(get_unaligned((__u16 *)ptr));
2571                                                 if (in_mss) {
2572                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2573                                                                 in_mss = opt_rx->user_mss;
2574                                                         opt_rx->mss_clamp = in_mss;
2575                                                 }
2576                                         }
2577                                         break;
2578                                 case TCPOPT_WINDOW:
2579                                         if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2580                                                 if (sysctl_tcp_window_scaling) {
2581                                                         __u8 snd_wscale = *(__u8 *) ptr;
2582                                                         opt_rx->wscale_ok = 1;
2583                                                         if (snd_wscale > 14) {
2584                                                                 if(net_ratelimit())
2585                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2586                                                                                "scaling value %d >14 received.\n",
2587                                                                                snd_wscale);
2588                                                                 snd_wscale = 14;
2589                                                         }
2590                                                         opt_rx->snd_wscale = snd_wscale;
2591                                                 }
2592                                         break;
2593                                 case TCPOPT_TIMESTAMP:
2594                                         if(opsize==TCPOLEN_TIMESTAMP) {
2595                                                 if ((estab && opt_rx->tstamp_ok) ||
2596                                                     (!estab && sysctl_tcp_timestamps)) {
2597                                                         opt_rx->saw_tstamp = 1;
2598                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr));
2599                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4)));
2600                                                 }
2601                                         }
2602                                         break;
2603                                 case TCPOPT_SACK_PERM:
2604                                         if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2605                                                 if (sysctl_tcp_sack) {
2606                                                         opt_rx->sack_ok = 1;
2607                                                         tcp_sack_reset(opt_rx);
2608                                                 }
2609                                         }
2610                                         break;
2611
2612                                 case TCPOPT_SACK:
2613                                         if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2614                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2615                                            opt_rx->sack_ok) {
2616                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2617                                         }
2618                                 };
2619                                 ptr+=opsize-2;
2620                                 length-=opsize;
2621                 };
2622         }
2623 }
2624
2625 /* Fast parse options. This hopes to only see timestamps.
2626  * If it is wrong it falls back on tcp_parse_options().
2627  */
2628 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2629                                   struct tcp_sock *tp)
2630 {
2631         if (th->doff == sizeof(struct tcphdr)>>2) {
2632                 tp->rx_opt.saw_tstamp = 0;
2633                 return 0;
2634         } else if (tp->rx_opt.tstamp_ok &&
2635                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2636                 __u32 *ptr = (__u32 *)(th + 1);
2637                 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2638                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2639                         tp->rx_opt.saw_tstamp = 1;
2640                         ++ptr;
2641                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2642                         ++ptr;
2643                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2644                         return 1;
2645                 }
2646         }
2647         tcp_parse_options(skb, &tp->rx_opt, 1);
2648         return 1;
2649 }
2650
2651 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2652 {
2653         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2654         tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2655 }
2656
2657 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2658 {
2659         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2660                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2661                  * extra check below makes sure this can only happen
2662                  * for pure ACK frames.  -DaveM
2663                  *
2664                  * Not only, also it occurs for expired timestamps.
2665                  */
2666
2667                 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2668                    xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2669                         tcp_store_ts_recent(tp);
2670         }
2671 }
2672
2673 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2674  *
2675  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2676  * it can pass through stack. So, the following predicate verifies that
2677  * this segment is not used for anything but congestion avoidance or
2678  * fast retransmit. Moreover, we even are able to eliminate most of such
2679  * second order effects, if we apply some small "replay" window (~RTO)
2680  * to timestamp space.
2681  *
2682  * All these measures still do not guarantee that we reject wrapped ACKs
2683  * on networks with high bandwidth, when sequence space is recycled fastly,
2684  * but it guarantees that such events will be very rare and do not affect
2685  * connection seriously. This doesn't look nice, but alas, PAWS is really
2686  * buggy extension.
2687  *
2688  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2689  * states that events when retransmit arrives after original data are rare.
2690  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2691  * the biggest problem on large power networks even with minor reordering.
2692  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2693  * up to bandwidth of 18Gigabit/sec. 8) ]
2694  */
2695
2696 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2697 {
2698         struct tcp_sock *tp = tcp_sk(sk);
2699         struct tcphdr *th = skb->h.th;
2700         u32 seq = TCP_SKB_CB(skb)->seq;
2701         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2702
2703         return (/* 1. Pure ACK with correct sequence number. */
2704                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2705
2706                 /* 2. ... and duplicate ACK. */
2707                 ack == tp->snd_una &&
2708
2709                 /* 3. ... and does not update window. */
2710                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2711
2712                 /* 4. ... and sits in replay window. */
2713                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
2714 }
2715
2716 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
2717 {
2718         const struct tcp_sock *tp = tcp_sk(sk);
2719         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2720                 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2721                 !tcp_disordered_ack(sk, skb));
2722 }
2723
2724 /* Check segment sequence number for validity.
2725  *
2726  * Segment controls are considered valid, if the segment
2727  * fits to the window after truncation to the window. Acceptability
2728  * of data (and SYN, FIN, of course) is checked separately.
2729  * See tcp_data_queue(), for example.
2730  *
2731  * Also, controls (RST is main one) are accepted using RCV.WUP instead
2732  * of RCV.NXT. Peer still did not advance his SND.UNA when we
2733  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2734  * (borrowed from freebsd)
2735  */
2736
2737 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2738 {
2739         return  !before(end_seq, tp->rcv_wup) &&
2740                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2741 }
2742
2743 /* When we get a reset we do this. */
2744 static void tcp_reset(struct sock *sk)
2745 {
2746         /* We want the right error as BSD sees it (and indeed as we do). */
2747         switch (sk->sk_state) {
2748                 case TCP_SYN_SENT:
2749                         sk->sk_err = ECONNREFUSED;
2750                         break;
2751                 case TCP_CLOSE_WAIT:
2752                         sk->sk_err = EPIPE;
2753                         break;
2754                 case TCP_CLOSE:
2755                         return;
2756                 default:
2757                         sk->sk_err = ECONNRESET;
2758         }
2759
2760         if (!sock_flag(sk, SOCK_DEAD))
2761                 sk->sk_error_report(sk);
2762
2763         tcp_done(sk);
2764 }
2765
2766 /*
2767  *      Process the FIN bit. This now behaves as it is supposed to work
2768  *      and the FIN takes effect when it is validly part of sequence
2769  *      space. Not before when we get holes.
2770  *
2771  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2772  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
2773  *      TIME-WAIT)
2774  *
2775  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
2776  *      close and we go into CLOSING (and later onto TIME-WAIT)
2777  *
2778  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2779  */
2780 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2781 {
2782         struct tcp_sock *tp = tcp_sk(sk);
2783
2784         inet_csk_schedule_ack(sk);
2785
2786         sk->sk_shutdown |= RCV_SHUTDOWN;
2787         sock_set_flag(sk, SOCK_DONE);
2788
2789         switch (sk->sk_state) {
2790                 case TCP_SYN_RECV:
2791                 case TCP_ESTABLISHED:
2792                         /* Move to CLOSE_WAIT */
2793                         tcp_set_state(sk, TCP_CLOSE_WAIT);
2794                         inet_csk(sk)->icsk_ack.pingpong = 1;
2795                         break;
2796
2797                 case TCP_CLOSE_WAIT:
2798                 case TCP_CLOSING:
2799                         /* Received a retransmission of the FIN, do
2800                          * nothing.
2801                          */
2802                         break;
2803                 case TCP_LAST_ACK:
2804                         /* RFC793: Remain in the LAST-ACK state. */
2805                         break;
2806
2807                 case TCP_FIN_WAIT1:
2808                         /* This case occurs when a simultaneous close
2809                          * happens, we must ack the received FIN and
2810                          * enter the CLOSING state.
2811                          */
2812                         tcp_send_ack(sk);
2813                         tcp_set_state(sk, TCP_CLOSING);
2814                         break;
2815                 case TCP_FIN_WAIT2:
2816                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
2817                         tcp_send_ack(sk);
2818                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2819                         break;
2820                 default:
2821                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2822                          * cases we should never reach this piece of code.
2823                          */
2824                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2825                                __FUNCTION__, sk->sk_state);
2826                         break;
2827         };
2828
2829         /* It _is_ possible, that we have something out-of-order _after_ FIN.
2830          * Probably, we should reset in this case. For now drop them.
2831          */
2832         __skb_queue_purge(&tp->out_of_order_queue);
2833         if (tp->rx_opt.sack_ok)
2834                 tcp_sack_reset(&tp->rx_opt);
2835         sk_stream_mem_reclaim(sk);
2836
2837         if (!sock_flag(sk, SOCK_DEAD)) {
2838                 sk->sk_state_change(sk);
2839
2840                 /* Do not send POLL_HUP for half duplex close. */
2841                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2842                     sk->sk_state == TCP_CLOSE)
2843                         sk_wake_async(sk, 1, POLL_HUP);
2844                 else
2845                         sk_wake_async(sk, 1, POLL_IN);
2846         }
2847 }
2848
2849 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2850 {
2851         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2852                 if (before(seq, sp->start_seq))
2853                         sp->start_seq = seq;
2854                 if (after(end_seq, sp->end_seq))
2855                         sp->end_seq = end_seq;
2856                 return 1;
2857         }
2858         return 0;
2859 }
2860
2861 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
2862 {
2863         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2864                 if (before(seq, tp->rcv_nxt))
2865                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
2866                 else
2867                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
2868
2869                 tp->rx_opt.dsack = 1;
2870                 tp->duplicate_sack[0].start_seq = seq;
2871                 tp->duplicate_sack[0].end_seq = end_seq;
2872                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
2873         }
2874 }
2875
2876 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
2877 {
2878         if (!tp->rx_opt.dsack)
2879                 tcp_dsack_set(tp, seq, end_seq);
2880         else
2881                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2882 }
2883
2884 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2885 {
2886         struct tcp_sock *tp = tcp_sk(sk);
2887
2888         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2889             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2890                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2891                 tcp_enter_quickack_mode(sk);
2892
2893                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2894                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2895
2896                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2897                                 end_seq = tp->rcv_nxt;
2898                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2899                 }
2900         }
2901
2902         tcp_send_ack(sk);
2903 }
2904
2905 /* These routines update the SACK block as out-of-order packets arrive or
2906  * in-order packets close up the sequence space.
2907  */
2908 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
2909 {
2910         int this_sack;
2911         struct tcp_sack_block *sp = &tp->selective_acks[0];
2912         struct tcp_sack_block *swalk = sp+1;
2913
2914         /* See if the recent change to the first SACK eats into
2915          * or hits the sequence space of other SACK blocks, if so coalesce.
2916          */
2917         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
2918                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2919                         int i;
2920
2921                         /* Zap SWALK, by moving every further SACK up by one slot.
2922                          * Decrease num_sacks.
2923                          */
2924                         tp->rx_opt.num_sacks--;
2925                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2926                         for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
2927                                 sp[i] = sp[i+1];
2928                         continue;
2929                 }
2930                 this_sack++, swalk++;
2931         }
2932 }
2933
2934 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2935 {
2936         __u32 tmp;
2937
2938         tmp = sack1->start_seq;
2939         sack1->start_seq = sack2->start_seq;
2940         sack2->start_seq = tmp;
2941
2942         tmp = sack1->end_seq;
2943         sack1->end_seq = sack2->end_seq;
2944         sack2->end_seq = tmp;
2945 }
2946
2947 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
2948 {
2949         struct tcp_sock *tp = tcp_sk(sk);
2950         struct tcp_sack_block *sp = &tp->selective_acks[0];
2951         int cur_sacks = tp->rx_opt.num_sacks;
2952         int this_sack;
2953
2954         if (!cur_sacks)
2955                 goto new_sack;
2956
2957         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
2958                 if (tcp_sack_extend(sp, seq, end_seq)) {
2959                         /* Rotate this_sack to the first one. */
2960                         for (; this_sack>0; this_sack--, sp--)
2961                                 tcp_sack_swap(sp, sp-1);
2962                         if (cur_sacks > 1)
2963                                 tcp_sack_maybe_coalesce(tp);
2964                         return;
2965                 }
2966         }
2967
2968         /* Could not find an adjacent existing SACK, build a new one,
2969          * put it at the front, and shift everyone else down.  We
2970          * always know there is at least one SACK present already here.
2971          *
2972          * If the sack array is full, forget about the last one.
2973          */
2974         if (this_sack >= 4) {
2975                 this_sack--;
2976                 tp->rx_opt.num_sacks--;
2977                 sp--;
2978         }
2979         for(; this_sack > 0; this_sack--, sp--)
2980                 *sp = *(sp-1);
2981
2982 new_sack:
2983         /* Build the new head SACK, and we're done. */
2984         sp->start_seq = seq;
2985         sp->end_seq = end_seq;
2986         tp->rx_opt.num_sacks++;
2987         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2988 }
2989
2990 /* RCV.NXT advances, some SACKs should be eaten. */
2991
2992 static void tcp_sack_remove(struct tcp_sock *tp)
2993 {
2994         struct tcp_sack_block *sp = &tp->selective_acks[0];
2995         int num_sacks = tp->rx_opt.num_sacks;
2996         int this_sack;
2997
2998         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2999         if (skb_queue_empty(&tp->out_of_order_queue)) {
3000                 tp->rx_opt.num_sacks = 0;
3001                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3002                 return;
3003         }
3004
3005         for(this_sack = 0; this_sack < num_sacks; ) {
3006                 /* Check if the start of the sack is covered by RCV.NXT. */
3007                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3008                         int i;
3009
3010                         /* RCV.NXT must cover all the block! */
3011                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3012
3013                         /* Zap this SACK, by moving forward any other SACKS. */
3014                         for (i=this_sack+1; i < num_sacks; i++)
3015                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3016                         num_sacks--;
3017                         continue;
3018                 }
3019                 this_sack++;
3020                 sp++;
3021         }
3022         if (num_sacks != tp->rx_opt.num_sacks) {
3023                 tp->rx_opt.num_sacks = num_sacks;
3024                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3025         }
3026 }
3027
3028 /* This one checks to see if we can put data from the
3029  * out_of_order queue into the receive_queue.
3030  */
3031 static void tcp_ofo_queue(struct sock *sk)
3032 {
3033         struct tcp_sock *tp = tcp_sk(sk);
3034         __u32 dsack_high = tp->rcv_nxt;
3035         struct sk_buff *skb;
3036
3037         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3038                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3039                         break;
3040
3041                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3042                         __u32 dsack = dsack_high;
3043                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3044                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3045                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3046                 }
3047
3048                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3049                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3050                         __skb_unlink(skb, &tp->out_of_order_queue);
3051                         __kfree_skb(skb);
3052                         continue;
3053                 }
3054                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3055                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3056                            TCP_SKB_CB(skb)->end_seq);
3057
3058                 __skb_unlink(skb, &tp->out_of_order_queue);
3059                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3060                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3061                 if(skb->h.th->fin)
3062                         tcp_fin(skb, sk, skb->h.th);
3063         }
3064 }
3065
3066 static int tcp_prune_queue(struct sock *sk);
3067
3068 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3069 {
3070         struct tcphdr *th = skb->h.th;
3071         struct tcp_sock *tp = tcp_sk(sk);
3072         int eaten = -1;
3073
3074         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3075                 goto drop;
3076
3077         __skb_pull(skb, th->doff*4);
3078
3079         TCP_ECN_accept_cwr(tp, skb);
3080
3081         if (tp->rx_opt.dsack) {
3082                 tp->rx_opt.dsack = 0;
3083                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3084                                                     4 - tp->rx_opt.tstamp_ok);
3085         }
3086
3087         /*  Queue data for delivery to the user.
3088          *  Packets in sequence go to the receive queue.
3089          *  Out of sequence packets to the out_of_order_queue.
3090          */
3091         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3092                 if (tcp_receive_window(tp) == 0)
3093                         goto out_of_window;
3094
3095                 /* Ok. In sequence. In window. */
3096                 if (tp->ucopy.task == current &&
3097                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3098                     sock_owned_by_user(sk) && !tp->urg_data) {
3099                         int chunk = min_t(unsigned int, skb->len,
3100                                                         tp->ucopy.len);
3101
3102                         __set_current_state(TASK_RUNNING);
3103
3104                         local_bh_enable();
3105                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3106                                 tp->ucopy.len -= chunk;
3107                                 tp->copied_seq += chunk;
3108                                 eaten = (chunk == skb->len && !th->fin);
3109                                 tcp_rcv_space_adjust(sk);
3110                         }
3111                         local_bh_disable();
3112                 }
3113
3114                 if (eaten <= 0) {
3115 queue_and_out:
3116                         if (eaten < 0 &&
3117                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3118                              !sk_stream_rmem_schedule(sk, skb))) {
3119                                 if (tcp_prune_queue(sk) < 0 ||
3120                                     !sk_stream_rmem_schedule(sk, skb))
3121                                         goto drop;
3122                         }
3123                         sk_stream_set_owner_r(skb, sk);
3124                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3125                 }
3126                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3127                 if(skb->len)
3128                         tcp_event_data_recv(sk, tp, skb);
3129                 if(th->fin)
3130                         tcp_fin(skb, sk, th);
3131
3132                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3133                         tcp_ofo_queue(sk);
3134
3135                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3136                          * gap in queue is filled.
3137                          */
3138                         if (skb_queue_empty(&tp->out_of_order_queue))
3139                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3140                 }
3141
3142                 if (tp->rx_opt.num_sacks)
3143                         tcp_sack_remove(tp);
3144
3145                 tcp_fast_path_check(sk, tp);
3146
3147                 if (eaten > 0)
3148                         __kfree_skb(skb);
3149                 else if (!sock_flag(sk, SOCK_DEAD))
3150                         sk->sk_data_ready(sk, 0);
3151                 return;
3152         }
3153
3154         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3155                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3156                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3157                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3158
3159 out_of_window:
3160                 tcp_enter_quickack_mode(sk);
3161                 inet_csk_schedule_ack(sk);
3162 drop:
3163                 __kfree_skb(skb);
3164                 return;
3165         }
3166
3167         /* Out of window. F.e. zero window probe. */
3168         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3169                 goto out_of_window;
3170
3171         tcp_enter_quickack_mode(sk);
3172
3173         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3174                 /* Partial packet, seq < rcv_next < end_seq */
3175                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3176                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3177                            TCP_SKB_CB(skb)->end_seq);
3178
3179                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3180                 
3181                 /* If window is closed, drop tail of packet. But after
3182                  * remembering D-SACK for its head made in previous line.
3183                  */
3184                 if (!tcp_receive_window(tp))
3185                         goto out_of_window;
3186                 goto queue_and_out;
3187         }
3188
3189         TCP_ECN_check_ce(tp, skb);
3190
3191         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3192             !sk_stream_rmem_schedule(sk, skb)) {
3193                 if (tcp_prune_queue(sk) < 0 ||
3194                     !sk_stream_rmem_schedule(sk, skb))
3195                         goto drop;
3196         }
3197
3198         /* Disable header prediction. */
3199         tp->pred_flags = 0;
3200         inet_csk_schedule_ack(sk);
3201
3202         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3203                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3204
3205         sk_stream_set_owner_r(skb, sk);
3206
3207         if (!skb_peek(&tp->out_of_order_queue)) {
3208                 /* Initial out of order segment, build 1 SACK. */
3209                 if (tp->rx_opt.sack_ok) {
3210                         tp->rx_opt.num_sacks = 1;
3211                         tp->rx_opt.dsack     = 0;
3212                         tp->rx_opt.eff_sacks = 1;
3213                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3214                         tp->selective_acks[0].end_seq =
3215                                                 TCP_SKB_CB(skb)->end_seq;
3216                 }
3217                 __skb_queue_head(&tp->out_of_order_queue,skb);
3218         } else {
3219                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3220                 u32 seq = TCP_SKB_CB(skb)->seq;
3221                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3222
3223                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3224                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3225
3226                         if (!tp->rx_opt.num_sacks ||
3227                             tp->selective_acks[0].end_seq != seq)
3228                                 goto add_sack;
3229
3230                         /* Common case: data arrive in order after hole. */
3231                         tp->selective_acks[0].end_seq = end_seq;
3232                         return;
3233                 }
3234
3235                 /* Find place to insert this segment. */
3236                 do {
3237                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3238                                 break;
3239                 } while ((skb1 = skb1->prev) !=
3240                          (struct sk_buff*)&tp->out_of_order_queue);
3241
3242                 /* Do skb overlap to previous one? */
3243                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3244                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3245                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3246                                 /* All the bits are present. Drop. */
3247                                 __kfree_skb(skb);
3248                                 tcp_dsack_set(tp, seq, end_seq);
3249                                 goto add_sack;
3250                         }
3251                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3252                                 /* Partial overlap. */
3253                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3254                         } else {
3255                                 skb1 = skb1->prev;
3256                         }
3257                 }
3258                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3259                 
3260                 /* And clean segments covered by new one as whole. */
3261                 while ((skb1 = skb->next) !=
3262                        (struct sk_buff*)&tp->out_of_order_queue &&
3263                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3264                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3265                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3266                                break;
3267                        }
3268                        __skb_unlink(skb1, &tp->out_of_order_queue);
3269                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3270                        __kfree_skb(skb1);
3271                 }
3272
3273 add_sack:
3274                 if (tp->rx_opt.sack_ok)
3275                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3276         }
3277 }
3278
3279 /* Collapse contiguous sequence of skbs head..tail with
3280  * sequence numbers start..end.
3281  * Segments with FIN/SYN are not collapsed (only because this
3282  * simplifies code)
3283  */
3284 static void
3285 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3286              struct sk_buff *head, struct sk_buff *tail,
3287              u32 start, u32 end)
3288 {
3289         struct sk_buff *skb;
3290
3291         /* First, check that queue is collapsible and find
3292          * the point where collapsing can be useful. */
3293         for (skb = head; skb != tail; ) {
3294                 /* No new bits? It is possible on ofo queue. */
3295                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3296                         struct sk_buff *next = skb->next;
3297                         __skb_unlink(skb, list);
3298                         __kfree_skb(skb);
3299                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3300                         skb = next;
3301                         continue;
3302                 }
3303
3304                 /* The first skb to collapse is:
3305                  * - not SYN/FIN and
3306                  * - bloated or contains data before "start" or
3307                  *   overlaps to the next one.
3308                  */
3309                 if (!skb->h.th->syn && !skb->h.th->fin &&
3310                     (tcp_win_from_space(skb->truesize) > skb->len ||
3311                      before(TCP_SKB_CB(skb)->seq, start) ||
3312                      (skb->next != tail &&
3313                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3314                         break;
3315
3316                 /* Decided to skip this, advance start seq. */
3317                 start = TCP_SKB_CB(skb)->end_seq;
3318                 skb = skb->next;
3319         }
3320         if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3321                 return;
3322
3323         while (before(start, end)) {
3324                 struct sk_buff *nskb;
3325                 int header = skb_headroom(skb);
3326                 int copy = SKB_MAX_ORDER(header, 0);
3327
3328                 /* Too big header? This can happen with IPv6. */
3329                 if (copy < 0)
3330                         return;
3331                 if (end-start < copy)
3332                         copy = end-start;
3333                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3334                 if (!nskb)
3335                         return;
3336                 skb_reserve(nskb, header);
3337                 memcpy(nskb->head, skb->head, header);
3338                 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3339                 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3340                 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3341                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3342                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3343                 __skb_insert(nskb, skb->prev, skb, list);
3344                 sk_stream_set_owner_r(nskb, sk);
3345
3346                 /* Copy data, releasing collapsed skbs. */
3347                 while (copy > 0) {
3348                         int offset = start - TCP_SKB_CB(skb)->seq;
3349                         int size = TCP_SKB_CB(skb)->end_seq - start;
3350
3351                         BUG_ON(offset < 0);
3352                         if (size > 0) {
3353                                 size = min(copy, size);
3354                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3355                                         BUG();
3356                                 TCP_SKB_CB(nskb)->end_seq += size;
3357                                 copy -= size;
3358                                 start += size;
3359                         }
3360                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3361                                 struct sk_buff *next = skb->next;
3362                                 __skb_unlink(skb, list);
3363                                 __kfree_skb(skb);
3364                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3365                                 skb = next;
3366                                 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3367                                         return;
3368                         }
3369                 }
3370         }
3371 }
3372
3373 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3374  * and tcp_collapse() them until all the queue is collapsed.
3375  */
3376 static void tcp_collapse_ofo_queue(struct sock *sk)
3377 {
3378         struct tcp_sock *tp = tcp_sk(sk);
3379         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3380         struct sk_buff *head;
3381         u32 start, end;
3382
3383         if (skb == NULL)
3384                 return;
3385
3386         start = TCP_SKB_CB(skb)->seq;
3387         end = TCP_SKB_CB(skb)->end_seq;
3388         head = skb;
3389
3390         for (;;) {
3391                 skb = skb->next;
3392
3393                 /* Segment is terminated when we see gap or when
3394                  * we are at the end of all the queue. */
3395                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3396                     after(TCP_SKB_CB(skb)->seq, end) ||
3397                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3398                         tcp_collapse(sk, &tp->out_of_order_queue,
3399                                      head, skb, start, end);
3400                         head = skb;
3401                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3402                                 break;
3403                         /* Start new segment */
3404                         start = TCP_SKB_CB(skb)->seq;
3405                         end = TCP_SKB_CB(skb)->end_seq;
3406                 } else {
3407                         if (before(TCP_SKB_CB(skb)->seq, start))
3408                                 start = TCP_SKB_CB(skb)->seq;
3409                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3410                                 end = TCP_SKB_CB(skb)->end_seq;
3411                 }
3412         }
3413 }
3414
3415 /* Reduce allocated memory if we can, trying to get
3416  * the socket within its memory limits again.
3417  *
3418  * Return less than zero if we should start dropping frames
3419  * until the socket owning process reads some of the data
3420  * to stabilize the situation.
3421  */
3422 static int tcp_prune_queue(struct sock *sk)
3423 {
3424         struct tcp_sock *tp = tcp_sk(sk); 
3425
3426         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3427
3428         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3429
3430         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3431                 tcp_clamp_window(sk, tp);
3432         else if (tcp_memory_pressure)
3433                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3434
3435         tcp_collapse_ofo_queue(sk);
3436         tcp_collapse(sk, &sk->sk_receive_queue,
3437                      sk->sk_receive_queue.next,
3438                      (struct sk_buff*)&sk->sk_receive_queue,
3439                      tp->copied_seq, tp->rcv_nxt);
3440         sk_stream_mem_reclaim(sk);
3441
3442         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3443                 return 0;
3444
3445         /* Collapsing did not help, destructive actions follow.
3446          * This must not ever occur. */
3447
3448         /* First, purge the out_of_order queue. */
3449         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3450                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3451                 __skb_queue_purge(&tp->out_of_order_queue);
3452
3453                 /* Reset SACK state.  A conforming SACK implementation will
3454                  * do the same at a timeout based retransmit.  When a connection
3455                  * is in a sad state like this, we care only about integrity
3456                  * of the connection not performance.
3457                  */
3458                 if (tp->rx_opt.sack_ok)
3459                         tcp_sack_reset(&tp->rx_opt);
3460                 sk_stream_mem_reclaim(sk);
3461         }
3462
3463         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3464                 return 0;
3465
3466         /* If we are really being abused, tell the caller to silently
3467          * drop receive data on the floor.  It will get retransmitted
3468          * and hopefully then we'll have sufficient space.
3469          */
3470         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3471
3472         /* Massive buffer overcommit. */
3473         tp->pred_flags = 0;
3474         return -1;
3475 }
3476
3477
3478 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3479  * As additional protections, we do not touch cwnd in retransmission phases,
3480  * and if application hit its sndbuf limit recently.
3481  */
3482 void tcp_cwnd_application_limited(struct sock *sk)
3483 {
3484         struct tcp_sock *tp = tcp_sk(sk);
3485
3486         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3487             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3488                 /* Limited by application or receiver window. */
3489                 u32 win_used = max(tp->snd_cwnd_used, 2U);
3490                 if (win_used < tp->snd_cwnd) {
3491                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3492                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3493                 }
3494                 tp->snd_cwnd_used = 0;
3495         }
3496         tp->snd_cwnd_stamp = tcp_time_stamp;
3497 }
3498
3499 static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3500 {
3501         /* If the user specified a specific send buffer setting, do
3502          * not modify it.
3503          */
3504         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3505                 return 0;
3506
3507         /* If we are under global TCP memory pressure, do not expand.  */
3508         if (tcp_memory_pressure)
3509                 return 0;
3510
3511         /* If we are under soft global TCP memory pressure, do not expand.  */
3512         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3513                 return 0;
3514
3515         /* If we filled the congestion window, do not expand.  */
3516         if (tp->packets_out >= tp->snd_cwnd)
3517                 return 0;
3518
3519         return 1;
3520 }
3521
3522 /* When incoming ACK allowed to free some skb from write_queue,
3523  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3524  * on the exit from tcp input handler.
3525  *
3526  * PROBLEM: sndbuf expansion does not work well with largesend.
3527  */
3528 static void tcp_new_space(struct sock *sk)
3529 {
3530         struct tcp_sock *tp = tcp_sk(sk);
3531
3532         if (tcp_should_expand_sndbuf(sk, tp)) {
3533                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3534                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3535                     demanded = max_t(unsigned int, tp->snd_cwnd,
3536                                                    tp->reordering + 1);
3537                 sndmem *= 2*demanded;
3538                 if (sndmem > sk->sk_sndbuf)
3539                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3540                 tp->snd_cwnd_stamp = tcp_time_stamp;
3541         }
3542
3543         sk->sk_write_space(sk);
3544 }
3545
3546 static void tcp_check_space(struct sock *sk)
3547 {
3548         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3549                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3550                 if (sk->sk_socket &&
3551                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3552                         tcp_new_space(sk);
3553         }
3554 }
3555
3556 static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3557 {
3558         tcp_push_pending_frames(sk, tp);
3559         tcp_check_space(sk);
3560 }
3561
3562 /*
3563  * Check if sending an ack is needed.
3564  */
3565 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3566 {
3567         struct tcp_sock *tp = tcp_sk(sk);
3568
3569             /* More than one full frame received... */
3570         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3571              /* ... and right edge of window advances far enough.
3572               * (tcp_recvmsg() will send ACK otherwise). Or...
3573               */
3574              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3575             /* We ACK each frame or... */
3576             tcp_in_quickack_mode(sk) ||
3577             /* We have out of order data. */
3578             (ofo_possible &&
3579              skb_peek(&tp->out_of_order_queue))) {
3580                 /* Then ack it now */
3581                 tcp_send_ack(sk);
3582         } else {
3583                 /* Else, send delayed ack. */
3584                 tcp_send_delayed_ack(sk);
3585         }
3586 }
3587
3588 static inline void tcp_ack_snd_check(struct sock *sk)
3589 {
3590         if (!inet_csk_ack_scheduled(sk)) {
3591                 /* We sent a data segment already. */
3592                 return;
3593         }
3594         __tcp_ack_snd_check(sk, 1);
3595 }
3596
3597 /*
3598  *      This routine is only called when we have urgent data
3599  *      signaled. Its the 'slow' part of tcp_urg. It could be
3600  *      moved inline now as tcp_urg is only called from one
3601  *      place. We handle URGent data wrong. We have to - as
3602  *      BSD still doesn't use the correction from RFC961.
3603  *      For 1003.1g we should support a new option TCP_STDURG to permit
3604  *      either form (or just set the sysctl tcp_stdurg).
3605  */
3606  
3607 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3608 {
3609         struct tcp_sock *tp = tcp_sk(sk);
3610         u32 ptr = ntohs(th->urg_ptr);
3611
3612         if (ptr && !sysctl_tcp_stdurg)
3613                 ptr--;
3614         ptr += ntohl(th->seq);
3615
3616         /* Ignore urgent data that we've already seen and read. */
3617         if (after(tp->copied_seq, ptr))
3618                 return;
3619
3620         /* Do not replay urg ptr.
3621          *
3622          * NOTE: interesting situation not covered by specs.
3623          * Misbehaving sender may send urg ptr, pointing to segment,
3624          * which we already have in ofo queue. We are not able to fetch
3625          * such data and will stay in TCP_URG_NOTYET until will be eaten
3626          * by recvmsg(). Seems, we are not obliged to handle such wicked
3627          * situations. But it is worth to think about possibility of some
3628          * DoSes using some hypothetical application level deadlock.
3629          */
3630         if (before(ptr, tp->rcv_nxt))
3631                 return;
3632
3633         /* Do we already have a newer (or duplicate) urgent pointer? */
3634         if (tp->urg_data && !after(ptr, tp->urg_seq))
3635                 return;
3636
3637         /* Tell the world about our new urgent pointer. */
3638         sk_send_sigurg(sk);
3639
3640         /* We may be adding urgent data when the last byte read was
3641          * urgent. To do this requires some care. We cannot just ignore
3642          * tp->copied_seq since we would read the last urgent byte again
3643          * as data, nor can we alter copied_seq until this data arrives
3644          * or we break the semantics of SIOCATMARK (and thus sockatmark())
3645          *
3646          * NOTE. Double Dutch. Rendering to plain English: author of comment
3647          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3648          * and expect that both A and B disappear from stream. This is _wrong_.
3649          * Though this happens in BSD with high probability, this is occasional.
3650          * Any application relying on this is buggy. Note also, that fix "works"
3651          * only in this artificial test. Insert some normal data between A and B and we will
3652          * decline of BSD again. Verdict: it is better to remove to trap
3653          * buggy users.
3654          */
3655         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3656             !sock_flag(sk, SOCK_URGINLINE) &&
3657             tp->copied_seq != tp->rcv_nxt) {
3658                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3659                 tp->copied_seq++;
3660                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3661                         __skb_unlink(skb, &sk->sk_receive_queue);
3662                         __kfree_skb(skb);
3663                 }
3664         }
3665
3666         tp->urg_data   = TCP_URG_NOTYET;
3667         tp->urg_seq    = ptr;
3668
3669         /* Disable header prediction. */
3670         tp->pred_flags = 0;
3671 }
3672
3673 /* This is the 'fast' part of urgent handling. */
3674 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3675 {
3676         struct tcp_sock *tp = tcp_sk(sk);
3677
3678         /* Check if we get a new urgent pointer - normally not. */
3679         if (th->urg)
3680                 tcp_check_urg(sk,th);
3681
3682         /* Do we wait for any urgent data? - normally not... */
3683         if (tp->urg_data == TCP_URG_NOTYET) {
3684                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3685                           th->syn;
3686
3687                 /* Is the urgent pointer pointing into this packet? */   
3688                 if (ptr < skb->len) {
3689                         u8 tmp;
3690                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3691                                 BUG();
3692                         tp->urg_data = TCP_URG_VALID | tmp;
3693                         if (!sock_flag(sk, SOCK_DEAD))
3694                                 sk->sk_data_ready(sk, 0);
3695                 }
3696         }
3697 }
3698
3699 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3700 {
3701         struct tcp_sock *tp = tcp_sk(sk);
3702         int chunk = skb->len - hlen;
3703         int err;
3704
3705         local_bh_enable();
3706         if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3707                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3708         else
3709                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3710                                                        tp->ucopy.iov);
3711
3712         if (!err) {
3713                 tp->ucopy.len -= chunk;
3714                 tp->copied_seq += chunk;
3715                 tcp_rcv_space_adjust(sk);
3716         }
3717
3718         local_bh_disable();
3719         return err;
3720 }
3721
3722 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3723 {
3724         int result;
3725
3726         if (sock_owned_by_user(sk)) {
3727                 local_bh_enable();
3728                 result = __tcp_checksum_complete(skb);
3729                 local_bh_disable();
3730         } else {
3731                 result = __tcp_checksum_complete(skb);
3732         }
3733         return result;
3734 }
3735
3736 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3737 {
3738         return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3739                 __tcp_checksum_complete_user(sk, skb);
3740 }
3741
3742 /*
3743  *      TCP receive function for the ESTABLISHED state. 
3744  *
3745  *      It is split into a fast path and a slow path. The fast path is 
3746  *      disabled when:
3747  *      - A zero window was announced from us - zero window probing
3748  *        is only handled properly in the slow path. 
3749  *      - Out of order segments arrived.
3750  *      - Urgent data is expected.
3751  *      - There is no buffer space left
3752  *      - Unexpected TCP flags/window values/header lengths are received
3753  *        (detected by checking the TCP header against pred_flags) 
3754  *      - Data is sent in both directions. Fast path only supports pure senders
3755  *        or pure receivers (this means either the sequence number or the ack
3756  *        value must stay constant)
3757  *      - Unexpected TCP option.
3758  *
3759  *      When these conditions are not satisfied it drops into a standard 
3760  *      receive procedure patterned after RFC793 to handle all cases.
3761  *      The first three cases are guaranteed by proper pred_flags setting,
3762  *      the rest is checked inline. Fast processing is turned on in 
3763  *      tcp_data_queue when everything is OK.
3764  */
3765 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3766                         struct tcphdr *th, unsigned len)
3767 {
3768         struct tcp_sock *tp = tcp_sk(sk);
3769
3770         /*
3771          *      Header prediction.
3772          *      The code loosely follows the one in the famous 
3773          *      "30 instruction TCP receive" Van Jacobson mail.
3774          *      
3775          *      Van's trick is to deposit buffers into socket queue 
3776          *      on a device interrupt, to call tcp_recv function
3777          *      on the receive process context and checksum and copy
3778          *      the buffer to user space. smart...
3779          *
3780          *      Our current scheme is not silly either but we take the 
3781          *      extra cost of the net_bh soft interrupt processing...
3782          *      We do checksum and copy also but from device to kernel.
3783          */
3784
3785         tp->rx_opt.saw_tstamp = 0;
3786
3787         /*      pred_flags is 0xS?10 << 16 + snd_wnd
3788          *      if header_prediction is to be made
3789          *      'S' will always be tp->tcp_header_len >> 2
3790          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3791          *  turn it off (when there are holes in the receive 
3792          *       space for instance)
3793          *      PSH flag is ignored.
3794          */
3795
3796         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3797                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3798                 int tcp_header_len = tp->tcp_header_len;
3799
3800                 /* Timestamp header prediction: tcp_header_len
3801                  * is automatically equal to th->doff*4 due to pred_flags
3802                  * match.
3803                  */
3804
3805                 /* Check timestamp */
3806                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3807                         __u32 *ptr = (__u32 *)(th + 1);
3808
3809                         /* No? Slow path! */
3810                         if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3811                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3812                                 goto slow_path;
3813
3814                         tp->rx_opt.saw_tstamp = 1;
3815                         ++ptr; 
3816                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3817                         ++ptr;
3818                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3819
3820                         /* If PAWS failed, check it more carefully in slow path */
3821                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
3822                                 goto slow_path;
3823
3824                         /* DO NOT update ts_recent here, if checksum fails
3825                          * and timestamp was corrupted part, it will result
3826                          * in a hung connection since we will drop all
3827                          * future packets due to the PAWS test.
3828                          */
3829                 }
3830
3831                 if (len <= tcp_header_len) {
3832                         /* Bulk data transfer: sender */
3833                         if (len == tcp_header_len) {
3834                                 /* Predicted packet is in window by definition.
3835                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3836                                  * Hence, check seq<=rcv_wup reduces to:
3837                                  */
3838                                 if (tcp_header_len ==
3839                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3840                                     tp->rcv_nxt == tp->rcv_wup)
3841                                         tcp_store_ts_recent(tp);
3842
3843                                 tcp_rcv_rtt_measure_ts(sk, skb);
3844
3845                                 /* We know that such packets are checksummed
3846                                  * on entry.
3847                                  */
3848                                 tcp_ack(sk, skb, 0);
3849                                 __kfree_skb(skb); 
3850                                 tcp_data_snd_check(sk, tp);
3851                                 return 0;
3852                         } else { /* Header too small */
3853                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3854                                 goto discard;
3855                         }
3856                 } else {
3857                         int eaten = 0;
3858
3859                         if (tp->ucopy.task == current &&
3860                             tp->copied_seq == tp->rcv_nxt &&
3861                             len - tcp_header_len <= tp->ucopy.len &&
3862                             sock_owned_by_user(sk)) {
3863                                 __set_current_state(TASK_RUNNING);
3864
3865                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
3866                                         /* Predicted packet is in window by definition.
3867                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3868                                          * Hence, check seq<=rcv_wup reduces to:
3869                                          */
3870                                         if (tcp_header_len ==
3871                                             (sizeof(struct tcphdr) +
3872                                              TCPOLEN_TSTAMP_ALIGNED) &&
3873                                             tp->rcv_nxt == tp->rcv_wup)
3874                                                 tcp_store_ts_recent(tp);
3875
3876                                         tcp_rcv_rtt_measure_ts(sk, skb);
3877
3878                                         __skb_pull(skb, tcp_header_len);
3879                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3880                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
3881                                         eaten = 1;
3882                                 }
3883                         }
3884                         if (!eaten) {
3885                                 if (tcp_checksum_complete_user(sk, skb))
3886                                         goto csum_error;
3887
3888                                 /* Predicted packet is in window by definition.
3889                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3890                                  * Hence, check seq<=rcv_wup reduces to:
3891                                  */
3892                                 if (tcp_header_len ==
3893                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3894                                     tp->rcv_nxt == tp->rcv_wup)
3895                                         tcp_store_ts_recent(tp);
3896
3897                                 tcp_rcv_rtt_measure_ts(sk, skb);
3898
3899                                 if ((int)skb->truesize > sk->sk_forward_alloc)
3900                                         goto step5;
3901
3902                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
3903
3904                                 /* Bulk data transfer: receiver */
3905                                 __skb_pull(skb,tcp_header_len);
3906                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3907                                 sk_stream_set_owner_r(skb, sk);
3908                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3909                         }
3910
3911                         tcp_event_data_recv(sk, tp, skb);
3912
3913                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
3914                                 /* Well, only one small jumplet in fast path... */
3915                                 tcp_ack(sk, skb, FLAG_DATA);
3916                                 tcp_data_snd_check(sk, tp);
3917                                 if (!inet_csk_ack_scheduled(sk))
3918                                         goto no_ack;
3919                         }
3920
3921                         __tcp_ack_snd_check(sk, 0);
3922 no_ack:
3923                         if (eaten)
3924                                 __kfree_skb(skb);
3925                         else
3926                                 sk->sk_data_ready(sk, 0);
3927                         return 0;
3928                 }
3929         }
3930
3931 slow_path:
3932         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
3933                 goto csum_error;
3934
3935         /*
3936          * RFC1323: H1. Apply PAWS check first.
3937          */
3938         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
3939             tcp_paws_discard(sk, skb)) {
3940                 if (!th->rst) {
3941                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
3942                         tcp_send_dupack(sk, skb);
3943                         goto discard;
3944                 }
3945                 /* Resets are accepted even if PAWS failed.
3946
3947                    ts_recent update must be made after we are sure
3948                    that the packet is in window.
3949                  */
3950         }
3951
3952         /*
3953          *      Standard slow path.
3954          */
3955
3956         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3957                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3958                  * (RST) segments are validated by checking their SEQ-fields."
3959                  * And page 69: "If an incoming segment is not acceptable,
3960                  * an acknowledgment should be sent in reply (unless the RST bit
3961                  * is set, if so drop the segment and return)".
3962                  */
3963                 if (!th->rst)
3964                         tcp_send_dupack(sk, skb);
3965                 goto discard;
3966         }
3967
3968         if(th->rst) {
3969                 tcp_reset(sk);
3970                 goto discard;
3971         }
3972
3973         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3974
3975         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3976                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3977                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
3978                 tcp_reset(sk);
3979                 return 1;
3980         }
3981
3982 step5:
3983         if(th->ack)
3984                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3985
3986         tcp_rcv_rtt_measure_ts(sk, skb);
3987
3988         /* Process urgent data. */
3989         tcp_urg(sk, skb, th);
3990
3991         /* step 7: process the segment text */
3992         tcp_data_queue(sk, skb);
3993
3994         tcp_data_snd_check(sk, tp);
3995         tcp_ack_snd_check(sk);
3996         return 0;
3997
3998 csum_error:
3999         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4000
4001 discard:
4002         __kfree_skb(skb);
4003         return 0;
4004 }
4005
4006 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4007                                          struct tcphdr *th, unsigned len)
4008 {
4009         struct tcp_sock *tp = tcp_sk(sk);
4010         struct inet_connection_sock *icsk = inet_csk(sk);
4011         int saved_clamp = tp->rx_opt.mss_clamp;
4012
4013         tcp_parse_options(skb, &tp->rx_opt, 0);
4014
4015         if (th->ack) {
4016                 /* rfc793:
4017                  * "If the state is SYN-SENT then
4018                  *    first check the ACK bit
4019                  *      If the ACK bit is set
4020                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4021                  *        a reset (unless the RST bit is set, if so drop
4022                  *        the segment and return)"
4023                  *
4024                  *  We do not send data with SYN, so that RFC-correct
4025                  *  test reduces to:
4026                  */
4027                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4028                         goto reset_and_undo;
4029
4030                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4031                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4032                              tcp_time_stamp)) {
4033                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4034                         goto reset_and_undo;
4035                 }
4036
4037                 /* Now ACK is acceptable.
4038                  *
4039                  * "If the RST bit is set
4040                  *    If the ACK was acceptable then signal the user "error:
4041                  *    connection reset", drop the segment, enter CLOSED state,
4042                  *    delete TCB, and return."
4043                  */
4044
4045                 if (th->rst) {
4046                         tcp_reset(sk);
4047                         goto discard;
4048                 }
4049
4050                 /* rfc793:
4051                  *   "fifth, if neither of the SYN or RST bits is set then
4052                  *    drop the segment and return."
4053                  *
4054                  *    See note below!
4055                  *                                        --ANK(990513)
4056                  */
4057                 if (!th->syn)
4058                         goto discard_and_undo;
4059
4060                 /* rfc793:
4061                  *   "If the SYN bit is on ...
4062                  *    are acceptable then ...
4063                  *    (our SYN has been ACKed), change the connection
4064                  *    state to ESTABLISHED..."
4065                  */
4066
4067                 TCP_ECN_rcv_synack(tp, th);
4068                 if (tp->ecn_flags&TCP_ECN_OK)
4069                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4070
4071                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4072                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4073
4074                 /* Ok.. it's good. Set up sequence numbers and
4075                  * move to established.
4076                  */
4077                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4078                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4079
4080                 /* RFC1323: The window in SYN & SYN/ACK segments is
4081                  * never scaled.
4082                  */
4083                 tp->snd_wnd = ntohs(th->window);
4084                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4085
4086                 if (!tp->rx_opt.wscale_ok) {
4087                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4088                         tp->window_clamp = min(tp->window_clamp, 65535U);
4089                 }
4090
4091                 if (tp->rx_opt.saw_tstamp) {
4092                         tp->rx_opt.tstamp_ok       = 1;
4093                         tp->tcp_header_len =
4094                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4095                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4096                         tcp_store_ts_recent(tp);
4097                 } else {
4098                         tp->tcp_header_len = sizeof(struct tcphdr);
4099                 }
4100
4101                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4102                         tp->rx_opt.sack_ok |= 2;
4103
4104                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4105                 tcp_initialize_rcv_mss(sk);
4106
4107                 /* Remember, tcp_poll() does not lock socket!
4108                  * Change state from SYN-SENT only after copied_seq
4109                  * is initialized. */
4110                 tp->copied_seq = tp->rcv_nxt;
4111                 mb();
4112                 tcp_set_state(sk, TCP_ESTABLISHED);
4113
4114                 /* Make sure socket is routed, for correct metrics.  */
4115                 icsk->icsk_af_ops->rebuild_header(sk);
4116
4117                 tcp_init_metrics(sk);
4118
4119                 tcp_init_congestion_control(sk);
4120
4121                 /* Prevent spurious tcp_cwnd_restart() on first data
4122                  * packet.
4123                  */
4124                 tp->lsndtime = tcp_time_stamp;
4125
4126                 tcp_init_buffer_space(sk);
4127
4128                 if (sock_flag(sk, SOCK_KEEPOPEN))
4129                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4130
4131                 if (!tp->rx_opt.snd_wscale)
4132                         __tcp_fast_path_on(tp, tp->snd_wnd);
4133                 else
4134                         tp->pred_flags = 0;
4135
4136                 if (!sock_flag(sk, SOCK_DEAD)) {
4137                         sk->sk_state_change(sk);
4138                         sk_wake_async(sk, 0, POLL_OUT);
4139                 }
4140
4141                 if (sk->sk_write_pending ||
4142                     icsk->icsk_accept_queue.rskq_defer_accept ||
4143                     icsk->icsk_ack.pingpong) {
4144                         /* Save one ACK. Data will be ready after
4145                          * several ticks, if write_pending is set.
4146                          *
4147                          * It may be deleted, but with this feature tcpdumps
4148                          * look so _wonderfully_ clever, that I was not able
4149                          * to stand against the temptation 8)     --ANK
4150                          */
4151                         inet_csk_schedule_ack(sk);
4152                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4153                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4154                         tcp_incr_quickack(sk);
4155                         tcp_enter_quickack_mode(sk);
4156                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4157                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4158
4159 discard:
4160                         __kfree_skb(skb);
4161                         return 0;
4162                 } else {
4163                         tcp_send_ack(sk);
4164                 }
4165                 return -1;
4166         }
4167
4168         /* No ACK in the segment */
4169
4170         if (th->rst) {
4171                 /* rfc793:
4172                  * "If the RST bit is set
4173                  *
4174                  *      Otherwise (no ACK) drop the segment and return."
4175                  */
4176
4177                 goto discard_and_undo;
4178         }
4179
4180         /* PAWS check. */
4181         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4182                 goto discard_and_undo;
4183
4184         if (th->syn) {
4185                 /* We see SYN without ACK. It is attempt of
4186                  * simultaneous connect with crossed SYNs.
4187                  * Particularly, it can be connect to self.
4188                  */
4189                 tcp_set_state(sk, TCP_SYN_RECV);
4190
4191                 if (tp->rx_opt.saw_tstamp) {
4192                         tp->rx_opt.tstamp_ok = 1;
4193                         tcp_store_ts_recent(tp);
4194                         tp->tcp_header_len =
4195                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4196                 } else {
4197                         tp->tcp_header_len = sizeof(struct tcphdr);
4198                 }
4199
4200                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4201                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4202
4203                 /* RFC1323: The window in SYN & SYN/ACK segments is
4204                  * never scaled.
4205                  */
4206                 tp->snd_wnd    = ntohs(th->window);
4207                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4208                 tp->max_window = tp->snd_wnd;
4209
4210                 TCP_ECN_rcv_syn(tp, th);
4211                 if (tp->ecn_flags&TCP_ECN_OK)
4212                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4213
4214                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4215                 tcp_initialize_rcv_mss(sk);
4216
4217
4218                 tcp_send_synack(sk);
4219 #if 0
4220                 /* Note, we could accept data and URG from this segment.
4221                  * There are no obstacles to make this.
4222                  *
4223                  * However, if we ignore data in ACKless segments sometimes,
4224                  * we have no reasons to accept it sometimes.
4225                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4226                  * is not flawless. So, discard packet for sanity.
4227                  * Uncomment this return to process the data.
4228                  */
4229                 return -1;
4230 #else
4231                 goto discard;
4232 #endif
4233         }
4234         /* "fifth, if neither of the SYN or RST bits is set then
4235          * drop the segment and return."
4236          */
4237
4238 discard_and_undo:
4239         tcp_clear_options(&tp->rx_opt);
4240         tp->rx_opt.mss_clamp = saved_clamp;
4241         goto discard;
4242
4243 reset_and_undo:
4244         tcp_clear_options(&tp->rx_opt);
4245         tp->rx_opt.mss_clamp = saved_clamp;
4246         return 1;
4247 }
4248
4249
4250 /*
4251  *      This function implements the receiving procedure of RFC 793 for
4252  *      all states except ESTABLISHED and TIME_WAIT. 
4253  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4254  *      address independent.
4255  */
4256         
4257 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4258                           struct tcphdr *th, unsigned len)
4259 {
4260         struct tcp_sock *tp = tcp_sk(sk);
4261         struct inet_connection_sock *icsk = inet_csk(sk);
4262         int queued = 0;
4263
4264         tp->rx_opt.saw_tstamp = 0;
4265
4266         switch (sk->sk_state) {
4267         case TCP_CLOSE:
4268                 goto discard;
4269
4270         case TCP_LISTEN:
4271                 if(th->ack)
4272                         return 1;
4273
4274                 if(th->rst)
4275                         goto discard;
4276
4277                 if(th->syn) {
4278                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4279                                 return 1;
4280
4281                         /* Now we have several options: In theory there is 
4282                          * nothing else in the frame. KA9Q has an option to 
4283                          * send data with the syn, BSD accepts data with the
4284                          * syn up to the [to be] advertised window and 
4285                          * Solaris 2.1 gives you a protocol error. For now 
4286                          * we just ignore it, that fits the spec precisely 
4287                          * and avoids incompatibilities. It would be nice in
4288                          * future to drop through and process the data.
4289                          *
4290                          * Now that TTCP is starting to be used we ought to 
4291                          * queue this data.
4292                          * But, this leaves one open to an easy denial of
4293                          * service attack, and SYN cookies can't defend
4294                          * against this problem. So, we drop the data
4295                          * in the interest of security over speed.
4296                          */
4297                         goto discard;
4298                 }
4299                 goto discard;
4300
4301         case TCP_SYN_SENT:
4302                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4303                 if (queued >= 0)
4304                         return queued;
4305
4306                 /* Do step6 onward by hand. */
4307                 tcp_urg(sk, skb, th);
4308                 __kfree_skb(skb);
4309                 tcp_data_snd_check(sk, tp);
4310                 return 0;
4311         }
4312
4313         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4314             tcp_paws_discard(sk, skb)) {
4315                 if (!th->rst) {
4316                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4317                         tcp_send_dupack(sk, skb);
4318                         goto discard;
4319                 }
4320                 /* Reset is accepted even if it did not pass PAWS. */
4321         }
4322
4323         /* step 1: check sequence number */
4324         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4325                 if (!th->rst)
4326                         tcp_send_dupack(sk, skb);
4327                 goto discard;
4328         }
4329
4330         /* step 2: check RST bit */
4331         if(th->rst) {
4332                 tcp_reset(sk);
4333                 goto discard;
4334         }
4335
4336         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4337
4338         /* step 3: check security and precedence [ignored] */
4339
4340         /*      step 4:
4341          *
4342          *      Check for a SYN in window.
4343          */
4344         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4345                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4346                 tcp_reset(sk);
4347                 return 1;
4348         }
4349
4350         /* step 5: check the ACK field */
4351         if (th->ack) {
4352                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4353
4354                 switch(sk->sk_state) {
4355                 case TCP_SYN_RECV:
4356                         if (acceptable) {
4357                                 tp->copied_seq = tp->rcv_nxt;
4358                                 mb();
4359                                 tcp_set_state(sk, TCP_ESTABLISHED);
4360                                 sk->sk_state_change(sk);
4361
4362                                 /* Note, that this wakeup is only for marginal
4363                                  * crossed SYN case. Passively open sockets
4364                                  * are not waked up, because sk->sk_sleep ==
4365                                  * NULL and sk->sk_socket == NULL.
4366                                  */
4367                                 if (sk->sk_socket) {
4368                                         sk_wake_async(sk,0,POLL_OUT);
4369                                 }
4370
4371                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4372                                 tp->snd_wnd = ntohs(th->window) <<
4373                                               tp->rx_opt.snd_wscale;
4374                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4375                                             TCP_SKB_CB(skb)->seq);
4376
4377                                 /* tcp_ack considers this ACK as duplicate
4378                                  * and does not calculate rtt.
4379                                  * Fix it at least with timestamps.
4380                                  */
4381                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4382                                     !tp->srtt)
4383                                         tcp_ack_saw_tstamp(sk, 0);
4384
4385                                 if (tp->rx_opt.tstamp_ok)
4386                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4387
4388                                 /* Make sure socket is routed, for
4389                                  * correct metrics.
4390                                  */
4391                                 icsk->icsk_af_ops->rebuild_header(sk);
4392
4393                                 tcp_init_metrics(sk);
4394
4395                                 tcp_init_congestion_control(sk);
4396
4397                                 /* Prevent spurious tcp_cwnd_restart() on
4398                                  * first data packet.
4399                                  */
4400                                 tp->lsndtime = tcp_time_stamp;
4401
4402                                 tcp_initialize_rcv_mss(sk);
4403                                 tcp_init_buffer_space(sk);
4404                                 tcp_fast_path_on(tp);
4405                         } else {
4406                                 return 1;
4407                         }
4408                         break;
4409
4410                 case TCP_FIN_WAIT1:
4411                         if (tp->snd_una == tp->write_seq) {
4412                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4413                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4414                                 dst_confirm(sk->sk_dst_cache);
4415
4416                                 if (!sock_flag(sk, SOCK_DEAD))
4417                                         /* Wake up lingering close() */
4418                                         sk->sk_state_change(sk);
4419                                 else {
4420                                         int tmo;
4421
4422                                         if (tp->linger2 < 0 ||
4423                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4424                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4425                                                 tcp_done(sk);
4426                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4427                                                 return 1;
4428                                         }
4429
4430                                         tmo = tcp_fin_time(sk);
4431                                         if (tmo > TCP_TIMEWAIT_LEN) {
4432                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4433                                         } else if (th->fin || sock_owned_by_user(sk)) {
4434                                                 /* Bad case. We could lose such FIN otherwise.
4435                                                  * It is not a big problem, but it looks confusing
4436                                                  * and not so rare event. We still can lose it now,
4437                                                  * if it spins in bh_lock_sock(), but it is really
4438                                                  * marginal case.
4439                                                  */
4440                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4441                                         } else {
4442                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4443                                                 goto discard;
4444                                         }
4445                                 }
4446                         }
4447                         break;
4448
4449                 case TCP_CLOSING:
4450                         if (tp->snd_una == tp->write_seq) {
4451                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4452                                 goto discard;
4453                         }
4454                         break;
4455
4456                 case TCP_LAST_ACK:
4457                         if (tp->snd_una == tp->write_seq) {
4458                                 tcp_update_metrics(sk);
4459                                 tcp_done(sk);
4460                                 goto discard;
4461                         }
4462                         break;
4463                 }
4464         } else
4465                 goto discard;
4466
4467         /* step 6: check the URG bit */
4468         tcp_urg(sk, skb, th);
4469
4470         /* step 7: process the segment text */
4471         switch (sk->sk_state) {
4472         case TCP_CLOSE_WAIT:
4473         case TCP_CLOSING:
4474         case TCP_LAST_ACK:
4475                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4476                         break;
4477         case TCP_FIN_WAIT1:
4478         case TCP_FIN_WAIT2:
4479                 /* RFC 793 says to queue data in these states,
4480                  * RFC 1122 says we MUST send a reset. 
4481                  * BSD 4.4 also does reset.
4482                  */
4483                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4484                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4485                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4486                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4487                                 tcp_reset(sk);
4488                                 return 1;
4489                         }
4490                 }
4491                 /* Fall through */
4492         case TCP_ESTABLISHED: 
4493                 tcp_data_queue(sk, skb);
4494                 queued = 1;
4495                 break;
4496         }
4497
4498         /* tcp_data could move socket to TIME-WAIT */
4499         if (sk->sk_state != TCP_CLOSE) {
4500                 tcp_data_snd_check(sk, tp);
4501                 tcp_ack_snd_check(sk);
4502         }
4503
4504         if (!queued) { 
4505 discard:
4506                 __kfree_skb(skb);
4507         }
4508         return 0;
4509 }
4510
4511 EXPORT_SYMBOL(sysctl_tcp_ecn);
4512 EXPORT_SYMBOL(sysctl_tcp_reordering);
4513 EXPORT_SYMBOL(sysctl_tcp_abc);
4514 EXPORT_SYMBOL(tcp_parse_options);
4515 EXPORT_SYMBOL(tcp_rcv_established);
4516 EXPORT_SYMBOL(tcp_rcv_state_process);
4517 EXPORT_SYMBOL(tcp_initialize_rcv_mss);