[ACPI] merge 3549 4320 4485 4588 4980 5483 5651 acpica asus fops pnpacpi branches...
[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                         int new_clamp = space;
461
462                         /* Receive space grows, normalize in order to
463                          * take into account packet headers and sk_buff
464                          * structure overhead.
465                          */
466                         space /= tp->advmss;
467                         if (!space)
468                                 space = 1;
469                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
470                                   16 + sizeof(struct sk_buff));
471                         while (tcp_win_from_space(rcvmem) < tp->advmss)
472                                 rcvmem += 128;
473                         space *= rcvmem;
474                         space = min(space, sysctl_tcp_rmem[2]);
475                         if (space > sk->sk_rcvbuf) {
476                                 sk->sk_rcvbuf = space;
477
478                                 /* Make the window clamp follow along.  */
479                                 tp->window_clamp = new_clamp;
480                         }
481                 }
482         }
483         
484 new_measure:
485         tp->rcvq_space.seq = tp->copied_seq;
486         tp->rcvq_space.time = tcp_time_stamp;
487 }
488
489 /* There is something which you must keep in mind when you analyze the
490  * behavior of the tp->ato delayed ack timeout interval.  When a
491  * connection starts up, we want to ack as quickly as possible.  The
492  * problem is that "good" TCP's do slow start at the beginning of data
493  * transmission.  The means that until we send the first few ACK's the
494  * sender will sit on his end and only queue most of his data, because
495  * he can only send snd_cwnd unacked packets at any given time.  For
496  * each ACK we send, he increments snd_cwnd and transmits more of his
497  * queue.  -DaveM
498  */
499 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
500 {
501         struct inet_connection_sock *icsk = inet_csk(sk);
502         u32 now;
503
504         inet_csk_schedule_ack(sk);
505
506         tcp_measure_rcv_mss(sk, skb);
507
508         tcp_rcv_rtt_measure(tp);
509         
510         now = tcp_time_stamp;
511
512         if (!icsk->icsk_ack.ato) {
513                 /* The _first_ data packet received, initialize
514                  * delayed ACK engine.
515                  */
516                 tcp_incr_quickack(sk);
517                 icsk->icsk_ack.ato = TCP_ATO_MIN;
518         } else {
519                 int m = now - icsk->icsk_ack.lrcvtime;
520
521                 if (m <= TCP_ATO_MIN/2) {
522                         /* The fastest case is the first. */
523                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
524                 } else if (m < icsk->icsk_ack.ato) {
525                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
526                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
527                                 icsk->icsk_ack.ato = icsk->icsk_rto;
528                 } else if (m > icsk->icsk_rto) {
529                         /* Too long gap. Apparently sender failed to
530                          * restart window, so that we send ACKs quickly.
531                          */
532                         tcp_incr_quickack(sk);
533                         sk_stream_mem_reclaim(sk);
534                 }
535         }
536         icsk->icsk_ack.lrcvtime = now;
537
538         TCP_ECN_check_ce(tp, skb);
539
540         if (skb->len >= 128)
541                 tcp_grow_window(sk, tp, skb);
542 }
543
544 /* Called to compute a smoothed rtt estimate. The data fed to this
545  * routine either comes from timestamps, or from segments that were
546  * known _not_ to have been retransmitted [see Karn/Partridge
547  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
548  * piece by Van Jacobson.
549  * NOTE: the next three routines used to be one big routine.
550  * To save cycles in the RFC 1323 implementation it was better to break
551  * it up into three procedures. -- erics
552  */
553 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
554 {
555         struct tcp_sock *tp = tcp_sk(sk);
556         long m = mrtt; /* RTT */
557
558         /*      The following amusing code comes from Jacobson's
559          *      article in SIGCOMM '88.  Note that rtt and mdev
560          *      are scaled versions of rtt and mean deviation.
561          *      This is designed to be as fast as possible 
562          *      m stands for "measurement".
563          *
564          *      On a 1990 paper the rto value is changed to:
565          *      RTO = rtt + 4 * mdev
566          *
567          * Funny. This algorithm seems to be very broken.
568          * These formulae increase RTO, when it should be decreased, increase
569          * too slowly, when it should be increased quickly, decrease too quickly
570          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
571          * does not matter how to _calculate_ it. Seems, it was trap
572          * that VJ failed to avoid. 8)
573          */
574         if(m == 0)
575                 m = 1;
576         if (tp->srtt != 0) {
577                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
578                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
579                 if (m < 0) {
580                         m = -m;         /* m is now abs(error) */
581                         m -= (tp->mdev >> 2);   /* similar update on mdev */
582                         /* This is similar to one of Eifel findings.
583                          * Eifel blocks mdev updates when rtt decreases.
584                          * This solution is a bit different: we use finer gain
585                          * for mdev in this case (alpha*beta).
586                          * Like Eifel it also prevents growth of rto,
587                          * but also it limits too fast rto decreases,
588                          * happening in pure Eifel.
589                          */
590                         if (m > 0)
591                                 m >>= 3;
592                 } else {
593                         m -= (tp->mdev >> 2);   /* similar update on mdev */
594                 }
595                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
596                 if (tp->mdev > tp->mdev_max) {
597                         tp->mdev_max = tp->mdev;
598                         if (tp->mdev_max > tp->rttvar)
599                                 tp->rttvar = tp->mdev_max;
600                 }
601                 if (after(tp->snd_una, tp->rtt_seq)) {
602                         if (tp->mdev_max < tp->rttvar)
603                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
604                         tp->rtt_seq = tp->snd_nxt;
605                         tp->mdev_max = TCP_RTO_MIN;
606                 }
607         } else {
608                 /* no previous measure. */
609                 tp->srtt = m<<3;        /* take the measured time to be rtt */
610                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
611                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
612                 tp->rtt_seq = tp->snd_nxt;
613         }
614 }
615
616 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
617  * routine referred to above.
618  */
619 static inline void tcp_set_rto(struct sock *sk)
620 {
621         const struct tcp_sock *tp = tcp_sk(sk);
622         /* Old crap is replaced with new one. 8)
623          *
624          * More seriously:
625          * 1. If rtt variance happened to be less 50msec, it is hallucination.
626          *    It cannot be less due to utterly erratic ACK generation made
627          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
628          *    to do with delayed acks, because at cwnd>2 true delack timeout
629          *    is invisible. Actually, Linux-2.4 also generates erratic
630          *    ACKs in some circumstances.
631          */
632         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
633
634         /* 2. Fixups made earlier cannot be right.
635          *    If we do not estimate RTO correctly without them,
636          *    all the algo is pure shit and should be replaced
637          *    with correct one. It is exactly, which we pretend to do.
638          */
639 }
640
641 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
642  * guarantees that rto is higher.
643  */
644 static inline void tcp_bound_rto(struct sock *sk)
645 {
646         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
647                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
648 }
649
650 /* Save metrics learned by this TCP session.
651    This function is called only, when TCP finishes successfully
652    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
653  */
654 void tcp_update_metrics(struct sock *sk)
655 {
656         struct tcp_sock *tp = tcp_sk(sk);
657         struct dst_entry *dst = __sk_dst_get(sk);
658
659         if (sysctl_tcp_nometrics_save)
660                 return;
661
662         dst_confirm(dst);
663
664         if (dst && (dst->flags&DST_HOST)) {
665                 const struct inet_connection_sock *icsk = inet_csk(sk);
666                 int m;
667
668                 if (icsk->icsk_backoff || !tp->srtt) {
669                         /* This session failed to estimate rtt. Why?
670                          * Probably, no packets returned in time.
671                          * Reset our results.
672                          */
673                         if (!(dst_metric_locked(dst, RTAX_RTT)))
674                                 dst->metrics[RTAX_RTT-1] = 0;
675                         return;
676                 }
677
678                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
679
680                 /* If newly calculated rtt larger than stored one,
681                  * store new one. Otherwise, use EWMA. Remember,
682                  * rtt overestimation is always better than underestimation.
683                  */
684                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
685                         if (m <= 0)
686                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
687                         else
688                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
689                 }
690
691                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
692                         if (m < 0)
693                                 m = -m;
694
695                         /* Scale deviation to rttvar fixed point */
696                         m >>= 1;
697                         if (m < tp->mdev)
698                                 m = tp->mdev;
699
700                         if (m >= dst_metric(dst, RTAX_RTTVAR))
701                                 dst->metrics[RTAX_RTTVAR-1] = m;
702                         else
703                                 dst->metrics[RTAX_RTTVAR-1] -=
704                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
705                 }
706
707                 if (tp->snd_ssthresh >= 0xFFFF) {
708                         /* Slow start still did not finish. */
709                         if (dst_metric(dst, RTAX_SSTHRESH) &&
710                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
711                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
712                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
713                         if (!dst_metric_locked(dst, RTAX_CWND) &&
714                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
715                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
716                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
717                            icsk->icsk_ca_state == TCP_CA_Open) {
718                         /* Cong. avoidance phase, cwnd is reliable. */
719                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
720                                 dst->metrics[RTAX_SSTHRESH-1] =
721                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
722                         if (!dst_metric_locked(dst, RTAX_CWND))
723                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
724                 } else {
725                         /* Else slow start did not finish, cwnd is non-sense,
726                            ssthresh may be also invalid.
727                          */
728                         if (!dst_metric_locked(dst, RTAX_CWND))
729                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
730                         if (dst->metrics[RTAX_SSTHRESH-1] &&
731                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
732                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
733                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
734                 }
735
736                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
737                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
738                             tp->reordering != sysctl_tcp_reordering)
739                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
740                 }
741         }
742 }
743
744 /* Numbers are taken from RFC2414.  */
745 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
746 {
747         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
748
749         if (!cwnd) {
750                 if (tp->mss_cache > 1460)
751                         cwnd = 2;
752                 else
753                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
754         }
755         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
756 }
757
758 /* Set slow start threshold and cwnd not falling to slow start */
759 void tcp_enter_cwr(struct sock *sk)
760 {
761         struct tcp_sock *tp = tcp_sk(sk);
762
763         tp->prior_ssthresh = 0;
764         tp->bytes_acked = 0;
765         if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
766                 tp->undo_marker = 0;
767                 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
768                 tp->snd_cwnd = min(tp->snd_cwnd,
769                                    tcp_packets_in_flight(tp) + 1U);
770                 tp->snd_cwnd_cnt = 0;
771                 tp->high_seq = tp->snd_nxt;
772                 tp->snd_cwnd_stamp = tcp_time_stamp;
773                 TCP_ECN_queue_cwr(tp);
774
775                 tcp_set_ca_state(sk, TCP_CA_CWR);
776         }
777 }
778
779 /* Initialize metrics on socket. */
780
781 static void tcp_init_metrics(struct sock *sk)
782 {
783         struct tcp_sock *tp = tcp_sk(sk);
784         struct dst_entry *dst = __sk_dst_get(sk);
785
786         if (dst == NULL)
787                 goto reset;
788
789         dst_confirm(dst);
790
791         if (dst_metric_locked(dst, RTAX_CWND))
792                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
793         if (dst_metric(dst, RTAX_SSTHRESH)) {
794                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
795                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
796                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
797         }
798         if (dst_metric(dst, RTAX_REORDERING) &&
799             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
800                 tp->rx_opt.sack_ok &= ~2;
801                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
802         }
803
804         if (dst_metric(dst, RTAX_RTT) == 0)
805                 goto reset;
806
807         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
808                 goto reset;
809
810         /* Initial rtt is determined from SYN,SYN-ACK.
811          * The segment is small and rtt may appear much
812          * less than real one. Use per-dst memory
813          * to make it more realistic.
814          *
815          * A bit of theory. RTT is time passed after "normal" sized packet
816          * is sent until it is ACKed. In normal circumstances sending small
817          * packets force peer to delay ACKs and calculation is correct too.
818          * The algorithm is adaptive and, provided we follow specs, it
819          * NEVER underestimate RTT. BUT! If peer tries to make some clever
820          * tricks sort of "quick acks" for time long enough to decrease RTT
821          * to low value, and then abruptly stops to do it and starts to delay
822          * ACKs, wait for troubles.
823          */
824         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
825                 tp->srtt = dst_metric(dst, RTAX_RTT);
826                 tp->rtt_seq = tp->snd_nxt;
827         }
828         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
829                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
830                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
831         }
832         tcp_set_rto(sk);
833         tcp_bound_rto(sk);
834         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
835                 goto reset;
836         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
837         tp->snd_cwnd_stamp = tcp_time_stamp;
838         return;
839
840 reset:
841         /* Play conservative. If timestamps are not
842          * supported, TCP will fail to recalculate correct
843          * rtt, if initial rto is too small. FORGET ALL AND RESET!
844          */
845         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
846                 tp->srtt = 0;
847                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
848                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
849         }
850 }
851
852 static void tcp_update_reordering(struct sock *sk, const int metric,
853                                   const int ts)
854 {
855         struct tcp_sock *tp = tcp_sk(sk);
856         if (metric > tp->reordering) {
857                 tp->reordering = min(TCP_MAX_REORDERING, metric);
858
859                 /* This exciting event is worth to be remembered. 8) */
860                 if (ts)
861                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
862                 else if (IsReno(tp))
863                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
864                 else if (IsFack(tp))
865                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
866                 else
867                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
868 #if FASTRETRANS_DEBUG > 1
869                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
870                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
871                        tp->reordering,
872                        tp->fackets_out,
873                        tp->sacked_out,
874                        tp->undo_marker ? tp->undo_retrans : 0);
875 #endif
876                 /* Disable FACK yet. */
877                 tp->rx_opt.sack_ok &= ~2;
878         }
879 }
880
881 /* This procedure tags the retransmission queue when SACKs arrive.
882  *
883  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
884  * Packets in queue with these bits set are counted in variables
885  * sacked_out, retrans_out and lost_out, correspondingly.
886  *
887  * Valid combinations are:
888  * Tag  InFlight        Description
889  * 0    1               - orig segment is in flight.
890  * S    0               - nothing flies, orig reached receiver.
891  * L    0               - nothing flies, orig lost by net.
892  * R    2               - both orig and retransmit are in flight.
893  * L|R  1               - orig is lost, retransmit is in flight.
894  * S|R  1               - orig reached receiver, retrans is still in flight.
895  * (L|S|R is logically valid, it could occur when L|R is sacked,
896  *  but it is equivalent to plain S and code short-curcuits it to S.
897  *  L|S is logically invalid, it would mean -1 packet in flight 8))
898  *
899  * These 6 states form finite state machine, controlled by the following events:
900  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
901  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
902  * 3. Loss detection event of one of three flavors:
903  *      A. Scoreboard estimator decided the packet is lost.
904  *         A'. Reno "three dupacks" marks head of queue lost.
905  *         A''. Its FACK modfication, head until snd.fack is lost.
906  *      B. SACK arrives sacking data transmitted after never retransmitted
907  *         hole was sent out.
908  *      C. SACK arrives sacking SND.NXT at the moment, when the
909  *         segment was retransmitted.
910  * 4. D-SACK added new rule: D-SACK changes any tag to S.
911  *
912  * It is pleasant to note, that state diagram turns out to be commutative,
913  * so that we are allowed not to be bothered by order of our actions,
914  * when multiple events arrive simultaneously. (see the function below).
915  *
916  * Reordering detection.
917  * --------------------
918  * Reordering metric is maximal distance, which a packet can be displaced
919  * in packet stream. With SACKs we can estimate it:
920  *
921  * 1. SACK fills old hole and the corresponding segment was not
922  *    ever retransmitted -> reordering. Alas, we cannot use it
923  *    when segment was retransmitted.
924  * 2. The last flaw is solved with D-SACK. D-SACK arrives
925  *    for retransmitted and already SACKed segment -> reordering..
926  * Both of these heuristics are not used in Loss state, when we cannot
927  * account for retransmits accurately.
928  */
929 static int
930 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
931 {
932         const struct inet_connection_sock *icsk = inet_csk(sk);
933         struct tcp_sock *tp = tcp_sk(sk);
934         unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
935         struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
936         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
937         int reord = tp->packets_out;
938         int prior_fackets;
939         u32 lost_retrans = 0;
940         int flag = 0;
941         int dup_sack = 0;
942         int i;
943
944         if (!tp->sacked_out)
945                 tp->fackets_out = 0;
946         prior_fackets = tp->fackets_out;
947
948         /* SACK fastpath:
949          * if the only SACK change is the increase of the end_seq of
950          * the first block then only apply that SACK block
951          * and use retrans queue hinting otherwise slowpath */
952         flag = 1;
953         for (i = 0; i< num_sacks; i++) {
954                 __u32 start_seq = ntohl(sp[i].start_seq);
955                 __u32 end_seq =  ntohl(sp[i].end_seq);
956
957                 if (i == 0){
958                         if (tp->recv_sack_cache[i].start_seq != start_seq)
959                                 flag = 0;
960                 } else {
961                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
962                             (tp->recv_sack_cache[i].end_seq != end_seq))
963                                 flag = 0;
964                 }
965                 tp->recv_sack_cache[i].start_seq = start_seq;
966                 tp->recv_sack_cache[i].end_seq = end_seq;
967
968                 /* Check for D-SACK. */
969                 if (i == 0) {
970                         u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
971
972                         if (before(start_seq, ack)) {
973                                 dup_sack = 1;
974                                 tp->rx_opt.sack_ok |= 4;
975                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
976                         } else if (num_sacks > 1 &&
977                                    !after(end_seq, ntohl(sp[1].end_seq)) &&
978                                    !before(start_seq, ntohl(sp[1].start_seq))) {
979                                 dup_sack = 1;
980                                 tp->rx_opt.sack_ok |= 4;
981                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
982                         }
983
984                         /* D-SACK for already forgotten data...
985                          * Do dumb counting. */
986                         if (dup_sack &&
987                             !after(end_seq, prior_snd_una) &&
988                             after(end_seq, tp->undo_marker))
989                                 tp->undo_retrans--;
990
991                         /* Eliminate too old ACKs, but take into
992                          * account more or less fresh ones, they can
993                          * contain valid SACK info.
994                          */
995                         if (before(ack, prior_snd_una - tp->max_window))
996                                 return 0;
997                 }
998         }
999
1000         if (flag)
1001                 num_sacks = 1;
1002         else {
1003                 int j;
1004                 tp->fastpath_skb_hint = NULL;
1005
1006                 /* order SACK blocks to allow in order walk of the retrans queue */
1007                 for (i = num_sacks-1; i > 0; i--) {
1008                         for (j = 0; j < i; j++){
1009                                 if (after(ntohl(sp[j].start_seq),
1010                                           ntohl(sp[j+1].start_seq))){
1011                                         sp[j].start_seq = htonl(tp->recv_sack_cache[j+1].start_seq);
1012                                         sp[j].end_seq = htonl(tp->recv_sack_cache[j+1].end_seq);
1013                                         sp[j+1].start_seq = htonl(tp->recv_sack_cache[j].start_seq);
1014                                         sp[j+1].end_seq = htonl(tp->recv_sack_cache[j].end_seq);
1015                                 }
1016
1017                         }
1018                 }
1019         }
1020
1021         /* clear flag as used for different purpose in following code */
1022         flag = 0;
1023
1024         for (i=0; i<num_sacks; i++, sp++) {
1025                 struct sk_buff *skb;
1026                 __u32 start_seq = ntohl(sp->start_seq);
1027                 __u32 end_seq = ntohl(sp->end_seq);
1028                 int fack_count;
1029
1030                 /* Use SACK fastpath hint if valid */
1031                 if (tp->fastpath_skb_hint) {
1032                         skb = tp->fastpath_skb_hint;
1033                         fack_count = tp->fastpath_cnt_hint;
1034                 } else {
1035                         skb = sk->sk_write_queue.next;
1036                         fack_count = 0;
1037                 }
1038
1039                 /* Event "B" in the comment above. */
1040                 if (after(end_seq, tp->high_seq))
1041                         flag |= FLAG_DATA_LOST;
1042
1043                 sk_stream_for_retrans_queue_from(skb, sk) {
1044                         int in_sack, pcount;
1045                         u8 sacked;
1046
1047                         tp->fastpath_skb_hint = skb;
1048                         tp->fastpath_cnt_hint = fack_count;
1049
1050                         /* The retransmission queue is always in order, so
1051                          * we can short-circuit the walk early.
1052                          */
1053                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1054                                 break;
1055
1056                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1057                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1058
1059                         pcount = tcp_skb_pcount(skb);
1060
1061                         if (pcount > 1 && !in_sack &&
1062                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1063                                 unsigned int pkt_len;
1064
1065                                 in_sack = !after(start_seq,
1066                                                  TCP_SKB_CB(skb)->seq);
1067
1068                                 if (!in_sack)
1069                                         pkt_len = (start_seq -
1070                                                    TCP_SKB_CB(skb)->seq);
1071                                 else
1072                                         pkt_len = (end_seq -
1073                                                    TCP_SKB_CB(skb)->seq);
1074                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->tso_size))
1075                                         break;
1076                                 pcount = tcp_skb_pcount(skb);
1077                         }
1078
1079                         fack_count += pcount;
1080
1081                         sacked = TCP_SKB_CB(skb)->sacked;
1082
1083                         /* Account D-SACK for retransmitted packet. */
1084                         if ((dup_sack && in_sack) &&
1085                             (sacked & TCPCB_RETRANS) &&
1086                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1087                                 tp->undo_retrans--;
1088
1089                         /* The frame is ACKed. */
1090                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1091                                 if (sacked&TCPCB_RETRANS) {
1092                                         if ((dup_sack && in_sack) &&
1093                                             (sacked&TCPCB_SACKED_ACKED))
1094                                                 reord = min(fack_count, reord);
1095                                 } else {
1096                                         /* If it was in a hole, we detected reordering. */
1097                                         if (fack_count < prior_fackets &&
1098                                             !(sacked&TCPCB_SACKED_ACKED))
1099                                                 reord = min(fack_count, reord);
1100                                 }
1101
1102                                 /* Nothing to do; acked frame is about to be dropped. */
1103                                 continue;
1104                         }
1105
1106                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1107                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1108                             (!lost_retrans || after(end_seq, lost_retrans)))
1109                                 lost_retrans = end_seq;
1110
1111                         if (!in_sack)
1112                                 continue;
1113
1114                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1115                                 if (sacked & TCPCB_SACKED_RETRANS) {
1116                                         /* If the segment is not tagged as lost,
1117                                          * we do not clear RETRANS, believing
1118                                          * that retransmission is still in flight.
1119                                          */
1120                                         if (sacked & TCPCB_LOST) {
1121                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1122                                                 tp->lost_out -= tcp_skb_pcount(skb);
1123                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1124
1125                                                 /* clear lost hint */
1126                                                 tp->retransmit_skb_hint = NULL;
1127                                         }
1128                                 } else {
1129                                         /* New sack for not retransmitted frame,
1130                                          * which was in hole. It is reordering.
1131                                          */
1132                                         if (!(sacked & TCPCB_RETRANS) &&
1133                                             fack_count < prior_fackets)
1134                                                 reord = min(fack_count, reord);
1135
1136                                         if (sacked & TCPCB_LOST) {
1137                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1138                                                 tp->lost_out -= tcp_skb_pcount(skb);
1139
1140                                                 /* clear lost hint */
1141                                                 tp->retransmit_skb_hint = NULL;
1142                                         }
1143                                 }
1144
1145                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1146                                 flag |= FLAG_DATA_SACKED;
1147                                 tp->sacked_out += tcp_skb_pcount(skb);
1148
1149                                 if (fack_count > tp->fackets_out)
1150                                         tp->fackets_out = fack_count;
1151                         } else {
1152                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1153                                         reord = min(fack_count, reord);
1154                         }
1155
1156                         /* D-SACK. We can detect redundant retransmission
1157                          * in S|R and plain R frames and clear it.
1158                          * undo_retrans is decreased above, L|R frames
1159                          * are accounted above as well.
1160                          */
1161                         if (dup_sack &&
1162                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1163                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1164                                 tp->retrans_out -= tcp_skb_pcount(skb);
1165                                 tp->retransmit_skb_hint = NULL;
1166                         }
1167                 }
1168         }
1169
1170         /* Check for lost retransmit. This superb idea is
1171          * borrowed from "ratehalving". Event "C".
1172          * Later note: FACK people cheated me again 8),
1173          * we have to account for reordering! Ugly,
1174          * but should help.
1175          */
1176         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1177                 struct sk_buff *skb;
1178
1179                 sk_stream_for_retrans_queue(skb, sk) {
1180                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1181                                 break;
1182                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1183                                 continue;
1184                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1185                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1186                             (IsFack(tp) ||
1187                              !before(lost_retrans,
1188                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1189                                      tp->mss_cache))) {
1190                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1191                                 tp->retrans_out -= tcp_skb_pcount(skb);
1192
1193                                 /* clear lost hint */
1194                                 tp->retransmit_skb_hint = NULL;
1195
1196                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1197                                         tp->lost_out += tcp_skb_pcount(skb);
1198                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1199                                         flag |= FLAG_DATA_SACKED;
1200                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1201                                 }
1202                         }
1203                 }
1204         }
1205
1206         tp->left_out = tp->sacked_out + tp->lost_out;
1207
1208         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
1209                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1210
1211 #if FASTRETRANS_DEBUG > 0
1212         BUG_TRAP((int)tp->sacked_out >= 0);
1213         BUG_TRAP((int)tp->lost_out >= 0);
1214         BUG_TRAP((int)tp->retrans_out >= 0);
1215         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1216 #endif
1217         return flag;
1218 }
1219
1220 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1221  * segments to see from the next ACKs whether any data was really missing.
1222  * If the RTO was spurious, new ACKs should arrive.
1223  */
1224 void tcp_enter_frto(struct sock *sk)
1225 {
1226         const struct inet_connection_sock *icsk = inet_csk(sk);
1227         struct tcp_sock *tp = tcp_sk(sk);
1228         struct sk_buff *skb;
1229
1230         tp->frto_counter = 1;
1231
1232         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1233             tp->snd_una == tp->high_seq ||
1234             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1235                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1236                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1237                 tcp_ca_event(sk, CA_EVENT_FRTO);
1238         }
1239
1240         /* Have to clear retransmission markers here to keep the bookkeeping
1241          * in shape, even though we are not yet in Loss state.
1242          * If something was really lost, it is eventually caught up
1243          * in tcp_enter_frto_loss.
1244          */
1245         tp->retrans_out = 0;
1246         tp->undo_marker = tp->snd_una;
1247         tp->undo_retrans = 0;
1248
1249         sk_stream_for_retrans_queue(skb, sk) {
1250                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1251         }
1252         tcp_sync_left_out(tp);
1253
1254         tcp_set_ca_state(sk, TCP_CA_Open);
1255         tp->frto_highmark = tp->snd_nxt;
1256 }
1257
1258 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1259  * which indicates that we should follow the traditional RTO recovery,
1260  * i.e. mark everything lost and do go-back-N retransmission.
1261  */
1262 static void tcp_enter_frto_loss(struct sock *sk)
1263 {
1264         struct tcp_sock *tp = tcp_sk(sk);
1265         struct sk_buff *skb;
1266         int cnt = 0;
1267
1268         tp->sacked_out = 0;
1269         tp->lost_out = 0;
1270         tp->fackets_out = 0;
1271
1272         sk_stream_for_retrans_queue(skb, sk) {
1273                 cnt += tcp_skb_pcount(skb);
1274                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1275                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1276
1277                         /* Do not mark those segments lost that were
1278                          * forward transmitted after RTO
1279                          */
1280                         if (!after(TCP_SKB_CB(skb)->end_seq,
1281                                    tp->frto_highmark)) {
1282                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1283                                 tp->lost_out += tcp_skb_pcount(skb);
1284                         }
1285                 } else {
1286                         tp->sacked_out += tcp_skb_pcount(skb);
1287                         tp->fackets_out = cnt;
1288                 }
1289         }
1290         tcp_sync_left_out(tp);
1291
1292         tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1293         tp->snd_cwnd_cnt = 0;
1294         tp->snd_cwnd_stamp = tcp_time_stamp;
1295         tp->undo_marker = 0;
1296         tp->frto_counter = 0;
1297
1298         tp->reordering = min_t(unsigned int, tp->reordering,
1299                                              sysctl_tcp_reordering);
1300         tcp_set_ca_state(sk, TCP_CA_Loss);
1301         tp->high_seq = tp->frto_highmark;
1302         TCP_ECN_queue_cwr(tp);
1303
1304         clear_all_retrans_hints(tp);
1305 }
1306
1307 void tcp_clear_retrans(struct tcp_sock *tp)
1308 {
1309         tp->left_out = 0;
1310         tp->retrans_out = 0;
1311
1312         tp->fackets_out = 0;
1313         tp->sacked_out = 0;
1314         tp->lost_out = 0;
1315
1316         tp->undo_marker = 0;
1317         tp->undo_retrans = 0;
1318 }
1319
1320 /* Enter Loss state. If "how" is not zero, forget all SACK information
1321  * and reset tags completely, otherwise preserve SACKs. If receiver
1322  * dropped its ofo queue, we will know this due to reneging detection.
1323  */
1324 void tcp_enter_loss(struct sock *sk, int how)
1325 {
1326         const struct inet_connection_sock *icsk = inet_csk(sk);
1327         struct tcp_sock *tp = tcp_sk(sk);
1328         struct sk_buff *skb;
1329         int cnt = 0;
1330
1331         /* Reduce ssthresh if it has not yet been made inside this window. */
1332         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1333             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1334                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1335                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1336                 tcp_ca_event(sk, CA_EVENT_LOSS);
1337         }
1338         tp->snd_cwnd       = 1;
1339         tp->snd_cwnd_cnt   = 0;
1340         tp->snd_cwnd_stamp = tcp_time_stamp;
1341
1342         tp->bytes_acked = 0;
1343         tcp_clear_retrans(tp);
1344
1345         /* Push undo marker, if it was plain RTO and nothing
1346          * was retransmitted. */
1347         if (!how)
1348                 tp->undo_marker = tp->snd_una;
1349
1350         sk_stream_for_retrans_queue(skb, sk) {
1351                 cnt += tcp_skb_pcount(skb);
1352                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1353                         tp->undo_marker = 0;
1354                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1355                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1356                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1357                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1358                         tp->lost_out += tcp_skb_pcount(skb);
1359                 } else {
1360                         tp->sacked_out += tcp_skb_pcount(skb);
1361                         tp->fackets_out = cnt;
1362                 }
1363         }
1364         tcp_sync_left_out(tp);
1365
1366         tp->reordering = min_t(unsigned int, tp->reordering,
1367                                              sysctl_tcp_reordering);
1368         tcp_set_ca_state(sk, TCP_CA_Loss);
1369         tp->high_seq = tp->snd_nxt;
1370         TCP_ECN_queue_cwr(tp);
1371
1372         clear_all_retrans_hints(tp);
1373 }
1374
1375 static int tcp_check_sack_reneging(struct sock *sk)
1376 {
1377         struct sk_buff *skb;
1378
1379         /* If ACK arrived pointing to a remembered SACK,
1380          * it means that our remembered SACKs do not reflect
1381          * real state of receiver i.e.
1382          * receiver _host_ is heavily congested (or buggy).
1383          * Do processing similar to RTO timeout.
1384          */
1385         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1386             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1387                 struct inet_connection_sock *icsk = inet_csk(sk);
1388                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1389
1390                 tcp_enter_loss(sk, 1);
1391                 icsk->icsk_retransmits++;
1392                 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1393                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1394                                           icsk->icsk_rto, TCP_RTO_MAX);
1395                 return 1;
1396         }
1397         return 0;
1398 }
1399
1400 static inline int tcp_fackets_out(struct tcp_sock *tp)
1401 {
1402         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1403 }
1404
1405 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1406 {
1407         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1408 }
1409
1410 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1411 {
1412         return tp->packets_out &&
1413                tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
1414 }
1415
1416 /* Linux NewReno/SACK/FACK/ECN state machine.
1417  * --------------------------------------
1418  *
1419  * "Open"       Normal state, no dubious events, fast path.
1420  * "Disorder"   In all the respects it is "Open",
1421  *              but requires a bit more attention. It is entered when
1422  *              we see some SACKs or dupacks. It is split of "Open"
1423  *              mainly to move some processing from fast path to slow one.
1424  * "CWR"        CWND was reduced due to some Congestion Notification event.
1425  *              It can be ECN, ICMP source quench, local device congestion.
1426  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1427  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1428  *
1429  * tcp_fastretrans_alert() is entered:
1430  * - each incoming ACK, if state is not "Open"
1431  * - when arrived ACK is unusual, namely:
1432  *      * SACK
1433  *      * Duplicate ACK.
1434  *      * ECN ECE.
1435  *
1436  * Counting packets in flight is pretty simple.
1437  *
1438  *      in_flight = packets_out - left_out + retrans_out
1439  *
1440  *      packets_out is SND.NXT-SND.UNA counted in packets.
1441  *
1442  *      retrans_out is number of retransmitted segments.
1443  *
1444  *      left_out is number of segments left network, but not ACKed yet.
1445  *
1446  *              left_out = sacked_out + lost_out
1447  *
1448  *     sacked_out: Packets, which arrived to receiver out of order
1449  *                 and hence not ACKed. With SACKs this number is simply
1450  *                 amount of SACKed data. Even without SACKs
1451  *                 it is easy to give pretty reliable estimate of this number,
1452  *                 counting duplicate ACKs.
1453  *
1454  *       lost_out: Packets lost by network. TCP has no explicit
1455  *                 "loss notification" feedback from network (for now).
1456  *                 It means that this number can be only _guessed_.
1457  *                 Actually, it is the heuristics to predict lossage that
1458  *                 distinguishes different algorithms.
1459  *
1460  *      F.e. after RTO, when all the queue is considered as lost,
1461  *      lost_out = packets_out and in_flight = retrans_out.
1462  *
1463  *              Essentially, we have now two algorithms counting
1464  *              lost packets.
1465  *
1466  *              FACK: It is the simplest heuristics. As soon as we decided
1467  *              that something is lost, we decide that _all_ not SACKed
1468  *              packets until the most forward SACK are lost. I.e.
1469  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1470  *              It is absolutely correct estimate, if network does not reorder
1471  *              packets. And it loses any connection to reality when reordering
1472  *              takes place. We use FACK by default until reordering
1473  *              is suspected on the path to this destination.
1474  *
1475  *              NewReno: when Recovery is entered, we assume that one segment
1476  *              is lost (classic Reno). While we are in Recovery and
1477  *              a partial ACK arrives, we assume that one more packet
1478  *              is lost (NewReno). This heuristics are the same in NewReno
1479  *              and SACK.
1480  *
1481  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1482  *  deflation etc. CWND is real congestion window, never inflated, changes
1483  *  only according to classic VJ rules.
1484  *
1485  * Really tricky (and requiring careful tuning) part of algorithm
1486  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1487  * The first determines the moment _when_ we should reduce CWND and,
1488  * hence, slow down forward transmission. In fact, it determines the moment
1489  * when we decide that hole is caused by loss, rather than by a reorder.
1490  *
1491  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1492  * holes, caused by lost packets.
1493  *
1494  * And the most logically complicated part of algorithm is undo
1495  * heuristics. We detect false retransmits due to both too early
1496  * fast retransmit (reordering) and underestimated RTO, analyzing
1497  * timestamps and D-SACKs. When we detect that some segments were
1498  * retransmitted by mistake and CWND reduction was wrong, we undo
1499  * window reduction and abort recovery phase. This logic is hidden
1500  * inside several functions named tcp_try_undo_<something>.
1501  */
1502
1503 /* This function decides, when we should leave Disordered state
1504  * and enter Recovery phase, reducing congestion window.
1505  *
1506  * Main question: may we further continue forward transmission
1507  * with the same cwnd?
1508  */
1509 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1510 {
1511         __u32 packets_out;
1512
1513         /* Trick#1: The loss is proven. */
1514         if (tp->lost_out)
1515                 return 1;
1516
1517         /* Not-A-Trick#2 : Classic rule... */
1518         if (tcp_fackets_out(tp) > tp->reordering)
1519                 return 1;
1520
1521         /* Trick#3 : when we use RFC2988 timer restart, fast
1522          * retransmit can be triggered by timeout of queue head.
1523          */
1524         if (tcp_head_timedout(sk, tp))
1525                 return 1;
1526
1527         /* Trick#4: It is still not OK... But will it be useful to delay
1528          * recovery more?
1529          */
1530         packets_out = tp->packets_out;
1531         if (packets_out <= tp->reordering &&
1532             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1533             !tcp_may_send_now(sk, tp)) {
1534                 /* We have nothing to send. This connection is limited
1535                  * either by receiver window or by application.
1536                  */
1537                 return 1;
1538         }
1539
1540         return 0;
1541 }
1542
1543 /* If we receive more dupacks than we expected counting segments
1544  * in assumption of absent reordering, interpret this as reordering.
1545  * The only another reason could be bug in receiver TCP.
1546  */
1547 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1548 {
1549         struct tcp_sock *tp = tcp_sk(sk);
1550         u32 holes;
1551
1552         holes = max(tp->lost_out, 1U);
1553         holes = min(holes, tp->packets_out);
1554
1555         if ((tp->sacked_out + holes) > tp->packets_out) {
1556                 tp->sacked_out = tp->packets_out - holes;
1557                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1558         }
1559 }
1560
1561 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1562
1563 static void tcp_add_reno_sack(struct sock *sk)
1564 {
1565         struct tcp_sock *tp = tcp_sk(sk);
1566         tp->sacked_out++;
1567         tcp_check_reno_reordering(sk, 0);
1568         tcp_sync_left_out(tp);
1569 }
1570
1571 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1572
1573 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1574 {
1575         if (acked > 0) {
1576                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1577                 if (acked-1 >= tp->sacked_out)
1578                         tp->sacked_out = 0;
1579                 else
1580                         tp->sacked_out -= acked-1;
1581         }
1582         tcp_check_reno_reordering(sk, acked);
1583         tcp_sync_left_out(tp);
1584 }
1585
1586 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1587 {
1588         tp->sacked_out = 0;
1589         tp->left_out = tp->lost_out;
1590 }
1591
1592 /* Mark head of queue up as lost. */
1593 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1594                                int packets, u32 high_seq)
1595 {
1596         struct sk_buff *skb;
1597         int cnt;
1598
1599         BUG_TRAP(packets <= tp->packets_out);
1600         if (tp->lost_skb_hint) {
1601                 skb = tp->lost_skb_hint;
1602                 cnt = tp->lost_cnt_hint;
1603         } else {
1604                 skb = sk->sk_write_queue.next;
1605                 cnt = 0;
1606         }
1607
1608         sk_stream_for_retrans_queue_from(skb, sk) {
1609                 /* TODO: do this better */
1610                 /* this is not the most efficient way to do this... */
1611                 tp->lost_skb_hint = skb;
1612                 tp->lost_cnt_hint = cnt;
1613                 cnt += tcp_skb_pcount(skb);
1614                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1615                         break;
1616                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1617                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1618                         tp->lost_out += tcp_skb_pcount(skb);
1619
1620                         /* clear xmit_retransmit_queue hints
1621                          *  if this is beyond hint */
1622                         if(tp->retransmit_skb_hint != NULL &&
1623                            before(TCP_SKB_CB(skb)->seq,
1624                                   TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) {
1625
1626                                 tp->retransmit_skb_hint = NULL;
1627                         }
1628                 }
1629         }
1630         tcp_sync_left_out(tp);
1631 }
1632
1633 /* Account newly detected lost packet(s) */
1634
1635 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1636 {
1637         if (IsFack(tp)) {
1638                 int lost = tp->fackets_out - tp->reordering;
1639                 if (lost <= 0)
1640                         lost = 1;
1641                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1642         } else {
1643                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1644         }
1645
1646         /* New heuristics: it is possible only after we switched
1647          * to restart timer each time when something is ACKed.
1648          * Hence, we can detect timed out packets during fast
1649          * retransmit without falling to slow start.
1650          */
1651         if (tcp_head_timedout(sk, tp)) {
1652                 struct sk_buff *skb;
1653
1654                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1655                         : sk->sk_write_queue.next;
1656
1657                 sk_stream_for_retrans_queue_from(skb, sk) {
1658                         if (!tcp_skb_timedout(sk, skb))
1659                                 break;
1660
1661                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1662                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1663                                 tp->lost_out += tcp_skb_pcount(skb);
1664
1665                                 /* clear xmit_retrans hint */
1666                                 if (tp->retransmit_skb_hint &&
1667                                     before(TCP_SKB_CB(skb)->seq,
1668                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1669
1670                                         tp->retransmit_skb_hint = NULL;
1671                         }
1672                 }
1673
1674                 tp->scoreboard_skb_hint = skb;
1675
1676                 tcp_sync_left_out(tp);
1677         }
1678 }
1679
1680 /* CWND moderation, preventing bursts due to too big ACKs
1681  * in dubious situations.
1682  */
1683 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1684 {
1685         tp->snd_cwnd = min(tp->snd_cwnd,
1686                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1687         tp->snd_cwnd_stamp = tcp_time_stamp;
1688 }
1689
1690 /* Decrease cwnd each second ack. */
1691 static void tcp_cwnd_down(struct sock *sk)
1692 {
1693         const struct inet_connection_sock *icsk = inet_csk(sk);
1694         struct tcp_sock *tp = tcp_sk(sk);
1695         int decr = tp->snd_cwnd_cnt + 1;
1696
1697         tp->snd_cwnd_cnt = decr&1;
1698         decr >>= 1;
1699
1700         if (decr && tp->snd_cwnd > icsk->icsk_ca_ops->min_cwnd(sk))
1701                 tp->snd_cwnd -= decr;
1702
1703         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1704         tp->snd_cwnd_stamp = tcp_time_stamp;
1705 }
1706
1707 /* Nothing was retransmitted or returned timestamp is less
1708  * than timestamp of the first retransmission.
1709  */
1710 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1711 {
1712         return !tp->retrans_stamp ||
1713                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1714                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1715 }
1716
1717 /* Undo procedures. */
1718
1719 #if FASTRETRANS_DEBUG > 1
1720 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1721 {
1722         struct inet_sock *inet = inet_sk(sk);
1723         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1724                msg,
1725                NIPQUAD(inet->daddr), ntohs(inet->dport),
1726                tp->snd_cwnd, tp->left_out,
1727                tp->snd_ssthresh, tp->prior_ssthresh,
1728                tp->packets_out);
1729 }
1730 #else
1731 #define DBGUNDO(x...) do { } while (0)
1732 #endif
1733
1734 static void tcp_undo_cwr(struct sock *sk, const int undo)
1735 {
1736         struct tcp_sock *tp = tcp_sk(sk);
1737
1738         if (tp->prior_ssthresh) {
1739                 const struct inet_connection_sock *icsk = inet_csk(sk);
1740
1741                 if (icsk->icsk_ca_ops->undo_cwnd)
1742                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1743                 else
1744                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1745
1746                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1747                         tp->snd_ssthresh = tp->prior_ssthresh;
1748                         TCP_ECN_withdraw_cwr(tp);
1749                 }
1750         } else {
1751                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1752         }
1753         tcp_moderate_cwnd(tp);
1754         tp->snd_cwnd_stamp = tcp_time_stamp;
1755
1756         /* There is something screwy going on with the retrans hints after
1757            an undo */
1758         clear_all_retrans_hints(tp);
1759 }
1760
1761 static inline int tcp_may_undo(struct tcp_sock *tp)
1762 {
1763         return tp->undo_marker &&
1764                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1765 }
1766
1767 /* People celebrate: "We love our President!" */
1768 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1769 {
1770         if (tcp_may_undo(tp)) {
1771                 /* Happy end! We did not retransmit anything
1772                  * or our original transmission succeeded.
1773                  */
1774                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1775                 tcp_undo_cwr(sk, 1);
1776                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1777                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1778                 else
1779                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1780                 tp->undo_marker = 0;
1781         }
1782         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1783                 /* Hold old state until something *above* high_seq
1784                  * is ACKed. For Reno it is MUST to prevent false
1785                  * fast retransmits (RFC2582). SACK TCP is safe. */
1786                 tcp_moderate_cwnd(tp);
1787                 return 1;
1788         }
1789         tcp_set_ca_state(sk, TCP_CA_Open);
1790         return 0;
1791 }
1792
1793 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1794 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1795 {
1796         if (tp->undo_marker && !tp->undo_retrans) {
1797                 DBGUNDO(sk, tp, "D-SACK");
1798                 tcp_undo_cwr(sk, 1);
1799                 tp->undo_marker = 0;
1800                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1801         }
1802 }
1803
1804 /* Undo during fast recovery after partial ACK. */
1805
1806 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1807                                 int acked)
1808 {
1809         /* Partial ACK arrived. Force Hoe's retransmit. */
1810         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1811
1812         if (tcp_may_undo(tp)) {
1813                 /* Plain luck! Hole if filled with delayed
1814                  * packet, rather than with a retransmit.
1815                  */
1816                 if (tp->retrans_out == 0)
1817                         tp->retrans_stamp = 0;
1818
1819                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1820
1821                 DBGUNDO(sk, tp, "Hoe");
1822                 tcp_undo_cwr(sk, 0);
1823                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1824
1825                 /* So... Do not make Hoe's retransmit yet.
1826                  * If the first packet was delayed, the rest
1827                  * ones are most probably delayed as well.
1828                  */
1829                 failed = 0;
1830         }
1831         return failed;
1832 }
1833
1834 /* Undo during loss recovery after partial ACK. */
1835 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1836 {
1837         if (tcp_may_undo(tp)) {
1838                 struct sk_buff *skb;
1839                 sk_stream_for_retrans_queue(skb, sk) {
1840                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1841                 }
1842
1843                 clear_all_retrans_hints(tp);
1844
1845                 DBGUNDO(sk, tp, "partial loss");
1846                 tp->lost_out = 0;
1847                 tp->left_out = tp->sacked_out;
1848                 tcp_undo_cwr(sk, 1);
1849                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1850                 inet_csk(sk)->icsk_retransmits = 0;
1851                 tp->undo_marker = 0;
1852                 if (!IsReno(tp))
1853                         tcp_set_ca_state(sk, TCP_CA_Open);
1854                 return 1;
1855         }
1856         return 0;
1857 }
1858
1859 static inline void tcp_complete_cwr(struct sock *sk)
1860 {
1861         struct tcp_sock *tp = tcp_sk(sk);
1862         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1863         tp->snd_cwnd_stamp = tcp_time_stamp;
1864         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
1865 }
1866
1867 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1868 {
1869         tp->left_out = tp->sacked_out;
1870
1871         if (tp->retrans_out == 0)
1872                 tp->retrans_stamp = 0;
1873
1874         if (flag&FLAG_ECE)
1875                 tcp_enter_cwr(sk);
1876
1877         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
1878                 int state = TCP_CA_Open;
1879
1880                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1881                         state = TCP_CA_Disorder;
1882
1883                 if (inet_csk(sk)->icsk_ca_state != state) {
1884                         tcp_set_ca_state(sk, state);
1885                         tp->high_seq = tp->snd_nxt;
1886                 }
1887                 tcp_moderate_cwnd(tp);
1888         } else {
1889                 tcp_cwnd_down(sk);
1890         }
1891 }
1892
1893 /* Process an event, which can update packets-in-flight not trivially.
1894  * Main goal of this function is to calculate new estimate for left_out,
1895  * taking into account both packets sitting in receiver's buffer and
1896  * packets lost by network.
1897  *
1898  * Besides that it does CWND reduction, when packet loss is detected
1899  * and changes state of machine.
1900  *
1901  * It does _not_ decide what to send, it is made in function
1902  * tcp_xmit_retransmit_queue().
1903  */
1904 static void
1905 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1906                       int prior_packets, int flag)
1907 {
1908         struct inet_connection_sock *icsk = inet_csk(sk);
1909         struct tcp_sock *tp = tcp_sk(sk);
1910         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1911
1912         /* Some technical things:
1913          * 1. Reno does not count dupacks (sacked_out) automatically. */
1914         if (!tp->packets_out)
1915                 tp->sacked_out = 0;
1916         /* 2. SACK counts snd_fack in packets inaccurately. */
1917         if (tp->sacked_out == 0)
1918                 tp->fackets_out = 0;
1919
1920         /* Now state machine starts.
1921          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1922         if (flag&FLAG_ECE)
1923                 tp->prior_ssthresh = 0;
1924
1925         /* B. In all the states check for reneging SACKs. */
1926         if (tp->sacked_out && tcp_check_sack_reneging(sk))
1927                 return;
1928
1929         /* C. Process data loss notification, provided it is valid. */
1930         if ((flag&FLAG_DATA_LOST) &&
1931             before(tp->snd_una, tp->high_seq) &&
1932             icsk->icsk_ca_state != TCP_CA_Open &&
1933             tp->fackets_out > tp->reordering) {
1934                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1935                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1936         }
1937
1938         /* D. Synchronize left_out to current state. */
1939         tcp_sync_left_out(tp);
1940
1941         /* E. Check state exit conditions. State can be terminated
1942          *    when high_seq is ACKed. */
1943         if (icsk->icsk_ca_state == TCP_CA_Open) {
1944                 if (!sysctl_tcp_frto)
1945                         BUG_TRAP(tp->retrans_out == 0);
1946                 tp->retrans_stamp = 0;
1947         } else if (!before(tp->snd_una, tp->high_seq)) {
1948                 switch (icsk->icsk_ca_state) {
1949                 case TCP_CA_Loss:
1950                         icsk->icsk_retransmits = 0;
1951                         if (tcp_try_undo_recovery(sk, tp))
1952                                 return;
1953                         break;
1954
1955                 case TCP_CA_CWR:
1956                         /* CWR is to be held something *above* high_seq
1957                          * is ACKed for CWR bit to reach receiver. */
1958                         if (tp->snd_una != tp->high_seq) {
1959                                 tcp_complete_cwr(sk);
1960                                 tcp_set_ca_state(sk, TCP_CA_Open);
1961                         }
1962                         break;
1963
1964                 case TCP_CA_Disorder:
1965                         tcp_try_undo_dsack(sk, tp);
1966                         if (!tp->undo_marker ||
1967                             /* For SACK case do not Open to allow to undo
1968                              * catching for all duplicate ACKs. */
1969                             IsReno(tp) || tp->snd_una != tp->high_seq) {
1970                                 tp->undo_marker = 0;
1971                                 tcp_set_ca_state(sk, TCP_CA_Open);
1972                         }
1973                         break;
1974
1975                 case TCP_CA_Recovery:
1976                         if (IsReno(tp))
1977                                 tcp_reset_reno_sack(tp);
1978                         if (tcp_try_undo_recovery(sk, tp))
1979                                 return;
1980                         tcp_complete_cwr(sk);
1981                         break;
1982                 }
1983         }
1984
1985         /* F. Process state. */
1986         switch (icsk->icsk_ca_state) {
1987         case TCP_CA_Recovery:
1988                 if (prior_snd_una == tp->snd_una) {
1989                         if (IsReno(tp) && is_dupack)
1990                                 tcp_add_reno_sack(sk);
1991                 } else {
1992                         int acked = prior_packets - tp->packets_out;
1993                         if (IsReno(tp))
1994                                 tcp_remove_reno_sacks(sk, tp, acked);
1995                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
1996                 }
1997                 break;
1998         case TCP_CA_Loss:
1999                 if (flag&FLAG_DATA_ACKED)
2000                         icsk->icsk_retransmits = 0;
2001                 if (!tcp_try_undo_loss(sk, tp)) {
2002                         tcp_moderate_cwnd(tp);
2003                         tcp_xmit_retransmit_queue(sk);
2004                         return;
2005                 }
2006                 if (icsk->icsk_ca_state != TCP_CA_Open)
2007                         return;
2008                 /* Loss is undone; fall through to processing in Open state. */
2009         default:
2010                 if (IsReno(tp)) {
2011                         if (tp->snd_una != prior_snd_una)
2012                                 tcp_reset_reno_sack(tp);
2013                         if (is_dupack)
2014                                 tcp_add_reno_sack(sk);
2015                 }
2016
2017                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2018                         tcp_try_undo_dsack(sk, tp);
2019
2020                 if (!tcp_time_to_recover(sk, tp)) {
2021                         tcp_try_to_open(sk, tp, flag);
2022                         return;
2023                 }
2024
2025                 /* Otherwise enter Recovery state */
2026
2027                 if (IsReno(tp))
2028                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2029                 else
2030                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2031
2032                 tp->high_seq = tp->snd_nxt;
2033                 tp->prior_ssthresh = 0;
2034                 tp->undo_marker = tp->snd_una;
2035                 tp->undo_retrans = tp->retrans_out;
2036
2037                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2038                         if (!(flag&FLAG_ECE))
2039                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2040                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2041                         TCP_ECN_queue_cwr(tp);
2042                 }
2043
2044                 tp->bytes_acked = 0;
2045                 tp->snd_cwnd_cnt = 0;
2046                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2047         }
2048
2049         if (is_dupack || tcp_head_timedout(sk, tp))
2050                 tcp_update_scoreboard(sk, tp);
2051         tcp_cwnd_down(sk);
2052         tcp_xmit_retransmit_queue(sk);
2053 }
2054
2055 /* Read draft-ietf-tcplw-high-performance before mucking
2056  * with this code. (Supersedes RFC1323)
2057  */
2058 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2059 {
2060         /* RTTM Rule: A TSecr value received in a segment is used to
2061          * update the averaged RTT measurement only if the segment
2062          * acknowledges some new data, i.e., only if it advances the
2063          * left edge of the send window.
2064          *
2065          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2066          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2067          *
2068          * Changed: reset backoff as soon as we see the first valid sample.
2069          * If we do not, we get strongly overestimated rto. With timestamps
2070          * samples are accepted even from very old segments: f.e., when rtt=1
2071          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2072          * answer arrives rto becomes 120 seconds! If at least one of segments
2073          * in window is lost... Voila.                          --ANK (010210)
2074          */
2075         struct tcp_sock *tp = tcp_sk(sk);
2076         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2077         tcp_rtt_estimator(sk, seq_rtt);
2078         tcp_set_rto(sk);
2079         inet_csk(sk)->icsk_backoff = 0;
2080         tcp_bound_rto(sk);
2081 }
2082
2083 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2084 {
2085         /* We don't have a timestamp. Can only use
2086          * packets that are not retransmitted to determine
2087          * rtt estimates. Also, we must not reset the
2088          * backoff for rto until we get a non-retransmitted
2089          * packet. This allows us to deal with a situation
2090          * where the network delay has increased suddenly.
2091          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2092          */
2093
2094         if (flag & FLAG_RETRANS_DATA_ACKED)
2095                 return;
2096
2097         tcp_rtt_estimator(sk, seq_rtt);
2098         tcp_set_rto(sk);
2099         inet_csk(sk)->icsk_backoff = 0;
2100         tcp_bound_rto(sk);
2101 }
2102
2103 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2104                                       const s32 seq_rtt)
2105 {
2106         const struct tcp_sock *tp = tcp_sk(sk);
2107         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2108         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2109                 tcp_ack_saw_tstamp(sk, flag);
2110         else if (seq_rtt >= 0)
2111                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2112 }
2113
2114 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2115                            u32 in_flight, int good)
2116 {
2117         const struct inet_connection_sock *icsk = inet_csk(sk);
2118         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2119         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2120 }
2121
2122 /* Restart timer after forward progress on connection.
2123  * RFC2988 recommends to restart timer to now+rto.
2124  */
2125
2126 static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2127 {
2128         if (!tp->packets_out) {
2129                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2130         } else {
2131                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2132         }
2133 }
2134
2135 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2136                          __u32 now, __s32 *seq_rtt)
2137 {
2138         struct tcp_sock *tp = tcp_sk(sk);
2139         struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2140         __u32 seq = tp->snd_una;
2141         __u32 packets_acked;
2142         int acked = 0;
2143
2144         /* If we get here, the whole TSO packet has not been
2145          * acked.
2146          */
2147         BUG_ON(!after(scb->end_seq, seq));
2148
2149         packets_acked = tcp_skb_pcount(skb);
2150         if (tcp_trim_head(sk, skb, seq - scb->seq))
2151                 return 0;
2152         packets_acked -= tcp_skb_pcount(skb);
2153
2154         if (packets_acked) {
2155                 __u8 sacked = scb->sacked;
2156
2157                 acked |= FLAG_DATA_ACKED;
2158                 if (sacked) {
2159                         if (sacked & TCPCB_RETRANS) {
2160                                 if (sacked & TCPCB_SACKED_RETRANS)
2161                                         tp->retrans_out -= packets_acked;
2162                                 acked |= FLAG_RETRANS_DATA_ACKED;
2163                                 *seq_rtt = -1;
2164                         } else if (*seq_rtt < 0)
2165                                 *seq_rtt = now - scb->when;
2166                         if (sacked & TCPCB_SACKED_ACKED)
2167                                 tp->sacked_out -= packets_acked;
2168                         if (sacked & TCPCB_LOST)
2169                                 tp->lost_out -= packets_acked;
2170                         if (sacked & TCPCB_URG) {
2171                                 if (tp->urg_mode &&
2172                                     !before(seq, tp->snd_up))
2173                                         tp->urg_mode = 0;
2174                         }
2175                 } else if (*seq_rtt < 0)
2176                         *seq_rtt = now - scb->when;
2177
2178                 if (tp->fackets_out) {
2179                         __u32 dval = min(tp->fackets_out, packets_acked);
2180                         tp->fackets_out -= dval;
2181                 }
2182                 tp->packets_out -= packets_acked;
2183
2184                 BUG_ON(tcp_skb_pcount(skb) == 0);
2185                 BUG_ON(!before(scb->seq, scb->end_seq));
2186         }
2187
2188         return acked;
2189 }
2190
2191 static u32 tcp_usrtt(const struct sk_buff *skb)
2192 {
2193         struct timeval tv, now;
2194
2195         do_gettimeofday(&now);
2196         skb_get_timestamp(skb, &tv);
2197         return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec);
2198 }
2199
2200 /* Remove acknowledged frames from the retransmission queue. */
2201 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2202 {
2203         struct tcp_sock *tp = tcp_sk(sk);
2204         const struct inet_connection_sock *icsk = inet_csk(sk);
2205         struct sk_buff *skb;
2206         __u32 now = tcp_time_stamp;
2207         int acked = 0;
2208         __s32 seq_rtt = -1;
2209         u32 pkts_acked = 0;
2210         void (*rtt_sample)(struct sock *sk, u32 usrtt)
2211                 = icsk->icsk_ca_ops->rtt_sample;
2212
2213         while ((skb = skb_peek(&sk->sk_write_queue)) &&
2214                skb != sk->sk_send_head) {
2215                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2216                 __u8 sacked = scb->sacked;
2217
2218                 /* If our packet is before the ack sequence we can
2219                  * discard it as it's confirmed to have arrived at
2220                  * the other end.
2221                  */
2222                 if (after(scb->end_seq, tp->snd_una)) {
2223                         if (tcp_skb_pcount(skb) > 1 &&
2224                             after(tp->snd_una, scb->seq))
2225                                 acked |= tcp_tso_acked(sk, skb,
2226                                                        now, &seq_rtt);
2227                         break;
2228                 }
2229
2230                 /* Initial outgoing SYN's get put onto the write_queue
2231                  * just like anything else we transmit.  It is not
2232                  * true data, and if we misinform our callers that
2233                  * this ACK acks real data, we will erroneously exit
2234                  * connection startup slow start one packet too
2235                  * quickly.  This is severely frowned upon behavior.
2236                  */
2237                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2238                         acked |= FLAG_DATA_ACKED;
2239                         ++pkts_acked;
2240                 } else {
2241                         acked |= FLAG_SYN_ACKED;
2242                         tp->retrans_stamp = 0;
2243                 }
2244
2245                 if (sacked) {
2246                         if (sacked & TCPCB_RETRANS) {
2247                                 if(sacked & TCPCB_SACKED_RETRANS)
2248                                         tp->retrans_out -= tcp_skb_pcount(skb);
2249                                 acked |= FLAG_RETRANS_DATA_ACKED;
2250                                 seq_rtt = -1;
2251                         } else if (seq_rtt < 0) {
2252                                 seq_rtt = now - scb->when;
2253                                 if (rtt_sample)
2254                                         (*rtt_sample)(sk, tcp_usrtt(skb));
2255                         }
2256                         if (sacked & TCPCB_SACKED_ACKED)
2257                                 tp->sacked_out -= tcp_skb_pcount(skb);
2258                         if (sacked & TCPCB_LOST)
2259                                 tp->lost_out -= tcp_skb_pcount(skb);
2260                         if (sacked & TCPCB_URG) {
2261                                 if (tp->urg_mode &&
2262                                     !before(scb->end_seq, tp->snd_up))
2263                                         tp->urg_mode = 0;
2264                         }
2265                 } else if (seq_rtt < 0) {
2266                         seq_rtt = now - scb->when;
2267                         if (rtt_sample)
2268                                 (*rtt_sample)(sk, tcp_usrtt(skb));
2269                 }
2270                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2271                 tcp_packets_out_dec(tp, skb);
2272                 __skb_unlink(skb, &sk->sk_write_queue);
2273                 sk_stream_free_skb(sk, skb);
2274                 clear_all_retrans_hints(tp);
2275         }
2276
2277         if (acked&FLAG_ACKED) {
2278                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2279                 tcp_ack_packets_out(sk, tp);
2280
2281                 if (icsk->icsk_ca_ops->pkts_acked)
2282                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2283         }
2284
2285 #if FASTRETRANS_DEBUG > 0
2286         BUG_TRAP((int)tp->sacked_out >= 0);
2287         BUG_TRAP((int)tp->lost_out >= 0);
2288         BUG_TRAP((int)tp->retrans_out >= 0);
2289         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2290                 const struct inet_connection_sock *icsk = inet_csk(sk);
2291                 if (tp->lost_out) {
2292                         printk(KERN_DEBUG "Leak l=%u %d\n",
2293                                tp->lost_out, icsk->icsk_ca_state);
2294                         tp->lost_out = 0;
2295                 }
2296                 if (tp->sacked_out) {
2297                         printk(KERN_DEBUG "Leak s=%u %d\n",
2298                                tp->sacked_out, icsk->icsk_ca_state);
2299                         tp->sacked_out = 0;
2300                 }
2301                 if (tp->retrans_out) {
2302                         printk(KERN_DEBUG "Leak r=%u %d\n",
2303                                tp->retrans_out, icsk->icsk_ca_state);
2304                         tp->retrans_out = 0;
2305                 }
2306         }
2307 #endif
2308         *seq_rtt_p = seq_rtt;
2309         return acked;
2310 }
2311
2312 static void tcp_ack_probe(struct sock *sk)
2313 {
2314         const struct tcp_sock *tp = tcp_sk(sk);
2315         struct inet_connection_sock *icsk = inet_csk(sk);
2316
2317         /* Was it a usable window open? */
2318
2319         if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2320                    tp->snd_una + tp->snd_wnd)) {
2321                 icsk->icsk_backoff = 0;
2322                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2323                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2324                  * This function is not for random using!
2325                  */
2326         } else {
2327                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2328                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2329                                           TCP_RTO_MAX);
2330         }
2331 }
2332
2333 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2334 {
2335         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2336                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2337 }
2338
2339 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2340 {
2341         const struct tcp_sock *tp = tcp_sk(sk);
2342         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2343                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2344 }
2345
2346 /* Check that window update is acceptable.
2347  * The function assumes that snd_una<=ack<=snd_next.
2348  */
2349 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2350                                         const u32 ack_seq, const u32 nwin)
2351 {
2352         return (after(ack, tp->snd_una) ||
2353                 after(ack_seq, tp->snd_wl1) ||
2354                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2355 }
2356
2357 /* Update our send window.
2358  *
2359  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2360  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2361  */
2362 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2363                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2364 {
2365         int flag = 0;
2366         u32 nwin = ntohs(skb->h.th->window);
2367
2368         if (likely(!skb->h.th->syn))
2369                 nwin <<= tp->rx_opt.snd_wscale;
2370
2371         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2372                 flag |= FLAG_WIN_UPDATE;
2373                 tcp_update_wl(tp, ack, ack_seq);
2374
2375                 if (tp->snd_wnd != nwin) {
2376                         tp->snd_wnd = nwin;
2377
2378                         /* Note, it is the only place, where
2379                          * fast path is recovered for sending TCP.
2380                          */
2381                         tp->pred_flags = 0;
2382                         tcp_fast_path_check(sk, tp);
2383
2384                         if (nwin > tp->max_window) {
2385                                 tp->max_window = nwin;
2386                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2387                         }
2388                 }
2389         }
2390
2391         tp->snd_una = ack;
2392
2393         return flag;
2394 }
2395
2396 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2397 {
2398         struct tcp_sock *tp = tcp_sk(sk);
2399         
2400         tcp_sync_left_out(tp);
2401         
2402         if (tp->snd_una == prior_snd_una ||
2403             !before(tp->snd_una, tp->frto_highmark)) {
2404                 /* RTO was caused by loss, start retransmitting in
2405                  * go-back-N slow start
2406                  */
2407                 tcp_enter_frto_loss(sk);
2408                 return;
2409         }
2410
2411         if (tp->frto_counter == 1) {
2412                 /* First ACK after RTO advances the window: allow two new
2413                  * segments out.
2414                  */
2415                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2416         } else {
2417                 /* Also the second ACK after RTO advances the window.
2418                  * The RTO was likely spurious. Reduce cwnd and continue
2419                  * in congestion avoidance
2420                  */
2421                 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2422                 tcp_moderate_cwnd(tp);
2423         }
2424
2425         /* F-RTO affects on two new ACKs following RTO.
2426          * At latest on third ACK the TCP behavior is back to normal.
2427          */
2428         tp->frto_counter = (tp->frto_counter + 1) % 3;
2429 }
2430
2431 /* This routine deals with incoming acks, but not outgoing ones. */
2432 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2433 {
2434         struct inet_connection_sock *icsk = inet_csk(sk);
2435         struct tcp_sock *tp = tcp_sk(sk);
2436         u32 prior_snd_una = tp->snd_una;
2437         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2438         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2439         u32 prior_in_flight;
2440         s32 seq_rtt;
2441         int prior_packets;
2442
2443         /* If the ack is newer than sent or older than previous acks
2444          * then we can probably ignore it.
2445          */
2446         if (after(ack, tp->snd_nxt))
2447                 goto uninteresting_ack;
2448
2449         if (before(ack, prior_snd_una))
2450                 goto old_ack;
2451
2452         if (sysctl_tcp_abc && icsk->icsk_ca_state < TCP_CA_CWR)
2453                 tp->bytes_acked += ack - prior_snd_una;
2454
2455         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2456                 /* Window is constant, pure forward advance.
2457                  * No more checks are required.
2458                  * Note, we use the fact that SND.UNA>=SND.WL2.
2459                  */
2460                 tcp_update_wl(tp, ack, ack_seq);
2461                 tp->snd_una = ack;
2462                 flag |= FLAG_WIN_UPDATE;
2463
2464                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2465
2466                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2467         } else {
2468                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2469                         flag |= FLAG_DATA;
2470                 else
2471                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2472
2473                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2474
2475                 if (TCP_SKB_CB(skb)->sacked)
2476                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2477
2478                 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2479                         flag |= FLAG_ECE;
2480
2481                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2482         }
2483
2484         /* We passed data and got it acked, remove any soft error
2485          * log. Something worked...
2486          */
2487         sk->sk_err_soft = 0;
2488         tp->rcv_tstamp = tcp_time_stamp;
2489         prior_packets = tp->packets_out;
2490         if (!prior_packets)
2491                 goto no_queue;
2492
2493         prior_in_flight = tcp_packets_in_flight(tp);
2494
2495         /* See if we can take anything off of the retransmit queue. */
2496         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2497
2498         if (tp->frto_counter)
2499                 tcp_process_frto(sk, prior_snd_una);
2500
2501         if (tcp_ack_is_dubious(sk, flag)) {
2502                 /* Advance CWND, if state allows this. */
2503                 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
2504                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2505                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2506         } else {
2507                 if ((flag & FLAG_DATA_ACKED))
2508                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2509         }
2510
2511         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2512                 dst_confirm(sk->sk_dst_cache);
2513
2514         return 1;
2515
2516 no_queue:
2517         icsk->icsk_probes_out = 0;
2518
2519         /* If this ack opens up a zero window, clear backoff.  It was
2520          * being used to time the probes, and is probably far higher than
2521          * it needs to be for normal retransmission.
2522          */
2523         if (sk->sk_send_head)
2524                 tcp_ack_probe(sk);
2525         return 1;
2526
2527 old_ack:
2528         if (TCP_SKB_CB(skb)->sacked)
2529                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2530
2531 uninteresting_ack:
2532         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2533         return 0;
2534 }
2535
2536
2537 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2538  * But, this can also be called on packets in the established flow when
2539  * the fast version below fails.
2540  */
2541 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2542 {
2543         unsigned char *ptr;
2544         struct tcphdr *th = skb->h.th;
2545         int length=(th->doff*4)-sizeof(struct tcphdr);
2546
2547         ptr = (unsigned char *)(th + 1);
2548         opt_rx->saw_tstamp = 0;
2549
2550         while(length>0) {
2551                 int opcode=*ptr++;
2552                 int opsize;
2553
2554                 switch (opcode) {
2555                         case TCPOPT_EOL:
2556                                 return;
2557                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2558                                 length--;
2559                                 continue;
2560                         default:
2561                                 opsize=*ptr++;
2562                                 if (opsize < 2) /* "silly options" */
2563                                         return;
2564                                 if (opsize > length)
2565                                         return; /* don't parse partial options */
2566                                 switch(opcode) {
2567                                 case TCPOPT_MSS:
2568                                         if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2569                                                 u16 in_mss = ntohs(get_unaligned((__u16 *)ptr));
2570                                                 if (in_mss) {
2571                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2572                                                                 in_mss = opt_rx->user_mss;
2573                                                         opt_rx->mss_clamp = in_mss;
2574                                                 }
2575                                         }
2576                                         break;
2577                                 case TCPOPT_WINDOW:
2578                                         if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2579                                                 if (sysctl_tcp_window_scaling) {
2580                                                         __u8 snd_wscale = *(__u8 *) ptr;
2581                                                         opt_rx->wscale_ok = 1;
2582                                                         if (snd_wscale > 14) {
2583                                                                 if(net_ratelimit())
2584                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2585                                                                                "scaling value %d >14 received.\n",
2586                                                                                snd_wscale);
2587                                                                 snd_wscale = 14;
2588                                                         }
2589                                                         opt_rx->snd_wscale = snd_wscale;
2590                                                 }
2591                                         break;
2592                                 case TCPOPT_TIMESTAMP:
2593                                         if(opsize==TCPOLEN_TIMESTAMP) {
2594                                                 if ((estab && opt_rx->tstamp_ok) ||
2595                                                     (!estab && sysctl_tcp_timestamps)) {
2596                                                         opt_rx->saw_tstamp = 1;
2597                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr));
2598                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4)));
2599                                                 }
2600                                         }
2601                                         break;
2602                                 case TCPOPT_SACK_PERM:
2603                                         if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2604                                                 if (sysctl_tcp_sack) {
2605                                                         opt_rx->sack_ok = 1;
2606                                                         tcp_sack_reset(opt_rx);
2607                                                 }
2608                                         }
2609                                         break;
2610
2611                                 case TCPOPT_SACK:
2612                                         if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2613                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2614                                            opt_rx->sack_ok) {
2615                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2616                                         }
2617                                 };
2618                                 ptr+=opsize-2;
2619                                 length-=opsize;
2620                 };
2621         }
2622 }
2623
2624 /* Fast parse options. This hopes to only see timestamps.
2625  * If it is wrong it falls back on tcp_parse_options().
2626  */
2627 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2628                                   struct tcp_sock *tp)
2629 {
2630         if (th->doff == sizeof(struct tcphdr)>>2) {
2631                 tp->rx_opt.saw_tstamp = 0;
2632                 return 0;
2633         } else if (tp->rx_opt.tstamp_ok &&
2634                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2635                 __u32 *ptr = (__u32 *)(th + 1);
2636                 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2637                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2638                         tp->rx_opt.saw_tstamp = 1;
2639                         ++ptr;
2640                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2641                         ++ptr;
2642                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2643                         return 1;
2644                 }
2645         }
2646         tcp_parse_options(skb, &tp->rx_opt, 1);
2647         return 1;
2648 }
2649
2650 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2651 {
2652         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2653         tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2654 }
2655
2656 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2657 {
2658         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2659                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2660                  * extra check below makes sure this can only happen
2661                  * for pure ACK frames.  -DaveM
2662                  *
2663                  * Not only, also it occurs for expired timestamps.
2664                  */
2665
2666                 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2667                    xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2668                         tcp_store_ts_recent(tp);
2669         }
2670 }
2671
2672 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2673  *
2674  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2675  * it can pass through stack. So, the following predicate verifies that
2676  * this segment is not used for anything but congestion avoidance or
2677  * fast retransmit. Moreover, we even are able to eliminate most of such
2678  * second order effects, if we apply some small "replay" window (~RTO)
2679  * to timestamp space.
2680  *
2681  * All these measures still do not guarantee that we reject wrapped ACKs
2682  * on networks with high bandwidth, when sequence space is recycled fastly,
2683  * but it guarantees that such events will be very rare and do not affect
2684  * connection seriously. This doesn't look nice, but alas, PAWS is really
2685  * buggy extension.
2686  *
2687  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2688  * states that events when retransmit arrives after original data are rare.
2689  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2690  * the biggest problem on large power networks even with minor reordering.
2691  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2692  * up to bandwidth of 18Gigabit/sec. 8) ]
2693  */
2694
2695 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2696 {
2697         struct tcp_sock *tp = tcp_sk(sk);
2698         struct tcphdr *th = skb->h.th;
2699         u32 seq = TCP_SKB_CB(skb)->seq;
2700         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2701
2702         return (/* 1. Pure ACK with correct sequence number. */
2703                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2704
2705                 /* 2. ... and duplicate ACK. */
2706                 ack == tp->snd_una &&
2707
2708                 /* 3. ... and does not update window. */
2709                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2710
2711                 /* 4. ... and sits in replay window. */
2712                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
2713 }
2714
2715 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
2716 {
2717         const struct tcp_sock *tp = tcp_sk(sk);
2718         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2719                 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2720                 !tcp_disordered_ack(sk, skb));
2721 }
2722
2723 /* Check segment sequence number for validity.
2724  *
2725  * Segment controls are considered valid, if the segment
2726  * fits to the window after truncation to the window. Acceptability
2727  * of data (and SYN, FIN, of course) is checked separately.
2728  * See tcp_data_queue(), for example.
2729  *
2730  * Also, controls (RST is main one) are accepted using RCV.WUP instead
2731  * of RCV.NXT. Peer still did not advance his SND.UNA when we
2732  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2733  * (borrowed from freebsd)
2734  */
2735
2736 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2737 {
2738         return  !before(end_seq, tp->rcv_wup) &&
2739                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2740 }
2741
2742 /* When we get a reset we do this. */
2743 static void tcp_reset(struct sock *sk)
2744 {
2745         /* We want the right error as BSD sees it (and indeed as we do). */
2746         switch (sk->sk_state) {
2747                 case TCP_SYN_SENT:
2748                         sk->sk_err = ECONNREFUSED;
2749                         break;
2750                 case TCP_CLOSE_WAIT:
2751                         sk->sk_err = EPIPE;
2752                         break;
2753                 case TCP_CLOSE:
2754                         return;
2755                 default:
2756                         sk->sk_err = ECONNRESET;
2757         }
2758
2759         if (!sock_flag(sk, SOCK_DEAD))
2760                 sk->sk_error_report(sk);
2761
2762         tcp_done(sk);
2763 }
2764
2765 /*
2766  *      Process the FIN bit. This now behaves as it is supposed to work
2767  *      and the FIN takes effect when it is validly part of sequence
2768  *      space. Not before when we get holes.
2769  *
2770  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2771  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
2772  *      TIME-WAIT)
2773  *
2774  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
2775  *      close and we go into CLOSING (and later onto TIME-WAIT)
2776  *
2777  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2778  */
2779 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2780 {
2781         struct tcp_sock *tp = tcp_sk(sk);
2782
2783         inet_csk_schedule_ack(sk);
2784
2785         sk->sk_shutdown |= RCV_SHUTDOWN;
2786         sock_set_flag(sk, SOCK_DONE);
2787
2788         switch (sk->sk_state) {
2789                 case TCP_SYN_RECV:
2790                 case TCP_ESTABLISHED:
2791                         /* Move to CLOSE_WAIT */
2792                         tcp_set_state(sk, TCP_CLOSE_WAIT);
2793                         inet_csk(sk)->icsk_ack.pingpong = 1;
2794                         break;
2795
2796                 case TCP_CLOSE_WAIT:
2797                 case TCP_CLOSING:
2798                         /* Received a retransmission of the FIN, do
2799                          * nothing.
2800                          */
2801                         break;
2802                 case TCP_LAST_ACK:
2803                         /* RFC793: Remain in the LAST-ACK state. */
2804                         break;
2805
2806                 case TCP_FIN_WAIT1:
2807                         /* This case occurs when a simultaneous close
2808                          * happens, we must ack the received FIN and
2809                          * enter the CLOSING state.
2810                          */
2811                         tcp_send_ack(sk);
2812                         tcp_set_state(sk, TCP_CLOSING);
2813                         break;
2814                 case TCP_FIN_WAIT2:
2815                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
2816                         tcp_send_ack(sk);
2817                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2818                         break;
2819                 default:
2820                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2821                          * cases we should never reach this piece of code.
2822                          */
2823                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2824                                __FUNCTION__, sk->sk_state);
2825                         break;
2826         };
2827
2828         /* It _is_ possible, that we have something out-of-order _after_ FIN.
2829          * Probably, we should reset in this case. For now drop them.
2830          */
2831         __skb_queue_purge(&tp->out_of_order_queue);
2832         if (tp->rx_opt.sack_ok)
2833                 tcp_sack_reset(&tp->rx_opt);
2834         sk_stream_mem_reclaim(sk);
2835
2836         if (!sock_flag(sk, SOCK_DEAD)) {
2837                 sk->sk_state_change(sk);
2838
2839                 /* Do not send POLL_HUP for half duplex close. */
2840                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2841                     sk->sk_state == TCP_CLOSE)
2842                         sk_wake_async(sk, 1, POLL_HUP);
2843                 else
2844                         sk_wake_async(sk, 1, POLL_IN);
2845         }
2846 }
2847
2848 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2849 {
2850         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2851                 if (before(seq, sp->start_seq))
2852                         sp->start_seq = seq;
2853                 if (after(end_seq, sp->end_seq))
2854                         sp->end_seq = end_seq;
2855                 return 1;
2856         }
2857         return 0;
2858 }
2859
2860 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
2861 {
2862         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2863                 if (before(seq, tp->rcv_nxt))
2864                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
2865                 else
2866                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
2867
2868                 tp->rx_opt.dsack = 1;
2869                 tp->duplicate_sack[0].start_seq = seq;
2870                 tp->duplicate_sack[0].end_seq = end_seq;
2871                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
2872         }
2873 }
2874
2875 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
2876 {
2877         if (!tp->rx_opt.dsack)
2878                 tcp_dsack_set(tp, seq, end_seq);
2879         else
2880                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2881 }
2882
2883 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2884 {
2885         struct tcp_sock *tp = tcp_sk(sk);
2886
2887         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2888             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2889                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2890                 tcp_enter_quickack_mode(sk);
2891
2892                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2893                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2894
2895                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2896                                 end_seq = tp->rcv_nxt;
2897                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2898                 }
2899         }
2900
2901         tcp_send_ack(sk);
2902 }
2903
2904 /* These routines update the SACK block as out-of-order packets arrive or
2905  * in-order packets close up the sequence space.
2906  */
2907 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
2908 {
2909         int this_sack;
2910         struct tcp_sack_block *sp = &tp->selective_acks[0];
2911         struct tcp_sack_block *swalk = sp+1;
2912
2913         /* See if the recent change to the first SACK eats into
2914          * or hits the sequence space of other SACK blocks, if so coalesce.
2915          */
2916         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
2917                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2918                         int i;
2919
2920                         /* Zap SWALK, by moving every further SACK up by one slot.
2921                          * Decrease num_sacks.
2922                          */
2923                         tp->rx_opt.num_sacks--;
2924                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2925                         for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
2926                                 sp[i] = sp[i+1];
2927                         continue;
2928                 }
2929                 this_sack++, swalk++;
2930         }
2931 }
2932
2933 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2934 {
2935         __u32 tmp;
2936
2937         tmp = sack1->start_seq;
2938         sack1->start_seq = sack2->start_seq;
2939         sack2->start_seq = tmp;
2940
2941         tmp = sack1->end_seq;
2942         sack1->end_seq = sack2->end_seq;
2943         sack2->end_seq = tmp;
2944 }
2945
2946 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
2947 {
2948         struct tcp_sock *tp = tcp_sk(sk);
2949         struct tcp_sack_block *sp = &tp->selective_acks[0];
2950         int cur_sacks = tp->rx_opt.num_sacks;
2951         int this_sack;
2952
2953         if (!cur_sacks)
2954                 goto new_sack;
2955
2956         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
2957                 if (tcp_sack_extend(sp, seq, end_seq)) {
2958                         /* Rotate this_sack to the first one. */
2959                         for (; this_sack>0; this_sack--, sp--)
2960                                 tcp_sack_swap(sp, sp-1);
2961                         if (cur_sacks > 1)
2962                                 tcp_sack_maybe_coalesce(tp);
2963                         return;
2964                 }
2965         }
2966
2967         /* Could not find an adjacent existing SACK, build a new one,
2968          * put it at the front, and shift everyone else down.  We
2969          * always know there is at least one SACK present already here.
2970          *
2971          * If the sack array is full, forget about the last one.
2972          */
2973         if (this_sack >= 4) {
2974                 this_sack--;
2975                 tp->rx_opt.num_sacks--;
2976                 sp--;
2977         }
2978         for(; this_sack > 0; this_sack--, sp--)
2979                 *sp = *(sp-1);
2980
2981 new_sack:
2982         /* Build the new head SACK, and we're done. */
2983         sp->start_seq = seq;
2984         sp->end_seq = end_seq;
2985         tp->rx_opt.num_sacks++;
2986         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2987 }
2988
2989 /* RCV.NXT advances, some SACKs should be eaten. */
2990
2991 static void tcp_sack_remove(struct tcp_sock *tp)
2992 {
2993         struct tcp_sack_block *sp = &tp->selective_acks[0];
2994         int num_sacks = tp->rx_opt.num_sacks;
2995         int this_sack;
2996
2997         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2998         if (skb_queue_empty(&tp->out_of_order_queue)) {
2999                 tp->rx_opt.num_sacks = 0;
3000                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3001                 return;
3002         }
3003
3004         for(this_sack = 0; this_sack < num_sacks; ) {
3005                 /* Check if the start of the sack is covered by RCV.NXT. */
3006                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3007                         int i;
3008
3009                         /* RCV.NXT must cover all the block! */
3010                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3011
3012                         /* Zap this SACK, by moving forward any other SACKS. */
3013                         for (i=this_sack+1; i < num_sacks; i++)
3014                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3015                         num_sacks--;
3016                         continue;
3017                 }
3018                 this_sack++;
3019                 sp++;
3020         }
3021         if (num_sacks != tp->rx_opt.num_sacks) {
3022                 tp->rx_opt.num_sacks = num_sacks;
3023                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3024         }
3025 }
3026
3027 /* This one checks to see if we can put data from the
3028  * out_of_order queue into the receive_queue.
3029  */
3030 static void tcp_ofo_queue(struct sock *sk)
3031 {
3032         struct tcp_sock *tp = tcp_sk(sk);
3033         __u32 dsack_high = tp->rcv_nxt;
3034         struct sk_buff *skb;
3035
3036         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3037                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3038                         break;
3039
3040                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3041                         __u32 dsack = dsack_high;
3042                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3043                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3044                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3045                 }
3046
3047                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3048                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3049                         __skb_unlink(skb, &tp->out_of_order_queue);
3050                         __kfree_skb(skb);
3051                         continue;
3052                 }
3053                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3054                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3055                            TCP_SKB_CB(skb)->end_seq);
3056
3057                 __skb_unlink(skb, &tp->out_of_order_queue);
3058                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3059                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3060                 if(skb->h.th->fin)
3061                         tcp_fin(skb, sk, skb->h.th);
3062         }
3063 }
3064
3065 static int tcp_prune_queue(struct sock *sk);
3066
3067 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3068 {
3069         struct tcphdr *th = skb->h.th;
3070         struct tcp_sock *tp = tcp_sk(sk);
3071         int eaten = -1;
3072
3073         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3074                 goto drop;
3075
3076         __skb_pull(skb, th->doff*4);
3077
3078         TCP_ECN_accept_cwr(tp, skb);
3079
3080         if (tp->rx_opt.dsack) {
3081                 tp->rx_opt.dsack = 0;
3082                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3083                                                     4 - tp->rx_opt.tstamp_ok);
3084         }
3085
3086         /*  Queue data for delivery to the user.
3087          *  Packets in sequence go to the receive queue.
3088          *  Out of sequence packets to the out_of_order_queue.
3089          */
3090         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3091                 if (tcp_receive_window(tp) == 0)
3092                         goto out_of_window;
3093
3094                 /* Ok. In sequence. In window. */
3095                 if (tp->ucopy.task == current &&
3096                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3097                     sock_owned_by_user(sk) && !tp->urg_data) {
3098                         int chunk = min_t(unsigned int, skb->len,
3099                                                         tp->ucopy.len);
3100
3101                         __set_current_state(TASK_RUNNING);
3102
3103                         local_bh_enable();
3104                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3105                                 tp->ucopy.len -= chunk;
3106                                 tp->copied_seq += chunk;
3107                                 eaten = (chunk == skb->len && !th->fin);
3108                                 tcp_rcv_space_adjust(sk);
3109                         }
3110                         local_bh_disable();
3111                 }
3112
3113                 if (eaten <= 0) {
3114 queue_and_out:
3115                         if (eaten < 0 &&
3116                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3117                              !sk_stream_rmem_schedule(sk, skb))) {
3118                                 if (tcp_prune_queue(sk) < 0 ||
3119                                     !sk_stream_rmem_schedule(sk, skb))
3120                                         goto drop;
3121                         }
3122                         sk_stream_set_owner_r(skb, sk);
3123                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3124                 }
3125                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3126                 if(skb->len)
3127                         tcp_event_data_recv(sk, tp, skb);
3128                 if(th->fin)
3129                         tcp_fin(skb, sk, th);
3130
3131                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3132                         tcp_ofo_queue(sk);
3133
3134                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3135                          * gap in queue is filled.
3136                          */
3137                         if (skb_queue_empty(&tp->out_of_order_queue))
3138                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3139                 }
3140
3141                 if (tp->rx_opt.num_sacks)
3142                         tcp_sack_remove(tp);
3143
3144                 tcp_fast_path_check(sk, tp);
3145
3146                 if (eaten > 0)
3147                         __kfree_skb(skb);
3148                 else if (!sock_flag(sk, SOCK_DEAD))
3149                         sk->sk_data_ready(sk, 0);
3150                 return;
3151         }
3152
3153         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3154                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3155                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3156                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3157
3158 out_of_window:
3159                 tcp_enter_quickack_mode(sk);
3160                 inet_csk_schedule_ack(sk);
3161 drop:
3162                 __kfree_skb(skb);
3163                 return;
3164         }
3165
3166         /* Out of window. F.e. zero window probe. */
3167         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3168                 goto out_of_window;
3169
3170         tcp_enter_quickack_mode(sk);
3171
3172         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3173                 /* Partial packet, seq < rcv_next < end_seq */
3174                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3175                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3176                            TCP_SKB_CB(skb)->end_seq);
3177
3178                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3179                 
3180                 /* If window is closed, drop tail of packet. But after
3181                  * remembering D-SACK for its head made in previous line.
3182                  */
3183                 if (!tcp_receive_window(tp))
3184                         goto out_of_window;
3185                 goto queue_and_out;
3186         }
3187
3188         TCP_ECN_check_ce(tp, skb);
3189
3190         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3191             !sk_stream_rmem_schedule(sk, skb)) {
3192                 if (tcp_prune_queue(sk) < 0 ||
3193                     !sk_stream_rmem_schedule(sk, skb))
3194                         goto drop;
3195         }
3196
3197         /* Disable header prediction. */
3198         tp->pred_flags = 0;
3199         inet_csk_schedule_ack(sk);
3200
3201         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3202                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3203
3204         sk_stream_set_owner_r(skb, sk);
3205
3206         if (!skb_peek(&tp->out_of_order_queue)) {
3207                 /* Initial out of order segment, build 1 SACK. */
3208                 if (tp->rx_opt.sack_ok) {
3209                         tp->rx_opt.num_sacks = 1;
3210                         tp->rx_opt.dsack     = 0;
3211                         tp->rx_opt.eff_sacks = 1;
3212                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3213                         tp->selective_acks[0].end_seq =
3214                                                 TCP_SKB_CB(skb)->end_seq;
3215                 }
3216                 __skb_queue_head(&tp->out_of_order_queue,skb);
3217         } else {
3218                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3219                 u32 seq = TCP_SKB_CB(skb)->seq;
3220                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3221
3222                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3223                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3224
3225                         if (!tp->rx_opt.num_sacks ||
3226                             tp->selective_acks[0].end_seq != seq)
3227                                 goto add_sack;
3228
3229                         /* Common case: data arrive in order after hole. */
3230                         tp->selective_acks[0].end_seq = end_seq;
3231                         return;
3232                 }
3233
3234                 /* Find place to insert this segment. */
3235                 do {
3236                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3237                                 break;
3238                 } while ((skb1 = skb1->prev) !=
3239                          (struct sk_buff*)&tp->out_of_order_queue);
3240
3241                 /* Do skb overlap to previous one? */
3242                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3243                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3244                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3245                                 /* All the bits are present. Drop. */
3246                                 __kfree_skb(skb);
3247                                 tcp_dsack_set(tp, seq, end_seq);
3248                                 goto add_sack;
3249                         }
3250                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3251                                 /* Partial overlap. */
3252                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3253                         } else {
3254                                 skb1 = skb1->prev;
3255                         }
3256                 }
3257                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3258                 
3259                 /* And clean segments covered by new one as whole. */
3260                 while ((skb1 = skb->next) !=
3261                        (struct sk_buff*)&tp->out_of_order_queue &&
3262                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3263                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3264                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3265                                break;
3266                        }
3267                        __skb_unlink(skb1, &tp->out_of_order_queue);
3268                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3269                        __kfree_skb(skb1);
3270                 }
3271
3272 add_sack:
3273                 if (tp->rx_opt.sack_ok)
3274                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3275         }
3276 }
3277
3278 /* Collapse contiguous sequence of skbs head..tail with
3279  * sequence numbers start..end.
3280  * Segments with FIN/SYN are not collapsed (only because this
3281  * simplifies code)
3282  */
3283 static void
3284 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3285              struct sk_buff *head, struct sk_buff *tail,
3286              u32 start, u32 end)
3287 {
3288         struct sk_buff *skb;
3289
3290         /* First, check that queue is collapsible and find
3291          * the point where collapsing can be useful. */
3292         for (skb = head; skb != tail; ) {
3293                 /* No new bits? It is possible on ofo queue. */
3294                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3295                         struct sk_buff *next = skb->next;
3296                         __skb_unlink(skb, list);
3297                         __kfree_skb(skb);
3298                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3299                         skb = next;
3300                         continue;
3301                 }
3302
3303                 /* The first skb to collapse is:
3304                  * - not SYN/FIN and
3305                  * - bloated or contains data before "start" or
3306                  *   overlaps to the next one.
3307                  */
3308                 if (!skb->h.th->syn && !skb->h.th->fin &&
3309                     (tcp_win_from_space(skb->truesize) > skb->len ||
3310                      before(TCP_SKB_CB(skb)->seq, start) ||
3311                      (skb->next != tail &&
3312                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3313                         break;
3314
3315                 /* Decided to skip this, advance start seq. */
3316                 start = TCP_SKB_CB(skb)->end_seq;
3317                 skb = skb->next;
3318         }
3319         if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3320                 return;
3321
3322         while (before(start, end)) {
3323                 struct sk_buff *nskb;
3324                 int header = skb_headroom(skb);
3325                 int copy = SKB_MAX_ORDER(header, 0);
3326
3327                 /* Too big header? This can happen with IPv6. */
3328                 if (copy < 0)
3329                         return;
3330                 if (end-start < copy)
3331                         copy = end-start;
3332                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3333                 if (!nskb)
3334                         return;
3335                 skb_reserve(nskb, header);
3336                 memcpy(nskb->head, skb->head, header);
3337                 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3338                 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3339                 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3340                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3341                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3342                 __skb_insert(nskb, skb->prev, skb, list);
3343                 sk_stream_set_owner_r(nskb, sk);
3344
3345                 /* Copy data, releasing collapsed skbs. */
3346                 while (copy > 0) {
3347                         int offset = start - TCP_SKB_CB(skb)->seq;
3348                         int size = TCP_SKB_CB(skb)->end_seq - start;
3349
3350                         BUG_ON(offset < 0);
3351                         if (size > 0) {
3352                                 size = min(copy, size);
3353                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3354                                         BUG();
3355                                 TCP_SKB_CB(nskb)->end_seq += size;
3356                                 copy -= size;
3357                                 start += size;
3358                         }
3359                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3360                                 struct sk_buff *next = skb->next;
3361                                 __skb_unlink(skb, list);
3362                                 __kfree_skb(skb);
3363                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3364                                 skb = next;
3365                                 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3366                                         return;
3367                         }
3368                 }
3369         }
3370 }
3371
3372 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3373  * and tcp_collapse() them until all the queue is collapsed.
3374  */
3375 static void tcp_collapse_ofo_queue(struct sock *sk)
3376 {
3377         struct tcp_sock *tp = tcp_sk(sk);
3378         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3379         struct sk_buff *head;
3380         u32 start, end;
3381
3382         if (skb == NULL)
3383                 return;
3384
3385         start = TCP_SKB_CB(skb)->seq;
3386         end = TCP_SKB_CB(skb)->end_seq;
3387         head = skb;
3388
3389         for (;;) {
3390                 skb = skb->next;
3391
3392                 /* Segment is terminated when we see gap or when
3393                  * we are at the end of all the queue. */
3394                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3395                     after(TCP_SKB_CB(skb)->seq, end) ||
3396                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3397                         tcp_collapse(sk, &tp->out_of_order_queue,
3398                                      head, skb, start, end);
3399                         head = skb;
3400                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3401                                 break;
3402                         /* Start new segment */
3403                         start = TCP_SKB_CB(skb)->seq;
3404                         end = TCP_SKB_CB(skb)->end_seq;
3405                 } else {
3406                         if (before(TCP_SKB_CB(skb)->seq, start))
3407                                 start = TCP_SKB_CB(skb)->seq;
3408                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3409                                 end = TCP_SKB_CB(skb)->end_seq;
3410                 }
3411         }
3412 }
3413
3414 /* Reduce allocated memory if we can, trying to get
3415  * the socket within its memory limits again.
3416  *
3417  * Return less than zero if we should start dropping frames
3418  * until the socket owning process reads some of the data
3419  * to stabilize the situation.
3420  */
3421 static int tcp_prune_queue(struct sock *sk)
3422 {
3423         struct tcp_sock *tp = tcp_sk(sk); 
3424
3425         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3426
3427         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3428
3429         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3430                 tcp_clamp_window(sk, tp);
3431         else if (tcp_memory_pressure)
3432                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3433
3434         tcp_collapse_ofo_queue(sk);
3435         tcp_collapse(sk, &sk->sk_receive_queue,
3436                      sk->sk_receive_queue.next,
3437                      (struct sk_buff*)&sk->sk_receive_queue,
3438                      tp->copied_seq, tp->rcv_nxt);
3439         sk_stream_mem_reclaim(sk);
3440
3441         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3442                 return 0;
3443
3444         /* Collapsing did not help, destructive actions follow.
3445          * This must not ever occur. */
3446
3447         /* First, purge the out_of_order queue. */
3448         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3449                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3450                 __skb_queue_purge(&tp->out_of_order_queue);
3451
3452                 /* Reset SACK state.  A conforming SACK implementation will
3453                  * do the same at a timeout based retransmit.  When a connection
3454                  * is in a sad state like this, we care only about integrity
3455                  * of the connection not performance.
3456                  */
3457                 if (tp->rx_opt.sack_ok)
3458                         tcp_sack_reset(&tp->rx_opt);
3459                 sk_stream_mem_reclaim(sk);
3460         }
3461
3462         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3463                 return 0;
3464
3465         /* If we are really being abused, tell the caller to silently
3466          * drop receive data on the floor.  It will get retransmitted
3467          * and hopefully then we'll have sufficient space.
3468          */
3469         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3470
3471         /* Massive buffer overcommit. */
3472         tp->pred_flags = 0;
3473         return -1;
3474 }
3475
3476
3477 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3478  * As additional protections, we do not touch cwnd in retransmission phases,
3479  * and if application hit its sndbuf limit recently.
3480  */
3481 void tcp_cwnd_application_limited(struct sock *sk)
3482 {
3483         struct tcp_sock *tp = tcp_sk(sk);
3484
3485         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3486             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3487                 /* Limited by application or receiver window. */
3488                 u32 win_used = max(tp->snd_cwnd_used, 2U);
3489                 if (win_used < tp->snd_cwnd) {
3490                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3491                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3492                 }
3493                 tp->snd_cwnd_used = 0;
3494         }
3495         tp->snd_cwnd_stamp = tcp_time_stamp;
3496 }
3497
3498 static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3499 {
3500         /* If the user specified a specific send buffer setting, do
3501          * not modify it.
3502          */
3503         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3504                 return 0;
3505
3506         /* If we are under global TCP memory pressure, do not expand.  */
3507         if (tcp_memory_pressure)
3508                 return 0;
3509
3510         /* If we are under soft global TCP memory pressure, do not expand.  */
3511         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3512                 return 0;
3513
3514         /* If we filled the congestion window, do not expand.  */
3515         if (tp->packets_out >= tp->snd_cwnd)
3516                 return 0;
3517
3518         return 1;
3519 }
3520
3521 /* When incoming ACK allowed to free some skb from write_queue,
3522  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3523  * on the exit from tcp input handler.
3524  *
3525  * PROBLEM: sndbuf expansion does not work well with largesend.
3526  */
3527 static void tcp_new_space(struct sock *sk)
3528 {
3529         struct tcp_sock *tp = tcp_sk(sk);
3530
3531         if (tcp_should_expand_sndbuf(sk, tp)) {
3532                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3533                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3534                     demanded = max_t(unsigned int, tp->snd_cwnd,
3535                                                    tp->reordering + 1);
3536                 sndmem *= 2*demanded;
3537                 if (sndmem > sk->sk_sndbuf)
3538                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3539                 tp->snd_cwnd_stamp = tcp_time_stamp;
3540         }
3541
3542         sk->sk_write_space(sk);
3543 }
3544
3545 static void tcp_check_space(struct sock *sk)
3546 {
3547         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3548                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3549                 if (sk->sk_socket &&
3550                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3551                         tcp_new_space(sk);
3552         }
3553 }
3554
3555 static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3556 {
3557         tcp_push_pending_frames(sk, tp);
3558         tcp_check_space(sk);
3559 }
3560
3561 /*
3562  * Check if sending an ack is needed.
3563  */
3564 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3565 {
3566         struct tcp_sock *tp = tcp_sk(sk);
3567
3568             /* More than one full frame received... */
3569         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3570              /* ... and right edge of window advances far enough.
3571               * (tcp_recvmsg() will send ACK otherwise). Or...
3572               */
3573              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3574             /* We ACK each frame or... */
3575             tcp_in_quickack_mode(sk) ||
3576             /* We have out of order data. */
3577             (ofo_possible &&
3578              skb_peek(&tp->out_of_order_queue))) {
3579                 /* Then ack it now */
3580                 tcp_send_ack(sk);
3581         } else {
3582                 /* Else, send delayed ack. */
3583                 tcp_send_delayed_ack(sk);
3584         }
3585 }
3586
3587 static inline void tcp_ack_snd_check(struct sock *sk)
3588 {
3589         if (!inet_csk_ack_scheduled(sk)) {
3590                 /* We sent a data segment already. */
3591                 return;
3592         }
3593         __tcp_ack_snd_check(sk, 1);
3594 }
3595
3596 /*
3597  *      This routine is only called when we have urgent data
3598  *      signaled. Its the 'slow' part of tcp_urg. It could be
3599  *      moved inline now as tcp_urg is only called from one
3600  *      place. We handle URGent data wrong. We have to - as
3601  *      BSD still doesn't use the correction from RFC961.
3602  *      For 1003.1g we should support a new option TCP_STDURG to permit
3603  *      either form (or just set the sysctl tcp_stdurg).
3604  */
3605  
3606 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3607 {
3608         struct tcp_sock *tp = tcp_sk(sk);
3609         u32 ptr = ntohs(th->urg_ptr);
3610
3611         if (ptr && !sysctl_tcp_stdurg)
3612                 ptr--;
3613         ptr += ntohl(th->seq);
3614
3615         /* Ignore urgent data that we've already seen and read. */
3616         if (after(tp->copied_seq, ptr))
3617                 return;
3618
3619         /* Do not replay urg ptr.
3620          *
3621          * NOTE: interesting situation not covered by specs.
3622          * Misbehaving sender may send urg ptr, pointing to segment,
3623          * which we already have in ofo queue. We are not able to fetch
3624          * such data and will stay in TCP_URG_NOTYET until will be eaten
3625          * by recvmsg(). Seems, we are not obliged to handle such wicked
3626          * situations. But it is worth to think about possibility of some
3627          * DoSes using some hypothetical application level deadlock.
3628          */
3629         if (before(ptr, tp->rcv_nxt))
3630                 return;
3631
3632         /* Do we already have a newer (or duplicate) urgent pointer? */
3633         if (tp->urg_data && !after(ptr, tp->urg_seq))
3634                 return;
3635
3636         /* Tell the world about our new urgent pointer. */
3637         sk_send_sigurg(sk);
3638
3639         /* We may be adding urgent data when the last byte read was
3640          * urgent. To do this requires some care. We cannot just ignore
3641          * tp->copied_seq since we would read the last urgent byte again
3642          * as data, nor can we alter copied_seq until this data arrives
3643          * or we break the semantics of SIOCATMARK (and thus sockatmark())
3644          *
3645          * NOTE. Double Dutch. Rendering to plain English: author of comment
3646          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3647          * and expect that both A and B disappear from stream. This is _wrong_.
3648          * Though this happens in BSD with high probability, this is occasional.
3649          * Any application relying on this is buggy. Note also, that fix "works"
3650          * only in this artificial test. Insert some normal data between A and B and we will
3651          * decline of BSD again. Verdict: it is better to remove to trap
3652          * buggy users.
3653          */
3654         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3655             !sock_flag(sk, SOCK_URGINLINE) &&
3656             tp->copied_seq != tp->rcv_nxt) {
3657                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3658                 tp->copied_seq++;
3659                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3660                         __skb_unlink(skb, &sk->sk_receive_queue);
3661                         __kfree_skb(skb);
3662                 }
3663         }
3664
3665         tp->urg_data   = TCP_URG_NOTYET;
3666         tp->urg_seq    = ptr;
3667
3668         /* Disable header prediction. */
3669         tp->pred_flags = 0;
3670 }
3671
3672 /* This is the 'fast' part of urgent handling. */
3673 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3674 {
3675         struct tcp_sock *tp = tcp_sk(sk);
3676
3677         /* Check if we get a new urgent pointer - normally not. */
3678         if (th->urg)
3679                 tcp_check_urg(sk,th);
3680
3681         /* Do we wait for any urgent data? - normally not... */
3682         if (tp->urg_data == TCP_URG_NOTYET) {
3683                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3684                           th->syn;
3685
3686                 /* Is the urgent pointer pointing into this packet? */   
3687                 if (ptr < skb->len) {
3688                         u8 tmp;
3689                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3690                                 BUG();
3691                         tp->urg_data = TCP_URG_VALID | tmp;
3692                         if (!sock_flag(sk, SOCK_DEAD))
3693                                 sk->sk_data_ready(sk, 0);
3694                 }
3695         }
3696 }
3697
3698 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3699 {
3700         struct tcp_sock *tp = tcp_sk(sk);
3701         int chunk = skb->len - hlen;
3702         int err;
3703
3704         local_bh_enable();
3705         if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3706                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3707         else
3708                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3709                                                        tp->ucopy.iov);
3710
3711         if (!err) {
3712                 tp->ucopy.len -= chunk;
3713                 tp->copied_seq += chunk;
3714                 tcp_rcv_space_adjust(sk);
3715         }
3716
3717         local_bh_disable();
3718         return err;
3719 }
3720
3721 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3722 {
3723         int result;
3724
3725         if (sock_owned_by_user(sk)) {
3726                 local_bh_enable();
3727                 result = __tcp_checksum_complete(skb);
3728                 local_bh_disable();
3729         } else {
3730                 result = __tcp_checksum_complete(skb);
3731         }
3732         return result;
3733 }
3734
3735 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3736 {
3737         return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3738                 __tcp_checksum_complete_user(sk, skb);
3739 }
3740
3741 /*
3742  *      TCP receive function for the ESTABLISHED state. 
3743  *
3744  *      It is split into a fast path and a slow path. The fast path is 
3745  *      disabled when:
3746  *      - A zero window was announced from us - zero window probing
3747  *        is only handled properly in the slow path. 
3748  *      - Out of order segments arrived.
3749  *      - Urgent data is expected.
3750  *      - There is no buffer space left
3751  *      - Unexpected TCP flags/window values/header lengths are received
3752  *        (detected by checking the TCP header against pred_flags) 
3753  *      - Data is sent in both directions. Fast path only supports pure senders
3754  *        or pure receivers (this means either the sequence number or the ack
3755  *        value must stay constant)
3756  *      - Unexpected TCP option.
3757  *
3758  *      When these conditions are not satisfied it drops into a standard 
3759  *      receive procedure patterned after RFC793 to handle all cases.
3760  *      The first three cases are guaranteed by proper pred_flags setting,
3761  *      the rest is checked inline. Fast processing is turned on in 
3762  *      tcp_data_queue when everything is OK.
3763  */
3764 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3765                         struct tcphdr *th, unsigned len)
3766 {
3767         struct tcp_sock *tp = tcp_sk(sk);
3768
3769         /*
3770          *      Header prediction.
3771          *      The code loosely follows the one in the famous 
3772          *      "30 instruction TCP receive" Van Jacobson mail.
3773          *      
3774          *      Van's trick is to deposit buffers into socket queue 
3775          *      on a device interrupt, to call tcp_recv function
3776          *      on the receive process context and checksum and copy
3777          *      the buffer to user space. smart...
3778          *
3779          *      Our current scheme is not silly either but we take the 
3780          *      extra cost of the net_bh soft interrupt processing...
3781          *      We do checksum and copy also but from device to kernel.
3782          */
3783
3784         tp->rx_opt.saw_tstamp = 0;
3785
3786         /*      pred_flags is 0xS?10 << 16 + snd_wnd
3787          *      if header_prediction is to be made
3788          *      'S' will always be tp->tcp_header_len >> 2
3789          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3790          *  turn it off (when there are holes in the receive 
3791          *       space for instance)
3792          *      PSH flag is ignored.
3793          */
3794
3795         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3796                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3797                 int tcp_header_len = tp->tcp_header_len;
3798
3799                 /* Timestamp header prediction: tcp_header_len
3800                  * is automatically equal to th->doff*4 due to pred_flags
3801                  * match.
3802                  */
3803
3804                 /* Check timestamp */
3805                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3806                         __u32 *ptr = (__u32 *)(th + 1);
3807
3808                         /* No? Slow path! */
3809                         if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3810                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3811                                 goto slow_path;
3812
3813                         tp->rx_opt.saw_tstamp = 1;
3814                         ++ptr; 
3815                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3816                         ++ptr;
3817                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3818
3819                         /* If PAWS failed, check it more carefully in slow path */
3820                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
3821                                 goto slow_path;
3822
3823                         /* DO NOT update ts_recent here, if checksum fails
3824                          * and timestamp was corrupted part, it will result
3825                          * in a hung connection since we will drop all
3826                          * future packets due to the PAWS test.
3827                          */
3828                 }
3829
3830                 if (len <= tcp_header_len) {
3831                         /* Bulk data transfer: sender */
3832                         if (len == tcp_header_len) {
3833                                 /* Predicted packet is in window by definition.
3834                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3835                                  * Hence, check seq<=rcv_wup reduces to:
3836                                  */
3837                                 if (tcp_header_len ==
3838                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3839                                     tp->rcv_nxt == tp->rcv_wup)
3840                                         tcp_store_ts_recent(tp);
3841
3842                                 tcp_rcv_rtt_measure_ts(sk, skb);
3843
3844                                 /* We know that such packets are checksummed
3845                                  * on entry.
3846                                  */
3847                                 tcp_ack(sk, skb, 0);
3848                                 __kfree_skb(skb); 
3849                                 tcp_data_snd_check(sk, tp);
3850                                 return 0;
3851                         } else { /* Header too small */
3852                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3853                                 goto discard;
3854                         }
3855                 } else {
3856                         int eaten = 0;
3857
3858                         if (tp->ucopy.task == current &&
3859                             tp->copied_seq == tp->rcv_nxt &&
3860                             len - tcp_header_len <= tp->ucopy.len &&
3861                             sock_owned_by_user(sk)) {
3862                                 __set_current_state(TASK_RUNNING);
3863
3864                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
3865                                         /* Predicted packet is in window by definition.
3866                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3867                                          * Hence, check seq<=rcv_wup reduces to:
3868                                          */
3869                                         if (tcp_header_len ==
3870                                             (sizeof(struct tcphdr) +
3871                                              TCPOLEN_TSTAMP_ALIGNED) &&
3872                                             tp->rcv_nxt == tp->rcv_wup)
3873                                                 tcp_store_ts_recent(tp);
3874
3875                                         tcp_rcv_rtt_measure_ts(sk, skb);
3876
3877                                         __skb_pull(skb, tcp_header_len);
3878                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3879                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
3880                                         eaten = 1;
3881                                 }
3882                         }
3883                         if (!eaten) {
3884                                 if (tcp_checksum_complete_user(sk, skb))
3885                                         goto csum_error;
3886
3887                                 /* Predicted packet is in window by definition.
3888                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3889                                  * Hence, check seq<=rcv_wup reduces to:
3890                                  */
3891                                 if (tcp_header_len ==
3892                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3893                                     tp->rcv_nxt == tp->rcv_wup)
3894                                         tcp_store_ts_recent(tp);
3895
3896                                 tcp_rcv_rtt_measure_ts(sk, skb);
3897
3898                                 if ((int)skb->truesize > sk->sk_forward_alloc)
3899                                         goto step5;
3900
3901                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
3902
3903                                 /* Bulk data transfer: receiver */
3904                                 __skb_pull(skb,tcp_header_len);
3905                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3906                                 sk_stream_set_owner_r(skb, sk);
3907                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3908                         }
3909
3910                         tcp_event_data_recv(sk, tp, skb);
3911
3912                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
3913                                 /* Well, only one small jumplet in fast path... */
3914                                 tcp_ack(sk, skb, FLAG_DATA);
3915                                 tcp_data_snd_check(sk, tp);
3916                                 if (!inet_csk_ack_scheduled(sk))
3917                                         goto no_ack;
3918                         }
3919
3920                         __tcp_ack_snd_check(sk, 0);
3921 no_ack:
3922                         if (eaten)
3923                                 __kfree_skb(skb);
3924                         else
3925                                 sk->sk_data_ready(sk, 0);
3926                         return 0;
3927                 }
3928         }
3929
3930 slow_path:
3931         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
3932                 goto csum_error;
3933
3934         /*
3935          * RFC1323: H1. Apply PAWS check first.
3936          */
3937         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
3938             tcp_paws_discard(sk, skb)) {
3939                 if (!th->rst) {
3940                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
3941                         tcp_send_dupack(sk, skb);
3942                         goto discard;
3943                 }
3944                 /* Resets are accepted even if PAWS failed.
3945
3946                    ts_recent update must be made after we are sure
3947                    that the packet is in window.
3948                  */
3949         }
3950
3951         /*
3952          *      Standard slow path.
3953          */
3954
3955         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3956                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3957                  * (RST) segments are validated by checking their SEQ-fields."
3958                  * And page 69: "If an incoming segment is not acceptable,
3959                  * an acknowledgment should be sent in reply (unless the RST bit
3960                  * is set, if so drop the segment and return)".
3961                  */
3962                 if (!th->rst)
3963                         tcp_send_dupack(sk, skb);
3964                 goto discard;
3965         }
3966
3967         if(th->rst) {
3968                 tcp_reset(sk);
3969                 goto discard;
3970         }
3971
3972         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3973
3974         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3975                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3976                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
3977                 tcp_reset(sk);
3978                 return 1;
3979         }
3980
3981 step5:
3982         if(th->ack)
3983                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3984
3985         tcp_rcv_rtt_measure_ts(sk, skb);
3986
3987         /* Process urgent data. */
3988         tcp_urg(sk, skb, th);
3989
3990         /* step 7: process the segment text */
3991         tcp_data_queue(sk, skb);
3992
3993         tcp_data_snd_check(sk, tp);
3994         tcp_ack_snd_check(sk);
3995         return 0;
3996
3997 csum_error:
3998         TCP_INC_STATS_BH(TCP_MIB_INERRS);
3999
4000 discard:
4001         __kfree_skb(skb);
4002         return 0;
4003 }
4004
4005 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4006                                          struct tcphdr *th, unsigned len)
4007 {
4008         struct tcp_sock *tp = tcp_sk(sk);
4009         struct inet_connection_sock *icsk = inet_csk(sk);
4010         int saved_clamp = tp->rx_opt.mss_clamp;
4011
4012         tcp_parse_options(skb, &tp->rx_opt, 0);
4013
4014         if (th->ack) {
4015                 /* rfc793:
4016                  * "If the state is SYN-SENT then
4017                  *    first check the ACK bit
4018                  *      If the ACK bit is set
4019                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4020                  *        a reset (unless the RST bit is set, if so drop
4021                  *        the segment and return)"
4022                  *
4023                  *  We do not send data with SYN, so that RFC-correct
4024                  *  test reduces to:
4025                  */
4026                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4027                         goto reset_and_undo;
4028
4029                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4030                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4031                              tcp_time_stamp)) {
4032                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4033                         goto reset_and_undo;
4034                 }
4035
4036                 /* Now ACK is acceptable.
4037                  *
4038                  * "If the RST bit is set
4039                  *    If the ACK was acceptable then signal the user "error:
4040                  *    connection reset", drop the segment, enter CLOSED state,
4041                  *    delete TCB, and return."
4042                  */
4043
4044                 if (th->rst) {
4045                         tcp_reset(sk);
4046                         goto discard;
4047                 }
4048
4049                 /* rfc793:
4050                  *   "fifth, if neither of the SYN or RST bits is set then
4051                  *    drop the segment and return."
4052                  *
4053                  *    See note below!
4054                  *                                        --ANK(990513)
4055                  */
4056                 if (!th->syn)
4057                         goto discard_and_undo;
4058
4059                 /* rfc793:
4060                  *   "If the SYN bit is on ...
4061                  *    are acceptable then ...
4062                  *    (our SYN has been ACKed), change the connection
4063                  *    state to ESTABLISHED..."
4064                  */
4065
4066                 TCP_ECN_rcv_synack(tp, th);
4067                 if (tp->ecn_flags&TCP_ECN_OK)
4068                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4069
4070                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4071                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4072
4073                 /* Ok.. it's good. Set up sequence numbers and
4074                  * move to established.
4075                  */
4076                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4077                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4078
4079                 /* RFC1323: The window in SYN & SYN/ACK segments is
4080                  * never scaled.
4081                  */
4082                 tp->snd_wnd = ntohs(th->window);
4083                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4084
4085                 if (!tp->rx_opt.wscale_ok) {
4086                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4087                         tp->window_clamp = min(tp->window_clamp, 65535U);
4088                 }
4089
4090                 if (tp->rx_opt.saw_tstamp) {
4091                         tp->rx_opt.tstamp_ok       = 1;
4092                         tp->tcp_header_len =
4093                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4094                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4095                         tcp_store_ts_recent(tp);
4096                 } else {
4097                         tp->tcp_header_len = sizeof(struct tcphdr);
4098                 }
4099
4100                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4101                         tp->rx_opt.sack_ok |= 2;
4102
4103                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4104                 tcp_initialize_rcv_mss(sk);
4105
4106                 /* Remember, tcp_poll() does not lock socket!
4107                  * Change state from SYN-SENT only after copied_seq
4108                  * is initialized. */
4109                 tp->copied_seq = tp->rcv_nxt;
4110                 mb();
4111                 tcp_set_state(sk, TCP_ESTABLISHED);
4112
4113                 /* Make sure socket is routed, for correct metrics.  */
4114                 icsk->icsk_af_ops->rebuild_header(sk);
4115
4116                 tcp_init_metrics(sk);
4117
4118                 tcp_init_congestion_control(sk);
4119
4120                 /* Prevent spurious tcp_cwnd_restart() on first data
4121                  * packet.
4122                  */
4123                 tp->lsndtime = tcp_time_stamp;
4124
4125                 tcp_init_buffer_space(sk);
4126
4127                 if (sock_flag(sk, SOCK_KEEPOPEN))
4128                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4129
4130                 if (!tp->rx_opt.snd_wscale)
4131                         __tcp_fast_path_on(tp, tp->snd_wnd);
4132                 else
4133                         tp->pred_flags = 0;
4134
4135                 if (!sock_flag(sk, SOCK_DEAD)) {
4136                         sk->sk_state_change(sk);
4137                         sk_wake_async(sk, 0, POLL_OUT);
4138                 }
4139
4140                 if (sk->sk_write_pending ||
4141                     icsk->icsk_accept_queue.rskq_defer_accept ||
4142                     icsk->icsk_ack.pingpong) {
4143                         /* Save one ACK. Data will be ready after
4144                          * several ticks, if write_pending is set.
4145                          *
4146                          * It may be deleted, but with this feature tcpdumps
4147                          * look so _wonderfully_ clever, that I was not able
4148                          * to stand against the temptation 8)     --ANK
4149                          */
4150                         inet_csk_schedule_ack(sk);
4151                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4152                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4153                         tcp_incr_quickack(sk);
4154                         tcp_enter_quickack_mode(sk);
4155                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4156                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4157
4158 discard:
4159                         __kfree_skb(skb);
4160                         return 0;
4161                 } else {
4162                         tcp_send_ack(sk);
4163                 }
4164                 return -1;
4165         }
4166
4167         /* No ACK in the segment */
4168
4169         if (th->rst) {
4170                 /* rfc793:
4171                  * "If the RST bit is set
4172                  *
4173                  *      Otherwise (no ACK) drop the segment and return."
4174                  */
4175
4176                 goto discard_and_undo;
4177         }
4178
4179         /* PAWS check. */
4180         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4181                 goto discard_and_undo;
4182
4183         if (th->syn) {
4184                 /* We see SYN without ACK. It is attempt of
4185                  * simultaneous connect with crossed SYNs.
4186                  * Particularly, it can be connect to self.
4187                  */
4188                 tcp_set_state(sk, TCP_SYN_RECV);
4189
4190                 if (tp->rx_opt.saw_tstamp) {
4191                         tp->rx_opt.tstamp_ok = 1;
4192                         tcp_store_ts_recent(tp);
4193                         tp->tcp_header_len =
4194                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4195                 } else {
4196                         tp->tcp_header_len = sizeof(struct tcphdr);
4197                 }
4198
4199                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4200                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4201
4202                 /* RFC1323: The window in SYN & SYN/ACK segments is
4203                  * never scaled.
4204                  */
4205                 tp->snd_wnd    = ntohs(th->window);
4206                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4207                 tp->max_window = tp->snd_wnd;
4208
4209                 TCP_ECN_rcv_syn(tp, th);
4210                 if (tp->ecn_flags&TCP_ECN_OK)
4211                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4212
4213                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4214                 tcp_initialize_rcv_mss(sk);
4215
4216
4217                 tcp_send_synack(sk);
4218 #if 0
4219                 /* Note, we could accept data and URG from this segment.
4220                  * There are no obstacles to make this.
4221                  *
4222                  * However, if we ignore data in ACKless segments sometimes,
4223                  * we have no reasons to accept it sometimes.
4224                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4225                  * is not flawless. So, discard packet for sanity.
4226                  * Uncomment this return to process the data.
4227                  */
4228                 return -1;
4229 #else
4230                 goto discard;
4231 #endif
4232         }
4233         /* "fifth, if neither of the SYN or RST bits is set then
4234          * drop the segment and return."
4235          */
4236
4237 discard_and_undo:
4238         tcp_clear_options(&tp->rx_opt);
4239         tp->rx_opt.mss_clamp = saved_clamp;
4240         goto discard;
4241
4242 reset_and_undo:
4243         tcp_clear_options(&tp->rx_opt);
4244         tp->rx_opt.mss_clamp = saved_clamp;
4245         return 1;
4246 }
4247
4248
4249 /*
4250  *      This function implements the receiving procedure of RFC 793 for
4251  *      all states except ESTABLISHED and TIME_WAIT. 
4252  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4253  *      address independent.
4254  */
4255         
4256 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4257                           struct tcphdr *th, unsigned len)
4258 {
4259         struct tcp_sock *tp = tcp_sk(sk);
4260         struct inet_connection_sock *icsk = inet_csk(sk);
4261         int queued = 0;
4262
4263         tp->rx_opt.saw_tstamp = 0;
4264
4265         switch (sk->sk_state) {
4266         case TCP_CLOSE:
4267                 goto discard;
4268
4269         case TCP_LISTEN:
4270                 if(th->ack)
4271                         return 1;
4272
4273                 if(th->rst)
4274                         goto discard;
4275
4276                 if(th->syn) {
4277                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4278                                 return 1;
4279
4280                         /* Now we have several options: In theory there is 
4281                          * nothing else in the frame. KA9Q has an option to 
4282                          * send data with the syn, BSD accepts data with the
4283                          * syn up to the [to be] advertised window and 
4284                          * Solaris 2.1 gives you a protocol error. For now 
4285                          * we just ignore it, that fits the spec precisely 
4286                          * and avoids incompatibilities. It would be nice in
4287                          * future to drop through and process the data.
4288                          *
4289                          * Now that TTCP is starting to be used we ought to 
4290                          * queue this data.
4291                          * But, this leaves one open to an easy denial of
4292                          * service attack, and SYN cookies can't defend
4293                          * against this problem. So, we drop the data
4294                          * in the interest of security over speed.
4295                          */
4296                         goto discard;
4297                 }
4298                 goto discard;
4299
4300         case TCP_SYN_SENT:
4301                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4302                 if (queued >= 0)
4303                         return queued;
4304
4305                 /* Do step6 onward by hand. */
4306                 tcp_urg(sk, skb, th);
4307                 __kfree_skb(skb);
4308                 tcp_data_snd_check(sk, tp);
4309                 return 0;
4310         }
4311
4312         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4313             tcp_paws_discard(sk, skb)) {
4314                 if (!th->rst) {
4315                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4316                         tcp_send_dupack(sk, skb);
4317                         goto discard;
4318                 }
4319                 /* Reset is accepted even if it did not pass PAWS. */
4320         }
4321
4322         /* step 1: check sequence number */
4323         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4324                 if (!th->rst)
4325                         tcp_send_dupack(sk, skb);
4326                 goto discard;
4327         }
4328
4329         /* step 2: check RST bit */
4330         if(th->rst) {
4331                 tcp_reset(sk);
4332                 goto discard;
4333         }
4334
4335         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4336
4337         /* step 3: check security and precedence [ignored] */
4338
4339         /*      step 4:
4340          *
4341          *      Check for a SYN in window.
4342          */
4343         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4344                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4345                 tcp_reset(sk);
4346                 return 1;
4347         }
4348
4349         /* step 5: check the ACK field */
4350         if (th->ack) {
4351                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4352
4353                 switch(sk->sk_state) {
4354                 case TCP_SYN_RECV:
4355                         if (acceptable) {
4356                                 tp->copied_seq = tp->rcv_nxt;
4357                                 mb();
4358                                 tcp_set_state(sk, TCP_ESTABLISHED);
4359                                 sk->sk_state_change(sk);
4360
4361                                 /* Note, that this wakeup is only for marginal
4362                                  * crossed SYN case. Passively open sockets
4363                                  * are not waked up, because sk->sk_sleep ==
4364                                  * NULL and sk->sk_socket == NULL.
4365                                  */
4366                                 if (sk->sk_socket) {
4367                                         sk_wake_async(sk,0,POLL_OUT);
4368                                 }
4369
4370                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4371                                 tp->snd_wnd = ntohs(th->window) <<
4372                                               tp->rx_opt.snd_wscale;
4373                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4374                                             TCP_SKB_CB(skb)->seq);
4375
4376                                 /* tcp_ack considers this ACK as duplicate
4377                                  * and does not calculate rtt.
4378                                  * Fix it at least with timestamps.
4379                                  */
4380                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4381                                     !tp->srtt)
4382                                         tcp_ack_saw_tstamp(sk, 0);
4383
4384                                 if (tp->rx_opt.tstamp_ok)
4385                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4386
4387                                 /* Make sure socket is routed, for
4388                                  * correct metrics.
4389                                  */
4390                                 icsk->icsk_af_ops->rebuild_header(sk);
4391
4392                                 tcp_init_metrics(sk);
4393
4394                                 tcp_init_congestion_control(sk);
4395
4396                                 /* Prevent spurious tcp_cwnd_restart() on
4397                                  * first data packet.
4398                                  */
4399                                 tp->lsndtime = tcp_time_stamp;
4400
4401                                 tcp_initialize_rcv_mss(sk);
4402                                 tcp_init_buffer_space(sk);
4403                                 tcp_fast_path_on(tp);
4404                         } else {
4405                                 return 1;
4406                         }
4407                         break;
4408
4409                 case TCP_FIN_WAIT1:
4410                         if (tp->snd_una == tp->write_seq) {
4411                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4412                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4413                                 dst_confirm(sk->sk_dst_cache);
4414
4415                                 if (!sock_flag(sk, SOCK_DEAD))
4416                                         /* Wake up lingering close() */
4417                                         sk->sk_state_change(sk);
4418                                 else {
4419                                         int tmo;
4420
4421                                         if (tp->linger2 < 0 ||
4422                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4423                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4424                                                 tcp_done(sk);
4425                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4426                                                 return 1;
4427                                         }
4428
4429                                         tmo = tcp_fin_time(sk);
4430                                         if (tmo > TCP_TIMEWAIT_LEN) {
4431                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4432                                         } else if (th->fin || sock_owned_by_user(sk)) {
4433                                                 /* Bad case. We could lose such FIN otherwise.
4434                                                  * It is not a big problem, but it looks confusing
4435                                                  * and not so rare event. We still can lose it now,
4436                                                  * if it spins in bh_lock_sock(), but it is really
4437                                                  * marginal case.
4438                                                  */
4439                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4440                                         } else {
4441                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4442                                                 goto discard;
4443                                         }
4444                                 }
4445                         }
4446                         break;
4447
4448                 case TCP_CLOSING:
4449                         if (tp->snd_una == tp->write_seq) {
4450                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4451                                 goto discard;
4452                         }
4453                         break;
4454
4455                 case TCP_LAST_ACK:
4456                         if (tp->snd_una == tp->write_seq) {
4457                                 tcp_update_metrics(sk);
4458                                 tcp_done(sk);
4459                                 goto discard;
4460                         }
4461                         break;
4462                 }
4463         } else
4464                 goto discard;
4465
4466         /* step 6: check the URG bit */
4467         tcp_urg(sk, skb, th);
4468
4469         /* step 7: process the segment text */
4470         switch (sk->sk_state) {
4471         case TCP_CLOSE_WAIT:
4472         case TCP_CLOSING:
4473         case TCP_LAST_ACK:
4474                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4475                         break;
4476         case TCP_FIN_WAIT1:
4477         case TCP_FIN_WAIT2:
4478                 /* RFC 793 says to queue data in these states,
4479                  * RFC 1122 says we MUST send a reset. 
4480                  * BSD 4.4 also does reset.
4481                  */
4482                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4483                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4484                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4485                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4486                                 tcp_reset(sk);
4487                                 return 1;
4488                         }
4489                 }
4490                 /* Fall through */
4491         case TCP_ESTABLISHED: 
4492                 tcp_data_queue(sk, skb);
4493                 queued = 1;
4494                 break;
4495         }
4496
4497         /* tcp_data could move socket to TIME-WAIT */
4498         if (sk->sk_state != TCP_CLOSE) {
4499                 tcp_data_snd_check(sk, tp);
4500                 tcp_ack_snd_check(sk);
4501         }
4502
4503         if (!queued) { 
4504 discard:
4505                 __kfree_skb(skb);
4506         }
4507         return 0;
4508 }
4509
4510 EXPORT_SYMBOL(sysctl_tcp_ecn);
4511 EXPORT_SYMBOL(sysctl_tcp_reordering);
4512 EXPORT_SYMBOL(sysctl_tcp_abc);
4513 EXPORT_SYMBOL(tcp_parse_options);
4514 EXPORT_SYMBOL(tcp_rcv_established);
4515 EXPORT_SYMBOL(tcp_rcv_state_process);
4516 EXPORT_SYMBOL(tcp_initialize_rcv_mss);