[TCP]: Appropriate Byte Count support
[linux-2.6.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presnce of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make 
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks. 
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
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 inline 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 spagetti. 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_full_space(sk)/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 inline 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 correspoding 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 iimediately 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         struct sk_buff *skb;
331         unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
332         int ofo_win = 0;
333
334         icsk->icsk_ack.quick = 0;
335
336         skb_queue_walk(&tp->out_of_order_queue, skb) {
337                 ofo_win += skb->len;
338         }
339
340         /* If overcommit is due to out of order segments,
341          * do not clamp window. Try to expand rcvbuf instead.
342          */
343         if (ofo_win) {
344                 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
345                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
346                     !tcp_memory_pressure &&
347                     atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
348                         sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
349                                             sysctl_tcp_rmem[2]);
350         }
351         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
352                 app_win += ofo_win;
353                 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
354                         app_win >>= 1;
355                 if (app_win > icsk->icsk_ack.rcv_mss)
356                         app_win -= icsk->icsk_ack.rcv_mss;
357                 app_win = max(app_win, 2U*tp->advmss);
358
359                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
360         }
361 }
362
363 /* Receiver "autotuning" code.
364  *
365  * The algorithm for RTT estimation w/o timestamps is based on
366  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
367  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
368  *
369  * More detail on this code can be found at
370  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
371  * though this reference is out of date.  A new paper
372  * is pending.
373  */
374 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
375 {
376         u32 new_sample = tp->rcv_rtt_est.rtt;
377         long m = sample;
378
379         if (m == 0)
380                 m = 1;
381
382         if (new_sample != 0) {
383                 /* If we sample in larger samples in the non-timestamp
384                  * case, we could grossly overestimate the RTT especially
385                  * with chatty applications or bulk transfer apps which
386                  * are stalled on filesystem I/O.
387                  *
388                  * Also, since we are only going for a minimum in the
389                  * non-timestamp case, we do not smoothe things out
390                  * else with timestamps disabled convergance takes too
391                  * long.
392                  */
393                 if (!win_dep) {
394                         m -= (new_sample >> 3);
395                         new_sample += m;
396                 } else if (m < new_sample)
397                         new_sample = m << 3;
398         } else {
399                 /* No previous mesaure. */
400                 new_sample = m << 3;
401         }
402
403         if (tp->rcv_rtt_est.rtt != new_sample)
404                 tp->rcv_rtt_est.rtt = new_sample;
405 }
406
407 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
408 {
409         if (tp->rcv_rtt_est.time == 0)
410                 goto new_measure;
411         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
412                 return;
413         tcp_rcv_rtt_update(tp,
414                            jiffies - tp->rcv_rtt_est.time,
415                            1);
416
417 new_measure:
418         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
419         tp->rcv_rtt_est.time = tcp_time_stamp;
420 }
421
422 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
423 {
424         struct tcp_sock *tp = tcp_sk(sk);
425         if (tp->rx_opt.rcv_tsecr &&
426             (TCP_SKB_CB(skb)->end_seq -
427              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
428                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
429 }
430
431 /*
432  * This function should be called every time data is copied to user space.
433  * It calculates the appropriate TCP receive buffer space.
434  */
435 void tcp_rcv_space_adjust(struct sock *sk)
436 {
437         struct tcp_sock *tp = tcp_sk(sk);
438         int time;
439         int space;
440         
441         if (tp->rcvq_space.time == 0)
442                 goto new_measure;
443         
444         time = tcp_time_stamp - tp->rcvq_space.time;
445         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
446             tp->rcv_rtt_est.rtt == 0)
447                 return;
448         
449         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
450
451         space = max(tp->rcvq_space.space, space);
452
453         if (tp->rcvq_space.space != space) {
454                 int rcvmem;
455
456                 tp->rcvq_space.space = space;
457
458                 if (sysctl_tcp_moderate_rcvbuf) {
459                         int new_clamp = space;
460
461                         /* Receive space grows, normalize in order to
462                          * take into account packet headers and sk_buff
463                          * structure overhead.
464                          */
465                         space /= tp->advmss;
466                         if (!space)
467                                 space = 1;
468                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
469                                   16 + sizeof(struct sk_buff));
470                         while (tcp_win_from_space(rcvmem) < tp->advmss)
471                                 rcvmem += 128;
472                         space *= rcvmem;
473                         space = min(space, sysctl_tcp_rmem[2]);
474                         if (space > sk->sk_rcvbuf) {
475                                 sk->sk_rcvbuf = space;
476
477                                 /* Make the window clamp follow along.  */
478                                 tp->window_clamp = new_clamp;
479                         }
480                 }
481         }
482         
483 new_measure:
484         tp->rcvq_space.seq = tp->copied_seq;
485         tp->rcvq_space.time = tcp_time_stamp;
486 }
487
488 /* There is something which you must keep in mind when you analyze the
489  * behavior of the tp->ato delayed ack timeout interval.  When a
490  * connection starts up, we want to ack as quickly as possible.  The
491  * problem is that "good" TCP's do slow start at the beginning of data
492  * transmission.  The means that until we send the first few ACK's the
493  * sender will sit on his end and only queue most of his data, because
494  * he can only send snd_cwnd unacked packets at any given time.  For
495  * each ACK we send, he increments snd_cwnd and transmits more of his
496  * queue.  -DaveM
497  */
498 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
499 {
500         struct inet_connection_sock *icsk = inet_csk(sk);
501         u32 now;
502
503         inet_csk_schedule_ack(sk);
504
505         tcp_measure_rcv_mss(sk, skb);
506
507         tcp_rcv_rtt_measure(tp);
508         
509         now = tcp_time_stamp;
510
511         if (!icsk->icsk_ack.ato) {
512                 /* The _first_ data packet received, initialize
513                  * delayed ACK engine.
514                  */
515                 tcp_incr_quickack(sk);
516                 icsk->icsk_ack.ato = TCP_ATO_MIN;
517         } else {
518                 int m = now - icsk->icsk_ack.lrcvtime;
519
520                 if (m <= TCP_ATO_MIN/2) {
521                         /* The fastest case is the first. */
522                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
523                 } else if (m < icsk->icsk_ack.ato) {
524                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
525                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
526                                 icsk->icsk_ack.ato = icsk->icsk_rto;
527                 } else if (m > icsk->icsk_rto) {
528                         /* Too long gap. Apparently sender falled to
529                          * restart window, so that we send ACKs quickly.
530                          */
531                         tcp_incr_quickack(sk);
532                         sk_stream_mem_reclaim(sk);
533                 }
534         }
535         icsk->icsk_ack.lrcvtime = now;
536
537         TCP_ECN_check_ce(tp, skb);
538
539         if (skb->len >= 128)
540                 tcp_grow_window(sk, tp, skb);
541 }
542
543 /* Called to compute a smoothed rtt estimate. The data fed to this
544  * routine either comes from timestamps, or from segments that were
545  * known _not_ to have been retransmitted [see Karn/Partridge
546  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
547  * piece by Van Jacobson.
548  * NOTE: the next three routines used to be one big routine.
549  * To save cycles in the RFC 1323 implementation it was better to break
550  * it up into three procedures. -- erics
551  */
552 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
553 {
554         struct tcp_sock *tp = tcp_sk(sk);
555         long m = mrtt; /* RTT */
556
557         /*      The following amusing code comes from Jacobson's
558          *      article in SIGCOMM '88.  Note that rtt and mdev
559          *      are scaled versions of rtt and mean deviation.
560          *      This is designed to be as fast as possible 
561          *      m stands for "measurement".
562          *
563          *      On a 1990 paper the rto value is changed to:
564          *      RTO = rtt + 4 * mdev
565          *
566          * Funny. This algorithm seems to be very broken.
567          * These formulae increase RTO, when it should be decreased, increase
568          * too slowly, when it should be incresed fastly, decrease too fastly
569          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
570          * does not matter how to _calculate_ it. Seems, it was trap
571          * that VJ failed to avoid. 8)
572          */
573         if(m == 0)
574                 m = 1;
575         if (tp->srtt != 0) {
576                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
577                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
578                 if (m < 0) {
579                         m = -m;         /* m is now abs(error) */
580                         m -= (tp->mdev >> 2);   /* similar update on mdev */
581                         /* This is similar to one of Eifel findings.
582                          * Eifel blocks mdev updates when rtt decreases.
583                          * This solution is a bit different: we use finer gain
584                          * for mdev in this case (alpha*beta).
585                          * Like Eifel it also prevents growth of rto,
586                          * but also it limits too fast rto decreases,
587                          * happening in pure Eifel.
588                          */
589                         if (m > 0)
590                                 m >>= 3;
591                 } else {
592                         m -= (tp->mdev >> 2);   /* similar update on mdev */
593                 }
594                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
595                 if (tp->mdev > tp->mdev_max) {
596                         tp->mdev_max = tp->mdev;
597                         if (tp->mdev_max > tp->rttvar)
598                                 tp->rttvar = tp->mdev_max;
599                 }
600                 if (after(tp->snd_una, tp->rtt_seq)) {
601                         if (tp->mdev_max < tp->rttvar)
602                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
603                         tp->rtt_seq = tp->snd_nxt;
604                         tp->mdev_max = TCP_RTO_MIN;
605                 }
606         } else {
607                 /* no previous measure. */
608                 tp->srtt = m<<3;        /* take the measured time to be rtt */
609                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
610                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
611                 tp->rtt_seq = tp->snd_nxt;
612         }
613 }
614
615 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
616  * routine referred to above.
617  */
618 static inline void tcp_set_rto(struct sock *sk)
619 {
620         const struct tcp_sock *tp = tcp_sk(sk);
621         /* Old crap is replaced with new one. 8)
622          *
623          * More seriously:
624          * 1. If rtt variance happened to be less 50msec, it is hallucination.
625          *    It cannot be less due to utterly erratic ACK generation made
626          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
627          *    to do with delayed acks, because at cwnd>2 true delack timeout
628          *    is invisible. Actually, Linux-2.4 also generates erratic
629          *    ACKs in some curcumstances.
630          */
631         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
632
633         /* 2. Fixups made earlier cannot be right.
634          *    If we do not estimate RTO correctly without them,
635          *    all the algo is pure shit and should be replaced
636          *    with correct one. It is exaclty, which we pretend to do.
637          */
638 }
639
640 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
641  * guarantees that rto is higher.
642  */
643 static inline void tcp_bound_rto(struct sock *sk)
644 {
645         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
646                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
647 }
648
649 /* Save metrics learned by this TCP session.
650    This function is called only, when TCP finishes successfully
651    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
652  */
653 void tcp_update_metrics(struct sock *sk)
654 {
655         struct tcp_sock *tp = tcp_sk(sk);
656         struct dst_entry *dst = __sk_dst_get(sk);
657
658         if (sysctl_tcp_nometrics_save)
659                 return;
660
661         dst_confirm(dst);
662
663         if (dst && (dst->flags&DST_HOST)) {
664                 const struct inet_connection_sock *icsk = inet_csk(sk);
665                 int m;
666
667                 if (icsk->icsk_backoff || !tp->srtt) {
668                         /* This session failed to estimate rtt. Why?
669                          * Probably, no packets returned in time.
670                          * Reset our results.
671                          */
672                         if (!(dst_metric_locked(dst, RTAX_RTT)))
673                                 dst->metrics[RTAX_RTT-1] = 0;
674                         return;
675                 }
676
677                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
678
679                 /* If newly calculated rtt larger than stored one,
680                  * store new one. Otherwise, use EWMA. Remember,
681                  * rtt overestimation is always better than underestimation.
682                  */
683                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
684                         if (m <= 0)
685                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
686                         else
687                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
688                 }
689
690                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
691                         if (m < 0)
692                                 m = -m;
693
694                         /* Scale deviation to rttvar fixed point */
695                         m >>= 1;
696                         if (m < tp->mdev)
697                                 m = tp->mdev;
698
699                         if (m >= dst_metric(dst, RTAX_RTTVAR))
700                                 dst->metrics[RTAX_RTTVAR-1] = m;
701                         else
702                                 dst->metrics[RTAX_RTTVAR-1] -=
703                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
704                 }
705
706                 if (tp->snd_ssthresh >= 0xFFFF) {
707                         /* Slow start still did not finish. */
708                         if (dst_metric(dst, RTAX_SSTHRESH) &&
709                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
710                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
711                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
712                         if (!dst_metric_locked(dst, RTAX_CWND) &&
713                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
714                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
715                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
716                            icsk->icsk_ca_state == TCP_CA_Open) {
717                         /* Cong. avoidance phase, cwnd is reliable. */
718                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
719                                 dst->metrics[RTAX_SSTHRESH-1] =
720                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
721                         if (!dst_metric_locked(dst, RTAX_CWND))
722                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
723                 } else {
724                         /* Else slow start did not finish, cwnd is non-sense,
725                            ssthresh may be also invalid.
726                          */
727                         if (!dst_metric_locked(dst, RTAX_CWND))
728                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
729                         if (dst->metrics[RTAX_SSTHRESH-1] &&
730                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
731                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
732                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
733                 }
734
735                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
736                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
737                             tp->reordering != sysctl_tcp_reordering)
738                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
739                 }
740         }
741 }
742
743 /* Numbers are taken from RFC2414.  */
744 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
745 {
746         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
747
748         if (!cwnd) {
749                 if (tp->mss_cache > 1460)
750                         cwnd = 2;
751                 else
752                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
753         }
754         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
755 }
756
757 /* Initialize metrics on socket. */
758
759 static void tcp_init_metrics(struct sock *sk)
760 {
761         struct tcp_sock *tp = tcp_sk(sk);
762         struct dst_entry *dst = __sk_dst_get(sk);
763
764         if (dst == NULL)
765                 goto reset;
766
767         dst_confirm(dst);
768
769         if (dst_metric_locked(dst, RTAX_CWND))
770                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
771         if (dst_metric(dst, RTAX_SSTHRESH)) {
772                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
773                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
774                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
775         }
776         if (dst_metric(dst, RTAX_REORDERING) &&
777             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
778                 tp->rx_opt.sack_ok &= ~2;
779                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
780         }
781
782         if (dst_metric(dst, RTAX_RTT) == 0)
783                 goto reset;
784
785         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
786                 goto reset;
787
788         /* Initial rtt is determined from SYN,SYN-ACK.
789          * The segment is small and rtt may appear much
790          * less than real one. Use per-dst memory
791          * to make it more realistic.
792          *
793          * A bit of theory. RTT is time passed after "normal" sized packet
794          * is sent until it is ACKed. In normal curcumstances sending small
795          * packets force peer to delay ACKs and calculation is correct too.
796          * The algorithm is adaptive and, provided we follow specs, it
797          * NEVER underestimate RTT. BUT! If peer tries to make some clever
798          * tricks sort of "quick acks" for time long enough to decrease RTT
799          * to low value, and then abruptly stops to do it and starts to delay
800          * ACKs, wait for troubles.
801          */
802         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
803                 tp->srtt = dst_metric(dst, RTAX_RTT);
804                 tp->rtt_seq = tp->snd_nxt;
805         }
806         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
807                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
808                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
809         }
810         tcp_set_rto(sk);
811         tcp_bound_rto(sk);
812         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
813                 goto reset;
814         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
815         tp->snd_cwnd_stamp = tcp_time_stamp;
816         return;
817
818 reset:
819         /* Play conservative. If timestamps are not
820          * supported, TCP will fail to recalculate correct
821          * rtt, if initial rto is too small. FORGET ALL AND RESET!
822          */
823         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
824                 tp->srtt = 0;
825                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
826                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
827         }
828 }
829
830 static void tcp_update_reordering(struct sock *sk, const int metric,
831                                   const int ts)
832 {
833         struct tcp_sock *tp = tcp_sk(sk);
834         if (metric > tp->reordering) {
835                 tp->reordering = min(TCP_MAX_REORDERING, metric);
836
837                 /* This exciting event is worth to be remembered. 8) */
838                 if (ts)
839                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
840                 else if (IsReno(tp))
841                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
842                 else if (IsFack(tp))
843                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
844                 else
845                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
846 #if FASTRETRANS_DEBUG > 1
847                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
848                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
849                        tp->reordering,
850                        tp->fackets_out,
851                        tp->sacked_out,
852                        tp->undo_marker ? tp->undo_retrans : 0);
853 #endif
854                 /* Disable FACK yet. */
855                 tp->rx_opt.sack_ok &= ~2;
856         }
857 }
858
859 /* This procedure tags the retransmission queue when SACKs arrive.
860  *
861  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
862  * Packets in queue with these bits set are counted in variables
863  * sacked_out, retrans_out and lost_out, correspondingly.
864  *
865  * Valid combinations are:
866  * Tag  InFlight        Description
867  * 0    1               - orig segment is in flight.
868  * S    0               - nothing flies, orig reached receiver.
869  * L    0               - nothing flies, orig lost by net.
870  * R    2               - both orig and retransmit are in flight.
871  * L|R  1               - orig is lost, retransmit is in flight.
872  * S|R  1               - orig reached receiver, retrans is still in flight.
873  * (L|S|R is logically valid, it could occur when L|R is sacked,
874  *  but it is equivalent to plain S and code short-curcuits it to S.
875  *  L|S is logically invalid, it would mean -1 packet in flight 8))
876  *
877  * These 6 states form finite state machine, controlled by the following events:
878  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
879  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
880  * 3. Loss detection event of one of three flavors:
881  *      A. Scoreboard estimator decided the packet is lost.
882  *         A'. Reno "three dupacks" marks head of queue lost.
883  *         A''. Its FACK modfication, head until snd.fack is lost.
884  *      B. SACK arrives sacking data transmitted after never retransmitted
885  *         hole was sent out.
886  *      C. SACK arrives sacking SND.NXT at the moment, when the
887  *         segment was retransmitted.
888  * 4. D-SACK added new rule: D-SACK changes any tag to S.
889  *
890  * It is pleasant to note, that state diagram turns out to be commutative,
891  * so that we are allowed not to be bothered by order of our actions,
892  * when multiple events arrive simultaneously. (see the function below).
893  *
894  * Reordering detection.
895  * --------------------
896  * Reordering metric is maximal distance, which a packet can be displaced
897  * in packet stream. With SACKs we can estimate it:
898  *
899  * 1. SACK fills old hole and the corresponding segment was not
900  *    ever retransmitted -> reordering. Alas, we cannot use it
901  *    when segment was retransmitted.
902  * 2. The last flaw is solved with D-SACK. D-SACK arrives
903  *    for retransmitted and already SACKed segment -> reordering..
904  * Both of these heuristics are not used in Loss state, when we cannot
905  * account for retransmits accurately.
906  */
907 static int
908 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
909 {
910         const struct inet_connection_sock *icsk = inet_csk(sk);
911         struct tcp_sock *tp = tcp_sk(sk);
912         unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
913         struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
914         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
915         int reord = tp->packets_out;
916         int prior_fackets;
917         u32 lost_retrans = 0;
918         int flag = 0;
919         int i;
920
921         if (!tp->sacked_out)
922                 tp->fackets_out = 0;
923         prior_fackets = tp->fackets_out;
924
925         for (i=0; i<num_sacks; i++, sp++) {
926                 struct sk_buff *skb;
927                 __u32 start_seq = ntohl(sp->start_seq);
928                 __u32 end_seq = ntohl(sp->end_seq);
929                 int fack_count = 0;
930                 int dup_sack = 0;
931
932                 /* Check for D-SACK. */
933                 if (i == 0) {
934                         u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
935
936                         if (before(start_seq, ack)) {
937                                 dup_sack = 1;
938                                 tp->rx_opt.sack_ok |= 4;
939                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
940                         } else if (num_sacks > 1 &&
941                                    !after(end_seq, ntohl(sp[1].end_seq)) &&
942                                    !before(start_seq, ntohl(sp[1].start_seq))) {
943                                 dup_sack = 1;
944                                 tp->rx_opt.sack_ok |= 4;
945                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
946                         }
947
948                         /* D-SACK for already forgotten data...
949                          * Do dumb counting. */
950                         if (dup_sack &&
951                             !after(end_seq, prior_snd_una) &&
952                             after(end_seq, tp->undo_marker))
953                                 tp->undo_retrans--;
954
955                         /* Eliminate too old ACKs, but take into
956                          * account more or less fresh ones, they can
957                          * contain valid SACK info.
958                          */
959                         if (before(ack, prior_snd_una - tp->max_window))
960                                 return 0;
961                 }
962
963                 /* Event "B" in the comment above. */
964                 if (after(end_seq, tp->high_seq))
965                         flag |= FLAG_DATA_LOST;
966
967                 sk_stream_for_retrans_queue(skb, sk) {
968                         int in_sack, pcount;
969                         u8 sacked;
970
971                         /* The retransmission queue is always in order, so
972                          * we can short-circuit the walk early.
973                          */
974                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
975                                 break;
976
977                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
978                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
979
980                         pcount = tcp_skb_pcount(skb);
981
982                         if (pcount > 1 && !in_sack &&
983                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
984                                 unsigned int pkt_len;
985
986                                 in_sack = !after(start_seq,
987                                                  TCP_SKB_CB(skb)->seq);
988
989                                 if (!in_sack)
990                                         pkt_len = (start_seq -
991                                                    TCP_SKB_CB(skb)->seq);
992                                 else
993                                         pkt_len = (end_seq -
994                                                    TCP_SKB_CB(skb)->seq);
995                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->tso_size))
996                                         break;
997                                 pcount = tcp_skb_pcount(skb);
998                         }
999
1000                         fack_count += pcount;
1001
1002                         sacked = TCP_SKB_CB(skb)->sacked;
1003
1004                         /* Account D-SACK for retransmitted packet. */
1005                         if ((dup_sack && in_sack) &&
1006                             (sacked & TCPCB_RETRANS) &&
1007                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1008                                 tp->undo_retrans--;
1009
1010                         /* The frame is ACKed. */
1011                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1012                                 if (sacked&TCPCB_RETRANS) {
1013                                         if ((dup_sack && in_sack) &&
1014                                             (sacked&TCPCB_SACKED_ACKED))
1015                                                 reord = min(fack_count, reord);
1016                                 } else {
1017                                         /* If it was in a hole, we detected reordering. */
1018                                         if (fack_count < prior_fackets &&
1019                                             !(sacked&TCPCB_SACKED_ACKED))
1020                                                 reord = min(fack_count, reord);
1021                                 }
1022
1023                                 /* Nothing to do; acked frame is about to be dropped. */
1024                                 continue;
1025                         }
1026
1027                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1028                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1029                             (!lost_retrans || after(end_seq, lost_retrans)))
1030                                 lost_retrans = end_seq;
1031
1032                         if (!in_sack)
1033                                 continue;
1034
1035                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1036                                 if (sacked & TCPCB_SACKED_RETRANS) {
1037                                         /* If the segment is not tagged as lost,
1038                                          * we do not clear RETRANS, believing
1039                                          * that retransmission is still in flight.
1040                                          */
1041                                         if (sacked & TCPCB_LOST) {
1042                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1043                                                 tp->lost_out -= tcp_skb_pcount(skb);
1044                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1045                                         }
1046                                 } else {
1047                                         /* New sack for not retransmitted frame,
1048                                          * which was in hole. It is reordering.
1049                                          */
1050                                         if (!(sacked & TCPCB_RETRANS) &&
1051                                             fack_count < prior_fackets)
1052                                                 reord = min(fack_count, reord);
1053
1054                                         if (sacked & TCPCB_LOST) {
1055                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1056                                                 tp->lost_out -= tcp_skb_pcount(skb);
1057                                         }
1058                                 }
1059
1060                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1061                                 flag |= FLAG_DATA_SACKED;
1062                                 tp->sacked_out += tcp_skb_pcount(skb);
1063
1064                                 if (fack_count > tp->fackets_out)
1065                                         tp->fackets_out = fack_count;
1066                         } else {
1067                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1068                                         reord = min(fack_count, reord);
1069                         }
1070
1071                         /* D-SACK. We can detect redundant retransmission
1072                          * in S|R and plain R frames and clear it.
1073                          * undo_retrans is decreased above, L|R frames
1074                          * are accounted above as well.
1075                          */
1076                         if (dup_sack &&
1077                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1078                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1079                                 tp->retrans_out -= tcp_skb_pcount(skb);
1080                         }
1081                 }
1082         }
1083
1084         /* Check for lost retransmit. This superb idea is
1085          * borrowed from "ratehalving". Event "C".
1086          * Later note: FACK people cheated me again 8),
1087          * we have to account for reordering! Ugly,
1088          * but should help.
1089          */
1090         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1091                 struct sk_buff *skb;
1092
1093                 sk_stream_for_retrans_queue(skb, sk) {
1094                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1095                                 break;
1096                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1097                                 continue;
1098                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1099                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1100                             (IsFack(tp) ||
1101                              !before(lost_retrans,
1102                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1103                                      tp->mss_cache))) {
1104                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1105                                 tp->retrans_out -= tcp_skb_pcount(skb);
1106
1107                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1108                                         tp->lost_out += tcp_skb_pcount(skb);
1109                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1110                                         flag |= FLAG_DATA_SACKED;
1111                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1112                                 }
1113                         }
1114                 }
1115         }
1116
1117         tp->left_out = tp->sacked_out + tp->lost_out;
1118
1119         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
1120                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1121
1122 #if FASTRETRANS_DEBUG > 0
1123         BUG_TRAP((int)tp->sacked_out >= 0);
1124         BUG_TRAP((int)tp->lost_out >= 0);
1125         BUG_TRAP((int)tp->retrans_out >= 0);
1126         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1127 #endif
1128         return flag;
1129 }
1130
1131 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1132  * segments to see from the next ACKs whether any data was really missing.
1133  * If the RTO was spurious, new ACKs should arrive.
1134  */
1135 void tcp_enter_frto(struct sock *sk)
1136 {
1137         const struct inet_connection_sock *icsk = inet_csk(sk);
1138         struct tcp_sock *tp = tcp_sk(sk);
1139         struct sk_buff *skb;
1140
1141         tp->frto_counter = 1;
1142
1143         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1144             tp->snd_una == tp->high_seq ||
1145             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1146                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1147                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1148                 tcp_ca_event(sk, CA_EVENT_FRTO);
1149         }
1150
1151         /* Have to clear retransmission markers here to keep the bookkeeping
1152          * in shape, even though we are not yet in Loss state.
1153          * If something was really lost, it is eventually caught up
1154          * in tcp_enter_frto_loss.
1155          */
1156         tp->retrans_out = 0;
1157         tp->undo_marker = tp->snd_una;
1158         tp->undo_retrans = 0;
1159
1160         sk_stream_for_retrans_queue(skb, sk) {
1161                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1162         }
1163         tcp_sync_left_out(tp);
1164
1165         tcp_set_ca_state(sk, TCP_CA_Open);
1166         tp->frto_highmark = tp->snd_nxt;
1167 }
1168
1169 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1170  * which indicates that we should follow the traditional RTO recovery,
1171  * i.e. mark everything lost and do go-back-N retransmission.
1172  */
1173 static void tcp_enter_frto_loss(struct sock *sk)
1174 {
1175         struct tcp_sock *tp = tcp_sk(sk);
1176         struct sk_buff *skb;
1177         int cnt = 0;
1178
1179         tp->sacked_out = 0;
1180         tp->lost_out = 0;
1181         tp->fackets_out = 0;
1182
1183         sk_stream_for_retrans_queue(skb, sk) {
1184                 cnt += tcp_skb_pcount(skb);
1185                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1186                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1187
1188                         /* Do not mark those segments lost that were
1189                          * forward transmitted after RTO
1190                          */
1191                         if (!after(TCP_SKB_CB(skb)->end_seq,
1192                                    tp->frto_highmark)) {
1193                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1194                                 tp->lost_out += tcp_skb_pcount(skb);
1195                         }
1196                 } else {
1197                         tp->sacked_out += tcp_skb_pcount(skb);
1198                         tp->fackets_out = cnt;
1199                 }
1200         }
1201         tcp_sync_left_out(tp);
1202
1203         tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1204         tp->snd_cwnd_cnt = 0;
1205         tp->snd_cwnd_stamp = tcp_time_stamp;
1206         tp->undo_marker = 0;
1207         tp->frto_counter = 0;
1208
1209         tp->reordering = min_t(unsigned int, tp->reordering,
1210                                              sysctl_tcp_reordering);
1211         tcp_set_ca_state(sk, TCP_CA_Loss);
1212         tp->high_seq = tp->frto_highmark;
1213         TCP_ECN_queue_cwr(tp);
1214 }
1215
1216 void tcp_clear_retrans(struct tcp_sock *tp)
1217 {
1218         tp->left_out = 0;
1219         tp->retrans_out = 0;
1220
1221         tp->fackets_out = 0;
1222         tp->sacked_out = 0;
1223         tp->lost_out = 0;
1224
1225         tp->undo_marker = 0;
1226         tp->undo_retrans = 0;
1227 }
1228
1229 /* Enter Loss state. If "how" is not zero, forget all SACK information
1230  * and reset tags completely, otherwise preserve SACKs. If receiver
1231  * dropped its ofo queue, we will know this due to reneging detection.
1232  */
1233 void tcp_enter_loss(struct sock *sk, int how)
1234 {
1235         const struct inet_connection_sock *icsk = inet_csk(sk);
1236         struct tcp_sock *tp = tcp_sk(sk);
1237         struct sk_buff *skb;
1238         int cnt = 0;
1239
1240         /* Reduce ssthresh if it has not yet been made inside this window. */
1241         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1242             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1243                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1244                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1245                 tcp_ca_event(sk, CA_EVENT_LOSS);
1246         }
1247         tp->snd_cwnd       = 1;
1248         tp->snd_cwnd_cnt   = 0;
1249         tp->snd_cwnd_stamp = tcp_time_stamp;
1250
1251         tp->bytes_acked = 0;
1252         tcp_clear_retrans(tp);
1253
1254         /* Push undo marker, if it was plain RTO and nothing
1255          * was retransmitted. */
1256         if (!how)
1257                 tp->undo_marker = tp->snd_una;
1258
1259         sk_stream_for_retrans_queue(skb, sk) {
1260                 cnt += tcp_skb_pcount(skb);
1261                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1262                         tp->undo_marker = 0;
1263                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1264                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1265                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1266                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1267                         tp->lost_out += tcp_skb_pcount(skb);
1268                 } else {
1269                         tp->sacked_out += tcp_skb_pcount(skb);
1270                         tp->fackets_out = cnt;
1271                 }
1272         }
1273         tcp_sync_left_out(tp);
1274
1275         tp->reordering = min_t(unsigned int, tp->reordering,
1276                                              sysctl_tcp_reordering);
1277         tcp_set_ca_state(sk, TCP_CA_Loss);
1278         tp->high_seq = tp->snd_nxt;
1279         TCP_ECN_queue_cwr(tp);
1280 }
1281
1282 static int tcp_check_sack_reneging(struct sock *sk)
1283 {
1284         struct sk_buff *skb;
1285
1286         /* If ACK arrived pointing to a remembered SACK,
1287          * it means that our remembered SACKs do not reflect
1288          * real state of receiver i.e.
1289          * receiver _host_ is heavily congested (or buggy).
1290          * Do processing similar to RTO timeout.
1291          */
1292         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1293             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1294                 struct inet_connection_sock *icsk = inet_csk(sk);
1295                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1296
1297                 tcp_enter_loss(sk, 1);
1298                 icsk->icsk_retransmits++;
1299                 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1300                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1301                                           icsk->icsk_rto, TCP_RTO_MAX);
1302                 return 1;
1303         }
1304         return 0;
1305 }
1306
1307 static inline int tcp_fackets_out(struct tcp_sock *tp)
1308 {
1309         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1310 }
1311
1312 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1313 {
1314         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1315 }
1316
1317 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1318 {
1319         return tp->packets_out &&
1320                tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
1321 }
1322
1323 /* Linux NewReno/SACK/FACK/ECN state machine.
1324  * --------------------------------------
1325  *
1326  * "Open"       Normal state, no dubious events, fast path.
1327  * "Disorder"   In all the respects it is "Open",
1328  *              but requires a bit more attention. It is entered when
1329  *              we see some SACKs or dupacks. It is split of "Open"
1330  *              mainly to move some processing from fast path to slow one.
1331  * "CWR"        CWND was reduced due to some Congestion Notification event.
1332  *              It can be ECN, ICMP source quench, local device congestion.
1333  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1334  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1335  *
1336  * tcp_fastretrans_alert() is entered:
1337  * - each incoming ACK, if state is not "Open"
1338  * - when arrived ACK is unusual, namely:
1339  *      * SACK
1340  *      * Duplicate ACK.
1341  *      * ECN ECE.
1342  *
1343  * Counting packets in flight is pretty simple.
1344  *
1345  *      in_flight = packets_out - left_out + retrans_out
1346  *
1347  *      packets_out is SND.NXT-SND.UNA counted in packets.
1348  *
1349  *      retrans_out is number of retransmitted segments.
1350  *
1351  *      left_out is number of segments left network, but not ACKed yet.
1352  *
1353  *              left_out = sacked_out + lost_out
1354  *
1355  *     sacked_out: Packets, which arrived to receiver out of order
1356  *                 and hence not ACKed. With SACKs this number is simply
1357  *                 amount of SACKed data. Even without SACKs
1358  *                 it is easy to give pretty reliable estimate of this number,
1359  *                 counting duplicate ACKs.
1360  *
1361  *       lost_out: Packets lost by network. TCP has no explicit
1362  *                 "loss notification" feedback from network (for now).
1363  *                 It means that this number can be only _guessed_.
1364  *                 Actually, it is the heuristics to predict lossage that
1365  *                 distinguishes different algorithms.
1366  *
1367  *      F.e. after RTO, when all the queue is considered as lost,
1368  *      lost_out = packets_out and in_flight = retrans_out.
1369  *
1370  *              Essentially, we have now two algorithms counting
1371  *              lost packets.
1372  *
1373  *              FACK: It is the simplest heuristics. As soon as we decided
1374  *              that something is lost, we decide that _all_ not SACKed
1375  *              packets until the most forward SACK are lost. I.e.
1376  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1377  *              It is absolutely correct estimate, if network does not reorder
1378  *              packets. And it loses any connection to reality when reordering
1379  *              takes place. We use FACK by default until reordering
1380  *              is suspected on the path to this destination.
1381  *
1382  *              NewReno: when Recovery is entered, we assume that one segment
1383  *              is lost (classic Reno). While we are in Recovery and
1384  *              a partial ACK arrives, we assume that one more packet
1385  *              is lost (NewReno). This heuristics are the same in NewReno
1386  *              and SACK.
1387  *
1388  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1389  *  deflation etc. CWND is real congestion window, never inflated, changes
1390  *  only according to classic VJ rules.
1391  *
1392  * Really tricky (and requiring careful tuning) part of algorithm
1393  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1394  * The first determines the moment _when_ we should reduce CWND and,
1395  * hence, slow down forward transmission. In fact, it determines the moment
1396  * when we decide that hole is caused by loss, rather than by a reorder.
1397  *
1398  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1399  * holes, caused by lost packets.
1400  *
1401  * And the most logically complicated part of algorithm is undo
1402  * heuristics. We detect false retransmits due to both too early
1403  * fast retransmit (reordering) and underestimated RTO, analyzing
1404  * timestamps and D-SACKs. When we detect that some segments were
1405  * retransmitted by mistake and CWND reduction was wrong, we undo
1406  * window reduction and abort recovery phase. This logic is hidden
1407  * inside several functions named tcp_try_undo_<something>.
1408  */
1409
1410 /* This function decides, when we should leave Disordered state
1411  * and enter Recovery phase, reducing congestion window.
1412  *
1413  * Main question: may we further continue forward transmission
1414  * with the same cwnd?
1415  */
1416 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1417 {
1418         __u32 packets_out;
1419
1420         /* Trick#1: The loss is proven. */
1421         if (tp->lost_out)
1422                 return 1;
1423
1424         /* Not-A-Trick#2 : Classic rule... */
1425         if (tcp_fackets_out(tp) > tp->reordering)
1426                 return 1;
1427
1428         /* Trick#3 : when we use RFC2988 timer restart, fast
1429          * retransmit can be triggered by timeout of queue head.
1430          */
1431         if (tcp_head_timedout(sk, tp))
1432                 return 1;
1433
1434         /* Trick#4: It is still not OK... But will it be useful to delay
1435          * recovery more?
1436          */
1437         packets_out = tp->packets_out;
1438         if (packets_out <= tp->reordering &&
1439             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1440             !tcp_may_send_now(sk, tp)) {
1441                 /* We have nothing to send. This connection is limited
1442                  * either by receiver window or by application.
1443                  */
1444                 return 1;
1445         }
1446
1447         return 0;
1448 }
1449
1450 /* If we receive more dupacks than we expected counting segments
1451  * in assumption of absent reordering, interpret this as reordering.
1452  * The only another reason could be bug in receiver TCP.
1453  */
1454 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1455 {
1456         struct tcp_sock *tp = tcp_sk(sk);
1457         u32 holes;
1458
1459         holes = max(tp->lost_out, 1U);
1460         holes = min(holes, tp->packets_out);
1461
1462         if ((tp->sacked_out + holes) > tp->packets_out) {
1463                 tp->sacked_out = tp->packets_out - holes;
1464                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1465         }
1466 }
1467
1468 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1469
1470 static void tcp_add_reno_sack(struct sock *sk)
1471 {
1472         struct tcp_sock *tp = tcp_sk(sk);
1473         tp->sacked_out++;
1474         tcp_check_reno_reordering(sk, 0);
1475         tcp_sync_left_out(tp);
1476 }
1477
1478 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1479
1480 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1481 {
1482         if (acked > 0) {
1483                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1484                 if (acked-1 >= tp->sacked_out)
1485                         tp->sacked_out = 0;
1486                 else
1487                         tp->sacked_out -= acked-1;
1488         }
1489         tcp_check_reno_reordering(sk, acked);
1490         tcp_sync_left_out(tp);
1491 }
1492
1493 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1494 {
1495         tp->sacked_out = 0;
1496         tp->left_out = tp->lost_out;
1497 }
1498
1499 /* Mark head of queue up as lost. */
1500 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1501                                int packets, u32 high_seq)
1502 {
1503         struct sk_buff *skb;
1504         int cnt = packets;
1505
1506         BUG_TRAP(cnt <= tp->packets_out);
1507
1508         sk_stream_for_retrans_queue(skb, sk) {
1509                 cnt -= tcp_skb_pcount(skb);
1510                 if (cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1511                         break;
1512                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1513                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1514                         tp->lost_out += tcp_skb_pcount(skb);
1515                 }
1516         }
1517         tcp_sync_left_out(tp);
1518 }
1519
1520 /* Account newly detected lost packet(s) */
1521
1522 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1523 {
1524         if (IsFack(tp)) {
1525                 int lost = tp->fackets_out - tp->reordering;
1526                 if (lost <= 0)
1527                         lost = 1;
1528                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1529         } else {
1530                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1531         }
1532
1533         /* New heuristics: it is possible only after we switched
1534          * to restart timer each time when something is ACKed.
1535          * Hence, we can detect timed out packets during fast
1536          * retransmit without falling to slow start.
1537          */
1538         if (tcp_head_timedout(sk, tp)) {
1539                 struct sk_buff *skb;
1540
1541                 sk_stream_for_retrans_queue(skb, sk) {
1542                         if (tcp_skb_timedout(sk, skb) &&
1543                             !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1544                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1545                                 tp->lost_out += tcp_skb_pcount(skb);
1546                         }
1547                 }
1548                 tcp_sync_left_out(tp);
1549         }
1550 }
1551
1552 /* CWND moderation, preventing bursts due to too big ACKs
1553  * in dubious situations.
1554  */
1555 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1556 {
1557         tp->snd_cwnd = min(tp->snd_cwnd,
1558                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1559         tp->snd_cwnd_stamp = tcp_time_stamp;
1560 }
1561
1562 /* Decrease cwnd each second ack. */
1563 static void tcp_cwnd_down(struct sock *sk)
1564 {
1565         const struct inet_connection_sock *icsk = inet_csk(sk);
1566         struct tcp_sock *tp = tcp_sk(sk);
1567         int decr = tp->snd_cwnd_cnt + 1;
1568
1569         tp->snd_cwnd_cnt = decr&1;
1570         decr >>= 1;
1571
1572         if (decr && tp->snd_cwnd > icsk->icsk_ca_ops->min_cwnd(sk))
1573                 tp->snd_cwnd -= decr;
1574
1575         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1576         tp->snd_cwnd_stamp = tcp_time_stamp;
1577 }
1578
1579 /* Nothing was retransmitted or returned timestamp is less
1580  * than timestamp of the first retransmission.
1581  */
1582 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1583 {
1584         return !tp->retrans_stamp ||
1585                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1586                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1587 }
1588
1589 /* Undo procedures. */
1590
1591 #if FASTRETRANS_DEBUG > 1
1592 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1593 {
1594         struct inet_sock *inet = inet_sk(sk);
1595         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1596                msg,
1597                NIPQUAD(inet->daddr), ntohs(inet->dport),
1598                tp->snd_cwnd, tp->left_out,
1599                tp->snd_ssthresh, tp->prior_ssthresh,
1600                tp->packets_out);
1601 }
1602 #else
1603 #define DBGUNDO(x...) do { } while (0)
1604 #endif
1605
1606 static void tcp_undo_cwr(struct sock *sk, const int undo)
1607 {
1608         struct tcp_sock *tp = tcp_sk(sk);
1609
1610         if (tp->prior_ssthresh) {
1611                 const struct inet_connection_sock *icsk = inet_csk(sk);
1612
1613                 if (icsk->icsk_ca_ops->undo_cwnd)
1614                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1615                 else
1616                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1617
1618                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1619                         tp->snd_ssthresh = tp->prior_ssthresh;
1620                         TCP_ECN_withdraw_cwr(tp);
1621                 }
1622         } else {
1623                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1624         }
1625         tcp_moderate_cwnd(tp);
1626         tp->snd_cwnd_stamp = tcp_time_stamp;
1627 }
1628
1629 static inline int tcp_may_undo(struct tcp_sock *tp)
1630 {
1631         return tp->undo_marker &&
1632                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1633 }
1634
1635 /* People celebrate: "We love our President!" */
1636 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1637 {
1638         if (tcp_may_undo(tp)) {
1639                 /* Happy end! We did not retransmit anything
1640                  * or our original transmission succeeded.
1641                  */
1642                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1643                 tcp_undo_cwr(sk, 1);
1644                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1645                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1646                 else
1647                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1648                 tp->undo_marker = 0;
1649         }
1650         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1651                 /* Hold old state until something *above* high_seq
1652                  * is ACKed. For Reno it is MUST to prevent false
1653                  * fast retransmits (RFC2582). SACK TCP is safe. */
1654                 tcp_moderate_cwnd(tp);
1655                 return 1;
1656         }
1657         tcp_set_ca_state(sk, TCP_CA_Open);
1658         return 0;
1659 }
1660
1661 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1662 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1663 {
1664         if (tp->undo_marker && !tp->undo_retrans) {
1665                 DBGUNDO(sk, tp, "D-SACK");
1666                 tcp_undo_cwr(sk, 1);
1667                 tp->undo_marker = 0;
1668                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1669         }
1670 }
1671
1672 /* Undo during fast recovery after partial ACK. */
1673
1674 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1675                                 int acked)
1676 {
1677         /* Partial ACK arrived. Force Hoe's retransmit. */
1678         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1679
1680         if (tcp_may_undo(tp)) {
1681                 /* Plain luck! Hole if filled with delayed
1682                  * packet, rather than with a retransmit.
1683                  */
1684                 if (tp->retrans_out == 0)
1685                         tp->retrans_stamp = 0;
1686
1687                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1688
1689                 DBGUNDO(sk, tp, "Hoe");
1690                 tcp_undo_cwr(sk, 0);
1691                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1692
1693                 /* So... Do not make Hoe's retransmit yet.
1694                  * If the first packet was delayed, the rest
1695                  * ones are most probably delayed as well.
1696                  */
1697                 failed = 0;
1698         }
1699         return failed;
1700 }
1701
1702 /* Undo during loss recovery after partial ACK. */
1703 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1704 {
1705         if (tcp_may_undo(tp)) {
1706                 struct sk_buff *skb;
1707                 sk_stream_for_retrans_queue(skb, sk) {
1708                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1709                 }
1710                 DBGUNDO(sk, tp, "partial loss");
1711                 tp->lost_out = 0;
1712                 tp->left_out = tp->sacked_out;
1713                 tcp_undo_cwr(sk, 1);
1714                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1715                 inet_csk(sk)->icsk_retransmits = 0;
1716                 tp->undo_marker = 0;
1717                 if (!IsReno(tp))
1718                         tcp_set_ca_state(sk, TCP_CA_Open);
1719                 return 1;
1720         }
1721         return 0;
1722 }
1723
1724 static inline void tcp_complete_cwr(struct sock *sk)
1725 {
1726         struct tcp_sock *tp = tcp_sk(sk);
1727         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1728         tp->snd_cwnd_stamp = tcp_time_stamp;
1729         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
1730 }
1731
1732 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1733 {
1734         tp->left_out = tp->sacked_out;
1735
1736         if (tp->retrans_out == 0)
1737                 tp->retrans_stamp = 0;
1738
1739         if (flag&FLAG_ECE)
1740                 tcp_enter_cwr(sk);
1741
1742         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
1743                 int state = TCP_CA_Open;
1744
1745                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1746                         state = TCP_CA_Disorder;
1747
1748                 if (inet_csk(sk)->icsk_ca_state != state) {
1749                         tcp_set_ca_state(sk, state);
1750                         tp->high_seq = tp->snd_nxt;
1751                 }
1752                 tcp_moderate_cwnd(tp);
1753         } else {
1754                 tcp_cwnd_down(sk);
1755         }
1756 }
1757
1758 /* Process an event, which can update packets-in-flight not trivially.
1759  * Main goal of this function is to calculate new estimate for left_out,
1760  * taking into account both packets sitting in receiver's buffer and
1761  * packets lost by network.
1762  *
1763  * Besides that it does CWND reduction, when packet loss is detected
1764  * and changes state of machine.
1765  *
1766  * It does _not_ decide what to send, it is made in function
1767  * tcp_xmit_retransmit_queue().
1768  */
1769 static void
1770 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1771                       int prior_packets, int flag)
1772 {
1773         struct inet_connection_sock *icsk = inet_csk(sk);
1774         struct tcp_sock *tp = tcp_sk(sk);
1775         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1776
1777         /* Some technical things:
1778          * 1. Reno does not count dupacks (sacked_out) automatically. */
1779         if (!tp->packets_out)
1780                 tp->sacked_out = 0;
1781         /* 2. SACK counts snd_fack in packets inaccurately. */
1782         if (tp->sacked_out == 0)
1783                 tp->fackets_out = 0;
1784
1785         /* Now state machine starts.
1786          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1787         if (flag&FLAG_ECE)
1788                 tp->prior_ssthresh = 0;
1789
1790         /* B. In all the states check for reneging SACKs. */
1791         if (tp->sacked_out && tcp_check_sack_reneging(sk))
1792                 return;
1793
1794         /* C. Process data loss notification, provided it is valid. */
1795         if ((flag&FLAG_DATA_LOST) &&
1796             before(tp->snd_una, tp->high_seq) &&
1797             icsk->icsk_ca_state != TCP_CA_Open &&
1798             tp->fackets_out > tp->reordering) {
1799                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1800                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1801         }
1802
1803         /* D. Synchronize left_out to current state. */
1804         tcp_sync_left_out(tp);
1805
1806         /* E. Check state exit conditions. State can be terminated
1807          *    when high_seq is ACKed. */
1808         if (icsk->icsk_ca_state == TCP_CA_Open) {
1809                 if (!sysctl_tcp_frto)
1810                         BUG_TRAP(tp->retrans_out == 0);
1811                 tp->retrans_stamp = 0;
1812         } else if (!before(tp->snd_una, tp->high_seq)) {
1813                 switch (icsk->icsk_ca_state) {
1814                 case TCP_CA_Loss:
1815                         icsk->icsk_retransmits = 0;
1816                         if (tcp_try_undo_recovery(sk, tp))
1817                                 return;
1818                         break;
1819
1820                 case TCP_CA_CWR:
1821                         /* CWR is to be held something *above* high_seq
1822                          * is ACKed for CWR bit to reach receiver. */
1823                         if (tp->snd_una != tp->high_seq) {
1824                                 tcp_complete_cwr(sk);
1825                                 tcp_set_ca_state(sk, TCP_CA_Open);
1826                         }
1827                         break;
1828
1829                 case TCP_CA_Disorder:
1830                         tcp_try_undo_dsack(sk, tp);
1831                         if (!tp->undo_marker ||
1832                             /* For SACK case do not Open to allow to undo
1833                              * catching for all duplicate ACKs. */
1834                             IsReno(tp) || tp->snd_una != tp->high_seq) {
1835                                 tp->undo_marker = 0;
1836                                 tcp_set_ca_state(sk, TCP_CA_Open);
1837                         }
1838                         break;
1839
1840                 case TCP_CA_Recovery:
1841                         if (IsReno(tp))
1842                                 tcp_reset_reno_sack(tp);
1843                         if (tcp_try_undo_recovery(sk, tp))
1844                                 return;
1845                         tcp_complete_cwr(sk);
1846                         break;
1847                 }
1848         }
1849
1850         /* F. Process state. */
1851         switch (icsk->icsk_ca_state) {
1852         case TCP_CA_Recovery:
1853                 if (prior_snd_una == tp->snd_una) {
1854                         if (IsReno(tp) && is_dupack)
1855                                 tcp_add_reno_sack(sk);
1856                 } else {
1857                         int acked = prior_packets - tp->packets_out;
1858                         if (IsReno(tp))
1859                                 tcp_remove_reno_sacks(sk, tp, acked);
1860                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
1861                 }
1862                 break;
1863         case TCP_CA_Loss:
1864                 if (flag&FLAG_DATA_ACKED)
1865                         icsk->icsk_retransmits = 0;
1866                 if (!tcp_try_undo_loss(sk, tp)) {
1867                         tcp_moderate_cwnd(tp);
1868                         tcp_xmit_retransmit_queue(sk);
1869                         return;
1870                 }
1871                 if (icsk->icsk_ca_state != TCP_CA_Open)
1872                         return;
1873                 /* Loss is undone; fall through to processing in Open state. */
1874         default:
1875                 if (IsReno(tp)) {
1876                         if (tp->snd_una != prior_snd_una)
1877                                 tcp_reset_reno_sack(tp);
1878                         if (is_dupack)
1879                                 tcp_add_reno_sack(sk);
1880                 }
1881
1882                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
1883                         tcp_try_undo_dsack(sk, tp);
1884
1885                 if (!tcp_time_to_recover(sk, tp)) {
1886                         tcp_try_to_open(sk, tp, flag);
1887                         return;
1888                 }
1889
1890                 /* Otherwise enter Recovery state */
1891
1892                 if (IsReno(tp))
1893                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
1894                 else
1895                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
1896
1897                 tp->high_seq = tp->snd_nxt;
1898                 tp->prior_ssthresh = 0;
1899                 tp->undo_marker = tp->snd_una;
1900                 tp->undo_retrans = tp->retrans_out;
1901
1902                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
1903                         if (!(flag&FLAG_ECE))
1904                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1905                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1906                         TCP_ECN_queue_cwr(tp);
1907                 }
1908
1909                 tp->bytes_acked = 0;
1910                 tp->snd_cwnd_cnt = 0;
1911                 tcp_set_ca_state(sk, TCP_CA_Recovery);
1912         }
1913
1914         if (is_dupack || tcp_head_timedout(sk, tp))
1915                 tcp_update_scoreboard(sk, tp);
1916         tcp_cwnd_down(sk);
1917         tcp_xmit_retransmit_queue(sk);
1918 }
1919
1920 /* Read draft-ietf-tcplw-high-performance before mucking
1921  * with this code. (Superceeds RFC1323)
1922  */
1923 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
1924 {
1925         /* RTTM Rule: A TSecr value received in a segment is used to
1926          * update the averaged RTT measurement only if the segment
1927          * acknowledges some new data, i.e., only if it advances the
1928          * left edge of the send window.
1929          *
1930          * See draft-ietf-tcplw-high-performance-00, section 3.3.
1931          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1932          *
1933          * Changed: reset backoff as soon as we see the first valid sample.
1934          * If we do not, we get strongly overstimated rto. With timestamps
1935          * samples are accepted even from very old segments: f.e., when rtt=1
1936          * increases to 8, we retransmit 5 times and after 8 seconds delayed
1937          * answer arrives rto becomes 120 seconds! If at least one of segments
1938          * in window is lost... Voila.                          --ANK (010210)
1939          */
1940         struct tcp_sock *tp = tcp_sk(sk);
1941         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
1942         tcp_rtt_estimator(sk, seq_rtt);
1943         tcp_set_rto(sk);
1944         inet_csk(sk)->icsk_backoff = 0;
1945         tcp_bound_rto(sk);
1946 }
1947
1948 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
1949 {
1950         /* We don't have a timestamp. Can only use
1951          * packets that are not retransmitted to determine
1952          * rtt estimates. Also, we must not reset the
1953          * backoff for rto until we get a non-retransmitted
1954          * packet. This allows us to deal with a situation
1955          * where the network delay has increased suddenly.
1956          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1957          */
1958
1959         if (flag & FLAG_RETRANS_DATA_ACKED)
1960                 return;
1961
1962         tcp_rtt_estimator(sk, seq_rtt);
1963         tcp_set_rto(sk);
1964         inet_csk(sk)->icsk_backoff = 0;
1965         tcp_bound_rto(sk);
1966 }
1967
1968 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
1969                                       const s32 seq_rtt)
1970 {
1971         const struct tcp_sock *tp = tcp_sk(sk);
1972         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1973         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
1974                 tcp_ack_saw_tstamp(sk, flag);
1975         else if (seq_rtt >= 0)
1976                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
1977 }
1978
1979 static inline void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
1980                                   u32 in_flight, int good)
1981 {
1982         const struct inet_connection_sock *icsk = inet_csk(sk);
1983         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
1984         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
1985 }
1986
1987 /* Restart timer after forward progress on connection.
1988  * RFC2988 recommends to restart timer to now+rto.
1989  */
1990
1991 static inline void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
1992 {
1993         if (!tp->packets_out) {
1994                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
1995         } else {
1996                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
1997         }
1998 }
1999
2000 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2001                          __u32 now, __s32 *seq_rtt)
2002 {
2003         struct tcp_sock *tp = tcp_sk(sk);
2004         struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2005         __u32 seq = tp->snd_una;
2006         __u32 packets_acked;
2007         int acked = 0;
2008
2009         /* If we get here, the whole TSO packet has not been
2010          * acked.
2011          */
2012         BUG_ON(!after(scb->end_seq, seq));
2013
2014         packets_acked = tcp_skb_pcount(skb);
2015         if (tcp_trim_head(sk, skb, seq - scb->seq))
2016                 return 0;
2017         packets_acked -= tcp_skb_pcount(skb);
2018
2019         if (packets_acked) {
2020                 __u8 sacked = scb->sacked;
2021
2022                 acked |= FLAG_DATA_ACKED;
2023                 if (sacked) {
2024                         if (sacked & TCPCB_RETRANS) {
2025                                 if (sacked & TCPCB_SACKED_RETRANS)
2026                                         tp->retrans_out -= packets_acked;
2027                                 acked |= FLAG_RETRANS_DATA_ACKED;
2028                                 *seq_rtt = -1;
2029                         } else if (*seq_rtt < 0)
2030                                 *seq_rtt = now - scb->when;
2031                         if (sacked & TCPCB_SACKED_ACKED)
2032                                 tp->sacked_out -= packets_acked;
2033                         if (sacked & TCPCB_LOST)
2034                                 tp->lost_out -= packets_acked;
2035                         if (sacked & TCPCB_URG) {
2036                                 if (tp->urg_mode &&
2037                                     !before(seq, tp->snd_up))
2038                                         tp->urg_mode = 0;
2039                         }
2040                 } else if (*seq_rtt < 0)
2041                         *seq_rtt = now - scb->when;
2042
2043                 if (tp->fackets_out) {
2044                         __u32 dval = min(tp->fackets_out, packets_acked);
2045                         tp->fackets_out -= dval;
2046                 }
2047                 tp->packets_out -= packets_acked;
2048
2049                 BUG_ON(tcp_skb_pcount(skb) == 0);
2050                 BUG_ON(!before(scb->seq, scb->end_seq));
2051         }
2052
2053         return acked;
2054 }
2055
2056 static inline u32 tcp_usrtt(const struct sk_buff *skb)
2057 {
2058         struct timeval tv, now;
2059
2060         do_gettimeofday(&now);
2061         skb_get_timestamp(skb, &tv);
2062         return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec);
2063 }
2064
2065 /* Remove acknowledged frames from the retransmission queue. */
2066 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2067 {
2068         struct tcp_sock *tp = tcp_sk(sk);
2069         const struct inet_connection_sock *icsk = inet_csk(sk);
2070         struct sk_buff *skb;
2071         __u32 now = tcp_time_stamp;
2072         int acked = 0;
2073         __s32 seq_rtt = -1;
2074         u32 pkts_acked = 0;
2075         void (*rtt_sample)(struct sock *sk, u32 usrtt)
2076                 = icsk->icsk_ca_ops->rtt_sample;
2077
2078         while ((skb = skb_peek(&sk->sk_write_queue)) &&
2079                skb != sk->sk_send_head) {
2080                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2081                 __u8 sacked = scb->sacked;
2082
2083                 /* If our packet is before the ack sequence we can
2084                  * discard it as it's confirmed to have arrived at
2085                  * the other end.
2086                  */
2087                 if (after(scb->end_seq, tp->snd_una)) {
2088                         if (tcp_skb_pcount(skb) > 1 &&
2089                             after(tp->snd_una, scb->seq))
2090                                 acked |= tcp_tso_acked(sk, skb,
2091                                                        now, &seq_rtt);
2092                         break;
2093                 }
2094
2095                 /* Initial outgoing SYN's get put onto the write_queue
2096                  * just like anything else we transmit.  It is not
2097                  * true data, and if we misinform our callers that
2098                  * this ACK acks real data, we will erroneously exit
2099                  * connection startup slow start one packet too
2100                  * quickly.  This is severely frowned upon behavior.
2101                  */
2102                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2103                         acked |= FLAG_DATA_ACKED;
2104                         ++pkts_acked;
2105                 } else {
2106                         acked |= FLAG_SYN_ACKED;
2107                         tp->retrans_stamp = 0;
2108                 }
2109
2110                 if (sacked) {
2111                         if (sacked & TCPCB_RETRANS) {
2112                                 if(sacked & TCPCB_SACKED_RETRANS)
2113                                         tp->retrans_out -= tcp_skb_pcount(skb);
2114                                 acked |= FLAG_RETRANS_DATA_ACKED;
2115                                 seq_rtt = -1;
2116                         } else if (seq_rtt < 0) {
2117                                 seq_rtt = now - scb->when;
2118                                 if (rtt_sample)
2119                                         (*rtt_sample)(sk, tcp_usrtt(skb));
2120                         }
2121                         if (sacked & TCPCB_SACKED_ACKED)
2122                                 tp->sacked_out -= tcp_skb_pcount(skb);
2123                         if (sacked & TCPCB_LOST)
2124                                 tp->lost_out -= tcp_skb_pcount(skb);
2125                         if (sacked & TCPCB_URG) {
2126                                 if (tp->urg_mode &&
2127                                     !before(scb->end_seq, tp->snd_up))
2128                                         tp->urg_mode = 0;
2129                         }
2130                 } else if (seq_rtt < 0) {
2131                         seq_rtt = now - scb->when;
2132                         if (rtt_sample)
2133                                 (*rtt_sample)(sk, tcp_usrtt(skb));
2134                 }
2135                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2136                 tcp_packets_out_dec(tp, skb);
2137                 __skb_unlink(skb, &sk->sk_write_queue);
2138                 sk_stream_free_skb(sk, skb);
2139         }
2140
2141         if (acked&FLAG_ACKED) {
2142                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2143                 tcp_ack_packets_out(sk, tp);
2144
2145                 if (icsk->icsk_ca_ops->pkts_acked)
2146                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2147         }
2148
2149 #if FASTRETRANS_DEBUG > 0
2150         BUG_TRAP((int)tp->sacked_out >= 0);
2151         BUG_TRAP((int)tp->lost_out >= 0);
2152         BUG_TRAP((int)tp->retrans_out >= 0);
2153         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2154                 const struct inet_connection_sock *icsk = inet_csk(sk);
2155                 if (tp->lost_out) {
2156                         printk(KERN_DEBUG "Leak l=%u %d\n",
2157                                tp->lost_out, icsk->icsk_ca_state);
2158                         tp->lost_out = 0;
2159                 }
2160                 if (tp->sacked_out) {
2161                         printk(KERN_DEBUG "Leak s=%u %d\n",
2162                                tp->sacked_out, icsk->icsk_ca_state);
2163                         tp->sacked_out = 0;
2164                 }
2165                 if (tp->retrans_out) {
2166                         printk(KERN_DEBUG "Leak r=%u %d\n",
2167                                tp->retrans_out, icsk->icsk_ca_state);
2168                         tp->retrans_out = 0;
2169                 }
2170         }
2171 #endif
2172         *seq_rtt_p = seq_rtt;
2173         return acked;
2174 }
2175
2176 static void tcp_ack_probe(struct sock *sk)
2177 {
2178         const struct tcp_sock *tp = tcp_sk(sk);
2179         struct inet_connection_sock *icsk = inet_csk(sk);
2180
2181         /* Was it a usable window open? */
2182
2183         if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2184                    tp->snd_una + tp->snd_wnd)) {
2185                 icsk->icsk_backoff = 0;
2186                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2187                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2188                  * This function is not for random using!
2189                  */
2190         } else {
2191                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2192                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2193                                           TCP_RTO_MAX);
2194         }
2195 }
2196
2197 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2198 {
2199         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2200                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2201 }
2202
2203 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2204 {
2205         const struct tcp_sock *tp = tcp_sk(sk);
2206         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2207                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2208 }
2209
2210 /* Check that window update is acceptable.
2211  * The function assumes that snd_una<=ack<=snd_next.
2212  */
2213 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2214                                         const u32 ack_seq, const u32 nwin)
2215 {
2216         return (after(ack, tp->snd_una) ||
2217                 after(ack_seq, tp->snd_wl1) ||
2218                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2219 }
2220
2221 /* Update our send window.
2222  *
2223  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2224  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2225  */
2226 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2227                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2228 {
2229         int flag = 0;
2230         u32 nwin = ntohs(skb->h.th->window);
2231
2232         if (likely(!skb->h.th->syn))
2233                 nwin <<= tp->rx_opt.snd_wscale;
2234
2235         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2236                 flag |= FLAG_WIN_UPDATE;
2237                 tcp_update_wl(tp, ack, ack_seq);
2238
2239                 if (tp->snd_wnd != nwin) {
2240                         tp->snd_wnd = nwin;
2241
2242                         /* Note, it is the only place, where
2243                          * fast path is recovered for sending TCP.
2244                          */
2245                         tp->pred_flags = 0;
2246                         tcp_fast_path_check(sk, tp);
2247
2248                         if (nwin > tp->max_window) {
2249                                 tp->max_window = nwin;
2250                                 tcp_sync_mss(sk, tp->pmtu_cookie);
2251                         }
2252                 }
2253         }
2254
2255         tp->snd_una = ack;
2256
2257         return flag;
2258 }
2259
2260 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2261 {
2262         struct tcp_sock *tp = tcp_sk(sk);
2263         
2264         tcp_sync_left_out(tp);
2265         
2266         if (tp->snd_una == prior_snd_una ||
2267             !before(tp->snd_una, tp->frto_highmark)) {
2268                 /* RTO was caused by loss, start retransmitting in
2269                  * go-back-N slow start
2270                  */
2271                 tcp_enter_frto_loss(sk);
2272                 return;
2273         }
2274
2275         if (tp->frto_counter == 1) {
2276                 /* First ACK after RTO advances the window: allow two new
2277                  * segments out.
2278                  */
2279                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2280         } else {
2281                 /* Also the second ACK after RTO advances the window.
2282                  * The RTO was likely spurious. Reduce cwnd and continue
2283                  * in congestion avoidance
2284                  */
2285                 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2286                 tcp_moderate_cwnd(tp);
2287         }
2288
2289         /* F-RTO affects on two new ACKs following RTO.
2290          * At latest on third ACK the TCP behavor is back to normal.
2291          */
2292         tp->frto_counter = (tp->frto_counter + 1) % 3;
2293 }
2294
2295 /* This routine deals with incoming acks, but not outgoing ones. */
2296 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2297 {
2298         struct inet_connection_sock *icsk = inet_csk(sk);
2299         struct tcp_sock *tp = tcp_sk(sk);
2300         u32 prior_snd_una = tp->snd_una;
2301         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2302         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2303         u32 prior_in_flight;
2304         s32 seq_rtt;
2305         int prior_packets;
2306
2307         /* If the ack is newer than sent or older than previous acks
2308          * then we can probably ignore it.
2309          */
2310         if (after(ack, tp->snd_nxt))
2311                 goto uninteresting_ack;
2312
2313         if (before(ack, prior_snd_una))
2314                 goto old_ack;
2315
2316         if (sysctl_tcp_abc && icsk->icsk_ca_state < TCP_CA_CWR)
2317                 tp->bytes_acked += ack - prior_snd_una;
2318
2319         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2320                 /* Window is constant, pure forward advance.
2321                  * No more checks are required.
2322                  * Note, we use the fact that SND.UNA>=SND.WL2.
2323                  */
2324                 tcp_update_wl(tp, ack, ack_seq);
2325                 tp->snd_una = ack;
2326                 flag |= FLAG_WIN_UPDATE;
2327
2328                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2329
2330                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2331         } else {
2332                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2333                         flag |= FLAG_DATA;
2334                 else
2335                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2336
2337                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2338
2339                 if (TCP_SKB_CB(skb)->sacked)
2340                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2341
2342                 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2343                         flag |= FLAG_ECE;
2344
2345                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2346         }
2347
2348         /* We passed data and got it acked, remove any soft error
2349          * log. Something worked...
2350          */
2351         sk->sk_err_soft = 0;
2352         tp->rcv_tstamp = tcp_time_stamp;
2353         prior_packets = tp->packets_out;
2354         if (!prior_packets)
2355                 goto no_queue;
2356
2357         prior_in_flight = tcp_packets_in_flight(tp);
2358
2359         /* See if we can take anything off of the retransmit queue. */
2360         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2361
2362         if (tp->frto_counter)
2363                 tcp_process_frto(sk, prior_snd_una);
2364
2365         if (tcp_ack_is_dubious(sk, flag)) {
2366                 /* Advanve CWND, if state allows this. */
2367                 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
2368                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2369                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2370         } else {
2371                 if ((flag & FLAG_DATA_ACKED))
2372                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2373         }
2374
2375         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2376                 dst_confirm(sk->sk_dst_cache);
2377
2378         return 1;
2379
2380 no_queue:
2381         icsk->icsk_probes_out = 0;
2382
2383         /* If this ack opens up a zero window, clear backoff.  It was
2384          * being used to time the probes, and is probably far higher than
2385          * it needs to be for normal retransmission.
2386          */
2387         if (sk->sk_send_head)
2388                 tcp_ack_probe(sk);
2389         return 1;
2390
2391 old_ack:
2392         if (TCP_SKB_CB(skb)->sacked)
2393                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2394
2395 uninteresting_ack:
2396         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2397         return 0;
2398 }
2399
2400
2401 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2402  * But, this can also be called on packets in the established flow when
2403  * the fast version below fails.
2404  */
2405 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2406 {
2407         unsigned char *ptr;
2408         struct tcphdr *th = skb->h.th;
2409         int length=(th->doff*4)-sizeof(struct tcphdr);
2410
2411         ptr = (unsigned char *)(th + 1);
2412         opt_rx->saw_tstamp = 0;
2413
2414         while(length>0) {
2415                 int opcode=*ptr++;
2416                 int opsize;
2417
2418                 switch (opcode) {
2419                         case TCPOPT_EOL:
2420                                 return;
2421                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2422                                 length--;
2423                                 continue;
2424                         default:
2425                                 opsize=*ptr++;
2426                                 if (opsize < 2) /* "silly options" */
2427                                         return;
2428                                 if (opsize > length)
2429                                         return; /* don't parse partial options */
2430                                 switch(opcode) {
2431                                 case TCPOPT_MSS:
2432                                         if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2433                                                 u16 in_mss = ntohs(get_unaligned((__u16 *)ptr));
2434                                                 if (in_mss) {
2435                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2436                                                                 in_mss = opt_rx->user_mss;
2437                                                         opt_rx->mss_clamp = in_mss;
2438                                                 }
2439                                         }
2440                                         break;
2441                                 case TCPOPT_WINDOW:
2442                                         if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2443                                                 if (sysctl_tcp_window_scaling) {
2444                                                         __u8 snd_wscale = *(__u8 *) ptr;
2445                                                         opt_rx->wscale_ok = 1;
2446                                                         if (snd_wscale > 14) {
2447                                                                 if(net_ratelimit())
2448                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2449                                                                                "scaling value %d >14 received.\n",
2450                                                                                snd_wscale);
2451                                                                 snd_wscale = 14;
2452                                                         }
2453                                                         opt_rx->snd_wscale = snd_wscale;
2454                                                 }
2455                                         break;
2456                                 case TCPOPT_TIMESTAMP:
2457                                         if(opsize==TCPOLEN_TIMESTAMP) {
2458                                                 if ((estab && opt_rx->tstamp_ok) ||
2459                                                     (!estab && sysctl_tcp_timestamps)) {
2460                                                         opt_rx->saw_tstamp = 1;
2461                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr));
2462                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4)));
2463                                                 }
2464                                         }
2465                                         break;
2466                                 case TCPOPT_SACK_PERM:
2467                                         if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2468                                                 if (sysctl_tcp_sack) {
2469                                                         opt_rx->sack_ok = 1;
2470                                                         tcp_sack_reset(opt_rx);
2471                                                 }
2472                                         }
2473                                         break;
2474
2475                                 case TCPOPT_SACK:
2476                                         if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2477                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2478                                            opt_rx->sack_ok) {
2479                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2480                                         }
2481                                 };
2482                                 ptr+=opsize-2;
2483                                 length-=opsize;
2484                 };
2485         }
2486 }
2487
2488 /* Fast parse options. This hopes to only see timestamps.
2489  * If it is wrong it falls back on tcp_parse_options().
2490  */
2491 static inline int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2492                                          struct tcp_sock *tp)
2493 {
2494         if (th->doff == sizeof(struct tcphdr)>>2) {
2495                 tp->rx_opt.saw_tstamp = 0;
2496                 return 0;
2497         } else if (tp->rx_opt.tstamp_ok &&
2498                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2499                 __u32 *ptr = (__u32 *)(th + 1);
2500                 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2501                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2502                         tp->rx_opt.saw_tstamp = 1;
2503                         ++ptr;
2504                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2505                         ++ptr;
2506                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2507                         return 1;
2508                 }
2509         }
2510         tcp_parse_options(skb, &tp->rx_opt, 1);
2511         return 1;
2512 }
2513
2514 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2515 {
2516         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2517         tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2518 }
2519
2520 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2521 {
2522         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2523                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2524                  * extra check below makes sure this can only happen
2525                  * for pure ACK frames.  -DaveM
2526                  *
2527                  * Not only, also it occurs for expired timestamps.
2528                  */
2529
2530                 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2531                    xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2532                         tcp_store_ts_recent(tp);
2533         }
2534 }
2535
2536 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2537  *
2538  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2539  * it can pass through stack. So, the following predicate verifies that
2540  * this segment is not used for anything but congestion avoidance or
2541  * fast retransmit. Moreover, we even are able to eliminate most of such
2542  * second order effects, if we apply some small "replay" window (~RTO)
2543  * to timestamp space.
2544  *
2545  * All these measures still do not guarantee that we reject wrapped ACKs
2546  * on networks with high bandwidth, when sequence space is recycled fastly,
2547  * but it guarantees that such events will be very rare and do not affect
2548  * connection seriously. This doesn't look nice, but alas, PAWS is really
2549  * buggy extension.
2550  *
2551  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2552  * states that events when retransmit arrives after original data are rare.
2553  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2554  * the biggest problem on large power networks even with minor reordering.
2555  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2556  * up to bandwidth of 18Gigabit/sec. 8) ]
2557  */
2558
2559 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2560 {
2561         struct tcp_sock *tp = tcp_sk(sk);
2562         struct tcphdr *th = skb->h.th;
2563         u32 seq = TCP_SKB_CB(skb)->seq;
2564         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2565
2566         return (/* 1. Pure ACK with correct sequence number. */
2567                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2568
2569                 /* 2. ... and duplicate ACK. */
2570                 ack == tp->snd_una &&
2571
2572                 /* 3. ... and does not update window. */
2573                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2574
2575                 /* 4. ... and sits in replay window. */
2576                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
2577 }
2578
2579 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
2580 {
2581         const struct tcp_sock *tp = tcp_sk(sk);
2582         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2583                 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2584                 !tcp_disordered_ack(sk, skb));
2585 }
2586
2587 /* Check segment sequence number for validity.
2588  *
2589  * Segment controls are considered valid, if the segment
2590  * fits to the window after truncation to the window. Acceptability
2591  * of data (and SYN, FIN, of course) is checked separately.
2592  * See tcp_data_queue(), for example.
2593  *
2594  * Also, controls (RST is main one) are accepted using RCV.WUP instead
2595  * of RCV.NXT. Peer still did not advance his SND.UNA when we
2596  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2597  * (borrowed from freebsd)
2598  */
2599
2600 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2601 {
2602         return  !before(end_seq, tp->rcv_wup) &&
2603                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2604 }
2605
2606 /* When we get a reset we do this. */
2607 static void tcp_reset(struct sock *sk)
2608 {
2609         /* We want the right error as BSD sees it (and indeed as we do). */
2610         switch (sk->sk_state) {
2611                 case TCP_SYN_SENT:
2612                         sk->sk_err = ECONNREFUSED;
2613                         break;
2614                 case TCP_CLOSE_WAIT:
2615                         sk->sk_err = EPIPE;
2616                         break;
2617                 case TCP_CLOSE:
2618                         return;
2619                 default:
2620                         sk->sk_err = ECONNRESET;
2621         }
2622
2623         if (!sock_flag(sk, SOCK_DEAD))
2624                 sk->sk_error_report(sk);
2625
2626         tcp_done(sk);
2627 }
2628
2629 /*
2630  *      Process the FIN bit. This now behaves as it is supposed to work
2631  *      and the FIN takes effect when it is validly part of sequence
2632  *      space. Not before when we get holes.
2633  *
2634  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2635  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
2636  *      TIME-WAIT)
2637  *
2638  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
2639  *      close and we go into CLOSING (and later onto TIME-WAIT)
2640  *
2641  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2642  */
2643 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2644 {
2645         struct tcp_sock *tp = tcp_sk(sk);
2646
2647         inet_csk_schedule_ack(sk);
2648
2649         sk->sk_shutdown |= RCV_SHUTDOWN;
2650         sock_set_flag(sk, SOCK_DONE);
2651
2652         switch (sk->sk_state) {
2653                 case TCP_SYN_RECV:
2654                 case TCP_ESTABLISHED:
2655                         /* Move to CLOSE_WAIT */
2656                         tcp_set_state(sk, TCP_CLOSE_WAIT);
2657                         inet_csk(sk)->icsk_ack.pingpong = 1;
2658                         break;
2659
2660                 case TCP_CLOSE_WAIT:
2661                 case TCP_CLOSING:
2662                         /* Received a retransmission of the FIN, do
2663                          * nothing.
2664                          */
2665                         break;
2666                 case TCP_LAST_ACK:
2667                         /* RFC793: Remain in the LAST-ACK state. */
2668                         break;
2669
2670                 case TCP_FIN_WAIT1:
2671                         /* This case occurs when a simultaneous close
2672                          * happens, we must ack the received FIN and
2673                          * enter the CLOSING state.
2674                          */
2675                         tcp_send_ack(sk);
2676                         tcp_set_state(sk, TCP_CLOSING);
2677                         break;
2678                 case TCP_FIN_WAIT2:
2679                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
2680                         tcp_send_ack(sk);
2681                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2682                         break;
2683                 default:
2684                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2685                          * cases we should never reach this piece of code.
2686                          */
2687                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2688                                __FUNCTION__, sk->sk_state);
2689                         break;
2690         };
2691
2692         /* It _is_ possible, that we have something out-of-order _after_ FIN.
2693          * Probably, we should reset in this case. For now drop them.
2694          */
2695         __skb_queue_purge(&tp->out_of_order_queue);
2696         if (tp->rx_opt.sack_ok)
2697                 tcp_sack_reset(&tp->rx_opt);
2698         sk_stream_mem_reclaim(sk);
2699
2700         if (!sock_flag(sk, SOCK_DEAD)) {
2701                 sk->sk_state_change(sk);
2702
2703                 /* Do not send POLL_HUP for half duplex close. */
2704                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2705                     sk->sk_state == TCP_CLOSE)
2706                         sk_wake_async(sk, 1, POLL_HUP);
2707                 else
2708                         sk_wake_async(sk, 1, POLL_IN);
2709         }
2710 }
2711
2712 static __inline__ int
2713 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2714 {
2715         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2716                 if (before(seq, sp->start_seq))
2717                         sp->start_seq = seq;
2718                 if (after(end_seq, sp->end_seq))
2719                         sp->end_seq = end_seq;
2720                 return 1;
2721         }
2722         return 0;
2723 }
2724
2725 static inline void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
2726 {
2727         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2728                 if (before(seq, tp->rcv_nxt))
2729                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
2730                 else
2731                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
2732
2733                 tp->rx_opt.dsack = 1;
2734                 tp->duplicate_sack[0].start_seq = seq;
2735                 tp->duplicate_sack[0].end_seq = end_seq;
2736                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
2737         }
2738 }
2739
2740 static inline void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
2741 {
2742         if (!tp->rx_opt.dsack)
2743                 tcp_dsack_set(tp, seq, end_seq);
2744         else
2745                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2746 }
2747
2748 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2749 {
2750         struct tcp_sock *tp = tcp_sk(sk);
2751
2752         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2753             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2754                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2755                 tcp_enter_quickack_mode(sk);
2756
2757                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2758                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2759
2760                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2761                                 end_seq = tp->rcv_nxt;
2762                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2763                 }
2764         }
2765
2766         tcp_send_ack(sk);
2767 }
2768
2769 /* These routines update the SACK block as out-of-order packets arrive or
2770  * in-order packets close up the sequence space.
2771  */
2772 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
2773 {
2774         int this_sack;
2775         struct tcp_sack_block *sp = &tp->selective_acks[0];
2776         struct tcp_sack_block *swalk = sp+1;
2777
2778         /* See if the recent change to the first SACK eats into
2779          * or hits the sequence space of other SACK blocks, if so coalesce.
2780          */
2781         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
2782                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2783                         int i;
2784
2785                         /* Zap SWALK, by moving every further SACK up by one slot.
2786                          * Decrease num_sacks.
2787                          */
2788                         tp->rx_opt.num_sacks--;
2789                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2790                         for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
2791                                 sp[i] = sp[i+1];
2792                         continue;
2793                 }
2794                 this_sack++, swalk++;
2795         }
2796 }
2797
2798 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2799 {
2800         __u32 tmp;
2801
2802         tmp = sack1->start_seq;
2803         sack1->start_seq = sack2->start_seq;
2804         sack2->start_seq = tmp;
2805
2806         tmp = sack1->end_seq;
2807         sack1->end_seq = sack2->end_seq;
2808         sack2->end_seq = tmp;
2809 }
2810
2811 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
2812 {
2813         struct tcp_sock *tp = tcp_sk(sk);
2814         struct tcp_sack_block *sp = &tp->selective_acks[0];
2815         int cur_sacks = tp->rx_opt.num_sacks;
2816         int this_sack;
2817
2818         if (!cur_sacks)
2819                 goto new_sack;
2820
2821         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
2822                 if (tcp_sack_extend(sp, seq, end_seq)) {
2823                         /* Rotate this_sack to the first one. */
2824                         for (; this_sack>0; this_sack--, sp--)
2825                                 tcp_sack_swap(sp, sp-1);
2826                         if (cur_sacks > 1)
2827                                 tcp_sack_maybe_coalesce(tp);
2828                         return;
2829                 }
2830         }
2831
2832         /* Could not find an adjacent existing SACK, build a new one,
2833          * put it at the front, and shift everyone else down.  We
2834          * always know there is at least one SACK present already here.
2835          *
2836          * If the sack array is full, forget about the last one.
2837          */
2838         if (this_sack >= 4) {
2839                 this_sack--;
2840                 tp->rx_opt.num_sacks--;
2841                 sp--;
2842         }
2843         for(; this_sack > 0; this_sack--, sp--)
2844                 *sp = *(sp-1);
2845
2846 new_sack:
2847         /* Build the new head SACK, and we're done. */
2848         sp->start_seq = seq;
2849         sp->end_seq = end_seq;
2850         tp->rx_opt.num_sacks++;
2851         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2852 }
2853
2854 /* RCV.NXT advances, some SACKs should be eaten. */
2855
2856 static void tcp_sack_remove(struct tcp_sock *tp)
2857 {
2858         struct tcp_sack_block *sp = &tp->selective_acks[0];
2859         int num_sacks = tp->rx_opt.num_sacks;
2860         int this_sack;
2861
2862         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2863         if (skb_queue_empty(&tp->out_of_order_queue)) {
2864                 tp->rx_opt.num_sacks = 0;
2865                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
2866                 return;
2867         }
2868
2869         for(this_sack = 0; this_sack < num_sacks; ) {
2870                 /* Check if the start of the sack is covered by RCV.NXT. */
2871                 if (!before(tp->rcv_nxt, sp->start_seq)) {
2872                         int i;
2873
2874                         /* RCV.NXT must cover all the block! */
2875                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
2876
2877                         /* Zap this SACK, by moving forward any other SACKS. */
2878                         for (i=this_sack+1; i < num_sacks; i++)
2879                                 tp->selective_acks[i-1] = tp->selective_acks[i];
2880                         num_sacks--;
2881                         continue;
2882                 }
2883                 this_sack++;
2884                 sp++;
2885         }
2886         if (num_sacks != tp->rx_opt.num_sacks) {
2887                 tp->rx_opt.num_sacks = num_sacks;
2888                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2889         }
2890 }
2891
2892 /* This one checks to see if we can put data from the
2893  * out_of_order queue into the receive_queue.
2894  */
2895 static void tcp_ofo_queue(struct sock *sk)
2896 {
2897         struct tcp_sock *tp = tcp_sk(sk);
2898         __u32 dsack_high = tp->rcv_nxt;
2899         struct sk_buff *skb;
2900
2901         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
2902                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
2903                         break;
2904
2905                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
2906                         __u32 dsack = dsack_high;
2907                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
2908                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
2909                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
2910                 }
2911
2912                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
2913                         SOCK_DEBUG(sk, "ofo packet was already received \n");
2914                         __skb_unlink(skb, &tp->out_of_order_queue);
2915                         __kfree_skb(skb);
2916                         continue;
2917                 }
2918                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
2919                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
2920                            TCP_SKB_CB(skb)->end_seq);
2921
2922                 __skb_unlink(skb, &tp->out_of_order_queue);
2923                 __skb_queue_tail(&sk->sk_receive_queue, skb);
2924                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2925                 if(skb->h.th->fin)
2926                         tcp_fin(skb, sk, skb->h.th);
2927         }
2928 }
2929
2930 static int tcp_prune_queue(struct sock *sk);
2931
2932 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
2933 {
2934         struct tcphdr *th = skb->h.th;
2935         struct tcp_sock *tp = tcp_sk(sk);
2936         int eaten = -1;
2937
2938         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
2939                 goto drop;
2940
2941         __skb_pull(skb, th->doff*4);
2942
2943         TCP_ECN_accept_cwr(tp, skb);
2944
2945         if (tp->rx_opt.dsack) {
2946                 tp->rx_opt.dsack = 0;
2947                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
2948                                                     4 - tp->rx_opt.tstamp_ok);
2949         }
2950
2951         /*  Queue data for delivery to the user.
2952          *  Packets in sequence go to the receive queue.
2953          *  Out of sequence packets to the out_of_order_queue.
2954          */
2955         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
2956                 if (tcp_receive_window(tp) == 0)
2957                         goto out_of_window;
2958
2959                 /* Ok. In sequence. In window. */
2960                 if (tp->ucopy.task == current &&
2961                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
2962                     sock_owned_by_user(sk) && !tp->urg_data) {
2963                         int chunk = min_t(unsigned int, skb->len,
2964                                                         tp->ucopy.len);
2965
2966                         __set_current_state(TASK_RUNNING);
2967
2968                         local_bh_enable();
2969                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
2970                                 tp->ucopy.len -= chunk;
2971                                 tp->copied_seq += chunk;
2972                                 eaten = (chunk == skb->len && !th->fin);
2973                                 tcp_rcv_space_adjust(sk);
2974                         }
2975                         local_bh_disable();
2976                 }
2977
2978                 if (eaten <= 0) {
2979 queue_and_out:
2980                         if (eaten < 0 &&
2981                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
2982                              !sk_stream_rmem_schedule(sk, skb))) {
2983                                 if (tcp_prune_queue(sk) < 0 ||
2984                                     !sk_stream_rmem_schedule(sk, skb))
2985                                         goto drop;
2986                         }
2987                         sk_stream_set_owner_r(skb, sk);
2988                         __skb_queue_tail(&sk->sk_receive_queue, skb);
2989                 }
2990                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2991                 if(skb->len)
2992                         tcp_event_data_recv(sk, tp, skb);
2993                 if(th->fin)
2994                         tcp_fin(skb, sk, th);
2995
2996                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
2997                         tcp_ofo_queue(sk);
2998
2999                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3000                          * gap in queue is filled.
3001                          */
3002                         if (skb_queue_empty(&tp->out_of_order_queue))
3003                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3004                 }
3005
3006                 if (tp->rx_opt.num_sacks)
3007                         tcp_sack_remove(tp);
3008
3009                 tcp_fast_path_check(sk, tp);
3010
3011                 if (eaten > 0)
3012                         __kfree_skb(skb);
3013                 else if (!sock_flag(sk, SOCK_DEAD))
3014                         sk->sk_data_ready(sk, 0);
3015                 return;
3016         }
3017
3018         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3019                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3020                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3021                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3022
3023 out_of_window:
3024                 tcp_enter_quickack_mode(sk);
3025                 inet_csk_schedule_ack(sk);
3026 drop:
3027                 __kfree_skb(skb);
3028                 return;
3029         }
3030
3031         /* Out of window. F.e. zero window probe. */
3032         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3033                 goto out_of_window;
3034
3035         tcp_enter_quickack_mode(sk);
3036
3037         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3038                 /* Partial packet, seq < rcv_next < end_seq */
3039                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3040                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3041                            TCP_SKB_CB(skb)->end_seq);
3042
3043                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3044                 
3045                 /* If window is closed, drop tail of packet. But after
3046                  * remembering D-SACK for its head made in previous line.
3047                  */
3048                 if (!tcp_receive_window(tp))
3049                         goto out_of_window;
3050                 goto queue_and_out;
3051         }
3052
3053         TCP_ECN_check_ce(tp, skb);
3054
3055         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3056             !sk_stream_rmem_schedule(sk, skb)) {
3057                 if (tcp_prune_queue(sk) < 0 ||
3058                     !sk_stream_rmem_schedule(sk, skb))
3059                         goto drop;
3060         }
3061
3062         /* Disable header prediction. */
3063         tp->pred_flags = 0;
3064         inet_csk_schedule_ack(sk);
3065
3066         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3067                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3068
3069         sk_stream_set_owner_r(skb, sk);
3070
3071         if (!skb_peek(&tp->out_of_order_queue)) {
3072                 /* Initial out of order segment, build 1 SACK. */
3073                 if (tp->rx_opt.sack_ok) {
3074                         tp->rx_opt.num_sacks = 1;
3075                         tp->rx_opt.dsack     = 0;
3076                         tp->rx_opt.eff_sacks = 1;
3077                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3078                         tp->selective_acks[0].end_seq =
3079                                                 TCP_SKB_CB(skb)->end_seq;
3080                 }
3081                 __skb_queue_head(&tp->out_of_order_queue,skb);
3082         } else {
3083                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3084                 u32 seq = TCP_SKB_CB(skb)->seq;
3085                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3086
3087                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3088                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3089
3090                         if (!tp->rx_opt.num_sacks ||
3091                             tp->selective_acks[0].end_seq != seq)
3092                                 goto add_sack;
3093
3094                         /* Common case: data arrive in order after hole. */
3095                         tp->selective_acks[0].end_seq = end_seq;
3096                         return;
3097                 }
3098
3099                 /* Find place to insert this segment. */
3100                 do {
3101                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3102                                 break;
3103                 } while ((skb1 = skb1->prev) !=
3104                          (struct sk_buff*)&tp->out_of_order_queue);
3105
3106                 /* Do skb overlap to previous one? */
3107                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3108                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3109                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3110                                 /* All the bits are present. Drop. */
3111                                 __kfree_skb(skb);
3112                                 tcp_dsack_set(tp, seq, end_seq);
3113                                 goto add_sack;
3114                         }
3115                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3116                                 /* Partial overlap. */
3117                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3118                         } else {
3119                                 skb1 = skb1->prev;
3120                         }
3121                 }
3122                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3123                 
3124                 /* And clean segments covered by new one as whole. */
3125                 while ((skb1 = skb->next) !=
3126                        (struct sk_buff*)&tp->out_of_order_queue &&
3127                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3128                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3129                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3130                                break;
3131                        }
3132                        __skb_unlink(skb1, &tp->out_of_order_queue);
3133                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3134                        __kfree_skb(skb1);
3135                 }
3136
3137 add_sack:
3138                 if (tp->rx_opt.sack_ok)
3139                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3140         }
3141 }
3142
3143 /* Collapse contiguous sequence of skbs head..tail with
3144  * sequence numbers start..end.
3145  * Segments with FIN/SYN are not collapsed (only because this
3146  * simplifies code)
3147  */
3148 static void
3149 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3150              struct sk_buff *head, struct sk_buff *tail,
3151              u32 start, u32 end)
3152 {
3153         struct sk_buff *skb;
3154
3155         /* First, check that queue is collapsable and find
3156          * the point where collapsing can be useful. */
3157         for (skb = head; skb != tail; ) {
3158                 /* No new bits? It is possible on ofo queue. */
3159                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3160                         struct sk_buff *next = skb->next;
3161                         __skb_unlink(skb, list);
3162                         __kfree_skb(skb);
3163                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3164                         skb = next;
3165                         continue;
3166                 }
3167
3168                 /* The first skb to collapse is:
3169                  * - not SYN/FIN and
3170                  * - bloated or contains data before "start" or
3171                  *   overlaps to the next one.
3172                  */
3173                 if (!skb->h.th->syn && !skb->h.th->fin &&
3174                     (tcp_win_from_space(skb->truesize) > skb->len ||
3175                      before(TCP_SKB_CB(skb)->seq, start) ||
3176                      (skb->next != tail &&
3177                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3178                         break;
3179
3180                 /* Decided to skip this, advance start seq. */
3181                 start = TCP_SKB_CB(skb)->end_seq;
3182                 skb = skb->next;
3183         }
3184         if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3185                 return;
3186
3187         while (before(start, end)) {
3188                 struct sk_buff *nskb;
3189                 int header = skb_headroom(skb);
3190                 int copy = SKB_MAX_ORDER(header, 0);
3191
3192                 /* Too big header? This can happen with IPv6. */
3193                 if (copy < 0)
3194                         return;
3195                 if (end-start < copy)
3196                         copy = end-start;
3197                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3198                 if (!nskb)
3199                         return;
3200                 skb_reserve(nskb, header);
3201                 memcpy(nskb->head, skb->head, header);
3202                 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3203                 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3204                 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3205                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3206                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3207                 __skb_insert(nskb, skb->prev, skb, list);
3208                 sk_stream_set_owner_r(nskb, sk);
3209
3210                 /* Copy data, releasing collapsed skbs. */
3211                 while (copy > 0) {
3212                         int offset = start - TCP_SKB_CB(skb)->seq;
3213                         int size = TCP_SKB_CB(skb)->end_seq - start;
3214
3215                         if (offset < 0) BUG();
3216                         if (size > 0) {
3217                                 size = min(copy, size);
3218                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3219                                         BUG();
3220                                 TCP_SKB_CB(nskb)->end_seq += size;
3221                                 copy -= size;
3222                                 start += size;
3223                         }
3224                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3225                                 struct sk_buff *next = skb->next;
3226                                 __skb_unlink(skb, list);
3227                                 __kfree_skb(skb);
3228                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3229                                 skb = next;
3230                                 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3231                                         return;
3232                         }
3233                 }
3234         }
3235 }
3236
3237 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3238  * and tcp_collapse() them until all the queue is collapsed.
3239  */
3240 static void tcp_collapse_ofo_queue(struct sock *sk)
3241 {
3242         struct tcp_sock *tp = tcp_sk(sk);
3243         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3244         struct sk_buff *head;
3245         u32 start, end;
3246
3247         if (skb == NULL)
3248                 return;
3249
3250         start = TCP_SKB_CB(skb)->seq;
3251         end = TCP_SKB_CB(skb)->end_seq;
3252         head = skb;
3253
3254         for (;;) {
3255                 skb = skb->next;
3256
3257                 /* Segment is terminated when we see gap or when
3258                  * we are at the end of all the queue. */
3259                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3260                     after(TCP_SKB_CB(skb)->seq, end) ||
3261                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3262                         tcp_collapse(sk, &tp->out_of_order_queue,
3263                                      head, skb, start, end);
3264                         head = skb;
3265                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3266                                 break;
3267                         /* Start new segment */
3268                         start = TCP_SKB_CB(skb)->seq;
3269                         end = TCP_SKB_CB(skb)->end_seq;
3270                 } else {
3271                         if (before(TCP_SKB_CB(skb)->seq, start))
3272                                 start = TCP_SKB_CB(skb)->seq;
3273                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3274                                 end = TCP_SKB_CB(skb)->end_seq;
3275                 }
3276         }
3277 }
3278
3279 /* Reduce allocated memory if we can, trying to get
3280  * the socket within its memory limits again.
3281  *
3282  * Return less than zero if we should start dropping frames
3283  * until the socket owning process reads some of the data
3284  * to stabilize the situation.
3285  */
3286 static int tcp_prune_queue(struct sock *sk)
3287 {
3288         struct tcp_sock *tp = tcp_sk(sk); 
3289
3290         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3291
3292         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3293
3294         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3295                 tcp_clamp_window(sk, tp);
3296         else if (tcp_memory_pressure)
3297                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3298
3299         tcp_collapse_ofo_queue(sk);
3300         tcp_collapse(sk, &sk->sk_receive_queue,
3301                      sk->sk_receive_queue.next,
3302                      (struct sk_buff*)&sk->sk_receive_queue,
3303                      tp->copied_seq, tp->rcv_nxt);
3304         sk_stream_mem_reclaim(sk);
3305
3306         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3307                 return 0;
3308
3309         /* Collapsing did not help, destructive actions follow.
3310          * This must not ever occur. */
3311
3312         /* First, purge the out_of_order queue. */
3313         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3314                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3315                 __skb_queue_purge(&tp->out_of_order_queue);
3316
3317                 /* Reset SACK state.  A conforming SACK implementation will
3318                  * do the same at a timeout based retransmit.  When a connection
3319                  * is in a sad state like this, we care only about integrity
3320                  * of the connection not performance.
3321                  */
3322                 if (tp->rx_opt.sack_ok)
3323                         tcp_sack_reset(&tp->rx_opt);
3324                 sk_stream_mem_reclaim(sk);
3325         }
3326
3327         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3328                 return 0;
3329
3330         /* If we are really being abused, tell the caller to silently
3331          * drop receive data on the floor.  It will get retransmitted
3332          * and hopefully then we'll have sufficient space.
3333          */
3334         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3335
3336         /* Massive buffer overcommit. */
3337         tp->pred_flags = 0;
3338         return -1;
3339 }
3340
3341
3342 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3343  * As additional protections, we do not touch cwnd in retransmission phases,
3344  * and if application hit its sndbuf limit recently.
3345  */
3346 void tcp_cwnd_application_limited(struct sock *sk)
3347 {
3348         struct tcp_sock *tp = tcp_sk(sk);
3349
3350         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3351             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3352                 /* Limited by application or receiver window. */
3353                 u32 win_used = max(tp->snd_cwnd_used, 2U);
3354                 if (win_used < tp->snd_cwnd) {
3355                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3356                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3357                 }
3358                 tp->snd_cwnd_used = 0;
3359         }
3360         tp->snd_cwnd_stamp = tcp_time_stamp;
3361 }
3362
3363 static inline int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3364 {
3365         /* If the user specified a specific send buffer setting, do
3366          * not modify it.
3367          */
3368         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3369                 return 0;
3370
3371         /* If we are under global TCP memory pressure, do not expand.  */
3372         if (tcp_memory_pressure)
3373                 return 0;
3374
3375         /* If we are under soft global TCP memory pressure, do not expand.  */
3376         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3377                 return 0;
3378
3379         /* If we filled the congestion window, do not expand.  */
3380         if (tp->packets_out >= tp->snd_cwnd)
3381                 return 0;
3382
3383         return 1;
3384 }
3385
3386 /* When incoming ACK allowed to free some skb from write_queue,
3387  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3388  * on the exit from tcp input handler.
3389  *
3390  * PROBLEM: sndbuf expansion does not work well with largesend.
3391  */
3392 static void tcp_new_space(struct sock *sk)
3393 {
3394         struct tcp_sock *tp = tcp_sk(sk);
3395
3396         if (tcp_should_expand_sndbuf(sk, tp)) {
3397                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3398                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3399                     demanded = max_t(unsigned int, tp->snd_cwnd,
3400                                                    tp->reordering + 1);
3401                 sndmem *= 2*demanded;
3402                 if (sndmem > sk->sk_sndbuf)
3403                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3404                 tp->snd_cwnd_stamp = tcp_time_stamp;
3405         }
3406
3407         sk->sk_write_space(sk);
3408 }
3409
3410 static inline void tcp_check_space(struct sock *sk)
3411 {
3412         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3413                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3414                 if (sk->sk_socket &&
3415                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3416                         tcp_new_space(sk);
3417         }
3418 }
3419
3420 static __inline__ void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3421 {
3422         tcp_push_pending_frames(sk, tp);
3423         tcp_check_space(sk);
3424 }
3425
3426 /*
3427  * Check if sending an ack is needed.
3428  */
3429 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3430 {
3431         struct tcp_sock *tp = tcp_sk(sk);
3432
3433             /* More than one full frame received... */
3434         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3435              /* ... and right edge of window advances far enough.
3436               * (tcp_recvmsg() will send ACK otherwise). Or...
3437               */
3438              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3439             /* We ACK each frame or... */
3440             tcp_in_quickack_mode(sk) ||
3441             /* We have out of order data. */
3442             (ofo_possible &&
3443              skb_peek(&tp->out_of_order_queue))) {
3444                 /* Then ack it now */
3445                 tcp_send_ack(sk);
3446         } else {
3447                 /* Else, send delayed ack. */
3448                 tcp_send_delayed_ack(sk);
3449         }
3450 }
3451
3452 static __inline__ void tcp_ack_snd_check(struct sock *sk)
3453 {
3454         if (!inet_csk_ack_scheduled(sk)) {
3455                 /* We sent a data segment already. */
3456                 return;
3457         }
3458         __tcp_ack_snd_check(sk, 1);
3459 }
3460
3461 /*
3462  *      This routine is only called when we have urgent data
3463  *      signalled. Its the 'slow' part of tcp_urg. It could be
3464  *      moved inline now as tcp_urg is only called from one
3465  *      place. We handle URGent data wrong. We have to - as
3466  *      BSD still doesn't use the correction from RFC961.
3467  *      For 1003.1g we should support a new option TCP_STDURG to permit
3468  *      either form (or just set the sysctl tcp_stdurg).
3469  */
3470  
3471 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3472 {
3473         struct tcp_sock *tp = tcp_sk(sk);
3474         u32 ptr = ntohs(th->urg_ptr);
3475
3476         if (ptr && !sysctl_tcp_stdurg)
3477                 ptr--;
3478         ptr += ntohl(th->seq);
3479
3480         /* Ignore urgent data that we've already seen and read. */
3481         if (after(tp->copied_seq, ptr))
3482                 return;
3483
3484         /* Do not replay urg ptr.
3485          *
3486          * NOTE: interesting situation not covered by specs.
3487          * Misbehaving sender may send urg ptr, pointing to segment,
3488          * which we already have in ofo queue. We are not able to fetch
3489          * such data and will stay in TCP_URG_NOTYET until will be eaten
3490          * by recvmsg(). Seems, we are not obliged to handle such wicked
3491          * situations. But it is worth to think about possibility of some
3492          * DoSes using some hypothetical application level deadlock.
3493          */
3494         if (before(ptr, tp->rcv_nxt))
3495                 return;
3496
3497         /* Do we already have a newer (or duplicate) urgent pointer? */
3498         if (tp->urg_data && !after(ptr, tp->urg_seq))
3499                 return;
3500
3501         /* Tell the world about our new urgent pointer. */
3502         sk_send_sigurg(sk);
3503
3504         /* We may be adding urgent data when the last byte read was
3505          * urgent. To do this requires some care. We cannot just ignore
3506          * tp->copied_seq since we would read the last urgent byte again
3507          * as data, nor can we alter copied_seq until this data arrives
3508          * or we break the sematics of SIOCATMARK (and thus sockatmark())
3509          *
3510          * NOTE. Double Dutch. Rendering to plain English: author of comment
3511          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3512          * and expect that both A and B disappear from stream. This is _wrong_.
3513          * Though this happens in BSD with high probability, this is occasional.
3514          * Any application relying on this is buggy. Note also, that fix "works"
3515          * only in this artificial test. Insert some normal data between A and B and we will
3516          * decline of BSD again. Verdict: it is better to remove to trap
3517          * buggy users.
3518          */
3519         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3520             !sock_flag(sk, SOCK_URGINLINE) &&
3521             tp->copied_seq != tp->rcv_nxt) {
3522                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3523                 tp->copied_seq++;
3524                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3525                         __skb_unlink(skb, &sk->sk_receive_queue);
3526                         __kfree_skb(skb);
3527                 }
3528         }
3529
3530         tp->urg_data   = TCP_URG_NOTYET;
3531         tp->urg_seq    = ptr;
3532
3533         /* Disable header prediction. */
3534         tp->pred_flags = 0;
3535 }
3536
3537 /* This is the 'fast' part of urgent handling. */
3538 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3539 {
3540         struct tcp_sock *tp = tcp_sk(sk);
3541
3542         /* Check if we get a new urgent pointer - normally not. */
3543         if (th->urg)
3544                 tcp_check_urg(sk,th);
3545
3546         /* Do we wait for any urgent data? - normally not... */
3547         if (tp->urg_data == TCP_URG_NOTYET) {
3548                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3549                           th->syn;
3550
3551                 /* Is the urgent pointer pointing into this packet? */   
3552                 if (ptr < skb->len) {
3553                         u8 tmp;
3554                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3555                                 BUG();
3556                         tp->urg_data = TCP_URG_VALID | tmp;
3557                         if (!sock_flag(sk, SOCK_DEAD))
3558                                 sk->sk_data_ready(sk, 0);
3559                 }
3560         }
3561 }
3562
3563 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3564 {
3565         struct tcp_sock *tp = tcp_sk(sk);
3566         int chunk = skb->len - hlen;
3567         int err;
3568
3569         local_bh_enable();
3570         if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3571                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3572         else
3573                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3574                                                        tp->ucopy.iov);
3575
3576         if (!err) {
3577                 tp->ucopy.len -= chunk;
3578                 tp->copied_seq += chunk;
3579                 tcp_rcv_space_adjust(sk);
3580         }
3581
3582         local_bh_disable();
3583         return err;
3584 }
3585
3586 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3587 {
3588         int result;
3589
3590         if (sock_owned_by_user(sk)) {
3591                 local_bh_enable();
3592                 result = __tcp_checksum_complete(skb);
3593                 local_bh_disable();
3594         } else {
3595                 result = __tcp_checksum_complete(skb);
3596         }
3597         return result;
3598 }
3599
3600 static __inline__ int
3601 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3602 {
3603         return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3604                 __tcp_checksum_complete_user(sk, skb);
3605 }
3606
3607 /*
3608  *      TCP receive function for the ESTABLISHED state. 
3609  *
3610  *      It is split into a fast path and a slow path. The fast path is 
3611  *      disabled when:
3612  *      - A zero window was announced from us - zero window probing
3613  *        is only handled properly in the slow path. 
3614  *      - Out of order segments arrived.
3615  *      - Urgent data is expected.
3616  *      - There is no buffer space left
3617  *      - Unexpected TCP flags/window values/header lengths are received
3618  *        (detected by checking the TCP header against pred_flags) 
3619  *      - Data is sent in both directions. Fast path only supports pure senders
3620  *        or pure receivers (this means either the sequence number or the ack
3621  *        value must stay constant)
3622  *      - Unexpected TCP option.
3623  *
3624  *      When these conditions are not satisfied it drops into a standard 
3625  *      receive procedure patterned after RFC793 to handle all cases.
3626  *      The first three cases are guaranteed by proper pred_flags setting,
3627  *      the rest is checked inline. Fast processing is turned on in 
3628  *      tcp_data_queue when everything is OK.
3629  */
3630 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3631                         struct tcphdr *th, unsigned len)
3632 {
3633         struct tcp_sock *tp = tcp_sk(sk);
3634
3635         /*
3636          *      Header prediction.
3637          *      The code loosely follows the one in the famous 
3638          *      "30 instruction TCP receive" Van Jacobson mail.
3639          *      
3640          *      Van's trick is to deposit buffers into socket queue 
3641          *      on a device interrupt, to call tcp_recv function
3642          *      on the receive process context and checksum and copy
3643          *      the buffer to user space. smart...
3644          *
3645          *      Our current scheme is not silly either but we take the 
3646          *      extra cost of the net_bh soft interrupt processing...
3647          *      We do checksum and copy also but from device to kernel.
3648          */
3649
3650         tp->rx_opt.saw_tstamp = 0;
3651
3652         /*      pred_flags is 0xS?10 << 16 + snd_wnd
3653          *      if header_predition is to be made
3654          *      'S' will always be tp->tcp_header_len >> 2
3655          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3656          *  turn it off (when there are holes in the receive 
3657          *       space for instance)
3658          *      PSH flag is ignored.
3659          */
3660
3661         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3662                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3663                 int tcp_header_len = tp->tcp_header_len;
3664
3665                 /* Timestamp header prediction: tcp_header_len
3666                  * is automatically equal to th->doff*4 due to pred_flags
3667                  * match.
3668                  */
3669
3670                 /* Check timestamp */
3671                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3672                         __u32 *ptr = (__u32 *)(th + 1);
3673
3674                         /* No? Slow path! */
3675                         if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3676                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3677                                 goto slow_path;
3678
3679                         tp->rx_opt.saw_tstamp = 1;
3680                         ++ptr; 
3681                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3682                         ++ptr;
3683                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3684
3685                         /* If PAWS failed, check it more carefully in slow path */
3686                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
3687                                 goto slow_path;
3688
3689                         /* DO NOT update ts_recent here, if checksum fails
3690                          * and timestamp was corrupted part, it will result
3691                          * in a hung connection since we will drop all
3692                          * future packets due to the PAWS test.
3693                          */
3694                 }
3695
3696                 if (len <= tcp_header_len) {
3697                         /* Bulk data transfer: sender */
3698                         if (len == tcp_header_len) {
3699                                 /* Predicted packet is in window by definition.
3700                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3701                                  * Hence, check seq<=rcv_wup reduces to:
3702                                  */
3703                                 if (tcp_header_len ==
3704                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3705                                     tp->rcv_nxt == tp->rcv_wup)
3706                                         tcp_store_ts_recent(tp);
3707
3708                                 tcp_rcv_rtt_measure_ts(sk, skb);
3709
3710                                 /* We know that such packets are checksummed
3711                                  * on entry.
3712                                  */
3713                                 tcp_ack(sk, skb, 0);
3714                                 __kfree_skb(skb); 
3715                                 tcp_data_snd_check(sk, tp);
3716                                 return 0;
3717                         } else { /* Header too small */
3718                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3719                                 goto discard;
3720                         }
3721                 } else {
3722                         int eaten = 0;
3723
3724                         if (tp->ucopy.task == current &&
3725                             tp->copied_seq == tp->rcv_nxt &&
3726                             len - tcp_header_len <= tp->ucopy.len &&
3727                             sock_owned_by_user(sk)) {
3728                                 __set_current_state(TASK_RUNNING);
3729
3730                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
3731                                         /* Predicted packet is in window by definition.
3732                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3733                                          * Hence, check seq<=rcv_wup reduces to:
3734                                          */
3735                                         if (tcp_header_len ==
3736                                             (sizeof(struct tcphdr) +
3737                                              TCPOLEN_TSTAMP_ALIGNED) &&
3738                                             tp->rcv_nxt == tp->rcv_wup)
3739                                                 tcp_store_ts_recent(tp);
3740
3741                                         tcp_rcv_rtt_measure_ts(sk, skb);
3742
3743                                         __skb_pull(skb, tcp_header_len);
3744                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3745                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
3746                                         eaten = 1;
3747                                 }
3748                         }
3749                         if (!eaten) {
3750                                 if (tcp_checksum_complete_user(sk, skb))
3751                                         goto csum_error;
3752
3753                                 /* Predicted packet is in window by definition.
3754                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3755                                  * Hence, check seq<=rcv_wup reduces to:
3756                                  */
3757                                 if (tcp_header_len ==
3758                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3759                                     tp->rcv_nxt == tp->rcv_wup)
3760                                         tcp_store_ts_recent(tp);
3761
3762                                 tcp_rcv_rtt_measure_ts(sk, skb);
3763
3764                                 if ((int)skb->truesize > sk->sk_forward_alloc)
3765                                         goto step5;
3766
3767                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
3768
3769                                 /* Bulk data transfer: receiver */
3770                                 __skb_pull(skb,tcp_header_len);
3771                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3772                                 sk_stream_set_owner_r(skb, sk);
3773                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3774                         }
3775
3776                         tcp_event_data_recv(sk, tp, skb);
3777
3778                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
3779                                 /* Well, only one small jumplet in fast path... */
3780                                 tcp_ack(sk, skb, FLAG_DATA);
3781                                 tcp_data_snd_check(sk, tp);
3782                                 if (!inet_csk_ack_scheduled(sk))
3783                                         goto no_ack;
3784                         }
3785
3786                         __tcp_ack_snd_check(sk, 0);
3787 no_ack:
3788                         if (eaten)
3789                                 __kfree_skb(skb);
3790                         else
3791                                 sk->sk_data_ready(sk, 0);
3792                         return 0;
3793                 }
3794         }
3795
3796 slow_path:
3797         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
3798                 goto csum_error;
3799
3800         /*
3801          * RFC1323: H1. Apply PAWS check first.
3802          */
3803         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
3804             tcp_paws_discard(sk, skb)) {
3805                 if (!th->rst) {
3806                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
3807                         tcp_send_dupack(sk, skb);
3808                         goto discard;
3809                 }
3810                 /* Resets are accepted even if PAWS failed.
3811
3812                    ts_recent update must be made after we are sure
3813                    that the packet is in window.
3814                  */
3815         }
3816
3817         /*
3818          *      Standard slow path.
3819          */
3820
3821         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3822                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3823                  * (RST) segments are validated by checking their SEQ-fields."
3824                  * And page 69: "If an incoming segment is not acceptable,
3825                  * an acknowledgment should be sent in reply (unless the RST bit
3826                  * is set, if so drop the segment and return)".
3827                  */
3828                 if (!th->rst)
3829                         tcp_send_dupack(sk, skb);
3830                 goto discard;
3831         }
3832
3833         if(th->rst) {
3834                 tcp_reset(sk);
3835                 goto discard;
3836         }
3837
3838         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3839
3840         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3841                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3842                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
3843                 tcp_reset(sk);
3844                 return 1;
3845         }
3846
3847 step5:
3848         if(th->ack)
3849                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3850
3851         tcp_rcv_rtt_measure_ts(sk, skb);
3852
3853         /* Process urgent data. */
3854         tcp_urg(sk, skb, th);
3855
3856         /* step 7: process the segment text */
3857         tcp_data_queue(sk, skb);
3858
3859         tcp_data_snd_check(sk, tp);
3860         tcp_ack_snd_check(sk);
3861         return 0;
3862
3863 csum_error:
3864         TCP_INC_STATS_BH(TCP_MIB_INERRS);
3865
3866 discard:
3867         __kfree_skb(skb);
3868         return 0;
3869 }
3870
3871 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
3872                                          struct tcphdr *th, unsigned len)
3873 {
3874         struct tcp_sock *tp = tcp_sk(sk);
3875         int saved_clamp = tp->rx_opt.mss_clamp;
3876
3877         tcp_parse_options(skb, &tp->rx_opt, 0);
3878
3879         if (th->ack) {
3880                 struct inet_connection_sock *icsk;
3881                 /* rfc793:
3882                  * "If the state is SYN-SENT then
3883                  *    first check the ACK bit
3884                  *      If the ACK bit is set
3885                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3886                  *        a reset (unless the RST bit is set, if so drop
3887                  *        the segment and return)"
3888                  *
3889                  *  We do not send data with SYN, so that RFC-correct
3890                  *  test reduces to:
3891                  */
3892                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
3893                         goto reset_and_undo;
3894
3895                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3896                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
3897                              tcp_time_stamp)) {
3898                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
3899                         goto reset_and_undo;
3900                 }
3901
3902                 /* Now ACK is acceptable.
3903                  *
3904                  * "If the RST bit is set
3905                  *    If the ACK was acceptable then signal the user "error:
3906                  *    connection reset", drop the segment, enter CLOSED state,
3907                  *    delete TCB, and return."
3908                  */
3909
3910                 if (th->rst) {
3911                         tcp_reset(sk);
3912                         goto discard;
3913                 }
3914
3915                 /* rfc793:
3916                  *   "fifth, if neither of the SYN or RST bits is set then
3917                  *    drop the segment and return."
3918                  *
3919                  *    See note below!
3920                  *                                        --ANK(990513)
3921                  */
3922                 if (!th->syn)
3923                         goto discard_and_undo;
3924
3925                 /* rfc793:
3926                  *   "If the SYN bit is on ...
3927                  *    are acceptable then ...
3928                  *    (our SYN has been ACKed), change the connection
3929                  *    state to ESTABLISHED..."
3930                  */
3931
3932                 TCP_ECN_rcv_synack(tp, th);
3933                 if (tp->ecn_flags&TCP_ECN_OK)
3934                         sock_set_flag(sk, SOCK_NO_LARGESEND);
3935
3936                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
3937                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3938
3939                 /* Ok.. it's good. Set up sequence numbers and
3940                  * move to established.
3941                  */
3942                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
3943                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
3944
3945                 /* RFC1323: The window in SYN & SYN/ACK segments is
3946                  * never scaled.
3947                  */
3948                 tp->snd_wnd = ntohs(th->window);
3949                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
3950
3951                 if (!tp->rx_opt.wscale_ok) {
3952                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
3953                         tp->window_clamp = min(tp->window_clamp, 65535U);
3954                 }
3955
3956                 if (tp->rx_opt.saw_tstamp) {
3957                         tp->rx_opt.tstamp_ok       = 1;
3958                         tp->tcp_header_len =
3959                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
3960                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
3961                         tcp_store_ts_recent(tp);
3962                 } else {
3963                         tp->tcp_header_len = sizeof(struct tcphdr);
3964                 }
3965
3966                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
3967                         tp->rx_opt.sack_ok |= 2;
3968
3969                 tcp_sync_mss(sk, tp->pmtu_cookie);
3970                 tcp_initialize_rcv_mss(sk);
3971
3972                 /* Remember, tcp_poll() does not lock socket!
3973                  * Change state from SYN-SENT only after copied_seq
3974                  * is initialized. */
3975                 tp->copied_seq = tp->rcv_nxt;
3976                 mb();
3977                 tcp_set_state(sk, TCP_ESTABLISHED);
3978
3979                 /* Make sure socket is routed, for correct metrics.  */
3980                 tp->af_specific->rebuild_header(sk);
3981
3982                 tcp_init_metrics(sk);
3983
3984                 tcp_init_congestion_control(sk);
3985
3986                 /* Prevent spurious tcp_cwnd_restart() on first data
3987                  * packet.
3988                  */
3989                 tp->lsndtime = tcp_time_stamp;
3990
3991                 tcp_init_buffer_space(sk);
3992
3993                 if (sock_flag(sk, SOCK_KEEPOPEN))
3994                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
3995
3996                 if (!tp->rx_opt.snd_wscale)
3997                         __tcp_fast_path_on(tp, tp->snd_wnd);
3998                 else
3999                         tp->pred_flags = 0;
4000
4001                 if (!sock_flag(sk, SOCK_DEAD)) {
4002                         sk->sk_state_change(sk);
4003                         sk_wake_async(sk, 0, POLL_OUT);
4004                 }
4005
4006                 icsk = inet_csk(sk);
4007
4008                 if (sk->sk_write_pending ||
4009                     icsk->icsk_accept_queue.rskq_defer_accept ||
4010                     icsk->icsk_ack.pingpong) {
4011                         /* Save one ACK. Data will be ready after
4012                          * several ticks, if write_pending is set.
4013                          *
4014                          * It may be deleted, but with this feature tcpdumps
4015                          * look so _wonderfully_ clever, that I was not able
4016                          * to stand against the temptation 8)     --ANK
4017                          */
4018                         inet_csk_schedule_ack(sk);
4019                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4020                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4021                         tcp_incr_quickack(sk);
4022                         tcp_enter_quickack_mode(sk);
4023                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4024                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4025
4026 discard:
4027                         __kfree_skb(skb);
4028                         return 0;
4029                 } else {
4030                         tcp_send_ack(sk);
4031                 }
4032                 return -1;
4033         }
4034
4035         /* No ACK in the segment */
4036
4037         if (th->rst) {
4038                 /* rfc793:
4039                  * "If the RST bit is set
4040                  *
4041                  *      Otherwise (no ACK) drop the segment and return."
4042                  */
4043
4044                 goto discard_and_undo;
4045         }
4046
4047         /* PAWS check. */
4048         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4049                 goto discard_and_undo;
4050
4051         if (th->syn) {
4052                 /* We see SYN without ACK. It is attempt of
4053                  * simultaneous connect with crossed SYNs.
4054                  * Particularly, it can be connect to self.
4055                  */
4056                 tcp_set_state(sk, TCP_SYN_RECV);
4057
4058                 if (tp->rx_opt.saw_tstamp) {
4059                         tp->rx_opt.tstamp_ok = 1;
4060                         tcp_store_ts_recent(tp);
4061                         tp->tcp_header_len =
4062                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4063                 } else {
4064                         tp->tcp_header_len = sizeof(struct tcphdr);
4065                 }
4066
4067                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4068                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4069
4070                 /* RFC1323: The window in SYN & SYN/ACK segments is
4071                  * never scaled.
4072                  */
4073                 tp->snd_wnd    = ntohs(th->window);
4074                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4075                 tp->max_window = tp->snd_wnd;
4076
4077                 TCP_ECN_rcv_syn(tp, th);
4078                 if (tp->ecn_flags&TCP_ECN_OK)
4079                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4080
4081                 tcp_sync_mss(sk, tp->pmtu_cookie);
4082                 tcp_initialize_rcv_mss(sk);
4083
4084
4085                 tcp_send_synack(sk);
4086 #if 0
4087                 /* Note, we could accept data and URG from this segment.
4088                  * There are no obstacles to make this.
4089                  *
4090                  * However, if we ignore data in ACKless segments sometimes,
4091                  * we have no reasons to accept it sometimes.
4092                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4093                  * is not flawless. So, discard packet for sanity.
4094                  * Uncomment this return to process the data.
4095                  */
4096                 return -1;
4097 #else
4098                 goto discard;
4099 #endif
4100         }
4101         /* "fifth, if neither of the SYN or RST bits is set then
4102          * drop the segment and return."
4103          */
4104
4105 discard_and_undo:
4106         tcp_clear_options(&tp->rx_opt);
4107         tp->rx_opt.mss_clamp = saved_clamp;
4108         goto discard;
4109
4110 reset_and_undo:
4111         tcp_clear_options(&tp->rx_opt);
4112         tp->rx_opt.mss_clamp = saved_clamp;
4113         return 1;
4114 }
4115
4116
4117 /*
4118  *      This function implements the receiving procedure of RFC 793 for
4119  *      all states except ESTABLISHED and TIME_WAIT. 
4120  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4121  *      address independent.
4122  */
4123         
4124 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4125                           struct tcphdr *th, unsigned len)
4126 {
4127         struct tcp_sock *tp = tcp_sk(sk);
4128         int queued = 0;
4129
4130         tp->rx_opt.saw_tstamp = 0;
4131
4132         switch (sk->sk_state) {
4133         case TCP_CLOSE:
4134                 goto discard;
4135
4136         case TCP_LISTEN:
4137                 if(th->ack)
4138                         return 1;
4139
4140                 if(th->rst)
4141                         goto discard;
4142
4143                 if(th->syn) {
4144                         if(tp->af_specific->conn_request(sk, skb) < 0)
4145                                 return 1;
4146
4147                         /* Now we have several options: In theory there is 
4148                          * nothing else in the frame. KA9Q has an option to 
4149                          * send data with the syn, BSD accepts data with the
4150                          * syn up to the [to be] advertised window and 
4151                          * Solaris 2.1 gives you a protocol error. For now 
4152                          * we just ignore it, that fits the spec precisely 
4153                          * and avoids incompatibilities. It would be nice in
4154                          * future to drop through and process the data.
4155                          *
4156                          * Now that TTCP is starting to be used we ought to 
4157                          * queue this data.
4158                          * But, this leaves one open to an easy denial of
4159                          * service attack, and SYN cookies can't defend
4160                          * against this problem. So, we drop the data
4161                          * in the interest of security over speed.
4162                          */
4163                         goto discard;
4164                 }
4165                 goto discard;
4166
4167         case TCP_SYN_SENT:
4168                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4169                 if (queued >= 0)
4170                         return queued;
4171
4172                 /* Do step6 onward by hand. */
4173                 tcp_urg(sk, skb, th);
4174                 __kfree_skb(skb);
4175                 tcp_data_snd_check(sk, tp);
4176                 return 0;
4177         }
4178
4179         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4180             tcp_paws_discard(sk, skb)) {
4181                 if (!th->rst) {
4182                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4183                         tcp_send_dupack(sk, skb);
4184                         goto discard;
4185                 }
4186                 /* Reset is accepted even if it did not pass PAWS. */
4187         }
4188
4189         /* step 1: check sequence number */
4190         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4191                 if (!th->rst)
4192                         tcp_send_dupack(sk, skb);
4193                 goto discard;
4194         }
4195
4196         /* step 2: check RST bit */
4197         if(th->rst) {
4198                 tcp_reset(sk);
4199                 goto discard;
4200         }
4201
4202         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4203
4204         /* step 3: check security and precedence [ignored] */
4205
4206         /*      step 4:
4207          *
4208          *      Check for a SYN in window.
4209          */
4210         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4211                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4212                 tcp_reset(sk);
4213                 return 1;
4214         }
4215
4216         /* step 5: check the ACK field */
4217         if (th->ack) {
4218                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4219
4220                 switch(sk->sk_state) {
4221                 case TCP_SYN_RECV:
4222                         if (acceptable) {
4223                                 tp->copied_seq = tp->rcv_nxt;
4224                                 mb();
4225                                 tcp_set_state(sk, TCP_ESTABLISHED);
4226                                 sk->sk_state_change(sk);
4227
4228                                 /* Note, that this wakeup is only for marginal
4229                                  * crossed SYN case. Passively open sockets
4230                                  * are not waked up, because sk->sk_sleep ==
4231                                  * NULL and sk->sk_socket == NULL.
4232                                  */
4233                                 if (sk->sk_socket) {
4234                                         sk_wake_async(sk,0,POLL_OUT);
4235                                 }
4236
4237                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4238                                 tp->snd_wnd = ntohs(th->window) <<
4239                                               tp->rx_opt.snd_wscale;
4240                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4241                                             TCP_SKB_CB(skb)->seq);
4242
4243                                 /* tcp_ack considers this ACK as duplicate
4244                                  * and does not calculate rtt.
4245                                  * Fix it at least with timestamps.
4246                                  */
4247                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4248                                     !tp->srtt)
4249                                         tcp_ack_saw_tstamp(sk, 0);
4250
4251                                 if (tp->rx_opt.tstamp_ok)
4252                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4253
4254                                 /* Make sure socket is routed, for
4255                                  * correct metrics.
4256                                  */
4257                                 tp->af_specific->rebuild_header(sk);
4258
4259                                 tcp_init_metrics(sk);
4260
4261                                 tcp_init_congestion_control(sk);
4262
4263                                 /* Prevent spurious tcp_cwnd_restart() on
4264                                  * first data packet.
4265                                  */
4266                                 tp->lsndtime = tcp_time_stamp;
4267
4268                                 tcp_initialize_rcv_mss(sk);
4269                                 tcp_init_buffer_space(sk);
4270                                 tcp_fast_path_on(tp);
4271                         } else {
4272                                 return 1;
4273                         }
4274                         break;
4275
4276                 case TCP_FIN_WAIT1:
4277                         if (tp->snd_una == tp->write_seq) {
4278                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4279                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4280                                 dst_confirm(sk->sk_dst_cache);
4281
4282                                 if (!sock_flag(sk, SOCK_DEAD))
4283                                         /* Wake up lingering close() */
4284                                         sk->sk_state_change(sk);
4285                                 else {
4286                                         int tmo;
4287
4288                                         if (tp->linger2 < 0 ||
4289                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4290                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4291                                                 tcp_done(sk);
4292                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4293                                                 return 1;
4294                                         }
4295
4296                                         tmo = tcp_fin_time(sk);
4297                                         if (tmo > TCP_TIMEWAIT_LEN) {
4298                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4299                                         } else if (th->fin || sock_owned_by_user(sk)) {
4300                                                 /* Bad case. We could lose such FIN otherwise.
4301                                                  * It is not a big problem, but it looks confusing
4302                                                  * and not so rare event. We still can lose it now,
4303                                                  * if it spins in bh_lock_sock(), but it is really
4304                                                  * marginal case.
4305                                                  */
4306                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4307                                         } else {
4308                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4309                                                 goto discard;
4310                                         }
4311                                 }
4312                         }
4313                         break;
4314
4315                 case TCP_CLOSING:
4316                         if (tp->snd_una == tp->write_seq) {
4317                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4318                                 goto discard;
4319                         }
4320                         break;
4321
4322                 case TCP_LAST_ACK:
4323                         if (tp->snd_una == tp->write_seq) {
4324                                 tcp_update_metrics(sk);
4325                                 tcp_done(sk);
4326                                 goto discard;
4327                         }
4328                         break;
4329                 }
4330         } else
4331                 goto discard;
4332
4333         /* step 6: check the URG bit */
4334         tcp_urg(sk, skb, th);
4335
4336         /* step 7: process the segment text */
4337         switch (sk->sk_state) {
4338         case TCP_CLOSE_WAIT:
4339         case TCP_CLOSING:
4340         case TCP_LAST_ACK:
4341                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4342                         break;
4343         case TCP_FIN_WAIT1:
4344         case TCP_FIN_WAIT2:
4345                 /* RFC 793 says to queue data in these states,
4346                  * RFC 1122 says we MUST send a reset. 
4347                  * BSD 4.4 also does reset.
4348                  */
4349                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4350                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4351                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4352                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4353                                 tcp_reset(sk);
4354                                 return 1;
4355                         }
4356                 }
4357                 /* Fall through */
4358         case TCP_ESTABLISHED: 
4359                 tcp_data_queue(sk, skb);
4360                 queued = 1;
4361                 break;
4362         }
4363
4364         /* tcp_data could move socket to TIME-WAIT */
4365         if (sk->sk_state != TCP_CLOSE) {
4366                 tcp_data_snd_check(sk, tp);
4367                 tcp_ack_snd_check(sk);
4368         }
4369
4370         if (!queued) { 
4371 discard:
4372                 __kfree_skb(skb);
4373         }
4374         return 0;
4375 }
4376
4377 EXPORT_SYMBOL(sysctl_tcp_ecn);
4378 EXPORT_SYMBOL(sysctl_tcp_reordering);
4379 EXPORT_SYMBOL(sysctl_tcp_abc);
4380 EXPORT_SYMBOL(tcp_parse_options);
4381 EXPORT_SYMBOL(tcp_rcv_established);
4382 EXPORT_SYMBOL(tcp_rcv_state_process);