Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/agpgart
[linux-2.6] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the 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
93 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
94 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
95 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
96 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
97 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
98 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
99 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
100 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
101 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
102
103 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
106 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
107
108 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
109 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
110 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
111
112 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
113
114 /* Adapt the MSS value used to make delayed ack decision to the 
115  * real world.
116  */ 
117 static inline void tcp_measure_rcv_mss(struct sock *sk,
118                                        const struct sk_buff *skb)
119 {
120         struct inet_connection_sock *icsk = inet_csk(sk);
121         const unsigned int lss = icsk->icsk_ack.last_seg_size; 
122         unsigned int len;
123
124         icsk->icsk_ack.last_seg_size = 0; 
125
126         /* skb->len may jitter because of SACKs, even if peer
127          * sends good full-sized frames.
128          */
129         len = skb->len;
130         if (len >= icsk->icsk_ack.rcv_mss) {
131                 icsk->icsk_ack.rcv_mss = len;
132         } else {
133                 /* Otherwise, we make more careful check taking into account,
134                  * that SACKs block is variable.
135                  *
136                  * "len" is invariant segment length, including TCP header.
137                  */
138                 len += skb->data - skb->h.raw;
139                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
140                     /* If PSH is not set, packet should be
141                      * full sized, provided peer TCP is not badly broken.
142                      * This observation (if it is correct 8)) allows
143                      * to handle super-low mtu links fairly.
144                      */
145                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
146                      !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
147                         /* Subtract also invariant (if peer is RFC compliant),
148                          * tcp header plus fixed timestamp option length.
149                          * Resulting "len" is MSS free of SACK jitter.
150                          */
151                         len -= tcp_sk(sk)->tcp_header_len;
152                         icsk->icsk_ack.last_seg_size = len;
153                         if (len == lss) {
154                                 icsk->icsk_ack.rcv_mss = len;
155                                 return;
156                         }
157                 }
158                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
159         }
160 }
161
162 static void tcp_incr_quickack(struct sock *sk)
163 {
164         struct inet_connection_sock *icsk = inet_csk(sk);
165         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
166
167         if (quickacks==0)
168                 quickacks=2;
169         if (quickacks > icsk->icsk_ack.quick)
170                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
171 }
172
173 void tcp_enter_quickack_mode(struct sock *sk)
174 {
175         struct inet_connection_sock *icsk = inet_csk(sk);
176         tcp_incr_quickack(sk);
177         icsk->icsk_ack.pingpong = 0;
178         icsk->icsk_ack.ato = TCP_ATO_MIN;
179 }
180
181 /* Send ACKs quickly, if "quick" count is not exhausted
182  * and the session is not interactive.
183  */
184
185 static inline int tcp_in_quickack_mode(const struct sock *sk)
186 {
187         const struct inet_connection_sock *icsk = inet_csk(sk);
188         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
189 }
190
191 /* Buffer size and advertised window tuning.
192  *
193  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
194  */
195
196 static void tcp_fixup_sndbuf(struct sock *sk)
197 {
198         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
199                      sizeof(struct sk_buff);
200
201         if (sk->sk_sndbuf < 3 * sndmem)
202                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
203 }
204
205 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
206  *
207  * All tcp_full_space() is split to two parts: "network" buffer, allocated
208  * forward and advertised in receiver window (tp->rcv_wnd) and
209  * "application buffer", required to isolate scheduling/application
210  * latencies from network.
211  * window_clamp is maximal advertised window. It can be less than
212  * tcp_full_space(), in this case tcp_full_space() - window_clamp
213  * is reserved for "application" buffer. The less window_clamp is
214  * the smoother our behaviour from viewpoint of network, but the lower
215  * throughput and the higher sensitivity of the connection to losses. 8)
216  *
217  * rcv_ssthresh is more strict window_clamp used at "slow start"
218  * phase to predict further behaviour of this connection.
219  * It is used for two goals:
220  * - to enforce header prediction at sender, even when application
221  *   requires some significant "application buffer". It is check #1.
222  * - to prevent pruning of receive queue because of misprediction
223  *   of receiver window. Check #2.
224  *
225  * The scheme does not work when sender sends good segments opening
226  * window and then starts to feed us spagetti. But it should work
227  * in common situations. Otherwise, we have to rely on queue collapsing.
228  */
229
230 /* Slow part of check#2. */
231 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
232                              const struct sk_buff *skb)
233 {
234         /* Optimize this! */
235         int truesize = tcp_win_from_space(skb->truesize)/2;
236         int window = tcp_full_space(sk)/2;
237
238         while (tp->rcv_ssthresh <= window) {
239                 if (truesize <= skb->len)
240                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
241
242                 truesize >>= 1;
243                 window >>= 1;
244         }
245         return 0;
246 }
247
248 static inline void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
249                                    struct sk_buff *skb)
250 {
251         /* Check #1 */
252         if (tp->rcv_ssthresh < tp->window_clamp &&
253             (int)tp->rcv_ssthresh < tcp_space(sk) &&
254             !tcp_memory_pressure) {
255                 int incr;
256
257                 /* Check #2. Increase window, if skb with such overhead
258                  * will fit to rcvbuf in future.
259                  */
260                 if (tcp_win_from_space(skb->truesize) <= skb->len)
261                         incr = 2*tp->advmss;
262                 else
263                         incr = __tcp_grow_window(sk, tp, skb);
264
265                 if (incr) {
266                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
267                         inet_csk(sk)->icsk_ack.quick |= 1;
268                 }
269         }
270 }
271
272 /* 3. Tuning rcvbuf, when connection enters established state. */
273
274 static void tcp_fixup_rcvbuf(struct sock *sk)
275 {
276         struct tcp_sock *tp = tcp_sk(sk);
277         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
278
279         /* Try to select rcvbuf so that 4 mss-sized segments
280          * will fit to window and correspoding skbs will fit to our rcvbuf.
281          * (was 3; 4 is minimum to allow fast retransmit to work.)
282          */
283         while (tcp_win_from_space(rcvmem) < tp->advmss)
284                 rcvmem += 128;
285         if (sk->sk_rcvbuf < 4 * rcvmem)
286                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
287 }
288
289 /* 4. Try to fixup all. It is made iimediately after connection enters
290  *    established state.
291  */
292 static void tcp_init_buffer_space(struct sock *sk)
293 {
294         struct tcp_sock *tp = tcp_sk(sk);
295         int maxwin;
296
297         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
298                 tcp_fixup_rcvbuf(sk);
299         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
300                 tcp_fixup_sndbuf(sk);
301
302         tp->rcvq_space.space = tp->rcv_wnd;
303
304         maxwin = tcp_full_space(sk);
305
306         if (tp->window_clamp >= maxwin) {
307                 tp->window_clamp = maxwin;
308
309                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
310                         tp->window_clamp = max(maxwin -
311                                                (maxwin >> sysctl_tcp_app_win),
312                                                4 * tp->advmss);
313         }
314
315         /* Force reservation of one segment. */
316         if (sysctl_tcp_app_win &&
317             tp->window_clamp > 2 * tp->advmss &&
318             tp->window_clamp + tp->advmss > maxwin)
319                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
320
321         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
322         tp->snd_cwnd_stamp = tcp_time_stamp;
323 }
324
325 /* 5. Recalculate window clamp after socket hit its memory bounds. */
326 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
327 {
328         struct inet_connection_sock *icsk = inet_csk(sk);
329         struct sk_buff *skb;
330         unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
331         int ofo_win = 0;
332
333         icsk->icsk_ack.quick = 0;
334
335         skb_queue_walk(&tp->out_of_order_queue, skb) {
336                 ofo_win += skb->len;
337         }
338
339         /* If overcommit is due to out of order segments,
340          * do not clamp window. Try to expand rcvbuf instead.
341          */
342         if (ofo_win) {
343                 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
344                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
345                     !tcp_memory_pressure &&
346                     atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
347                         sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
348                                             sysctl_tcp_rmem[2]);
349         }
350         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
351                 app_win += ofo_win;
352                 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
353                         app_win >>= 1;
354                 if (app_win > icsk->icsk_ack.rcv_mss)
355                         app_win -= icsk->icsk_ack.rcv_mss;
356                 app_win = max(app_win, 2U*tp->advmss);
357
358                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
359         }
360 }
361
362 /* Receiver "autotuning" code.
363  *
364  * The algorithm for RTT estimation w/o timestamps is based on
365  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
366  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
367  *
368  * More detail on this code can be found at
369  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
370  * though this reference is out of date.  A new paper
371  * is pending.
372  */
373 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
374 {
375         u32 new_sample = tp->rcv_rtt_est.rtt;
376         long m = sample;
377
378         if (m == 0)
379                 m = 1;
380
381         if (new_sample != 0) {
382                 /* If we sample in larger samples in the non-timestamp
383                  * case, we could grossly overestimate the RTT especially
384                  * with chatty applications or bulk transfer apps which
385                  * are stalled on filesystem I/O.
386                  *
387                  * Also, since we are only going for a minimum in the
388                  * non-timestamp case, we do not smoothe things out
389                  * else with timestamps disabled convergance takes too
390                  * long.
391                  */
392                 if (!win_dep) {
393                         m -= (new_sample >> 3);
394                         new_sample += m;
395                 } else if (m < new_sample)
396                         new_sample = m << 3;
397         } else {
398                 /* No previous mesaure. */
399                 new_sample = m << 3;
400         }
401
402         if (tp->rcv_rtt_est.rtt != new_sample)
403                 tp->rcv_rtt_est.rtt = new_sample;
404 }
405
406 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
407 {
408         if (tp->rcv_rtt_est.time == 0)
409                 goto new_measure;
410         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
411                 return;
412         tcp_rcv_rtt_update(tp,
413                            jiffies - tp->rcv_rtt_est.time,
414                            1);
415
416 new_measure:
417         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
418         tp->rcv_rtt_est.time = tcp_time_stamp;
419 }
420
421 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
422 {
423         struct tcp_sock *tp = tcp_sk(sk);
424         if (tp->rx_opt.rcv_tsecr &&
425             (TCP_SKB_CB(skb)->end_seq -
426              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
427                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
428 }
429
430 /*
431  * This function should be called every time data is copied to user space.
432  * It calculates the appropriate TCP receive buffer space.
433  */
434 void tcp_rcv_space_adjust(struct sock *sk)
435 {
436         struct tcp_sock *tp = tcp_sk(sk);
437         int time;
438         int space;
439         
440         if (tp->rcvq_space.time == 0)
441                 goto new_measure;
442         
443         time = tcp_time_stamp - tp->rcvq_space.time;
444         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
445             tp->rcv_rtt_est.rtt == 0)
446                 return;
447         
448         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
449
450         space = max(tp->rcvq_space.space, space);
451
452         if (tp->rcvq_space.space != space) {
453                 int rcvmem;
454
455                 tp->rcvq_space.space = space;
456
457                 if (sysctl_tcp_moderate_rcvbuf) {
458                         int new_clamp = space;
459
460                         /* Receive space grows, normalize in order to
461                          * take into account packet headers and sk_buff
462                          * structure overhead.
463                          */
464                         space /= tp->advmss;
465                         if (!space)
466                                 space = 1;
467                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
468                                   16 + sizeof(struct sk_buff));
469                         while (tcp_win_from_space(rcvmem) < tp->advmss)
470                                 rcvmem += 128;
471                         space *= rcvmem;
472                         space = min(space, sysctl_tcp_rmem[2]);
473                         if (space > sk->sk_rcvbuf) {
474                                 sk->sk_rcvbuf = space;
475
476                                 /* Make the window clamp follow along.  */
477                                 tp->window_clamp = new_clamp;
478                         }
479                 }
480         }
481         
482 new_measure:
483         tp->rcvq_space.seq = tp->copied_seq;
484         tp->rcvq_space.time = tcp_time_stamp;
485 }
486
487 /* There is something which you must keep in mind when you analyze the
488  * behavior of the tp->ato delayed ack timeout interval.  When a
489  * connection starts up, we want to ack as quickly as possible.  The
490  * problem is that "good" TCP's do slow start at the beginning of data
491  * transmission.  The means that until we send the first few ACK's the
492  * sender will sit on his end and only queue most of his data, because
493  * he can only send snd_cwnd unacked packets at any given time.  For
494  * each ACK we send, he increments snd_cwnd and transmits more of his
495  * queue.  -DaveM
496  */
497 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
498 {
499         struct inet_connection_sock *icsk = inet_csk(sk);
500         u32 now;
501
502         inet_csk_schedule_ack(sk);
503
504         tcp_measure_rcv_mss(sk, skb);
505
506         tcp_rcv_rtt_measure(tp);
507         
508         now = tcp_time_stamp;
509
510         if (!icsk->icsk_ack.ato) {
511                 /* The _first_ data packet received, initialize
512                  * delayed ACK engine.
513                  */
514                 tcp_incr_quickack(sk);
515                 icsk->icsk_ack.ato = TCP_ATO_MIN;
516         } else {
517                 int m = now - icsk->icsk_ack.lrcvtime;
518
519                 if (m <= TCP_ATO_MIN/2) {
520                         /* The fastest case is the first. */
521                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
522                 } else if (m < icsk->icsk_ack.ato) {
523                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
524                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
525                                 icsk->icsk_ack.ato = icsk->icsk_rto;
526                 } else if (m > icsk->icsk_rto) {
527                         /* Too long gap. Apparently sender falled to
528                          * restart window, so that we send ACKs quickly.
529                          */
530                         tcp_incr_quickack(sk);
531                         sk_stream_mem_reclaim(sk);
532                 }
533         }
534         icsk->icsk_ack.lrcvtime = now;
535
536         TCP_ECN_check_ce(tp, skb);
537
538         if (skb->len >= 128)
539                 tcp_grow_window(sk, tp, skb);
540 }
541
542 /* Called to compute a smoothed rtt estimate. The data fed to this
543  * routine either comes from timestamps, or from segments that were
544  * known _not_ to have been retransmitted [see Karn/Partridge
545  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
546  * piece by Van Jacobson.
547  * NOTE: the next three routines used to be one big routine.
548  * To save cycles in the RFC 1323 implementation it was better to break
549  * it up into three procedures. -- erics
550  */
551 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt, u32 *usrtt)
552 {
553         struct tcp_sock *tp = tcp_sk(sk);
554         const struct inet_connection_sock *icsk = inet_csk(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         if (icsk->icsk_ca_ops->rtt_sample)
615                 icsk->icsk_ca_ops->rtt_sample(sk, *usrtt);
616 }
617
618 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
619  * routine referred to above.
620  */
621 static inline void tcp_set_rto(struct sock *sk)
622 {
623         const struct tcp_sock *tp = tcp_sk(sk);
624         /* Old crap is replaced with new one. 8)
625          *
626          * More seriously:
627          * 1. If rtt variance happened to be less 50msec, it is hallucination.
628          *    It cannot be less due to utterly erratic ACK generation made
629          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
630          *    to do with delayed acks, because at cwnd>2 true delack timeout
631          *    is invisible. Actually, Linux-2.4 also generates erratic
632          *    ACKs in some curcumstances.
633          */
634         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
635
636         /* 2. Fixups made earlier cannot be right.
637          *    If we do not estimate RTO correctly without them,
638          *    all the algo is pure shit and should be replaced
639          *    with correct one. It is exaclty, which we pretend to do.
640          */
641 }
642
643 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
644  * guarantees that rto is higher.
645  */
646 static inline void tcp_bound_rto(struct sock *sk)
647 {
648         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
649                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
650 }
651
652 /* Save metrics learned by this TCP session.
653    This function is called only, when TCP finishes successfully
654    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
655  */
656 void tcp_update_metrics(struct sock *sk)
657 {
658         struct tcp_sock *tp = tcp_sk(sk);
659         struct dst_entry *dst = __sk_dst_get(sk);
660
661         if (sysctl_tcp_nometrics_save)
662                 return;
663
664         dst_confirm(dst);
665
666         if (dst && (dst->flags&DST_HOST)) {
667                 const struct inet_connection_sock *icsk = inet_csk(sk);
668                 int m;
669
670                 if (icsk->icsk_backoff || !tp->srtt) {
671                         /* This session failed to estimate rtt. Why?
672                          * Probably, no packets returned in time.
673                          * Reset our results.
674                          */
675                         if (!(dst_metric_locked(dst, RTAX_RTT)))
676                                 dst->metrics[RTAX_RTT-1] = 0;
677                         return;
678                 }
679
680                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
681
682                 /* If newly calculated rtt larger than stored one,
683                  * store new one. Otherwise, use EWMA. Remember,
684                  * rtt overestimation is always better than underestimation.
685                  */
686                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
687                         if (m <= 0)
688                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
689                         else
690                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
691                 }
692
693                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
694                         if (m < 0)
695                                 m = -m;
696
697                         /* Scale deviation to rttvar fixed point */
698                         m >>= 1;
699                         if (m < tp->mdev)
700                                 m = tp->mdev;
701
702                         if (m >= dst_metric(dst, RTAX_RTTVAR))
703                                 dst->metrics[RTAX_RTTVAR-1] = m;
704                         else
705                                 dst->metrics[RTAX_RTTVAR-1] -=
706                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
707                 }
708
709                 if (tp->snd_ssthresh >= 0xFFFF) {
710                         /* Slow start still did not finish. */
711                         if (dst_metric(dst, RTAX_SSTHRESH) &&
712                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
713                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
714                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
715                         if (!dst_metric_locked(dst, RTAX_CWND) &&
716                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
717                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
718                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
719                            icsk->icsk_ca_state == TCP_CA_Open) {
720                         /* Cong. avoidance phase, cwnd is reliable. */
721                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
722                                 dst->metrics[RTAX_SSTHRESH-1] =
723                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
724                         if (!dst_metric_locked(dst, RTAX_CWND))
725                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
726                 } else {
727                         /* Else slow start did not finish, cwnd is non-sense,
728                            ssthresh may be also invalid.
729                          */
730                         if (!dst_metric_locked(dst, RTAX_CWND))
731                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
732                         if (dst->metrics[RTAX_SSTHRESH-1] &&
733                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
734                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
735                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
736                 }
737
738                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
739                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
740                             tp->reordering != sysctl_tcp_reordering)
741                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
742                 }
743         }
744 }
745
746 /* Numbers are taken from RFC2414.  */
747 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
748 {
749         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
750
751         if (!cwnd) {
752                 if (tp->mss_cache > 1460)
753                         cwnd = 2;
754                 else
755                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
756         }
757         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
758 }
759
760 /* Initialize metrics on socket. */
761
762 static void tcp_init_metrics(struct sock *sk)
763 {
764         struct tcp_sock *tp = tcp_sk(sk);
765         struct dst_entry *dst = __sk_dst_get(sk);
766
767         if (dst == NULL)
768                 goto reset;
769
770         dst_confirm(dst);
771
772         if (dst_metric_locked(dst, RTAX_CWND))
773                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
774         if (dst_metric(dst, RTAX_SSTHRESH)) {
775                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
776                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
777                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
778         }
779         if (dst_metric(dst, RTAX_REORDERING) &&
780             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
781                 tp->rx_opt.sack_ok &= ~2;
782                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
783         }
784
785         if (dst_metric(dst, RTAX_RTT) == 0)
786                 goto reset;
787
788         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
789                 goto reset;
790
791         /* Initial rtt is determined from SYN,SYN-ACK.
792          * The segment is small and rtt may appear much
793          * less than real one. Use per-dst memory
794          * to make it more realistic.
795          *
796          * A bit of theory. RTT is time passed after "normal" sized packet
797          * is sent until it is ACKed. In normal curcumstances sending small
798          * packets force peer to delay ACKs and calculation is correct too.
799          * The algorithm is adaptive and, provided we follow specs, it
800          * NEVER underestimate RTT. BUT! If peer tries to make some clever
801          * tricks sort of "quick acks" for time long enough to decrease RTT
802          * to low value, and then abruptly stops to do it and starts to delay
803          * ACKs, wait for troubles.
804          */
805         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
806                 tp->srtt = dst_metric(dst, RTAX_RTT);
807                 tp->rtt_seq = tp->snd_nxt;
808         }
809         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
810                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
811                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
812         }
813         tcp_set_rto(sk);
814         tcp_bound_rto(sk);
815         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
816                 goto reset;
817         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
818         tp->snd_cwnd_stamp = tcp_time_stamp;
819         return;
820
821 reset:
822         /* Play conservative. If timestamps are not
823          * supported, TCP will fail to recalculate correct
824          * rtt, if initial rto is too small. FORGET ALL AND RESET!
825          */
826         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
827                 tp->srtt = 0;
828                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
829                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
830         }
831 }
832
833 static void tcp_update_reordering(struct sock *sk, const int metric,
834                                   const int ts)
835 {
836         struct tcp_sock *tp = tcp_sk(sk);
837         if (metric > tp->reordering) {
838                 tp->reordering = min(TCP_MAX_REORDERING, metric);
839
840                 /* This exciting event is worth to be remembered. 8) */
841                 if (ts)
842                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
843                 else if (IsReno(tp))
844                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
845                 else if (IsFack(tp))
846                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
847                 else
848                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
849 #if FASTRETRANS_DEBUG > 1
850                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
851                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
852                        tp->reordering,
853                        tp->fackets_out,
854                        tp->sacked_out,
855                        tp->undo_marker ? tp->undo_retrans : 0);
856 #endif
857                 /* Disable FACK yet. */
858                 tp->rx_opt.sack_ok &= ~2;
859         }
860 }
861
862 /* This procedure tags the retransmission queue when SACKs arrive.
863  *
864  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
865  * Packets in queue with these bits set are counted in variables
866  * sacked_out, retrans_out and lost_out, correspondingly.
867  *
868  * Valid combinations are:
869  * Tag  InFlight        Description
870  * 0    1               - orig segment is in flight.
871  * S    0               - nothing flies, orig reached receiver.
872  * L    0               - nothing flies, orig lost by net.
873  * R    2               - both orig and retransmit are in flight.
874  * L|R  1               - orig is lost, retransmit is in flight.
875  * S|R  1               - orig reached receiver, retrans is still in flight.
876  * (L|S|R is logically valid, it could occur when L|R is sacked,
877  *  but it is equivalent to plain S and code short-curcuits it to S.
878  *  L|S is logically invalid, it would mean -1 packet in flight 8))
879  *
880  * These 6 states form finite state machine, controlled by the following events:
881  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
882  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
883  * 3. Loss detection event of one of three flavors:
884  *      A. Scoreboard estimator decided the packet is lost.
885  *         A'. Reno "three dupacks" marks head of queue lost.
886  *         A''. Its FACK modfication, head until snd.fack is lost.
887  *      B. SACK arrives sacking data transmitted after never retransmitted
888  *         hole was sent out.
889  *      C. SACK arrives sacking SND.NXT at the moment, when the
890  *         segment was retransmitted.
891  * 4. D-SACK added new rule: D-SACK changes any tag to S.
892  *
893  * It is pleasant to note, that state diagram turns out to be commutative,
894  * so that we are allowed not to be bothered by order of our actions,
895  * when multiple events arrive simultaneously. (see the function below).
896  *
897  * Reordering detection.
898  * --------------------
899  * Reordering metric is maximal distance, which a packet can be displaced
900  * in packet stream. With SACKs we can estimate it:
901  *
902  * 1. SACK fills old hole and the corresponding segment was not
903  *    ever retransmitted -> reordering. Alas, we cannot use it
904  *    when segment was retransmitted.
905  * 2. The last flaw is solved with D-SACK. D-SACK arrives
906  *    for retransmitted and already SACKed segment -> reordering..
907  * Both of these heuristics are not used in Loss state, when we cannot
908  * account for retransmits accurately.
909  */
910 static int
911 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
912 {
913         const struct inet_connection_sock *icsk = inet_csk(sk);
914         struct tcp_sock *tp = tcp_sk(sk);
915         unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
916         struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
917         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
918         int reord = tp->packets_out;
919         int prior_fackets;
920         u32 lost_retrans = 0;
921         int flag = 0;
922         int i;
923
924         if (!tp->sacked_out)
925                 tp->fackets_out = 0;
926         prior_fackets = tp->fackets_out;
927
928         for (i=0; i<num_sacks; i++, sp++) {
929                 struct sk_buff *skb;
930                 __u32 start_seq = ntohl(sp->start_seq);
931                 __u32 end_seq = ntohl(sp->end_seq);
932                 int fack_count = 0;
933                 int dup_sack = 0;
934
935                 /* Check for D-SACK. */
936                 if (i == 0) {
937                         u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
938
939                         if (before(start_seq, ack)) {
940                                 dup_sack = 1;
941                                 tp->rx_opt.sack_ok |= 4;
942                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
943                         } else if (num_sacks > 1 &&
944                                    !after(end_seq, ntohl(sp[1].end_seq)) &&
945                                    !before(start_seq, ntohl(sp[1].start_seq))) {
946                                 dup_sack = 1;
947                                 tp->rx_opt.sack_ok |= 4;
948                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
949                         }
950
951                         /* D-SACK for already forgotten data...
952                          * Do dumb counting. */
953                         if (dup_sack &&
954                             !after(end_seq, prior_snd_una) &&
955                             after(end_seq, tp->undo_marker))
956                                 tp->undo_retrans--;
957
958                         /* Eliminate too old ACKs, but take into
959                          * account more or less fresh ones, they can
960                          * contain valid SACK info.
961                          */
962                         if (before(ack, prior_snd_una - tp->max_window))
963                                 return 0;
964                 }
965
966                 /* Event "B" in the comment above. */
967                 if (after(end_seq, tp->high_seq))
968                         flag |= FLAG_DATA_LOST;
969
970                 sk_stream_for_retrans_queue(skb, sk) {
971                         int in_sack, pcount;
972                         u8 sacked;
973
974                         /* The retransmission queue is always in order, so
975                          * we can short-circuit the walk early.
976                          */
977                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
978                                 break;
979
980                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
981                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
982
983                         pcount = tcp_skb_pcount(skb);
984
985                         if (pcount > 1 && !in_sack &&
986                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
987                                 unsigned int pkt_len;
988
989                                 in_sack = !after(start_seq,
990                                                  TCP_SKB_CB(skb)->seq);
991
992                                 if (!in_sack)
993                                         pkt_len = (start_seq -
994                                                    TCP_SKB_CB(skb)->seq);
995                                 else
996                                         pkt_len = (end_seq -
997                                                    TCP_SKB_CB(skb)->seq);
998                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->tso_size))
999                                         break;
1000                                 pcount = tcp_skb_pcount(skb);
1001                         }
1002
1003                         fack_count += pcount;
1004
1005                         sacked = TCP_SKB_CB(skb)->sacked;
1006
1007                         /* Account D-SACK for retransmitted packet. */
1008                         if ((dup_sack && in_sack) &&
1009                             (sacked & TCPCB_RETRANS) &&
1010                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1011                                 tp->undo_retrans--;
1012
1013                         /* The frame is ACKed. */
1014                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1015                                 if (sacked&TCPCB_RETRANS) {
1016                                         if ((dup_sack && in_sack) &&
1017                                             (sacked&TCPCB_SACKED_ACKED))
1018                                                 reord = min(fack_count, reord);
1019                                 } else {
1020                                         /* If it was in a hole, we detected reordering. */
1021                                         if (fack_count < prior_fackets &&
1022                                             !(sacked&TCPCB_SACKED_ACKED))
1023                                                 reord = min(fack_count, reord);
1024                                 }
1025
1026                                 /* Nothing to do; acked frame is about to be dropped. */
1027                                 continue;
1028                         }
1029
1030                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1031                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1032                             (!lost_retrans || after(end_seq, lost_retrans)))
1033                                 lost_retrans = end_seq;
1034
1035                         if (!in_sack)
1036                                 continue;
1037
1038                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1039                                 if (sacked & TCPCB_SACKED_RETRANS) {
1040                                         /* If the segment is not tagged as lost,
1041                                          * we do not clear RETRANS, believing
1042                                          * that retransmission is still in flight.
1043                                          */
1044                                         if (sacked & TCPCB_LOST) {
1045                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1046                                                 tp->lost_out -= tcp_skb_pcount(skb);
1047                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1048                                         }
1049                                 } else {
1050                                         /* New sack for not retransmitted frame,
1051                                          * which was in hole. It is reordering.
1052                                          */
1053                                         if (!(sacked & TCPCB_RETRANS) &&
1054                                             fack_count < prior_fackets)
1055                                                 reord = min(fack_count, reord);
1056
1057                                         if (sacked & TCPCB_LOST) {
1058                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1059                                                 tp->lost_out -= tcp_skb_pcount(skb);
1060                                         }
1061                                 }
1062
1063                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1064                                 flag |= FLAG_DATA_SACKED;
1065                                 tp->sacked_out += tcp_skb_pcount(skb);
1066
1067                                 if (fack_count > tp->fackets_out)
1068                                         tp->fackets_out = fack_count;
1069                         } else {
1070                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1071                                         reord = min(fack_count, reord);
1072                         }
1073
1074                         /* D-SACK. We can detect redundant retransmission
1075                          * in S|R and plain R frames and clear it.
1076                          * undo_retrans is decreased above, L|R frames
1077                          * are accounted above as well.
1078                          */
1079                         if (dup_sack &&
1080                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1081                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1082                                 tp->retrans_out -= tcp_skb_pcount(skb);
1083                         }
1084                 }
1085         }
1086
1087         /* Check for lost retransmit. This superb idea is
1088          * borrowed from "ratehalving". Event "C".
1089          * Later note: FACK people cheated me again 8),
1090          * we have to account for reordering! Ugly,
1091          * but should help.
1092          */
1093         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1094                 struct sk_buff *skb;
1095
1096                 sk_stream_for_retrans_queue(skb, sk) {
1097                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1098                                 break;
1099                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1100                                 continue;
1101                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1102                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1103                             (IsFack(tp) ||
1104                              !before(lost_retrans,
1105                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1106                                      tp->mss_cache))) {
1107                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1108                                 tp->retrans_out -= tcp_skb_pcount(skb);
1109
1110                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1111                                         tp->lost_out += tcp_skb_pcount(skb);
1112                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1113                                         flag |= FLAG_DATA_SACKED;
1114                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1115                                 }
1116                         }
1117                 }
1118         }
1119
1120         tp->left_out = tp->sacked_out + tp->lost_out;
1121
1122         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
1123                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1124
1125 #if FASTRETRANS_DEBUG > 0
1126         BUG_TRAP((int)tp->sacked_out >= 0);
1127         BUG_TRAP((int)tp->lost_out >= 0);
1128         BUG_TRAP((int)tp->retrans_out >= 0);
1129         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1130 #endif
1131         return flag;
1132 }
1133
1134 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1135  * segments to see from the next ACKs whether any data was really missing.
1136  * If the RTO was spurious, new ACKs should arrive.
1137  */
1138 void tcp_enter_frto(struct sock *sk)
1139 {
1140         const struct inet_connection_sock *icsk = inet_csk(sk);
1141         struct tcp_sock *tp = tcp_sk(sk);
1142         struct sk_buff *skb;
1143
1144         tp->frto_counter = 1;
1145
1146         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1147             tp->snd_una == tp->high_seq ||
1148             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1149                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1150                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1151                 tcp_ca_event(sk, CA_EVENT_FRTO);
1152         }
1153
1154         /* Have to clear retransmission markers here to keep the bookkeeping
1155          * in shape, even though we are not yet in Loss state.
1156          * If something was really lost, it is eventually caught up
1157          * in tcp_enter_frto_loss.
1158          */
1159         tp->retrans_out = 0;
1160         tp->undo_marker = tp->snd_una;
1161         tp->undo_retrans = 0;
1162
1163         sk_stream_for_retrans_queue(skb, sk) {
1164                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1165         }
1166         tcp_sync_left_out(tp);
1167
1168         tcp_set_ca_state(sk, TCP_CA_Open);
1169         tp->frto_highmark = tp->snd_nxt;
1170 }
1171
1172 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1173  * which indicates that we should follow the traditional RTO recovery,
1174  * i.e. mark everything lost and do go-back-N retransmission.
1175  */
1176 static void tcp_enter_frto_loss(struct sock *sk)
1177 {
1178         struct tcp_sock *tp = tcp_sk(sk);
1179         struct sk_buff *skb;
1180         int cnt = 0;
1181
1182         tp->sacked_out = 0;
1183         tp->lost_out = 0;
1184         tp->fackets_out = 0;
1185
1186         sk_stream_for_retrans_queue(skb, sk) {
1187                 cnt += tcp_skb_pcount(skb);
1188                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1189                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1190
1191                         /* Do not mark those segments lost that were
1192                          * forward transmitted after RTO
1193                          */
1194                         if (!after(TCP_SKB_CB(skb)->end_seq,
1195                                    tp->frto_highmark)) {
1196                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1197                                 tp->lost_out += tcp_skb_pcount(skb);
1198                         }
1199                 } else {
1200                         tp->sacked_out += tcp_skb_pcount(skb);
1201                         tp->fackets_out = cnt;
1202                 }
1203         }
1204         tcp_sync_left_out(tp);
1205
1206         tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1207         tp->snd_cwnd_cnt = 0;
1208         tp->snd_cwnd_stamp = tcp_time_stamp;
1209         tp->undo_marker = 0;
1210         tp->frto_counter = 0;
1211
1212         tp->reordering = min_t(unsigned int, tp->reordering,
1213                                              sysctl_tcp_reordering);
1214         tcp_set_ca_state(sk, TCP_CA_Loss);
1215         tp->high_seq = tp->frto_highmark;
1216         TCP_ECN_queue_cwr(tp);
1217 }
1218
1219 void tcp_clear_retrans(struct tcp_sock *tp)
1220 {
1221         tp->left_out = 0;
1222         tp->retrans_out = 0;
1223
1224         tp->fackets_out = 0;
1225         tp->sacked_out = 0;
1226         tp->lost_out = 0;
1227
1228         tp->undo_marker = 0;
1229         tp->undo_retrans = 0;
1230 }
1231
1232 /* Enter Loss state. If "how" is not zero, forget all SACK information
1233  * and reset tags completely, otherwise preserve SACKs. If receiver
1234  * dropped its ofo queue, we will know this due to reneging detection.
1235  */
1236 void tcp_enter_loss(struct sock *sk, int how)
1237 {
1238         const struct inet_connection_sock *icsk = inet_csk(sk);
1239         struct tcp_sock *tp = tcp_sk(sk);
1240         struct sk_buff *skb;
1241         int cnt = 0;
1242
1243         /* Reduce ssthresh if it has not yet been made inside this window. */
1244         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1245             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1246                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1247                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1248                 tcp_ca_event(sk, CA_EVENT_LOSS);
1249         }
1250         tp->snd_cwnd       = 1;
1251         tp->snd_cwnd_cnt   = 0;
1252         tp->snd_cwnd_stamp = tcp_time_stamp;
1253
1254         tcp_clear_retrans(tp);
1255
1256         /* Push undo marker, if it was plain RTO and nothing
1257          * was retransmitted. */
1258         if (!how)
1259                 tp->undo_marker = tp->snd_una;
1260
1261         sk_stream_for_retrans_queue(skb, sk) {
1262                 cnt += tcp_skb_pcount(skb);
1263                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1264                         tp->undo_marker = 0;
1265                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1266                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1267                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1268                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1269                         tp->lost_out += tcp_skb_pcount(skb);
1270                 } else {
1271                         tp->sacked_out += tcp_skb_pcount(skb);
1272                         tp->fackets_out = cnt;
1273                 }
1274         }
1275         tcp_sync_left_out(tp);
1276
1277         tp->reordering = min_t(unsigned int, tp->reordering,
1278                                              sysctl_tcp_reordering);
1279         tcp_set_ca_state(sk, TCP_CA_Loss);
1280         tp->high_seq = tp->snd_nxt;
1281         TCP_ECN_queue_cwr(tp);
1282 }
1283
1284 static int tcp_check_sack_reneging(struct sock *sk)
1285 {
1286         struct sk_buff *skb;
1287
1288         /* If ACK arrived pointing to a remembered SACK,
1289          * it means that our remembered SACKs do not reflect
1290          * real state of receiver i.e.
1291          * receiver _host_ is heavily congested (or buggy).
1292          * Do processing similar to RTO timeout.
1293          */
1294         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1295             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1296                 struct inet_connection_sock *icsk = inet_csk(sk);
1297                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1298
1299                 tcp_enter_loss(sk, 1);
1300                 icsk->icsk_retransmits++;
1301                 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1302                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1303                                           icsk->icsk_rto, TCP_RTO_MAX);
1304                 return 1;
1305         }
1306         return 0;
1307 }
1308
1309 static inline int tcp_fackets_out(struct tcp_sock *tp)
1310 {
1311         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1312 }
1313
1314 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1315 {
1316         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1317 }
1318
1319 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1320 {
1321         return tp->packets_out &&
1322                tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
1323 }
1324
1325 /* Linux NewReno/SACK/FACK/ECN state machine.
1326  * --------------------------------------
1327  *
1328  * "Open"       Normal state, no dubious events, fast path.
1329  * "Disorder"   In all the respects it is "Open",
1330  *              but requires a bit more attention. It is entered when
1331  *              we see some SACKs or dupacks. It is split of "Open"
1332  *              mainly to move some processing from fast path to slow one.
1333  * "CWR"        CWND was reduced due to some Congestion Notification event.
1334  *              It can be ECN, ICMP source quench, local device congestion.
1335  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1336  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1337  *
1338  * tcp_fastretrans_alert() is entered:
1339  * - each incoming ACK, if state is not "Open"
1340  * - when arrived ACK is unusual, namely:
1341  *      * SACK
1342  *      * Duplicate ACK.
1343  *      * ECN ECE.
1344  *
1345  * Counting packets in flight is pretty simple.
1346  *
1347  *      in_flight = packets_out - left_out + retrans_out
1348  *
1349  *      packets_out is SND.NXT-SND.UNA counted in packets.
1350  *
1351  *      retrans_out is number of retransmitted segments.
1352  *
1353  *      left_out is number of segments left network, but not ACKed yet.
1354  *
1355  *              left_out = sacked_out + lost_out
1356  *
1357  *     sacked_out: Packets, which arrived to receiver out of order
1358  *                 and hence not ACKed. With SACKs this number is simply
1359  *                 amount of SACKed data. Even without SACKs
1360  *                 it is easy to give pretty reliable estimate of this number,
1361  *                 counting duplicate ACKs.
1362  *
1363  *       lost_out: Packets lost by network. TCP has no explicit
1364  *                 "loss notification" feedback from network (for now).
1365  *                 It means that this number can be only _guessed_.
1366  *                 Actually, it is the heuristics to predict lossage that
1367  *                 distinguishes different algorithms.
1368  *
1369  *      F.e. after RTO, when all the queue is considered as lost,
1370  *      lost_out = packets_out and in_flight = retrans_out.
1371  *
1372  *              Essentially, we have now two algorithms counting
1373  *              lost packets.
1374  *
1375  *              FACK: It is the simplest heuristics. As soon as we decided
1376  *              that something is lost, we decide that _all_ not SACKed
1377  *              packets until the most forward SACK are lost. I.e.
1378  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1379  *              It is absolutely correct estimate, if network does not reorder
1380  *              packets. And it loses any connection to reality when reordering
1381  *              takes place. We use FACK by default until reordering
1382  *              is suspected on the path to this destination.
1383  *
1384  *              NewReno: when Recovery is entered, we assume that one segment
1385  *              is lost (classic Reno). While we are in Recovery and
1386  *              a partial ACK arrives, we assume that one more packet
1387  *              is lost (NewReno). This heuristics are the same in NewReno
1388  *              and SACK.
1389  *
1390  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1391  *  deflation etc. CWND is real congestion window, never inflated, changes
1392  *  only according to classic VJ rules.
1393  *
1394  * Really tricky (and requiring careful tuning) part of algorithm
1395  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1396  * The first determines the moment _when_ we should reduce CWND and,
1397  * hence, slow down forward transmission. In fact, it determines the moment
1398  * when we decide that hole is caused by loss, rather than by a reorder.
1399  *
1400  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1401  * holes, caused by lost packets.
1402  *
1403  * And the most logically complicated part of algorithm is undo
1404  * heuristics. We detect false retransmits due to both too early
1405  * fast retransmit (reordering) and underestimated RTO, analyzing
1406  * timestamps and D-SACKs. When we detect that some segments were
1407  * retransmitted by mistake and CWND reduction was wrong, we undo
1408  * window reduction and abort recovery phase. This logic is hidden
1409  * inside several functions named tcp_try_undo_<something>.
1410  */
1411
1412 /* This function decides, when we should leave Disordered state
1413  * and enter Recovery phase, reducing congestion window.
1414  *
1415  * Main question: may we further continue forward transmission
1416  * with the same cwnd?
1417  */
1418 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1419 {
1420         __u32 packets_out;
1421
1422         /* Trick#1: The loss is proven. */
1423         if (tp->lost_out)
1424                 return 1;
1425
1426         /* Not-A-Trick#2 : Classic rule... */
1427         if (tcp_fackets_out(tp) > tp->reordering)
1428                 return 1;
1429
1430         /* Trick#3 : when we use RFC2988 timer restart, fast
1431          * retransmit can be triggered by timeout of queue head.
1432          */
1433         if (tcp_head_timedout(sk, tp))
1434                 return 1;
1435
1436         /* Trick#4: It is still not OK... But will it be useful to delay
1437          * recovery more?
1438          */
1439         packets_out = tp->packets_out;
1440         if (packets_out <= tp->reordering &&
1441             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1442             !tcp_may_send_now(sk, tp)) {
1443                 /* We have nothing to send. This connection is limited
1444                  * either by receiver window or by application.
1445                  */
1446                 return 1;
1447         }
1448
1449         return 0;
1450 }
1451
1452 /* If we receive more dupacks than we expected counting segments
1453  * in assumption of absent reordering, interpret this as reordering.
1454  * The only another reason could be bug in receiver TCP.
1455  */
1456 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1457 {
1458         struct tcp_sock *tp = tcp_sk(sk);
1459         u32 holes;
1460
1461         holes = max(tp->lost_out, 1U);
1462         holes = min(holes, tp->packets_out);
1463
1464         if ((tp->sacked_out + holes) > tp->packets_out) {
1465                 tp->sacked_out = tp->packets_out - holes;
1466                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1467         }
1468 }
1469
1470 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1471
1472 static void tcp_add_reno_sack(struct sock *sk)
1473 {
1474         struct tcp_sock *tp = tcp_sk(sk);
1475         tp->sacked_out++;
1476         tcp_check_reno_reordering(sk, 0);
1477         tcp_sync_left_out(tp);
1478 }
1479
1480 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1481
1482 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1483 {
1484         if (acked > 0) {
1485                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1486                 if (acked-1 >= tp->sacked_out)
1487                         tp->sacked_out = 0;
1488                 else
1489                         tp->sacked_out -= acked-1;
1490         }
1491         tcp_check_reno_reordering(sk, acked);
1492         tcp_sync_left_out(tp);
1493 }
1494
1495 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1496 {
1497         tp->sacked_out = 0;
1498         tp->left_out = tp->lost_out;
1499 }
1500
1501 /* Mark head of queue up as lost. */
1502 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1503                                int packets, u32 high_seq)
1504 {
1505         struct sk_buff *skb;
1506         int cnt = packets;
1507
1508         BUG_TRAP(cnt <= tp->packets_out);
1509
1510         sk_stream_for_retrans_queue(skb, sk) {
1511                 cnt -= tcp_skb_pcount(skb);
1512                 if (cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1513                         break;
1514                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1515                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1516                         tp->lost_out += tcp_skb_pcount(skb);
1517                 }
1518         }
1519         tcp_sync_left_out(tp);
1520 }
1521
1522 /* Account newly detected lost packet(s) */
1523
1524 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1525 {
1526         if (IsFack(tp)) {
1527                 int lost = tp->fackets_out - tp->reordering;
1528                 if (lost <= 0)
1529                         lost = 1;
1530                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1531         } else {
1532                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1533         }
1534
1535         /* New heuristics: it is possible only after we switched
1536          * to restart timer each time when something is ACKed.
1537          * Hence, we can detect timed out packets during fast
1538          * retransmit without falling to slow start.
1539          */
1540         if (tcp_head_timedout(sk, tp)) {
1541                 struct sk_buff *skb;
1542
1543                 sk_stream_for_retrans_queue(skb, sk) {
1544                         if (tcp_skb_timedout(sk, skb) &&
1545                             !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1546                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1547                                 tp->lost_out += tcp_skb_pcount(skb);
1548                         }
1549                 }
1550                 tcp_sync_left_out(tp);
1551         }
1552 }
1553
1554 /* CWND moderation, preventing bursts due to too big ACKs
1555  * in dubious situations.
1556  */
1557 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1558 {
1559         tp->snd_cwnd = min(tp->snd_cwnd,
1560                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1561         tp->snd_cwnd_stamp = tcp_time_stamp;
1562 }
1563
1564 /* Decrease cwnd each second ack. */
1565 static void tcp_cwnd_down(struct sock *sk)
1566 {
1567         const struct inet_connection_sock *icsk = inet_csk(sk);
1568         struct tcp_sock *tp = tcp_sk(sk);
1569         int decr = tp->snd_cwnd_cnt + 1;
1570
1571         tp->snd_cwnd_cnt = decr&1;
1572         decr >>= 1;
1573
1574         if (decr && tp->snd_cwnd > icsk->icsk_ca_ops->min_cwnd(sk))
1575                 tp->snd_cwnd -= decr;
1576
1577         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1578         tp->snd_cwnd_stamp = tcp_time_stamp;
1579 }
1580
1581 /* Nothing was retransmitted or returned timestamp is less
1582  * than timestamp of the first retransmission.
1583  */
1584 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1585 {
1586         return !tp->retrans_stamp ||
1587                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1588                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1589 }
1590
1591 /* Undo procedures. */
1592
1593 #if FASTRETRANS_DEBUG > 1
1594 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1595 {
1596         struct inet_sock *inet = inet_sk(sk);
1597         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1598                msg,
1599                NIPQUAD(inet->daddr), ntohs(inet->dport),
1600                tp->snd_cwnd, tp->left_out,
1601                tp->snd_ssthresh, tp->prior_ssthresh,
1602                tp->packets_out);
1603 }
1604 #else
1605 #define DBGUNDO(x...) do { } while (0)
1606 #endif
1607
1608 static void tcp_undo_cwr(struct sock *sk, const int undo)
1609 {
1610         struct tcp_sock *tp = tcp_sk(sk);
1611
1612         if (tp->prior_ssthresh) {
1613                 const struct inet_connection_sock *icsk = inet_csk(sk);
1614
1615                 if (icsk->icsk_ca_ops->undo_cwnd)
1616                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1617                 else
1618                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1619
1620                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1621                         tp->snd_ssthresh = tp->prior_ssthresh;
1622                         TCP_ECN_withdraw_cwr(tp);
1623                 }
1624         } else {
1625                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1626         }
1627         tcp_moderate_cwnd(tp);
1628         tp->snd_cwnd_stamp = tcp_time_stamp;
1629 }
1630
1631 static inline int tcp_may_undo(struct tcp_sock *tp)
1632 {
1633         return tp->undo_marker &&
1634                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1635 }
1636
1637 /* People celebrate: "We love our President!" */
1638 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1639 {
1640         if (tcp_may_undo(tp)) {
1641                 /* Happy end! We did not retransmit anything
1642                  * or our original transmission succeeded.
1643                  */
1644                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1645                 tcp_undo_cwr(sk, 1);
1646                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1647                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1648                 else
1649                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1650                 tp->undo_marker = 0;
1651         }
1652         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1653                 /* Hold old state until something *above* high_seq
1654                  * is ACKed. For Reno it is MUST to prevent false
1655                  * fast retransmits (RFC2582). SACK TCP is safe. */
1656                 tcp_moderate_cwnd(tp);
1657                 return 1;
1658         }
1659         tcp_set_ca_state(sk, TCP_CA_Open);
1660         return 0;
1661 }
1662
1663 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1664 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1665 {
1666         if (tp->undo_marker && !tp->undo_retrans) {
1667                 DBGUNDO(sk, tp, "D-SACK");
1668                 tcp_undo_cwr(sk, 1);
1669                 tp->undo_marker = 0;
1670                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1671         }
1672 }
1673
1674 /* Undo during fast recovery after partial ACK. */
1675
1676 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1677                                 int acked)
1678 {
1679         /* Partial ACK arrived. Force Hoe's retransmit. */
1680         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1681
1682         if (tcp_may_undo(tp)) {
1683                 /* Plain luck! Hole if filled with delayed
1684                  * packet, rather than with a retransmit.
1685                  */
1686                 if (tp->retrans_out == 0)
1687                         tp->retrans_stamp = 0;
1688
1689                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1690
1691                 DBGUNDO(sk, tp, "Hoe");
1692                 tcp_undo_cwr(sk, 0);
1693                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1694
1695                 /* So... Do not make Hoe's retransmit yet.
1696                  * If the first packet was delayed, the rest
1697                  * ones are most probably delayed as well.
1698                  */
1699                 failed = 0;
1700         }
1701         return failed;
1702 }
1703
1704 /* Undo during loss recovery after partial ACK. */
1705 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1706 {
1707         if (tcp_may_undo(tp)) {
1708                 struct sk_buff *skb;
1709                 sk_stream_for_retrans_queue(skb, sk) {
1710                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1711                 }
1712                 DBGUNDO(sk, tp, "partial loss");
1713                 tp->lost_out = 0;
1714                 tp->left_out = tp->sacked_out;
1715                 tcp_undo_cwr(sk, 1);
1716                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1717                 inet_csk(sk)->icsk_retransmits = 0;
1718                 tp->undo_marker = 0;
1719                 if (!IsReno(tp))
1720                         tcp_set_ca_state(sk, TCP_CA_Open);
1721                 return 1;
1722         }
1723         return 0;
1724 }
1725
1726 static inline void tcp_complete_cwr(struct sock *sk)
1727 {
1728         struct tcp_sock *tp = tcp_sk(sk);
1729         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1730         tp->snd_cwnd_stamp = tcp_time_stamp;
1731         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
1732 }
1733
1734 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1735 {
1736         tp->left_out = tp->sacked_out;
1737
1738         if (tp->retrans_out == 0)
1739                 tp->retrans_stamp = 0;
1740
1741         if (flag&FLAG_ECE)
1742                 tcp_enter_cwr(sk);
1743
1744         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
1745                 int state = TCP_CA_Open;
1746
1747                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1748                         state = TCP_CA_Disorder;
1749
1750                 if (inet_csk(sk)->icsk_ca_state != state) {
1751                         tcp_set_ca_state(sk, state);
1752                         tp->high_seq = tp->snd_nxt;
1753                 }
1754                 tcp_moderate_cwnd(tp);
1755         } else {
1756                 tcp_cwnd_down(sk);
1757         }
1758 }
1759
1760 /* Process an event, which can update packets-in-flight not trivially.
1761  * Main goal of this function is to calculate new estimate for left_out,
1762  * taking into account both packets sitting in receiver's buffer and
1763  * packets lost by network.
1764  *
1765  * Besides that it does CWND reduction, when packet loss is detected
1766  * and changes state of machine.
1767  *
1768  * It does _not_ decide what to send, it is made in function
1769  * tcp_xmit_retransmit_queue().
1770  */
1771 static void
1772 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1773                       int prior_packets, int flag)
1774 {
1775         struct inet_connection_sock *icsk = inet_csk(sk);
1776         struct tcp_sock *tp = tcp_sk(sk);
1777         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1778
1779         /* Some technical things:
1780          * 1. Reno does not count dupacks (sacked_out) automatically. */
1781         if (!tp->packets_out)
1782                 tp->sacked_out = 0;
1783         /* 2. SACK counts snd_fack in packets inaccurately. */
1784         if (tp->sacked_out == 0)
1785                 tp->fackets_out = 0;
1786
1787         /* Now state machine starts.
1788          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1789         if (flag&FLAG_ECE)
1790                 tp->prior_ssthresh = 0;
1791
1792         /* B. In all the states check for reneging SACKs. */
1793         if (tp->sacked_out && tcp_check_sack_reneging(sk))
1794                 return;
1795
1796         /* C. Process data loss notification, provided it is valid. */
1797         if ((flag&FLAG_DATA_LOST) &&
1798             before(tp->snd_una, tp->high_seq) &&
1799             icsk->icsk_ca_state != TCP_CA_Open &&
1800             tp->fackets_out > tp->reordering) {
1801                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1802                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1803         }
1804
1805         /* D. Synchronize left_out to current state. */
1806         tcp_sync_left_out(tp);
1807
1808         /* E. Check state exit conditions. State can be terminated
1809          *    when high_seq is ACKed. */
1810         if (icsk->icsk_ca_state == TCP_CA_Open) {
1811                 if (!sysctl_tcp_frto)
1812                         BUG_TRAP(tp->retrans_out == 0);
1813                 tp->retrans_stamp = 0;
1814         } else if (!before(tp->snd_una, tp->high_seq)) {
1815                 switch (icsk->icsk_ca_state) {
1816                 case TCP_CA_Loss:
1817                         icsk->icsk_retransmits = 0;
1818                         if (tcp_try_undo_recovery(sk, tp))
1819                                 return;
1820                         break;
1821
1822                 case TCP_CA_CWR:
1823                         /* CWR is to be held something *above* high_seq
1824                          * is ACKed for CWR bit to reach receiver. */
1825                         if (tp->snd_una != tp->high_seq) {
1826                                 tcp_complete_cwr(sk);
1827                                 tcp_set_ca_state(sk, TCP_CA_Open);
1828                         }
1829                         break;
1830
1831                 case TCP_CA_Disorder:
1832                         tcp_try_undo_dsack(sk, tp);
1833                         if (!tp->undo_marker ||
1834                             /* For SACK case do not Open to allow to undo
1835                              * catching for all duplicate ACKs. */
1836                             IsReno(tp) || tp->snd_una != tp->high_seq) {
1837                                 tp->undo_marker = 0;
1838                                 tcp_set_ca_state(sk, TCP_CA_Open);
1839                         }
1840                         break;
1841
1842                 case TCP_CA_Recovery:
1843                         if (IsReno(tp))
1844                                 tcp_reset_reno_sack(tp);
1845                         if (tcp_try_undo_recovery(sk, tp))
1846                                 return;
1847                         tcp_complete_cwr(sk);
1848                         break;
1849                 }
1850         }
1851
1852         /* F. Process state. */
1853         switch (icsk->icsk_ca_state) {
1854         case TCP_CA_Recovery:
1855                 if (prior_snd_una == tp->snd_una) {
1856                         if (IsReno(tp) && is_dupack)
1857                                 tcp_add_reno_sack(sk);
1858                 } else {
1859                         int acked = prior_packets - tp->packets_out;
1860                         if (IsReno(tp))
1861                                 tcp_remove_reno_sacks(sk, tp, acked);
1862                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
1863                 }
1864                 break;
1865         case TCP_CA_Loss:
1866                 if (flag&FLAG_DATA_ACKED)
1867                         icsk->icsk_retransmits = 0;
1868                 if (!tcp_try_undo_loss(sk, tp)) {
1869                         tcp_moderate_cwnd(tp);
1870                         tcp_xmit_retransmit_queue(sk);
1871                         return;
1872                 }
1873                 if (icsk->icsk_ca_state != TCP_CA_Open)
1874                         return;
1875                 /* Loss is undone; fall through to processing in Open state. */
1876         default:
1877                 if (IsReno(tp)) {
1878                         if (tp->snd_una != prior_snd_una)
1879                                 tcp_reset_reno_sack(tp);
1880                         if (is_dupack)
1881                                 tcp_add_reno_sack(sk);
1882                 }
1883
1884                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
1885                         tcp_try_undo_dsack(sk, tp);
1886
1887                 if (!tcp_time_to_recover(sk, tp)) {
1888                         tcp_try_to_open(sk, tp, flag);
1889                         return;
1890                 }
1891
1892                 /* Otherwise enter Recovery state */
1893
1894                 if (IsReno(tp))
1895                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
1896                 else
1897                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
1898
1899                 tp->high_seq = tp->snd_nxt;
1900                 tp->prior_ssthresh = 0;
1901                 tp->undo_marker = tp->snd_una;
1902                 tp->undo_retrans = tp->retrans_out;
1903
1904                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
1905                         if (!(flag&FLAG_ECE))
1906                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1907                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1908                         TCP_ECN_queue_cwr(tp);
1909                 }
1910
1911                 tp->snd_cwnd_cnt = 0;
1912                 tcp_set_ca_state(sk, TCP_CA_Recovery);
1913         }
1914
1915         if (is_dupack || tcp_head_timedout(sk, tp))
1916                 tcp_update_scoreboard(sk, tp);
1917         tcp_cwnd_down(sk);
1918         tcp_xmit_retransmit_queue(sk);
1919 }
1920
1921 /* Read draft-ietf-tcplw-high-performance before mucking
1922  * with this code. (Superceeds RFC1323)
1923  */
1924 static void tcp_ack_saw_tstamp(struct sock *sk, u32 *usrtt, int flag)
1925 {
1926         /* RTTM Rule: A TSecr value received in a segment is used to
1927          * update the averaged RTT measurement only if the segment
1928          * acknowledges some new data, i.e., only if it advances the
1929          * left edge of the send window.
1930          *
1931          * See draft-ietf-tcplw-high-performance-00, section 3.3.
1932          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1933          *
1934          * Changed: reset backoff as soon as we see the first valid sample.
1935          * If we do not, we get strongly overstimated rto. With timestamps
1936          * samples are accepted even from very old segments: f.e., when rtt=1
1937          * increases to 8, we retransmit 5 times and after 8 seconds delayed
1938          * answer arrives rto becomes 120 seconds! If at least one of segments
1939          * in window is lost... Voila.                          --ANK (010210)
1940          */
1941         struct tcp_sock *tp = tcp_sk(sk);
1942         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
1943         tcp_rtt_estimator(sk, seq_rtt, usrtt);
1944         tcp_set_rto(sk);
1945         inet_csk(sk)->icsk_backoff = 0;
1946         tcp_bound_rto(sk);
1947 }
1948
1949 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, u32 *usrtt, int flag)
1950 {
1951         /* We don't have a timestamp. Can only use
1952          * packets that are not retransmitted to determine
1953          * rtt estimates. Also, we must not reset the
1954          * backoff for rto until we get a non-retransmitted
1955          * packet. This allows us to deal with a situation
1956          * where the network delay has increased suddenly.
1957          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1958          */
1959
1960         if (flag & FLAG_RETRANS_DATA_ACKED)
1961                 return;
1962
1963         tcp_rtt_estimator(sk, seq_rtt, usrtt);
1964         tcp_set_rto(sk);
1965         inet_csk(sk)->icsk_backoff = 0;
1966         tcp_bound_rto(sk);
1967 }
1968
1969 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
1970                                       const s32 seq_rtt, u32 *usrtt)
1971 {
1972         const struct tcp_sock *tp = tcp_sk(sk);
1973         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1974         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
1975                 tcp_ack_saw_tstamp(sk, usrtt, flag);
1976         else if (seq_rtt >= 0)
1977                 tcp_ack_no_tstamp(sk, seq_rtt, usrtt, flag);
1978 }
1979
1980 static inline void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
1981                                   u32 in_flight, int good)
1982 {
1983         const struct inet_connection_sock *icsk = inet_csk(sk);
1984         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
1985         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
1986 }
1987
1988 /* Restart timer after forward progress on connection.
1989  * RFC2988 recommends to restart timer to now+rto.
1990  */
1991
1992 static inline void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
1993 {
1994         if (!tp->packets_out) {
1995                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
1996         } else {
1997                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
1998         }
1999 }
2000
2001 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2002                          __u32 now, __s32 *seq_rtt)
2003 {
2004         struct tcp_sock *tp = tcp_sk(sk);
2005         struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2006         __u32 seq = tp->snd_una;
2007         __u32 packets_acked;
2008         int acked = 0;
2009
2010         /* If we get here, the whole TSO packet has not been
2011          * acked.
2012          */
2013         BUG_ON(!after(scb->end_seq, seq));
2014
2015         packets_acked = tcp_skb_pcount(skb);
2016         if (tcp_trim_head(sk, skb, seq - scb->seq))
2017                 return 0;
2018         packets_acked -= tcp_skb_pcount(skb);
2019
2020         if (packets_acked) {
2021                 __u8 sacked = scb->sacked;
2022
2023                 acked |= FLAG_DATA_ACKED;
2024                 if (sacked) {
2025                         if (sacked & TCPCB_RETRANS) {
2026                                 if (sacked & TCPCB_SACKED_RETRANS)
2027                                         tp->retrans_out -= packets_acked;
2028                                 acked |= FLAG_RETRANS_DATA_ACKED;
2029                                 *seq_rtt = -1;
2030                         } else if (*seq_rtt < 0)
2031                                 *seq_rtt = now - scb->when;
2032                         if (sacked & TCPCB_SACKED_ACKED)
2033                                 tp->sacked_out -= packets_acked;
2034                         if (sacked & TCPCB_LOST)
2035                                 tp->lost_out -= packets_acked;
2036                         if (sacked & TCPCB_URG) {
2037                                 if (tp->urg_mode &&
2038                                     !before(seq, tp->snd_up))
2039                                         tp->urg_mode = 0;
2040                         }
2041                 } else if (*seq_rtt < 0)
2042                         *seq_rtt = now - scb->when;
2043
2044                 if (tp->fackets_out) {
2045                         __u32 dval = min(tp->fackets_out, packets_acked);
2046                         tp->fackets_out -= dval;
2047                 }
2048                 tp->packets_out -= packets_acked;
2049
2050                 BUG_ON(tcp_skb_pcount(skb) == 0);
2051                 BUG_ON(!before(scb->seq, scb->end_seq));
2052         }
2053
2054         return acked;
2055 }
2056
2057
2058 /* Remove acknowledged frames from the retransmission queue. */
2059 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p, s32 *seq_usrtt)
2060 {
2061         struct tcp_sock *tp = tcp_sk(sk);
2062         struct sk_buff *skb;
2063         __u32 now = tcp_time_stamp;
2064         int acked = 0;
2065         __s32 seq_rtt = -1;
2066         struct timeval usnow;
2067         u32 pkts_acked = 0;
2068
2069         if (seq_usrtt)
2070                 do_gettimeofday(&usnow);
2071
2072         while ((skb = skb_peek(&sk->sk_write_queue)) &&
2073                skb != sk->sk_send_head) {
2074                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 
2075                 __u8 sacked = scb->sacked;
2076
2077                 /* If our packet is before the ack sequence we can
2078                  * discard it as it's confirmed to have arrived at
2079                  * the other end.
2080                  */
2081                 if (after(scb->end_seq, tp->snd_una)) {
2082                         if (tcp_skb_pcount(skb) > 1 &&
2083                             after(tp->snd_una, scb->seq))
2084                                 acked |= tcp_tso_acked(sk, skb,
2085                                                        now, &seq_rtt);
2086                         break;
2087                 }
2088
2089                 /* Initial outgoing SYN's get put onto the write_queue
2090                  * just like anything else we transmit.  It is not
2091                  * true data, and if we misinform our callers that
2092                  * this ACK acks real data, we will erroneously exit
2093                  * connection startup slow start one packet too
2094                  * quickly.  This is severely frowned upon behavior.
2095                  */
2096                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2097                         acked |= FLAG_DATA_ACKED;
2098                         ++pkts_acked;
2099                 } else {
2100                         acked |= FLAG_SYN_ACKED;
2101                         tp->retrans_stamp = 0;
2102                 }
2103
2104                 if (sacked) {
2105                         if (sacked & TCPCB_RETRANS) {
2106                                 if(sacked & TCPCB_SACKED_RETRANS)
2107                                         tp->retrans_out -= tcp_skb_pcount(skb);
2108                                 acked |= FLAG_RETRANS_DATA_ACKED;
2109                                 seq_rtt = -1;
2110                         } else if (seq_rtt < 0)
2111                                 seq_rtt = now - scb->when;
2112                         if (seq_usrtt) {
2113                                 struct timeval tv;
2114                         
2115                                 skb_get_timestamp(skb, &tv);
2116                                 *seq_usrtt = (usnow.tv_sec - tv.tv_sec) * 1000000
2117                                         + (usnow.tv_usec - tv.tv_usec);
2118                         }
2119
2120                         if (sacked & TCPCB_SACKED_ACKED)
2121                                 tp->sacked_out -= tcp_skb_pcount(skb);
2122                         if (sacked & TCPCB_LOST)
2123                                 tp->lost_out -= tcp_skb_pcount(skb);
2124                         if (sacked & TCPCB_URG) {
2125                                 if (tp->urg_mode &&
2126                                     !before(scb->end_seq, tp->snd_up))
2127                                         tp->urg_mode = 0;
2128                         }
2129                 } else if (seq_rtt < 0)
2130                         seq_rtt = now - scb->when;
2131                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2132                 tcp_packets_out_dec(tp, skb);
2133                 __skb_unlink(skb, &sk->sk_write_queue);
2134                 sk_stream_free_skb(sk, skb);
2135         }
2136
2137         if (acked&FLAG_ACKED) {
2138                 const struct inet_connection_sock *icsk = inet_csk(sk);
2139                 tcp_ack_update_rtt(sk, acked, seq_rtt, seq_usrtt);
2140                 tcp_ack_packets_out(sk, tp);
2141
2142                 if (icsk->icsk_ca_ops->pkts_acked)
2143                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2144         }
2145
2146 #if FASTRETRANS_DEBUG > 0
2147         BUG_TRAP((int)tp->sacked_out >= 0);
2148         BUG_TRAP((int)tp->lost_out >= 0);
2149         BUG_TRAP((int)tp->retrans_out >= 0);
2150         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2151                 const struct inet_connection_sock *icsk = inet_csk(sk);
2152                 if (tp->lost_out) {
2153                         printk(KERN_DEBUG "Leak l=%u %d\n",
2154                                tp->lost_out, icsk->icsk_ca_state);
2155                         tp->lost_out = 0;
2156                 }
2157                 if (tp->sacked_out) {
2158                         printk(KERN_DEBUG "Leak s=%u %d\n",
2159                                tp->sacked_out, icsk->icsk_ca_state);
2160                         tp->sacked_out = 0;
2161                 }
2162                 if (tp->retrans_out) {
2163                         printk(KERN_DEBUG "Leak r=%u %d\n",
2164                                tp->retrans_out, icsk->icsk_ca_state);
2165                         tp->retrans_out = 0;
2166                 }
2167         }
2168 #endif
2169         *seq_rtt_p = seq_rtt;
2170         return acked;
2171 }
2172
2173 static void tcp_ack_probe(struct sock *sk)
2174 {
2175         const struct tcp_sock *tp = tcp_sk(sk);
2176         struct inet_connection_sock *icsk = inet_csk(sk);
2177
2178         /* Was it a usable window open? */
2179
2180         if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2181                    tp->snd_una + tp->snd_wnd)) {
2182                 icsk->icsk_backoff = 0;
2183                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2184                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2185                  * This function is not for random using!
2186                  */
2187         } else {
2188                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2189                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2190                                           TCP_RTO_MAX);
2191         }
2192 }
2193
2194 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2195 {
2196         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2197                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2198 }
2199
2200 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2201 {
2202         const struct tcp_sock *tp = tcp_sk(sk);
2203         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2204                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2205 }
2206
2207 /* Check that window update is acceptable.
2208  * The function assumes that snd_una<=ack<=snd_next.
2209  */
2210 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2211                                         const u32 ack_seq, const u32 nwin)
2212 {
2213         return (after(ack, tp->snd_una) ||
2214                 after(ack_seq, tp->snd_wl1) ||
2215                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2216 }
2217
2218 /* Update our send window.
2219  *
2220  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2221  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2222  */
2223 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2224                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2225 {
2226         int flag = 0;
2227         u32 nwin = ntohs(skb->h.th->window);
2228
2229         if (likely(!skb->h.th->syn))
2230                 nwin <<= tp->rx_opt.snd_wscale;
2231
2232         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2233                 flag |= FLAG_WIN_UPDATE;
2234                 tcp_update_wl(tp, ack, ack_seq);
2235
2236                 if (tp->snd_wnd != nwin) {
2237                         tp->snd_wnd = nwin;
2238
2239                         /* Note, it is the only place, where
2240                          * fast path is recovered for sending TCP.
2241                          */
2242                         tp->pred_flags = 0;
2243                         tcp_fast_path_check(sk, tp);
2244
2245                         if (nwin > tp->max_window) {
2246                                 tp->max_window = nwin;
2247                                 tcp_sync_mss(sk, tp->pmtu_cookie);
2248                         }
2249                 }
2250         }
2251
2252         tp->snd_una = ack;
2253
2254         return flag;
2255 }
2256
2257 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2258 {
2259         struct tcp_sock *tp = tcp_sk(sk);
2260         
2261         tcp_sync_left_out(tp);
2262         
2263         if (tp->snd_una == prior_snd_una ||
2264             !before(tp->snd_una, tp->frto_highmark)) {
2265                 /* RTO was caused by loss, start retransmitting in
2266                  * go-back-N slow start
2267                  */
2268                 tcp_enter_frto_loss(sk);
2269                 return;
2270         }
2271
2272         if (tp->frto_counter == 1) {
2273                 /* First ACK after RTO advances the window: allow two new
2274                  * segments out.
2275                  */
2276                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2277         } else {
2278                 /* Also the second ACK after RTO advances the window.
2279                  * The RTO was likely spurious. Reduce cwnd and continue
2280                  * in congestion avoidance
2281                  */
2282                 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2283                 tcp_moderate_cwnd(tp);
2284         }
2285
2286         /* F-RTO affects on two new ACKs following RTO.
2287          * At latest on third ACK the TCP behavor is back to normal.
2288          */
2289         tp->frto_counter = (tp->frto_counter + 1) % 3;
2290 }
2291
2292 /* This routine deals with incoming acks, but not outgoing ones. */
2293 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2294 {
2295         struct inet_connection_sock *icsk = inet_csk(sk);
2296         struct tcp_sock *tp = tcp_sk(sk);
2297         u32 prior_snd_una = tp->snd_una;
2298         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2299         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2300         u32 prior_in_flight;
2301         s32 seq_rtt;
2302         s32 seq_usrtt = 0;
2303         int prior_packets;
2304
2305         /* If the ack is newer than sent or older than previous acks
2306          * then we can probably ignore it.
2307          */
2308         if (after(ack, tp->snd_nxt))
2309                 goto uninteresting_ack;
2310
2311         if (before(ack, prior_snd_una))
2312                 goto old_ack;
2313
2314         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2315                 /* Window is constant, pure forward advance.
2316                  * No more checks are required.
2317                  * Note, we use the fact that SND.UNA>=SND.WL2.
2318                  */
2319                 tcp_update_wl(tp, ack, ack_seq);
2320                 tp->snd_una = ack;
2321                 flag |= FLAG_WIN_UPDATE;
2322
2323                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2324
2325                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2326         } else {
2327                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2328                         flag |= FLAG_DATA;
2329                 else
2330                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2331
2332                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2333
2334                 if (TCP_SKB_CB(skb)->sacked)
2335                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2336
2337                 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2338                         flag |= FLAG_ECE;
2339
2340                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2341         }
2342
2343         /* We passed data and got it acked, remove any soft error
2344          * log. Something worked...
2345          */
2346         sk->sk_err_soft = 0;
2347         tp->rcv_tstamp = tcp_time_stamp;
2348         prior_packets = tp->packets_out;
2349         if (!prior_packets)
2350                 goto no_queue;
2351
2352         prior_in_flight = tcp_packets_in_flight(tp);
2353
2354         /* See if we can take anything off of the retransmit queue. */
2355         flag |= tcp_clean_rtx_queue(sk, &seq_rtt,
2356                                     icsk->icsk_ca_ops->rtt_sample ? &seq_usrtt : NULL);
2357
2358         if (tp->frto_counter)
2359                 tcp_process_frto(sk, prior_snd_una);
2360
2361         if (tcp_ack_is_dubious(sk, flag)) {
2362                 /* Advanve CWND, if state allows this. */
2363                 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
2364                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2365                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2366         } else {
2367                 if ((flag & FLAG_DATA_ACKED))
2368                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2369         }
2370
2371         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2372                 dst_confirm(sk->sk_dst_cache);
2373
2374         return 1;
2375
2376 no_queue:
2377         icsk->icsk_probes_out = 0;
2378
2379         /* If this ack opens up a zero window, clear backoff.  It was
2380          * being used to time the probes, and is probably far higher than
2381          * it needs to be for normal retransmission.
2382          */
2383         if (sk->sk_send_head)
2384                 tcp_ack_probe(sk);
2385         return 1;
2386
2387 old_ack:
2388         if (TCP_SKB_CB(skb)->sacked)
2389                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2390
2391 uninteresting_ack:
2392         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2393         return 0;
2394 }
2395
2396
2397 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2398  * But, this can also be called on packets in the established flow when
2399  * the fast version below fails.
2400  */
2401 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2402 {
2403         unsigned char *ptr;
2404         struct tcphdr *th = skb->h.th;
2405         int length=(th->doff*4)-sizeof(struct tcphdr);
2406
2407         ptr = (unsigned char *)(th + 1);
2408         opt_rx->saw_tstamp = 0;
2409
2410         while(length>0) {
2411                 int opcode=*ptr++;
2412                 int opsize;
2413
2414                 switch (opcode) {
2415                         case TCPOPT_EOL:
2416                                 return;
2417                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2418                                 length--;
2419                                 continue;
2420                         default:
2421                                 opsize=*ptr++;
2422                                 if (opsize < 2) /* "silly options" */
2423                                         return;
2424                                 if (opsize > length)
2425                                         return; /* don't parse partial options */
2426                                 switch(opcode) {
2427                                 case TCPOPT_MSS:
2428                                         if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2429                                                 u16 in_mss = ntohs(get_unaligned((__u16 *)ptr));
2430                                                 if (in_mss) {
2431                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2432                                                                 in_mss = opt_rx->user_mss;
2433                                                         opt_rx->mss_clamp = in_mss;
2434                                                 }
2435                                         }
2436                                         break;
2437                                 case TCPOPT_WINDOW:
2438                                         if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2439                                                 if (sysctl_tcp_window_scaling) {
2440                                                         __u8 snd_wscale = *(__u8 *) ptr;
2441                                                         opt_rx->wscale_ok = 1;
2442                                                         if (snd_wscale > 14) {
2443                                                                 if(net_ratelimit())
2444                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2445                                                                                "scaling value %d >14 received.\n",
2446                                                                                snd_wscale);
2447                                                                 snd_wscale = 14;
2448                                                         }
2449                                                         opt_rx->snd_wscale = snd_wscale;
2450                                                 }
2451                                         break;
2452                                 case TCPOPT_TIMESTAMP:
2453                                         if(opsize==TCPOLEN_TIMESTAMP) {
2454                                                 if ((estab && opt_rx->tstamp_ok) ||
2455                                                     (!estab && sysctl_tcp_timestamps)) {
2456                                                         opt_rx->saw_tstamp = 1;
2457                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr));
2458                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4)));
2459                                                 }
2460                                         }
2461                                         break;
2462                                 case TCPOPT_SACK_PERM:
2463                                         if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2464                                                 if (sysctl_tcp_sack) {
2465                                                         opt_rx->sack_ok = 1;
2466                                                         tcp_sack_reset(opt_rx);
2467                                                 }
2468                                         }
2469                                         break;
2470
2471                                 case TCPOPT_SACK:
2472                                         if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2473                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2474                                            opt_rx->sack_ok) {
2475                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2476                                         }
2477                                 };
2478                                 ptr+=opsize-2;
2479                                 length-=opsize;
2480                 };
2481         }
2482 }
2483
2484 /* Fast parse options. This hopes to only see timestamps.
2485  * If it is wrong it falls back on tcp_parse_options().
2486  */
2487 static inline int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2488                                          struct tcp_sock *tp)
2489 {
2490         if (th->doff == sizeof(struct tcphdr)>>2) {
2491                 tp->rx_opt.saw_tstamp = 0;
2492                 return 0;
2493         } else if (tp->rx_opt.tstamp_ok &&
2494                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2495                 __u32 *ptr = (__u32 *)(th + 1);
2496                 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2497                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2498                         tp->rx_opt.saw_tstamp = 1;
2499                         ++ptr;
2500                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2501                         ++ptr;
2502                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2503                         return 1;
2504                 }
2505         }
2506         tcp_parse_options(skb, &tp->rx_opt, 1);
2507         return 1;
2508 }
2509
2510 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2511 {
2512         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2513         tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2514 }
2515
2516 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2517 {
2518         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2519                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2520                  * extra check below makes sure this can only happen
2521                  * for pure ACK frames.  -DaveM
2522                  *
2523                  * Not only, also it occurs for expired timestamps.
2524                  */
2525
2526                 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2527                    xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2528                         tcp_store_ts_recent(tp);
2529         }
2530 }
2531
2532 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2533  *
2534  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2535  * it can pass through stack. So, the following predicate verifies that
2536  * this segment is not used for anything but congestion avoidance or
2537  * fast retransmit. Moreover, we even are able to eliminate most of such
2538  * second order effects, if we apply some small "replay" window (~RTO)
2539  * to timestamp space.
2540  *
2541  * All these measures still do not guarantee that we reject wrapped ACKs
2542  * on networks with high bandwidth, when sequence space is recycled fastly,
2543  * but it guarantees that such events will be very rare and do not affect
2544  * connection seriously. This doesn't look nice, but alas, PAWS is really
2545  * buggy extension.
2546  *
2547  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2548  * states that events when retransmit arrives after original data are rare.
2549  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2550  * the biggest problem on large power networks even with minor reordering.
2551  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2552  * up to bandwidth of 18Gigabit/sec. 8) ]
2553  */
2554
2555 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2556 {
2557         struct tcp_sock *tp = tcp_sk(sk);
2558         struct tcphdr *th = skb->h.th;
2559         u32 seq = TCP_SKB_CB(skb)->seq;
2560         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2561
2562         return (/* 1. Pure ACK with correct sequence number. */
2563                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2564
2565                 /* 2. ... and duplicate ACK. */
2566                 ack == tp->snd_una &&
2567
2568                 /* 3. ... and does not update window. */
2569                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2570
2571                 /* 4. ... and sits in replay window. */
2572                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
2573 }
2574
2575 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
2576 {
2577         const struct tcp_sock *tp = tcp_sk(sk);
2578         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2579                 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2580                 !tcp_disordered_ack(sk, skb));
2581 }
2582
2583 /* Check segment sequence number for validity.
2584  *
2585  * Segment controls are considered valid, if the segment
2586  * fits to the window after truncation to the window. Acceptability
2587  * of data (and SYN, FIN, of course) is checked separately.
2588  * See tcp_data_queue(), for example.
2589  *
2590  * Also, controls (RST is main one) are accepted using RCV.WUP instead
2591  * of RCV.NXT. Peer still did not advance his SND.UNA when we
2592  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2593  * (borrowed from freebsd)
2594  */
2595
2596 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2597 {
2598         return  !before(end_seq, tp->rcv_wup) &&
2599                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2600 }
2601
2602 /* When we get a reset we do this. */
2603 static void tcp_reset(struct sock *sk)
2604 {
2605         /* We want the right error as BSD sees it (and indeed as we do). */
2606         switch (sk->sk_state) {
2607                 case TCP_SYN_SENT:
2608                         sk->sk_err = ECONNREFUSED;
2609                         break;
2610                 case TCP_CLOSE_WAIT:
2611                         sk->sk_err = EPIPE;
2612                         break;
2613                 case TCP_CLOSE:
2614                         return;
2615                 default:
2616                         sk->sk_err = ECONNRESET;
2617         }
2618
2619         if (!sock_flag(sk, SOCK_DEAD))
2620                 sk->sk_error_report(sk);
2621
2622         tcp_done(sk);
2623 }
2624
2625 /*
2626  *      Process the FIN bit. This now behaves as it is supposed to work
2627  *      and the FIN takes effect when it is validly part of sequence
2628  *      space. Not before when we get holes.
2629  *
2630  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2631  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
2632  *      TIME-WAIT)
2633  *
2634  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
2635  *      close and we go into CLOSING (and later onto TIME-WAIT)
2636  *
2637  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2638  */
2639 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2640 {
2641         struct tcp_sock *tp = tcp_sk(sk);
2642
2643         inet_csk_schedule_ack(sk);
2644
2645         sk->sk_shutdown |= RCV_SHUTDOWN;
2646         sock_set_flag(sk, SOCK_DONE);
2647
2648         switch (sk->sk_state) {
2649                 case TCP_SYN_RECV:
2650                 case TCP_ESTABLISHED:
2651                         /* Move to CLOSE_WAIT */
2652                         tcp_set_state(sk, TCP_CLOSE_WAIT);
2653                         inet_csk(sk)->icsk_ack.pingpong = 1;
2654                         break;
2655
2656                 case TCP_CLOSE_WAIT:
2657                 case TCP_CLOSING:
2658                         /* Received a retransmission of the FIN, do
2659                          * nothing.
2660                          */
2661                         break;
2662                 case TCP_LAST_ACK:
2663                         /* RFC793: Remain in the LAST-ACK state. */
2664                         break;
2665
2666                 case TCP_FIN_WAIT1:
2667                         /* This case occurs when a simultaneous close
2668                          * happens, we must ack the received FIN and
2669                          * enter the CLOSING state.
2670                          */
2671                         tcp_send_ack(sk);
2672                         tcp_set_state(sk, TCP_CLOSING);
2673                         break;
2674                 case TCP_FIN_WAIT2:
2675                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
2676                         tcp_send_ack(sk);
2677                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2678                         break;
2679                 default:
2680                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2681                          * cases we should never reach this piece of code.
2682                          */
2683                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2684                                __FUNCTION__, sk->sk_state);
2685                         break;
2686         };
2687
2688         /* It _is_ possible, that we have something out-of-order _after_ FIN.
2689          * Probably, we should reset in this case. For now drop them.
2690          */
2691         __skb_queue_purge(&tp->out_of_order_queue);
2692         if (tp->rx_opt.sack_ok)
2693                 tcp_sack_reset(&tp->rx_opt);
2694         sk_stream_mem_reclaim(sk);
2695
2696         if (!sock_flag(sk, SOCK_DEAD)) {
2697                 sk->sk_state_change(sk);
2698
2699                 /* Do not send POLL_HUP for half duplex close. */
2700                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2701                     sk->sk_state == TCP_CLOSE)
2702                         sk_wake_async(sk, 1, POLL_HUP);
2703                 else
2704                         sk_wake_async(sk, 1, POLL_IN);
2705         }
2706 }
2707
2708 static __inline__ int
2709 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2710 {
2711         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2712                 if (before(seq, sp->start_seq))
2713                         sp->start_seq = seq;
2714                 if (after(end_seq, sp->end_seq))
2715                         sp->end_seq = end_seq;
2716                 return 1;
2717         }
2718         return 0;
2719 }
2720
2721 static inline void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
2722 {
2723         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2724                 if (before(seq, tp->rcv_nxt))
2725                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
2726                 else
2727                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
2728
2729                 tp->rx_opt.dsack = 1;
2730                 tp->duplicate_sack[0].start_seq = seq;
2731                 tp->duplicate_sack[0].end_seq = end_seq;
2732                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
2733         }
2734 }
2735
2736 static inline void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
2737 {
2738         if (!tp->rx_opt.dsack)
2739                 tcp_dsack_set(tp, seq, end_seq);
2740         else
2741                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2742 }
2743
2744 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2745 {
2746         struct tcp_sock *tp = tcp_sk(sk);
2747
2748         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2749             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2750                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2751                 tcp_enter_quickack_mode(sk);
2752
2753                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2754                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2755
2756                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2757                                 end_seq = tp->rcv_nxt;
2758                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2759                 }
2760         }
2761
2762         tcp_send_ack(sk);
2763 }
2764
2765 /* These routines update the SACK block as out-of-order packets arrive or
2766  * in-order packets close up the sequence space.
2767  */
2768 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
2769 {
2770         int this_sack;
2771         struct tcp_sack_block *sp = &tp->selective_acks[0];
2772         struct tcp_sack_block *swalk = sp+1;
2773
2774         /* See if the recent change to the first SACK eats into
2775          * or hits the sequence space of other SACK blocks, if so coalesce.
2776          */
2777         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
2778                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2779                         int i;
2780
2781                         /* Zap SWALK, by moving every further SACK up by one slot.
2782                          * Decrease num_sacks.
2783                          */
2784                         tp->rx_opt.num_sacks--;
2785                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2786                         for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
2787                                 sp[i] = sp[i+1];
2788                         continue;
2789                 }
2790                 this_sack++, swalk++;
2791         }
2792 }
2793
2794 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2795 {
2796         __u32 tmp;
2797
2798         tmp = sack1->start_seq;
2799         sack1->start_seq = sack2->start_seq;
2800         sack2->start_seq = tmp;
2801
2802         tmp = sack1->end_seq;
2803         sack1->end_seq = sack2->end_seq;
2804         sack2->end_seq = tmp;
2805 }
2806
2807 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
2808 {
2809         struct tcp_sock *tp = tcp_sk(sk);
2810         struct tcp_sack_block *sp = &tp->selective_acks[0];
2811         int cur_sacks = tp->rx_opt.num_sacks;
2812         int this_sack;
2813
2814         if (!cur_sacks)
2815                 goto new_sack;
2816
2817         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
2818                 if (tcp_sack_extend(sp, seq, end_seq)) {
2819                         /* Rotate this_sack to the first one. */
2820                         for (; this_sack>0; this_sack--, sp--)
2821                                 tcp_sack_swap(sp, sp-1);
2822                         if (cur_sacks > 1)
2823                                 tcp_sack_maybe_coalesce(tp);
2824                         return;
2825                 }
2826         }
2827
2828         /* Could not find an adjacent existing SACK, build a new one,
2829          * put it at the front, and shift everyone else down.  We
2830          * always know there is at least one SACK present already here.
2831          *
2832          * If the sack array is full, forget about the last one.
2833          */
2834         if (this_sack >= 4) {
2835                 this_sack--;
2836                 tp->rx_opt.num_sacks--;
2837                 sp--;
2838         }
2839         for(; this_sack > 0; this_sack--, sp--)
2840                 *sp = *(sp-1);
2841
2842 new_sack:
2843         /* Build the new head SACK, and we're done. */
2844         sp->start_seq = seq;
2845         sp->end_seq = end_seq;
2846         tp->rx_opt.num_sacks++;
2847         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2848 }
2849
2850 /* RCV.NXT advances, some SACKs should be eaten. */
2851
2852 static void tcp_sack_remove(struct tcp_sock *tp)
2853 {
2854         struct tcp_sack_block *sp = &tp->selective_acks[0];
2855         int num_sacks = tp->rx_opt.num_sacks;
2856         int this_sack;
2857
2858         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2859         if (skb_queue_empty(&tp->out_of_order_queue)) {
2860                 tp->rx_opt.num_sacks = 0;
2861                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
2862                 return;
2863         }
2864
2865         for(this_sack = 0; this_sack < num_sacks; ) {
2866                 /* Check if the start of the sack is covered by RCV.NXT. */
2867                 if (!before(tp->rcv_nxt, sp->start_seq)) {
2868                         int i;
2869
2870                         /* RCV.NXT must cover all the block! */
2871                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
2872
2873                         /* Zap this SACK, by moving forward any other SACKS. */
2874                         for (i=this_sack+1; i < num_sacks; i++)
2875                                 tp->selective_acks[i-1] = tp->selective_acks[i];
2876                         num_sacks--;
2877                         continue;
2878                 }
2879                 this_sack++;
2880                 sp++;
2881         }
2882         if (num_sacks != tp->rx_opt.num_sacks) {
2883                 tp->rx_opt.num_sacks = num_sacks;
2884                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2885         }
2886 }
2887
2888 /* This one checks to see if we can put data from the
2889  * out_of_order queue into the receive_queue.
2890  */
2891 static void tcp_ofo_queue(struct sock *sk)
2892 {
2893         struct tcp_sock *tp = tcp_sk(sk);
2894         __u32 dsack_high = tp->rcv_nxt;
2895         struct sk_buff *skb;
2896
2897         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
2898                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
2899                         break;
2900
2901                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
2902                         __u32 dsack = dsack_high;
2903                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
2904                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
2905                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
2906                 }
2907
2908                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
2909                         SOCK_DEBUG(sk, "ofo packet was already received \n");
2910                         __skb_unlink(skb, &tp->out_of_order_queue);
2911                         __kfree_skb(skb);
2912                         continue;
2913                 }
2914                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
2915                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
2916                            TCP_SKB_CB(skb)->end_seq);
2917
2918                 __skb_unlink(skb, &tp->out_of_order_queue);
2919                 __skb_queue_tail(&sk->sk_receive_queue, skb);
2920                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2921                 if(skb->h.th->fin)
2922                         tcp_fin(skb, sk, skb->h.th);
2923         }
2924 }
2925
2926 static int tcp_prune_queue(struct sock *sk);
2927
2928 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
2929 {
2930         struct tcphdr *th = skb->h.th;
2931         struct tcp_sock *tp = tcp_sk(sk);
2932         int eaten = -1;
2933
2934         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
2935                 goto drop;
2936
2937         __skb_pull(skb, th->doff*4);
2938
2939         TCP_ECN_accept_cwr(tp, skb);
2940
2941         if (tp->rx_opt.dsack) {
2942                 tp->rx_opt.dsack = 0;
2943                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
2944                                                     4 - tp->rx_opt.tstamp_ok);
2945         }
2946
2947         /*  Queue data for delivery to the user.
2948          *  Packets in sequence go to the receive queue.
2949          *  Out of sequence packets to the out_of_order_queue.
2950          */
2951         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
2952                 if (tcp_receive_window(tp) == 0)
2953                         goto out_of_window;
2954
2955                 /* Ok. In sequence. In window. */
2956                 if (tp->ucopy.task == current &&
2957                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
2958                     sock_owned_by_user(sk) && !tp->urg_data) {
2959                         int chunk = min_t(unsigned int, skb->len,
2960                                                         tp->ucopy.len);
2961
2962                         __set_current_state(TASK_RUNNING);
2963
2964                         local_bh_enable();
2965                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
2966                                 tp->ucopy.len -= chunk;
2967                                 tp->copied_seq += chunk;
2968                                 eaten = (chunk == skb->len && !th->fin);
2969                                 tcp_rcv_space_adjust(sk);
2970                         }
2971                         local_bh_disable();
2972                 }
2973
2974                 if (eaten <= 0) {
2975 queue_and_out:
2976                         if (eaten < 0 &&
2977                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
2978                              !sk_stream_rmem_schedule(sk, skb))) {
2979                                 if (tcp_prune_queue(sk) < 0 ||
2980                                     !sk_stream_rmem_schedule(sk, skb))
2981                                         goto drop;
2982                         }
2983                         sk_stream_set_owner_r(skb, sk);
2984                         __skb_queue_tail(&sk->sk_receive_queue, skb);
2985                 }
2986                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2987                 if(skb->len)
2988                         tcp_event_data_recv(sk, tp, skb);
2989                 if(th->fin)
2990                         tcp_fin(skb, sk, th);
2991
2992                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
2993                         tcp_ofo_queue(sk);
2994
2995                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
2996                          * gap in queue is filled.
2997                          */
2998                         if (skb_queue_empty(&tp->out_of_order_queue))
2999                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3000                 }
3001
3002                 if (tp->rx_opt.num_sacks)
3003                         tcp_sack_remove(tp);
3004
3005                 tcp_fast_path_check(sk, tp);
3006
3007                 if (eaten > 0)
3008                         __kfree_skb(skb);
3009                 else if (!sock_flag(sk, SOCK_DEAD))
3010                         sk->sk_data_ready(sk, 0);
3011                 return;
3012         }
3013
3014         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3015                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3016                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3017                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3018
3019 out_of_window:
3020                 tcp_enter_quickack_mode(sk);
3021                 inet_csk_schedule_ack(sk);
3022 drop:
3023                 __kfree_skb(skb);
3024                 return;
3025         }
3026
3027         /* Out of window. F.e. zero window probe. */
3028         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3029                 goto out_of_window;
3030
3031         tcp_enter_quickack_mode(sk);
3032
3033         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3034                 /* Partial packet, seq < rcv_next < end_seq */
3035                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3036                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3037                            TCP_SKB_CB(skb)->end_seq);
3038
3039                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3040                 
3041                 /* If window is closed, drop tail of packet. But after
3042                  * remembering D-SACK for its head made in previous line.
3043                  */
3044                 if (!tcp_receive_window(tp))
3045                         goto out_of_window;
3046                 goto queue_and_out;
3047         }
3048
3049         TCP_ECN_check_ce(tp, skb);
3050
3051         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3052             !sk_stream_rmem_schedule(sk, skb)) {
3053                 if (tcp_prune_queue(sk) < 0 ||
3054                     !sk_stream_rmem_schedule(sk, skb))
3055                         goto drop;
3056         }
3057
3058         /* Disable header prediction. */
3059         tp->pred_flags = 0;
3060         inet_csk_schedule_ack(sk);
3061
3062         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3063                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3064
3065         sk_stream_set_owner_r(skb, sk);
3066
3067         if (!skb_peek(&tp->out_of_order_queue)) {
3068                 /* Initial out of order segment, build 1 SACK. */
3069                 if (tp->rx_opt.sack_ok) {
3070                         tp->rx_opt.num_sacks = 1;
3071                         tp->rx_opt.dsack     = 0;
3072                         tp->rx_opt.eff_sacks = 1;
3073                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3074                         tp->selective_acks[0].end_seq =
3075                                                 TCP_SKB_CB(skb)->end_seq;
3076                 }
3077                 __skb_queue_head(&tp->out_of_order_queue,skb);
3078         } else {
3079                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3080                 u32 seq = TCP_SKB_CB(skb)->seq;
3081                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3082
3083                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3084                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3085
3086                         if (!tp->rx_opt.num_sacks ||
3087                             tp->selective_acks[0].end_seq != seq)
3088                                 goto add_sack;
3089
3090                         /* Common case: data arrive in order after hole. */
3091                         tp->selective_acks[0].end_seq = end_seq;
3092                         return;
3093                 }
3094
3095                 /* Find place to insert this segment. */
3096                 do {
3097                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3098                                 break;
3099                 } while ((skb1 = skb1->prev) !=
3100                          (struct sk_buff*)&tp->out_of_order_queue);
3101
3102                 /* Do skb overlap to previous one? */
3103                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3104                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3105                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3106                                 /* All the bits are present. Drop. */
3107                                 __kfree_skb(skb);
3108                                 tcp_dsack_set(tp, seq, end_seq);
3109                                 goto add_sack;
3110                         }
3111                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3112                                 /* Partial overlap. */
3113                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3114                         } else {
3115                                 skb1 = skb1->prev;
3116                         }
3117                 }
3118                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3119                 
3120                 /* And clean segments covered by new one as whole. */
3121                 while ((skb1 = skb->next) !=
3122                        (struct sk_buff*)&tp->out_of_order_queue &&
3123                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3124                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3125                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3126                                break;
3127                        }
3128                        __skb_unlink(skb1, &tp->out_of_order_queue);
3129                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3130                        __kfree_skb(skb1);
3131                 }
3132
3133 add_sack:
3134                 if (tp->rx_opt.sack_ok)
3135                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3136         }
3137 }
3138
3139 /* Collapse contiguous sequence of skbs head..tail with
3140  * sequence numbers start..end.
3141  * Segments with FIN/SYN are not collapsed (only because this
3142  * simplifies code)
3143  */
3144 static void
3145 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3146              struct sk_buff *head, struct sk_buff *tail,
3147              u32 start, u32 end)
3148 {
3149         struct sk_buff *skb;
3150
3151         /* First, check that queue is collapsable and find
3152          * the point where collapsing can be useful. */
3153         for (skb = head; skb != tail; ) {
3154                 /* No new bits? It is possible on ofo queue. */
3155                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3156                         struct sk_buff *next = skb->next;
3157                         __skb_unlink(skb, list);
3158                         __kfree_skb(skb);
3159                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3160                         skb = next;
3161                         continue;
3162                 }
3163
3164                 /* The first skb to collapse is:
3165                  * - not SYN/FIN and
3166                  * - bloated or contains data before "start" or
3167                  *   overlaps to the next one.
3168                  */
3169                 if (!skb->h.th->syn && !skb->h.th->fin &&
3170                     (tcp_win_from_space(skb->truesize) > skb->len ||
3171                      before(TCP_SKB_CB(skb)->seq, start) ||
3172                      (skb->next != tail &&
3173                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3174                         break;
3175
3176                 /* Decided to skip this, advance start seq. */
3177                 start = TCP_SKB_CB(skb)->end_seq;
3178                 skb = skb->next;
3179         }
3180         if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3181                 return;
3182
3183         while (before(start, end)) {
3184                 struct sk_buff *nskb;
3185                 int header = skb_headroom(skb);
3186                 int copy = SKB_MAX_ORDER(header, 0);
3187
3188                 /* Too big header? This can happen with IPv6. */
3189                 if (copy < 0)
3190                         return;
3191                 if (end-start < copy)
3192                         copy = end-start;
3193                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3194                 if (!nskb)
3195                         return;
3196                 skb_reserve(nskb, header);
3197                 memcpy(nskb->head, skb->head, header);
3198                 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3199                 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3200                 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3201                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3202                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3203                 __skb_insert(nskb, skb->prev, skb, list);
3204                 sk_stream_set_owner_r(nskb, sk);
3205
3206                 /* Copy data, releasing collapsed skbs. */
3207                 while (copy > 0) {
3208                         int offset = start - TCP_SKB_CB(skb)->seq;
3209                         int size = TCP_SKB_CB(skb)->end_seq - start;
3210
3211                         if (offset < 0) BUG();
3212                         if (size > 0) {
3213                                 size = min(copy, size);
3214                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3215                                         BUG();
3216                                 TCP_SKB_CB(nskb)->end_seq += size;
3217                                 copy -= size;
3218                                 start += size;
3219                         }
3220                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3221                                 struct sk_buff *next = skb->next;
3222                                 __skb_unlink(skb, list);
3223                                 __kfree_skb(skb);
3224                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3225                                 skb = next;
3226                                 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3227                                         return;
3228                         }
3229                 }
3230         }
3231 }
3232
3233 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3234  * and tcp_collapse() them until all the queue is collapsed.
3235  */
3236 static void tcp_collapse_ofo_queue(struct sock *sk)
3237 {
3238         struct tcp_sock *tp = tcp_sk(sk);
3239         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3240         struct sk_buff *head;
3241         u32 start, end;
3242
3243         if (skb == NULL)
3244                 return;
3245
3246         start = TCP_SKB_CB(skb)->seq;
3247         end = TCP_SKB_CB(skb)->end_seq;
3248         head = skb;
3249
3250         for (;;) {
3251                 skb = skb->next;
3252
3253                 /* Segment is terminated when we see gap or when
3254                  * we are at the end of all the queue. */
3255                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3256                     after(TCP_SKB_CB(skb)->seq, end) ||
3257                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3258                         tcp_collapse(sk, &tp->out_of_order_queue,
3259                                      head, skb, start, end);
3260                         head = skb;
3261                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3262                                 break;
3263                         /* Start new segment */
3264                         start = TCP_SKB_CB(skb)->seq;
3265                         end = TCP_SKB_CB(skb)->end_seq;
3266                 } else {
3267                         if (before(TCP_SKB_CB(skb)->seq, start))
3268                                 start = TCP_SKB_CB(skb)->seq;
3269                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3270                                 end = TCP_SKB_CB(skb)->end_seq;
3271                 }
3272         }
3273 }
3274
3275 /* Reduce allocated memory if we can, trying to get
3276  * the socket within its memory limits again.
3277  *
3278  * Return less than zero if we should start dropping frames
3279  * until the socket owning process reads some of the data
3280  * to stabilize the situation.
3281  */
3282 static int tcp_prune_queue(struct sock *sk)
3283 {
3284         struct tcp_sock *tp = tcp_sk(sk); 
3285
3286         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3287
3288         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3289
3290         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3291                 tcp_clamp_window(sk, tp);
3292         else if (tcp_memory_pressure)
3293                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3294
3295         tcp_collapse_ofo_queue(sk);
3296         tcp_collapse(sk, &sk->sk_receive_queue,
3297                      sk->sk_receive_queue.next,
3298                      (struct sk_buff*)&sk->sk_receive_queue,
3299                      tp->copied_seq, tp->rcv_nxt);
3300         sk_stream_mem_reclaim(sk);
3301
3302         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3303                 return 0;
3304
3305         /* Collapsing did not help, destructive actions follow.
3306          * This must not ever occur. */
3307
3308         /* First, purge the out_of_order queue. */
3309         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3310                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3311                 __skb_queue_purge(&tp->out_of_order_queue);
3312
3313                 /* Reset SACK state.  A conforming SACK implementation will
3314                  * do the same at a timeout based retransmit.  When a connection
3315                  * is in a sad state like this, we care only about integrity
3316                  * of the connection not performance.
3317                  */
3318                 if (tp->rx_opt.sack_ok)
3319                         tcp_sack_reset(&tp->rx_opt);
3320                 sk_stream_mem_reclaim(sk);
3321         }
3322
3323         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3324                 return 0;
3325
3326         /* If we are really being abused, tell the caller to silently
3327          * drop receive data on the floor.  It will get retransmitted
3328          * and hopefully then we'll have sufficient space.
3329          */
3330         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3331
3332         /* Massive buffer overcommit. */
3333         tp->pred_flags = 0;
3334         return -1;
3335 }
3336
3337
3338 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3339  * As additional protections, we do not touch cwnd in retransmission phases,
3340  * and if application hit its sndbuf limit recently.
3341  */
3342 void tcp_cwnd_application_limited(struct sock *sk)
3343 {
3344         struct tcp_sock *tp = tcp_sk(sk);
3345
3346         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3347             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3348                 /* Limited by application or receiver window. */
3349                 u32 win_used = max(tp->snd_cwnd_used, 2U);
3350                 if (win_used < tp->snd_cwnd) {
3351                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3352                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3353                 }
3354                 tp->snd_cwnd_used = 0;
3355         }
3356         tp->snd_cwnd_stamp = tcp_time_stamp;
3357 }
3358
3359 static inline int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3360 {
3361         /* If the user specified a specific send buffer setting, do
3362          * not modify it.
3363          */
3364         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3365                 return 0;
3366
3367         /* If we are under global TCP memory pressure, do not expand.  */
3368         if (tcp_memory_pressure)
3369                 return 0;
3370
3371         /* If we are under soft global TCP memory pressure, do not expand.  */
3372         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3373                 return 0;
3374
3375         /* If we filled the congestion window, do not expand.  */
3376         if (tp->packets_out >= tp->snd_cwnd)
3377                 return 0;
3378
3379         return 1;
3380 }
3381
3382 /* When incoming ACK allowed to free some skb from write_queue,
3383  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3384  * on the exit from tcp input handler.
3385  *
3386  * PROBLEM: sndbuf expansion does not work well with largesend.
3387  */
3388 static void tcp_new_space(struct sock *sk)
3389 {
3390         struct tcp_sock *tp = tcp_sk(sk);
3391
3392         if (tcp_should_expand_sndbuf(sk, tp)) {
3393                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3394                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3395                     demanded = max_t(unsigned int, tp->snd_cwnd,
3396                                                    tp->reordering + 1);
3397                 sndmem *= 2*demanded;
3398                 if (sndmem > sk->sk_sndbuf)
3399                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3400                 tp->snd_cwnd_stamp = tcp_time_stamp;
3401         }
3402
3403         sk->sk_write_space(sk);
3404 }
3405
3406 static inline void tcp_check_space(struct sock *sk)
3407 {
3408         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3409                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3410                 if (sk->sk_socket &&
3411                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3412                         tcp_new_space(sk);
3413         }
3414 }
3415
3416 static __inline__ void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3417 {
3418         tcp_push_pending_frames(sk, tp);
3419         tcp_check_space(sk);
3420 }
3421
3422 /*
3423  * Check if sending an ack is needed.
3424  */
3425 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3426 {
3427         struct tcp_sock *tp = tcp_sk(sk);
3428
3429             /* More than one full frame received... */
3430         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3431              /* ... and right edge of window advances far enough.
3432               * (tcp_recvmsg() will send ACK otherwise). Or...
3433               */
3434              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3435             /* We ACK each frame or... */
3436             tcp_in_quickack_mode(sk) ||
3437             /* We have out of order data. */
3438             (ofo_possible &&
3439              skb_peek(&tp->out_of_order_queue))) {
3440                 /* Then ack it now */
3441                 tcp_send_ack(sk);
3442         } else {
3443                 /* Else, send delayed ack. */
3444                 tcp_send_delayed_ack(sk);
3445         }
3446 }
3447
3448 static __inline__ void tcp_ack_snd_check(struct sock *sk)
3449 {
3450         if (!inet_csk_ack_scheduled(sk)) {
3451                 /* We sent a data segment already. */
3452                 return;
3453         }
3454         __tcp_ack_snd_check(sk, 1);
3455 }
3456
3457 /*
3458  *      This routine is only called when we have urgent data
3459  *      signalled. Its the 'slow' part of tcp_urg. It could be
3460  *      moved inline now as tcp_urg is only called from one
3461  *      place. We handle URGent data wrong. We have to - as
3462  *      BSD still doesn't use the correction from RFC961.
3463  *      For 1003.1g we should support a new option TCP_STDURG to permit
3464  *      either form (or just set the sysctl tcp_stdurg).
3465  */
3466  
3467 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3468 {
3469         struct tcp_sock *tp = tcp_sk(sk);
3470         u32 ptr = ntohs(th->urg_ptr);
3471
3472         if (ptr && !sysctl_tcp_stdurg)
3473                 ptr--;
3474         ptr += ntohl(th->seq);
3475
3476         /* Ignore urgent data that we've already seen and read. */
3477         if (after(tp->copied_seq, ptr))
3478                 return;
3479
3480         /* Do not replay urg ptr.
3481          *
3482          * NOTE: interesting situation not covered by specs.
3483          * Misbehaving sender may send urg ptr, pointing to segment,
3484          * which we already have in ofo queue. We are not able to fetch
3485          * such data and will stay in TCP_URG_NOTYET until will be eaten
3486          * by recvmsg(). Seems, we are not obliged to handle such wicked
3487          * situations. But it is worth to think about possibility of some
3488          * DoSes using some hypothetical application level deadlock.
3489          */
3490         if (before(ptr, tp->rcv_nxt))
3491                 return;
3492
3493         /* Do we already have a newer (or duplicate) urgent pointer? */
3494         if (tp->urg_data && !after(ptr, tp->urg_seq))
3495                 return;
3496
3497         /* Tell the world about our new urgent pointer. */
3498         sk_send_sigurg(sk);
3499
3500         /* We may be adding urgent data when the last byte read was
3501          * urgent. To do this requires some care. We cannot just ignore
3502          * tp->copied_seq since we would read the last urgent byte again
3503          * as data, nor can we alter copied_seq until this data arrives
3504          * or we break the sematics of SIOCATMARK (and thus sockatmark())
3505          *
3506          * NOTE. Double Dutch. Rendering to plain English: author of comment
3507          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3508          * and expect that both A and B disappear from stream. This is _wrong_.
3509          * Though this happens in BSD with high probability, this is occasional.
3510          * Any application relying on this is buggy. Note also, that fix "works"
3511          * only in this artificial test. Insert some normal data between A and B and we will
3512          * decline of BSD again. Verdict: it is better to remove to trap
3513          * buggy users.
3514          */
3515         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3516             !sock_flag(sk, SOCK_URGINLINE) &&
3517             tp->copied_seq != tp->rcv_nxt) {
3518                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3519                 tp->copied_seq++;
3520                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3521                         __skb_unlink(skb, &sk->sk_receive_queue);
3522                         __kfree_skb(skb);
3523                 }
3524         }
3525
3526         tp->urg_data   = TCP_URG_NOTYET;
3527         tp->urg_seq    = ptr;
3528
3529         /* Disable header prediction. */
3530         tp->pred_flags = 0;
3531 }
3532
3533 /* This is the 'fast' part of urgent handling. */
3534 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3535 {
3536         struct tcp_sock *tp = tcp_sk(sk);
3537
3538         /* Check if we get a new urgent pointer - normally not. */
3539         if (th->urg)
3540                 tcp_check_urg(sk,th);
3541
3542         /* Do we wait for any urgent data? - normally not... */
3543         if (tp->urg_data == TCP_URG_NOTYET) {
3544                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3545                           th->syn;
3546
3547                 /* Is the urgent pointer pointing into this packet? */   
3548                 if (ptr < skb->len) {
3549                         u8 tmp;
3550                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3551                                 BUG();
3552                         tp->urg_data = TCP_URG_VALID | tmp;
3553                         if (!sock_flag(sk, SOCK_DEAD))
3554                                 sk->sk_data_ready(sk, 0);
3555                 }
3556         }
3557 }
3558
3559 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3560 {
3561         struct tcp_sock *tp = tcp_sk(sk);
3562         int chunk = skb->len - hlen;
3563         int err;
3564
3565         local_bh_enable();
3566         if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3567                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3568         else
3569                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3570                                                        tp->ucopy.iov);
3571
3572         if (!err) {
3573                 tp->ucopy.len -= chunk;
3574                 tp->copied_seq += chunk;
3575                 tcp_rcv_space_adjust(sk);
3576         }
3577
3578         local_bh_disable();
3579         return err;
3580 }
3581
3582 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3583 {
3584         int result;
3585
3586         if (sock_owned_by_user(sk)) {
3587                 local_bh_enable();
3588                 result = __tcp_checksum_complete(skb);
3589                 local_bh_disable();
3590         } else {
3591                 result = __tcp_checksum_complete(skb);
3592         }
3593         return result;
3594 }
3595
3596 static __inline__ int
3597 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3598 {
3599         return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3600                 __tcp_checksum_complete_user(sk, skb);
3601 }
3602
3603 /*
3604  *      TCP receive function for the ESTABLISHED state. 
3605  *
3606  *      It is split into a fast path and a slow path. The fast path is 
3607  *      disabled when:
3608  *      - A zero window was announced from us - zero window probing
3609  *        is only handled properly in the slow path. 
3610  *      - Out of order segments arrived.
3611  *      - Urgent data is expected.
3612  *      - There is no buffer space left
3613  *      - Unexpected TCP flags/window values/header lengths are received
3614  *        (detected by checking the TCP header against pred_flags) 
3615  *      - Data is sent in both directions. Fast path only supports pure senders
3616  *        or pure receivers (this means either the sequence number or the ack
3617  *        value must stay constant)
3618  *      - Unexpected TCP option.
3619  *
3620  *      When these conditions are not satisfied it drops into a standard 
3621  *      receive procedure patterned after RFC793 to handle all cases.
3622  *      The first three cases are guaranteed by proper pred_flags setting,
3623  *      the rest is checked inline. Fast processing is turned on in 
3624  *      tcp_data_queue when everything is OK.
3625  */
3626 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3627                         struct tcphdr *th, unsigned len)
3628 {
3629         struct tcp_sock *tp = tcp_sk(sk);
3630
3631         /*
3632          *      Header prediction.
3633          *      The code loosely follows the one in the famous 
3634          *      "30 instruction TCP receive" Van Jacobson mail.
3635          *      
3636          *      Van's trick is to deposit buffers into socket queue 
3637          *      on a device interrupt, to call tcp_recv function
3638          *      on the receive process context and checksum and copy
3639          *      the buffer to user space. smart...
3640          *
3641          *      Our current scheme is not silly either but we take the 
3642          *      extra cost of the net_bh soft interrupt processing...
3643          *      We do checksum and copy also but from device to kernel.
3644          */
3645
3646         tp->rx_opt.saw_tstamp = 0;
3647
3648         /*      pred_flags is 0xS?10 << 16 + snd_wnd
3649          *      if header_predition is to be made
3650          *      'S' will always be tp->tcp_header_len >> 2
3651          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3652          *  turn it off (when there are holes in the receive 
3653          *       space for instance)
3654          *      PSH flag is ignored.
3655          */
3656
3657         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3658                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3659                 int tcp_header_len = tp->tcp_header_len;
3660
3661                 /* Timestamp header prediction: tcp_header_len
3662                  * is automatically equal to th->doff*4 due to pred_flags
3663                  * match.
3664                  */
3665
3666                 /* Check timestamp */
3667                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3668                         __u32 *ptr = (__u32 *)(th + 1);
3669
3670                         /* No? Slow path! */
3671                         if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3672                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3673                                 goto slow_path;
3674
3675                         tp->rx_opt.saw_tstamp = 1;
3676                         ++ptr; 
3677                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3678                         ++ptr;
3679                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3680
3681                         /* If PAWS failed, check it more carefully in slow path */
3682                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
3683                                 goto slow_path;
3684
3685                         /* DO NOT update ts_recent here, if checksum fails
3686                          * and timestamp was corrupted part, it will result
3687                          * in a hung connection since we will drop all
3688                          * future packets due to the PAWS test.
3689                          */
3690                 }
3691
3692                 if (len <= tcp_header_len) {
3693                         /* Bulk data transfer: sender */
3694                         if (len == tcp_header_len) {
3695                                 /* Predicted packet is in window by definition.
3696                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3697                                  * Hence, check seq<=rcv_wup reduces to:
3698                                  */
3699                                 if (tcp_header_len ==
3700                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3701                                     tp->rcv_nxt == tp->rcv_wup)
3702                                         tcp_store_ts_recent(tp);
3703
3704                                 tcp_rcv_rtt_measure_ts(sk, skb);
3705
3706                                 /* We know that such packets are checksummed
3707                                  * on entry.
3708                                  */
3709                                 tcp_ack(sk, skb, 0);
3710                                 __kfree_skb(skb); 
3711                                 tcp_data_snd_check(sk, tp);
3712                                 return 0;
3713                         } else { /* Header too small */
3714                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3715                                 goto discard;
3716                         }
3717                 } else {
3718                         int eaten = 0;
3719
3720                         if (tp->ucopy.task == current &&
3721                             tp->copied_seq == tp->rcv_nxt &&
3722                             len - tcp_header_len <= tp->ucopy.len &&
3723                             sock_owned_by_user(sk)) {
3724                                 __set_current_state(TASK_RUNNING);
3725
3726                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
3727                                         /* Predicted packet is in window by definition.
3728                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3729                                          * Hence, check seq<=rcv_wup reduces to:
3730                                          */
3731                                         if (tcp_header_len ==
3732                                             (sizeof(struct tcphdr) +
3733                                              TCPOLEN_TSTAMP_ALIGNED) &&
3734                                             tp->rcv_nxt == tp->rcv_wup)
3735                                                 tcp_store_ts_recent(tp);
3736
3737                                         tcp_rcv_rtt_measure_ts(sk, skb);
3738
3739                                         __skb_pull(skb, tcp_header_len);
3740                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3741                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
3742                                         eaten = 1;
3743                                 }
3744                         }
3745                         if (!eaten) {
3746                                 if (tcp_checksum_complete_user(sk, skb))
3747                                         goto csum_error;
3748
3749                                 /* Predicted packet is in window by definition.
3750                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3751                                  * Hence, check seq<=rcv_wup reduces to:
3752                                  */
3753                                 if (tcp_header_len ==
3754                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3755                                     tp->rcv_nxt == tp->rcv_wup)
3756                                         tcp_store_ts_recent(tp);
3757
3758                                 tcp_rcv_rtt_measure_ts(sk, skb);
3759
3760                                 if ((int)skb->truesize > sk->sk_forward_alloc)
3761                                         goto step5;
3762
3763                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
3764
3765                                 /* Bulk data transfer: receiver */
3766                                 __skb_pull(skb,tcp_header_len);
3767                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3768                                 sk_stream_set_owner_r(skb, sk);
3769                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3770                         }
3771
3772                         tcp_event_data_recv(sk, tp, skb);
3773
3774                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
3775                                 /* Well, only one small jumplet in fast path... */
3776                                 tcp_ack(sk, skb, FLAG_DATA);
3777                                 tcp_data_snd_check(sk, tp);
3778                                 if (!inet_csk_ack_scheduled(sk))
3779                                         goto no_ack;
3780                         }
3781
3782                         __tcp_ack_snd_check(sk, 0);
3783 no_ack:
3784                         if (eaten)
3785                                 __kfree_skb(skb);
3786                         else
3787                                 sk->sk_data_ready(sk, 0);
3788                         return 0;
3789                 }
3790         }
3791
3792 slow_path:
3793         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
3794                 goto csum_error;
3795
3796         /*
3797          * RFC1323: H1. Apply PAWS check first.
3798          */
3799         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
3800             tcp_paws_discard(sk, skb)) {
3801                 if (!th->rst) {
3802                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
3803                         tcp_send_dupack(sk, skb);
3804                         goto discard;
3805                 }
3806                 /* Resets are accepted even if PAWS failed.
3807
3808                    ts_recent update must be made after we are sure
3809                    that the packet is in window.
3810                  */
3811         }
3812
3813         /*
3814          *      Standard slow path.
3815          */
3816
3817         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3818                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3819                  * (RST) segments are validated by checking their SEQ-fields."
3820                  * And page 69: "If an incoming segment is not acceptable,
3821                  * an acknowledgment should be sent in reply (unless the RST bit
3822                  * is set, if so drop the segment and return)".
3823                  */
3824                 if (!th->rst)
3825                         tcp_send_dupack(sk, skb);
3826                 goto discard;
3827         }
3828
3829         if(th->rst) {
3830                 tcp_reset(sk);
3831                 goto discard;
3832         }
3833
3834         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3835
3836         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3837                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3838                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
3839                 tcp_reset(sk);
3840                 return 1;
3841         }
3842
3843 step5:
3844         if(th->ack)
3845                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3846
3847         tcp_rcv_rtt_measure_ts(sk, skb);
3848
3849         /* Process urgent data. */
3850         tcp_urg(sk, skb, th);
3851
3852         /* step 7: process the segment text */
3853         tcp_data_queue(sk, skb);
3854
3855         tcp_data_snd_check(sk, tp);
3856         tcp_ack_snd_check(sk);
3857         return 0;
3858
3859 csum_error:
3860         TCP_INC_STATS_BH(TCP_MIB_INERRS);
3861
3862 discard:
3863         __kfree_skb(skb);
3864         return 0;
3865 }
3866
3867 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
3868                                          struct tcphdr *th, unsigned len)
3869 {
3870         struct tcp_sock *tp = tcp_sk(sk);
3871         int saved_clamp = tp->rx_opt.mss_clamp;
3872
3873         tcp_parse_options(skb, &tp->rx_opt, 0);
3874
3875         if (th->ack) {
3876                 struct inet_connection_sock *icsk;
3877                 /* rfc793:
3878                  * "If the state is SYN-SENT then
3879                  *    first check the ACK bit
3880                  *      If the ACK bit is set
3881                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3882                  *        a reset (unless the RST bit is set, if so drop
3883                  *        the segment and return)"
3884                  *
3885                  *  We do not send data with SYN, so that RFC-correct
3886                  *  test reduces to:
3887                  */
3888                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
3889                         goto reset_and_undo;
3890
3891                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3892                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
3893                              tcp_time_stamp)) {
3894                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
3895                         goto reset_and_undo;
3896                 }
3897
3898                 /* Now ACK is acceptable.
3899                  *
3900                  * "If the RST bit is set
3901                  *    If the ACK was acceptable then signal the user "error:
3902                  *    connection reset", drop the segment, enter CLOSED state,
3903                  *    delete TCB, and return."
3904                  */
3905
3906                 if (th->rst) {
3907                         tcp_reset(sk);
3908                         goto discard;
3909                 }
3910
3911                 /* rfc793:
3912                  *   "fifth, if neither of the SYN or RST bits is set then
3913                  *    drop the segment and return."
3914                  *
3915                  *    See note below!
3916                  *                                        --ANK(990513)
3917                  */
3918                 if (!th->syn)
3919                         goto discard_and_undo;
3920
3921                 /* rfc793:
3922                  *   "If the SYN bit is on ...
3923                  *    are acceptable then ...
3924                  *    (our SYN has been ACKed), change the connection
3925                  *    state to ESTABLISHED..."
3926                  */
3927
3928                 TCP_ECN_rcv_synack(tp, th);
3929                 if (tp->ecn_flags&TCP_ECN_OK)
3930                         sock_set_flag(sk, SOCK_NO_LARGESEND);
3931
3932                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
3933                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3934
3935                 /* Ok.. it's good. Set up sequence numbers and
3936                  * move to established.
3937                  */
3938                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
3939                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
3940
3941                 /* RFC1323: The window in SYN & SYN/ACK segments is
3942                  * never scaled.
3943                  */
3944                 tp->snd_wnd = ntohs(th->window);
3945                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
3946
3947                 if (!tp->rx_opt.wscale_ok) {
3948                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
3949                         tp->window_clamp = min(tp->window_clamp, 65535U);
3950                 }
3951
3952                 if (tp->rx_opt.saw_tstamp) {
3953                         tp->rx_opt.tstamp_ok       = 1;
3954                         tp->tcp_header_len =
3955                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
3956                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
3957                         tcp_store_ts_recent(tp);
3958                 } else {
3959                         tp->tcp_header_len = sizeof(struct tcphdr);
3960                 }
3961
3962                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
3963                         tp->rx_opt.sack_ok |= 2;
3964
3965                 tcp_sync_mss(sk, tp->pmtu_cookie);
3966                 tcp_initialize_rcv_mss(sk);
3967
3968                 /* Remember, tcp_poll() does not lock socket!
3969                  * Change state from SYN-SENT only after copied_seq
3970                  * is initialized. */
3971                 tp->copied_seq = tp->rcv_nxt;
3972                 mb();
3973                 tcp_set_state(sk, TCP_ESTABLISHED);
3974
3975                 /* Make sure socket is routed, for correct metrics.  */
3976                 tp->af_specific->rebuild_header(sk);
3977
3978                 tcp_init_metrics(sk);
3979
3980                 tcp_init_congestion_control(sk);
3981
3982                 /* Prevent spurious tcp_cwnd_restart() on first data
3983                  * packet.
3984                  */
3985                 tp->lsndtime = tcp_time_stamp;
3986
3987                 tcp_init_buffer_space(sk);
3988
3989                 if (sock_flag(sk, SOCK_KEEPOPEN))
3990                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
3991
3992                 if (!tp->rx_opt.snd_wscale)
3993                         __tcp_fast_path_on(tp, tp->snd_wnd);
3994                 else
3995                         tp->pred_flags = 0;
3996
3997                 if (!sock_flag(sk, SOCK_DEAD)) {
3998                         sk->sk_state_change(sk);
3999                         sk_wake_async(sk, 0, POLL_OUT);
4000                 }
4001
4002                 icsk = inet_csk(sk);
4003
4004                 if (sk->sk_write_pending ||
4005                     icsk->icsk_accept_queue.rskq_defer_accept ||
4006                     icsk->icsk_ack.pingpong) {
4007                         /* Save one ACK. Data will be ready after
4008                          * several ticks, if write_pending is set.
4009                          *
4010                          * It may be deleted, but with this feature tcpdumps
4011                          * look so _wonderfully_ clever, that I was not able
4012                          * to stand against the temptation 8)     --ANK
4013                          */
4014                         inet_csk_schedule_ack(sk);
4015                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4016                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4017                         tcp_incr_quickack(sk);
4018                         tcp_enter_quickack_mode(sk);
4019                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4020                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4021
4022 discard:
4023                         __kfree_skb(skb);
4024                         return 0;
4025                 } else {
4026                         tcp_send_ack(sk);
4027                 }
4028                 return -1;
4029         }
4030
4031         /* No ACK in the segment */
4032
4033         if (th->rst) {
4034                 /* rfc793:
4035                  * "If the RST bit is set
4036                  *
4037                  *      Otherwise (no ACK) drop the segment and return."
4038                  */
4039
4040                 goto discard_and_undo;
4041         }
4042
4043         /* PAWS check. */
4044         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4045                 goto discard_and_undo;
4046
4047         if (th->syn) {
4048                 /* We see SYN without ACK. It is attempt of
4049                  * simultaneous connect with crossed SYNs.
4050                  * Particularly, it can be connect to self.
4051                  */
4052                 tcp_set_state(sk, TCP_SYN_RECV);
4053
4054                 if (tp->rx_opt.saw_tstamp) {
4055                         tp->rx_opt.tstamp_ok = 1;
4056                         tcp_store_ts_recent(tp);
4057                         tp->tcp_header_len =
4058                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4059                 } else {
4060                         tp->tcp_header_len = sizeof(struct tcphdr);
4061                 }
4062
4063                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4064                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4065
4066                 /* RFC1323: The window in SYN & SYN/ACK segments is
4067                  * never scaled.
4068                  */
4069                 tp->snd_wnd    = ntohs(th->window);
4070                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4071                 tp->max_window = tp->snd_wnd;
4072
4073                 TCP_ECN_rcv_syn(tp, th);
4074                 if (tp->ecn_flags&TCP_ECN_OK)
4075                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4076
4077                 tcp_sync_mss(sk, tp->pmtu_cookie);
4078                 tcp_initialize_rcv_mss(sk);
4079
4080
4081                 tcp_send_synack(sk);
4082 #if 0
4083                 /* Note, we could accept data and URG from this segment.
4084                  * There are no obstacles to make this.
4085                  *
4086                  * However, if we ignore data in ACKless segments sometimes,
4087                  * we have no reasons to accept it sometimes.
4088                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4089                  * is not flawless. So, discard packet for sanity.
4090                  * Uncomment this return to process the data.
4091                  */
4092                 return -1;
4093 #else
4094                 goto discard;
4095 #endif
4096         }
4097         /* "fifth, if neither of the SYN or RST bits is set then
4098          * drop the segment and return."
4099          */
4100
4101 discard_and_undo:
4102         tcp_clear_options(&tp->rx_opt);
4103         tp->rx_opt.mss_clamp = saved_clamp;
4104         goto discard;
4105
4106 reset_and_undo:
4107         tcp_clear_options(&tp->rx_opt);
4108         tp->rx_opt.mss_clamp = saved_clamp;
4109         return 1;
4110 }
4111
4112
4113 /*
4114  *      This function implements the receiving procedure of RFC 793 for
4115  *      all states except ESTABLISHED and TIME_WAIT. 
4116  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4117  *      address independent.
4118  */
4119         
4120 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4121                           struct tcphdr *th, unsigned len)
4122 {
4123         struct tcp_sock *tp = tcp_sk(sk);
4124         int queued = 0;
4125
4126         tp->rx_opt.saw_tstamp = 0;
4127
4128         switch (sk->sk_state) {
4129         case TCP_CLOSE:
4130                 goto discard;
4131
4132         case TCP_LISTEN:
4133                 if(th->ack)
4134                         return 1;
4135
4136                 if(th->rst)
4137                         goto discard;
4138
4139                 if(th->syn) {
4140                         if(tp->af_specific->conn_request(sk, skb) < 0)
4141                                 return 1;
4142
4143                         /* Now we have several options: In theory there is 
4144                          * nothing else in the frame. KA9Q has an option to 
4145                          * send data with the syn, BSD accepts data with the
4146                          * syn up to the [to be] advertised window and 
4147                          * Solaris 2.1 gives you a protocol error. For now 
4148                          * we just ignore it, that fits the spec precisely 
4149                          * and avoids incompatibilities. It would be nice in
4150                          * future to drop through and process the data.
4151                          *
4152                          * Now that TTCP is starting to be used we ought to 
4153                          * queue this data.
4154                          * But, this leaves one open to an easy denial of
4155                          * service attack, and SYN cookies can't defend
4156                          * against this problem. So, we drop the data
4157                          * in the interest of security over speed.
4158                          */
4159                         goto discard;
4160                 }
4161                 goto discard;
4162
4163         case TCP_SYN_SENT:
4164                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4165                 if (queued >= 0)
4166                         return queued;
4167
4168                 /* Do step6 onward by hand. */
4169                 tcp_urg(sk, skb, th);
4170                 __kfree_skb(skb);
4171                 tcp_data_snd_check(sk, tp);
4172                 return 0;
4173         }
4174
4175         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4176             tcp_paws_discard(sk, skb)) {
4177                 if (!th->rst) {
4178                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4179                         tcp_send_dupack(sk, skb);
4180                         goto discard;
4181                 }
4182                 /* Reset is accepted even if it did not pass PAWS. */
4183         }
4184
4185         /* step 1: check sequence number */
4186         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4187                 if (!th->rst)
4188                         tcp_send_dupack(sk, skb);
4189                 goto discard;
4190         }
4191
4192         /* step 2: check RST bit */
4193         if(th->rst) {
4194                 tcp_reset(sk);
4195                 goto discard;
4196         }
4197
4198         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4199
4200         /* step 3: check security and precedence [ignored] */
4201
4202         /*      step 4:
4203          *
4204          *      Check for a SYN in window.
4205          */
4206         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4207                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4208                 tcp_reset(sk);
4209                 return 1;
4210         }
4211
4212         /* step 5: check the ACK field */
4213         if (th->ack) {
4214                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4215
4216                 switch(sk->sk_state) {
4217                 case TCP_SYN_RECV:
4218                         if (acceptable) {
4219                                 tp->copied_seq = tp->rcv_nxt;
4220                                 mb();
4221                                 tcp_set_state(sk, TCP_ESTABLISHED);
4222                                 sk->sk_state_change(sk);
4223
4224                                 /* Note, that this wakeup is only for marginal
4225                                  * crossed SYN case. Passively open sockets
4226                                  * are not waked up, because sk->sk_sleep ==
4227                                  * NULL and sk->sk_socket == NULL.
4228                                  */
4229                                 if (sk->sk_socket) {
4230                                         sk_wake_async(sk,0,POLL_OUT);
4231                                 }
4232
4233                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4234                                 tp->snd_wnd = ntohs(th->window) <<
4235                                               tp->rx_opt.snd_wscale;
4236                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4237                                             TCP_SKB_CB(skb)->seq);
4238
4239                                 /* tcp_ack considers this ACK as duplicate
4240                                  * and does not calculate rtt.
4241                                  * Fix it at least with timestamps.
4242                                  */
4243                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4244                                     !tp->srtt)
4245                                         tcp_ack_saw_tstamp(sk, NULL, 0);
4246
4247                                 if (tp->rx_opt.tstamp_ok)
4248                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4249
4250                                 /* Make sure socket is routed, for
4251                                  * correct metrics.
4252                                  */
4253                                 tp->af_specific->rebuild_header(sk);
4254
4255                                 tcp_init_metrics(sk);
4256
4257                                 tcp_init_congestion_control(sk);
4258
4259                                 /* Prevent spurious tcp_cwnd_restart() on
4260                                  * first data packet.
4261                                  */
4262                                 tp->lsndtime = tcp_time_stamp;
4263
4264                                 tcp_initialize_rcv_mss(sk);
4265                                 tcp_init_buffer_space(sk);
4266                                 tcp_fast_path_on(tp);
4267                         } else {
4268                                 return 1;
4269                         }
4270                         break;
4271
4272                 case TCP_FIN_WAIT1:
4273                         if (tp->snd_una == tp->write_seq) {
4274                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4275                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4276                                 dst_confirm(sk->sk_dst_cache);
4277
4278                                 if (!sock_flag(sk, SOCK_DEAD))
4279                                         /* Wake up lingering close() */
4280                                         sk->sk_state_change(sk);
4281                                 else {
4282                                         int tmo;
4283
4284                                         if (tp->linger2 < 0 ||
4285                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4286                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4287                                                 tcp_done(sk);
4288                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4289                                                 return 1;
4290                                         }
4291
4292                                         tmo = tcp_fin_time(sk);
4293                                         if (tmo > TCP_TIMEWAIT_LEN) {
4294                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4295                                         } else if (th->fin || sock_owned_by_user(sk)) {
4296                                                 /* Bad case. We could lose such FIN otherwise.
4297                                                  * It is not a big problem, but it looks confusing
4298                                                  * and not so rare event. We still can lose it now,
4299                                                  * if it spins in bh_lock_sock(), but it is really
4300                                                  * marginal case.
4301                                                  */
4302                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4303                                         } else {
4304                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4305                                                 goto discard;
4306                                         }
4307                                 }
4308                         }
4309                         break;
4310
4311                 case TCP_CLOSING:
4312                         if (tp->snd_una == tp->write_seq) {
4313                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4314                                 goto discard;
4315                         }
4316                         break;
4317
4318                 case TCP_LAST_ACK:
4319                         if (tp->snd_una == tp->write_seq) {
4320                                 tcp_update_metrics(sk);
4321                                 tcp_done(sk);
4322                                 goto discard;
4323                         }
4324                         break;
4325                 }
4326         } else
4327                 goto discard;
4328
4329         /* step 6: check the URG bit */
4330         tcp_urg(sk, skb, th);
4331
4332         /* step 7: process the segment text */
4333         switch (sk->sk_state) {
4334         case TCP_CLOSE_WAIT:
4335         case TCP_CLOSING:
4336         case TCP_LAST_ACK:
4337                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4338                         break;
4339         case TCP_FIN_WAIT1:
4340         case TCP_FIN_WAIT2:
4341                 /* RFC 793 says to queue data in these states,
4342                  * RFC 1122 says we MUST send a reset. 
4343                  * BSD 4.4 also does reset.
4344                  */
4345                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4346                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4347                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4348                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4349                                 tcp_reset(sk);
4350                                 return 1;
4351                         }
4352                 }
4353                 /* Fall through */
4354         case TCP_ESTABLISHED: 
4355                 tcp_data_queue(sk, skb);
4356                 queued = 1;
4357                 break;
4358         }
4359
4360         /* tcp_data could move socket to TIME-WAIT */
4361         if (sk->sk_state != TCP_CLOSE) {
4362                 tcp_data_snd_check(sk, tp);
4363                 tcp_ack_snd_check(sk);
4364         }
4365
4366         if (!queued) { 
4367 discard:
4368                 __kfree_skb(skb);
4369         }
4370         return 0;
4371 }
4372
4373 EXPORT_SYMBOL(sysctl_tcp_ecn);
4374 EXPORT_SYMBOL(sysctl_tcp_reordering);
4375 EXPORT_SYMBOL(tcp_parse_options);
4376 EXPORT_SYMBOL(tcp_rcv_established);
4377 EXPORT_SYMBOL(tcp_rcv_state_process);