[TCP]: whitespace cleanup
[linux-2.6.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105
106 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
109 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
110
111 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
112 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
113 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
114
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
116
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118
119 /* Adapt the MSS value used to make delayed ack decision to the
120  * real world.
121  */
122 static void tcp_measure_rcv_mss(struct sock *sk,
123                                 const struct sk_buff *skb)
124 {
125         struct inet_connection_sock *icsk = inet_csk(sk);
126         const unsigned int lss = icsk->icsk_ack.last_seg_size;
127         unsigned int len;
128
129         icsk->icsk_ack.last_seg_size = 0;
130
131         /* skb->len may jitter because of SACKs, even if peer
132          * sends good full-sized frames.
133          */
134         len = skb_shinfo(skb)->gso_size ?: skb->len;
135         if (len >= icsk->icsk_ack.rcv_mss) {
136                 icsk->icsk_ack.rcv_mss = len;
137         } else {
138                 /* Otherwise, we make more careful check taking into account,
139                  * that SACKs block is variable.
140                  *
141                  * "len" is invariant segment length, including TCP header.
142                  */
143                 len += skb->data - skb->h.raw;
144                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
145                     /* If PSH is not set, packet should be
146                      * full sized, provided peer TCP is not badly broken.
147                      * This observation (if it is correct 8)) allows
148                      * to handle super-low mtu links fairly.
149                      */
150                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
151                      !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
152                         /* Subtract also invariant (if peer is RFC compliant),
153                          * tcp header plus fixed timestamp option length.
154                          * Resulting "len" is MSS free of SACK jitter.
155                          */
156                         len -= tcp_sk(sk)->tcp_header_len;
157                         icsk->icsk_ack.last_seg_size = len;
158                         if (len == lss) {
159                                 icsk->icsk_ack.rcv_mss = len;
160                                 return;
161                         }
162                 }
163                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
164                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
165                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
166         }
167 }
168
169 static void tcp_incr_quickack(struct sock *sk)
170 {
171         struct inet_connection_sock *icsk = inet_csk(sk);
172         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
173
174         if (quickacks==0)
175                 quickacks=2;
176         if (quickacks > icsk->icsk_ack.quick)
177                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
178 }
179
180 void tcp_enter_quickack_mode(struct sock *sk)
181 {
182         struct inet_connection_sock *icsk = inet_csk(sk);
183         tcp_incr_quickack(sk);
184         icsk->icsk_ack.pingpong = 0;
185         icsk->icsk_ack.ato = TCP_ATO_MIN;
186 }
187
188 /* Send ACKs quickly, if "quick" count is not exhausted
189  * and the session is not interactive.
190  */
191
192 static inline int tcp_in_quickack_mode(const struct sock *sk)
193 {
194         const struct inet_connection_sock *icsk = inet_csk(sk);
195         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
196 }
197
198 /* Buffer size and advertised window tuning.
199  *
200  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
201  */
202
203 static void tcp_fixup_sndbuf(struct sock *sk)
204 {
205         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
206                      sizeof(struct sk_buff);
207
208         if (sk->sk_sndbuf < 3 * sndmem)
209                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
210 }
211
212 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
213  *
214  * All tcp_full_space() is split to two parts: "network" buffer, allocated
215  * forward and advertised in receiver window (tp->rcv_wnd) and
216  * "application buffer", required to isolate scheduling/application
217  * latencies from network.
218  * window_clamp is maximal advertised window. It can be less than
219  * tcp_full_space(), in this case tcp_full_space() - window_clamp
220  * is reserved for "application" buffer. The less window_clamp is
221  * the smoother our behaviour from viewpoint of network, but the lower
222  * throughput and the higher sensitivity of the connection to losses. 8)
223  *
224  * rcv_ssthresh is more strict window_clamp used at "slow start"
225  * phase to predict further behaviour of this connection.
226  * It is used for two goals:
227  * - to enforce header prediction at sender, even when application
228  *   requires some significant "application buffer". It is check #1.
229  * - to prevent pruning of receive queue because of misprediction
230  *   of receiver window. Check #2.
231  *
232  * The scheme does not work when sender sends good segments opening
233  * window and then starts to feed us spaghetti. But it should work
234  * in common situations. Otherwise, we have to rely on queue collapsing.
235  */
236
237 /* Slow part of check#2. */
238 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
239                              const struct sk_buff *skb)
240 {
241         /* Optimize this! */
242         int truesize = tcp_win_from_space(skb->truesize)/2;
243         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
244
245         while (tp->rcv_ssthresh <= window) {
246                 if (truesize <= skb->len)
247                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
248
249                 truesize >>= 1;
250                 window >>= 1;
251         }
252         return 0;
253 }
254
255 static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
256                             struct sk_buff *skb)
257 {
258         /* Check #1 */
259         if (tp->rcv_ssthresh < tp->window_clamp &&
260             (int)tp->rcv_ssthresh < tcp_space(sk) &&
261             !tcp_memory_pressure) {
262                 int incr;
263
264                 /* Check #2. Increase window, if skb with such overhead
265                  * will fit to rcvbuf in future.
266                  */
267                 if (tcp_win_from_space(skb->truesize) <= skb->len)
268                         incr = 2*tp->advmss;
269                 else
270                         incr = __tcp_grow_window(sk, tp, skb);
271
272                 if (incr) {
273                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
274                         inet_csk(sk)->icsk_ack.quick |= 1;
275                 }
276         }
277 }
278
279 /* 3. Tuning rcvbuf, when connection enters established state. */
280
281 static void tcp_fixup_rcvbuf(struct sock *sk)
282 {
283         struct tcp_sock *tp = tcp_sk(sk);
284         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
285
286         /* Try to select rcvbuf so that 4 mss-sized segments
287          * will fit to window and corresponding skbs will fit to our rcvbuf.
288          * (was 3; 4 is minimum to allow fast retransmit to work.)
289          */
290         while (tcp_win_from_space(rcvmem) < tp->advmss)
291                 rcvmem += 128;
292         if (sk->sk_rcvbuf < 4 * rcvmem)
293                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
294 }
295
296 /* 4. Try to fixup all. It is made immediately after connection enters
297  *    established state.
298  */
299 static void tcp_init_buffer_space(struct sock *sk)
300 {
301         struct tcp_sock *tp = tcp_sk(sk);
302         int maxwin;
303
304         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
305                 tcp_fixup_rcvbuf(sk);
306         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
307                 tcp_fixup_sndbuf(sk);
308
309         tp->rcvq_space.space = tp->rcv_wnd;
310
311         maxwin = tcp_full_space(sk);
312
313         if (tp->window_clamp >= maxwin) {
314                 tp->window_clamp = maxwin;
315
316                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
317                         tp->window_clamp = max(maxwin -
318                                                (maxwin >> sysctl_tcp_app_win),
319                                                4 * tp->advmss);
320         }
321
322         /* Force reservation of one segment. */
323         if (sysctl_tcp_app_win &&
324             tp->window_clamp > 2 * tp->advmss &&
325             tp->window_clamp + tp->advmss > maxwin)
326                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
327
328         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
329         tp->snd_cwnd_stamp = tcp_time_stamp;
330 }
331
332 /* 5. Recalculate window clamp after socket hit its memory bounds. */
333 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
334 {
335         struct inet_connection_sock *icsk = inet_csk(sk);
336
337         icsk->icsk_ack.quick = 0;
338
339         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
340             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
341             !tcp_memory_pressure &&
342             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
343                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
344                                     sysctl_tcp_rmem[2]);
345         }
346         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
347                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
348 }
349
350
351 /* Initialize RCV_MSS value.
352  * RCV_MSS is an our guess about MSS used by the peer.
353  * We haven't any direct information about the MSS.
354  * It's better to underestimate the RCV_MSS rather than overestimate.
355  * Overestimations make us ACKing less frequently than needed.
356  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
357  */
358 void tcp_initialize_rcv_mss(struct sock *sk)
359 {
360         struct tcp_sock *tp = tcp_sk(sk);
361         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
362
363         hint = min(hint, tp->rcv_wnd/2);
364         hint = min(hint, TCP_MIN_RCVMSS);
365         hint = max(hint, TCP_MIN_MSS);
366
367         inet_csk(sk)->icsk_ack.rcv_mss = hint;
368 }
369
370 /* Receiver "autotuning" code.
371  *
372  * The algorithm for RTT estimation w/o timestamps is based on
373  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
374  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
375  *
376  * More detail on this code can be found at
377  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
378  * though this reference is out of date.  A new paper
379  * is pending.
380  */
381 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
382 {
383         u32 new_sample = tp->rcv_rtt_est.rtt;
384         long m = sample;
385
386         if (m == 0)
387                 m = 1;
388
389         if (new_sample != 0) {
390                 /* If we sample in larger samples in the non-timestamp
391                  * case, we could grossly overestimate the RTT especially
392                  * with chatty applications or bulk transfer apps which
393                  * are stalled on filesystem I/O.
394                  *
395                  * Also, since we are only going for a minimum in the
396                  * non-timestamp case, we do not smooth things out
397                  * else with timestamps disabled convergence takes too
398                  * long.
399                  */
400                 if (!win_dep) {
401                         m -= (new_sample >> 3);
402                         new_sample += m;
403                 } else if (m < new_sample)
404                         new_sample = m << 3;
405         } else {
406                 /* No previous measure. */
407                 new_sample = m << 3;
408         }
409
410         if (tp->rcv_rtt_est.rtt != new_sample)
411                 tp->rcv_rtt_est.rtt = new_sample;
412 }
413
414 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
415 {
416         if (tp->rcv_rtt_est.time == 0)
417                 goto new_measure;
418         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
419                 return;
420         tcp_rcv_rtt_update(tp,
421                            jiffies - tp->rcv_rtt_est.time,
422                            1);
423
424 new_measure:
425         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
426         tp->rcv_rtt_est.time = tcp_time_stamp;
427 }
428
429 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
430 {
431         struct tcp_sock *tp = tcp_sk(sk);
432         if (tp->rx_opt.rcv_tsecr &&
433             (TCP_SKB_CB(skb)->end_seq -
434              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
435                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
436 }
437
438 /*
439  * This function should be called every time data is copied to user space.
440  * It calculates the appropriate TCP receive buffer space.
441  */
442 void tcp_rcv_space_adjust(struct sock *sk)
443 {
444         struct tcp_sock *tp = tcp_sk(sk);
445         int time;
446         int space;
447
448         if (tp->rcvq_space.time == 0)
449                 goto new_measure;
450
451         time = tcp_time_stamp - tp->rcvq_space.time;
452         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
453             tp->rcv_rtt_est.rtt == 0)
454                 return;
455
456         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
457
458         space = max(tp->rcvq_space.space, space);
459
460         if (tp->rcvq_space.space != space) {
461                 int rcvmem;
462
463                 tp->rcvq_space.space = space;
464
465                 if (sysctl_tcp_moderate_rcvbuf &&
466                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
467                         int new_clamp = space;
468
469                         /* Receive space grows, normalize in order to
470                          * take into account packet headers and sk_buff
471                          * structure overhead.
472                          */
473                         space /= tp->advmss;
474                         if (!space)
475                                 space = 1;
476                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
477                                   16 + sizeof(struct sk_buff));
478                         while (tcp_win_from_space(rcvmem) < tp->advmss)
479                                 rcvmem += 128;
480                         space *= rcvmem;
481                         space = min(space, sysctl_tcp_rmem[2]);
482                         if (space > sk->sk_rcvbuf) {
483                                 sk->sk_rcvbuf = space;
484
485                                 /* Make the window clamp follow along.  */
486                                 tp->window_clamp = new_clamp;
487                         }
488                 }
489         }
490
491 new_measure:
492         tp->rcvq_space.seq = tp->copied_seq;
493         tp->rcvq_space.time = tcp_time_stamp;
494 }
495
496 /* There is something which you must keep in mind when you analyze the
497  * behavior of the tp->ato delayed ack timeout interval.  When a
498  * connection starts up, we want to ack as quickly as possible.  The
499  * problem is that "good" TCP's do slow start at the beginning of data
500  * transmission.  The means that until we send the first few ACK's the
501  * sender will sit on his end and only queue most of his data, because
502  * he can only send snd_cwnd unacked packets at any given time.  For
503  * each ACK we send, he increments snd_cwnd and transmits more of his
504  * queue.  -DaveM
505  */
506 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
507 {
508         struct inet_connection_sock *icsk = inet_csk(sk);
509         u32 now;
510
511         inet_csk_schedule_ack(sk);
512
513         tcp_measure_rcv_mss(sk, skb);
514
515         tcp_rcv_rtt_measure(tp);
516
517         now = tcp_time_stamp;
518
519         if (!icsk->icsk_ack.ato) {
520                 /* The _first_ data packet received, initialize
521                  * delayed ACK engine.
522                  */
523                 tcp_incr_quickack(sk);
524                 icsk->icsk_ack.ato = TCP_ATO_MIN;
525         } else {
526                 int m = now - icsk->icsk_ack.lrcvtime;
527
528                 if (m <= TCP_ATO_MIN/2) {
529                         /* The fastest case is the first. */
530                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
531                 } else if (m < icsk->icsk_ack.ato) {
532                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
533                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
534                                 icsk->icsk_ack.ato = icsk->icsk_rto;
535                 } else if (m > icsk->icsk_rto) {
536                         /* Too long gap. Apparently sender failed to
537                          * restart window, so that we send ACKs quickly.
538                          */
539                         tcp_incr_quickack(sk);
540                         sk_stream_mem_reclaim(sk);
541                 }
542         }
543         icsk->icsk_ack.lrcvtime = now;
544
545         TCP_ECN_check_ce(tp, skb);
546
547         if (skb->len >= 128)
548                 tcp_grow_window(sk, tp, skb);
549 }
550
551 /* Called to compute a smoothed rtt estimate. The data fed to this
552  * routine either comes from timestamps, or from segments that were
553  * known _not_ to have been retransmitted [see Karn/Partridge
554  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
555  * piece by Van Jacobson.
556  * NOTE: the next three routines used to be one big routine.
557  * To save cycles in the RFC 1323 implementation it was better to break
558  * it up into three procedures. -- erics
559  */
560 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
561 {
562         struct tcp_sock *tp = tcp_sk(sk);
563         long m = mrtt; /* RTT */
564
565         /*      The following amusing code comes from Jacobson's
566          *      article in SIGCOMM '88.  Note that rtt and mdev
567          *      are scaled versions of rtt and mean deviation.
568          *      This is designed to be as fast as possible
569          *      m stands for "measurement".
570          *
571          *      On a 1990 paper the rto value is changed to:
572          *      RTO = rtt + 4 * mdev
573          *
574          * Funny. This algorithm seems to be very broken.
575          * These formulae increase RTO, when it should be decreased, increase
576          * too slowly, when it should be increased quickly, decrease too quickly
577          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
578          * does not matter how to _calculate_ it. Seems, it was trap
579          * that VJ failed to avoid. 8)
580          */
581         if (m == 0)
582                 m = 1;
583         if (tp->srtt != 0) {
584                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
585                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
586                 if (m < 0) {
587                         m = -m;         /* m is now abs(error) */
588                         m -= (tp->mdev >> 2);   /* similar update on mdev */
589                         /* This is similar to one of Eifel findings.
590                          * Eifel blocks mdev updates when rtt decreases.
591                          * This solution is a bit different: we use finer gain
592                          * for mdev in this case (alpha*beta).
593                          * Like Eifel it also prevents growth of rto,
594                          * but also it limits too fast rto decreases,
595                          * happening in pure Eifel.
596                          */
597                         if (m > 0)
598                                 m >>= 3;
599                 } else {
600                         m -= (tp->mdev >> 2);   /* similar update on mdev */
601                 }
602                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
603                 if (tp->mdev > tp->mdev_max) {
604                         tp->mdev_max = tp->mdev;
605                         if (tp->mdev_max > tp->rttvar)
606                                 tp->rttvar = tp->mdev_max;
607                 }
608                 if (after(tp->snd_una, tp->rtt_seq)) {
609                         if (tp->mdev_max < tp->rttvar)
610                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
611                         tp->rtt_seq = tp->snd_nxt;
612                         tp->mdev_max = TCP_RTO_MIN;
613                 }
614         } else {
615                 /* no previous measure. */
616                 tp->srtt = m<<3;        /* take the measured time to be rtt */
617                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
618                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
619                 tp->rtt_seq = tp->snd_nxt;
620         }
621 }
622
623 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
624  * routine referred to above.
625  */
626 static inline void tcp_set_rto(struct sock *sk)
627 {
628         const struct tcp_sock *tp = tcp_sk(sk);
629         /* Old crap is replaced with new one. 8)
630          *
631          * More seriously:
632          * 1. If rtt variance happened to be less 50msec, it is hallucination.
633          *    It cannot be less due to utterly erratic ACK generation made
634          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
635          *    to do with delayed acks, because at cwnd>2 true delack timeout
636          *    is invisible. Actually, Linux-2.4 also generates erratic
637          *    ACKs in some circumstances.
638          */
639         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
640
641         /* 2. Fixups made earlier cannot be right.
642          *    If we do not estimate RTO correctly without them,
643          *    all the algo is pure shit and should be replaced
644          *    with correct one. It is exactly, which we pretend to do.
645          */
646 }
647
648 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
649  * guarantees that rto is higher.
650  */
651 static inline void tcp_bound_rto(struct sock *sk)
652 {
653         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
654                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
655 }
656
657 /* Save metrics learned by this TCP session.
658    This function is called only, when TCP finishes successfully
659    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
660  */
661 void tcp_update_metrics(struct sock *sk)
662 {
663         struct tcp_sock *tp = tcp_sk(sk);
664         struct dst_entry *dst = __sk_dst_get(sk);
665
666         if (sysctl_tcp_nometrics_save)
667                 return;
668
669         dst_confirm(dst);
670
671         if (dst && (dst->flags&DST_HOST)) {
672                 const struct inet_connection_sock *icsk = inet_csk(sk);
673                 int m;
674
675                 if (icsk->icsk_backoff || !tp->srtt) {
676                         /* This session failed to estimate rtt. Why?
677                          * Probably, no packets returned in time.
678                          * Reset our results.
679                          */
680                         if (!(dst_metric_locked(dst, RTAX_RTT)))
681                                 dst->metrics[RTAX_RTT-1] = 0;
682                         return;
683                 }
684
685                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
686
687                 /* If newly calculated rtt larger than stored one,
688                  * store new one. Otherwise, use EWMA. Remember,
689                  * rtt overestimation is always better than underestimation.
690                  */
691                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
692                         if (m <= 0)
693                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
694                         else
695                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
696                 }
697
698                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
699                         if (m < 0)
700                                 m = -m;
701
702                         /* Scale deviation to rttvar fixed point */
703                         m >>= 1;
704                         if (m < tp->mdev)
705                                 m = tp->mdev;
706
707                         if (m >= dst_metric(dst, RTAX_RTTVAR))
708                                 dst->metrics[RTAX_RTTVAR-1] = m;
709                         else
710                                 dst->metrics[RTAX_RTTVAR-1] -=
711                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
712                 }
713
714                 if (tp->snd_ssthresh >= 0xFFFF) {
715                         /* Slow start still did not finish. */
716                         if (dst_metric(dst, RTAX_SSTHRESH) &&
717                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
718                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
719                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
720                         if (!dst_metric_locked(dst, RTAX_CWND) &&
721                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
722                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
723                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
724                            icsk->icsk_ca_state == TCP_CA_Open) {
725                         /* Cong. avoidance phase, cwnd is reliable. */
726                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
727                                 dst->metrics[RTAX_SSTHRESH-1] =
728                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
729                         if (!dst_metric_locked(dst, RTAX_CWND))
730                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
731                 } else {
732                         /* Else slow start did not finish, cwnd is non-sense,
733                            ssthresh may be also invalid.
734                          */
735                         if (!dst_metric_locked(dst, RTAX_CWND))
736                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
737                         if (dst->metrics[RTAX_SSTHRESH-1] &&
738                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
739                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
740                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
741                 }
742
743                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
744                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
745                             tp->reordering != sysctl_tcp_reordering)
746                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
747                 }
748         }
749 }
750
751 /* Numbers are taken from RFC2414.  */
752 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
753 {
754         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
755
756         if (!cwnd) {
757                 if (tp->mss_cache > 1460)
758                         cwnd = 2;
759                 else
760                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
761         }
762         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
763 }
764
765 /* Set slow start threshold and cwnd not falling to slow start */
766 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
767 {
768         struct tcp_sock *tp = tcp_sk(sk);
769         const struct inet_connection_sock *icsk = inet_csk(sk);
770
771         tp->prior_ssthresh = 0;
772         tp->bytes_acked = 0;
773         if (icsk->icsk_ca_state < TCP_CA_CWR) {
774                 tp->undo_marker = 0;
775                 if (set_ssthresh)
776                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
777                 tp->snd_cwnd = min(tp->snd_cwnd,
778                                    tcp_packets_in_flight(tp) + 1U);
779                 tp->snd_cwnd_cnt = 0;
780                 tp->high_seq = tp->snd_nxt;
781                 tp->snd_cwnd_stamp = tcp_time_stamp;
782                 TCP_ECN_queue_cwr(tp);
783
784                 tcp_set_ca_state(sk, TCP_CA_CWR);
785         }
786 }
787
788 /* Initialize metrics on socket. */
789
790 static void tcp_init_metrics(struct sock *sk)
791 {
792         struct tcp_sock *tp = tcp_sk(sk);
793         struct dst_entry *dst = __sk_dst_get(sk);
794
795         if (dst == NULL)
796                 goto reset;
797
798         dst_confirm(dst);
799
800         if (dst_metric_locked(dst, RTAX_CWND))
801                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
802         if (dst_metric(dst, RTAX_SSTHRESH)) {
803                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
804                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
805                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
806         }
807         if (dst_metric(dst, RTAX_REORDERING) &&
808             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
809                 tp->rx_opt.sack_ok &= ~2;
810                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
811         }
812
813         if (dst_metric(dst, RTAX_RTT) == 0)
814                 goto reset;
815
816         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
817                 goto reset;
818
819         /* Initial rtt is determined from SYN,SYN-ACK.
820          * The segment is small and rtt may appear much
821          * less than real one. Use per-dst memory
822          * to make it more realistic.
823          *
824          * A bit of theory. RTT is time passed after "normal" sized packet
825          * is sent until it is ACKed. In normal circumstances sending small
826          * packets force peer to delay ACKs and calculation is correct too.
827          * The algorithm is adaptive and, provided we follow specs, it
828          * NEVER underestimate RTT. BUT! If peer tries to make some clever
829          * tricks sort of "quick acks" for time long enough to decrease RTT
830          * to low value, and then abruptly stops to do it and starts to delay
831          * ACKs, wait for troubles.
832          */
833         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
834                 tp->srtt = dst_metric(dst, RTAX_RTT);
835                 tp->rtt_seq = tp->snd_nxt;
836         }
837         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
838                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
839                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
840         }
841         tcp_set_rto(sk);
842         tcp_bound_rto(sk);
843         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
844                 goto reset;
845         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
846         tp->snd_cwnd_stamp = tcp_time_stamp;
847         return;
848
849 reset:
850         /* Play conservative. If timestamps are not
851          * supported, TCP will fail to recalculate correct
852          * rtt, if initial rto is too small. FORGET ALL AND RESET!
853          */
854         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
855                 tp->srtt = 0;
856                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
857                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
858         }
859 }
860
861 static void tcp_update_reordering(struct sock *sk, const int metric,
862                                   const int ts)
863 {
864         struct tcp_sock *tp = tcp_sk(sk);
865         if (metric > tp->reordering) {
866                 tp->reordering = min(TCP_MAX_REORDERING, metric);
867
868                 /* This exciting event is worth to be remembered. 8) */
869                 if (ts)
870                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
871                 else if (IsReno(tp))
872                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
873                 else if (IsFack(tp))
874                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
875                 else
876                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
877 #if FASTRETRANS_DEBUG > 1
878                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
879                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
880                        tp->reordering,
881                        tp->fackets_out,
882                        tp->sacked_out,
883                        tp->undo_marker ? tp->undo_retrans : 0);
884 #endif
885                 /* Disable FACK yet. */
886                 tp->rx_opt.sack_ok &= ~2;
887         }
888 }
889
890 /* This procedure tags the retransmission queue when SACKs arrive.
891  *
892  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
893  * Packets in queue with these bits set are counted in variables
894  * sacked_out, retrans_out and lost_out, correspondingly.
895  *
896  * Valid combinations are:
897  * Tag  InFlight        Description
898  * 0    1               - orig segment is in flight.
899  * S    0               - nothing flies, orig reached receiver.
900  * L    0               - nothing flies, orig lost by net.
901  * R    2               - both orig and retransmit are in flight.
902  * L|R  1               - orig is lost, retransmit is in flight.
903  * S|R  1               - orig reached receiver, retrans is still in flight.
904  * (L|S|R is logically valid, it could occur when L|R is sacked,
905  *  but it is equivalent to plain S and code short-curcuits it to S.
906  *  L|S is logically invalid, it would mean -1 packet in flight 8))
907  *
908  * These 6 states form finite state machine, controlled by the following events:
909  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
910  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
911  * 3. Loss detection event of one of three flavors:
912  *      A. Scoreboard estimator decided the packet is lost.
913  *         A'. Reno "three dupacks" marks head of queue lost.
914  *         A''. Its FACK modfication, head until snd.fack is lost.
915  *      B. SACK arrives sacking data transmitted after never retransmitted
916  *         hole was sent out.
917  *      C. SACK arrives sacking SND.NXT at the moment, when the
918  *         segment was retransmitted.
919  * 4. D-SACK added new rule: D-SACK changes any tag to S.
920  *
921  * It is pleasant to note, that state diagram turns out to be commutative,
922  * so that we are allowed not to be bothered by order of our actions,
923  * when multiple events arrive simultaneously. (see the function below).
924  *
925  * Reordering detection.
926  * --------------------
927  * Reordering metric is maximal distance, which a packet can be displaced
928  * in packet stream. With SACKs we can estimate it:
929  *
930  * 1. SACK fills old hole and the corresponding segment was not
931  *    ever retransmitted -> reordering. Alas, we cannot use it
932  *    when segment was retransmitted.
933  * 2. The last flaw is solved with D-SACK. D-SACK arrives
934  *    for retransmitted and already SACKed segment -> reordering..
935  * Both of these heuristics are not used in Loss state, when we cannot
936  * account for retransmits accurately.
937  */
938 static int
939 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
940 {
941         const struct inet_connection_sock *icsk = inet_csk(sk);
942         struct tcp_sock *tp = tcp_sk(sk);
943         unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
944         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
945         struct sk_buff *cached_skb;
946         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
947         int reord = tp->packets_out;
948         int prior_fackets;
949         u32 lost_retrans = 0;
950         int flag = 0;
951         int dup_sack = 0;
952         int cached_fack_count;
953         int i;
954         int first_sack_index;
955
956         if (!tp->sacked_out)
957                 tp->fackets_out = 0;
958         prior_fackets = tp->fackets_out;
959
960         /* Check for D-SACK. */
961         if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
962                 dup_sack = 1;
963                 tp->rx_opt.sack_ok |= 4;
964                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
965         } else if (num_sacks > 1 &&
966                         !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
967                         !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
968                 dup_sack = 1;
969                 tp->rx_opt.sack_ok |= 4;
970                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
971         }
972
973         /* D-SACK for already forgotten data...
974          * Do dumb counting. */
975         if (dup_sack &&
976                         !after(ntohl(sp[0].end_seq), prior_snd_una) &&
977                         after(ntohl(sp[0].end_seq), tp->undo_marker))
978                 tp->undo_retrans--;
979
980         /* Eliminate too old ACKs, but take into
981          * account more or less fresh ones, they can
982          * contain valid SACK info.
983          */
984         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
985                 return 0;
986
987         /* SACK fastpath:
988          * if the only SACK change is the increase of the end_seq of
989          * the first block then only apply that SACK block
990          * and use retrans queue hinting otherwise slowpath */
991         flag = 1;
992         for (i = 0; i < num_sacks; i++) {
993                 __be32 start_seq = sp[i].start_seq;
994                 __be32 end_seq = sp[i].end_seq;
995
996                 if (i == 0) {
997                         if (tp->recv_sack_cache[i].start_seq != start_seq)
998                                 flag = 0;
999                 } else {
1000                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1001                             (tp->recv_sack_cache[i].end_seq != end_seq))
1002                                 flag = 0;
1003                 }
1004                 tp->recv_sack_cache[i].start_seq = start_seq;
1005                 tp->recv_sack_cache[i].end_seq = end_seq;
1006         }
1007         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1008         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1009                 tp->recv_sack_cache[i].start_seq = 0;
1010                 tp->recv_sack_cache[i].end_seq = 0;
1011         }
1012
1013         first_sack_index = 0;
1014         if (flag)
1015                 num_sacks = 1;
1016         else {
1017                 int j;
1018                 tp->fastpath_skb_hint = NULL;
1019
1020                 /* order SACK blocks to allow in order walk of the retrans queue */
1021                 for (i = num_sacks-1; i > 0; i--) {
1022                         for (j = 0; j < i; j++){
1023                                 if (after(ntohl(sp[j].start_seq),
1024                                           ntohl(sp[j+1].start_seq))){
1025                                         struct tcp_sack_block_wire tmp;
1026
1027                                         tmp = sp[j];
1028                                         sp[j] = sp[j+1];
1029                                         sp[j+1] = tmp;
1030
1031                                         /* Track where the first SACK block goes to */
1032                                         if (j == first_sack_index)
1033                                                 first_sack_index = j+1;
1034                                 }
1035
1036                         }
1037                 }
1038         }
1039
1040         /* clear flag as used for different purpose in following code */
1041         flag = 0;
1042
1043         /* Use SACK fastpath hint if valid */
1044         cached_skb = tp->fastpath_skb_hint;
1045         cached_fack_count = tp->fastpath_cnt_hint;
1046         if (!cached_skb) {
1047                 cached_skb = tcp_write_queue_head(sk);
1048                 cached_fack_count = 0;
1049         }
1050
1051         for (i=0; i<num_sacks; i++, sp++) {
1052                 struct sk_buff *skb;
1053                 __u32 start_seq = ntohl(sp->start_seq);
1054                 __u32 end_seq = ntohl(sp->end_seq);
1055                 int fack_count;
1056
1057                 skb = cached_skb;
1058                 fack_count = cached_fack_count;
1059
1060                 /* Event "B" in the comment above. */
1061                 if (after(end_seq, tp->high_seq))
1062                         flag |= FLAG_DATA_LOST;
1063
1064                 tcp_for_write_queue_from(skb, sk) {
1065                         int in_sack, pcount;
1066                         u8 sacked;
1067
1068                         if (skb == tcp_send_head(sk))
1069                                 break;
1070
1071                         cached_skb = skb;
1072                         cached_fack_count = fack_count;
1073                         if (i == first_sack_index) {
1074                                 tp->fastpath_skb_hint = skb;
1075                                 tp->fastpath_cnt_hint = fack_count;
1076                         }
1077
1078                         /* The retransmission queue is always in order, so
1079                          * we can short-circuit the walk early.
1080                          */
1081                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1082                                 break;
1083
1084                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1085                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1086
1087                         pcount = tcp_skb_pcount(skb);
1088
1089                         if (pcount > 1 && !in_sack &&
1090                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1091                                 unsigned int pkt_len;
1092
1093                                 in_sack = !after(start_seq,
1094                                                  TCP_SKB_CB(skb)->seq);
1095
1096                                 if (!in_sack)
1097                                         pkt_len = (start_seq -
1098                                                    TCP_SKB_CB(skb)->seq);
1099                                 else
1100                                         pkt_len = (end_seq -
1101                                                    TCP_SKB_CB(skb)->seq);
1102                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1103                                         break;
1104                                 pcount = tcp_skb_pcount(skb);
1105                         }
1106
1107                         fack_count += pcount;
1108
1109                         sacked = TCP_SKB_CB(skb)->sacked;
1110
1111                         /* Account D-SACK for retransmitted packet. */
1112                         if ((dup_sack && in_sack) &&
1113                             (sacked & TCPCB_RETRANS) &&
1114                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1115                                 tp->undo_retrans--;
1116
1117                         /* The frame is ACKed. */
1118                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1119                                 if (sacked&TCPCB_RETRANS) {
1120                                         if ((dup_sack && in_sack) &&
1121                                             (sacked&TCPCB_SACKED_ACKED))
1122                                                 reord = min(fack_count, reord);
1123                                 } else {
1124                                         /* If it was in a hole, we detected reordering. */
1125                                         if (fack_count < prior_fackets &&
1126                                             !(sacked&TCPCB_SACKED_ACKED))
1127                                                 reord = min(fack_count, reord);
1128                                 }
1129
1130                                 /* Nothing to do; acked frame is about to be dropped. */
1131                                 continue;
1132                         }
1133
1134                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1135                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1136                             (!lost_retrans || after(end_seq, lost_retrans)))
1137                                 lost_retrans = end_seq;
1138
1139                         if (!in_sack)
1140                                 continue;
1141
1142                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1143                                 if (sacked & TCPCB_SACKED_RETRANS) {
1144                                         /* If the segment is not tagged as lost,
1145                                          * we do not clear RETRANS, believing
1146                                          * that retransmission is still in flight.
1147                                          */
1148                                         if (sacked & TCPCB_LOST) {
1149                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1150                                                 tp->lost_out -= tcp_skb_pcount(skb);
1151                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1152
1153                                                 /* clear lost hint */
1154                                                 tp->retransmit_skb_hint = NULL;
1155                                         }
1156                                 } else {
1157                                         /* New sack for not retransmitted frame,
1158                                          * which was in hole. It is reordering.
1159                                          */
1160                                         if (!(sacked & TCPCB_RETRANS) &&
1161                                             fack_count < prior_fackets)
1162                                                 reord = min(fack_count, reord);
1163
1164                                         if (sacked & TCPCB_LOST) {
1165                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1166                                                 tp->lost_out -= tcp_skb_pcount(skb);
1167
1168                                                 /* clear lost hint */
1169                                                 tp->retransmit_skb_hint = NULL;
1170                                         }
1171                                         /* SACK enhanced F-RTO detection.
1172                                          * Set flag if and only if non-rexmitted
1173                                          * segments below frto_highmark are
1174                                          * SACKed (RFC4138; Appendix B).
1175                                          * Clearing correct due to in-order walk
1176                                          */
1177                                         if (after(end_seq, tp->frto_highmark)) {
1178                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1179                                         } else {
1180                                                 if (!(sacked & TCPCB_RETRANS))
1181                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1182                                         }
1183                                 }
1184
1185                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1186                                 flag |= FLAG_DATA_SACKED;
1187                                 tp->sacked_out += tcp_skb_pcount(skb);
1188
1189                                 if (fack_count > tp->fackets_out)
1190                                         tp->fackets_out = fack_count;
1191                         } else {
1192                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1193                                         reord = min(fack_count, reord);
1194                         }
1195
1196                         /* D-SACK. We can detect redundant retransmission
1197                          * in S|R and plain R frames and clear it.
1198                          * undo_retrans is decreased above, L|R frames
1199                          * are accounted above as well.
1200                          */
1201                         if (dup_sack &&
1202                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1203                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1204                                 tp->retrans_out -= tcp_skb_pcount(skb);
1205                                 tp->retransmit_skb_hint = NULL;
1206                         }
1207                 }
1208         }
1209
1210         /* Check for lost retransmit. This superb idea is
1211          * borrowed from "ratehalving". Event "C".
1212          * Later note: FACK people cheated me again 8),
1213          * we have to account for reordering! Ugly,
1214          * but should help.
1215          */
1216         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1217                 struct sk_buff *skb;
1218
1219                 tcp_for_write_queue(skb, sk) {
1220                         if (skb == tcp_send_head(sk))
1221                                 break;
1222                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1223                                 break;
1224                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1225                                 continue;
1226                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1227                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1228                             (IsFack(tp) ||
1229                              !before(lost_retrans,
1230                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1231                                      tp->mss_cache))) {
1232                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1233                                 tp->retrans_out -= tcp_skb_pcount(skb);
1234
1235                                 /* clear lost hint */
1236                                 tp->retransmit_skb_hint = NULL;
1237
1238                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1239                                         tp->lost_out += tcp_skb_pcount(skb);
1240                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1241                                         flag |= FLAG_DATA_SACKED;
1242                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1243                                 }
1244                         }
1245                 }
1246         }
1247
1248         tp->left_out = tp->sacked_out + tp->lost_out;
1249
1250         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1251             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1252                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1253
1254 #if FASTRETRANS_DEBUG > 0
1255         BUG_TRAP((int)tp->sacked_out >= 0);
1256         BUG_TRAP((int)tp->lost_out >= 0);
1257         BUG_TRAP((int)tp->retrans_out >= 0);
1258         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1259 #endif
1260         return flag;
1261 }
1262
1263 /* F-RTO can only be used if these conditions are satisfied:
1264  *  - there must be some unsent new data
1265  *  - the advertised window should allow sending it
1266  *  - TCP has never retransmitted anything other than head (SACK enhanced
1267  *    variant from Appendix B of RFC4138 is more robust here)
1268  */
1269 int tcp_use_frto(struct sock *sk)
1270 {
1271         const struct tcp_sock *tp = tcp_sk(sk);
1272         struct sk_buff *skb;
1273
1274         if (!sysctl_tcp_frto || !tcp_send_head(sk) ||
1275                 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
1276                       tp->snd_una + tp->snd_wnd))
1277                 return 0;
1278
1279         if (IsSackFrto())
1280                 return 1;
1281
1282         /* Avoid expensive walking of rexmit queue if possible */
1283         if (tp->retrans_out > 1)
1284                 return 0;
1285
1286         skb = tcp_write_queue_head(sk);
1287         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1288         tcp_for_write_queue_from(skb, sk) {
1289                 if (skb == tcp_send_head(sk))
1290                         break;
1291                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1292                         return 0;
1293                 /* Short-circuit when first non-SACKed skb has been checked */
1294                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1295                         break;
1296         }
1297         return 1;
1298 }
1299
1300 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1301  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1302  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1303  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1304  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1305  * bits are handled if the Loss state is really to be entered (in
1306  * tcp_enter_frto_loss).
1307  *
1308  * Do like tcp_enter_loss() would; when RTO expires the second time it
1309  * does:
1310  *  "Reduce ssthresh if it has not yet been made inside this window."
1311  */
1312 void tcp_enter_frto(struct sock *sk)
1313 {
1314         const struct inet_connection_sock *icsk = inet_csk(sk);
1315         struct tcp_sock *tp = tcp_sk(sk);
1316         struct sk_buff *skb;
1317
1318         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1319             tp->snd_una == tp->high_seq ||
1320             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1321              !icsk->icsk_retransmits)) {
1322                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1323                 /* Our state is too optimistic in ssthresh() call because cwnd
1324                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1325                  * recovery has not yet completed. Pattern would be this: RTO,
1326                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1327                  * up here twice).
1328                  * RFC4138 should be more specific on what to do, even though
1329                  * RTO is quite unlikely to occur after the first Cumulative ACK
1330                  * due to back-off and complexity of triggering events ...
1331                  */
1332                 if (tp->frto_counter) {
1333                         u32 stored_cwnd;
1334                         stored_cwnd = tp->snd_cwnd;
1335                         tp->snd_cwnd = 2;
1336                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1337                         tp->snd_cwnd = stored_cwnd;
1338                 } else {
1339                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1340                 }
1341                 /* ... in theory, cong.control module could do "any tricks" in
1342                  * ssthresh(), which means that ca_state, lost bits and lost_out
1343                  * counter would have to be faked before the call occurs. We
1344                  * consider that too expensive, unlikely and hacky, so modules
1345                  * using these in ssthresh() must deal these incompatibility
1346                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1347                  */
1348                 tcp_ca_event(sk, CA_EVENT_FRTO);
1349         }
1350
1351         tp->undo_marker = tp->snd_una;
1352         tp->undo_retrans = 0;
1353
1354         skb = tcp_write_queue_head(sk);
1355         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1356                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1357                 tp->retrans_out -= tcp_skb_pcount(skb);
1358         }
1359         tcp_sync_left_out(tp);
1360
1361         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1362          * The last condition is necessary at least in tp->frto_counter case.
1363          */
1364         if (IsSackFrto() && (tp->frto_counter ||
1365             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1366             after(tp->high_seq, tp->snd_una)) {
1367                 tp->frto_highmark = tp->high_seq;
1368         } else {
1369                 tp->frto_highmark = tp->snd_nxt;
1370         }
1371         tcp_set_ca_state(sk, TCP_CA_Disorder);
1372         tp->high_seq = tp->snd_nxt;
1373         tp->frto_counter = 1;
1374 }
1375
1376 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1377  * which indicates that we should follow the traditional RTO recovery,
1378  * i.e. mark everything lost and do go-back-N retransmission.
1379  */
1380 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1381 {
1382         struct tcp_sock *tp = tcp_sk(sk);
1383         struct sk_buff *skb;
1384         int cnt = 0;
1385
1386         tp->sacked_out = 0;
1387         tp->lost_out = 0;
1388         tp->fackets_out = 0;
1389         tp->retrans_out = 0;
1390
1391         tcp_for_write_queue(skb, sk) {
1392                 if (skb == tcp_send_head(sk))
1393                         break;
1394                 cnt += tcp_skb_pcount(skb);
1395                 /*
1396                  * Count the retransmission made on RTO correctly (only when
1397                  * waiting for the first ACK and did not get it)...
1398                  */
1399                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1400                         tp->retrans_out += tcp_skb_pcount(skb);
1401                         /* ...enter this if branch just for the first segment */
1402                         flag |= FLAG_DATA_ACKED;
1403                 } else {
1404                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1405                 }
1406                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1407
1408                         /* Do not mark those segments lost that were
1409                          * forward transmitted after RTO
1410                          */
1411                         if (!after(TCP_SKB_CB(skb)->end_seq,
1412                                    tp->frto_highmark)) {
1413                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1414                                 tp->lost_out += tcp_skb_pcount(skb);
1415                         }
1416                 } else {
1417                         tp->sacked_out += tcp_skb_pcount(skb);
1418                         tp->fackets_out = cnt;
1419                 }
1420         }
1421         tcp_sync_left_out(tp);
1422
1423         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1424         tp->snd_cwnd_cnt = 0;
1425         tp->snd_cwnd_stamp = tcp_time_stamp;
1426         tp->undo_marker = 0;
1427         tp->frto_counter = 0;
1428
1429         tp->reordering = min_t(unsigned int, tp->reordering,
1430                                              sysctl_tcp_reordering);
1431         tcp_set_ca_state(sk, TCP_CA_Loss);
1432         tp->high_seq = tp->frto_highmark;
1433         TCP_ECN_queue_cwr(tp);
1434
1435         clear_all_retrans_hints(tp);
1436 }
1437
1438 void tcp_clear_retrans(struct tcp_sock *tp)
1439 {
1440         tp->left_out = 0;
1441         tp->retrans_out = 0;
1442
1443         tp->fackets_out = 0;
1444         tp->sacked_out = 0;
1445         tp->lost_out = 0;
1446
1447         tp->undo_marker = 0;
1448         tp->undo_retrans = 0;
1449 }
1450
1451 /* Enter Loss state. If "how" is not zero, forget all SACK information
1452  * and reset tags completely, otherwise preserve SACKs. If receiver
1453  * dropped its ofo queue, we will know this due to reneging detection.
1454  */
1455 void tcp_enter_loss(struct sock *sk, int how)
1456 {
1457         const struct inet_connection_sock *icsk = inet_csk(sk);
1458         struct tcp_sock *tp = tcp_sk(sk);
1459         struct sk_buff *skb;
1460         int cnt = 0;
1461
1462         /* Reduce ssthresh if it has not yet been made inside this window. */
1463         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1464             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1465                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1466                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1467                 tcp_ca_event(sk, CA_EVENT_LOSS);
1468         }
1469         tp->snd_cwnd       = 1;
1470         tp->snd_cwnd_cnt   = 0;
1471         tp->snd_cwnd_stamp = tcp_time_stamp;
1472
1473         tp->bytes_acked = 0;
1474         tcp_clear_retrans(tp);
1475
1476         /* Push undo marker, if it was plain RTO and nothing
1477          * was retransmitted. */
1478         if (!how)
1479                 tp->undo_marker = tp->snd_una;
1480
1481         tcp_for_write_queue(skb, sk) {
1482                 if (skb == tcp_send_head(sk))
1483                         break;
1484                 cnt += tcp_skb_pcount(skb);
1485                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1486                         tp->undo_marker = 0;
1487                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1488                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1489                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1490                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1491                         tp->lost_out += tcp_skb_pcount(skb);
1492                 } else {
1493                         tp->sacked_out += tcp_skb_pcount(skb);
1494                         tp->fackets_out = cnt;
1495                 }
1496         }
1497         tcp_sync_left_out(tp);
1498
1499         tp->reordering = min_t(unsigned int, tp->reordering,
1500                                              sysctl_tcp_reordering);
1501         tcp_set_ca_state(sk, TCP_CA_Loss);
1502         tp->high_seq = tp->snd_nxt;
1503         TCP_ECN_queue_cwr(tp);
1504
1505         clear_all_retrans_hints(tp);
1506 }
1507
1508 static int tcp_check_sack_reneging(struct sock *sk)
1509 {
1510         struct sk_buff *skb;
1511
1512         /* If ACK arrived pointing to a remembered SACK,
1513          * it means that our remembered SACKs do not reflect
1514          * real state of receiver i.e.
1515          * receiver _host_ is heavily congested (or buggy).
1516          * Do processing similar to RTO timeout.
1517          */
1518         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1519             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1520                 struct inet_connection_sock *icsk = inet_csk(sk);
1521                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1522
1523                 tcp_enter_loss(sk, 1);
1524                 icsk->icsk_retransmits++;
1525                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1526                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1527                                           icsk->icsk_rto, TCP_RTO_MAX);
1528                 return 1;
1529         }
1530         return 0;
1531 }
1532
1533 static inline int tcp_fackets_out(struct tcp_sock *tp)
1534 {
1535         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1536 }
1537
1538 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1539 {
1540         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1541 }
1542
1543 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1544 {
1545         return tp->packets_out &&
1546                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1547 }
1548
1549 /* Linux NewReno/SACK/FACK/ECN state machine.
1550  * --------------------------------------
1551  *
1552  * "Open"       Normal state, no dubious events, fast path.
1553  * "Disorder"   In all the respects it is "Open",
1554  *              but requires a bit more attention. It is entered when
1555  *              we see some SACKs or dupacks. It is split of "Open"
1556  *              mainly to move some processing from fast path to slow one.
1557  * "CWR"        CWND was reduced due to some Congestion Notification event.
1558  *              It can be ECN, ICMP source quench, local device congestion.
1559  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1560  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1561  *
1562  * tcp_fastretrans_alert() is entered:
1563  * - each incoming ACK, if state is not "Open"
1564  * - when arrived ACK is unusual, namely:
1565  *      * SACK
1566  *      * Duplicate ACK.
1567  *      * ECN ECE.
1568  *
1569  * Counting packets in flight is pretty simple.
1570  *
1571  *      in_flight = packets_out - left_out + retrans_out
1572  *
1573  *      packets_out is SND.NXT-SND.UNA counted in packets.
1574  *
1575  *      retrans_out is number of retransmitted segments.
1576  *
1577  *      left_out is number of segments left network, but not ACKed yet.
1578  *
1579  *              left_out = sacked_out + lost_out
1580  *
1581  *     sacked_out: Packets, which arrived to receiver out of order
1582  *                 and hence not ACKed. With SACKs this number is simply
1583  *                 amount of SACKed data. Even without SACKs
1584  *                 it is easy to give pretty reliable estimate of this number,
1585  *                 counting duplicate ACKs.
1586  *
1587  *       lost_out: Packets lost by network. TCP has no explicit
1588  *                 "loss notification" feedback from network (for now).
1589  *                 It means that this number can be only _guessed_.
1590  *                 Actually, it is the heuristics to predict lossage that
1591  *                 distinguishes different algorithms.
1592  *
1593  *      F.e. after RTO, when all the queue is considered as lost,
1594  *      lost_out = packets_out and in_flight = retrans_out.
1595  *
1596  *              Essentially, we have now two algorithms counting
1597  *              lost packets.
1598  *
1599  *              FACK: It is the simplest heuristics. As soon as we decided
1600  *              that something is lost, we decide that _all_ not SACKed
1601  *              packets until the most forward SACK are lost. I.e.
1602  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1603  *              It is absolutely correct estimate, if network does not reorder
1604  *              packets. And it loses any connection to reality when reordering
1605  *              takes place. We use FACK by default until reordering
1606  *              is suspected on the path to this destination.
1607  *
1608  *              NewReno: when Recovery is entered, we assume that one segment
1609  *              is lost (classic Reno). While we are in Recovery and
1610  *              a partial ACK arrives, we assume that one more packet
1611  *              is lost (NewReno). This heuristics are the same in NewReno
1612  *              and SACK.
1613  *
1614  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1615  *  deflation etc. CWND is real congestion window, never inflated, changes
1616  *  only according to classic VJ rules.
1617  *
1618  * Really tricky (and requiring careful tuning) part of algorithm
1619  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1620  * The first determines the moment _when_ we should reduce CWND and,
1621  * hence, slow down forward transmission. In fact, it determines the moment
1622  * when we decide that hole is caused by loss, rather than by a reorder.
1623  *
1624  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1625  * holes, caused by lost packets.
1626  *
1627  * And the most logically complicated part of algorithm is undo
1628  * heuristics. We detect false retransmits due to both too early
1629  * fast retransmit (reordering) and underestimated RTO, analyzing
1630  * timestamps and D-SACKs. When we detect that some segments were
1631  * retransmitted by mistake and CWND reduction was wrong, we undo
1632  * window reduction and abort recovery phase. This logic is hidden
1633  * inside several functions named tcp_try_undo_<something>.
1634  */
1635
1636 /* This function decides, when we should leave Disordered state
1637  * and enter Recovery phase, reducing congestion window.
1638  *
1639  * Main question: may we further continue forward transmission
1640  * with the same cwnd?
1641  */
1642 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1643 {
1644         __u32 packets_out;
1645
1646         /* Do not perform any recovery during FRTO algorithm */
1647         if (tp->frto_counter)
1648                 return 0;
1649
1650         /* Trick#1: The loss is proven. */
1651         if (tp->lost_out)
1652                 return 1;
1653
1654         /* Not-A-Trick#2 : Classic rule... */
1655         if (tcp_fackets_out(tp) > tp->reordering)
1656                 return 1;
1657
1658         /* Trick#3 : when we use RFC2988 timer restart, fast
1659          * retransmit can be triggered by timeout of queue head.
1660          */
1661         if (tcp_head_timedout(sk, tp))
1662                 return 1;
1663
1664         /* Trick#4: It is still not OK... But will it be useful to delay
1665          * recovery more?
1666          */
1667         packets_out = tp->packets_out;
1668         if (packets_out <= tp->reordering &&
1669             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1670             !tcp_may_send_now(sk, tp)) {
1671                 /* We have nothing to send. This connection is limited
1672                  * either by receiver window or by application.
1673                  */
1674                 return 1;
1675         }
1676
1677         return 0;
1678 }
1679
1680 /* If we receive more dupacks than we expected counting segments
1681  * in assumption of absent reordering, interpret this as reordering.
1682  * The only another reason could be bug in receiver TCP.
1683  */
1684 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1685 {
1686         struct tcp_sock *tp = tcp_sk(sk);
1687         u32 holes;
1688
1689         holes = max(tp->lost_out, 1U);
1690         holes = min(holes, tp->packets_out);
1691
1692         if ((tp->sacked_out + holes) > tp->packets_out) {
1693                 tp->sacked_out = tp->packets_out - holes;
1694                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1695         }
1696 }
1697
1698 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1699
1700 static void tcp_add_reno_sack(struct sock *sk)
1701 {
1702         struct tcp_sock *tp = tcp_sk(sk);
1703         tp->sacked_out++;
1704         tcp_check_reno_reordering(sk, 0);
1705         tcp_sync_left_out(tp);
1706 }
1707
1708 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1709
1710 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1711 {
1712         if (acked > 0) {
1713                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1714                 if (acked-1 >= tp->sacked_out)
1715                         tp->sacked_out = 0;
1716                 else
1717                         tp->sacked_out -= acked-1;
1718         }
1719         tcp_check_reno_reordering(sk, acked);
1720         tcp_sync_left_out(tp);
1721 }
1722
1723 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1724 {
1725         tp->sacked_out = 0;
1726         tp->left_out = tp->lost_out;
1727 }
1728
1729 /* Mark head of queue up as lost. */
1730 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1731                                int packets, u32 high_seq)
1732 {
1733         struct sk_buff *skb;
1734         int cnt;
1735
1736         BUG_TRAP(packets <= tp->packets_out);
1737         if (tp->lost_skb_hint) {
1738                 skb = tp->lost_skb_hint;
1739                 cnt = tp->lost_cnt_hint;
1740         } else {
1741                 skb = tcp_write_queue_head(sk);
1742                 cnt = 0;
1743         }
1744
1745         tcp_for_write_queue_from(skb, sk) {
1746                 if (skb == tcp_send_head(sk))
1747                         break;
1748                 /* TODO: do this better */
1749                 /* this is not the most efficient way to do this... */
1750                 tp->lost_skb_hint = skb;
1751                 tp->lost_cnt_hint = cnt;
1752                 cnt += tcp_skb_pcount(skb);
1753                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1754                         break;
1755                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1756                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1757                         tp->lost_out += tcp_skb_pcount(skb);
1758
1759                         /* clear xmit_retransmit_queue hints
1760                          *  if this is beyond hint */
1761                         if (tp->retransmit_skb_hint != NULL &&
1762                             before(TCP_SKB_CB(skb)->seq,
1763                                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1764                                 tp->retransmit_skb_hint = NULL;
1765
1766                 }
1767         }
1768         tcp_sync_left_out(tp);
1769 }
1770
1771 /* Account newly detected lost packet(s) */
1772
1773 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1774 {
1775         if (IsFack(tp)) {
1776                 int lost = tp->fackets_out - tp->reordering;
1777                 if (lost <= 0)
1778                         lost = 1;
1779                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1780         } else {
1781                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1782         }
1783
1784         /* New heuristics: it is possible only after we switched
1785          * to restart timer each time when something is ACKed.
1786          * Hence, we can detect timed out packets during fast
1787          * retransmit without falling to slow start.
1788          */
1789         if (!IsReno(tp) && tcp_head_timedout(sk, tp)) {
1790                 struct sk_buff *skb;
1791
1792                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1793                         : tcp_write_queue_head(sk);
1794
1795                 tcp_for_write_queue_from(skb, sk) {
1796                         if (skb == tcp_send_head(sk))
1797                                 break;
1798                         if (!tcp_skb_timedout(sk, skb))
1799                                 break;
1800
1801                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1802                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1803                                 tp->lost_out += tcp_skb_pcount(skb);
1804
1805                                 /* clear xmit_retrans hint */
1806                                 if (tp->retransmit_skb_hint &&
1807                                     before(TCP_SKB_CB(skb)->seq,
1808                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1809
1810                                         tp->retransmit_skb_hint = NULL;
1811                         }
1812                 }
1813
1814                 tp->scoreboard_skb_hint = skb;
1815
1816                 tcp_sync_left_out(tp);
1817         }
1818 }
1819
1820 /* CWND moderation, preventing bursts due to too big ACKs
1821  * in dubious situations.
1822  */
1823 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1824 {
1825         tp->snd_cwnd = min(tp->snd_cwnd,
1826                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1827         tp->snd_cwnd_stamp = tcp_time_stamp;
1828 }
1829
1830 /* Lower bound on congestion window is slow start threshold
1831  * unless congestion avoidance choice decides to overide it.
1832  */
1833 static inline u32 tcp_cwnd_min(const struct sock *sk)
1834 {
1835         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1836
1837         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1838 }
1839
1840 /* Decrease cwnd each second ack. */
1841 static void tcp_cwnd_down(struct sock *sk)
1842 {
1843         struct tcp_sock *tp = tcp_sk(sk);
1844         int decr = tp->snd_cwnd_cnt + 1;
1845
1846         tp->snd_cwnd_cnt = decr&1;
1847         decr >>= 1;
1848
1849         if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1850                 tp->snd_cwnd -= decr;
1851
1852         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1853         tp->snd_cwnd_stamp = tcp_time_stamp;
1854 }
1855
1856 /* Nothing was retransmitted or returned timestamp is less
1857  * than timestamp of the first retransmission.
1858  */
1859 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1860 {
1861         return !tp->retrans_stamp ||
1862                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1863                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1864 }
1865
1866 /* Undo procedures. */
1867
1868 #if FASTRETRANS_DEBUG > 1
1869 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1870 {
1871         struct inet_sock *inet = inet_sk(sk);
1872         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1873                msg,
1874                NIPQUAD(inet->daddr), ntohs(inet->dport),
1875                tp->snd_cwnd, tp->left_out,
1876                tp->snd_ssthresh, tp->prior_ssthresh,
1877                tp->packets_out);
1878 }
1879 #else
1880 #define DBGUNDO(x...) do { } while (0)
1881 #endif
1882
1883 static void tcp_undo_cwr(struct sock *sk, const int undo)
1884 {
1885         struct tcp_sock *tp = tcp_sk(sk);
1886
1887         if (tp->prior_ssthresh) {
1888                 const struct inet_connection_sock *icsk = inet_csk(sk);
1889
1890                 if (icsk->icsk_ca_ops->undo_cwnd)
1891                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1892                 else
1893                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1894
1895                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1896                         tp->snd_ssthresh = tp->prior_ssthresh;
1897                         TCP_ECN_withdraw_cwr(tp);
1898                 }
1899         } else {
1900                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1901         }
1902         tcp_moderate_cwnd(tp);
1903         tp->snd_cwnd_stamp = tcp_time_stamp;
1904
1905         /* There is something screwy going on with the retrans hints after
1906            an undo */
1907         clear_all_retrans_hints(tp);
1908 }
1909
1910 static inline int tcp_may_undo(struct tcp_sock *tp)
1911 {
1912         return tp->undo_marker &&
1913                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1914 }
1915
1916 /* People celebrate: "We love our President!" */
1917 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1918 {
1919         if (tcp_may_undo(tp)) {
1920                 /* Happy end! We did not retransmit anything
1921                  * or our original transmission succeeded.
1922                  */
1923                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1924                 tcp_undo_cwr(sk, 1);
1925                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1926                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1927                 else
1928                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1929                 tp->undo_marker = 0;
1930         }
1931         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1932                 /* Hold old state until something *above* high_seq
1933                  * is ACKed. For Reno it is MUST to prevent false
1934                  * fast retransmits (RFC2582). SACK TCP is safe. */
1935                 tcp_moderate_cwnd(tp);
1936                 return 1;
1937         }
1938         tcp_set_ca_state(sk, TCP_CA_Open);
1939         return 0;
1940 }
1941
1942 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1943 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1944 {
1945         if (tp->undo_marker && !tp->undo_retrans) {
1946                 DBGUNDO(sk, tp, "D-SACK");
1947                 tcp_undo_cwr(sk, 1);
1948                 tp->undo_marker = 0;
1949                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1950         }
1951 }
1952
1953 /* Undo during fast recovery after partial ACK. */
1954
1955 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1956                                 int acked)
1957 {
1958         /* Partial ACK arrived. Force Hoe's retransmit. */
1959         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1960
1961         if (tcp_may_undo(tp)) {
1962                 /* Plain luck! Hole if filled with delayed
1963                  * packet, rather than with a retransmit.
1964                  */
1965                 if (tp->retrans_out == 0)
1966                         tp->retrans_stamp = 0;
1967
1968                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1969
1970                 DBGUNDO(sk, tp, "Hoe");
1971                 tcp_undo_cwr(sk, 0);
1972                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1973
1974                 /* So... Do not make Hoe's retransmit yet.
1975                  * If the first packet was delayed, the rest
1976                  * ones are most probably delayed as well.
1977                  */
1978                 failed = 0;
1979         }
1980         return failed;
1981 }
1982
1983 /* Undo during loss recovery after partial ACK. */
1984 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1985 {
1986         if (tcp_may_undo(tp)) {
1987                 struct sk_buff *skb;
1988                 tcp_for_write_queue(skb, sk) {
1989                         if (skb == tcp_send_head(sk))
1990                                 break;
1991                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1992                 }
1993
1994                 clear_all_retrans_hints(tp);
1995
1996                 DBGUNDO(sk, tp, "partial loss");
1997                 tp->lost_out = 0;
1998                 tp->left_out = tp->sacked_out;
1999                 tcp_undo_cwr(sk, 1);
2000                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2001                 inet_csk(sk)->icsk_retransmits = 0;
2002                 tp->undo_marker = 0;
2003                 if (!IsReno(tp))
2004                         tcp_set_ca_state(sk, TCP_CA_Open);
2005                 return 1;
2006         }
2007         return 0;
2008 }
2009
2010 static inline void tcp_complete_cwr(struct sock *sk)
2011 {
2012         struct tcp_sock *tp = tcp_sk(sk);
2013         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2014         tp->snd_cwnd_stamp = tcp_time_stamp;
2015         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2016 }
2017
2018 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
2019 {
2020         tp->left_out = tp->sacked_out;
2021
2022         if (tp->retrans_out == 0)
2023                 tp->retrans_stamp = 0;
2024
2025         if (flag&FLAG_ECE)
2026                 tcp_enter_cwr(sk, 1);
2027
2028         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2029                 int state = TCP_CA_Open;
2030
2031                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2032                         state = TCP_CA_Disorder;
2033
2034                 if (inet_csk(sk)->icsk_ca_state != state) {
2035                         tcp_set_ca_state(sk, state);
2036                         tp->high_seq = tp->snd_nxt;
2037                 }
2038                 tcp_moderate_cwnd(tp);
2039         } else {
2040                 tcp_cwnd_down(sk);
2041         }
2042 }
2043
2044 static void tcp_mtup_probe_failed(struct sock *sk)
2045 {
2046         struct inet_connection_sock *icsk = inet_csk(sk);
2047
2048         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2049         icsk->icsk_mtup.probe_size = 0;
2050 }
2051
2052 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2053 {
2054         struct tcp_sock *tp = tcp_sk(sk);
2055         struct inet_connection_sock *icsk = inet_csk(sk);
2056
2057         /* FIXME: breaks with very large cwnd */
2058         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2059         tp->snd_cwnd = tp->snd_cwnd *
2060                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2061                        icsk->icsk_mtup.probe_size;
2062         tp->snd_cwnd_cnt = 0;
2063         tp->snd_cwnd_stamp = tcp_time_stamp;
2064         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2065
2066         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2067         icsk->icsk_mtup.probe_size = 0;
2068         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2069 }
2070
2071
2072 /* Process an event, which can update packets-in-flight not trivially.
2073  * Main goal of this function is to calculate new estimate for left_out,
2074  * taking into account both packets sitting in receiver's buffer and
2075  * packets lost by network.
2076  *
2077  * Besides that it does CWND reduction, when packet loss is detected
2078  * and changes state of machine.
2079  *
2080  * It does _not_ decide what to send, it is made in function
2081  * tcp_xmit_retransmit_queue().
2082  */
2083 static void
2084 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
2085                       int prior_packets, int flag)
2086 {
2087         struct inet_connection_sock *icsk = inet_csk(sk);
2088         struct tcp_sock *tp = tcp_sk(sk);
2089         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
2090
2091         /* Some technical things:
2092          * 1. Reno does not count dupacks (sacked_out) automatically. */
2093         if (!tp->packets_out)
2094                 tp->sacked_out = 0;
2095         /* 2. SACK counts snd_fack in packets inaccurately. */
2096         if (tp->sacked_out == 0)
2097                 tp->fackets_out = 0;
2098
2099         /* Now state machine starts.
2100          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2101         if (flag&FLAG_ECE)
2102                 tp->prior_ssthresh = 0;
2103
2104         /* B. In all the states check for reneging SACKs. */
2105         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2106                 return;
2107
2108         /* C. Process data loss notification, provided it is valid. */
2109         if ((flag&FLAG_DATA_LOST) &&
2110             before(tp->snd_una, tp->high_seq) &&
2111             icsk->icsk_ca_state != TCP_CA_Open &&
2112             tp->fackets_out > tp->reordering) {
2113                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
2114                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2115         }
2116
2117         /* D. Synchronize left_out to current state. */
2118         tcp_sync_left_out(tp);
2119
2120         /* E. Check state exit conditions. State can be terminated
2121          *    when high_seq is ACKed. */
2122         if (icsk->icsk_ca_state == TCP_CA_Open) {
2123                 BUG_TRAP(tp->retrans_out == 0);
2124                 tp->retrans_stamp = 0;
2125         } else if (!before(tp->snd_una, tp->high_seq)) {
2126                 switch (icsk->icsk_ca_state) {
2127                 case TCP_CA_Loss:
2128                         icsk->icsk_retransmits = 0;
2129                         if (tcp_try_undo_recovery(sk, tp))
2130                                 return;
2131                         break;
2132
2133                 case TCP_CA_CWR:
2134                         /* CWR is to be held something *above* high_seq
2135                          * is ACKed for CWR bit to reach receiver. */
2136                         if (tp->snd_una != tp->high_seq) {
2137                                 tcp_complete_cwr(sk);
2138                                 tcp_set_ca_state(sk, TCP_CA_Open);
2139                         }
2140                         break;
2141
2142                 case TCP_CA_Disorder:
2143                         tcp_try_undo_dsack(sk, tp);
2144                         if (!tp->undo_marker ||
2145                             /* For SACK case do not Open to allow to undo
2146                              * catching for all duplicate ACKs. */
2147                             IsReno(tp) || tp->snd_una != tp->high_seq) {
2148                                 tp->undo_marker = 0;
2149                                 tcp_set_ca_state(sk, TCP_CA_Open);
2150                         }
2151                         break;
2152
2153                 case TCP_CA_Recovery:
2154                         if (IsReno(tp))
2155                                 tcp_reset_reno_sack(tp);
2156                         if (tcp_try_undo_recovery(sk, tp))
2157                                 return;
2158                         tcp_complete_cwr(sk);
2159                         break;
2160                 }
2161         }
2162
2163         /* F. Process state. */
2164         switch (icsk->icsk_ca_state) {
2165         case TCP_CA_Recovery:
2166                 if (prior_snd_una == tp->snd_una) {
2167                         if (IsReno(tp) && is_dupack)
2168                                 tcp_add_reno_sack(sk);
2169                 } else {
2170                         int acked = prior_packets - tp->packets_out;
2171                         if (IsReno(tp))
2172                                 tcp_remove_reno_sacks(sk, tp, acked);
2173                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
2174                 }
2175                 break;
2176         case TCP_CA_Loss:
2177                 if (flag&FLAG_DATA_ACKED)
2178                         icsk->icsk_retransmits = 0;
2179                 if (!tcp_try_undo_loss(sk, tp)) {
2180                         tcp_moderate_cwnd(tp);
2181                         tcp_xmit_retransmit_queue(sk);
2182                         return;
2183                 }
2184                 if (icsk->icsk_ca_state != TCP_CA_Open)
2185                         return;
2186                 /* Loss is undone; fall through to processing in Open state. */
2187         default:
2188                 if (IsReno(tp)) {
2189                         if (tp->snd_una != prior_snd_una)
2190                                 tcp_reset_reno_sack(tp);
2191                         if (is_dupack)
2192                                 tcp_add_reno_sack(sk);
2193                 }
2194
2195                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2196                         tcp_try_undo_dsack(sk, tp);
2197
2198                 if (!tcp_time_to_recover(sk, tp)) {
2199                         tcp_try_to_open(sk, tp, flag);
2200                         return;
2201                 }
2202
2203                 /* MTU probe failure: don't reduce cwnd */
2204                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2205                     icsk->icsk_mtup.probe_size &&
2206                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2207                         tcp_mtup_probe_failed(sk);
2208                         /* Restores the reduction we did in tcp_mtup_probe() */
2209                         tp->snd_cwnd++;
2210                         tcp_simple_retransmit(sk);
2211                         return;
2212                 }
2213
2214                 /* Otherwise enter Recovery state */
2215
2216                 if (IsReno(tp))
2217                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2218                 else
2219                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2220
2221                 tp->high_seq = tp->snd_nxt;
2222                 tp->prior_ssthresh = 0;
2223                 tp->undo_marker = tp->snd_una;
2224                 tp->undo_retrans = tp->retrans_out;
2225
2226                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2227                         if (!(flag&FLAG_ECE))
2228                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2229                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2230                         TCP_ECN_queue_cwr(tp);
2231                 }
2232
2233                 tp->bytes_acked = 0;
2234                 tp->snd_cwnd_cnt = 0;
2235                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2236         }
2237
2238         if (is_dupack || tcp_head_timedout(sk, tp))
2239                 tcp_update_scoreboard(sk, tp);
2240         tcp_cwnd_down(sk);
2241         tcp_xmit_retransmit_queue(sk);
2242 }
2243
2244 /* Read draft-ietf-tcplw-high-performance before mucking
2245  * with this code. (Supersedes RFC1323)
2246  */
2247 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2248 {
2249         /* RTTM Rule: A TSecr value received in a segment is used to
2250          * update the averaged RTT measurement only if the segment
2251          * acknowledges some new data, i.e., only if it advances the
2252          * left edge of the send window.
2253          *
2254          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2255          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2256          *
2257          * Changed: reset backoff as soon as we see the first valid sample.
2258          * If we do not, we get strongly overestimated rto. With timestamps
2259          * samples are accepted even from very old segments: f.e., when rtt=1
2260          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2261          * answer arrives rto becomes 120 seconds! If at least one of segments
2262          * in window is lost... Voila.                          --ANK (010210)
2263          */
2264         struct tcp_sock *tp = tcp_sk(sk);
2265         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2266         tcp_rtt_estimator(sk, seq_rtt);
2267         tcp_set_rto(sk);
2268         inet_csk(sk)->icsk_backoff = 0;
2269         tcp_bound_rto(sk);
2270 }
2271
2272 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2273 {
2274         /* We don't have a timestamp. Can only use
2275          * packets that are not retransmitted to determine
2276          * rtt estimates. Also, we must not reset the
2277          * backoff for rto until we get a non-retransmitted
2278          * packet. This allows us to deal with a situation
2279          * where the network delay has increased suddenly.
2280          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2281          */
2282
2283         if (flag & FLAG_RETRANS_DATA_ACKED)
2284                 return;
2285
2286         tcp_rtt_estimator(sk, seq_rtt);
2287         tcp_set_rto(sk);
2288         inet_csk(sk)->icsk_backoff = 0;
2289         tcp_bound_rto(sk);
2290 }
2291
2292 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2293                                       const s32 seq_rtt)
2294 {
2295         const struct tcp_sock *tp = tcp_sk(sk);
2296         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2297         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2298                 tcp_ack_saw_tstamp(sk, flag);
2299         else if (seq_rtt >= 0)
2300                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2301 }
2302
2303 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2304                            u32 in_flight, int good)
2305 {
2306         const struct inet_connection_sock *icsk = inet_csk(sk);
2307         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2308         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2309 }
2310
2311 /* Restart timer after forward progress on connection.
2312  * RFC2988 recommends to restart timer to now+rto.
2313  */
2314
2315 static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2316 {
2317         if (!tp->packets_out) {
2318                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2319         } else {
2320                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2321         }
2322 }
2323
2324 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2325                          __u32 now, __s32 *seq_rtt)
2326 {
2327         struct tcp_sock *tp = tcp_sk(sk);
2328         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2329         __u32 seq = tp->snd_una;
2330         __u32 packets_acked;
2331         int acked = 0;
2332
2333         /* If we get here, the whole TSO packet has not been
2334          * acked.
2335          */
2336         BUG_ON(!after(scb->end_seq, seq));
2337
2338         packets_acked = tcp_skb_pcount(skb);
2339         if (tcp_trim_head(sk, skb, seq - scb->seq))
2340                 return 0;
2341         packets_acked -= tcp_skb_pcount(skb);
2342
2343         if (packets_acked) {
2344                 __u8 sacked = scb->sacked;
2345
2346                 acked |= FLAG_DATA_ACKED;
2347                 if (sacked) {
2348                         if (sacked & TCPCB_RETRANS) {
2349                                 if (sacked & TCPCB_SACKED_RETRANS)
2350                                         tp->retrans_out -= packets_acked;
2351                                 acked |= FLAG_RETRANS_DATA_ACKED;
2352                                 *seq_rtt = -1;
2353                         } else if (*seq_rtt < 0)
2354                                 *seq_rtt = now - scb->when;
2355                         if (sacked & TCPCB_SACKED_ACKED)
2356                                 tp->sacked_out -= packets_acked;
2357                         if (sacked & TCPCB_LOST)
2358                                 tp->lost_out -= packets_acked;
2359                         if (sacked & TCPCB_URG) {
2360                                 if (tp->urg_mode &&
2361                                     !before(seq, tp->snd_up))
2362                                         tp->urg_mode = 0;
2363                         }
2364                 } else if (*seq_rtt < 0)
2365                         *seq_rtt = now - scb->when;
2366
2367                 if (tp->fackets_out) {
2368                         __u32 dval = min(tp->fackets_out, packets_acked);
2369                         tp->fackets_out -= dval;
2370                 }
2371                 tp->packets_out -= packets_acked;
2372
2373                 BUG_ON(tcp_skb_pcount(skb) == 0);
2374                 BUG_ON(!before(scb->seq, scb->end_seq));
2375         }
2376
2377         return acked;
2378 }
2379
2380 static u32 tcp_usrtt(struct timeval *tv)
2381 {
2382         struct timeval now;
2383
2384         do_gettimeofday(&now);
2385         return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec);
2386 }
2387
2388 /* Remove acknowledged frames from the retransmission queue. */
2389 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2390 {
2391         struct tcp_sock *tp = tcp_sk(sk);
2392         const struct inet_connection_sock *icsk = inet_csk(sk);
2393         struct sk_buff *skb;
2394         __u32 now = tcp_time_stamp;
2395         int acked = 0;
2396         __s32 seq_rtt = -1;
2397         u32 pkts_acked = 0;
2398         void (*rtt_sample)(struct sock *sk, u32 usrtt)
2399                 = icsk->icsk_ca_ops->rtt_sample;
2400         struct timeval tv = { .tv_sec = 0, .tv_usec = 0 };
2401
2402         while ((skb = tcp_write_queue_head(sk)) &&
2403                skb != tcp_send_head(sk)) {
2404                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2405                 __u8 sacked = scb->sacked;
2406
2407                 /* If our packet is before the ack sequence we can
2408                  * discard it as it's confirmed to have arrived at
2409                  * the other end.
2410                  */
2411                 if (after(scb->end_seq, tp->snd_una)) {
2412                         if (tcp_skb_pcount(skb) > 1 &&
2413                             after(tp->snd_una, scb->seq))
2414                                 acked |= tcp_tso_acked(sk, skb,
2415                                                        now, &seq_rtt);
2416                         break;
2417                 }
2418
2419                 /* Initial outgoing SYN's get put onto the write_queue
2420                  * just like anything else we transmit.  It is not
2421                  * true data, and if we misinform our callers that
2422                  * this ACK acks real data, we will erroneously exit
2423                  * connection startup slow start one packet too
2424                  * quickly.  This is severely frowned upon behavior.
2425                  */
2426                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2427                         acked |= FLAG_DATA_ACKED;
2428                         ++pkts_acked;
2429                 } else {
2430                         acked |= FLAG_SYN_ACKED;
2431                         tp->retrans_stamp = 0;
2432                 }
2433
2434                 /* MTU probing checks */
2435                 if (icsk->icsk_mtup.probe_size) {
2436                         if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2437                                 tcp_mtup_probe_success(sk, skb);
2438                         }
2439                 }
2440
2441                 if (sacked) {
2442                         if (sacked & TCPCB_RETRANS) {
2443                                 if (sacked & TCPCB_SACKED_RETRANS)
2444                                         tp->retrans_out -= tcp_skb_pcount(skb);
2445                                 acked |= FLAG_RETRANS_DATA_ACKED;
2446                                 seq_rtt = -1;
2447                         } else if (seq_rtt < 0) {
2448                                 seq_rtt = now - scb->when;
2449                                 skb_get_timestamp(skb, &tv);
2450                         }
2451                         if (sacked & TCPCB_SACKED_ACKED)
2452                                 tp->sacked_out -= tcp_skb_pcount(skb);
2453                         if (sacked & TCPCB_LOST)
2454                                 tp->lost_out -= tcp_skb_pcount(skb);
2455                         if (sacked & TCPCB_URG) {
2456                                 if (tp->urg_mode &&
2457                                     !before(scb->end_seq, tp->snd_up))
2458                                         tp->urg_mode = 0;
2459                         }
2460                 } else if (seq_rtt < 0) {
2461                         seq_rtt = now - scb->when;
2462                         skb_get_timestamp(skb, &tv);
2463                 }
2464                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2465                 tcp_packets_out_dec(tp, skb);
2466                 tcp_unlink_write_queue(skb, sk);
2467                 sk_stream_free_skb(sk, skb);
2468                 clear_all_retrans_hints(tp);
2469         }
2470
2471         if (acked&FLAG_ACKED) {
2472                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2473                 tcp_ack_packets_out(sk, tp);
2474                 if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
2475                         (*rtt_sample)(sk, tcp_usrtt(&tv));
2476
2477                 if (icsk->icsk_ca_ops->pkts_acked)
2478                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2479         }
2480
2481 #if FASTRETRANS_DEBUG > 0
2482         BUG_TRAP((int)tp->sacked_out >= 0);
2483         BUG_TRAP((int)tp->lost_out >= 0);
2484         BUG_TRAP((int)tp->retrans_out >= 0);
2485         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2486                 const struct inet_connection_sock *icsk = inet_csk(sk);
2487                 if (tp->lost_out) {
2488                         printk(KERN_DEBUG "Leak l=%u %d\n",
2489                                tp->lost_out, icsk->icsk_ca_state);
2490                         tp->lost_out = 0;
2491                 }
2492                 if (tp->sacked_out) {
2493                         printk(KERN_DEBUG "Leak s=%u %d\n",
2494                                tp->sacked_out, icsk->icsk_ca_state);
2495                         tp->sacked_out = 0;
2496                 }
2497                 if (tp->retrans_out) {
2498                         printk(KERN_DEBUG "Leak r=%u %d\n",
2499                                tp->retrans_out, icsk->icsk_ca_state);
2500                         tp->retrans_out = 0;
2501                 }
2502         }
2503 #endif
2504         *seq_rtt_p = seq_rtt;
2505         return acked;
2506 }
2507
2508 static void tcp_ack_probe(struct sock *sk)
2509 {
2510         const struct tcp_sock *tp = tcp_sk(sk);
2511         struct inet_connection_sock *icsk = inet_csk(sk);
2512
2513         /* Was it a usable window open? */
2514
2515         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2516                    tp->snd_una + tp->snd_wnd)) {
2517                 icsk->icsk_backoff = 0;
2518                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2519                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2520                  * This function is not for random using!
2521                  */
2522         } else {
2523                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2524                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2525                                           TCP_RTO_MAX);
2526         }
2527 }
2528
2529 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2530 {
2531         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2532                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2533 }
2534
2535 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2536 {
2537         const struct tcp_sock *tp = tcp_sk(sk);
2538         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2539                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2540 }
2541
2542 /* Check that window update is acceptable.
2543  * The function assumes that snd_una<=ack<=snd_next.
2544  */
2545 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2546                                         const u32 ack_seq, const u32 nwin)
2547 {
2548         return (after(ack, tp->snd_una) ||
2549                 after(ack_seq, tp->snd_wl1) ||
2550                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2551 }
2552
2553 /* Update our send window.
2554  *
2555  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2556  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2557  */
2558 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2559                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2560 {
2561         int flag = 0;
2562         u32 nwin = ntohs(skb->h.th->window);
2563
2564         if (likely(!skb->h.th->syn))
2565                 nwin <<= tp->rx_opt.snd_wscale;
2566
2567         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2568                 flag |= FLAG_WIN_UPDATE;
2569                 tcp_update_wl(tp, ack, ack_seq);
2570
2571                 if (tp->snd_wnd != nwin) {
2572                         tp->snd_wnd = nwin;
2573
2574                         /* Note, it is the only place, where
2575                          * fast path is recovered for sending TCP.
2576                          */
2577                         tp->pred_flags = 0;
2578                         tcp_fast_path_check(sk, tp);
2579
2580                         if (nwin > tp->max_window) {
2581                                 tp->max_window = nwin;
2582                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2583                         }
2584                 }
2585         }
2586
2587         tp->snd_una = ack;
2588
2589         return flag;
2590 }
2591
2592 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2593  * continue in congestion avoidance.
2594  */
2595 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2596 {
2597         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2598         tp->snd_cwnd_cnt = 0;
2599         tcp_moderate_cwnd(tp);
2600 }
2601
2602 /* A conservative spurious RTO response algorithm: reduce cwnd using
2603  * rate halving and continue in congestion avoidance.
2604  */
2605 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2606 {
2607         tcp_enter_cwr(sk, 0);
2608 }
2609
2610 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2611 {
2612         if (flag&FLAG_ECE)
2613                 tcp_ratehalving_spur_to_response(sk);
2614         else
2615                 tcp_undo_cwr(sk, 1);
2616 }
2617
2618 /* F-RTO spurious RTO detection algorithm (RFC4138)
2619  *
2620  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2621  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2622  * window (but not to or beyond highest sequence sent before RTO):
2623  *   On First ACK,  send two new segments out.
2624  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2625  *                  algorithm is not part of the F-RTO detection algorithm
2626  *                  given in RFC4138 but can be selected separately).
2627  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2628  * and TCP falls back to conventional RTO recovery.
2629  *
2630  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2631  * original window even after we transmit two new data segments.
2632  *
2633  * SACK version:
2634  *   on first step, wait until first cumulative ACK arrives, then move to
2635  *   the second step. In second step, the next ACK decides.
2636  *
2637  * F-RTO is implemented (mainly) in four functions:
2638  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2639  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2640  *     called when tcp_use_frto() showed green light
2641  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2642  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2643  *     to prove that the RTO is indeed spurious. It transfers the control
2644  *     from F-RTO to the conventional RTO recovery
2645  */
2646 static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
2647 {
2648         struct tcp_sock *tp = tcp_sk(sk);
2649
2650         tcp_sync_left_out(tp);
2651
2652         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2653         if (flag&FLAG_DATA_ACKED)
2654                 inet_csk(sk)->icsk_retransmits = 0;
2655
2656         if (!before(tp->snd_una, tp->frto_highmark)) {
2657                 tcp_enter_frto_loss(sk, tp->frto_counter + 1, flag);
2658                 return 1;
2659         }
2660
2661         if (!IsSackFrto() || IsReno(tp)) {
2662                 /* RFC4138 shortcoming in step 2; should also have case c):
2663                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2664                  * data, winupdate
2665                  */
2666                 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
2667                     !(flag&FLAG_FORWARD_PROGRESS))
2668                         return 1;
2669
2670                 if (!(flag&FLAG_DATA_ACKED)) {
2671                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2672                                             flag);
2673                         return 1;
2674                 }
2675         } else {
2676                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2677                         /* Prevent sending of new data. */
2678                         tp->snd_cwnd = min(tp->snd_cwnd,
2679                                            tcp_packets_in_flight(tp));
2680                         return 1;
2681                 }
2682
2683                 if ((tp->frto_counter == 2) &&
2684                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2685                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2686                         /* RFC4138 shortcoming (see comment above) */
2687                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2688                                 return 1;
2689
2690                         tcp_enter_frto_loss(sk, 3, flag);
2691                         return 1;
2692                 }
2693         }
2694
2695         if (tp->frto_counter == 1) {
2696                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2697                 tp->frto_counter = 2;
2698                 return 1;
2699         } else /* frto_counter == 2 */ {
2700                 switch (sysctl_tcp_frto_response) {
2701                 case 2:
2702                         tcp_undo_spur_to_response(sk, flag);
2703                         break;
2704                 case 1:
2705                         tcp_conservative_spur_to_response(tp);
2706                         break;
2707                 default:
2708                         tcp_ratehalving_spur_to_response(sk);
2709                         break;
2710                 };
2711                 tp->frto_counter = 0;
2712         }
2713         return 0;
2714 }
2715
2716 /* This routine deals with incoming acks, but not outgoing ones. */
2717 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2718 {
2719         struct inet_connection_sock *icsk = inet_csk(sk);
2720         struct tcp_sock *tp = tcp_sk(sk);
2721         u32 prior_snd_una = tp->snd_una;
2722         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2723         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2724         u32 prior_in_flight;
2725         s32 seq_rtt;
2726         int prior_packets;
2727         int frto_cwnd = 0;
2728
2729         /* If the ack is newer than sent or older than previous acks
2730          * then we can probably ignore it.
2731          */
2732         if (after(ack, tp->snd_nxt))
2733                 goto uninteresting_ack;
2734
2735         if (before(ack, prior_snd_una))
2736                 goto old_ack;
2737
2738         if (sysctl_tcp_abc) {
2739                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2740                         tp->bytes_acked += ack - prior_snd_una;
2741                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2742                         /* we assume just one segment left network */
2743                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2744         }
2745
2746         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2747                 /* Window is constant, pure forward advance.
2748                  * No more checks are required.
2749                  * Note, we use the fact that SND.UNA>=SND.WL2.
2750                  */
2751                 tcp_update_wl(tp, ack, ack_seq);
2752                 tp->snd_una = ack;
2753                 flag |= FLAG_WIN_UPDATE;
2754
2755                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2756
2757                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2758         } else {
2759                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2760                         flag |= FLAG_DATA;
2761                 else
2762                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2763
2764                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2765
2766                 if (TCP_SKB_CB(skb)->sacked)
2767                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2768
2769                 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2770                         flag |= FLAG_ECE;
2771
2772                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2773         }
2774
2775         /* We passed data and got it acked, remove any soft error
2776          * log. Something worked...
2777          */
2778         sk->sk_err_soft = 0;
2779         tp->rcv_tstamp = tcp_time_stamp;
2780         prior_packets = tp->packets_out;
2781         if (!prior_packets)
2782                 goto no_queue;
2783
2784         prior_in_flight = tcp_packets_in_flight(tp);
2785
2786         /* See if we can take anything off of the retransmit queue. */
2787         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2788
2789         if (tp->frto_counter)
2790                 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
2791
2792         if (tcp_ack_is_dubious(sk, flag)) {
2793                 /* Advance CWND, if state allows this. */
2794                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2795                     tcp_may_raise_cwnd(sk, flag))
2796                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2797                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2798         } else {
2799                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2800                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2801         }
2802
2803         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2804                 dst_confirm(sk->sk_dst_cache);
2805
2806         return 1;
2807
2808 no_queue:
2809         icsk->icsk_probes_out = 0;
2810
2811         /* If this ack opens up a zero window, clear backoff.  It was
2812          * being used to time the probes, and is probably far higher than
2813          * it needs to be for normal retransmission.
2814          */
2815         if (tcp_send_head(sk))
2816                 tcp_ack_probe(sk);
2817         return 1;
2818
2819 old_ack:
2820         if (TCP_SKB_CB(skb)->sacked)
2821                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2822
2823 uninteresting_ack:
2824         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2825         return 0;
2826 }
2827
2828
2829 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2830  * But, this can also be called on packets in the established flow when
2831  * the fast version below fails.
2832  */
2833 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2834 {
2835         unsigned char *ptr;
2836         struct tcphdr *th = skb->h.th;
2837         int length=(th->doff*4)-sizeof(struct tcphdr);
2838
2839         ptr = (unsigned char *)(th + 1);
2840         opt_rx->saw_tstamp = 0;
2841
2842         while (length > 0) {
2843                 int opcode=*ptr++;
2844                 int opsize;
2845
2846                 switch (opcode) {
2847                         case TCPOPT_EOL:
2848                                 return;
2849                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2850                                 length--;
2851                                 continue;
2852                         default:
2853                                 opsize=*ptr++;
2854                                 if (opsize < 2) /* "silly options" */
2855                                         return;
2856                                 if (opsize > length)
2857                                         return; /* don't parse partial options */
2858                                 switch (opcode) {
2859                                 case TCPOPT_MSS:
2860                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2861                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2862                                                 if (in_mss) {
2863                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2864                                                                 in_mss = opt_rx->user_mss;
2865                                                         opt_rx->mss_clamp = in_mss;
2866                                                 }
2867                                         }
2868                                         break;
2869                                 case TCPOPT_WINDOW:
2870                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2871                                                 if (sysctl_tcp_window_scaling) {
2872                                                         __u8 snd_wscale = *(__u8 *) ptr;
2873                                                         opt_rx->wscale_ok = 1;
2874                                                         if (snd_wscale > 14) {
2875                                                                 if (net_ratelimit())
2876                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2877                                                                                "scaling value %d >14 received.\n",
2878                                                                                snd_wscale);
2879                                                                 snd_wscale = 14;
2880                                                         }
2881                                                         opt_rx->snd_wscale = snd_wscale;
2882                                                 }
2883                                         break;
2884                                 case TCPOPT_TIMESTAMP:
2885                                         if (opsize==TCPOLEN_TIMESTAMP) {
2886                                                 if ((estab && opt_rx->tstamp_ok) ||
2887                                                     (!estab && sysctl_tcp_timestamps)) {
2888                                                         opt_rx->saw_tstamp = 1;
2889                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2890                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2891                                                 }
2892                                         }
2893                                         break;
2894                                 case TCPOPT_SACK_PERM:
2895                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2896                                                 if (sysctl_tcp_sack) {
2897                                                         opt_rx->sack_ok = 1;
2898                                                         tcp_sack_reset(opt_rx);
2899                                                 }
2900                                         }
2901                                         break;
2902
2903                                 case TCPOPT_SACK:
2904                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2905                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2906                                            opt_rx->sack_ok) {
2907                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2908                                         }
2909 #ifdef CONFIG_TCP_MD5SIG
2910                                 case TCPOPT_MD5SIG:
2911                                         /*
2912                                          * The MD5 Hash has already been
2913                                          * checked (see tcp_v{4,6}_do_rcv()).
2914                                          */
2915                                         break;
2916 #endif
2917                                 };
2918                                 ptr+=opsize-2;
2919                                 length-=opsize;
2920                 };
2921         }
2922 }
2923
2924 /* Fast parse options. This hopes to only see timestamps.
2925  * If it is wrong it falls back on tcp_parse_options().
2926  */
2927 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2928                                   struct tcp_sock *tp)
2929 {
2930         if (th->doff == sizeof(struct tcphdr)>>2) {
2931                 tp->rx_opt.saw_tstamp = 0;
2932                 return 0;
2933         } else if (tp->rx_opt.tstamp_ok &&
2934                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2935                 __be32 *ptr = (__be32 *)(th + 1);
2936                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2937                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2938                         tp->rx_opt.saw_tstamp = 1;
2939                         ++ptr;
2940                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2941                         ++ptr;
2942                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2943                         return 1;
2944                 }
2945         }
2946         tcp_parse_options(skb, &tp->rx_opt, 1);
2947         return 1;
2948 }
2949
2950 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2951 {
2952         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2953         tp->rx_opt.ts_recent_stamp = get_seconds();
2954 }
2955
2956 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2957 {
2958         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2959                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2960                  * extra check below makes sure this can only happen
2961                  * for pure ACK frames.  -DaveM
2962                  *
2963                  * Not only, also it occurs for expired timestamps.
2964                  */
2965
2966                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2967                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2968                         tcp_store_ts_recent(tp);
2969         }
2970 }
2971
2972 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2973  *
2974  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2975  * it can pass through stack. So, the following predicate verifies that
2976  * this segment is not used for anything but congestion avoidance or
2977  * fast retransmit. Moreover, we even are able to eliminate most of such
2978  * second order effects, if we apply some small "replay" window (~RTO)
2979  * to timestamp space.
2980  *
2981  * All these measures still do not guarantee that we reject wrapped ACKs
2982  * on networks with high bandwidth, when sequence space is recycled fastly,
2983  * but it guarantees that such events will be very rare and do not affect
2984  * connection seriously. This doesn't look nice, but alas, PAWS is really
2985  * buggy extension.
2986  *
2987  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2988  * states that events when retransmit arrives after original data are rare.
2989  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2990  * the biggest problem on large power networks even with minor reordering.
2991  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2992  * up to bandwidth of 18Gigabit/sec. 8) ]
2993  */
2994
2995 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2996 {
2997         struct tcp_sock *tp = tcp_sk(sk);
2998         struct tcphdr *th = skb->h.th;
2999         u32 seq = TCP_SKB_CB(skb)->seq;
3000         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3001
3002         return (/* 1. Pure ACK with correct sequence number. */
3003                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3004
3005                 /* 2. ... and duplicate ACK. */
3006                 ack == tp->snd_una &&
3007
3008                 /* 3. ... and does not update window. */
3009                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3010
3011                 /* 4. ... and sits in replay window. */
3012                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3013 }
3014
3015 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3016 {
3017         const struct tcp_sock *tp = tcp_sk(sk);
3018         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3019                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3020                 !tcp_disordered_ack(sk, skb));
3021 }
3022
3023 /* Check segment sequence number for validity.
3024  *
3025  * Segment controls are considered valid, if the segment
3026  * fits to the window after truncation to the window. Acceptability
3027  * of data (and SYN, FIN, of course) is checked separately.
3028  * See tcp_data_queue(), for example.
3029  *
3030  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3031  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3032  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3033  * (borrowed from freebsd)
3034  */
3035
3036 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3037 {
3038         return  !before(end_seq, tp->rcv_wup) &&
3039                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3040 }
3041
3042 /* When we get a reset we do this. */
3043 static void tcp_reset(struct sock *sk)
3044 {
3045         /* We want the right error as BSD sees it (and indeed as we do). */
3046         switch (sk->sk_state) {
3047                 case TCP_SYN_SENT:
3048                         sk->sk_err = ECONNREFUSED;
3049                         break;
3050                 case TCP_CLOSE_WAIT:
3051                         sk->sk_err = EPIPE;
3052                         break;
3053                 case TCP_CLOSE:
3054                         return;
3055                 default:
3056                         sk->sk_err = ECONNRESET;
3057         }
3058
3059         if (!sock_flag(sk, SOCK_DEAD))
3060                 sk->sk_error_report(sk);
3061
3062         tcp_done(sk);
3063 }
3064
3065 /*
3066  *      Process the FIN bit. This now behaves as it is supposed to work
3067  *      and the FIN takes effect when it is validly part of sequence
3068  *      space. Not before when we get holes.
3069  *
3070  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3071  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3072  *      TIME-WAIT)
3073  *
3074  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3075  *      close and we go into CLOSING (and later onto TIME-WAIT)
3076  *
3077  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3078  */
3079 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3080 {
3081         struct tcp_sock *tp = tcp_sk(sk);
3082
3083         inet_csk_schedule_ack(sk);
3084
3085         sk->sk_shutdown |= RCV_SHUTDOWN;
3086         sock_set_flag(sk, SOCK_DONE);
3087
3088         switch (sk->sk_state) {
3089                 case TCP_SYN_RECV:
3090                 case TCP_ESTABLISHED:
3091                         /* Move to CLOSE_WAIT */
3092                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3093                         inet_csk(sk)->icsk_ack.pingpong = 1;
3094                         break;
3095
3096                 case TCP_CLOSE_WAIT:
3097                 case TCP_CLOSING:
3098                         /* Received a retransmission of the FIN, do
3099                          * nothing.
3100                          */
3101                         break;
3102                 case TCP_LAST_ACK:
3103                         /* RFC793: Remain in the LAST-ACK state. */
3104                         break;
3105
3106                 case TCP_FIN_WAIT1:
3107                         /* This case occurs when a simultaneous close
3108                          * happens, we must ack the received FIN and
3109                          * enter the CLOSING state.
3110                          */
3111                         tcp_send_ack(sk);
3112                         tcp_set_state(sk, TCP_CLOSING);
3113                         break;
3114                 case TCP_FIN_WAIT2:
3115                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3116                         tcp_send_ack(sk);
3117                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3118                         break;
3119                 default:
3120                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3121                          * cases we should never reach this piece of code.
3122                          */
3123                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3124                                __FUNCTION__, sk->sk_state);
3125                         break;
3126         };
3127
3128         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3129          * Probably, we should reset in this case. For now drop them.
3130          */
3131         __skb_queue_purge(&tp->out_of_order_queue);
3132         if (tp->rx_opt.sack_ok)
3133                 tcp_sack_reset(&tp->rx_opt);
3134         sk_stream_mem_reclaim(sk);
3135
3136         if (!sock_flag(sk, SOCK_DEAD)) {
3137                 sk->sk_state_change(sk);
3138
3139                 /* Do not send POLL_HUP for half duplex close. */
3140                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3141                     sk->sk_state == TCP_CLOSE)
3142                         sk_wake_async(sk, 1, POLL_HUP);
3143                 else
3144                         sk_wake_async(sk, 1, POLL_IN);
3145         }
3146 }
3147
3148 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3149 {
3150         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3151                 if (before(seq, sp->start_seq))
3152                         sp->start_seq = seq;
3153                 if (after(end_seq, sp->end_seq))
3154                         sp->end_seq = end_seq;
3155                 return 1;
3156         }
3157         return 0;
3158 }
3159
3160 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3161 {
3162         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3163                 if (before(seq, tp->rcv_nxt))
3164                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3165                 else
3166                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3167
3168                 tp->rx_opt.dsack = 1;
3169                 tp->duplicate_sack[0].start_seq = seq;
3170                 tp->duplicate_sack[0].end_seq = end_seq;
3171                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3172         }
3173 }
3174
3175 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3176 {
3177         if (!tp->rx_opt.dsack)
3178                 tcp_dsack_set(tp, seq, end_seq);
3179         else
3180                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3181 }
3182
3183 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3184 {
3185         struct tcp_sock *tp = tcp_sk(sk);
3186
3187         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3188             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3189                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3190                 tcp_enter_quickack_mode(sk);
3191
3192                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3193                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3194
3195                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3196                                 end_seq = tp->rcv_nxt;
3197                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3198                 }
3199         }
3200
3201         tcp_send_ack(sk);
3202 }
3203
3204 /* These routines update the SACK block as out-of-order packets arrive or
3205  * in-order packets close up the sequence space.
3206  */
3207 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3208 {
3209         int this_sack;
3210         struct tcp_sack_block *sp = &tp->selective_acks[0];
3211         struct tcp_sack_block *swalk = sp+1;
3212
3213         /* See if the recent change to the first SACK eats into
3214          * or hits the sequence space of other SACK blocks, if so coalesce.
3215          */
3216         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3217                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3218                         int i;
3219
3220                         /* Zap SWALK, by moving every further SACK up by one slot.
3221                          * Decrease num_sacks.
3222                          */
3223                         tp->rx_opt.num_sacks--;
3224                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3225                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3226                                 sp[i] = sp[i+1];
3227                         continue;
3228                 }
3229                 this_sack++, swalk++;
3230         }
3231 }
3232
3233 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3234 {
3235         __u32 tmp;
3236
3237         tmp = sack1->start_seq;
3238         sack1->start_seq = sack2->start_seq;
3239         sack2->start_seq = tmp;
3240
3241         tmp = sack1->end_seq;
3242         sack1->end_seq = sack2->end_seq;
3243         sack2->end_seq = tmp;
3244 }
3245
3246 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3247 {
3248         struct tcp_sock *tp = tcp_sk(sk);
3249         struct tcp_sack_block *sp = &tp->selective_acks[0];
3250         int cur_sacks = tp->rx_opt.num_sacks;
3251         int this_sack;
3252
3253         if (!cur_sacks)
3254                 goto new_sack;
3255
3256         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3257                 if (tcp_sack_extend(sp, seq, end_seq)) {
3258                         /* Rotate this_sack to the first one. */
3259                         for (; this_sack>0; this_sack--, sp--)
3260                                 tcp_sack_swap(sp, sp-1);
3261                         if (cur_sacks > 1)
3262                                 tcp_sack_maybe_coalesce(tp);
3263                         return;
3264                 }
3265         }
3266
3267         /* Could not find an adjacent existing SACK, build a new one,
3268          * put it at the front, and shift everyone else down.  We
3269          * always know there is at least one SACK present already here.
3270          *
3271          * If the sack array is full, forget about the last one.
3272          */
3273         if (this_sack >= 4) {
3274                 this_sack--;
3275                 tp->rx_opt.num_sacks--;
3276                 sp--;
3277         }
3278         for (; this_sack > 0; this_sack--, sp--)
3279                 *sp = *(sp-1);
3280
3281 new_sack:
3282         /* Build the new head SACK, and we're done. */
3283         sp->start_seq = seq;
3284         sp->end_seq = end_seq;
3285         tp->rx_opt.num_sacks++;
3286         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3287 }
3288
3289 /* RCV.NXT advances, some SACKs should be eaten. */
3290
3291 static void tcp_sack_remove(struct tcp_sock *tp)
3292 {
3293         struct tcp_sack_block *sp = &tp->selective_acks[0];
3294         int num_sacks = tp->rx_opt.num_sacks;
3295         int this_sack;
3296
3297         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3298         if (skb_queue_empty(&tp->out_of_order_queue)) {
3299                 tp->rx_opt.num_sacks = 0;
3300                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3301                 return;
3302         }
3303
3304         for (this_sack = 0; this_sack < num_sacks; ) {
3305                 /* Check if the start of the sack is covered by RCV.NXT. */
3306                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3307                         int i;
3308
3309                         /* RCV.NXT must cover all the block! */
3310                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3311
3312                         /* Zap this SACK, by moving forward any other SACKS. */
3313                         for (i=this_sack+1; i < num_sacks; i++)
3314                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3315                         num_sacks--;
3316                         continue;
3317                 }
3318                 this_sack++;
3319                 sp++;
3320         }
3321         if (num_sacks != tp->rx_opt.num_sacks) {
3322                 tp->rx_opt.num_sacks = num_sacks;
3323                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3324         }
3325 }
3326
3327 /* This one checks to see if we can put data from the
3328  * out_of_order queue into the receive_queue.
3329  */
3330 static void tcp_ofo_queue(struct sock *sk)
3331 {
3332         struct tcp_sock *tp = tcp_sk(sk);
3333         __u32 dsack_high = tp->rcv_nxt;
3334         struct sk_buff *skb;
3335
3336         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3337                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3338                         break;
3339
3340                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3341                         __u32 dsack = dsack_high;
3342                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3343                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3344                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3345                 }
3346
3347                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3348                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3349                         __skb_unlink(skb, &tp->out_of_order_queue);
3350                         __kfree_skb(skb);
3351                         continue;
3352                 }
3353                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3354                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3355                            TCP_SKB_CB(skb)->end_seq);
3356
3357                 __skb_unlink(skb, &tp->out_of_order_queue);
3358                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3359                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3360                 if (skb->h.th->fin)
3361                         tcp_fin(skb, sk, skb->h.th);
3362         }
3363 }
3364
3365 static int tcp_prune_queue(struct sock *sk);
3366
3367 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3368 {
3369         struct tcphdr *th = skb->h.th;
3370         struct tcp_sock *tp = tcp_sk(sk);
3371         int eaten = -1;
3372
3373         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3374                 goto drop;
3375
3376         __skb_pull(skb, th->doff*4);
3377
3378         TCP_ECN_accept_cwr(tp, skb);
3379
3380         if (tp->rx_opt.dsack) {
3381                 tp->rx_opt.dsack = 0;
3382                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3383                                                     4 - tp->rx_opt.tstamp_ok);
3384         }
3385
3386         /*  Queue data for delivery to the user.
3387          *  Packets in sequence go to the receive queue.
3388          *  Out of sequence packets to the out_of_order_queue.
3389          */
3390         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3391                 if (tcp_receive_window(tp) == 0)
3392                         goto out_of_window;
3393
3394                 /* Ok. In sequence. In window. */
3395                 if (tp->ucopy.task == current &&
3396                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3397                     sock_owned_by_user(sk) && !tp->urg_data) {
3398                         int chunk = min_t(unsigned int, skb->len,
3399                                                         tp->ucopy.len);
3400
3401                         __set_current_state(TASK_RUNNING);
3402
3403                         local_bh_enable();
3404                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3405                                 tp->ucopy.len -= chunk;
3406                                 tp->copied_seq += chunk;
3407                                 eaten = (chunk == skb->len && !th->fin);
3408                                 tcp_rcv_space_adjust(sk);
3409                         }
3410                         local_bh_disable();
3411                 }
3412
3413                 if (eaten <= 0) {
3414 queue_and_out:
3415                         if (eaten < 0 &&
3416                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3417                              !sk_stream_rmem_schedule(sk, skb))) {
3418                                 if (tcp_prune_queue(sk) < 0 ||
3419                                     !sk_stream_rmem_schedule(sk, skb))
3420                                         goto drop;
3421                         }
3422                         sk_stream_set_owner_r(skb, sk);
3423                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3424                 }
3425                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3426                 if (skb->len)
3427                         tcp_event_data_recv(sk, tp, skb);
3428                 if (th->fin)
3429                         tcp_fin(skb, sk, th);
3430
3431                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3432                         tcp_ofo_queue(sk);
3433
3434                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3435                          * gap in queue is filled.
3436                          */
3437                         if (skb_queue_empty(&tp->out_of_order_queue))
3438                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3439                 }
3440
3441                 if (tp->rx_opt.num_sacks)
3442                         tcp_sack_remove(tp);
3443
3444                 tcp_fast_path_check(sk, tp);
3445
3446                 if (eaten > 0)
3447                         __kfree_skb(skb);
3448                 else if (!sock_flag(sk, SOCK_DEAD))
3449                         sk->sk_data_ready(sk, 0);
3450                 return;
3451         }
3452
3453         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3454                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3455                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3456                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3457
3458 out_of_window:
3459                 tcp_enter_quickack_mode(sk);
3460                 inet_csk_schedule_ack(sk);
3461 drop:
3462                 __kfree_skb(skb);
3463                 return;
3464         }
3465
3466         /* Out of window. F.e. zero window probe. */
3467         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3468                 goto out_of_window;
3469
3470         tcp_enter_quickack_mode(sk);
3471
3472         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3473                 /* Partial packet, seq < rcv_next < end_seq */
3474                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3475                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3476                            TCP_SKB_CB(skb)->end_seq);
3477
3478                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3479
3480                 /* If window is closed, drop tail of packet. But after
3481                  * remembering D-SACK for its head made in previous line.
3482                  */
3483                 if (!tcp_receive_window(tp))
3484                         goto out_of_window;
3485                 goto queue_and_out;
3486         }
3487
3488         TCP_ECN_check_ce(tp, skb);
3489
3490         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3491             !sk_stream_rmem_schedule(sk, skb)) {
3492                 if (tcp_prune_queue(sk) < 0 ||
3493                     !sk_stream_rmem_schedule(sk, skb))
3494                         goto drop;
3495         }
3496
3497         /* Disable header prediction. */
3498         tp->pred_flags = 0;
3499         inet_csk_schedule_ack(sk);
3500
3501         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3502                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3503
3504         sk_stream_set_owner_r(skb, sk);
3505
3506         if (!skb_peek(&tp->out_of_order_queue)) {
3507                 /* Initial out of order segment, build 1 SACK. */
3508                 if (tp->rx_opt.sack_ok) {
3509                         tp->rx_opt.num_sacks = 1;
3510                         tp->rx_opt.dsack     = 0;
3511                         tp->rx_opt.eff_sacks = 1;
3512                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3513                         tp->selective_acks[0].end_seq =
3514                                                 TCP_SKB_CB(skb)->end_seq;
3515                 }
3516                 __skb_queue_head(&tp->out_of_order_queue,skb);
3517         } else {
3518                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3519                 u32 seq = TCP_SKB_CB(skb)->seq;
3520                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3521
3522                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3523                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3524
3525                         if (!tp->rx_opt.num_sacks ||
3526                             tp->selective_acks[0].end_seq != seq)
3527                                 goto add_sack;
3528
3529                         /* Common case: data arrive in order after hole. */
3530                         tp->selective_acks[0].end_seq = end_seq;
3531                         return;
3532                 }
3533
3534                 /* Find place to insert this segment. */
3535                 do {
3536                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3537                                 break;
3538                 } while ((skb1 = skb1->prev) !=
3539                          (struct sk_buff*)&tp->out_of_order_queue);
3540
3541                 /* Do skb overlap to previous one? */
3542                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3543                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3544                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3545                                 /* All the bits are present. Drop. */
3546                                 __kfree_skb(skb);
3547                                 tcp_dsack_set(tp, seq, end_seq);
3548                                 goto add_sack;
3549                         }
3550                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3551                                 /* Partial overlap. */
3552                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3553                         } else {
3554                                 skb1 = skb1->prev;
3555                         }
3556                 }
3557                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3558
3559                 /* And clean segments covered by new one as whole. */
3560                 while ((skb1 = skb->next) !=
3561                        (struct sk_buff*)&tp->out_of_order_queue &&
3562                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3563                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3564                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3565                                break;
3566                        }
3567                        __skb_unlink(skb1, &tp->out_of_order_queue);
3568                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3569                        __kfree_skb(skb1);
3570                 }
3571
3572 add_sack:
3573                 if (tp->rx_opt.sack_ok)
3574                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3575         }
3576 }
3577
3578 /* Collapse contiguous sequence of skbs head..tail with
3579  * sequence numbers start..end.
3580  * Segments with FIN/SYN are not collapsed (only because this
3581  * simplifies code)
3582  */
3583 static void
3584 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3585              struct sk_buff *head, struct sk_buff *tail,
3586              u32 start, u32 end)
3587 {
3588         struct sk_buff *skb;
3589
3590         /* First, check that queue is collapsible and find
3591          * the point where collapsing can be useful. */
3592         for (skb = head; skb != tail; ) {
3593                 /* No new bits? It is possible on ofo queue. */
3594                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3595                         struct sk_buff *next = skb->next;
3596                         __skb_unlink(skb, list);
3597                         __kfree_skb(skb);
3598                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3599                         skb = next;
3600                         continue;
3601                 }
3602
3603                 /* The first skb to collapse is:
3604                  * - not SYN/FIN and
3605                  * - bloated or contains data before "start" or
3606                  *   overlaps to the next one.
3607                  */
3608                 if (!skb->h.th->syn && !skb->h.th->fin &&
3609                     (tcp_win_from_space(skb->truesize) > skb->len ||
3610                      before(TCP_SKB_CB(skb)->seq, start) ||
3611                      (skb->next != tail &&
3612                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3613                         break;
3614
3615                 /* Decided to skip this, advance start seq. */
3616                 start = TCP_SKB_CB(skb)->end_seq;
3617                 skb = skb->next;
3618         }
3619         if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3620                 return;
3621
3622         while (before(start, end)) {
3623                 struct sk_buff *nskb;
3624                 int header = skb_headroom(skb);
3625                 int copy = SKB_MAX_ORDER(header, 0);
3626
3627                 /* Too big header? This can happen with IPv6. */
3628                 if (copy < 0)
3629                         return;
3630                 if (end-start < copy)
3631                         copy = end-start;
3632                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3633                 if (!nskb)
3634                         return;
3635                 skb_reserve(nskb, header);
3636                 memcpy(nskb->head, skb->head, header);
3637                 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3638                 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3639                 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3640                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3641                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3642                 __skb_insert(nskb, skb->prev, skb, list);
3643                 sk_stream_set_owner_r(nskb, sk);
3644
3645                 /* Copy data, releasing collapsed skbs. */
3646                 while (copy > 0) {
3647                         int offset = start - TCP_SKB_CB(skb)->seq;
3648                         int size = TCP_SKB_CB(skb)->end_seq - start;
3649
3650                         BUG_ON(offset < 0);
3651                         if (size > 0) {
3652                                 size = min(copy, size);
3653                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3654                                         BUG();
3655                                 TCP_SKB_CB(nskb)->end_seq += size;
3656                                 copy -= size;
3657                                 start += size;
3658                         }
3659                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3660                                 struct sk_buff *next = skb->next;
3661                                 __skb_unlink(skb, list);
3662                                 __kfree_skb(skb);
3663                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3664                                 skb = next;
3665                                 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3666                                         return;
3667                         }
3668                 }
3669         }
3670 }
3671
3672 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3673  * and tcp_collapse() them until all the queue is collapsed.
3674  */
3675 static void tcp_collapse_ofo_queue(struct sock *sk)
3676 {
3677         struct tcp_sock *tp = tcp_sk(sk);
3678         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3679         struct sk_buff *head;
3680         u32 start, end;
3681
3682         if (skb == NULL)
3683                 return;
3684
3685         start = TCP_SKB_CB(skb)->seq;
3686         end = TCP_SKB_CB(skb)->end_seq;
3687         head = skb;
3688
3689         for (;;) {
3690                 skb = skb->next;
3691
3692                 /* Segment is terminated when we see gap or when
3693                  * we are at the end of all the queue. */
3694                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3695                     after(TCP_SKB_CB(skb)->seq, end) ||
3696                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3697                         tcp_collapse(sk, &tp->out_of_order_queue,
3698                                      head, skb, start, end);
3699                         head = skb;
3700                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3701                                 break;
3702                         /* Start new segment */
3703                         start = TCP_SKB_CB(skb)->seq;
3704                         end = TCP_SKB_CB(skb)->end_seq;
3705                 } else {
3706                         if (before(TCP_SKB_CB(skb)->seq, start))
3707                                 start = TCP_SKB_CB(skb)->seq;
3708                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3709                                 end = TCP_SKB_CB(skb)->end_seq;
3710                 }
3711         }
3712 }
3713
3714 /* Reduce allocated memory if we can, trying to get
3715  * the socket within its memory limits again.
3716  *
3717  * Return less than zero if we should start dropping frames
3718  * until the socket owning process reads some of the data
3719  * to stabilize the situation.
3720  */
3721 static int tcp_prune_queue(struct sock *sk)
3722 {
3723         struct tcp_sock *tp = tcp_sk(sk);
3724
3725         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3726
3727         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3728
3729         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3730                 tcp_clamp_window(sk, tp);
3731         else if (tcp_memory_pressure)
3732                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3733
3734         tcp_collapse_ofo_queue(sk);
3735         tcp_collapse(sk, &sk->sk_receive_queue,
3736                      sk->sk_receive_queue.next,
3737                      (struct sk_buff*)&sk->sk_receive_queue,
3738                      tp->copied_seq, tp->rcv_nxt);
3739         sk_stream_mem_reclaim(sk);
3740
3741         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3742                 return 0;
3743
3744         /* Collapsing did not help, destructive actions follow.
3745          * This must not ever occur. */
3746
3747         /* First, purge the out_of_order queue. */
3748         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3749                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3750                 __skb_queue_purge(&tp->out_of_order_queue);
3751
3752                 /* Reset SACK state.  A conforming SACK implementation will
3753                  * do the same at a timeout based retransmit.  When a connection
3754                  * is in a sad state like this, we care only about integrity
3755                  * of the connection not performance.
3756                  */
3757                 if (tp->rx_opt.sack_ok)
3758                         tcp_sack_reset(&tp->rx_opt);
3759                 sk_stream_mem_reclaim(sk);
3760         }
3761
3762         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3763                 return 0;
3764
3765         /* If we are really being abused, tell the caller to silently
3766          * drop receive data on the floor.  It will get retransmitted
3767          * and hopefully then we'll have sufficient space.
3768          */
3769         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3770
3771         /* Massive buffer overcommit. */
3772         tp->pred_flags = 0;
3773         return -1;
3774 }
3775
3776
3777 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3778  * As additional protections, we do not touch cwnd in retransmission phases,
3779  * and if application hit its sndbuf limit recently.
3780  */
3781 void tcp_cwnd_application_limited(struct sock *sk)
3782 {
3783         struct tcp_sock *tp = tcp_sk(sk);
3784
3785         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3786             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3787                 /* Limited by application or receiver window. */
3788                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3789                 u32 win_used = max(tp->snd_cwnd_used, init_win);
3790                 if (win_used < tp->snd_cwnd) {
3791                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3792                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3793                 }
3794                 tp->snd_cwnd_used = 0;
3795         }
3796         tp->snd_cwnd_stamp = tcp_time_stamp;
3797 }
3798
3799 static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3800 {
3801         /* If the user specified a specific send buffer setting, do
3802          * not modify it.
3803          */
3804         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3805                 return 0;
3806
3807         /* If we are under global TCP memory pressure, do not expand.  */
3808         if (tcp_memory_pressure)
3809                 return 0;
3810
3811         /* If we are under soft global TCP memory pressure, do not expand.  */
3812         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3813                 return 0;
3814
3815         /* If we filled the congestion window, do not expand.  */
3816         if (tp->packets_out >= tp->snd_cwnd)
3817                 return 0;
3818
3819         return 1;
3820 }
3821
3822 /* When incoming ACK allowed to free some skb from write_queue,
3823  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3824  * on the exit from tcp input handler.
3825  *
3826  * PROBLEM: sndbuf expansion does not work well with largesend.
3827  */
3828 static void tcp_new_space(struct sock *sk)
3829 {
3830         struct tcp_sock *tp = tcp_sk(sk);
3831
3832         if (tcp_should_expand_sndbuf(sk, tp)) {
3833                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3834                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3835                     demanded = max_t(unsigned int, tp->snd_cwnd,
3836                                                    tp->reordering + 1);
3837                 sndmem *= 2*demanded;
3838                 if (sndmem > sk->sk_sndbuf)
3839                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3840                 tp->snd_cwnd_stamp = tcp_time_stamp;
3841         }
3842
3843         sk->sk_write_space(sk);
3844 }
3845
3846 static void tcp_check_space(struct sock *sk)
3847 {
3848         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3849                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3850                 if (sk->sk_socket &&
3851                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3852                         tcp_new_space(sk);
3853         }
3854 }
3855
3856 static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3857 {
3858         tcp_push_pending_frames(sk, tp);
3859         tcp_check_space(sk);
3860 }
3861
3862 /*
3863  * Check if sending an ack is needed.
3864  */
3865 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3866 {
3867         struct tcp_sock *tp = tcp_sk(sk);
3868
3869             /* More than one full frame received... */
3870         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3871              /* ... and right edge of window advances far enough.
3872               * (tcp_recvmsg() will send ACK otherwise). Or...
3873               */
3874              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3875             /* We ACK each frame or... */
3876             tcp_in_quickack_mode(sk) ||
3877             /* We have out of order data. */
3878             (ofo_possible &&
3879              skb_peek(&tp->out_of_order_queue))) {
3880                 /* Then ack it now */
3881                 tcp_send_ack(sk);
3882         } else {
3883                 /* Else, send delayed ack. */
3884                 tcp_send_delayed_ack(sk);
3885         }
3886 }
3887
3888 static inline void tcp_ack_snd_check(struct sock *sk)
3889 {
3890         if (!inet_csk_ack_scheduled(sk)) {
3891                 /* We sent a data segment already. */
3892                 return;
3893         }
3894         __tcp_ack_snd_check(sk, 1);
3895 }
3896
3897 /*
3898  *      This routine is only called when we have urgent data
3899  *      signaled. Its the 'slow' part of tcp_urg. It could be
3900  *      moved inline now as tcp_urg is only called from one
3901  *      place. We handle URGent data wrong. We have to - as
3902  *      BSD still doesn't use the correction from RFC961.
3903  *      For 1003.1g we should support a new option TCP_STDURG to permit
3904  *      either form (or just set the sysctl tcp_stdurg).
3905  */
3906
3907 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3908 {
3909         struct tcp_sock *tp = tcp_sk(sk);
3910         u32 ptr = ntohs(th->urg_ptr);
3911
3912         if (ptr && !sysctl_tcp_stdurg)
3913                 ptr--;
3914         ptr += ntohl(th->seq);
3915
3916         /* Ignore urgent data that we've already seen and read. */
3917         if (after(tp->copied_seq, ptr))
3918                 return;
3919
3920         /* Do not replay urg ptr.
3921          *
3922          * NOTE: interesting situation not covered by specs.
3923          * Misbehaving sender may send urg ptr, pointing to segment,
3924          * which we already have in ofo queue. We are not able to fetch
3925          * such data and will stay in TCP_URG_NOTYET until will be eaten
3926          * by recvmsg(). Seems, we are not obliged to handle such wicked
3927          * situations. But it is worth to think about possibility of some
3928          * DoSes using some hypothetical application level deadlock.
3929          */
3930         if (before(ptr, tp->rcv_nxt))
3931                 return;
3932
3933         /* Do we already have a newer (or duplicate) urgent pointer? */
3934         if (tp->urg_data && !after(ptr, tp->urg_seq))
3935                 return;
3936
3937         /* Tell the world about our new urgent pointer. */
3938         sk_send_sigurg(sk);
3939
3940         /* We may be adding urgent data when the last byte read was
3941          * urgent. To do this requires some care. We cannot just ignore
3942          * tp->copied_seq since we would read the last urgent byte again
3943          * as data, nor can we alter copied_seq until this data arrives
3944          * or we break the semantics of SIOCATMARK (and thus sockatmark())
3945          *
3946          * NOTE. Double Dutch. Rendering to plain English: author of comment
3947          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3948          * and expect that both A and B disappear from stream. This is _wrong_.
3949          * Though this happens in BSD with high probability, this is occasional.
3950          * Any application relying on this is buggy. Note also, that fix "works"
3951          * only in this artificial test. Insert some normal data between A and B and we will
3952          * decline of BSD again. Verdict: it is better to remove to trap
3953          * buggy users.
3954          */
3955         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3956             !sock_flag(sk, SOCK_URGINLINE) &&
3957             tp->copied_seq != tp->rcv_nxt) {
3958                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3959                 tp->copied_seq++;
3960                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3961                         __skb_unlink(skb, &sk->sk_receive_queue);
3962                         __kfree_skb(skb);
3963                 }
3964         }
3965
3966         tp->urg_data   = TCP_URG_NOTYET;
3967         tp->urg_seq    = ptr;
3968
3969         /* Disable header prediction. */
3970         tp->pred_flags = 0;
3971 }
3972
3973 /* This is the 'fast' part of urgent handling. */
3974 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3975 {
3976         struct tcp_sock *tp = tcp_sk(sk);
3977
3978         /* Check if we get a new urgent pointer - normally not. */
3979         if (th->urg)
3980                 tcp_check_urg(sk,th);
3981
3982         /* Do we wait for any urgent data? - normally not... */
3983         if (tp->urg_data == TCP_URG_NOTYET) {
3984                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3985                           th->syn;
3986
3987                 /* Is the urgent pointer pointing into this packet? */
3988                 if (ptr < skb->len) {
3989                         u8 tmp;
3990                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3991                                 BUG();
3992                         tp->urg_data = TCP_URG_VALID | tmp;
3993                         if (!sock_flag(sk, SOCK_DEAD))
3994                                 sk->sk_data_ready(sk, 0);
3995                 }
3996         }
3997 }
3998
3999 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4000 {
4001         struct tcp_sock *tp = tcp_sk(sk);
4002         int chunk = skb->len - hlen;
4003         int err;
4004
4005         local_bh_enable();
4006         if (skb->ip_summed==CHECKSUM_UNNECESSARY)
4007                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4008         else
4009                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4010                                                        tp->ucopy.iov);
4011
4012         if (!err) {
4013                 tp->ucopy.len -= chunk;
4014                 tp->copied_seq += chunk;
4015                 tcp_rcv_space_adjust(sk);
4016         }
4017
4018         local_bh_disable();
4019         return err;
4020 }
4021
4022 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4023 {
4024         __sum16 result;
4025
4026         if (sock_owned_by_user(sk)) {
4027                 local_bh_enable();
4028                 result = __tcp_checksum_complete(skb);
4029                 local_bh_disable();
4030         } else {
4031                 result = __tcp_checksum_complete(skb);
4032         }
4033         return result;
4034 }
4035
4036 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4037 {
4038         return skb->ip_summed != CHECKSUM_UNNECESSARY &&
4039                 __tcp_checksum_complete_user(sk, skb);
4040 }
4041
4042 #ifdef CONFIG_NET_DMA
4043 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4044 {
4045         struct tcp_sock *tp = tcp_sk(sk);
4046         int chunk = skb->len - hlen;
4047         int dma_cookie;
4048         int copied_early = 0;
4049
4050         if (tp->ucopy.wakeup)
4051                 return 0;
4052
4053         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4054                 tp->ucopy.dma_chan = get_softnet_dma();
4055
4056         if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) {
4057
4058                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4059                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4060
4061                 if (dma_cookie < 0)
4062                         goto out;
4063
4064                 tp->ucopy.dma_cookie = dma_cookie;
4065                 copied_early = 1;
4066
4067                 tp->ucopy.len -= chunk;
4068                 tp->copied_seq += chunk;
4069                 tcp_rcv_space_adjust(sk);
4070
4071                 if ((tp->ucopy.len == 0) ||
4072                     (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) ||
4073                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4074                         tp->ucopy.wakeup = 1;
4075                         sk->sk_data_ready(sk, 0);
4076                 }
4077         } else if (chunk > 0) {
4078                 tp->ucopy.wakeup = 1;
4079                 sk->sk_data_ready(sk, 0);
4080         }
4081 out:
4082         return copied_early;
4083 }
4084 #endif /* CONFIG_NET_DMA */
4085
4086 /*
4087  *      TCP receive function for the ESTABLISHED state.
4088  *
4089  *      It is split into a fast path and a slow path. The fast path is
4090  *      disabled when:
4091  *      - A zero window was announced from us - zero window probing
4092  *        is only handled properly in the slow path.
4093  *      - Out of order segments arrived.
4094  *      - Urgent data is expected.
4095  *      - There is no buffer space left
4096  *      - Unexpected TCP flags/window values/header lengths are received
4097  *        (detected by checking the TCP header against pred_flags)
4098  *      - Data is sent in both directions. Fast path only supports pure senders
4099  *        or pure receivers (this means either the sequence number or the ack
4100  *        value must stay constant)
4101  *      - Unexpected TCP option.
4102  *
4103  *      When these conditions are not satisfied it drops into a standard
4104  *      receive procedure patterned after RFC793 to handle all cases.
4105  *      The first three cases are guaranteed by proper pred_flags setting,
4106  *      the rest is checked inline. Fast processing is turned on in
4107  *      tcp_data_queue when everything is OK.
4108  */
4109 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4110                         struct tcphdr *th, unsigned len)
4111 {
4112         struct tcp_sock *tp = tcp_sk(sk);
4113
4114         /*
4115          *      Header prediction.
4116          *      The code loosely follows the one in the famous
4117          *      "30 instruction TCP receive" Van Jacobson mail.
4118          *
4119          *      Van's trick is to deposit buffers into socket queue
4120          *      on a device interrupt, to call tcp_recv function
4121          *      on the receive process context and checksum and copy
4122          *      the buffer to user space. smart...
4123          *
4124          *      Our current scheme is not silly either but we take the
4125          *      extra cost of the net_bh soft interrupt processing...
4126          *      We do checksum and copy also but from device to kernel.
4127          */
4128
4129         tp->rx_opt.saw_tstamp = 0;
4130
4131         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4132          *      if header_prediction is to be made
4133          *      'S' will always be tp->tcp_header_len >> 2
4134          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4135          *  turn it off (when there are holes in the receive
4136          *       space for instance)
4137          *      PSH flag is ignored.
4138          */
4139
4140         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4141                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4142                 int tcp_header_len = tp->tcp_header_len;
4143
4144                 /* Timestamp header prediction: tcp_header_len
4145                  * is automatically equal to th->doff*4 due to pred_flags
4146                  * match.
4147                  */
4148
4149                 /* Check timestamp */
4150                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4151                         __be32 *ptr = (__be32 *)(th + 1);
4152
4153                         /* No? Slow path! */
4154                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4155                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4156                                 goto slow_path;
4157
4158                         tp->rx_opt.saw_tstamp = 1;
4159                         ++ptr;
4160                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4161                         ++ptr;
4162                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4163
4164                         /* If PAWS failed, check it more carefully in slow path */
4165                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4166                                 goto slow_path;
4167
4168                         /* DO NOT update ts_recent here, if checksum fails
4169                          * and timestamp was corrupted part, it will result
4170                          * in a hung connection since we will drop all
4171                          * future packets due to the PAWS test.
4172                          */
4173                 }
4174
4175                 if (len <= tcp_header_len) {
4176                         /* Bulk data transfer: sender */
4177                         if (len == tcp_header_len) {
4178                                 /* Predicted packet is in window by definition.
4179                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4180                                  * Hence, check seq<=rcv_wup reduces to:
4181                                  */
4182                                 if (tcp_header_len ==
4183                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4184                                     tp->rcv_nxt == tp->rcv_wup)
4185                                         tcp_store_ts_recent(tp);
4186
4187                                 /* We know that such packets are checksummed
4188                                  * on entry.
4189                                  */
4190                                 tcp_ack(sk, skb, 0);
4191                                 __kfree_skb(skb);
4192                                 tcp_data_snd_check(sk, tp);
4193                                 return 0;
4194                         } else { /* Header too small */
4195                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4196                                 goto discard;
4197                         }
4198                 } else {
4199                         int eaten = 0;
4200                         int copied_early = 0;
4201
4202                         if (tp->copied_seq == tp->rcv_nxt &&
4203                             len - tcp_header_len <= tp->ucopy.len) {
4204 #ifdef CONFIG_NET_DMA
4205                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4206                                         copied_early = 1;
4207                                         eaten = 1;
4208                                 }
4209 #endif
4210                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4211                                         __set_current_state(TASK_RUNNING);
4212
4213                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4214                                                 eaten = 1;
4215                                 }
4216                                 if (eaten) {
4217                                         /* Predicted packet is in window by definition.
4218                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4219                                          * Hence, check seq<=rcv_wup reduces to:
4220                                          */
4221                                         if (tcp_header_len ==
4222                                             (sizeof(struct tcphdr) +
4223                                              TCPOLEN_TSTAMP_ALIGNED) &&
4224                                             tp->rcv_nxt == tp->rcv_wup)
4225                                                 tcp_store_ts_recent(tp);
4226
4227                                         tcp_rcv_rtt_measure_ts(sk, skb);
4228
4229                                         __skb_pull(skb, tcp_header_len);
4230                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4231                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4232                                 }
4233                                 if (copied_early)
4234                                         tcp_cleanup_rbuf(sk, skb->len);
4235                         }
4236                         if (!eaten) {
4237                                 if (tcp_checksum_complete_user(sk, skb))
4238                                         goto csum_error;
4239
4240                                 /* Predicted packet is in window by definition.
4241                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4242                                  * Hence, check seq<=rcv_wup reduces to:
4243                                  */
4244                                 if (tcp_header_len ==
4245                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4246                                     tp->rcv_nxt == tp->rcv_wup)
4247                                         tcp_store_ts_recent(tp);
4248
4249                                 tcp_rcv_rtt_measure_ts(sk, skb);
4250
4251                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4252                                         goto step5;
4253
4254                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4255
4256                                 /* Bulk data transfer: receiver */
4257                                 __skb_pull(skb,tcp_header_len);
4258                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4259                                 sk_stream_set_owner_r(skb, sk);
4260                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4261                         }
4262
4263                         tcp_event_data_recv(sk, tp, skb);
4264
4265                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4266                                 /* Well, only one small jumplet in fast path... */
4267                                 tcp_ack(sk, skb, FLAG_DATA);
4268                                 tcp_data_snd_check(sk, tp);
4269                                 if (!inet_csk_ack_scheduled(sk))
4270                                         goto no_ack;
4271                         }
4272
4273                         __tcp_ack_snd_check(sk, 0);
4274 no_ack:
4275 #ifdef CONFIG_NET_DMA
4276                         if (copied_early)
4277                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4278                         else
4279 #endif
4280                         if (eaten)
4281                                 __kfree_skb(skb);
4282                         else
4283                                 sk->sk_data_ready(sk, 0);
4284                         return 0;
4285                 }
4286         }
4287
4288 slow_path:
4289         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4290                 goto csum_error;
4291
4292         /*
4293          * RFC1323: H1. Apply PAWS check first.
4294          */
4295         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4296             tcp_paws_discard(sk, skb)) {
4297                 if (!th->rst) {
4298                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4299                         tcp_send_dupack(sk, skb);
4300                         goto discard;
4301                 }
4302                 /* Resets are accepted even if PAWS failed.
4303
4304                    ts_recent update must be made after we are sure
4305                    that the packet is in window.
4306                  */
4307         }
4308
4309         /*
4310          *      Standard slow path.
4311          */
4312
4313         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4314                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4315                  * (RST) segments are validated by checking their SEQ-fields."
4316                  * And page 69: "If an incoming segment is not acceptable,
4317                  * an acknowledgment should be sent in reply (unless the RST bit
4318                  * is set, if so drop the segment and return)".
4319                  */
4320                 if (!th->rst)
4321                         tcp_send_dupack(sk, skb);
4322                 goto discard;
4323         }
4324
4325         if (th->rst) {
4326                 tcp_reset(sk);
4327                 goto discard;
4328         }
4329
4330         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4331
4332         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4333                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4334                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4335                 tcp_reset(sk);
4336                 return 1;
4337         }
4338
4339 step5:
4340         if (th->ack)
4341                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4342
4343         tcp_rcv_rtt_measure_ts(sk, skb);
4344
4345         /* Process urgent data. */
4346         tcp_urg(sk, skb, th);
4347
4348         /* step 7: process the segment text */
4349         tcp_data_queue(sk, skb);
4350
4351         tcp_data_snd_check(sk, tp);
4352         tcp_ack_snd_check(sk);
4353         return 0;
4354
4355 csum_error:
4356         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4357
4358 discard:
4359         __kfree_skb(skb);
4360         return 0;
4361 }
4362
4363 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4364                                          struct tcphdr *th, unsigned len)
4365 {
4366         struct tcp_sock *tp = tcp_sk(sk);
4367         struct inet_connection_sock *icsk = inet_csk(sk);
4368         int saved_clamp = tp->rx_opt.mss_clamp;
4369
4370         tcp_parse_options(skb, &tp->rx_opt, 0);
4371
4372         if (th->ack) {
4373                 /* rfc793:
4374                  * "If the state is SYN-SENT then
4375                  *    first check the ACK bit
4376                  *      If the ACK bit is set
4377                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4378                  *        a reset (unless the RST bit is set, if so drop
4379                  *        the segment and return)"
4380                  *
4381                  *  We do not send data with SYN, so that RFC-correct
4382                  *  test reduces to:
4383                  */
4384                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4385                         goto reset_and_undo;
4386
4387                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4388                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4389                              tcp_time_stamp)) {
4390                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4391                         goto reset_and_undo;
4392                 }
4393
4394                 /* Now ACK is acceptable.
4395                  *
4396                  * "If the RST bit is set
4397                  *    If the ACK was acceptable then signal the user "error:
4398                  *    connection reset", drop the segment, enter CLOSED state,
4399                  *    delete TCB, and return."
4400                  */
4401
4402                 if (th->rst) {
4403                         tcp_reset(sk);
4404                         goto discard;
4405                 }
4406
4407                 /* rfc793:
4408                  *   "fifth, if neither of the SYN or RST bits is set then
4409                  *    drop the segment and return."
4410                  *
4411                  *    See note below!
4412                  *                                        --ANK(990513)
4413                  */
4414                 if (!th->syn)
4415                         goto discard_and_undo;
4416
4417                 /* rfc793:
4418                  *   "If the SYN bit is on ...
4419                  *    are acceptable then ...
4420                  *    (our SYN has been ACKed), change the connection
4421                  *    state to ESTABLISHED..."
4422                  */
4423
4424                 TCP_ECN_rcv_synack(tp, th);
4425
4426                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4427                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4428
4429                 /* Ok.. it's good. Set up sequence numbers and
4430                  * move to established.
4431                  */
4432                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4433                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4434
4435                 /* RFC1323: The window in SYN & SYN/ACK segments is
4436                  * never scaled.
4437                  */
4438                 tp->snd_wnd = ntohs(th->window);
4439                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4440
4441                 if (!tp->rx_opt.wscale_ok) {
4442                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4443                         tp->window_clamp = min(tp->window_clamp, 65535U);
4444                 }
4445
4446                 if (tp->rx_opt.saw_tstamp) {
4447                         tp->rx_opt.tstamp_ok       = 1;
4448                         tp->tcp_header_len =
4449                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4450                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4451                         tcp_store_ts_recent(tp);
4452                 } else {
4453                         tp->tcp_header_len = sizeof(struct tcphdr);
4454                 }
4455
4456                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4457                         tp->rx_opt.sack_ok |= 2;
4458
4459                 tcp_mtup_init(sk);
4460                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4461                 tcp_initialize_rcv_mss(sk);
4462
4463                 /* Remember, tcp_poll() does not lock socket!
4464                  * Change state from SYN-SENT only after copied_seq
4465                  * is initialized. */
4466                 tp->copied_seq = tp->rcv_nxt;
4467                 smp_mb();
4468                 tcp_set_state(sk, TCP_ESTABLISHED);
4469
4470                 security_inet_conn_established(sk, skb);
4471
4472                 /* Make sure socket is routed, for correct metrics.  */
4473                 icsk->icsk_af_ops->rebuild_header(sk);
4474
4475                 tcp_init_metrics(sk);
4476
4477                 tcp_init_congestion_control(sk);
4478
4479                 /* Prevent spurious tcp_cwnd_restart() on first data
4480                  * packet.
4481                  */
4482                 tp->lsndtime = tcp_time_stamp;
4483
4484                 tcp_init_buffer_space(sk);
4485
4486                 if (sock_flag(sk, SOCK_KEEPOPEN))
4487                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4488
4489                 if (!tp->rx_opt.snd_wscale)
4490                         __tcp_fast_path_on(tp, tp->snd_wnd);
4491                 else
4492                         tp->pred_flags = 0;
4493
4494                 if (!sock_flag(sk, SOCK_DEAD)) {
4495                         sk->sk_state_change(sk);
4496                         sk_wake_async(sk, 0, POLL_OUT);
4497                 }
4498
4499                 if (sk->sk_write_pending ||
4500                     icsk->icsk_accept_queue.rskq_defer_accept ||
4501                     icsk->icsk_ack.pingpong) {
4502                         /* Save one ACK. Data will be ready after
4503                          * several ticks, if write_pending is set.
4504                          *
4505                          * It may be deleted, but with this feature tcpdumps
4506                          * look so _wonderfully_ clever, that I was not able
4507                          * to stand against the temptation 8)     --ANK
4508                          */
4509                         inet_csk_schedule_ack(sk);
4510                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4511                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4512                         tcp_incr_quickack(sk);
4513                         tcp_enter_quickack_mode(sk);
4514                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4515                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4516
4517 discard:
4518                         __kfree_skb(skb);
4519                         return 0;
4520                 } else {
4521                         tcp_send_ack(sk);
4522                 }
4523                 return -1;
4524         }
4525
4526         /* No ACK in the segment */
4527
4528         if (th->rst) {
4529                 /* rfc793:
4530                  * "If the RST bit is set
4531                  *
4532                  *      Otherwise (no ACK) drop the segment and return."
4533                  */
4534
4535                 goto discard_and_undo;
4536         }
4537
4538         /* PAWS check. */
4539         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4540                 goto discard_and_undo;
4541
4542         if (th->syn) {
4543                 /* We see SYN without ACK. It is attempt of
4544                  * simultaneous connect with crossed SYNs.
4545                  * Particularly, it can be connect to self.
4546                  */
4547                 tcp_set_state(sk, TCP_SYN_RECV);
4548
4549                 if (tp->rx_opt.saw_tstamp) {
4550                         tp->rx_opt.tstamp_ok = 1;
4551                         tcp_store_ts_recent(tp);
4552                         tp->tcp_header_len =
4553                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4554                 } else {
4555                         tp->tcp_header_len = sizeof(struct tcphdr);
4556                 }
4557
4558                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4559                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4560
4561                 /* RFC1323: The window in SYN & SYN/ACK segments is
4562                  * never scaled.
4563                  */
4564                 tp->snd_wnd    = ntohs(th->window);
4565                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4566                 tp->max_window = tp->snd_wnd;
4567
4568                 TCP_ECN_rcv_syn(tp, th);
4569
4570                 tcp_mtup_init(sk);
4571                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4572                 tcp_initialize_rcv_mss(sk);
4573
4574
4575                 tcp_send_synack(sk);
4576 #if 0
4577                 /* Note, we could accept data and URG from this segment.
4578                  * There are no obstacles to make this.
4579                  *
4580                  * However, if we ignore data in ACKless segments sometimes,
4581                  * we have no reasons to accept it sometimes.
4582                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4583                  * is not flawless. So, discard packet for sanity.
4584                  * Uncomment this return to process the data.
4585                  */
4586                 return -1;
4587 #else
4588                 goto discard;
4589 #endif
4590         }
4591         /* "fifth, if neither of the SYN or RST bits is set then
4592          * drop the segment and return."
4593          */
4594
4595 discard_and_undo:
4596         tcp_clear_options(&tp->rx_opt);
4597         tp->rx_opt.mss_clamp = saved_clamp;
4598         goto discard;
4599
4600 reset_and_undo:
4601         tcp_clear_options(&tp->rx_opt);
4602         tp->rx_opt.mss_clamp = saved_clamp;
4603         return 1;
4604 }
4605
4606
4607 /*
4608  *      This function implements the receiving procedure of RFC 793 for
4609  *      all states except ESTABLISHED and TIME_WAIT.
4610  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4611  *      address independent.
4612  */
4613
4614 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4615                           struct tcphdr *th, unsigned len)
4616 {
4617         struct tcp_sock *tp = tcp_sk(sk);
4618         struct inet_connection_sock *icsk = inet_csk(sk);
4619         int queued = 0;
4620
4621         tp->rx_opt.saw_tstamp = 0;
4622
4623         switch (sk->sk_state) {
4624         case TCP_CLOSE:
4625                 goto discard;
4626
4627         case TCP_LISTEN:
4628                 if (th->ack)
4629                         return 1;
4630
4631                 if (th->rst)
4632                         goto discard;
4633
4634                 if (th->syn) {
4635                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4636                                 return 1;
4637
4638                         /* Now we have several options: In theory there is
4639                          * nothing else in the frame. KA9Q has an option to
4640                          * send data with the syn, BSD accepts data with the
4641                          * syn up to the [to be] advertised window and
4642                          * Solaris 2.1 gives you a protocol error. For now
4643                          * we just ignore it, that fits the spec precisely
4644                          * and avoids incompatibilities. It would be nice in
4645                          * future to drop through and process the data.
4646                          *
4647                          * Now that TTCP is starting to be used we ought to
4648                          * queue this data.
4649                          * But, this leaves one open to an easy denial of
4650                          * service attack, and SYN cookies can't defend
4651                          * against this problem. So, we drop the data
4652                          * in the interest of security over speed unless
4653                          * it's still in use.
4654                          */
4655                         kfree_skb(skb);
4656                         return 0;
4657                 }
4658                 goto discard;
4659
4660         case TCP_SYN_SENT:
4661                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4662                 if (queued >= 0)
4663                         return queued;
4664
4665                 /* Do step6 onward by hand. */
4666                 tcp_urg(sk, skb, th);
4667                 __kfree_skb(skb);
4668                 tcp_data_snd_check(sk, tp);
4669                 return 0;
4670         }
4671
4672         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4673             tcp_paws_discard(sk, skb)) {
4674                 if (!th->rst) {
4675                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4676                         tcp_send_dupack(sk, skb);
4677                         goto discard;
4678                 }
4679                 /* Reset is accepted even if it did not pass PAWS. */
4680         }
4681
4682         /* step 1: check sequence number */
4683         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4684                 if (!th->rst)
4685                         tcp_send_dupack(sk, skb);
4686                 goto discard;
4687         }
4688
4689         /* step 2: check RST bit */
4690         if (th->rst) {
4691                 tcp_reset(sk);
4692                 goto discard;
4693         }
4694
4695         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4696
4697         /* step 3: check security and precedence [ignored] */
4698
4699         /*      step 4:
4700          *
4701          *      Check for a SYN in window.
4702          */
4703         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4704                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4705                 tcp_reset(sk);
4706                 return 1;
4707         }
4708
4709         /* step 5: check the ACK field */
4710         if (th->ack) {
4711                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4712
4713                 switch (sk->sk_state) {
4714                 case TCP_SYN_RECV:
4715                         if (acceptable) {
4716                                 tp->copied_seq = tp->rcv_nxt;
4717                                 smp_mb();
4718                                 tcp_set_state(sk, TCP_ESTABLISHED);
4719                                 sk->sk_state_change(sk);
4720
4721                                 /* Note, that this wakeup is only for marginal
4722                                  * crossed SYN case. Passively open sockets
4723                                  * are not waked up, because sk->sk_sleep ==
4724                                  * NULL and sk->sk_socket == NULL.
4725                                  */
4726                                 if (sk->sk_socket) {
4727                                         sk_wake_async(sk,0,POLL_OUT);
4728                                 }
4729
4730                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4731                                 tp->snd_wnd = ntohs(th->window) <<
4732                                               tp->rx_opt.snd_wscale;
4733                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4734                                             TCP_SKB_CB(skb)->seq);
4735
4736                                 /* tcp_ack considers this ACK as duplicate
4737                                  * and does not calculate rtt.
4738                                  * Fix it at least with timestamps.
4739                                  */
4740                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4741                                     !tp->srtt)
4742                                         tcp_ack_saw_tstamp(sk, 0);
4743
4744                                 if (tp->rx_opt.tstamp_ok)
4745                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4746
4747                                 /* Make sure socket is routed, for
4748                                  * correct metrics.
4749                                  */
4750                                 icsk->icsk_af_ops->rebuild_header(sk);
4751
4752                                 tcp_init_metrics(sk);
4753
4754                                 tcp_init_congestion_control(sk);
4755
4756                                 /* Prevent spurious tcp_cwnd_restart() on
4757                                  * first data packet.
4758                                  */
4759                                 tp->lsndtime = tcp_time_stamp;
4760
4761                                 tcp_mtup_init(sk);
4762                                 tcp_initialize_rcv_mss(sk);
4763                                 tcp_init_buffer_space(sk);
4764                                 tcp_fast_path_on(tp);
4765                         } else {
4766                                 return 1;
4767                         }
4768                         break;
4769
4770                 case TCP_FIN_WAIT1:
4771                         if (tp->snd_una == tp->write_seq) {
4772                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4773                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4774                                 dst_confirm(sk->sk_dst_cache);
4775
4776                                 if (!sock_flag(sk, SOCK_DEAD))
4777                                         /* Wake up lingering close() */
4778                                         sk->sk_state_change(sk);
4779                                 else {
4780                                         int tmo;
4781
4782                                         if (tp->linger2 < 0 ||
4783                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4784                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4785                                                 tcp_done(sk);
4786                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4787                                                 return 1;
4788                                         }
4789
4790                                         tmo = tcp_fin_time(sk);
4791                                         if (tmo > TCP_TIMEWAIT_LEN) {
4792                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4793                                         } else if (th->fin || sock_owned_by_user(sk)) {
4794                                                 /* Bad case. We could lose such FIN otherwise.
4795                                                  * It is not a big problem, but it looks confusing
4796                                                  * and not so rare event. We still can lose it now,
4797                                                  * if it spins in bh_lock_sock(), but it is really
4798                                                  * marginal case.
4799                                                  */
4800                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4801                                         } else {
4802                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4803                                                 goto discard;
4804                                         }
4805                                 }
4806                         }
4807                         break;
4808
4809                 case TCP_CLOSING:
4810                         if (tp->snd_una == tp->write_seq) {
4811                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4812                                 goto discard;
4813                         }
4814                         break;
4815
4816                 case TCP_LAST_ACK:
4817                         if (tp->snd_una == tp->write_seq) {
4818                                 tcp_update_metrics(sk);
4819                                 tcp_done(sk);
4820                                 goto discard;
4821                         }
4822                         break;
4823                 }
4824         } else
4825                 goto discard;
4826
4827         /* step 6: check the URG bit */
4828         tcp_urg(sk, skb, th);
4829
4830         /* step 7: process the segment text */
4831         switch (sk->sk_state) {
4832         case TCP_CLOSE_WAIT:
4833         case TCP_CLOSING:
4834         case TCP_LAST_ACK:
4835                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4836                         break;
4837         case TCP_FIN_WAIT1:
4838         case TCP_FIN_WAIT2:
4839                 /* RFC 793 says to queue data in these states,
4840                  * RFC 1122 says we MUST send a reset.
4841                  * BSD 4.4 also does reset.
4842                  */
4843                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4844                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4845                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4846                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4847                                 tcp_reset(sk);
4848                                 return 1;
4849                         }
4850                 }
4851                 /* Fall through */
4852         case TCP_ESTABLISHED:
4853                 tcp_data_queue(sk, skb);
4854                 queued = 1;
4855                 break;
4856         }
4857
4858         /* tcp_data could move socket to TIME-WAIT */
4859         if (sk->sk_state != TCP_CLOSE) {
4860                 tcp_data_snd_check(sk, tp);
4861                 tcp_ack_snd_check(sk);
4862         }
4863
4864         if (!queued) {
4865 discard:
4866                 __kfree_skb(skb);
4867         }
4868         return 0;
4869 }
4870
4871 EXPORT_SYMBOL(sysctl_tcp_ecn);
4872 EXPORT_SYMBOL(sysctl_tcp_reordering);
4873 EXPORT_SYMBOL(tcp_parse_options);
4874 EXPORT_SYMBOL(tcp_rcv_established);
4875 EXPORT_SYMBOL(tcp_rcv_state_process);
4876 EXPORT_SYMBOL(tcp_initialize_rcv_mss);