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