Pull video into release branch
[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_transport_header(skb);
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(tcp_hdr(skb)) & 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, const struct sk_buff *skb)
239 {
240         struct tcp_sock *tp = tcp_sk(sk);
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,
256                             struct sk_buff *skb)
257 {
258         struct tcp_sock *tp = tcp_sk(sk);
259
260         /* Check #1 */
261         if (tp->rcv_ssthresh < tp->window_clamp &&
262             (int)tp->rcv_ssthresh < tcp_space(sk) &&
263             !tcp_memory_pressure) {
264                 int incr;
265
266                 /* Check #2. Increase window, if skb with such overhead
267                  * will fit to rcvbuf in future.
268                  */
269                 if (tcp_win_from_space(skb->truesize) <= skb->len)
270                         incr = 2*tp->advmss;
271                 else
272                         incr = __tcp_grow_window(sk, skb);
273
274                 if (incr) {
275                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
276                         inet_csk(sk)->icsk_ack.quick |= 1;
277                 }
278         }
279 }
280
281 /* 3. Tuning rcvbuf, when connection enters established state. */
282
283 static void tcp_fixup_rcvbuf(struct sock *sk)
284 {
285         struct tcp_sock *tp = tcp_sk(sk);
286         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
287
288         /* Try to select rcvbuf so that 4 mss-sized segments
289          * will fit to window and corresponding skbs will fit to our rcvbuf.
290          * (was 3; 4 is minimum to allow fast retransmit to work.)
291          */
292         while (tcp_win_from_space(rcvmem) < tp->advmss)
293                 rcvmem += 128;
294         if (sk->sk_rcvbuf < 4 * rcvmem)
295                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
296 }
297
298 /* 4. Try to fixup all. It is made immediately after connection enters
299  *    established state.
300  */
301 static void tcp_init_buffer_space(struct sock *sk)
302 {
303         struct tcp_sock *tp = tcp_sk(sk);
304         int maxwin;
305
306         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
307                 tcp_fixup_rcvbuf(sk);
308         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
309                 tcp_fixup_sndbuf(sk);
310
311         tp->rcvq_space.space = tp->rcv_wnd;
312
313         maxwin = tcp_full_space(sk);
314
315         if (tp->window_clamp >= maxwin) {
316                 tp->window_clamp = maxwin;
317
318                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
319                         tp->window_clamp = max(maxwin -
320                                                (maxwin >> sysctl_tcp_app_win),
321                                                4 * tp->advmss);
322         }
323
324         /* Force reservation of one segment. */
325         if (sysctl_tcp_app_win &&
326             tp->window_clamp > 2 * tp->advmss &&
327             tp->window_clamp + tp->advmss > maxwin)
328                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
329
330         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
331         tp->snd_cwnd_stamp = tcp_time_stamp;
332 }
333
334 /* 5. Recalculate window clamp after socket hit its memory bounds. */
335 static void tcp_clamp_window(struct sock *sk)
336 {
337         struct tcp_sock *tp = tcp_sk(sk);
338         struct inet_connection_sock *icsk = inet_csk(sk);
339
340         icsk->icsk_ack.quick = 0;
341
342         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
343             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
344             !tcp_memory_pressure &&
345             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
346                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
347                                     sysctl_tcp_rmem[2]);
348         }
349         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
350                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
351 }
352
353
354 /* Initialize RCV_MSS value.
355  * RCV_MSS is an our guess about MSS used by the peer.
356  * We haven't any direct information about the MSS.
357  * It's better to underestimate the RCV_MSS rather than overestimate.
358  * Overestimations make us ACKing less frequently than needed.
359  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
360  */
361 void tcp_initialize_rcv_mss(struct sock *sk)
362 {
363         struct tcp_sock *tp = tcp_sk(sk);
364         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
365
366         hint = min(hint, tp->rcv_wnd/2);
367         hint = min(hint, TCP_MIN_RCVMSS);
368         hint = max(hint, TCP_MIN_MSS);
369
370         inet_csk(sk)->icsk_ack.rcv_mss = hint;
371 }
372
373 /* Receiver "autotuning" code.
374  *
375  * The algorithm for RTT estimation w/o timestamps is based on
376  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
377  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
378  *
379  * More detail on this code can be found at
380  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
381  * though this reference is out of date.  A new paper
382  * is pending.
383  */
384 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
385 {
386         u32 new_sample = tp->rcv_rtt_est.rtt;
387         long m = sample;
388
389         if (m == 0)
390                 m = 1;
391
392         if (new_sample != 0) {
393                 /* If we sample in larger samples in the non-timestamp
394                  * case, we could grossly overestimate the RTT especially
395                  * with chatty applications or bulk transfer apps which
396                  * are stalled on filesystem I/O.
397                  *
398                  * Also, since we are only going for a minimum in the
399                  * non-timestamp case, we do not smooth things out
400                  * else with timestamps disabled convergence takes too
401                  * long.
402                  */
403                 if (!win_dep) {
404                         m -= (new_sample >> 3);
405                         new_sample += m;
406                 } else if (m < new_sample)
407                         new_sample = m << 3;
408         } else {
409                 /* No previous measure. */
410                 new_sample = m << 3;
411         }
412
413         if (tp->rcv_rtt_est.rtt != new_sample)
414                 tp->rcv_rtt_est.rtt = new_sample;
415 }
416
417 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
418 {
419         if (tp->rcv_rtt_est.time == 0)
420                 goto new_measure;
421         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
422                 return;
423         tcp_rcv_rtt_update(tp,
424                            jiffies - tp->rcv_rtt_est.time,
425                            1);
426
427 new_measure:
428         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
429         tp->rcv_rtt_est.time = tcp_time_stamp;
430 }
431
432 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
433 {
434         struct tcp_sock *tp = tcp_sk(sk);
435         if (tp->rx_opt.rcv_tsecr &&
436             (TCP_SKB_CB(skb)->end_seq -
437              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
438                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
439 }
440
441 /*
442  * This function should be called every time data is copied to user space.
443  * It calculates the appropriate TCP receive buffer space.
444  */
445 void tcp_rcv_space_adjust(struct sock *sk)
446 {
447         struct tcp_sock *tp = tcp_sk(sk);
448         int time;
449         int space;
450
451         if (tp->rcvq_space.time == 0)
452                 goto new_measure;
453
454         time = tcp_time_stamp - tp->rcvq_space.time;
455         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
456             tp->rcv_rtt_est.rtt == 0)
457                 return;
458
459         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
460
461         space = max(tp->rcvq_space.space, space);
462
463         if (tp->rcvq_space.space != space) {
464                 int rcvmem;
465
466                 tp->rcvq_space.space = space;
467
468                 if (sysctl_tcp_moderate_rcvbuf &&
469                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
470                         int new_clamp = space;
471
472                         /* Receive space grows, normalize in order to
473                          * take into account packet headers and sk_buff
474                          * structure overhead.
475                          */
476                         space /= tp->advmss;
477                         if (!space)
478                                 space = 1;
479                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
480                                   16 + sizeof(struct sk_buff));
481                         while (tcp_win_from_space(rcvmem) < tp->advmss)
482                                 rcvmem += 128;
483                         space *= rcvmem;
484                         space = min(space, sysctl_tcp_rmem[2]);
485                         if (space > sk->sk_rcvbuf) {
486                                 sk->sk_rcvbuf = space;
487
488                                 /* Make the window clamp follow along.  */
489                                 tp->window_clamp = new_clamp;
490                         }
491                 }
492         }
493
494 new_measure:
495         tp->rcvq_space.seq = tp->copied_seq;
496         tp->rcvq_space.time = tcp_time_stamp;
497 }
498
499 /* There is something which you must keep in mind when you analyze the
500  * behavior of the tp->ato delayed ack timeout interval.  When a
501  * connection starts up, we want to ack as quickly as possible.  The
502  * problem is that "good" TCP's do slow start at the beginning of data
503  * transmission.  The means that until we send the first few ACK's the
504  * sender will sit on his end and only queue most of his data, because
505  * he can only send snd_cwnd unacked packets at any given time.  For
506  * each ACK we send, he increments snd_cwnd and transmits more of his
507  * queue.  -DaveM
508  */
509 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
510 {
511         struct tcp_sock *tp = tcp_sk(sk);
512         struct inet_connection_sock *icsk = inet_csk(sk);
513         u32 now;
514
515         inet_csk_schedule_ack(sk);
516
517         tcp_measure_rcv_mss(sk, skb);
518
519         tcp_rcv_rtt_measure(tp);
520
521         now = tcp_time_stamp;
522
523         if (!icsk->icsk_ack.ato) {
524                 /* The _first_ data packet received, initialize
525                  * delayed ACK engine.
526                  */
527                 tcp_incr_quickack(sk);
528                 icsk->icsk_ack.ato = TCP_ATO_MIN;
529         } else {
530                 int m = now - icsk->icsk_ack.lrcvtime;
531
532                 if (m <= TCP_ATO_MIN/2) {
533                         /* The fastest case is the first. */
534                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
535                 } else if (m < icsk->icsk_ack.ato) {
536                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
537                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
538                                 icsk->icsk_ack.ato = icsk->icsk_rto;
539                 } else if (m > icsk->icsk_rto) {
540                         /* Too long gap. Apparently sender failed to
541                          * restart window, so that we send ACKs quickly.
542                          */
543                         tcp_incr_quickack(sk);
544                         sk_stream_mem_reclaim(sk);
545                 }
546         }
547         icsk->icsk_ack.lrcvtime = now;
548
549         TCP_ECN_check_ce(tp, skb);
550
551         if (skb->len >= 128)
552                 tcp_grow_window(sk, skb);
553 }
554
555 /* Called to compute a smoothed rtt estimate. The data fed to this
556  * routine either comes from timestamps, or from segments that were
557  * known _not_ to have been retransmitted [see Karn/Partridge
558  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
559  * piece by Van Jacobson.
560  * NOTE: the next three routines used to be one big routine.
561  * To save cycles in the RFC 1323 implementation it was better to break
562  * it up into three procedures. -- erics
563  */
564 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
565 {
566         struct tcp_sock *tp = tcp_sk(sk);
567         long m = mrtt; /* RTT */
568
569         /*      The following amusing code comes from Jacobson's
570          *      article in SIGCOMM '88.  Note that rtt and mdev
571          *      are scaled versions of rtt and mean deviation.
572          *      This is designed to be as fast as possible
573          *      m stands for "measurement".
574          *
575          *      On a 1990 paper the rto value is changed to:
576          *      RTO = rtt + 4 * mdev
577          *
578          * Funny. This algorithm seems to be very broken.
579          * These formulae increase RTO, when it should be decreased, increase
580          * too slowly, when it should be increased quickly, decrease too quickly
581          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
582          * does not matter how to _calculate_ it. Seems, it was trap
583          * that VJ failed to avoid. 8)
584          */
585         if (m == 0)
586                 m = 1;
587         if (tp->srtt != 0) {
588                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
589                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
590                 if (m < 0) {
591                         m = -m;         /* m is now abs(error) */
592                         m -= (tp->mdev >> 2);   /* similar update on mdev */
593                         /* This is similar to one of Eifel findings.
594                          * Eifel blocks mdev updates when rtt decreases.
595                          * This solution is a bit different: we use finer gain
596                          * for mdev in this case (alpha*beta).
597                          * Like Eifel it also prevents growth of rto,
598                          * but also it limits too fast rto decreases,
599                          * happening in pure Eifel.
600                          */
601                         if (m > 0)
602                                 m >>= 3;
603                 } else {
604                         m -= (tp->mdev >> 2);   /* similar update on mdev */
605                 }
606                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
607                 if (tp->mdev > tp->mdev_max) {
608                         tp->mdev_max = tp->mdev;
609                         if (tp->mdev_max > tp->rttvar)
610                                 tp->rttvar = tp->mdev_max;
611                 }
612                 if (after(tp->snd_una, tp->rtt_seq)) {
613                         if (tp->mdev_max < tp->rttvar)
614                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
615                         tp->rtt_seq = tp->snd_nxt;
616                         tp->mdev_max = TCP_RTO_MIN;
617                 }
618         } else {
619                 /* no previous measure. */
620                 tp->srtt = m<<3;        /* take the measured time to be rtt */
621                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
622                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
623                 tp->rtt_seq = tp->snd_nxt;
624         }
625 }
626
627 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
628  * routine referred to above.
629  */
630 static inline void tcp_set_rto(struct sock *sk)
631 {
632         const struct tcp_sock *tp = tcp_sk(sk);
633         /* Old crap is replaced with new one. 8)
634          *
635          * More seriously:
636          * 1. If rtt variance happened to be less 50msec, it is hallucination.
637          *    It cannot be less due to utterly erratic ACK generation made
638          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
639          *    to do with delayed acks, because at cwnd>2 true delack timeout
640          *    is invisible. Actually, Linux-2.4 also generates erratic
641          *    ACKs in some circumstances.
642          */
643         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
644
645         /* 2. Fixups made earlier cannot be right.
646          *    If we do not estimate RTO correctly without them,
647          *    all the algo is pure shit and should be replaced
648          *    with correct one. It is exactly, which we pretend to do.
649          */
650 }
651
652 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
653  * guarantees that rto is higher.
654  */
655 static inline void tcp_bound_rto(struct sock *sk)
656 {
657         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
658                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
659 }
660
661 /* Save metrics learned by this TCP session.
662    This function is called only, when TCP finishes successfully
663    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
664  */
665 void tcp_update_metrics(struct sock *sk)
666 {
667         struct tcp_sock *tp = tcp_sk(sk);
668         struct dst_entry *dst = __sk_dst_get(sk);
669
670         if (sysctl_tcp_nometrics_save)
671                 return;
672
673         dst_confirm(dst);
674
675         if (dst && (dst->flags&DST_HOST)) {
676                 const struct inet_connection_sock *icsk = inet_csk(sk);
677                 int m;
678
679                 if (icsk->icsk_backoff || !tp->srtt) {
680                         /* This session failed to estimate rtt. Why?
681                          * Probably, no packets returned in time.
682                          * Reset our results.
683                          */
684                         if (!(dst_metric_locked(dst, RTAX_RTT)))
685                                 dst->metrics[RTAX_RTT-1] = 0;
686                         return;
687                 }
688
689                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
690
691                 /* If newly calculated rtt larger than stored one,
692                  * store new one. Otherwise, use EWMA. Remember,
693                  * rtt overestimation is always better than underestimation.
694                  */
695                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
696                         if (m <= 0)
697                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
698                         else
699                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
700                 }
701
702                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
703                         if (m < 0)
704                                 m = -m;
705
706                         /* Scale deviation to rttvar fixed point */
707                         m >>= 1;
708                         if (m < tp->mdev)
709                                 m = tp->mdev;
710
711                         if (m >= dst_metric(dst, RTAX_RTTVAR))
712                                 dst->metrics[RTAX_RTTVAR-1] = m;
713                         else
714                                 dst->metrics[RTAX_RTTVAR-1] -=
715                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
716                 }
717
718                 if (tp->snd_ssthresh >= 0xFFFF) {
719                         /* Slow start still did not finish. */
720                         if (dst_metric(dst, RTAX_SSTHRESH) &&
721                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
722                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
723                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
724                         if (!dst_metric_locked(dst, RTAX_CWND) &&
725                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
726                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
727                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
728                            icsk->icsk_ca_state == TCP_CA_Open) {
729                         /* Cong. avoidance phase, cwnd is reliable. */
730                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
731                                 dst->metrics[RTAX_SSTHRESH-1] =
732                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
733                         if (!dst_metric_locked(dst, RTAX_CWND))
734                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
735                 } else {
736                         /* Else slow start did not finish, cwnd is non-sense,
737                            ssthresh may be also invalid.
738                          */
739                         if (!dst_metric_locked(dst, RTAX_CWND))
740                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
741                         if (dst->metrics[RTAX_SSTHRESH-1] &&
742                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
743                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
744                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
745                 }
746
747                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
748                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
749                             tp->reordering != sysctl_tcp_reordering)
750                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
751                 }
752         }
753 }
754
755 /* Numbers are taken from RFC2414.  */
756 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
757 {
758         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
759
760         if (!cwnd) {
761                 if (tp->mss_cache > 1460)
762                         cwnd = 2;
763                 else
764                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
765         }
766         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
767 }
768
769 /* Set slow start threshold and cwnd not falling to slow start */
770 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
771 {
772         struct tcp_sock *tp = tcp_sk(sk);
773         const struct inet_connection_sock *icsk = inet_csk(sk);
774
775         tp->prior_ssthresh = 0;
776         tp->bytes_acked = 0;
777         if (icsk->icsk_ca_state < TCP_CA_CWR) {
778                 tp->undo_marker = 0;
779                 if (set_ssthresh)
780                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
781                 tp->snd_cwnd = min(tp->snd_cwnd,
782                                    tcp_packets_in_flight(tp) + 1U);
783                 tp->snd_cwnd_cnt = 0;
784                 tp->high_seq = tp->snd_nxt;
785                 tp->snd_cwnd_stamp = tcp_time_stamp;
786                 TCP_ECN_queue_cwr(tp);
787
788                 tcp_set_ca_state(sk, TCP_CA_CWR);
789         }
790 }
791
792 /* Initialize metrics on socket. */
793
794 static void tcp_init_metrics(struct sock *sk)
795 {
796         struct tcp_sock *tp = tcp_sk(sk);
797         struct dst_entry *dst = __sk_dst_get(sk);
798
799         if (dst == NULL)
800                 goto reset;
801
802         dst_confirm(dst);
803
804         if (dst_metric_locked(dst, RTAX_CWND))
805                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
806         if (dst_metric(dst, RTAX_SSTHRESH)) {
807                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
808                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
809                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
810         }
811         if (dst_metric(dst, RTAX_REORDERING) &&
812             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
813                 tp->rx_opt.sack_ok &= ~2;
814                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
815         }
816
817         if (dst_metric(dst, RTAX_RTT) == 0)
818                 goto reset;
819
820         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
821                 goto reset;
822
823         /* Initial rtt is determined from SYN,SYN-ACK.
824          * The segment is small and rtt may appear much
825          * less than real one. Use per-dst memory
826          * to make it more realistic.
827          *
828          * A bit of theory. RTT is time passed after "normal" sized packet
829          * is sent until it is ACKed. In normal circumstances sending small
830          * packets force peer to delay ACKs and calculation is correct too.
831          * The algorithm is adaptive and, provided we follow specs, it
832          * NEVER underestimate RTT. BUT! If peer tries to make some clever
833          * tricks sort of "quick acks" for time long enough to decrease RTT
834          * to low value, and then abruptly stops to do it and starts to delay
835          * ACKs, wait for troubles.
836          */
837         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
838                 tp->srtt = dst_metric(dst, RTAX_RTT);
839                 tp->rtt_seq = tp->snd_nxt;
840         }
841         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
842                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
843                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
844         }
845         tcp_set_rto(sk);
846         tcp_bound_rto(sk);
847         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
848                 goto reset;
849         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
850         tp->snd_cwnd_stamp = tcp_time_stamp;
851         return;
852
853 reset:
854         /* Play conservative. If timestamps are not
855          * supported, TCP will fail to recalculate correct
856          * rtt, if initial rto is too small. FORGET ALL AND RESET!
857          */
858         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
859                 tp->srtt = 0;
860                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
861                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
862         }
863 }
864
865 static void tcp_update_reordering(struct sock *sk, const int metric,
866                                   const int ts)
867 {
868         struct tcp_sock *tp = tcp_sk(sk);
869         if (metric > tp->reordering) {
870                 tp->reordering = min(TCP_MAX_REORDERING, metric);
871
872                 /* This exciting event is worth to be remembered. 8) */
873                 if (ts)
874                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
875                 else if (IsReno(tp))
876                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
877                 else if (IsFack(tp))
878                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
879                 else
880                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
881 #if FASTRETRANS_DEBUG > 1
882                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
883                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
884                        tp->reordering,
885                        tp->fackets_out,
886                        tp->sacked_out,
887                        tp->undo_marker ? tp->undo_retrans : 0);
888 #endif
889                 /* Disable FACK yet. */
890                 tp->rx_opt.sack_ok &= ~2;
891         }
892 }
893
894 /* This procedure tags the retransmission queue when SACKs arrive.
895  *
896  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
897  * Packets in queue with these bits set are counted in variables
898  * sacked_out, retrans_out and lost_out, correspondingly.
899  *
900  * Valid combinations are:
901  * Tag  InFlight        Description
902  * 0    1               - orig segment is in flight.
903  * S    0               - nothing flies, orig reached receiver.
904  * L    0               - nothing flies, orig lost by net.
905  * R    2               - both orig and retransmit are in flight.
906  * L|R  1               - orig is lost, retransmit is in flight.
907  * S|R  1               - orig reached receiver, retrans is still in flight.
908  * (L|S|R is logically valid, it could occur when L|R is sacked,
909  *  but it is equivalent to plain S and code short-curcuits it to S.
910  *  L|S is logically invalid, it would mean -1 packet in flight 8))
911  *
912  * These 6 states form finite state machine, controlled by the following events:
913  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
914  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
915  * 3. Loss detection event of one of three flavors:
916  *      A. Scoreboard estimator decided the packet is lost.
917  *         A'. Reno "three dupacks" marks head of queue lost.
918  *         A''. Its FACK modfication, head until snd.fack is lost.
919  *      B. SACK arrives sacking data transmitted after never retransmitted
920  *         hole was sent out.
921  *      C. SACK arrives sacking SND.NXT at the moment, when the
922  *         segment was retransmitted.
923  * 4. D-SACK added new rule: D-SACK changes any tag to S.
924  *
925  * It is pleasant to note, that state diagram turns out to be commutative,
926  * so that we are allowed not to be bothered by order of our actions,
927  * when multiple events arrive simultaneously. (see the function below).
928  *
929  * Reordering detection.
930  * --------------------
931  * Reordering metric is maximal distance, which a packet can be displaced
932  * in packet stream. With SACKs we can estimate it:
933  *
934  * 1. SACK fills old hole and the corresponding segment was not
935  *    ever retransmitted -> reordering. Alas, we cannot use it
936  *    when segment was retransmitted.
937  * 2. The last flaw is solved with D-SACK. D-SACK arrives
938  *    for retransmitted and already SACKed segment -> reordering..
939  * Both of these heuristics are not used in Loss state, when we cannot
940  * account for retransmits accurately.
941  */
942 static int
943 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
944 {
945         const struct inet_connection_sock *icsk = inet_csk(sk);
946         struct tcp_sock *tp = tcp_sk(sk);
947         unsigned char *ptr = (skb_transport_header(ack_skb) +
948                               TCP_SKB_CB(ack_skb)->sacked);
949         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
950         struct sk_buff *cached_skb;
951         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
952         int reord = tp->packets_out;
953         int prior_fackets;
954         u32 lost_retrans = 0;
955         int flag = 0;
956         int found_dup_sack = 0;
957         int cached_fack_count;
958         int i;
959         int first_sack_index;
960
961         if (!tp->sacked_out)
962                 tp->fackets_out = 0;
963         prior_fackets = tp->fackets_out;
964
965         /* Check for D-SACK. */
966         if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
967                 found_dup_sack = 1;
968                 tp->rx_opt.sack_ok |= 4;
969                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
970         } else if (num_sacks > 1 &&
971                         !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
972                         !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
973                 found_dup_sack = 1;
974                 tp->rx_opt.sack_ok |= 4;
975                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
976         }
977
978         /* D-SACK for already forgotten data...
979          * Do dumb counting. */
980         if (found_dup_sack &&
981                         !after(ntohl(sp[0].end_seq), prior_snd_una) &&
982                         after(ntohl(sp[0].end_seq), tp->undo_marker))
983                 tp->undo_retrans--;
984
985         /* Eliminate too old ACKs, but take into
986          * account more or less fresh ones, they can
987          * contain valid SACK info.
988          */
989         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
990                 return 0;
991
992         /* SACK fastpath:
993          * if the only SACK change is the increase of the end_seq of
994          * the first block then only apply that SACK block
995          * and use retrans queue hinting otherwise slowpath */
996         flag = 1;
997         for (i = 0; i < num_sacks; i++) {
998                 __be32 start_seq = sp[i].start_seq;
999                 __be32 end_seq = sp[i].end_seq;
1000
1001                 if (i == 0) {
1002                         if (tp->recv_sack_cache[i].start_seq != start_seq)
1003                                 flag = 0;
1004                 } else {
1005                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1006                             (tp->recv_sack_cache[i].end_seq != end_seq))
1007                                 flag = 0;
1008                 }
1009                 tp->recv_sack_cache[i].start_seq = start_seq;
1010                 tp->recv_sack_cache[i].end_seq = end_seq;
1011         }
1012         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1013         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1014                 tp->recv_sack_cache[i].start_seq = 0;
1015                 tp->recv_sack_cache[i].end_seq = 0;
1016         }
1017
1018         first_sack_index = 0;
1019         if (flag)
1020                 num_sacks = 1;
1021         else {
1022                 int j;
1023                 tp->fastpath_skb_hint = NULL;
1024
1025                 /* order SACK blocks to allow in order walk of the retrans queue */
1026                 for (i = num_sacks-1; i > 0; i--) {
1027                         for (j = 0; j < i; j++){
1028                                 if (after(ntohl(sp[j].start_seq),
1029                                           ntohl(sp[j+1].start_seq))){
1030                                         struct tcp_sack_block_wire tmp;
1031
1032                                         tmp = sp[j];
1033                                         sp[j] = sp[j+1];
1034                                         sp[j+1] = tmp;
1035
1036                                         /* Track where the first SACK block goes to */
1037                                         if (j == first_sack_index)
1038                                                 first_sack_index = j+1;
1039                                 }
1040
1041                         }
1042                 }
1043         }
1044
1045         /* clear flag as used for different purpose in following code */
1046         flag = 0;
1047
1048         /* Use SACK fastpath hint if valid */
1049         cached_skb = tp->fastpath_skb_hint;
1050         cached_fack_count = tp->fastpath_cnt_hint;
1051         if (!cached_skb) {
1052                 cached_skb = tcp_write_queue_head(sk);
1053                 cached_fack_count = 0;
1054         }
1055
1056         for (i=0; i<num_sacks; i++, sp++) {
1057                 struct sk_buff *skb;
1058                 __u32 start_seq = ntohl(sp->start_seq);
1059                 __u32 end_seq = ntohl(sp->end_seq);
1060                 int fack_count;
1061                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1062
1063                 skb = cached_skb;
1064                 fack_count = cached_fack_count;
1065
1066                 /* Event "B" in the comment above. */
1067                 if (after(end_seq, tp->high_seq))
1068                         flag |= FLAG_DATA_LOST;
1069
1070                 tcp_for_write_queue_from(skb, sk) {
1071                         int in_sack, pcount;
1072                         u8 sacked;
1073
1074                         if (skb == tcp_send_head(sk))
1075                                 break;
1076
1077                         cached_skb = skb;
1078                         cached_fack_count = fack_count;
1079                         if (i == first_sack_index) {
1080                                 tp->fastpath_skb_hint = skb;
1081                                 tp->fastpath_cnt_hint = fack_count;
1082                         }
1083
1084                         /* The retransmission queue is always in order, so
1085                          * we can short-circuit the walk early.
1086                          */
1087                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1088                                 break;
1089
1090                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1091                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1092
1093                         pcount = tcp_skb_pcount(skb);
1094
1095                         if (pcount > 1 && !in_sack &&
1096                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1097                                 unsigned int pkt_len;
1098
1099                                 in_sack = !after(start_seq,
1100                                                  TCP_SKB_CB(skb)->seq);
1101
1102                                 if (!in_sack)
1103                                         pkt_len = (start_seq -
1104                                                    TCP_SKB_CB(skb)->seq);
1105                                 else
1106                                         pkt_len = (end_seq -
1107                                                    TCP_SKB_CB(skb)->seq);
1108                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1109                                         break;
1110                                 pcount = tcp_skb_pcount(skb);
1111                         }
1112
1113                         fack_count += pcount;
1114
1115                         sacked = TCP_SKB_CB(skb)->sacked;
1116
1117                         /* Account D-SACK for retransmitted packet. */
1118                         if ((dup_sack && in_sack) &&
1119                             (sacked & TCPCB_RETRANS) &&
1120                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1121                                 tp->undo_retrans--;
1122
1123                         /* The frame is ACKed. */
1124                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1125                                 if (sacked&TCPCB_RETRANS) {
1126                                         if ((dup_sack && in_sack) &&
1127                                             (sacked&TCPCB_SACKED_ACKED))
1128                                                 reord = min(fack_count, reord);
1129                                 } else {
1130                                         /* If it was in a hole, we detected reordering. */
1131                                         if (fack_count < prior_fackets &&
1132                                             !(sacked&TCPCB_SACKED_ACKED))
1133                                                 reord = min(fack_count, reord);
1134                                 }
1135
1136                                 /* Nothing to do; acked frame is about to be dropped. */
1137                                 continue;
1138                         }
1139
1140                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1141                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1142                             (!lost_retrans || after(end_seq, lost_retrans)))
1143                                 lost_retrans = end_seq;
1144
1145                         if (!in_sack)
1146                                 continue;
1147
1148                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1149                                 if (sacked & TCPCB_SACKED_RETRANS) {
1150                                         /* If the segment is not tagged as lost,
1151                                          * we do not clear RETRANS, believing
1152                                          * that retransmission is still in flight.
1153                                          */
1154                                         if (sacked & TCPCB_LOST) {
1155                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1156                                                 tp->lost_out -= tcp_skb_pcount(skb);
1157                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1158
1159                                                 /* clear lost hint */
1160                                                 tp->retransmit_skb_hint = NULL;
1161                                         }
1162                                 } else {
1163                                         /* New sack for not retransmitted frame,
1164                                          * which was in hole. It is reordering.
1165                                          */
1166                                         if (!(sacked & TCPCB_RETRANS) &&
1167                                             fack_count < prior_fackets)
1168                                                 reord = min(fack_count, reord);
1169
1170                                         if (sacked & TCPCB_LOST) {
1171                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1172                                                 tp->lost_out -= tcp_skb_pcount(skb);
1173
1174                                                 /* clear lost hint */
1175                                                 tp->retransmit_skb_hint = NULL;
1176                                         }
1177                                         /* SACK enhanced F-RTO detection.
1178                                          * Set flag if and only if non-rexmitted
1179                                          * segments below frto_highmark are
1180                                          * SACKed (RFC4138; Appendix B).
1181                                          * Clearing correct due to in-order walk
1182                                          */
1183                                         if (after(end_seq, tp->frto_highmark)) {
1184                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1185                                         } else {
1186                                                 if (!(sacked & TCPCB_RETRANS))
1187                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1188                                         }
1189                                 }
1190
1191                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1192                                 flag |= FLAG_DATA_SACKED;
1193                                 tp->sacked_out += tcp_skb_pcount(skb);
1194
1195                                 if (fack_count > tp->fackets_out)
1196                                         tp->fackets_out = fack_count;
1197                         } else {
1198                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1199                                         reord = min(fack_count, reord);
1200                         }
1201
1202                         /* D-SACK. We can detect redundant retransmission
1203                          * in S|R and plain R frames and clear it.
1204                          * undo_retrans is decreased above, L|R frames
1205                          * are accounted above as well.
1206                          */
1207                         if (dup_sack &&
1208                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1209                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1210                                 tp->retrans_out -= tcp_skb_pcount(skb);
1211                                 tp->retransmit_skb_hint = NULL;
1212                         }
1213                 }
1214         }
1215
1216         /* Check for lost retransmit. This superb idea is
1217          * borrowed from "ratehalving". Event "C".
1218          * Later note: FACK people cheated me again 8),
1219          * we have to account for reordering! Ugly,
1220          * but should help.
1221          */
1222         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1223                 struct sk_buff *skb;
1224
1225                 tcp_for_write_queue(skb, sk) {
1226                         if (skb == tcp_send_head(sk))
1227                                 break;
1228                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1229                                 break;
1230                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1231                                 continue;
1232                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1233                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1234                             (IsFack(tp) ||
1235                              !before(lost_retrans,
1236                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1237                                      tp->mss_cache))) {
1238                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1239                                 tp->retrans_out -= tcp_skb_pcount(skb);
1240
1241                                 /* clear lost hint */
1242                                 tp->retransmit_skb_hint = NULL;
1243
1244                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1245                                         tp->lost_out += tcp_skb_pcount(skb);
1246                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1247                                         flag |= FLAG_DATA_SACKED;
1248                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1249                                 }
1250                         }
1251                 }
1252         }
1253
1254         tp->left_out = tp->sacked_out + tp->lost_out;
1255
1256         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1257             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1258                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1259
1260 #if FASTRETRANS_DEBUG > 0
1261         BUG_TRAP((int)tp->sacked_out >= 0);
1262         BUG_TRAP((int)tp->lost_out >= 0);
1263         BUG_TRAP((int)tp->retrans_out >= 0);
1264         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1265 #endif
1266         return flag;
1267 }
1268
1269 /* F-RTO can only be used if TCP has never retransmitted anything other than
1270  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1271  */
1272 int tcp_use_frto(struct sock *sk)
1273 {
1274         const struct tcp_sock *tp = tcp_sk(sk);
1275         struct sk_buff *skb;
1276
1277         if (!sysctl_tcp_frto)
1278                 return 0;
1279
1280         if (IsSackFrto())
1281                 return 1;
1282
1283         /* Avoid expensive walking of rexmit queue if possible */
1284         if (tp->retrans_out > 1)
1285                 return 0;
1286
1287         skb = tcp_write_queue_head(sk);
1288         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1289         tcp_for_write_queue_from(skb, sk) {
1290                 if (skb == tcp_send_head(sk))
1291                         break;
1292                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1293                         return 0;
1294                 /* Short-circuit when first non-SACKed skb has been checked */
1295                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1296                         break;
1297         }
1298         return 1;
1299 }
1300
1301 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1302  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1303  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1304  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1305  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1306  * bits are handled if the Loss state is really to be entered (in
1307  * tcp_enter_frto_loss).
1308  *
1309  * Do like tcp_enter_loss() would; when RTO expires the second time it
1310  * does:
1311  *  "Reduce ssthresh if it has not yet been made inside this window."
1312  */
1313 void tcp_enter_frto(struct sock *sk)
1314 {
1315         const struct inet_connection_sock *icsk = inet_csk(sk);
1316         struct tcp_sock *tp = tcp_sk(sk);
1317         struct sk_buff *skb;
1318
1319         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1320             tp->snd_una == tp->high_seq ||
1321             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1322              !icsk->icsk_retransmits)) {
1323                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1324                 /* Our state is too optimistic in ssthresh() call because cwnd
1325                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1326                  * recovery has not yet completed. Pattern would be this: RTO,
1327                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1328                  * up here twice).
1329                  * RFC4138 should be more specific on what to do, even though
1330                  * RTO is quite unlikely to occur after the first Cumulative ACK
1331                  * due to back-off and complexity of triggering events ...
1332                  */
1333                 if (tp->frto_counter) {
1334                         u32 stored_cwnd;
1335                         stored_cwnd = tp->snd_cwnd;
1336                         tp->snd_cwnd = 2;
1337                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1338                         tp->snd_cwnd = stored_cwnd;
1339                 } else {
1340                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1341                 }
1342                 /* ... in theory, cong.control module could do "any tricks" in
1343                  * ssthresh(), which means that ca_state, lost bits and lost_out
1344                  * counter would have to be faked before the call occurs. We
1345                  * consider that too expensive, unlikely and hacky, so modules
1346                  * using these in ssthresh() must deal these incompatibility
1347                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1348                  */
1349                 tcp_ca_event(sk, CA_EVENT_FRTO);
1350         }
1351
1352         tp->undo_marker = tp->snd_una;
1353         tp->undo_retrans = 0;
1354
1355         skb = tcp_write_queue_head(sk);
1356         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1357                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1358                 tp->retrans_out -= tcp_skb_pcount(skb);
1359         }
1360         tcp_sync_left_out(tp);
1361
1362         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1363          * The last condition is necessary at least in tp->frto_counter case.
1364          */
1365         if (IsSackFrto() && (tp->frto_counter ||
1366             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1367             after(tp->high_seq, tp->snd_una)) {
1368                 tp->frto_highmark = tp->high_seq;
1369         } else {
1370                 tp->frto_highmark = tp->snd_nxt;
1371         }
1372         tcp_set_ca_state(sk, TCP_CA_Disorder);
1373         tp->high_seq = tp->snd_nxt;
1374         tp->frto_counter = 1;
1375 }
1376
1377 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1378  * which indicates that we should follow the traditional RTO recovery,
1379  * i.e. mark everything lost and do go-back-N retransmission.
1380  */
1381 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1382 {
1383         struct tcp_sock *tp = tcp_sk(sk);
1384         struct sk_buff *skb;
1385         int cnt = 0;
1386
1387         tp->sacked_out = 0;
1388         tp->lost_out = 0;
1389         tp->fackets_out = 0;
1390         tp->retrans_out = 0;
1391
1392         tcp_for_write_queue(skb, sk) {
1393                 if (skb == tcp_send_head(sk))
1394                         break;
1395                 cnt += tcp_skb_pcount(skb);
1396                 /*
1397                  * Count the retransmission made on RTO correctly (only when
1398                  * waiting for the first ACK and did not get it)...
1399                  */
1400                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1401                         /* For some reason this R-bit might get cleared? */
1402                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1403                                 tp->retrans_out += tcp_skb_pcount(skb);
1404                         /* ...enter this if branch just for the first segment */
1405                         flag |= FLAG_DATA_ACKED;
1406                 } else {
1407                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1408                 }
1409                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1410
1411                         /* Do not mark those segments lost that were
1412                          * forward transmitted after RTO
1413                          */
1414                         if (!after(TCP_SKB_CB(skb)->end_seq,
1415                                    tp->frto_highmark)) {
1416                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1417                                 tp->lost_out += tcp_skb_pcount(skb);
1418                         }
1419                 } else {
1420                         tp->sacked_out += tcp_skb_pcount(skb);
1421                         tp->fackets_out = cnt;
1422                 }
1423         }
1424         tcp_sync_left_out(tp);
1425
1426         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1427         tp->snd_cwnd_cnt = 0;
1428         tp->snd_cwnd_stamp = tcp_time_stamp;
1429         tp->undo_marker = 0;
1430         tp->frto_counter = 0;
1431
1432         tp->reordering = min_t(unsigned int, tp->reordering,
1433                                              sysctl_tcp_reordering);
1434         tcp_set_ca_state(sk, TCP_CA_Loss);
1435         tp->high_seq = tp->frto_highmark;
1436         TCP_ECN_queue_cwr(tp);
1437
1438         clear_all_retrans_hints(tp);
1439 }
1440
1441 void tcp_clear_retrans(struct tcp_sock *tp)
1442 {
1443         tp->left_out = 0;
1444         tp->retrans_out = 0;
1445
1446         tp->fackets_out = 0;
1447         tp->sacked_out = 0;
1448         tp->lost_out = 0;
1449
1450         tp->undo_marker = 0;
1451         tp->undo_retrans = 0;
1452 }
1453
1454 /* Enter Loss state. If "how" is not zero, forget all SACK information
1455  * and reset tags completely, otherwise preserve SACKs. If receiver
1456  * dropped its ofo queue, we will know this due to reneging detection.
1457  */
1458 void tcp_enter_loss(struct sock *sk, int how)
1459 {
1460         const struct inet_connection_sock *icsk = inet_csk(sk);
1461         struct tcp_sock *tp = tcp_sk(sk);
1462         struct sk_buff *skb;
1463         int cnt = 0;
1464
1465         /* Reduce ssthresh if it has not yet been made inside this window. */
1466         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1467             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1468                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1469                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1470                 tcp_ca_event(sk, CA_EVENT_LOSS);
1471         }
1472         tp->snd_cwnd       = 1;
1473         tp->snd_cwnd_cnt   = 0;
1474         tp->snd_cwnd_stamp = tcp_time_stamp;
1475
1476         tp->bytes_acked = 0;
1477         tcp_clear_retrans(tp);
1478
1479         /* Push undo marker, if it was plain RTO and nothing
1480          * was retransmitted. */
1481         if (!how)
1482                 tp->undo_marker = tp->snd_una;
1483
1484         tcp_for_write_queue(skb, sk) {
1485                 if (skb == tcp_send_head(sk))
1486                         break;
1487                 cnt += tcp_skb_pcount(skb);
1488                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1489                         tp->undo_marker = 0;
1490                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1491                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1492                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1493                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1494                         tp->lost_out += tcp_skb_pcount(skb);
1495                 } else {
1496                         tp->sacked_out += tcp_skb_pcount(skb);
1497                         tp->fackets_out = cnt;
1498                 }
1499         }
1500         tcp_sync_left_out(tp);
1501
1502         tp->reordering = min_t(unsigned int, tp->reordering,
1503                                              sysctl_tcp_reordering);
1504         tcp_set_ca_state(sk, TCP_CA_Loss);
1505         tp->high_seq = tp->snd_nxt;
1506         TCP_ECN_queue_cwr(tp);
1507         /* Abort FRTO algorithm if one is in progress */
1508         tp->frto_counter = 0;
1509
1510         clear_all_retrans_hints(tp);
1511 }
1512
1513 static int tcp_check_sack_reneging(struct sock *sk)
1514 {
1515         struct sk_buff *skb;
1516
1517         /* If ACK arrived pointing to a remembered SACK,
1518          * it means that our remembered SACKs do not reflect
1519          * real state of receiver i.e.
1520          * receiver _host_ is heavily congested (or buggy).
1521          * Do processing similar to RTO timeout.
1522          */
1523         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1524             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1525                 struct inet_connection_sock *icsk = inet_csk(sk);
1526                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1527
1528                 tcp_enter_loss(sk, 1);
1529                 icsk->icsk_retransmits++;
1530                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1531                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1532                                           icsk->icsk_rto, TCP_RTO_MAX);
1533                 return 1;
1534         }
1535         return 0;
1536 }
1537
1538 static inline int tcp_fackets_out(struct tcp_sock *tp)
1539 {
1540         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1541 }
1542
1543 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1544 {
1545         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1546 }
1547
1548 static inline int tcp_head_timedout(struct sock *sk)
1549 {
1550         struct tcp_sock *tp = tcp_sk(sk);
1551
1552         return tp->packets_out &&
1553                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1554 }
1555
1556 /* Linux NewReno/SACK/FACK/ECN state machine.
1557  * --------------------------------------
1558  *
1559  * "Open"       Normal state, no dubious events, fast path.
1560  * "Disorder"   In all the respects it is "Open",
1561  *              but requires a bit more attention. It is entered when
1562  *              we see some SACKs or dupacks. It is split of "Open"
1563  *              mainly to move some processing from fast path to slow one.
1564  * "CWR"        CWND was reduced due to some Congestion Notification event.
1565  *              It can be ECN, ICMP source quench, local device congestion.
1566  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1567  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1568  *
1569  * tcp_fastretrans_alert() is entered:
1570  * - each incoming ACK, if state is not "Open"
1571  * - when arrived ACK is unusual, namely:
1572  *      * SACK
1573  *      * Duplicate ACK.
1574  *      * ECN ECE.
1575  *
1576  * Counting packets in flight is pretty simple.
1577  *
1578  *      in_flight = packets_out - left_out + retrans_out
1579  *
1580  *      packets_out is SND.NXT-SND.UNA counted in packets.
1581  *
1582  *      retrans_out is number of retransmitted segments.
1583  *
1584  *      left_out is number of segments left network, but not ACKed yet.
1585  *
1586  *              left_out = sacked_out + lost_out
1587  *
1588  *     sacked_out: Packets, which arrived to receiver out of order
1589  *                 and hence not ACKed. With SACKs this number is simply
1590  *                 amount of SACKed data. Even without SACKs
1591  *                 it is easy to give pretty reliable estimate of this number,
1592  *                 counting duplicate ACKs.
1593  *
1594  *       lost_out: Packets lost by network. TCP has no explicit
1595  *                 "loss notification" feedback from network (for now).
1596  *                 It means that this number can be only _guessed_.
1597  *                 Actually, it is the heuristics to predict lossage that
1598  *                 distinguishes different algorithms.
1599  *
1600  *      F.e. after RTO, when all the queue is considered as lost,
1601  *      lost_out = packets_out and in_flight = retrans_out.
1602  *
1603  *              Essentially, we have now two algorithms counting
1604  *              lost packets.
1605  *
1606  *              FACK: It is the simplest heuristics. As soon as we decided
1607  *              that something is lost, we decide that _all_ not SACKed
1608  *              packets until the most forward SACK are lost. I.e.
1609  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1610  *              It is absolutely correct estimate, if network does not reorder
1611  *              packets. And it loses any connection to reality when reordering
1612  *              takes place. We use FACK by default until reordering
1613  *              is suspected on the path to this destination.
1614  *
1615  *              NewReno: when Recovery is entered, we assume that one segment
1616  *              is lost (classic Reno). While we are in Recovery and
1617  *              a partial ACK arrives, we assume that one more packet
1618  *              is lost (NewReno). This heuristics are the same in NewReno
1619  *              and SACK.
1620  *
1621  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1622  *  deflation etc. CWND is real congestion window, never inflated, changes
1623  *  only according to classic VJ rules.
1624  *
1625  * Really tricky (and requiring careful tuning) part of algorithm
1626  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1627  * The first determines the moment _when_ we should reduce CWND and,
1628  * hence, slow down forward transmission. In fact, it determines the moment
1629  * when we decide that hole is caused by loss, rather than by a reorder.
1630  *
1631  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1632  * holes, caused by lost packets.
1633  *
1634  * And the most logically complicated part of algorithm is undo
1635  * heuristics. We detect false retransmits due to both too early
1636  * fast retransmit (reordering) and underestimated RTO, analyzing
1637  * timestamps and D-SACKs. When we detect that some segments were
1638  * retransmitted by mistake and CWND reduction was wrong, we undo
1639  * window reduction and abort recovery phase. This logic is hidden
1640  * inside several functions named tcp_try_undo_<something>.
1641  */
1642
1643 /* This function decides, when we should leave Disordered state
1644  * and enter Recovery phase, reducing congestion window.
1645  *
1646  * Main question: may we further continue forward transmission
1647  * with the same cwnd?
1648  */
1649 static int tcp_time_to_recover(struct sock *sk)
1650 {
1651         struct tcp_sock *tp = tcp_sk(sk);
1652         __u32 packets_out;
1653
1654         /* Do not perform any recovery during FRTO algorithm */
1655         if (tp->frto_counter)
1656                 return 0;
1657
1658         /* Trick#1: The loss is proven. */
1659         if (tp->lost_out)
1660                 return 1;
1661
1662         /* Not-A-Trick#2 : Classic rule... */
1663         if (tcp_fackets_out(tp) > tp->reordering)
1664                 return 1;
1665
1666         /* Trick#3 : when we use RFC2988 timer restart, fast
1667          * retransmit can be triggered by timeout of queue head.
1668          */
1669         if (tcp_head_timedout(sk))
1670                 return 1;
1671
1672         /* Trick#4: It is still not OK... But will it be useful to delay
1673          * recovery more?
1674          */
1675         packets_out = tp->packets_out;
1676         if (packets_out <= tp->reordering &&
1677             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1678             !tcp_may_send_now(sk)) {
1679                 /* We have nothing to send. This connection is limited
1680                  * either by receiver window or by application.
1681                  */
1682                 return 1;
1683         }
1684
1685         return 0;
1686 }
1687
1688 /* If we receive more dupacks than we expected counting segments
1689  * in assumption of absent reordering, interpret this as reordering.
1690  * The only another reason could be bug in receiver TCP.
1691  */
1692 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1693 {
1694         struct tcp_sock *tp = tcp_sk(sk);
1695         u32 holes;
1696
1697         holes = max(tp->lost_out, 1U);
1698         holes = min(holes, tp->packets_out);
1699
1700         if ((tp->sacked_out + holes) > tp->packets_out) {
1701                 tp->sacked_out = tp->packets_out - holes;
1702                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1703         }
1704 }
1705
1706 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1707
1708 static void tcp_add_reno_sack(struct sock *sk)
1709 {
1710         struct tcp_sock *tp = tcp_sk(sk);
1711         tp->sacked_out++;
1712         tcp_check_reno_reordering(sk, 0);
1713         tcp_sync_left_out(tp);
1714 }
1715
1716 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1717
1718 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1719 {
1720         struct tcp_sock *tp = tcp_sk(sk);
1721
1722         if (acked > 0) {
1723                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1724                 if (acked-1 >= tp->sacked_out)
1725                         tp->sacked_out = 0;
1726                 else
1727                         tp->sacked_out -= acked-1;
1728         }
1729         tcp_check_reno_reordering(sk, acked);
1730         tcp_sync_left_out(tp);
1731 }
1732
1733 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1734 {
1735         tp->sacked_out = 0;
1736         tp->left_out = tp->lost_out;
1737 }
1738
1739 /* Mark head of queue up as lost. */
1740 static void tcp_mark_head_lost(struct sock *sk,
1741                                int packets, u32 high_seq)
1742 {
1743         struct tcp_sock *tp = tcp_sk(sk);
1744         struct sk_buff *skb;
1745         int cnt;
1746
1747         BUG_TRAP(packets <= tp->packets_out);
1748         if (tp->lost_skb_hint) {
1749                 skb = tp->lost_skb_hint;
1750                 cnt = tp->lost_cnt_hint;
1751         } else {
1752                 skb = tcp_write_queue_head(sk);
1753                 cnt = 0;
1754         }
1755
1756         tcp_for_write_queue_from(skb, sk) {
1757                 if (skb == tcp_send_head(sk))
1758                         break;
1759                 /* TODO: do this better */
1760                 /* this is not the most efficient way to do this... */
1761                 tp->lost_skb_hint = skb;
1762                 tp->lost_cnt_hint = cnt;
1763                 cnt += tcp_skb_pcount(skb);
1764                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1765                         break;
1766                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1767                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1768                         tp->lost_out += tcp_skb_pcount(skb);
1769
1770                         /* clear xmit_retransmit_queue hints
1771                          *  if this is beyond hint */
1772                         if (tp->retransmit_skb_hint != NULL &&
1773                             before(TCP_SKB_CB(skb)->seq,
1774                                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1775                                 tp->retransmit_skb_hint = NULL;
1776
1777                 }
1778         }
1779         tcp_sync_left_out(tp);
1780 }
1781
1782 /* Account newly detected lost packet(s) */
1783
1784 static void tcp_update_scoreboard(struct sock *sk)
1785 {
1786         struct tcp_sock *tp = tcp_sk(sk);
1787
1788         if (IsFack(tp)) {
1789                 int lost = tp->fackets_out - tp->reordering;
1790                 if (lost <= 0)
1791                         lost = 1;
1792                 tcp_mark_head_lost(sk, lost, tp->high_seq);
1793         } else {
1794                 tcp_mark_head_lost(sk, 1, tp->high_seq);
1795         }
1796
1797         /* New heuristics: it is possible only after we switched
1798          * to restart timer each time when something is ACKed.
1799          * Hence, we can detect timed out packets during fast
1800          * retransmit without falling to slow start.
1801          */
1802         if (!IsReno(tp) && tcp_head_timedout(sk)) {
1803                 struct sk_buff *skb;
1804
1805                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1806                         : tcp_write_queue_head(sk);
1807
1808                 tcp_for_write_queue_from(skb, sk) {
1809                         if (skb == tcp_send_head(sk))
1810                                 break;
1811                         if (!tcp_skb_timedout(sk, skb))
1812                                 break;
1813
1814                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1815                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1816                                 tp->lost_out += tcp_skb_pcount(skb);
1817
1818                                 /* clear xmit_retrans hint */
1819                                 if (tp->retransmit_skb_hint &&
1820                                     before(TCP_SKB_CB(skb)->seq,
1821                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1822
1823                                         tp->retransmit_skb_hint = NULL;
1824                         }
1825                 }
1826
1827                 tp->scoreboard_skb_hint = skb;
1828
1829                 tcp_sync_left_out(tp);
1830         }
1831 }
1832
1833 /* CWND moderation, preventing bursts due to too big ACKs
1834  * in dubious situations.
1835  */
1836 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1837 {
1838         tp->snd_cwnd = min(tp->snd_cwnd,
1839                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1840         tp->snd_cwnd_stamp = tcp_time_stamp;
1841 }
1842
1843 /* Lower bound on congestion window is slow start threshold
1844  * unless congestion avoidance choice decides to overide it.
1845  */
1846 static inline u32 tcp_cwnd_min(const struct sock *sk)
1847 {
1848         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1849
1850         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1851 }
1852
1853 /* Decrease cwnd each second ack. */
1854 static void tcp_cwnd_down(struct sock *sk)
1855 {
1856         struct tcp_sock *tp = tcp_sk(sk);
1857         int decr = tp->snd_cwnd_cnt + 1;
1858
1859         tp->snd_cwnd_cnt = decr&1;
1860         decr >>= 1;
1861
1862         if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1863                 tp->snd_cwnd -= decr;
1864
1865         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1866         tp->snd_cwnd_stamp = tcp_time_stamp;
1867 }
1868
1869 /* Nothing was retransmitted or returned timestamp is less
1870  * than timestamp of the first retransmission.
1871  */
1872 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1873 {
1874         return !tp->retrans_stamp ||
1875                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1876                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1877 }
1878
1879 /* Undo procedures. */
1880
1881 #if FASTRETRANS_DEBUG > 1
1882 static void DBGUNDO(struct sock *sk, const char *msg)
1883 {
1884         struct tcp_sock *tp = tcp_sk(sk);
1885         struct inet_sock *inet = inet_sk(sk);
1886
1887         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1888                msg,
1889                NIPQUAD(inet->daddr), ntohs(inet->dport),
1890                tp->snd_cwnd, tp->left_out,
1891                tp->snd_ssthresh, tp->prior_ssthresh,
1892                tp->packets_out);
1893 }
1894 #else
1895 #define DBGUNDO(x...) do { } while (0)
1896 #endif
1897
1898 static void tcp_undo_cwr(struct sock *sk, const int undo)
1899 {
1900         struct tcp_sock *tp = tcp_sk(sk);
1901
1902         if (tp->prior_ssthresh) {
1903                 const struct inet_connection_sock *icsk = inet_csk(sk);
1904
1905                 if (icsk->icsk_ca_ops->undo_cwnd)
1906                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1907                 else
1908                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1909
1910                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1911                         tp->snd_ssthresh = tp->prior_ssthresh;
1912                         TCP_ECN_withdraw_cwr(tp);
1913                 }
1914         } else {
1915                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1916         }
1917         tcp_moderate_cwnd(tp);
1918         tp->snd_cwnd_stamp = tcp_time_stamp;
1919
1920         /* There is something screwy going on with the retrans hints after
1921            an undo */
1922         clear_all_retrans_hints(tp);
1923 }
1924
1925 static inline int tcp_may_undo(struct tcp_sock *tp)
1926 {
1927         return tp->undo_marker &&
1928                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1929 }
1930
1931 /* People celebrate: "We love our President!" */
1932 static int tcp_try_undo_recovery(struct sock *sk)
1933 {
1934         struct tcp_sock *tp = tcp_sk(sk);
1935
1936         if (tcp_may_undo(tp)) {
1937                 /* Happy end! We did not retransmit anything
1938                  * or our original transmission succeeded.
1939                  */
1940                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1941                 tcp_undo_cwr(sk, 1);
1942                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1943                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1944                 else
1945                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1946                 tp->undo_marker = 0;
1947         }
1948         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1949                 /* Hold old state until something *above* high_seq
1950                  * is ACKed. For Reno it is MUST to prevent false
1951                  * fast retransmits (RFC2582). SACK TCP is safe. */
1952                 tcp_moderate_cwnd(tp);
1953                 return 1;
1954         }
1955         tcp_set_ca_state(sk, TCP_CA_Open);
1956         return 0;
1957 }
1958
1959 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1960 static void tcp_try_undo_dsack(struct sock *sk)
1961 {
1962         struct tcp_sock *tp = tcp_sk(sk);
1963
1964         if (tp->undo_marker && !tp->undo_retrans) {
1965                 DBGUNDO(sk, "D-SACK");
1966                 tcp_undo_cwr(sk, 1);
1967                 tp->undo_marker = 0;
1968                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1969         }
1970 }
1971
1972 /* Undo during fast recovery after partial ACK. */
1973
1974 static int tcp_try_undo_partial(struct sock *sk, int acked)
1975 {
1976         struct tcp_sock *tp = tcp_sk(sk);
1977         /* Partial ACK arrived. Force Hoe's retransmit. */
1978         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1979
1980         if (tcp_may_undo(tp)) {
1981                 /* Plain luck! Hole if filled with delayed
1982                  * packet, rather than with a retransmit.
1983                  */
1984                 if (tp->retrans_out == 0)
1985                         tp->retrans_stamp = 0;
1986
1987                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1988
1989                 DBGUNDO(sk, "Hoe");
1990                 tcp_undo_cwr(sk, 0);
1991                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1992
1993                 /* So... Do not make Hoe's retransmit yet.
1994                  * If the first packet was delayed, the rest
1995                  * ones are most probably delayed as well.
1996                  */
1997                 failed = 0;
1998         }
1999         return failed;
2000 }
2001
2002 /* Undo during loss recovery after partial ACK. */
2003 static int tcp_try_undo_loss(struct sock *sk)
2004 {
2005         struct tcp_sock *tp = tcp_sk(sk);
2006
2007         if (tcp_may_undo(tp)) {
2008                 struct sk_buff *skb;
2009                 tcp_for_write_queue(skb, sk) {
2010                         if (skb == tcp_send_head(sk))
2011                                 break;
2012                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2013                 }
2014
2015                 clear_all_retrans_hints(tp);
2016
2017                 DBGUNDO(sk, "partial loss");
2018                 tp->lost_out = 0;
2019                 tp->left_out = tp->sacked_out;
2020                 tcp_undo_cwr(sk, 1);
2021                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2022                 inet_csk(sk)->icsk_retransmits = 0;
2023                 tp->undo_marker = 0;
2024                 if (!IsReno(tp))
2025                         tcp_set_ca_state(sk, TCP_CA_Open);
2026                 return 1;
2027         }
2028         return 0;
2029 }
2030
2031 static inline void tcp_complete_cwr(struct sock *sk)
2032 {
2033         struct tcp_sock *tp = tcp_sk(sk);
2034         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2035         tp->snd_cwnd_stamp = tcp_time_stamp;
2036         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2037 }
2038
2039 static void tcp_try_to_open(struct sock *sk, int flag)
2040 {
2041         struct tcp_sock *tp = tcp_sk(sk);
2042
2043         tcp_sync_left_out(tp);
2044
2045         if (tp->retrans_out == 0)
2046                 tp->retrans_stamp = 0;
2047
2048         if (flag&FLAG_ECE)
2049                 tcp_enter_cwr(sk, 1);
2050
2051         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2052                 int state = TCP_CA_Open;
2053
2054                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2055                         state = TCP_CA_Disorder;
2056
2057                 if (inet_csk(sk)->icsk_ca_state != state) {
2058                         tcp_set_ca_state(sk, state);
2059                         tp->high_seq = tp->snd_nxt;
2060                 }
2061                 tcp_moderate_cwnd(tp);
2062         } else {
2063                 tcp_cwnd_down(sk);
2064         }
2065 }
2066
2067 static void tcp_mtup_probe_failed(struct sock *sk)
2068 {
2069         struct inet_connection_sock *icsk = inet_csk(sk);
2070
2071         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2072         icsk->icsk_mtup.probe_size = 0;
2073 }
2074
2075 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2076 {
2077         struct tcp_sock *tp = tcp_sk(sk);
2078         struct inet_connection_sock *icsk = inet_csk(sk);
2079
2080         /* FIXME: breaks with very large cwnd */
2081         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2082         tp->snd_cwnd = tp->snd_cwnd *
2083                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2084                        icsk->icsk_mtup.probe_size;
2085         tp->snd_cwnd_cnt = 0;
2086         tp->snd_cwnd_stamp = tcp_time_stamp;
2087         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2088
2089         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2090         icsk->icsk_mtup.probe_size = 0;
2091         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2092 }
2093
2094
2095 /* Process an event, which can update packets-in-flight not trivially.
2096  * Main goal of this function is to calculate new estimate for left_out,
2097  * taking into account both packets sitting in receiver's buffer and
2098  * packets lost by network.
2099  *
2100  * Besides that it does CWND reduction, when packet loss is detected
2101  * and changes state of machine.
2102  *
2103  * It does _not_ decide what to send, it is made in function
2104  * tcp_xmit_retransmit_queue().
2105  */
2106 static void
2107 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
2108                       int prior_packets, int flag)
2109 {
2110         struct inet_connection_sock *icsk = inet_csk(sk);
2111         struct tcp_sock *tp = tcp_sk(sk);
2112         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
2113
2114         /* Some technical things:
2115          * 1. Reno does not count dupacks (sacked_out) automatically. */
2116         if (!tp->packets_out)
2117                 tp->sacked_out = 0;
2118         /* 2. SACK counts snd_fack in packets inaccurately. */
2119         if (tp->sacked_out == 0)
2120                 tp->fackets_out = 0;
2121
2122         /* Now state machine starts.
2123          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2124         if (flag&FLAG_ECE)
2125                 tp->prior_ssthresh = 0;
2126
2127         /* B. In all the states check for reneging SACKs. */
2128         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2129                 return;
2130
2131         /* C. Process data loss notification, provided it is valid. */
2132         if ((flag&FLAG_DATA_LOST) &&
2133             before(tp->snd_una, tp->high_seq) &&
2134             icsk->icsk_ca_state != TCP_CA_Open &&
2135             tp->fackets_out > tp->reordering) {
2136                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2137                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2138         }
2139
2140         /* D. Synchronize left_out to current state. */
2141         tcp_sync_left_out(tp);
2142
2143         /* E. Check state exit conditions. State can be terminated
2144          *    when high_seq is ACKed. */
2145         if (icsk->icsk_ca_state == TCP_CA_Open) {
2146                 BUG_TRAP(tp->retrans_out == 0);
2147                 tp->retrans_stamp = 0;
2148         } else if (!before(tp->snd_una, tp->high_seq)) {
2149                 switch (icsk->icsk_ca_state) {
2150                 case TCP_CA_Loss:
2151                         icsk->icsk_retransmits = 0;
2152                         if (tcp_try_undo_recovery(sk))
2153                                 return;
2154                         break;
2155
2156                 case TCP_CA_CWR:
2157                         /* CWR is to be held something *above* high_seq
2158                          * is ACKed for CWR bit to reach receiver. */
2159                         if (tp->snd_una != tp->high_seq) {
2160                                 tcp_complete_cwr(sk);
2161                                 tcp_set_ca_state(sk, TCP_CA_Open);
2162                         }
2163                         break;
2164
2165                 case TCP_CA_Disorder:
2166                         tcp_try_undo_dsack(sk);
2167                         if (!tp->undo_marker ||
2168                             /* For SACK case do not Open to allow to undo
2169                              * catching for all duplicate ACKs. */
2170                             IsReno(tp) || tp->snd_una != tp->high_seq) {
2171                                 tp->undo_marker = 0;
2172                                 tcp_set_ca_state(sk, TCP_CA_Open);
2173                         }
2174                         break;
2175
2176                 case TCP_CA_Recovery:
2177                         if (IsReno(tp))
2178                                 tcp_reset_reno_sack(tp);
2179                         if (tcp_try_undo_recovery(sk))
2180                                 return;
2181                         tcp_complete_cwr(sk);
2182                         break;
2183                 }
2184         }
2185
2186         /* F. Process state. */
2187         switch (icsk->icsk_ca_state) {
2188         case TCP_CA_Recovery:
2189                 if (prior_snd_una == tp->snd_una) {
2190                         if (IsReno(tp) && is_dupack)
2191                                 tcp_add_reno_sack(sk);
2192                 } else {
2193                         int acked = prior_packets - tp->packets_out;
2194                         if (IsReno(tp))
2195                                 tcp_remove_reno_sacks(sk, acked);
2196                         is_dupack = tcp_try_undo_partial(sk, acked);
2197                 }
2198                 break;
2199         case TCP_CA_Loss:
2200                 if (flag&FLAG_DATA_ACKED)
2201                         icsk->icsk_retransmits = 0;
2202                 if (!tcp_try_undo_loss(sk)) {
2203                         tcp_moderate_cwnd(tp);
2204                         tcp_xmit_retransmit_queue(sk);
2205                         return;
2206                 }
2207                 if (icsk->icsk_ca_state != TCP_CA_Open)
2208                         return;
2209                 /* Loss is undone; fall through to processing in Open state. */
2210         default:
2211                 if (IsReno(tp)) {
2212                         if (tp->snd_una != prior_snd_una)
2213                                 tcp_reset_reno_sack(tp);
2214                         if (is_dupack)
2215                                 tcp_add_reno_sack(sk);
2216                 }
2217
2218                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2219                         tcp_try_undo_dsack(sk);
2220
2221                 if (!tcp_time_to_recover(sk)) {
2222                         tcp_try_to_open(sk, flag);
2223                         return;
2224                 }
2225
2226                 /* MTU probe failure: don't reduce cwnd */
2227                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2228                     icsk->icsk_mtup.probe_size &&
2229                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2230                         tcp_mtup_probe_failed(sk);
2231                         /* Restores the reduction we did in tcp_mtup_probe() */
2232                         tp->snd_cwnd++;
2233                         tcp_simple_retransmit(sk);
2234                         return;
2235                 }
2236
2237                 /* Otherwise enter Recovery state */
2238
2239                 if (IsReno(tp))
2240                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2241                 else
2242                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2243
2244                 tp->high_seq = tp->snd_nxt;
2245                 tp->prior_ssthresh = 0;
2246                 tp->undo_marker = tp->snd_una;
2247                 tp->undo_retrans = tp->retrans_out;
2248
2249                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2250                         if (!(flag&FLAG_ECE))
2251                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2252                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2253                         TCP_ECN_queue_cwr(tp);
2254                 }
2255
2256                 tp->bytes_acked = 0;
2257                 tp->snd_cwnd_cnt = 0;
2258                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2259         }
2260
2261         if (is_dupack || tcp_head_timedout(sk))
2262                 tcp_update_scoreboard(sk);
2263         tcp_cwnd_down(sk);
2264         tcp_xmit_retransmit_queue(sk);
2265 }
2266
2267 /* Read draft-ietf-tcplw-high-performance before mucking
2268  * with this code. (Supersedes RFC1323)
2269  */
2270 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2271 {
2272         /* RTTM Rule: A TSecr value received in a segment is used to
2273          * update the averaged RTT measurement only if the segment
2274          * acknowledges some new data, i.e., only if it advances the
2275          * left edge of the send window.
2276          *
2277          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2278          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2279          *
2280          * Changed: reset backoff as soon as we see the first valid sample.
2281          * If we do not, we get strongly overestimated rto. With timestamps
2282          * samples are accepted even from very old segments: f.e., when rtt=1
2283          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2284          * answer arrives rto becomes 120 seconds! If at least one of segments
2285          * in window is lost... Voila.                          --ANK (010210)
2286          */
2287         struct tcp_sock *tp = tcp_sk(sk);
2288         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2289         tcp_rtt_estimator(sk, seq_rtt);
2290         tcp_set_rto(sk);
2291         inet_csk(sk)->icsk_backoff = 0;
2292         tcp_bound_rto(sk);
2293 }
2294
2295 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2296 {
2297         /* We don't have a timestamp. Can only use
2298          * packets that are not retransmitted to determine
2299          * rtt estimates. Also, we must not reset the
2300          * backoff for rto until we get a non-retransmitted
2301          * packet. This allows us to deal with a situation
2302          * where the network delay has increased suddenly.
2303          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2304          */
2305
2306         if (flag & FLAG_RETRANS_DATA_ACKED)
2307                 return;
2308
2309         tcp_rtt_estimator(sk, seq_rtt);
2310         tcp_set_rto(sk);
2311         inet_csk(sk)->icsk_backoff = 0;
2312         tcp_bound_rto(sk);
2313 }
2314
2315 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2316                                       const s32 seq_rtt)
2317 {
2318         const struct tcp_sock *tp = tcp_sk(sk);
2319         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2320         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2321                 tcp_ack_saw_tstamp(sk, flag);
2322         else if (seq_rtt >= 0)
2323                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2324 }
2325
2326 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2327                            u32 in_flight, int good)
2328 {
2329         const struct inet_connection_sock *icsk = inet_csk(sk);
2330         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2331         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2332 }
2333
2334 /* Restart timer after forward progress on connection.
2335  * RFC2988 recommends to restart timer to now+rto.
2336  */
2337
2338 static void tcp_ack_packets_out(struct sock *sk)
2339 {
2340         struct tcp_sock *tp = tcp_sk(sk);
2341
2342         if (!tp->packets_out) {
2343                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2344         } else {
2345                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2346         }
2347 }
2348
2349 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2350                          __u32 now, __s32 *seq_rtt)
2351 {
2352         struct tcp_sock *tp = tcp_sk(sk);
2353         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2354         __u32 seq = tp->snd_una;
2355         __u32 packets_acked;
2356         int acked = 0;
2357
2358         /* If we get here, the whole TSO packet has not been
2359          * acked.
2360          */
2361         BUG_ON(!after(scb->end_seq, seq));
2362
2363         packets_acked = tcp_skb_pcount(skb);
2364         if (tcp_trim_head(sk, skb, seq - scb->seq))
2365                 return 0;
2366         packets_acked -= tcp_skb_pcount(skb);
2367
2368         if (packets_acked) {
2369                 __u8 sacked = scb->sacked;
2370
2371                 acked |= FLAG_DATA_ACKED;
2372                 if (sacked) {
2373                         if (sacked & TCPCB_RETRANS) {
2374                                 if (sacked & TCPCB_SACKED_RETRANS)
2375                                         tp->retrans_out -= packets_acked;
2376                                 acked |= FLAG_RETRANS_DATA_ACKED;
2377                                 *seq_rtt = -1;
2378                         } else if (*seq_rtt < 0)
2379                                 *seq_rtt = now - scb->when;
2380                         if (sacked & TCPCB_SACKED_ACKED)
2381                                 tp->sacked_out -= packets_acked;
2382                         if (sacked & TCPCB_LOST)
2383                                 tp->lost_out -= packets_acked;
2384                         if (sacked & TCPCB_URG) {
2385                                 if (tp->urg_mode &&
2386                                     !before(seq, tp->snd_up))
2387                                         tp->urg_mode = 0;
2388                         }
2389                 } else if (*seq_rtt < 0)
2390                         *seq_rtt = now - scb->when;
2391
2392                 if (tp->fackets_out) {
2393                         __u32 dval = min(tp->fackets_out, packets_acked);
2394                         tp->fackets_out -= dval;
2395                 }
2396                 tp->packets_out -= packets_acked;
2397
2398                 BUG_ON(tcp_skb_pcount(skb) == 0);
2399                 BUG_ON(!before(scb->seq, scb->end_seq));
2400         }
2401
2402         return acked;
2403 }
2404
2405 /* Remove acknowledged frames from the retransmission queue. */
2406 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2407 {
2408         struct tcp_sock *tp = tcp_sk(sk);
2409         const struct inet_connection_sock *icsk = inet_csk(sk);
2410         struct sk_buff *skb;
2411         __u32 now = tcp_time_stamp;
2412         int acked = 0;
2413         int prior_packets = tp->packets_out;
2414         __s32 seq_rtt = -1;
2415         ktime_t last_ackt = net_invalid_timestamp();
2416
2417         while ((skb = tcp_write_queue_head(sk)) &&
2418                skb != tcp_send_head(sk)) {
2419                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2420                 __u8 sacked = scb->sacked;
2421
2422                 /* If our packet is before the ack sequence we can
2423                  * discard it as it's confirmed to have arrived at
2424                  * the other end.
2425                  */
2426                 if (after(scb->end_seq, tp->snd_una)) {
2427                         if (tcp_skb_pcount(skb) > 1 &&
2428                             after(tp->snd_una, scb->seq))
2429                                 acked |= tcp_tso_acked(sk, skb,
2430                                                        now, &seq_rtt);
2431                         break;
2432                 }
2433
2434                 /* Initial outgoing SYN's get put onto the write_queue
2435                  * just like anything else we transmit.  It is not
2436                  * true data, and if we misinform our callers that
2437                  * this ACK acks real data, we will erroneously exit
2438                  * connection startup slow start one packet too
2439                  * quickly.  This is severely frowned upon behavior.
2440                  */
2441                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2442                         acked |= FLAG_DATA_ACKED;
2443                 } else {
2444                         acked |= FLAG_SYN_ACKED;
2445                         tp->retrans_stamp = 0;
2446                 }
2447
2448                 /* MTU probing checks */
2449                 if (icsk->icsk_mtup.probe_size) {
2450                         if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2451                                 tcp_mtup_probe_success(sk, skb);
2452                         }
2453                 }
2454
2455                 if (sacked) {
2456                         if (sacked & TCPCB_RETRANS) {
2457                                 if (sacked & TCPCB_SACKED_RETRANS)
2458                                         tp->retrans_out -= tcp_skb_pcount(skb);
2459                                 acked |= FLAG_RETRANS_DATA_ACKED;
2460                                 seq_rtt = -1;
2461                         } else if (seq_rtt < 0) {
2462                                 seq_rtt = now - scb->when;
2463                                 last_ackt = skb->tstamp;
2464                         }
2465                         if (sacked & TCPCB_SACKED_ACKED)
2466                                 tp->sacked_out -= tcp_skb_pcount(skb);
2467                         if (sacked & TCPCB_LOST)
2468                                 tp->lost_out -= tcp_skb_pcount(skb);
2469                         if (sacked & TCPCB_URG) {
2470                                 if (tp->urg_mode &&
2471                                     !before(scb->end_seq, tp->snd_up))
2472                                         tp->urg_mode = 0;
2473                         }
2474                 } else if (seq_rtt < 0) {
2475                         seq_rtt = now - scb->when;
2476                         last_ackt = skb->tstamp;
2477                 }
2478                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2479                 tcp_packets_out_dec(tp, skb);
2480                 tcp_unlink_write_queue(skb, sk);
2481                 sk_stream_free_skb(sk, skb);
2482                 clear_all_retrans_hints(tp);
2483         }
2484
2485         if (acked&FLAG_ACKED) {
2486                 u32 pkts_acked = prior_packets - tp->packets_out;
2487                 const struct tcp_congestion_ops *ca_ops
2488                         = inet_csk(sk)->icsk_ca_ops;
2489
2490                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2491                 tcp_ack_packets_out(sk);
2492
2493                 /* Is the ACK triggering packet unambiguous? */
2494                 if (acked & FLAG_RETRANS_DATA_ACKED)
2495                         last_ackt = net_invalid_timestamp();
2496
2497                 if (ca_ops->pkts_acked)
2498                         ca_ops->pkts_acked(sk, pkts_acked, last_ackt);
2499         }
2500
2501 #if FASTRETRANS_DEBUG > 0
2502         BUG_TRAP((int)tp->sacked_out >= 0);
2503         BUG_TRAP((int)tp->lost_out >= 0);
2504         BUG_TRAP((int)tp->retrans_out >= 0);
2505         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2506                 const struct inet_connection_sock *icsk = inet_csk(sk);
2507                 if (tp->lost_out) {
2508                         printk(KERN_DEBUG "Leak l=%u %d\n",
2509                                tp->lost_out, icsk->icsk_ca_state);
2510                         tp->lost_out = 0;
2511                 }
2512                 if (tp->sacked_out) {
2513                         printk(KERN_DEBUG "Leak s=%u %d\n",
2514                                tp->sacked_out, icsk->icsk_ca_state);
2515                         tp->sacked_out = 0;
2516                 }
2517                 if (tp->retrans_out) {
2518                         printk(KERN_DEBUG "Leak r=%u %d\n",
2519                                tp->retrans_out, icsk->icsk_ca_state);
2520                         tp->retrans_out = 0;
2521                 }
2522         }
2523 #endif
2524         *seq_rtt_p = seq_rtt;
2525         return acked;
2526 }
2527
2528 static void tcp_ack_probe(struct sock *sk)
2529 {
2530         const struct tcp_sock *tp = tcp_sk(sk);
2531         struct inet_connection_sock *icsk = inet_csk(sk);
2532
2533         /* Was it a usable window open? */
2534
2535         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2536                    tp->snd_una + tp->snd_wnd)) {
2537                 icsk->icsk_backoff = 0;
2538                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2539                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2540                  * This function is not for random using!
2541                  */
2542         } else {
2543                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2544                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2545                                           TCP_RTO_MAX);
2546         }
2547 }
2548
2549 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2550 {
2551         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2552                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2553 }
2554
2555 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2556 {
2557         const struct tcp_sock *tp = tcp_sk(sk);
2558         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2559                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2560 }
2561
2562 /* Check that window update is acceptable.
2563  * The function assumes that snd_una<=ack<=snd_next.
2564  */
2565 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2566                                         const u32 ack_seq, const u32 nwin)
2567 {
2568         return (after(ack, tp->snd_una) ||
2569                 after(ack_seq, tp->snd_wl1) ||
2570                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2571 }
2572
2573 /* Update our send window.
2574  *
2575  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2576  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2577  */
2578 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2579                                  u32 ack_seq)
2580 {
2581         struct tcp_sock *tp = tcp_sk(sk);
2582         int flag = 0;
2583         u32 nwin = ntohs(tcp_hdr(skb)->window);
2584
2585         if (likely(!tcp_hdr(skb)->syn))
2586                 nwin <<= tp->rx_opt.snd_wscale;
2587
2588         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2589                 flag |= FLAG_WIN_UPDATE;
2590                 tcp_update_wl(tp, ack, ack_seq);
2591
2592                 if (tp->snd_wnd != nwin) {
2593                         tp->snd_wnd = nwin;
2594
2595                         /* Note, it is the only place, where
2596                          * fast path is recovered for sending TCP.
2597                          */
2598                         tp->pred_flags = 0;
2599                         tcp_fast_path_check(sk);
2600
2601                         if (nwin > tp->max_window) {
2602                                 tp->max_window = nwin;
2603                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2604                         }
2605                 }
2606         }
2607
2608         tp->snd_una = ack;
2609
2610         return flag;
2611 }
2612
2613 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2614  * continue in congestion avoidance.
2615  */
2616 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2617 {
2618         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2619         tp->snd_cwnd_cnt = 0;
2620         TCP_ECN_queue_cwr(tp);
2621         tcp_moderate_cwnd(tp);
2622 }
2623
2624 /* A conservative spurious RTO response algorithm: reduce cwnd using
2625  * rate halving and continue in congestion avoidance.
2626  */
2627 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2628 {
2629         tcp_enter_cwr(sk, 0);
2630 }
2631
2632 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2633 {
2634         if (flag&FLAG_ECE)
2635                 tcp_ratehalving_spur_to_response(sk);
2636         else
2637                 tcp_undo_cwr(sk, 1);
2638 }
2639
2640 /* F-RTO spurious RTO detection algorithm (RFC4138)
2641  *
2642  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2643  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2644  * window (but not to or beyond highest sequence sent before RTO):
2645  *   On First ACK,  send two new segments out.
2646  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2647  *                  algorithm is not part of the F-RTO detection algorithm
2648  *                  given in RFC4138 but can be selected separately).
2649  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2650  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2651  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2652  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2653  *
2654  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2655  * original window even after we transmit two new data segments.
2656  *
2657  * SACK version:
2658  *   on first step, wait until first cumulative ACK arrives, then move to
2659  *   the second step. In second step, the next ACK decides.
2660  *
2661  * F-RTO is implemented (mainly) in four functions:
2662  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2663  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2664  *     called when tcp_use_frto() showed green light
2665  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2666  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2667  *     to prove that the RTO is indeed spurious. It transfers the control
2668  *     from F-RTO to the conventional RTO recovery
2669  */
2670 static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
2671 {
2672         struct tcp_sock *tp = tcp_sk(sk);
2673
2674         tcp_sync_left_out(tp);
2675
2676         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2677         if (flag&FLAG_DATA_ACKED)
2678                 inet_csk(sk)->icsk_retransmits = 0;
2679
2680         if (!before(tp->snd_una, tp->frto_highmark)) {
2681                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2682                 return 1;
2683         }
2684
2685         if (!IsSackFrto() || IsReno(tp)) {
2686                 /* RFC4138 shortcoming in step 2; should also have case c):
2687                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2688                  * data, winupdate
2689                  */
2690                 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
2691                     !(flag&FLAG_FORWARD_PROGRESS))
2692                         return 1;
2693
2694                 if (!(flag&FLAG_DATA_ACKED)) {
2695                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2696                                             flag);
2697                         return 1;
2698                 }
2699         } else {
2700                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2701                         /* Prevent sending of new data. */
2702                         tp->snd_cwnd = min(tp->snd_cwnd,
2703                                            tcp_packets_in_flight(tp));
2704                         return 1;
2705                 }
2706
2707                 if ((tp->frto_counter >= 2) &&
2708                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2709                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2710                         /* RFC4138 shortcoming (see comment above) */
2711                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2712                                 return 1;
2713
2714                         tcp_enter_frto_loss(sk, 3, flag);
2715                         return 1;
2716                 }
2717         }
2718
2719         if (tp->frto_counter == 1) {
2720                 /* Sending of the next skb must be allowed or no FRTO */
2721                 if (!tcp_send_head(sk) ||
2722                     after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2723                                      tp->snd_una + tp->snd_wnd)) {
2724                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2725                                             flag);
2726                         return 1;
2727                 }
2728
2729                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2730                 tp->frto_counter = 2;
2731                 return 1;
2732         } else {
2733                 switch (sysctl_tcp_frto_response) {
2734                 case 2:
2735                         tcp_undo_spur_to_response(sk, flag);
2736                         break;
2737                 case 1:
2738                         tcp_conservative_spur_to_response(tp);
2739                         break;
2740                 default:
2741                         tcp_ratehalving_spur_to_response(sk);
2742                         break;
2743                 }
2744                 tp->frto_counter = 0;
2745         }
2746         return 0;
2747 }
2748
2749 /* This routine deals with incoming acks, but not outgoing ones. */
2750 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2751 {
2752         struct inet_connection_sock *icsk = inet_csk(sk);
2753         struct tcp_sock *tp = tcp_sk(sk);
2754         u32 prior_snd_una = tp->snd_una;
2755         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2756         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2757         u32 prior_in_flight;
2758         s32 seq_rtt;
2759         int prior_packets;
2760         int frto_cwnd = 0;
2761
2762         /* If the ack is newer than sent or older than previous acks
2763          * then we can probably ignore it.
2764          */
2765         if (after(ack, tp->snd_nxt))
2766                 goto uninteresting_ack;
2767
2768         if (before(ack, prior_snd_una))
2769                 goto old_ack;
2770
2771         if (sysctl_tcp_abc) {
2772                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2773                         tp->bytes_acked += ack - prior_snd_una;
2774                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2775                         /* we assume just one segment left network */
2776                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2777         }
2778
2779         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2780                 /* Window is constant, pure forward advance.
2781                  * No more checks are required.
2782                  * Note, we use the fact that SND.UNA>=SND.WL2.
2783                  */
2784                 tcp_update_wl(tp, ack, ack_seq);
2785                 tp->snd_una = ack;
2786                 flag |= FLAG_WIN_UPDATE;
2787
2788                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2789
2790                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2791         } else {
2792                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2793                         flag |= FLAG_DATA;
2794                 else
2795                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2796
2797                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2798
2799                 if (TCP_SKB_CB(skb)->sacked)
2800                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2801
2802                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2803                         flag |= FLAG_ECE;
2804
2805                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2806         }
2807
2808         /* We passed data and got it acked, remove any soft error
2809          * log. Something worked...
2810          */
2811         sk->sk_err_soft = 0;
2812         tp->rcv_tstamp = tcp_time_stamp;
2813         prior_packets = tp->packets_out;
2814         if (!prior_packets)
2815                 goto no_queue;
2816
2817         prior_in_flight = tcp_packets_in_flight(tp);
2818
2819         /* See if we can take anything off of the retransmit queue. */
2820         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2821
2822         if (tp->frto_counter)
2823                 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
2824
2825         if (tcp_ack_is_dubious(sk, flag)) {
2826                 /* Advance CWND, if state allows this. */
2827                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2828                     tcp_may_raise_cwnd(sk, flag))
2829                         tcp_cong_avoid(sk, ack, prior_in_flight, 0);
2830                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2831         } else {
2832                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2833                         tcp_cong_avoid(sk, ack, prior_in_flight, 1);
2834         }
2835
2836         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2837                 dst_confirm(sk->sk_dst_cache);
2838
2839         return 1;
2840
2841 no_queue:
2842         icsk->icsk_probes_out = 0;
2843
2844         /* If this ack opens up a zero window, clear backoff.  It was
2845          * being used to time the probes, and is probably far higher than
2846          * it needs to be for normal retransmission.
2847          */
2848         if (tcp_send_head(sk))
2849                 tcp_ack_probe(sk);
2850         return 1;
2851
2852 old_ack:
2853         if (TCP_SKB_CB(skb)->sacked)
2854                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2855
2856 uninteresting_ack:
2857         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2858         return 0;
2859 }
2860
2861
2862 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2863  * But, this can also be called on packets in the established flow when
2864  * the fast version below fails.
2865  */
2866 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2867 {
2868         unsigned char *ptr;
2869         struct tcphdr *th = tcp_hdr(skb);
2870         int length=(th->doff*4)-sizeof(struct tcphdr);
2871
2872         ptr = (unsigned char *)(th + 1);
2873         opt_rx->saw_tstamp = 0;
2874
2875         while (length > 0) {
2876                 int opcode=*ptr++;
2877                 int opsize;
2878
2879                 switch (opcode) {
2880                         case TCPOPT_EOL:
2881                                 return;
2882                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2883                                 length--;
2884                                 continue;
2885                         default:
2886                                 opsize=*ptr++;
2887                                 if (opsize < 2) /* "silly options" */
2888                                         return;
2889                                 if (opsize > length)
2890                                         return; /* don't parse partial options */
2891                                 switch (opcode) {
2892                                 case TCPOPT_MSS:
2893                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2894                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2895                                                 if (in_mss) {
2896                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2897                                                                 in_mss = opt_rx->user_mss;
2898                                                         opt_rx->mss_clamp = in_mss;
2899                                                 }
2900                                         }
2901                                         break;
2902                                 case TCPOPT_WINDOW:
2903                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2904                                                 if (sysctl_tcp_window_scaling) {
2905                                                         __u8 snd_wscale = *(__u8 *) ptr;
2906                                                         opt_rx->wscale_ok = 1;
2907                                                         if (snd_wscale > 14) {
2908                                                                 if (net_ratelimit())
2909                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2910                                                                                "scaling value %d >14 received.\n",
2911                                                                                snd_wscale);
2912                                                                 snd_wscale = 14;
2913                                                         }
2914                                                         opt_rx->snd_wscale = snd_wscale;
2915                                                 }
2916                                         break;
2917                                 case TCPOPT_TIMESTAMP:
2918                                         if (opsize==TCPOLEN_TIMESTAMP) {
2919                                                 if ((estab && opt_rx->tstamp_ok) ||
2920                                                     (!estab && sysctl_tcp_timestamps)) {
2921                                                         opt_rx->saw_tstamp = 1;
2922                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2923                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2924                                                 }
2925                                         }
2926                                         break;
2927                                 case TCPOPT_SACK_PERM:
2928                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2929                                                 if (sysctl_tcp_sack) {
2930                                                         opt_rx->sack_ok = 1;
2931                                                         tcp_sack_reset(opt_rx);
2932                                                 }
2933                                         }
2934                                         break;
2935
2936                                 case TCPOPT_SACK:
2937                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2938                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2939                                            opt_rx->sack_ok) {
2940                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2941                                         }
2942                                         break;
2943 #ifdef CONFIG_TCP_MD5SIG
2944                                 case TCPOPT_MD5SIG:
2945                                         /*
2946                                          * The MD5 Hash has already been
2947                                          * checked (see tcp_v{4,6}_do_rcv()).
2948                                          */
2949                                         break;
2950 #endif
2951                                 }
2952
2953                                 ptr+=opsize-2;
2954                                 length-=opsize;
2955                 }
2956         }
2957 }
2958
2959 /* Fast parse options. This hopes to only see timestamps.
2960  * If it is wrong it falls back on tcp_parse_options().
2961  */
2962 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2963                                   struct tcp_sock *tp)
2964 {
2965         if (th->doff == sizeof(struct tcphdr)>>2) {
2966                 tp->rx_opt.saw_tstamp = 0;
2967                 return 0;
2968         } else if (tp->rx_opt.tstamp_ok &&
2969                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2970                 __be32 *ptr = (__be32 *)(th + 1);
2971                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2972                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2973                         tp->rx_opt.saw_tstamp = 1;
2974                         ++ptr;
2975                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2976                         ++ptr;
2977                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2978                         return 1;
2979                 }
2980         }
2981         tcp_parse_options(skb, &tp->rx_opt, 1);
2982         return 1;
2983 }
2984
2985 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2986 {
2987         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2988         tp->rx_opt.ts_recent_stamp = get_seconds();
2989 }
2990
2991 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2992 {
2993         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2994                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2995                  * extra check below makes sure this can only happen
2996                  * for pure ACK frames.  -DaveM
2997                  *
2998                  * Not only, also it occurs for expired timestamps.
2999                  */
3000
3001                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3002                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3003                         tcp_store_ts_recent(tp);
3004         }
3005 }
3006
3007 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3008  *
3009  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3010  * it can pass through stack. So, the following predicate verifies that
3011  * this segment is not used for anything but congestion avoidance or
3012  * fast retransmit. Moreover, we even are able to eliminate most of such
3013  * second order effects, if we apply some small "replay" window (~RTO)
3014  * to timestamp space.
3015  *
3016  * All these measures still do not guarantee that we reject wrapped ACKs
3017  * on networks with high bandwidth, when sequence space is recycled fastly,
3018  * but it guarantees that such events will be very rare and do not affect
3019  * connection seriously. This doesn't look nice, but alas, PAWS is really
3020  * buggy extension.
3021  *
3022  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3023  * states that events when retransmit arrives after original data are rare.
3024  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3025  * the biggest problem on large power networks even with minor reordering.
3026  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3027  * up to bandwidth of 18Gigabit/sec. 8) ]
3028  */
3029
3030 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3031 {
3032         struct tcp_sock *tp = tcp_sk(sk);
3033         struct tcphdr *th = tcp_hdr(skb);
3034         u32 seq = TCP_SKB_CB(skb)->seq;
3035         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3036
3037         return (/* 1. Pure ACK with correct sequence number. */
3038                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3039
3040                 /* 2. ... and duplicate ACK. */
3041                 ack == tp->snd_una &&
3042
3043                 /* 3. ... and does not update window. */
3044                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3045
3046                 /* 4. ... and sits in replay window. */
3047                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3048 }
3049
3050 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3051 {
3052         const struct tcp_sock *tp = tcp_sk(sk);
3053         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3054                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3055                 !tcp_disordered_ack(sk, skb));
3056 }
3057
3058 /* Check segment sequence number for validity.
3059  *
3060  * Segment controls are considered valid, if the segment
3061  * fits to the window after truncation to the window. Acceptability
3062  * of data (and SYN, FIN, of course) is checked separately.
3063  * See tcp_data_queue(), for example.
3064  *
3065  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3066  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3067  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3068  * (borrowed from freebsd)
3069  */
3070
3071 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3072 {
3073         return  !before(end_seq, tp->rcv_wup) &&
3074                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3075 }
3076
3077 /* When we get a reset we do this. */
3078 static void tcp_reset(struct sock *sk)
3079 {
3080         /* We want the right error as BSD sees it (and indeed as we do). */
3081         switch (sk->sk_state) {
3082                 case TCP_SYN_SENT:
3083                         sk->sk_err = ECONNREFUSED;
3084                         break;
3085                 case TCP_CLOSE_WAIT:
3086                         sk->sk_err = EPIPE;
3087                         break;
3088                 case TCP_CLOSE:
3089                         return;
3090                 default:
3091                         sk->sk_err = ECONNRESET;
3092         }
3093
3094         if (!sock_flag(sk, SOCK_DEAD))
3095                 sk->sk_error_report(sk);
3096
3097         tcp_done(sk);
3098 }
3099
3100 /*
3101  *      Process the FIN bit. This now behaves as it is supposed to work
3102  *      and the FIN takes effect when it is validly part of sequence
3103  *      space. Not before when we get holes.
3104  *
3105  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3106  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3107  *      TIME-WAIT)
3108  *
3109  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3110  *      close and we go into CLOSING (and later onto TIME-WAIT)
3111  *
3112  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3113  */
3114 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3115 {
3116         struct tcp_sock *tp = tcp_sk(sk);
3117
3118         inet_csk_schedule_ack(sk);
3119
3120         sk->sk_shutdown |= RCV_SHUTDOWN;
3121         sock_set_flag(sk, SOCK_DONE);
3122
3123         switch (sk->sk_state) {
3124                 case TCP_SYN_RECV:
3125                 case TCP_ESTABLISHED:
3126                         /* Move to CLOSE_WAIT */
3127                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3128                         inet_csk(sk)->icsk_ack.pingpong = 1;
3129                         break;
3130
3131                 case TCP_CLOSE_WAIT:
3132                 case TCP_CLOSING:
3133                         /* Received a retransmission of the FIN, do
3134                          * nothing.
3135                          */
3136                         break;
3137                 case TCP_LAST_ACK:
3138                         /* RFC793: Remain in the LAST-ACK state. */
3139                         break;
3140
3141                 case TCP_FIN_WAIT1:
3142                         /* This case occurs when a simultaneous close
3143                          * happens, we must ack the received FIN and
3144                          * enter the CLOSING state.
3145                          */
3146                         tcp_send_ack(sk);
3147                         tcp_set_state(sk, TCP_CLOSING);
3148                         break;
3149                 case TCP_FIN_WAIT2:
3150                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3151                         tcp_send_ack(sk);
3152                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3153                         break;
3154                 default:
3155                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3156                          * cases we should never reach this piece of code.
3157                          */
3158                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3159                                __FUNCTION__, sk->sk_state);
3160                         break;
3161         }
3162
3163         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3164          * Probably, we should reset in this case. For now drop them.
3165          */
3166         __skb_queue_purge(&tp->out_of_order_queue);
3167         if (tp->rx_opt.sack_ok)
3168                 tcp_sack_reset(&tp->rx_opt);
3169         sk_stream_mem_reclaim(sk);
3170
3171         if (!sock_flag(sk, SOCK_DEAD)) {
3172                 sk->sk_state_change(sk);
3173
3174                 /* Do not send POLL_HUP for half duplex close. */
3175                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3176                     sk->sk_state == TCP_CLOSE)
3177                         sk_wake_async(sk, 1, POLL_HUP);
3178                 else
3179                         sk_wake_async(sk, 1, POLL_IN);
3180         }
3181 }
3182
3183 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3184 {
3185         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3186                 if (before(seq, sp->start_seq))
3187                         sp->start_seq = seq;
3188                 if (after(end_seq, sp->end_seq))
3189                         sp->end_seq = end_seq;
3190                 return 1;
3191         }
3192         return 0;
3193 }
3194
3195 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3196 {
3197         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3198                 if (before(seq, tp->rcv_nxt))
3199                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3200                 else
3201                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3202
3203                 tp->rx_opt.dsack = 1;
3204                 tp->duplicate_sack[0].start_seq = seq;
3205                 tp->duplicate_sack[0].end_seq = end_seq;
3206                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3207         }
3208 }
3209
3210 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3211 {
3212         if (!tp->rx_opt.dsack)
3213                 tcp_dsack_set(tp, seq, end_seq);
3214         else
3215                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3216 }
3217
3218 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3219 {
3220         struct tcp_sock *tp = tcp_sk(sk);
3221
3222         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3223             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3224                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3225                 tcp_enter_quickack_mode(sk);
3226
3227                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3228                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3229
3230                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3231                                 end_seq = tp->rcv_nxt;
3232                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3233                 }
3234         }
3235
3236         tcp_send_ack(sk);
3237 }
3238
3239 /* These routines update the SACK block as out-of-order packets arrive or
3240  * in-order packets close up the sequence space.
3241  */
3242 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3243 {
3244         int this_sack;
3245         struct tcp_sack_block *sp = &tp->selective_acks[0];
3246         struct tcp_sack_block *swalk = sp+1;
3247
3248         /* See if the recent change to the first SACK eats into
3249          * or hits the sequence space of other SACK blocks, if so coalesce.
3250          */
3251         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3252                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3253                         int i;
3254
3255                         /* Zap SWALK, by moving every further SACK up by one slot.
3256                          * Decrease num_sacks.
3257                          */
3258                         tp->rx_opt.num_sacks--;
3259                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3260                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3261                                 sp[i] = sp[i+1];
3262                         continue;
3263                 }
3264                 this_sack++, swalk++;
3265         }
3266 }
3267
3268 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3269 {
3270         __u32 tmp;
3271
3272         tmp = sack1->start_seq;
3273         sack1->start_seq = sack2->start_seq;
3274         sack2->start_seq = tmp;
3275
3276         tmp = sack1->end_seq;
3277         sack1->end_seq = sack2->end_seq;
3278         sack2->end_seq = tmp;
3279 }
3280
3281 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3282 {
3283         struct tcp_sock *tp = tcp_sk(sk);
3284         struct tcp_sack_block *sp = &tp->selective_acks[0];
3285         int cur_sacks = tp->rx_opt.num_sacks;
3286         int this_sack;
3287
3288         if (!cur_sacks)
3289                 goto new_sack;
3290
3291         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3292                 if (tcp_sack_extend(sp, seq, end_seq)) {
3293                         /* Rotate this_sack to the first one. */
3294                         for (; this_sack>0; this_sack--, sp--)
3295                                 tcp_sack_swap(sp, sp-1);
3296                         if (cur_sacks > 1)
3297                                 tcp_sack_maybe_coalesce(tp);
3298                         return;
3299                 }
3300         }
3301
3302         /* Could not find an adjacent existing SACK, build a new one,
3303          * put it at the front, and shift everyone else down.  We
3304          * always know there is at least one SACK present already here.
3305          *
3306          * If the sack array is full, forget about the last one.
3307          */
3308         if (this_sack >= 4) {
3309                 this_sack--;
3310                 tp->rx_opt.num_sacks--;
3311                 sp--;
3312         }
3313         for (; this_sack > 0; this_sack--, sp--)
3314                 *sp = *(sp-1);
3315
3316 new_sack:
3317         /* Build the new head SACK, and we're done. */
3318         sp->start_seq = seq;
3319         sp->end_seq = end_seq;
3320         tp->rx_opt.num_sacks++;
3321         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3322 }
3323
3324 /* RCV.NXT advances, some SACKs should be eaten. */
3325
3326 static void tcp_sack_remove(struct tcp_sock *tp)
3327 {
3328         struct tcp_sack_block *sp = &tp->selective_acks[0];
3329         int num_sacks = tp->rx_opt.num_sacks;
3330         int this_sack;
3331
3332         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3333         if (skb_queue_empty(&tp->out_of_order_queue)) {
3334                 tp->rx_opt.num_sacks = 0;
3335                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3336                 return;
3337         }
3338
3339         for (this_sack = 0; this_sack < num_sacks; ) {
3340                 /* Check if the start of the sack is covered by RCV.NXT. */
3341                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3342                         int i;
3343
3344                         /* RCV.NXT must cover all the block! */
3345                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3346
3347                         /* Zap this SACK, by moving forward any other SACKS. */
3348                         for (i=this_sack+1; i < num_sacks; i++)
3349                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3350                         num_sacks--;
3351                         continue;
3352                 }
3353                 this_sack++;
3354                 sp++;
3355         }
3356         if (num_sacks != tp->rx_opt.num_sacks) {
3357                 tp->rx_opt.num_sacks = num_sacks;
3358                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3359         }
3360 }
3361
3362 /* This one checks to see if we can put data from the
3363  * out_of_order queue into the receive_queue.
3364  */
3365 static void tcp_ofo_queue(struct sock *sk)
3366 {
3367         struct tcp_sock *tp = tcp_sk(sk);
3368         __u32 dsack_high = tp->rcv_nxt;
3369         struct sk_buff *skb;
3370
3371         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3372                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3373                         break;
3374
3375                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3376                         __u32 dsack = dsack_high;
3377                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3378                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3379                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3380                 }
3381
3382                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3383                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3384                         __skb_unlink(skb, &tp->out_of_order_queue);
3385                         __kfree_skb(skb);
3386                         continue;
3387                 }
3388                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3389                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3390                            TCP_SKB_CB(skb)->end_seq);
3391
3392                 __skb_unlink(skb, &tp->out_of_order_queue);
3393                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3394                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3395                 if (tcp_hdr(skb)->fin)
3396                         tcp_fin(skb, sk, tcp_hdr(skb));
3397         }
3398 }
3399
3400 static int tcp_prune_queue(struct sock *sk);
3401
3402 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3403 {
3404         struct tcphdr *th = tcp_hdr(skb);
3405         struct tcp_sock *tp = tcp_sk(sk);
3406         int eaten = -1;
3407
3408         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3409                 goto drop;
3410
3411         __skb_pull(skb, th->doff*4);
3412
3413         TCP_ECN_accept_cwr(tp, skb);
3414
3415         if (tp->rx_opt.dsack) {
3416                 tp->rx_opt.dsack = 0;
3417                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3418                                                     4 - tp->rx_opt.tstamp_ok);
3419         }
3420
3421         /*  Queue data for delivery to the user.
3422          *  Packets in sequence go to the receive queue.
3423          *  Out of sequence packets to the out_of_order_queue.
3424          */
3425         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3426                 if (tcp_receive_window(tp) == 0)
3427                         goto out_of_window;
3428
3429                 /* Ok. In sequence. In window. */
3430                 if (tp->ucopy.task == current &&
3431                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3432                     sock_owned_by_user(sk) && !tp->urg_data) {
3433                         int chunk = min_t(unsigned int, skb->len,
3434                                                         tp->ucopy.len);
3435
3436                         __set_current_state(TASK_RUNNING);
3437
3438                         local_bh_enable();
3439                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3440                                 tp->ucopy.len -= chunk;
3441                                 tp->copied_seq += chunk;
3442                                 eaten = (chunk == skb->len && !th->fin);
3443                                 tcp_rcv_space_adjust(sk);
3444                         }
3445                         local_bh_disable();
3446                 }
3447
3448                 if (eaten <= 0) {
3449 queue_and_out:
3450                         if (eaten < 0 &&
3451                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3452                              !sk_stream_rmem_schedule(sk, skb))) {
3453                                 if (tcp_prune_queue(sk) < 0 ||
3454                                     !sk_stream_rmem_schedule(sk, skb))
3455                                         goto drop;
3456                         }
3457                         sk_stream_set_owner_r(skb, sk);
3458                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3459                 }
3460                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3461                 if (skb->len)
3462                         tcp_event_data_recv(sk, skb);
3463                 if (th->fin)
3464                         tcp_fin(skb, sk, th);
3465
3466                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3467                         tcp_ofo_queue(sk);
3468
3469                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3470                          * gap in queue is filled.
3471                          */
3472                         if (skb_queue_empty(&tp->out_of_order_queue))
3473                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3474                 }
3475
3476                 if (tp->rx_opt.num_sacks)
3477                         tcp_sack_remove(tp);
3478
3479                 tcp_fast_path_check(sk);
3480
3481                 if (eaten > 0)
3482                         __kfree_skb(skb);
3483                 else if (!sock_flag(sk, SOCK_DEAD))
3484                         sk->sk_data_ready(sk, 0);
3485                 return;
3486         }
3487
3488         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3489                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3490                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3491                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3492
3493 out_of_window:
3494                 tcp_enter_quickack_mode(sk);
3495                 inet_csk_schedule_ack(sk);
3496 drop:
3497                 __kfree_skb(skb);
3498                 return;
3499         }
3500
3501         /* Out of window. F.e. zero window probe. */
3502         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3503                 goto out_of_window;
3504
3505         tcp_enter_quickack_mode(sk);
3506
3507         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3508                 /* Partial packet, seq < rcv_next < end_seq */
3509                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3510                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3511                            TCP_SKB_CB(skb)->end_seq);
3512
3513                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3514
3515                 /* If window is closed, drop tail of packet. But after
3516                  * remembering D-SACK for its head made in previous line.
3517                  */
3518                 if (!tcp_receive_window(tp))
3519                         goto out_of_window;
3520                 goto queue_and_out;
3521         }
3522
3523         TCP_ECN_check_ce(tp, skb);
3524
3525         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3526             !sk_stream_rmem_schedule(sk, skb)) {
3527                 if (tcp_prune_queue(sk) < 0 ||
3528                     !sk_stream_rmem_schedule(sk, skb))
3529                         goto drop;
3530         }
3531
3532         /* Disable header prediction. */
3533         tp->pred_flags = 0;
3534         inet_csk_schedule_ack(sk);
3535
3536         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3537                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3538
3539         sk_stream_set_owner_r(skb, sk);
3540
3541         if (!skb_peek(&tp->out_of_order_queue)) {
3542                 /* Initial out of order segment, build 1 SACK. */
3543                 if (tp->rx_opt.sack_ok) {
3544                         tp->rx_opt.num_sacks = 1;
3545                         tp->rx_opt.dsack     = 0;
3546                         tp->rx_opt.eff_sacks = 1;
3547                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3548                         tp->selective_acks[0].end_seq =
3549                                                 TCP_SKB_CB(skb)->end_seq;
3550                 }
3551                 __skb_queue_head(&tp->out_of_order_queue,skb);
3552         } else {
3553                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3554                 u32 seq = TCP_SKB_CB(skb)->seq;
3555                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3556
3557                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3558                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3559
3560                         if (!tp->rx_opt.num_sacks ||
3561                             tp->selective_acks[0].end_seq != seq)
3562                                 goto add_sack;
3563
3564                         /* Common case: data arrive in order after hole. */
3565                         tp->selective_acks[0].end_seq = end_seq;
3566                         return;
3567                 }
3568
3569                 /* Find place to insert this segment. */
3570                 do {
3571                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3572                                 break;
3573                 } while ((skb1 = skb1->prev) !=
3574                          (struct sk_buff*)&tp->out_of_order_queue);
3575
3576                 /* Do skb overlap to previous one? */
3577                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3578                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3579                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3580                                 /* All the bits are present. Drop. */
3581                                 __kfree_skb(skb);
3582                                 tcp_dsack_set(tp, seq, end_seq);
3583                                 goto add_sack;
3584                         }
3585                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3586                                 /* Partial overlap. */
3587                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3588                         } else {
3589                                 skb1 = skb1->prev;
3590                         }
3591                 }
3592                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3593
3594                 /* And clean segments covered by new one as whole. */
3595                 while ((skb1 = skb->next) !=
3596                        (struct sk_buff*)&tp->out_of_order_queue &&
3597                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3598                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3599                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3600                                break;
3601                        }
3602                        __skb_unlink(skb1, &tp->out_of_order_queue);
3603                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3604                        __kfree_skb(skb1);
3605                 }
3606
3607 add_sack:
3608                 if (tp->rx_opt.sack_ok)
3609                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3610         }
3611 }
3612
3613 /* Collapse contiguous sequence of skbs head..tail with
3614  * sequence numbers start..end.
3615  * Segments with FIN/SYN are not collapsed (only because this
3616  * simplifies code)
3617  */
3618 static void
3619 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3620              struct sk_buff *head, struct sk_buff *tail,
3621              u32 start, u32 end)
3622 {
3623         struct sk_buff *skb;
3624
3625         /* First, check that queue is collapsible and find
3626          * the point where collapsing can be useful. */
3627         for (skb = head; skb != tail; ) {
3628                 /* No new bits? It is possible on ofo queue. */
3629                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3630                         struct sk_buff *next = skb->next;
3631                         __skb_unlink(skb, list);
3632                         __kfree_skb(skb);
3633                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3634                         skb = next;
3635                         continue;
3636                 }
3637
3638                 /* The first skb to collapse is:
3639                  * - not SYN/FIN and
3640                  * - bloated or contains data before "start" or
3641                  *   overlaps to the next one.
3642                  */
3643                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3644                     (tcp_win_from_space(skb->truesize) > skb->len ||
3645                      before(TCP_SKB_CB(skb)->seq, start) ||
3646                      (skb->next != tail &&
3647                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3648                         break;
3649
3650                 /* Decided to skip this, advance start seq. */
3651                 start = TCP_SKB_CB(skb)->end_seq;
3652                 skb = skb->next;
3653         }
3654         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3655                 return;
3656
3657         while (before(start, end)) {
3658                 struct sk_buff *nskb;
3659                 int header = skb_headroom(skb);
3660                 int copy = SKB_MAX_ORDER(header, 0);
3661
3662                 /* Too big header? This can happen with IPv6. */
3663                 if (copy < 0)
3664                         return;
3665                 if (end-start < copy)
3666                         copy = end-start;
3667                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3668                 if (!nskb)
3669                         return;
3670
3671                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3672                 skb_set_network_header(nskb, (skb_network_header(skb) -
3673                                               skb->head));
3674                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3675                                                 skb->head));
3676                 skb_reserve(nskb, header);
3677                 memcpy(nskb->head, skb->head, header);
3678                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3679                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3680                 __skb_insert(nskb, skb->prev, skb, list);
3681                 sk_stream_set_owner_r(nskb, sk);
3682
3683                 /* Copy data, releasing collapsed skbs. */
3684                 while (copy > 0) {
3685                         int offset = start - TCP_SKB_CB(skb)->seq;
3686                         int size = TCP_SKB_CB(skb)->end_seq - start;
3687
3688                         BUG_ON(offset < 0);
3689                         if (size > 0) {
3690                                 size = min(copy, size);
3691                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3692                                         BUG();
3693                                 TCP_SKB_CB(nskb)->end_seq += size;
3694                                 copy -= size;
3695                                 start += size;
3696                         }
3697                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3698                                 struct sk_buff *next = skb->next;
3699                                 __skb_unlink(skb, list);
3700                                 __kfree_skb(skb);
3701                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3702                                 skb = next;
3703                                 if (skb == tail ||
3704                                     tcp_hdr(skb)->syn ||
3705                                     tcp_hdr(skb)->fin)
3706                                         return;
3707                         }
3708                 }
3709         }
3710 }
3711
3712 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3713  * and tcp_collapse() them until all the queue is collapsed.
3714  */
3715 static void tcp_collapse_ofo_queue(struct sock *sk)
3716 {
3717         struct tcp_sock *tp = tcp_sk(sk);
3718         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3719         struct sk_buff *head;
3720         u32 start, end;
3721
3722         if (skb == NULL)
3723                 return;
3724
3725         start = TCP_SKB_CB(skb)->seq;
3726         end = TCP_SKB_CB(skb)->end_seq;
3727         head = skb;
3728
3729         for (;;) {
3730                 skb = skb->next;
3731
3732                 /* Segment is terminated when we see gap or when
3733                  * we are at the end of all the queue. */
3734                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3735                     after(TCP_SKB_CB(skb)->seq, end) ||
3736                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3737                         tcp_collapse(sk, &tp->out_of_order_queue,
3738                                      head, skb, start, end);
3739                         head = skb;
3740                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3741                                 break;
3742                         /* Start new segment */
3743                         start = TCP_SKB_CB(skb)->seq;
3744                         end = TCP_SKB_CB(skb)->end_seq;
3745                 } else {
3746                         if (before(TCP_SKB_CB(skb)->seq, start))
3747                                 start = TCP_SKB_CB(skb)->seq;
3748                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3749                                 end = TCP_SKB_CB(skb)->end_seq;
3750                 }
3751         }
3752 }
3753
3754 /* Reduce allocated memory if we can, trying to get
3755  * the socket within its memory limits again.
3756  *
3757  * Return less than zero if we should start dropping frames
3758  * until the socket owning process reads some of the data
3759  * to stabilize the situation.
3760  */
3761 static int tcp_prune_queue(struct sock *sk)
3762 {
3763         struct tcp_sock *tp = tcp_sk(sk);
3764
3765         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3766
3767         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3768
3769         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3770                 tcp_clamp_window(sk);
3771         else if (tcp_memory_pressure)
3772                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3773
3774         tcp_collapse_ofo_queue(sk);
3775         tcp_collapse(sk, &sk->sk_receive_queue,
3776                      sk->sk_receive_queue.next,
3777                      (struct sk_buff*)&sk->sk_receive_queue,
3778                      tp->copied_seq, tp->rcv_nxt);
3779         sk_stream_mem_reclaim(sk);
3780
3781         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3782                 return 0;
3783
3784         /* Collapsing did not help, destructive actions follow.
3785          * This must not ever occur. */
3786
3787         /* First, purge the out_of_order queue. */
3788         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3789                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3790                 __skb_queue_purge(&tp->out_of_order_queue);
3791
3792                 /* Reset SACK state.  A conforming SACK implementation will
3793                  * do the same at a timeout based retransmit.  When a connection
3794                  * is in a sad state like this, we care only about integrity
3795                  * of the connection not performance.
3796                  */
3797                 if (tp->rx_opt.sack_ok)
3798                         tcp_sack_reset(&tp->rx_opt);
3799                 sk_stream_mem_reclaim(sk);
3800         }
3801
3802         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3803                 return 0;
3804
3805         /* If we are really being abused, tell the caller to silently
3806          * drop receive data on the floor.  It will get retransmitted
3807          * and hopefully then we'll have sufficient space.
3808          */
3809         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3810
3811         /* Massive buffer overcommit. */
3812         tp->pred_flags = 0;
3813         return -1;
3814 }
3815
3816
3817 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3818  * As additional protections, we do not touch cwnd in retransmission phases,
3819  * and if application hit its sndbuf limit recently.
3820  */
3821 void tcp_cwnd_application_limited(struct sock *sk)
3822 {
3823         struct tcp_sock *tp = tcp_sk(sk);
3824
3825         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3826             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3827                 /* Limited by application or receiver window. */
3828                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3829                 u32 win_used = max(tp->snd_cwnd_used, init_win);
3830                 if (win_used < tp->snd_cwnd) {
3831                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3832                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3833                 }
3834                 tp->snd_cwnd_used = 0;
3835         }
3836         tp->snd_cwnd_stamp = tcp_time_stamp;
3837 }
3838
3839 static int tcp_should_expand_sndbuf(struct sock *sk)
3840 {
3841         struct tcp_sock *tp = tcp_sk(sk);
3842
3843         /* If the user specified a specific send buffer setting, do
3844          * not modify it.
3845          */
3846         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3847                 return 0;
3848
3849         /* If we are under global TCP memory pressure, do not expand.  */
3850         if (tcp_memory_pressure)
3851                 return 0;
3852
3853         /* If we are under soft global TCP memory pressure, do not expand.  */
3854         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3855                 return 0;
3856
3857         /* If we filled the congestion window, do not expand.  */
3858         if (tp->packets_out >= tp->snd_cwnd)
3859                 return 0;
3860
3861         return 1;
3862 }
3863
3864 /* When incoming ACK allowed to free some skb from write_queue,
3865  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3866  * on the exit from tcp input handler.
3867  *
3868  * PROBLEM: sndbuf expansion does not work well with largesend.
3869  */
3870 static void tcp_new_space(struct sock *sk)
3871 {
3872         struct tcp_sock *tp = tcp_sk(sk);
3873
3874         if (tcp_should_expand_sndbuf(sk)) {
3875                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3876                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3877                     demanded = max_t(unsigned int, tp->snd_cwnd,
3878                                                    tp->reordering + 1);
3879                 sndmem *= 2*demanded;
3880                 if (sndmem > sk->sk_sndbuf)
3881                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3882                 tp->snd_cwnd_stamp = tcp_time_stamp;
3883         }
3884
3885         sk->sk_write_space(sk);
3886 }
3887
3888 static void tcp_check_space(struct sock *sk)
3889 {
3890         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3891                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3892                 if (sk->sk_socket &&
3893                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3894                         tcp_new_space(sk);
3895         }
3896 }
3897
3898 static inline void tcp_data_snd_check(struct sock *sk)
3899 {
3900         tcp_push_pending_frames(sk);
3901         tcp_check_space(sk);
3902 }
3903
3904 /*
3905  * Check if sending an ack is needed.
3906  */
3907 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3908 {
3909         struct tcp_sock *tp = tcp_sk(sk);
3910
3911             /* More than one full frame received... */
3912         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3913              /* ... and right edge of window advances far enough.
3914               * (tcp_recvmsg() will send ACK otherwise). Or...
3915               */
3916              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3917             /* We ACK each frame or... */
3918             tcp_in_quickack_mode(sk) ||
3919             /* We have out of order data. */
3920             (ofo_possible &&
3921              skb_peek(&tp->out_of_order_queue))) {
3922                 /* Then ack it now */
3923                 tcp_send_ack(sk);
3924         } else {
3925                 /* Else, send delayed ack. */
3926                 tcp_send_delayed_ack(sk);
3927         }
3928 }
3929
3930 static inline void tcp_ack_snd_check(struct sock *sk)
3931 {
3932         if (!inet_csk_ack_scheduled(sk)) {
3933                 /* We sent a data segment already. */
3934                 return;
3935         }
3936         __tcp_ack_snd_check(sk, 1);
3937 }
3938
3939 /*
3940  *      This routine is only called when we have urgent data
3941  *      signaled. Its the 'slow' part of tcp_urg. It could be
3942  *      moved inline now as tcp_urg is only called from one
3943  *      place. We handle URGent data wrong. We have to - as
3944  *      BSD still doesn't use the correction from RFC961.
3945  *      For 1003.1g we should support a new option TCP_STDURG to permit
3946  *      either form (or just set the sysctl tcp_stdurg).
3947  */
3948
3949 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3950 {
3951         struct tcp_sock *tp = tcp_sk(sk);
3952         u32 ptr = ntohs(th->urg_ptr);
3953
3954         if (ptr && !sysctl_tcp_stdurg)
3955                 ptr--;
3956         ptr += ntohl(th->seq);
3957
3958         /* Ignore urgent data that we've already seen and read. */
3959         if (after(tp->copied_seq, ptr))
3960                 return;
3961
3962         /* Do not replay urg ptr.
3963          *
3964          * NOTE: interesting situation not covered by specs.
3965          * Misbehaving sender may send urg ptr, pointing to segment,
3966          * which we already have in ofo queue. We are not able to fetch
3967          * such data and will stay in TCP_URG_NOTYET until will be eaten
3968          * by recvmsg(). Seems, we are not obliged to handle such wicked
3969          * situations. But it is worth to think about possibility of some
3970          * DoSes using some hypothetical application level deadlock.
3971          */
3972         if (before(ptr, tp->rcv_nxt))
3973                 return;
3974
3975         /* Do we already have a newer (or duplicate) urgent pointer? */
3976         if (tp->urg_data && !after(ptr, tp->urg_seq))
3977                 return;
3978
3979         /* Tell the world about our new urgent pointer. */
3980         sk_send_sigurg(sk);
3981
3982         /* We may be adding urgent data when the last byte read was
3983          * urgent. To do this requires some care. We cannot just ignore
3984          * tp->copied_seq since we would read the last urgent byte again
3985          * as data, nor can we alter copied_seq until this data arrives
3986          * or we break the semantics of SIOCATMARK (and thus sockatmark())
3987          *
3988          * NOTE. Double Dutch. Rendering to plain English: author of comment
3989          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3990          * and expect that both A and B disappear from stream. This is _wrong_.
3991          * Though this happens in BSD with high probability, this is occasional.
3992          * Any application relying on this is buggy. Note also, that fix "works"
3993          * only in this artificial test. Insert some normal data between A and B and we will
3994          * decline of BSD again. Verdict: it is better to remove to trap
3995          * buggy users.
3996          */
3997         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3998             !sock_flag(sk, SOCK_URGINLINE) &&
3999             tp->copied_seq != tp->rcv_nxt) {
4000                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4001                 tp->copied_seq++;
4002                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4003                         __skb_unlink(skb, &sk->sk_receive_queue);
4004                         __kfree_skb(skb);
4005                 }
4006         }
4007
4008         tp->urg_data   = TCP_URG_NOTYET;
4009         tp->urg_seq    = ptr;
4010
4011         /* Disable header prediction. */
4012         tp->pred_flags = 0;
4013 }
4014
4015 /* This is the 'fast' part of urgent handling. */
4016 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4017 {
4018         struct tcp_sock *tp = tcp_sk(sk);
4019
4020         /* Check if we get a new urgent pointer - normally not. */
4021         if (th->urg)
4022                 tcp_check_urg(sk,th);
4023
4024         /* Do we wait for any urgent data? - normally not... */
4025         if (tp->urg_data == TCP_URG_NOTYET) {
4026                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4027                           th->syn;
4028
4029                 /* Is the urgent pointer pointing into this packet? */
4030                 if (ptr < skb->len) {
4031                         u8 tmp;
4032                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4033                                 BUG();
4034                         tp->urg_data = TCP_URG_VALID | tmp;
4035                         if (!sock_flag(sk, SOCK_DEAD))
4036                                 sk->sk_data_ready(sk, 0);
4037                 }
4038         }
4039 }
4040
4041 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4042 {
4043         struct tcp_sock *tp = tcp_sk(sk);
4044         int chunk = skb->len - hlen;
4045         int err;
4046
4047         local_bh_enable();
4048         if (skb_csum_unnecessary(skb))
4049                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4050         else
4051                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4052                                                        tp->ucopy.iov);
4053
4054         if (!err) {
4055                 tp->ucopy.len -= chunk;
4056                 tp->copied_seq += chunk;
4057                 tcp_rcv_space_adjust(sk);
4058         }
4059
4060         local_bh_disable();
4061         return err;
4062 }
4063
4064 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4065 {
4066         __sum16 result;
4067
4068         if (sock_owned_by_user(sk)) {
4069                 local_bh_enable();
4070                 result = __tcp_checksum_complete(skb);
4071                 local_bh_disable();
4072         } else {
4073                 result = __tcp_checksum_complete(skb);
4074         }
4075         return result;
4076 }
4077
4078 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4079 {
4080         return !skb_csum_unnecessary(skb) &&
4081                 __tcp_checksum_complete_user(sk, skb);
4082 }
4083
4084 #ifdef CONFIG_NET_DMA
4085 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4086 {
4087         struct tcp_sock *tp = tcp_sk(sk);
4088         int chunk = skb->len - hlen;
4089         int dma_cookie;
4090         int copied_early = 0;
4091
4092         if (tp->ucopy.wakeup)
4093                 return 0;
4094
4095         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4096                 tp->ucopy.dma_chan = get_softnet_dma();
4097
4098         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4099
4100                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4101                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4102
4103                 if (dma_cookie < 0)
4104                         goto out;
4105
4106                 tp->ucopy.dma_cookie = dma_cookie;
4107                 copied_early = 1;
4108
4109                 tp->ucopy.len -= chunk;
4110                 tp->copied_seq += chunk;
4111                 tcp_rcv_space_adjust(sk);
4112
4113                 if ((tp->ucopy.len == 0) ||
4114                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4115                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4116                         tp->ucopy.wakeup = 1;
4117                         sk->sk_data_ready(sk, 0);
4118                 }
4119         } else if (chunk > 0) {
4120                 tp->ucopy.wakeup = 1;
4121                 sk->sk_data_ready(sk, 0);
4122         }
4123 out:
4124         return copied_early;
4125 }
4126 #endif /* CONFIG_NET_DMA */
4127
4128 /*
4129  *      TCP receive function for the ESTABLISHED state.
4130  *
4131  *      It is split into a fast path and a slow path. The fast path is
4132  *      disabled when:
4133  *      - A zero window was announced from us - zero window probing
4134  *        is only handled properly in the slow path.
4135  *      - Out of order segments arrived.
4136  *      - Urgent data is expected.
4137  *      - There is no buffer space left
4138  *      - Unexpected TCP flags/window values/header lengths are received
4139  *        (detected by checking the TCP header against pred_flags)
4140  *      - Data is sent in both directions. Fast path only supports pure senders
4141  *        or pure receivers (this means either the sequence number or the ack
4142  *        value must stay constant)
4143  *      - Unexpected TCP option.
4144  *
4145  *      When these conditions are not satisfied it drops into a standard
4146  *      receive procedure patterned after RFC793 to handle all cases.
4147  *      The first three cases are guaranteed by proper pred_flags setting,
4148  *      the rest is checked inline. Fast processing is turned on in
4149  *      tcp_data_queue when everything is OK.
4150  */
4151 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4152                         struct tcphdr *th, unsigned len)
4153 {
4154         struct tcp_sock *tp = tcp_sk(sk);
4155
4156         /*
4157          *      Header prediction.
4158          *      The code loosely follows the one in the famous
4159          *      "30 instruction TCP receive" Van Jacobson mail.
4160          *
4161          *      Van's trick is to deposit buffers into socket queue
4162          *      on a device interrupt, to call tcp_recv function
4163          *      on the receive process context and checksum and copy
4164          *      the buffer to user space. smart...
4165          *
4166          *      Our current scheme is not silly either but we take the
4167          *      extra cost of the net_bh soft interrupt processing...
4168          *      We do checksum and copy also but from device to kernel.
4169          */
4170
4171         tp->rx_opt.saw_tstamp = 0;
4172
4173         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4174          *      if header_prediction is to be made
4175          *      'S' will always be tp->tcp_header_len >> 2
4176          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4177          *  turn it off (when there are holes in the receive
4178          *       space for instance)
4179          *      PSH flag is ignored.
4180          */
4181
4182         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4183                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4184                 int tcp_header_len = tp->tcp_header_len;
4185
4186                 /* Timestamp header prediction: tcp_header_len
4187                  * is automatically equal to th->doff*4 due to pred_flags
4188                  * match.
4189                  */
4190
4191                 /* Check timestamp */
4192                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4193                         __be32 *ptr = (__be32 *)(th + 1);
4194
4195                         /* No? Slow path! */
4196                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4197                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4198                                 goto slow_path;
4199
4200                         tp->rx_opt.saw_tstamp = 1;
4201                         ++ptr;
4202                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4203                         ++ptr;
4204                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4205
4206                         /* If PAWS failed, check it more carefully in slow path */
4207                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4208                                 goto slow_path;
4209
4210                         /* DO NOT update ts_recent here, if checksum fails
4211                          * and timestamp was corrupted part, it will result
4212                          * in a hung connection since we will drop all
4213                          * future packets due to the PAWS test.
4214                          */
4215                 }
4216
4217                 if (len <= tcp_header_len) {
4218                         /* Bulk data transfer: sender */
4219                         if (len == tcp_header_len) {
4220                                 /* Predicted packet is in window by definition.
4221                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4222                                  * Hence, check seq<=rcv_wup reduces to:
4223                                  */
4224                                 if (tcp_header_len ==
4225                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4226                                     tp->rcv_nxt == tp->rcv_wup)
4227                                         tcp_store_ts_recent(tp);
4228
4229                                 /* We know that such packets are checksummed
4230                                  * on entry.
4231                                  */
4232                                 tcp_ack(sk, skb, 0);
4233                                 __kfree_skb(skb);
4234                                 tcp_data_snd_check(sk);
4235                                 return 0;
4236                         } else { /* Header too small */
4237                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4238                                 goto discard;
4239                         }
4240                 } else {
4241                         int eaten = 0;
4242                         int copied_early = 0;
4243
4244                         if (tp->copied_seq == tp->rcv_nxt &&
4245                             len - tcp_header_len <= tp->ucopy.len) {
4246 #ifdef CONFIG_NET_DMA
4247                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4248                                         copied_early = 1;
4249                                         eaten = 1;
4250                                 }
4251 #endif
4252                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4253                                         __set_current_state(TASK_RUNNING);
4254
4255                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4256                                                 eaten = 1;
4257                                 }
4258                                 if (eaten) {
4259                                         /* Predicted packet is in window by definition.
4260                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4261                                          * Hence, check seq<=rcv_wup reduces to:
4262                                          */
4263                                         if (tcp_header_len ==
4264                                             (sizeof(struct tcphdr) +
4265                                              TCPOLEN_TSTAMP_ALIGNED) &&
4266                                             tp->rcv_nxt == tp->rcv_wup)
4267                                                 tcp_store_ts_recent(tp);
4268
4269                                         tcp_rcv_rtt_measure_ts(sk, skb);
4270
4271                                         __skb_pull(skb, tcp_header_len);
4272                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4273                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4274                                 }
4275                                 if (copied_early)
4276                                         tcp_cleanup_rbuf(sk, skb->len);
4277                         }
4278                         if (!eaten) {
4279                                 if (tcp_checksum_complete_user(sk, skb))
4280                                         goto csum_error;
4281
4282                                 /* Predicted packet is in window by definition.
4283                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4284                                  * Hence, check seq<=rcv_wup reduces to:
4285                                  */
4286                                 if (tcp_header_len ==
4287                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4288                                     tp->rcv_nxt == tp->rcv_wup)
4289                                         tcp_store_ts_recent(tp);
4290
4291                                 tcp_rcv_rtt_measure_ts(sk, skb);
4292
4293                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4294                                         goto step5;
4295
4296                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4297
4298                                 /* Bulk data transfer: receiver */
4299                                 __skb_pull(skb,tcp_header_len);
4300                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4301                                 sk_stream_set_owner_r(skb, sk);
4302                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4303                         }
4304
4305                         tcp_event_data_recv(sk, skb);
4306
4307                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4308                                 /* Well, only one small jumplet in fast path... */
4309                                 tcp_ack(sk, skb, FLAG_DATA);
4310                                 tcp_data_snd_check(sk);
4311                                 if (!inet_csk_ack_scheduled(sk))
4312                                         goto no_ack;
4313                         }
4314
4315                         __tcp_ack_snd_check(sk, 0);
4316 no_ack:
4317 #ifdef CONFIG_NET_DMA
4318                         if (copied_early)
4319                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4320                         else
4321 #endif
4322                         if (eaten)
4323                                 __kfree_skb(skb);
4324                         else
4325                                 sk->sk_data_ready(sk, 0);
4326                         return 0;
4327                 }
4328         }
4329
4330 slow_path:
4331         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4332                 goto csum_error;
4333
4334         /*
4335          * RFC1323: H1. Apply PAWS check first.
4336          */
4337         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4338             tcp_paws_discard(sk, skb)) {
4339                 if (!th->rst) {
4340                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4341                         tcp_send_dupack(sk, skb);
4342                         goto discard;
4343                 }
4344                 /* Resets are accepted even if PAWS failed.
4345
4346                    ts_recent update must be made after we are sure
4347                    that the packet is in window.
4348                  */
4349         }
4350
4351         /*
4352          *      Standard slow path.
4353          */
4354
4355         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4356                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4357                  * (RST) segments are validated by checking their SEQ-fields."
4358                  * And page 69: "If an incoming segment is not acceptable,
4359                  * an acknowledgment should be sent in reply (unless the RST bit
4360                  * is set, if so drop the segment and return)".
4361                  */
4362                 if (!th->rst)
4363                         tcp_send_dupack(sk, skb);
4364                 goto discard;
4365         }
4366
4367         if (th->rst) {
4368                 tcp_reset(sk);
4369                 goto discard;
4370         }
4371
4372         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4373
4374         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4375                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4376                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4377                 tcp_reset(sk);
4378                 return 1;
4379         }
4380
4381 step5:
4382         if (th->ack)
4383                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4384
4385         tcp_rcv_rtt_measure_ts(sk, skb);
4386
4387         /* Process urgent data. */
4388         tcp_urg(sk, skb, th);
4389
4390         /* step 7: process the segment text */
4391         tcp_data_queue(sk, skb);
4392
4393         tcp_data_snd_check(sk);
4394         tcp_ack_snd_check(sk);
4395         return 0;
4396
4397 csum_error:
4398         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4399
4400 discard:
4401         __kfree_skb(skb);
4402         return 0;
4403 }
4404
4405 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4406                                          struct tcphdr *th, unsigned len)
4407 {
4408         struct tcp_sock *tp = tcp_sk(sk);
4409         struct inet_connection_sock *icsk = inet_csk(sk);
4410         int saved_clamp = tp->rx_opt.mss_clamp;
4411
4412         tcp_parse_options(skb, &tp->rx_opt, 0);
4413
4414         if (th->ack) {
4415                 /* rfc793:
4416                  * "If the state is SYN-SENT then
4417                  *    first check the ACK bit
4418                  *      If the ACK bit is set
4419                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4420                  *        a reset (unless the RST bit is set, if so drop
4421                  *        the segment and return)"
4422                  *
4423                  *  We do not send data with SYN, so that RFC-correct
4424                  *  test reduces to:
4425                  */
4426                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4427                         goto reset_and_undo;
4428
4429                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4430                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4431                              tcp_time_stamp)) {
4432                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4433                         goto reset_and_undo;
4434                 }
4435
4436                 /* Now ACK is acceptable.
4437                  *
4438                  * "If the RST bit is set
4439                  *    If the ACK was acceptable then signal the user "error:
4440                  *    connection reset", drop the segment, enter CLOSED state,
4441                  *    delete TCB, and return."
4442                  */
4443
4444                 if (th->rst) {
4445                         tcp_reset(sk);
4446                         goto discard;
4447                 }
4448
4449                 /* rfc793:
4450                  *   "fifth, if neither of the SYN or RST bits is set then
4451                  *    drop the segment and return."
4452                  *
4453                  *    See note below!
4454                  *                                        --ANK(990513)
4455                  */
4456                 if (!th->syn)
4457                         goto discard_and_undo;
4458
4459                 /* rfc793:
4460                  *   "If the SYN bit is on ...
4461                  *    are acceptable then ...
4462                  *    (our SYN has been ACKed), change the connection
4463                  *    state to ESTABLISHED..."
4464                  */
4465
4466                 TCP_ECN_rcv_synack(tp, th);
4467
4468                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4469                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4470
4471                 /* Ok.. it's good. Set up sequence numbers and
4472                  * move to established.
4473                  */
4474                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4475                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4476
4477                 /* RFC1323: The window in SYN & SYN/ACK segments is
4478                  * never scaled.
4479                  */
4480                 tp->snd_wnd = ntohs(th->window);
4481                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4482
4483                 if (!tp->rx_opt.wscale_ok) {
4484                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4485                         tp->window_clamp = min(tp->window_clamp, 65535U);
4486                 }
4487
4488                 if (tp->rx_opt.saw_tstamp) {
4489                         tp->rx_opt.tstamp_ok       = 1;
4490                         tp->tcp_header_len =
4491                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4492                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4493                         tcp_store_ts_recent(tp);
4494                 } else {
4495                         tp->tcp_header_len = sizeof(struct tcphdr);
4496                 }
4497
4498                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4499                         tp->rx_opt.sack_ok |= 2;
4500
4501                 tcp_mtup_init(sk);
4502                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4503                 tcp_initialize_rcv_mss(sk);
4504
4505                 /* Remember, tcp_poll() does not lock socket!
4506                  * Change state from SYN-SENT only after copied_seq
4507                  * is initialized. */
4508                 tp->copied_seq = tp->rcv_nxt;
4509                 smp_mb();
4510                 tcp_set_state(sk, TCP_ESTABLISHED);
4511
4512                 security_inet_conn_established(sk, skb);
4513
4514                 /* Make sure socket is routed, for correct metrics.  */
4515                 icsk->icsk_af_ops->rebuild_header(sk);
4516
4517                 tcp_init_metrics(sk);
4518
4519                 tcp_init_congestion_control(sk);
4520
4521                 /* Prevent spurious tcp_cwnd_restart() on first data
4522                  * packet.
4523                  */
4524                 tp->lsndtime = tcp_time_stamp;
4525
4526                 tcp_init_buffer_space(sk);
4527
4528                 if (sock_flag(sk, SOCK_KEEPOPEN))
4529                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4530
4531                 if (!tp->rx_opt.snd_wscale)
4532                         __tcp_fast_path_on(tp, tp->snd_wnd);
4533                 else
4534                         tp->pred_flags = 0;
4535
4536                 if (!sock_flag(sk, SOCK_DEAD)) {
4537                         sk->sk_state_change(sk);
4538                         sk_wake_async(sk, 0, POLL_OUT);
4539                 }
4540
4541                 if (sk->sk_write_pending ||
4542                     icsk->icsk_accept_queue.rskq_defer_accept ||
4543                     icsk->icsk_ack.pingpong) {
4544                         /* Save one ACK. Data will be ready after
4545                          * several ticks, if write_pending is set.
4546                          *
4547                          * It may be deleted, but with this feature tcpdumps
4548                          * look so _wonderfully_ clever, that I was not able
4549                          * to stand against the temptation 8)     --ANK
4550                          */
4551                         inet_csk_schedule_ack(sk);
4552                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4553                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4554                         tcp_incr_quickack(sk);
4555                         tcp_enter_quickack_mode(sk);
4556                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4557                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4558
4559 discard:
4560                         __kfree_skb(skb);
4561                         return 0;
4562                 } else {
4563                         tcp_send_ack(sk);
4564                 }
4565                 return -1;
4566         }
4567
4568         /* No ACK in the segment */
4569
4570         if (th->rst) {
4571                 /* rfc793:
4572                  * "If the RST bit is set
4573                  *
4574                  *      Otherwise (no ACK) drop the segment and return."
4575                  */
4576
4577                 goto discard_and_undo;
4578         }
4579
4580         /* PAWS check. */
4581         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4582                 goto discard_and_undo;
4583
4584         if (th->syn) {
4585                 /* We see SYN without ACK. It is attempt of
4586                  * simultaneous connect with crossed SYNs.
4587                  * Particularly, it can be connect to self.
4588                  */
4589                 tcp_set_state(sk, TCP_SYN_RECV);
4590
4591                 if (tp->rx_opt.saw_tstamp) {
4592                         tp->rx_opt.tstamp_ok = 1;
4593                         tcp_store_ts_recent(tp);
4594                         tp->tcp_header_len =
4595                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4596                 } else {
4597                         tp->tcp_header_len = sizeof(struct tcphdr);
4598                 }
4599
4600                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4601                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4602
4603                 /* RFC1323: The window in SYN & SYN/ACK segments is
4604                  * never scaled.
4605                  */
4606                 tp->snd_wnd    = ntohs(th->window);
4607                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4608                 tp->max_window = tp->snd_wnd;
4609
4610                 TCP_ECN_rcv_syn(tp, th);
4611
4612                 tcp_mtup_init(sk);
4613                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4614                 tcp_initialize_rcv_mss(sk);
4615
4616
4617                 tcp_send_synack(sk);
4618 #if 0
4619                 /* Note, we could accept data and URG from this segment.
4620                  * There are no obstacles to make this.
4621                  *
4622                  * However, if we ignore data in ACKless segments sometimes,
4623                  * we have no reasons to accept it sometimes.
4624                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4625                  * is not flawless. So, discard packet for sanity.
4626                  * Uncomment this return to process the data.
4627                  */
4628                 return -1;
4629 #else
4630                 goto discard;
4631 #endif
4632         }
4633         /* "fifth, if neither of the SYN or RST bits is set then
4634          * drop the segment and return."
4635          */
4636
4637 discard_and_undo:
4638         tcp_clear_options(&tp->rx_opt);
4639         tp->rx_opt.mss_clamp = saved_clamp;
4640         goto discard;
4641
4642 reset_and_undo:
4643         tcp_clear_options(&tp->rx_opt);
4644         tp->rx_opt.mss_clamp = saved_clamp;
4645         return 1;
4646 }
4647
4648
4649 /*
4650  *      This function implements the receiving procedure of RFC 793 for
4651  *      all states except ESTABLISHED and TIME_WAIT.
4652  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4653  *      address independent.
4654  */
4655
4656 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4657                           struct tcphdr *th, unsigned len)
4658 {
4659         struct tcp_sock *tp = tcp_sk(sk);
4660         struct inet_connection_sock *icsk = inet_csk(sk);
4661         int queued = 0;
4662
4663         tp->rx_opt.saw_tstamp = 0;
4664
4665         switch (sk->sk_state) {
4666         case TCP_CLOSE:
4667                 goto discard;
4668
4669         case TCP_LISTEN:
4670                 if (th->ack)
4671                         return 1;
4672
4673                 if (th->rst)
4674                         goto discard;
4675
4676                 if (th->syn) {
4677                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4678                                 return 1;
4679
4680                         /* Now we have several options: In theory there is
4681                          * nothing else in the frame. KA9Q has an option to
4682                          * send data with the syn, BSD accepts data with the
4683                          * syn up to the [to be] advertised window and
4684                          * Solaris 2.1 gives you a protocol error. For now
4685                          * we just ignore it, that fits the spec precisely
4686                          * and avoids incompatibilities. It would be nice in
4687                          * future to drop through and process the data.
4688                          *
4689                          * Now that TTCP is starting to be used we ought to
4690                          * queue this data.
4691                          * But, this leaves one open to an easy denial of
4692                          * service attack, and SYN cookies can't defend
4693                          * against this problem. So, we drop the data
4694                          * in the interest of security over speed unless
4695                          * it's still in use.
4696                          */
4697                         kfree_skb(skb);
4698                         return 0;
4699                 }
4700                 goto discard;
4701
4702         case TCP_SYN_SENT:
4703                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4704                 if (queued >= 0)
4705                         return queued;
4706
4707                 /* Do step6 onward by hand. */
4708                 tcp_urg(sk, skb, th);
4709                 __kfree_skb(skb);
4710                 tcp_data_snd_check(sk);
4711                 return 0;
4712         }
4713
4714         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4715             tcp_paws_discard(sk, skb)) {
4716                 if (!th->rst) {
4717                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4718                         tcp_send_dupack(sk, skb);
4719                         goto discard;
4720                 }
4721                 /* Reset is accepted even if it did not pass PAWS. */
4722         }
4723
4724         /* step 1: check sequence number */
4725         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4726                 if (!th->rst)
4727                         tcp_send_dupack(sk, skb);
4728                 goto discard;
4729         }
4730
4731         /* step 2: check RST bit */
4732         if (th->rst) {
4733                 tcp_reset(sk);
4734                 goto discard;
4735         }
4736
4737         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4738
4739         /* step 3: check security and precedence [ignored] */
4740
4741         /*      step 4:
4742          *
4743          *      Check for a SYN in window.
4744          */
4745         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4746                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4747                 tcp_reset(sk);
4748                 return 1;
4749         }
4750
4751         /* step 5: check the ACK field */
4752         if (th->ack) {
4753                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4754
4755                 switch (sk->sk_state) {
4756                 case TCP_SYN_RECV:
4757                         if (acceptable) {
4758                                 tp->copied_seq = tp->rcv_nxt;
4759                                 smp_mb();
4760                                 tcp_set_state(sk, TCP_ESTABLISHED);
4761                                 sk->sk_state_change(sk);
4762
4763                                 /* Note, that this wakeup is only for marginal
4764                                  * crossed SYN case. Passively open sockets
4765                                  * are not waked up, because sk->sk_sleep ==
4766                                  * NULL and sk->sk_socket == NULL.
4767                                  */
4768                                 if (sk->sk_socket) {
4769                                         sk_wake_async(sk,0,POLL_OUT);
4770                                 }
4771
4772                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4773                                 tp->snd_wnd = ntohs(th->window) <<
4774                                               tp->rx_opt.snd_wscale;
4775                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4776                                             TCP_SKB_CB(skb)->seq);
4777
4778                                 /* tcp_ack considers this ACK as duplicate
4779                                  * and does not calculate rtt.
4780                                  * Fix it at least with timestamps.
4781                                  */
4782                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4783                                     !tp->srtt)
4784                                         tcp_ack_saw_tstamp(sk, 0);
4785
4786                                 if (tp->rx_opt.tstamp_ok)
4787                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4788
4789                                 /* Make sure socket is routed, for
4790                                  * correct metrics.
4791                                  */
4792                                 icsk->icsk_af_ops->rebuild_header(sk);
4793
4794                                 tcp_init_metrics(sk);
4795
4796                                 tcp_init_congestion_control(sk);
4797
4798                                 /* Prevent spurious tcp_cwnd_restart() on
4799                                  * first data packet.
4800                                  */
4801                                 tp->lsndtime = tcp_time_stamp;
4802
4803                                 tcp_mtup_init(sk);
4804                                 tcp_initialize_rcv_mss(sk);
4805                                 tcp_init_buffer_space(sk);
4806                                 tcp_fast_path_on(tp);
4807                         } else {
4808                                 return 1;
4809                         }
4810                         break;
4811
4812                 case TCP_FIN_WAIT1:
4813                         if (tp->snd_una == tp->write_seq) {
4814                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4815                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4816                                 dst_confirm(sk->sk_dst_cache);
4817
4818                                 if (!sock_flag(sk, SOCK_DEAD))
4819                                         /* Wake up lingering close() */
4820                                         sk->sk_state_change(sk);
4821                                 else {
4822                                         int tmo;
4823
4824                                         if (tp->linger2 < 0 ||
4825                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4826                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4827                                                 tcp_done(sk);
4828                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4829                                                 return 1;
4830                                         }
4831
4832                                         tmo = tcp_fin_time(sk);
4833                                         if (tmo > TCP_TIMEWAIT_LEN) {
4834                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4835                                         } else if (th->fin || sock_owned_by_user(sk)) {
4836                                                 /* Bad case. We could lose such FIN otherwise.
4837                                                  * It is not a big problem, but it looks confusing
4838                                                  * and not so rare event. We still can lose it now,
4839                                                  * if it spins in bh_lock_sock(), but it is really
4840                                                  * marginal case.
4841                                                  */
4842                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4843                                         } else {
4844                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4845                                                 goto discard;
4846                                         }
4847                                 }
4848                         }
4849                         break;
4850
4851                 case TCP_CLOSING:
4852                         if (tp->snd_una == tp->write_seq) {
4853                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4854                                 goto discard;
4855                         }
4856                         break;
4857
4858                 case TCP_LAST_ACK:
4859                         if (tp->snd_una == tp->write_seq) {
4860                                 tcp_update_metrics(sk);
4861                                 tcp_done(sk);
4862                                 goto discard;
4863                         }
4864                         break;
4865                 }
4866         } else
4867                 goto discard;
4868
4869         /* step 6: check the URG bit */
4870         tcp_urg(sk, skb, th);
4871
4872         /* step 7: process the segment text */
4873         switch (sk->sk_state) {
4874         case TCP_CLOSE_WAIT:
4875         case TCP_CLOSING:
4876         case TCP_LAST_ACK:
4877                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4878                         break;
4879         case TCP_FIN_WAIT1:
4880         case TCP_FIN_WAIT2:
4881                 /* RFC 793 says to queue data in these states,
4882                  * RFC 1122 says we MUST send a reset.
4883                  * BSD 4.4 also does reset.
4884                  */
4885                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4886                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4887                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4888                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4889                                 tcp_reset(sk);
4890                                 return 1;
4891                         }
4892                 }
4893                 /* Fall through */
4894         case TCP_ESTABLISHED:
4895                 tcp_data_queue(sk, skb);
4896                 queued = 1;
4897                 break;
4898         }
4899
4900         /* tcp_data could move socket to TIME-WAIT */
4901         if (sk->sk_state != TCP_CLOSE) {
4902                 tcp_data_snd_check(sk);
4903                 tcp_ack_snd_check(sk);
4904         }
4905
4906         if (!queued) {
4907 discard:
4908                 __kfree_skb(skb);
4909         }
4910         return 0;
4911 }
4912
4913 EXPORT_SYMBOL(sysctl_tcp_ecn);
4914 EXPORT_SYMBOL(sysctl_tcp_reordering);
4915 EXPORT_SYMBOL(tcp_parse_options);
4916 EXPORT_SYMBOL(tcp_rcv_established);
4917 EXPORT_SYMBOL(tcp_rcv_state_process);
4918 EXPORT_SYMBOL(tcp_initialize_rcv_mss);