Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[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 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                 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                 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 (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
1062                 skb = cached_skb;
1063                 fack_count = cached_fack_count;
1064
1065                 /* Event "B" in the comment above. */
1066                 if (after(end_seq, tp->high_seq))
1067                         flag |= FLAG_DATA_LOST;
1068
1069                 tcp_for_write_queue_from(skb, sk) {
1070                         int in_sack, pcount;
1071                         u8 sacked;
1072
1073                         if (skb == tcp_send_head(sk))
1074                                 break;
1075
1076                         cached_skb = skb;
1077                         cached_fack_count = fack_count;
1078                         if (i == first_sack_index) {
1079                                 tp->fastpath_skb_hint = skb;
1080                                 tp->fastpath_cnt_hint = fack_count;
1081                         }
1082
1083                         /* The retransmission queue is always in order, so
1084                          * we can short-circuit the walk early.
1085                          */
1086                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1087                                 break;
1088
1089                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1090                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1091
1092                         pcount = tcp_skb_pcount(skb);
1093
1094                         if (pcount > 1 && !in_sack &&
1095                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1096                                 unsigned int pkt_len;
1097
1098                                 in_sack = !after(start_seq,
1099                                                  TCP_SKB_CB(skb)->seq);
1100
1101                                 if (!in_sack)
1102                                         pkt_len = (start_seq -
1103                                                    TCP_SKB_CB(skb)->seq);
1104                                 else
1105                                         pkt_len = (end_seq -
1106                                                    TCP_SKB_CB(skb)->seq);
1107                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1108                                         break;
1109                                 pcount = tcp_skb_pcount(skb);
1110                         }
1111
1112                         fack_count += pcount;
1113
1114                         sacked = TCP_SKB_CB(skb)->sacked;
1115
1116                         /* Account D-SACK for retransmitted packet. */
1117                         if ((dup_sack && in_sack) &&
1118                             (sacked & TCPCB_RETRANS) &&
1119                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1120                                 tp->undo_retrans--;
1121
1122                         /* The frame is ACKed. */
1123                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1124                                 if (sacked&TCPCB_RETRANS) {
1125                                         if ((dup_sack && in_sack) &&
1126                                             (sacked&TCPCB_SACKED_ACKED))
1127                                                 reord = min(fack_count, reord);
1128                                 } else {
1129                                         /* If it was in a hole, we detected reordering. */
1130                                         if (fack_count < prior_fackets &&
1131                                             !(sacked&TCPCB_SACKED_ACKED))
1132                                                 reord = min(fack_count, reord);
1133                                 }
1134
1135                                 /* Nothing to do; acked frame is about to be dropped. */
1136                                 continue;
1137                         }
1138
1139                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1140                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1141                             (!lost_retrans || after(end_seq, lost_retrans)))
1142                                 lost_retrans = end_seq;
1143
1144                         if (!in_sack)
1145                                 continue;
1146
1147                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1148                                 if (sacked & TCPCB_SACKED_RETRANS) {
1149                                         /* If the segment is not tagged as lost,
1150                                          * we do not clear RETRANS, believing
1151                                          * that retransmission is still in flight.
1152                                          */
1153                                         if (sacked & TCPCB_LOST) {
1154                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1155                                                 tp->lost_out -= tcp_skb_pcount(skb);
1156                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1157
1158                                                 /* clear lost hint */
1159                                                 tp->retransmit_skb_hint = NULL;
1160                                         }
1161                                 } else {
1162                                         /* New sack for not retransmitted frame,
1163                                          * which was in hole. It is reordering.
1164                                          */
1165                                         if (!(sacked & TCPCB_RETRANS) &&
1166                                             fack_count < prior_fackets)
1167                                                 reord = min(fack_count, reord);
1168
1169                                         if (sacked & TCPCB_LOST) {
1170                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1171                                                 tp->lost_out -= tcp_skb_pcount(skb);
1172
1173                                                 /* clear lost hint */
1174                                                 tp->retransmit_skb_hint = NULL;
1175                                         }
1176                                         /* SACK enhanced F-RTO detection.
1177                                          * Set flag if and only if non-rexmitted
1178                                          * segments below frto_highmark are
1179                                          * SACKed (RFC4138; Appendix B).
1180                                          * Clearing correct due to in-order walk
1181                                          */
1182                                         if (after(end_seq, tp->frto_highmark)) {
1183                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1184                                         } else {
1185                                                 if (!(sacked & TCPCB_RETRANS))
1186                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1187                                         }
1188                                 }
1189
1190                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1191                                 flag |= FLAG_DATA_SACKED;
1192                                 tp->sacked_out += tcp_skb_pcount(skb);
1193
1194                                 if (fack_count > tp->fackets_out)
1195                                         tp->fackets_out = fack_count;
1196                         } else {
1197                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1198                                         reord = min(fack_count, reord);
1199                         }
1200
1201                         /* D-SACK. We can detect redundant retransmission
1202                          * in S|R and plain R frames and clear it.
1203                          * undo_retrans is decreased above, L|R frames
1204                          * are accounted above as well.
1205                          */
1206                         if (dup_sack &&
1207                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1208                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1209                                 tp->retrans_out -= tcp_skb_pcount(skb);
1210                                 tp->retransmit_skb_hint = NULL;
1211                         }
1212                 }
1213         }
1214
1215         /* Check for lost retransmit. This superb idea is
1216          * borrowed from "ratehalving". Event "C".
1217          * Later note: FACK people cheated me again 8),
1218          * we have to account for reordering! Ugly,
1219          * but should help.
1220          */
1221         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1222                 struct sk_buff *skb;
1223
1224                 tcp_for_write_queue(skb, sk) {
1225                         if (skb == tcp_send_head(sk))
1226                                 break;
1227                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1228                                 break;
1229                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1230                                 continue;
1231                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1232                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1233                             (IsFack(tp) ||
1234                              !before(lost_retrans,
1235                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1236                                      tp->mss_cache))) {
1237                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1238                                 tp->retrans_out -= tcp_skb_pcount(skb);
1239
1240                                 /* clear lost hint */
1241                                 tp->retransmit_skb_hint = NULL;
1242
1243                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1244                                         tp->lost_out += tcp_skb_pcount(skb);
1245                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1246                                         flag |= FLAG_DATA_SACKED;
1247                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1248                                 }
1249                         }
1250                 }
1251         }
1252
1253         tp->left_out = tp->sacked_out + tp->lost_out;
1254
1255         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1256             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1257                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1258
1259 #if FASTRETRANS_DEBUG > 0
1260         BUG_TRAP((int)tp->sacked_out >= 0);
1261         BUG_TRAP((int)tp->lost_out >= 0);
1262         BUG_TRAP((int)tp->retrans_out >= 0);
1263         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1264 #endif
1265         return flag;
1266 }
1267
1268 /* F-RTO can only be used if TCP has never retransmitted anything other than
1269  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1270  */
1271 int tcp_use_frto(struct sock *sk)
1272 {
1273         const struct tcp_sock *tp = tcp_sk(sk);
1274         struct sk_buff *skb;
1275
1276         if (!sysctl_tcp_frto)
1277                 return 0;
1278
1279         if (IsSackFrto())
1280                 return 1;
1281
1282         /* Avoid expensive walking of rexmit queue if possible */
1283         if (tp->retrans_out > 1)
1284                 return 0;
1285
1286         skb = tcp_write_queue_head(sk);
1287         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1288         tcp_for_write_queue_from(skb, sk) {
1289                 if (skb == tcp_send_head(sk))
1290                         break;
1291                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1292                         return 0;
1293                 /* Short-circuit when first non-SACKed skb has been checked */
1294                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1295                         break;
1296         }
1297         return 1;
1298 }
1299
1300 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1301  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1302  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1303  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1304  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1305  * bits are handled if the Loss state is really to be entered (in
1306  * tcp_enter_frto_loss).
1307  *
1308  * Do like tcp_enter_loss() would; when RTO expires the second time it
1309  * does:
1310  *  "Reduce ssthresh if it has not yet been made inside this window."
1311  */
1312 void tcp_enter_frto(struct sock *sk)
1313 {
1314         const struct inet_connection_sock *icsk = inet_csk(sk);
1315         struct tcp_sock *tp = tcp_sk(sk);
1316         struct sk_buff *skb;
1317
1318         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1319             tp->snd_una == tp->high_seq ||
1320             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1321              !icsk->icsk_retransmits)) {
1322                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1323                 /* Our state is too optimistic in ssthresh() call because cwnd
1324                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1325                  * recovery has not yet completed. Pattern would be this: RTO,
1326                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1327                  * up here twice).
1328                  * RFC4138 should be more specific on what to do, even though
1329                  * RTO is quite unlikely to occur after the first Cumulative ACK
1330                  * due to back-off and complexity of triggering events ...
1331                  */
1332                 if (tp->frto_counter) {
1333                         u32 stored_cwnd;
1334                         stored_cwnd = tp->snd_cwnd;
1335                         tp->snd_cwnd = 2;
1336                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1337                         tp->snd_cwnd = stored_cwnd;
1338                 } else {
1339                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1340                 }
1341                 /* ... in theory, cong.control module could do "any tricks" in
1342                  * ssthresh(), which means that ca_state, lost bits and lost_out
1343                  * counter would have to be faked before the call occurs. We
1344                  * consider that too expensive, unlikely and hacky, so modules
1345                  * using these in ssthresh() must deal these incompatibility
1346                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1347                  */
1348                 tcp_ca_event(sk, CA_EVENT_FRTO);
1349         }
1350
1351         tp->undo_marker = tp->snd_una;
1352         tp->undo_retrans = 0;
1353
1354         skb = tcp_write_queue_head(sk);
1355         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1356                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1357                 tp->retrans_out -= tcp_skb_pcount(skb);
1358         }
1359         tcp_sync_left_out(tp);
1360
1361         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1362          * The last condition is necessary at least in tp->frto_counter case.
1363          */
1364         if (IsSackFrto() && (tp->frto_counter ||
1365             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1366             after(tp->high_seq, tp->snd_una)) {
1367                 tp->frto_highmark = tp->high_seq;
1368         } else {
1369                 tp->frto_highmark = tp->snd_nxt;
1370         }
1371         tcp_set_ca_state(sk, TCP_CA_Disorder);
1372         tp->high_seq = tp->snd_nxt;
1373         tp->frto_counter = 1;
1374 }
1375
1376 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1377  * which indicates that we should follow the traditional RTO recovery,
1378  * i.e. mark everything lost and do go-back-N retransmission.
1379  */
1380 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1381 {
1382         struct tcp_sock *tp = tcp_sk(sk);
1383         struct sk_buff *skb;
1384         int cnt = 0;
1385
1386         tp->sacked_out = 0;
1387         tp->lost_out = 0;
1388         tp->fackets_out = 0;
1389         tp->retrans_out = 0;
1390
1391         tcp_for_write_queue(skb, sk) {
1392                 if (skb == tcp_send_head(sk))
1393                         break;
1394                 cnt += tcp_skb_pcount(skb);
1395                 /*
1396                  * Count the retransmission made on RTO correctly (only when
1397                  * waiting for the first ACK and did not get it)...
1398                  */
1399                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1400                         tp->retrans_out += tcp_skb_pcount(skb);
1401                         /* ...enter this if branch just for the first segment */
1402                         flag |= FLAG_DATA_ACKED;
1403                 } else {
1404                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1405                 }
1406                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1407
1408                         /* Do not mark those segments lost that were
1409                          * forward transmitted after RTO
1410                          */
1411                         if (!after(TCP_SKB_CB(skb)->end_seq,
1412                                    tp->frto_highmark)) {
1413                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1414                                 tp->lost_out += tcp_skb_pcount(skb);
1415                         }
1416                 } else {
1417                         tp->sacked_out += tcp_skb_pcount(skb);
1418                         tp->fackets_out = cnt;
1419                 }
1420         }
1421         tcp_sync_left_out(tp);
1422
1423         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1424         tp->snd_cwnd_cnt = 0;
1425         tp->snd_cwnd_stamp = tcp_time_stamp;
1426         tp->undo_marker = 0;
1427         tp->frto_counter = 0;
1428
1429         tp->reordering = min_t(unsigned int, tp->reordering,
1430                                              sysctl_tcp_reordering);
1431         tcp_set_ca_state(sk, TCP_CA_Loss);
1432         tp->high_seq = tp->frto_highmark;
1433         TCP_ECN_queue_cwr(tp);
1434
1435         clear_all_retrans_hints(tp);
1436 }
1437
1438 void tcp_clear_retrans(struct tcp_sock *tp)
1439 {
1440         tp->left_out = 0;
1441         tp->retrans_out = 0;
1442
1443         tp->fackets_out = 0;
1444         tp->sacked_out = 0;
1445         tp->lost_out = 0;
1446
1447         tp->undo_marker = 0;
1448         tp->undo_retrans = 0;
1449 }
1450
1451 /* Enter Loss state. If "how" is not zero, forget all SACK information
1452  * and reset tags completely, otherwise preserve SACKs. If receiver
1453  * dropped its ofo queue, we will know this due to reneging detection.
1454  */
1455 void tcp_enter_loss(struct sock *sk, int how)
1456 {
1457         const struct inet_connection_sock *icsk = inet_csk(sk);
1458         struct tcp_sock *tp = tcp_sk(sk);
1459         struct sk_buff *skb;
1460         int cnt = 0;
1461
1462         /* Reduce ssthresh if it has not yet been made inside this window. */
1463         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1464             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1465                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1466                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1467                 tcp_ca_event(sk, CA_EVENT_LOSS);
1468         }
1469         tp->snd_cwnd       = 1;
1470         tp->snd_cwnd_cnt   = 0;
1471         tp->snd_cwnd_stamp = tcp_time_stamp;
1472
1473         tp->bytes_acked = 0;
1474         tcp_clear_retrans(tp);
1475
1476         /* Push undo marker, if it was plain RTO and nothing
1477          * was retransmitted. */
1478         if (!how)
1479                 tp->undo_marker = tp->snd_una;
1480
1481         tcp_for_write_queue(skb, sk) {
1482                 if (skb == tcp_send_head(sk))
1483                         break;
1484                 cnt += tcp_skb_pcount(skb);
1485                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1486                         tp->undo_marker = 0;
1487                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1488                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1489                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1490                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1491                         tp->lost_out += tcp_skb_pcount(skb);
1492                 } else {
1493                         tp->sacked_out += tcp_skb_pcount(skb);
1494                         tp->fackets_out = cnt;
1495                 }
1496         }
1497         tcp_sync_left_out(tp);
1498
1499         tp->reordering = min_t(unsigned int, tp->reordering,
1500                                              sysctl_tcp_reordering);
1501         tcp_set_ca_state(sk, TCP_CA_Loss);
1502         tp->high_seq = tp->snd_nxt;
1503         TCP_ECN_queue_cwr(tp);
1504         /* Abort FRTO algorithm if one is in progress */
1505         tp->frto_counter = 0;
1506
1507         clear_all_retrans_hints(tp);
1508 }
1509
1510 static int tcp_check_sack_reneging(struct sock *sk)
1511 {
1512         struct sk_buff *skb;
1513
1514         /* If ACK arrived pointing to a remembered SACK,
1515          * it means that our remembered SACKs do not reflect
1516          * real state of receiver i.e.
1517          * receiver _host_ is heavily congested (or buggy).
1518          * Do processing similar to RTO timeout.
1519          */
1520         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1521             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1522                 struct inet_connection_sock *icsk = inet_csk(sk);
1523                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1524
1525                 tcp_enter_loss(sk, 1);
1526                 icsk->icsk_retransmits++;
1527                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1528                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1529                                           icsk->icsk_rto, TCP_RTO_MAX);
1530                 return 1;
1531         }
1532         return 0;
1533 }
1534
1535 static inline int tcp_fackets_out(struct tcp_sock *tp)
1536 {
1537         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1538 }
1539
1540 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1541 {
1542         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1543 }
1544
1545 static inline int tcp_head_timedout(struct sock *sk)
1546 {
1547         struct tcp_sock *tp = tcp_sk(sk);
1548
1549         return tp->packets_out &&
1550                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1551 }
1552
1553 /* Linux NewReno/SACK/FACK/ECN state machine.
1554  * --------------------------------------
1555  *
1556  * "Open"       Normal state, no dubious events, fast path.
1557  * "Disorder"   In all the respects it is "Open",
1558  *              but requires a bit more attention. It is entered when
1559  *              we see some SACKs or dupacks. It is split of "Open"
1560  *              mainly to move some processing from fast path to slow one.
1561  * "CWR"        CWND was reduced due to some Congestion Notification event.
1562  *              It can be ECN, ICMP source quench, local device congestion.
1563  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1564  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1565  *
1566  * tcp_fastretrans_alert() is entered:
1567  * - each incoming ACK, if state is not "Open"
1568  * - when arrived ACK is unusual, namely:
1569  *      * SACK
1570  *      * Duplicate ACK.
1571  *      * ECN ECE.
1572  *
1573  * Counting packets in flight is pretty simple.
1574  *
1575  *      in_flight = packets_out - left_out + retrans_out
1576  *
1577  *      packets_out is SND.NXT-SND.UNA counted in packets.
1578  *
1579  *      retrans_out is number of retransmitted segments.
1580  *
1581  *      left_out is number of segments left network, but not ACKed yet.
1582  *
1583  *              left_out = sacked_out + lost_out
1584  *
1585  *     sacked_out: Packets, which arrived to receiver out of order
1586  *                 and hence not ACKed. With SACKs this number is simply
1587  *                 amount of SACKed data. Even without SACKs
1588  *                 it is easy to give pretty reliable estimate of this number,
1589  *                 counting duplicate ACKs.
1590  *
1591  *       lost_out: Packets lost by network. TCP has no explicit
1592  *                 "loss notification" feedback from network (for now).
1593  *                 It means that this number can be only _guessed_.
1594  *                 Actually, it is the heuristics to predict lossage that
1595  *                 distinguishes different algorithms.
1596  *
1597  *      F.e. after RTO, when all the queue is considered as lost,
1598  *      lost_out = packets_out and in_flight = retrans_out.
1599  *
1600  *              Essentially, we have now two algorithms counting
1601  *              lost packets.
1602  *
1603  *              FACK: It is the simplest heuristics. As soon as we decided
1604  *              that something is lost, we decide that _all_ not SACKed
1605  *              packets until the most forward SACK are lost. I.e.
1606  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1607  *              It is absolutely correct estimate, if network does not reorder
1608  *              packets. And it loses any connection to reality when reordering
1609  *              takes place. We use FACK by default until reordering
1610  *              is suspected on the path to this destination.
1611  *
1612  *              NewReno: when Recovery is entered, we assume that one segment
1613  *              is lost (classic Reno). While we are in Recovery and
1614  *              a partial ACK arrives, we assume that one more packet
1615  *              is lost (NewReno). This heuristics are the same in NewReno
1616  *              and SACK.
1617  *
1618  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1619  *  deflation etc. CWND is real congestion window, never inflated, changes
1620  *  only according to classic VJ rules.
1621  *
1622  * Really tricky (and requiring careful tuning) part of algorithm
1623  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1624  * The first determines the moment _when_ we should reduce CWND and,
1625  * hence, slow down forward transmission. In fact, it determines the moment
1626  * when we decide that hole is caused by loss, rather than by a reorder.
1627  *
1628  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1629  * holes, caused by lost packets.
1630  *
1631  * And the most logically complicated part of algorithm is undo
1632  * heuristics. We detect false retransmits due to both too early
1633  * fast retransmit (reordering) and underestimated RTO, analyzing
1634  * timestamps and D-SACKs. When we detect that some segments were
1635  * retransmitted by mistake and CWND reduction was wrong, we undo
1636  * window reduction and abort recovery phase. This logic is hidden
1637  * inside several functions named tcp_try_undo_<something>.
1638  */
1639
1640 /* This function decides, when we should leave Disordered state
1641  * and enter Recovery phase, reducing congestion window.
1642  *
1643  * Main question: may we further continue forward transmission
1644  * with the same cwnd?
1645  */
1646 static int tcp_time_to_recover(struct sock *sk)
1647 {
1648         struct tcp_sock *tp = tcp_sk(sk);
1649         __u32 packets_out;
1650
1651         /* Do not perform any recovery during FRTO algorithm */
1652         if (tp->frto_counter)
1653                 return 0;
1654
1655         /* Trick#1: The loss is proven. */
1656         if (tp->lost_out)
1657                 return 1;
1658
1659         /* Not-A-Trick#2 : Classic rule... */
1660         if (tcp_fackets_out(tp) > tp->reordering)
1661                 return 1;
1662
1663         /* Trick#3 : when we use RFC2988 timer restart, fast
1664          * retransmit can be triggered by timeout of queue head.
1665          */
1666         if (tcp_head_timedout(sk))
1667                 return 1;
1668
1669         /* Trick#4: It is still not OK... But will it be useful to delay
1670          * recovery more?
1671          */
1672         packets_out = tp->packets_out;
1673         if (packets_out <= tp->reordering &&
1674             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1675             !tcp_may_send_now(sk)) {
1676                 /* We have nothing to send. This connection is limited
1677                  * either by receiver window or by application.
1678                  */
1679                 return 1;
1680         }
1681
1682         return 0;
1683 }
1684
1685 /* If we receive more dupacks than we expected counting segments
1686  * in assumption of absent reordering, interpret this as reordering.
1687  * The only another reason could be bug in receiver TCP.
1688  */
1689 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1690 {
1691         struct tcp_sock *tp = tcp_sk(sk);
1692         u32 holes;
1693
1694         holes = max(tp->lost_out, 1U);
1695         holes = min(holes, tp->packets_out);
1696
1697         if ((tp->sacked_out + holes) > tp->packets_out) {
1698                 tp->sacked_out = tp->packets_out - holes;
1699                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1700         }
1701 }
1702
1703 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1704
1705 static void tcp_add_reno_sack(struct sock *sk)
1706 {
1707         struct tcp_sock *tp = tcp_sk(sk);
1708         tp->sacked_out++;
1709         tcp_check_reno_reordering(sk, 0);
1710         tcp_sync_left_out(tp);
1711 }
1712
1713 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1714
1715 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1716 {
1717         struct tcp_sock *tp = tcp_sk(sk);
1718
1719         if (acked > 0) {
1720                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1721                 if (acked-1 >= tp->sacked_out)
1722                         tp->sacked_out = 0;
1723                 else
1724                         tp->sacked_out -= acked-1;
1725         }
1726         tcp_check_reno_reordering(sk, acked);
1727         tcp_sync_left_out(tp);
1728 }
1729
1730 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1731 {
1732         tp->sacked_out = 0;
1733         tp->left_out = tp->lost_out;
1734 }
1735
1736 /* Mark head of queue up as lost. */
1737 static void tcp_mark_head_lost(struct sock *sk,
1738                                int packets, u32 high_seq)
1739 {
1740         struct tcp_sock *tp = tcp_sk(sk);
1741         struct sk_buff *skb;
1742         int cnt;
1743
1744         BUG_TRAP(packets <= tp->packets_out);
1745         if (tp->lost_skb_hint) {
1746                 skb = tp->lost_skb_hint;
1747                 cnt = tp->lost_cnt_hint;
1748         } else {
1749                 skb = tcp_write_queue_head(sk);
1750                 cnt = 0;
1751         }
1752
1753         tcp_for_write_queue_from(skb, sk) {
1754                 if (skb == tcp_send_head(sk))
1755                         break;
1756                 /* TODO: do this better */
1757                 /* this is not the most efficient way to do this... */
1758                 tp->lost_skb_hint = skb;
1759                 tp->lost_cnt_hint = cnt;
1760                 cnt += tcp_skb_pcount(skb);
1761                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1762                         break;
1763                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1764                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1765                         tp->lost_out += tcp_skb_pcount(skb);
1766
1767                         /* clear xmit_retransmit_queue hints
1768                          *  if this is beyond hint */
1769                         if (tp->retransmit_skb_hint != NULL &&
1770                             before(TCP_SKB_CB(skb)->seq,
1771                                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1772                                 tp->retransmit_skb_hint = NULL;
1773
1774                 }
1775         }
1776         tcp_sync_left_out(tp);
1777 }
1778
1779 /* Account newly detected lost packet(s) */
1780
1781 static void tcp_update_scoreboard(struct sock *sk)
1782 {
1783         struct tcp_sock *tp = tcp_sk(sk);
1784
1785         if (IsFack(tp)) {
1786                 int lost = tp->fackets_out - tp->reordering;
1787                 if (lost <= 0)
1788                         lost = 1;
1789                 tcp_mark_head_lost(sk, lost, tp->high_seq);
1790         } else {
1791                 tcp_mark_head_lost(sk, 1, tp->high_seq);
1792         }
1793
1794         /* New heuristics: it is possible only after we switched
1795          * to restart timer each time when something is ACKed.
1796          * Hence, we can detect timed out packets during fast
1797          * retransmit without falling to slow start.
1798          */
1799         if (!IsReno(tp) && tcp_head_timedout(sk)) {
1800                 struct sk_buff *skb;
1801
1802                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1803                         : tcp_write_queue_head(sk);
1804
1805                 tcp_for_write_queue_from(skb, sk) {
1806                         if (skb == tcp_send_head(sk))
1807                                 break;
1808                         if (!tcp_skb_timedout(sk, skb))
1809                                 break;
1810
1811                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1812                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1813                                 tp->lost_out += tcp_skb_pcount(skb);
1814
1815                                 /* clear xmit_retrans hint */
1816                                 if (tp->retransmit_skb_hint &&
1817                                     before(TCP_SKB_CB(skb)->seq,
1818                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1819
1820                                         tp->retransmit_skb_hint = NULL;
1821                         }
1822                 }
1823
1824                 tp->scoreboard_skb_hint = skb;
1825
1826                 tcp_sync_left_out(tp);
1827         }
1828 }
1829
1830 /* CWND moderation, preventing bursts due to too big ACKs
1831  * in dubious situations.
1832  */
1833 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1834 {
1835         tp->snd_cwnd = min(tp->snd_cwnd,
1836                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1837         tp->snd_cwnd_stamp = tcp_time_stamp;
1838 }
1839
1840 /* Lower bound on congestion window is slow start threshold
1841  * unless congestion avoidance choice decides to overide it.
1842  */
1843 static inline u32 tcp_cwnd_min(const struct sock *sk)
1844 {
1845         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1846
1847         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1848 }
1849
1850 /* Decrease cwnd each second ack. */
1851 static void tcp_cwnd_down(struct sock *sk)
1852 {
1853         struct tcp_sock *tp = tcp_sk(sk);
1854         int decr = tp->snd_cwnd_cnt + 1;
1855
1856         tp->snd_cwnd_cnt = decr&1;
1857         decr >>= 1;
1858
1859         if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1860                 tp->snd_cwnd -= decr;
1861
1862         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1863         tp->snd_cwnd_stamp = tcp_time_stamp;
1864 }
1865
1866 /* Nothing was retransmitted or returned timestamp is less
1867  * than timestamp of the first retransmission.
1868  */
1869 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1870 {
1871         return !tp->retrans_stamp ||
1872                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1873                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1874 }
1875
1876 /* Undo procedures. */
1877
1878 #if FASTRETRANS_DEBUG > 1
1879 static void DBGUNDO(struct sock *sk, const char *msg)
1880 {
1881         struct tcp_sock *tp = tcp_sk(sk);
1882         struct inet_sock *inet = inet_sk(sk);
1883
1884         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1885                msg,
1886                NIPQUAD(inet->daddr), ntohs(inet->dport),
1887                tp->snd_cwnd, tp->left_out,
1888                tp->snd_ssthresh, tp->prior_ssthresh,
1889                tp->packets_out);
1890 }
1891 #else
1892 #define DBGUNDO(x...) do { } while (0)
1893 #endif
1894
1895 static void tcp_undo_cwr(struct sock *sk, const int undo)
1896 {
1897         struct tcp_sock *tp = tcp_sk(sk);
1898
1899         if (tp->prior_ssthresh) {
1900                 const struct inet_connection_sock *icsk = inet_csk(sk);
1901
1902                 if (icsk->icsk_ca_ops->undo_cwnd)
1903                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1904                 else
1905                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1906
1907                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1908                         tp->snd_ssthresh = tp->prior_ssthresh;
1909                         TCP_ECN_withdraw_cwr(tp);
1910                 }
1911         } else {
1912                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1913         }
1914         tcp_moderate_cwnd(tp);
1915         tp->snd_cwnd_stamp = tcp_time_stamp;
1916
1917         /* There is something screwy going on with the retrans hints after
1918            an undo */
1919         clear_all_retrans_hints(tp);
1920 }
1921
1922 static inline int tcp_may_undo(struct tcp_sock *tp)
1923 {
1924         return tp->undo_marker &&
1925                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1926 }
1927
1928 /* People celebrate: "We love our President!" */
1929 static int tcp_try_undo_recovery(struct sock *sk)
1930 {
1931         struct tcp_sock *tp = tcp_sk(sk);
1932
1933         if (tcp_may_undo(tp)) {
1934                 /* Happy end! We did not retransmit anything
1935                  * or our original transmission succeeded.
1936                  */
1937                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1938                 tcp_undo_cwr(sk, 1);
1939                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1940                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1941                 else
1942                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1943                 tp->undo_marker = 0;
1944         }
1945         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1946                 /* Hold old state until something *above* high_seq
1947                  * is ACKed. For Reno it is MUST to prevent false
1948                  * fast retransmits (RFC2582). SACK TCP is safe. */
1949                 tcp_moderate_cwnd(tp);
1950                 return 1;
1951         }
1952         tcp_set_ca_state(sk, TCP_CA_Open);
1953         return 0;
1954 }
1955
1956 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1957 static void tcp_try_undo_dsack(struct sock *sk)
1958 {
1959         struct tcp_sock *tp = tcp_sk(sk);
1960
1961         if (tp->undo_marker && !tp->undo_retrans) {
1962                 DBGUNDO(sk, "D-SACK");
1963                 tcp_undo_cwr(sk, 1);
1964                 tp->undo_marker = 0;
1965                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1966         }
1967 }
1968
1969 /* Undo during fast recovery after partial ACK. */
1970
1971 static int tcp_try_undo_partial(struct sock *sk, int acked)
1972 {
1973         struct tcp_sock *tp = tcp_sk(sk);
1974         /* Partial ACK arrived. Force Hoe's retransmit. */
1975         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1976
1977         if (tcp_may_undo(tp)) {
1978                 /* Plain luck! Hole if filled with delayed
1979                  * packet, rather than with a retransmit.
1980                  */
1981                 if (tp->retrans_out == 0)
1982                         tp->retrans_stamp = 0;
1983
1984                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1985
1986                 DBGUNDO(sk, "Hoe");
1987                 tcp_undo_cwr(sk, 0);
1988                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1989
1990                 /* So... Do not make Hoe's retransmit yet.
1991                  * If the first packet was delayed, the rest
1992                  * ones are most probably delayed as well.
1993                  */
1994                 failed = 0;
1995         }
1996         return failed;
1997 }
1998
1999 /* Undo during loss recovery after partial ACK. */
2000 static int tcp_try_undo_loss(struct sock *sk)
2001 {
2002         struct tcp_sock *tp = tcp_sk(sk);
2003
2004         if (tcp_may_undo(tp)) {
2005                 struct sk_buff *skb;
2006                 tcp_for_write_queue(skb, sk) {
2007                         if (skb == tcp_send_head(sk))
2008                                 break;
2009                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2010                 }
2011
2012                 clear_all_retrans_hints(tp);
2013
2014                 DBGUNDO(sk, "partial loss");
2015                 tp->lost_out = 0;
2016                 tp->left_out = tp->sacked_out;
2017                 tcp_undo_cwr(sk, 1);
2018                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2019                 inet_csk(sk)->icsk_retransmits = 0;
2020                 tp->undo_marker = 0;
2021                 if (!IsReno(tp))
2022                         tcp_set_ca_state(sk, TCP_CA_Open);
2023                 return 1;
2024         }
2025         return 0;
2026 }
2027
2028 static inline void tcp_complete_cwr(struct sock *sk)
2029 {
2030         struct tcp_sock *tp = tcp_sk(sk);
2031         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2032         tp->snd_cwnd_stamp = tcp_time_stamp;
2033         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2034 }
2035
2036 static void tcp_try_to_open(struct sock *sk, int flag)
2037 {
2038         struct tcp_sock *tp = tcp_sk(sk);
2039
2040         tp->left_out = tp->sacked_out;
2041
2042         if (tp->retrans_out == 0)
2043                 tp->retrans_stamp = 0;
2044
2045         if (flag&FLAG_ECE)
2046                 tcp_enter_cwr(sk, 1);
2047
2048         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2049                 int state = TCP_CA_Open;
2050
2051                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2052                         state = TCP_CA_Disorder;
2053
2054                 if (inet_csk(sk)->icsk_ca_state != state) {
2055                         tcp_set_ca_state(sk, state);
2056                         tp->high_seq = tp->snd_nxt;
2057                 }
2058                 tcp_moderate_cwnd(tp);
2059         } else {
2060                 tcp_cwnd_down(sk);
2061         }
2062 }
2063
2064 static void tcp_mtup_probe_failed(struct sock *sk)
2065 {
2066         struct inet_connection_sock *icsk = inet_csk(sk);
2067
2068         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2069         icsk->icsk_mtup.probe_size = 0;
2070 }
2071
2072 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2073 {
2074         struct tcp_sock *tp = tcp_sk(sk);
2075         struct inet_connection_sock *icsk = inet_csk(sk);
2076
2077         /* FIXME: breaks with very large cwnd */
2078         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2079         tp->snd_cwnd = tp->snd_cwnd *
2080                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2081                        icsk->icsk_mtup.probe_size;
2082         tp->snd_cwnd_cnt = 0;
2083         tp->snd_cwnd_stamp = tcp_time_stamp;
2084         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2085
2086         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2087         icsk->icsk_mtup.probe_size = 0;
2088         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2089 }
2090
2091
2092 /* Process an event, which can update packets-in-flight not trivially.
2093  * Main goal of this function is to calculate new estimate for left_out,
2094  * taking into account both packets sitting in receiver's buffer and
2095  * packets lost by network.
2096  *
2097  * Besides that it does CWND reduction, when packet loss is detected
2098  * and changes state of machine.
2099  *
2100  * It does _not_ decide what to send, it is made in function
2101  * tcp_xmit_retransmit_queue().
2102  */
2103 static void
2104 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
2105                       int prior_packets, int flag)
2106 {
2107         struct inet_connection_sock *icsk = inet_csk(sk);
2108         struct tcp_sock *tp = tcp_sk(sk);
2109         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
2110
2111         /* Some technical things:
2112          * 1. Reno does not count dupacks (sacked_out) automatically. */
2113         if (!tp->packets_out)
2114                 tp->sacked_out = 0;
2115         /* 2. SACK counts snd_fack in packets inaccurately. */
2116         if (tp->sacked_out == 0)
2117                 tp->fackets_out = 0;
2118
2119         /* Now state machine starts.
2120          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2121         if (flag&FLAG_ECE)
2122                 tp->prior_ssthresh = 0;
2123
2124         /* B. In all the states check for reneging SACKs. */
2125         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2126                 return;
2127
2128         /* C. Process data loss notification, provided it is valid. */
2129         if ((flag&FLAG_DATA_LOST) &&
2130             before(tp->snd_una, tp->high_seq) &&
2131             icsk->icsk_ca_state != TCP_CA_Open &&
2132             tp->fackets_out > tp->reordering) {
2133                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2134                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2135         }
2136
2137         /* D. Synchronize left_out to current state. */
2138         tcp_sync_left_out(tp);
2139
2140         /* E. Check state exit conditions. State can be terminated
2141          *    when high_seq is ACKed. */
2142         if (icsk->icsk_ca_state == TCP_CA_Open) {
2143                 BUG_TRAP(tp->retrans_out == 0);
2144                 tp->retrans_stamp = 0;
2145         } else if (!before(tp->snd_una, tp->high_seq)) {
2146                 switch (icsk->icsk_ca_state) {
2147                 case TCP_CA_Loss:
2148                         icsk->icsk_retransmits = 0;
2149                         if (tcp_try_undo_recovery(sk))
2150                                 return;
2151                         break;
2152
2153                 case TCP_CA_CWR:
2154                         /* CWR is to be held something *above* high_seq
2155                          * is ACKed for CWR bit to reach receiver. */
2156                         if (tp->snd_una != tp->high_seq) {
2157                                 tcp_complete_cwr(sk);
2158                                 tcp_set_ca_state(sk, TCP_CA_Open);
2159                         }
2160                         break;
2161
2162                 case TCP_CA_Disorder:
2163                         tcp_try_undo_dsack(sk);
2164                         if (!tp->undo_marker ||
2165                             /* For SACK case do not Open to allow to undo
2166                              * catching for all duplicate ACKs. */
2167                             IsReno(tp) || tp->snd_una != tp->high_seq) {
2168                                 tp->undo_marker = 0;
2169                                 tcp_set_ca_state(sk, TCP_CA_Open);
2170                         }
2171                         break;
2172
2173                 case TCP_CA_Recovery:
2174                         if (IsReno(tp))
2175                                 tcp_reset_reno_sack(tp);
2176                         if (tcp_try_undo_recovery(sk))
2177                                 return;
2178                         tcp_complete_cwr(sk);
2179                         break;
2180                 }
2181         }
2182
2183         /* F. Process state. */
2184         switch (icsk->icsk_ca_state) {
2185         case TCP_CA_Recovery:
2186                 if (prior_snd_una == tp->snd_una) {
2187                         if (IsReno(tp) && is_dupack)
2188                                 tcp_add_reno_sack(sk);
2189                 } else {
2190                         int acked = prior_packets - tp->packets_out;
2191                         if (IsReno(tp))
2192                                 tcp_remove_reno_sacks(sk, acked);
2193                         is_dupack = tcp_try_undo_partial(sk, acked);
2194                 }
2195                 break;
2196         case TCP_CA_Loss:
2197                 if (flag&FLAG_DATA_ACKED)
2198                         icsk->icsk_retransmits = 0;
2199                 if (!tcp_try_undo_loss(sk)) {
2200                         tcp_moderate_cwnd(tp);
2201                         tcp_xmit_retransmit_queue(sk);
2202                         return;
2203                 }
2204                 if (icsk->icsk_ca_state != TCP_CA_Open)
2205                         return;
2206                 /* Loss is undone; fall through to processing in Open state. */
2207         default:
2208                 if (IsReno(tp)) {
2209                         if (tp->snd_una != prior_snd_una)
2210                                 tcp_reset_reno_sack(tp);
2211                         if (is_dupack)
2212                                 tcp_add_reno_sack(sk);
2213                 }
2214
2215                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2216                         tcp_try_undo_dsack(sk);
2217
2218                 if (!tcp_time_to_recover(sk)) {
2219                         tcp_try_to_open(sk, flag);
2220                         return;
2221                 }
2222
2223                 /* MTU probe failure: don't reduce cwnd */
2224                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2225                     icsk->icsk_mtup.probe_size &&
2226                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2227                         tcp_mtup_probe_failed(sk);
2228                         /* Restores the reduction we did in tcp_mtup_probe() */
2229                         tp->snd_cwnd++;
2230                         tcp_simple_retransmit(sk);
2231                         return;
2232                 }
2233
2234                 /* Otherwise enter Recovery state */
2235
2236                 if (IsReno(tp))
2237                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2238                 else
2239                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2240
2241                 tp->high_seq = tp->snd_nxt;
2242                 tp->prior_ssthresh = 0;
2243                 tp->undo_marker = tp->snd_una;
2244                 tp->undo_retrans = tp->retrans_out;
2245
2246                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2247                         if (!(flag&FLAG_ECE))
2248                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2249                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2250                         TCP_ECN_queue_cwr(tp);
2251                 }
2252
2253                 tp->bytes_acked = 0;
2254                 tp->snd_cwnd_cnt = 0;
2255                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2256         }
2257
2258         if (is_dupack || tcp_head_timedout(sk))
2259                 tcp_update_scoreboard(sk);
2260         tcp_cwnd_down(sk);
2261         tcp_xmit_retransmit_queue(sk);
2262 }
2263
2264 /* Read draft-ietf-tcplw-high-performance before mucking
2265  * with this code. (Supersedes RFC1323)
2266  */
2267 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2268 {
2269         /* RTTM Rule: A TSecr value received in a segment is used to
2270          * update the averaged RTT measurement only if the segment
2271          * acknowledges some new data, i.e., only if it advances the
2272          * left edge of the send window.
2273          *
2274          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2275          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2276          *
2277          * Changed: reset backoff as soon as we see the first valid sample.
2278          * If we do not, we get strongly overestimated rto. With timestamps
2279          * samples are accepted even from very old segments: f.e., when rtt=1
2280          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2281          * answer arrives rto becomes 120 seconds! If at least one of segments
2282          * in window is lost... Voila.                          --ANK (010210)
2283          */
2284         struct tcp_sock *tp = tcp_sk(sk);
2285         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2286         tcp_rtt_estimator(sk, seq_rtt);
2287         tcp_set_rto(sk);
2288         inet_csk(sk)->icsk_backoff = 0;
2289         tcp_bound_rto(sk);
2290 }
2291
2292 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2293 {
2294         /* We don't have a timestamp. Can only use
2295          * packets that are not retransmitted to determine
2296          * rtt estimates. Also, we must not reset the
2297          * backoff for rto until we get a non-retransmitted
2298          * packet. This allows us to deal with a situation
2299          * where the network delay has increased suddenly.
2300          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2301          */
2302
2303         if (flag & FLAG_RETRANS_DATA_ACKED)
2304                 return;
2305
2306         tcp_rtt_estimator(sk, seq_rtt);
2307         tcp_set_rto(sk);
2308         inet_csk(sk)->icsk_backoff = 0;
2309         tcp_bound_rto(sk);
2310 }
2311
2312 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2313                                       const s32 seq_rtt)
2314 {
2315         const struct tcp_sock *tp = tcp_sk(sk);
2316         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2317         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2318                 tcp_ack_saw_tstamp(sk, flag);
2319         else if (seq_rtt >= 0)
2320                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2321 }
2322
2323 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2324                            u32 in_flight, int good)
2325 {
2326         const struct inet_connection_sock *icsk = inet_csk(sk);
2327         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2328         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2329 }
2330
2331 /* Restart timer after forward progress on connection.
2332  * RFC2988 recommends to restart timer to now+rto.
2333  */
2334
2335 static void tcp_ack_packets_out(struct sock *sk)
2336 {
2337         struct tcp_sock *tp = tcp_sk(sk);
2338
2339         if (!tp->packets_out) {
2340                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2341         } else {
2342                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2343         }
2344 }
2345
2346 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2347                          __u32 now, __s32 *seq_rtt)
2348 {
2349         struct tcp_sock *tp = tcp_sk(sk);
2350         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2351         __u32 seq = tp->snd_una;
2352         __u32 packets_acked;
2353         int acked = 0;
2354
2355         /* If we get here, the whole TSO packet has not been
2356          * acked.
2357          */
2358         BUG_ON(!after(scb->end_seq, seq));
2359
2360         packets_acked = tcp_skb_pcount(skb);
2361         if (tcp_trim_head(sk, skb, seq - scb->seq))
2362                 return 0;
2363         packets_acked -= tcp_skb_pcount(skb);
2364
2365         if (packets_acked) {
2366                 __u8 sacked = scb->sacked;
2367
2368                 acked |= FLAG_DATA_ACKED;
2369                 if (sacked) {
2370                         if (sacked & TCPCB_RETRANS) {
2371                                 if (sacked & TCPCB_SACKED_RETRANS)
2372                                         tp->retrans_out -= packets_acked;
2373                                 acked |= FLAG_RETRANS_DATA_ACKED;
2374                                 *seq_rtt = -1;
2375                         } else if (*seq_rtt < 0)
2376                                 *seq_rtt = now - scb->when;
2377                         if (sacked & TCPCB_SACKED_ACKED)
2378                                 tp->sacked_out -= packets_acked;
2379                         if (sacked & TCPCB_LOST)
2380                                 tp->lost_out -= packets_acked;
2381                         if (sacked & TCPCB_URG) {
2382                                 if (tp->urg_mode &&
2383                                     !before(seq, tp->snd_up))
2384                                         tp->urg_mode = 0;
2385                         }
2386                 } else if (*seq_rtt < 0)
2387                         *seq_rtt = now - scb->when;
2388
2389                 if (tp->fackets_out) {
2390                         __u32 dval = min(tp->fackets_out, packets_acked);
2391                         tp->fackets_out -= dval;
2392                 }
2393                 tp->packets_out -= packets_acked;
2394
2395                 BUG_ON(tcp_skb_pcount(skb) == 0);
2396                 BUG_ON(!before(scb->seq, scb->end_seq));
2397         }
2398
2399         return acked;
2400 }
2401
2402 /* Remove acknowledged frames from the retransmission queue. */
2403 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2404 {
2405         struct tcp_sock *tp = tcp_sk(sk);
2406         const struct inet_connection_sock *icsk = inet_csk(sk);
2407         struct sk_buff *skb;
2408         __u32 now = tcp_time_stamp;
2409         int acked = 0;
2410         __s32 seq_rtt = -1;
2411         u32 pkts_acked = 0;
2412         ktime_t last_ackt = ktime_set(0,0);
2413
2414         while ((skb = tcp_write_queue_head(sk)) &&
2415                skb != tcp_send_head(sk)) {
2416                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2417                 __u8 sacked = scb->sacked;
2418
2419                 /* If our packet is before the ack sequence we can
2420                  * discard it as it's confirmed to have arrived at
2421                  * the other end.
2422                  */
2423                 if (after(scb->end_seq, tp->snd_una)) {
2424                         if (tcp_skb_pcount(skb) > 1 &&
2425                             after(tp->snd_una, scb->seq))
2426                                 acked |= tcp_tso_acked(sk, skb,
2427                                                        now, &seq_rtt);
2428                         break;
2429                 }
2430
2431                 /* Initial outgoing SYN's get put onto the write_queue
2432                  * just like anything else we transmit.  It is not
2433                  * true data, and if we misinform our callers that
2434                  * this ACK acks real data, we will erroneously exit
2435                  * connection startup slow start one packet too
2436                  * quickly.  This is severely frowned upon behavior.
2437                  */
2438                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2439                         acked |= FLAG_DATA_ACKED;
2440                         ++pkts_acked;
2441                 } else {
2442                         acked |= FLAG_SYN_ACKED;
2443                         tp->retrans_stamp = 0;
2444                 }
2445
2446                 /* MTU probing checks */
2447                 if (icsk->icsk_mtup.probe_size) {
2448                         if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2449                                 tcp_mtup_probe_success(sk, skb);
2450                         }
2451                 }
2452
2453                 if (sacked) {
2454                         if (sacked & TCPCB_RETRANS) {
2455                                 if (sacked & TCPCB_SACKED_RETRANS)
2456                                         tp->retrans_out -= tcp_skb_pcount(skb);
2457                                 acked |= FLAG_RETRANS_DATA_ACKED;
2458                                 seq_rtt = -1;
2459                         } else if (seq_rtt < 0) {
2460                                 seq_rtt = now - scb->when;
2461                                 last_ackt = skb->tstamp;
2462                         }
2463                         if (sacked & TCPCB_SACKED_ACKED)
2464                                 tp->sacked_out -= tcp_skb_pcount(skb);
2465                         if (sacked & TCPCB_LOST)
2466                                 tp->lost_out -= tcp_skb_pcount(skb);
2467                         if (sacked & TCPCB_URG) {
2468                                 if (tp->urg_mode &&
2469                                     !before(scb->end_seq, tp->snd_up))
2470                                         tp->urg_mode = 0;
2471                         }
2472                 } else if (seq_rtt < 0) {
2473                         seq_rtt = now - scb->when;
2474                         last_ackt = skb->tstamp;
2475                 }
2476                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2477                 tcp_packets_out_dec(tp, skb);
2478                 tcp_unlink_write_queue(skb, sk);
2479                 sk_stream_free_skb(sk, skb);
2480                 clear_all_retrans_hints(tp);
2481         }
2482
2483         if (acked&FLAG_ACKED) {
2484                 const struct tcp_congestion_ops *ca_ops
2485                         = inet_csk(sk)->icsk_ca_ops;
2486
2487                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2488                 tcp_ack_packets_out(sk);
2489
2490                 if (ca_ops->pkts_acked)
2491                         ca_ops->pkts_acked(sk, pkts_acked, last_ackt);
2492         }
2493
2494 #if FASTRETRANS_DEBUG > 0
2495         BUG_TRAP((int)tp->sacked_out >= 0);
2496         BUG_TRAP((int)tp->lost_out >= 0);
2497         BUG_TRAP((int)tp->retrans_out >= 0);
2498         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2499                 const struct inet_connection_sock *icsk = inet_csk(sk);
2500                 if (tp->lost_out) {
2501                         printk(KERN_DEBUG "Leak l=%u %d\n",
2502                                tp->lost_out, icsk->icsk_ca_state);
2503                         tp->lost_out = 0;
2504                 }
2505                 if (tp->sacked_out) {
2506                         printk(KERN_DEBUG "Leak s=%u %d\n",
2507                                tp->sacked_out, icsk->icsk_ca_state);
2508                         tp->sacked_out = 0;
2509                 }
2510                 if (tp->retrans_out) {
2511                         printk(KERN_DEBUG "Leak r=%u %d\n",
2512                                tp->retrans_out, icsk->icsk_ca_state);
2513                         tp->retrans_out = 0;
2514                 }
2515         }
2516 #endif
2517         *seq_rtt_p = seq_rtt;
2518         return acked;
2519 }
2520
2521 static void tcp_ack_probe(struct sock *sk)
2522 {
2523         const struct tcp_sock *tp = tcp_sk(sk);
2524         struct inet_connection_sock *icsk = inet_csk(sk);
2525
2526         /* Was it a usable window open? */
2527
2528         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2529                    tp->snd_una + tp->snd_wnd)) {
2530                 icsk->icsk_backoff = 0;
2531                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2532                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2533                  * This function is not for random using!
2534                  */
2535         } else {
2536                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2537                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2538                                           TCP_RTO_MAX);
2539         }
2540 }
2541
2542 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2543 {
2544         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2545                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2546 }
2547
2548 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2549 {
2550         const struct tcp_sock *tp = tcp_sk(sk);
2551         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2552                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2553 }
2554
2555 /* Check that window update is acceptable.
2556  * The function assumes that snd_una<=ack<=snd_next.
2557  */
2558 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2559                                         const u32 ack_seq, const u32 nwin)
2560 {
2561         return (after(ack, tp->snd_una) ||
2562                 after(ack_seq, tp->snd_wl1) ||
2563                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2564 }
2565
2566 /* Update our send window.
2567  *
2568  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2569  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2570  */
2571 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2572                                  u32 ack_seq)
2573 {
2574         struct tcp_sock *tp = tcp_sk(sk);
2575         int flag = 0;
2576         u32 nwin = ntohs(tcp_hdr(skb)->window);
2577
2578         if (likely(!tcp_hdr(skb)->syn))
2579                 nwin <<= tp->rx_opt.snd_wscale;
2580
2581         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2582                 flag |= FLAG_WIN_UPDATE;
2583                 tcp_update_wl(tp, ack, ack_seq);
2584
2585                 if (tp->snd_wnd != nwin) {
2586                         tp->snd_wnd = nwin;
2587
2588                         /* Note, it is the only place, where
2589                          * fast path is recovered for sending TCP.
2590                          */
2591                         tp->pred_flags = 0;
2592                         tcp_fast_path_check(sk);
2593
2594                         if (nwin > tp->max_window) {
2595                                 tp->max_window = nwin;
2596                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2597                         }
2598                 }
2599         }
2600
2601         tp->snd_una = ack;
2602
2603         return flag;
2604 }
2605
2606 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2607  * continue in congestion avoidance.
2608  */
2609 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2610 {
2611         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2612         tp->snd_cwnd_cnt = 0;
2613         TCP_ECN_queue_cwr(tp);
2614         tcp_moderate_cwnd(tp);
2615 }
2616
2617 /* A conservative spurious RTO response algorithm: reduce cwnd using
2618  * rate halving and continue in congestion avoidance.
2619  */
2620 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2621 {
2622         tcp_enter_cwr(sk, 0);
2623 }
2624
2625 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2626 {
2627         if (flag&FLAG_ECE)
2628                 tcp_ratehalving_spur_to_response(sk);
2629         else
2630                 tcp_undo_cwr(sk, 1);
2631 }
2632
2633 /* F-RTO spurious RTO detection algorithm (RFC4138)
2634  *
2635  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2636  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2637  * window (but not to or beyond highest sequence sent before RTO):
2638  *   On First ACK,  send two new segments out.
2639  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2640  *                  algorithm is not part of the F-RTO detection algorithm
2641  *                  given in RFC4138 but can be selected separately).
2642  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2643  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2644  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2645  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2646  *
2647  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2648  * original window even after we transmit two new data segments.
2649  *
2650  * SACK version:
2651  *   on first step, wait until first cumulative ACK arrives, then move to
2652  *   the second step. In second step, the next ACK decides.
2653  *
2654  * F-RTO is implemented (mainly) in four functions:
2655  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2656  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2657  *     called when tcp_use_frto() showed green light
2658  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2659  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2660  *     to prove that the RTO is indeed spurious. It transfers the control
2661  *     from F-RTO to the conventional RTO recovery
2662  */
2663 static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
2664 {
2665         struct tcp_sock *tp = tcp_sk(sk);
2666
2667         tcp_sync_left_out(tp);
2668
2669         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2670         if (flag&FLAG_DATA_ACKED)
2671                 inet_csk(sk)->icsk_retransmits = 0;
2672
2673         if (!before(tp->snd_una, tp->frto_highmark)) {
2674                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2675                 return 1;
2676         }
2677
2678         if (!IsSackFrto() || IsReno(tp)) {
2679                 /* RFC4138 shortcoming in step 2; should also have case c):
2680                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2681                  * data, winupdate
2682                  */
2683                 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
2684                     !(flag&FLAG_FORWARD_PROGRESS))
2685                         return 1;
2686
2687                 if (!(flag&FLAG_DATA_ACKED)) {
2688                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2689                                             flag);
2690                         return 1;
2691                 }
2692         } else {
2693                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2694                         /* Prevent sending of new data. */
2695                         tp->snd_cwnd = min(tp->snd_cwnd,
2696                                            tcp_packets_in_flight(tp));
2697                         return 1;
2698                 }
2699
2700                 if ((tp->frto_counter >= 2) &&
2701                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2702                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2703                         /* RFC4138 shortcoming (see comment above) */
2704                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2705                                 return 1;
2706
2707                         tcp_enter_frto_loss(sk, 3, flag);
2708                         return 1;
2709                 }
2710         }
2711
2712         if (tp->frto_counter == 1) {
2713                 /* Sending of the next skb must be allowed or no FRTO */
2714                 if (!tcp_send_head(sk) ||
2715                     after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2716                                      tp->snd_una + tp->snd_wnd)) {
2717                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2718                                             flag);
2719                         return 1;
2720                 }
2721
2722                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2723                 tp->frto_counter = 2;
2724                 return 1;
2725         } else {
2726                 switch (sysctl_tcp_frto_response) {
2727                 case 2:
2728                         tcp_undo_spur_to_response(sk, flag);
2729                         break;
2730                 case 1:
2731                         tcp_conservative_spur_to_response(tp);
2732                         break;
2733                 default:
2734                         tcp_ratehalving_spur_to_response(sk);
2735                         break;
2736                 }
2737                 tp->frto_counter = 0;
2738         }
2739         return 0;
2740 }
2741
2742 /* This routine deals with incoming acks, but not outgoing ones. */
2743 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2744 {
2745         struct inet_connection_sock *icsk = inet_csk(sk);
2746         struct tcp_sock *tp = tcp_sk(sk);
2747         u32 prior_snd_una = tp->snd_una;
2748         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2749         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2750         u32 prior_in_flight;
2751         s32 seq_rtt;
2752         int prior_packets;
2753         int frto_cwnd = 0;
2754
2755         /* If the ack is newer than sent or older than previous acks
2756          * then we can probably ignore it.
2757          */
2758         if (after(ack, tp->snd_nxt))
2759                 goto uninteresting_ack;
2760
2761         if (before(ack, prior_snd_una))
2762                 goto old_ack;
2763
2764         if (sysctl_tcp_abc) {
2765                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2766                         tp->bytes_acked += ack - prior_snd_una;
2767                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2768                         /* we assume just one segment left network */
2769                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2770         }
2771
2772         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2773                 /* Window is constant, pure forward advance.
2774                  * No more checks are required.
2775                  * Note, we use the fact that SND.UNA>=SND.WL2.
2776                  */
2777                 tcp_update_wl(tp, ack, ack_seq);
2778                 tp->snd_una = ack;
2779                 flag |= FLAG_WIN_UPDATE;
2780
2781                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2782
2783                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2784         } else {
2785                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2786                         flag |= FLAG_DATA;
2787                 else
2788                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2789
2790                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2791
2792                 if (TCP_SKB_CB(skb)->sacked)
2793                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2794
2795                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2796                         flag |= FLAG_ECE;
2797
2798                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2799         }
2800
2801         /* We passed data and got it acked, remove any soft error
2802          * log. Something worked...
2803          */
2804         sk->sk_err_soft = 0;
2805         tp->rcv_tstamp = tcp_time_stamp;
2806         prior_packets = tp->packets_out;
2807         if (!prior_packets)
2808                 goto no_queue;
2809
2810         prior_in_flight = tcp_packets_in_flight(tp);
2811
2812         /* See if we can take anything off of the retransmit queue. */
2813         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2814
2815         if (tp->frto_counter)
2816                 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
2817
2818         if (tcp_ack_is_dubious(sk, flag)) {
2819                 /* Advance CWND, if state allows this. */
2820                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2821                     tcp_may_raise_cwnd(sk, flag))
2822                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2823                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2824         } else {
2825                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2826                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2827         }
2828
2829         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2830                 dst_confirm(sk->sk_dst_cache);
2831
2832         return 1;
2833
2834 no_queue:
2835         icsk->icsk_probes_out = 0;
2836
2837         /* If this ack opens up a zero window, clear backoff.  It was
2838          * being used to time the probes, and is probably far higher than
2839          * it needs to be for normal retransmission.
2840          */
2841         if (tcp_send_head(sk))
2842                 tcp_ack_probe(sk);
2843         return 1;
2844
2845 old_ack:
2846         if (TCP_SKB_CB(skb)->sacked)
2847                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2848
2849 uninteresting_ack:
2850         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2851         return 0;
2852 }
2853
2854
2855 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2856  * But, this can also be called on packets in the established flow when
2857  * the fast version below fails.
2858  */
2859 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2860 {
2861         unsigned char *ptr;
2862         struct tcphdr *th = tcp_hdr(skb);
2863         int length=(th->doff*4)-sizeof(struct tcphdr);
2864
2865         ptr = (unsigned char *)(th + 1);
2866         opt_rx->saw_tstamp = 0;
2867
2868         while (length > 0) {
2869                 int opcode=*ptr++;
2870                 int opsize;
2871
2872                 switch (opcode) {
2873                         case TCPOPT_EOL:
2874                                 return;
2875                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2876                                 length--;
2877                                 continue;
2878                         default:
2879                                 opsize=*ptr++;
2880                                 if (opsize < 2) /* "silly options" */
2881                                         return;
2882                                 if (opsize > length)
2883                                         return; /* don't parse partial options */
2884                                 switch (opcode) {
2885                                 case TCPOPT_MSS:
2886                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2887                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2888                                                 if (in_mss) {
2889                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2890                                                                 in_mss = opt_rx->user_mss;
2891                                                         opt_rx->mss_clamp = in_mss;
2892                                                 }
2893                                         }
2894                                         break;
2895                                 case TCPOPT_WINDOW:
2896                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2897                                                 if (sysctl_tcp_window_scaling) {
2898                                                         __u8 snd_wscale = *(__u8 *) ptr;
2899                                                         opt_rx->wscale_ok = 1;
2900                                                         if (snd_wscale > 14) {
2901                                                                 if (net_ratelimit())
2902                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2903                                                                                "scaling value %d >14 received.\n",
2904                                                                                snd_wscale);
2905                                                                 snd_wscale = 14;
2906                                                         }
2907                                                         opt_rx->snd_wscale = snd_wscale;
2908                                                 }
2909                                         break;
2910                                 case TCPOPT_TIMESTAMP:
2911                                         if (opsize==TCPOLEN_TIMESTAMP) {
2912                                                 if ((estab && opt_rx->tstamp_ok) ||
2913                                                     (!estab && sysctl_tcp_timestamps)) {
2914                                                         opt_rx->saw_tstamp = 1;
2915                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2916                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2917                                                 }
2918                                         }
2919                                         break;
2920                                 case TCPOPT_SACK_PERM:
2921                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2922                                                 if (sysctl_tcp_sack) {
2923                                                         opt_rx->sack_ok = 1;
2924                                                         tcp_sack_reset(opt_rx);
2925                                                 }
2926                                         }
2927                                         break;
2928
2929                                 case TCPOPT_SACK:
2930                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2931                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2932                                            opt_rx->sack_ok) {
2933                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2934                                         }
2935 #ifdef CONFIG_TCP_MD5SIG
2936                                 case TCPOPT_MD5SIG:
2937                                         /*
2938                                          * The MD5 Hash has already been
2939                                          * checked (see tcp_v{4,6}_do_rcv()).
2940                                          */
2941                                         break;
2942 #endif
2943                                 }
2944
2945                                 ptr+=opsize-2;
2946                                 length-=opsize;
2947                 }
2948         }
2949 }
2950
2951 /* Fast parse options. This hopes to only see timestamps.
2952  * If it is wrong it falls back on tcp_parse_options().
2953  */
2954 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2955                                   struct tcp_sock *tp)
2956 {
2957         if (th->doff == sizeof(struct tcphdr)>>2) {
2958                 tp->rx_opt.saw_tstamp = 0;
2959                 return 0;
2960         } else if (tp->rx_opt.tstamp_ok &&
2961                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2962                 __be32 *ptr = (__be32 *)(th + 1);
2963                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2964                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2965                         tp->rx_opt.saw_tstamp = 1;
2966                         ++ptr;
2967                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2968                         ++ptr;
2969                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2970                         return 1;
2971                 }
2972         }
2973         tcp_parse_options(skb, &tp->rx_opt, 1);
2974         return 1;
2975 }
2976
2977 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2978 {
2979         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2980         tp->rx_opt.ts_recent_stamp = get_seconds();
2981 }
2982
2983 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2984 {
2985         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2986                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2987                  * extra check below makes sure this can only happen
2988                  * for pure ACK frames.  -DaveM
2989                  *
2990                  * Not only, also it occurs for expired timestamps.
2991                  */
2992
2993                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2994                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2995                         tcp_store_ts_recent(tp);
2996         }
2997 }
2998
2999 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3000  *
3001  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3002  * it can pass through stack. So, the following predicate verifies that
3003  * this segment is not used for anything but congestion avoidance or
3004  * fast retransmit. Moreover, we even are able to eliminate most of such
3005  * second order effects, if we apply some small "replay" window (~RTO)
3006  * to timestamp space.
3007  *
3008  * All these measures still do not guarantee that we reject wrapped ACKs
3009  * on networks with high bandwidth, when sequence space is recycled fastly,
3010  * but it guarantees that such events will be very rare and do not affect
3011  * connection seriously. This doesn't look nice, but alas, PAWS is really
3012  * buggy extension.
3013  *
3014  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3015  * states that events when retransmit arrives after original data are rare.
3016  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3017  * the biggest problem on large power networks even with minor reordering.
3018  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3019  * up to bandwidth of 18Gigabit/sec. 8) ]
3020  */
3021
3022 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3023 {
3024         struct tcp_sock *tp = tcp_sk(sk);
3025         struct tcphdr *th = tcp_hdr(skb);
3026         u32 seq = TCP_SKB_CB(skb)->seq;
3027         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3028
3029         return (/* 1. Pure ACK with correct sequence number. */
3030                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3031
3032                 /* 2. ... and duplicate ACK. */
3033                 ack == tp->snd_una &&
3034
3035                 /* 3. ... and does not update window. */
3036                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3037
3038                 /* 4. ... and sits in replay window. */
3039                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3040 }
3041
3042 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3043 {
3044         const struct tcp_sock *tp = tcp_sk(sk);
3045         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3046                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3047                 !tcp_disordered_ack(sk, skb));
3048 }
3049
3050 /* Check segment sequence number for validity.
3051  *
3052  * Segment controls are considered valid, if the segment
3053  * fits to the window after truncation to the window. Acceptability
3054  * of data (and SYN, FIN, of course) is checked separately.
3055  * See tcp_data_queue(), for example.
3056  *
3057  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3058  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3059  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3060  * (borrowed from freebsd)
3061  */
3062
3063 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3064 {
3065         return  !before(end_seq, tp->rcv_wup) &&
3066                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3067 }
3068
3069 /* When we get a reset we do this. */
3070 static void tcp_reset(struct sock *sk)
3071 {
3072         /* We want the right error as BSD sees it (and indeed as we do). */
3073         switch (sk->sk_state) {
3074                 case TCP_SYN_SENT:
3075                         sk->sk_err = ECONNREFUSED;
3076                         break;
3077                 case TCP_CLOSE_WAIT:
3078                         sk->sk_err = EPIPE;
3079                         break;
3080                 case TCP_CLOSE:
3081                         return;
3082                 default:
3083                         sk->sk_err = ECONNRESET;
3084         }
3085
3086         if (!sock_flag(sk, SOCK_DEAD))
3087                 sk->sk_error_report(sk);
3088
3089         tcp_done(sk);
3090 }
3091
3092 /*
3093  *      Process the FIN bit. This now behaves as it is supposed to work
3094  *      and the FIN takes effect when it is validly part of sequence
3095  *      space. Not before when we get holes.
3096  *
3097  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3098  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3099  *      TIME-WAIT)
3100  *
3101  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3102  *      close and we go into CLOSING (and later onto TIME-WAIT)
3103  *
3104  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3105  */
3106 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3107 {
3108         struct tcp_sock *tp = tcp_sk(sk);
3109
3110         inet_csk_schedule_ack(sk);
3111
3112         sk->sk_shutdown |= RCV_SHUTDOWN;
3113         sock_set_flag(sk, SOCK_DONE);
3114
3115         switch (sk->sk_state) {
3116                 case TCP_SYN_RECV:
3117                 case TCP_ESTABLISHED:
3118                         /* Move to CLOSE_WAIT */
3119                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3120                         inet_csk(sk)->icsk_ack.pingpong = 1;
3121                         break;
3122
3123                 case TCP_CLOSE_WAIT:
3124                 case TCP_CLOSING:
3125                         /* Received a retransmission of the FIN, do
3126                          * nothing.
3127                          */
3128                         break;
3129                 case TCP_LAST_ACK:
3130                         /* RFC793: Remain in the LAST-ACK state. */
3131                         break;
3132
3133                 case TCP_FIN_WAIT1:
3134                         /* This case occurs when a simultaneous close
3135                          * happens, we must ack the received FIN and
3136                          * enter the CLOSING state.
3137                          */
3138                         tcp_send_ack(sk);
3139                         tcp_set_state(sk, TCP_CLOSING);
3140                         break;
3141                 case TCP_FIN_WAIT2:
3142                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3143                         tcp_send_ack(sk);
3144                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3145                         break;
3146                 default:
3147                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3148                          * cases we should never reach this piece of code.
3149                          */
3150                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3151                                __FUNCTION__, sk->sk_state);
3152                         break;
3153         }
3154
3155         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3156          * Probably, we should reset in this case. For now drop them.
3157          */
3158         __skb_queue_purge(&tp->out_of_order_queue);
3159         if (tp->rx_opt.sack_ok)
3160                 tcp_sack_reset(&tp->rx_opt);
3161         sk_stream_mem_reclaim(sk);
3162
3163         if (!sock_flag(sk, SOCK_DEAD)) {
3164                 sk->sk_state_change(sk);
3165
3166                 /* Do not send POLL_HUP for half duplex close. */
3167                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3168                     sk->sk_state == TCP_CLOSE)
3169                         sk_wake_async(sk, 1, POLL_HUP);
3170                 else
3171                         sk_wake_async(sk, 1, POLL_IN);
3172         }
3173 }
3174
3175 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3176 {
3177         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3178                 if (before(seq, sp->start_seq))
3179                         sp->start_seq = seq;
3180                 if (after(end_seq, sp->end_seq))
3181                         sp->end_seq = end_seq;
3182                 return 1;
3183         }
3184         return 0;
3185 }
3186
3187 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3188 {
3189         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3190                 if (before(seq, tp->rcv_nxt))
3191                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3192                 else
3193                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3194
3195                 tp->rx_opt.dsack = 1;
3196                 tp->duplicate_sack[0].start_seq = seq;
3197                 tp->duplicate_sack[0].end_seq = end_seq;
3198                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3199         }
3200 }
3201
3202 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3203 {
3204         if (!tp->rx_opt.dsack)
3205                 tcp_dsack_set(tp, seq, end_seq);
3206         else
3207                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3208 }
3209
3210 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3211 {
3212         struct tcp_sock *tp = tcp_sk(sk);
3213
3214         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3215             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3216                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3217                 tcp_enter_quickack_mode(sk);
3218
3219                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3220                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3221
3222                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3223                                 end_seq = tp->rcv_nxt;
3224                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3225                 }
3226         }
3227
3228         tcp_send_ack(sk);
3229 }
3230
3231 /* These routines update the SACK block as out-of-order packets arrive or
3232  * in-order packets close up the sequence space.
3233  */
3234 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3235 {
3236         int this_sack;
3237         struct tcp_sack_block *sp = &tp->selective_acks[0];
3238         struct tcp_sack_block *swalk = sp+1;
3239
3240         /* See if the recent change to the first SACK eats into
3241          * or hits the sequence space of other SACK blocks, if so coalesce.
3242          */
3243         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3244                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3245                         int i;
3246
3247                         /* Zap SWALK, by moving every further SACK up by one slot.
3248                          * Decrease num_sacks.
3249                          */
3250                         tp->rx_opt.num_sacks--;
3251                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3252                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3253                                 sp[i] = sp[i+1];
3254                         continue;
3255                 }
3256                 this_sack++, swalk++;
3257         }
3258 }
3259
3260 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3261 {
3262         __u32 tmp;
3263
3264         tmp = sack1->start_seq;
3265         sack1->start_seq = sack2->start_seq;
3266         sack2->start_seq = tmp;
3267
3268         tmp = sack1->end_seq;
3269         sack1->end_seq = sack2->end_seq;
3270         sack2->end_seq = tmp;
3271 }
3272
3273 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3274 {
3275         struct tcp_sock *tp = tcp_sk(sk);
3276         struct tcp_sack_block *sp = &tp->selective_acks[0];
3277         int cur_sacks = tp->rx_opt.num_sacks;
3278         int this_sack;
3279
3280         if (!cur_sacks)
3281                 goto new_sack;
3282
3283         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3284                 if (tcp_sack_extend(sp, seq, end_seq)) {
3285                         /* Rotate this_sack to the first one. */
3286                         for (; this_sack>0; this_sack--, sp--)
3287                                 tcp_sack_swap(sp, sp-1);
3288                         if (cur_sacks > 1)
3289                                 tcp_sack_maybe_coalesce(tp);
3290                         return;
3291                 }
3292         }
3293
3294         /* Could not find an adjacent existing SACK, build a new one,
3295          * put it at the front, and shift everyone else down.  We
3296          * always know there is at least one SACK present already here.
3297          *
3298          * If the sack array is full, forget about the last one.
3299          */
3300         if (this_sack >= 4) {
3301                 this_sack--;
3302                 tp->rx_opt.num_sacks--;
3303                 sp--;
3304         }
3305         for (; this_sack > 0; this_sack--, sp--)
3306                 *sp = *(sp-1);
3307
3308 new_sack:
3309         /* Build the new head SACK, and we're done. */
3310         sp->start_seq = seq;
3311         sp->end_seq = end_seq;
3312         tp->rx_opt.num_sacks++;
3313         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3314 }
3315
3316 /* RCV.NXT advances, some SACKs should be eaten. */
3317
3318 static void tcp_sack_remove(struct tcp_sock *tp)
3319 {
3320         struct tcp_sack_block *sp = &tp->selective_acks[0];
3321         int num_sacks = tp->rx_opt.num_sacks;
3322         int this_sack;
3323
3324         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3325         if (skb_queue_empty(&tp->out_of_order_queue)) {
3326                 tp->rx_opt.num_sacks = 0;
3327                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3328                 return;
3329         }
3330
3331         for (this_sack = 0; this_sack < num_sacks; ) {
3332                 /* Check if the start of the sack is covered by RCV.NXT. */
3333                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3334                         int i;
3335
3336                         /* RCV.NXT must cover all the block! */
3337                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3338
3339                         /* Zap this SACK, by moving forward any other SACKS. */
3340                         for (i=this_sack+1; i < num_sacks; i++)
3341                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3342                         num_sacks--;
3343                         continue;
3344                 }
3345                 this_sack++;
3346                 sp++;
3347         }
3348         if (num_sacks != tp->rx_opt.num_sacks) {
3349                 tp->rx_opt.num_sacks = num_sacks;
3350                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3351         }
3352 }
3353
3354 /* This one checks to see if we can put data from the
3355  * out_of_order queue into the receive_queue.
3356  */
3357 static void tcp_ofo_queue(struct sock *sk)
3358 {
3359         struct tcp_sock *tp = tcp_sk(sk);
3360         __u32 dsack_high = tp->rcv_nxt;
3361         struct sk_buff *skb;
3362
3363         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3364                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3365                         break;
3366
3367                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3368                         __u32 dsack = dsack_high;
3369                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3370                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3371                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3372                 }
3373
3374                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3375                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3376                         __skb_unlink(skb, &tp->out_of_order_queue);
3377                         __kfree_skb(skb);
3378                         continue;
3379                 }
3380                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3381                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3382                            TCP_SKB_CB(skb)->end_seq);
3383
3384                 __skb_unlink(skb, &tp->out_of_order_queue);
3385                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3386                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3387                 if (tcp_hdr(skb)->fin)
3388                         tcp_fin(skb, sk, tcp_hdr(skb));
3389         }
3390 }
3391
3392 static int tcp_prune_queue(struct sock *sk);
3393
3394 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3395 {
3396         struct tcphdr *th = tcp_hdr(skb);
3397         struct tcp_sock *tp = tcp_sk(sk);
3398         int eaten = -1;
3399
3400         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3401                 goto drop;
3402
3403         __skb_pull(skb, th->doff*4);
3404
3405         TCP_ECN_accept_cwr(tp, skb);
3406
3407         if (tp->rx_opt.dsack) {
3408                 tp->rx_opt.dsack = 0;
3409                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3410                                                     4 - tp->rx_opt.tstamp_ok);
3411         }
3412
3413         /*  Queue data for delivery to the user.
3414          *  Packets in sequence go to the receive queue.
3415          *  Out of sequence packets to the out_of_order_queue.
3416          */
3417         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3418                 if (tcp_receive_window(tp) == 0)
3419                         goto out_of_window;
3420
3421                 /* Ok. In sequence. In window. */
3422                 if (tp->ucopy.task == current &&
3423                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3424                     sock_owned_by_user(sk) && !tp->urg_data) {
3425                         int chunk = min_t(unsigned int, skb->len,
3426                                                         tp->ucopy.len);
3427
3428                         __set_current_state(TASK_RUNNING);
3429
3430                         local_bh_enable();
3431                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3432                                 tp->ucopy.len -= chunk;
3433                                 tp->copied_seq += chunk;
3434                                 eaten = (chunk == skb->len && !th->fin);
3435                                 tcp_rcv_space_adjust(sk);
3436                         }
3437                         local_bh_disable();
3438                 }
3439
3440                 if (eaten <= 0) {
3441 queue_and_out:
3442                         if (eaten < 0 &&
3443                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3444                              !sk_stream_rmem_schedule(sk, skb))) {
3445                                 if (tcp_prune_queue(sk) < 0 ||
3446                                     !sk_stream_rmem_schedule(sk, skb))
3447                                         goto drop;
3448                         }
3449                         sk_stream_set_owner_r(skb, sk);
3450                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3451                 }
3452                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3453                 if (skb->len)
3454                         tcp_event_data_recv(sk, skb);
3455                 if (th->fin)
3456                         tcp_fin(skb, sk, th);
3457
3458                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3459                         tcp_ofo_queue(sk);
3460
3461                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3462                          * gap in queue is filled.
3463                          */
3464                         if (skb_queue_empty(&tp->out_of_order_queue))
3465                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3466                 }
3467
3468                 if (tp->rx_opt.num_sacks)
3469                         tcp_sack_remove(tp);
3470
3471                 tcp_fast_path_check(sk);
3472
3473                 if (eaten > 0)
3474                         __kfree_skb(skb);
3475                 else if (!sock_flag(sk, SOCK_DEAD))
3476                         sk->sk_data_ready(sk, 0);
3477                 return;
3478         }
3479
3480         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3481                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3482                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3483                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3484
3485 out_of_window:
3486                 tcp_enter_quickack_mode(sk);
3487                 inet_csk_schedule_ack(sk);
3488 drop:
3489                 __kfree_skb(skb);
3490                 return;
3491         }
3492
3493         /* Out of window. F.e. zero window probe. */
3494         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3495                 goto out_of_window;
3496
3497         tcp_enter_quickack_mode(sk);
3498
3499         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3500                 /* Partial packet, seq < rcv_next < end_seq */
3501                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3502                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3503                            TCP_SKB_CB(skb)->end_seq);
3504
3505                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3506
3507                 /* If window is closed, drop tail of packet. But after
3508                  * remembering D-SACK for its head made in previous line.
3509                  */
3510                 if (!tcp_receive_window(tp))
3511                         goto out_of_window;
3512                 goto queue_and_out;
3513         }
3514
3515         TCP_ECN_check_ce(tp, skb);
3516
3517         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3518             !sk_stream_rmem_schedule(sk, skb)) {
3519                 if (tcp_prune_queue(sk) < 0 ||
3520                     !sk_stream_rmem_schedule(sk, skb))
3521                         goto drop;
3522         }
3523
3524         /* Disable header prediction. */
3525         tp->pred_flags = 0;
3526         inet_csk_schedule_ack(sk);
3527
3528         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3529                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3530
3531         sk_stream_set_owner_r(skb, sk);
3532
3533         if (!skb_peek(&tp->out_of_order_queue)) {
3534                 /* Initial out of order segment, build 1 SACK. */
3535                 if (tp->rx_opt.sack_ok) {
3536                         tp->rx_opt.num_sacks = 1;
3537                         tp->rx_opt.dsack     = 0;
3538                         tp->rx_opt.eff_sacks = 1;
3539                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3540                         tp->selective_acks[0].end_seq =
3541                                                 TCP_SKB_CB(skb)->end_seq;
3542                 }
3543                 __skb_queue_head(&tp->out_of_order_queue,skb);
3544         } else {
3545                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3546                 u32 seq = TCP_SKB_CB(skb)->seq;
3547                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3548
3549                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3550                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3551
3552                         if (!tp->rx_opt.num_sacks ||
3553                             tp->selective_acks[0].end_seq != seq)
3554                                 goto add_sack;
3555
3556                         /* Common case: data arrive in order after hole. */
3557                         tp->selective_acks[0].end_seq = end_seq;
3558                         return;
3559                 }
3560
3561                 /* Find place to insert this segment. */
3562                 do {
3563                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3564                                 break;
3565                 } while ((skb1 = skb1->prev) !=
3566                          (struct sk_buff*)&tp->out_of_order_queue);
3567
3568                 /* Do skb overlap to previous one? */
3569                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3570                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3571                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3572                                 /* All the bits are present. Drop. */
3573                                 __kfree_skb(skb);
3574                                 tcp_dsack_set(tp, seq, end_seq);
3575                                 goto add_sack;
3576                         }
3577                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3578                                 /* Partial overlap. */
3579                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3580                         } else {
3581                                 skb1 = skb1->prev;
3582                         }
3583                 }
3584                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3585
3586                 /* And clean segments covered by new one as whole. */
3587                 while ((skb1 = skb->next) !=
3588                        (struct sk_buff*)&tp->out_of_order_queue &&
3589                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3590                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3591                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3592                                break;
3593                        }
3594                        __skb_unlink(skb1, &tp->out_of_order_queue);
3595                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3596                        __kfree_skb(skb1);
3597                 }
3598
3599 add_sack:
3600                 if (tp->rx_opt.sack_ok)
3601                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3602         }
3603 }
3604
3605 /* Collapse contiguous sequence of skbs head..tail with
3606  * sequence numbers start..end.
3607  * Segments with FIN/SYN are not collapsed (only because this
3608  * simplifies code)
3609  */
3610 static void
3611 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3612              struct sk_buff *head, struct sk_buff *tail,
3613              u32 start, u32 end)
3614 {
3615         struct sk_buff *skb;
3616
3617         /* First, check that queue is collapsible and find
3618          * the point where collapsing can be useful. */
3619         for (skb = head; skb != tail; ) {
3620                 /* No new bits? It is possible on ofo queue. */
3621                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3622                         struct sk_buff *next = skb->next;
3623                         __skb_unlink(skb, list);
3624                         __kfree_skb(skb);
3625                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3626                         skb = next;
3627                         continue;
3628                 }
3629
3630                 /* The first skb to collapse is:
3631                  * - not SYN/FIN and
3632                  * - bloated or contains data before "start" or
3633                  *   overlaps to the next one.
3634                  */
3635                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3636                     (tcp_win_from_space(skb->truesize) > skb->len ||
3637                      before(TCP_SKB_CB(skb)->seq, start) ||
3638                      (skb->next != tail &&
3639                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3640                         break;
3641
3642                 /* Decided to skip this, advance start seq. */
3643                 start = TCP_SKB_CB(skb)->end_seq;
3644                 skb = skb->next;
3645         }
3646         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3647                 return;
3648
3649         while (before(start, end)) {
3650                 struct sk_buff *nskb;
3651                 int header = skb_headroom(skb);
3652                 int copy = SKB_MAX_ORDER(header, 0);
3653
3654                 /* Too big header? This can happen with IPv6. */
3655                 if (copy < 0)
3656                         return;
3657                 if (end-start < copy)
3658                         copy = end-start;
3659                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3660                 if (!nskb)
3661                         return;
3662
3663                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3664                 skb_set_network_header(nskb, (skb_network_header(skb) -
3665                                               skb->head));
3666                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3667                                                 skb->head));
3668                 skb_reserve(nskb, header);
3669                 memcpy(nskb->head, skb->head, header);
3670                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3671                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3672                 __skb_insert(nskb, skb->prev, skb, list);
3673                 sk_stream_set_owner_r(nskb, sk);
3674
3675                 /* Copy data, releasing collapsed skbs. */
3676                 while (copy > 0) {
3677                         int offset = start - TCP_SKB_CB(skb)->seq;
3678                         int size = TCP_SKB_CB(skb)->end_seq - start;
3679
3680                         BUG_ON(offset < 0);
3681                         if (size > 0) {
3682                                 size = min(copy, size);
3683                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3684                                         BUG();
3685                                 TCP_SKB_CB(nskb)->end_seq += size;
3686                                 copy -= size;
3687                                 start += size;
3688                         }
3689                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3690                                 struct sk_buff *next = skb->next;
3691                                 __skb_unlink(skb, list);
3692                                 __kfree_skb(skb);
3693                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3694                                 skb = next;
3695                                 if (skb == tail ||
3696                                     tcp_hdr(skb)->syn ||
3697                                     tcp_hdr(skb)->fin)
3698                                         return;
3699                         }
3700                 }
3701         }
3702 }
3703
3704 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3705  * and tcp_collapse() them until all the queue is collapsed.
3706  */
3707 static void tcp_collapse_ofo_queue(struct sock *sk)
3708 {
3709         struct tcp_sock *tp = tcp_sk(sk);
3710         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3711         struct sk_buff *head;
3712         u32 start, end;
3713
3714         if (skb == NULL)
3715                 return;
3716
3717         start = TCP_SKB_CB(skb)->seq;
3718         end = TCP_SKB_CB(skb)->end_seq;
3719         head = skb;
3720
3721         for (;;) {
3722                 skb = skb->next;
3723
3724                 /* Segment is terminated when we see gap or when
3725                  * we are at the end of all the queue. */
3726                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3727                     after(TCP_SKB_CB(skb)->seq, end) ||
3728                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3729                         tcp_collapse(sk, &tp->out_of_order_queue,
3730                                      head, skb, start, end);
3731                         head = skb;
3732                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3733                                 break;
3734                         /* Start new segment */
3735                         start = TCP_SKB_CB(skb)->seq;
3736                         end = TCP_SKB_CB(skb)->end_seq;
3737                 } else {
3738                         if (before(TCP_SKB_CB(skb)->seq, start))
3739                                 start = TCP_SKB_CB(skb)->seq;
3740                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3741                                 end = TCP_SKB_CB(skb)->end_seq;
3742                 }
3743         }
3744 }
3745
3746 /* Reduce allocated memory if we can, trying to get
3747  * the socket within its memory limits again.
3748  *
3749  * Return less than zero if we should start dropping frames
3750  * until the socket owning process reads some of the data
3751  * to stabilize the situation.
3752  */
3753 static int tcp_prune_queue(struct sock *sk)
3754 {
3755         struct tcp_sock *tp = tcp_sk(sk);
3756
3757         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3758
3759         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3760
3761         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3762                 tcp_clamp_window(sk);
3763         else if (tcp_memory_pressure)
3764                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3765
3766         tcp_collapse_ofo_queue(sk);
3767         tcp_collapse(sk, &sk->sk_receive_queue,
3768                      sk->sk_receive_queue.next,
3769                      (struct sk_buff*)&sk->sk_receive_queue,
3770                      tp->copied_seq, tp->rcv_nxt);
3771         sk_stream_mem_reclaim(sk);
3772
3773         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3774                 return 0;
3775
3776         /* Collapsing did not help, destructive actions follow.
3777          * This must not ever occur. */
3778
3779         /* First, purge the out_of_order queue. */
3780         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3781                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3782                 __skb_queue_purge(&tp->out_of_order_queue);
3783
3784                 /* Reset SACK state.  A conforming SACK implementation will
3785                  * do the same at a timeout based retransmit.  When a connection
3786                  * is in a sad state like this, we care only about integrity
3787                  * of the connection not performance.
3788                  */
3789                 if (tp->rx_opt.sack_ok)
3790                         tcp_sack_reset(&tp->rx_opt);
3791                 sk_stream_mem_reclaim(sk);
3792         }
3793
3794         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3795                 return 0;
3796
3797         /* If we are really being abused, tell the caller to silently
3798          * drop receive data on the floor.  It will get retransmitted
3799          * and hopefully then we'll have sufficient space.
3800          */
3801         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3802
3803         /* Massive buffer overcommit. */
3804         tp->pred_flags = 0;
3805         return -1;
3806 }
3807
3808
3809 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3810  * As additional protections, we do not touch cwnd in retransmission phases,
3811  * and if application hit its sndbuf limit recently.
3812  */
3813 void tcp_cwnd_application_limited(struct sock *sk)
3814 {
3815         struct tcp_sock *tp = tcp_sk(sk);
3816
3817         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3818             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3819                 /* Limited by application or receiver window. */
3820                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3821                 u32 win_used = max(tp->snd_cwnd_used, init_win);
3822                 if (win_used < tp->snd_cwnd) {
3823                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3824                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3825                 }
3826                 tp->snd_cwnd_used = 0;
3827         }
3828         tp->snd_cwnd_stamp = tcp_time_stamp;
3829 }
3830
3831 static int tcp_should_expand_sndbuf(struct sock *sk)
3832 {
3833         struct tcp_sock *tp = tcp_sk(sk);
3834
3835         /* If the user specified a specific send buffer setting, do
3836          * not modify it.
3837          */
3838         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3839                 return 0;
3840
3841         /* If we are under global TCP memory pressure, do not expand.  */
3842         if (tcp_memory_pressure)
3843                 return 0;
3844
3845         /* If we are under soft global TCP memory pressure, do not expand.  */
3846         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3847                 return 0;
3848
3849         /* If we filled the congestion window, do not expand.  */
3850         if (tp->packets_out >= tp->snd_cwnd)
3851                 return 0;
3852
3853         return 1;
3854 }
3855
3856 /* When incoming ACK allowed to free some skb from write_queue,
3857  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3858  * on the exit from tcp input handler.
3859  *
3860  * PROBLEM: sndbuf expansion does not work well with largesend.
3861  */
3862 static void tcp_new_space(struct sock *sk)
3863 {
3864         struct tcp_sock *tp = tcp_sk(sk);
3865
3866         if (tcp_should_expand_sndbuf(sk)) {
3867                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3868                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3869                     demanded = max_t(unsigned int, tp->snd_cwnd,
3870                                                    tp->reordering + 1);
3871                 sndmem *= 2*demanded;
3872                 if (sndmem > sk->sk_sndbuf)
3873                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3874                 tp->snd_cwnd_stamp = tcp_time_stamp;
3875         }
3876
3877         sk->sk_write_space(sk);
3878 }
3879
3880 static void tcp_check_space(struct sock *sk)
3881 {
3882         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3883                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3884                 if (sk->sk_socket &&
3885                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3886                         tcp_new_space(sk);
3887         }
3888 }
3889
3890 static inline void tcp_data_snd_check(struct sock *sk)
3891 {
3892         tcp_push_pending_frames(sk);
3893         tcp_check_space(sk);
3894 }
3895
3896 /*
3897  * Check if sending an ack is needed.
3898  */
3899 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3900 {
3901         struct tcp_sock *tp = tcp_sk(sk);
3902
3903             /* More than one full frame received... */
3904         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3905              /* ... and right edge of window advances far enough.
3906               * (tcp_recvmsg() will send ACK otherwise). Or...
3907               */
3908              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3909             /* We ACK each frame or... */
3910             tcp_in_quickack_mode(sk) ||
3911             /* We have out of order data. */
3912             (ofo_possible &&
3913              skb_peek(&tp->out_of_order_queue))) {
3914                 /* Then ack it now */
3915                 tcp_send_ack(sk);
3916         } else {
3917                 /* Else, send delayed ack. */
3918                 tcp_send_delayed_ack(sk);
3919         }
3920 }
3921
3922 static inline void tcp_ack_snd_check(struct sock *sk)
3923 {
3924         if (!inet_csk_ack_scheduled(sk)) {
3925                 /* We sent a data segment already. */
3926                 return;
3927         }
3928         __tcp_ack_snd_check(sk, 1);
3929 }
3930
3931 /*
3932  *      This routine is only called when we have urgent data
3933  *      signaled. Its the 'slow' part of tcp_urg. It could be
3934  *      moved inline now as tcp_urg is only called from one
3935  *      place. We handle URGent data wrong. We have to - as
3936  *      BSD still doesn't use the correction from RFC961.
3937  *      For 1003.1g we should support a new option TCP_STDURG to permit
3938  *      either form (or just set the sysctl tcp_stdurg).
3939  */
3940
3941 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3942 {
3943         struct tcp_sock *tp = tcp_sk(sk);
3944         u32 ptr = ntohs(th->urg_ptr);
3945
3946         if (ptr && !sysctl_tcp_stdurg)
3947                 ptr--;
3948         ptr += ntohl(th->seq);
3949
3950         /* Ignore urgent data that we've already seen and read. */
3951         if (after(tp->copied_seq, ptr))
3952                 return;
3953
3954         /* Do not replay urg ptr.
3955          *
3956          * NOTE: interesting situation not covered by specs.
3957          * Misbehaving sender may send urg ptr, pointing to segment,
3958          * which we already have in ofo queue. We are not able to fetch
3959          * such data and will stay in TCP_URG_NOTYET until will be eaten
3960          * by recvmsg(). Seems, we are not obliged to handle such wicked
3961          * situations. But it is worth to think about possibility of some
3962          * DoSes using some hypothetical application level deadlock.
3963          */
3964         if (before(ptr, tp->rcv_nxt))
3965                 return;
3966
3967         /* Do we already have a newer (or duplicate) urgent pointer? */
3968         if (tp->urg_data && !after(ptr, tp->urg_seq))
3969                 return;
3970
3971         /* Tell the world about our new urgent pointer. */
3972         sk_send_sigurg(sk);
3973
3974         /* We may be adding urgent data when the last byte read was
3975          * urgent. To do this requires some care. We cannot just ignore
3976          * tp->copied_seq since we would read the last urgent byte again
3977          * as data, nor can we alter copied_seq until this data arrives
3978          * or we break the semantics of SIOCATMARK (and thus sockatmark())
3979          *
3980          * NOTE. Double Dutch. Rendering to plain English: author of comment
3981          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3982          * and expect that both A and B disappear from stream. This is _wrong_.
3983          * Though this happens in BSD with high probability, this is occasional.
3984          * Any application relying on this is buggy. Note also, that fix "works"
3985          * only in this artificial test. Insert some normal data between A and B and we will
3986          * decline of BSD again. Verdict: it is better to remove to trap
3987          * buggy users.
3988          */
3989         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3990             !sock_flag(sk, SOCK_URGINLINE) &&
3991             tp->copied_seq != tp->rcv_nxt) {
3992                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3993                 tp->copied_seq++;
3994                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3995                         __skb_unlink(skb, &sk->sk_receive_queue);
3996                         __kfree_skb(skb);
3997                 }
3998         }
3999
4000         tp->urg_data   = TCP_URG_NOTYET;
4001         tp->urg_seq    = ptr;
4002
4003         /* Disable header prediction. */
4004         tp->pred_flags = 0;
4005 }
4006
4007 /* This is the 'fast' part of urgent handling. */
4008 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4009 {
4010         struct tcp_sock *tp = tcp_sk(sk);
4011
4012         /* Check if we get a new urgent pointer - normally not. */
4013         if (th->urg)
4014                 tcp_check_urg(sk,th);
4015
4016         /* Do we wait for any urgent data? - normally not... */
4017         if (tp->urg_data == TCP_URG_NOTYET) {
4018                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4019                           th->syn;
4020
4021                 /* Is the urgent pointer pointing into this packet? */
4022                 if (ptr < skb->len) {
4023                         u8 tmp;
4024                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4025                                 BUG();
4026                         tp->urg_data = TCP_URG_VALID | tmp;
4027                         if (!sock_flag(sk, SOCK_DEAD))
4028                                 sk->sk_data_ready(sk, 0);
4029                 }
4030         }
4031 }
4032
4033 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4034 {
4035         struct tcp_sock *tp = tcp_sk(sk);
4036         int chunk = skb->len - hlen;
4037         int err;
4038
4039         local_bh_enable();
4040         if (skb_csum_unnecessary(skb))
4041                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4042         else
4043                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4044                                                        tp->ucopy.iov);
4045
4046         if (!err) {
4047                 tp->ucopy.len -= chunk;
4048                 tp->copied_seq += chunk;
4049                 tcp_rcv_space_adjust(sk);
4050         }
4051
4052         local_bh_disable();
4053         return err;
4054 }
4055
4056 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4057 {
4058         __sum16 result;
4059
4060         if (sock_owned_by_user(sk)) {
4061                 local_bh_enable();
4062                 result = __tcp_checksum_complete(skb);
4063                 local_bh_disable();
4064         } else {
4065                 result = __tcp_checksum_complete(skb);
4066         }
4067         return result;
4068 }
4069
4070 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4071 {
4072         return !skb_csum_unnecessary(skb) &&
4073                 __tcp_checksum_complete_user(sk, skb);
4074 }
4075
4076 #ifdef CONFIG_NET_DMA
4077 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4078 {
4079         struct tcp_sock *tp = tcp_sk(sk);
4080         int chunk = skb->len - hlen;
4081         int dma_cookie;
4082         int copied_early = 0;
4083
4084         if (tp->ucopy.wakeup)
4085                 return 0;
4086
4087         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4088                 tp->ucopy.dma_chan = get_softnet_dma();
4089
4090         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4091
4092                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4093                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4094
4095                 if (dma_cookie < 0)
4096                         goto out;
4097
4098                 tp->ucopy.dma_cookie = dma_cookie;
4099                 copied_early = 1;
4100
4101                 tp->ucopy.len -= chunk;
4102                 tp->copied_seq += chunk;
4103                 tcp_rcv_space_adjust(sk);
4104
4105                 if ((tp->ucopy.len == 0) ||
4106                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4107                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4108                         tp->ucopy.wakeup = 1;
4109                         sk->sk_data_ready(sk, 0);
4110                 }
4111         } else if (chunk > 0) {
4112                 tp->ucopy.wakeup = 1;
4113                 sk->sk_data_ready(sk, 0);
4114         }
4115 out:
4116         return copied_early;
4117 }
4118 #endif /* CONFIG_NET_DMA */
4119
4120 /*
4121  *      TCP receive function for the ESTABLISHED state.
4122  *
4123  *      It is split into a fast path and a slow path. The fast path is
4124  *      disabled when:
4125  *      - A zero window was announced from us - zero window probing
4126  *        is only handled properly in the slow path.
4127  *      - Out of order segments arrived.
4128  *      - Urgent data is expected.
4129  *      - There is no buffer space left
4130  *      - Unexpected TCP flags/window values/header lengths are received
4131  *        (detected by checking the TCP header against pred_flags)
4132  *      - Data is sent in both directions. Fast path only supports pure senders
4133  *        or pure receivers (this means either the sequence number or the ack
4134  *        value must stay constant)
4135  *      - Unexpected TCP option.
4136  *
4137  *      When these conditions are not satisfied it drops into a standard
4138  *      receive procedure patterned after RFC793 to handle all cases.
4139  *      The first three cases are guaranteed by proper pred_flags setting,
4140  *      the rest is checked inline. Fast processing is turned on in
4141  *      tcp_data_queue when everything is OK.
4142  */
4143 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4144                         struct tcphdr *th, unsigned len)
4145 {
4146         struct tcp_sock *tp = tcp_sk(sk);
4147
4148         /*
4149          *      Header prediction.
4150          *      The code loosely follows the one in the famous
4151          *      "30 instruction TCP receive" Van Jacobson mail.
4152          *
4153          *      Van's trick is to deposit buffers into socket queue
4154          *      on a device interrupt, to call tcp_recv function
4155          *      on the receive process context and checksum and copy
4156          *      the buffer to user space. smart...
4157          *
4158          *      Our current scheme is not silly either but we take the
4159          *      extra cost of the net_bh soft interrupt processing...
4160          *      We do checksum and copy also but from device to kernel.
4161          */
4162
4163         tp->rx_opt.saw_tstamp = 0;
4164
4165         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4166          *      if header_prediction is to be made
4167          *      'S' will always be tp->tcp_header_len >> 2
4168          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4169          *  turn it off (when there are holes in the receive
4170          *       space for instance)
4171          *      PSH flag is ignored.
4172          */
4173
4174         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4175                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4176                 int tcp_header_len = tp->tcp_header_len;
4177
4178                 /* Timestamp header prediction: tcp_header_len
4179                  * is automatically equal to th->doff*4 due to pred_flags
4180                  * match.
4181                  */
4182
4183                 /* Check timestamp */
4184                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4185                         __be32 *ptr = (__be32 *)(th + 1);
4186
4187                         /* No? Slow path! */
4188                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4189                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4190                                 goto slow_path;
4191
4192                         tp->rx_opt.saw_tstamp = 1;
4193                         ++ptr;
4194                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4195                         ++ptr;
4196                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4197
4198                         /* If PAWS failed, check it more carefully in slow path */
4199                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4200                                 goto slow_path;
4201
4202                         /* DO NOT update ts_recent here, if checksum fails
4203                          * and timestamp was corrupted part, it will result
4204                          * in a hung connection since we will drop all
4205                          * future packets due to the PAWS test.
4206                          */
4207                 }
4208
4209                 if (len <= tcp_header_len) {
4210                         /* Bulk data transfer: sender */
4211                         if (len == tcp_header_len) {
4212                                 /* Predicted packet is in window by definition.
4213                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4214                                  * Hence, check seq<=rcv_wup reduces to:
4215                                  */
4216                                 if (tcp_header_len ==
4217                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4218                                     tp->rcv_nxt == tp->rcv_wup)
4219                                         tcp_store_ts_recent(tp);
4220
4221                                 /* We know that such packets are checksummed
4222                                  * on entry.
4223                                  */
4224                                 tcp_ack(sk, skb, 0);
4225                                 __kfree_skb(skb);
4226                                 tcp_data_snd_check(sk);
4227                                 return 0;
4228                         } else { /* Header too small */
4229                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4230                                 goto discard;
4231                         }
4232                 } else {
4233                         int eaten = 0;
4234                         int copied_early = 0;
4235
4236                         if (tp->copied_seq == tp->rcv_nxt &&
4237                             len - tcp_header_len <= tp->ucopy.len) {
4238 #ifdef CONFIG_NET_DMA
4239                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4240                                         copied_early = 1;
4241                                         eaten = 1;
4242                                 }
4243 #endif
4244                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4245                                         __set_current_state(TASK_RUNNING);
4246
4247                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4248                                                 eaten = 1;
4249                                 }
4250                                 if (eaten) {
4251                                         /* Predicted packet is in window by definition.
4252                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4253                                          * Hence, check seq<=rcv_wup reduces to:
4254                                          */
4255                                         if (tcp_header_len ==
4256                                             (sizeof(struct tcphdr) +
4257                                              TCPOLEN_TSTAMP_ALIGNED) &&
4258                                             tp->rcv_nxt == tp->rcv_wup)
4259                                                 tcp_store_ts_recent(tp);
4260
4261                                         tcp_rcv_rtt_measure_ts(sk, skb);
4262
4263                                         __skb_pull(skb, tcp_header_len);
4264                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4265                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4266                                 }
4267                                 if (copied_early)
4268                                         tcp_cleanup_rbuf(sk, skb->len);
4269                         }
4270                         if (!eaten) {
4271                                 if (tcp_checksum_complete_user(sk, skb))
4272                                         goto csum_error;
4273
4274                                 /* Predicted packet is in window by definition.
4275                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4276                                  * Hence, check seq<=rcv_wup reduces to:
4277                                  */
4278                                 if (tcp_header_len ==
4279                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4280                                     tp->rcv_nxt == tp->rcv_wup)
4281                                         tcp_store_ts_recent(tp);
4282
4283                                 tcp_rcv_rtt_measure_ts(sk, skb);
4284
4285                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4286                                         goto step5;
4287
4288                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4289
4290                                 /* Bulk data transfer: receiver */
4291                                 __skb_pull(skb,tcp_header_len);
4292                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4293                                 sk_stream_set_owner_r(skb, sk);
4294                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4295                         }
4296
4297                         tcp_event_data_recv(sk, skb);
4298
4299                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4300                                 /* Well, only one small jumplet in fast path... */
4301                                 tcp_ack(sk, skb, FLAG_DATA);
4302                                 tcp_data_snd_check(sk);
4303                                 if (!inet_csk_ack_scheduled(sk))
4304                                         goto no_ack;
4305                         }
4306
4307                         __tcp_ack_snd_check(sk, 0);
4308 no_ack:
4309 #ifdef CONFIG_NET_DMA
4310                         if (copied_early)
4311                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4312                         else
4313 #endif
4314                         if (eaten)
4315                                 __kfree_skb(skb);
4316                         else
4317                                 sk->sk_data_ready(sk, 0);
4318                         return 0;
4319                 }
4320         }
4321
4322 slow_path:
4323         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4324                 goto csum_error;
4325
4326         /*
4327          * RFC1323: H1. Apply PAWS check first.
4328          */
4329         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4330             tcp_paws_discard(sk, skb)) {
4331                 if (!th->rst) {
4332                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4333                         tcp_send_dupack(sk, skb);
4334                         goto discard;
4335                 }
4336                 /* Resets are accepted even if PAWS failed.
4337
4338                    ts_recent update must be made after we are sure
4339                    that the packet is in window.
4340                  */
4341         }
4342
4343         /*
4344          *      Standard slow path.
4345          */
4346
4347         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4348                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4349                  * (RST) segments are validated by checking their SEQ-fields."
4350                  * And page 69: "If an incoming segment is not acceptable,
4351                  * an acknowledgment should be sent in reply (unless the RST bit
4352                  * is set, if so drop the segment and return)".
4353                  */
4354                 if (!th->rst)
4355                         tcp_send_dupack(sk, skb);
4356                 goto discard;
4357         }
4358
4359         if (th->rst) {
4360                 tcp_reset(sk);
4361                 goto discard;
4362         }
4363
4364         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4365
4366         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4367                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4368                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4369                 tcp_reset(sk);
4370                 return 1;
4371         }
4372
4373 step5:
4374         if (th->ack)
4375                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4376
4377         tcp_rcv_rtt_measure_ts(sk, skb);
4378
4379         /* Process urgent data. */
4380         tcp_urg(sk, skb, th);
4381
4382         /* step 7: process the segment text */
4383         tcp_data_queue(sk, skb);
4384
4385         tcp_data_snd_check(sk);
4386         tcp_ack_snd_check(sk);
4387         return 0;
4388
4389 csum_error:
4390         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4391
4392 discard:
4393         __kfree_skb(skb);
4394         return 0;
4395 }
4396
4397 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4398                                          struct tcphdr *th, unsigned len)
4399 {
4400         struct tcp_sock *tp = tcp_sk(sk);
4401         struct inet_connection_sock *icsk = inet_csk(sk);
4402         int saved_clamp = tp->rx_opt.mss_clamp;
4403
4404         tcp_parse_options(skb, &tp->rx_opt, 0);
4405
4406         if (th->ack) {
4407                 /* rfc793:
4408                  * "If the state is SYN-SENT then
4409                  *    first check the ACK bit
4410                  *      If the ACK bit is set
4411                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4412                  *        a reset (unless the RST bit is set, if so drop
4413                  *        the segment and return)"
4414                  *
4415                  *  We do not send data with SYN, so that RFC-correct
4416                  *  test reduces to:
4417                  */
4418                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4419                         goto reset_and_undo;
4420
4421                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4422                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4423                              tcp_time_stamp)) {
4424                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4425                         goto reset_and_undo;
4426                 }
4427
4428                 /* Now ACK is acceptable.
4429                  *
4430                  * "If the RST bit is set
4431                  *    If the ACK was acceptable then signal the user "error:
4432                  *    connection reset", drop the segment, enter CLOSED state,
4433                  *    delete TCB, and return."
4434                  */
4435
4436                 if (th->rst) {
4437                         tcp_reset(sk);
4438                         goto discard;
4439                 }
4440
4441                 /* rfc793:
4442                  *   "fifth, if neither of the SYN or RST bits is set then
4443                  *    drop the segment and return."
4444                  *
4445                  *    See note below!
4446                  *                                        --ANK(990513)
4447                  */
4448                 if (!th->syn)
4449                         goto discard_and_undo;
4450
4451                 /* rfc793:
4452                  *   "If the SYN bit is on ...
4453                  *    are acceptable then ...
4454                  *    (our SYN has been ACKed), change the connection
4455                  *    state to ESTABLISHED..."
4456                  */
4457
4458                 TCP_ECN_rcv_synack(tp, th);
4459
4460                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4461                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4462
4463                 /* Ok.. it's good. Set up sequence numbers and
4464                  * move to established.
4465                  */
4466                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4467                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4468
4469                 /* RFC1323: The window in SYN & SYN/ACK segments is
4470                  * never scaled.
4471                  */
4472                 tp->snd_wnd = ntohs(th->window);
4473                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4474
4475                 if (!tp->rx_opt.wscale_ok) {
4476                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4477                         tp->window_clamp = min(tp->window_clamp, 65535U);
4478                 }
4479
4480                 if (tp->rx_opt.saw_tstamp) {
4481                         tp->rx_opt.tstamp_ok       = 1;
4482                         tp->tcp_header_len =
4483                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4484                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4485                         tcp_store_ts_recent(tp);
4486                 } else {
4487                         tp->tcp_header_len = sizeof(struct tcphdr);
4488                 }
4489
4490                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4491                         tp->rx_opt.sack_ok |= 2;
4492
4493                 tcp_mtup_init(sk);
4494                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4495                 tcp_initialize_rcv_mss(sk);
4496
4497                 /* Remember, tcp_poll() does not lock socket!
4498                  * Change state from SYN-SENT only after copied_seq
4499                  * is initialized. */
4500                 tp->copied_seq = tp->rcv_nxt;
4501                 smp_mb();
4502                 tcp_set_state(sk, TCP_ESTABLISHED);
4503
4504                 security_inet_conn_established(sk, skb);
4505
4506                 /* Make sure socket is routed, for correct metrics.  */
4507                 icsk->icsk_af_ops->rebuild_header(sk);
4508
4509                 tcp_init_metrics(sk);
4510
4511                 tcp_init_congestion_control(sk);
4512
4513                 /* Prevent spurious tcp_cwnd_restart() on first data
4514                  * packet.
4515                  */
4516                 tp->lsndtime = tcp_time_stamp;
4517
4518                 tcp_init_buffer_space(sk);
4519
4520                 if (sock_flag(sk, SOCK_KEEPOPEN))
4521                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4522
4523                 if (!tp->rx_opt.snd_wscale)
4524                         __tcp_fast_path_on(tp, tp->snd_wnd);
4525                 else
4526                         tp->pred_flags = 0;
4527
4528                 if (!sock_flag(sk, SOCK_DEAD)) {
4529                         sk->sk_state_change(sk);
4530                         sk_wake_async(sk, 0, POLL_OUT);
4531                 }
4532
4533                 if (sk->sk_write_pending ||
4534                     icsk->icsk_accept_queue.rskq_defer_accept ||
4535                     icsk->icsk_ack.pingpong) {
4536                         /* Save one ACK. Data will be ready after
4537                          * several ticks, if write_pending is set.
4538                          *
4539                          * It may be deleted, but with this feature tcpdumps
4540                          * look so _wonderfully_ clever, that I was not able
4541                          * to stand against the temptation 8)     --ANK
4542                          */
4543                         inet_csk_schedule_ack(sk);
4544                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4545                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4546                         tcp_incr_quickack(sk);
4547                         tcp_enter_quickack_mode(sk);
4548                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4549                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4550
4551 discard:
4552                         __kfree_skb(skb);
4553                         return 0;
4554                 } else {
4555                         tcp_send_ack(sk);
4556                 }
4557                 return -1;
4558         }
4559
4560         /* No ACK in the segment */
4561
4562         if (th->rst) {
4563                 /* rfc793:
4564                  * "If the RST bit is set
4565                  *
4566                  *      Otherwise (no ACK) drop the segment and return."
4567                  */
4568
4569                 goto discard_and_undo;
4570         }
4571
4572         /* PAWS check. */
4573         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4574                 goto discard_and_undo;
4575
4576         if (th->syn) {
4577                 /* We see SYN without ACK. It is attempt of
4578                  * simultaneous connect with crossed SYNs.
4579                  * Particularly, it can be connect to self.
4580                  */
4581                 tcp_set_state(sk, TCP_SYN_RECV);
4582
4583                 if (tp->rx_opt.saw_tstamp) {
4584                         tp->rx_opt.tstamp_ok = 1;
4585                         tcp_store_ts_recent(tp);
4586                         tp->tcp_header_len =
4587                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4588                 } else {
4589                         tp->tcp_header_len = sizeof(struct tcphdr);
4590                 }
4591
4592                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4593                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4594
4595                 /* RFC1323: The window in SYN & SYN/ACK segments is
4596                  * never scaled.
4597                  */
4598                 tp->snd_wnd    = ntohs(th->window);
4599                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4600                 tp->max_window = tp->snd_wnd;
4601
4602                 TCP_ECN_rcv_syn(tp, th);
4603
4604                 tcp_mtup_init(sk);
4605                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4606                 tcp_initialize_rcv_mss(sk);
4607
4608
4609                 tcp_send_synack(sk);
4610 #if 0
4611                 /* Note, we could accept data and URG from this segment.
4612                  * There are no obstacles to make this.
4613                  *
4614                  * However, if we ignore data in ACKless segments sometimes,
4615                  * we have no reasons to accept it sometimes.
4616                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4617                  * is not flawless. So, discard packet for sanity.
4618                  * Uncomment this return to process the data.
4619                  */
4620                 return -1;
4621 #else
4622                 goto discard;
4623 #endif
4624         }
4625         /* "fifth, if neither of the SYN or RST bits is set then
4626          * drop the segment and return."
4627          */
4628
4629 discard_and_undo:
4630         tcp_clear_options(&tp->rx_opt);
4631         tp->rx_opt.mss_clamp = saved_clamp;
4632         goto discard;
4633
4634 reset_and_undo:
4635         tcp_clear_options(&tp->rx_opt);
4636         tp->rx_opt.mss_clamp = saved_clamp;
4637         return 1;
4638 }
4639
4640
4641 /*
4642  *      This function implements the receiving procedure of RFC 793 for
4643  *      all states except ESTABLISHED and TIME_WAIT.
4644  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4645  *      address independent.
4646  */
4647
4648 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4649                           struct tcphdr *th, unsigned len)
4650 {
4651         struct tcp_sock *tp = tcp_sk(sk);
4652         struct inet_connection_sock *icsk = inet_csk(sk);
4653         int queued = 0;
4654
4655         tp->rx_opt.saw_tstamp = 0;
4656
4657         switch (sk->sk_state) {
4658         case TCP_CLOSE:
4659                 goto discard;
4660
4661         case TCP_LISTEN:
4662                 if (th->ack)
4663                         return 1;
4664
4665                 if (th->rst)
4666                         goto discard;
4667
4668                 if (th->syn) {
4669                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4670                                 return 1;
4671
4672                         /* Now we have several options: In theory there is
4673                          * nothing else in the frame. KA9Q has an option to
4674                          * send data with the syn, BSD accepts data with the
4675                          * syn up to the [to be] advertised window and
4676                          * Solaris 2.1 gives you a protocol error. For now
4677                          * we just ignore it, that fits the spec precisely
4678                          * and avoids incompatibilities. It would be nice in
4679                          * future to drop through and process the data.
4680                          *
4681                          * Now that TTCP is starting to be used we ought to
4682                          * queue this data.
4683                          * But, this leaves one open to an easy denial of
4684                          * service attack, and SYN cookies can't defend
4685                          * against this problem. So, we drop the data
4686                          * in the interest of security over speed unless
4687                          * it's still in use.
4688                          */
4689                         kfree_skb(skb);
4690                         return 0;
4691                 }
4692                 goto discard;
4693
4694         case TCP_SYN_SENT:
4695                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4696                 if (queued >= 0)
4697                         return queued;
4698
4699                 /* Do step6 onward by hand. */
4700                 tcp_urg(sk, skb, th);
4701                 __kfree_skb(skb);
4702                 tcp_data_snd_check(sk);
4703                 return 0;
4704         }
4705
4706         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4707             tcp_paws_discard(sk, skb)) {
4708                 if (!th->rst) {
4709                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4710                         tcp_send_dupack(sk, skb);
4711                         goto discard;
4712                 }
4713                 /* Reset is accepted even if it did not pass PAWS. */
4714         }
4715
4716         /* step 1: check sequence number */
4717         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4718                 if (!th->rst)
4719                         tcp_send_dupack(sk, skb);
4720                 goto discard;
4721         }
4722
4723         /* step 2: check RST bit */
4724         if (th->rst) {
4725                 tcp_reset(sk);
4726                 goto discard;
4727         }
4728
4729         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4730
4731         /* step 3: check security and precedence [ignored] */
4732
4733         /*      step 4:
4734          *
4735          *      Check for a SYN in window.
4736          */
4737         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4738                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4739                 tcp_reset(sk);
4740                 return 1;
4741         }
4742
4743         /* step 5: check the ACK field */
4744         if (th->ack) {
4745                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4746
4747                 switch (sk->sk_state) {
4748                 case TCP_SYN_RECV:
4749                         if (acceptable) {
4750                                 tp->copied_seq = tp->rcv_nxt;
4751                                 smp_mb();
4752                                 tcp_set_state(sk, TCP_ESTABLISHED);
4753                                 sk->sk_state_change(sk);
4754
4755                                 /* Note, that this wakeup is only for marginal
4756                                  * crossed SYN case. Passively open sockets
4757                                  * are not waked up, because sk->sk_sleep ==
4758                                  * NULL and sk->sk_socket == NULL.
4759                                  */
4760                                 if (sk->sk_socket) {
4761                                         sk_wake_async(sk,0,POLL_OUT);
4762                                 }
4763
4764                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4765                                 tp->snd_wnd = ntohs(th->window) <<
4766                                               tp->rx_opt.snd_wscale;
4767                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4768                                             TCP_SKB_CB(skb)->seq);
4769
4770                                 /* tcp_ack considers this ACK as duplicate
4771                                  * and does not calculate rtt.
4772                                  * Fix it at least with timestamps.
4773                                  */
4774                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4775                                     !tp->srtt)
4776                                         tcp_ack_saw_tstamp(sk, 0);
4777
4778                                 if (tp->rx_opt.tstamp_ok)
4779                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4780
4781                                 /* Make sure socket is routed, for
4782                                  * correct metrics.
4783                                  */
4784                                 icsk->icsk_af_ops->rebuild_header(sk);
4785
4786                                 tcp_init_metrics(sk);
4787
4788                                 tcp_init_congestion_control(sk);
4789
4790                                 /* Prevent spurious tcp_cwnd_restart() on
4791                                  * first data packet.
4792                                  */
4793                                 tp->lsndtime = tcp_time_stamp;
4794
4795                                 tcp_mtup_init(sk);
4796                                 tcp_initialize_rcv_mss(sk);
4797                                 tcp_init_buffer_space(sk);
4798                                 tcp_fast_path_on(tp);
4799                         } else {
4800                                 return 1;
4801                         }
4802                         break;
4803
4804                 case TCP_FIN_WAIT1:
4805                         if (tp->snd_una == tp->write_seq) {
4806                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4807                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4808                                 dst_confirm(sk->sk_dst_cache);
4809
4810                                 if (!sock_flag(sk, SOCK_DEAD))
4811                                         /* Wake up lingering close() */
4812                                         sk->sk_state_change(sk);
4813                                 else {
4814                                         int tmo;
4815
4816                                         if (tp->linger2 < 0 ||
4817                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4818                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4819                                                 tcp_done(sk);
4820                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4821                                                 return 1;
4822                                         }
4823
4824                                         tmo = tcp_fin_time(sk);
4825                                         if (tmo > TCP_TIMEWAIT_LEN) {
4826                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4827                                         } else if (th->fin || sock_owned_by_user(sk)) {
4828                                                 /* Bad case. We could lose such FIN otherwise.
4829                                                  * It is not a big problem, but it looks confusing
4830                                                  * and not so rare event. We still can lose it now,
4831                                                  * if it spins in bh_lock_sock(), but it is really
4832                                                  * marginal case.
4833                                                  */
4834                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4835                                         } else {
4836                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4837                                                 goto discard;
4838                                         }
4839                                 }
4840                         }
4841                         break;
4842
4843                 case TCP_CLOSING:
4844                         if (tp->snd_una == tp->write_seq) {
4845                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4846                                 goto discard;
4847                         }
4848                         break;
4849
4850                 case TCP_LAST_ACK:
4851                         if (tp->snd_una == tp->write_seq) {
4852                                 tcp_update_metrics(sk);
4853                                 tcp_done(sk);
4854                                 goto discard;
4855                         }
4856                         break;
4857                 }
4858         } else
4859                 goto discard;
4860
4861         /* step 6: check the URG bit */
4862         tcp_urg(sk, skb, th);
4863
4864         /* step 7: process the segment text */
4865         switch (sk->sk_state) {
4866         case TCP_CLOSE_WAIT:
4867         case TCP_CLOSING:
4868         case TCP_LAST_ACK:
4869                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4870                         break;
4871         case TCP_FIN_WAIT1:
4872         case TCP_FIN_WAIT2:
4873                 /* RFC 793 says to queue data in these states,
4874                  * RFC 1122 says we MUST send a reset.
4875                  * BSD 4.4 also does reset.
4876                  */
4877                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4878                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4879                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4880                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4881                                 tcp_reset(sk);
4882                                 return 1;
4883                         }
4884                 }
4885                 /* Fall through */
4886         case TCP_ESTABLISHED:
4887                 tcp_data_queue(sk, skb);
4888                 queued = 1;
4889                 break;
4890         }
4891
4892         /* tcp_data could move socket to TIME-WAIT */
4893         if (sk->sk_state != TCP_CLOSE) {
4894                 tcp_data_snd_check(sk);
4895                 tcp_ack_snd_check(sk);
4896         }
4897
4898         if (!queued) {
4899 discard:
4900                 __kfree_skb(skb);
4901         }
4902         return 0;
4903 }
4904
4905 EXPORT_SYMBOL(sysctl_tcp_ecn);
4906 EXPORT_SYMBOL(sysctl_tcp_reordering);
4907 EXPORT_SYMBOL(tcp_parse_options);
4908 EXPORT_SYMBOL(tcp_rcv_established);
4909 EXPORT_SYMBOL(tcp_rcv_state_process);
4910 EXPORT_SYMBOL(tcp_initialize_rcv_mss);