Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[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 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained DSACK info */
107
108 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
109 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
110 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
111 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
112 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
113
114 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
115 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
116 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
117
118 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
119
120 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
121
122 /* Adapt the MSS value used to make delayed ack decision to the
123  * real world.
124  */
125 static void tcp_measure_rcv_mss(struct sock *sk,
126                                 const struct sk_buff *skb)
127 {
128         struct inet_connection_sock *icsk = inet_csk(sk);
129         const unsigned int lss = icsk->icsk_ack.last_seg_size;
130         unsigned int len;
131
132         icsk->icsk_ack.last_seg_size = 0;
133
134         /* skb->len may jitter because of SACKs, even if peer
135          * sends good full-sized frames.
136          */
137         len = skb_shinfo(skb)->gso_size ?: skb->len;
138         if (len >= icsk->icsk_ack.rcv_mss) {
139                 icsk->icsk_ack.rcv_mss = len;
140         } else {
141                 /* Otherwise, we make more careful check taking into account,
142                  * that SACKs block is variable.
143                  *
144                  * "len" is invariant segment length, including TCP header.
145                  */
146                 len += skb->data - skb_transport_header(skb);
147                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
148                     /* If PSH is not set, packet should be
149                      * full sized, provided peer TCP is not badly broken.
150                      * This observation (if it is correct 8)) allows
151                      * to handle super-low mtu links fairly.
152                      */
153                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
154                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
155                         /* Subtract also invariant (if peer is RFC compliant),
156                          * tcp header plus fixed timestamp option length.
157                          * Resulting "len" is MSS free of SACK jitter.
158                          */
159                         len -= tcp_sk(sk)->tcp_header_len;
160                         icsk->icsk_ack.last_seg_size = len;
161                         if (len == lss) {
162                                 icsk->icsk_ack.rcv_mss = len;
163                                 return;
164                         }
165                 }
166                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
167                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
168                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
169         }
170 }
171
172 static void tcp_incr_quickack(struct sock *sk)
173 {
174         struct inet_connection_sock *icsk = inet_csk(sk);
175         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
176
177         if (quickacks==0)
178                 quickacks=2;
179         if (quickacks > icsk->icsk_ack.quick)
180                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
181 }
182
183 void tcp_enter_quickack_mode(struct sock *sk)
184 {
185         struct inet_connection_sock *icsk = inet_csk(sk);
186         tcp_incr_quickack(sk);
187         icsk->icsk_ack.pingpong = 0;
188         icsk->icsk_ack.ato = TCP_ATO_MIN;
189 }
190
191 /* Send ACKs quickly, if "quick" count is not exhausted
192  * and the session is not interactive.
193  */
194
195 static inline int tcp_in_quickack_mode(const struct sock *sk)
196 {
197         const struct inet_connection_sock *icsk = inet_csk(sk);
198         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
199 }
200
201 /* Buffer size and advertised window tuning.
202  *
203  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
204  */
205
206 static void tcp_fixup_sndbuf(struct sock *sk)
207 {
208         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
209                      sizeof(struct sk_buff);
210
211         if (sk->sk_sndbuf < 3 * sndmem)
212                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
213 }
214
215 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
216  *
217  * All tcp_full_space() is split to two parts: "network" buffer, allocated
218  * forward and advertised in receiver window (tp->rcv_wnd) and
219  * "application buffer", required to isolate scheduling/application
220  * latencies from network.
221  * window_clamp is maximal advertised window. It can be less than
222  * tcp_full_space(), in this case tcp_full_space() - window_clamp
223  * is reserved for "application" buffer. The less window_clamp is
224  * the smoother our behaviour from viewpoint of network, but the lower
225  * throughput and the higher sensitivity of the connection to losses. 8)
226  *
227  * rcv_ssthresh is more strict window_clamp used at "slow start"
228  * phase to predict further behaviour of this connection.
229  * It is used for two goals:
230  * - to enforce header prediction at sender, even when application
231  *   requires some significant "application buffer". It is check #1.
232  * - to prevent pruning of receive queue because of misprediction
233  *   of receiver window. Check #2.
234  *
235  * The scheme does not work when sender sends good segments opening
236  * window and then starts to feed us spaghetti. But it should work
237  * in common situations. Otherwise, we have to rely on queue collapsing.
238  */
239
240 /* Slow part of check#2. */
241 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
242 {
243         struct tcp_sock *tp = tcp_sk(sk);
244         /* Optimize this! */
245         int truesize = tcp_win_from_space(skb->truesize)/2;
246         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
247
248         while (tp->rcv_ssthresh <= window) {
249                 if (truesize <= skb->len)
250                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
251
252                 truesize >>= 1;
253                 window >>= 1;
254         }
255         return 0;
256 }
257
258 static void tcp_grow_window(struct sock *sk,
259                             struct sk_buff *skb)
260 {
261         struct tcp_sock *tp = tcp_sk(sk);
262
263         /* Check #1 */
264         if (tp->rcv_ssthresh < tp->window_clamp &&
265             (int)tp->rcv_ssthresh < tcp_space(sk) &&
266             !tcp_memory_pressure) {
267                 int incr;
268
269                 /* Check #2. Increase window, if skb with such overhead
270                  * will fit to rcvbuf in future.
271                  */
272                 if (tcp_win_from_space(skb->truesize) <= skb->len)
273                         incr = 2*tp->advmss;
274                 else
275                         incr = __tcp_grow_window(sk, skb);
276
277                 if (incr) {
278                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
279                         inet_csk(sk)->icsk_ack.quick |= 1;
280                 }
281         }
282 }
283
284 /* 3. Tuning rcvbuf, when connection enters established state. */
285
286 static void tcp_fixup_rcvbuf(struct sock *sk)
287 {
288         struct tcp_sock *tp = tcp_sk(sk);
289         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
290
291         /* Try to select rcvbuf so that 4 mss-sized segments
292          * will fit to window and corresponding skbs will fit to our rcvbuf.
293          * (was 3; 4 is minimum to allow fast retransmit to work.)
294          */
295         while (tcp_win_from_space(rcvmem) < tp->advmss)
296                 rcvmem += 128;
297         if (sk->sk_rcvbuf < 4 * rcvmem)
298                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
299 }
300
301 /* 4. Try to fixup all. It is made immediately after connection enters
302  *    established state.
303  */
304 static void tcp_init_buffer_space(struct sock *sk)
305 {
306         struct tcp_sock *tp = tcp_sk(sk);
307         int maxwin;
308
309         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
310                 tcp_fixup_rcvbuf(sk);
311         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
312                 tcp_fixup_sndbuf(sk);
313
314         tp->rcvq_space.space = tp->rcv_wnd;
315
316         maxwin = tcp_full_space(sk);
317
318         if (tp->window_clamp >= maxwin) {
319                 tp->window_clamp = maxwin;
320
321                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
322                         tp->window_clamp = max(maxwin -
323                                                (maxwin >> sysctl_tcp_app_win),
324                                                4 * tp->advmss);
325         }
326
327         /* Force reservation of one segment. */
328         if (sysctl_tcp_app_win &&
329             tp->window_clamp > 2 * tp->advmss &&
330             tp->window_clamp + tp->advmss > maxwin)
331                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
332
333         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
334         tp->snd_cwnd_stamp = tcp_time_stamp;
335 }
336
337 /* 5. Recalculate window clamp after socket hit its memory bounds. */
338 static void tcp_clamp_window(struct sock *sk)
339 {
340         struct tcp_sock *tp = tcp_sk(sk);
341         struct inet_connection_sock *icsk = inet_csk(sk);
342
343         icsk->icsk_ack.quick = 0;
344
345         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
346             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
347             !tcp_memory_pressure &&
348             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
349                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
350                                     sysctl_tcp_rmem[2]);
351         }
352         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
353                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
354 }
355
356
357 /* Initialize RCV_MSS value.
358  * RCV_MSS is an our guess about MSS used by the peer.
359  * We haven't any direct information about the MSS.
360  * It's better to underestimate the RCV_MSS rather than overestimate.
361  * Overestimations make us ACKing less frequently than needed.
362  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
363  */
364 void tcp_initialize_rcv_mss(struct sock *sk)
365 {
366         struct tcp_sock *tp = tcp_sk(sk);
367         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
368
369         hint = min(hint, tp->rcv_wnd/2);
370         hint = min(hint, TCP_MIN_RCVMSS);
371         hint = max(hint, TCP_MIN_MSS);
372
373         inet_csk(sk)->icsk_ack.rcv_mss = hint;
374 }
375
376 /* Receiver "autotuning" code.
377  *
378  * The algorithm for RTT estimation w/o timestamps is based on
379  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
380  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
381  *
382  * More detail on this code can be found at
383  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
384  * though this reference is out of date.  A new paper
385  * is pending.
386  */
387 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
388 {
389         u32 new_sample = tp->rcv_rtt_est.rtt;
390         long m = sample;
391
392         if (m == 0)
393                 m = 1;
394
395         if (new_sample != 0) {
396                 /* If we sample in larger samples in the non-timestamp
397                  * case, we could grossly overestimate the RTT especially
398                  * with chatty applications or bulk transfer apps which
399                  * are stalled on filesystem I/O.
400                  *
401                  * Also, since we are only going for a minimum in the
402                  * non-timestamp case, we do not smooth things out
403                  * else with timestamps disabled convergence takes too
404                  * long.
405                  */
406                 if (!win_dep) {
407                         m -= (new_sample >> 3);
408                         new_sample += m;
409                 } else if (m < new_sample)
410                         new_sample = m << 3;
411         } else {
412                 /* No previous measure. */
413                 new_sample = m << 3;
414         }
415
416         if (tp->rcv_rtt_est.rtt != new_sample)
417                 tp->rcv_rtt_est.rtt = new_sample;
418 }
419
420 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
421 {
422         if (tp->rcv_rtt_est.time == 0)
423                 goto new_measure;
424         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
425                 return;
426         tcp_rcv_rtt_update(tp,
427                            jiffies - tp->rcv_rtt_est.time,
428                            1);
429
430 new_measure:
431         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
432         tp->rcv_rtt_est.time = tcp_time_stamp;
433 }
434
435 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
436 {
437         struct tcp_sock *tp = tcp_sk(sk);
438         if (tp->rx_opt.rcv_tsecr &&
439             (TCP_SKB_CB(skb)->end_seq -
440              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
441                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
442 }
443
444 /*
445  * This function should be called every time data is copied to user space.
446  * It calculates the appropriate TCP receive buffer space.
447  */
448 void tcp_rcv_space_adjust(struct sock *sk)
449 {
450         struct tcp_sock *tp = tcp_sk(sk);
451         int time;
452         int space;
453
454         if (tp->rcvq_space.time == 0)
455                 goto new_measure;
456
457         time = tcp_time_stamp - tp->rcvq_space.time;
458         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
459             tp->rcv_rtt_est.rtt == 0)
460                 return;
461
462         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
463
464         space = max(tp->rcvq_space.space, space);
465
466         if (tp->rcvq_space.space != space) {
467                 int rcvmem;
468
469                 tp->rcvq_space.space = space;
470
471                 if (sysctl_tcp_moderate_rcvbuf &&
472                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
473                         int new_clamp = space;
474
475                         /* Receive space grows, normalize in order to
476                          * take into account packet headers and sk_buff
477                          * structure overhead.
478                          */
479                         space /= tp->advmss;
480                         if (!space)
481                                 space = 1;
482                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
483                                   16 + sizeof(struct sk_buff));
484                         while (tcp_win_from_space(rcvmem) < tp->advmss)
485                                 rcvmem += 128;
486                         space *= rcvmem;
487                         space = min(space, sysctl_tcp_rmem[2]);
488                         if (space > sk->sk_rcvbuf) {
489                                 sk->sk_rcvbuf = space;
490
491                                 /* Make the window clamp follow along.  */
492                                 tp->window_clamp = new_clamp;
493                         }
494                 }
495         }
496
497 new_measure:
498         tp->rcvq_space.seq = tp->copied_seq;
499         tp->rcvq_space.time = tcp_time_stamp;
500 }
501
502 /* There is something which you must keep in mind when you analyze the
503  * behavior of the tp->ato delayed ack timeout interval.  When a
504  * connection starts up, we want to ack as quickly as possible.  The
505  * problem is that "good" TCP's do slow start at the beginning of data
506  * transmission.  The means that until we send the first few ACK's the
507  * sender will sit on his end and only queue most of his data, because
508  * he can only send snd_cwnd unacked packets at any given time.  For
509  * each ACK we send, he increments snd_cwnd and transmits more of his
510  * queue.  -DaveM
511  */
512 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
513 {
514         struct tcp_sock *tp = tcp_sk(sk);
515         struct inet_connection_sock *icsk = inet_csk(sk);
516         u32 now;
517
518         inet_csk_schedule_ack(sk);
519
520         tcp_measure_rcv_mss(sk, skb);
521
522         tcp_rcv_rtt_measure(tp);
523
524         now = tcp_time_stamp;
525
526         if (!icsk->icsk_ack.ato) {
527                 /* The _first_ data packet received, initialize
528                  * delayed ACK engine.
529                  */
530                 tcp_incr_quickack(sk);
531                 icsk->icsk_ack.ato = TCP_ATO_MIN;
532         } else {
533                 int m = now - icsk->icsk_ack.lrcvtime;
534
535                 if (m <= TCP_ATO_MIN/2) {
536                         /* The fastest case is the first. */
537                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
538                 } else if (m < icsk->icsk_ack.ato) {
539                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
540                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
541                                 icsk->icsk_ack.ato = icsk->icsk_rto;
542                 } else if (m > icsk->icsk_rto) {
543                         /* Too long gap. Apparently sender failed to
544                          * restart window, so that we send ACKs quickly.
545                          */
546                         tcp_incr_quickack(sk);
547                         sk_stream_mem_reclaim(sk);
548                 }
549         }
550         icsk->icsk_ack.lrcvtime = now;
551
552         TCP_ECN_check_ce(tp, skb);
553
554         if (skb->len >= 128)
555                 tcp_grow_window(sk, skb);
556 }
557
558 static u32 tcp_rto_min(struct sock *sk)
559 {
560         struct dst_entry *dst = __sk_dst_get(sk);
561         u32 rto_min = TCP_RTO_MIN;
562
563         if (dst_metric_locked(dst, RTAX_RTO_MIN))
564                 rto_min = dst->metrics[RTAX_RTO_MIN-1];
565         return rto_min;
566 }
567
568 /* Called to compute a smoothed rtt estimate. The data fed to this
569  * routine either comes from timestamps, or from segments that were
570  * known _not_ to have been retransmitted [see Karn/Partridge
571  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
572  * piece by Van Jacobson.
573  * NOTE: the next three routines used to be one big routine.
574  * To save cycles in the RFC 1323 implementation it was better to break
575  * it up into three procedures. -- erics
576  */
577 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
578 {
579         struct tcp_sock *tp = tcp_sk(sk);
580         long m = mrtt; /* RTT */
581
582         /*      The following amusing code comes from Jacobson's
583          *      article in SIGCOMM '88.  Note that rtt and mdev
584          *      are scaled versions of rtt and mean deviation.
585          *      This is designed to be as fast as possible
586          *      m stands for "measurement".
587          *
588          *      On a 1990 paper the rto value is changed to:
589          *      RTO = rtt + 4 * mdev
590          *
591          * Funny. This algorithm seems to be very broken.
592          * These formulae increase RTO, when it should be decreased, increase
593          * too slowly, when it should be increased quickly, decrease too quickly
594          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
595          * does not matter how to _calculate_ it. Seems, it was trap
596          * that VJ failed to avoid. 8)
597          */
598         if (m == 0)
599                 m = 1;
600         if (tp->srtt != 0) {
601                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
602                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
603                 if (m < 0) {
604                         m = -m;         /* m is now abs(error) */
605                         m -= (tp->mdev >> 2);   /* similar update on mdev */
606                         /* This is similar to one of Eifel findings.
607                          * Eifel blocks mdev updates when rtt decreases.
608                          * This solution is a bit different: we use finer gain
609                          * for mdev in this case (alpha*beta).
610                          * Like Eifel it also prevents growth of rto,
611                          * but also it limits too fast rto decreases,
612                          * happening in pure Eifel.
613                          */
614                         if (m > 0)
615                                 m >>= 3;
616                 } else {
617                         m -= (tp->mdev >> 2);   /* similar update on mdev */
618                 }
619                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
620                 if (tp->mdev > tp->mdev_max) {
621                         tp->mdev_max = tp->mdev;
622                         if (tp->mdev_max > tp->rttvar)
623                                 tp->rttvar = tp->mdev_max;
624                 }
625                 if (after(tp->snd_una, tp->rtt_seq)) {
626                         if (tp->mdev_max < tp->rttvar)
627                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
628                         tp->rtt_seq = tp->snd_nxt;
629                         tp->mdev_max = tcp_rto_min(sk);
630                 }
631         } else {
632                 /* no previous measure. */
633                 tp->srtt = m<<3;        /* take the measured time to be rtt */
634                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
635                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
636                 tp->rtt_seq = tp->snd_nxt;
637         }
638 }
639
640 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
641  * routine referred to above.
642  */
643 static inline void tcp_set_rto(struct sock *sk)
644 {
645         const struct tcp_sock *tp = tcp_sk(sk);
646         /* Old crap is replaced with new one. 8)
647          *
648          * More seriously:
649          * 1. If rtt variance happened to be less 50msec, it is hallucination.
650          *    It cannot be less due to utterly erratic ACK generation made
651          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
652          *    to do with delayed acks, because at cwnd>2 true delack timeout
653          *    is invisible. Actually, Linux-2.4 also generates erratic
654          *    ACKs in some circumstances.
655          */
656         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
657
658         /* 2. Fixups made earlier cannot be right.
659          *    If we do not estimate RTO correctly without them,
660          *    all the algo is pure shit and should be replaced
661          *    with correct one. It is exactly, which we pretend to do.
662          */
663 }
664
665 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
666  * guarantees that rto is higher.
667  */
668 static inline void tcp_bound_rto(struct sock *sk)
669 {
670         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
671                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
672 }
673
674 /* Save metrics learned by this TCP session.
675    This function is called only, when TCP finishes successfully
676    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
677  */
678 void tcp_update_metrics(struct sock *sk)
679 {
680         struct tcp_sock *tp = tcp_sk(sk);
681         struct dst_entry *dst = __sk_dst_get(sk);
682
683         if (sysctl_tcp_nometrics_save)
684                 return;
685
686         dst_confirm(dst);
687
688         if (dst && (dst->flags&DST_HOST)) {
689                 const struct inet_connection_sock *icsk = inet_csk(sk);
690                 int m;
691
692                 if (icsk->icsk_backoff || !tp->srtt) {
693                         /* This session failed to estimate rtt. Why?
694                          * Probably, no packets returned in time.
695                          * Reset our results.
696                          */
697                         if (!(dst_metric_locked(dst, RTAX_RTT)))
698                                 dst->metrics[RTAX_RTT-1] = 0;
699                         return;
700                 }
701
702                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
703
704                 /* If newly calculated rtt larger than stored one,
705                  * store new one. Otherwise, use EWMA. Remember,
706                  * rtt overestimation is always better than underestimation.
707                  */
708                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
709                         if (m <= 0)
710                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
711                         else
712                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
713                 }
714
715                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
716                         if (m < 0)
717                                 m = -m;
718
719                         /* Scale deviation to rttvar fixed point */
720                         m >>= 1;
721                         if (m < tp->mdev)
722                                 m = tp->mdev;
723
724                         if (m >= dst_metric(dst, RTAX_RTTVAR))
725                                 dst->metrics[RTAX_RTTVAR-1] = m;
726                         else
727                                 dst->metrics[RTAX_RTTVAR-1] -=
728                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
729                 }
730
731                 if (tp->snd_ssthresh >= 0xFFFF) {
732                         /* Slow start still did not finish. */
733                         if (dst_metric(dst, RTAX_SSTHRESH) &&
734                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
735                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
736                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
737                         if (!dst_metric_locked(dst, RTAX_CWND) &&
738                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
739                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
740                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
741                            icsk->icsk_ca_state == TCP_CA_Open) {
742                         /* Cong. avoidance phase, cwnd is reliable. */
743                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
744                                 dst->metrics[RTAX_SSTHRESH-1] =
745                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
746                         if (!dst_metric_locked(dst, RTAX_CWND))
747                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
748                 } else {
749                         /* Else slow start did not finish, cwnd is non-sense,
750                            ssthresh may be also invalid.
751                          */
752                         if (!dst_metric_locked(dst, RTAX_CWND))
753                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
754                         if (dst->metrics[RTAX_SSTHRESH-1] &&
755                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
756                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
757                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
758                 }
759
760                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
761                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
762                             tp->reordering != sysctl_tcp_reordering)
763                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
764                 }
765         }
766 }
767
768 /* Numbers are taken from RFC3390.
769  *
770  * John Heffner states:
771  *
772  *      The RFC specifies a window of no more than 4380 bytes
773  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
774  *      is a bit misleading because they use a clamp at 4380 bytes
775  *      rather than use a multiplier in the relevant range.
776  */
777 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
778 {
779         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
780
781         if (!cwnd) {
782                 if (tp->mss_cache > 1460)
783                         cwnd = 2;
784                 else
785                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
786         }
787         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
788 }
789
790 /* Set slow start threshold and cwnd not falling to slow start */
791 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
792 {
793         struct tcp_sock *tp = tcp_sk(sk);
794         const struct inet_connection_sock *icsk = inet_csk(sk);
795
796         tp->prior_ssthresh = 0;
797         tp->bytes_acked = 0;
798         if (icsk->icsk_ca_state < TCP_CA_CWR) {
799                 tp->undo_marker = 0;
800                 if (set_ssthresh)
801                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
802                 tp->snd_cwnd = min(tp->snd_cwnd,
803                                    tcp_packets_in_flight(tp) + 1U);
804                 tp->snd_cwnd_cnt = 0;
805                 tp->high_seq = tp->snd_nxt;
806                 tp->snd_cwnd_stamp = tcp_time_stamp;
807                 TCP_ECN_queue_cwr(tp);
808
809                 tcp_set_ca_state(sk, TCP_CA_CWR);
810         }
811 }
812
813 /* Initialize metrics on socket. */
814
815 static void tcp_init_metrics(struct sock *sk)
816 {
817         struct tcp_sock *tp = tcp_sk(sk);
818         struct dst_entry *dst = __sk_dst_get(sk);
819
820         if (dst == NULL)
821                 goto reset;
822
823         dst_confirm(dst);
824
825         if (dst_metric_locked(dst, RTAX_CWND))
826                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
827         if (dst_metric(dst, RTAX_SSTHRESH)) {
828                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
829                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
830                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
831         }
832         if (dst_metric(dst, RTAX_REORDERING) &&
833             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
834                 tp->rx_opt.sack_ok &= ~2;
835                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
836         }
837
838         if (dst_metric(dst, RTAX_RTT) == 0)
839                 goto reset;
840
841         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
842                 goto reset;
843
844         /* Initial rtt is determined from SYN,SYN-ACK.
845          * The segment is small and rtt may appear much
846          * less than real one. Use per-dst memory
847          * to make it more realistic.
848          *
849          * A bit of theory. RTT is time passed after "normal" sized packet
850          * is sent until it is ACKed. In normal circumstances sending small
851          * packets force peer to delay ACKs and calculation is correct too.
852          * The algorithm is adaptive and, provided we follow specs, it
853          * NEVER underestimate RTT. BUT! If peer tries to make some clever
854          * tricks sort of "quick acks" for time long enough to decrease RTT
855          * to low value, and then abruptly stops to do it and starts to delay
856          * ACKs, wait for troubles.
857          */
858         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
859                 tp->srtt = dst_metric(dst, RTAX_RTT);
860                 tp->rtt_seq = tp->snd_nxt;
861         }
862         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
863                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
864                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
865         }
866         tcp_set_rto(sk);
867         tcp_bound_rto(sk);
868         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
869                 goto reset;
870         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
871         tp->snd_cwnd_stamp = tcp_time_stamp;
872         return;
873
874 reset:
875         /* Play conservative. If timestamps are not
876          * supported, TCP will fail to recalculate correct
877          * rtt, if initial rto is too small. FORGET ALL AND RESET!
878          */
879         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
880                 tp->srtt = 0;
881                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
882                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
883         }
884 }
885
886 static void tcp_update_reordering(struct sock *sk, const int metric,
887                                   const int ts)
888 {
889         struct tcp_sock *tp = tcp_sk(sk);
890         if (metric > tp->reordering) {
891                 tp->reordering = min(TCP_MAX_REORDERING, metric);
892
893                 /* This exciting event is worth to be remembered. 8) */
894                 if (ts)
895                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
896                 else if (IsReno(tp))
897                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
898                 else if (IsFack(tp))
899                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
900                 else
901                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
902 #if FASTRETRANS_DEBUG > 1
903                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
904                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
905                        tp->reordering,
906                        tp->fackets_out,
907                        tp->sacked_out,
908                        tp->undo_marker ? tp->undo_retrans : 0);
909 #endif
910                 /* Disable FACK yet. */
911                 tp->rx_opt.sack_ok &= ~2;
912         }
913 }
914
915 /* This procedure tags the retransmission queue when SACKs arrive.
916  *
917  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
918  * Packets in queue with these bits set are counted in variables
919  * sacked_out, retrans_out and lost_out, correspondingly.
920  *
921  * Valid combinations are:
922  * Tag  InFlight        Description
923  * 0    1               - orig segment is in flight.
924  * S    0               - nothing flies, orig reached receiver.
925  * L    0               - nothing flies, orig lost by net.
926  * R    2               - both orig and retransmit are in flight.
927  * L|R  1               - orig is lost, retransmit is in flight.
928  * S|R  1               - orig reached receiver, retrans is still in flight.
929  * (L|S|R is logically valid, it could occur when L|R is sacked,
930  *  but it is equivalent to plain S and code short-curcuits it to S.
931  *  L|S is logically invalid, it would mean -1 packet in flight 8))
932  *
933  * These 6 states form finite state machine, controlled by the following events:
934  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
935  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
936  * 3. Loss detection event of one of three flavors:
937  *      A. Scoreboard estimator decided the packet is lost.
938  *         A'. Reno "three dupacks" marks head of queue lost.
939  *         A''. Its FACK modfication, head until snd.fack is lost.
940  *      B. SACK arrives sacking data transmitted after never retransmitted
941  *         hole was sent out.
942  *      C. SACK arrives sacking SND.NXT at the moment, when the
943  *         segment was retransmitted.
944  * 4. D-SACK added new rule: D-SACK changes any tag to S.
945  *
946  * It is pleasant to note, that state diagram turns out to be commutative,
947  * so that we are allowed not to be bothered by order of our actions,
948  * when multiple events arrive simultaneously. (see the function below).
949  *
950  * Reordering detection.
951  * --------------------
952  * Reordering metric is maximal distance, which a packet can be displaced
953  * in packet stream. With SACKs we can estimate it:
954  *
955  * 1. SACK fills old hole and the corresponding segment was not
956  *    ever retransmitted -> reordering. Alas, we cannot use it
957  *    when segment was retransmitted.
958  * 2. The last flaw is solved with D-SACK. D-SACK arrives
959  *    for retransmitted and already SACKed segment -> reordering..
960  * Both of these heuristics are not used in Loss state, when we cannot
961  * account for retransmits accurately.
962  */
963 static int
964 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
965 {
966         const struct inet_connection_sock *icsk = inet_csk(sk);
967         struct tcp_sock *tp = tcp_sk(sk);
968         unsigned char *ptr = (skb_transport_header(ack_skb) +
969                               TCP_SKB_CB(ack_skb)->sacked);
970         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
971         struct sk_buff *cached_skb;
972         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
973         int reord = tp->packets_out;
974         int prior_fackets;
975         u32 lost_retrans = 0;
976         int flag = 0;
977         int found_dup_sack = 0;
978         int cached_fack_count;
979         int i;
980         int first_sack_index;
981
982         if (!tp->sacked_out)
983                 tp->fackets_out = 0;
984         prior_fackets = tp->fackets_out;
985
986         /* Check for D-SACK. */
987         if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
988                 flag |= FLAG_DSACKING_ACK;
989                 found_dup_sack = 1;
990                 tp->rx_opt.sack_ok |= 4;
991                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
992         } else if (num_sacks > 1 &&
993                         !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
994                         !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
995                 flag |= FLAG_DSACKING_ACK;
996                 found_dup_sack = 1;
997                 tp->rx_opt.sack_ok |= 4;
998                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
999         }
1000
1001         /* D-SACK for already forgotten data...
1002          * Do dumb counting. */
1003         if (found_dup_sack &&
1004                         !after(ntohl(sp[0].end_seq), prior_snd_una) &&
1005                         after(ntohl(sp[0].end_seq), tp->undo_marker))
1006                 tp->undo_retrans--;
1007
1008         /* Eliminate too old ACKs, but take into
1009          * account more or less fresh ones, they can
1010          * contain valid SACK info.
1011          */
1012         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1013                 return 0;
1014
1015         /* SACK fastpath:
1016          * if the only SACK change is the increase of the end_seq of
1017          * the first block then only apply that SACK block
1018          * and use retrans queue hinting otherwise slowpath */
1019         flag = 1;
1020         for (i = 0; i < num_sacks; i++) {
1021                 __be32 start_seq = sp[i].start_seq;
1022                 __be32 end_seq = sp[i].end_seq;
1023
1024                 if (i == 0) {
1025                         if (tp->recv_sack_cache[i].start_seq != start_seq)
1026                                 flag = 0;
1027                 } else {
1028                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1029                             (tp->recv_sack_cache[i].end_seq != end_seq))
1030                                 flag = 0;
1031                 }
1032                 tp->recv_sack_cache[i].start_seq = start_seq;
1033                 tp->recv_sack_cache[i].end_seq = end_seq;
1034         }
1035         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1036         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1037                 tp->recv_sack_cache[i].start_seq = 0;
1038                 tp->recv_sack_cache[i].end_seq = 0;
1039         }
1040
1041         first_sack_index = 0;
1042         if (flag)
1043                 num_sacks = 1;
1044         else {
1045                 int j;
1046                 tp->fastpath_skb_hint = NULL;
1047
1048                 /* order SACK blocks to allow in order walk of the retrans queue */
1049                 for (i = num_sacks-1; i > 0; i--) {
1050                         for (j = 0; j < i; j++){
1051                                 if (after(ntohl(sp[j].start_seq),
1052                                           ntohl(sp[j+1].start_seq))){
1053                                         struct tcp_sack_block_wire tmp;
1054
1055                                         tmp = sp[j];
1056                                         sp[j] = sp[j+1];
1057                                         sp[j+1] = tmp;
1058
1059                                         /* Track where the first SACK block goes to */
1060                                         if (j == first_sack_index)
1061                                                 first_sack_index = j+1;
1062                                 }
1063
1064                         }
1065                 }
1066         }
1067
1068         /* clear flag as used for different purpose in following code */
1069         flag = 0;
1070
1071         /* Use SACK fastpath hint if valid */
1072         cached_skb = tp->fastpath_skb_hint;
1073         cached_fack_count = tp->fastpath_cnt_hint;
1074         if (!cached_skb) {
1075                 cached_skb = tcp_write_queue_head(sk);
1076                 cached_fack_count = 0;
1077         }
1078
1079         for (i=0; i<num_sacks; i++, sp++) {
1080                 struct sk_buff *skb;
1081                 __u32 start_seq = ntohl(sp->start_seq);
1082                 __u32 end_seq = ntohl(sp->end_seq);
1083                 int fack_count;
1084                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1085
1086                 skb = cached_skb;
1087                 fack_count = cached_fack_count;
1088
1089                 /* Event "B" in the comment above. */
1090                 if (after(end_seq, tp->high_seq))
1091                         flag |= FLAG_DATA_LOST;
1092
1093                 tcp_for_write_queue_from(skb, sk) {
1094                         int in_sack, pcount;
1095                         u8 sacked;
1096
1097                         if (skb == tcp_send_head(sk))
1098                                 break;
1099
1100                         cached_skb = skb;
1101                         cached_fack_count = fack_count;
1102                         if (i == first_sack_index) {
1103                                 tp->fastpath_skb_hint = skb;
1104                                 tp->fastpath_cnt_hint = fack_count;
1105                         }
1106
1107                         /* The retransmission queue is always in order, so
1108                          * we can short-circuit the walk early.
1109                          */
1110                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1111                                 break;
1112
1113                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1114                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1115
1116                         pcount = tcp_skb_pcount(skb);
1117
1118                         if (pcount > 1 && !in_sack &&
1119                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1120                                 unsigned int pkt_len;
1121
1122                                 in_sack = !after(start_seq,
1123                                                  TCP_SKB_CB(skb)->seq);
1124
1125                                 if (!in_sack)
1126                                         pkt_len = (start_seq -
1127                                                    TCP_SKB_CB(skb)->seq);
1128                                 else
1129                                         pkt_len = (end_seq -
1130                                                    TCP_SKB_CB(skb)->seq);
1131                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1132                                         break;
1133                                 pcount = tcp_skb_pcount(skb);
1134                         }
1135
1136                         fack_count += pcount;
1137
1138                         sacked = TCP_SKB_CB(skb)->sacked;
1139
1140                         /* Account D-SACK for retransmitted packet. */
1141                         if ((dup_sack && in_sack) &&
1142                             (sacked & TCPCB_RETRANS) &&
1143                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1144                                 tp->undo_retrans--;
1145
1146                         /* The frame is ACKed. */
1147                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1148                                 if (sacked&TCPCB_RETRANS) {
1149                                         if ((dup_sack && in_sack) &&
1150                                             (sacked&TCPCB_SACKED_ACKED))
1151                                                 reord = min(fack_count, reord);
1152                                 } else {
1153                                         /* If it was in a hole, we detected reordering. */
1154                                         if (fack_count < prior_fackets &&
1155                                             !(sacked&TCPCB_SACKED_ACKED))
1156                                                 reord = min(fack_count, reord);
1157                                 }
1158
1159                                 /* Nothing to do; acked frame is about to be dropped. */
1160                                 continue;
1161                         }
1162
1163                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1164                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1165                             (!lost_retrans || after(end_seq, lost_retrans)))
1166                                 lost_retrans = end_seq;
1167
1168                         if (!in_sack)
1169                                 continue;
1170
1171                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1172                                 if (sacked & TCPCB_SACKED_RETRANS) {
1173                                         /* If the segment is not tagged as lost,
1174                                          * we do not clear RETRANS, believing
1175                                          * that retransmission is still in flight.
1176                                          */
1177                                         if (sacked & TCPCB_LOST) {
1178                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1179                                                 tp->lost_out -= tcp_skb_pcount(skb);
1180                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1181
1182                                                 /* clear lost hint */
1183                                                 tp->retransmit_skb_hint = NULL;
1184                                         }
1185                                 } else {
1186                                         /* New sack for not retransmitted frame,
1187                                          * which was in hole. It is reordering.
1188                                          */
1189                                         if (!(sacked & TCPCB_RETRANS) &&
1190                                             fack_count < prior_fackets)
1191                                                 reord = min(fack_count, reord);
1192
1193                                         if (sacked & TCPCB_LOST) {
1194                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1195                                                 tp->lost_out -= tcp_skb_pcount(skb);
1196
1197                                                 /* clear lost hint */
1198                                                 tp->retransmit_skb_hint = NULL;
1199                                         }
1200                                         /* SACK enhanced F-RTO detection.
1201                                          * Set flag if and only if non-rexmitted
1202                                          * segments below frto_highmark are
1203                                          * SACKed (RFC4138; Appendix B).
1204                                          * Clearing correct due to in-order walk
1205                                          */
1206                                         if (after(end_seq, tp->frto_highmark)) {
1207                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1208                                         } else {
1209                                                 if (!(sacked & TCPCB_RETRANS))
1210                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1211                                         }
1212                                 }
1213
1214                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1215                                 flag |= FLAG_DATA_SACKED;
1216                                 tp->sacked_out += tcp_skb_pcount(skb);
1217
1218                                 if (fack_count > tp->fackets_out)
1219                                         tp->fackets_out = fack_count;
1220                         } else {
1221                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1222                                         reord = min(fack_count, reord);
1223                         }
1224
1225                         /* D-SACK. We can detect redundant retransmission
1226                          * in S|R and plain R frames and clear it.
1227                          * undo_retrans is decreased above, L|R frames
1228                          * are accounted above as well.
1229                          */
1230                         if (dup_sack &&
1231                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1232                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1233                                 tp->retrans_out -= tcp_skb_pcount(skb);
1234                                 tp->retransmit_skb_hint = NULL;
1235                         }
1236                 }
1237         }
1238
1239         /* Check for lost retransmit. This superb idea is
1240          * borrowed from "ratehalving". Event "C".
1241          * Later note: FACK people cheated me again 8),
1242          * we have to account for reordering! Ugly,
1243          * but should help.
1244          */
1245         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1246                 struct sk_buff *skb;
1247
1248                 tcp_for_write_queue(skb, sk) {
1249                         if (skb == tcp_send_head(sk))
1250                                 break;
1251                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1252                                 break;
1253                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1254                                 continue;
1255                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1256                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1257                             (IsFack(tp) ||
1258                              !before(lost_retrans,
1259                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1260                                      tp->mss_cache))) {
1261                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1262                                 tp->retrans_out -= tcp_skb_pcount(skb);
1263
1264                                 /* clear lost hint */
1265                                 tp->retransmit_skb_hint = NULL;
1266
1267                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1268                                         tp->lost_out += tcp_skb_pcount(skb);
1269                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1270                                         flag |= FLAG_DATA_SACKED;
1271                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1272                                 }
1273                         }
1274                 }
1275         }
1276
1277         tp->left_out = tp->sacked_out + tp->lost_out;
1278
1279         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1280             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1281                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1282
1283 #if FASTRETRANS_DEBUG > 0
1284         BUG_TRAP((int)tp->sacked_out >= 0);
1285         BUG_TRAP((int)tp->lost_out >= 0);
1286         BUG_TRAP((int)tp->retrans_out >= 0);
1287         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1288 #endif
1289         return flag;
1290 }
1291
1292 /* F-RTO can only be used if TCP has never retransmitted anything other than
1293  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1294  */
1295 int tcp_use_frto(struct sock *sk)
1296 {
1297         const struct tcp_sock *tp = tcp_sk(sk);
1298         struct sk_buff *skb;
1299
1300         if (!sysctl_tcp_frto)
1301                 return 0;
1302
1303         if (IsSackFrto())
1304                 return 1;
1305
1306         /* Avoid expensive walking of rexmit queue if possible */
1307         if (tp->retrans_out > 1)
1308                 return 0;
1309
1310         skb = tcp_write_queue_head(sk);
1311         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1312         tcp_for_write_queue_from(skb, sk) {
1313                 if (skb == tcp_send_head(sk))
1314                         break;
1315                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1316                         return 0;
1317                 /* Short-circuit when first non-SACKed skb has been checked */
1318                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1319                         break;
1320         }
1321         return 1;
1322 }
1323
1324 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1325  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1326  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1327  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1328  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1329  * bits are handled if the Loss state is really to be entered (in
1330  * tcp_enter_frto_loss).
1331  *
1332  * Do like tcp_enter_loss() would; when RTO expires the second time it
1333  * does:
1334  *  "Reduce ssthresh if it has not yet been made inside this window."
1335  */
1336 void tcp_enter_frto(struct sock *sk)
1337 {
1338         const struct inet_connection_sock *icsk = inet_csk(sk);
1339         struct tcp_sock *tp = tcp_sk(sk);
1340         struct sk_buff *skb;
1341
1342         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1343             tp->snd_una == tp->high_seq ||
1344             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1345              !icsk->icsk_retransmits)) {
1346                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1347                 /* Our state is too optimistic in ssthresh() call because cwnd
1348                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1349                  * recovery has not yet completed. Pattern would be this: RTO,
1350                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1351                  * up here twice).
1352                  * RFC4138 should be more specific on what to do, even though
1353                  * RTO is quite unlikely to occur after the first Cumulative ACK
1354                  * due to back-off and complexity of triggering events ...
1355                  */
1356                 if (tp->frto_counter) {
1357                         u32 stored_cwnd;
1358                         stored_cwnd = tp->snd_cwnd;
1359                         tp->snd_cwnd = 2;
1360                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1361                         tp->snd_cwnd = stored_cwnd;
1362                 } else {
1363                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1364                 }
1365                 /* ... in theory, cong.control module could do "any tricks" in
1366                  * ssthresh(), which means that ca_state, lost bits and lost_out
1367                  * counter would have to be faked before the call occurs. We
1368                  * consider that too expensive, unlikely and hacky, so modules
1369                  * using these in ssthresh() must deal these incompatibility
1370                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1371                  */
1372                 tcp_ca_event(sk, CA_EVENT_FRTO);
1373         }
1374
1375         tp->undo_marker = tp->snd_una;
1376         tp->undo_retrans = 0;
1377
1378         skb = tcp_write_queue_head(sk);
1379         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1380                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1381                 tp->retrans_out -= tcp_skb_pcount(skb);
1382         }
1383         tcp_sync_left_out(tp);
1384
1385         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1386          * The last condition is necessary at least in tp->frto_counter case.
1387          */
1388         if (IsSackFrto() && (tp->frto_counter ||
1389             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1390             after(tp->high_seq, tp->snd_una)) {
1391                 tp->frto_highmark = tp->high_seq;
1392         } else {
1393                 tp->frto_highmark = tp->snd_nxt;
1394         }
1395         tcp_set_ca_state(sk, TCP_CA_Disorder);
1396         tp->high_seq = tp->snd_nxt;
1397         tp->frto_counter = 1;
1398 }
1399
1400 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1401  * which indicates that we should follow the traditional RTO recovery,
1402  * i.e. mark everything lost and do go-back-N retransmission.
1403  */
1404 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1405 {
1406         struct tcp_sock *tp = tcp_sk(sk);
1407         struct sk_buff *skb;
1408         int cnt = 0;
1409
1410         tp->sacked_out = 0;
1411         tp->lost_out = 0;
1412         tp->fackets_out = 0;
1413         tp->retrans_out = 0;
1414
1415         tcp_for_write_queue(skb, sk) {
1416                 if (skb == tcp_send_head(sk))
1417                         break;
1418                 cnt += tcp_skb_pcount(skb);
1419                 /*
1420                  * Count the retransmission made on RTO correctly (only when
1421                  * waiting for the first ACK and did not get it)...
1422                  */
1423                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1424                         /* For some reason this R-bit might get cleared? */
1425                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1426                                 tp->retrans_out += tcp_skb_pcount(skb);
1427                         /* ...enter this if branch just for the first segment */
1428                         flag |= FLAG_DATA_ACKED;
1429                 } else {
1430                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1431                 }
1432                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1433
1434                         /* Do not mark those segments lost that were
1435                          * forward transmitted after RTO
1436                          */
1437                         if (!after(TCP_SKB_CB(skb)->end_seq,
1438                                    tp->frto_highmark)) {
1439                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1440                                 tp->lost_out += tcp_skb_pcount(skb);
1441                         }
1442                 } else {
1443                         tp->sacked_out += tcp_skb_pcount(skb);
1444                         tp->fackets_out = cnt;
1445                 }
1446         }
1447         tcp_sync_left_out(tp);
1448
1449         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1450         tp->snd_cwnd_cnt = 0;
1451         tp->snd_cwnd_stamp = tcp_time_stamp;
1452         tp->undo_marker = 0;
1453         tp->frto_counter = 0;
1454
1455         tp->reordering = min_t(unsigned int, tp->reordering,
1456                                              sysctl_tcp_reordering);
1457         tcp_set_ca_state(sk, TCP_CA_Loss);
1458         tp->high_seq = tp->frto_highmark;
1459         TCP_ECN_queue_cwr(tp);
1460
1461         clear_all_retrans_hints(tp);
1462 }
1463
1464 void tcp_clear_retrans(struct tcp_sock *tp)
1465 {
1466         tp->left_out = 0;
1467         tp->retrans_out = 0;
1468
1469         tp->fackets_out = 0;
1470         tp->sacked_out = 0;
1471         tp->lost_out = 0;
1472
1473         tp->undo_marker = 0;
1474         tp->undo_retrans = 0;
1475 }
1476
1477 /* Enter Loss state. If "how" is not zero, forget all SACK information
1478  * and reset tags completely, otherwise preserve SACKs. If receiver
1479  * dropped its ofo queue, we will know this due to reneging detection.
1480  */
1481 void tcp_enter_loss(struct sock *sk, int how)
1482 {
1483         const struct inet_connection_sock *icsk = inet_csk(sk);
1484         struct tcp_sock *tp = tcp_sk(sk);
1485         struct sk_buff *skb;
1486         int cnt = 0;
1487
1488         /* Reduce ssthresh if it has not yet been made inside this window. */
1489         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1490             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1491                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1492                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1493                 tcp_ca_event(sk, CA_EVENT_LOSS);
1494         }
1495         tp->snd_cwnd       = 1;
1496         tp->snd_cwnd_cnt   = 0;
1497         tp->snd_cwnd_stamp = tcp_time_stamp;
1498
1499         tp->bytes_acked = 0;
1500         tcp_clear_retrans(tp);
1501
1502         /* Push undo marker, if it was plain RTO and nothing
1503          * was retransmitted. */
1504         if (!how)
1505                 tp->undo_marker = tp->snd_una;
1506
1507         tcp_for_write_queue(skb, sk) {
1508                 if (skb == tcp_send_head(sk))
1509                         break;
1510                 cnt += tcp_skb_pcount(skb);
1511                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1512                         tp->undo_marker = 0;
1513                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1514                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1515                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1516                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1517                         tp->lost_out += tcp_skb_pcount(skb);
1518                 } else {
1519                         tp->sacked_out += tcp_skb_pcount(skb);
1520                         tp->fackets_out = cnt;
1521                 }
1522         }
1523         tcp_sync_left_out(tp);
1524
1525         tp->reordering = min_t(unsigned int, tp->reordering,
1526                                              sysctl_tcp_reordering);
1527         tcp_set_ca_state(sk, TCP_CA_Loss);
1528         tp->high_seq = tp->snd_nxt;
1529         TCP_ECN_queue_cwr(tp);
1530         /* Abort FRTO algorithm if one is in progress */
1531         tp->frto_counter = 0;
1532
1533         clear_all_retrans_hints(tp);
1534 }
1535
1536 static int tcp_check_sack_reneging(struct sock *sk)
1537 {
1538         struct sk_buff *skb;
1539
1540         /* If ACK arrived pointing to a remembered SACK,
1541          * it means that our remembered SACKs do not reflect
1542          * real state of receiver i.e.
1543          * receiver _host_ is heavily congested (or buggy).
1544          * Do processing similar to RTO timeout.
1545          */
1546         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1547             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1548                 struct inet_connection_sock *icsk = inet_csk(sk);
1549                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1550
1551                 tcp_enter_loss(sk, 1);
1552                 icsk->icsk_retransmits++;
1553                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1554                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1555                                           icsk->icsk_rto, TCP_RTO_MAX);
1556                 return 1;
1557         }
1558         return 0;
1559 }
1560
1561 static inline int tcp_fackets_out(struct tcp_sock *tp)
1562 {
1563         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1564 }
1565
1566 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1567 {
1568         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1569 }
1570
1571 static inline int tcp_head_timedout(struct sock *sk)
1572 {
1573         struct tcp_sock *tp = tcp_sk(sk);
1574
1575         return tp->packets_out &&
1576                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1577 }
1578
1579 /* Linux NewReno/SACK/FACK/ECN state machine.
1580  * --------------------------------------
1581  *
1582  * "Open"       Normal state, no dubious events, fast path.
1583  * "Disorder"   In all the respects it is "Open",
1584  *              but requires a bit more attention. It is entered when
1585  *              we see some SACKs or dupacks. It is split of "Open"
1586  *              mainly to move some processing from fast path to slow one.
1587  * "CWR"        CWND was reduced due to some Congestion Notification event.
1588  *              It can be ECN, ICMP source quench, local device congestion.
1589  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1590  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1591  *
1592  * tcp_fastretrans_alert() is entered:
1593  * - each incoming ACK, if state is not "Open"
1594  * - when arrived ACK is unusual, namely:
1595  *      * SACK
1596  *      * Duplicate ACK.
1597  *      * ECN ECE.
1598  *
1599  * Counting packets in flight is pretty simple.
1600  *
1601  *      in_flight = packets_out - left_out + retrans_out
1602  *
1603  *      packets_out is SND.NXT-SND.UNA counted in packets.
1604  *
1605  *      retrans_out is number of retransmitted segments.
1606  *
1607  *      left_out is number of segments left network, but not ACKed yet.
1608  *
1609  *              left_out = sacked_out + lost_out
1610  *
1611  *     sacked_out: Packets, which arrived to receiver out of order
1612  *                 and hence not ACKed. With SACKs this number is simply
1613  *                 amount of SACKed data. Even without SACKs
1614  *                 it is easy to give pretty reliable estimate of this number,
1615  *                 counting duplicate ACKs.
1616  *
1617  *       lost_out: Packets lost by network. TCP has no explicit
1618  *                 "loss notification" feedback from network (for now).
1619  *                 It means that this number can be only _guessed_.
1620  *                 Actually, it is the heuristics to predict lossage that
1621  *                 distinguishes different algorithms.
1622  *
1623  *      F.e. after RTO, when all the queue is considered as lost,
1624  *      lost_out = packets_out and in_flight = retrans_out.
1625  *
1626  *              Essentially, we have now two algorithms counting
1627  *              lost packets.
1628  *
1629  *              FACK: It is the simplest heuristics. As soon as we decided
1630  *              that something is lost, we decide that _all_ not SACKed
1631  *              packets until the most forward SACK are lost. I.e.
1632  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1633  *              It is absolutely correct estimate, if network does not reorder
1634  *              packets. And it loses any connection to reality when reordering
1635  *              takes place. We use FACK by default until reordering
1636  *              is suspected on the path to this destination.
1637  *
1638  *              NewReno: when Recovery is entered, we assume that one segment
1639  *              is lost (classic Reno). While we are in Recovery and
1640  *              a partial ACK arrives, we assume that one more packet
1641  *              is lost (NewReno). This heuristics are the same in NewReno
1642  *              and SACK.
1643  *
1644  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1645  *  deflation etc. CWND is real congestion window, never inflated, changes
1646  *  only according to classic VJ rules.
1647  *
1648  * Really tricky (and requiring careful tuning) part of algorithm
1649  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1650  * The first determines the moment _when_ we should reduce CWND and,
1651  * hence, slow down forward transmission. In fact, it determines the moment
1652  * when we decide that hole is caused by loss, rather than by a reorder.
1653  *
1654  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1655  * holes, caused by lost packets.
1656  *
1657  * And the most logically complicated part of algorithm is undo
1658  * heuristics. We detect false retransmits due to both too early
1659  * fast retransmit (reordering) and underestimated RTO, analyzing
1660  * timestamps and D-SACKs. When we detect that some segments were
1661  * retransmitted by mistake and CWND reduction was wrong, we undo
1662  * window reduction and abort recovery phase. This logic is hidden
1663  * inside several functions named tcp_try_undo_<something>.
1664  */
1665
1666 /* This function decides, when we should leave Disordered state
1667  * and enter Recovery phase, reducing congestion window.
1668  *
1669  * Main question: may we further continue forward transmission
1670  * with the same cwnd?
1671  */
1672 static int tcp_time_to_recover(struct sock *sk)
1673 {
1674         struct tcp_sock *tp = tcp_sk(sk);
1675         __u32 packets_out;
1676
1677         /* Do not perform any recovery during FRTO algorithm */
1678         if (tp->frto_counter)
1679                 return 0;
1680
1681         /* Trick#1: The loss is proven. */
1682         if (tp->lost_out)
1683                 return 1;
1684
1685         /* Not-A-Trick#2 : Classic rule... */
1686         if (tcp_fackets_out(tp) > tp->reordering)
1687                 return 1;
1688
1689         /* Trick#3 : when we use RFC2988 timer restart, fast
1690          * retransmit can be triggered by timeout of queue head.
1691          */
1692         if (tcp_head_timedout(sk))
1693                 return 1;
1694
1695         /* Trick#4: It is still not OK... But will it be useful to delay
1696          * recovery more?
1697          */
1698         packets_out = tp->packets_out;
1699         if (packets_out <= tp->reordering &&
1700             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1701             !tcp_may_send_now(sk)) {
1702                 /* We have nothing to send. This connection is limited
1703                  * either by receiver window or by application.
1704                  */
1705                 return 1;
1706         }
1707
1708         return 0;
1709 }
1710
1711 /* If we receive more dupacks than we expected counting segments
1712  * in assumption of absent reordering, interpret this as reordering.
1713  * The only another reason could be bug in receiver TCP.
1714  */
1715 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1716 {
1717         struct tcp_sock *tp = tcp_sk(sk);
1718         u32 holes;
1719
1720         holes = max(tp->lost_out, 1U);
1721         holes = min(holes, tp->packets_out);
1722
1723         if ((tp->sacked_out + holes) > tp->packets_out) {
1724                 tp->sacked_out = tp->packets_out - holes;
1725                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1726         }
1727 }
1728
1729 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1730
1731 static void tcp_add_reno_sack(struct sock *sk)
1732 {
1733         struct tcp_sock *tp = tcp_sk(sk);
1734         tp->sacked_out++;
1735         tcp_check_reno_reordering(sk, 0);
1736         tcp_sync_left_out(tp);
1737 }
1738
1739 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1740
1741 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1742 {
1743         struct tcp_sock *tp = tcp_sk(sk);
1744
1745         if (acked > 0) {
1746                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1747                 if (acked-1 >= tp->sacked_out)
1748                         tp->sacked_out = 0;
1749                 else
1750                         tp->sacked_out -= acked-1;
1751         }
1752         tcp_check_reno_reordering(sk, acked);
1753         tcp_sync_left_out(tp);
1754 }
1755
1756 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1757 {
1758         tp->sacked_out = 0;
1759         tp->left_out = tp->lost_out;
1760 }
1761
1762 /* Mark head of queue up as lost. */
1763 static void tcp_mark_head_lost(struct sock *sk,
1764                                int packets, u32 high_seq)
1765 {
1766         struct tcp_sock *tp = tcp_sk(sk);
1767         struct sk_buff *skb;
1768         int cnt;
1769
1770         BUG_TRAP(packets <= tp->packets_out);
1771         if (tp->lost_skb_hint) {
1772                 skb = tp->lost_skb_hint;
1773                 cnt = tp->lost_cnt_hint;
1774         } else {
1775                 skb = tcp_write_queue_head(sk);
1776                 cnt = 0;
1777         }
1778
1779         tcp_for_write_queue_from(skb, sk) {
1780                 if (skb == tcp_send_head(sk))
1781                         break;
1782                 /* TODO: do this better */
1783                 /* this is not the most efficient way to do this... */
1784                 tp->lost_skb_hint = skb;
1785                 tp->lost_cnt_hint = cnt;
1786                 cnt += tcp_skb_pcount(skb);
1787                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1788                         break;
1789                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1790                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1791                         tp->lost_out += tcp_skb_pcount(skb);
1792
1793                         /* clear xmit_retransmit_queue hints
1794                          *  if this is beyond hint */
1795                         if (tp->retransmit_skb_hint != NULL &&
1796                             before(TCP_SKB_CB(skb)->seq,
1797                                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1798                                 tp->retransmit_skb_hint = NULL;
1799
1800                 }
1801         }
1802         tcp_sync_left_out(tp);
1803 }
1804
1805 /* Account newly detected lost packet(s) */
1806
1807 static void tcp_update_scoreboard(struct sock *sk)
1808 {
1809         struct tcp_sock *tp = tcp_sk(sk);
1810
1811         if (IsFack(tp)) {
1812                 int lost = tp->fackets_out - tp->reordering;
1813                 if (lost <= 0)
1814                         lost = 1;
1815                 tcp_mark_head_lost(sk, lost, tp->high_seq);
1816         } else {
1817                 tcp_mark_head_lost(sk, 1, tp->high_seq);
1818         }
1819
1820         /* New heuristics: it is possible only after we switched
1821          * to restart timer each time when something is ACKed.
1822          * Hence, we can detect timed out packets during fast
1823          * retransmit without falling to slow start.
1824          */
1825         if (!IsReno(tp) && tcp_head_timedout(sk)) {
1826                 struct sk_buff *skb;
1827
1828                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1829                         : tcp_write_queue_head(sk);
1830
1831                 tcp_for_write_queue_from(skb, sk) {
1832                         if (skb == tcp_send_head(sk))
1833                                 break;
1834                         if (!tcp_skb_timedout(sk, skb))
1835                                 break;
1836
1837                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1838                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1839                                 tp->lost_out += tcp_skb_pcount(skb);
1840
1841                                 /* clear xmit_retrans hint */
1842                                 if (tp->retransmit_skb_hint &&
1843                                     before(TCP_SKB_CB(skb)->seq,
1844                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1845
1846                                         tp->retransmit_skb_hint = NULL;
1847                         }
1848                 }
1849
1850                 tp->scoreboard_skb_hint = skb;
1851
1852                 tcp_sync_left_out(tp);
1853         }
1854 }
1855
1856 /* CWND moderation, preventing bursts due to too big ACKs
1857  * in dubious situations.
1858  */
1859 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1860 {
1861         tp->snd_cwnd = min(tp->snd_cwnd,
1862                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1863         tp->snd_cwnd_stamp = tcp_time_stamp;
1864 }
1865
1866 /* Lower bound on congestion window is slow start threshold
1867  * unless congestion avoidance choice decides to overide it.
1868  */
1869 static inline u32 tcp_cwnd_min(const struct sock *sk)
1870 {
1871         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1872
1873         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1874 }
1875
1876 /* Decrease cwnd each second ack. */
1877 static void tcp_cwnd_down(struct sock *sk, int flag)
1878 {
1879         struct tcp_sock *tp = tcp_sk(sk);
1880         int decr = tp->snd_cwnd_cnt + 1;
1881
1882         if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
1883             (IsReno(tp) && !(flag&FLAG_NOT_DUP))) {
1884                 tp->snd_cwnd_cnt = decr&1;
1885                 decr >>= 1;
1886
1887                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1888                         tp->snd_cwnd -= decr;
1889
1890                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1891                 tp->snd_cwnd_stamp = tcp_time_stamp;
1892         }
1893 }
1894
1895 /* Nothing was retransmitted or returned timestamp is less
1896  * than timestamp of the first retransmission.
1897  */
1898 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1899 {
1900         return !tp->retrans_stamp ||
1901                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1902                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1903 }
1904
1905 /* Undo procedures. */
1906
1907 #if FASTRETRANS_DEBUG > 1
1908 static void DBGUNDO(struct sock *sk, const char *msg)
1909 {
1910         struct tcp_sock *tp = tcp_sk(sk);
1911         struct inet_sock *inet = inet_sk(sk);
1912
1913         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1914                msg,
1915                NIPQUAD(inet->daddr), ntohs(inet->dport),
1916                tp->snd_cwnd, tp->left_out,
1917                tp->snd_ssthresh, tp->prior_ssthresh,
1918                tp->packets_out);
1919 }
1920 #else
1921 #define DBGUNDO(x...) do { } while (0)
1922 #endif
1923
1924 static void tcp_undo_cwr(struct sock *sk, const int undo)
1925 {
1926         struct tcp_sock *tp = tcp_sk(sk);
1927
1928         if (tp->prior_ssthresh) {
1929                 const struct inet_connection_sock *icsk = inet_csk(sk);
1930
1931                 if (icsk->icsk_ca_ops->undo_cwnd)
1932                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1933                 else
1934                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1935
1936                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1937                         tp->snd_ssthresh = tp->prior_ssthresh;
1938                         TCP_ECN_withdraw_cwr(tp);
1939                 }
1940         } else {
1941                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1942         }
1943         tcp_moderate_cwnd(tp);
1944         tp->snd_cwnd_stamp = tcp_time_stamp;
1945
1946         /* There is something screwy going on with the retrans hints after
1947            an undo */
1948         clear_all_retrans_hints(tp);
1949 }
1950
1951 static inline int tcp_may_undo(struct tcp_sock *tp)
1952 {
1953         return tp->undo_marker &&
1954                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1955 }
1956
1957 /* People celebrate: "We love our President!" */
1958 static int tcp_try_undo_recovery(struct sock *sk)
1959 {
1960         struct tcp_sock *tp = tcp_sk(sk);
1961
1962         if (tcp_may_undo(tp)) {
1963                 /* Happy end! We did not retransmit anything
1964                  * or our original transmission succeeded.
1965                  */
1966                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1967                 tcp_undo_cwr(sk, 1);
1968                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1969                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1970                 else
1971                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1972                 tp->undo_marker = 0;
1973         }
1974         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1975                 /* Hold old state until something *above* high_seq
1976                  * is ACKed. For Reno it is MUST to prevent false
1977                  * fast retransmits (RFC2582). SACK TCP is safe. */
1978                 tcp_moderate_cwnd(tp);
1979                 return 1;
1980         }
1981         tcp_set_ca_state(sk, TCP_CA_Open);
1982         return 0;
1983 }
1984
1985 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1986 static void tcp_try_undo_dsack(struct sock *sk)
1987 {
1988         struct tcp_sock *tp = tcp_sk(sk);
1989
1990         if (tp->undo_marker && !tp->undo_retrans) {
1991                 DBGUNDO(sk, "D-SACK");
1992                 tcp_undo_cwr(sk, 1);
1993                 tp->undo_marker = 0;
1994                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1995         }
1996 }
1997
1998 /* Undo during fast recovery after partial ACK. */
1999
2000 static int tcp_try_undo_partial(struct sock *sk, int acked)
2001 {
2002         struct tcp_sock *tp = tcp_sk(sk);
2003         /* Partial ACK arrived. Force Hoe's retransmit. */
2004         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
2005
2006         if (tcp_may_undo(tp)) {
2007                 /* Plain luck! Hole if filled with delayed
2008                  * packet, rather than with a retransmit.
2009                  */
2010                 if (tp->retrans_out == 0)
2011                         tp->retrans_stamp = 0;
2012
2013                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2014
2015                 DBGUNDO(sk, "Hoe");
2016                 tcp_undo_cwr(sk, 0);
2017                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2018
2019                 /* So... Do not make Hoe's retransmit yet.
2020                  * If the first packet was delayed, the rest
2021                  * ones are most probably delayed as well.
2022                  */
2023                 failed = 0;
2024         }
2025         return failed;
2026 }
2027
2028 /* Undo during loss recovery after partial ACK. */
2029 static int tcp_try_undo_loss(struct sock *sk)
2030 {
2031         struct tcp_sock *tp = tcp_sk(sk);
2032
2033         if (tcp_may_undo(tp)) {
2034                 struct sk_buff *skb;
2035                 tcp_for_write_queue(skb, sk) {
2036                         if (skb == tcp_send_head(sk))
2037                                 break;
2038                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2039                 }
2040
2041                 clear_all_retrans_hints(tp);
2042
2043                 DBGUNDO(sk, "partial loss");
2044                 tp->lost_out = 0;
2045                 tp->left_out = tp->sacked_out;
2046                 tcp_undo_cwr(sk, 1);
2047                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2048                 inet_csk(sk)->icsk_retransmits = 0;
2049                 tp->undo_marker = 0;
2050                 if (!IsReno(tp))
2051                         tcp_set_ca_state(sk, TCP_CA_Open);
2052                 return 1;
2053         }
2054         return 0;
2055 }
2056
2057 static inline void tcp_complete_cwr(struct sock *sk)
2058 {
2059         struct tcp_sock *tp = tcp_sk(sk);
2060         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2061         tp->snd_cwnd_stamp = tcp_time_stamp;
2062         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2063 }
2064
2065 static void tcp_try_to_open(struct sock *sk, int flag)
2066 {
2067         struct tcp_sock *tp = tcp_sk(sk);
2068
2069         tcp_sync_left_out(tp);
2070
2071         if (tp->retrans_out == 0)
2072                 tp->retrans_stamp = 0;
2073
2074         if (flag&FLAG_ECE)
2075                 tcp_enter_cwr(sk, 1);
2076
2077         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2078                 int state = TCP_CA_Open;
2079
2080                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2081                         state = TCP_CA_Disorder;
2082
2083                 if (inet_csk(sk)->icsk_ca_state != state) {
2084                         tcp_set_ca_state(sk, state);
2085                         tp->high_seq = tp->snd_nxt;
2086                 }
2087                 tcp_moderate_cwnd(tp);
2088         } else {
2089                 tcp_cwnd_down(sk, flag);
2090         }
2091 }
2092
2093 static void tcp_mtup_probe_failed(struct sock *sk)
2094 {
2095         struct inet_connection_sock *icsk = inet_csk(sk);
2096
2097         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2098         icsk->icsk_mtup.probe_size = 0;
2099 }
2100
2101 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2102 {
2103         struct tcp_sock *tp = tcp_sk(sk);
2104         struct inet_connection_sock *icsk = inet_csk(sk);
2105
2106         /* FIXME: breaks with very large cwnd */
2107         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2108         tp->snd_cwnd = tp->snd_cwnd *
2109                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2110                        icsk->icsk_mtup.probe_size;
2111         tp->snd_cwnd_cnt = 0;
2112         tp->snd_cwnd_stamp = tcp_time_stamp;
2113         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2114
2115         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2116         icsk->icsk_mtup.probe_size = 0;
2117         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2118 }
2119
2120
2121 /* Process an event, which can update packets-in-flight not trivially.
2122  * Main goal of this function is to calculate new estimate for left_out,
2123  * taking into account both packets sitting in receiver's buffer and
2124  * packets lost by network.
2125  *
2126  * Besides that it does CWND reduction, when packet loss is detected
2127  * and changes state of machine.
2128  *
2129  * It does _not_ decide what to send, it is made in function
2130  * tcp_xmit_retransmit_queue().
2131  */
2132 static void
2133 tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag)
2134 {
2135         struct inet_connection_sock *icsk = inet_csk(sk);
2136         struct tcp_sock *tp = tcp_sk(sk);
2137         int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2138         int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2139                                     (tp->fackets_out > tp->reordering));
2140
2141         /* Some technical things:
2142          * 1. Reno does not count dupacks (sacked_out) automatically. */
2143         if (!tp->packets_out)
2144                 tp->sacked_out = 0;
2145         /* 2. SACK counts snd_fack in packets inaccurately. */
2146         if (tp->sacked_out == 0)
2147                 tp->fackets_out = 0;
2148
2149         /* Now state machine starts.
2150          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2151         if (flag&FLAG_ECE)
2152                 tp->prior_ssthresh = 0;
2153
2154         /* B. In all the states check for reneging SACKs. */
2155         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2156                 return;
2157
2158         /* C. Process data loss notification, provided it is valid. */
2159         if ((flag&FLAG_DATA_LOST) &&
2160             before(tp->snd_una, tp->high_seq) &&
2161             icsk->icsk_ca_state != TCP_CA_Open &&
2162             tp->fackets_out > tp->reordering) {
2163                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2164                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2165         }
2166
2167         /* D. Synchronize left_out to current state. */
2168         tcp_sync_left_out(tp);
2169
2170         /* E. Check state exit conditions. State can be terminated
2171          *    when high_seq is ACKed. */
2172         if (icsk->icsk_ca_state == TCP_CA_Open) {
2173                 BUG_TRAP(tp->retrans_out == 0);
2174                 tp->retrans_stamp = 0;
2175         } else if (!before(tp->snd_una, tp->high_seq)) {
2176                 switch (icsk->icsk_ca_state) {
2177                 case TCP_CA_Loss:
2178                         icsk->icsk_retransmits = 0;
2179                         if (tcp_try_undo_recovery(sk))
2180                                 return;
2181                         break;
2182
2183                 case TCP_CA_CWR:
2184                         /* CWR is to be held something *above* high_seq
2185                          * is ACKed for CWR bit to reach receiver. */
2186                         if (tp->snd_una != tp->high_seq) {
2187                                 tcp_complete_cwr(sk);
2188                                 tcp_set_ca_state(sk, TCP_CA_Open);
2189                         }
2190                         break;
2191
2192                 case TCP_CA_Disorder:
2193                         tcp_try_undo_dsack(sk);
2194                         if (!tp->undo_marker ||
2195                             /* For SACK case do not Open to allow to undo
2196                              * catching for all duplicate ACKs. */
2197                             IsReno(tp) || tp->snd_una != tp->high_seq) {
2198                                 tp->undo_marker = 0;
2199                                 tcp_set_ca_state(sk, TCP_CA_Open);
2200                         }
2201                         break;
2202
2203                 case TCP_CA_Recovery:
2204                         if (IsReno(tp))
2205                                 tcp_reset_reno_sack(tp);
2206                         if (tcp_try_undo_recovery(sk))
2207                                 return;
2208                         tcp_complete_cwr(sk);
2209                         break;
2210                 }
2211         }
2212
2213         /* F. Process state. */
2214         switch (icsk->icsk_ca_state) {
2215         case TCP_CA_Recovery:
2216                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2217                         if (IsReno(tp) && is_dupack)
2218                                 tcp_add_reno_sack(sk);
2219                 } else {
2220                         int acked = prior_packets - tp->packets_out;
2221                         if (IsReno(tp))
2222                                 tcp_remove_reno_sacks(sk, acked);
2223                         do_lost = tcp_try_undo_partial(sk, acked);
2224                 }
2225                 break;
2226         case TCP_CA_Loss:
2227                 if (flag&FLAG_DATA_ACKED)
2228                         icsk->icsk_retransmits = 0;
2229                 if (!tcp_try_undo_loss(sk)) {
2230                         tcp_moderate_cwnd(tp);
2231                         tcp_xmit_retransmit_queue(sk);
2232                         return;
2233                 }
2234                 if (icsk->icsk_ca_state != TCP_CA_Open)
2235                         return;
2236                 /* Loss is undone; fall through to processing in Open state. */
2237         default:
2238                 if (IsReno(tp)) {
2239                         if (flag & FLAG_SND_UNA_ADVANCED)
2240                                 tcp_reset_reno_sack(tp);
2241                         if (is_dupack)
2242                                 tcp_add_reno_sack(sk);
2243                 }
2244
2245                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2246                         tcp_try_undo_dsack(sk);
2247
2248                 if (!tcp_time_to_recover(sk)) {
2249                         tcp_try_to_open(sk, flag);
2250                         return;
2251                 }
2252
2253                 /* MTU probe failure: don't reduce cwnd */
2254                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2255                     icsk->icsk_mtup.probe_size &&
2256                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2257                         tcp_mtup_probe_failed(sk);
2258                         /* Restores the reduction we did in tcp_mtup_probe() */
2259                         tp->snd_cwnd++;
2260                         tcp_simple_retransmit(sk);
2261                         return;
2262                 }
2263
2264                 /* Otherwise enter Recovery state */
2265
2266                 if (IsReno(tp))
2267                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2268                 else
2269                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2270
2271                 tp->high_seq = tp->snd_nxt;
2272                 tp->prior_ssthresh = 0;
2273                 tp->undo_marker = tp->snd_una;
2274                 tp->undo_retrans = tp->retrans_out;
2275
2276                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2277                         if (!(flag&FLAG_ECE))
2278                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2279                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2280                         TCP_ECN_queue_cwr(tp);
2281                 }
2282
2283                 tp->bytes_acked = 0;
2284                 tp->snd_cwnd_cnt = 0;
2285                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2286         }
2287
2288         if (do_lost || tcp_head_timedout(sk))
2289                 tcp_update_scoreboard(sk);
2290         tcp_cwnd_down(sk, flag);
2291         tcp_xmit_retransmit_queue(sk);
2292 }
2293
2294 /* Read draft-ietf-tcplw-high-performance before mucking
2295  * with this code. (Supersedes RFC1323)
2296  */
2297 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2298 {
2299         /* RTTM Rule: A TSecr value received in a segment is used to
2300          * update the averaged RTT measurement only if the segment
2301          * acknowledges some new data, i.e., only if it advances the
2302          * left edge of the send window.
2303          *
2304          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2305          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2306          *
2307          * Changed: reset backoff as soon as we see the first valid sample.
2308          * If we do not, we get strongly overestimated rto. With timestamps
2309          * samples are accepted even from very old segments: f.e., when rtt=1
2310          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2311          * answer arrives rto becomes 120 seconds! If at least one of segments
2312          * in window is lost... Voila.                          --ANK (010210)
2313          */
2314         struct tcp_sock *tp = tcp_sk(sk);
2315         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2316         tcp_rtt_estimator(sk, seq_rtt);
2317         tcp_set_rto(sk);
2318         inet_csk(sk)->icsk_backoff = 0;
2319         tcp_bound_rto(sk);
2320 }
2321
2322 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2323 {
2324         /* We don't have a timestamp. Can only use
2325          * packets that are not retransmitted to determine
2326          * rtt estimates. Also, we must not reset the
2327          * backoff for rto until we get a non-retransmitted
2328          * packet. This allows us to deal with a situation
2329          * where the network delay has increased suddenly.
2330          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2331          */
2332
2333         if (flag & FLAG_RETRANS_DATA_ACKED)
2334                 return;
2335
2336         tcp_rtt_estimator(sk, seq_rtt);
2337         tcp_set_rto(sk);
2338         inet_csk(sk)->icsk_backoff = 0;
2339         tcp_bound_rto(sk);
2340 }
2341
2342 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2343                                       const s32 seq_rtt)
2344 {
2345         const struct tcp_sock *tp = tcp_sk(sk);
2346         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2347         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2348                 tcp_ack_saw_tstamp(sk, flag);
2349         else if (seq_rtt >= 0)
2350                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2351 }
2352
2353 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2354                            u32 in_flight, int good)
2355 {
2356         const struct inet_connection_sock *icsk = inet_csk(sk);
2357         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2358         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2359 }
2360
2361 /* Restart timer after forward progress on connection.
2362  * RFC2988 recommends to restart timer to now+rto.
2363  */
2364
2365 static void tcp_ack_packets_out(struct sock *sk)
2366 {
2367         struct tcp_sock *tp = tcp_sk(sk);
2368
2369         if (!tp->packets_out) {
2370                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2371         } else {
2372                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2373         }
2374 }
2375
2376 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2377                          __u32 now, __s32 *seq_rtt)
2378 {
2379         struct tcp_sock *tp = tcp_sk(sk);
2380         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2381         __u32 seq = tp->snd_una;
2382         __u32 packets_acked;
2383         int acked = 0;
2384
2385         /* If we get here, the whole TSO packet has not been
2386          * acked.
2387          */
2388         BUG_ON(!after(scb->end_seq, seq));
2389
2390         packets_acked = tcp_skb_pcount(skb);
2391         if (tcp_trim_head(sk, skb, seq - scb->seq))
2392                 return 0;
2393         packets_acked -= tcp_skb_pcount(skb);
2394
2395         if (packets_acked) {
2396                 __u8 sacked = scb->sacked;
2397
2398                 acked |= FLAG_DATA_ACKED;
2399                 if (sacked) {
2400                         if (sacked & TCPCB_RETRANS) {
2401                                 if (sacked & TCPCB_SACKED_RETRANS)
2402                                         tp->retrans_out -= packets_acked;
2403                                 acked |= FLAG_RETRANS_DATA_ACKED;
2404                                 *seq_rtt = -1;
2405                         } else if (*seq_rtt < 0)
2406                                 *seq_rtt = now - scb->when;
2407                         if (sacked & TCPCB_SACKED_ACKED)
2408                                 tp->sacked_out -= packets_acked;
2409                         if (sacked & TCPCB_LOST)
2410                                 tp->lost_out -= packets_acked;
2411                         if (sacked & TCPCB_URG) {
2412                                 if (tp->urg_mode &&
2413                                     !before(seq, tp->snd_up))
2414                                         tp->urg_mode = 0;
2415                         }
2416                 } else if (*seq_rtt < 0)
2417                         *seq_rtt = now - scb->when;
2418
2419                 if (tp->fackets_out) {
2420                         __u32 dval = min(tp->fackets_out, packets_acked);
2421                         tp->fackets_out -= dval;
2422                 }
2423                 tp->packets_out -= packets_acked;
2424
2425                 BUG_ON(tcp_skb_pcount(skb) == 0);
2426                 BUG_ON(!before(scb->seq, scb->end_seq));
2427         }
2428
2429         return acked;
2430 }
2431
2432 /* Remove acknowledged frames from the retransmission queue. */
2433 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2434 {
2435         struct tcp_sock *tp = tcp_sk(sk);
2436         const struct inet_connection_sock *icsk = inet_csk(sk);
2437         struct sk_buff *skb;
2438         __u32 now = tcp_time_stamp;
2439         int acked = 0;
2440         int prior_packets = tp->packets_out;
2441         __s32 seq_rtt = -1;
2442         ktime_t last_ackt = net_invalid_timestamp();
2443
2444         while ((skb = tcp_write_queue_head(sk)) &&
2445                skb != tcp_send_head(sk)) {
2446                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2447                 __u8 sacked = scb->sacked;
2448
2449                 /* If our packet is before the ack sequence we can
2450                  * discard it as it's confirmed to have arrived at
2451                  * the other end.
2452                  */
2453                 if (after(scb->end_seq, tp->snd_una)) {
2454                         if (tcp_skb_pcount(skb) > 1 &&
2455                             after(tp->snd_una, scb->seq))
2456                                 acked |= tcp_tso_acked(sk, skb,
2457                                                        now, &seq_rtt);
2458                         break;
2459                 }
2460
2461                 /* Initial outgoing SYN's get put onto the write_queue
2462                  * just like anything else we transmit.  It is not
2463                  * true data, and if we misinform our callers that
2464                  * this ACK acks real data, we will erroneously exit
2465                  * connection startup slow start one packet too
2466                  * quickly.  This is severely frowned upon behavior.
2467                  */
2468                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2469                         acked |= FLAG_DATA_ACKED;
2470                 } else {
2471                         acked |= FLAG_SYN_ACKED;
2472                         tp->retrans_stamp = 0;
2473                 }
2474
2475                 /* MTU probing checks */
2476                 if (icsk->icsk_mtup.probe_size) {
2477                         if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2478                                 tcp_mtup_probe_success(sk, skb);
2479                         }
2480                 }
2481
2482                 if (sacked) {
2483                         if (sacked & TCPCB_RETRANS) {
2484                                 if (sacked & TCPCB_SACKED_RETRANS)
2485                                         tp->retrans_out -= tcp_skb_pcount(skb);
2486                                 acked |= FLAG_RETRANS_DATA_ACKED;
2487                                 seq_rtt = -1;
2488                         } else if (seq_rtt < 0) {
2489                                 seq_rtt = now - scb->when;
2490                                 last_ackt = skb->tstamp;
2491                         }
2492                         if (sacked & TCPCB_SACKED_ACKED)
2493                                 tp->sacked_out -= tcp_skb_pcount(skb);
2494                         if (sacked & TCPCB_LOST)
2495                                 tp->lost_out -= tcp_skb_pcount(skb);
2496                         if (sacked & TCPCB_URG) {
2497                                 if (tp->urg_mode &&
2498                                     !before(scb->end_seq, tp->snd_up))
2499                                         tp->urg_mode = 0;
2500                         }
2501                 } else if (seq_rtt < 0) {
2502                         seq_rtt = now - scb->when;
2503                         last_ackt = skb->tstamp;
2504                 }
2505                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2506                 tcp_packets_out_dec(tp, skb);
2507                 tcp_unlink_write_queue(skb, sk);
2508                 sk_stream_free_skb(sk, skb);
2509                 clear_all_retrans_hints(tp);
2510         }
2511
2512         if (acked&FLAG_ACKED) {
2513                 u32 pkts_acked = prior_packets - tp->packets_out;
2514                 const struct tcp_congestion_ops *ca_ops
2515                         = inet_csk(sk)->icsk_ca_ops;
2516
2517                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2518                 tcp_ack_packets_out(sk);
2519
2520                 if (ca_ops->pkts_acked) {
2521                         s32 rtt_us = -1;
2522
2523                         /* Is the ACK triggering packet unambiguous? */
2524                         if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
2525                                 /* High resolution needed and available? */
2526                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2527                                     !ktime_equal(last_ackt,
2528                                                  net_invalid_timestamp()))
2529                                         rtt_us = ktime_us_delta(ktime_get_real(),
2530                                                                 last_ackt);
2531                                 else if (seq_rtt > 0)
2532                                         rtt_us = jiffies_to_usecs(seq_rtt);
2533                         }
2534
2535                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2536                 }
2537         }
2538
2539 #if FASTRETRANS_DEBUG > 0
2540         BUG_TRAP((int)tp->sacked_out >= 0);
2541         BUG_TRAP((int)tp->lost_out >= 0);
2542         BUG_TRAP((int)tp->retrans_out >= 0);
2543         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2544                 const struct inet_connection_sock *icsk = inet_csk(sk);
2545                 if (tp->lost_out) {
2546                         printk(KERN_DEBUG "Leak l=%u %d\n",
2547                                tp->lost_out, icsk->icsk_ca_state);
2548                         tp->lost_out = 0;
2549                 }
2550                 if (tp->sacked_out) {
2551                         printk(KERN_DEBUG "Leak s=%u %d\n",
2552                                tp->sacked_out, icsk->icsk_ca_state);
2553                         tp->sacked_out = 0;
2554                 }
2555                 if (tp->retrans_out) {
2556                         printk(KERN_DEBUG "Leak r=%u %d\n",
2557                                tp->retrans_out, icsk->icsk_ca_state);
2558                         tp->retrans_out = 0;
2559                 }
2560         }
2561 #endif
2562         *seq_rtt_p = seq_rtt;
2563         return acked;
2564 }
2565
2566 static void tcp_ack_probe(struct sock *sk)
2567 {
2568         const struct tcp_sock *tp = tcp_sk(sk);
2569         struct inet_connection_sock *icsk = inet_csk(sk);
2570
2571         /* Was it a usable window open? */
2572
2573         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2574                    tp->snd_una + tp->snd_wnd)) {
2575                 icsk->icsk_backoff = 0;
2576                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2577                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2578                  * This function is not for random using!
2579                  */
2580         } else {
2581                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2582                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2583                                           TCP_RTO_MAX);
2584         }
2585 }
2586
2587 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2588 {
2589         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2590                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2591 }
2592
2593 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2594 {
2595         const struct tcp_sock *tp = tcp_sk(sk);
2596         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2597                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2598 }
2599
2600 /* Check that window update is acceptable.
2601  * The function assumes that snd_una<=ack<=snd_next.
2602  */
2603 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2604                                         const u32 ack_seq, const u32 nwin)
2605 {
2606         return (after(ack, tp->snd_una) ||
2607                 after(ack_seq, tp->snd_wl1) ||
2608                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2609 }
2610
2611 /* Update our send window.
2612  *
2613  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2614  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2615  */
2616 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2617                                  u32 ack_seq)
2618 {
2619         struct tcp_sock *tp = tcp_sk(sk);
2620         int flag = 0;
2621         u32 nwin = ntohs(tcp_hdr(skb)->window);
2622
2623         if (likely(!tcp_hdr(skb)->syn))
2624                 nwin <<= tp->rx_opt.snd_wscale;
2625
2626         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2627                 flag |= FLAG_WIN_UPDATE;
2628                 tcp_update_wl(tp, ack, ack_seq);
2629
2630                 if (tp->snd_wnd != nwin) {
2631                         tp->snd_wnd = nwin;
2632
2633                         /* Note, it is the only place, where
2634                          * fast path is recovered for sending TCP.
2635                          */
2636                         tp->pred_flags = 0;
2637                         tcp_fast_path_check(sk);
2638
2639                         if (nwin > tp->max_window) {
2640                                 tp->max_window = nwin;
2641                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2642                         }
2643                 }
2644         }
2645
2646         tp->snd_una = ack;
2647
2648         return flag;
2649 }
2650
2651 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2652  * continue in congestion avoidance.
2653  */
2654 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2655 {
2656         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2657         tp->snd_cwnd_cnt = 0;
2658         TCP_ECN_queue_cwr(tp);
2659         tcp_moderate_cwnd(tp);
2660 }
2661
2662 /* A conservative spurious RTO response algorithm: reduce cwnd using
2663  * rate halving and continue in congestion avoidance.
2664  */
2665 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2666 {
2667         tcp_enter_cwr(sk, 0);
2668 }
2669
2670 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2671 {
2672         if (flag&FLAG_ECE)
2673                 tcp_ratehalving_spur_to_response(sk);
2674         else
2675                 tcp_undo_cwr(sk, 1);
2676 }
2677
2678 /* F-RTO spurious RTO detection algorithm (RFC4138)
2679  *
2680  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2681  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2682  * window (but not to or beyond highest sequence sent before RTO):
2683  *   On First ACK,  send two new segments out.
2684  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2685  *                  algorithm is not part of the F-RTO detection algorithm
2686  *                  given in RFC4138 but can be selected separately).
2687  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2688  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2689  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2690  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2691  *
2692  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2693  * original window even after we transmit two new data segments.
2694  *
2695  * SACK version:
2696  *   on first step, wait until first cumulative ACK arrives, then move to
2697  *   the second step. In second step, the next ACK decides.
2698  *
2699  * F-RTO is implemented (mainly) in four functions:
2700  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2701  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2702  *     called when tcp_use_frto() showed green light
2703  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2704  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2705  *     to prove that the RTO is indeed spurious. It transfers the control
2706  *     from F-RTO to the conventional RTO recovery
2707  */
2708 static int tcp_process_frto(struct sock *sk, int flag)
2709 {
2710         struct tcp_sock *tp = tcp_sk(sk);
2711
2712         tcp_sync_left_out(tp);
2713
2714         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2715         if (flag&FLAG_DATA_ACKED)
2716                 inet_csk(sk)->icsk_retransmits = 0;
2717
2718         if (!before(tp->snd_una, tp->frto_highmark)) {
2719                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2720                 return 1;
2721         }
2722
2723         if (!IsSackFrto() || IsReno(tp)) {
2724                 /* RFC4138 shortcoming in step 2; should also have case c):
2725                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2726                  * data, winupdate
2727                  */
2728                 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2729                         return 1;
2730
2731                 if (!(flag&FLAG_DATA_ACKED)) {
2732                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2733                                             flag);
2734                         return 1;
2735                 }
2736         } else {
2737                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2738                         /* Prevent sending of new data. */
2739                         tp->snd_cwnd = min(tp->snd_cwnd,
2740                                            tcp_packets_in_flight(tp));
2741                         return 1;
2742                 }
2743
2744                 if ((tp->frto_counter >= 2) &&
2745                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2746                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2747                         /* RFC4138 shortcoming (see comment above) */
2748                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2749                                 return 1;
2750
2751                         tcp_enter_frto_loss(sk, 3, flag);
2752                         return 1;
2753                 }
2754         }
2755
2756         if (tp->frto_counter == 1) {
2757                 /* Sending of the next skb must be allowed or no FRTO */
2758                 if (!tcp_send_head(sk) ||
2759                     after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2760                                      tp->snd_una + tp->snd_wnd)) {
2761                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2762                                             flag);
2763                         return 1;
2764                 }
2765
2766                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2767                 tp->frto_counter = 2;
2768                 return 1;
2769         } else {
2770                 switch (sysctl_tcp_frto_response) {
2771                 case 2:
2772                         tcp_undo_spur_to_response(sk, flag);
2773                         break;
2774                 case 1:
2775                         tcp_conservative_spur_to_response(tp);
2776                         break;
2777                 default:
2778                         tcp_ratehalving_spur_to_response(sk);
2779                         break;
2780                 }
2781                 tp->frto_counter = 0;
2782         }
2783         return 0;
2784 }
2785
2786 /* This routine deals with incoming acks, but not outgoing ones. */
2787 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2788 {
2789         struct inet_connection_sock *icsk = inet_csk(sk);
2790         struct tcp_sock *tp = tcp_sk(sk);
2791         u32 prior_snd_una = tp->snd_una;
2792         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2793         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2794         u32 prior_in_flight;
2795         s32 seq_rtt;
2796         int prior_packets;
2797         int frto_cwnd = 0;
2798
2799         /* If the ack is newer than sent or older than previous acks
2800          * then we can probably ignore it.
2801          */
2802         if (after(ack, tp->snd_nxt))
2803                 goto uninteresting_ack;
2804
2805         if (before(ack, prior_snd_una))
2806                 goto old_ack;
2807
2808         if (after(ack, prior_snd_una))
2809                 flag |= FLAG_SND_UNA_ADVANCED;
2810
2811         if (sysctl_tcp_abc) {
2812                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2813                         tp->bytes_acked += ack - prior_snd_una;
2814                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2815                         /* we assume just one segment left network */
2816                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2817         }
2818
2819         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2820                 /* Window is constant, pure forward advance.
2821                  * No more checks are required.
2822                  * Note, we use the fact that SND.UNA>=SND.WL2.
2823                  */
2824                 tcp_update_wl(tp, ack, ack_seq);
2825                 tp->snd_una = ack;
2826                 flag |= FLAG_WIN_UPDATE;
2827
2828                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2829
2830                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2831         } else {
2832                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2833                         flag |= FLAG_DATA;
2834                 else
2835                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2836
2837                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2838
2839                 if (TCP_SKB_CB(skb)->sacked)
2840                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2841
2842                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2843                         flag |= FLAG_ECE;
2844
2845                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2846         }
2847
2848         /* We passed data and got it acked, remove any soft error
2849          * log. Something worked...
2850          */
2851         sk->sk_err_soft = 0;
2852         tp->rcv_tstamp = tcp_time_stamp;
2853         prior_packets = tp->packets_out;
2854         if (!prior_packets)
2855                 goto no_queue;
2856
2857         prior_in_flight = tcp_packets_in_flight(tp);
2858
2859         /* See if we can take anything off of the retransmit queue. */
2860         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2861
2862         if (tp->frto_counter)
2863                 frto_cwnd = tcp_process_frto(sk, flag);
2864
2865         if (tcp_ack_is_dubious(sk, flag)) {
2866                 /* Advance CWND, if state allows this. */
2867                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2868                     tcp_may_raise_cwnd(sk, flag))
2869                         tcp_cong_avoid(sk, ack, prior_in_flight, 0);
2870                 tcp_fastretrans_alert(sk, prior_packets, flag);
2871         } else {
2872                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2873                         tcp_cong_avoid(sk, ack, prior_in_flight, 1);
2874         }
2875
2876         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2877                 dst_confirm(sk->sk_dst_cache);
2878
2879         return 1;
2880
2881 no_queue:
2882         icsk->icsk_probes_out = 0;
2883
2884         /* If this ack opens up a zero window, clear backoff.  It was
2885          * being used to time the probes, and is probably far higher than
2886          * it needs to be for normal retransmission.
2887          */
2888         if (tcp_send_head(sk))
2889                 tcp_ack_probe(sk);
2890         return 1;
2891
2892 old_ack:
2893         if (TCP_SKB_CB(skb)->sacked)
2894                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2895
2896 uninteresting_ack:
2897         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2898         return 0;
2899 }
2900
2901
2902 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2903  * But, this can also be called on packets in the established flow when
2904  * the fast version below fails.
2905  */
2906 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2907 {
2908         unsigned char *ptr;
2909         struct tcphdr *th = tcp_hdr(skb);
2910         int length=(th->doff*4)-sizeof(struct tcphdr);
2911
2912         ptr = (unsigned char *)(th + 1);
2913         opt_rx->saw_tstamp = 0;
2914
2915         while (length > 0) {
2916                 int opcode=*ptr++;
2917                 int opsize;
2918
2919                 switch (opcode) {
2920                         case TCPOPT_EOL:
2921                                 return;
2922                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2923                                 length--;
2924                                 continue;
2925                         default:
2926                                 opsize=*ptr++;
2927                                 if (opsize < 2) /* "silly options" */
2928                                         return;
2929                                 if (opsize > length)
2930                                         return; /* don't parse partial options */
2931                                 switch (opcode) {
2932                                 case TCPOPT_MSS:
2933                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2934                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2935                                                 if (in_mss) {
2936                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2937                                                                 in_mss = opt_rx->user_mss;
2938                                                         opt_rx->mss_clamp = in_mss;
2939                                                 }
2940                                         }
2941                                         break;
2942                                 case TCPOPT_WINDOW:
2943                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2944                                                 if (sysctl_tcp_window_scaling) {
2945                                                         __u8 snd_wscale = *(__u8 *) ptr;
2946                                                         opt_rx->wscale_ok = 1;
2947                                                         if (snd_wscale > 14) {
2948                                                                 if (net_ratelimit())
2949                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2950                                                                                "scaling value %d >14 received.\n",
2951                                                                                snd_wscale);
2952                                                                 snd_wscale = 14;
2953                                                         }
2954                                                         opt_rx->snd_wscale = snd_wscale;
2955                                                 }
2956                                         break;
2957                                 case TCPOPT_TIMESTAMP:
2958                                         if (opsize==TCPOLEN_TIMESTAMP) {
2959                                                 if ((estab && opt_rx->tstamp_ok) ||
2960                                                     (!estab && sysctl_tcp_timestamps)) {
2961                                                         opt_rx->saw_tstamp = 1;
2962                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2963                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2964                                                 }
2965                                         }
2966                                         break;
2967                                 case TCPOPT_SACK_PERM:
2968                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2969                                                 if (sysctl_tcp_sack) {
2970                                                         opt_rx->sack_ok = 1;
2971                                                         tcp_sack_reset(opt_rx);
2972                                                 }
2973                                         }
2974                                         break;
2975
2976                                 case TCPOPT_SACK:
2977                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2978                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2979                                            opt_rx->sack_ok) {
2980                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2981                                         }
2982                                         break;
2983 #ifdef CONFIG_TCP_MD5SIG
2984                                 case TCPOPT_MD5SIG:
2985                                         /*
2986                                          * The MD5 Hash has already been
2987                                          * checked (see tcp_v{4,6}_do_rcv()).
2988                                          */
2989                                         break;
2990 #endif
2991                                 }
2992
2993                                 ptr+=opsize-2;
2994                                 length-=opsize;
2995                 }
2996         }
2997 }
2998
2999 /* Fast parse options. This hopes to only see timestamps.
3000  * If it is wrong it falls back on tcp_parse_options().
3001  */
3002 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3003                                   struct tcp_sock *tp)
3004 {
3005         if (th->doff == sizeof(struct tcphdr)>>2) {
3006                 tp->rx_opt.saw_tstamp = 0;
3007                 return 0;
3008         } else if (tp->rx_opt.tstamp_ok &&
3009                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3010                 __be32 *ptr = (__be32 *)(th + 1);
3011                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3012                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3013                         tp->rx_opt.saw_tstamp = 1;
3014                         ++ptr;
3015                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3016                         ++ptr;
3017                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3018                         return 1;
3019                 }
3020         }
3021         tcp_parse_options(skb, &tp->rx_opt, 1);
3022         return 1;
3023 }
3024
3025 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3026 {
3027         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3028         tp->rx_opt.ts_recent_stamp = get_seconds();
3029 }
3030
3031 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3032 {
3033         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3034                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3035                  * extra check below makes sure this can only happen
3036                  * for pure ACK frames.  -DaveM
3037                  *
3038                  * Not only, also it occurs for expired timestamps.
3039                  */
3040
3041                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3042                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3043                         tcp_store_ts_recent(tp);
3044         }
3045 }
3046
3047 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3048  *
3049  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3050  * it can pass through stack. So, the following predicate verifies that
3051  * this segment is not used for anything but congestion avoidance or
3052  * fast retransmit. Moreover, we even are able to eliminate most of such
3053  * second order effects, if we apply some small "replay" window (~RTO)
3054  * to timestamp space.
3055  *
3056  * All these measures still do not guarantee that we reject wrapped ACKs
3057  * on networks with high bandwidth, when sequence space is recycled fastly,
3058  * but it guarantees that such events will be very rare and do not affect
3059  * connection seriously. This doesn't look nice, but alas, PAWS is really
3060  * buggy extension.
3061  *
3062  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3063  * states that events when retransmit arrives after original data are rare.
3064  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3065  * the biggest problem on large power networks even with minor reordering.
3066  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3067  * up to bandwidth of 18Gigabit/sec. 8) ]
3068  */
3069
3070 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3071 {
3072         struct tcp_sock *tp = tcp_sk(sk);
3073         struct tcphdr *th = tcp_hdr(skb);
3074         u32 seq = TCP_SKB_CB(skb)->seq;
3075         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3076
3077         return (/* 1. Pure ACK with correct sequence number. */
3078                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3079
3080                 /* 2. ... and duplicate ACK. */
3081                 ack == tp->snd_una &&
3082
3083                 /* 3. ... and does not update window. */
3084                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3085
3086                 /* 4. ... and sits in replay window. */
3087                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3088 }
3089
3090 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3091 {
3092         const struct tcp_sock *tp = tcp_sk(sk);
3093         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3094                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3095                 !tcp_disordered_ack(sk, skb));
3096 }
3097
3098 /* Check segment sequence number for validity.
3099  *
3100  * Segment controls are considered valid, if the segment
3101  * fits to the window after truncation to the window. Acceptability
3102  * of data (and SYN, FIN, of course) is checked separately.
3103  * See tcp_data_queue(), for example.
3104  *
3105  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3106  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3107  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3108  * (borrowed from freebsd)
3109  */
3110
3111 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3112 {
3113         return  !before(end_seq, tp->rcv_wup) &&
3114                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3115 }
3116
3117 /* When we get a reset we do this. */
3118 static void tcp_reset(struct sock *sk)
3119 {
3120         /* We want the right error as BSD sees it (and indeed as we do). */
3121         switch (sk->sk_state) {
3122                 case TCP_SYN_SENT:
3123                         sk->sk_err = ECONNREFUSED;
3124                         break;
3125                 case TCP_CLOSE_WAIT:
3126                         sk->sk_err = EPIPE;
3127                         break;
3128                 case TCP_CLOSE:
3129                         return;
3130                 default:
3131                         sk->sk_err = ECONNRESET;
3132         }
3133
3134         if (!sock_flag(sk, SOCK_DEAD))
3135                 sk->sk_error_report(sk);
3136
3137         tcp_done(sk);
3138 }
3139
3140 /*
3141  *      Process the FIN bit. This now behaves as it is supposed to work
3142  *      and the FIN takes effect when it is validly part of sequence
3143  *      space. Not before when we get holes.
3144  *
3145  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3146  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3147  *      TIME-WAIT)
3148  *
3149  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3150  *      close and we go into CLOSING (and later onto TIME-WAIT)
3151  *
3152  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3153  */
3154 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3155 {
3156         struct tcp_sock *tp = tcp_sk(sk);
3157
3158         inet_csk_schedule_ack(sk);
3159
3160         sk->sk_shutdown |= RCV_SHUTDOWN;
3161         sock_set_flag(sk, SOCK_DONE);
3162
3163         switch (sk->sk_state) {
3164                 case TCP_SYN_RECV:
3165                 case TCP_ESTABLISHED:
3166                         /* Move to CLOSE_WAIT */
3167                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3168                         inet_csk(sk)->icsk_ack.pingpong = 1;
3169                         break;
3170
3171                 case TCP_CLOSE_WAIT:
3172                 case TCP_CLOSING:
3173                         /* Received a retransmission of the FIN, do
3174                          * nothing.
3175                          */
3176                         break;
3177                 case TCP_LAST_ACK:
3178                         /* RFC793: Remain in the LAST-ACK state. */
3179                         break;
3180
3181                 case TCP_FIN_WAIT1:
3182                         /* This case occurs when a simultaneous close
3183                          * happens, we must ack the received FIN and
3184                          * enter the CLOSING state.
3185                          */
3186                         tcp_send_ack(sk);
3187                         tcp_set_state(sk, TCP_CLOSING);
3188                         break;
3189                 case TCP_FIN_WAIT2:
3190                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3191                         tcp_send_ack(sk);
3192                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3193                         break;
3194                 default:
3195                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3196                          * cases we should never reach this piece of code.
3197                          */
3198                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3199                                __FUNCTION__, sk->sk_state);
3200                         break;
3201         }
3202
3203         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3204          * Probably, we should reset in this case. For now drop them.
3205          */
3206         __skb_queue_purge(&tp->out_of_order_queue);
3207         if (tp->rx_opt.sack_ok)
3208                 tcp_sack_reset(&tp->rx_opt);
3209         sk_stream_mem_reclaim(sk);
3210
3211         if (!sock_flag(sk, SOCK_DEAD)) {
3212                 sk->sk_state_change(sk);
3213
3214                 /* Do not send POLL_HUP for half duplex close. */
3215                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3216                     sk->sk_state == TCP_CLOSE)
3217                         sk_wake_async(sk, 1, POLL_HUP);
3218                 else
3219                         sk_wake_async(sk, 1, POLL_IN);
3220         }
3221 }
3222
3223 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3224 {
3225         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3226                 if (before(seq, sp->start_seq))
3227                         sp->start_seq = seq;
3228                 if (after(end_seq, sp->end_seq))
3229                         sp->end_seq = end_seq;
3230                 return 1;
3231         }
3232         return 0;
3233 }
3234
3235 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3236 {
3237         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3238                 if (before(seq, tp->rcv_nxt))
3239                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3240                 else
3241                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3242
3243                 tp->rx_opt.dsack = 1;
3244                 tp->duplicate_sack[0].start_seq = seq;
3245                 tp->duplicate_sack[0].end_seq = end_seq;
3246                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3247         }
3248 }
3249
3250 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3251 {
3252         if (!tp->rx_opt.dsack)
3253                 tcp_dsack_set(tp, seq, end_seq);
3254         else
3255                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3256 }
3257
3258 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3259 {
3260         struct tcp_sock *tp = tcp_sk(sk);
3261
3262         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3263             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3264                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3265                 tcp_enter_quickack_mode(sk);
3266
3267                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3268                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3269
3270                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3271                                 end_seq = tp->rcv_nxt;
3272                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3273                 }
3274         }
3275
3276         tcp_send_ack(sk);
3277 }
3278
3279 /* These routines update the SACK block as out-of-order packets arrive or
3280  * in-order packets close up the sequence space.
3281  */
3282 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3283 {
3284         int this_sack;
3285         struct tcp_sack_block *sp = &tp->selective_acks[0];
3286         struct tcp_sack_block *swalk = sp+1;
3287
3288         /* See if the recent change to the first SACK eats into
3289          * or hits the sequence space of other SACK blocks, if so coalesce.
3290          */
3291         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3292                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3293                         int i;
3294
3295                         /* Zap SWALK, by moving every further SACK up by one slot.
3296                          * Decrease num_sacks.
3297                          */
3298                         tp->rx_opt.num_sacks--;
3299                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3300                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3301                                 sp[i] = sp[i+1];
3302                         continue;
3303                 }
3304                 this_sack++, swalk++;
3305         }
3306 }
3307
3308 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3309 {
3310         __u32 tmp;
3311
3312         tmp = sack1->start_seq;
3313         sack1->start_seq = sack2->start_seq;
3314         sack2->start_seq = tmp;
3315
3316         tmp = sack1->end_seq;
3317         sack1->end_seq = sack2->end_seq;
3318         sack2->end_seq = tmp;
3319 }
3320
3321 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3322 {
3323         struct tcp_sock *tp = tcp_sk(sk);
3324         struct tcp_sack_block *sp = &tp->selective_acks[0];
3325         int cur_sacks = tp->rx_opt.num_sacks;
3326         int this_sack;
3327
3328         if (!cur_sacks)
3329                 goto new_sack;
3330
3331         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3332                 if (tcp_sack_extend(sp, seq, end_seq)) {
3333                         /* Rotate this_sack to the first one. */
3334                         for (; this_sack>0; this_sack--, sp--)
3335                                 tcp_sack_swap(sp, sp-1);
3336                         if (cur_sacks > 1)
3337                                 tcp_sack_maybe_coalesce(tp);
3338                         return;
3339                 }
3340         }
3341
3342         /* Could not find an adjacent existing SACK, build a new one,
3343          * put it at the front, and shift everyone else down.  We
3344          * always know there is at least one SACK present already here.
3345          *
3346          * If the sack array is full, forget about the last one.
3347          */
3348         if (this_sack >= 4) {
3349                 this_sack--;
3350                 tp->rx_opt.num_sacks--;
3351                 sp--;
3352         }
3353         for (; this_sack > 0; this_sack--, sp--)
3354                 *sp = *(sp-1);
3355
3356 new_sack:
3357         /* Build the new head SACK, and we're done. */
3358         sp->start_seq = seq;
3359         sp->end_seq = end_seq;
3360         tp->rx_opt.num_sacks++;
3361         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3362 }
3363
3364 /* RCV.NXT advances, some SACKs should be eaten. */
3365
3366 static void tcp_sack_remove(struct tcp_sock *tp)
3367 {
3368         struct tcp_sack_block *sp = &tp->selective_acks[0];
3369         int num_sacks = tp->rx_opt.num_sacks;
3370         int this_sack;
3371
3372         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3373         if (skb_queue_empty(&tp->out_of_order_queue)) {
3374                 tp->rx_opt.num_sacks = 0;
3375                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3376                 return;
3377         }
3378
3379         for (this_sack = 0; this_sack < num_sacks; ) {
3380                 /* Check if the start of the sack is covered by RCV.NXT. */
3381                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3382                         int i;
3383
3384                         /* RCV.NXT must cover all the block! */
3385                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3386
3387                         /* Zap this SACK, by moving forward any other SACKS. */
3388                         for (i=this_sack+1; i < num_sacks; i++)
3389                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3390                         num_sacks--;
3391                         continue;
3392                 }
3393                 this_sack++;
3394                 sp++;
3395         }
3396         if (num_sacks != tp->rx_opt.num_sacks) {
3397                 tp->rx_opt.num_sacks = num_sacks;
3398                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3399         }
3400 }
3401
3402 /* This one checks to see if we can put data from the
3403  * out_of_order queue into the receive_queue.
3404  */
3405 static void tcp_ofo_queue(struct sock *sk)
3406 {
3407         struct tcp_sock *tp = tcp_sk(sk);
3408         __u32 dsack_high = tp->rcv_nxt;
3409         struct sk_buff *skb;
3410
3411         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3412                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3413                         break;
3414
3415                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3416                         __u32 dsack = dsack_high;
3417                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3418                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3419                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3420                 }
3421
3422                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3423                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3424                         __skb_unlink(skb, &tp->out_of_order_queue);
3425                         __kfree_skb(skb);
3426                         continue;
3427                 }
3428                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3429                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3430                            TCP_SKB_CB(skb)->end_seq);
3431
3432                 __skb_unlink(skb, &tp->out_of_order_queue);
3433                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3434                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3435                 if (tcp_hdr(skb)->fin)
3436                         tcp_fin(skb, sk, tcp_hdr(skb));
3437         }
3438 }
3439
3440 static int tcp_prune_queue(struct sock *sk);
3441
3442 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3443 {
3444         struct tcphdr *th = tcp_hdr(skb);
3445         struct tcp_sock *tp = tcp_sk(sk);
3446         int eaten = -1;
3447
3448         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3449                 goto drop;
3450
3451         __skb_pull(skb, th->doff*4);
3452
3453         TCP_ECN_accept_cwr(tp, skb);
3454
3455         if (tp->rx_opt.dsack) {
3456                 tp->rx_opt.dsack = 0;
3457                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3458                                                     4 - tp->rx_opt.tstamp_ok);
3459         }
3460
3461         /*  Queue data for delivery to the user.
3462          *  Packets in sequence go to the receive queue.
3463          *  Out of sequence packets to the out_of_order_queue.
3464          */
3465         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3466                 if (tcp_receive_window(tp) == 0)
3467                         goto out_of_window;
3468
3469                 /* Ok. In sequence. In window. */
3470                 if (tp->ucopy.task == current &&
3471                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3472                     sock_owned_by_user(sk) && !tp->urg_data) {
3473                         int chunk = min_t(unsigned int, skb->len,
3474                                                         tp->ucopy.len);
3475
3476                         __set_current_state(TASK_RUNNING);
3477
3478                         local_bh_enable();
3479                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3480                                 tp->ucopy.len -= chunk;
3481                                 tp->copied_seq += chunk;
3482                                 eaten = (chunk == skb->len && !th->fin);
3483                                 tcp_rcv_space_adjust(sk);
3484                         }
3485                         local_bh_disable();
3486                 }
3487
3488                 if (eaten <= 0) {
3489 queue_and_out:
3490                         if (eaten < 0 &&
3491                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3492                              !sk_stream_rmem_schedule(sk, skb))) {
3493                                 if (tcp_prune_queue(sk) < 0 ||
3494                                     !sk_stream_rmem_schedule(sk, skb))
3495                                         goto drop;
3496                         }
3497                         sk_stream_set_owner_r(skb, sk);
3498                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3499                 }
3500                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3501                 if (skb->len)
3502                         tcp_event_data_recv(sk, skb);
3503                 if (th->fin)
3504                         tcp_fin(skb, sk, th);
3505
3506                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3507                         tcp_ofo_queue(sk);
3508
3509                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3510                          * gap in queue is filled.
3511                          */
3512                         if (skb_queue_empty(&tp->out_of_order_queue))
3513                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3514                 }
3515
3516                 if (tp->rx_opt.num_sacks)
3517                         tcp_sack_remove(tp);
3518
3519                 tcp_fast_path_check(sk);
3520
3521                 if (eaten > 0)
3522                         __kfree_skb(skb);
3523                 else if (!sock_flag(sk, SOCK_DEAD))
3524                         sk->sk_data_ready(sk, 0);
3525                 return;
3526         }
3527
3528         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3529                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3530                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3531                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3532
3533 out_of_window:
3534                 tcp_enter_quickack_mode(sk);
3535                 inet_csk_schedule_ack(sk);
3536 drop:
3537                 __kfree_skb(skb);
3538                 return;
3539         }
3540
3541         /* Out of window. F.e. zero window probe. */
3542         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3543                 goto out_of_window;
3544
3545         tcp_enter_quickack_mode(sk);
3546
3547         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3548                 /* Partial packet, seq < rcv_next < end_seq */
3549                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3550                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3551                            TCP_SKB_CB(skb)->end_seq);
3552
3553                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3554
3555                 /* If window is closed, drop tail of packet. But after
3556                  * remembering D-SACK for its head made in previous line.
3557                  */
3558                 if (!tcp_receive_window(tp))
3559                         goto out_of_window;
3560                 goto queue_and_out;
3561         }
3562
3563         TCP_ECN_check_ce(tp, skb);
3564
3565         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3566             !sk_stream_rmem_schedule(sk, skb)) {
3567                 if (tcp_prune_queue(sk) < 0 ||
3568                     !sk_stream_rmem_schedule(sk, skb))
3569                         goto drop;
3570         }
3571
3572         /* Disable header prediction. */
3573         tp->pred_flags = 0;
3574         inet_csk_schedule_ack(sk);
3575
3576         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3577                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3578
3579         sk_stream_set_owner_r(skb, sk);
3580
3581         if (!skb_peek(&tp->out_of_order_queue)) {
3582                 /* Initial out of order segment, build 1 SACK. */
3583                 if (tp->rx_opt.sack_ok) {
3584                         tp->rx_opt.num_sacks = 1;
3585                         tp->rx_opt.dsack     = 0;
3586                         tp->rx_opt.eff_sacks = 1;
3587                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3588                         tp->selective_acks[0].end_seq =
3589                                                 TCP_SKB_CB(skb)->end_seq;
3590                 }
3591                 __skb_queue_head(&tp->out_of_order_queue,skb);
3592         } else {
3593                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3594                 u32 seq = TCP_SKB_CB(skb)->seq;
3595                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3596
3597                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3598                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3599
3600                         if (!tp->rx_opt.num_sacks ||
3601                             tp->selective_acks[0].end_seq != seq)
3602                                 goto add_sack;
3603
3604                         /* Common case: data arrive in order after hole. */
3605                         tp->selective_acks[0].end_seq = end_seq;
3606                         return;
3607                 }
3608
3609                 /* Find place to insert this segment. */
3610                 do {
3611                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3612                                 break;
3613                 } while ((skb1 = skb1->prev) !=
3614                          (struct sk_buff*)&tp->out_of_order_queue);
3615
3616                 /* Do skb overlap to previous one? */
3617                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3618                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3619                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3620                                 /* All the bits are present. Drop. */
3621                                 __kfree_skb(skb);
3622                                 tcp_dsack_set(tp, seq, end_seq);
3623                                 goto add_sack;
3624                         }
3625                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3626                                 /* Partial overlap. */
3627                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3628                         } else {
3629                                 skb1 = skb1->prev;
3630                         }
3631                 }
3632                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3633
3634                 /* And clean segments covered by new one as whole. */
3635                 while ((skb1 = skb->next) !=
3636                        (struct sk_buff*)&tp->out_of_order_queue &&
3637                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3638                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3639                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3640                                break;
3641                        }
3642                        __skb_unlink(skb1, &tp->out_of_order_queue);
3643                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3644                        __kfree_skb(skb1);
3645                 }
3646
3647 add_sack:
3648                 if (tp->rx_opt.sack_ok)
3649                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3650         }
3651 }
3652
3653 /* Collapse contiguous sequence of skbs head..tail with
3654  * sequence numbers start..end.
3655  * Segments with FIN/SYN are not collapsed (only because this
3656  * simplifies code)
3657  */
3658 static void
3659 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3660              struct sk_buff *head, struct sk_buff *tail,
3661              u32 start, u32 end)
3662 {
3663         struct sk_buff *skb;
3664
3665         /* First, check that queue is collapsible and find
3666          * the point where collapsing can be useful. */
3667         for (skb = head; skb != tail; ) {
3668                 /* No new bits? It is possible on ofo queue. */
3669                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3670                         struct sk_buff *next = skb->next;
3671                         __skb_unlink(skb, list);
3672                         __kfree_skb(skb);
3673                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3674                         skb = next;
3675                         continue;
3676                 }
3677
3678                 /* The first skb to collapse is:
3679                  * - not SYN/FIN and
3680                  * - bloated or contains data before "start" or
3681                  *   overlaps to the next one.
3682                  */
3683                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3684                     (tcp_win_from_space(skb->truesize) > skb->len ||
3685                      before(TCP_SKB_CB(skb)->seq, start) ||
3686                      (skb->next != tail &&
3687                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3688                         break;
3689
3690                 /* Decided to skip this, advance start seq. */
3691                 start = TCP_SKB_CB(skb)->end_seq;
3692                 skb = skb->next;
3693         }
3694         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3695                 return;
3696
3697         while (before(start, end)) {
3698                 struct sk_buff *nskb;
3699                 int header = skb_headroom(skb);
3700                 int copy = SKB_MAX_ORDER(header, 0);
3701
3702                 /* Too big header? This can happen with IPv6. */
3703                 if (copy < 0)
3704                         return;
3705                 if (end-start < copy)
3706                         copy = end-start;
3707                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3708                 if (!nskb)
3709                         return;
3710
3711                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3712                 skb_set_network_header(nskb, (skb_network_header(skb) -
3713                                               skb->head));
3714                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3715                                                 skb->head));
3716                 skb_reserve(nskb, header);
3717                 memcpy(nskb->head, skb->head, header);
3718                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3719                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3720                 __skb_insert(nskb, skb->prev, skb, list);
3721                 sk_stream_set_owner_r(nskb, sk);
3722
3723                 /* Copy data, releasing collapsed skbs. */
3724                 while (copy > 0) {
3725                         int offset = start - TCP_SKB_CB(skb)->seq;
3726                         int size = TCP_SKB_CB(skb)->end_seq - start;
3727
3728                         BUG_ON(offset < 0);
3729                         if (size > 0) {
3730                                 size = min(copy, size);
3731                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3732                                         BUG();
3733                                 TCP_SKB_CB(nskb)->end_seq += size;
3734                                 copy -= size;
3735                                 start += size;
3736                         }
3737                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3738                                 struct sk_buff *next = skb->next;
3739                                 __skb_unlink(skb, list);
3740                                 __kfree_skb(skb);
3741                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3742                                 skb = next;
3743                                 if (skb == tail ||
3744                                     tcp_hdr(skb)->syn ||
3745                                     tcp_hdr(skb)->fin)
3746                                         return;
3747                         }
3748                 }
3749         }
3750 }
3751
3752 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3753  * and tcp_collapse() them until all the queue is collapsed.
3754  */
3755 static void tcp_collapse_ofo_queue(struct sock *sk)
3756 {
3757         struct tcp_sock *tp = tcp_sk(sk);
3758         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3759         struct sk_buff *head;
3760         u32 start, end;
3761
3762         if (skb == NULL)
3763                 return;
3764
3765         start = TCP_SKB_CB(skb)->seq;
3766         end = TCP_SKB_CB(skb)->end_seq;
3767         head = skb;
3768
3769         for (;;) {
3770                 skb = skb->next;
3771
3772                 /* Segment is terminated when we see gap or when
3773                  * we are at the end of all the queue. */
3774                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3775                     after(TCP_SKB_CB(skb)->seq, end) ||
3776                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3777                         tcp_collapse(sk, &tp->out_of_order_queue,
3778                                      head, skb, start, end);
3779                         head = skb;
3780                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3781                                 break;
3782                         /* Start new segment */
3783                         start = TCP_SKB_CB(skb)->seq;
3784                         end = TCP_SKB_CB(skb)->end_seq;
3785                 } else {
3786                         if (before(TCP_SKB_CB(skb)->seq, start))
3787                                 start = TCP_SKB_CB(skb)->seq;
3788                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3789                                 end = TCP_SKB_CB(skb)->end_seq;
3790                 }
3791         }
3792 }
3793
3794 /* Reduce allocated memory if we can, trying to get
3795  * the socket within its memory limits again.
3796  *
3797  * Return less than zero if we should start dropping frames
3798  * until the socket owning process reads some of the data
3799  * to stabilize the situation.
3800  */
3801 static int tcp_prune_queue(struct sock *sk)
3802 {
3803         struct tcp_sock *tp = tcp_sk(sk);
3804
3805         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3806
3807         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3808
3809         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3810                 tcp_clamp_window(sk);
3811         else if (tcp_memory_pressure)
3812                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3813
3814         tcp_collapse_ofo_queue(sk);
3815         tcp_collapse(sk, &sk->sk_receive_queue,
3816                      sk->sk_receive_queue.next,
3817                      (struct sk_buff*)&sk->sk_receive_queue,
3818                      tp->copied_seq, tp->rcv_nxt);
3819         sk_stream_mem_reclaim(sk);
3820
3821         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3822                 return 0;
3823
3824         /* Collapsing did not help, destructive actions follow.
3825          * This must not ever occur. */
3826
3827         /* First, purge the out_of_order queue. */
3828         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3829                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3830                 __skb_queue_purge(&tp->out_of_order_queue);
3831
3832                 /* Reset SACK state.  A conforming SACK implementation will
3833                  * do the same at a timeout based retransmit.  When a connection
3834                  * is in a sad state like this, we care only about integrity
3835                  * of the connection not performance.
3836                  */
3837                 if (tp->rx_opt.sack_ok)
3838                         tcp_sack_reset(&tp->rx_opt);
3839                 sk_stream_mem_reclaim(sk);
3840         }
3841
3842         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3843                 return 0;
3844
3845         /* If we are really being abused, tell the caller to silently
3846          * drop receive data on the floor.  It will get retransmitted
3847          * and hopefully then we'll have sufficient space.
3848          */
3849         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3850
3851         /* Massive buffer overcommit. */
3852         tp->pred_flags = 0;
3853         return -1;
3854 }
3855
3856
3857 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3858  * As additional protections, we do not touch cwnd in retransmission phases,
3859  * and if application hit its sndbuf limit recently.
3860  */
3861 void tcp_cwnd_application_limited(struct sock *sk)
3862 {
3863         struct tcp_sock *tp = tcp_sk(sk);
3864
3865         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3866             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3867                 /* Limited by application or receiver window. */
3868                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3869                 u32 win_used = max(tp->snd_cwnd_used, init_win);
3870                 if (win_used < tp->snd_cwnd) {
3871                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3872                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3873                 }
3874                 tp->snd_cwnd_used = 0;
3875         }
3876         tp->snd_cwnd_stamp = tcp_time_stamp;
3877 }
3878
3879 static int tcp_should_expand_sndbuf(struct sock *sk)
3880 {
3881         struct tcp_sock *tp = tcp_sk(sk);
3882
3883         /* If the user specified a specific send buffer setting, do
3884          * not modify it.
3885          */
3886         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3887                 return 0;
3888
3889         /* If we are under global TCP memory pressure, do not expand.  */
3890         if (tcp_memory_pressure)
3891                 return 0;
3892
3893         /* If we are under soft global TCP memory pressure, do not expand.  */
3894         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3895                 return 0;
3896
3897         /* If we filled the congestion window, do not expand.  */
3898         if (tp->packets_out >= tp->snd_cwnd)
3899                 return 0;
3900
3901         return 1;
3902 }
3903
3904 /* When incoming ACK allowed to free some skb from write_queue,
3905  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3906  * on the exit from tcp input handler.
3907  *
3908  * PROBLEM: sndbuf expansion does not work well with largesend.
3909  */
3910 static void tcp_new_space(struct sock *sk)
3911 {
3912         struct tcp_sock *tp = tcp_sk(sk);
3913
3914         if (tcp_should_expand_sndbuf(sk)) {
3915                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3916                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3917                     demanded = max_t(unsigned int, tp->snd_cwnd,
3918                                                    tp->reordering + 1);
3919                 sndmem *= 2*demanded;
3920                 if (sndmem > sk->sk_sndbuf)
3921                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3922                 tp->snd_cwnd_stamp = tcp_time_stamp;
3923         }
3924
3925         sk->sk_write_space(sk);
3926 }
3927
3928 static void tcp_check_space(struct sock *sk)
3929 {
3930         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3931                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3932                 if (sk->sk_socket &&
3933                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3934                         tcp_new_space(sk);
3935         }
3936 }
3937
3938 static inline void tcp_data_snd_check(struct sock *sk)
3939 {
3940         tcp_push_pending_frames(sk);
3941         tcp_check_space(sk);
3942 }
3943
3944 /*
3945  * Check if sending an ack is needed.
3946  */
3947 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3948 {
3949         struct tcp_sock *tp = tcp_sk(sk);
3950
3951             /* More than one full frame received... */
3952         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3953              /* ... and right edge of window advances far enough.
3954               * (tcp_recvmsg() will send ACK otherwise). Or...
3955               */
3956              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3957             /* We ACK each frame or... */
3958             tcp_in_quickack_mode(sk) ||
3959             /* We have out of order data. */
3960             (ofo_possible &&
3961              skb_peek(&tp->out_of_order_queue))) {
3962                 /* Then ack it now */
3963                 tcp_send_ack(sk);
3964         } else {
3965                 /* Else, send delayed ack. */
3966                 tcp_send_delayed_ack(sk);
3967         }
3968 }
3969
3970 static inline void tcp_ack_snd_check(struct sock *sk)
3971 {
3972         if (!inet_csk_ack_scheduled(sk)) {
3973                 /* We sent a data segment already. */
3974                 return;
3975         }
3976         __tcp_ack_snd_check(sk, 1);
3977 }
3978
3979 /*
3980  *      This routine is only called when we have urgent data
3981  *      signaled. Its the 'slow' part of tcp_urg. It could be
3982  *      moved inline now as tcp_urg is only called from one
3983  *      place. We handle URGent data wrong. We have to - as
3984  *      BSD still doesn't use the correction from RFC961.
3985  *      For 1003.1g we should support a new option TCP_STDURG to permit
3986  *      either form (or just set the sysctl tcp_stdurg).
3987  */
3988
3989 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3990 {
3991         struct tcp_sock *tp = tcp_sk(sk);
3992         u32 ptr = ntohs(th->urg_ptr);
3993
3994         if (ptr && !sysctl_tcp_stdurg)
3995                 ptr--;
3996         ptr += ntohl(th->seq);
3997
3998         /* Ignore urgent data that we've already seen and read. */
3999         if (after(tp->copied_seq, ptr))
4000                 return;
4001
4002         /* Do not replay urg ptr.
4003          *
4004          * NOTE: interesting situation not covered by specs.
4005          * Misbehaving sender may send urg ptr, pointing to segment,
4006          * which we already have in ofo queue. We are not able to fetch
4007          * such data and will stay in TCP_URG_NOTYET until will be eaten
4008          * by recvmsg(). Seems, we are not obliged to handle such wicked
4009          * situations. But it is worth to think about possibility of some
4010          * DoSes using some hypothetical application level deadlock.
4011          */
4012         if (before(ptr, tp->rcv_nxt))
4013                 return;
4014
4015         /* Do we already have a newer (or duplicate) urgent pointer? */
4016         if (tp->urg_data && !after(ptr, tp->urg_seq))
4017                 return;
4018
4019         /* Tell the world about our new urgent pointer. */
4020         sk_send_sigurg(sk);
4021
4022         /* We may be adding urgent data when the last byte read was
4023          * urgent. To do this requires some care. We cannot just ignore
4024          * tp->copied_seq since we would read the last urgent byte again
4025          * as data, nor can we alter copied_seq until this data arrives
4026          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4027          *
4028          * NOTE. Double Dutch. Rendering to plain English: author of comment
4029          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4030          * and expect that both A and B disappear from stream. This is _wrong_.
4031          * Though this happens in BSD with high probability, this is occasional.
4032          * Any application relying on this is buggy. Note also, that fix "works"
4033          * only in this artificial test. Insert some normal data between A and B and we will
4034          * decline of BSD again. Verdict: it is better to remove to trap
4035          * buggy users.
4036          */
4037         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4038             !sock_flag(sk, SOCK_URGINLINE) &&
4039             tp->copied_seq != tp->rcv_nxt) {
4040                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4041                 tp->copied_seq++;
4042                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4043                         __skb_unlink(skb, &sk->sk_receive_queue);
4044                         __kfree_skb(skb);
4045                 }
4046         }
4047
4048         tp->urg_data   = TCP_URG_NOTYET;
4049         tp->urg_seq    = ptr;
4050
4051         /* Disable header prediction. */
4052         tp->pred_flags = 0;
4053 }
4054
4055 /* This is the 'fast' part of urgent handling. */
4056 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4057 {
4058         struct tcp_sock *tp = tcp_sk(sk);
4059
4060         /* Check if we get a new urgent pointer - normally not. */
4061         if (th->urg)
4062                 tcp_check_urg(sk,th);
4063
4064         /* Do we wait for any urgent data? - normally not... */
4065         if (tp->urg_data == TCP_URG_NOTYET) {
4066                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4067                           th->syn;
4068
4069                 /* Is the urgent pointer pointing into this packet? */
4070                 if (ptr < skb->len) {
4071                         u8 tmp;
4072                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4073                                 BUG();
4074                         tp->urg_data = TCP_URG_VALID | tmp;
4075                         if (!sock_flag(sk, SOCK_DEAD))
4076                                 sk->sk_data_ready(sk, 0);
4077                 }
4078         }
4079 }
4080
4081 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4082 {
4083         struct tcp_sock *tp = tcp_sk(sk);
4084         int chunk = skb->len - hlen;
4085         int err;
4086
4087         local_bh_enable();
4088         if (skb_csum_unnecessary(skb))
4089                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4090         else
4091                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4092                                                        tp->ucopy.iov);
4093
4094         if (!err) {
4095                 tp->ucopy.len -= chunk;
4096                 tp->copied_seq += chunk;
4097                 tcp_rcv_space_adjust(sk);
4098         }
4099
4100         local_bh_disable();
4101         return err;
4102 }
4103
4104 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4105 {
4106         __sum16 result;
4107
4108         if (sock_owned_by_user(sk)) {
4109                 local_bh_enable();
4110                 result = __tcp_checksum_complete(skb);
4111                 local_bh_disable();
4112         } else {
4113                 result = __tcp_checksum_complete(skb);
4114         }
4115         return result;
4116 }
4117
4118 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4119 {
4120         return !skb_csum_unnecessary(skb) &&
4121                 __tcp_checksum_complete_user(sk, skb);
4122 }
4123
4124 #ifdef CONFIG_NET_DMA
4125 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4126 {
4127         struct tcp_sock *tp = tcp_sk(sk);
4128         int chunk = skb->len - hlen;
4129         int dma_cookie;
4130         int copied_early = 0;
4131
4132         if (tp->ucopy.wakeup)
4133                 return 0;
4134
4135         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4136                 tp->ucopy.dma_chan = get_softnet_dma();
4137
4138         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4139
4140                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4141                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4142
4143                 if (dma_cookie < 0)
4144                         goto out;
4145
4146                 tp->ucopy.dma_cookie = dma_cookie;
4147                 copied_early = 1;
4148
4149                 tp->ucopy.len -= chunk;
4150                 tp->copied_seq += chunk;
4151                 tcp_rcv_space_adjust(sk);
4152
4153                 if ((tp->ucopy.len == 0) ||
4154                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4155                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4156                         tp->ucopy.wakeup = 1;
4157                         sk->sk_data_ready(sk, 0);
4158                 }
4159         } else if (chunk > 0) {
4160                 tp->ucopy.wakeup = 1;
4161                 sk->sk_data_ready(sk, 0);
4162         }
4163 out:
4164         return copied_early;
4165 }
4166 #endif /* CONFIG_NET_DMA */
4167
4168 /*
4169  *      TCP receive function for the ESTABLISHED state.
4170  *
4171  *      It is split into a fast path and a slow path. The fast path is
4172  *      disabled when:
4173  *      - A zero window was announced from us - zero window probing
4174  *        is only handled properly in the slow path.
4175  *      - Out of order segments arrived.
4176  *      - Urgent data is expected.
4177  *      - There is no buffer space left
4178  *      - Unexpected TCP flags/window values/header lengths are received
4179  *        (detected by checking the TCP header against pred_flags)
4180  *      - Data is sent in both directions. Fast path only supports pure senders
4181  *        or pure receivers (this means either the sequence number or the ack
4182  *        value must stay constant)
4183  *      - Unexpected TCP option.
4184  *
4185  *      When these conditions are not satisfied it drops into a standard
4186  *      receive procedure patterned after RFC793 to handle all cases.
4187  *      The first three cases are guaranteed by proper pred_flags setting,
4188  *      the rest is checked inline. Fast processing is turned on in
4189  *      tcp_data_queue when everything is OK.
4190  */
4191 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4192                         struct tcphdr *th, unsigned len)
4193 {
4194         struct tcp_sock *tp = tcp_sk(sk);
4195
4196         /*
4197          *      Header prediction.
4198          *      The code loosely follows the one in the famous
4199          *      "30 instruction TCP receive" Van Jacobson mail.
4200          *
4201          *      Van's trick is to deposit buffers into socket queue
4202          *      on a device interrupt, to call tcp_recv function
4203          *      on the receive process context and checksum and copy
4204          *      the buffer to user space. smart...
4205          *
4206          *      Our current scheme is not silly either but we take the
4207          *      extra cost of the net_bh soft interrupt processing...
4208          *      We do checksum and copy also but from device to kernel.
4209          */
4210
4211         tp->rx_opt.saw_tstamp = 0;
4212
4213         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4214          *      if header_prediction is to be made
4215          *      'S' will always be tp->tcp_header_len >> 2
4216          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4217          *  turn it off (when there are holes in the receive
4218          *       space for instance)
4219          *      PSH flag is ignored.
4220          */
4221
4222         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4223                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4224                 int tcp_header_len = tp->tcp_header_len;
4225
4226                 /* Timestamp header prediction: tcp_header_len
4227                  * is automatically equal to th->doff*4 due to pred_flags
4228                  * match.
4229                  */
4230
4231                 /* Check timestamp */
4232                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4233                         __be32 *ptr = (__be32 *)(th + 1);
4234
4235                         /* No? Slow path! */
4236                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4237                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4238                                 goto slow_path;
4239
4240                         tp->rx_opt.saw_tstamp = 1;
4241                         ++ptr;
4242                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4243                         ++ptr;
4244                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4245
4246                         /* If PAWS failed, check it more carefully in slow path */
4247                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4248                                 goto slow_path;
4249
4250                         /* DO NOT update ts_recent here, if checksum fails
4251                          * and timestamp was corrupted part, it will result
4252                          * in a hung connection since we will drop all
4253                          * future packets due to the PAWS test.
4254                          */
4255                 }
4256
4257                 if (len <= tcp_header_len) {
4258                         /* Bulk data transfer: sender */
4259                         if (len == tcp_header_len) {
4260                                 /* Predicted packet is in window by definition.
4261                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4262                                  * Hence, check seq<=rcv_wup reduces to:
4263                                  */
4264                                 if (tcp_header_len ==
4265                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4266                                     tp->rcv_nxt == tp->rcv_wup)
4267                                         tcp_store_ts_recent(tp);
4268
4269                                 /* We know that such packets are checksummed
4270                                  * on entry.
4271                                  */
4272                                 tcp_ack(sk, skb, 0);
4273                                 __kfree_skb(skb);
4274                                 tcp_data_snd_check(sk);
4275                                 return 0;
4276                         } else { /* Header too small */
4277                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4278                                 goto discard;
4279                         }
4280                 } else {
4281                         int eaten = 0;
4282                         int copied_early = 0;
4283
4284                         if (tp->copied_seq == tp->rcv_nxt &&
4285                             len - tcp_header_len <= tp->ucopy.len) {
4286 #ifdef CONFIG_NET_DMA
4287                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4288                                         copied_early = 1;
4289                                         eaten = 1;
4290                                 }
4291 #endif
4292                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4293                                         __set_current_state(TASK_RUNNING);
4294
4295                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4296                                                 eaten = 1;
4297                                 }
4298                                 if (eaten) {
4299                                         /* Predicted packet is in window by definition.
4300                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4301                                          * Hence, check seq<=rcv_wup reduces to:
4302                                          */
4303                                         if (tcp_header_len ==
4304                                             (sizeof(struct tcphdr) +
4305                                              TCPOLEN_TSTAMP_ALIGNED) &&
4306                                             tp->rcv_nxt == tp->rcv_wup)
4307                                                 tcp_store_ts_recent(tp);
4308
4309                                         tcp_rcv_rtt_measure_ts(sk, skb);
4310
4311                                         __skb_pull(skb, tcp_header_len);
4312                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4313                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4314                                 }
4315                                 if (copied_early)
4316                                         tcp_cleanup_rbuf(sk, skb->len);
4317                         }
4318                         if (!eaten) {
4319                                 if (tcp_checksum_complete_user(sk, skb))
4320                                         goto csum_error;
4321
4322                                 /* Predicted packet is in window by definition.
4323                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4324                                  * Hence, check seq<=rcv_wup reduces to:
4325                                  */
4326                                 if (tcp_header_len ==
4327                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4328                                     tp->rcv_nxt == tp->rcv_wup)
4329                                         tcp_store_ts_recent(tp);
4330
4331                                 tcp_rcv_rtt_measure_ts(sk, skb);
4332
4333                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4334                                         goto step5;
4335
4336                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4337
4338                                 /* Bulk data transfer: receiver */
4339                                 __skb_pull(skb,tcp_header_len);
4340                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4341                                 sk_stream_set_owner_r(skb, sk);
4342                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4343                         }
4344
4345                         tcp_event_data_recv(sk, skb);
4346
4347                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4348                                 /* Well, only one small jumplet in fast path... */
4349                                 tcp_ack(sk, skb, FLAG_DATA);
4350                                 tcp_data_snd_check(sk);
4351                                 if (!inet_csk_ack_scheduled(sk))
4352                                         goto no_ack;
4353                         }
4354
4355                         __tcp_ack_snd_check(sk, 0);
4356 no_ack:
4357 #ifdef CONFIG_NET_DMA
4358                         if (copied_early)
4359                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4360                         else
4361 #endif
4362                         if (eaten)
4363                                 __kfree_skb(skb);
4364                         else
4365                                 sk->sk_data_ready(sk, 0);
4366                         return 0;
4367                 }
4368         }
4369
4370 slow_path:
4371         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4372                 goto csum_error;
4373
4374         /*
4375          * RFC1323: H1. Apply PAWS check first.
4376          */
4377         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4378             tcp_paws_discard(sk, skb)) {
4379                 if (!th->rst) {
4380                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4381                         tcp_send_dupack(sk, skb);
4382                         goto discard;
4383                 }
4384                 /* Resets are accepted even if PAWS failed.
4385
4386                    ts_recent update must be made after we are sure
4387                    that the packet is in window.
4388                  */
4389         }
4390
4391         /*
4392          *      Standard slow path.
4393          */
4394
4395         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4396                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4397                  * (RST) segments are validated by checking their SEQ-fields."
4398                  * And page 69: "If an incoming segment is not acceptable,
4399                  * an acknowledgment should be sent in reply (unless the RST bit
4400                  * is set, if so drop the segment and return)".
4401                  */
4402                 if (!th->rst)
4403                         tcp_send_dupack(sk, skb);
4404                 goto discard;
4405         }
4406
4407         if (th->rst) {
4408                 tcp_reset(sk);
4409                 goto discard;
4410         }
4411
4412         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4413
4414         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4415                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4416                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4417                 tcp_reset(sk);
4418                 return 1;
4419         }
4420
4421 step5:
4422         if (th->ack)
4423                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4424
4425         tcp_rcv_rtt_measure_ts(sk, skb);
4426
4427         /* Process urgent data. */
4428         tcp_urg(sk, skb, th);
4429
4430         /* step 7: process the segment text */
4431         tcp_data_queue(sk, skb);
4432
4433         tcp_data_snd_check(sk);
4434         tcp_ack_snd_check(sk);
4435         return 0;
4436
4437 csum_error:
4438         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4439
4440 discard:
4441         __kfree_skb(skb);
4442         return 0;
4443 }
4444
4445 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4446                                          struct tcphdr *th, unsigned len)
4447 {
4448         struct tcp_sock *tp = tcp_sk(sk);
4449         struct inet_connection_sock *icsk = inet_csk(sk);
4450         int saved_clamp = tp->rx_opt.mss_clamp;
4451
4452         tcp_parse_options(skb, &tp->rx_opt, 0);
4453
4454         if (th->ack) {
4455                 /* rfc793:
4456                  * "If the state is SYN-SENT then
4457                  *    first check the ACK bit
4458                  *      If the ACK bit is set
4459                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4460                  *        a reset (unless the RST bit is set, if so drop
4461                  *        the segment and return)"
4462                  *
4463                  *  We do not send data with SYN, so that RFC-correct
4464                  *  test reduces to:
4465                  */
4466                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4467                         goto reset_and_undo;
4468
4469                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4470                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4471                              tcp_time_stamp)) {
4472                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4473                         goto reset_and_undo;
4474                 }
4475
4476                 /* Now ACK is acceptable.
4477                  *
4478                  * "If the RST bit is set
4479                  *    If the ACK was acceptable then signal the user "error:
4480                  *    connection reset", drop the segment, enter CLOSED state,
4481                  *    delete TCB, and return."
4482                  */
4483
4484                 if (th->rst) {
4485                         tcp_reset(sk);
4486                         goto discard;
4487                 }
4488
4489                 /* rfc793:
4490                  *   "fifth, if neither of the SYN or RST bits is set then
4491                  *    drop the segment and return."
4492                  *
4493                  *    See note below!
4494                  *                                        --ANK(990513)
4495                  */
4496                 if (!th->syn)
4497                         goto discard_and_undo;
4498
4499                 /* rfc793:
4500                  *   "If the SYN bit is on ...
4501                  *    are acceptable then ...
4502                  *    (our SYN has been ACKed), change the connection
4503                  *    state to ESTABLISHED..."
4504                  */
4505
4506                 TCP_ECN_rcv_synack(tp, th);
4507
4508                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4509                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4510
4511                 /* Ok.. it's good. Set up sequence numbers and
4512                  * move to established.
4513                  */
4514                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4515                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4516
4517                 /* RFC1323: The window in SYN & SYN/ACK segments is
4518                  * never scaled.
4519                  */
4520                 tp->snd_wnd = ntohs(th->window);
4521                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4522
4523                 if (!tp->rx_opt.wscale_ok) {
4524                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4525                         tp->window_clamp = min(tp->window_clamp, 65535U);
4526                 }
4527
4528                 if (tp->rx_opt.saw_tstamp) {
4529                         tp->rx_opt.tstamp_ok       = 1;
4530                         tp->tcp_header_len =
4531                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4532                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4533                         tcp_store_ts_recent(tp);
4534                 } else {
4535                         tp->tcp_header_len = sizeof(struct tcphdr);
4536                 }
4537
4538                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4539                         tp->rx_opt.sack_ok |= 2;
4540
4541                 tcp_mtup_init(sk);
4542                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4543                 tcp_initialize_rcv_mss(sk);
4544
4545                 /* Remember, tcp_poll() does not lock socket!
4546                  * Change state from SYN-SENT only after copied_seq
4547                  * is initialized. */
4548                 tp->copied_seq = tp->rcv_nxt;
4549                 smp_mb();
4550                 tcp_set_state(sk, TCP_ESTABLISHED);
4551
4552                 security_inet_conn_established(sk, skb);
4553
4554                 /* Make sure socket is routed, for correct metrics.  */
4555                 icsk->icsk_af_ops->rebuild_header(sk);
4556
4557                 tcp_init_metrics(sk);
4558
4559                 tcp_init_congestion_control(sk);
4560
4561                 /* Prevent spurious tcp_cwnd_restart() on first data
4562                  * packet.
4563                  */
4564                 tp->lsndtime = tcp_time_stamp;
4565
4566                 tcp_init_buffer_space(sk);
4567
4568                 if (sock_flag(sk, SOCK_KEEPOPEN))
4569                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4570
4571                 if (!tp->rx_opt.snd_wscale)
4572                         __tcp_fast_path_on(tp, tp->snd_wnd);
4573                 else
4574                         tp->pred_flags = 0;
4575
4576                 if (!sock_flag(sk, SOCK_DEAD)) {
4577                         sk->sk_state_change(sk);
4578                         sk_wake_async(sk, 0, POLL_OUT);
4579                 }
4580
4581                 if (sk->sk_write_pending ||
4582                     icsk->icsk_accept_queue.rskq_defer_accept ||
4583                     icsk->icsk_ack.pingpong) {
4584                         /* Save one ACK. Data will be ready after
4585                          * several ticks, if write_pending is set.
4586                          *
4587                          * It may be deleted, but with this feature tcpdumps
4588                          * look so _wonderfully_ clever, that I was not able
4589                          * to stand against the temptation 8)     --ANK
4590                          */
4591                         inet_csk_schedule_ack(sk);
4592                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4593                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4594                         tcp_incr_quickack(sk);
4595                         tcp_enter_quickack_mode(sk);
4596                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4597                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4598
4599 discard:
4600                         __kfree_skb(skb);
4601                         return 0;
4602                 } else {
4603                         tcp_send_ack(sk);
4604                 }
4605                 return -1;
4606         }
4607
4608         /* No ACK in the segment */
4609
4610         if (th->rst) {
4611                 /* rfc793:
4612                  * "If the RST bit is set
4613                  *
4614                  *      Otherwise (no ACK) drop the segment and return."
4615                  */
4616
4617                 goto discard_and_undo;
4618         }
4619
4620         /* PAWS check. */
4621         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4622                 goto discard_and_undo;
4623
4624         if (th->syn) {
4625                 /* We see SYN without ACK. It is attempt of
4626                  * simultaneous connect with crossed SYNs.
4627                  * Particularly, it can be connect to self.
4628                  */
4629                 tcp_set_state(sk, TCP_SYN_RECV);
4630
4631                 if (tp->rx_opt.saw_tstamp) {
4632                         tp->rx_opt.tstamp_ok = 1;
4633                         tcp_store_ts_recent(tp);
4634                         tp->tcp_header_len =
4635                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4636                 } else {
4637                         tp->tcp_header_len = sizeof(struct tcphdr);
4638                 }
4639
4640                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4641                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4642
4643                 /* RFC1323: The window in SYN & SYN/ACK segments is
4644                  * never scaled.
4645                  */
4646                 tp->snd_wnd    = ntohs(th->window);
4647                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4648                 tp->max_window = tp->snd_wnd;
4649
4650                 TCP_ECN_rcv_syn(tp, th);
4651
4652                 tcp_mtup_init(sk);
4653                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4654                 tcp_initialize_rcv_mss(sk);
4655
4656
4657                 tcp_send_synack(sk);
4658 #if 0
4659                 /* Note, we could accept data and URG from this segment.
4660                  * There are no obstacles to make this.
4661                  *
4662                  * However, if we ignore data in ACKless segments sometimes,
4663                  * we have no reasons to accept it sometimes.
4664                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4665                  * is not flawless. So, discard packet for sanity.
4666                  * Uncomment this return to process the data.
4667                  */
4668                 return -1;
4669 #else
4670                 goto discard;
4671 #endif
4672         }
4673         /* "fifth, if neither of the SYN or RST bits is set then
4674          * drop the segment and return."
4675          */
4676
4677 discard_and_undo:
4678         tcp_clear_options(&tp->rx_opt);
4679         tp->rx_opt.mss_clamp = saved_clamp;
4680         goto discard;
4681
4682 reset_and_undo:
4683         tcp_clear_options(&tp->rx_opt);
4684         tp->rx_opt.mss_clamp = saved_clamp;
4685         return 1;
4686 }
4687
4688
4689 /*
4690  *      This function implements the receiving procedure of RFC 793 for
4691  *      all states except ESTABLISHED and TIME_WAIT.
4692  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4693  *      address independent.
4694  */
4695
4696 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4697                           struct tcphdr *th, unsigned len)
4698 {
4699         struct tcp_sock *tp = tcp_sk(sk);
4700         struct inet_connection_sock *icsk = inet_csk(sk);
4701         int queued = 0;
4702
4703         tp->rx_opt.saw_tstamp = 0;
4704
4705         switch (sk->sk_state) {
4706         case TCP_CLOSE:
4707                 goto discard;
4708
4709         case TCP_LISTEN:
4710                 if (th->ack)
4711                         return 1;
4712
4713                 if (th->rst)
4714                         goto discard;
4715
4716                 if (th->syn) {
4717                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4718                                 return 1;
4719
4720                         /* Now we have several options: In theory there is
4721                          * nothing else in the frame. KA9Q has an option to
4722                          * send data with the syn, BSD accepts data with the
4723                          * syn up to the [to be] advertised window and
4724                          * Solaris 2.1 gives you a protocol error. For now
4725                          * we just ignore it, that fits the spec precisely
4726                          * and avoids incompatibilities. It would be nice in
4727                          * future to drop through and process the data.
4728                          *
4729                          * Now that TTCP is starting to be used we ought to
4730                          * queue this data.
4731                          * But, this leaves one open to an easy denial of
4732                          * service attack, and SYN cookies can't defend
4733                          * against this problem. So, we drop the data
4734                          * in the interest of security over speed unless
4735                          * it's still in use.
4736                          */
4737                         kfree_skb(skb);
4738                         return 0;
4739                 }
4740                 goto discard;
4741
4742         case TCP_SYN_SENT:
4743                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4744                 if (queued >= 0)
4745                         return queued;
4746
4747                 /* Do step6 onward by hand. */
4748                 tcp_urg(sk, skb, th);
4749                 __kfree_skb(skb);
4750                 tcp_data_snd_check(sk);
4751                 return 0;
4752         }
4753
4754         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4755             tcp_paws_discard(sk, skb)) {
4756                 if (!th->rst) {
4757                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4758                         tcp_send_dupack(sk, skb);
4759                         goto discard;
4760                 }
4761                 /* Reset is accepted even if it did not pass PAWS. */
4762         }
4763
4764         /* step 1: check sequence number */
4765         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4766                 if (!th->rst)
4767                         tcp_send_dupack(sk, skb);
4768                 goto discard;
4769         }
4770
4771         /* step 2: check RST bit */
4772         if (th->rst) {
4773                 tcp_reset(sk);
4774                 goto discard;
4775         }
4776
4777         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4778
4779         /* step 3: check security and precedence [ignored] */
4780
4781         /*      step 4:
4782          *
4783          *      Check for a SYN in window.
4784          */
4785         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4786                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4787                 tcp_reset(sk);
4788                 return 1;
4789         }
4790
4791         /* step 5: check the ACK field */
4792         if (th->ack) {
4793                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4794
4795                 switch (sk->sk_state) {
4796                 case TCP_SYN_RECV:
4797                         if (acceptable) {
4798                                 tp->copied_seq = tp->rcv_nxt;
4799                                 smp_mb();
4800                                 tcp_set_state(sk, TCP_ESTABLISHED);
4801                                 sk->sk_state_change(sk);
4802
4803                                 /* Note, that this wakeup is only for marginal
4804                                  * crossed SYN case. Passively open sockets
4805                                  * are not waked up, because sk->sk_sleep ==
4806                                  * NULL and sk->sk_socket == NULL.
4807                                  */
4808                                 if (sk->sk_socket) {
4809                                         sk_wake_async(sk,0,POLL_OUT);
4810                                 }
4811
4812                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4813                                 tp->snd_wnd = ntohs(th->window) <<
4814                                               tp->rx_opt.snd_wscale;
4815                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4816                                             TCP_SKB_CB(skb)->seq);
4817
4818                                 /* tcp_ack considers this ACK as duplicate
4819                                  * and does not calculate rtt.
4820                                  * Fix it at least with timestamps.
4821                                  */
4822                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4823                                     !tp->srtt)
4824                                         tcp_ack_saw_tstamp(sk, 0);
4825
4826                                 if (tp->rx_opt.tstamp_ok)
4827                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4828
4829                                 /* Make sure socket is routed, for
4830                                  * correct metrics.
4831                                  */
4832                                 icsk->icsk_af_ops->rebuild_header(sk);
4833
4834                                 tcp_init_metrics(sk);
4835
4836                                 tcp_init_congestion_control(sk);
4837
4838                                 /* Prevent spurious tcp_cwnd_restart() on
4839                                  * first data packet.
4840                                  */
4841                                 tp->lsndtime = tcp_time_stamp;
4842
4843                                 tcp_mtup_init(sk);
4844                                 tcp_initialize_rcv_mss(sk);
4845                                 tcp_init_buffer_space(sk);
4846                                 tcp_fast_path_on(tp);
4847                         } else {
4848                                 return 1;
4849                         }
4850                         break;
4851
4852                 case TCP_FIN_WAIT1:
4853                         if (tp->snd_una == tp->write_seq) {
4854                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4855                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4856                                 dst_confirm(sk->sk_dst_cache);
4857
4858                                 if (!sock_flag(sk, SOCK_DEAD))
4859                                         /* Wake up lingering close() */
4860                                         sk->sk_state_change(sk);
4861                                 else {
4862                                         int tmo;
4863
4864                                         if (tp->linger2 < 0 ||
4865                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4866                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4867                                                 tcp_done(sk);
4868                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4869                                                 return 1;
4870                                         }
4871
4872                                         tmo = tcp_fin_time(sk);
4873                                         if (tmo > TCP_TIMEWAIT_LEN) {
4874                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4875                                         } else if (th->fin || sock_owned_by_user(sk)) {
4876                                                 /* Bad case. We could lose such FIN otherwise.
4877                                                  * It is not a big problem, but it looks confusing
4878                                                  * and not so rare event. We still can lose it now,
4879                                                  * if it spins in bh_lock_sock(), but it is really
4880                                                  * marginal case.
4881                                                  */
4882                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4883                                         } else {
4884                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4885                                                 goto discard;
4886                                         }
4887                                 }
4888                         }
4889                         break;
4890
4891                 case TCP_CLOSING:
4892                         if (tp->snd_una == tp->write_seq) {
4893                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4894                                 goto discard;
4895                         }
4896                         break;
4897
4898                 case TCP_LAST_ACK:
4899                         if (tp->snd_una == tp->write_seq) {
4900                                 tcp_update_metrics(sk);
4901                                 tcp_done(sk);
4902                                 goto discard;
4903                         }
4904                         break;
4905                 }
4906         } else
4907                 goto discard;
4908
4909         /* step 6: check the URG bit */
4910         tcp_urg(sk, skb, th);
4911
4912         /* step 7: process the segment text */
4913         switch (sk->sk_state) {
4914         case TCP_CLOSE_WAIT:
4915         case TCP_CLOSING:
4916         case TCP_LAST_ACK:
4917                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4918                         break;
4919         case TCP_FIN_WAIT1:
4920         case TCP_FIN_WAIT2:
4921                 /* RFC 793 says to queue data in these states,
4922                  * RFC 1122 says we MUST send a reset.
4923                  * BSD 4.4 also does reset.
4924                  */
4925                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4926                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4927                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4928                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4929                                 tcp_reset(sk);
4930                                 return 1;
4931                         }
4932                 }
4933                 /* Fall through */
4934         case TCP_ESTABLISHED:
4935                 tcp_data_queue(sk, skb);
4936                 queued = 1;
4937                 break;
4938         }
4939
4940         /* tcp_data could move socket to TIME-WAIT */
4941         if (sk->sk_state != TCP_CLOSE) {
4942                 tcp_data_snd_check(sk);
4943                 tcp_ack_snd_check(sk);
4944         }
4945
4946         if (!queued) {
4947 discard:
4948                 __kfree_skb(skb);
4949         }
4950         return 0;
4951 }
4952
4953 EXPORT_SYMBOL(sysctl_tcp_ecn);
4954 EXPORT_SYMBOL(sysctl_tcp_reordering);
4955 EXPORT_SYMBOL(tcp_parse_options);
4956 EXPORT_SYMBOL(tcp_rcv_established);
4957 EXPORT_SYMBOL(tcp_rcv_state_process);
4958 EXPORT_SYMBOL(tcp_initialize_rcv_mss);