Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #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 /* Called to compute a smoothed rtt estimate. The data fed to this
559  * routine either comes from timestamps, or from segments that were
560  * known _not_ to have been retransmitted [see Karn/Partridge
561  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
562  * piece by Van Jacobson.
563  * NOTE: the next three routines used to be one big routine.
564  * To save cycles in the RFC 1323 implementation it was better to break
565  * it up into three procedures. -- erics
566  */
567 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
568 {
569         struct tcp_sock *tp = tcp_sk(sk);
570         long m = mrtt; /* RTT */
571
572         /*      The following amusing code comes from Jacobson's
573          *      article in SIGCOMM '88.  Note that rtt and mdev
574          *      are scaled versions of rtt and mean deviation.
575          *      This is designed to be as fast as possible
576          *      m stands for "measurement".
577          *
578          *      On a 1990 paper the rto value is changed to:
579          *      RTO = rtt + 4 * mdev
580          *
581          * Funny. This algorithm seems to be very broken.
582          * These formulae increase RTO, when it should be decreased, increase
583          * too slowly, when it should be increased quickly, decrease too quickly
584          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
585          * does not matter how to _calculate_ it. Seems, it was trap
586          * that VJ failed to avoid. 8)
587          */
588         if (m == 0)
589                 m = 1;
590         if (tp->srtt != 0) {
591                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
592                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
593                 if (m < 0) {
594                         m = -m;         /* m is now abs(error) */
595                         m -= (tp->mdev >> 2);   /* similar update on mdev */
596                         /* This is similar to one of Eifel findings.
597                          * Eifel blocks mdev updates when rtt decreases.
598                          * This solution is a bit different: we use finer gain
599                          * for mdev in this case (alpha*beta).
600                          * Like Eifel it also prevents growth of rto,
601                          * but also it limits too fast rto decreases,
602                          * happening in pure Eifel.
603                          */
604                         if (m > 0)
605                                 m >>= 3;
606                 } else {
607                         m -= (tp->mdev >> 2);   /* similar update on mdev */
608                 }
609                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
610                 if (tp->mdev > tp->mdev_max) {
611                         tp->mdev_max = tp->mdev;
612                         if (tp->mdev_max > tp->rttvar)
613                                 tp->rttvar = tp->mdev_max;
614                 }
615                 if (after(tp->snd_una, tp->rtt_seq)) {
616                         if (tp->mdev_max < tp->rttvar)
617                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
618                         tp->rtt_seq = tp->snd_nxt;
619                         tp->mdev_max = TCP_RTO_MIN;
620                 }
621         } else {
622                 /* no previous measure. */
623                 tp->srtt = m<<3;        /* take the measured time to be rtt */
624                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
625                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
626                 tp->rtt_seq = tp->snd_nxt;
627         }
628 }
629
630 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
631  * routine referred to above.
632  */
633 static inline void tcp_set_rto(struct sock *sk)
634 {
635         const struct tcp_sock *tp = tcp_sk(sk);
636         /* Old crap is replaced with new one. 8)
637          *
638          * More seriously:
639          * 1. If rtt variance happened to be less 50msec, it is hallucination.
640          *    It cannot be less due to utterly erratic ACK generation made
641          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
642          *    to do with delayed acks, because at cwnd>2 true delack timeout
643          *    is invisible. Actually, Linux-2.4 also generates erratic
644          *    ACKs in some circumstances.
645          */
646         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
647
648         /* 2. Fixups made earlier cannot be right.
649          *    If we do not estimate RTO correctly without them,
650          *    all the algo is pure shit and should be replaced
651          *    with correct one. It is exactly, which we pretend to do.
652          */
653 }
654
655 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
656  * guarantees that rto is higher.
657  */
658 static inline void tcp_bound_rto(struct sock *sk)
659 {
660         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
661                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
662 }
663
664 /* Save metrics learned by this TCP session.
665    This function is called only, when TCP finishes successfully
666    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
667  */
668 void tcp_update_metrics(struct sock *sk)
669 {
670         struct tcp_sock *tp = tcp_sk(sk);
671         struct dst_entry *dst = __sk_dst_get(sk);
672
673         if (sysctl_tcp_nometrics_save)
674                 return;
675
676         dst_confirm(dst);
677
678         if (dst && (dst->flags&DST_HOST)) {
679                 const struct inet_connection_sock *icsk = inet_csk(sk);
680                 int m;
681
682                 if (icsk->icsk_backoff || !tp->srtt) {
683                         /* This session failed to estimate rtt. Why?
684                          * Probably, no packets returned in time.
685                          * Reset our results.
686                          */
687                         if (!(dst_metric_locked(dst, RTAX_RTT)))
688                                 dst->metrics[RTAX_RTT-1] = 0;
689                         return;
690                 }
691
692                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
693
694                 /* If newly calculated rtt larger than stored one,
695                  * store new one. Otherwise, use EWMA. Remember,
696                  * rtt overestimation is always better than underestimation.
697                  */
698                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
699                         if (m <= 0)
700                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
701                         else
702                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
703                 }
704
705                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
706                         if (m < 0)
707                                 m = -m;
708
709                         /* Scale deviation to rttvar fixed point */
710                         m >>= 1;
711                         if (m < tp->mdev)
712                                 m = tp->mdev;
713
714                         if (m >= dst_metric(dst, RTAX_RTTVAR))
715                                 dst->metrics[RTAX_RTTVAR-1] = m;
716                         else
717                                 dst->metrics[RTAX_RTTVAR-1] -=
718                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
719                 }
720
721                 if (tp->snd_ssthresh >= 0xFFFF) {
722                         /* Slow start still did not finish. */
723                         if (dst_metric(dst, RTAX_SSTHRESH) &&
724                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
725                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
726                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
727                         if (!dst_metric_locked(dst, RTAX_CWND) &&
728                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
729                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
730                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
731                            icsk->icsk_ca_state == TCP_CA_Open) {
732                         /* Cong. avoidance phase, cwnd is reliable. */
733                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
734                                 dst->metrics[RTAX_SSTHRESH-1] =
735                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
736                         if (!dst_metric_locked(dst, RTAX_CWND))
737                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
738                 } else {
739                         /* Else slow start did not finish, cwnd is non-sense,
740                            ssthresh may be also invalid.
741                          */
742                         if (!dst_metric_locked(dst, RTAX_CWND))
743                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
744                         if (dst->metrics[RTAX_SSTHRESH-1] &&
745                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
746                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
747                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
748                 }
749
750                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
751                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
752                             tp->reordering != sysctl_tcp_reordering)
753                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
754                 }
755         }
756 }
757
758 /* Numbers are taken from RFC3390.
759  *
760  * John Heffner states:
761  *
762  *      The RFC specifies a window of no more than 4380 bytes
763  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
764  *      is a bit misleading because they use a clamp at 4380 bytes
765  *      rather than use a multiplier in the relevant range.
766  */
767 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
768 {
769         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
770
771         if (!cwnd) {
772                 if (tp->mss_cache > 1460)
773                         cwnd = 2;
774                 else
775                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
776         }
777         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
778 }
779
780 /* Set slow start threshold and cwnd not falling to slow start */
781 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
782 {
783         struct tcp_sock *tp = tcp_sk(sk);
784         const struct inet_connection_sock *icsk = inet_csk(sk);
785
786         tp->prior_ssthresh = 0;
787         tp->bytes_acked = 0;
788         if (icsk->icsk_ca_state < TCP_CA_CWR) {
789                 tp->undo_marker = 0;
790                 if (set_ssthresh)
791                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
792                 tp->snd_cwnd = min(tp->snd_cwnd,
793                                    tcp_packets_in_flight(tp) + 1U);
794                 tp->snd_cwnd_cnt = 0;
795                 tp->high_seq = tp->snd_nxt;
796                 tp->snd_cwnd_stamp = tcp_time_stamp;
797                 TCP_ECN_queue_cwr(tp);
798
799                 tcp_set_ca_state(sk, TCP_CA_CWR);
800         }
801 }
802
803 /* Initialize metrics on socket. */
804
805 static void tcp_init_metrics(struct sock *sk)
806 {
807         struct tcp_sock *tp = tcp_sk(sk);
808         struct dst_entry *dst = __sk_dst_get(sk);
809
810         if (dst == NULL)
811                 goto reset;
812
813         dst_confirm(dst);
814
815         if (dst_metric_locked(dst, RTAX_CWND))
816                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
817         if (dst_metric(dst, RTAX_SSTHRESH)) {
818                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
819                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
820                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
821         }
822         if (dst_metric(dst, RTAX_REORDERING) &&
823             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
824                 tp->rx_opt.sack_ok &= ~2;
825                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
826         }
827
828         if (dst_metric(dst, RTAX_RTT) == 0)
829                 goto reset;
830
831         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
832                 goto reset;
833
834         /* Initial rtt is determined from SYN,SYN-ACK.
835          * The segment is small and rtt may appear much
836          * less than real one. Use per-dst memory
837          * to make it more realistic.
838          *
839          * A bit of theory. RTT is time passed after "normal" sized packet
840          * is sent until it is ACKed. In normal circumstances sending small
841          * packets force peer to delay ACKs and calculation is correct too.
842          * The algorithm is adaptive and, provided we follow specs, it
843          * NEVER underestimate RTT. BUT! If peer tries to make some clever
844          * tricks sort of "quick acks" for time long enough to decrease RTT
845          * to low value, and then abruptly stops to do it and starts to delay
846          * ACKs, wait for troubles.
847          */
848         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
849                 tp->srtt = dst_metric(dst, RTAX_RTT);
850                 tp->rtt_seq = tp->snd_nxt;
851         }
852         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
853                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
854                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
855         }
856         tcp_set_rto(sk);
857         tcp_bound_rto(sk);
858         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
859                 goto reset;
860         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
861         tp->snd_cwnd_stamp = tcp_time_stamp;
862         return;
863
864 reset:
865         /* Play conservative. If timestamps are not
866          * supported, TCP will fail to recalculate correct
867          * rtt, if initial rto is too small. FORGET ALL AND RESET!
868          */
869         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
870                 tp->srtt = 0;
871                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
872                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
873         }
874 }
875
876 static void tcp_update_reordering(struct sock *sk, const int metric,
877                                   const int ts)
878 {
879         struct tcp_sock *tp = tcp_sk(sk);
880         if (metric > tp->reordering) {
881                 tp->reordering = min(TCP_MAX_REORDERING, metric);
882
883                 /* This exciting event is worth to be remembered. 8) */
884                 if (ts)
885                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
886                 else if (IsReno(tp))
887                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
888                 else if (IsFack(tp))
889                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
890                 else
891                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
892 #if FASTRETRANS_DEBUG > 1
893                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
894                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
895                        tp->reordering,
896                        tp->fackets_out,
897                        tp->sacked_out,
898                        tp->undo_marker ? tp->undo_retrans : 0);
899 #endif
900                 /* Disable FACK yet. */
901                 tp->rx_opt.sack_ok &= ~2;
902         }
903 }
904
905 /* This procedure tags the retransmission queue when SACKs arrive.
906  *
907  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
908  * Packets in queue with these bits set are counted in variables
909  * sacked_out, retrans_out and lost_out, correspondingly.
910  *
911  * Valid combinations are:
912  * Tag  InFlight        Description
913  * 0    1               - orig segment is in flight.
914  * S    0               - nothing flies, orig reached receiver.
915  * L    0               - nothing flies, orig lost by net.
916  * R    2               - both orig and retransmit are in flight.
917  * L|R  1               - orig is lost, retransmit is in flight.
918  * S|R  1               - orig reached receiver, retrans is still in flight.
919  * (L|S|R is logically valid, it could occur when L|R is sacked,
920  *  but it is equivalent to plain S and code short-curcuits it to S.
921  *  L|S is logically invalid, it would mean -1 packet in flight 8))
922  *
923  * These 6 states form finite state machine, controlled by the following events:
924  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
925  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
926  * 3. Loss detection event of one of three flavors:
927  *      A. Scoreboard estimator decided the packet is lost.
928  *         A'. Reno "three dupacks" marks head of queue lost.
929  *         A''. Its FACK modfication, head until snd.fack is lost.
930  *      B. SACK arrives sacking data transmitted after never retransmitted
931  *         hole was sent out.
932  *      C. SACK arrives sacking SND.NXT at the moment, when the
933  *         segment was retransmitted.
934  * 4. D-SACK added new rule: D-SACK changes any tag to S.
935  *
936  * It is pleasant to note, that state diagram turns out to be commutative,
937  * so that we are allowed not to be bothered by order of our actions,
938  * when multiple events arrive simultaneously. (see the function below).
939  *
940  * Reordering detection.
941  * --------------------
942  * Reordering metric is maximal distance, which a packet can be displaced
943  * in packet stream. With SACKs we can estimate it:
944  *
945  * 1. SACK fills old hole and the corresponding segment was not
946  *    ever retransmitted -> reordering. Alas, we cannot use it
947  *    when segment was retransmitted.
948  * 2. The last flaw is solved with D-SACK. D-SACK arrives
949  *    for retransmitted and already SACKed segment -> reordering..
950  * Both of these heuristics are not used in Loss state, when we cannot
951  * account for retransmits accurately.
952  */
953 static int
954 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
955 {
956         const struct inet_connection_sock *icsk = inet_csk(sk);
957         struct tcp_sock *tp = tcp_sk(sk);
958         unsigned char *ptr = (skb_transport_header(ack_skb) +
959                               TCP_SKB_CB(ack_skb)->sacked);
960         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
961         struct sk_buff *cached_skb;
962         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
963         int reord = tp->packets_out;
964         int prior_fackets;
965         u32 lost_retrans = 0;
966         int flag = 0;
967         int found_dup_sack = 0;
968         int cached_fack_count;
969         int i;
970         int first_sack_index;
971
972         if (!tp->sacked_out)
973                 tp->fackets_out = 0;
974         prior_fackets = tp->fackets_out;
975
976         /* Check for D-SACK. */
977         if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
978                 flag |= FLAG_DSACKING_ACK;
979                 found_dup_sack = 1;
980                 tp->rx_opt.sack_ok |= 4;
981                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
982         } else if (num_sacks > 1 &&
983                         !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
984                         !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
985                 flag |= FLAG_DSACKING_ACK;
986                 found_dup_sack = 1;
987                 tp->rx_opt.sack_ok |= 4;
988                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
989         }
990
991         /* D-SACK for already forgotten data...
992          * Do dumb counting. */
993         if (found_dup_sack &&
994                         !after(ntohl(sp[0].end_seq), prior_snd_una) &&
995                         after(ntohl(sp[0].end_seq), tp->undo_marker))
996                 tp->undo_retrans--;
997
998         /* Eliminate too old ACKs, but take into
999          * account more or less fresh ones, they can
1000          * contain valid SACK info.
1001          */
1002         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1003                 return 0;
1004
1005         /* SACK fastpath:
1006          * if the only SACK change is the increase of the end_seq of
1007          * the first block then only apply that SACK block
1008          * and use retrans queue hinting otherwise slowpath */
1009         flag = 1;
1010         for (i = 0; i < num_sacks; i++) {
1011                 __be32 start_seq = sp[i].start_seq;
1012                 __be32 end_seq = sp[i].end_seq;
1013
1014                 if (i == 0) {
1015                         if (tp->recv_sack_cache[i].start_seq != start_seq)
1016                                 flag = 0;
1017                 } else {
1018                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1019                             (tp->recv_sack_cache[i].end_seq != end_seq))
1020                                 flag = 0;
1021                 }
1022                 tp->recv_sack_cache[i].start_seq = start_seq;
1023                 tp->recv_sack_cache[i].end_seq = end_seq;
1024         }
1025         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1026         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1027                 tp->recv_sack_cache[i].start_seq = 0;
1028                 tp->recv_sack_cache[i].end_seq = 0;
1029         }
1030
1031         first_sack_index = 0;
1032         if (flag)
1033                 num_sacks = 1;
1034         else {
1035                 int j;
1036                 tp->fastpath_skb_hint = NULL;
1037
1038                 /* order SACK blocks to allow in order walk of the retrans queue */
1039                 for (i = num_sacks-1; i > 0; i--) {
1040                         for (j = 0; j < i; j++){
1041                                 if (after(ntohl(sp[j].start_seq),
1042                                           ntohl(sp[j+1].start_seq))){
1043                                         struct tcp_sack_block_wire tmp;
1044
1045                                         tmp = sp[j];
1046                                         sp[j] = sp[j+1];
1047                                         sp[j+1] = tmp;
1048
1049                                         /* Track where the first SACK block goes to */
1050                                         if (j == first_sack_index)
1051                                                 first_sack_index = j+1;
1052                                 }
1053
1054                         }
1055                 }
1056         }
1057
1058         /* clear flag as used for different purpose in following code */
1059         flag = 0;
1060
1061         /* Use SACK fastpath hint if valid */
1062         cached_skb = tp->fastpath_skb_hint;
1063         cached_fack_count = tp->fastpath_cnt_hint;
1064         if (!cached_skb) {
1065                 cached_skb = tcp_write_queue_head(sk);
1066                 cached_fack_count = 0;
1067         }
1068
1069         for (i=0; i<num_sacks; i++, sp++) {
1070                 struct sk_buff *skb;
1071                 __u32 start_seq = ntohl(sp->start_seq);
1072                 __u32 end_seq = ntohl(sp->end_seq);
1073                 int fack_count;
1074                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1075
1076                 skb = cached_skb;
1077                 fack_count = cached_fack_count;
1078
1079                 /* Event "B" in the comment above. */
1080                 if (after(end_seq, tp->high_seq))
1081                         flag |= FLAG_DATA_LOST;
1082
1083                 tcp_for_write_queue_from(skb, sk) {
1084                         int in_sack, pcount;
1085                         u8 sacked;
1086
1087                         if (skb == tcp_send_head(sk))
1088                                 break;
1089
1090                         cached_skb = skb;
1091                         cached_fack_count = fack_count;
1092                         if (i == first_sack_index) {
1093                                 tp->fastpath_skb_hint = skb;
1094                                 tp->fastpath_cnt_hint = fack_count;
1095                         }
1096
1097                         /* The retransmission queue is always in order, so
1098                          * we can short-circuit the walk early.
1099                          */
1100                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1101                                 break;
1102
1103                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1104                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1105
1106                         pcount = tcp_skb_pcount(skb);
1107
1108                         if (pcount > 1 && !in_sack &&
1109                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1110                                 unsigned int pkt_len;
1111
1112                                 in_sack = !after(start_seq,
1113                                                  TCP_SKB_CB(skb)->seq);
1114
1115                                 if (!in_sack)
1116                                         pkt_len = (start_seq -
1117                                                    TCP_SKB_CB(skb)->seq);
1118                                 else
1119                                         pkt_len = (end_seq -
1120                                                    TCP_SKB_CB(skb)->seq);
1121                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1122                                         break;
1123                                 pcount = tcp_skb_pcount(skb);
1124                         }
1125
1126                         fack_count += pcount;
1127
1128                         sacked = TCP_SKB_CB(skb)->sacked;
1129
1130                         /* Account D-SACK for retransmitted packet. */
1131                         if ((dup_sack && in_sack) &&
1132                             (sacked & TCPCB_RETRANS) &&
1133                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1134                                 tp->undo_retrans--;
1135
1136                         /* The frame is ACKed. */
1137                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1138                                 if (sacked&TCPCB_RETRANS) {
1139                                         if ((dup_sack && in_sack) &&
1140                                             (sacked&TCPCB_SACKED_ACKED))
1141                                                 reord = min(fack_count, reord);
1142                                 } else {
1143                                         /* If it was in a hole, we detected reordering. */
1144                                         if (fack_count < prior_fackets &&
1145                                             !(sacked&TCPCB_SACKED_ACKED))
1146                                                 reord = min(fack_count, reord);
1147                                 }
1148
1149                                 /* Nothing to do; acked frame is about to be dropped. */
1150                                 continue;
1151                         }
1152
1153                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1154                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1155                             (!lost_retrans || after(end_seq, lost_retrans)))
1156                                 lost_retrans = end_seq;
1157
1158                         if (!in_sack)
1159                                 continue;
1160
1161                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1162                                 if (sacked & TCPCB_SACKED_RETRANS) {
1163                                         /* If the segment is not tagged as lost,
1164                                          * we do not clear RETRANS, believing
1165                                          * that retransmission is still in flight.
1166                                          */
1167                                         if (sacked & TCPCB_LOST) {
1168                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1169                                                 tp->lost_out -= tcp_skb_pcount(skb);
1170                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1171
1172                                                 /* clear lost hint */
1173                                                 tp->retransmit_skb_hint = NULL;
1174                                         }
1175                                 } else {
1176                                         /* New sack for not retransmitted frame,
1177                                          * which was in hole. It is reordering.
1178                                          */
1179                                         if (!(sacked & TCPCB_RETRANS) &&
1180                                             fack_count < prior_fackets)
1181                                                 reord = min(fack_count, reord);
1182
1183                                         if (sacked & TCPCB_LOST) {
1184                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1185                                                 tp->lost_out -= tcp_skb_pcount(skb);
1186
1187                                                 /* clear lost hint */
1188                                                 tp->retransmit_skb_hint = NULL;
1189                                         }
1190                                         /* SACK enhanced F-RTO detection.
1191                                          * Set flag if and only if non-rexmitted
1192                                          * segments below frto_highmark are
1193                                          * SACKed (RFC4138; Appendix B).
1194                                          * Clearing correct due to in-order walk
1195                                          */
1196                                         if (after(end_seq, tp->frto_highmark)) {
1197                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1198                                         } else {
1199                                                 if (!(sacked & TCPCB_RETRANS))
1200                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1201                                         }
1202                                 }
1203
1204                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1205                                 flag |= FLAG_DATA_SACKED;
1206                                 tp->sacked_out += tcp_skb_pcount(skb);
1207
1208                                 if (fack_count > tp->fackets_out)
1209                                         tp->fackets_out = fack_count;
1210                         } else {
1211                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1212                                         reord = min(fack_count, reord);
1213                         }
1214
1215                         /* D-SACK. We can detect redundant retransmission
1216                          * in S|R and plain R frames and clear it.
1217                          * undo_retrans is decreased above, L|R frames
1218                          * are accounted above as well.
1219                          */
1220                         if (dup_sack &&
1221                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1222                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1223                                 tp->retrans_out -= tcp_skb_pcount(skb);
1224                                 tp->retransmit_skb_hint = NULL;
1225                         }
1226                 }
1227         }
1228
1229         /* Check for lost retransmit. This superb idea is
1230          * borrowed from "ratehalving". Event "C".
1231          * Later note: FACK people cheated me again 8),
1232          * we have to account for reordering! Ugly,
1233          * but should help.
1234          */
1235         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1236                 struct sk_buff *skb;
1237
1238                 tcp_for_write_queue(skb, sk) {
1239                         if (skb == tcp_send_head(sk))
1240                                 break;
1241                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1242                                 break;
1243                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1244                                 continue;
1245                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1246                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1247                             (IsFack(tp) ||
1248                              !before(lost_retrans,
1249                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1250                                      tp->mss_cache))) {
1251                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1252                                 tp->retrans_out -= tcp_skb_pcount(skb);
1253
1254                                 /* clear lost hint */
1255                                 tp->retransmit_skb_hint = NULL;
1256
1257                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1258                                         tp->lost_out += tcp_skb_pcount(skb);
1259                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1260                                         flag |= FLAG_DATA_SACKED;
1261                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1262                                 }
1263                         }
1264                 }
1265         }
1266
1267         tp->left_out = tp->sacked_out + tp->lost_out;
1268
1269         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1270             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1271                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1272
1273 #if FASTRETRANS_DEBUG > 0
1274         BUG_TRAP((int)tp->sacked_out >= 0);
1275         BUG_TRAP((int)tp->lost_out >= 0);
1276         BUG_TRAP((int)tp->retrans_out >= 0);
1277         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1278 #endif
1279         return flag;
1280 }
1281
1282 /* F-RTO can only be used if TCP has never retransmitted anything other than
1283  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1284  */
1285 int tcp_use_frto(struct sock *sk)
1286 {
1287         const struct tcp_sock *tp = tcp_sk(sk);
1288         struct sk_buff *skb;
1289
1290         if (!sysctl_tcp_frto)
1291                 return 0;
1292
1293         if (IsSackFrto())
1294                 return 1;
1295
1296         /* Avoid expensive walking of rexmit queue if possible */
1297         if (tp->retrans_out > 1)
1298                 return 0;
1299
1300         skb = tcp_write_queue_head(sk);
1301         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1302         tcp_for_write_queue_from(skb, sk) {
1303                 if (skb == tcp_send_head(sk))
1304                         break;
1305                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1306                         return 0;
1307                 /* Short-circuit when first non-SACKed skb has been checked */
1308                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1309                         break;
1310         }
1311         return 1;
1312 }
1313
1314 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1315  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1316  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1317  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1318  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1319  * bits are handled if the Loss state is really to be entered (in
1320  * tcp_enter_frto_loss).
1321  *
1322  * Do like tcp_enter_loss() would; when RTO expires the second time it
1323  * does:
1324  *  "Reduce ssthresh if it has not yet been made inside this window."
1325  */
1326 void tcp_enter_frto(struct sock *sk)
1327 {
1328         const struct inet_connection_sock *icsk = inet_csk(sk);
1329         struct tcp_sock *tp = tcp_sk(sk);
1330         struct sk_buff *skb;
1331
1332         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1333             tp->snd_una == tp->high_seq ||
1334             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1335              !icsk->icsk_retransmits)) {
1336                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1337                 /* Our state is too optimistic in ssthresh() call because cwnd
1338                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1339                  * recovery has not yet completed. Pattern would be this: RTO,
1340                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1341                  * up here twice).
1342                  * RFC4138 should be more specific on what to do, even though
1343                  * RTO is quite unlikely to occur after the first Cumulative ACK
1344                  * due to back-off and complexity of triggering events ...
1345                  */
1346                 if (tp->frto_counter) {
1347                         u32 stored_cwnd;
1348                         stored_cwnd = tp->snd_cwnd;
1349                         tp->snd_cwnd = 2;
1350                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1351                         tp->snd_cwnd = stored_cwnd;
1352                 } else {
1353                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1354                 }
1355                 /* ... in theory, cong.control module could do "any tricks" in
1356                  * ssthresh(), which means that ca_state, lost bits and lost_out
1357                  * counter would have to be faked before the call occurs. We
1358                  * consider that too expensive, unlikely and hacky, so modules
1359                  * using these in ssthresh() must deal these incompatibility
1360                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1361                  */
1362                 tcp_ca_event(sk, CA_EVENT_FRTO);
1363         }
1364
1365         tp->undo_marker = tp->snd_una;
1366         tp->undo_retrans = 0;
1367
1368         skb = tcp_write_queue_head(sk);
1369         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1370                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1371                 tp->retrans_out -= tcp_skb_pcount(skb);
1372         }
1373         tcp_sync_left_out(tp);
1374
1375         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1376          * The last condition is necessary at least in tp->frto_counter case.
1377          */
1378         if (IsSackFrto() && (tp->frto_counter ||
1379             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1380             after(tp->high_seq, tp->snd_una)) {
1381                 tp->frto_highmark = tp->high_seq;
1382         } else {
1383                 tp->frto_highmark = tp->snd_nxt;
1384         }
1385         tcp_set_ca_state(sk, TCP_CA_Disorder);
1386         tp->high_seq = tp->snd_nxt;
1387         tp->frto_counter = 1;
1388 }
1389
1390 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1391  * which indicates that we should follow the traditional RTO recovery,
1392  * i.e. mark everything lost and do go-back-N retransmission.
1393  */
1394 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1395 {
1396         struct tcp_sock *tp = tcp_sk(sk);
1397         struct sk_buff *skb;
1398         int cnt = 0;
1399
1400         tp->sacked_out = 0;
1401         tp->lost_out = 0;
1402         tp->fackets_out = 0;
1403         tp->retrans_out = 0;
1404
1405         tcp_for_write_queue(skb, sk) {
1406                 if (skb == tcp_send_head(sk))
1407                         break;
1408                 cnt += tcp_skb_pcount(skb);
1409                 /*
1410                  * Count the retransmission made on RTO correctly (only when
1411                  * waiting for the first ACK and did not get it)...
1412                  */
1413                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1414                         /* For some reason this R-bit might get cleared? */
1415                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1416                                 tp->retrans_out += tcp_skb_pcount(skb);
1417                         /* ...enter this if branch just for the first segment */
1418                         flag |= FLAG_DATA_ACKED;
1419                 } else {
1420                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1421                 }
1422                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1423
1424                         /* Do not mark those segments lost that were
1425                          * forward transmitted after RTO
1426                          */
1427                         if (!after(TCP_SKB_CB(skb)->end_seq,
1428                                    tp->frto_highmark)) {
1429                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1430                                 tp->lost_out += tcp_skb_pcount(skb);
1431                         }
1432                 } else {
1433                         tp->sacked_out += tcp_skb_pcount(skb);
1434                         tp->fackets_out = cnt;
1435                 }
1436         }
1437         tcp_sync_left_out(tp);
1438
1439         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1440         tp->snd_cwnd_cnt = 0;
1441         tp->snd_cwnd_stamp = tcp_time_stamp;
1442         tp->undo_marker = 0;
1443         tp->frto_counter = 0;
1444
1445         tp->reordering = min_t(unsigned int, tp->reordering,
1446                                              sysctl_tcp_reordering);
1447         tcp_set_ca_state(sk, TCP_CA_Loss);
1448         tp->high_seq = tp->frto_highmark;
1449         TCP_ECN_queue_cwr(tp);
1450
1451         clear_all_retrans_hints(tp);
1452 }
1453
1454 void tcp_clear_retrans(struct tcp_sock *tp)
1455 {
1456         tp->left_out = 0;
1457         tp->retrans_out = 0;
1458
1459         tp->fackets_out = 0;
1460         tp->sacked_out = 0;
1461         tp->lost_out = 0;
1462
1463         tp->undo_marker = 0;
1464         tp->undo_retrans = 0;
1465 }
1466
1467 /* Enter Loss state. If "how" is not zero, forget all SACK information
1468  * and reset tags completely, otherwise preserve SACKs. If receiver
1469  * dropped its ofo queue, we will know this due to reneging detection.
1470  */
1471 void tcp_enter_loss(struct sock *sk, int how)
1472 {
1473         const struct inet_connection_sock *icsk = inet_csk(sk);
1474         struct tcp_sock *tp = tcp_sk(sk);
1475         struct sk_buff *skb;
1476         int cnt = 0;
1477
1478         /* Reduce ssthresh if it has not yet been made inside this window. */
1479         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1480             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1481                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1482                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1483                 tcp_ca_event(sk, CA_EVENT_LOSS);
1484         }
1485         tp->snd_cwnd       = 1;
1486         tp->snd_cwnd_cnt   = 0;
1487         tp->snd_cwnd_stamp = tcp_time_stamp;
1488
1489         tp->bytes_acked = 0;
1490         tcp_clear_retrans(tp);
1491
1492         /* Push undo marker, if it was plain RTO and nothing
1493          * was retransmitted. */
1494         if (!how)
1495                 tp->undo_marker = tp->snd_una;
1496
1497         tcp_for_write_queue(skb, sk) {
1498                 if (skb == tcp_send_head(sk))
1499                         break;
1500                 cnt += tcp_skb_pcount(skb);
1501                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1502                         tp->undo_marker = 0;
1503                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1504                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1505                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1506                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1507                         tp->lost_out += tcp_skb_pcount(skb);
1508                 } else {
1509                         tp->sacked_out += tcp_skb_pcount(skb);
1510                         tp->fackets_out = cnt;
1511                 }
1512         }
1513         tcp_sync_left_out(tp);
1514
1515         tp->reordering = min_t(unsigned int, tp->reordering,
1516                                              sysctl_tcp_reordering);
1517         tcp_set_ca_state(sk, TCP_CA_Loss);
1518         tp->high_seq = tp->snd_nxt;
1519         TCP_ECN_queue_cwr(tp);
1520         /* Abort FRTO algorithm if one is in progress */
1521         tp->frto_counter = 0;
1522
1523         clear_all_retrans_hints(tp);
1524 }
1525
1526 static int tcp_check_sack_reneging(struct sock *sk)
1527 {
1528         struct sk_buff *skb;
1529
1530         /* If ACK arrived pointing to a remembered SACK,
1531          * it means that our remembered SACKs do not reflect
1532          * real state of receiver i.e.
1533          * receiver _host_ is heavily congested (or buggy).
1534          * Do processing similar to RTO timeout.
1535          */
1536         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1537             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1538                 struct inet_connection_sock *icsk = inet_csk(sk);
1539                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1540
1541                 tcp_enter_loss(sk, 1);
1542                 icsk->icsk_retransmits++;
1543                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1544                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1545                                           icsk->icsk_rto, TCP_RTO_MAX);
1546                 return 1;
1547         }
1548         return 0;
1549 }
1550
1551 static inline int tcp_fackets_out(struct tcp_sock *tp)
1552 {
1553         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1554 }
1555
1556 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1557 {
1558         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1559 }
1560
1561 static inline int tcp_head_timedout(struct sock *sk)
1562 {
1563         struct tcp_sock *tp = tcp_sk(sk);
1564
1565         return tp->packets_out &&
1566                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1567 }
1568
1569 /* Linux NewReno/SACK/FACK/ECN state machine.
1570  * --------------------------------------
1571  *
1572  * "Open"       Normal state, no dubious events, fast path.
1573  * "Disorder"   In all the respects it is "Open",
1574  *              but requires a bit more attention. It is entered when
1575  *              we see some SACKs or dupacks. It is split of "Open"
1576  *              mainly to move some processing from fast path to slow one.
1577  * "CWR"        CWND was reduced due to some Congestion Notification event.
1578  *              It can be ECN, ICMP source quench, local device congestion.
1579  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1580  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1581  *
1582  * tcp_fastretrans_alert() is entered:
1583  * - each incoming ACK, if state is not "Open"
1584  * - when arrived ACK is unusual, namely:
1585  *      * SACK
1586  *      * Duplicate ACK.
1587  *      * ECN ECE.
1588  *
1589  * Counting packets in flight is pretty simple.
1590  *
1591  *      in_flight = packets_out - left_out + retrans_out
1592  *
1593  *      packets_out is SND.NXT-SND.UNA counted in packets.
1594  *
1595  *      retrans_out is number of retransmitted segments.
1596  *
1597  *      left_out is number of segments left network, but not ACKed yet.
1598  *
1599  *              left_out = sacked_out + lost_out
1600  *
1601  *     sacked_out: Packets, which arrived to receiver out of order
1602  *                 and hence not ACKed. With SACKs this number is simply
1603  *                 amount of SACKed data. Even without SACKs
1604  *                 it is easy to give pretty reliable estimate of this number,
1605  *                 counting duplicate ACKs.
1606  *
1607  *       lost_out: Packets lost by network. TCP has no explicit
1608  *                 "loss notification" feedback from network (for now).
1609  *                 It means that this number can be only _guessed_.
1610  *                 Actually, it is the heuristics to predict lossage that
1611  *                 distinguishes different algorithms.
1612  *
1613  *      F.e. after RTO, when all the queue is considered as lost,
1614  *      lost_out = packets_out and in_flight = retrans_out.
1615  *
1616  *              Essentially, we have now two algorithms counting
1617  *              lost packets.
1618  *
1619  *              FACK: It is the simplest heuristics. As soon as we decided
1620  *              that something is lost, we decide that _all_ not SACKed
1621  *              packets until the most forward SACK are lost. I.e.
1622  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1623  *              It is absolutely correct estimate, if network does not reorder
1624  *              packets. And it loses any connection to reality when reordering
1625  *              takes place. We use FACK by default until reordering
1626  *              is suspected on the path to this destination.
1627  *
1628  *              NewReno: when Recovery is entered, we assume that one segment
1629  *              is lost (classic Reno). While we are in Recovery and
1630  *              a partial ACK arrives, we assume that one more packet
1631  *              is lost (NewReno). This heuristics are the same in NewReno
1632  *              and SACK.
1633  *
1634  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1635  *  deflation etc. CWND is real congestion window, never inflated, changes
1636  *  only according to classic VJ rules.
1637  *
1638  * Really tricky (and requiring careful tuning) part of algorithm
1639  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1640  * The first determines the moment _when_ we should reduce CWND and,
1641  * hence, slow down forward transmission. In fact, it determines the moment
1642  * when we decide that hole is caused by loss, rather than by a reorder.
1643  *
1644  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1645  * holes, caused by lost packets.
1646  *
1647  * And the most logically complicated part of algorithm is undo
1648  * heuristics. We detect false retransmits due to both too early
1649  * fast retransmit (reordering) and underestimated RTO, analyzing
1650  * timestamps and D-SACKs. When we detect that some segments were
1651  * retransmitted by mistake and CWND reduction was wrong, we undo
1652  * window reduction and abort recovery phase. This logic is hidden
1653  * inside several functions named tcp_try_undo_<something>.
1654  */
1655
1656 /* This function decides, when we should leave Disordered state
1657  * and enter Recovery phase, reducing congestion window.
1658  *
1659  * Main question: may we further continue forward transmission
1660  * with the same cwnd?
1661  */
1662 static int tcp_time_to_recover(struct sock *sk)
1663 {
1664         struct tcp_sock *tp = tcp_sk(sk);
1665         __u32 packets_out;
1666
1667         /* Do not perform any recovery during FRTO algorithm */
1668         if (tp->frto_counter)
1669                 return 0;
1670
1671         /* Trick#1: The loss is proven. */
1672         if (tp->lost_out)
1673                 return 1;
1674
1675         /* Not-A-Trick#2 : Classic rule... */
1676         if (tcp_fackets_out(tp) > tp->reordering)
1677                 return 1;
1678
1679         /* Trick#3 : when we use RFC2988 timer restart, fast
1680          * retransmit can be triggered by timeout of queue head.
1681          */
1682         if (tcp_head_timedout(sk))
1683                 return 1;
1684
1685         /* Trick#4: It is still not OK... But will it be useful to delay
1686          * recovery more?
1687          */
1688         packets_out = tp->packets_out;
1689         if (packets_out <= tp->reordering &&
1690             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1691             !tcp_may_send_now(sk)) {
1692                 /* We have nothing to send. This connection is limited
1693                  * either by receiver window or by application.
1694                  */
1695                 return 1;
1696         }
1697
1698         return 0;
1699 }
1700
1701 /* If we receive more dupacks than we expected counting segments
1702  * in assumption of absent reordering, interpret this as reordering.
1703  * The only another reason could be bug in receiver TCP.
1704  */
1705 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1706 {
1707         struct tcp_sock *tp = tcp_sk(sk);
1708         u32 holes;
1709
1710         holes = max(tp->lost_out, 1U);
1711         holes = min(holes, tp->packets_out);
1712
1713         if ((tp->sacked_out + holes) > tp->packets_out) {
1714                 tp->sacked_out = tp->packets_out - holes;
1715                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1716         }
1717 }
1718
1719 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1720
1721 static void tcp_add_reno_sack(struct sock *sk)
1722 {
1723         struct tcp_sock *tp = tcp_sk(sk);
1724         tp->sacked_out++;
1725         tcp_check_reno_reordering(sk, 0);
1726         tcp_sync_left_out(tp);
1727 }
1728
1729 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1730
1731 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1732 {
1733         struct tcp_sock *tp = tcp_sk(sk);
1734
1735         if (acked > 0) {
1736                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1737                 if (acked-1 >= tp->sacked_out)
1738                         tp->sacked_out = 0;
1739                 else
1740                         tp->sacked_out -= acked-1;
1741         }
1742         tcp_check_reno_reordering(sk, acked);
1743         tcp_sync_left_out(tp);
1744 }
1745
1746 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1747 {
1748         tp->sacked_out = 0;
1749         tp->left_out = tp->lost_out;
1750 }
1751
1752 /* Mark head of queue up as lost. */
1753 static void tcp_mark_head_lost(struct sock *sk,
1754                                int packets, u32 high_seq)
1755 {
1756         struct tcp_sock *tp = tcp_sk(sk);
1757         struct sk_buff *skb;
1758         int cnt;
1759
1760         BUG_TRAP(packets <= tp->packets_out);
1761         if (tp->lost_skb_hint) {
1762                 skb = tp->lost_skb_hint;
1763                 cnt = tp->lost_cnt_hint;
1764         } else {
1765                 skb = tcp_write_queue_head(sk);
1766                 cnt = 0;
1767         }
1768
1769         tcp_for_write_queue_from(skb, sk) {
1770                 if (skb == tcp_send_head(sk))
1771                         break;
1772                 /* TODO: do this better */
1773                 /* this is not the most efficient way to do this... */
1774                 tp->lost_skb_hint = skb;
1775                 tp->lost_cnt_hint = cnt;
1776                 cnt += tcp_skb_pcount(skb);
1777                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1778                         break;
1779                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1780                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1781                         tp->lost_out += tcp_skb_pcount(skb);
1782
1783                         /* clear xmit_retransmit_queue hints
1784                          *  if this is beyond hint */
1785                         if (tp->retransmit_skb_hint != NULL &&
1786                             before(TCP_SKB_CB(skb)->seq,
1787                                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1788                                 tp->retransmit_skb_hint = NULL;
1789
1790                 }
1791         }
1792         tcp_sync_left_out(tp);
1793 }
1794
1795 /* Account newly detected lost packet(s) */
1796
1797 static void tcp_update_scoreboard(struct sock *sk)
1798 {
1799         struct tcp_sock *tp = tcp_sk(sk);
1800
1801         if (IsFack(tp)) {
1802                 int lost = tp->fackets_out - tp->reordering;
1803                 if (lost <= 0)
1804                         lost = 1;
1805                 tcp_mark_head_lost(sk, lost, tp->high_seq);
1806         } else {
1807                 tcp_mark_head_lost(sk, 1, tp->high_seq);
1808         }
1809
1810         /* New heuristics: it is possible only after we switched
1811          * to restart timer each time when something is ACKed.
1812          * Hence, we can detect timed out packets during fast
1813          * retransmit without falling to slow start.
1814          */
1815         if (!IsReno(tp) && tcp_head_timedout(sk)) {
1816                 struct sk_buff *skb;
1817
1818                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1819                         : tcp_write_queue_head(sk);
1820
1821                 tcp_for_write_queue_from(skb, sk) {
1822                         if (skb == tcp_send_head(sk))
1823                                 break;
1824                         if (!tcp_skb_timedout(sk, skb))
1825                                 break;
1826
1827                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1828                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1829                                 tp->lost_out += tcp_skb_pcount(skb);
1830
1831                                 /* clear xmit_retrans hint */
1832                                 if (tp->retransmit_skb_hint &&
1833                                     before(TCP_SKB_CB(skb)->seq,
1834                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1835
1836                                         tp->retransmit_skb_hint = NULL;
1837                         }
1838                 }
1839
1840                 tp->scoreboard_skb_hint = skb;
1841
1842                 tcp_sync_left_out(tp);
1843         }
1844 }
1845
1846 /* CWND moderation, preventing bursts due to too big ACKs
1847  * in dubious situations.
1848  */
1849 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1850 {
1851         tp->snd_cwnd = min(tp->snd_cwnd,
1852                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1853         tp->snd_cwnd_stamp = tcp_time_stamp;
1854 }
1855
1856 /* Lower bound on congestion window is slow start threshold
1857  * unless congestion avoidance choice decides to overide it.
1858  */
1859 static inline u32 tcp_cwnd_min(const struct sock *sk)
1860 {
1861         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1862
1863         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1864 }
1865
1866 /* Decrease cwnd each second ack. */
1867 static void tcp_cwnd_down(struct sock *sk, int flag)
1868 {
1869         struct tcp_sock *tp = tcp_sk(sk);
1870         int decr = tp->snd_cwnd_cnt + 1;
1871
1872         if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
1873             (IsReno(tp) && !(flag&FLAG_NOT_DUP))) {
1874                 tp->snd_cwnd_cnt = decr&1;
1875                 decr >>= 1;
1876
1877                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1878                         tp->snd_cwnd -= decr;
1879
1880                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1881                 tp->snd_cwnd_stamp = tcp_time_stamp;
1882         }
1883 }
1884
1885 /* Nothing was retransmitted or returned timestamp is less
1886  * than timestamp of the first retransmission.
1887  */
1888 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1889 {
1890         return !tp->retrans_stamp ||
1891                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1892                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1893 }
1894
1895 /* Undo procedures. */
1896
1897 #if FASTRETRANS_DEBUG > 1
1898 static void DBGUNDO(struct sock *sk, const char *msg)
1899 {
1900         struct tcp_sock *tp = tcp_sk(sk);
1901         struct inet_sock *inet = inet_sk(sk);
1902
1903         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1904                msg,
1905                NIPQUAD(inet->daddr), ntohs(inet->dport),
1906                tp->snd_cwnd, tp->left_out,
1907                tp->snd_ssthresh, tp->prior_ssthresh,
1908                tp->packets_out);
1909 }
1910 #else
1911 #define DBGUNDO(x...) do { } while (0)
1912 #endif
1913
1914 static void tcp_undo_cwr(struct sock *sk, const int undo)
1915 {
1916         struct tcp_sock *tp = tcp_sk(sk);
1917
1918         if (tp->prior_ssthresh) {
1919                 const struct inet_connection_sock *icsk = inet_csk(sk);
1920
1921                 if (icsk->icsk_ca_ops->undo_cwnd)
1922                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1923                 else
1924                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1925
1926                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1927                         tp->snd_ssthresh = tp->prior_ssthresh;
1928                         TCP_ECN_withdraw_cwr(tp);
1929                 }
1930         } else {
1931                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1932         }
1933         tcp_moderate_cwnd(tp);
1934         tp->snd_cwnd_stamp = tcp_time_stamp;
1935
1936         /* There is something screwy going on with the retrans hints after
1937            an undo */
1938         clear_all_retrans_hints(tp);
1939 }
1940
1941 static inline int tcp_may_undo(struct tcp_sock *tp)
1942 {
1943         return tp->undo_marker &&
1944                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1945 }
1946
1947 /* People celebrate: "We love our President!" */
1948 static int tcp_try_undo_recovery(struct sock *sk)
1949 {
1950         struct tcp_sock *tp = tcp_sk(sk);
1951
1952         if (tcp_may_undo(tp)) {
1953                 /* Happy end! We did not retransmit anything
1954                  * or our original transmission succeeded.
1955                  */
1956                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1957                 tcp_undo_cwr(sk, 1);
1958                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1959                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1960                 else
1961                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1962                 tp->undo_marker = 0;
1963         }
1964         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1965                 /* Hold old state until something *above* high_seq
1966                  * is ACKed. For Reno it is MUST to prevent false
1967                  * fast retransmits (RFC2582). SACK TCP is safe. */
1968                 tcp_moderate_cwnd(tp);
1969                 return 1;
1970         }
1971         tcp_set_ca_state(sk, TCP_CA_Open);
1972         return 0;
1973 }
1974
1975 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1976 static void tcp_try_undo_dsack(struct sock *sk)
1977 {
1978         struct tcp_sock *tp = tcp_sk(sk);
1979
1980         if (tp->undo_marker && !tp->undo_retrans) {
1981                 DBGUNDO(sk, "D-SACK");
1982                 tcp_undo_cwr(sk, 1);
1983                 tp->undo_marker = 0;
1984                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1985         }
1986 }
1987
1988 /* Undo during fast recovery after partial ACK. */
1989
1990 static int tcp_try_undo_partial(struct sock *sk, int acked)
1991 {
1992         struct tcp_sock *tp = tcp_sk(sk);
1993         /* Partial ACK arrived. Force Hoe's retransmit. */
1994         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1995
1996         if (tcp_may_undo(tp)) {
1997                 /* Plain luck! Hole if filled with delayed
1998                  * packet, rather than with a retransmit.
1999                  */
2000                 if (tp->retrans_out == 0)
2001                         tp->retrans_stamp = 0;
2002
2003                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2004
2005                 DBGUNDO(sk, "Hoe");
2006                 tcp_undo_cwr(sk, 0);
2007                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2008
2009                 /* So... Do not make Hoe's retransmit yet.
2010                  * If the first packet was delayed, the rest
2011                  * ones are most probably delayed as well.
2012                  */
2013                 failed = 0;
2014         }
2015         return failed;
2016 }
2017
2018 /* Undo during loss recovery after partial ACK. */
2019 static int tcp_try_undo_loss(struct sock *sk)
2020 {
2021         struct tcp_sock *tp = tcp_sk(sk);
2022
2023         if (tcp_may_undo(tp)) {
2024                 struct sk_buff *skb;
2025                 tcp_for_write_queue(skb, sk) {
2026                         if (skb == tcp_send_head(sk))
2027                                 break;
2028                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2029                 }
2030
2031                 clear_all_retrans_hints(tp);
2032
2033                 DBGUNDO(sk, "partial loss");
2034                 tp->lost_out = 0;
2035                 tp->left_out = tp->sacked_out;
2036                 tcp_undo_cwr(sk, 1);
2037                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2038                 inet_csk(sk)->icsk_retransmits = 0;
2039                 tp->undo_marker = 0;
2040                 if (!IsReno(tp))
2041                         tcp_set_ca_state(sk, TCP_CA_Open);
2042                 return 1;
2043         }
2044         return 0;
2045 }
2046
2047 static inline void tcp_complete_cwr(struct sock *sk)
2048 {
2049         struct tcp_sock *tp = tcp_sk(sk);
2050         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2051         tp->snd_cwnd_stamp = tcp_time_stamp;
2052         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2053 }
2054
2055 static void tcp_try_to_open(struct sock *sk, int flag)
2056 {
2057         struct tcp_sock *tp = tcp_sk(sk);
2058
2059         tcp_sync_left_out(tp);
2060
2061         if (tp->retrans_out == 0)
2062                 tp->retrans_stamp = 0;
2063
2064         if (flag&FLAG_ECE)
2065                 tcp_enter_cwr(sk, 1);
2066
2067         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2068                 int state = TCP_CA_Open;
2069
2070                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2071                         state = TCP_CA_Disorder;
2072
2073                 if (inet_csk(sk)->icsk_ca_state != state) {
2074                         tcp_set_ca_state(sk, state);
2075                         tp->high_seq = tp->snd_nxt;
2076                 }
2077                 tcp_moderate_cwnd(tp);
2078         } else {
2079                 tcp_cwnd_down(sk, flag);
2080         }
2081 }
2082
2083 static void tcp_mtup_probe_failed(struct sock *sk)
2084 {
2085         struct inet_connection_sock *icsk = inet_csk(sk);
2086
2087         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2088         icsk->icsk_mtup.probe_size = 0;
2089 }
2090
2091 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2092 {
2093         struct tcp_sock *tp = tcp_sk(sk);
2094         struct inet_connection_sock *icsk = inet_csk(sk);
2095
2096         /* FIXME: breaks with very large cwnd */
2097         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2098         tp->snd_cwnd = tp->snd_cwnd *
2099                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2100                        icsk->icsk_mtup.probe_size;
2101         tp->snd_cwnd_cnt = 0;
2102         tp->snd_cwnd_stamp = tcp_time_stamp;
2103         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2104
2105         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2106         icsk->icsk_mtup.probe_size = 0;
2107         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2108 }
2109
2110
2111 /* Process an event, which can update packets-in-flight not trivially.
2112  * Main goal of this function is to calculate new estimate for left_out,
2113  * taking into account both packets sitting in receiver's buffer and
2114  * packets lost by network.
2115  *
2116  * Besides that it does CWND reduction, when packet loss is detected
2117  * and changes state of machine.
2118  *
2119  * It does _not_ decide what to send, it is made in function
2120  * tcp_xmit_retransmit_queue().
2121  */
2122 static void
2123 tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag)
2124 {
2125         struct inet_connection_sock *icsk = inet_csk(sk);
2126         struct tcp_sock *tp = tcp_sk(sk);
2127         int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2128         int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2129                                     (tp->fackets_out > tp->reordering));
2130
2131         /* Some technical things:
2132          * 1. Reno does not count dupacks (sacked_out) automatically. */
2133         if (!tp->packets_out)
2134                 tp->sacked_out = 0;
2135         /* 2. SACK counts snd_fack in packets inaccurately. */
2136         if (tp->sacked_out == 0)
2137                 tp->fackets_out = 0;
2138
2139         /* Now state machine starts.
2140          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2141         if (flag&FLAG_ECE)
2142                 tp->prior_ssthresh = 0;
2143
2144         /* B. In all the states check for reneging SACKs. */
2145         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2146                 return;
2147
2148         /* C. Process data loss notification, provided it is valid. */
2149         if ((flag&FLAG_DATA_LOST) &&
2150             before(tp->snd_una, tp->high_seq) &&
2151             icsk->icsk_ca_state != TCP_CA_Open &&
2152             tp->fackets_out > tp->reordering) {
2153                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2154                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2155         }
2156
2157         /* D. Synchronize left_out to current state. */
2158         tcp_sync_left_out(tp);
2159
2160         /* E. Check state exit conditions. State can be terminated
2161          *    when high_seq is ACKed. */
2162         if (icsk->icsk_ca_state == TCP_CA_Open) {
2163                 BUG_TRAP(tp->retrans_out == 0);
2164                 tp->retrans_stamp = 0;
2165         } else if (!before(tp->snd_una, tp->high_seq)) {
2166                 switch (icsk->icsk_ca_state) {
2167                 case TCP_CA_Loss:
2168                         icsk->icsk_retransmits = 0;
2169                         if (tcp_try_undo_recovery(sk))
2170                                 return;
2171                         break;
2172
2173                 case TCP_CA_CWR:
2174                         /* CWR is to be held something *above* high_seq
2175                          * is ACKed for CWR bit to reach receiver. */
2176                         if (tp->snd_una != tp->high_seq) {
2177                                 tcp_complete_cwr(sk);
2178                                 tcp_set_ca_state(sk, TCP_CA_Open);
2179                         }
2180                         break;
2181
2182                 case TCP_CA_Disorder:
2183                         tcp_try_undo_dsack(sk);
2184                         if (!tp->undo_marker ||
2185                             /* For SACK case do not Open to allow to undo
2186                              * catching for all duplicate ACKs. */
2187                             IsReno(tp) || tp->snd_una != tp->high_seq) {
2188                                 tp->undo_marker = 0;
2189                                 tcp_set_ca_state(sk, TCP_CA_Open);
2190                         }
2191                         break;
2192
2193                 case TCP_CA_Recovery:
2194                         if (IsReno(tp))
2195                                 tcp_reset_reno_sack(tp);
2196                         if (tcp_try_undo_recovery(sk))
2197                                 return;
2198                         tcp_complete_cwr(sk);
2199                         break;
2200                 }
2201         }
2202
2203         /* F. Process state. */
2204         switch (icsk->icsk_ca_state) {
2205         case TCP_CA_Recovery:
2206                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2207                         if (IsReno(tp) && is_dupack)
2208                                 tcp_add_reno_sack(sk);
2209                 } else {
2210                         int acked = prior_packets - tp->packets_out;
2211                         if (IsReno(tp))
2212                                 tcp_remove_reno_sacks(sk, acked);
2213                         do_lost = tcp_try_undo_partial(sk, acked);
2214                 }
2215                 break;
2216         case TCP_CA_Loss:
2217                 if (flag&FLAG_DATA_ACKED)
2218                         icsk->icsk_retransmits = 0;
2219                 if (!tcp_try_undo_loss(sk)) {
2220                         tcp_moderate_cwnd(tp);
2221                         tcp_xmit_retransmit_queue(sk);
2222                         return;
2223                 }
2224                 if (icsk->icsk_ca_state != TCP_CA_Open)
2225                         return;
2226                 /* Loss is undone; fall through to processing in Open state. */
2227         default:
2228                 if (IsReno(tp)) {
2229                         if (flag & FLAG_SND_UNA_ADVANCED)
2230                                 tcp_reset_reno_sack(tp);
2231                         if (is_dupack)
2232                                 tcp_add_reno_sack(sk);
2233                 }
2234
2235                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2236                         tcp_try_undo_dsack(sk);
2237
2238                 if (!tcp_time_to_recover(sk)) {
2239                         tcp_try_to_open(sk, flag);
2240                         return;
2241                 }
2242
2243                 /* MTU probe failure: don't reduce cwnd */
2244                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2245                     icsk->icsk_mtup.probe_size &&
2246                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2247                         tcp_mtup_probe_failed(sk);
2248                         /* Restores the reduction we did in tcp_mtup_probe() */
2249                         tp->snd_cwnd++;
2250                         tcp_simple_retransmit(sk);
2251                         return;
2252                 }
2253
2254                 /* Otherwise enter Recovery state */
2255
2256                 if (IsReno(tp))
2257                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2258                 else
2259                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2260
2261                 tp->high_seq = tp->snd_nxt;
2262                 tp->prior_ssthresh = 0;
2263                 tp->undo_marker = tp->snd_una;
2264                 tp->undo_retrans = tp->retrans_out;
2265
2266                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2267                         if (!(flag&FLAG_ECE))
2268                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2269                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2270                         TCP_ECN_queue_cwr(tp);
2271                 }
2272
2273                 tp->bytes_acked = 0;
2274                 tp->snd_cwnd_cnt = 0;
2275                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2276         }
2277
2278         if (do_lost || tcp_head_timedout(sk))
2279                 tcp_update_scoreboard(sk);
2280         tcp_cwnd_down(sk, flag);
2281         tcp_xmit_retransmit_queue(sk);
2282 }
2283
2284 /* Read draft-ietf-tcplw-high-performance before mucking
2285  * with this code. (Supersedes RFC1323)
2286  */
2287 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2288 {
2289         /* RTTM Rule: A TSecr value received in a segment is used to
2290          * update the averaged RTT measurement only if the segment
2291          * acknowledges some new data, i.e., only if it advances the
2292          * left edge of the send window.
2293          *
2294          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2295          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2296          *
2297          * Changed: reset backoff as soon as we see the first valid sample.
2298          * If we do not, we get strongly overestimated rto. With timestamps
2299          * samples are accepted even from very old segments: f.e., when rtt=1
2300          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2301          * answer arrives rto becomes 120 seconds! If at least one of segments
2302          * in window is lost... Voila.                          --ANK (010210)
2303          */
2304         struct tcp_sock *tp = tcp_sk(sk);
2305         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2306         tcp_rtt_estimator(sk, seq_rtt);
2307         tcp_set_rto(sk);
2308         inet_csk(sk)->icsk_backoff = 0;
2309         tcp_bound_rto(sk);
2310 }
2311
2312 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2313 {
2314         /* We don't have a timestamp. Can only use
2315          * packets that are not retransmitted to determine
2316          * rtt estimates. Also, we must not reset the
2317          * backoff for rto until we get a non-retransmitted
2318          * packet. This allows us to deal with a situation
2319          * where the network delay has increased suddenly.
2320          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2321          */
2322
2323         if (flag & FLAG_RETRANS_DATA_ACKED)
2324                 return;
2325
2326         tcp_rtt_estimator(sk, seq_rtt);
2327         tcp_set_rto(sk);
2328         inet_csk(sk)->icsk_backoff = 0;
2329         tcp_bound_rto(sk);
2330 }
2331
2332 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2333                                       const s32 seq_rtt)
2334 {
2335         const struct tcp_sock *tp = tcp_sk(sk);
2336         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2337         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2338                 tcp_ack_saw_tstamp(sk, flag);
2339         else if (seq_rtt >= 0)
2340                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2341 }
2342
2343 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2344                            u32 in_flight, int good)
2345 {
2346         const struct inet_connection_sock *icsk = inet_csk(sk);
2347         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2348         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2349 }
2350
2351 /* Restart timer after forward progress on connection.
2352  * RFC2988 recommends to restart timer to now+rto.
2353  */
2354
2355 static void tcp_ack_packets_out(struct sock *sk)
2356 {
2357         struct tcp_sock *tp = tcp_sk(sk);
2358
2359         if (!tp->packets_out) {
2360                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2361         } else {
2362                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2363         }
2364 }
2365
2366 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2367                          __u32 now, __s32 *seq_rtt)
2368 {
2369         struct tcp_sock *tp = tcp_sk(sk);
2370         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2371         __u32 seq = tp->snd_una;
2372         __u32 packets_acked;
2373         int acked = 0;
2374
2375         /* If we get here, the whole TSO packet has not been
2376          * acked.
2377          */
2378         BUG_ON(!after(scb->end_seq, seq));
2379
2380         packets_acked = tcp_skb_pcount(skb);
2381         if (tcp_trim_head(sk, skb, seq - scb->seq))
2382                 return 0;
2383         packets_acked -= tcp_skb_pcount(skb);
2384
2385         if (packets_acked) {
2386                 __u8 sacked = scb->sacked;
2387
2388                 acked |= FLAG_DATA_ACKED;
2389                 if (sacked) {
2390                         if (sacked & TCPCB_RETRANS) {
2391                                 if (sacked & TCPCB_SACKED_RETRANS)
2392                                         tp->retrans_out -= packets_acked;
2393                                 acked |= FLAG_RETRANS_DATA_ACKED;
2394                                 *seq_rtt = -1;
2395                         } else if (*seq_rtt < 0)
2396                                 *seq_rtt = now - scb->when;
2397                         if (sacked & TCPCB_SACKED_ACKED)
2398                                 tp->sacked_out -= packets_acked;
2399                         if (sacked & TCPCB_LOST)
2400                                 tp->lost_out -= packets_acked;
2401                         if (sacked & TCPCB_URG) {
2402                                 if (tp->urg_mode &&
2403                                     !before(seq, tp->snd_up))
2404                                         tp->urg_mode = 0;
2405                         }
2406                 } else if (*seq_rtt < 0)
2407                         *seq_rtt = now - scb->when;
2408
2409                 if (tp->fackets_out) {
2410                         __u32 dval = min(tp->fackets_out, packets_acked);
2411                         tp->fackets_out -= dval;
2412                 }
2413                 tp->packets_out -= packets_acked;
2414
2415                 BUG_ON(tcp_skb_pcount(skb) == 0);
2416                 BUG_ON(!before(scb->seq, scb->end_seq));
2417         }
2418
2419         return acked;
2420 }
2421
2422 /* Remove acknowledged frames from the retransmission queue. */
2423 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2424 {
2425         struct tcp_sock *tp = tcp_sk(sk);
2426         const struct inet_connection_sock *icsk = inet_csk(sk);
2427         struct sk_buff *skb;
2428         __u32 now = tcp_time_stamp;
2429         int acked = 0;
2430         int prior_packets = tp->packets_out;
2431         __s32 seq_rtt = -1;
2432         ktime_t last_ackt = net_invalid_timestamp();
2433
2434         while ((skb = tcp_write_queue_head(sk)) &&
2435                skb != tcp_send_head(sk)) {
2436                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2437                 __u8 sacked = scb->sacked;
2438
2439                 /* If our packet is before the ack sequence we can
2440                  * discard it as it's confirmed to have arrived at
2441                  * the other end.
2442                  */
2443                 if (after(scb->end_seq, tp->snd_una)) {
2444                         if (tcp_skb_pcount(skb) > 1 &&
2445                             after(tp->snd_una, scb->seq))
2446                                 acked |= tcp_tso_acked(sk, skb,
2447                                                        now, &seq_rtt);
2448                         break;
2449                 }
2450
2451                 /* Initial outgoing SYN's get put onto the write_queue
2452                  * just like anything else we transmit.  It is not
2453                  * true data, and if we misinform our callers that
2454                  * this ACK acks real data, we will erroneously exit
2455                  * connection startup slow start one packet too
2456                  * quickly.  This is severely frowned upon behavior.
2457                  */
2458                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2459                         acked |= FLAG_DATA_ACKED;
2460                 } else {
2461                         acked |= FLAG_SYN_ACKED;
2462                         tp->retrans_stamp = 0;
2463                 }
2464
2465                 /* MTU probing checks */
2466                 if (icsk->icsk_mtup.probe_size) {
2467                         if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2468                                 tcp_mtup_probe_success(sk, skb);
2469                         }
2470                 }
2471
2472                 if (sacked) {
2473                         if (sacked & TCPCB_RETRANS) {
2474                                 if (sacked & TCPCB_SACKED_RETRANS)
2475                                         tp->retrans_out -= tcp_skb_pcount(skb);
2476                                 acked |= FLAG_RETRANS_DATA_ACKED;
2477                                 seq_rtt = -1;
2478                         } else if (seq_rtt < 0) {
2479                                 seq_rtt = now - scb->when;
2480                                 last_ackt = skb->tstamp;
2481                         }
2482                         if (sacked & TCPCB_SACKED_ACKED)
2483                                 tp->sacked_out -= tcp_skb_pcount(skb);
2484                         if (sacked & TCPCB_LOST)
2485                                 tp->lost_out -= tcp_skb_pcount(skb);
2486                         if (sacked & TCPCB_URG) {
2487                                 if (tp->urg_mode &&
2488                                     !before(scb->end_seq, tp->snd_up))
2489                                         tp->urg_mode = 0;
2490                         }
2491                 } else if (seq_rtt < 0) {
2492                         seq_rtt = now - scb->when;
2493                         last_ackt = skb->tstamp;
2494                 }
2495                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2496                 tcp_packets_out_dec(tp, skb);
2497                 tcp_unlink_write_queue(skb, sk);
2498                 sk_stream_free_skb(sk, skb);
2499                 clear_all_retrans_hints(tp);
2500         }
2501
2502         if (acked&FLAG_ACKED) {
2503                 u32 pkts_acked = prior_packets - tp->packets_out;
2504                 const struct tcp_congestion_ops *ca_ops
2505                         = inet_csk(sk)->icsk_ca_ops;
2506
2507                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2508                 tcp_ack_packets_out(sk);
2509
2510                 if (ca_ops->pkts_acked) {
2511                         s32 rtt_us = -1;
2512
2513                         /* Is the ACK triggering packet unambiguous? */
2514                         if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
2515                                 /* High resolution needed and available? */
2516                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2517                                     !ktime_equal(last_ackt,
2518                                                  net_invalid_timestamp()))
2519                                         rtt_us = ktime_us_delta(ktime_get_real(),
2520                                                                 last_ackt);
2521                                 else if (seq_rtt > 0)
2522                                         rtt_us = jiffies_to_usecs(seq_rtt);
2523                         }
2524
2525                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2526                 }
2527         }
2528
2529 #if FASTRETRANS_DEBUG > 0
2530         BUG_TRAP((int)tp->sacked_out >= 0);
2531         BUG_TRAP((int)tp->lost_out >= 0);
2532         BUG_TRAP((int)tp->retrans_out >= 0);
2533         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2534                 const struct inet_connection_sock *icsk = inet_csk(sk);
2535                 if (tp->lost_out) {
2536                         printk(KERN_DEBUG "Leak l=%u %d\n",
2537                                tp->lost_out, icsk->icsk_ca_state);
2538                         tp->lost_out = 0;
2539                 }
2540                 if (tp->sacked_out) {
2541                         printk(KERN_DEBUG "Leak s=%u %d\n",
2542                                tp->sacked_out, icsk->icsk_ca_state);
2543                         tp->sacked_out = 0;
2544                 }
2545                 if (tp->retrans_out) {
2546                         printk(KERN_DEBUG "Leak r=%u %d\n",
2547                                tp->retrans_out, icsk->icsk_ca_state);
2548                         tp->retrans_out = 0;
2549                 }
2550         }
2551 #endif
2552         *seq_rtt_p = seq_rtt;
2553         return acked;
2554 }
2555
2556 static void tcp_ack_probe(struct sock *sk)
2557 {
2558         const struct tcp_sock *tp = tcp_sk(sk);
2559         struct inet_connection_sock *icsk = inet_csk(sk);
2560
2561         /* Was it a usable window open? */
2562
2563         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2564                    tp->snd_una + tp->snd_wnd)) {
2565                 icsk->icsk_backoff = 0;
2566                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2567                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2568                  * This function is not for random using!
2569                  */
2570         } else {
2571                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2572                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2573                                           TCP_RTO_MAX);
2574         }
2575 }
2576
2577 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2578 {
2579         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2580                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2581 }
2582
2583 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2584 {
2585         const struct tcp_sock *tp = tcp_sk(sk);
2586         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2587                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2588 }
2589
2590 /* Check that window update is acceptable.
2591  * The function assumes that snd_una<=ack<=snd_next.
2592  */
2593 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2594                                         const u32 ack_seq, const u32 nwin)
2595 {
2596         return (after(ack, tp->snd_una) ||
2597                 after(ack_seq, tp->snd_wl1) ||
2598                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2599 }
2600
2601 /* Update our send window.
2602  *
2603  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2604  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2605  */
2606 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2607                                  u32 ack_seq)
2608 {
2609         struct tcp_sock *tp = tcp_sk(sk);
2610         int flag = 0;
2611         u32 nwin = ntohs(tcp_hdr(skb)->window);
2612
2613         if (likely(!tcp_hdr(skb)->syn))
2614                 nwin <<= tp->rx_opt.snd_wscale;
2615
2616         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2617                 flag |= FLAG_WIN_UPDATE;
2618                 tcp_update_wl(tp, ack, ack_seq);
2619
2620                 if (tp->snd_wnd != nwin) {
2621                         tp->snd_wnd = nwin;
2622
2623                         /* Note, it is the only place, where
2624                          * fast path is recovered for sending TCP.
2625                          */
2626                         tp->pred_flags = 0;
2627                         tcp_fast_path_check(sk);
2628
2629                         if (nwin > tp->max_window) {
2630                                 tp->max_window = nwin;
2631                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2632                         }
2633                 }
2634         }
2635
2636         tp->snd_una = ack;
2637
2638         return flag;
2639 }
2640
2641 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2642  * continue in congestion avoidance.
2643  */
2644 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2645 {
2646         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2647         tp->snd_cwnd_cnt = 0;
2648         TCP_ECN_queue_cwr(tp);
2649         tcp_moderate_cwnd(tp);
2650 }
2651
2652 /* A conservative spurious RTO response algorithm: reduce cwnd using
2653  * rate halving and continue in congestion avoidance.
2654  */
2655 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2656 {
2657         tcp_enter_cwr(sk, 0);
2658 }
2659
2660 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2661 {
2662         if (flag&FLAG_ECE)
2663                 tcp_ratehalving_spur_to_response(sk);
2664         else
2665                 tcp_undo_cwr(sk, 1);
2666 }
2667
2668 /* F-RTO spurious RTO detection algorithm (RFC4138)
2669  *
2670  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2671  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2672  * window (but not to or beyond highest sequence sent before RTO):
2673  *   On First ACK,  send two new segments out.
2674  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2675  *                  algorithm is not part of the F-RTO detection algorithm
2676  *                  given in RFC4138 but can be selected separately).
2677  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2678  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2679  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2680  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2681  *
2682  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2683  * original window even after we transmit two new data segments.
2684  *
2685  * SACK version:
2686  *   on first step, wait until first cumulative ACK arrives, then move to
2687  *   the second step. In second step, the next ACK decides.
2688  *
2689  * F-RTO is implemented (mainly) in four functions:
2690  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2691  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2692  *     called when tcp_use_frto() showed green light
2693  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2694  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2695  *     to prove that the RTO is indeed spurious. It transfers the control
2696  *     from F-RTO to the conventional RTO recovery
2697  */
2698 static int tcp_process_frto(struct sock *sk, int flag)
2699 {
2700         struct tcp_sock *tp = tcp_sk(sk);
2701
2702         tcp_sync_left_out(tp);
2703
2704         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2705         if (flag&FLAG_DATA_ACKED)
2706                 inet_csk(sk)->icsk_retransmits = 0;
2707
2708         if (!before(tp->snd_una, tp->frto_highmark)) {
2709                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2710                 return 1;
2711         }
2712
2713         if (!IsSackFrto() || IsReno(tp)) {
2714                 /* RFC4138 shortcoming in step 2; should also have case c):
2715                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2716                  * data, winupdate
2717                  */
2718                 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2719                         return 1;
2720
2721                 if (!(flag&FLAG_DATA_ACKED)) {
2722                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2723                                             flag);
2724                         return 1;
2725                 }
2726         } else {
2727                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2728                         /* Prevent sending of new data. */
2729                         tp->snd_cwnd = min(tp->snd_cwnd,
2730                                            tcp_packets_in_flight(tp));
2731                         return 1;
2732                 }
2733
2734                 if ((tp->frto_counter >= 2) &&
2735                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2736                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2737                         /* RFC4138 shortcoming (see comment above) */
2738                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2739                                 return 1;
2740
2741                         tcp_enter_frto_loss(sk, 3, flag);
2742                         return 1;
2743                 }
2744         }
2745
2746         if (tp->frto_counter == 1) {
2747                 /* Sending of the next skb must be allowed or no FRTO */
2748                 if (!tcp_send_head(sk) ||
2749                     after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2750                                      tp->snd_una + tp->snd_wnd)) {
2751                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2752                                             flag);
2753                         return 1;
2754                 }
2755
2756                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2757                 tp->frto_counter = 2;
2758                 return 1;
2759         } else {
2760                 switch (sysctl_tcp_frto_response) {
2761                 case 2:
2762                         tcp_undo_spur_to_response(sk, flag);
2763                         break;
2764                 case 1:
2765                         tcp_conservative_spur_to_response(tp);
2766                         break;
2767                 default:
2768                         tcp_ratehalving_spur_to_response(sk);
2769                         break;
2770                 }
2771                 tp->frto_counter = 0;
2772         }
2773         return 0;
2774 }
2775
2776 /* This routine deals with incoming acks, but not outgoing ones. */
2777 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2778 {
2779         struct inet_connection_sock *icsk = inet_csk(sk);
2780         struct tcp_sock *tp = tcp_sk(sk);
2781         u32 prior_snd_una = tp->snd_una;
2782         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2783         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2784         u32 prior_in_flight;
2785         s32 seq_rtt;
2786         int prior_packets;
2787         int frto_cwnd = 0;
2788
2789         /* If the ack is newer than sent or older than previous acks
2790          * then we can probably ignore it.
2791          */
2792         if (after(ack, tp->snd_nxt))
2793                 goto uninteresting_ack;
2794
2795         if (before(ack, prior_snd_una))
2796                 goto old_ack;
2797
2798         if (after(ack, prior_snd_una))
2799                 flag |= FLAG_SND_UNA_ADVANCED;
2800
2801         if (sysctl_tcp_abc) {
2802                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2803                         tp->bytes_acked += ack - prior_snd_una;
2804                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2805                         /* we assume just one segment left network */
2806                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2807         }
2808
2809         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2810                 /* Window is constant, pure forward advance.
2811                  * No more checks are required.
2812                  * Note, we use the fact that SND.UNA>=SND.WL2.
2813                  */
2814                 tcp_update_wl(tp, ack, ack_seq);
2815                 tp->snd_una = ack;
2816                 flag |= FLAG_WIN_UPDATE;
2817
2818                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2819
2820                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2821         } else {
2822                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2823                         flag |= FLAG_DATA;
2824                 else
2825                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2826
2827                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2828
2829                 if (TCP_SKB_CB(skb)->sacked)
2830                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2831
2832                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2833                         flag |= FLAG_ECE;
2834
2835                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2836         }
2837
2838         /* We passed data and got it acked, remove any soft error
2839          * log. Something worked...
2840          */
2841         sk->sk_err_soft = 0;
2842         tp->rcv_tstamp = tcp_time_stamp;
2843         prior_packets = tp->packets_out;
2844         if (!prior_packets)
2845                 goto no_queue;
2846
2847         prior_in_flight = tcp_packets_in_flight(tp);
2848
2849         /* See if we can take anything off of the retransmit queue. */
2850         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2851
2852         if (tp->frto_counter)
2853                 frto_cwnd = tcp_process_frto(sk, flag);
2854
2855         if (tcp_ack_is_dubious(sk, flag)) {
2856                 /* Advance CWND, if state allows this. */
2857                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2858                     tcp_may_raise_cwnd(sk, flag))
2859                         tcp_cong_avoid(sk, ack, prior_in_flight, 0);
2860                 tcp_fastretrans_alert(sk, prior_packets, flag);
2861         } else {
2862                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2863                         tcp_cong_avoid(sk, ack, prior_in_flight, 1);
2864         }
2865
2866         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2867                 dst_confirm(sk->sk_dst_cache);
2868
2869         return 1;
2870
2871 no_queue:
2872         icsk->icsk_probes_out = 0;
2873
2874         /* If this ack opens up a zero window, clear backoff.  It was
2875          * being used to time the probes, and is probably far higher than
2876          * it needs to be for normal retransmission.
2877          */
2878         if (tcp_send_head(sk))
2879                 tcp_ack_probe(sk);
2880         return 1;
2881
2882 old_ack:
2883         if (TCP_SKB_CB(skb)->sacked)
2884                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2885
2886 uninteresting_ack:
2887         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2888         return 0;
2889 }
2890
2891
2892 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2893  * But, this can also be called on packets in the established flow when
2894  * the fast version below fails.
2895  */
2896 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2897 {
2898         unsigned char *ptr;
2899         struct tcphdr *th = tcp_hdr(skb);
2900         int length=(th->doff*4)-sizeof(struct tcphdr);
2901
2902         ptr = (unsigned char *)(th + 1);
2903         opt_rx->saw_tstamp = 0;
2904
2905         while (length > 0) {
2906                 int opcode=*ptr++;
2907                 int opsize;
2908
2909                 switch (opcode) {
2910                         case TCPOPT_EOL:
2911                                 return;
2912                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2913                                 length--;
2914                                 continue;
2915                         default:
2916                                 opsize=*ptr++;
2917                                 if (opsize < 2) /* "silly options" */
2918                                         return;
2919                                 if (opsize > length)
2920                                         return; /* don't parse partial options */
2921                                 switch (opcode) {
2922                                 case TCPOPT_MSS:
2923                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2924                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2925                                                 if (in_mss) {
2926                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2927                                                                 in_mss = opt_rx->user_mss;
2928                                                         opt_rx->mss_clamp = in_mss;
2929                                                 }
2930                                         }
2931                                         break;
2932                                 case TCPOPT_WINDOW:
2933                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2934                                                 if (sysctl_tcp_window_scaling) {
2935                                                         __u8 snd_wscale = *(__u8 *) ptr;
2936                                                         opt_rx->wscale_ok = 1;
2937                                                         if (snd_wscale > 14) {
2938                                                                 if (net_ratelimit())
2939                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2940                                                                                "scaling value %d >14 received.\n",
2941                                                                                snd_wscale);
2942                                                                 snd_wscale = 14;
2943                                                         }
2944                                                         opt_rx->snd_wscale = snd_wscale;
2945                                                 }
2946                                         break;
2947                                 case TCPOPT_TIMESTAMP:
2948                                         if (opsize==TCPOLEN_TIMESTAMP) {
2949                                                 if ((estab && opt_rx->tstamp_ok) ||
2950                                                     (!estab && sysctl_tcp_timestamps)) {
2951                                                         opt_rx->saw_tstamp = 1;
2952                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2953                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2954                                                 }
2955                                         }
2956                                         break;
2957                                 case TCPOPT_SACK_PERM:
2958                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2959                                                 if (sysctl_tcp_sack) {
2960                                                         opt_rx->sack_ok = 1;
2961                                                         tcp_sack_reset(opt_rx);
2962                                                 }
2963                                         }
2964                                         break;
2965
2966                                 case TCPOPT_SACK:
2967                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2968                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2969                                            opt_rx->sack_ok) {
2970                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2971                                         }
2972                                         break;
2973 #ifdef CONFIG_TCP_MD5SIG
2974                                 case TCPOPT_MD5SIG:
2975                                         /*
2976                                          * The MD5 Hash has already been
2977                                          * checked (see tcp_v{4,6}_do_rcv()).
2978                                          */
2979                                         break;
2980 #endif
2981                                 }
2982
2983                                 ptr+=opsize-2;
2984                                 length-=opsize;
2985                 }
2986         }
2987 }
2988
2989 /* Fast parse options. This hopes to only see timestamps.
2990  * If it is wrong it falls back on tcp_parse_options().
2991  */
2992 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2993                                   struct tcp_sock *tp)
2994 {
2995         if (th->doff == sizeof(struct tcphdr)>>2) {
2996                 tp->rx_opt.saw_tstamp = 0;
2997                 return 0;
2998         } else if (tp->rx_opt.tstamp_ok &&
2999                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3000                 __be32 *ptr = (__be32 *)(th + 1);
3001                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3002                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3003                         tp->rx_opt.saw_tstamp = 1;
3004                         ++ptr;
3005                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3006                         ++ptr;
3007                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3008                         return 1;
3009                 }
3010         }
3011         tcp_parse_options(skb, &tp->rx_opt, 1);
3012         return 1;
3013 }
3014
3015 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3016 {
3017         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3018         tp->rx_opt.ts_recent_stamp = get_seconds();
3019 }
3020
3021 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3022 {
3023         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3024                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3025                  * extra check below makes sure this can only happen
3026                  * for pure ACK frames.  -DaveM
3027                  *
3028                  * Not only, also it occurs for expired timestamps.
3029                  */
3030
3031                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3032                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3033                         tcp_store_ts_recent(tp);
3034         }
3035 }
3036
3037 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3038  *
3039  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3040  * it can pass through stack. So, the following predicate verifies that
3041  * this segment is not used for anything but congestion avoidance or
3042  * fast retransmit. Moreover, we even are able to eliminate most of such
3043  * second order effects, if we apply some small "replay" window (~RTO)
3044  * to timestamp space.
3045  *
3046  * All these measures still do not guarantee that we reject wrapped ACKs
3047  * on networks with high bandwidth, when sequence space is recycled fastly,
3048  * but it guarantees that such events will be very rare and do not affect
3049  * connection seriously. This doesn't look nice, but alas, PAWS is really
3050  * buggy extension.
3051  *
3052  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3053  * states that events when retransmit arrives after original data are rare.
3054  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3055  * the biggest problem on large power networks even with minor reordering.
3056  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3057  * up to bandwidth of 18Gigabit/sec. 8) ]
3058  */
3059
3060 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3061 {
3062         struct tcp_sock *tp = tcp_sk(sk);
3063         struct tcphdr *th = tcp_hdr(skb);
3064         u32 seq = TCP_SKB_CB(skb)->seq;
3065         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3066
3067         return (/* 1. Pure ACK with correct sequence number. */
3068                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3069
3070                 /* 2. ... and duplicate ACK. */
3071                 ack == tp->snd_una &&
3072
3073                 /* 3. ... and does not update window. */
3074                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3075
3076                 /* 4. ... and sits in replay window. */
3077                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3078 }
3079
3080 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3081 {
3082         const struct tcp_sock *tp = tcp_sk(sk);
3083         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3084                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3085                 !tcp_disordered_ack(sk, skb));
3086 }
3087
3088 /* Check segment sequence number for validity.
3089  *
3090  * Segment controls are considered valid, if the segment
3091  * fits to the window after truncation to the window. Acceptability
3092  * of data (and SYN, FIN, of course) is checked separately.
3093  * See tcp_data_queue(), for example.
3094  *
3095  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3096  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3097  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3098  * (borrowed from freebsd)
3099  */
3100
3101 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3102 {
3103         return  !before(end_seq, tp->rcv_wup) &&
3104                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3105 }
3106
3107 /* When we get a reset we do this. */
3108 static void tcp_reset(struct sock *sk)
3109 {
3110         /* We want the right error as BSD sees it (and indeed as we do). */
3111         switch (sk->sk_state) {
3112                 case TCP_SYN_SENT:
3113                         sk->sk_err = ECONNREFUSED;
3114                         break;
3115                 case TCP_CLOSE_WAIT:
3116                         sk->sk_err = EPIPE;
3117                         break;
3118                 case TCP_CLOSE:
3119                         return;
3120                 default:
3121                         sk->sk_err = ECONNRESET;
3122         }
3123
3124         if (!sock_flag(sk, SOCK_DEAD))
3125                 sk->sk_error_report(sk);
3126
3127         tcp_done(sk);
3128 }
3129
3130 /*
3131  *      Process the FIN bit. This now behaves as it is supposed to work
3132  *      and the FIN takes effect when it is validly part of sequence
3133  *      space. Not before when we get holes.
3134  *
3135  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3136  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3137  *      TIME-WAIT)
3138  *
3139  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3140  *      close and we go into CLOSING (and later onto TIME-WAIT)
3141  *
3142  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3143  */
3144 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3145 {
3146         struct tcp_sock *tp = tcp_sk(sk);
3147
3148         inet_csk_schedule_ack(sk);
3149
3150         sk->sk_shutdown |= RCV_SHUTDOWN;
3151         sock_set_flag(sk, SOCK_DONE);
3152
3153         switch (sk->sk_state) {
3154                 case TCP_SYN_RECV:
3155                 case TCP_ESTABLISHED:
3156                         /* Move to CLOSE_WAIT */
3157                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3158                         inet_csk(sk)->icsk_ack.pingpong = 1;
3159                         break;
3160
3161                 case TCP_CLOSE_WAIT:
3162                 case TCP_CLOSING:
3163                         /* Received a retransmission of the FIN, do
3164                          * nothing.
3165                          */
3166                         break;
3167                 case TCP_LAST_ACK:
3168                         /* RFC793: Remain in the LAST-ACK state. */
3169                         break;
3170
3171                 case TCP_FIN_WAIT1:
3172                         /* This case occurs when a simultaneous close
3173                          * happens, we must ack the received FIN and
3174                          * enter the CLOSING state.
3175                          */
3176                         tcp_send_ack(sk);
3177                         tcp_set_state(sk, TCP_CLOSING);
3178                         break;
3179                 case TCP_FIN_WAIT2:
3180                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3181                         tcp_send_ack(sk);
3182                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3183                         break;
3184                 default:
3185                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3186                          * cases we should never reach this piece of code.
3187                          */
3188                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3189                                __FUNCTION__, sk->sk_state);
3190                         break;
3191         }
3192
3193         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3194          * Probably, we should reset in this case. For now drop them.
3195          */
3196         __skb_queue_purge(&tp->out_of_order_queue);
3197         if (tp->rx_opt.sack_ok)
3198                 tcp_sack_reset(&tp->rx_opt);
3199         sk_stream_mem_reclaim(sk);
3200
3201         if (!sock_flag(sk, SOCK_DEAD)) {
3202                 sk->sk_state_change(sk);
3203
3204                 /* Do not send POLL_HUP for half duplex close. */
3205                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3206                     sk->sk_state == TCP_CLOSE)
3207                         sk_wake_async(sk, 1, POLL_HUP);
3208                 else
3209                         sk_wake_async(sk, 1, POLL_IN);
3210         }
3211 }
3212
3213 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3214 {
3215         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3216                 if (before(seq, sp->start_seq))
3217                         sp->start_seq = seq;
3218                 if (after(end_seq, sp->end_seq))
3219                         sp->end_seq = end_seq;
3220                 return 1;
3221         }
3222         return 0;
3223 }
3224
3225 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3226 {
3227         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3228                 if (before(seq, tp->rcv_nxt))
3229                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3230                 else
3231                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3232
3233                 tp->rx_opt.dsack = 1;
3234                 tp->duplicate_sack[0].start_seq = seq;
3235                 tp->duplicate_sack[0].end_seq = end_seq;
3236                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3237         }
3238 }
3239
3240 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3241 {
3242         if (!tp->rx_opt.dsack)
3243                 tcp_dsack_set(tp, seq, end_seq);
3244         else
3245                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3246 }
3247
3248 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3249 {
3250         struct tcp_sock *tp = tcp_sk(sk);
3251
3252         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3253             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3254                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3255                 tcp_enter_quickack_mode(sk);
3256
3257                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3258                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3259
3260                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3261                                 end_seq = tp->rcv_nxt;
3262                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3263                 }
3264         }
3265
3266         tcp_send_ack(sk);
3267 }
3268
3269 /* These routines update the SACK block as out-of-order packets arrive or
3270  * in-order packets close up the sequence space.
3271  */
3272 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3273 {
3274         int this_sack;
3275         struct tcp_sack_block *sp = &tp->selective_acks[0];
3276         struct tcp_sack_block *swalk = sp+1;
3277
3278         /* See if the recent change to the first SACK eats into
3279          * or hits the sequence space of other SACK blocks, if so coalesce.
3280          */
3281         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3282                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3283                         int i;
3284
3285                         /* Zap SWALK, by moving every further SACK up by one slot.
3286                          * Decrease num_sacks.
3287                          */
3288                         tp->rx_opt.num_sacks--;
3289                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3290                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3291                                 sp[i] = sp[i+1];
3292                         continue;
3293                 }
3294                 this_sack++, swalk++;
3295         }
3296 }
3297
3298 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3299 {
3300         __u32 tmp;
3301
3302         tmp = sack1->start_seq;
3303         sack1->start_seq = sack2->start_seq;
3304         sack2->start_seq = tmp;
3305
3306         tmp = sack1->end_seq;
3307         sack1->end_seq = sack2->end_seq;
3308         sack2->end_seq = tmp;
3309 }
3310
3311 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3312 {
3313         struct tcp_sock *tp = tcp_sk(sk);
3314         struct tcp_sack_block *sp = &tp->selective_acks[0];
3315         int cur_sacks = tp->rx_opt.num_sacks;
3316         int this_sack;
3317
3318         if (!cur_sacks)
3319                 goto new_sack;
3320
3321         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3322                 if (tcp_sack_extend(sp, seq, end_seq)) {
3323                         /* Rotate this_sack to the first one. */
3324                         for (; this_sack>0; this_sack--, sp--)
3325                                 tcp_sack_swap(sp, sp-1);
3326                         if (cur_sacks > 1)
3327                                 tcp_sack_maybe_coalesce(tp);
3328                         return;
3329                 }
3330         }
3331
3332         /* Could not find an adjacent existing SACK, build a new one,
3333          * put it at the front, and shift everyone else down.  We
3334          * always know there is at least one SACK present already here.
3335          *
3336          * If the sack array is full, forget about the last one.
3337          */
3338         if (this_sack >= 4) {
3339                 this_sack--;
3340                 tp->rx_opt.num_sacks--;
3341                 sp--;
3342         }
3343         for (; this_sack > 0; this_sack--, sp--)
3344                 *sp = *(sp-1);
3345
3346 new_sack:
3347         /* Build the new head SACK, and we're done. */
3348         sp->start_seq = seq;
3349         sp->end_seq = end_seq;
3350         tp->rx_opt.num_sacks++;
3351         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3352 }
3353
3354 /* RCV.NXT advances, some SACKs should be eaten. */
3355
3356 static void tcp_sack_remove(struct tcp_sock *tp)
3357 {
3358         struct tcp_sack_block *sp = &tp->selective_acks[0];
3359         int num_sacks = tp->rx_opt.num_sacks;
3360         int this_sack;
3361
3362         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3363         if (skb_queue_empty(&tp->out_of_order_queue)) {
3364                 tp->rx_opt.num_sacks = 0;
3365                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3366                 return;
3367         }
3368
3369         for (this_sack = 0; this_sack < num_sacks; ) {
3370                 /* Check if the start of the sack is covered by RCV.NXT. */
3371                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3372                         int i;
3373
3374                         /* RCV.NXT must cover all the block! */
3375                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3376
3377                         /* Zap this SACK, by moving forward any other SACKS. */
3378                         for (i=this_sack+1; i < num_sacks; i++)
3379                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3380                         num_sacks--;
3381                         continue;
3382                 }
3383                 this_sack++;
3384                 sp++;
3385         }
3386         if (num_sacks != tp->rx_opt.num_sacks) {
3387                 tp->rx_opt.num_sacks = num_sacks;
3388                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3389         }
3390 }
3391
3392 /* This one checks to see if we can put data from the
3393  * out_of_order queue into the receive_queue.
3394  */
3395 static void tcp_ofo_queue(struct sock *sk)
3396 {
3397         struct tcp_sock *tp = tcp_sk(sk);
3398         __u32 dsack_high = tp->rcv_nxt;
3399         struct sk_buff *skb;
3400
3401         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3402                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3403                         break;
3404
3405                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3406                         __u32 dsack = dsack_high;
3407                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3408                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3409                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3410                 }
3411
3412                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3413                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3414                         __skb_unlink(skb, &tp->out_of_order_queue);
3415                         __kfree_skb(skb);
3416                         continue;
3417                 }
3418                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3419                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3420                            TCP_SKB_CB(skb)->end_seq);
3421
3422                 __skb_unlink(skb, &tp->out_of_order_queue);
3423                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3424                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3425                 if (tcp_hdr(skb)->fin)
3426                         tcp_fin(skb, sk, tcp_hdr(skb));
3427         }
3428 }
3429
3430 static int tcp_prune_queue(struct sock *sk);
3431
3432 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3433 {
3434         struct tcphdr *th = tcp_hdr(skb);
3435         struct tcp_sock *tp = tcp_sk(sk);
3436         int eaten = -1;
3437
3438         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3439                 goto drop;
3440
3441         __skb_pull(skb, th->doff*4);
3442
3443         TCP_ECN_accept_cwr(tp, skb);
3444
3445         if (tp->rx_opt.dsack) {
3446                 tp->rx_opt.dsack = 0;
3447                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3448                                                     4 - tp->rx_opt.tstamp_ok);
3449         }
3450
3451         /*  Queue data for delivery to the user.
3452          *  Packets in sequence go to the receive queue.
3453          *  Out of sequence packets to the out_of_order_queue.
3454          */
3455         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3456                 if (tcp_receive_window(tp) == 0)
3457                         goto out_of_window;
3458
3459                 /* Ok. In sequence. In window. */
3460                 if (tp->ucopy.task == current &&
3461                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3462                     sock_owned_by_user(sk) && !tp->urg_data) {
3463                         int chunk = min_t(unsigned int, skb->len,
3464                                                         tp->ucopy.len);
3465
3466                         __set_current_state(TASK_RUNNING);
3467
3468                         local_bh_enable();
3469                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3470                                 tp->ucopy.len -= chunk;
3471                                 tp->copied_seq += chunk;
3472                                 eaten = (chunk == skb->len && !th->fin);
3473                                 tcp_rcv_space_adjust(sk);
3474                         }
3475                         local_bh_disable();
3476                 }
3477
3478                 if (eaten <= 0) {
3479 queue_and_out:
3480                         if (eaten < 0 &&
3481                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3482                              !sk_stream_rmem_schedule(sk, skb))) {
3483                                 if (tcp_prune_queue(sk) < 0 ||
3484                                     !sk_stream_rmem_schedule(sk, skb))
3485                                         goto drop;
3486                         }
3487                         sk_stream_set_owner_r(skb, sk);
3488                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3489                 }
3490                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3491                 if (skb->len)
3492                         tcp_event_data_recv(sk, skb);
3493                 if (th->fin)
3494                         tcp_fin(skb, sk, th);
3495
3496                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3497                         tcp_ofo_queue(sk);
3498
3499                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3500                          * gap in queue is filled.
3501                          */
3502                         if (skb_queue_empty(&tp->out_of_order_queue))
3503                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3504                 }
3505
3506                 if (tp->rx_opt.num_sacks)
3507                         tcp_sack_remove(tp);
3508
3509                 tcp_fast_path_check(sk);
3510
3511                 if (eaten > 0)
3512                         __kfree_skb(skb);
3513                 else if (!sock_flag(sk, SOCK_DEAD))
3514                         sk->sk_data_ready(sk, 0);
3515                 return;
3516         }
3517
3518         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3519                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3520                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3521                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3522
3523 out_of_window:
3524                 tcp_enter_quickack_mode(sk);
3525                 inet_csk_schedule_ack(sk);
3526 drop:
3527                 __kfree_skb(skb);
3528                 return;
3529         }
3530
3531         /* Out of window. F.e. zero window probe. */
3532         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3533                 goto out_of_window;
3534
3535         tcp_enter_quickack_mode(sk);
3536
3537         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3538                 /* Partial packet, seq < rcv_next < end_seq */
3539                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3540                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3541                            TCP_SKB_CB(skb)->end_seq);
3542
3543                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3544
3545                 /* If window is closed, drop tail of packet. But after
3546                  * remembering D-SACK for its head made in previous line.
3547                  */
3548                 if (!tcp_receive_window(tp))
3549                         goto out_of_window;
3550                 goto queue_and_out;
3551         }
3552
3553         TCP_ECN_check_ce(tp, skb);
3554
3555         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3556             !sk_stream_rmem_schedule(sk, skb)) {
3557                 if (tcp_prune_queue(sk) < 0 ||
3558                     !sk_stream_rmem_schedule(sk, skb))
3559                         goto drop;
3560         }
3561
3562         /* Disable header prediction. */
3563         tp->pred_flags = 0;
3564         inet_csk_schedule_ack(sk);
3565
3566         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3567                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3568
3569         sk_stream_set_owner_r(skb, sk);
3570
3571         if (!skb_peek(&tp->out_of_order_queue)) {
3572                 /* Initial out of order segment, build 1 SACK. */
3573                 if (tp->rx_opt.sack_ok) {
3574                         tp->rx_opt.num_sacks = 1;
3575                         tp->rx_opt.dsack     = 0;
3576                         tp->rx_opt.eff_sacks = 1;
3577                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3578                         tp->selective_acks[0].end_seq =
3579                                                 TCP_SKB_CB(skb)->end_seq;
3580                 }
3581                 __skb_queue_head(&tp->out_of_order_queue,skb);
3582         } else {
3583                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3584                 u32 seq = TCP_SKB_CB(skb)->seq;
3585                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3586
3587                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3588                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3589
3590                         if (!tp->rx_opt.num_sacks ||
3591                             tp->selective_acks[0].end_seq != seq)
3592                                 goto add_sack;
3593
3594                         /* Common case: data arrive in order after hole. */
3595                         tp->selective_acks[0].end_seq = end_seq;
3596                         return;
3597                 }
3598
3599                 /* Find place to insert this segment. */
3600                 do {
3601                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3602                                 break;
3603                 } while ((skb1 = skb1->prev) !=
3604                          (struct sk_buff*)&tp->out_of_order_queue);
3605
3606                 /* Do skb overlap to previous one? */
3607                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3608                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3609                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3610                                 /* All the bits are present. Drop. */
3611                                 __kfree_skb(skb);
3612                                 tcp_dsack_set(tp, seq, end_seq);
3613                                 goto add_sack;
3614                         }
3615                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3616                                 /* Partial overlap. */
3617                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3618                         } else {
3619                                 skb1 = skb1->prev;
3620                         }
3621                 }
3622                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3623
3624                 /* And clean segments covered by new one as whole. */
3625                 while ((skb1 = skb->next) !=
3626                        (struct sk_buff*)&tp->out_of_order_queue &&
3627                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3628                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3629                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3630                                break;
3631                        }
3632                        __skb_unlink(skb1, &tp->out_of_order_queue);
3633                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3634                        __kfree_skb(skb1);
3635                 }
3636
3637 add_sack:
3638                 if (tp->rx_opt.sack_ok)
3639                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3640         }
3641 }
3642
3643 /* Collapse contiguous sequence of skbs head..tail with
3644  * sequence numbers start..end.
3645  * Segments with FIN/SYN are not collapsed (only because this
3646  * simplifies code)
3647  */
3648 static void
3649 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3650              struct sk_buff *head, struct sk_buff *tail,
3651              u32 start, u32 end)
3652 {
3653         struct sk_buff *skb;
3654
3655         /* First, check that queue is collapsible and find
3656          * the point where collapsing can be useful. */
3657         for (skb = head; skb != tail; ) {
3658                 /* No new bits? It is possible on ofo queue. */
3659                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3660                         struct sk_buff *next = skb->next;
3661                         __skb_unlink(skb, list);
3662                         __kfree_skb(skb);
3663                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3664                         skb = next;
3665                         continue;
3666                 }
3667
3668                 /* The first skb to collapse is:
3669                  * - not SYN/FIN and
3670                  * - bloated or contains data before "start" or
3671                  *   overlaps to the next one.
3672                  */
3673                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3674                     (tcp_win_from_space(skb->truesize) > skb->len ||
3675                      before(TCP_SKB_CB(skb)->seq, start) ||
3676                      (skb->next != tail &&
3677                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3678                         break;
3679
3680                 /* Decided to skip this, advance start seq. */
3681                 start = TCP_SKB_CB(skb)->end_seq;
3682                 skb = skb->next;
3683         }
3684         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3685                 return;
3686
3687         while (before(start, end)) {
3688                 struct sk_buff *nskb;
3689                 int header = skb_headroom(skb);
3690                 int copy = SKB_MAX_ORDER(header, 0);
3691
3692                 /* Too big header? This can happen with IPv6. */
3693                 if (copy < 0)
3694                         return;
3695                 if (end-start < copy)
3696                         copy = end-start;
3697                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3698                 if (!nskb)
3699                         return;
3700
3701                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3702                 skb_set_network_header(nskb, (skb_network_header(skb) -
3703                                               skb->head));
3704                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3705                                                 skb->head));
3706                 skb_reserve(nskb, header);
3707                 memcpy(nskb->head, skb->head, header);
3708                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3709                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3710                 __skb_insert(nskb, skb->prev, skb, list);
3711                 sk_stream_set_owner_r(nskb, sk);
3712
3713                 /* Copy data, releasing collapsed skbs. */
3714                 while (copy > 0) {
3715                         int offset = start - TCP_SKB_CB(skb)->seq;
3716                         int size = TCP_SKB_CB(skb)->end_seq - start;
3717
3718                         BUG_ON(offset < 0);
3719                         if (size > 0) {
3720                                 size = min(copy, size);
3721                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3722                                         BUG();
3723                                 TCP_SKB_CB(nskb)->end_seq += size;
3724                                 copy -= size;
3725                                 start += size;
3726                         }
3727                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3728                                 struct sk_buff *next = skb->next;
3729                                 __skb_unlink(skb, list);
3730                                 __kfree_skb(skb);
3731                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3732                                 skb = next;
3733                                 if (skb == tail ||
3734                                     tcp_hdr(skb)->syn ||
3735                                     tcp_hdr(skb)->fin)
3736                                         return;
3737                         }
3738                 }
3739         }
3740 }
3741
3742 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3743  * and tcp_collapse() them until all the queue is collapsed.
3744  */
3745 static void tcp_collapse_ofo_queue(struct sock *sk)
3746 {
3747         struct tcp_sock *tp = tcp_sk(sk);
3748         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3749         struct sk_buff *head;
3750         u32 start, end;
3751
3752         if (skb == NULL)
3753                 return;
3754
3755         start = TCP_SKB_CB(skb)->seq;
3756         end = TCP_SKB_CB(skb)->end_seq;
3757         head = skb;
3758
3759         for (;;) {
3760                 skb = skb->next;
3761
3762                 /* Segment is terminated when we see gap or when
3763                  * we are at the end of all the queue. */
3764                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3765                     after(TCP_SKB_CB(skb)->seq, end) ||
3766                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3767                         tcp_collapse(sk, &tp->out_of_order_queue,
3768                                      head, skb, start, end);
3769                         head = skb;
3770                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3771                                 break;
3772                         /* Start new segment */
3773                         start = TCP_SKB_CB(skb)->seq;
3774                         end = TCP_SKB_CB(skb)->end_seq;
3775                 } else {
3776                         if (before(TCP_SKB_CB(skb)->seq, start))
3777                                 start = TCP_SKB_CB(skb)->seq;
3778                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3779                                 end = TCP_SKB_CB(skb)->end_seq;
3780                 }
3781         }
3782 }
3783
3784 /* Reduce allocated memory if we can, trying to get
3785  * the socket within its memory limits again.
3786  *
3787  * Return less than zero if we should start dropping frames
3788  * until the socket owning process reads some of the data
3789  * to stabilize the situation.
3790  */
3791 static int tcp_prune_queue(struct sock *sk)
3792 {
3793         struct tcp_sock *tp = tcp_sk(sk);
3794
3795         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3796
3797         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3798
3799         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3800                 tcp_clamp_window(sk);
3801         else if (tcp_memory_pressure)
3802                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3803
3804         tcp_collapse_ofo_queue(sk);
3805         tcp_collapse(sk, &sk->sk_receive_queue,
3806                      sk->sk_receive_queue.next,
3807                      (struct sk_buff*)&sk->sk_receive_queue,
3808                      tp->copied_seq, tp->rcv_nxt);
3809         sk_stream_mem_reclaim(sk);
3810
3811         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3812                 return 0;
3813
3814         /* Collapsing did not help, destructive actions follow.
3815          * This must not ever occur. */
3816
3817         /* First, purge the out_of_order queue. */
3818         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3819                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3820                 __skb_queue_purge(&tp->out_of_order_queue);
3821
3822                 /* Reset SACK state.  A conforming SACK implementation will
3823                  * do the same at a timeout based retransmit.  When a connection
3824                  * is in a sad state like this, we care only about integrity
3825                  * of the connection not performance.
3826                  */
3827                 if (tp->rx_opt.sack_ok)
3828                         tcp_sack_reset(&tp->rx_opt);
3829                 sk_stream_mem_reclaim(sk);
3830         }
3831
3832         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3833                 return 0;
3834
3835         /* If we are really being abused, tell the caller to silently
3836          * drop receive data on the floor.  It will get retransmitted
3837          * and hopefully then we'll have sufficient space.
3838          */
3839         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3840
3841         /* Massive buffer overcommit. */
3842         tp->pred_flags = 0;
3843         return -1;
3844 }
3845
3846
3847 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3848  * As additional protections, we do not touch cwnd in retransmission phases,
3849  * and if application hit its sndbuf limit recently.
3850  */
3851 void tcp_cwnd_application_limited(struct sock *sk)
3852 {
3853         struct tcp_sock *tp = tcp_sk(sk);
3854
3855         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3856             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3857                 /* Limited by application or receiver window. */
3858                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3859                 u32 win_used = max(tp->snd_cwnd_used, init_win);
3860                 if (win_used < tp->snd_cwnd) {
3861                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3862                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3863                 }
3864                 tp->snd_cwnd_used = 0;
3865         }
3866         tp->snd_cwnd_stamp = tcp_time_stamp;
3867 }
3868
3869 static int tcp_should_expand_sndbuf(struct sock *sk)
3870 {
3871         struct tcp_sock *tp = tcp_sk(sk);
3872
3873         /* If the user specified a specific send buffer setting, do
3874          * not modify it.
3875          */
3876         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3877                 return 0;
3878
3879         /* If we are under global TCP memory pressure, do not expand.  */
3880         if (tcp_memory_pressure)
3881                 return 0;
3882
3883         /* If we are under soft global TCP memory pressure, do not expand.  */
3884         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3885                 return 0;
3886
3887         /* If we filled the congestion window, do not expand.  */
3888         if (tp->packets_out >= tp->snd_cwnd)
3889                 return 0;
3890
3891         return 1;
3892 }
3893
3894 /* When incoming ACK allowed to free some skb from write_queue,
3895  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3896  * on the exit from tcp input handler.
3897  *
3898  * PROBLEM: sndbuf expansion does not work well with largesend.
3899  */
3900 static void tcp_new_space(struct sock *sk)
3901 {
3902         struct tcp_sock *tp = tcp_sk(sk);
3903
3904         if (tcp_should_expand_sndbuf(sk)) {
3905                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3906                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3907                     demanded = max_t(unsigned int, tp->snd_cwnd,
3908                                                    tp->reordering + 1);
3909                 sndmem *= 2*demanded;
3910                 if (sndmem > sk->sk_sndbuf)
3911                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3912                 tp->snd_cwnd_stamp = tcp_time_stamp;
3913         }
3914
3915         sk->sk_write_space(sk);
3916 }
3917
3918 static void tcp_check_space(struct sock *sk)
3919 {
3920         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3921                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3922                 if (sk->sk_socket &&
3923                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3924                         tcp_new_space(sk);
3925         }
3926 }
3927
3928 static inline void tcp_data_snd_check(struct sock *sk)
3929 {
3930         tcp_push_pending_frames(sk);
3931         tcp_check_space(sk);
3932 }
3933
3934 /*
3935  * Check if sending an ack is needed.
3936  */
3937 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3938 {
3939         struct tcp_sock *tp = tcp_sk(sk);
3940
3941             /* More than one full frame received... */
3942         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3943              /* ... and right edge of window advances far enough.
3944               * (tcp_recvmsg() will send ACK otherwise). Or...
3945               */
3946              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3947             /* We ACK each frame or... */
3948             tcp_in_quickack_mode(sk) ||
3949             /* We have out of order data. */
3950             (ofo_possible &&
3951              skb_peek(&tp->out_of_order_queue))) {
3952                 /* Then ack it now */
3953                 tcp_send_ack(sk);
3954         } else {
3955                 /* Else, send delayed ack. */
3956                 tcp_send_delayed_ack(sk);
3957         }
3958 }
3959
3960 static inline void tcp_ack_snd_check(struct sock *sk)
3961 {
3962         if (!inet_csk_ack_scheduled(sk)) {
3963                 /* We sent a data segment already. */
3964                 return;
3965         }
3966         __tcp_ack_snd_check(sk, 1);
3967 }
3968
3969 /*
3970  *      This routine is only called when we have urgent data
3971  *      signaled. Its the 'slow' part of tcp_urg. It could be
3972  *      moved inline now as tcp_urg is only called from one
3973  *      place. We handle URGent data wrong. We have to - as
3974  *      BSD still doesn't use the correction from RFC961.
3975  *      For 1003.1g we should support a new option TCP_STDURG to permit
3976  *      either form (or just set the sysctl tcp_stdurg).
3977  */
3978
3979 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3980 {
3981         struct tcp_sock *tp = tcp_sk(sk);
3982         u32 ptr = ntohs(th->urg_ptr);
3983
3984         if (ptr && !sysctl_tcp_stdurg)
3985                 ptr--;
3986         ptr += ntohl(th->seq);
3987
3988         /* Ignore urgent data that we've already seen and read. */
3989         if (after(tp->copied_seq, ptr))
3990                 return;
3991
3992         /* Do not replay urg ptr.
3993          *
3994          * NOTE: interesting situation not covered by specs.
3995          * Misbehaving sender may send urg ptr, pointing to segment,
3996          * which we already have in ofo queue. We are not able to fetch
3997          * such data and will stay in TCP_URG_NOTYET until will be eaten
3998          * by recvmsg(). Seems, we are not obliged to handle such wicked
3999          * situations. But it is worth to think about possibility of some
4000          * DoSes using some hypothetical application level deadlock.
4001          */
4002         if (before(ptr, tp->rcv_nxt))
4003                 return;
4004
4005         /* Do we already have a newer (or duplicate) urgent pointer? */
4006         if (tp->urg_data && !after(ptr, tp->urg_seq))
4007                 return;
4008
4009         /* Tell the world about our new urgent pointer. */
4010         sk_send_sigurg(sk);
4011
4012         /* We may be adding urgent data when the last byte read was
4013          * urgent. To do this requires some care. We cannot just ignore
4014          * tp->copied_seq since we would read the last urgent byte again
4015          * as data, nor can we alter copied_seq until this data arrives
4016          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4017          *
4018          * NOTE. Double Dutch. Rendering to plain English: author of comment
4019          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4020          * and expect that both A and B disappear from stream. This is _wrong_.
4021          * Though this happens in BSD with high probability, this is occasional.
4022          * Any application relying on this is buggy. Note also, that fix "works"
4023          * only in this artificial test. Insert some normal data between A and B and we will
4024          * decline of BSD again. Verdict: it is better to remove to trap
4025          * buggy users.
4026          */
4027         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4028             !sock_flag(sk, SOCK_URGINLINE) &&
4029             tp->copied_seq != tp->rcv_nxt) {
4030                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4031                 tp->copied_seq++;
4032                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4033                         __skb_unlink(skb, &sk->sk_receive_queue);
4034                         __kfree_skb(skb);
4035                 }
4036         }
4037
4038         tp->urg_data   = TCP_URG_NOTYET;
4039         tp->urg_seq    = ptr;
4040
4041         /* Disable header prediction. */
4042         tp->pred_flags = 0;
4043 }
4044
4045 /* This is the 'fast' part of urgent handling. */
4046 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4047 {
4048         struct tcp_sock *tp = tcp_sk(sk);
4049
4050         /* Check if we get a new urgent pointer - normally not. */
4051         if (th->urg)
4052                 tcp_check_urg(sk,th);
4053
4054         /* Do we wait for any urgent data? - normally not... */
4055         if (tp->urg_data == TCP_URG_NOTYET) {
4056                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4057                           th->syn;
4058
4059                 /* Is the urgent pointer pointing into this packet? */
4060                 if (ptr < skb->len) {
4061                         u8 tmp;
4062                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4063                                 BUG();
4064                         tp->urg_data = TCP_URG_VALID | tmp;
4065                         if (!sock_flag(sk, SOCK_DEAD))
4066                                 sk->sk_data_ready(sk, 0);
4067                 }
4068         }
4069 }
4070
4071 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4072 {
4073         struct tcp_sock *tp = tcp_sk(sk);
4074         int chunk = skb->len - hlen;
4075         int err;
4076
4077         local_bh_enable();
4078         if (skb_csum_unnecessary(skb))
4079                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4080         else
4081                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4082                                                        tp->ucopy.iov);
4083
4084         if (!err) {
4085                 tp->ucopy.len -= chunk;
4086                 tp->copied_seq += chunk;
4087                 tcp_rcv_space_adjust(sk);
4088         }
4089
4090         local_bh_disable();
4091         return err;
4092 }
4093
4094 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4095 {
4096         __sum16 result;
4097
4098         if (sock_owned_by_user(sk)) {
4099                 local_bh_enable();
4100                 result = __tcp_checksum_complete(skb);
4101                 local_bh_disable();
4102         } else {
4103                 result = __tcp_checksum_complete(skb);
4104         }
4105         return result;
4106 }
4107
4108 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4109 {
4110         return !skb_csum_unnecessary(skb) &&
4111                 __tcp_checksum_complete_user(sk, skb);
4112 }
4113
4114 #ifdef CONFIG_NET_DMA
4115 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4116 {
4117         struct tcp_sock *tp = tcp_sk(sk);
4118         int chunk = skb->len - hlen;
4119         int dma_cookie;
4120         int copied_early = 0;
4121
4122         if (tp->ucopy.wakeup)
4123                 return 0;
4124
4125         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4126                 tp->ucopy.dma_chan = get_softnet_dma();
4127
4128         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4129
4130                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4131                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4132
4133                 if (dma_cookie < 0)
4134                         goto out;
4135
4136                 tp->ucopy.dma_cookie = dma_cookie;
4137                 copied_early = 1;
4138
4139                 tp->ucopy.len -= chunk;
4140                 tp->copied_seq += chunk;
4141                 tcp_rcv_space_adjust(sk);
4142
4143                 if ((tp->ucopy.len == 0) ||
4144                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4145                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4146                         tp->ucopy.wakeup = 1;
4147                         sk->sk_data_ready(sk, 0);
4148                 }
4149         } else if (chunk > 0) {
4150                 tp->ucopy.wakeup = 1;
4151                 sk->sk_data_ready(sk, 0);
4152         }
4153 out:
4154         return copied_early;
4155 }
4156 #endif /* CONFIG_NET_DMA */
4157
4158 /*
4159  *      TCP receive function for the ESTABLISHED state.
4160  *
4161  *      It is split into a fast path and a slow path. The fast path is
4162  *      disabled when:
4163  *      - A zero window was announced from us - zero window probing
4164  *        is only handled properly in the slow path.
4165  *      - Out of order segments arrived.
4166  *      - Urgent data is expected.
4167  *      - There is no buffer space left
4168  *      - Unexpected TCP flags/window values/header lengths are received
4169  *        (detected by checking the TCP header against pred_flags)
4170  *      - Data is sent in both directions. Fast path only supports pure senders
4171  *        or pure receivers (this means either the sequence number or the ack
4172  *        value must stay constant)
4173  *      - Unexpected TCP option.
4174  *
4175  *      When these conditions are not satisfied it drops into a standard
4176  *      receive procedure patterned after RFC793 to handle all cases.
4177  *      The first three cases are guaranteed by proper pred_flags setting,
4178  *      the rest is checked inline. Fast processing is turned on in
4179  *      tcp_data_queue when everything is OK.
4180  */
4181 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4182                         struct tcphdr *th, unsigned len)
4183 {
4184         struct tcp_sock *tp = tcp_sk(sk);
4185
4186         /*
4187          *      Header prediction.
4188          *      The code loosely follows the one in the famous
4189          *      "30 instruction TCP receive" Van Jacobson mail.
4190          *
4191          *      Van's trick is to deposit buffers into socket queue
4192          *      on a device interrupt, to call tcp_recv function
4193          *      on the receive process context and checksum and copy
4194          *      the buffer to user space. smart...
4195          *
4196          *      Our current scheme is not silly either but we take the
4197          *      extra cost of the net_bh soft interrupt processing...
4198          *      We do checksum and copy also but from device to kernel.
4199          */
4200
4201         tp->rx_opt.saw_tstamp = 0;
4202
4203         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4204          *      if header_prediction is to be made
4205          *      'S' will always be tp->tcp_header_len >> 2
4206          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4207          *  turn it off (when there are holes in the receive
4208          *       space for instance)
4209          *      PSH flag is ignored.
4210          */
4211
4212         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4213                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4214                 int tcp_header_len = tp->tcp_header_len;
4215
4216                 /* Timestamp header prediction: tcp_header_len
4217                  * is automatically equal to th->doff*4 due to pred_flags
4218                  * match.
4219                  */
4220
4221                 /* Check timestamp */
4222                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4223                         __be32 *ptr = (__be32 *)(th + 1);
4224
4225                         /* No? Slow path! */
4226                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4227                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4228                                 goto slow_path;
4229
4230                         tp->rx_opt.saw_tstamp = 1;
4231                         ++ptr;
4232                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4233                         ++ptr;
4234                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4235
4236                         /* If PAWS failed, check it more carefully in slow path */
4237                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4238                                 goto slow_path;
4239
4240                         /* DO NOT update ts_recent here, if checksum fails
4241                          * and timestamp was corrupted part, it will result
4242                          * in a hung connection since we will drop all
4243                          * future packets due to the PAWS test.
4244                          */
4245                 }
4246
4247                 if (len <= tcp_header_len) {
4248                         /* Bulk data transfer: sender */
4249                         if (len == tcp_header_len) {
4250                                 /* Predicted packet is in window by definition.
4251                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4252                                  * Hence, check seq<=rcv_wup reduces to:
4253                                  */
4254                                 if (tcp_header_len ==
4255                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4256                                     tp->rcv_nxt == tp->rcv_wup)
4257                                         tcp_store_ts_recent(tp);
4258
4259                                 /* We know that such packets are checksummed
4260                                  * on entry.
4261                                  */
4262                                 tcp_ack(sk, skb, 0);
4263                                 __kfree_skb(skb);
4264                                 tcp_data_snd_check(sk);
4265                                 return 0;
4266                         } else { /* Header too small */
4267                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4268                                 goto discard;
4269                         }
4270                 } else {
4271                         int eaten = 0;
4272                         int copied_early = 0;
4273
4274                         if (tp->copied_seq == tp->rcv_nxt &&
4275                             len - tcp_header_len <= tp->ucopy.len) {
4276 #ifdef CONFIG_NET_DMA
4277                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4278                                         copied_early = 1;
4279                                         eaten = 1;
4280                                 }
4281 #endif
4282                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4283                                         __set_current_state(TASK_RUNNING);
4284
4285                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4286                                                 eaten = 1;
4287                                 }
4288                                 if (eaten) {
4289                                         /* Predicted packet is in window by definition.
4290                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4291                                          * Hence, check seq<=rcv_wup reduces to:
4292                                          */
4293                                         if (tcp_header_len ==
4294                                             (sizeof(struct tcphdr) +
4295                                              TCPOLEN_TSTAMP_ALIGNED) &&
4296                                             tp->rcv_nxt == tp->rcv_wup)
4297                                                 tcp_store_ts_recent(tp);
4298
4299                                         tcp_rcv_rtt_measure_ts(sk, skb);
4300
4301                                         __skb_pull(skb, tcp_header_len);
4302                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4303                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4304                                 }
4305                                 if (copied_early)
4306                                         tcp_cleanup_rbuf(sk, skb->len);
4307                         }
4308                         if (!eaten) {
4309                                 if (tcp_checksum_complete_user(sk, skb))
4310                                         goto csum_error;
4311
4312                                 /* Predicted packet is in window by definition.
4313                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4314                                  * Hence, check seq<=rcv_wup reduces to:
4315                                  */
4316                                 if (tcp_header_len ==
4317                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4318                                     tp->rcv_nxt == tp->rcv_wup)
4319                                         tcp_store_ts_recent(tp);
4320
4321                                 tcp_rcv_rtt_measure_ts(sk, skb);
4322
4323                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4324                                         goto step5;
4325
4326                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4327
4328                                 /* Bulk data transfer: receiver */
4329                                 __skb_pull(skb,tcp_header_len);
4330                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4331                                 sk_stream_set_owner_r(skb, sk);
4332                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4333                         }
4334
4335                         tcp_event_data_recv(sk, skb);
4336
4337                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4338                                 /* Well, only one small jumplet in fast path... */
4339                                 tcp_ack(sk, skb, FLAG_DATA);
4340                                 tcp_data_snd_check(sk);
4341                                 if (!inet_csk_ack_scheduled(sk))
4342                                         goto no_ack;
4343                         }
4344
4345                         __tcp_ack_snd_check(sk, 0);
4346 no_ack:
4347 #ifdef CONFIG_NET_DMA
4348                         if (copied_early)
4349                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4350                         else
4351 #endif
4352                         if (eaten)
4353                                 __kfree_skb(skb);
4354                         else
4355                                 sk->sk_data_ready(sk, 0);
4356                         return 0;
4357                 }
4358         }
4359
4360 slow_path:
4361         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4362                 goto csum_error;
4363
4364         /*
4365          * RFC1323: H1. Apply PAWS check first.
4366          */
4367         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4368             tcp_paws_discard(sk, skb)) {
4369                 if (!th->rst) {
4370                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4371                         tcp_send_dupack(sk, skb);
4372                         goto discard;
4373                 }
4374                 /* Resets are accepted even if PAWS failed.
4375
4376                    ts_recent update must be made after we are sure
4377                    that the packet is in window.
4378                  */
4379         }
4380
4381         /*
4382          *      Standard slow path.
4383          */
4384
4385         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4386                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4387                  * (RST) segments are validated by checking their SEQ-fields."
4388                  * And page 69: "If an incoming segment is not acceptable,
4389                  * an acknowledgment should be sent in reply (unless the RST bit
4390                  * is set, if so drop the segment and return)".
4391                  */
4392                 if (!th->rst)
4393                         tcp_send_dupack(sk, skb);
4394                 goto discard;
4395         }
4396
4397         if (th->rst) {
4398                 tcp_reset(sk);
4399                 goto discard;
4400         }
4401
4402         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4403
4404         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4405                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4406                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4407                 tcp_reset(sk);
4408                 return 1;
4409         }
4410
4411 step5:
4412         if (th->ack)
4413                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4414
4415         tcp_rcv_rtt_measure_ts(sk, skb);
4416
4417         /* Process urgent data. */
4418         tcp_urg(sk, skb, th);
4419
4420         /* step 7: process the segment text */
4421         tcp_data_queue(sk, skb);
4422
4423         tcp_data_snd_check(sk);
4424         tcp_ack_snd_check(sk);
4425         return 0;
4426
4427 csum_error:
4428         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4429
4430 discard:
4431         __kfree_skb(skb);
4432         return 0;
4433 }
4434
4435 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4436                                          struct tcphdr *th, unsigned len)
4437 {
4438         struct tcp_sock *tp = tcp_sk(sk);
4439         struct inet_connection_sock *icsk = inet_csk(sk);
4440         int saved_clamp = tp->rx_opt.mss_clamp;
4441
4442         tcp_parse_options(skb, &tp->rx_opt, 0);
4443
4444         if (th->ack) {
4445                 /* rfc793:
4446                  * "If the state is SYN-SENT then
4447                  *    first check the ACK bit
4448                  *      If the ACK bit is set
4449                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4450                  *        a reset (unless the RST bit is set, if so drop
4451                  *        the segment and return)"
4452                  *
4453                  *  We do not send data with SYN, so that RFC-correct
4454                  *  test reduces to:
4455                  */
4456                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4457                         goto reset_and_undo;
4458
4459                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4460                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4461                              tcp_time_stamp)) {
4462                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4463                         goto reset_and_undo;
4464                 }
4465
4466                 /* Now ACK is acceptable.
4467                  *
4468                  * "If the RST bit is set
4469                  *    If the ACK was acceptable then signal the user "error:
4470                  *    connection reset", drop the segment, enter CLOSED state,
4471                  *    delete TCB, and return."
4472                  */
4473
4474                 if (th->rst) {
4475                         tcp_reset(sk);
4476                         goto discard;
4477                 }
4478
4479                 /* rfc793:
4480                  *   "fifth, if neither of the SYN or RST bits is set then
4481                  *    drop the segment and return."
4482                  *
4483                  *    See note below!
4484                  *                                        --ANK(990513)
4485                  */
4486                 if (!th->syn)
4487                         goto discard_and_undo;
4488
4489                 /* rfc793:
4490                  *   "If the SYN bit is on ...
4491                  *    are acceptable then ...
4492                  *    (our SYN has been ACKed), change the connection
4493                  *    state to ESTABLISHED..."
4494                  */
4495
4496                 TCP_ECN_rcv_synack(tp, th);
4497
4498                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4499                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4500
4501                 /* Ok.. it's good. Set up sequence numbers and
4502                  * move to established.
4503                  */
4504                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4505                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4506
4507                 /* RFC1323: The window in SYN & SYN/ACK segments is
4508                  * never scaled.
4509                  */
4510                 tp->snd_wnd = ntohs(th->window);
4511                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4512
4513                 if (!tp->rx_opt.wscale_ok) {
4514                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4515                         tp->window_clamp = min(tp->window_clamp, 65535U);
4516                 }
4517
4518                 if (tp->rx_opt.saw_tstamp) {
4519                         tp->rx_opt.tstamp_ok       = 1;
4520                         tp->tcp_header_len =
4521                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4522                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4523                         tcp_store_ts_recent(tp);
4524                 } else {
4525                         tp->tcp_header_len = sizeof(struct tcphdr);
4526                 }
4527
4528                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4529                         tp->rx_opt.sack_ok |= 2;
4530
4531                 tcp_mtup_init(sk);
4532                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4533                 tcp_initialize_rcv_mss(sk);
4534
4535                 /* Remember, tcp_poll() does not lock socket!
4536                  * Change state from SYN-SENT only after copied_seq
4537                  * is initialized. */
4538                 tp->copied_seq = tp->rcv_nxt;
4539                 smp_mb();
4540                 tcp_set_state(sk, TCP_ESTABLISHED);
4541
4542                 security_inet_conn_established(sk, skb);
4543
4544                 /* Make sure socket is routed, for correct metrics.  */
4545                 icsk->icsk_af_ops->rebuild_header(sk);
4546
4547                 tcp_init_metrics(sk);
4548
4549                 tcp_init_congestion_control(sk);
4550
4551                 /* Prevent spurious tcp_cwnd_restart() on first data
4552                  * packet.
4553                  */
4554                 tp->lsndtime = tcp_time_stamp;
4555
4556                 tcp_init_buffer_space(sk);
4557
4558                 if (sock_flag(sk, SOCK_KEEPOPEN))
4559                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4560
4561                 if (!tp->rx_opt.snd_wscale)
4562                         __tcp_fast_path_on(tp, tp->snd_wnd);
4563                 else
4564                         tp->pred_flags = 0;
4565
4566                 if (!sock_flag(sk, SOCK_DEAD)) {
4567                         sk->sk_state_change(sk);
4568                         sk_wake_async(sk, 0, POLL_OUT);
4569                 }
4570
4571                 if (sk->sk_write_pending ||
4572                     icsk->icsk_accept_queue.rskq_defer_accept ||
4573                     icsk->icsk_ack.pingpong) {
4574                         /* Save one ACK. Data will be ready after
4575                          * several ticks, if write_pending is set.
4576                          *
4577                          * It may be deleted, but with this feature tcpdumps
4578                          * look so _wonderfully_ clever, that I was not able
4579                          * to stand against the temptation 8)     --ANK
4580                          */
4581                         inet_csk_schedule_ack(sk);
4582                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4583                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4584                         tcp_incr_quickack(sk);
4585                         tcp_enter_quickack_mode(sk);
4586                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4587                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4588
4589 discard:
4590                         __kfree_skb(skb);
4591                         return 0;
4592                 } else {
4593                         tcp_send_ack(sk);
4594                 }
4595                 return -1;
4596         }
4597
4598         /* No ACK in the segment */
4599
4600         if (th->rst) {
4601                 /* rfc793:
4602                  * "If the RST bit is set
4603                  *
4604                  *      Otherwise (no ACK) drop the segment and return."
4605                  */
4606
4607                 goto discard_and_undo;
4608         }
4609
4610         /* PAWS check. */
4611         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4612                 goto discard_and_undo;
4613
4614         if (th->syn) {
4615                 /* We see SYN without ACK. It is attempt of
4616                  * simultaneous connect with crossed SYNs.
4617                  * Particularly, it can be connect to self.
4618                  */
4619                 tcp_set_state(sk, TCP_SYN_RECV);
4620
4621                 if (tp->rx_opt.saw_tstamp) {
4622                         tp->rx_opt.tstamp_ok = 1;
4623                         tcp_store_ts_recent(tp);
4624                         tp->tcp_header_len =
4625                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4626                 } else {
4627                         tp->tcp_header_len = sizeof(struct tcphdr);
4628                 }
4629
4630                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4631                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4632
4633                 /* RFC1323: The window in SYN & SYN/ACK segments is
4634                  * never scaled.
4635                  */
4636                 tp->snd_wnd    = ntohs(th->window);
4637                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4638                 tp->max_window = tp->snd_wnd;
4639
4640                 TCP_ECN_rcv_syn(tp, th);
4641
4642                 tcp_mtup_init(sk);
4643                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4644                 tcp_initialize_rcv_mss(sk);
4645
4646
4647                 tcp_send_synack(sk);
4648 #if 0
4649                 /* Note, we could accept data and URG from this segment.
4650                  * There are no obstacles to make this.
4651                  *
4652                  * However, if we ignore data in ACKless segments sometimes,
4653                  * we have no reasons to accept it sometimes.
4654                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4655                  * is not flawless. So, discard packet for sanity.
4656                  * Uncomment this return to process the data.
4657                  */
4658                 return -1;
4659 #else
4660                 goto discard;
4661 #endif
4662         }
4663         /* "fifth, if neither of the SYN or RST bits is set then
4664          * drop the segment and return."
4665          */
4666
4667 discard_and_undo:
4668         tcp_clear_options(&tp->rx_opt);
4669         tp->rx_opt.mss_clamp = saved_clamp;
4670         goto discard;
4671
4672 reset_and_undo:
4673         tcp_clear_options(&tp->rx_opt);
4674         tp->rx_opt.mss_clamp = saved_clamp;
4675         return 1;
4676 }
4677
4678
4679 /*
4680  *      This function implements the receiving procedure of RFC 793 for
4681  *      all states except ESTABLISHED and TIME_WAIT.
4682  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4683  *      address independent.
4684  */
4685
4686 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4687                           struct tcphdr *th, unsigned len)
4688 {
4689         struct tcp_sock *tp = tcp_sk(sk);
4690         struct inet_connection_sock *icsk = inet_csk(sk);
4691         int queued = 0;
4692
4693         tp->rx_opt.saw_tstamp = 0;
4694
4695         switch (sk->sk_state) {
4696         case TCP_CLOSE:
4697                 goto discard;
4698
4699         case TCP_LISTEN:
4700                 if (th->ack)
4701                         return 1;
4702
4703                 if (th->rst)
4704                         goto discard;
4705
4706                 if (th->syn) {
4707                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4708                                 return 1;
4709
4710                         /* Now we have several options: In theory there is
4711                          * nothing else in the frame. KA9Q has an option to
4712                          * send data with the syn, BSD accepts data with the
4713                          * syn up to the [to be] advertised window and
4714                          * Solaris 2.1 gives you a protocol error. For now
4715                          * we just ignore it, that fits the spec precisely
4716                          * and avoids incompatibilities. It would be nice in
4717                          * future to drop through and process the data.
4718                          *
4719                          * Now that TTCP is starting to be used we ought to
4720                          * queue this data.
4721                          * But, this leaves one open to an easy denial of
4722                          * service attack, and SYN cookies can't defend
4723                          * against this problem. So, we drop the data
4724                          * in the interest of security over speed unless
4725                          * it's still in use.
4726                          */
4727                         kfree_skb(skb);
4728                         return 0;
4729                 }
4730                 goto discard;
4731
4732         case TCP_SYN_SENT:
4733                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4734                 if (queued >= 0)
4735                         return queued;
4736
4737                 /* Do step6 onward by hand. */
4738                 tcp_urg(sk, skb, th);
4739                 __kfree_skb(skb);
4740                 tcp_data_snd_check(sk);
4741                 return 0;
4742         }
4743
4744         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4745             tcp_paws_discard(sk, skb)) {
4746                 if (!th->rst) {
4747                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4748                         tcp_send_dupack(sk, skb);
4749                         goto discard;
4750                 }
4751                 /* Reset is accepted even if it did not pass PAWS. */
4752         }
4753
4754         /* step 1: check sequence number */
4755         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4756                 if (!th->rst)
4757                         tcp_send_dupack(sk, skb);
4758                 goto discard;
4759         }
4760
4761         /* step 2: check RST bit */
4762         if (th->rst) {
4763                 tcp_reset(sk);
4764                 goto discard;
4765         }
4766
4767         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4768
4769         /* step 3: check security and precedence [ignored] */
4770
4771         /*      step 4:
4772          *
4773          *      Check for a SYN in window.
4774          */
4775         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4776                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4777                 tcp_reset(sk);
4778                 return 1;
4779         }
4780
4781         /* step 5: check the ACK field */
4782         if (th->ack) {
4783                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4784
4785                 switch (sk->sk_state) {
4786                 case TCP_SYN_RECV:
4787                         if (acceptable) {
4788                                 tp->copied_seq = tp->rcv_nxt;
4789                                 smp_mb();
4790                                 tcp_set_state(sk, TCP_ESTABLISHED);
4791                                 sk->sk_state_change(sk);
4792
4793                                 /* Note, that this wakeup is only for marginal
4794                                  * crossed SYN case. Passively open sockets
4795                                  * are not waked up, because sk->sk_sleep ==
4796                                  * NULL and sk->sk_socket == NULL.
4797                                  */
4798                                 if (sk->sk_socket) {
4799                                         sk_wake_async(sk,0,POLL_OUT);
4800                                 }
4801
4802                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4803                                 tp->snd_wnd = ntohs(th->window) <<
4804                                               tp->rx_opt.snd_wscale;
4805                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4806                                             TCP_SKB_CB(skb)->seq);
4807
4808                                 /* tcp_ack considers this ACK as duplicate
4809                                  * and does not calculate rtt.
4810                                  * Fix it at least with timestamps.
4811                                  */
4812                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4813                                     !tp->srtt)
4814                                         tcp_ack_saw_tstamp(sk, 0);
4815
4816                                 if (tp->rx_opt.tstamp_ok)
4817                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4818
4819                                 /* Make sure socket is routed, for
4820                                  * correct metrics.
4821                                  */
4822                                 icsk->icsk_af_ops->rebuild_header(sk);
4823
4824                                 tcp_init_metrics(sk);
4825
4826                                 tcp_init_congestion_control(sk);
4827
4828                                 /* Prevent spurious tcp_cwnd_restart() on
4829                                  * first data packet.
4830                                  */
4831                                 tp->lsndtime = tcp_time_stamp;
4832
4833                                 tcp_mtup_init(sk);
4834                                 tcp_initialize_rcv_mss(sk);
4835                                 tcp_init_buffer_space(sk);
4836                                 tcp_fast_path_on(tp);
4837                         } else {
4838                                 return 1;
4839                         }
4840                         break;
4841
4842                 case TCP_FIN_WAIT1:
4843                         if (tp->snd_una == tp->write_seq) {
4844                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4845                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4846                                 dst_confirm(sk->sk_dst_cache);
4847
4848                                 if (!sock_flag(sk, SOCK_DEAD))
4849                                         /* Wake up lingering close() */
4850                                         sk->sk_state_change(sk);
4851                                 else {
4852                                         int tmo;
4853
4854                                         if (tp->linger2 < 0 ||
4855                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4856                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4857                                                 tcp_done(sk);
4858                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4859                                                 return 1;
4860                                         }
4861
4862                                         tmo = tcp_fin_time(sk);
4863                                         if (tmo > TCP_TIMEWAIT_LEN) {
4864                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4865                                         } else if (th->fin || sock_owned_by_user(sk)) {
4866                                                 /* Bad case. We could lose such FIN otherwise.
4867                                                  * It is not a big problem, but it looks confusing
4868                                                  * and not so rare event. We still can lose it now,
4869                                                  * if it spins in bh_lock_sock(), but it is really
4870                                                  * marginal case.
4871                                                  */
4872                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4873                                         } else {
4874                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4875                                                 goto discard;
4876                                         }
4877                                 }
4878                         }
4879                         break;
4880
4881                 case TCP_CLOSING:
4882                         if (tp->snd_una == tp->write_seq) {
4883                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4884                                 goto discard;
4885                         }
4886                         break;
4887
4888                 case TCP_LAST_ACK:
4889                         if (tp->snd_una == tp->write_seq) {
4890                                 tcp_update_metrics(sk);
4891                                 tcp_done(sk);
4892                                 goto discard;
4893                         }
4894                         break;
4895                 }
4896         } else
4897                 goto discard;
4898
4899         /* step 6: check the URG bit */
4900         tcp_urg(sk, skb, th);
4901
4902         /* step 7: process the segment text */
4903         switch (sk->sk_state) {
4904         case TCP_CLOSE_WAIT:
4905         case TCP_CLOSING:
4906         case TCP_LAST_ACK:
4907                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4908                         break;
4909         case TCP_FIN_WAIT1:
4910         case TCP_FIN_WAIT2:
4911                 /* RFC 793 says to queue data in these states,
4912                  * RFC 1122 says we MUST send a reset.
4913                  * BSD 4.4 also does reset.
4914                  */
4915                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4916                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4917                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4918                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4919                                 tcp_reset(sk);
4920                                 return 1;
4921                         }
4922                 }
4923                 /* Fall through */
4924         case TCP_ESTABLISHED:
4925                 tcp_data_queue(sk, skb);
4926                 queued = 1;
4927                 break;
4928         }
4929
4930         /* tcp_data could move socket to TIME-WAIT */
4931         if (sk->sk_state != TCP_CLOSE) {
4932                 tcp_data_snd_check(sk);
4933                 tcp_ack_snd_check(sk);
4934         }
4935
4936         if (!queued) {
4937 discard:
4938                 __kfree_skb(skb);
4939         }
4940         return 0;
4941 }
4942
4943 EXPORT_SYMBOL(sysctl_tcp_ecn);
4944 EXPORT_SYMBOL(sysctl_tcp_reordering);
4945 EXPORT_SYMBOL(tcp_parse_options);
4946 EXPORT_SYMBOL(tcp_rcv_established);
4947 EXPORT_SYMBOL(tcp_rcv_state_process);
4948 EXPORT_SYMBOL(tcp_initialize_rcv_mss);