d1604f59d77e36fdc0d14a839807b0e3fe442f8c
[linux-2.6.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105
106 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
109 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
110
111 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
112 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
113 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
114
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
116
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118
119 /* Adapt the MSS value used to make delayed ack decision to the
120  * real world.
121  */
122 static void tcp_measure_rcv_mss(struct sock *sk,
123                                 const struct sk_buff *skb)
124 {
125         struct inet_connection_sock *icsk = inet_csk(sk);
126         const unsigned int lss = icsk->icsk_ack.last_seg_size;
127         unsigned int len;
128
129         icsk->icsk_ack.last_seg_size = 0;
130
131         /* skb->len may jitter because of SACKs, even if peer
132          * sends good full-sized frames.
133          */
134         len = skb_shinfo(skb)->gso_size ?: skb->len;
135         if (len >= icsk->icsk_ack.rcv_mss) {
136                 icsk->icsk_ack.rcv_mss = len;
137         } else {
138                 /* Otherwise, we make more careful check taking into account,
139                  * that SACKs block is variable.
140                  *
141                  * "len" is invariant segment length, including TCP header.
142                  */
143                 len += skb->data - skb_transport_header(skb);
144                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
145                     /* If PSH is not set, packet should be
146                      * full sized, provided peer TCP is not badly broken.
147                      * This observation (if it is correct 8)) allows
148                      * to handle super-low mtu links fairly.
149                      */
150                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
151                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
152                         /* Subtract also invariant (if peer is RFC compliant),
153                          * tcp header plus fixed timestamp option length.
154                          * Resulting "len" is MSS free of SACK jitter.
155                          */
156                         len -= tcp_sk(sk)->tcp_header_len;
157                         icsk->icsk_ack.last_seg_size = len;
158                         if (len == lss) {
159                                 icsk->icsk_ack.rcv_mss = len;
160                                 return;
161                         }
162                 }
163                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
164                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
165                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
166         }
167 }
168
169 static void tcp_incr_quickack(struct sock *sk)
170 {
171         struct inet_connection_sock *icsk = inet_csk(sk);
172         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
173
174         if (quickacks==0)
175                 quickacks=2;
176         if (quickacks > icsk->icsk_ack.quick)
177                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
178 }
179
180 void tcp_enter_quickack_mode(struct sock *sk)
181 {
182         struct inet_connection_sock *icsk = inet_csk(sk);
183         tcp_incr_quickack(sk);
184         icsk->icsk_ack.pingpong = 0;
185         icsk->icsk_ack.ato = TCP_ATO_MIN;
186 }
187
188 /* Send ACKs quickly, if "quick" count is not exhausted
189  * and the session is not interactive.
190  */
191
192 static inline int tcp_in_quickack_mode(const struct sock *sk)
193 {
194         const struct inet_connection_sock *icsk = inet_csk(sk);
195         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
196 }
197
198 /* Buffer size and advertised window tuning.
199  *
200  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
201  */
202
203 static void tcp_fixup_sndbuf(struct sock *sk)
204 {
205         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
206                      sizeof(struct sk_buff);
207
208         if (sk->sk_sndbuf < 3 * sndmem)
209                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
210 }
211
212 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
213  *
214  * All tcp_full_space() is split to two parts: "network" buffer, allocated
215  * forward and advertised in receiver window (tp->rcv_wnd) and
216  * "application buffer", required to isolate scheduling/application
217  * latencies from network.
218  * window_clamp is maximal advertised window. It can be less than
219  * tcp_full_space(), in this case tcp_full_space() - window_clamp
220  * is reserved for "application" buffer. The less window_clamp is
221  * the smoother our behaviour from viewpoint of network, but the lower
222  * throughput and the higher sensitivity of the connection to losses. 8)
223  *
224  * rcv_ssthresh is more strict window_clamp used at "slow start"
225  * phase to predict further behaviour of this connection.
226  * It is used for two goals:
227  * - to enforce header prediction at sender, even when application
228  *   requires some significant "application buffer". It is check #1.
229  * - to prevent pruning of receive queue because of misprediction
230  *   of receiver window. Check #2.
231  *
232  * The scheme does not work when sender sends good segments opening
233  * window and then starts to feed us spaghetti. But it should work
234  * in common situations. Otherwise, we have to rely on queue collapsing.
235  */
236
237 /* Slow part of check#2. */
238 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
239                              const struct sk_buff *skb)
240 {
241         /* Optimize this! */
242         int truesize = tcp_win_from_space(skb->truesize)/2;
243         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
244
245         while (tp->rcv_ssthresh <= window) {
246                 if (truesize <= skb->len)
247                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
248
249                 truesize >>= 1;
250                 window >>= 1;
251         }
252         return 0;
253 }
254
255 static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
256                             struct sk_buff *skb)
257 {
258         /* Check #1 */
259         if (tp->rcv_ssthresh < tp->window_clamp &&
260             (int)tp->rcv_ssthresh < tcp_space(sk) &&
261             !tcp_memory_pressure) {
262                 int incr;
263
264                 /* Check #2. Increase window, if skb with such overhead
265                  * will fit to rcvbuf in future.
266                  */
267                 if (tcp_win_from_space(skb->truesize) <= skb->len)
268                         incr = 2*tp->advmss;
269                 else
270                         incr = __tcp_grow_window(sk, tp, skb);
271
272                 if (incr) {
273                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
274                         inet_csk(sk)->icsk_ack.quick |= 1;
275                 }
276         }
277 }
278
279 /* 3. Tuning rcvbuf, when connection enters established state. */
280
281 static void tcp_fixup_rcvbuf(struct sock *sk)
282 {
283         struct tcp_sock *tp = tcp_sk(sk);
284         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
285
286         /* Try to select rcvbuf so that 4 mss-sized segments
287          * will fit to window and corresponding skbs will fit to our rcvbuf.
288          * (was 3; 4 is minimum to allow fast retransmit to work.)
289          */
290         while (tcp_win_from_space(rcvmem) < tp->advmss)
291                 rcvmem += 128;
292         if (sk->sk_rcvbuf < 4 * rcvmem)
293                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
294 }
295
296 /* 4. Try to fixup all. It is made immediately after connection enters
297  *    established state.
298  */
299 static void tcp_init_buffer_space(struct sock *sk)
300 {
301         struct tcp_sock *tp = tcp_sk(sk);
302         int maxwin;
303
304         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
305                 tcp_fixup_rcvbuf(sk);
306         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
307                 tcp_fixup_sndbuf(sk);
308
309         tp->rcvq_space.space = tp->rcv_wnd;
310
311         maxwin = tcp_full_space(sk);
312
313         if (tp->window_clamp >= maxwin) {
314                 tp->window_clamp = maxwin;
315
316                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
317                         tp->window_clamp = max(maxwin -
318                                                (maxwin >> sysctl_tcp_app_win),
319                                                4 * tp->advmss);
320         }
321
322         /* Force reservation of one segment. */
323         if (sysctl_tcp_app_win &&
324             tp->window_clamp > 2 * tp->advmss &&
325             tp->window_clamp + tp->advmss > maxwin)
326                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
327
328         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
329         tp->snd_cwnd_stamp = tcp_time_stamp;
330 }
331
332 /* 5. Recalculate window clamp after socket hit its memory bounds. */
333 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
334 {
335         struct inet_connection_sock *icsk = inet_csk(sk);
336
337         icsk->icsk_ack.quick = 0;
338
339         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
340             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
341             !tcp_memory_pressure &&
342             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
343                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
344                                     sysctl_tcp_rmem[2]);
345         }
346         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
347                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
348 }
349
350
351 /* Initialize RCV_MSS value.
352  * RCV_MSS is an our guess about MSS used by the peer.
353  * We haven't any direct information about the MSS.
354  * It's better to underestimate the RCV_MSS rather than overestimate.
355  * Overestimations make us ACKing less frequently than needed.
356  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
357  */
358 void tcp_initialize_rcv_mss(struct sock *sk)
359 {
360         struct tcp_sock *tp = tcp_sk(sk);
361         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
362
363         hint = min(hint, tp->rcv_wnd/2);
364         hint = min(hint, TCP_MIN_RCVMSS);
365         hint = max(hint, TCP_MIN_MSS);
366
367         inet_csk(sk)->icsk_ack.rcv_mss = hint;
368 }
369
370 /* Receiver "autotuning" code.
371  *
372  * The algorithm for RTT estimation w/o timestamps is based on
373  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
374  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
375  *
376  * More detail on this code can be found at
377  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
378  * though this reference is out of date.  A new paper
379  * is pending.
380  */
381 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
382 {
383         u32 new_sample = tp->rcv_rtt_est.rtt;
384         long m = sample;
385
386         if (m == 0)
387                 m = 1;
388
389         if (new_sample != 0) {
390                 /* If we sample in larger samples in the non-timestamp
391                  * case, we could grossly overestimate the RTT especially
392                  * with chatty applications or bulk transfer apps which
393                  * are stalled on filesystem I/O.
394                  *
395                  * Also, since we are only going for a minimum in the
396                  * non-timestamp case, we do not smooth things out
397                  * else with timestamps disabled convergence takes too
398                  * long.
399                  */
400                 if (!win_dep) {
401                         m -= (new_sample >> 3);
402                         new_sample += m;
403                 } else if (m < new_sample)
404                         new_sample = m << 3;
405         } else {
406                 /* No previous measure. */
407                 new_sample = m << 3;
408         }
409
410         if (tp->rcv_rtt_est.rtt != new_sample)
411                 tp->rcv_rtt_est.rtt = new_sample;
412 }
413
414 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
415 {
416         if (tp->rcv_rtt_est.time == 0)
417                 goto new_measure;
418         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
419                 return;
420         tcp_rcv_rtt_update(tp,
421                            jiffies - tp->rcv_rtt_est.time,
422                            1);
423
424 new_measure:
425         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
426         tp->rcv_rtt_est.time = tcp_time_stamp;
427 }
428
429 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
430 {
431         struct tcp_sock *tp = tcp_sk(sk);
432         if (tp->rx_opt.rcv_tsecr &&
433             (TCP_SKB_CB(skb)->end_seq -
434              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
435                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
436 }
437
438 /*
439  * This function should be called every time data is copied to user space.
440  * It calculates the appropriate TCP receive buffer space.
441  */
442 void tcp_rcv_space_adjust(struct sock *sk)
443 {
444         struct tcp_sock *tp = tcp_sk(sk);
445         int time;
446         int space;
447
448         if (tp->rcvq_space.time == 0)
449                 goto new_measure;
450
451         time = tcp_time_stamp - tp->rcvq_space.time;
452         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
453             tp->rcv_rtt_est.rtt == 0)
454                 return;
455
456         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
457
458         space = max(tp->rcvq_space.space, space);
459
460         if (tp->rcvq_space.space != space) {
461                 int rcvmem;
462
463                 tp->rcvq_space.space = space;
464
465                 if (sysctl_tcp_moderate_rcvbuf &&
466                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
467                         int new_clamp = space;
468
469                         /* Receive space grows, normalize in order to
470                          * take into account packet headers and sk_buff
471                          * structure overhead.
472                          */
473                         space /= tp->advmss;
474                         if (!space)
475                                 space = 1;
476                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
477                                   16 + sizeof(struct sk_buff));
478                         while (tcp_win_from_space(rcvmem) < tp->advmss)
479                                 rcvmem += 128;
480                         space *= rcvmem;
481                         space = min(space, sysctl_tcp_rmem[2]);
482                         if (space > sk->sk_rcvbuf) {
483                                 sk->sk_rcvbuf = space;
484
485                                 /* Make the window clamp follow along.  */
486                                 tp->window_clamp = new_clamp;
487                         }
488                 }
489         }
490
491 new_measure:
492         tp->rcvq_space.seq = tp->copied_seq;
493         tp->rcvq_space.time = tcp_time_stamp;
494 }
495
496 /* There is something which you must keep in mind when you analyze the
497  * behavior of the tp->ato delayed ack timeout interval.  When a
498  * connection starts up, we want to ack as quickly as possible.  The
499  * problem is that "good" TCP's do slow start at the beginning of data
500  * transmission.  The means that until we send the first few ACK's the
501  * sender will sit on his end and only queue most of his data, because
502  * he can only send snd_cwnd unacked packets at any given time.  For
503  * each ACK we send, he increments snd_cwnd and transmits more of his
504  * queue.  -DaveM
505  */
506 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
507 {
508         struct inet_connection_sock *icsk = inet_csk(sk);
509         u32 now;
510
511         inet_csk_schedule_ack(sk);
512
513         tcp_measure_rcv_mss(sk, skb);
514
515         tcp_rcv_rtt_measure(tp);
516
517         now = tcp_time_stamp;
518
519         if (!icsk->icsk_ack.ato) {
520                 /* The _first_ data packet received, initialize
521                  * delayed ACK engine.
522                  */
523                 tcp_incr_quickack(sk);
524                 icsk->icsk_ack.ato = TCP_ATO_MIN;
525         } else {
526                 int m = now - icsk->icsk_ack.lrcvtime;
527
528                 if (m <= TCP_ATO_MIN/2) {
529                         /* The fastest case is the first. */
530                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
531                 } else if (m < icsk->icsk_ack.ato) {
532                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
533                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
534                                 icsk->icsk_ack.ato = icsk->icsk_rto;
535                 } else if (m > icsk->icsk_rto) {
536                         /* Too long gap. Apparently sender failed to
537                          * restart window, so that we send ACKs quickly.
538                          */
539                         tcp_incr_quickack(sk);
540                         sk_stream_mem_reclaim(sk);
541                 }
542         }
543         icsk->icsk_ack.lrcvtime = now;
544
545         TCP_ECN_check_ce(tp, skb);
546
547         if (skb->len >= 128)
548                 tcp_grow_window(sk, tp, skb);
549 }
550
551 /* Called to compute a smoothed rtt estimate. The data fed to this
552  * routine either comes from timestamps, or from segments that were
553  * known _not_ to have been retransmitted [see Karn/Partridge
554  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
555  * piece by Van Jacobson.
556  * NOTE: the next three routines used to be one big routine.
557  * To save cycles in the RFC 1323 implementation it was better to break
558  * it up into three procedures. -- erics
559  */
560 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
561 {
562         struct tcp_sock *tp = tcp_sk(sk);
563         long m = mrtt; /* RTT */
564
565         /*      The following amusing code comes from Jacobson's
566          *      article in SIGCOMM '88.  Note that rtt and mdev
567          *      are scaled versions of rtt and mean deviation.
568          *      This is designed to be as fast as possible
569          *      m stands for "measurement".
570          *
571          *      On a 1990 paper the rto value is changed to:
572          *      RTO = rtt + 4 * mdev
573          *
574          * Funny. This algorithm seems to be very broken.
575          * These formulae increase RTO, when it should be decreased, increase
576          * too slowly, when it should be increased quickly, decrease too quickly
577          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
578          * does not matter how to _calculate_ it. Seems, it was trap
579          * that VJ failed to avoid. 8)
580          */
581         if (m == 0)
582                 m = 1;
583         if (tp->srtt != 0) {
584                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
585                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
586                 if (m < 0) {
587                         m = -m;         /* m is now abs(error) */
588                         m -= (tp->mdev >> 2);   /* similar update on mdev */
589                         /* This is similar to one of Eifel findings.
590                          * Eifel blocks mdev updates when rtt decreases.
591                          * This solution is a bit different: we use finer gain
592                          * for mdev in this case (alpha*beta).
593                          * Like Eifel it also prevents growth of rto,
594                          * but also it limits too fast rto decreases,
595                          * happening in pure Eifel.
596                          */
597                         if (m > 0)
598                                 m >>= 3;
599                 } else {
600                         m -= (tp->mdev >> 2);   /* similar update on mdev */
601                 }
602                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
603                 if (tp->mdev > tp->mdev_max) {
604                         tp->mdev_max = tp->mdev;
605                         if (tp->mdev_max > tp->rttvar)
606                                 tp->rttvar = tp->mdev_max;
607                 }
608                 if (after(tp->snd_una, tp->rtt_seq)) {
609                         if (tp->mdev_max < tp->rttvar)
610                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
611                         tp->rtt_seq = tp->snd_nxt;
612                         tp->mdev_max = TCP_RTO_MIN;
613                 }
614         } else {
615                 /* no previous measure. */
616                 tp->srtt = m<<3;        /* take the measured time to be rtt */
617                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
618                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
619                 tp->rtt_seq = tp->snd_nxt;
620         }
621 }
622
623 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
624  * routine referred to above.
625  */
626 static inline void tcp_set_rto(struct sock *sk)
627 {
628         const struct tcp_sock *tp = tcp_sk(sk);
629         /* Old crap is replaced with new one. 8)
630          *
631          * More seriously:
632          * 1. If rtt variance happened to be less 50msec, it is hallucination.
633          *    It cannot be less due to utterly erratic ACK generation made
634          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
635          *    to do with delayed acks, because at cwnd>2 true delack timeout
636          *    is invisible. Actually, Linux-2.4 also generates erratic
637          *    ACKs in some circumstances.
638          */
639         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
640
641         /* 2. Fixups made earlier cannot be right.
642          *    If we do not estimate RTO correctly without them,
643          *    all the algo is pure shit and should be replaced
644          *    with correct one. It is exactly, which we pretend to do.
645          */
646 }
647
648 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
649  * guarantees that rto is higher.
650  */
651 static inline void tcp_bound_rto(struct sock *sk)
652 {
653         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
654                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
655 }
656
657 /* Save metrics learned by this TCP session.
658    This function is called only, when TCP finishes successfully
659    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
660  */
661 void tcp_update_metrics(struct sock *sk)
662 {
663         struct tcp_sock *tp = tcp_sk(sk);
664         struct dst_entry *dst = __sk_dst_get(sk);
665
666         if (sysctl_tcp_nometrics_save)
667                 return;
668
669         dst_confirm(dst);
670
671         if (dst && (dst->flags&DST_HOST)) {
672                 const struct inet_connection_sock *icsk = inet_csk(sk);
673                 int m;
674
675                 if (icsk->icsk_backoff || !tp->srtt) {
676                         /* This session failed to estimate rtt. Why?
677                          * Probably, no packets returned in time.
678                          * Reset our results.
679                          */
680                         if (!(dst_metric_locked(dst, RTAX_RTT)))
681                                 dst->metrics[RTAX_RTT-1] = 0;
682                         return;
683                 }
684
685                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
686
687                 /* If newly calculated rtt larger than stored one,
688                  * store new one. Otherwise, use EWMA. Remember,
689                  * rtt overestimation is always better than underestimation.
690                  */
691                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
692                         if (m <= 0)
693                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
694                         else
695                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
696                 }
697
698                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
699                         if (m < 0)
700                                 m = -m;
701
702                         /* Scale deviation to rttvar fixed point */
703                         m >>= 1;
704                         if (m < tp->mdev)
705                                 m = tp->mdev;
706
707                         if (m >= dst_metric(dst, RTAX_RTTVAR))
708                                 dst->metrics[RTAX_RTTVAR-1] = m;
709                         else
710                                 dst->metrics[RTAX_RTTVAR-1] -=
711                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
712                 }
713
714                 if (tp->snd_ssthresh >= 0xFFFF) {
715                         /* Slow start still did not finish. */
716                         if (dst_metric(dst, RTAX_SSTHRESH) &&
717                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
718                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
719                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
720                         if (!dst_metric_locked(dst, RTAX_CWND) &&
721                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
722                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
723                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
724                            icsk->icsk_ca_state == TCP_CA_Open) {
725                         /* Cong. avoidance phase, cwnd is reliable. */
726                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
727                                 dst->metrics[RTAX_SSTHRESH-1] =
728                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
729                         if (!dst_metric_locked(dst, RTAX_CWND))
730                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
731                 } else {
732                         /* Else slow start did not finish, cwnd is non-sense,
733                            ssthresh may be also invalid.
734                          */
735                         if (!dst_metric_locked(dst, RTAX_CWND))
736                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
737                         if (dst->metrics[RTAX_SSTHRESH-1] &&
738                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
739                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
740                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
741                 }
742
743                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
744                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
745                             tp->reordering != sysctl_tcp_reordering)
746                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
747                 }
748         }
749 }
750
751 /* Numbers are taken from RFC2414.  */
752 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
753 {
754         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
755
756         if (!cwnd) {
757                 if (tp->mss_cache > 1460)
758                         cwnd = 2;
759                 else
760                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
761         }
762         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
763 }
764
765 /* Set slow start threshold and cwnd not falling to slow start */
766 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
767 {
768         struct tcp_sock *tp = tcp_sk(sk);
769         const struct inet_connection_sock *icsk = inet_csk(sk);
770
771         tp->prior_ssthresh = 0;
772         tp->bytes_acked = 0;
773         if (icsk->icsk_ca_state < TCP_CA_CWR) {
774                 tp->undo_marker = 0;
775                 if (set_ssthresh)
776                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
777                 tp->snd_cwnd = min(tp->snd_cwnd,
778                                    tcp_packets_in_flight(tp) + 1U);
779                 tp->snd_cwnd_cnt = 0;
780                 tp->high_seq = tp->snd_nxt;
781                 tp->snd_cwnd_stamp = tcp_time_stamp;
782                 TCP_ECN_queue_cwr(tp);
783
784                 tcp_set_ca_state(sk, TCP_CA_CWR);
785         }
786 }
787
788 /* Initialize metrics on socket. */
789
790 static void tcp_init_metrics(struct sock *sk)
791 {
792         struct tcp_sock *tp = tcp_sk(sk);
793         struct dst_entry *dst = __sk_dst_get(sk);
794
795         if (dst == NULL)
796                 goto reset;
797
798         dst_confirm(dst);
799
800         if (dst_metric_locked(dst, RTAX_CWND))
801                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
802         if (dst_metric(dst, RTAX_SSTHRESH)) {
803                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
804                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
805                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
806         }
807         if (dst_metric(dst, RTAX_REORDERING) &&
808             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
809                 tp->rx_opt.sack_ok &= ~2;
810                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
811         }
812
813         if (dst_metric(dst, RTAX_RTT) == 0)
814                 goto reset;
815
816         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
817                 goto reset;
818
819         /* Initial rtt is determined from SYN,SYN-ACK.
820          * The segment is small and rtt may appear much
821          * less than real one. Use per-dst memory
822          * to make it more realistic.
823          *
824          * A bit of theory. RTT is time passed after "normal" sized packet
825          * is sent until it is ACKed. In normal circumstances sending small
826          * packets force peer to delay ACKs and calculation is correct too.
827          * The algorithm is adaptive and, provided we follow specs, it
828          * NEVER underestimate RTT. BUT! If peer tries to make some clever
829          * tricks sort of "quick acks" for time long enough to decrease RTT
830          * to low value, and then abruptly stops to do it and starts to delay
831          * ACKs, wait for troubles.
832          */
833         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
834                 tp->srtt = dst_metric(dst, RTAX_RTT);
835                 tp->rtt_seq = tp->snd_nxt;
836         }
837         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
838                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
839                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
840         }
841         tcp_set_rto(sk);
842         tcp_bound_rto(sk);
843         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
844                 goto reset;
845         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
846         tp->snd_cwnd_stamp = tcp_time_stamp;
847         return;
848
849 reset:
850         /* Play conservative. If timestamps are not
851          * supported, TCP will fail to recalculate correct
852          * rtt, if initial rto is too small. FORGET ALL AND RESET!
853          */
854         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
855                 tp->srtt = 0;
856                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
857                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
858         }
859 }
860
861 static void tcp_update_reordering(struct sock *sk, const int metric,
862                                   const int ts)
863 {
864         struct tcp_sock *tp = tcp_sk(sk);
865         if (metric > tp->reordering) {
866                 tp->reordering = min(TCP_MAX_REORDERING, metric);
867
868                 /* This exciting event is worth to be remembered. 8) */
869                 if (ts)
870                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
871                 else if (IsReno(tp))
872                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
873                 else if (IsFack(tp))
874                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
875                 else
876                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
877 #if FASTRETRANS_DEBUG > 1
878                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
879                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
880                        tp->reordering,
881                        tp->fackets_out,
882                        tp->sacked_out,
883                        tp->undo_marker ? tp->undo_retrans : 0);
884 #endif
885                 /* Disable FACK yet. */
886                 tp->rx_opt.sack_ok &= ~2;
887         }
888 }
889
890 /* This procedure tags the retransmission queue when SACKs arrive.
891  *
892  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
893  * Packets in queue with these bits set are counted in variables
894  * sacked_out, retrans_out and lost_out, correspondingly.
895  *
896  * Valid combinations are:
897  * Tag  InFlight        Description
898  * 0    1               - orig segment is in flight.
899  * S    0               - nothing flies, orig reached receiver.
900  * L    0               - nothing flies, orig lost by net.
901  * R    2               - both orig and retransmit are in flight.
902  * L|R  1               - orig is lost, retransmit is in flight.
903  * S|R  1               - orig reached receiver, retrans is still in flight.
904  * (L|S|R is logically valid, it could occur when L|R is sacked,
905  *  but it is equivalent to plain S and code short-curcuits it to S.
906  *  L|S is logically invalid, it would mean -1 packet in flight 8))
907  *
908  * These 6 states form finite state machine, controlled by the following events:
909  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
910  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
911  * 3. Loss detection event of one of three flavors:
912  *      A. Scoreboard estimator decided the packet is lost.
913  *         A'. Reno "three dupacks" marks head of queue lost.
914  *         A''. Its FACK modfication, head until snd.fack is lost.
915  *      B. SACK arrives sacking data transmitted after never retransmitted
916  *         hole was sent out.
917  *      C. SACK arrives sacking SND.NXT at the moment, when the
918  *         segment was retransmitted.
919  * 4. D-SACK added new rule: D-SACK changes any tag to S.
920  *
921  * It is pleasant to note, that state diagram turns out to be commutative,
922  * so that we are allowed not to be bothered by order of our actions,
923  * when multiple events arrive simultaneously. (see the function below).
924  *
925  * Reordering detection.
926  * --------------------
927  * Reordering metric is maximal distance, which a packet can be displaced
928  * in packet stream. With SACKs we can estimate it:
929  *
930  * 1. SACK fills old hole and the corresponding segment was not
931  *    ever retransmitted -> reordering. Alas, we cannot use it
932  *    when segment was retransmitted.
933  * 2. The last flaw is solved with D-SACK. D-SACK arrives
934  *    for retransmitted and already SACKed segment -> reordering..
935  * Both of these heuristics are not used in Loss state, when we cannot
936  * account for retransmits accurately.
937  */
938 static int
939 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
940 {
941         const struct inet_connection_sock *icsk = inet_csk(sk);
942         struct tcp_sock *tp = tcp_sk(sk);
943         unsigned char *ptr = (skb_transport_header(ack_skb) +
944                               TCP_SKB_CB(ack_skb)->sacked);
945         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
946         struct sk_buff *cached_skb;
947         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
948         int reord = tp->packets_out;
949         int prior_fackets;
950         u32 lost_retrans = 0;
951         int flag = 0;
952         int dup_sack = 0;
953         int cached_fack_count;
954         int i;
955         int first_sack_index;
956
957         if (!tp->sacked_out)
958                 tp->fackets_out = 0;
959         prior_fackets = tp->fackets_out;
960
961         /* Check for D-SACK. */
962         if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
963                 dup_sack = 1;
964                 tp->rx_opt.sack_ok |= 4;
965                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
966         } else if (num_sacks > 1 &&
967                         !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
968                         !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
969                 dup_sack = 1;
970                 tp->rx_opt.sack_ok |= 4;
971                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
972         }
973
974         /* D-SACK for already forgotten data...
975          * Do dumb counting. */
976         if (dup_sack &&
977                         !after(ntohl(sp[0].end_seq), prior_snd_una) &&
978                         after(ntohl(sp[0].end_seq), tp->undo_marker))
979                 tp->undo_retrans--;
980
981         /* Eliminate too old ACKs, but take into
982          * account more or less fresh ones, they can
983          * contain valid SACK info.
984          */
985         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
986                 return 0;
987
988         /* SACK fastpath:
989          * if the only SACK change is the increase of the end_seq of
990          * the first block then only apply that SACK block
991          * and use retrans queue hinting otherwise slowpath */
992         flag = 1;
993         for (i = 0; i < num_sacks; i++) {
994                 __be32 start_seq = sp[i].start_seq;
995                 __be32 end_seq = sp[i].end_seq;
996
997                 if (i == 0) {
998                         if (tp->recv_sack_cache[i].start_seq != start_seq)
999                                 flag = 0;
1000                 } else {
1001                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1002                             (tp->recv_sack_cache[i].end_seq != end_seq))
1003                                 flag = 0;
1004                 }
1005                 tp->recv_sack_cache[i].start_seq = start_seq;
1006                 tp->recv_sack_cache[i].end_seq = end_seq;
1007         }
1008         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1009         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1010                 tp->recv_sack_cache[i].start_seq = 0;
1011                 tp->recv_sack_cache[i].end_seq = 0;
1012         }
1013
1014         first_sack_index = 0;
1015         if (flag)
1016                 num_sacks = 1;
1017         else {
1018                 int j;
1019                 tp->fastpath_skb_hint = NULL;
1020
1021                 /* order SACK blocks to allow in order walk of the retrans queue */
1022                 for (i = num_sacks-1; i > 0; i--) {
1023                         for (j = 0; j < i; j++){
1024                                 if (after(ntohl(sp[j].start_seq),
1025                                           ntohl(sp[j+1].start_seq))){
1026                                         struct tcp_sack_block_wire tmp;
1027
1028                                         tmp = sp[j];
1029                                         sp[j] = sp[j+1];
1030                                         sp[j+1] = tmp;
1031
1032                                         /* Track where the first SACK block goes to */
1033                                         if (j == first_sack_index)
1034                                                 first_sack_index = j+1;
1035                                 }
1036
1037                         }
1038                 }
1039         }
1040
1041         /* clear flag as used for different purpose in following code */
1042         flag = 0;
1043
1044         /* Use SACK fastpath hint if valid */
1045         cached_skb = tp->fastpath_skb_hint;
1046         cached_fack_count = tp->fastpath_cnt_hint;
1047         if (!cached_skb) {
1048                 cached_skb = tcp_write_queue_head(sk);
1049                 cached_fack_count = 0;
1050         }
1051
1052         for (i=0; i<num_sacks; i++, sp++) {
1053                 struct sk_buff *skb;
1054                 __u32 start_seq = ntohl(sp->start_seq);
1055                 __u32 end_seq = ntohl(sp->end_seq);
1056                 int fack_count;
1057
1058                 skb = cached_skb;
1059                 fack_count = cached_fack_count;
1060
1061                 /* Event "B" in the comment above. */
1062                 if (after(end_seq, tp->high_seq))
1063                         flag |= FLAG_DATA_LOST;
1064
1065                 tcp_for_write_queue_from(skb, sk) {
1066                         int in_sack, pcount;
1067                         u8 sacked;
1068
1069                         if (skb == tcp_send_head(sk))
1070                                 break;
1071
1072                         cached_skb = skb;
1073                         cached_fack_count = fack_count;
1074                         if (i == first_sack_index) {
1075                                 tp->fastpath_skb_hint = skb;
1076                                 tp->fastpath_cnt_hint = fack_count;
1077                         }
1078
1079                         /* The retransmission queue is always in order, so
1080                          * we can short-circuit the walk early.
1081                          */
1082                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1083                                 break;
1084
1085                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1086                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1087
1088                         pcount = tcp_skb_pcount(skb);
1089
1090                         if (pcount > 1 && !in_sack &&
1091                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1092                                 unsigned int pkt_len;
1093
1094                                 in_sack = !after(start_seq,
1095                                                  TCP_SKB_CB(skb)->seq);
1096
1097                                 if (!in_sack)
1098                                         pkt_len = (start_seq -
1099                                                    TCP_SKB_CB(skb)->seq);
1100                                 else
1101                                         pkt_len = (end_seq -
1102                                                    TCP_SKB_CB(skb)->seq);
1103                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1104                                         break;
1105                                 pcount = tcp_skb_pcount(skb);
1106                         }
1107
1108                         fack_count += pcount;
1109
1110                         sacked = TCP_SKB_CB(skb)->sacked;
1111
1112                         /* Account D-SACK for retransmitted packet. */
1113                         if ((dup_sack && in_sack) &&
1114                             (sacked & TCPCB_RETRANS) &&
1115                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1116                                 tp->undo_retrans--;
1117
1118                         /* The frame is ACKed. */
1119                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1120                                 if (sacked&TCPCB_RETRANS) {
1121                                         if ((dup_sack && in_sack) &&
1122                                             (sacked&TCPCB_SACKED_ACKED))
1123                                                 reord = min(fack_count, reord);
1124                                 } else {
1125                                         /* If it was in a hole, we detected reordering. */
1126                                         if (fack_count < prior_fackets &&
1127                                             !(sacked&TCPCB_SACKED_ACKED))
1128                                                 reord = min(fack_count, reord);
1129                                 }
1130
1131                                 /* Nothing to do; acked frame is about to be dropped. */
1132                                 continue;
1133                         }
1134
1135                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1136                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1137                             (!lost_retrans || after(end_seq, lost_retrans)))
1138                                 lost_retrans = end_seq;
1139
1140                         if (!in_sack)
1141                                 continue;
1142
1143                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1144                                 if (sacked & TCPCB_SACKED_RETRANS) {
1145                                         /* If the segment is not tagged as lost,
1146                                          * we do not clear RETRANS, believing
1147                                          * that retransmission is still in flight.
1148                                          */
1149                                         if (sacked & TCPCB_LOST) {
1150                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1151                                                 tp->lost_out -= tcp_skb_pcount(skb);
1152                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1153
1154                                                 /* clear lost hint */
1155                                                 tp->retransmit_skb_hint = NULL;
1156                                         }
1157                                 } else {
1158                                         /* New sack for not retransmitted frame,
1159                                          * which was in hole. It is reordering.
1160                                          */
1161                                         if (!(sacked & TCPCB_RETRANS) &&
1162                                             fack_count < prior_fackets)
1163                                                 reord = min(fack_count, reord);
1164
1165                                         if (sacked & TCPCB_LOST) {
1166                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1167                                                 tp->lost_out -= tcp_skb_pcount(skb);
1168
1169                                                 /* clear lost hint */
1170                                                 tp->retransmit_skb_hint = NULL;
1171                                         }
1172                                         /* SACK enhanced F-RTO detection.
1173                                          * Set flag if and only if non-rexmitted
1174                                          * segments below frto_highmark are
1175                                          * SACKed (RFC4138; Appendix B).
1176                                          * Clearing correct due to in-order walk
1177                                          */
1178                                         if (after(end_seq, tp->frto_highmark)) {
1179                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1180                                         } else {
1181                                                 if (!(sacked & TCPCB_RETRANS))
1182                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1183                                         }
1184                                 }
1185
1186                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1187                                 flag |= FLAG_DATA_SACKED;
1188                                 tp->sacked_out += tcp_skb_pcount(skb);
1189
1190                                 if (fack_count > tp->fackets_out)
1191                                         tp->fackets_out = fack_count;
1192                         } else {
1193                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1194                                         reord = min(fack_count, reord);
1195                         }
1196
1197                         /* D-SACK. We can detect redundant retransmission
1198                          * in S|R and plain R frames and clear it.
1199                          * undo_retrans is decreased above, L|R frames
1200                          * are accounted above as well.
1201                          */
1202                         if (dup_sack &&
1203                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1204                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1205                                 tp->retrans_out -= tcp_skb_pcount(skb);
1206                                 tp->retransmit_skb_hint = NULL;
1207                         }
1208                 }
1209         }
1210
1211         /* Check for lost retransmit. This superb idea is
1212          * borrowed from "ratehalving". Event "C".
1213          * Later note: FACK people cheated me again 8),
1214          * we have to account for reordering! Ugly,
1215          * but should help.
1216          */
1217         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1218                 struct sk_buff *skb;
1219
1220                 tcp_for_write_queue(skb, sk) {
1221                         if (skb == tcp_send_head(sk))
1222                                 break;
1223                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1224                                 break;
1225                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1226                                 continue;
1227                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1228                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1229                             (IsFack(tp) ||
1230                              !before(lost_retrans,
1231                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1232                                      tp->mss_cache))) {
1233                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1234                                 tp->retrans_out -= tcp_skb_pcount(skb);
1235
1236                                 /* clear lost hint */
1237                                 tp->retransmit_skb_hint = NULL;
1238
1239                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1240                                         tp->lost_out += tcp_skb_pcount(skb);
1241                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1242                                         flag |= FLAG_DATA_SACKED;
1243                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1244                                 }
1245                         }
1246                 }
1247         }
1248
1249         tp->left_out = tp->sacked_out + tp->lost_out;
1250
1251         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1252             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1253                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1254
1255 #if FASTRETRANS_DEBUG > 0
1256         BUG_TRAP((int)tp->sacked_out >= 0);
1257         BUG_TRAP((int)tp->lost_out >= 0);
1258         BUG_TRAP((int)tp->retrans_out >= 0);
1259         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1260 #endif
1261         return flag;
1262 }
1263
1264 /* F-RTO can only be used if these conditions are satisfied:
1265  *  - there must be some unsent new data
1266  *  - the advertised window should allow sending it
1267  *  - TCP has never retransmitted anything other than head (SACK enhanced
1268  *    variant from Appendix B of RFC4138 is more robust here)
1269  */
1270 int tcp_use_frto(struct sock *sk)
1271 {
1272         const struct tcp_sock *tp = tcp_sk(sk);
1273         struct sk_buff *skb;
1274
1275         if (!sysctl_tcp_frto || !tcp_send_head(sk) ||
1276                 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
1277                       tp->snd_una + tp->snd_wnd))
1278                 return 0;
1279
1280         if (IsSackFrto())
1281                 return 1;
1282
1283         /* Avoid expensive walking of rexmit queue if possible */
1284         if (tp->retrans_out > 1)
1285                 return 0;
1286
1287         skb = tcp_write_queue_head(sk);
1288         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1289         tcp_for_write_queue_from(skb, sk) {
1290                 if (skb == tcp_send_head(sk))
1291                         break;
1292                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1293                         return 0;
1294                 /* Short-circuit when first non-SACKed skb has been checked */
1295                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1296                         break;
1297         }
1298         return 1;
1299 }
1300
1301 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1302  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1303  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1304  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1305  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1306  * bits are handled if the Loss state is really to be entered (in
1307  * tcp_enter_frto_loss).
1308  *
1309  * Do like tcp_enter_loss() would; when RTO expires the second time it
1310  * does:
1311  *  "Reduce ssthresh if it has not yet been made inside this window."
1312  */
1313 void tcp_enter_frto(struct sock *sk)
1314 {
1315         const struct inet_connection_sock *icsk = inet_csk(sk);
1316         struct tcp_sock *tp = tcp_sk(sk);
1317         struct sk_buff *skb;
1318
1319         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1320             tp->snd_una == tp->high_seq ||
1321             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1322              !icsk->icsk_retransmits)) {
1323                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1324                 /* Our state is too optimistic in ssthresh() call because cwnd
1325                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1326                  * recovery has not yet completed. Pattern would be this: RTO,
1327                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1328                  * up here twice).
1329                  * RFC4138 should be more specific on what to do, even though
1330                  * RTO is quite unlikely to occur after the first Cumulative ACK
1331                  * due to back-off and complexity of triggering events ...
1332                  */
1333                 if (tp->frto_counter) {
1334                         u32 stored_cwnd;
1335                         stored_cwnd = tp->snd_cwnd;
1336                         tp->snd_cwnd = 2;
1337                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1338                         tp->snd_cwnd = stored_cwnd;
1339                 } else {
1340                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1341                 }
1342                 /* ... in theory, cong.control module could do "any tricks" in
1343                  * ssthresh(), which means that ca_state, lost bits and lost_out
1344                  * counter would have to be faked before the call occurs. We
1345                  * consider that too expensive, unlikely and hacky, so modules
1346                  * using these in ssthresh() must deal these incompatibility
1347                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1348                  */
1349                 tcp_ca_event(sk, CA_EVENT_FRTO);
1350         }
1351
1352         tp->undo_marker = tp->snd_una;
1353         tp->undo_retrans = 0;
1354
1355         skb = tcp_write_queue_head(sk);
1356         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1357                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1358                 tp->retrans_out -= tcp_skb_pcount(skb);
1359         }
1360         tcp_sync_left_out(tp);
1361
1362         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1363          * The last condition is necessary at least in tp->frto_counter case.
1364          */
1365         if (IsSackFrto() && (tp->frto_counter ||
1366             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1367             after(tp->high_seq, tp->snd_una)) {
1368                 tp->frto_highmark = tp->high_seq;
1369         } else {
1370                 tp->frto_highmark = tp->snd_nxt;
1371         }
1372         tcp_set_ca_state(sk, TCP_CA_Disorder);
1373         tp->high_seq = tp->snd_nxt;
1374         tp->frto_counter = 1;
1375 }
1376
1377 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1378  * which indicates that we should follow the traditional RTO recovery,
1379  * i.e. mark everything lost and do go-back-N retransmission.
1380  */
1381 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1382 {
1383         struct tcp_sock *tp = tcp_sk(sk);
1384         struct sk_buff *skb;
1385         int cnt = 0;
1386
1387         tp->sacked_out = 0;
1388         tp->lost_out = 0;
1389         tp->fackets_out = 0;
1390         tp->retrans_out = 0;
1391
1392         tcp_for_write_queue(skb, sk) {
1393                 if (skb == tcp_send_head(sk))
1394                         break;
1395                 cnt += tcp_skb_pcount(skb);
1396                 /*
1397                  * Count the retransmission made on RTO correctly (only when
1398                  * waiting for the first ACK and did not get it)...
1399                  */
1400                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1401                         tp->retrans_out += tcp_skb_pcount(skb);
1402                         /* ...enter this if branch just for the first segment */
1403                         flag |= FLAG_DATA_ACKED;
1404                 } else {
1405                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1406                 }
1407                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1408
1409                         /* Do not mark those segments lost that were
1410                          * forward transmitted after RTO
1411                          */
1412                         if (!after(TCP_SKB_CB(skb)->end_seq,
1413                                    tp->frto_highmark)) {
1414                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1415                                 tp->lost_out += tcp_skb_pcount(skb);
1416                         }
1417                 } else {
1418                         tp->sacked_out += tcp_skb_pcount(skb);
1419                         tp->fackets_out = cnt;
1420                 }
1421         }
1422         tcp_sync_left_out(tp);
1423
1424         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1425         tp->snd_cwnd_cnt = 0;
1426         tp->snd_cwnd_stamp = tcp_time_stamp;
1427         tp->undo_marker = 0;
1428         tp->frto_counter = 0;
1429
1430         tp->reordering = min_t(unsigned int, tp->reordering,
1431                                              sysctl_tcp_reordering);
1432         tcp_set_ca_state(sk, TCP_CA_Loss);
1433         tp->high_seq = tp->frto_highmark;
1434         TCP_ECN_queue_cwr(tp);
1435
1436         clear_all_retrans_hints(tp);
1437 }
1438
1439 void tcp_clear_retrans(struct tcp_sock *tp)
1440 {
1441         tp->left_out = 0;
1442         tp->retrans_out = 0;
1443
1444         tp->fackets_out = 0;
1445         tp->sacked_out = 0;
1446         tp->lost_out = 0;
1447
1448         tp->undo_marker = 0;
1449         tp->undo_retrans = 0;
1450 }
1451
1452 /* Enter Loss state. If "how" is not zero, forget all SACK information
1453  * and reset tags completely, otherwise preserve SACKs. If receiver
1454  * dropped its ofo queue, we will know this due to reneging detection.
1455  */
1456 void tcp_enter_loss(struct sock *sk, int how)
1457 {
1458         const struct inet_connection_sock *icsk = inet_csk(sk);
1459         struct tcp_sock *tp = tcp_sk(sk);
1460         struct sk_buff *skb;
1461         int cnt = 0;
1462
1463         /* Reduce ssthresh if it has not yet been made inside this window. */
1464         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1465             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1466                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1467                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1468                 tcp_ca_event(sk, CA_EVENT_LOSS);
1469         }
1470         tp->snd_cwnd       = 1;
1471         tp->snd_cwnd_cnt   = 0;
1472         tp->snd_cwnd_stamp = tcp_time_stamp;
1473
1474         tp->bytes_acked = 0;
1475         tcp_clear_retrans(tp);
1476
1477         /* Push undo marker, if it was plain RTO and nothing
1478          * was retransmitted. */
1479         if (!how)
1480                 tp->undo_marker = tp->snd_una;
1481
1482         tcp_for_write_queue(skb, sk) {
1483                 if (skb == tcp_send_head(sk))
1484                         break;
1485                 cnt += tcp_skb_pcount(skb);
1486                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1487                         tp->undo_marker = 0;
1488                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1489                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1490                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1491                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1492                         tp->lost_out += tcp_skb_pcount(skb);
1493                 } else {
1494                         tp->sacked_out += tcp_skb_pcount(skb);
1495                         tp->fackets_out = cnt;
1496                 }
1497         }
1498         tcp_sync_left_out(tp);
1499
1500         tp->reordering = min_t(unsigned int, tp->reordering,
1501                                              sysctl_tcp_reordering);
1502         tcp_set_ca_state(sk, TCP_CA_Loss);
1503         tp->high_seq = tp->snd_nxt;
1504         TCP_ECN_queue_cwr(tp);
1505
1506         clear_all_retrans_hints(tp);
1507 }
1508
1509 static int tcp_check_sack_reneging(struct sock *sk)
1510 {
1511         struct sk_buff *skb;
1512
1513         /* If ACK arrived pointing to a remembered SACK,
1514          * it means that our remembered SACKs do not reflect
1515          * real state of receiver i.e.
1516          * receiver _host_ is heavily congested (or buggy).
1517          * Do processing similar to RTO timeout.
1518          */
1519         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1520             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1521                 struct inet_connection_sock *icsk = inet_csk(sk);
1522                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1523
1524                 tcp_enter_loss(sk, 1);
1525                 icsk->icsk_retransmits++;
1526                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1527                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1528                                           icsk->icsk_rto, TCP_RTO_MAX);
1529                 return 1;
1530         }
1531         return 0;
1532 }
1533
1534 static inline int tcp_fackets_out(struct tcp_sock *tp)
1535 {
1536         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1537 }
1538
1539 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1540 {
1541         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1542 }
1543
1544 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1545 {
1546         return tp->packets_out &&
1547                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1548 }
1549
1550 /* Linux NewReno/SACK/FACK/ECN state machine.
1551  * --------------------------------------
1552  *
1553  * "Open"       Normal state, no dubious events, fast path.
1554  * "Disorder"   In all the respects it is "Open",
1555  *              but requires a bit more attention. It is entered when
1556  *              we see some SACKs or dupacks. It is split of "Open"
1557  *              mainly to move some processing from fast path to slow one.
1558  * "CWR"        CWND was reduced due to some Congestion Notification event.
1559  *              It can be ECN, ICMP source quench, local device congestion.
1560  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1561  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1562  *
1563  * tcp_fastretrans_alert() is entered:
1564  * - each incoming ACK, if state is not "Open"
1565  * - when arrived ACK is unusual, namely:
1566  *      * SACK
1567  *      * Duplicate ACK.
1568  *      * ECN ECE.
1569  *
1570  * Counting packets in flight is pretty simple.
1571  *
1572  *      in_flight = packets_out - left_out + retrans_out
1573  *
1574  *      packets_out is SND.NXT-SND.UNA counted in packets.
1575  *
1576  *      retrans_out is number of retransmitted segments.
1577  *
1578  *      left_out is number of segments left network, but not ACKed yet.
1579  *
1580  *              left_out = sacked_out + lost_out
1581  *
1582  *     sacked_out: Packets, which arrived to receiver out of order
1583  *                 and hence not ACKed. With SACKs this number is simply
1584  *                 amount of SACKed data. Even without SACKs
1585  *                 it is easy to give pretty reliable estimate of this number,
1586  *                 counting duplicate ACKs.
1587  *
1588  *       lost_out: Packets lost by network. TCP has no explicit
1589  *                 "loss notification" feedback from network (for now).
1590  *                 It means that this number can be only _guessed_.
1591  *                 Actually, it is the heuristics to predict lossage that
1592  *                 distinguishes different algorithms.
1593  *
1594  *      F.e. after RTO, when all the queue is considered as lost,
1595  *      lost_out = packets_out and in_flight = retrans_out.
1596  *
1597  *              Essentially, we have now two algorithms counting
1598  *              lost packets.
1599  *
1600  *              FACK: It is the simplest heuristics. As soon as we decided
1601  *              that something is lost, we decide that _all_ not SACKed
1602  *              packets until the most forward SACK are lost. I.e.
1603  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1604  *              It is absolutely correct estimate, if network does not reorder
1605  *              packets. And it loses any connection to reality when reordering
1606  *              takes place. We use FACK by default until reordering
1607  *              is suspected on the path to this destination.
1608  *
1609  *              NewReno: when Recovery is entered, we assume that one segment
1610  *              is lost (classic Reno). While we are in Recovery and
1611  *              a partial ACK arrives, we assume that one more packet
1612  *              is lost (NewReno). This heuristics are the same in NewReno
1613  *              and SACK.
1614  *
1615  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1616  *  deflation etc. CWND is real congestion window, never inflated, changes
1617  *  only according to classic VJ rules.
1618  *
1619  * Really tricky (and requiring careful tuning) part of algorithm
1620  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1621  * The first determines the moment _when_ we should reduce CWND and,
1622  * hence, slow down forward transmission. In fact, it determines the moment
1623  * when we decide that hole is caused by loss, rather than by a reorder.
1624  *
1625  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1626  * holes, caused by lost packets.
1627  *
1628  * And the most logically complicated part of algorithm is undo
1629  * heuristics. We detect false retransmits due to both too early
1630  * fast retransmit (reordering) and underestimated RTO, analyzing
1631  * timestamps and D-SACKs. When we detect that some segments were
1632  * retransmitted by mistake and CWND reduction was wrong, we undo
1633  * window reduction and abort recovery phase. This logic is hidden
1634  * inside several functions named tcp_try_undo_<something>.
1635  */
1636
1637 /* This function decides, when we should leave Disordered state
1638  * and enter Recovery phase, reducing congestion window.
1639  *
1640  * Main question: may we further continue forward transmission
1641  * with the same cwnd?
1642  */
1643 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1644 {
1645         __u32 packets_out;
1646
1647         /* Do not perform any recovery during FRTO algorithm */
1648         if (tp->frto_counter)
1649                 return 0;
1650
1651         /* Trick#1: The loss is proven. */
1652         if (tp->lost_out)
1653                 return 1;
1654
1655         /* Not-A-Trick#2 : Classic rule... */
1656         if (tcp_fackets_out(tp) > tp->reordering)
1657                 return 1;
1658
1659         /* Trick#3 : when we use RFC2988 timer restart, fast
1660          * retransmit can be triggered by timeout of queue head.
1661          */
1662         if (tcp_head_timedout(sk, tp))
1663                 return 1;
1664
1665         /* Trick#4: It is still not OK... But will it be useful to delay
1666          * recovery more?
1667          */
1668         packets_out = tp->packets_out;
1669         if (packets_out <= tp->reordering &&
1670             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1671             !tcp_may_send_now(sk, tp)) {
1672                 /* We have nothing to send. This connection is limited
1673                  * either by receiver window or by application.
1674                  */
1675                 return 1;
1676         }
1677
1678         return 0;
1679 }
1680
1681 /* If we receive more dupacks than we expected counting segments
1682  * in assumption of absent reordering, interpret this as reordering.
1683  * The only another reason could be bug in receiver TCP.
1684  */
1685 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1686 {
1687         struct tcp_sock *tp = tcp_sk(sk);
1688         u32 holes;
1689
1690         holes = max(tp->lost_out, 1U);
1691         holes = min(holes, tp->packets_out);
1692
1693         if ((tp->sacked_out + holes) > tp->packets_out) {
1694                 tp->sacked_out = tp->packets_out - holes;
1695                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1696         }
1697 }
1698
1699 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1700
1701 static void tcp_add_reno_sack(struct sock *sk)
1702 {
1703         struct tcp_sock *tp = tcp_sk(sk);
1704         tp->sacked_out++;
1705         tcp_check_reno_reordering(sk, 0);
1706         tcp_sync_left_out(tp);
1707 }
1708
1709 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1710
1711 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1712 {
1713         if (acked > 0) {
1714                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1715                 if (acked-1 >= tp->sacked_out)
1716                         tp->sacked_out = 0;
1717                 else
1718                         tp->sacked_out -= acked-1;
1719         }
1720         tcp_check_reno_reordering(sk, acked);
1721         tcp_sync_left_out(tp);
1722 }
1723
1724 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1725 {
1726         tp->sacked_out = 0;
1727         tp->left_out = tp->lost_out;
1728 }
1729
1730 /* Mark head of queue up as lost. */
1731 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1732                                int packets, u32 high_seq)
1733 {
1734         struct sk_buff *skb;
1735         int cnt;
1736
1737         BUG_TRAP(packets <= tp->packets_out);
1738         if (tp->lost_skb_hint) {
1739                 skb = tp->lost_skb_hint;
1740                 cnt = tp->lost_cnt_hint;
1741         } else {
1742                 skb = tcp_write_queue_head(sk);
1743                 cnt = 0;
1744         }
1745
1746         tcp_for_write_queue_from(skb, sk) {
1747                 if (skb == tcp_send_head(sk))
1748                         break;
1749                 /* TODO: do this better */
1750                 /* this is not the most efficient way to do this... */
1751                 tp->lost_skb_hint = skb;
1752                 tp->lost_cnt_hint = cnt;
1753                 cnt += tcp_skb_pcount(skb);
1754                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1755                         break;
1756                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1757                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1758                         tp->lost_out += tcp_skb_pcount(skb);
1759
1760                         /* clear xmit_retransmit_queue hints
1761                          *  if this is beyond hint */
1762                         if (tp->retransmit_skb_hint != NULL &&
1763                             before(TCP_SKB_CB(skb)->seq,
1764                                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1765                                 tp->retransmit_skb_hint = NULL;
1766
1767                 }
1768         }
1769         tcp_sync_left_out(tp);
1770 }
1771
1772 /* Account newly detected lost packet(s) */
1773
1774 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1775 {
1776         if (IsFack(tp)) {
1777                 int lost = tp->fackets_out - tp->reordering;
1778                 if (lost <= 0)
1779                         lost = 1;
1780                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1781         } else {
1782                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1783         }
1784
1785         /* New heuristics: it is possible only after we switched
1786          * to restart timer each time when something is ACKed.
1787          * Hence, we can detect timed out packets during fast
1788          * retransmit without falling to slow start.
1789          */
1790         if (!IsReno(tp) && tcp_head_timedout(sk, tp)) {
1791                 struct sk_buff *skb;
1792
1793                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1794                         : tcp_write_queue_head(sk);
1795
1796                 tcp_for_write_queue_from(skb, sk) {
1797                         if (skb == tcp_send_head(sk))
1798                                 break;
1799                         if (!tcp_skb_timedout(sk, skb))
1800                                 break;
1801
1802                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1803                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1804                                 tp->lost_out += tcp_skb_pcount(skb);
1805
1806                                 /* clear xmit_retrans hint */
1807                                 if (tp->retransmit_skb_hint &&
1808                                     before(TCP_SKB_CB(skb)->seq,
1809                                            TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1810
1811                                         tp->retransmit_skb_hint = NULL;
1812                         }
1813                 }
1814
1815                 tp->scoreboard_skb_hint = skb;
1816
1817                 tcp_sync_left_out(tp);
1818         }
1819 }
1820
1821 /* CWND moderation, preventing bursts due to too big ACKs
1822  * in dubious situations.
1823  */
1824 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1825 {
1826         tp->snd_cwnd = min(tp->snd_cwnd,
1827                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1828         tp->snd_cwnd_stamp = tcp_time_stamp;
1829 }
1830
1831 /* Lower bound on congestion window is slow start threshold
1832  * unless congestion avoidance choice decides to overide it.
1833  */
1834 static inline u32 tcp_cwnd_min(const struct sock *sk)
1835 {
1836         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1837
1838         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1839 }
1840
1841 /* Decrease cwnd each second ack. */
1842 static void tcp_cwnd_down(struct sock *sk)
1843 {
1844         struct tcp_sock *tp = tcp_sk(sk);
1845         int decr = tp->snd_cwnd_cnt + 1;
1846
1847         tp->snd_cwnd_cnt = decr&1;
1848         decr >>= 1;
1849
1850         if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1851                 tp->snd_cwnd -= decr;
1852
1853         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1854         tp->snd_cwnd_stamp = tcp_time_stamp;
1855 }
1856
1857 /* Nothing was retransmitted or returned timestamp is less
1858  * than timestamp of the first retransmission.
1859  */
1860 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1861 {
1862         return !tp->retrans_stamp ||
1863                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1864                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1865 }
1866
1867 /* Undo procedures. */
1868
1869 #if FASTRETRANS_DEBUG > 1
1870 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1871 {
1872         struct inet_sock *inet = inet_sk(sk);
1873         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1874                msg,
1875                NIPQUAD(inet->daddr), ntohs(inet->dport),
1876                tp->snd_cwnd, tp->left_out,
1877                tp->snd_ssthresh, tp->prior_ssthresh,
1878                tp->packets_out);
1879 }
1880 #else
1881 #define DBGUNDO(x...) do { } while (0)
1882 #endif
1883
1884 static void tcp_undo_cwr(struct sock *sk, const int undo)
1885 {
1886         struct tcp_sock *tp = tcp_sk(sk);
1887
1888         if (tp->prior_ssthresh) {
1889                 const struct inet_connection_sock *icsk = inet_csk(sk);
1890
1891                 if (icsk->icsk_ca_ops->undo_cwnd)
1892                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1893                 else
1894                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1895
1896                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1897                         tp->snd_ssthresh = tp->prior_ssthresh;
1898                         TCP_ECN_withdraw_cwr(tp);
1899                 }
1900         } else {
1901                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1902         }
1903         tcp_moderate_cwnd(tp);
1904         tp->snd_cwnd_stamp = tcp_time_stamp;
1905
1906         /* There is something screwy going on with the retrans hints after
1907            an undo */
1908         clear_all_retrans_hints(tp);
1909 }
1910
1911 static inline int tcp_may_undo(struct tcp_sock *tp)
1912 {
1913         return tp->undo_marker &&
1914                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1915 }
1916
1917 /* People celebrate: "We love our President!" */
1918 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1919 {
1920         if (tcp_may_undo(tp)) {
1921                 /* Happy end! We did not retransmit anything
1922                  * or our original transmission succeeded.
1923                  */
1924                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1925                 tcp_undo_cwr(sk, 1);
1926                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1927                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1928                 else
1929                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1930                 tp->undo_marker = 0;
1931         }
1932         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1933                 /* Hold old state until something *above* high_seq
1934                  * is ACKed. For Reno it is MUST to prevent false
1935                  * fast retransmits (RFC2582). SACK TCP is safe. */
1936                 tcp_moderate_cwnd(tp);
1937                 return 1;
1938         }
1939         tcp_set_ca_state(sk, TCP_CA_Open);
1940         return 0;
1941 }
1942
1943 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1944 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1945 {
1946         if (tp->undo_marker && !tp->undo_retrans) {
1947                 DBGUNDO(sk, tp, "D-SACK");
1948                 tcp_undo_cwr(sk, 1);
1949                 tp->undo_marker = 0;
1950                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1951         }
1952 }
1953
1954 /* Undo during fast recovery after partial ACK. */
1955
1956 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1957                                 int acked)
1958 {
1959         /* Partial ACK arrived. Force Hoe's retransmit. */
1960         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1961
1962         if (tcp_may_undo(tp)) {
1963                 /* Plain luck! Hole if filled with delayed
1964                  * packet, rather than with a retransmit.
1965                  */
1966                 if (tp->retrans_out == 0)
1967                         tp->retrans_stamp = 0;
1968
1969                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1970
1971                 DBGUNDO(sk, tp, "Hoe");
1972                 tcp_undo_cwr(sk, 0);
1973                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1974
1975                 /* So... Do not make Hoe's retransmit yet.
1976                  * If the first packet was delayed, the rest
1977                  * ones are most probably delayed as well.
1978                  */
1979                 failed = 0;
1980         }
1981         return failed;
1982 }
1983
1984 /* Undo during loss recovery after partial ACK. */
1985 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1986 {
1987         if (tcp_may_undo(tp)) {
1988                 struct sk_buff *skb;
1989                 tcp_for_write_queue(skb, sk) {
1990                         if (skb == tcp_send_head(sk))
1991                                 break;
1992                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1993                 }
1994
1995                 clear_all_retrans_hints(tp);
1996
1997                 DBGUNDO(sk, tp, "partial loss");
1998                 tp->lost_out = 0;
1999                 tp->left_out = tp->sacked_out;
2000                 tcp_undo_cwr(sk, 1);
2001                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2002                 inet_csk(sk)->icsk_retransmits = 0;
2003                 tp->undo_marker = 0;
2004                 if (!IsReno(tp))
2005                         tcp_set_ca_state(sk, TCP_CA_Open);
2006                 return 1;
2007         }
2008         return 0;
2009 }
2010
2011 static inline void tcp_complete_cwr(struct sock *sk)
2012 {
2013         struct tcp_sock *tp = tcp_sk(sk);
2014         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2015         tp->snd_cwnd_stamp = tcp_time_stamp;
2016         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2017 }
2018
2019 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
2020 {
2021         tp->left_out = tp->sacked_out;
2022
2023         if (tp->retrans_out == 0)
2024                 tp->retrans_stamp = 0;
2025
2026         if (flag&FLAG_ECE)
2027                 tcp_enter_cwr(sk, 1);
2028
2029         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2030                 int state = TCP_CA_Open;
2031
2032                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2033                         state = TCP_CA_Disorder;
2034
2035                 if (inet_csk(sk)->icsk_ca_state != state) {
2036                         tcp_set_ca_state(sk, state);
2037                         tp->high_seq = tp->snd_nxt;
2038                 }
2039                 tcp_moderate_cwnd(tp);
2040         } else {
2041                 tcp_cwnd_down(sk);
2042         }
2043 }
2044
2045 static void tcp_mtup_probe_failed(struct sock *sk)
2046 {
2047         struct inet_connection_sock *icsk = inet_csk(sk);
2048
2049         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2050         icsk->icsk_mtup.probe_size = 0;
2051 }
2052
2053 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2054 {
2055         struct tcp_sock *tp = tcp_sk(sk);
2056         struct inet_connection_sock *icsk = inet_csk(sk);
2057
2058         /* FIXME: breaks with very large cwnd */
2059         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2060         tp->snd_cwnd = tp->snd_cwnd *
2061                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2062                        icsk->icsk_mtup.probe_size;
2063         tp->snd_cwnd_cnt = 0;
2064         tp->snd_cwnd_stamp = tcp_time_stamp;
2065         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2066
2067         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2068         icsk->icsk_mtup.probe_size = 0;
2069         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2070 }
2071
2072
2073 /* Process an event, which can update packets-in-flight not trivially.
2074  * Main goal of this function is to calculate new estimate for left_out,
2075  * taking into account both packets sitting in receiver's buffer and
2076  * packets lost by network.
2077  *
2078  * Besides that it does CWND reduction, when packet loss is detected
2079  * and changes state of machine.
2080  *
2081  * It does _not_ decide what to send, it is made in function
2082  * tcp_xmit_retransmit_queue().
2083  */
2084 static void
2085 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
2086                       int prior_packets, int flag)
2087 {
2088         struct inet_connection_sock *icsk = inet_csk(sk);
2089         struct tcp_sock *tp = tcp_sk(sk);
2090         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
2091
2092         /* Some technical things:
2093          * 1. Reno does not count dupacks (sacked_out) automatically. */
2094         if (!tp->packets_out)
2095                 tp->sacked_out = 0;
2096         /* 2. SACK counts snd_fack in packets inaccurately. */
2097         if (tp->sacked_out == 0)
2098                 tp->fackets_out = 0;
2099
2100         /* Now state machine starts.
2101          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2102         if (flag&FLAG_ECE)
2103                 tp->prior_ssthresh = 0;
2104
2105         /* B. In all the states check for reneging SACKs. */
2106         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2107                 return;
2108
2109         /* C. Process data loss notification, provided it is valid. */
2110         if ((flag&FLAG_DATA_LOST) &&
2111             before(tp->snd_una, tp->high_seq) &&
2112             icsk->icsk_ca_state != TCP_CA_Open &&
2113             tp->fackets_out > tp->reordering) {
2114                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
2115                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2116         }
2117
2118         /* D. Synchronize left_out to current state. */
2119         tcp_sync_left_out(tp);
2120
2121         /* E. Check state exit conditions. State can be terminated
2122          *    when high_seq is ACKed. */
2123         if (icsk->icsk_ca_state == TCP_CA_Open) {
2124                 BUG_TRAP(tp->retrans_out == 0);
2125                 tp->retrans_stamp = 0;
2126         } else if (!before(tp->snd_una, tp->high_seq)) {
2127                 switch (icsk->icsk_ca_state) {
2128                 case TCP_CA_Loss:
2129                         icsk->icsk_retransmits = 0;
2130                         if (tcp_try_undo_recovery(sk, tp))
2131                                 return;
2132                         break;
2133
2134                 case TCP_CA_CWR:
2135                         /* CWR is to be held something *above* high_seq
2136                          * is ACKed for CWR bit to reach receiver. */
2137                         if (tp->snd_una != tp->high_seq) {
2138                                 tcp_complete_cwr(sk);
2139                                 tcp_set_ca_state(sk, TCP_CA_Open);
2140                         }
2141                         break;
2142
2143                 case TCP_CA_Disorder:
2144                         tcp_try_undo_dsack(sk, tp);
2145                         if (!tp->undo_marker ||
2146                             /* For SACK case do not Open to allow to undo
2147                              * catching for all duplicate ACKs. */
2148                             IsReno(tp) || tp->snd_una != tp->high_seq) {
2149                                 tp->undo_marker = 0;
2150                                 tcp_set_ca_state(sk, TCP_CA_Open);
2151                         }
2152                         break;
2153
2154                 case TCP_CA_Recovery:
2155                         if (IsReno(tp))
2156                                 tcp_reset_reno_sack(tp);
2157                         if (tcp_try_undo_recovery(sk, tp))
2158                                 return;
2159                         tcp_complete_cwr(sk);
2160                         break;
2161                 }
2162         }
2163
2164         /* F. Process state. */
2165         switch (icsk->icsk_ca_state) {
2166         case TCP_CA_Recovery:
2167                 if (prior_snd_una == tp->snd_una) {
2168                         if (IsReno(tp) && is_dupack)
2169                                 tcp_add_reno_sack(sk);
2170                 } else {
2171                         int acked = prior_packets - tp->packets_out;
2172                         if (IsReno(tp))
2173                                 tcp_remove_reno_sacks(sk, tp, acked);
2174                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
2175                 }
2176                 break;
2177         case TCP_CA_Loss:
2178                 if (flag&FLAG_DATA_ACKED)
2179                         icsk->icsk_retransmits = 0;
2180                 if (!tcp_try_undo_loss(sk, tp)) {
2181                         tcp_moderate_cwnd(tp);
2182                         tcp_xmit_retransmit_queue(sk);
2183                         return;
2184                 }
2185                 if (icsk->icsk_ca_state != TCP_CA_Open)
2186                         return;
2187                 /* Loss is undone; fall through to processing in Open state. */
2188         default:
2189                 if (IsReno(tp)) {
2190                         if (tp->snd_una != prior_snd_una)
2191                                 tcp_reset_reno_sack(tp);
2192                         if (is_dupack)
2193                                 tcp_add_reno_sack(sk);
2194                 }
2195
2196                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2197                         tcp_try_undo_dsack(sk, tp);
2198
2199                 if (!tcp_time_to_recover(sk, tp)) {
2200                         tcp_try_to_open(sk, tp, flag);
2201                         return;
2202                 }
2203
2204                 /* MTU probe failure: don't reduce cwnd */
2205                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2206                     icsk->icsk_mtup.probe_size &&
2207                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2208                         tcp_mtup_probe_failed(sk);
2209                         /* Restores the reduction we did in tcp_mtup_probe() */
2210                         tp->snd_cwnd++;
2211                         tcp_simple_retransmit(sk);
2212                         return;
2213                 }
2214
2215                 /* Otherwise enter Recovery state */
2216
2217                 if (IsReno(tp))
2218                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2219                 else
2220                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2221
2222                 tp->high_seq = tp->snd_nxt;
2223                 tp->prior_ssthresh = 0;
2224                 tp->undo_marker = tp->snd_una;
2225                 tp->undo_retrans = tp->retrans_out;
2226
2227                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2228                         if (!(flag&FLAG_ECE))
2229                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2230                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2231                         TCP_ECN_queue_cwr(tp);
2232                 }
2233
2234                 tp->bytes_acked = 0;
2235                 tp->snd_cwnd_cnt = 0;
2236                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2237         }
2238
2239         if (is_dupack || tcp_head_timedout(sk, tp))
2240                 tcp_update_scoreboard(sk, tp);
2241         tcp_cwnd_down(sk);
2242         tcp_xmit_retransmit_queue(sk);
2243 }
2244
2245 /* Read draft-ietf-tcplw-high-performance before mucking
2246  * with this code. (Supersedes RFC1323)
2247  */
2248 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2249 {
2250         /* RTTM Rule: A TSecr value received in a segment is used to
2251          * update the averaged RTT measurement only if the segment
2252          * acknowledges some new data, i.e., only if it advances the
2253          * left edge of the send window.
2254          *
2255          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2256          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2257          *
2258          * Changed: reset backoff as soon as we see the first valid sample.
2259          * If we do not, we get strongly overestimated rto. With timestamps
2260          * samples are accepted even from very old segments: f.e., when rtt=1
2261          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2262          * answer arrives rto becomes 120 seconds! If at least one of segments
2263          * in window is lost... Voila.                          --ANK (010210)
2264          */
2265         struct tcp_sock *tp = tcp_sk(sk);
2266         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2267         tcp_rtt_estimator(sk, seq_rtt);
2268         tcp_set_rto(sk);
2269         inet_csk(sk)->icsk_backoff = 0;
2270         tcp_bound_rto(sk);
2271 }
2272
2273 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2274 {
2275         /* We don't have a timestamp. Can only use
2276          * packets that are not retransmitted to determine
2277          * rtt estimates. Also, we must not reset the
2278          * backoff for rto until we get a non-retransmitted
2279          * packet. This allows us to deal with a situation
2280          * where the network delay has increased suddenly.
2281          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2282          */
2283
2284         if (flag & FLAG_RETRANS_DATA_ACKED)
2285                 return;
2286
2287         tcp_rtt_estimator(sk, seq_rtt);
2288         tcp_set_rto(sk);
2289         inet_csk(sk)->icsk_backoff = 0;
2290         tcp_bound_rto(sk);
2291 }
2292
2293 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2294                                       const s32 seq_rtt)
2295 {
2296         const struct tcp_sock *tp = tcp_sk(sk);
2297         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2298         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2299                 tcp_ack_saw_tstamp(sk, flag);
2300         else if (seq_rtt >= 0)
2301                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2302 }
2303
2304 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2305                            u32 in_flight, int good)
2306 {
2307         const struct inet_connection_sock *icsk = inet_csk(sk);
2308         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2309         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2310 }
2311
2312 /* Restart timer after forward progress on connection.
2313  * RFC2988 recommends to restart timer to now+rto.
2314  */
2315
2316 static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2317 {
2318         if (!tp->packets_out) {
2319                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2320         } else {
2321                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2322         }
2323 }
2324
2325 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2326                          __u32 now, __s32 *seq_rtt)
2327 {
2328         struct tcp_sock *tp = tcp_sk(sk);
2329         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2330         __u32 seq = tp->snd_una;
2331         __u32 packets_acked;
2332         int acked = 0;
2333
2334         /* If we get here, the whole TSO packet has not been
2335          * acked.
2336          */
2337         BUG_ON(!after(scb->end_seq, seq));
2338
2339         packets_acked = tcp_skb_pcount(skb);
2340         if (tcp_trim_head(sk, skb, seq - scb->seq))
2341                 return 0;
2342         packets_acked -= tcp_skb_pcount(skb);
2343
2344         if (packets_acked) {
2345                 __u8 sacked = scb->sacked;
2346
2347                 acked |= FLAG_DATA_ACKED;
2348                 if (sacked) {
2349                         if (sacked & TCPCB_RETRANS) {
2350                                 if (sacked & TCPCB_SACKED_RETRANS)
2351                                         tp->retrans_out -= packets_acked;
2352                                 acked |= FLAG_RETRANS_DATA_ACKED;
2353                                 *seq_rtt = -1;
2354                         } else if (*seq_rtt < 0)
2355                                 *seq_rtt = now - scb->when;
2356                         if (sacked & TCPCB_SACKED_ACKED)
2357                                 tp->sacked_out -= packets_acked;
2358                         if (sacked & TCPCB_LOST)
2359                                 tp->lost_out -= packets_acked;
2360                         if (sacked & TCPCB_URG) {
2361                                 if (tp->urg_mode &&
2362                                     !before(seq, tp->snd_up))
2363                                         tp->urg_mode = 0;
2364                         }
2365                 } else if (*seq_rtt < 0)
2366                         *seq_rtt = now - scb->when;
2367
2368                 if (tp->fackets_out) {
2369                         __u32 dval = min(tp->fackets_out, packets_acked);
2370                         tp->fackets_out -= dval;
2371                 }
2372                 tp->packets_out -= packets_acked;
2373
2374                 BUG_ON(tcp_skb_pcount(skb) == 0);
2375                 BUG_ON(!before(scb->seq, scb->end_seq));
2376         }
2377
2378         return acked;
2379 }
2380
2381 static u32 tcp_usrtt(struct timeval *tv)
2382 {
2383         struct timeval now;
2384
2385         do_gettimeofday(&now);
2386         return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec);
2387 }
2388
2389 /* Remove acknowledged frames from the retransmission queue. */
2390 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2391 {
2392         struct tcp_sock *tp = tcp_sk(sk);
2393         const struct inet_connection_sock *icsk = inet_csk(sk);
2394         struct sk_buff *skb;
2395         __u32 now = tcp_time_stamp;
2396         int acked = 0;
2397         __s32 seq_rtt = -1;
2398         u32 pkts_acked = 0;
2399         void (*rtt_sample)(struct sock *sk, u32 usrtt)
2400                 = icsk->icsk_ca_ops->rtt_sample;
2401         struct timeval tv = { .tv_sec = 0, .tv_usec = 0 };
2402
2403         while ((skb = tcp_write_queue_head(sk)) &&
2404                skb != tcp_send_head(sk)) {
2405                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2406                 __u8 sacked = scb->sacked;
2407
2408                 /* If our packet is before the ack sequence we can
2409                  * discard it as it's confirmed to have arrived at
2410                  * the other end.
2411                  */
2412                 if (after(scb->end_seq, tp->snd_una)) {
2413                         if (tcp_skb_pcount(skb) > 1 &&
2414                             after(tp->snd_una, scb->seq))
2415                                 acked |= tcp_tso_acked(sk, skb,
2416                                                        now, &seq_rtt);
2417                         break;
2418                 }
2419
2420                 /* Initial outgoing SYN's get put onto the write_queue
2421                  * just like anything else we transmit.  It is not
2422                  * true data, and if we misinform our callers that
2423                  * this ACK acks real data, we will erroneously exit
2424                  * connection startup slow start one packet too
2425                  * quickly.  This is severely frowned upon behavior.
2426                  */
2427                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2428                         acked |= FLAG_DATA_ACKED;
2429                         ++pkts_acked;
2430                 } else {
2431                         acked |= FLAG_SYN_ACKED;
2432                         tp->retrans_stamp = 0;
2433                 }
2434
2435                 /* MTU probing checks */
2436                 if (icsk->icsk_mtup.probe_size) {
2437                         if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2438                                 tcp_mtup_probe_success(sk, skb);
2439                         }
2440                 }
2441
2442                 if (sacked) {
2443                         if (sacked & TCPCB_RETRANS) {
2444                                 if (sacked & TCPCB_SACKED_RETRANS)
2445                                         tp->retrans_out -= tcp_skb_pcount(skb);
2446                                 acked |= FLAG_RETRANS_DATA_ACKED;
2447                                 seq_rtt = -1;
2448                         } else if (seq_rtt < 0) {
2449                                 seq_rtt = now - scb->when;
2450                                 skb_get_timestamp(skb, &tv);
2451                         }
2452                         if (sacked & TCPCB_SACKED_ACKED)
2453                                 tp->sacked_out -= tcp_skb_pcount(skb);
2454                         if (sacked & TCPCB_LOST)
2455                                 tp->lost_out -= tcp_skb_pcount(skb);
2456                         if (sacked & TCPCB_URG) {
2457                                 if (tp->urg_mode &&
2458                                     !before(scb->end_seq, tp->snd_up))
2459                                         tp->urg_mode = 0;
2460                         }
2461                 } else if (seq_rtt < 0) {
2462                         seq_rtt = now - scb->when;
2463                         skb_get_timestamp(skb, &tv);
2464                 }
2465                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2466                 tcp_packets_out_dec(tp, skb);
2467                 tcp_unlink_write_queue(skb, sk);
2468                 sk_stream_free_skb(sk, skb);
2469                 clear_all_retrans_hints(tp);
2470         }
2471
2472         if (acked&FLAG_ACKED) {
2473                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2474                 tcp_ack_packets_out(sk, tp);
2475                 if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
2476                         (*rtt_sample)(sk, tcp_usrtt(&tv));
2477
2478                 if (icsk->icsk_ca_ops->pkts_acked)
2479                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2480         }
2481
2482 #if FASTRETRANS_DEBUG > 0
2483         BUG_TRAP((int)tp->sacked_out >= 0);
2484         BUG_TRAP((int)tp->lost_out >= 0);
2485         BUG_TRAP((int)tp->retrans_out >= 0);
2486         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2487                 const struct inet_connection_sock *icsk = inet_csk(sk);
2488                 if (tp->lost_out) {
2489                         printk(KERN_DEBUG "Leak l=%u %d\n",
2490                                tp->lost_out, icsk->icsk_ca_state);
2491                         tp->lost_out = 0;
2492                 }
2493                 if (tp->sacked_out) {
2494                         printk(KERN_DEBUG "Leak s=%u %d\n",
2495                                tp->sacked_out, icsk->icsk_ca_state);
2496                         tp->sacked_out = 0;
2497                 }
2498                 if (tp->retrans_out) {
2499                         printk(KERN_DEBUG "Leak r=%u %d\n",
2500                                tp->retrans_out, icsk->icsk_ca_state);
2501                         tp->retrans_out = 0;
2502                 }
2503         }
2504 #endif
2505         *seq_rtt_p = seq_rtt;
2506         return acked;
2507 }
2508
2509 static void tcp_ack_probe(struct sock *sk)
2510 {
2511         const struct tcp_sock *tp = tcp_sk(sk);
2512         struct inet_connection_sock *icsk = inet_csk(sk);
2513
2514         /* Was it a usable window open? */
2515
2516         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2517                    tp->snd_una + tp->snd_wnd)) {
2518                 icsk->icsk_backoff = 0;
2519                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2520                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2521                  * This function is not for random using!
2522                  */
2523         } else {
2524                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2525                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2526                                           TCP_RTO_MAX);
2527         }
2528 }
2529
2530 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2531 {
2532         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2533                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2534 }
2535
2536 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2537 {
2538         const struct tcp_sock *tp = tcp_sk(sk);
2539         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2540                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2541 }
2542
2543 /* Check that window update is acceptable.
2544  * The function assumes that snd_una<=ack<=snd_next.
2545  */
2546 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2547                                         const u32 ack_seq, const u32 nwin)
2548 {
2549         return (after(ack, tp->snd_una) ||
2550                 after(ack_seq, tp->snd_wl1) ||
2551                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2552 }
2553
2554 /* Update our send window.
2555  *
2556  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2557  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2558  */
2559 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2560                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2561 {
2562         int flag = 0;
2563         u32 nwin = ntohs(tcp_hdr(skb)->window);
2564
2565         if (likely(!tcp_hdr(skb)->syn))
2566                 nwin <<= tp->rx_opt.snd_wscale;
2567
2568         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2569                 flag |= FLAG_WIN_UPDATE;
2570                 tcp_update_wl(tp, ack, ack_seq);
2571
2572                 if (tp->snd_wnd != nwin) {
2573                         tp->snd_wnd = nwin;
2574
2575                         /* Note, it is the only place, where
2576                          * fast path is recovered for sending TCP.
2577                          */
2578                         tp->pred_flags = 0;
2579                         tcp_fast_path_check(sk, tp);
2580
2581                         if (nwin > tp->max_window) {
2582                                 tp->max_window = nwin;
2583                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2584                         }
2585                 }
2586         }
2587
2588         tp->snd_una = ack;
2589
2590         return flag;
2591 }
2592
2593 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2594  * continue in congestion avoidance.
2595  */
2596 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2597 {
2598         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2599         tp->snd_cwnd_cnt = 0;
2600         tcp_moderate_cwnd(tp);
2601 }
2602
2603 /* A conservative spurious RTO response algorithm: reduce cwnd using
2604  * rate halving and continue in congestion avoidance.
2605  */
2606 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2607 {
2608         tcp_enter_cwr(sk, 0);
2609 }
2610
2611 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2612 {
2613         if (flag&FLAG_ECE)
2614                 tcp_ratehalving_spur_to_response(sk);
2615         else
2616                 tcp_undo_cwr(sk, 1);
2617 }
2618
2619 /* F-RTO spurious RTO detection algorithm (RFC4138)
2620  *
2621  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2622  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2623  * window (but not to or beyond highest sequence sent before RTO):
2624  *   On First ACK,  send two new segments out.
2625  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2626  *                  algorithm is not part of the F-RTO detection algorithm
2627  *                  given in RFC4138 but can be selected separately).
2628  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2629  * and TCP falls back to conventional RTO recovery.
2630  *
2631  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2632  * original window even after we transmit two new data segments.
2633  *
2634  * SACK version:
2635  *   on first step, wait until first cumulative ACK arrives, then move to
2636  *   the second step. In second step, the next ACK decides.
2637  *
2638  * F-RTO is implemented (mainly) in four functions:
2639  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2640  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2641  *     called when tcp_use_frto() showed green light
2642  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2643  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2644  *     to prove that the RTO is indeed spurious. It transfers the control
2645  *     from F-RTO to the conventional RTO recovery
2646  */
2647 static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
2648 {
2649         struct tcp_sock *tp = tcp_sk(sk);
2650
2651         tcp_sync_left_out(tp);
2652
2653         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2654         if (flag&FLAG_DATA_ACKED)
2655                 inet_csk(sk)->icsk_retransmits = 0;
2656
2657         if (!before(tp->snd_una, tp->frto_highmark)) {
2658                 tcp_enter_frto_loss(sk, tp->frto_counter + 1, flag);
2659                 return 1;
2660         }
2661
2662         if (!IsSackFrto() || IsReno(tp)) {
2663                 /* RFC4138 shortcoming in step 2; should also have case c):
2664                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2665                  * data, winupdate
2666                  */
2667                 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
2668                     !(flag&FLAG_FORWARD_PROGRESS))
2669                         return 1;
2670
2671                 if (!(flag&FLAG_DATA_ACKED)) {
2672                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2673                                             flag);
2674                         return 1;
2675                 }
2676         } else {
2677                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2678                         /* Prevent sending of new data. */
2679                         tp->snd_cwnd = min(tp->snd_cwnd,
2680                                            tcp_packets_in_flight(tp));
2681                         return 1;
2682                 }
2683
2684                 if ((tp->frto_counter == 2) &&
2685                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2686                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2687                         /* RFC4138 shortcoming (see comment above) */
2688                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2689                                 return 1;
2690
2691                         tcp_enter_frto_loss(sk, 3, flag);
2692                         return 1;
2693                 }
2694         }
2695
2696         if (tp->frto_counter == 1) {
2697                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2698                 tp->frto_counter = 2;
2699                 return 1;
2700         } else /* frto_counter == 2 */ {
2701                 switch (sysctl_tcp_frto_response) {
2702                 case 2:
2703                         tcp_undo_spur_to_response(sk, flag);
2704                         break;
2705                 case 1:
2706                         tcp_conservative_spur_to_response(tp);
2707                         break;
2708                 default:
2709                         tcp_ratehalving_spur_to_response(sk);
2710                         break;
2711                 };
2712                 tp->frto_counter = 0;
2713         }
2714         return 0;
2715 }
2716
2717 /* This routine deals with incoming acks, but not outgoing ones. */
2718 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2719 {
2720         struct inet_connection_sock *icsk = inet_csk(sk);
2721         struct tcp_sock *tp = tcp_sk(sk);
2722         u32 prior_snd_una = tp->snd_una;
2723         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2724         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2725         u32 prior_in_flight;
2726         s32 seq_rtt;
2727         int prior_packets;
2728         int frto_cwnd = 0;
2729
2730         /* If the ack is newer than sent or older than previous acks
2731          * then we can probably ignore it.
2732          */
2733         if (after(ack, tp->snd_nxt))
2734                 goto uninteresting_ack;
2735
2736         if (before(ack, prior_snd_una))
2737                 goto old_ack;
2738
2739         if (sysctl_tcp_abc) {
2740                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2741                         tp->bytes_acked += ack - prior_snd_una;
2742                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2743                         /* we assume just one segment left network */
2744                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2745         }
2746
2747         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2748                 /* Window is constant, pure forward advance.
2749                  * No more checks are required.
2750                  * Note, we use the fact that SND.UNA>=SND.WL2.
2751                  */
2752                 tcp_update_wl(tp, ack, ack_seq);
2753                 tp->snd_una = ack;
2754                 flag |= FLAG_WIN_UPDATE;
2755
2756                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2757
2758                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2759         } else {
2760                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2761                         flag |= FLAG_DATA;
2762                 else
2763                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2764
2765                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2766
2767                 if (TCP_SKB_CB(skb)->sacked)
2768                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2769
2770                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2771                         flag |= FLAG_ECE;
2772
2773                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2774         }
2775
2776         /* We passed data and got it acked, remove any soft error
2777          * log. Something worked...
2778          */
2779         sk->sk_err_soft = 0;
2780         tp->rcv_tstamp = tcp_time_stamp;
2781         prior_packets = tp->packets_out;
2782         if (!prior_packets)
2783                 goto no_queue;
2784
2785         prior_in_flight = tcp_packets_in_flight(tp);
2786
2787         /* See if we can take anything off of the retransmit queue. */
2788         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2789
2790         if (tp->frto_counter)
2791                 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
2792
2793         if (tcp_ack_is_dubious(sk, flag)) {
2794                 /* Advance CWND, if state allows this. */
2795                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2796                     tcp_may_raise_cwnd(sk, flag))
2797                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2798                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2799         } else {
2800                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2801                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2802         }
2803
2804         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2805                 dst_confirm(sk->sk_dst_cache);
2806
2807         return 1;
2808
2809 no_queue:
2810         icsk->icsk_probes_out = 0;
2811
2812         /* If this ack opens up a zero window, clear backoff.  It was
2813          * being used to time the probes, and is probably far higher than
2814          * it needs to be for normal retransmission.
2815          */
2816         if (tcp_send_head(sk))
2817                 tcp_ack_probe(sk);
2818         return 1;
2819
2820 old_ack:
2821         if (TCP_SKB_CB(skb)->sacked)
2822                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2823
2824 uninteresting_ack:
2825         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2826         return 0;
2827 }
2828
2829
2830 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2831  * But, this can also be called on packets in the established flow when
2832  * the fast version below fails.
2833  */
2834 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2835 {
2836         unsigned char *ptr;
2837         struct tcphdr *th = tcp_hdr(skb);
2838         int length=(th->doff*4)-sizeof(struct tcphdr);
2839
2840         ptr = (unsigned char *)(th + 1);
2841         opt_rx->saw_tstamp = 0;
2842
2843         while (length > 0) {
2844                 int opcode=*ptr++;
2845                 int opsize;
2846
2847                 switch (opcode) {
2848                         case TCPOPT_EOL:
2849                                 return;
2850                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2851                                 length--;
2852                                 continue;
2853                         default:
2854                                 opsize=*ptr++;
2855                                 if (opsize < 2) /* "silly options" */
2856                                         return;
2857                                 if (opsize > length)
2858                                         return; /* don't parse partial options */
2859                                 switch (opcode) {
2860                                 case TCPOPT_MSS:
2861                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2862                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2863                                                 if (in_mss) {
2864                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2865                                                                 in_mss = opt_rx->user_mss;
2866                                                         opt_rx->mss_clamp = in_mss;
2867                                                 }
2868                                         }
2869                                         break;
2870                                 case TCPOPT_WINDOW:
2871                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2872                                                 if (sysctl_tcp_window_scaling) {
2873                                                         __u8 snd_wscale = *(__u8 *) ptr;
2874                                                         opt_rx->wscale_ok = 1;
2875                                                         if (snd_wscale > 14) {
2876                                                                 if (net_ratelimit())
2877                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2878                                                                                "scaling value %d >14 received.\n",
2879                                                                                snd_wscale);
2880                                                                 snd_wscale = 14;
2881                                                         }
2882                                                         opt_rx->snd_wscale = snd_wscale;
2883                                                 }
2884                                         break;
2885                                 case TCPOPT_TIMESTAMP:
2886                                         if (opsize==TCPOLEN_TIMESTAMP) {
2887                                                 if ((estab && opt_rx->tstamp_ok) ||
2888                                                     (!estab && sysctl_tcp_timestamps)) {
2889                                                         opt_rx->saw_tstamp = 1;
2890                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2891                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2892                                                 }
2893                                         }
2894                                         break;
2895                                 case TCPOPT_SACK_PERM:
2896                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2897                                                 if (sysctl_tcp_sack) {
2898                                                         opt_rx->sack_ok = 1;
2899                                                         tcp_sack_reset(opt_rx);
2900                                                 }
2901                                         }
2902                                         break;
2903
2904                                 case TCPOPT_SACK:
2905                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2906                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2907                                            opt_rx->sack_ok) {
2908                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2909                                         }
2910 #ifdef CONFIG_TCP_MD5SIG
2911                                 case TCPOPT_MD5SIG:
2912                                         /*
2913                                          * The MD5 Hash has already been
2914                                          * checked (see tcp_v{4,6}_do_rcv()).
2915                                          */
2916                                         break;
2917 #endif
2918                                 };
2919                                 ptr+=opsize-2;
2920                                 length-=opsize;
2921                 };
2922         }
2923 }
2924
2925 /* Fast parse options. This hopes to only see timestamps.
2926  * If it is wrong it falls back on tcp_parse_options().
2927  */
2928 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2929                                   struct tcp_sock *tp)
2930 {
2931         if (th->doff == sizeof(struct tcphdr)>>2) {
2932                 tp->rx_opt.saw_tstamp = 0;
2933                 return 0;
2934         } else if (tp->rx_opt.tstamp_ok &&
2935                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2936                 __be32 *ptr = (__be32 *)(th + 1);
2937                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2938                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2939                         tp->rx_opt.saw_tstamp = 1;
2940                         ++ptr;
2941                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2942                         ++ptr;
2943                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2944                         return 1;
2945                 }
2946         }
2947         tcp_parse_options(skb, &tp->rx_opt, 1);
2948         return 1;
2949 }
2950
2951 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2952 {
2953         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2954         tp->rx_opt.ts_recent_stamp = get_seconds();
2955 }
2956
2957 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2958 {
2959         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2960                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2961                  * extra check below makes sure this can only happen
2962                  * for pure ACK frames.  -DaveM
2963                  *
2964                  * Not only, also it occurs for expired timestamps.
2965                  */
2966
2967                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2968                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2969                         tcp_store_ts_recent(tp);
2970         }
2971 }
2972
2973 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2974  *
2975  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2976  * it can pass through stack. So, the following predicate verifies that
2977  * this segment is not used for anything but congestion avoidance or
2978  * fast retransmit. Moreover, we even are able to eliminate most of such
2979  * second order effects, if we apply some small "replay" window (~RTO)
2980  * to timestamp space.
2981  *
2982  * All these measures still do not guarantee that we reject wrapped ACKs
2983  * on networks with high bandwidth, when sequence space is recycled fastly,
2984  * but it guarantees that such events will be very rare and do not affect
2985  * connection seriously. This doesn't look nice, but alas, PAWS is really
2986  * buggy extension.
2987  *
2988  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2989  * states that events when retransmit arrives after original data are rare.
2990  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2991  * the biggest problem on large power networks even with minor reordering.
2992  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2993  * up to bandwidth of 18Gigabit/sec. 8) ]
2994  */
2995
2996 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2997 {
2998         struct tcp_sock *tp = tcp_sk(sk);
2999         struct tcphdr *th = tcp_hdr(skb);
3000         u32 seq = TCP_SKB_CB(skb)->seq;
3001         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3002
3003         return (/* 1. Pure ACK with correct sequence number. */
3004                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3005
3006                 /* 2. ... and duplicate ACK. */
3007                 ack == tp->snd_una &&
3008
3009                 /* 3. ... and does not update window. */
3010                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3011
3012                 /* 4. ... and sits in replay window. */
3013                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3014 }
3015
3016 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3017 {
3018         const struct tcp_sock *tp = tcp_sk(sk);
3019         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3020                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3021                 !tcp_disordered_ack(sk, skb));
3022 }
3023
3024 /* Check segment sequence number for validity.
3025  *
3026  * Segment controls are considered valid, if the segment
3027  * fits to the window after truncation to the window. Acceptability
3028  * of data (and SYN, FIN, of course) is checked separately.
3029  * See tcp_data_queue(), for example.
3030  *
3031  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3032  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3033  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3034  * (borrowed from freebsd)
3035  */
3036
3037 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3038 {
3039         return  !before(end_seq, tp->rcv_wup) &&
3040                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3041 }
3042
3043 /* When we get a reset we do this. */
3044 static void tcp_reset(struct sock *sk)
3045 {
3046         /* We want the right error as BSD sees it (and indeed as we do). */
3047         switch (sk->sk_state) {
3048                 case TCP_SYN_SENT:
3049                         sk->sk_err = ECONNREFUSED;
3050                         break;
3051                 case TCP_CLOSE_WAIT:
3052                         sk->sk_err = EPIPE;
3053                         break;
3054                 case TCP_CLOSE:
3055                         return;
3056                 default:
3057                         sk->sk_err = ECONNRESET;
3058         }
3059
3060         if (!sock_flag(sk, SOCK_DEAD))
3061                 sk->sk_error_report(sk);
3062
3063         tcp_done(sk);
3064 }
3065
3066 /*
3067  *      Process the FIN bit. This now behaves as it is supposed to work
3068  *      and the FIN takes effect when it is validly part of sequence
3069  *      space. Not before when we get holes.
3070  *
3071  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3072  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3073  *      TIME-WAIT)
3074  *
3075  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3076  *      close and we go into CLOSING (and later onto TIME-WAIT)
3077  *
3078  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3079  */
3080 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3081 {
3082         struct tcp_sock *tp = tcp_sk(sk);
3083
3084         inet_csk_schedule_ack(sk);
3085
3086         sk->sk_shutdown |= RCV_SHUTDOWN;
3087         sock_set_flag(sk, SOCK_DONE);
3088
3089         switch (sk->sk_state) {
3090                 case TCP_SYN_RECV:
3091                 case TCP_ESTABLISHED:
3092                         /* Move to CLOSE_WAIT */
3093                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3094                         inet_csk(sk)->icsk_ack.pingpong = 1;
3095                         break;
3096
3097                 case TCP_CLOSE_WAIT:
3098                 case TCP_CLOSING:
3099                         /* Received a retransmission of the FIN, do
3100                          * nothing.
3101                          */
3102                         break;
3103                 case TCP_LAST_ACK:
3104                         /* RFC793: Remain in the LAST-ACK state. */
3105                         break;
3106
3107                 case TCP_FIN_WAIT1:
3108                         /* This case occurs when a simultaneous close
3109                          * happens, we must ack the received FIN and
3110                          * enter the CLOSING state.
3111                          */
3112                         tcp_send_ack(sk);
3113                         tcp_set_state(sk, TCP_CLOSING);
3114                         break;
3115                 case TCP_FIN_WAIT2:
3116                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3117                         tcp_send_ack(sk);
3118                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3119                         break;
3120                 default:
3121                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3122                          * cases we should never reach this piece of code.
3123                          */
3124                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3125                                __FUNCTION__, sk->sk_state);
3126                         break;
3127         };
3128
3129         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3130          * Probably, we should reset in this case. For now drop them.
3131          */
3132         __skb_queue_purge(&tp->out_of_order_queue);
3133         if (tp->rx_opt.sack_ok)
3134                 tcp_sack_reset(&tp->rx_opt);
3135         sk_stream_mem_reclaim(sk);
3136
3137         if (!sock_flag(sk, SOCK_DEAD)) {
3138                 sk->sk_state_change(sk);
3139
3140                 /* Do not send POLL_HUP for half duplex close. */
3141                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3142                     sk->sk_state == TCP_CLOSE)
3143                         sk_wake_async(sk, 1, POLL_HUP);
3144                 else
3145                         sk_wake_async(sk, 1, POLL_IN);
3146         }
3147 }
3148
3149 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3150 {
3151         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3152                 if (before(seq, sp->start_seq))
3153                         sp->start_seq = seq;
3154                 if (after(end_seq, sp->end_seq))
3155                         sp->end_seq = end_seq;
3156                 return 1;
3157         }
3158         return 0;
3159 }
3160
3161 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3162 {
3163         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3164                 if (before(seq, tp->rcv_nxt))
3165                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3166                 else
3167                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3168
3169                 tp->rx_opt.dsack = 1;
3170                 tp->duplicate_sack[0].start_seq = seq;
3171                 tp->duplicate_sack[0].end_seq = end_seq;
3172                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3173         }
3174 }
3175
3176 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3177 {
3178         if (!tp->rx_opt.dsack)
3179                 tcp_dsack_set(tp, seq, end_seq);
3180         else
3181                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3182 }
3183
3184 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3185 {
3186         struct tcp_sock *tp = tcp_sk(sk);
3187
3188         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3189             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3190                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3191                 tcp_enter_quickack_mode(sk);
3192
3193                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3194                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3195
3196                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3197                                 end_seq = tp->rcv_nxt;
3198                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3199                 }
3200         }
3201
3202         tcp_send_ack(sk);
3203 }
3204
3205 /* These routines update the SACK block as out-of-order packets arrive or
3206  * in-order packets close up the sequence space.
3207  */
3208 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3209 {
3210         int this_sack;
3211         struct tcp_sack_block *sp = &tp->selective_acks[0];
3212         struct tcp_sack_block *swalk = sp+1;
3213
3214         /* See if the recent change to the first SACK eats into
3215          * or hits the sequence space of other SACK blocks, if so coalesce.
3216          */
3217         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3218                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3219                         int i;
3220
3221                         /* Zap SWALK, by moving every further SACK up by one slot.
3222                          * Decrease num_sacks.
3223                          */
3224                         tp->rx_opt.num_sacks--;
3225                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3226                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3227                                 sp[i] = sp[i+1];
3228                         continue;
3229                 }
3230                 this_sack++, swalk++;
3231         }
3232 }
3233
3234 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3235 {
3236         __u32 tmp;
3237
3238         tmp = sack1->start_seq;
3239         sack1->start_seq = sack2->start_seq;
3240         sack2->start_seq = tmp;
3241
3242         tmp = sack1->end_seq;
3243         sack1->end_seq = sack2->end_seq;
3244         sack2->end_seq = tmp;
3245 }
3246
3247 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3248 {
3249         struct tcp_sock *tp = tcp_sk(sk);
3250         struct tcp_sack_block *sp = &tp->selective_acks[0];
3251         int cur_sacks = tp->rx_opt.num_sacks;
3252         int this_sack;
3253
3254         if (!cur_sacks)
3255                 goto new_sack;
3256
3257         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3258                 if (tcp_sack_extend(sp, seq, end_seq)) {
3259                         /* Rotate this_sack to the first one. */
3260                         for (; this_sack>0; this_sack--, sp--)
3261                                 tcp_sack_swap(sp, sp-1);
3262                         if (cur_sacks > 1)
3263                                 tcp_sack_maybe_coalesce(tp);
3264                         return;
3265                 }
3266         }
3267
3268         /* Could not find an adjacent existing SACK, build a new one,
3269          * put it at the front, and shift everyone else down.  We
3270          * always know there is at least one SACK present already here.
3271          *
3272          * If the sack array is full, forget about the last one.
3273          */
3274         if (this_sack >= 4) {
3275                 this_sack--;
3276                 tp->rx_opt.num_sacks--;
3277                 sp--;
3278         }
3279         for (; this_sack > 0; this_sack--, sp--)
3280                 *sp = *(sp-1);
3281
3282 new_sack:
3283         /* Build the new head SACK, and we're done. */
3284         sp->start_seq = seq;
3285         sp->end_seq = end_seq;
3286         tp->rx_opt.num_sacks++;
3287         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3288 }
3289
3290 /* RCV.NXT advances, some SACKs should be eaten. */
3291
3292 static void tcp_sack_remove(struct tcp_sock *tp)
3293 {
3294         struct tcp_sack_block *sp = &tp->selective_acks[0];
3295         int num_sacks = tp->rx_opt.num_sacks;
3296         int this_sack;
3297
3298         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3299         if (skb_queue_empty(&tp->out_of_order_queue)) {
3300                 tp->rx_opt.num_sacks = 0;
3301                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3302                 return;
3303         }
3304
3305         for (this_sack = 0; this_sack < num_sacks; ) {
3306                 /* Check if the start of the sack is covered by RCV.NXT. */
3307                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3308                         int i;
3309
3310                         /* RCV.NXT must cover all the block! */
3311                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3312
3313                         /* Zap this SACK, by moving forward any other SACKS. */
3314                         for (i=this_sack+1; i < num_sacks; i++)
3315                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3316                         num_sacks--;
3317                         continue;
3318                 }
3319                 this_sack++;
3320                 sp++;
3321         }
3322         if (num_sacks != tp->rx_opt.num_sacks) {
3323                 tp->rx_opt.num_sacks = num_sacks;
3324                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3325         }
3326 }
3327
3328 /* This one checks to see if we can put data from the
3329  * out_of_order queue into the receive_queue.
3330  */
3331 static void tcp_ofo_queue(struct sock *sk)
3332 {
3333         struct tcp_sock *tp = tcp_sk(sk);
3334         __u32 dsack_high = tp->rcv_nxt;
3335         struct sk_buff *skb;
3336
3337         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3338                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3339                         break;
3340
3341                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3342                         __u32 dsack = dsack_high;
3343                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3344                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3345                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3346                 }
3347
3348                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3349                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3350                         __skb_unlink(skb, &tp->out_of_order_queue);
3351                         __kfree_skb(skb);
3352                         continue;
3353                 }
3354                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3355                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3356                            TCP_SKB_CB(skb)->end_seq);
3357
3358                 __skb_unlink(skb, &tp->out_of_order_queue);
3359                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3360                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3361                 if (tcp_hdr(skb)->fin)
3362                         tcp_fin(skb, sk, tcp_hdr(skb));
3363         }
3364 }
3365
3366 static int tcp_prune_queue(struct sock *sk);
3367
3368 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3369 {
3370         struct tcphdr *th = tcp_hdr(skb);
3371         struct tcp_sock *tp = tcp_sk(sk);
3372         int eaten = -1;
3373
3374         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3375                 goto drop;
3376
3377         __skb_pull(skb, th->doff*4);
3378
3379         TCP_ECN_accept_cwr(tp, skb);
3380
3381         if (tp->rx_opt.dsack) {
3382                 tp->rx_opt.dsack = 0;
3383                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3384                                                     4 - tp->rx_opt.tstamp_ok);
3385         }
3386
3387         /*  Queue data for delivery to the user.
3388          *  Packets in sequence go to the receive queue.
3389          *  Out of sequence packets to the out_of_order_queue.
3390          */
3391         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3392                 if (tcp_receive_window(tp) == 0)
3393                         goto out_of_window;
3394
3395                 /* Ok. In sequence. In window. */
3396                 if (tp->ucopy.task == current &&
3397                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3398                     sock_owned_by_user(sk) && !tp->urg_data) {
3399                         int chunk = min_t(unsigned int, skb->len,
3400                                                         tp->ucopy.len);
3401
3402                         __set_current_state(TASK_RUNNING);
3403
3404                         local_bh_enable();
3405                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3406                                 tp->ucopy.len -= chunk;
3407                                 tp->copied_seq += chunk;
3408                                 eaten = (chunk == skb->len && !th->fin);
3409                                 tcp_rcv_space_adjust(sk);
3410                         }
3411                         local_bh_disable();
3412                 }
3413
3414                 if (eaten <= 0) {
3415 queue_and_out:
3416                         if (eaten < 0 &&
3417                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3418                              !sk_stream_rmem_schedule(sk, skb))) {
3419                                 if (tcp_prune_queue(sk) < 0 ||
3420                                     !sk_stream_rmem_schedule(sk, skb))
3421                                         goto drop;
3422                         }
3423                         sk_stream_set_owner_r(skb, sk);
3424                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3425                 }
3426                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3427                 if (skb->len)
3428                         tcp_event_data_recv(sk, tp, skb);
3429                 if (th->fin)
3430                         tcp_fin(skb, sk, th);
3431
3432                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3433                         tcp_ofo_queue(sk);
3434
3435                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3436                          * gap in queue is filled.
3437                          */
3438                         if (skb_queue_empty(&tp->out_of_order_queue))
3439                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3440                 }
3441
3442                 if (tp->rx_opt.num_sacks)
3443                         tcp_sack_remove(tp);
3444
3445                 tcp_fast_path_check(sk, tp);
3446
3447                 if (eaten > 0)
3448                         __kfree_skb(skb);
3449                 else if (!sock_flag(sk, SOCK_DEAD))
3450                         sk->sk_data_ready(sk, 0);
3451                 return;
3452         }
3453
3454         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3455                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3456                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3457                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3458
3459 out_of_window:
3460                 tcp_enter_quickack_mode(sk);
3461                 inet_csk_schedule_ack(sk);
3462 drop:
3463                 __kfree_skb(skb);
3464                 return;
3465         }
3466
3467         /* Out of window. F.e. zero window probe. */
3468         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3469                 goto out_of_window;
3470
3471         tcp_enter_quickack_mode(sk);
3472
3473         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3474                 /* Partial packet, seq < rcv_next < end_seq */
3475                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3476                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3477                            TCP_SKB_CB(skb)->end_seq);
3478
3479                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3480
3481                 /* If window is closed, drop tail of packet. But after
3482                  * remembering D-SACK for its head made in previous line.
3483                  */
3484                 if (!tcp_receive_window(tp))
3485                         goto out_of_window;
3486                 goto queue_and_out;
3487         }
3488
3489         TCP_ECN_check_ce(tp, skb);
3490
3491         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3492             !sk_stream_rmem_schedule(sk, skb)) {
3493                 if (tcp_prune_queue(sk) < 0 ||
3494                     !sk_stream_rmem_schedule(sk, skb))
3495                         goto drop;
3496         }
3497
3498         /* Disable header prediction. */
3499         tp->pred_flags = 0;
3500         inet_csk_schedule_ack(sk);
3501
3502         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3503                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3504
3505         sk_stream_set_owner_r(skb, sk);
3506
3507         if (!skb_peek(&tp->out_of_order_queue)) {
3508                 /* Initial out of order segment, build 1 SACK. */
3509                 if (tp->rx_opt.sack_ok) {
3510                         tp->rx_opt.num_sacks = 1;
3511                         tp->rx_opt.dsack     = 0;
3512                         tp->rx_opt.eff_sacks = 1;
3513                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3514                         tp->selective_acks[0].end_seq =
3515                                                 TCP_SKB_CB(skb)->end_seq;
3516                 }
3517                 __skb_queue_head(&tp->out_of_order_queue,skb);
3518         } else {
3519                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3520                 u32 seq = TCP_SKB_CB(skb)->seq;
3521                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3522
3523                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3524                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3525
3526                         if (!tp->rx_opt.num_sacks ||
3527                             tp->selective_acks[0].end_seq != seq)
3528                                 goto add_sack;
3529
3530                         /* Common case: data arrive in order after hole. */
3531                         tp->selective_acks[0].end_seq = end_seq;
3532                         return;
3533                 }
3534
3535                 /* Find place to insert this segment. */
3536                 do {
3537                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3538                                 break;
3539                 } while ((skb1 = skb1->prev) !=
3540                          (struct sk_buff*)&tp->out_of_order_queue);
3541
3542                 /* Do skb overlap to previous one? */
3543                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3544                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3545                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3546                                 /* All the bits are present. Drop. */
3547                                 __kfree_skb(skb);
3548                                 tcp_dsack_set(tp, seq, end_seq);
3549                                 goto add_sack;
3550                         }
3551                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3552                                 /* Partial overlap. */
3553                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3554                         } else {
3555                                 skb1 = skb1->prev;
3556                         }
3557                 }
3558                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3559
3560                 /* And clean segments covered by new one as whole. */
3561                 while ((skb1 = skb->next) !=
3562                        (struct sk_buff*)&tp->out_of_order_queue &&
3563                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3564                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3565                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3566                                break;
3567                        }
3568                        __skb_unlink(skb1, &tp->out_of_order_queue);
3569                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3570                        __kfree_skb(skb1);
3571                 }
3572
3573 add_sack:
3574                 if (tp->rx_opt.sack_ok)
3575                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3576         }
3577 }
3578
3579 /* Collapse contiguous sequence of skbs head..tail with
3580  * sequence numbers start..end.
3581  * Segments with FIN/SYN are not collapsed (only because this
3582  * simplifies code)
3583  */
3584 static void
3585 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3586              struct sk_buff *head, struct sk_buff *tail,
3587              u32 start, u32 end)
3588 {
3589         struct sk_buff *skb;
3590
3591         /* First, check that queue is collapsible and find
3592          * the point where collapsing can be useful. */
3593         for (skb = head; skb != tail; ) {
3594                 /* No new bits? It is possible on ofo queue. */
3595                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3596                         struct sk_buff *next = skb->next;
3597                         __skb_unlink(skb, list);
3598                         __kfree_skb(skb);
3599                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3600                         skb = next;
3601                         continue;
3602                 }
3603
3604                 /* The first skb to collapse is:
3605                  * - not SYN/FIN and
3606                  * - bloated or contains data before "start" or
3607                  *   overlaps to the next one.
3608                  */
3609                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3610                     (tcp_win_from_space(skb->truesize) > skb->len ||
3611                      before(TCP_SKB_CB(skb)->seq, start) ||
3612                      (skb->next != tail &&
3613                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3614                         break;
3615
3616                 /* Decided to skip this, advance start seq. */
3617                 start = TCP_SKB_CB(skb)->end_seq;
3618                 skb = skb->next;
3619         }
3620         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3621                 return;
3622
3623         while (before(start, end)) {
3624                 struct sk_buff *nskb;
3625                 int header = skb_headroom(skb);
3626                 int copy = SKB_MAX_ORDER(header, 0);
3627
3628                 /* Too big header? This can happen with IPv6. */
3629                 if (copy < 0)
3630                         return;
3631                 if (end-start < copy)
3632                         copy = end-start;
3633                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3634                 if (!nskb)
3635                         return;
3636
3637                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3638                 skb_set_network_header(nskb, (skb_network_header(skb) -
3639                                               skb->head));
3640                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3641                                                 skb->head));
3642                 skb_reserve(nskb, header);
3643                 memcpy(nskb->head, skb->head, header);
3644                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3645                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3646                 __skb_insert(nskb, skb->prev, skb, list);
3647                 sk_stream_set_owner_r(nskb, sk);
3648
3649                 /* Copy data, releasing collapsed skbs. */
3650                 while (copy > 0) {
3651                         int offset = start - TCP_SKB_CB(skb)->seq;
3652                         int size = TCP_SKB_CB(skb)->end_seq - start;
3653
3654                         BUG_ON(offset < 0);
3655                         if (size > 0) {
3656                                 size = min(copy, size);
3657                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3658                                         BUG();
3659                                 TCP_SKB_CB(nskb)->end_seq += size;
3660                                 copy -= size;
3661                                 start += size;
3662                         }
3663                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3664                                 struct sk_buff *next = skb->next;
3665                                 __skb_unlink(skb, list);
3666                                 __kfree_skb(skb);
3667                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3668                                 skb = next;
3669                                 if (skb == tail ||
3670                                     tcp_hdr(skb)->syn ||
3671                                     tcp_hdr(skb)->fin)
3672                                         return;
3673                         }
3674                 }
3675         }
3676 }
3677
3678 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3679  * and tcp_collapse() them until all the queue is collapsed.
3680  */
3681 static void tcp_collapse_ofo_queue(struct sock *sk)
3682 {
3683         struct tcp_sock *tp = tcp_sk(sk);
3684         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3685         struct sk_buff *head;
3686         u32 start, end;
3687
3688         if (skb == NULL)
3689                 return;
3690
3691         start = TCP_SKB_CB(skb)->seq;
3692         end = TCP_SKB_CB(skb)->end_seq;
3693         head = skb;
3694
3695         for (;;) {
3696                 skb = skb->next;
3697
3698                 /* Segment is terminated when we see gap or when
3699                  * we are at the end of all the queue. */
3700                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3701                     after(TCP_SKB_CB(skb)->seq, end) ||
3702                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3703                         tcp_collapse(sk, &tp->out_of_order_queue,
3704                                      head, skb, start, end);
3705                         head = skb;
3706                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3707                                 break;
3708                         /* Start new segment */
3709                         start = TCP_SKB_CB(skb)->seq;
3710                         end = TCP_SKB_CB(skb)->end_seq;
3711                 } else {
3712                         if (before(TCP_SKB_CB(skb)->seq, start))
3713                                 start = TCP_SKB_CB(skb)->seq;
3714                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3715                                 end = TCP_SKB_CB(skb)->end_seq;
3716                 }
3717         }
3718 }
3719
3720 /* Reduce allocated memory if we can, trying to get
3721  * the socket within its memory limits again.
3722  *
3723  * Return less than zero if we should start dropping frames
3724  * until the socket owning process reads some of the data
3725  * to stabilize the situation.
3726  */
3727 static int tcp_prune_queue(struct sock *sk)
3728 {
3729         struct tcp_sock *tp = tcp_sk(sk);
3730
3731         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3732
3733         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3734
3735         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3736                 tcp_clamp_window(sk, tp);
3737         else if (tcp_memory_pressure)
3738                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3739
3740         tcp_collapse_ofo_queue(sk);
3741         tcp_collapse(sk, &sk->sk_receive_queue,
3742                      sk->sk_receive_queue.next,
3743                      (struct sk_buff*)&sk->sk_receive_queue,
3744                      tp->copied_seq, tp->rcv_nxt);
3745         sk_stream_mem_reclaim(sk);
3746
3747         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3748                 return 0;
3749
3750         /* Collapsing did not help, destructive actions follow.
3751          * This must not ever occur. */
3752
3753         /* First, purge the out_of_order queue. */
3754         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3755                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3756                 __skb_queue_purge(&tp->out_of_order_queue);
3757
3758                 /* Reset SACK state.  A conforming SACK implementation will
3759                  * do the same at a timeout based retransmit.  When a connection
3760                  * is in a sad state like this, we care only about integrity
3761                  * of the connection not performance.
3762                  */
3763                 if (tp->rx_opt.sack_ok)
3764                         tcp_sack_reset(&tp->rx_opt);
3765                 sk_stream_mem_reclaim(sk);
3766         }
3767
3768         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3769                 return 0;
3770
3771         /* If we are really being abused, tell the caller to silently
3772          * drop receive data on the floor.  It will get retransmitted
3773          * and hopefully then we'll have sufficient space.
3774          */
3775         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3776
3777         /* Massive buffer overcommit. */
3778         tp->pred_flags = 0;
3779         return -1;
3780 }
3781
3782
3783 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3784  * As additional protections, we do not touch cwnd in retransmission phases,
3785  * and if application hit its sndbuf limit recently.
3786  */
3787 void tcp_cwnd_application_limited(struct sock *sk)
3788 {
3789         struct tcp_sock *tp = tcp_sk(sk);
3790
3791         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3792             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3793                 /* Limited by application or receiver window. */
3794                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3795                 u32 win_used = max(tp->snd_cwnd_used, init_win);
3796                 if (win_used < tp->snd_cwnd) {
3797                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3798                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3799                 }
3800                 tp->snd_cwnd_used = 0;
3801         }
3802         tp->snd_cwnd_stamp = tcp_time_stamp;
3803 }
3804
3805 static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3806 {
3807         /* If the user specified a specific send buffer setting, do
3808          * not modify it.
3809          */
3810         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3811                 return 0;
3812
3813         /* If we are under global TCP memory pressure, do not expand.  */
3814         if (tcp_memory_pressure)
3815                 return 0;
3816
3817         /* If we are under soft global TCP memory pressure, do not expand.  */
3818         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3819                 return 0;
3820
3821         /* If we filled the congestion window, do not expand.  */
3822         if (tp->packets_out >= tp->snd_cwnd)
3823                 return 0;
3824
3825         return 1;
3826 }
3827
3828 /* When incoming ACK allowed to free some skb from write_queue,
3829  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3830  * on the exit from tcp input handler.
3831  *
3832  * PROBLEM: sndbuf expansion does not work well with largesend.
3833  */
3834 static void tcp_new_space(struct sock *sk)
3835 {
3836         struct tcp_sock *tp = tcp_sk(sk);
3837
3838         if (tcp_should_expand_sndbuf(sk, tp)) {
3839                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3840                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3841                     demanded = max_t(unsigned int, tp->snd_cwnd,
3842                                                    tp->reordering + 1);
3843                 sndmem *= 2*demanded;
3844                 if (sndmem > sk->sk_sndbuf)
3845                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3846                 tp->snd_cwnd_stamp = tcp_time_stamp;
3847         }
3848
3849         sk->sk_write_space(sk);
3850 }
3851
3852 static void tcp_check_space(struct sock *sk)
3853 {
3854         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3855                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3856                 if (sk->sk_socket &&
3857                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3858                         tcp_new_space(sk);
3859         }
3860 }
3861
3862 static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3863 {
3864         tcp_push_pending_frames(sk, tp);
3865         tcp_check_space(sk);
3866 }
3867
3868 /*
3869  * Check if sending an ack is needed.
3870  */
3871 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3872 {
3873         struct tcp_sock *tp = tcp_sk(sk);
3874
3875             /* More than one full frame received... */
3876         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3877              /* ... and right edge of window advances far enough.
3878               * (tcp_recvmsg() will send ACK otherwise). Or...
3879               */
3880              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3881             /* We ACK each frame or... */
3882             tcp_in_quickack_mode(sk) ||
3883             /* We have out of order data. */
3884             (ofo_possible &&
3885              skb_peek(&tp->out_of_order_queue))) {
3886                 /* Then ack it now */
3887                 tcp_send_ack(sk);
3888         } else {
3889                 /* Else, send delayed ack. */
3890                 tcp_send_delayed_ack(sk);
3891         }
3892 }
3893
3894 static inline void tcp_ack_snd_check(struct sock *sk)
3895 {
3896         if (!inet_csk_ack_scheduled(sk)) {
3897                 /* We sent a data segment already. */
3898                 return;
3899         }
3900         __tcp_ack_snd_check(sk, 1);
3901 }
3902
3903 /*
3904  *      This routine is only called when we have urgent data
3905  *      signaled. Its the 'slow' part of tcp_urg. It could be
3906  *      moved inline now as tcp_urg is only called from one
3907  *      place. We handle URGent data wrong. We have to - as
3908  *      BSD still doesn't use the correction from RFC961.
3909  *      For 1003.1g we should support a new option TCP_STDURG to permit
3910  *      either form (or just set the sysctl tcp_stdurg).
3911  */
3912
3913 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3914 {
3915         struct tcp_sock *tp = tcp_sk(sk);
3916         u32 ptr = ntohs(th->urg_ptr);
3917
3918         if (ptr && !sysctl_tcp_stdurg)
3919                 ptr--;
3920         ptr += ntohl(th->seq);
3921
3922         /* Ignore urgent data that we've already seen and read. */
3923         if (after(tp->copied_seq, ptr))
3924                 return;
3925
3926         /* Do not replay urg ptr.
3927          *
3928          * NOTE: interesting situation not covered by specs.
3929          * Misbehaving sender may send urg ptr, pointing to segment,
3930          * which we already have in ofo queue. We are not able to fetch
3931          * such data and will stay in TCP_URG_NOTYET until will be eaten
3932          * by recvmsg(). Seems, we are not obliged to handle such wicked
3933          * situations. But it is worth to think about possibility of some
3934          * DoSes using some hypothetical application level deadlock.
3935          */
3936         if (before(ptr, tp->rcv_nxt))
3937                 return;
3938
3939         /* Do we already have a newer (or duplicate) urgent pointer? */
3940         if (tp->urg_data && !after(ptr, tp->urg_seq))
3941                 return;
3942
3943         /* Tell the world about our new urgent pointer. */
3944         sk_send_sigurg(sk);
3945
3946         /* We may be adding urgent data when the last byte read was
3947          * urgent. To do this requires some care. We cannot just ignore
3948          * tp->copied_seq since we would read the last urgent byte again
3949          * as data, nor can we alter copied_seq until this data arrives
3950          * or we break the semantics of SIOCATMARK (and thus sockatmark())
3951          *
3952          * NOTE. Double Dutch. Rendering to plain English: author of comment
3953          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3954          * and expect that both A and B disappear from stream. This is _wrong_.
3955          * Though this happens in BSD with high probability, this is occasional.
3956          * Any application relying on this is buggy. Note also, that fix "works"
3957          * only in this artificial test. Insert some normal data between A and B and we will
3958          * decline of BSD again. Verdict: it is better to remove to trap
3959          * buggy users.
3960          */
3961         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3962             !sock_flag(sk, SOCK_URGINLINE) &&
3963             tp->copied_seq != tp->rcv_nxt) {
3964                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3965                 tp->copied_seq++;
3966                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3967                         __skb_unlink(skb, &sk->sk_receive_queue);
3968                         __kfree_skb(skb);
3969                 }
3970         }
3971
3972         tp->urg_data   = TCP_URG_NOTYET;
3973         tp->urg_seq    = ptr;
3974
3975         /* Disable header prediction. */
3976         tp->pred_flags = 0;
3977 }
3978
3979 /* This is the 'fast' part of urgent handling. */
3980 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3981 {
3982         struct tcp_sock *tp = tcp_sk(sk);
3983
3984         /* Check if we get a new urgent pointer - normally not. */
3985         if (th->urg)
3986                 tcp_check_urg(sk,th);
3987
3988         /* Do we wait for any urgent data? - normally not... */
3989         if (tp->urg_data == TCP_URG_NOTYET) {
3990                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3991                           th->syn;
3992
3993                 /* Is the urgent pointer pointing into this packet? */
3994                 if (ptr < skb->len) {
3995                         u8 tmp;
3996                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3997                                 BUG();
3998                         tp->urg_data = TCP_URG_VALID | tmp;
3999                         if (!sock_flag(sk, SOCK_DEAD))
4000                                 sk->sk_data_ready(sk, 0);
4001                 }
4002         }
4003 }
4004
4005 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4006 {
4007         struct tcp_sock *tp = tcp_sk(sk);
4008         int chunk = skb->len - hlen;
4009         int err;
4010
4011         local_bh_enable();
4012         if (skb_csum_unnecessary(skb))
4013                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4014         else
4015                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4016                                                        tp->ucopy.iov);
4017
4018         if (!err) {
4019                 tp->ucopy.len -= chunk;
4020                 tp->copied_seq += chunk;
4021                 tcp_rcv_space_adjust(sk);
4022         }
4023
4024         local_bh_disable();
4025         return err;
4026 }
4027
4028 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4029 {
4030         __sum16 result;
4031
4032         if (sock_owned_by_user(sk)) {
4033                 local_bh_enable();
4034                 result = __tcp_checksum_complete(skb);
4035                 local_bh_disable();
4036         } else {
4037                 result = __tcp_checksum_complete(skb);
4038         }
4039         return result;
4040 }
4041
4042 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4043 {
4044         return !skb_csum_unnecessary(skb) &&
4045                 __tcp_checksum_complete_user(sk, skb);
4046 }
4047
4048 #ifdef CONFIG_NET_DMA
4049 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4050 {
4051         struct tcp_sock *tp = tcp_sk(sk);
4052         int chunk = skb->len - hlen;
4053         int dma_cookie;
4054         int copied_early = 0;
4055
4056         if (tp->ucopy.wakeup)
4057                 return 0;
4058
4059         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4060                 tp->ucopy.dma_chan = get_softnet_dma();
4061
4062         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4063
4064                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4065                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4066
4067                 if (dma_cookie < 0)
4068                         goto out;
4069
4070                 tp->ucopy.dma_cookie = dma_cookie;
4071                 copied_early = 1;
4072
4073                 tp->ucopy.len -= chunk;
4074                 tp->copied_seq += chunk;
4075                 tcp_rcv_space_adjust(sk);
4076
4077                 if ((tp->ucopy.len == 0) ||
4078                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4079                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4080                         tp->ucopy.wakeup = 1;
4081                         sk->sk_data_ready(sk, 0);
4082                 }
4083         } else if (chunk > 0) {
4084                 tp->ucopy.wakeup = 1;
4085                 sk->sk_data_ready(sk, 0);
4086         }
4087 out:
4088         return copied_early;
4089 }
4090 #endif /* CONFIG_NET_DMA */
4091
4092 /*
4093  *      TCP receive function for the ESTABLISHED state.
4094  *
4095  *      It is split into a fast path and a slow path. The fast path is
4096  *      disabled when:
4097  *      - A zero window was announced from us - zero window probing
4098  *        is only handled properly in the slow path.
4099  *      - Out of order segments arrived.
4100  *      - Urgent data is expected.
4101  *      - There is no buffer space left
4102  *      - Unexpected TCP flags/window values/header lengths are received
4103  *        (detected by checking the TCP header against pred_flags)
4104  *      - Data is sent in both directions. Fast path only supports pure senders
4105  *        or pure receivers (this means either the sequence number or the ack
4106  *        value must stay constant)
4107  *      - Unexpected TCP option.
4108  *
4109  *      When these conditions are not satisfied it drops into a standard
4110  *      receive procedure patterned after RFC793 to handle all cases.
4111  *      The first three cases are guaranteed by proper pred_flags setting,
4112  *      the rest is checked inline. Fast processing is turned on in
4113  *      tcp_data_queue when everything is OK.
4114  */
4115 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4116                         struct tcphdr *th, unsigned len)
4117 {
4118         struct tcp_sock *tp = tcp_sk(sk);
4119
4120         /*
4121          *      Header prediction.
4122          *      The code loosely follows the one in the famous
4123          *      "30 instruction TCP receive" Van Jacobson mail.
4124          *
4125          *      Van's trick is to deposit buffers into socket queue
4126          *      on a device interrupt, to call tcp_recv function
4127          *      on the receive process context and checksum and copy
4128          *      the buffer to user space. smart...
4129          *
4130          *      Our current scheme is not silly either but we take the
4131          *      extra cost of the net_bh soft interrupt processing...
4132          *      We do checksum and copy also but from device to kernel.
4133          */
4134
4135         tp->rx_opt.saw_tstamp = 0;
4136
4137         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4138          *      if header_prediction is to be made
4139          *      'S' will always be tp->tcp_header_len >> 2
4140          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4141          *  turn it off (when there are holes in the receive
4142          *       space for instance)
4143          *      PSH flag is ignored.
4144          */
4145
4146         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4147                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4148                 int tcp_header_len = tp->tcp_header_len;
4149
4150                 /* Timestamp header prediction: tcp_header_len
4151                  * is automatically equal to th->doff*4 due to pred_flags
4152                  * match.
4153                  */
4154
4155                 /* Check timestamp */
4156                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4157                         __be32 *ptr = (__be32 *)(th + 1);
4158
4159                         /* No? Slow path! */
4160                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4161                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4162                                 goto slow_path;
4163
4164                         tp->rx_opt.saw_tstamp = 1;
4165                         ++ptr;
4166                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4167                         ++ptr;
4168                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4169
4170                         /* If PAWS failed, check it more carefully in slow path */
4171                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4172                                 goto slow_path;
4173
4174                         /* DO NOT update ts_recent here, if checksum fails
4175                          * and timestamp was corrupted part, it will result
4176                          * in a hung connection since we will drop all
4177                          * future packets due to the PAWS test.
4178                          */
4179                 }
4180
4181                 if (len <= tcp_header_len) {
4182                         /* Bulk data transfer: sender */
4183                         if (len == tcp_header_len) {
4184                                 /* Predicted packet is in window by definition.
4185                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4186                                  * Hence, check seq<=rcv_wup reduces to:
4187                                  */
4188                                 if (tcp_header_len ==
4189                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4190                                     tp->rcv_nxt == tp->rcv_wup)
4191                                         tcp_store_ts_recent(tp);
4192
4193                                 /* We know that such packets are checksummed
4194                                  * on entry.
4195                                  */
4196                                 tcp_ack(sk, skb, 0);
4197                                 __kfree_skb(skb);
4198                                 tcp_data_snd_check(sk, tp);
4199                                 return 0;
4200                         } else { /* Header too small */
4201                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4202                                 goto discard;
4203                         }
4204                 } else {
4205                         int eaten = 0;
4206                         int copied_early = 0;
4207
4208                         if (tp->copied_seq == tp->rcv_nxt &&
4209                             len - tcp_header_len <= tp->ucopy.len) {
4210 #ifdef CONFIG_NET_DMA
4211                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4212                                         copied_early = 1;
4213                                         eaten = 1;
4214                                 }
4215 #endif
4216                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4217                                         __set_current_state(TASK_RUNNING);
4218
4219                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4220                                                 eaten = 1;
4221                                 }
4222                                 if (eaten) {
4223                                         /* Predicted packet is in window by definition.
4224                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4225                                          * Hence, check seq<=rcv_wup reduces to:
4226                                          */
4227                                         if (tcp_header_len ==
4228                                             (sizeof(struct tcphdr) +
4229                                              TCPOLEN_TSTAMP_ALIGNED) &&
4230                                             tp->rcv_nxt == tp->rcv_wup)
4231                                                 tcp_store_ts_recent(tp);
4232
4233                                         tcp_rcv_rtt_measure_ts(sk, skb);
4234
4235                                         __skb_pull(skb, tcp_header_len);
4236                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4237                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4238                                 }
4239                                 if (copied_early)
4240                                         tcp_cleanup_rbuf(sk, skb->len);
4241                         }
4242                         if (!eaten) {
4243                                 if (tcp_checksum_complete_user(sk, skb))
4244                                         goto csum_error;
4245
4246                                 /* Predicted packet is in window by definition.
4247                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4248                                  * Hence, check seq<=rcv_wup reduces to:
4249                                  */
4250                                 if (tcp_header_len ==
4251                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4252                                     tp->rcv_nxt == tp->rcv_wup)
4253                                         tcp_store_ts_recent(tp);
4254
4255                                 tcp_rcv_rtt_measure_ts(sk, skb);
4256
4257                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4258                                         goto step5;
4259
4260                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4261
4262                                 /* Bulk data transfer: receiver */
4263                                 __skb_pull(skb,tcp_header_len);
4264                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4265                                 sk_stream_set_owner_r(skb, sk);
4266                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4267                         }
4268
4269                         tcp_event_data_recv(sk, tp, skb);
4270
4271                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4272                                 /* Well, only one small jumplet in fast path... */
4273                                 tcp_ack(sk, skb, FLAG_DATA);
4274                                 tcp_data_snd_check(sk, tp);
4275                                 if (!inet_csk_ack_scheduled(sk))
4276                                         goto no_ack;
4277                         }
4278
4279                         __tcp_ack_snd_check(sk, 0);
4280 no_ack:
4281 #ifdef CONFIG_NET_DMA
4282                         if (copied_early)
4283                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4284                         else
4285 #endif
4286                         if (eaten)
4287                                 __kfree_skb(skb);
4288                         else
4289                                 sk->sk_data_ready(sk, 0);
4290                         return 0;
4291                 }
4292         }
4293
4294 slow_path:
4295         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4296                 goto csum_error;
4297
4298         /*
4299          * RFC1323: H1. Apply PAWS check first.
4300          */
4301         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4302             tcp_paws_discard(sk, skb)) {
4303                 if (!th->rst) {
4304                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4305                         tcp_send_dupack(sk, skb);
4306                         goto discard;
4307                 }
4308                 /* Resets are accepted even if PAWS failed.
4309
4310                    ts_recent update must be made after we are sure
4311                    that the packet is in window.
4312                  */
4313         }
4314
4315         /*
4316          *      Standard slow path.
4317          */
4318
4319         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4320                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4321                  * (RST) segments are validated by checking their SEQ-fields."
4322                  * And page 69: "If an incoming segment is not acceptable,
4323                  * an acknowledgment should be sent in reply (unless the RST bit
4324                  * is set, if so drop the segment and return)".
4325                  */
4326                 if (!th->rst)
4327                         tcp_send_dupack(sk, skb);
4328                 goto discard;
4329         }
4330
4331         if (th->rst) {
4332                 tcp_reset(sk);
4333                 goto discard;
4334         }
4335
4336         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4337
4338         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4339                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4340                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4341                 tcp_reset(sk);
4342                 return 1;
4343         }
4344
4345 step5:
4346         if (th->ack)
4347                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4348
4349         tcp_rcv_rtt_measure_ts(sk, skb);
4350
4351         /* Process urgent data. */
4352         tcp_urg(sk, skb, th);
4353
4354         /* step 7: process the segment text */
4355         tcp_data_queue(sk, skb);
4356
4357         tcp_data_snd_check(sk, tp);
4358         tcp_ack_snd_check(sk);
4359         return 0;
4360
4361 csum_error:
4362         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4363
4364 discard:
4365         __kfree_skb(skb);
4366         return 0;
4367 }
4368
4369 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4370                                          struct tcphdr *th, unsigned len)
4371 {
4372         struct tcp_sock *tp = tcp_sk(sk);
4373         struct inet_connection_sock *icsk = inet_csk(sk);
4374         int saved_clamp = tp->rx_opt.mss_clamp;
4375
4376         tcp_parse_options(skb, &tp->rx_opt, 0);
4377
4378         if (th->ack) {
4379                 /* rfc793:
4380                  * "If the state is SYN-SENT then
4381                  *    first check the ACK bit
4382                  *      If the ACK bit is set
4383                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4384                  *        a reset (unless the RST bit is set, if so drop
4385                  *        the segment and return)"
4386                  *
4387                  *  We do not send data with SYN, so that RFC-correct
4388                  *  test reduces to:
4389                  */
4390                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4391                         goto reset_and_undo;
4392
4393                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4394                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4395                              tcp_time_stamp)) {
4396                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4397                         goto reset_and_undo;
4398                 }
4399
4400                 /* Now ACK is acceptable.
4401                  *
4402                  * "If the RST bit is set
4403                  *    If the ACK was acceptable then signal the user "error:
4404                  *    connection reset", drop the segment, enter CLOSED state,
4405                  *    delete TCB, and return."
4406                  */
4407
4408                 if (th->rst) {
4409                         tcp_reset(sk);
4410                         goto discard;
4411                 }
4412
4413                 /* rfc793:
4414                  *   "fifth, if neither of the SYN or RST bits is set then
4415                  *    drop the segment and return."
4416                  *
4417                  *    See note below!
4418                  *                                        --ANK(990513)
4419                  */
4420                 if (!th->syn)
4421                         goto discard_and_undo;
4422
4423                 /* rfc793:
4424                  *   "If the SYN bit is on ...
4425                  *    are acceptable then ...
4426                  *    (our SYN has been ACKed), change the connection
4427                  *    state to ESTABLISHED..."
4428                  */
4429
4430                 TCP_ECN_rcv_synack(tp, th);
4431
4432                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4433                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4434
4435                 /* Ok.. it's good. Set up sequence numbers and
4436                  * move to established.
4437                  */
4438                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4439                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4440
4441                 /* RFC1323: The window in SYN & SYN/ACK segments is
4442                  * never scaled.
4443                  */
4444                 tp->snd_wnd = ntohs(th->window);
4445                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4446
4447                 if (!tp->rx_opt.wscale_ok) {
4448                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4449                         tp->window_clamp = min(tp->window_clamp, 65535U);
4450                 }
4451
4452                 if (tp->rx_opt.saw_tstamp) {
4453                         tp->rx_opt.tstamp_ok       = 1;
4454                         tp->tcp_header_len =
4455                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4456                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4457                         tcp_store_ts_recent(tp);
4458                 } else {
4459                         tp->tcp_header_len = sizeof(struct tcphdr);
4460                 }
4461
4462                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4463                         tp->rx_opt.sack_ok |= 2;
4464
4465                 tcp_mtup_init(sk);
4466                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4467                 tcp_initialize_rcv_mss(sk);
4468
4469                 /* Remember, tcp_poll() does not lock socket!
4470                  * Change state from SYN-SENT only after copied_seq
4471                  * is initialized. */
4472                 tp->copied_seq = tp->rcv_nxt;
4473                 smp_mb();
4474                 tcp_set_state(sk, TCP_ESTABLISHED);
4475
4476                 security_inet_conn_established(sk, skb);
4477
4478                 /* Make sure socket is routed, for correct metrics.  */
4479                 icsk->icsk_af_ops->rebuild_header(sk);
4480
4481                 tcp_init_metrics(sk);
4482
4483                 tcp_init_congestion_control(sk);
4484
4485                 /* Prevent spurious tcp_cwnd_restart() on first data
4486                  * packet.
4487                  */
4488                 tp->lsndtime = tcp_time_stamp;
4489
4490                 tcp_init_buffer_space(sk);
4491
4492                 if (sock_flag(sk, SOCK_KEEPOPEN))
4493                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4494
4495                 if (!tp->rx_opt.snd_wscale)
4496                         __tcp_fast_path_on(tp, tp->snd_wnd);
4497                 else
4498                         tp->pred_flags = 0;
4499
4500                 if (!sock_flag(sk, SOCK_DEAD)) {
4501                         sk->sk_state_change(sk);
4502                         sk_wake_async(sk, 0, POLL_OUT);
4503                 }
4504
4505                 if (sk->sk_write_pending ||
4506                     icsk->icsk_accept_queue.rskq_defer_accept ||
4507                     icsk->icsk_ack.pingpong) {
4508                         /* Save one ACK. Data will be ready after
4509                          * several ticks, if write_pending is set.
4510                          *
4511                          * It may be deleted, but with this feature tcpdumps
4512                          * look so _wonderfully_ clever, that I was not able
4513                          * to stand against the temptation 8)     --ANK
4514                          */
4515                         inet_csk_schedule_ack(sk);
4516                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4517                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4518                         tcp_incr_quickack(sk);
4519                         tcp_enter_quickack_mode(sk);
4520                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4521                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4522
4523 discard:
4524                         __kfree_skb(skb);
4525                         return 0;
4526                 } else {
4527                         tcp_send_ack(sk);
4528                 }
4529                 return -1;
4530         }
4531
4532         /* No ACK in the segment */
4533
4534         if (th->rst) {
4535                 /* rfc793:
4536                  * "If the RST bit is set
4537                  *
4538                  *      Otherwise (no ACK) drop the segment and return."
4539                  */
4540
4541                 goto discard_and_undo;
4542         }
4543
4544         /* PAWS check. */
4545         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4546                 goto discard_and_undo;
4547
4548         if (th->syn) {
4549                 /* We see SYN without ACK. It is attempt of
4550                  * simultaneous connect with crossed SYNs.
4551                  * Particularly, it can be connect to self.
4552                  */
4553                 tcp_set_state(sk, TCP_SYN_RECV);
4554
4555                 if (tp->rx_opt.saw_tstamp) {
4556                         tp->rx_opt.tstamp_ok = 1;
4557                         tcp_store_ts_recent(tp);
4558                         tp->tcp_header_len =
4559                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4560                 } else {
4561                         tp->tcp_header_len = sizeof(struct tcphdr);
4562                 }
4563
4564                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4565                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4566
4567                 /* RFC1323: The window in SYN & SYN/ACK segments is
4568                  * never scaled.
4569                  */
4570                 tp->snd_wnd    = ntohs(th->window);
4571                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4572                 tp->max_window = tp->snd_wnd;
4573
4574                 TCP_ECN_rcv_syn(tp, th);
4575
4576                 tcp_mtup_init(sk);
4577                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4578                 tcp_initialize_rcv_mss(sk);
4579
4580
4581                 tcp_send_synack(sk);
4582 #if 0
4583                 /* Note, we could accept data and URG from this segment.
4584                  * There are no obstacles to make this.
4585                  *
4586                  * However, if we ignore data in ACKless segments sometimes,
4587                  * we have no reasons to accept it sometimes.
4588                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4589                  * is not flawless. So, discard packet for sanity.
4590                  * Uncomment this return to process the data.
4591                  */
4592                 return -1;
4593 #else
4594                 goto discard;
4595 #endif
4596         }
4597         /* "fifth, if neither of the SYN or RST bits is set then
4598          * drop the segment and return."
4599          */
4600
4601 discard_and_undo:
4602         tcp_clear_options(&tp->rx_opt);
4603         tp->rx_opt.mss_clamp = saved_clamp;
4604         goto discard;
4605
4606 reset_and_undo:
4607         tcp_clear_options(&tp->rx_opt);
4608         tp->rx_opt.mss_clamp = saved_clamp;
4609         return 1;
4610 }
4611
4612
4613 /*
4614  *      This function implements the receiving procedure of RFC 793 for
4615  *      all states except ESTABLISHED and TIME_WAIT.
4616  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4617  *      address independent.
4618  */
4619
4620 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4621                           struct tcphdr *th, unsigned len)
4622 {
4623         struct tcp_sock *tp = tcp_sk(sk);
4624         struct inet_connection_sock *icsk = inet_csk(sk);
4625         int queued = 0;
4626
4627         tp->rx_opt.saw_tstamp = 0;
4628
4629         switch (sk->sk_state) {
4630         case TCP_CLOSE:
4631                 goto discard;
4632
4633         case TCP_LISTEN:
4634                 if (th->ack)
4635                         return 1;
4636
4637                 if (th->rst)
4638                         goto discard;
4639
4640                 if (th->syn) {
4641                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4642                                 return 1;
4643
4644                         /* Now we have several options: In theory there is
4645                          * nothing else in the frame. KA9Q has an option to
4646                          * send data with the syn, BSD accepts data with the
4647                          * syn up to the [to be] advertised window and
4648                          * Solaris 2.1 gives you a protocol error. For now
4649                          * we just ignore it, that fits the spec precisely
4650                          * and avoids incompatibilities. It would be nice in
4651                          * future to drop through and process the data.
4652                          *
4653                          * Now that TTCP is starting to be used we ought to
4654                          * queue this data.
4655                          * But, this leaves one open to an easy denial of
4656                          * service attack, and SYN cookies can't defend
4657                          * against this problem. So, we drop the data
4658                          * in the interest of security over speed unless
4659                          * it's still in use.
4660                          */
4661                         kfree_skb(skb);
4662                         return 0;
4663                 }
4664                 goto discard;
4665
4666         case TCP_SYN_SENT:
4667                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4668                 if (queued >= 0)
4669                         return queued;
4670
4671                 /* Do step6 onward by hand. */
4672                 tcp_urg(sk, skb, th);
4673                 __kfree_skb(skb);
4674                 tcp_data_snd_check(sk, tp);
4675                 return 0;
4676         }
4677
4678         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4679             tcp_paws_discard(sk, skb)) {
4680                 if (!th->rst) {
4681                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4682                         tcp_send_dupack(sk, skb);
4683                         goto discard;
4684                 }
4685                 /* Reset is accepted even if it did not pass PAWS. */
4686         }
4687
4688         /* step 1: check sequence number */
4689         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4690                 if (!th->rst)
4691                         tcp_send_dupack(sk, skb);
4692                 goto discard;
4693         }
4694
4695         /* step 2: check RST bit */
4696         if (th->rst) {
4697                 tcp_reset(sk);
4698                 goto discard;
4699         }
4700
4701         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4702
4703         /* step 3: check security and precedence [ignored] */
4704
4705         /*      step 4:
4706          *
4707          *      Check for a SYN in window.
4708          */
4709         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4710                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4711                 tcp_reset(sk);
4712                 return 1;
4713         }
4714
4715         /* step 5: check the ACK field */
4716         if (th->ack) {
4717                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4718
4719                 switch (sk->sk_state) {
4720                 case TCP_SYN_RECV:
4721                         if (acceptable) {
4722                                 tp->copied_seq = tp->rcv_nxt;
4723                                 smp_mb();
4724                                 tcp_set_state(sk, TCP_ESTABLISHED);
4725                                 sk->sk_state_change(sk);
4726
4727                                 /* Note, that this wakeup is only for marginal
4728                                  * crossed SYN case. Passively open sockets
4729                                  * are not waked up, because sk->sk_sleep ==
4730                                  * NULL and sk->sk_socket == NULL.
4731                                  */
4732                                 if (sk->sk_socket) {
4733                                         sk_wake_async(sk,0,POLL_OUT);
4734                                 }
4735
4736                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4737                                 tp->snd_wnd = ntohs(th->window) <<
4738                                               tp->rx_opt.snd_wscale;
4739                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4740                                             TCP_SKB_CB(skb)->seq);
4741
4742                                 /* tcp_ack considers this ACK as duplicate
4743                                  * and does not calculate rtt.
4744                                  * Fix it at least with timestamps.
4745                                  */
4746                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4747                                     !tp->srtt)
4748                                         tcp_ack_saw_tstamp(sk, 0);
4749
4750                                 if (tp->rx_opt.tstamp_ok)
4751                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4752
4753                                 /* Make sure socket is routed, for
4754                                  * correct metrics.
4755                                  */
4756                                 icsk->icsk_af_ops->rebuild_header(sk);
4757
4758                                 tcp_init_metrics(sk);
4759
4760                                 tcp_init_congestion_control(sk);
4761
4762                                 /* Prevent spurious tcp_cwnd_restart() on
4763                                  * first data packet.
4764                                  */
4765                                 tp->lsndtime = tcp_time_stamp;
4766
4767                                 tcp_mtup_init(sk);
4768                                 tcp_initialize_rcv_mss(sk);
4769                                 tcp_init_buffer_space(sk);
4770                                 tcp_fast_path_on(tp);
4771                         } else {
4772                                 return 1;
4773                         }
4774                         break;
4775
4776                 case TCP_FIN_WAIT1:
4777                         if (tp->snd_una == tp->write_seq) {
4778                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4779                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4780                                 dst_confirm(sk->sk_dst_cache);
4781
4782                                 if (!sock_flag(sk, SOCK_DEAD))
4783                                         /* Wake up lingering close() */
4784                                         sk->sk_state_change(sk);
4785                                 else {
4786                                         int tmo;
4787
4788                                         if (tp->linger2 < 0 ||
4789                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4790                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4791                                                 tcp_done(sk);
4792                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4793                                                 return 1;
4794                                         }
4795
4796                                         tmo = tcp_fin_time(sk);
4797                                         if (tmo > TCP_TIMEWAIT_LEN) {
4798                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4799                                         } else if (th->fin || sock_owned_by_user(sk)) {
4800                                                 /* Bad case. We could lose such FIN otherwise.
4801                                                  * It is not a big problem, but it looks confusing
4802                                                  * and not so rare event. We still can lose it now,
4803                                                  * if it spins in bh_lock_sock(), but it is really
4804                                                  * marginal case.
4805                                                  */
4806                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4807                                         } else {
4808                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4809                                                 goto discard;
4810                                         }
4811                                 }
4812                         }
4813                         break;
4814
4815                 case TCP_CLOSING:
4816                         if (tp->snd_una == tp->write_seq) {
4817                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4818                                 goto discard;
4819                         }
4820                         break;
4821
4822                 case TCP_LAST_ACK:
4823                         if (tp->snd_una == tp->write_seq) {
4824                                 tcp_update_metrics(sk);
4825                                 tcp_done(sk);
4826                                 goto discard;
4827                         }
4828                         break;
4829                 }
4830         } else
4831                 goto discard;
4832
4833         /* step 6: check the URG bit */
4834         tcp_urg(sk, skb, th);
4835
4836         /* step 7: process the segment text */
4837         switch (sk->sk_state) {
4838         case TCP_CLOSE_WAIT:
4839         case TCP_CLOSING:
4840         case TCP_LAST_ACK:
4841                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4842                         break;
4843         case TCP_FIN_WAIT1:
4844         case TCP_FIN_WAIT2:
4845                 /* RFC 793 says to queue data in these states,
4846                  * RFC 1122 says we MUST send a reset.
4847                  * BSD 4.4 also does reset.
4848                  */
4849                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4850                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4851                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4852                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4853                                 tcp_reset(sk);
4854                                 return 1;
4855                         }
4856                 }
4857                 /* Fall through */
4858         case TCP_ESTABLISHED:
4859                 tcp_data_queue(sk, skb);
4860                 queued = 1;
4861                 break;
4862         }
4863
4864         /* tcp_data could move socket to TIME-WAIT */
4865         if (sk->sk_state != TCP_CLOSE) {
4866                 tcp_data_snd_check(sk, tp);
4867                 tcp_ack_snd_check(sk);
4868         }
4869
4870         if (!queued) {
4871 discard:
4872                 __kfree_skb(skb);
4873         }
4874         return 0;
4875 }
4876
4877 EXPORT_SYMBOL(sysctl_tcp_ecn);
4878 EXPORT_SYMBOL(sysctl_tcp_reordering);
4879 EXPORT_SYMBOL(tcp_parse_options);
4880 EXPORT_SYMBOL(tcp_rcv_established);
4881 EXPORT_SYMBOL(tcp_rcv_state_process);
4882 EXPORT_SYMBOL(tcp_initialize_rcv_mss);