Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux...
[linux-2.6] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
108
109 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
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 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
119
120 /* Adapt the MSS value used to make delayed ack decision to the
121  * real world.
122  */
123 static void tcp_measure_rcv_mss(struct sock *sk,
124                                 const struct sk_buff *skb)
125 {
126         struct inet_connection_sock *icsk = inet_csk(sk);
127         const unsigned int lss = icsk->icsk_ack.last_seg_size;
128         unsigned int len;
129
130         icsk->icsk_ack.last_seg_size = 0;
131
132         /* skb->len may jitter because of SACKs, even if peer
133          * sends good full-sized frames.
134          */
135         len = skb_shinfo(skb)->gso_size ?: skb->len;
136         if (len >= icsk->icsk_ack.rcv_mss) {
137                 icsk->icsk_ack.rcv_mss = len;
138         } else {
139                 /* Otherwise, we make more careful check taking into account,
140                  * that SACKs block is variable.
141                  *
142                  * "len" is invariant segment length, including TCP header.
143                  */
144                 len += skb->data - skb_transport_header(skb);
145                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
146                     /* If PSH is not set, packet should be
147                      * full sized, provided peer TCP is not badly broken.
148                      * This observation (if it is correct 8)) allows
149                      * to handle super-low mtu links fairly.
150                      */
151                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
152                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
153                         /* Subtract also invariant (if peer is RFC compliant),
154                          * tcp header plus fixed timestamp option length.
155                          * Resulting "len" is MSS free of SACK jitter.
156                          */
157                         len -= tcp_sk(sk)->tcp_header_len;
158                         icsk->icsk_ack.last_seg_size = len;
159                         if (len == lss) {
160                                 icsk->icsk_ack.rcv_mss = len;
161                                 return;
162                         }
163                 }
164                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
166                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
167         }
168 }
169
170 static void tcp_incr_quickack(struct sock *sk)
171 {
172         struct inet_connection_sock *icsk = inet_csk(sk);
173         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
174
175         if (quickacks==0)
176                 quickacks=2;
177         if (quickacks > icsk->icsk_ack.quick)
178                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
179 }
180
181 void tcp_enter_quickack_mode(struct sock *sk)
182 {
183         struct inet_connection_sock *icsk = inet_csk(sk);
184         tcp_incr_quickack(sk);
185         icsk->icsk_ack.pingpong = 0;
186         icsk->icsk_ack.ato = TCP_ATO_MIN;
187 }
188
189 /* Send ACKs quickly, if "quick" count is not exhausted
190  * and the session is not interactive.
191  */
192
193 static inline int tcp_in_quickack_mode(const struct sock *sk)
194 {
195         const struct inet_connection_sock *icsk = inet_csk(sk);
196         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
197 }
198
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
200 {
201         if (tp->ecn_flags&TCP_ECN_OK)
202                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
203 }
204
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
206 {
207         if (tcp_hdr(skb)->cwr)
208                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
209 }
210
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
212 {
213         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 }
215
216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
217 {
218         if (tp->ecn_flags&TCP_ECN_OK) {
219                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221                 /* Funny extension: if ECT is not set on a segment,
222                  * it is surely retransmit. It is not in ECN RFC,
223                  * but Linux follows this rule. */
224                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225                         tcp_enter_quickack_mode((struct sock *)tp);
226         }
227 }
228
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
230 {
231         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
232                 tp->ecn_flags &= ~TCP_ECN_OK;
233 }
234
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
236 {
237         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
238                 tp->ecn_flags &= ~TCP_ECN_OK;
239 }
240
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
242 {
243         if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
244                 return 1;
245         return 0;
246 }
247
248 /* Buffer size and advertised window tuning.
249  *
250  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
251  */
252
253 static void tcp_fixup_sndbuf(struct sock *sk)
254 {
255         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256                      sizeof(struct sk_buff);
257
258         if (sk->sk_sndbuf < 3 * sndmem)
259                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
260 }
261
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263  *
264  * All tcp_full_space() is split to two parts: "network" buffer, allocated
265  * forward and advertised in receiver window (tp->rcv_wnd) and
266  * "application buffer", required to isolate scheduling/application
267  * latencies from network.
268  * window_clamp is maximal advertised window. It can be less than
269  * tcp_full_space(), in this case tcp_full_space() - window_clamp
270  * is reserved for "application" buffer. The less window_clamp is
271  * the smoother our behaviour from viewpoint of network, but the lower
272  * throughput and the higher sensitivity of the connection to losses. 8)
273  *
274  * rcv_ssthresh is more strict window_clamp used at "slow start"
275  * phase to predict further behaviour of this connection.
276  * It is used for two goals:
277  * - to enforce header prediction at sender, even when application
278  *   requires some significant "application buffer". It is check #1.
279  * - to prevent pruning of receive queue because of misprediction
280  *   of receiver window. Check #2.
281  *
282  * The scheme does not work when sender sends good segments opening
283  * window and then starts to feed us spaghetti. But it should work
284  * in common situations. Otherwise, we have to rely on queue collapsing.
285  */
286
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
289 {
290         struct tcp_sock *tp = tcp_sk(sk);
291         /* Optimize this! */
292         int truesize = tcp_win_from_space(skb->truesize)/2;
293         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
294
295         while (tp->rcv_ssthresh <= window) {
296                 if (truesize <= skb->len)
297                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
298
299                 truesize >>= 1;
300                 window >>= 1;
301         }
302         return 0;
303 }
304
305 static void tcp_grow_window(struct sock *sk,
306                             struct sk_buff *skb)
307 {
308         struct tcp_sock *tp = tcp_sk(sk);
309
310         /* Check #1 */
311         if (tp->rcv_ssthresh < tp->window_clamp &&
312             (int)tp->rcv_ssthresh < tcp_space(sk) &&
313             !tcp_memory_pressure) {
314                 int incr;
315
316                 /* Check #2. Increase window, if skb with such overhead
317                  * will fit to rcvbuf in future.
318                  */
319                 if (tcp_win_from_space(skb->truesize) <= skb->len)
320                         incr = 2*tp->advmss;
321                 else
322                         incr = __tcp_grow_window(sk, skb);
323
324                 if (incr) {
325                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
326                         inet_csk(sk)->icsk_ack.quick |= 1;
327                 }
328         }
329 }
330
331 /* 3. Tuning rcvbuf, when connection enters established state. */
332
333 static void tcp_fixup_rcvbuf(struct sock *sk)
334 {
335         struct tcp_sock *tp = tcp_sk(sk);
336         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
337
338         /* Try to select rcvbuf so that 4 mss-sized segments
339          * will fit to window and corresponding skbs will fit to our rcvbuf.
340          * (was 3; 4 is minimum to allow fast retransmit to work.)
341          */
342         while (tcp_win_from_space(rcvmem) < tp->advmss)
343                 rcvmem += 128;
344         if (sk->sk_rcvbuf < 4 * rcvmem)
345                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
346 }
347
348 /* 4. Try to fixup all. It is made immediately after connection enters
349  *    established state.
350  */
351 static void tcp_init_buffer_space(struct sock *sk)
352 {
353         struct tcp_sock *tp = tcp_sk(sk);
354         int maxwin;
355
356         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357                 tcp_fixup_rcvbuf(sk);
358         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359                 tcp_fixup_sndbuf(sk);
360
361         tp->rcvq_space.space = tp->rcv_wnd;
362
363         maxwin = tcp_full_space(sk);
364
365         if (tp->window_clamp >= maxwin) {
366                 tp->window_clamp = maxwin;
367
368                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369                         tp->window_clamp = max(maxwin -
370                                                (maxwin >> sysctl_tcp_app_win),
371                                                4 * tp->advmss);
372         }
373
374         /* Force reservation of one segment. */
375         if (sysctl_tcp_app_win &&
376             tp->window_clamp > 2 * tp->advmss &&
377             tp->window_clamp + tp->advmss > maxwin)
378                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
379
380         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381         tp->snd_cwnd_stamp = tcp_time_stamp;
382 }
383
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock *sk)
386 {
387         struct tcp_sock *tp = tcp_sk(sk);
388         struct inet_connection_sock *icsk = inet_csk(sk);
389
390         icsk->icsk_ack.quick = 0;
391
392         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394             !tcp_memory_pressure &&
395             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
397                                     sysctl_tcp_rmem[2]);
398         }
399         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
400                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
401 }
402
403
404 /* Initialize RCV_MSS value.
405  * RCV_MSS is an our guess about MSS used by the peer.
406  * We haven't any direct information about the MSS.
407  * It's better to underestimate the RCV_MSS rather than overestimate.
408  * Overestimations make us ACKing less frequently than needed.
409  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
410  */
411 void tcp_initialize_rcv_mss(struct sock *sk)
412 {
413         struct tcp_sock *tp = tcp_sk(sk);
414         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
415
416         hint = min(hint, tp->rcv_wnd/2);
417         hint = min(hint, TCP_MIN_RCVMSS);
418         hint = max(hint, TCP_MIN_MSS);
419
420         inet_csk(sk)->icsk_ack.rcv_mss = hint;
421 }
422
423 /* Receiver "autotuning" code.
424  *
425  * The algorithm for RTT estimation w/o timestamps is based on
426  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
427  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
428  *
429  * More detail on this code can be found at
430  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
431  * though this reference is out of date.  A new paper
432  * is pending.
433  */
434 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
435 {
436         u32 new_sample = tp->rcv_rtt_est.rtt;
437         long m = sample;
438
439         if (m == 0)
440                 m = 1;
441
442         if (new_sample != 0) {
443                 /* If we sample in larger samples in the non-timestamp
444                  * case, we could grossly overestimate the RTT especially
445                  * with chatty applications or bulk transfer apps which
446                  * are stalled on filesystem I/O.
447                  *
448                  * Also, since we are only going for a minimum in the
449                  * non-timestamp case, we do not smooth things out
450                  * else with timestamps disabled convergence takes too
451                  * long.
452                  */
453                 if (!win_dep) {
454                         m -= (new_sample >> 3);
455                         new_sample += m;
456                 } else if (m < new_sample)
457                         new_sample = m << 3;
458         } else {
459                 /* No previous measure. */
460                 new_sample = m << 3;
461         }
462
463         if (tp->rcv_rtt_est.rtt != new_sample)
464                 tp->rcv_rtt_est.rtt = new_sample;
465 }
466
467 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
468 {
469         if (tp->rcv_rtt_est.time == 0)
470                 goto new_measure;
471         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
472                 return;
473         tcp_rcv_rtt_update(tp,
474                            jiffies - tp->rcv_rtt_est.time,
475                            1);
476
477 new_measure:
478         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
479         tp->rcv_rtt_est.time = tcp_time_stamp;
480 }
481
482 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
483 {
484         struct tcp_sock *tp = tcp_sk(sk);
485         if (tp->rx_opt.rcv_tsecr &&
486             (TCP_SKB_CB(skb)->end_seq -
487              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
488                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
489 }
490
491 /*
492  * This function should be called every time data is copied to user space.
493  * It calculates the appropriate TCP receive buffer space.
494  */
495 void tcp_rcv_space_adjust(struct sock *sk)
496 {
497         struct tcp_sock *tp = tcp_sk(sk);
498         int time;
499         int space;
500
501         if (tp->rcvq_space.time == 0)
502                 goto new_measure;
503
504         time = tcp_time_stamp - tp->rcvq_space.time;
505         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
506             tp->rcv_rtt_est.rtt == 0)
507                 return;
508
509         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
510
511         space = max(tp->rcvq_space.space, space);
512
513         if (tp->rcvq_space.space != space) {
514                 int rcvmem;
515
516                 tp->rcvq_space.space = space;
517
518                 if (sysctl_tcp_moderate_rcvbuf &&
519                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
520                         int new_clamp = space;
521
522                         /* Receive space grows, normalize in order to
523                          * take into account packet headers and sk_buff
524                          * structure overhead.
525                          */
526                         space /= tp->advmss;
527                         if (!space)
528                                 space = 1;
529                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
530                                   16 + sizeof(struct sk_buff));
531                         while (tcp_win_from_space(rcvmem) < tp->advmss)
532                                 rcvmem += 128;
533                         space *= rcvmem;
534                         space = min(space, sysctl_tcp_rmem[2]);
535                         if (space > sk->sk_rcvbuf) {
536                                 sk->sk_rcvbuf = space;
537
538                                 /* Make the window clamp follow along.  */
539                                 tp->window_clamp = new_clamp;
540                         }
541                 }
542         }
543
544 new_measure:
545         tp->rcvq_space.seq = tp->copied_seq;
546         tp->rcvq_space.time = tcp_time_stamp;
547 }
548
549 /* There is something which you must keep in mind when you analyze the
550  * behavior of the tp->ato delayed ack timeout interval.  When a
551  * connection starts up, we want to ack as quickly as possible.  The
552  * problem is that "good" TCP's do slow start at the beginning of data
553  * transmission.  The means that until we send the first few ACK's the
554  * sender will sit on his end and only queue most of his data, because
555  * he can only send snd_cwnd unacked packets at any given time.  For
556  * each ACK we send, he increments snd_cwnd and transmits more of his
557  * queue.  -DaveM
558  */
559 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
560 {
561         struct tcp_sock *tp = tcp_sk(sk);
562         struct inet_connection_sock *icsk = inet_csk(sk);
563         u32 now;
564
565         inet_csk_schedule_ack(sk);
566
567         tcp_measure_rcv_mss(sk, skb);
568
569         tcp_rcv_rtt_measure(tp);
570
571         now = tcp_time_stamp;
572
573         if (!icsk->icsk_ack.ato) {
574                 /* The _first_ data packet received, initialize
575                  * delayed ACK engine.
576                  */
577                 tcp_incr_quickack(sk);
578                 icsk->icsk_ack.ato = TCP_ATO_MIN;
579         } else {
580                 int m = now - icsk->icsk_ack.lrcvtime;
581
582                 if (m <= TCP_ATO_MIN/2) {
583                         /* The fastest case is the first. */
584                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
585                 } else if (m < icsk->icsk_ack.ato) {
586                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
587                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
588                                 icsk->icsk_ack.ato = icsk->icsk_rto;
589                 } else if (m > icsk->icsk_rto) {
590                         /* Too long gap. Apparently sender failed to
591                          * restart window, so that we send ACKs quickly.
592                          */
593                         tcp_incr_quickack(sk);
594                         sk_stream_mem_reclaim(sk);
595                 }
596         }
597         icsk->icsk_ack.lrcvtime = now;
598
599         TCP_ECN_check_ce(tp, skb);
600
601         if (skb->len >= 128)
602                 tcp_grow_window(sk, skb);
603 }
604
605 static u32 tcp_rto_min(struct sock *sk)
606 {
607         struct dst_entry *dst = __sk_dst_get(sk);
608         u32 rto_min = TCP_RTO_MIN;
609
610         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
611                 rto_min = dst->metrics[RTAX_RTO_MIN-1];
612         return rto_min;
613 }
614
615 /* Called to compute a smoothed rtt estimate. The data fed to this
616  * routine either comes from timestamps, or from segments that were
617  * known _not_ to have been retransmitted [see Karn/Partridge
618  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
619  * piece by Van Jacobson.
620  * NOTE: the next three routines used to be one big routine.
621  * To save cycles in the RFC 1323 implementation it was better to break
622  * it up into three procedures. -- erics
623  */
624 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
625 {
626         struct tcp_sock *tp = tcp_sk(sk);
627         long m = mrtt; /* RTT */
628
629         /*      The following amusing code comes from Jacobson's
630          *      article in SIGCOMM '88.  Note that rtt and mdev
631          *      are scaled versions of rtt and mean deviation.
632          *      This is designed to be as fast as possible
633          *      m stands for "measurement".
634          *
635          *      On a 1990 paper the rto value is changed to:
636          *      RTO = rtt + 4 * mdev
637          *
638          * Funny. This algorithm seems to be very broken.
639          * These formulae increase RTO, when it should be decreased, increase
640          * too slowly, when it should be increased quickly, decrease too quickly
641          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
642          * does not matter how to _calculate_ it. Seems, it was trap
643          * that VJ failed to avoid. 8)
644          */
645         if (m == 0)
646                 m = 1;
647         if (tp->srtt != 0) {
648                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
649                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
650                 if (m < 0) {
651                         m = -m;         /* m is now abs(error) */
652                         m -= (tp->mdev >> 2);   /* similar update on mdev */
653                         /* This is similar to one of Eifel findings.
654                          * Eifel blocks mdev updates when rtt decreases.
655                          * This solution is a bit different: we use finer gain
656                          * for mdev in this case (alpha*beta).
657                          * Like Eifel it also prevents growth of rto,
658                          * but also it limits too fast rto decreases,
659                          * happening in pure Eifel.
660                          */
661                         if (m > 0)
662                                 m >>= 3;
663                 } else {
664                         m -= (tp->mdev >> 2);   /* similar update on mdev */
665                 }
666                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
667                 if (tp->mdev > tp->mdev_max) {
668                         tp->mdev_max = tp->mdev;
669                         if (tp->mdev_max > tp->rttvar)
670                                 tp->rttvar = tp->mdev_max;
671                 }
672                 if (after(tp->snd_una, tp->rtt_seq)) {
673                         if (tp->mdev_max < tp->rttvar)
674                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
675                         tp->rtt_seq = tp->snd_nxt;
676                         tp->mdev_max = tcp_rto_min(sk);
677                 }
678         } else {
679                 /* no previous measure. */
680                 tp->srtt = m<<3;        /* take the measured time to be rtt */
681                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
682                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
683                 tp->rtt_seq = tp->snd_nxt;
684         }
685 }
686
687 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
688  * routine referred to above.
689  */
690 static inline void tcp_set_rto(struct sock *sk)
691 {
692         const struct tcp_sock *tp = tcp_sk(sk);
693         /* Old crap is replaced with new one. 8)
694          *
695          * More seriously:
696          * 1. If rtt variance happened to be less 50msec, it is hallucination.
697          *    It cannot be less due to utterly erratic ACK generation made
698          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
699          *    to do with delayed acks, because at cwnd>2 true delack timeout
700          *    is invisible. Actually, Linux-2.4 also generates erratic
701          *    ACKs in some circumstances.
702          */
703         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
704
705         /* 2. Fixups made earlier cannot be right.
706          *    If we do not estimate RTO correctly without them,
707          *    all the algo is pure shit and should be replaced
708          *    with correct one. It is exactly, which we pretend to do.
709          */
710 }
711
712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
713  * guarantees that rto is higher.
714  */
715 static inline void tcp_bound_rto(struct sock *sk)
716 {
717         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
718                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
719 }
720
721 /* Save metrics learned by this TCP session.
722    This function is called only, when TCP finishes successfully
723    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
724  */
725 void tcp_update_metrics(struct sock *sk)
726 {
727         struct tcp_sock *tp = tcp_sk(sk);
728         struct dst_entry *dst = __sk_dst_get(sk);
729
730         if (sysctl_tcp_nometrics_save)
731                 return;
732
733         dst_confirm(dst);
734
735         if (dst && (dst->flags&DST_HOST)) {
736                 const struct inet_connection_sock *icsk = inet_csk(sk);
737                 int m;
738
739                 if (icsk->icsk_backoff || !tp->srtt) {
740                         /* This session failed to estimate rtt. Why?
741                          * Probably, no packets returned in time.
742                          * Reset our results.
743                          */
744                         if (!(dst_metric_locked(dst, RTAX_RTT)))
745                                 dst->metrics[RTAX_RTT-1] = 0;
746                         return;
747                 }
748
749                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
750
751                 /* If newly calculated rtt larger than stored one,
752                  * store new one. Otherwise, use EWMA. Remember,
753                  * rtt overestimation is always better than underestimation.
754                  */
755                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
756                         if (m <= 0)
757                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
758                         else
759                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
760                 }
761
762                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
763                         if (m < 0)
764                                 m = -m;
765
766                         /* Scale deviation to rttvar fixed point */
767                         m >>= 1;
768                         if (m < tp->mdev)
769                                 m = tp->mdev;
770
771                         if (m >= dst_metric(dst, RTAX_RTTVAR))
772                                 dst->metrics[RTAX_RTTVAR-1] = m;
773                         else
774                                 dst->metrics[RTAX_RTTVAR-1] -=
775                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
776                 }
777
778                 if (tp->snd_ssthresh >= 0xFFFF) {
779                         /* Slow start still did not finish. */
780                         if (dst_metric(dst, RTAX_SSTHRESH) &&
781                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
782                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
783                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
784                         if (!dst_metric_locked(dst, RTAX_CWND) &&
785                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
786                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
787                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
788                            icsk->icsk_ca_state == TCP_CA_Open) {
789                         /* Cong. avoidance phase, cwnd is reliable. */
790                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
791                                 dst->metrics[RTAX_SSTHRESH-1] =
792                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
793                         if (!dst_metric_locked(dst, RTAX_CWND))
794                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
795                 } else {
796                         /* Else slow start did not finish, cwnd is non-sense,
797                            ssthresh may be also invalid.
798                          */
799                         if (!dst_metric_locked(dst, RTAX_CWND))
800                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
801                         if (dst->metrics[RTAX_SSTHRESH-1] &&
802                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
803                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
804                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
805                 }
806
807                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
808                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
809                             tp->reordering != sysctl_tcp_reordering)
810                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
811                 }
812         }
813 }
814
815 /* Numbers are taken from RFC3390.
816  *
817  * John Heffner states:
818  *
819  *      The RFC specifies a window of no more than 4380 bytes
820  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
821  *      is a bit misleading because they use a clamp at 4380 bytes
822  *      rather than use a multiplier in the relevant range.
823  */
824 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
825 {
826         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
827
828         if (!cwnd) {
829                 if (tp->mss_cache > 1460)
830                         cwnd = 2;
831                 else
832                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
833         }
834         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
835 }
836
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
839 {
840         struct tcp_sock *tp = tcp_sk(sk);
841         const struct inet_connection_sock *icsk = inet_csk(sk);
842
843         tp->prior_ssthresh = 0;
844         tp->bytes_acked = 0;
845         if (icsk->icsk_ca_state < TCP_CA_CWR) {
846                 tp->undo_marker = 0;
847                 if (set_ssthresh)
848                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
849                 tp->snd_cwnd = min(tp->snd_cwnd,
850                                    tcp_packets_in_flight(tp) + 1U);
851                 tp->snd_cwnd_cnt = 0;
852                 tp->high_seq = tp->snd_nxt;
853                 tp->snd_cwnd_stamp = tcp_time_stamp;
854                 TCP_ECN_queue_cwr(tp);
855
856                 tcp_set_ca_state(sk, TCP_CA_CWR);
857         }
858 }
859
860 /*
861  * Packet counting of FACK is based on in-order assumptions, therefore TCP
862  * disables it when reordering is detected
863  */
864 static void tcp_disable_fack(struct tcp_sock *tp)
865 {
866         tp->rx_opt.sack_ok &= ~2;
867 }
868
869 /* Take a notice that peer is sending D-SACKs */
870 static void tcp_dsack_seen(struct tcp_sock *tp)
871 {
872         tp->rx_opt.sack_ok |= 4;
873 }
874
875 /* Initialize metrics on socket. */
876
877 static void tcp_init_metrics(struct sock *sk)
878 {
879         struct tcp_sock *tp = tcp_sk(sk);
880         struct dst_entry *dst = __sk_dst_get(sk);
881
882         if (dst == NULL)
883                 goto reset;
884
885         dst_confirm(dst);
886
887         if (dst_metric_locked(dst, RTAX_CWND))
888                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889         if (dst_metric(dst, RTAX_SSTHRESH)) {
890                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
893         }
894         if (dst_metric(dst, RTAX_REORDERING) &&
895             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
896                 tcp_disable_fack(tp);
897                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
898         }
899
900         if (dst_metric(dst, RTAX_RTT) == 0)
901                 goto reset;
902
903         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
904                 goto reset;
905
906         /* Initial rtt is determined from SYN,SYN-ACK.
907          * The segment is small and rtt may appear much
908          * less than real one. Use per-dst memory
909          * to make it more realistic.
910          *
911          * A bit of theory. RTT is time passed after "normal" sized packet
912          * is sent until it is ACKed. In normal circumstances sending small
913          * packets force peer to delay ACKs and calculation is correct too.
914          * The algorithm is adaptive and, provided we follow specs, it
915          * NEVER underestimate RTT. BUT! If peer tries to make some clever
916          * tricks sort of "quick acks" for time long enough to decrease RTT
917          * to low value, and then abruptly stops to do it and starts to delay
918          * ACKs, wait for troubles.
919          */
920         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
921                 tp->srtt = dst_metric(dst, RTAX_RTT);
922                 tp->rtt_seq = tp->snd_nxt;
923         }
924         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
925                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
926                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
927         }
928         tcp_set_rto(sk);
929         tcp_bound_rto(sk);
930         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
931                 goto reset;
932         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
933         tp->snd_cwnd_stamp = tcp_time_stamp;
934         return;
935
936 reset:
937         /* Play conservative. If timestamps are not
938          * supported, TCP will fail to recalculate correct
939          * rtt, if initial rto is too small. FORGET ALL AND RESET!
940          */
941         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
942                 tp->srtt = 0;
943                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
944                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
945         }
946 }
947
948 static void tcp_update_reordering(struct sock *sk, const int metric,
949                                   const int ts)
950 {
951         struct tcp_sock *tp = tcp_sk(sk);
952         if (metric > tp->reordering) {
953                 tp->reordering = min(TCP_MAX_REORDERING, metric);
954
955                 /* This exciting event is worth to be remembered. 8) */
956                 if (ts)
957                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
958                 else if (tcp_is_reno(tp))
959                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
960                 else if (tcp_is_fack(tp))
961                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
962                 else
963                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
964 #if FASTRETRANS_DEBUG > 1
965                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
966                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
967                        tp->reordering,
968                        tp->fackets_out,
969                        tp->sacked_out,
970                        tp->undo_marker ? tp->undo_retrans : 0);
971 #endif
972                 tcp_disable_fack(tp);
973         }
974 }
975
976 /* This procedure tags the retransmission queue when SACKs arrive.
977  *
978  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
979  * Packets in queue with these bits set are counted in variables
980  * sacked_out, retrans_out and lost_out, correspondingly.
981  *
982  * Valid combinations are:
983  * Tag  InFlight        Description
984  * 0    1               - orig segment is in flight.
985  * S    0               - nothing flies, orig reached receiver.
986  * L    0               - nothing flies, orig lost by net.
987  * R    2               - both orig and retransmit are in flight.
988  * L|R  1               - orig is lost, retransmit is in flight.
989  * S|R  1               - orig reached receiver, retrans is still in flight.
990  * (L|S|R is logically valid, it could occur when L|R is sacked,
991  *  but it is equivalent to plain S and code short-curcuits it to S.
992  *  L|S is logically invalid, it would mean -1 packet in flight 8))
993  *
994  * These 6 states form finite state machine, controlled by the following events:
995  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
996  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
997  * 3. Loss detection event of one of three flavors:
998  *      A. Scoreboard estimator decided the packet is lost.
999  *         A'. Reno "three dupacks" marks head of queue lost.
1000  *         A''. Its FACK modfication, head until snd.fack is lost.
1001  *      B. SACK arrives sacking data transmitted after never retransmitted
1002  *         hole was sent out.
1003  *      C. SACK arrives sacking SND.NXT at the moment, when the
1004  *         segment was retransmitted.
1005  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1006  *
1007  * It is pleasant to note, that state diagram turns out to be commutative,
1008  * so that we are allowed not to be bothered by order of our actions,
1009  * when multiple events arrive simultaneously. (see the function below).
1010  *
1011  * Reordering detection.
1012  * --------------------
1013  * Reordering metric is maximal distance, which a packet can be displaced
1014  * in packet stream. With SACKs we can estimate it:
1015  *
1016  * 1. SACK fills old hole and the corresponding segment was not
1017  *    ever retransmitted -> reordering. Alas, we cannot use it
1018  *    when segment was retransmitted.
1019  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1020  *    for retransmitted and already SACKed segment -> reordering..
1021  * Both of these heuristics are not used in Loss state, when we cannot
1022  * account for retransmits accurately.
1023  *
1024  * SACK block validation.
1025  * ----------------------
1026  *
1027  * SACK block range validation checks that the received SACK block fits to
1028  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1029  * Note that SND.UNA is not included to the range though being valid because
1030  * it means that the receiver is rather inconsistent with itself reporting
1031  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1032  * perfectly valid, however, in light of RFC2018 which explicitly states
1033  * that "SACK block MUST reflect the newest segment.  Even if the newest
1034  * segment is going to be discarded ...", not that it looks very clever
1035  * in case of head skb. Due to potentional receiver driven attacks, we
1036  * choose to avoid immediate execution of a walk in write queue due to
1037  * reneging and defer head skb's loss recovery to standard loss recovery
1038  * procedure that will eventually trigger (nothing forbids us doing this).
1039  *
1040  * Implements also blockage to start_seq wrap-around. Problem lies in the
1041  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1042  * there's no guarantee that it will be before snd_nxt (n). The problem
1043  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1044  * wrap (s_w):
1045  *
1046  *         <- outs wnd ->                          <- wrapzone ->
1047  *         u     e      n                         u_w   e_w  s n_w
1048  *         |     |      |                          |     |   |  |
1049  * |<------------+------+----- TCP seqno space --------------+---------->|
1050  * ...-- <2^31 ->|                                           |<--------...
1051  * ...---- >2^31 ------>|                                    |<--------...
1052  *
1053  * Current code wouldn't be vulnerable but it's better still to discard such
1054  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1055  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1056  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1057  * equal to the ideal case (infinite seqno space without wrap caused issues).
1058  *
1059  * With D-SACK the lower bound is extended to cover sequence space below
1060  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1061  * again, D-SACK block must not to go across snd_una (for the same reason as
1062  * for the normal SACK blocks, explained above). But there all simplicity
1063  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1064  * fully below undo_marker they do not affect behavior in anyway and can
1065  * therefore be safely ignored. In rare cases (which are more or less
1066  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1067  * fragmentation and packet reordering past skb's retransmission. To consider
1068  * them correctly, the acceptable range must be extended even more though
1069  * the exact amount is rather hard to quantify. However, tp->max_window can
1070  * be used as an exaggerated estimate.
1071  */
1072 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1073                                   u32 start_seq, u32 end_seq)
1074 {
1075         /* Too far in future, or reversed (interpretation is ambiguous) */
1076         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1077                 return 0;
1078
1079         /* Nasty start_seq wrap-around check (see comments above) */
1080         if (!before(start_seq, tp->snd_nxt))
1081                 return 0;
1082
1083         /* In outstanding window? ...This is valid exit for D-SACKs too.
1084          * start_seq == snd_una is non-sensical (see comments above)
1085          */
1086         if (after(start_seq, tp->snd_una))
1087                 return 1;
1088
1089         if (!is_dsack || !tp->undo_marker)
1090                 return 0;
1091
1092         /* ...Then it's D-SACK, and must reside below snd_una completely */
1093         if (!after(end_seq, tp->snd_una))
1094                 return 0;
1095
1096         if (!before(start_seq, tp->undo_marker))
1097                 return 1;
1098
1099         /* Too old */
1100         if (!after(end_seq, tp->undo_marker))
1101                 return 0;
1102
1103         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1104          *   start_seq < undo_marker and end_seq >= undo_marker.
1105          */
1106         return !before(start_seq, end_seq - tp->max_window);
1107 }
1108
1109 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1110  * Event "C". Later note: FACK people cheated me again 8), we have to account
1111  * for reordering! Ugly, but should help.
1112  *
1113  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1114  * less than what is now known to be received by the other end (derived from
1115  * SACK blocks by the caller). Also calculate the lowest snd_nxt among the
1116  * remaining retransmitted skbs to avoid some costly processing per ACKs.
1117  */
1118 static int tcp_mark_lost_retrans(struct sock *sk, u32 received_upto)
1119 {
1120         struct tcp_sock *tp = tcp_sk(sk);
1121         struct sk_buff *skb;
1122         int flag = 0;
1123         int cnt = 0;
1124         u32 new_low_seq = tp->snd_nxt;
1125
1126         tcp_for_write_queue(skb, sk) {
1127                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1128
1129                 if (skb == tcp_send_head(sk))
1130                         break;
1131                 if (cnt == tp->retrans_out)
1132                         break;
1133                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1134                         continue;
1135
1136                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1137                         continue;
1138
1139                 if (after(received_upto, ack_seq) &&
1140                     (tcp_is_fack(tp) ||
1141                      !before(received_upto,
1142                              ack_seq + tp->reordering * tp->mss_cache))) {
1143                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1144                         tp->retrans_out -= tcp_skb_pcount(skb);
1145
1146                         /* clear lost hint */
1147                         tp->retransmit_skb_hint = NULL;
1148
1149                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1150                                 tp->lost_out += tcp_skb_pcount(skb);
1151                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1152                                 flag |= FLAG_DATA_SACKED;
1153                                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1154                         }
1155                 } else {
1156                         if (before(ack_seq, new_low_seq))
1157                                 new_low_seq = ack_seq;
1158                         cnt += tcp_skb_pcount(skb);
1159                 }
1160         }
1161
1162         if (tp->retrans_out)
1163                 tp->lost_retrans_low = new_low_seq;
1164
1165         return flag;
1166 }
1167
1168 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1169                            struct tcp_sack_block_wire *sp, int num_sacks,
1170                            u32 prior_snd_una)
1171 {
1172         u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1173         u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1174         int dup_sack = 0;
1175
1176         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1177                 dup_sack = 1;
1178                 tcp_dsack_seen(tp);
1179                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1180         } else if (num_sacks > 1) {
1181                 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1182                 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1183
1184                 if (!after(end_seq_0, end_seq_1) &&
1185                     !before(start_seq_0, start_seq_1)) {
1186                         dup_sack = 1;
1187                         tcp_dsack_seen(tp);
1188                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1189                 }
1190         }
1191
1192         /* D-SACK for already forgotten data... Do dumb counting. */
1193         if (dup_sack &&
1194             !after(end_seq_0, prior_snd_una) &&
1195             after(end_seq_0, tp->undo_marker))
1196                 tp->undo_retrans--;
1197
1198         return dup_sack;
1199 }
1200
1201 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1202  * the incoming SACK may not exactly match but we can find smaller MSS
1203  * aligned portion of it that matches. Therefore we might need to fragment
1204  * which may fail and creates some hassle (caller must handle error case
1205  * returns).
1206  */
1207 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1208                                  u32 start_seq, u32 end_seq)
1209 {
1210         int in_sack, err;
1211         unsigned int pkt_len;
1212
1213         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1214                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1215
1216         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1217             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1218
1219                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1220
1221                 if (!in_sack)
1222                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1223                 else
1224                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1225                 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1226                 if (err < 0)
1227                         return err;
1228         }
1229
1230         return in_sack;
1231 }
1232
1233 static int
1234 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1235 {
1236         const struct inet_connection_sock *icsk = inet_csk(sk);
1237         struct tcp_sock *tp = tcp_sk(sk);
1238         unsigned char *ptr = (skb_transport_header(ack_skb) +
1239                               TCP_SKB_CB(ack_skb)->sacked);
1240         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1241         struct sk_buff *cached_skb;
1242         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1243         int reord = tp->packets_out;
1244         int prior_fackets;
1245         u32 highest_sack_end_seq = tp->lost_retrans_low;
1246         int flag = 0;
1247         int found_dup_sack = 0;
1248         int cached_fack_count;
1249         int i;
1250         int first_sack_index;
1251         int force_one_sack;
1252
1253         if (!tp->sacked_out) {
1254                 if (WARN_ON(tp->fackets_out))
1255                         tp->fackets_out = 0;
1256                 tp->highest_sack = tp->snd_una;
1257         }
1258         prior_fackets = tp->fackets_out;
1259
1260         found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1261                                          num_sacks, prior_snd_una);
1262         if (found_dup_sack)
1263                 flag |= FLAG_DSACKING_ACK;
1264
1265         /* Eliminate too old ACKs, but take into
1266          * account more or less fresh ones, they can
1267          * contain valid SACK info.
1268          */
1269         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1270                 return 0;
1271
1272         if (!tp->packets_out)
1273                 goto out;
1274
1275         /* SACK fastpath:
1276          * if the only SACK change is the increase of the end_seq of
1277          * the first block then only apply that SACK block
1278          * and use retrans queue hinting otherwise slowpath */
1279         force_one_sack = 1;
1280         for (i = 0; i < num_sacks; i++) {
1281                 __be32 start_seq = sp[i].start_seq;
1282                 __be32 end_seq = sp[i].end_seq;
1283
1284                 if (i == 0) {
1285                         if (tp->recv_sack_cache[i].start_seq != start_seq)
1286                                 force_one_sack = 0;
1287                 } else {
1288                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1289                             (tp->recv_sack_cache[i].end_seq != end_seq))
1290                                 force_one_sack = 0;
1291                 }
1292                 tp->recv_sack_cache[i].start_seq = start_seq;
1293                 tp->recv_sack_cache[i].end_seq = end_seq;
1294         }
1295         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1296         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1297                 tp->recv_sack_cache[i].start_seq = 0;
1298                 tp->recv_sack_cache[i].end_seq = 0;
1299         }
1300
1301         first_sack_index = 0;
1302         if (force_one_sack)
1303                 num_sacks = 1;
1304         else {
1305                 int j;
1306                 tp->fastpath_skb_hint = NULL;
1307
1308                 /* order SACK blocks to allow in order walk of the retrans queue */
1309                 for (i = num_sacks-1; i > 0; i--) {
1310                         for (j = 0; j < i; j++){
1311                                 if (after(ntohl(sp[j].start_seq),
1312                                           ntohl(sp[j+1].start_seq))){
1313                                         struct tcp_sack_block_wire tmp;
1314
1315                                         tmp = sp[j];
1316                                         sp[j] = sp[j+1];
1317                                         sp[j+1] = tmp;
1318
1319                                         /* Track where the first SACK block goes to */
1320                                         if (j == first_sack_index)
1321                                                 first_sack_index = j+1;
1322                                 }
1323
1324                         }
1325                 }
1326         }
1327
1328         /* Use SACK fastpath hint if valid */
1329         cached_skb = tp->fastpath_skb_hint;
1330         cached_fack_count = tp->fastpath_cnt_hint;
1331         if (!cached_skb) {
1332                 cached_skb = tcp_write_queue_head(sk);
1333                 cached_fack_count = 0;
1334         }
1335
1336         for (i = 0; i < num_sacks; i++) {
1337                 struct sk_buff *skb;
1338                 __u32 start_seq = ntohl(sp->start_seq);
1339                 __u32 end_seq = ntohl(sp->end_seq);
1340                 int fack_count;
1341                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1342                 int next_dup = (found_dup_sack && (i+1 == first_sack_index));
1343
1344                 sp++;
1345
1346                 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
1347                         if (dup_sack) {
1348                                 if (!tp->undo_marker)
1349                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1350                                 else
1351                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1352                         } else {
1353                                 /* Don't count olds caused by ACK reordering */
1354                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1355                                     !after(end_seq, tp->snd_una))
1356                                         continue;
1357                                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1358                         }
1359                         continue;
1360                 }
1361
1362                 skb = cached_skb;
1363                 fack_count = cached_fack_count;
1364
1365                 /* Event "B" in the comment above. */
1366                 if (after(end_seq, tp->high_seq))
1367                         flag |= FLAG_DATA_LOST;
1368
1369                 tcp_for_write_queue_from(skb, sk) {
1370                         int in_sack = 0;
1371                         u8 sacked;
1372
1373                         if (skb == tcp_send_head(sk))
1374                                 break;
1375
1376                         cached_skb = skb;
1377                         cached_fack_count = fack_count;
1378                         if (i == first_sack_index) {
1379                                 tp->fastpath_skb_hint = skb;
1380                                 tp->fastpath_cnt_hint = fack_count;
1381                         }
1382
1383                         /* The retransmission queue is always in order, so
1384                          * we can short-circuit the walk early.
1385                          */
1386                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1387                                 break;
1388
1389                         dup_sack = (found_dup_sack && (i == first_sack_index));
1390
1391                         /* Due to sorting DSACK may reside within this SACK block! */
1392                         if (next_dup) {
1393                                 u32 dup_start = ntohl(sp->start_seq);
1394                                 u32 dup_end = ntohl(sp->end_seq);
1395
1396                                 if (before(TCP_SKB_CB(skb)->seq, dup_end)) {
1397                                         in_sack = tcp_match_skb_to_sack(sk, skb, dup_start, dup_end);
1398                                         if (in_sack > 0)
1399                                                 dup_sack = 1;
1400                                 }
1401                         }
1402
1403                         /* DSACK info lost if out-of-mem, try SACK still */
1404                         if (in_sack <= 0)
1405                                 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq);
1406                         if (unlikely(in_sack < 0))
1407                                 break;
1408
1409                         sacked = TCP_SKB_CB(skb)->sacked;
1410
1411                         /* Account D-SACK for retransmitted packet. */
1412                         if ((dup_sack && in_sack) &&
1413                             (sacked & TCPCB_RETRANS) &&
1414                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1415                                 tp->undo_retrans--;
1416
1417                         /* The frame is ACKed. */
1418                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1419                                 if (sacked&TCPCB_RETRANS) {
1420                                         if ((dup_sack && in_sack) &&
1421                                             (sacked&TCPCB_SACKED_ACKED))
1422                                                 reord = min(fack_count, reord);
1423                                 }
1424
1425                                 /* Nothing to do; acked frame is about to be dropped. */
1426                                 fack_count += tcp_skb_pcount(skb);
1427                                 continue;
1428                         }
1429
1430                         if (!in_sack) {
1431                                 fack_count += tcp_skb_pcount(skb);
1432                                 continue;
1433                         }
1434
1435                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1436                                 if (sacked & TCPCB_SACKED_RETRANS) {
1437                                         /* If the segment is not tagged as lost,
1438                                          * we do not clear RETRANS, believing
1439                                          * that retransmission is still in flight.
1440                                          */
1441                                         if (sacked & TCPCB_LOST) {
1442                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1443                                                 tp->lost_out -= tcp_skb_pcount(skb);
1444                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1445
1446                                                 /* clear lost hint */
1447                                                 tp->retransmit_skb_hint = NULL;
1448                                         }
1449                                 } else {
1450                                         if (!(sacked & TCPCB_RETRANS)) {
1451                                                 /* New sack for not retransmitted frame,
1452                                                  * which was in hole. It is reordering.
1453                                                  */
1454                                                 if (fack_count < prior_fackets)
1455                                                         reord = min(fack_count, reord);
1456
1457                                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1458                                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1459                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1460                                         }
1461
1462                                         if (sacked & TCPCB_LOST) {
1463                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1464                                                 tp->lost_out -= tcp_skb_pcount(skb);
1465
1466                                                 /* clear lost hint */
1467                                                 tp->retransmit_skb_hint = NULL;
1468                                         }
1469                                 }
1470
1471                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1472                                 flag |= FLAG_DATA_SACKED;
1473                                 tp->sacked_out += tcp_skb_pcount(skb);
1474
1475                                 fack_count += tcp_skb_pcount(skb);
1476                                 if (fack_count > tp->fackets_out)
1477                                         tp->fackets_out = fack_count;
1478
1479                                 if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) {
1480                                         tp->highest_sack = TCP_SKB_CB(skb)->seq;
1481                                         highest_sack_end_seq = TCP_SKB_CB(skb)->end_seq;
1482                                 }
1483                         } else {
1484                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1485                                         reord = min(fack_count, reord);
1486
1487                                 fack_count += tcp_skb_pcount(skb);
1488                         }
1489
1490                         /* D-SACK. We can detect redundant retransmission
1491                          * in S|R and plain R frames and clear it.
1492                          * undo_retrans is decreased above, L|R frames
1493                          * are accounted above as well.
1494                          */
1495                         if (dup_sack &&
1496                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1497                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1498                                 tp->retrans_out -= tcp_skb_pcount(skb);
1499                                 tp->retransmit_skb_hint = NULL;
1500                         }
1501                 }
1502
1503                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1504                  * due to in-order walk
1505                  */
1506                 if (after(end_seq, tp->frto_highmark))
1507                         flag &= ~FLAG_ONLY_ORIG_SACKED;
1508         }
1509
1510         if (tp->retrans_out &&
1511             after(highest_sack_end_seq, tp->lost_retrans_low) &&
1512             icsk->icsk_ca_state == TCP_CA_Recovery)
1513                 flag |= tcp_mark_lost_retrans(sk, highest_sack_end_seq);
1514
1515         tcp_verify_left_out(tp);
1516
1517         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1518             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1519                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1520
1521 out:
1522
1523 #if FASTRETRANS_DEBUG > 0
1524         BUG_TRAP((int)tp->sacked_out >= 0);
1525         BUG_TRAP((int)tp->lost_out >= 0);
1526         BUG_TRAP((int)tp->retrans_out >= 0);
1527         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1528 #endif
1529         return flag;
1530 }
1531
1532 /* If we receive more dupacks than we expected counting segments
1533  * in assumption of absent reordering, interpret this as reordering.
1534  * The only another reason could be bug in receiver TCP.
1535  */
1536 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1537 {
1538         struct tcp_sock *tp = tcp_sk(sk);
1539         u32 holes;
1540
1541         holes = max(tp->lost_out, 1U);
1542         holes = min(holes, tp->packets_out);
1543
1544         if ((tp->sacked_out + holes) > tp->packets_out) {
1545                 tp->sacked_out = tp->packets_out - holes;
1546                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1547         }
1548 }
1549
1550 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1551
1552 static void tcp_add_reno_sack(struct sock *sk)
1553 {
1554         struct tcp_sock *tp = tcp_sk(sk);
1555         tp->sacked_out++;
1556         tcp_check_reno_reordering(sk, 0);
1557         tcp_verify_left_out(tp);
1558 }
1559
1560 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1561
1562 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1563 {
1564         struct tcp_sock *tp = tcp_sk(sk);
1565
1566         if (acked > 0) {
1567                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1568                 if (acked-1 >= tp->sacked_out)
1569                         tp->sacked_out = 0;
1570                 else
1571                         tp->sacked_out -= acked-1;
1572         }
1573         tcp_check_reno_reordering(sk, acked);
1574         tcp_verify_left_out(tp);
1575 }
1576
1577 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1578 {
1579         tp->sacked_out = 0;
1580 }
1581
1582 /* F-RTO can only be used if TCP has never retransmitted anything other than
1583  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1584  */
1585 int tcp_use_frto(struct sock *sk)
1586 {
1587         const struct tcp_sock *tp = tcp_sk(sk);
1588         struct sk_buff *skb;
1589
1590         if (!sysctl_tcp_frto)
1591                 return 0;
1592
1593         if (IsSackFrto())
1594                 return 1;
1595
1596         /* Avoid expensive walking of rexmit queue if possible */
1597         if (tp->retrans_out > 1)
1598                 return 0;
1599
1600         skb = tcp_write_queue_head(sk);
1601         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1602         tcp_for_write_queue_from(skb, sk) {
1603                 if (skb == tcp_send_head(sk))
1604                         break;
1605                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1606                         return 0;
1607                 /* Short-circuit when first non-SACKed skb has been checked */
1608                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1609                         break;
1610         }
1611         return 1;
1612 }
1613
1614 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1615  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1616  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1617  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1618  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1619  * bits are handled if the Loss state is really to be entered (in
1620  * tcp_enter_frto_loss).
1621  *
1622  * Do like tcp_enter_loss() would; when RTO expires the second time it
1623  * does:
1624  *  "Reduce ssthresh if it has not yet been made inside this window."
1625  */
1626 void tcp_enter_frto(struct sock *sk)
1627 {
1628         const struct inet_connection_sock *icsk = inet_csk(sk);
1629         struct tcp_sock *tp = tcp_sk(sk);
1630         struct sk_buff *skb;
1631
1632         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1633             tp->snd_una == tp->high_seq ||
1634             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1635              !icsk->icsk_retransmits)) {
1636                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1637                 /* Our state is too optimistic in ssthresh() call because cwnd
1638                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1639                  * recovery has not yet completed. Pattern would be this: RTO,
1640                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1641                  * up here twice).
1642                  * RFC4138 should be more specific on what to do, even though
1643                  * RTO is quite unlikely to occur after the first Cumulative ACK
1644                  * due to back-off and complexity of triggering events ...
1645                  */
1646                 if (tp->frto_counter) {
1647                         u32 stored_cwnd;
1648                         stored_cwnd = tp->snd_cwnd;
1649                         tp->snd_cwnd = 2;
1650                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1651                         tp->snd_cwnd = stored_cwnd;
1652                 } else {
1653                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1654                 }
1655                 /* ... in theory, cong.control module could do "any tricks" in
1656                  * ssthresh(), which means that ca_state, lost bits and lost_out
1657                  * counter would have to be faked before the call occurs. We
1658                  * consider that too expensive, unlikely and hacky, so modules
1659                  * using these in ssthresh() must deal these incompatibility
1660                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1661                  */
1662                 tcp_ca_event(sk, CA_EVENT_FRTO);
1663         }
1664
1665         tp->undo_marker = tp->snd_una;
1666         tp->undo_retrans = 0;
1667
1668         skb = tcp_write_queue_head(sk);
1669         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1670                 tp->undo_marker = 0;
1671         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1672                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1673                 tp->retrans_out -= tcp_skb_pcount(skb);
1674         }
1675         tcp_verify_left_out(tp);
1676
1677         /* Too bad if TCP was application limited */
1678         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1679
1680         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1681          * The last condition is necessary at least in tp->frto_counter case.
1682          */
1683         if (IsSackFrto() && (tp->frto_counter ||
1684             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1685             after(tp->high_seq, tp->snd_una)) {
1686                 tp->frto_highmark = tp->high_seq;
1687         } else {
1688                 tp->frto_highmark = tp->snd_nxt;
1689         }
1690         tcp_set_ca_state(sk, TCP_CA_Disorder);
1691         tp->high_seq = tp->snd_nxt;
1692         tp->frto_counter = 1;
1693 }
1694
1695 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1696  * which indicates that we should follow the traditional RTO recovery,
1697  * i.e. mark everything lost and do go-back-N retransmission.
1698  */
1699 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1700 {
1701         struct tcp_sock *tp = tcp_sk(sk);
1702         struct sk_buff *skb;
1703
1704         tp->lost_out = 0;
1705         tp->retrans_out = 0;
1706         if (tcp_is_reno(tp))
1707                 tcp_reset_reno_sack(tp);
1708
1709         tcp_for_write_queue(skb, sk) {
1710                 if (skb == tcp_send_head(sk))
1711                         break;
1712
1713                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1714                 /*
1715                  * Count the retransmission made on RTO correctly (only when
1716                  * waiting for the first ACK and did not get it)...
1717                  */
1718                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1719                         /* For some reason this R-bit might get cleared? */
1720                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1721                                 tp->retrans_out += tcp_skb_pcount(skb);
1722                         /* ...enter this if branch just for the first segment */
1723                         flag |= FLAG_DATA_ACKED;
1724                 } else {
1725                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1726                                 tp->undo_marker = 0;
1727                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1728                 }
1729
1730                 /* Don't lost mark skbs that were fwd transmitted after RTO */
1731                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1732                     !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1733                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1734                         tp->lost_out += tcp_skb_pcount(skb);
1735                 }
1736         }
1737         tcp_verify_left_out(tp);
1738
1739         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1740         tp->snd_cwnd_cnt = 0;
1741         tp->snd_cwnd_stamp = tcp_time_stamp;
1742         tp->frto_counter = 0;
1743         tp->bytes_acked = 0;
1744
1745         tp->reordering = min_t(unsigned int, tp->reordering,
1746                                              sysctl_tcp_reordering);
1747         tcp_set_ca_state(sk, TCP_CA_Loss);
1748         tp->high_seq = tp->frto_highmark;
1749         TCP_ECN_queue_cwr(tp);
1750
1751         tcp_clear_retrans_hints_partial(tp);
1752 }
1753
1754 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1755 {
1756         tp->retrans_out = 0;
1757         tp->lost_out = 0;
1758
1759         tp->undo_marker = 0;
1760         tp->undo_retrans = 0;
1761 }
1762
1763 void tcp_clear_retrans(struct tcp_sock *tp)
1764 {
1765         tcp_clear_retrans_partial(tp);
1766
1767         tp->fackets_out = 0;
1768         tp->sacked_out = 0;
1769 }
1770
1771 /* Enter Loss state. If "how" is not zero, forget all SACK information
1772  * and reset tags completely, otherwise preserve SACKs. If receiver
1773  * dropped its ofo queue, we will know this due to reneging detection.
1774  */
1775 void tcp_enter_loss(struct sock *sk, int how)
1776 {
1777         const struct inet_connection_sock *icsk = inet_csk(sk);
1778         struct tcp_sock *tp = tcp_sk(sk);
1779         struct sk_buff *skb;
1780
1781         /* Reduce ssthresh if it has not yet been made inside this window. */
1782         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1783             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1784                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1785                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1786                 tcp_ca_event(sk, CA_EVENT_LOSS);
1787         }
1788         tp->snd_cwnd       = 1;
1789         tp->snd_cwnd_cnt   = 0;
1790         tp->snd_cwnd_stamp = tcp_time_stamp;
1791
1792         tp->bytes_acked = 0;
1793         tcp_clear_retrans_partial(tp);
1794
1795         if (tcp_is_reno(tp))
1796                 tcp_reset_reno_sack(tp);
1797
1798         if (!how) {
1799                 /* Push undo marker, if it was plain RTO and nothing
1800                  * was retransmitted. */
1801                 tp->undo_marker = tp->snd_una;
1802                 tcp_clear_retrans_hints_partial(tp);
1803         } else {
1804                 tp->sacked_out = 0;
1805                 tp->fackets_out = 0;
1806                 tcp_clear_all_retrans_hints(tp);
1807         }
1808
1809         tcp_for_write_queue(skb, sk) {
1810                 if (skb == tcp_send_head(sk))
1811                         break;
1812
1813                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1814                         tp->undo_marker = 0;
1815                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1816                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1817                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1818                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1819                         tp->lost_out += tcp_skb_pcount(skb);
1820                 }
1821         }
1822         tcp_verify_left_out(tp);
1823
1824         tp->reordering = min_t(unsigned int, tp->reordering,
1825                                              sysctl_tcp_reordering);
1826         tcp_set_ca_state(sk, TCP_CA_Loss);
1827         tp->high_seq = tp->snd_nxt;
1828         TCP_ECN_queue_cwr(tp);
1829         /* Abort F-RTO algorithm if one is in progress */
1830         tp->frto_counter = 0;
1831 }
1832
1833 static int tcp_check_sack_reneging(struct sock *sk)
1834 {
1835         struct sk_buff *skb;
1836
1837         /* If ACK arrived pointing to a remembered SACK,
1838          * it means that our remembered SACKs do not reflect
1839          * real state of receiver i.e.
1840          * receiver _host_ is heavily congested (or buggy).
1841          * Do processing similar to RTO timeout.
1842          */
1843         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1844             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1845                 struct inet_connection_sock *icsk = inet_csk(sk);
1846                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1847
1848                 tcp_enter_loss(sk, 1);
1849                 icsk->icsk_retransmits++;
1850                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1851                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1852                                           icsk->icsk_rto, TCP_RTO_MAX);
1853                 return 1;
1854         }
1855         return 0;
1856 }
1857
1858 static inline int tcp_fackets_out(struct tcp_sock *tp)
1859 {
1860         return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1861 }
1862
1863 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1864 {
1865         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1866 }
1867
1868 static inline int tcp_head_timedout(struct sock *sk)
1869 {
1870         struct tcp_sock *tp = tcp_sk(sk);
1871
1872         return tp->packets_out &&
1873                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1874 }
1875
1876 /* Linux NewReno/SACK/FACK/ECN state machine.
1877  * --------------------------------------
1878  *
1879  * "Open"       Normal state, no dubious events, fast path.
1880  * "Disorder"   In all the respects it is "Open",
1881  *              but requires a bit more attention. It is entered when
1882  *              we see some SACKs or dupacks. It is split of "Open"
1883  *              mainly to move some processing from fast path to slow one.
1884  * "CWR"        CWND was reduced due to some Congestion Notification event.
1885  *              It can be ECN, ICMP source quench, local device congestion.
1886  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1887  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1888  *
1889  * tcp_fastretrans_alert() is entered:
1890  * - each incoming ACK, if state is not "Open"
1891  * - when arrived ACK is unusual, namely:
1892  *      * SACK
1893  *      * Duplicate ACK.
1894  *      * ECN ECE.
1895  *
1896  * Counting packets in flight is pretty simple.
1897  *
1898  *      in_flight = packets_out - left_out + retrans_out
1899  *
1900  *      packets_out is SND.NXT-SND.UNA counted in packets.
1901  *
1902  *      retrans_out is number of retransmitted segments.
1903  *
1904  *      left_out is number of segments left network, but not ACKed yet.
1905  *
1906  *              left_out = sacked_out + lost_out
1907  *
1908  *     sacked_out: Packets, which arrived to receiver out of order
1909  *                 and hence not ACKed. With SACKs this number is simply
1910  *                 amount of SACKed data. Even without SACKs
1911  *                 it is easy to give pretty reliable estimate of this number,
1912  *                 counting duplicate ACKs.
1913  *
1914  *       lost_out: Packets lost by network. TCP has no explicit
1915  *                 "loss notification" feedback from network (for now).
1916  *                 It means that this number can be only _guessed_.
1917  *                 Actually, it is the heuristics to predict lossage that
1918  *                 distinguishes different algorithms.
1919  *
1920  *      F.e. after RTO, when all the queue is considered as lost,
1921  *      lost_out = packets_out and in_flight = retrans_out.
1922  *
1923  *              Essentially, we have now two algorithms counting
1924  *              lost packets.
1925  *
1926  *              FACK: It is the simplest heuristics. As soon as we decided
1927  *              that something is lost, we decide that _all_ not SACKed
1928  *              packets until the most forward SACK are lost. I.e.
1929  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1930  *              It is absolutely correct estimate, if network does not reorder
1931  *              packets. And it loses any connection to reality when reordering
1932  *              takes place. We use FACK by default until reordering
1933  *              is suspected on the path to this destination.
1934  *
1935  *              NewReno: when Recovery is entered, we assume that one segment
1936  *              is lost (classic Reno). While we are in Recovery and
1937  *              a partial ACK arrives, we assume that one more packet
1938  *              is lost (NewReno). This heuristics are the same in NewReno
1939  *              and SACK.
1940  *
1941  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1942  *  deflation etc. CWND is real congestion window, never inflated, changes
1943  *  only according to classic VJ rules.
1944  *
1945  * Really tricky (and requiring careful tuning) part of algorithm
1946  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1947  * The first determines the moment _when_ we should reduce CWND and,
1948  * hence, slow down forward transmission. In fact, it determines the moment
1949  * when we decide that hole is caused by loss, rather than by a reorder.
1950  *
1951  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1952  * holes, caused by lost packets.
1953  *
1954  * And the most logically complicated part of algorithm is undo
1955  * heuristics. We detect false retransmits due to both too early
1956  * fast retransmit (reordering) and underestimated RTO, analyzing
1957  * timestamps and D-SACKs. When we detect that some segments were
1958  * retransmitted by mistake and CWND reduction was wrong, we undo
1959  * window reduction and abort recovery phase. This logic is hidden
1960  * inside several functions named tcp_try_undo_<something>.
1961  */
1962
1963 /* This function decides, when we should leave Disordered state
1964  * and enter Recovery phase, reducing congestion window.
1965  *
1966  * Main question: may we further continue forward transmission
1967  * with the same cwnd?
1968  */
1969 static int tcp_time_to_recover(struct sock *sk)
1970 {
1971         struct tcp_sock *tp = tcp_sk(sk);
1972         __u32 packets_out;
1973
1974         /* Do not perform any recovery during F-RTO algorithm */
1975         if (tp->frto_counter)
1976                 return 0;
1977
1978         /* Trick#1: The loss is proven. */
1979         if (tp->lost_out)
1980                 return 1;
1981
1982         /* Not-A-Trick#2 : Classic rule... */
1983         if (tcp_fackets_out(tp) > tp->reordering)
1984                 return 1;
1985
1986         /* Trick#3 : when we use RFC2988 timer restart, fast
1987          * retransmit can be triggered by timeout of queue head.
1988          */
1989         if (tcp_head_timedout(sk))
1990                 return 1;
1991
1992         /* Trick#4: It is still not OK... But will it be useful to delay
1993          * recovery more?
1994          */
1995         packets_out = tp->packets_out;
1996         if (packets_out <= tp->reordering &&
1997             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1998             !tcp_may_send_now(sk)) {
1999                 /* We have nothing to send. This connection is limited
2000                  * either by receiver window or by application.
2001                  */
2002                 return 1;
2003         }
2004
2005         return 0;
2006 }
2007
2008 /* RFC: This is from the original, I doubt that this is necessary at all:
2009  * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2010  * retransmitted past LOST markings in the first place? I'm not fully sure
2011  * about undo and end of connection cases, which can cause R without L?
2012  */
2013 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
2014                                        struct sk_buff *skb)
2015 {
2016         if ((tp->retransmit_skb_hint != NULL) &&
2017             before(TCP_SKB_CB(skb)->seq,
2018             TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
2019                 tp->retransmit_skb_hint = NULL;
2020 }
2021
2022 /* Mark head of queue up as lost. */
2023 static void tcp_mark_head_lost(struct sock *sk, int packets)
2024 {
2025         struct tcp_sock *tp = tcp_sk(sk);
2026         struct sk_buff *skb;
2027         int cnt;
2028
2029         BUG_TRAP(packets <= tp->packets_out);
2030         if (tp->lost_skb_hint) {
2031                 skb = tp->lost_skb_hint;
2032                 cnt = tp->lost_cnt_hint;
2033         } else {
2034                 skb = tcp_write_queue_head(sk);
2035                 cnt = 0;
2036         }
2037
2038         tcp_for_write_queue_from(skb, sk) {
2039                 if (skb == tcp_send_head(sk))
2040                         break;
2041                 /* TODO: do this better */
2042                 /* this is not the most efficient way to do this... */
2043                 tp->lost_skb_hint = skb;
2044                 tp->lost_cnt_hint = cnt;
2045                 cnt += tcp_skb_pcount(skb);
2046                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2047                         break;
2048                 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2049                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2050                         tp->lost_out += tcp_skb_pcount(skb);
2051                         tcp_verify_retransmit_hint(tp, skb);
2052                 }
2053         }
2054         tcp_verify_left_out(tp);
2055 }
2056
2057 /* Account newly detected lost packet(s) */
2058
2059 static void tcp_update_scoreboard(struct sock *sk)
2060 {
2061         struct tcp_sock *tp = tcp_sk(sk);
2062
2063         if (tcp_is_fack(tp)) {
2064                 int lost = tp->fackets_out - tp->reordering;
2065                 if (lost <= 0)
2066                         lost = 1;
2067                 tcp_mark_head_lost(sk, lost);
2068         } else {
2069                 tcp_mark_head_lost(sk, 1);
2070         }
2071
2072         /* New heuristics: it is possible only after we switched
2073          * to restart timer each time when something is ACKed.
2074          * Hence, we can detect timed out packets during fast
2075          * retransmit without falling to slow start.
2076          */
2077         if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) {
2078                 struct sk_buff *skb;
2079
2080                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2081                         : tcp_write_queue_head(sk);
2082
2083                 tcp_for_write_queue_from(skb, sk) {
2084                         if (skb == tcp_send_head(sk))
2085                                 break;
2086                         if (!tcp_skb_timedout(sk, skb))
2087                                 break;
2088
2089                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2090                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2091                                 tp->lost_out += tcp_skb_pcount(skb);
2092                                 tcp_verify_retransmit_hint(tp, skb);
2093                         }
2094                 }
2095
2096                 tp->scoreboard_skb_hint = skb;
2097
2098                 tcp_verify_left_out(tp);
2099         }
2100 }
2101
2102 /* CWND moderation, preventing bursts due to too big ACKs
2103  * in dubious situations.
2104  */
2105 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2106 {
2107         tp->snd_cwnd = min(tp->snd_cwnd,
2108                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2109         tp->snd_cwnd_stamp = tcp_time_stamp;
2110 }
2111
2112 /* Lower bound on congestion window is slow start threshold
2113  * unless congestion avoidance choice decides to overide it.
2114  */
2115 static inline u32 tcp_cwnd_min(const struct sock *sk)
2116 {
2117         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2118
2119         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2120 }
2121
2122 /* Decrease cwnd each second ack. */
2123 static void tcp_cwnd_down(struct sock *sk, int flag)
2124 {
2125         struct tcp_sock *tp = tcp_sk(sk);
2126         int decr = tp->snd_cwnd_cnt + 1;
2127
2128         if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2129             (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2130                 tp->snd_cwnd_cnt = decr&1;
2131                 decr >>= 1;
2132
2133                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2134                         tp->snd_cwnd -= decr;
2135
2136                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2137                 tp->snd_cwnd_stamp = tcp_time_stamp;
2138         }
2139 }
2140
2141 /* Nothing was retransmitted or returned timestamp is less
2142  * than timestamp of the first retransmission.
2143  */
2144 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2145 {
2146         return !tp->retrans_stamp ||
2147                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2148                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2149 }
2150
2151 /* Undo procedures. */
2152
2153 #if FASTRETRANS_DEBUG > 1
2154 static void DBGUNDO(struct sock *sk, const char *msg)
2155 {
2156         struct tcp_sock *tp = tcp_sk(sk);
2157         struct inet_sock *inet = inet_sk(sk);
2158
2159         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2160                msg,
2161                NIPQUAD(inet->daddr), ntohs(inet->dport),
2162                tp->snd_cwnd, tcp_left_out(tp),
2163                tp->snd_ssthresh, tp->prior_ssthresh,
2164                tp->packets_out);
2165 }
2166 #else
2167 #define DBGUNDO(x...) do { } while (0)
2168 #endif
2169
2170 static void tcp_undo_cwr(struct sock *sk, const int undo)
2171 {
2172         struct tcp_sock *tp = tcp_sk(sk);
2173
2174         if (tp->prior_ssthresh) {
2175                 const struct inet_connection_sock *icsk = inet_csk(sk);
2176
2177                 if (icsk->icsk_ca_ops->undo_cwnd)
2178                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2179                 else
2180                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2181
2182                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2183                         tp->snd_ssthresh = tp->prior_ssthresh;
2184                         TCP_ECN_withdraw_cwr(tp);
2185                 }
2186         } else {
2187                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2188         }
2189         tcp_moderate_cwnd(tp);
2190         tp->snd_cwnd_stamp = tcp_time_stamp;
2191
2192         /* There is something screwy going on with the retrans hints after
2193            an undo */
2194         tcp_clear_all_retrans_hints(tp);
2195 }
2196
2197 static inline int tcp_may_undo(struct tcp_sock *tp)
2198 {
2199         return tp->undo_marker &&
2200                 (!tp->undo_retrans || tcp_packet_delayed(tp));
2201 }
2202
2203 /* People celebrate: "We love our President!" */
2204 static int tcp_try_undo_recovery(struct sock *sk)
2205 {
2206         struct tcp_sock *tp = tcp_sk(sk);
2207
2208         if (tcp_may_undo(tp)) {
2209                 /* Happy end! We did not retransmit anything
2210                  * or our original transmission succeeded.
2211                  */
2212                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2213                 tcp_undo_cwr(sk, 1);
2214                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2215                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2216                 else
2217                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2218                 tp->undo_marker = 0;
2219         }
2220         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2221                 /* Hold old state until something *above* high_seq
2222                  * is ACKed. For Reno it is MUST to prevent false
2223                  * fast retransmits (RFC2582). SACK TCP is safe. */
2224                 tcp_moderate_cwnd(tp);
2225                 return 1;
2226         }
2227         tcp_set_ca_state(sk, TCP_CA_Open);
2228         return 0;
2229 }
2230
2231 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2232 static void tcp_try_undo_dsack(struct sock *sk)
2233 {
2234         struct tcp_sock *tp = tcp_sk(sk);
2235
2236         if (tp->undo_marker && !tp->undo_retrans) {
2237                 DBGUNDO(sk, "D-SACK");
2238                 tcp_undo_cwr(sk, 1);
2239                 tp->undo_marker = 0;
2240                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2241         }
2242 }
2243
2244 /* Undo during fast recovery after partial ACK. */
2245
2246 static int tcp_try_undo_partial(struct sock *sk, int acked)
2247 {
2248         struct tcp_sock *tp = tcp_sk(sk);
2249         /* Partial ACK arrived. Force Hoe's retransmit. */
2250         int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering;
2251
2252         if (tcp_may_undo(tp)) {
2253                 /* Plain luck! Hole if filled with delayed
2254                  * packet, rather than with a retransmit.
2255                  */
2256                 if (tp->retrans_out == 0)
2257                         tp->retrans_stamp = 0;
2258
2259                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2260
2261                 DBGUNDO(sk, "Hoe");
2262                 tcp_undo_cwr(sk, 0);
2263                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2264
2265                 /* So... Do not make Hoe's retransmit yet.
2266                  * If the first packet was delayed, the rest
2267                  * ones are most probably delayed as well.
2268                  */
2269                 failed = 0;
2270         }
2271         return failed;
2272 }
2273
2274 /* Undo during loss recovery after partial ACK. */
2275 static int tcp_try_undo_loss(struct sock *sk)
2276 {
2277         struct tcp_sock *tp = tcp_sk(sk);
2278
2279         if (tcp_may_undo(tp)) {
2280                 struct sk_buff *skb;
2281                 tcp_for_write_queue(skb, sk) {
2282                         if (skb == tcp_send_head(sk))
2283                                 break;
2284                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2285                 }
2286
2287                 tcp_clear_all_retrans_hints(tp);
2288
2289                 DBGUNDO(sk, "partial loss");
2290                 tp->lost_out = 0;
2291                 tcp_undo_cwr(sk, 1);
2292                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2293                 inet_csk(sk)->icsk_retransmits = 0;
2294                 tp->undo_marker = 0;
2295                 if (tcp_is_sack(tp))
2296                         tcp_set_ca_state(sk, TCP_CA_Open);
2297                 return 1;
2298         }
2299         return 0;
2300 }
2301
2302 static inline void tcp_complete_cwr(struct sock *sk)
2303 {
2304         struct tcp_sock *tp = tcp_sk(sk);
2305         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2306         tp->snd_cwnd_stamp = tcp_time_stamp;
2307         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2308 }
2309
2310 static void tcp_try_to_open(struct sock *sk, int flag)
2311 {
2312         struct tcp_sock *tp = tcp_sk(sk);
2313
2314         tcp_verify_left_out(tp);
2315
2316         if (tp->retrans_out == 0)
2317                 tp->retrans_stamp = 0;
2318
2319         if (flag&FLAG_ECE)
2320                 tcp_enter_cwr(sk, 1);
2321
2322         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2323                 int state = TCP_CA_Open;
2324
2325                 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2326                         state = TCP_CA_Disorder;
2327
2328                 if (inet_csk(sk)->icsk_ca_state != state) {
2329                         tcp_set_ca_state(sk, state);
2330                         tp->high_seq = tp->snd_nxt;
2331                 }
2332                 tcp_moderate_cwnd(tp);
2333         } else {
2334                 tcp_cwnd_down(sk, flag);
2335         }
2336 }
2337
2338 static void tcp_mtup_probe_failed(struct sock *sk)
2339 {
2340         struct inet_connection_sock *icsk = inet_csk(sk);
2341
2342         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2343         icsk->icsk_mtup.probe_size = 0;
2344 }
2345
2346 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2347 {
2348         struct tcp_sock *tp = tcp_sk(sk);
2349         struct inet_connection_sock *icsk = inet_csk(sk);
2350
2351         /* FIXME: breaks with very large cwnd */
2352         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2353         tp->snd_cwnd = tp->snd_cwnd *
2354                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2355                        icsk->icsk_mtup.probe_size;
2356         tp->snd_cwnd_cnt = 0;
2357         tp->snd_cwnd_stamp = tcp_time_stamp;
2358         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2359
2360         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2361         icsk->icsk_mtup.probe_size = 0;
2362         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2363 }
2364
2365
2366 /* Process an event, which can update packets-in-flight not trivially.
2367  * Main goal of this function is to calculate new estimate for left_out,
2368  * taking into account both packets sitting in receiver's buffer and
2369  * packets lost by network.
2370  *
2371  * Besides that it does CWND reduction, when packet loss is detected
2372  * and changes state of machine.
2373  *
2374  * It does _not_ decide what to send, it is made in function
2375  * tcp_xmit_retransmit_queue().
2376  */
2377 static void
2378 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2379 {
2380         struct inet_connection_sock *icsk = inet_csk(sk);
2381         struct tcp_sock *tp = tcp_sk(sk);
2382         int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2383         int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2384                                     (tp->fackets_out > tp->reordering));
2385
2386         /* Some technical things:
2387          * 1. Reno does not count dupacks (sacked_out) automatically. */
2388         if (!tp->packets_out)
2389                 tp->sacked_out = 0;
2390
2391         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2392                 tp->fackets_out = 0;
2393
2394         /* Now state machine starts.
2395          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2396         if (flag&FLAG_ECE)
2397                 tp->prior_ssthresh = 0;
2398
2399         /* B. In all the states check for reneging SACKs. */
2400         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2401                 return;
2402
2403         /* C. Process data loss notification, provided it is valid. */
2404         if ((flag&FLAG_DATA_LOST) &&
2405             before(tp->snd_una, tp->high_seq) &&
2406             icsk->icsk_ca_state != TCP_CA_Open &&
2407             tp->fackets_out > tp->reordering) {
2408                 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2409                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2410         }
2411
2412         /* D. Check consistency of the current state. */
2413         tcp_verify_left_out(tp);
2414
2415         /* E. Check state exit conditions. State can be terminated
2416          *    when high_seq is ACKed. */
2417         if (icsk->icsk_ca_state == TCP_CA_Open) {
2418                 BUG_TRAP(tp->retrans_out == 0);
2419                 tp->retrans_stamp = 0;
2420         } else if (!before(tp->snd_una, tp->high_seq)) {
2421                 switch (icsk->icsk_ca_state) {
2422                 case TCP_CA_Loss:
2423                         icsk->icsk_retransmits = 0;
2424                         if (tcp_try_undo_recovery(sk))
2425                                 return;
2426                         break;
2427
2428                 case TCP_CA_CWR:
2429                         /* CWR is to be held something *above* high_seq
2430                          * is ACKed for CWR bit to reach receiver. */
2431                         if (tp->snd_una != tp->high_seq) {
2432                                 tcp_complete_cwr(sk);
2433                                 tcp_set_ca_state(sk, TCP_CA_Open);
2434                         }
2435                         break;
2436
2437                 case TCP_CA_Disorder:
2438                         tcp_try_undo_dsack(sk);
2439                         if (!tp->undo_marker ||
2440                             /* For SACK case do not Open to allow to undo
2441                              * catching for all duplicate ACKs. */
2442                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2443                                 tp->undo_marker = 0;
2444                                 tcp_set_ca_state(sk, TCP_CA_Open);
2445                         }
2446                         break;
2447
2448                 case TCP_CA_Recovery:
2449                         if (tcp_is_reno(tp))
2450                                 tcp_reset_reno_sack(tp);
2451                         if (tcp_try_undo_recovery(sk))
2452                                 return;
2453                         tcp_complete_cwr(sk);
2454                         break;
2455                 }
2456         }
2457
2458         /* F. Process state. */
2459         switch (icsk->icsk_ca_state) {
2460         case TCP_CA_Recovery:
2461                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2462                         if (tcp_is_reno(tp) && is_dupack)
2463                                 tcp_add_reno_sack(sk);
2464                 } else
2465                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2466                 break;
2467         case TCP_CA_Loss:
2468                 if (flag&FLAG_DATA_ACKED)
2469                         icsk->icsk_retransmits = 0;
2470                 if (!tcp_try_undo_loss(sk)) {
2471                         tcp_moderate_cwnd(tp);
2472                         tcp_xmit_retransmit_queue(sk);
2473                         return;
2474                 }
2475                 if (icsk->icsk_ca_state != TCP_CA_Open)
2476                         return;
2477                 /* Loss is undone; fall through to processing in Open state. */
2478         default:
2479                 if (tcp_is_reno(tp)) {
2480                         if (flag & FLAG_SND_UNA_ADVANCED)
2481                                 tcp_reset_reno_sack(tp);
2482                         if (is_dupack)
2483                                 tcp_add_reno_sack(sk);
2484                 }
2485
2486                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2487                         tcp_try_undo_dsack(sk);
2488
2489                 if (!tcp_time_to_recover(sk)) {
2490                         tcp_try_to_open(sk, flag);
2491                         return;
2492                 }
2493
2494                 /* MTU probe failure: don't reduce cwnd */
2495                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2496                     icsk->icsk_mtup.probe_size &&
2497                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2498                         tcp_mtup_probe_failed(sk);
2499                         /* Restores the reduction we did in tcp_mtup_probe() */
2500                         tp->snd_cwnd++;
2501                         tcp_simple_retransmit(sk);
2502                         return;
2503                 }
2504
2505                 /* Otherwise enter Recovery state */
2506
2507                 if (tcp_is_reno(tp))
2508                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2509                 else
2510                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2511
2512                 tp->high_seq = tp->snd_nxt;
2513                 tp->prior_ssthresh = 0;
2514                 tp->undo_marker = tp->snd_una;
2515                 tp->undo_retrans = tp->retrans_out;
2516
2517                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2518                         if (!(flag&FLAG_ECE))
2519                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2520                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2521                         TCP_ECN_queue_cwr(tp);
2522                 }
2523
2524                 tp->bytes_acked = 0;
2525                 tp->snd_cwnd_cnt = 0;
2526                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2527         }
2528
2529         if (do_lost || tcp_head_timedout(sk))
2530                 tcp_update_scoreboard(sk);
2531         tcp_cwnd_down(sk, flag);
2532         tcp_xmit_retransmit_queue(sk);
2533 }
2534
2535 /* Read draft-ietf-tcplw-high-performance before mucking
2536  * with this code. (Supersedes RFC1323)
2537  */
2538 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2539 {
2540         /* RTTM Rule: A TSecr value received in a segment is used to
2541          * update the averaged RTT measurement only if the segment
2542          * acknowledges some new data, i.e., only if it advances the
2543          * left edge of the send window.
2544          *
2545          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2546          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2547          *
2548          * Changed: reset backoff as soon as we see the first valid sample.
2549          * If we do not, we get strongly overestimated rto. With timestamps
2550          * samples are accepted even from very old segments: f.e., when rtt=1
2551          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2552          * answer arrives rto becomes 120 seconds! If at least one of segments
2553          * in window is lost... Voila.                          --ANK (010210)
2554          */
2555         struct tcp_sock *tp = tcp_sk(sk);
2556         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2557         tcp_rtt_estimator(sk, seq_rtt);
2558         tcp_set_rto(sk);
2559         inet_csk(sk)->icsk_backoff = 0;
2560         tcp_bound_rto(sk);
2561 }
2562
2563 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2564 {
2565         /* We don't have a timestamp. Can only use
2566          * packets that are not retransmitted to determine
2567          * rtt estimates. Also, we must not reset the
2568          * backoff for rto until we get a non-retransmitted
2569          * packet. This allows us to deal with a situation
2570          * where the network delay has increased suddenly.
2571          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2572          */
2573
2574         if (flag & FLAG_RETRANS_DATA_ACKED)
2575                 return;
2576
2577         tcp_rtt_estimator(sk, seq_rtt);
2578         tcp_set_rto(sk);
2579         inet_csk(sk)->icsk_backoff = 0;
2580         tcp_bound_rto(sk);
2581 }
2582
2583 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2584                                       const s32 seq_rtt)
2585 {
2586         const struct tcp_sock *tp = tcp_sk(sk);
2587         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2588         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2589                 tcp_ack_saw_tstamp(sk, flag);
2590         else if (seq_rtt >= 0)
2591                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2592 }
2593
2594 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2595                            u32 in_flight, int good)
2596 {
2597         const struct inet_connection_sock *icsk = inet_csk(sk);
2598         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2599         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2600 }
2601
2602 /* Restart timer after forward progress on connection.
2603  * RFC2988 recommends to restart timer to now+rto.
2604  */
2605 static void tcp_rearm_rto(struct sock *sk)
2606 {
2607         struct tcp_sock *tp = tcp_sk(sk);
2608
2609         if (!tp->packets_out) {
2610                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2611         } else {
2612                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2613         }
2614 }
2615
2616 /* If we get here, the whole TSO packet has not been acked. */
2617 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2618 {
2619         struct tcp_sock *tp = tcp_sk(sk);
2620         u32 packets_acked;
2621
2622         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2623
2624         packets_acked = tcp_skb_pcount(skb);
2625         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2626                 return 0;
2627         packets_acked -= tcp_skb_pcount(skb);
2628
2629         if (packets_acked) {
2630                 BUG_ON(tcp_skb_pcount(skb) == 0);
2631                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2632         }
2633
2634         return packets_acked;
2635 }
2636
2637 /* Remove acknowledged frames from the retransmission queue. If our packet
2638  * is before the ack sequence we can discard it as it's confirmed to have
2639  * arrived at the other end.
2640  */
2641 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p,
2642                                int prior_fackets)
2643 {
2644         struct tcp_sock *tp = tcp_sk(sk);
2645         const struct inet_connection_sock *icsk = inet_csk(sk);
2646         struct sk_buff *skb;
2647         u32 now = tcp_time_stamp;
2648         int fully_acked = 1;
2649         int flag = 0;
2650         int prior_packets = tp->packets_out;
2651         u32 cnt = 0;
2652         u32 reord = tp->packets_out;
2653         s32 seq_rtt = -1;
2654         ktime_t last_ackt = net_invalid_timestamp();
2655
2656         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2657                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2658                 u32 end_seq;
2659                 u32 packets_acked;
2660                 u8 sacked = scb->sacked;
2661
2662                 if (after(scb->end_seq, tp->snd_una)) {
2663                         if (tcp_skb_pcount(skb) == 1 ||
2664                             !after(tp->snd_una, scb->seq))
2665                                 break;
2666
2667                         packets_acked = tcp_tso_acked(sk, skb);
2668                         if (!packets_acked)
2669                                 break;
2670
2671                         fully_acked = 0;
2672                         end_seq = tp->snd_una;
2673                 } else {
2674                         packets_acked = tcp_skb_pcount(skb);
2675                         end_seq = scb->end_seq;
2676                 }
2677
2678                 /* MTU probing checks */
2679                 if (fully_acked && icsk->icsk_mtup.probe_size &&
2680                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2681                         tcp_mtup_probe_success(sk, skb);
2682                 }
2683
2684                 if (sacked) {
2685                         if (sacked & TCPCB_RETRANS) {
2686                                 if (sacked & TCPCB_SACKED_RETRANS)
2687                                         tp->retrans_out -= packets_acked;
2688                                 flag |= FLAG_RETRANS_DATA_ACKED;
2689                                 seq_rtt = -1;
2690                                 if ((flag & FLAG_DATA_ACKED) ||
2691                                     (packets_acked > 1))
2692                                         flag |= FLAG_NONHEAD_RETRANS_ACKED;
2693                         } else {
2694                                 if (seq_rtt < 0) {
2695                                         seq_rtt = now - scb->when;
2696                                         if (fully_acked)
2697                                                 last_ackt = skb->tstamp;
2698                                 }
2699                                 if (!(sacked & TCPCB_SACKED_ACKED))
2700                                         reord = min(cnt, reord);
2701                         }
2702
2703                         if (sacked & TCPCB_SACKED_ACKED)
2704                                 tp->sacked_out -= packets_acked;
2705                         if (sacked & TCPCB_LOST)
2706                                 tp->lost_out -= packets_acked;
2707
2708                         if ((sacked & TCPCB_URG) && tp->urg_mode &&
2709                             !before(end_seq, tp->snd_up))
2710                                 tp->urg_mode = 0;
2711                 } else {
2712                         if (seq_rtt < 0) {
2713                                 seq_rtt = now - scb->when;
2714                                 if (fully_acked)
2715                                         last_ackt = skb->tstamp;
2716                         }
2717                         reord = min(cnt, reord);
2718                 }
2719                 tp->packets_out -= packets_acked;
2720                 cnt += packets_acked;
2721
2722                 /* Initial outgoing SYN's get put onto the write_queue
2723                  * just like anything else we transmit.  It is not
2724                  * true data, and if we misinform our callers that
2725                  * this ACK acks real data, we will erroneously exit
2726                  * connection startup slow start one packet too
2727                  * quickly.  This is severely frowned upon behavior.
2728                  */
2729                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2730                         flag |= FLAG_DATA_ACKED;
2731                 } else {
2732                         flag |= FLAG_SYN_ACKED;
2733                         tp->retrans_stamp = 0;
2734                 }
2735
2736                 if (!fully_acked)
2737                         break;
2738
2739                 tcp_unlink_write_queue(skb, sk);
2740                 sk_stream_free_skb(sk, skb);
2741                 tcp_clear_all_retrans_hints(tp);
2742         }
2743
2744         if (flag & FLAG_ACKED) {
2745                 u32 pkts_acked = prior_packets - tp->packets_out;
2746                 const struct tcp_congestion_ops *ca_ops
2747                         = inet_csk(sk)->icsk_ca_ops;
2748
2749                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2750                 tcp_rearm_rto(sk);
2751
2752                 if (tcp_is_reno(tp)) {
2753                         tcp_remove_reno_sacks(sk, pkts_acked);
2754                 } else {
2755                         /* Non-retransmitted hole got filled? That's reordering */
2756                         if (reord < prior_fackets)
2757                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2758                 }
2759
2760                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2761                 /* hint's skb might be NULL but we don't need to care */
2762                 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2763                                                tp->fastpath_cnt_hint);
2764                 if (ca_ops->pkts_acked) {
2765                         s32 rtt_us = -1;
2766
2767                         /* Is the ACK triggering packet unambiguous? */
2768                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2769                                 /* High resolution needed and available? */
2770                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2771                                     !ktime_equal(last_ackt,
2772                                                  net_invalid_timestamp()))
2773                                         rtt_us = ktime_us_delta(ktime_get_real(),
2774                                                                 last_ackt);
2775                                 else if (seq_rtt > 0)
2776                                         rtt_us = jiffies_to_usecs(seq_rtt);
2777                         }
2778
2779                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2780                 }
2781         }
2782
2783 #if FASTRETRANS_DEBUG > 0
2784         BUG_TRAP((int)tp->sacked_out >= 0);
2785         BUG_TRAP((int)tp->lost_out >= 0);
2786         BUG_TRAP((int)tp->retrans_out >= 0);
2787         if (!tp->packets_out && tcp_is_sack(tp)) {
2788                 icsk = inet_csk(sk);
2789                 if (tp->lost_out) {
2790                         printk(KERN_DEBUG "Leak l=%u %d\n",
2791                                tp->lost_out, icsk->icsk_ca_state);
2792                         tp->lost_out = 0;
2793                 }
2794                 if (tp->sacked_out) {
2795                         printk(KERN_DEBUG "Leak s=%u %d\n",
2796                                tp->sacked_out, icsk->icsk_ca_state);
2797                         tp->sacked_out = 0;
2798                 }
2799                 if (tp->retrans_out) {
2800                         printk(KERN_DEBUG "Leak r=%u %d\n",
2801                                tp->retrans_out, icsk->icsk_ca_state);
2802                         tp->retrans_out = 0;
2803                 }
2804         }
2805 #endif
2806         *seq_rtt_p = seq_rtt;
2807         return flag;
2808 }
2809
2810 static void tcp_ack_probe(struct sock *sk)
2811 {
2812         const struct tcp_sock *tp = tcp_sk(sk);
2813         struct inet_connection_sock *icsk = inet_csk(sk);
2814
2815         /* Was it a usable window open? */
2816
2817         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2818                    tp->snd_una + tp->snd_wnd)) {
2819                 icsk->icsk_backoff = 0;
2820                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2821                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2822                  * This function is not for random using!
2823                  */
2824         } else {
2825                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2826                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2827                                           TCP_RTO_MAX);
2828         }
2829 }
2830
2831 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2832 {
2833         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2834                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2835 }
2836
2837 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2838 {
2839         const struct tcp_sock *tp = tcp_sk(sk);
2840         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2841                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2842 }
2843
2844 /* Check that window update is acceptable.
2845  * The function assumes that snd_una<=ack<=snd_next.
2846  */
2847 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2848                                         const u32 ack_seq, const u32 nwin)
2849 {
2850         return (after(ack, tp->snd_una) ||
2851                 after(ack_seq, tp->snd_wl1) ||
2852                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2853 }
2854
2855 /* Update our send window.
2856  *
2857  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2858  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2859  */
2860 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2861                                  u32 ack_seq)
2862 {
2863         struct tcp_sock *tp = tcp_sk(sk);
2864         int flag = 0;
2865         u32 nwin = ntohs(tcp_hdr(skb)->window);
2866
2867         if (likely(!tcp_hdr(skb)->syn))
2868                 nwin <<= tp->rx_opt.snd_wscale;
2869
2870         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2871                 flag |= FLAG_WIN_UPDATE;
2872                 tcp_update_wl(tp, ack, ack_seq);
2873
2874                 if (tp->snd_wnd != nwin) {
2875                         tp->snd_wnd = nwin;
2876
2877                         /* Note, it is the only place, where
2878                          * fast path is recovered for sending TCP.
2879                          */
2880                         tp->pred_flags = 0;
2881                         tcp_fast_path_check(sk);
2882
2883                         if (nwin > tp->max_window) {
2884                                 tp->max_window = nwin;
2885                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2886                         }
2887                 }
2888         }
2889
2890         tp->snd_una = ack;
2891
2892         return flag;
2893 }
2894
2895 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2896  * continue in congestion avoidance.
2897  */
2898 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2899 {
2900         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2901         tp->snd_cwnd_cnt = 0;
2902         tp->bytes_acked = 0;
2903         TCP_ECN_queue_cwr(tp);
2904         tcp_moderate_cwnd(tp);
2905 }
2906
2907 /* A conservative spurious RTO response algorithm: reduce cwnd using
2908  * rate halving and continue in congestion avoidance.
2909  */
2910 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2911 {
2912         tcp_enter_cwr(sk, 0);
2913 }
2914
2915 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2916 {
2917         if (flag&FLAG_ECE)
2918                 tcp_ratehalving_spur_to_response(sk);
2919         else
2920                 tcp_undo_cwr(sk, 1);
2921 }
2922
2923 /* F-RTO spurious RTO detection algorithm (RFC4138)
2924  *
2925  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2926  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2927  * window (but not to or beyond highest sequence sent before RTO):
2928  *   On First ACK,  send two new segments out.
2929  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2930  *                  algorithm is not part of the F-RTO detection algorithm
2931  *                  given in RFC4138 but can be selected separately).
2932  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2933  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2934  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2935  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2936  *
2937  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2938  * original window even after we transmit two new data segments.
2939  *
2940  * SACK version:
2941  *   on first step, wait until first cumulative ACK arrives, then move to
2942  *   the second step. In second step, the next ACK decides.
2943  *
2944  * F-RTO is implemented (mainly) in four functions:
2945  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2946  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2947  *     called when tcp_use_frto() showed green light
2948  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2949  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2950  *     to prove that the RTO is indeed spurious. It transfers the control
2951  *     from F-RTO to the conventional RTO recovery
2952  */
2953 static int tcp_process_frto(struct sock *sk, int flag)
2954 {
2955         struct tcp_sock *tp = tcp_sk(sk);
2956
2957         tcp_verify_left_out(tp);
2958
2959         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2960         if (flag&FLAG_DATA_ACKED)
2961                 inet_csk(sk)->icsk_retransmits = 0;
2962
2963         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
2964             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
2965                 tp->undo_marker = 0;
2966
2967         if (!before(tp->snd_una, tp->frto_highmark)) {
2968                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2969                 return 1;
2970         }
2971
2972         if (!IsSackFrto() || tcp_is_reno(tp)) {
2973                 /* RFC4138 shortcoming in step 2; should also have case c):
2974                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2975                  * data, winupdate
2976                  */
2977                 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2978                         return 1;
2979
2980                 if (!(flag&FLAG_DATA_ACKED)) {
2981                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2982                                             flag);
2983                         return 1;
2984                 }
2985         } else {
2986                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2987                         /* Prevent sending of new data. */
2988                         tp->snd_cwnd = min(tp->snd_cwnd,
2989                                            tcp_packets_in_flight(tp));
2990                         return 1;
2991                 }
2992
2993                 if ((tp->frto_counter >= 2) &&
2994                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2995                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2996                         /* RFC4138 shortcoming (see comment above) */
2997                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2998                                 return 1;
2999
3000                         tcp_enter_frto_loss(sk, 3, flag);
3001                         return 1;
3002                 }
3003         }
3004
3005         if (tp->frto_counter == 1) {
3006                 /* Sending of the next skb must be allowed or no F-RTO */
3007                 if (!tcp_send_head(sk) ||
3008                     after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
3009                                      tp->snd_una + tp->snd_wnd)) {
3010                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
3011                                             flag);
3012                         return 1;
3013                 }
3014
3015                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3016                 tp->frto_counter = 2;
3017                 return 1;
3018         } else {
3019                 switch (sysctl_tcp_frto_response) {
3020                 case 2:
3021                         tcp_undo_spur_to_response(sk, flag);
3022                         break;
3023                 case 1:
3024                         tcp_conservative_spur_to_response(tp);
3025                         break;
3026                 default:
3027                         tcp_ratehalving_spur_to_response(sk);
3028                         break;
3029                 }
3030                 tp->frto_counter = 0;
3031                 tp->undo_marker = 0;
3032                 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
3033         }
3034         return 0;
3035 }
3036
3037 /* This routine deals with incoming acks, but not outgoing ones. */
3038 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3039 {
3040         struct inet_connection_sock *icsk = inet_csk(sk);
3041         struct tcp_sock *tp = tcp_sk(sk);
3042         u32 prior_snd_una = tp->snd_una;
3043         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3044         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3045         u32 prior_in_flight;
3046         u32 prior_fackets;
3047         s32 seq_rtt;
3048         int prior_packets;
3049         int frto_cwnd = 0;
3050
3051         /* If the ack is newer than sent or older than previous acks
3052          * then we can probably ignore it.
3053          */
3054         if (after(ack, tp->snd_nxt))
3055                 goto uninteresting_ack;
3056
3057         if (before(ack, prior_snd_una))
3058                 goto old_ack;
3059
3060         if (after(ack, prior_snd_una))
3061                 flag |= FLAG_SND_UNA_ADVANCED;
3062
3063         if (sysctl_tcp_abc) {
3064                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3065                         tp->bytes_acked += ack - prior_snd_una;
3066                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3067                         /* we assume just one segment left network */
3068                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
3069         }
3070
3071         prior_fackets = tp->fackets_out;
3072
3073         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3074                 /* Window is constant, pure forward advance.
3075                  * No more checks are required.
3076                  * Note, we use the fact that SND.UNA>=SND.WL2.
3077                  */
3078                 tcp_update_wl(tp, ack, ack_seq);
3079                 tp->snd_una = ack;
3080                 flag |= FLAG_WIN_UPDATE;
3081
3082                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3083
3084                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3085         } else {
3086                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3087                         flag |= FLAG_DATA;
3088                 else
3089                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3090
3091                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3092
3093                 if (TCP_SKB_CB(skb)->sacked)
3094                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3095
3096                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3097                         flag |= FLAG_ECE;
3098
3099                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3100         }
3101
3102         /* We passed data and got it acked, remove any soft error
3103          * log. Something worked...
3104          */
3105         sk->sk_err_soft = 0;
3106         tp->rcv_tstamp = tcp_time_stamp;
3107         prior_packets = tp->packets_out;
3108         if (!prior_packets)
3109                 goto no_queue;
3110
3111         prior_in_flight = tcp_packets_in_flight(tp);
3112
3113         /* See if we can take anything off of the retransmit queue. */
3114         flag |= tcp_clean_rtx_queue(sk, &seq_rtt, prior_fackets);
3115
3116         if (tp->frto_counter)
3117                 frto_cwnd = tcp_process_frto(sk, flag);
3118         /* Guarantee sacktag reordering detection against wrap-arounds */
3119         if (before(tp->frto_highmark, tp->snd_una))
3120                 tp->frto_highmark = 0;
3121
3122         if (tcp_ack_is_dubious(sk, flag)) {
3123                 /* Advance CWND, if state allows this. */
3124                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3125                     tcp_may_raise_cwnd(sk, flag))
3126                         tcp_cong_avoid(sk, ack, prior_in_flight, 0);
3127                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3128         } else {
3129                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3130                         tcp_cong_avoid(sk, ack, prior_in_flight, 1);
3131         }
3132
3133         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3134                 dst_confirm(sk->sk_dst_cache);
3135
3136         return 1;
3137
3138 no_queue:
3139         icsk->icsk_probes_out = 0;
3140
3141         /* If this ack opens up a zero window, clear backoff.  It was
3142          * being used to time the probes, and is probably far higher than
3143          * it needs to be for normal retransmission.
3144          */
3145         if (tcp_send_head(sk))
3146                 tcp_ack_probe(sk);
3147         return 1;
3148
3149 old_ack:
3150         if (TCP_SKB_CB(skb)->sacked)
3151                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3152
3153 uninteresting_ack:
3154         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3155         return 0;
3156 }
3157
3158
3159 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3160  * But, this can also be called on packets in the established flow when
3161  * the fast version below fails.
3162  */
3163 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3164 {
3165         unsigned char *ptr;
3166         struct tcphdr *th = tcp_hdr(skb);
3167         int length=(th->doff*4)-sizeof(struct tcphdr);
3168
3169         ptr = (unsigned char *)(th + 1);
3170         opt_rx->saw_tstamp = 0;
3171
3172         while (length > 0) {
3173                 int opcode=*ptr++;
3174                 int opsize;
3175
3176                 switch (opcode) {
3177                         case TCPOPT_EOL:
3178                                 return;
3179                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3180                                 length--;
3181                                 continue;
3182                         default:
3183                                 opsize=*ptr++;
3184                                 if (opsize < 2) /* "silly options" */
3185                                         return;
3186                                 if (opsize > length)
3187                                         return; /* don't parse partial options */
3188                                 switch (opcode) {
3189                                 case TCPOPT_MSS:
3190                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3191                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3192                                                 if (in_mss) {
3193                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3194                                                                 in_mss = opt_rx->user_mss;
3195                                                         opt_rx->mss_clamp = in_mss;
3196                                                 }
3197                                         }
3198                                         break;
3199                                 case TCPOPT_WINDOW:
3200                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3201                                                 if (sysctl_tcp_window_scaling) {
3202                                                         __u8 snd_wscale = *(__u8 *) ptr;
3203                                                         opt_rx->wscale_ok = 1;
3204                                                         if (snd_wscale > 14) {
3205                                                                 if (net_ratelimit())
3206                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3207                                                                                "scaling value %d >14 received.\n",
3208                                                                                snd_wscale);
3209                                                                 snd_wscale = 14;
3210                                                         }
3211                                                         opt_rx->snd_wscale = snd_wscale;
3212                                                 }
3213                                         break;
3214                                 case TCPOPT_TIMESTAMP:
3215                                         if (opsize==TCPOLEN_TIMESTAMP) {
3216                                                 if ((estab && opt_rx->tstamp_ok) ||
3217                                                     (!estab && sysctl_tcp_timestamps)) {
3218                                                         opt_rx->saw_tstamp = 1;
3219                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3220                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3221                                                 }
3222                                         }
3223                                         break;
3224                                 case TCPOPT_SACK_PERM:
3225                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3226                                                 if (sysctl_tcp_sack) {
3227                                                         opt_rx->sack_ok = 1;
3228                                                         tcp_sack_reset(opt_rx);
3229                                                 }
3230                                         }
3231                                         break;
3232
3233                                 case TCPOPT_SACK:
3234                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3235                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3236                                            opt_rx->sack_ok) {
3237                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3238                                         }
3239                                         break;
3240 #ifdef CONFIG_TCP_MD5SIG
3241                                 case TCPOPT_MD5SIG:
3242                                         /*
3243                                          * The MD5 Hash has already been
3244                                          * checked (see tcp_v{4,6}_do_rcv()).
3245                                          */
3246                                         break;
3247 #endif
3248                                 }
3249
3250                                 ptr+=opsize-2;
3251                                 length-=opsize;
3252                 }
3253         }
3254 }
3255
3256 /* Fast parse options. This hopes to only see timestamps.
3257  * If it is wrong it falls back on tcp_parse_options().
3258  */
3259 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3260                                   struct tcp_sock *tp)
3261 {
3262         if (th->doff == sizeof(struct tcphdr)>>2) {
3263                 tp->rx_opt.saw_tstamp = 0;
3264                 return 0;
3265         } else if (tp->rx_opt.tstamp_ok &&
3266                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3267                 __be32 *ptr = (__be32 *)(th + 1);
3268                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3269                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3270                         tp->rx_opt.saw_tstamp = 1;
3271                         ++ptr;
3272                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3273                         ++ptr;
3274                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3275                         return 1;
3276                 }
3277         }
3278         tcp_parse_options(skb, &tp->rx_opt, 1);
3279         return 1;
3280 }
3281
3282 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3283 {
3284         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3285         tp->rx_opt.ts_recent_stamp = get_seconds();
3286 }
3287
3288 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3289 {
3290         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3291                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3292                  * extra check below makes sure this can only happen
3293                  * for pure ACK frames.  -DaveM
3294                  *
3295                  * Not only, also it occurs for expired timestamps.
3296                  */
3297
3298                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3299                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3300                         tcp_store_ts_recent(tp);
3301         }
3302 }
3303
3304 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3305  *
3306  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3307  * it can pass through stack. So, the following predicate verifies that
3308  * this segment is not used for anything but congestion avoidance or
3309  * fast retransmit. Moreover, we even are able to eliminate most of such
3310  * second order effects, if we apply some small "replay" window (~RTO)
3311  * to timestamp space.
3312  *
3313  * All these measures still do not guarantee that we reject wrapped ACKs
3314  * on networks with high bandwidth, when sequence space is recycled fastly,
3315  * but it guarantees that such events will be very rare and do not affect
3316  * connection seriously. This doesn't look nice, but alas, PAWS is really
3317  * buggy extension.
3318  *
3319  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3320  * states that events when retransmit arrives after original data are rare.
3321  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3322  * the biggest problem on large power networks even with minor reordering.
3323  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3324  * up to bandwidth of 18Gigabit/sec. 8) ]
3325  */
3326
3327 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3328 {
3329         struct tcp_sock *tp = tcp_sk(sk);
3330         struct tcphdr *th = tcp_hdr(skb);
3331         u32 seq = TCP_SKB_CB(skb)->seq;
3332         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3333
3334         return (/* 1. Pure ACK with correct sequence number. */
3335                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3336
3337                 /* 2. ... and duplicate ACK. */
3338                 ack == tp->snd_una &&
3339
3340                 /* 3. ... and does not update window. */
3341                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3342
3343                 /* 4. ... and sits in replay window. */
3344                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3345 }
3346
3347 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3348 {
3349         const struct tcp_sock *tp = tcp_sk(sk);
3350         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3351                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3352                 !tcp_disordered_ack(sk, skb));
3353 }
3354
3355 /* Check segment sequence number for validity.
3356  *
3357  * Segment controls are considered valid, if the segment
3358  * fits to the window after truncation to the window. Acceptability
3359  * of data (and SYN, FIN, of course) is checked separately.
3360  * See tcp_data_queue(), for example.
3361  *
3362  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3363  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3364  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3365  * (borrowed from freebsd)
3366  */
3367
3368 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3369 {
3370         return  !before(end_seq, tp->rcv_wup) &&
3371                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3372 }
3373
3374 /* When we get a reset we do this. */
3375 static void tcp_reset(struct sock *sk)
3376 {
3377         /* We want the right error as BSD sees it (and indeed as we do). */
3378         switch (sk->sk_state) {
3379                 case TCP_SYN_SENT:
3380                         sk->sk_err = ECONNREFUSED;
3381                         break;
3382                 case TCP_CLOSE_WAIT:
3383                         sk->sk_err = EPIPE;
3384                         break;
3385                 case TCP_CLOSE:
3386                         return;
3387                 default:
3388                         sk->sk_err = ECONNRESET;
3389         }
3390
3391         if (!sock_flag(sk, SOCK_DEAD))
3392                 sk->sk_error_report(sk);
3393
3394         tcp_done(sk);
3395 }
3396
3397 /*
3398  *      Process the FIN bit. This now behaves as it is supposed to work
3399  *      and the FIN takes effect when it is validly part of sequence
3400  *      space. Not before when we get holes.
3401  *
3402  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3403  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3404  *      TIME-WAIT)
3405  *
3406  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3407  *      close and we go into CLOSING (and later onto TIME-WAIT)
3408  *
3409  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3410  */
3411 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3412 {
3413         struct tcp_sock *tp = tcp_sk(sk);
3414
3415         inet_csk_schedule_ack(sk);
3416
3417         sk->sk_shutdown |= RCV_SHUTDOWN;
3418         sock_set_flag(sk, SOCK_DONE);
3419
3420         switch (sk->sk_state) {
3421                 case TCP_SYN_RECV:
3422                 case TCP_ESTABLISHED:
3423                         /* Move to CLOSE_WAIT */
3424                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3425                         inet_csk(sk)->icsk_ack.pingpong = 1;
3426                         break;
3427
3428                 case TCP_CLOSE_WAIT:
3429                 case TCP_CLOSING:
3430                         /* Received a retransmission of the FIN, do
3431                          * nothing.
3432                          */
3433                         break;
3434                 case TCP_LAST_ACK:
3435                         /* RFC793: Remain in the LAST-ACK state. */
3436                         break;
3437
3438                 case TCP_FIN_WAIT1:
3439                         /* This case occurs when a simultaneous close
3440                          * happens, we must ack the received FIN and
3441                          * enter the CLOSING state.
3442                          */
3443                         tcp_send_ack(sk);
3444                         tcp_set_state(sk, TCP_CLOSING);
3445                         break;
3446                 case TCP_FIN_WAIT2:
3447                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3448                         tcp_send_ack(sk);
3449                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3450                         break;
3451                 default:
3452                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3453                          * cases we should never reach this piece of code.
3454                          */
3455                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3456                                __FUNCTION__, sk->sk_state);
3457                         break;
3458         }
3459
3460         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3461          * Probably, we should reset in this case. For now drop them.
3462          */
3463         __skb_queue_purge(&tp->out_of_order_queue);
3464         if (tcp_is_sack(tp))
3465                 tcp_sack_reset(&tp->rx_opt);
3466         sk_stream_mem_reclaim(sk);
3467
3468         if (!sock_flag(sk, SOCK_DEAD)) {
3469                 sk->sk_state_change(sk);
3470
3471                 /* Do not send POLL_HUP for half duplex close. */
3472                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3473                     sk->sk_state == TCP_CLOSE)
3474                         sk_wake_async(sk, 1, POLL_HUP);
3475                 else
3476                         sk_wake_async(sk, 1, POLL_IN);
3477         }
3478 }
3479
3480 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3481 {
3482         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3483                 if (before(seq, sp->start_seq))
3484                         sp->start_seq = seq;
3485                 if (after(end_seq, sp->end_seq))
3486                         sp->end_seq = end_seq;
3487                 return 1;
3488         }
3489         return 0;
3490 }
3491
3492 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3493 {
3494         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3495                 if (before(seq, tp->rcv_nxt))
3496                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3497                 else
3498                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3499
3500                 tp->rx_opt.dsack = 1;
3501                 tp->duplicate_sack[0].start_seq = seq;
3502                 tp->duplicate_sack[0].end_seq = end_seq;
3503                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3504         }
3505 }
3506
3507 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3508 {
3509         if (!tp->rx_opt.dsack)
3510                 tcp_dsack_set(tp, seq, end_seq);
3511         else
3512                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3513 }
3514
3515 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3516 {
3517         struct tcp_sock *tp = tcp_sk(sk);
3518
3519         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3520             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3521                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3522                 tcp_enter_quickack_mode(sk);
3523
3524                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3525                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3526
3527                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3528                                 end_seq = tp->rcv_nxt;
3529                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3530                 }
3531         }
3532
3533         tcp_send_ack(sk);
3534 }
3535
3536 /* These routines update the SACK block as out-of-order packets arrive or
3537  * in-order packets close up the sequence space.
3538  */
3539 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3540 {
3541         int this_sack;
3542         struct tcp_sack_block *sp = &tp->selective_acks[0];
3543         struct tcp_sack_block *swalk = sp+1;
3544
3545         /* See if the recent change to the first SACK eats into
3546          * or hits the sequence space of other SACK blocks, if so coalesce.
3547          */
3548         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3549                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3550                         int i;
3551
3552                         /* Zap SWALK, by moving every further SACK up by one slot.
3553                          * Decrease num_sacks.
3554                          */
3555                         tp->rx_opt.num_sacks--;
3556                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3557                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3558                                 sp[i] = sp[i+1];
3559                         continue;
3560                 }
3561                 this_sack++, swalk++;
3562         }
3563 }
3564
3565 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3566 {
3567         __u32 tmp;
3568
3569         tmp = sack1->start_seq;
3570         sack1->start_seq = sack2->start_seq;
3571         sack2->start_seq = tmp;
3572
3573         tmp = sack1->end_seq;
3574         sack1->end_seq = sack2->end_seq;
3575         sack2->end_seq = tmp;
3576 }
3577
3578 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3579 {
3580         struct tcp_sock *tp = tcp_sk(sk);
3581         struct tcp_sack_block *sp = &tp->selective_acks[0];
3582         int cur_sacks = tp->rx_opt.num_sacks;
3583         int this_sack;
3584
3585         if (!cur_sacks)
3586                 goto new_sack;
3587
3588         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3589                 if (tcp_sack_extend(sp, seq, end_seq)) {
3590                         /* Rotate this_sack to the first one. */
3591                         for (; this_sack>0; this_sack--, sp--)
3592                                 tcp_sack_swap(sp, sp-1);
3593                         if (cur_sacks > 1)
3594                                 tcp_sack_maybe_coalesce(tp);
3595                         return;
3596                 }
3597         }
3598
3599         /* Could not find an adjacent existing SACK, build a new one,
3600          * put it at the front, and shift everyone else down.  We
3601          * always know there is at least one SACK present already here.
3602          *
3603          * If the sack array is full, forget about the last one.
3604          */
3605         if (this_sack >= 4) {
3606                 this_sack--;
3607                 tp->rx_opt.num_sacks--;
3608                 sp--;
3609         }
3610         for (; this_sack > 0; this_sack--, sp--)
3611                 *sp = *(sp-1);
3612
3613 new_sack:
3614         /* Build the new head SACK, and we're done. */
3615         sp->start_seq = seq;
3616         sp->end_seq = end_seq;
3617         tp->rx_opt.num_sacks++;
3618         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3619 }
3620
3621 /* RCV.NXT advances, some SACKs should be eaten. */
3622
3623 static void tcp_sack_remove(struct tcp_sock *tp)
3624 {
3625         struct tcp_sack_block *sp = &tp->selective_acks[0];
3626         int num_sacks = tp->rx_opt.num_sacks;
3627         int this_sack;
3628
3629         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3630         if (skb_queue_empty(&tp->out_of_order_queue)) {
3631                 tp->rx_opt.num_sacks = 0;
3632                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3633                 return;
3634         }
3635
3636         for (this_sack = 0; this_sack < num_sacks; ) {
3637                 /* Check if the start of the sack is covered by RCV.NXT. */
3638                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3639                         int i;
3640
3641                         /* RCV.NXT must cover all the block! */
3642                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3643
3644                         /* Zap this SACK, by moving forward any other SACKS. */
3645                         for (i=this_sack+1; i < num_sacks; i++)
3646                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3647                         num_sacks--;
3648                         continue;
3649                 }
3650                 this_sack++;
3651                 sp++;
3652         }
3653         if (num_sacks != tp->rx_opt.num_sacks) {
3654                 tp->rx_opt.num_sacks = num_sacks;
3655                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3656         }
3657 }
3658
3659 /* This one checks to see if we can put data from the
3660  * out_of_order queue into the receive_queue.
3661  */
3662 static void tcp_ofo_queue(struct sock *sk)
3663 {
3664         struct tcp_sock *tp = tcp_sk(sk);
3665         __u32 dsack_high = tp->rcv_nxt;
3666         struct sk_buff *skb;
3667
3668         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3669                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3670                         break;
3671
3672                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3673                         __u32 dsack = dsack_high;
3674                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3675                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3676                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3677                 }
3678
3679                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3680                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3681                         __skb_unlink(skb, &tp->out_of_order_queue);
3682                         __kfree_skb(skb);
3683                         continue;
3684                 }
3685                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3686                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3687                            TCP_SKB_CB(skb)->end_seq);
3688
3689                 __skb_unlink(skb, &tp->out_of_order_queue);
3690                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3691                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3692                 if (tcp_hdr(skb)->fin)
3693                         tcp_fin(skb, sk, tcp_hdr(skb));
3694         }
3695 }
3696
3697 static int tcp_prune_queue(struct sock *sk);
3698
3699 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3700 {
3701         struct tcphdr *th = tcp_hdr(skb);
3702         struct tcp_sock *tp = tcp_sk(sk);
3703         int eaten = -1;
3704
3705         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3706                 goto drop;
3707
3708         __skb_pull(skb, th->doff*4);
3709
3710         TCP_ECN_accept_cwr(tp, skb);
3711
3712         if (tp->rx_opt.dsack) {
3713                 tp->rx_opt.dsack = 0;
3714                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3715                                                     4 - tp->rx_opt.tstamp_ok);
3716         }
3717
3718         /*  Queue data for delivery to the user.
3719          *  Packets in sequence go to the receive queue.
3720          *  Out of sequence packets to the out_of_order_queue.
3721          */
3722         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3723                 if (tcp_receive_window(tp) == 0)
3724                         goto out_of_window;
3725
3726                 /* Ok. In sequence. In window. */
3727                 if (tp->ucopy.task == current &&
3728                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3729                     sock_owned_by_user(sk) && !tp->urg_data) {
3730                         int chunk = min_t(unsigned int, skb->len,
3731                                                         tp->ucopy.len);
3732
3733                         __set_current_state(TASK_RUNNING);
3734
3735                         local_bh_enable();
3736                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3737                                 tp->ucopy.len -= chunk;
3738                                 tp->copied_seq += chunk;
3739                                 eaten = (chunk == skb->len && !th->fin);
3740                                 tcp_rcv_space_adjust(sk);
3741                         }
3742                         local_bh_disable();
3743                 }
3744
3745                 if (eaten <= 0) {
3746 queue_and_out:
3747                         if (eaten < 0 &&
3748                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3749                              !sk_stream_rmem_schedule(sk, skb))) {
3750                                 if (tcp_prune_queue(sk) < 0 ||
3751                                     !sk_stream_rmem_schedule(sk, skb))
3752                                         goto drop;
3753                         }
3754                         sk_stream_set_owner_r(skb, sk);
3755                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3756                 }
3757                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3758                 if (skb->len)
3759                         tcp_event_data_recv(sk, skb);
3760                 if (th->fin)
3761                         tcp_fin(skb, sk, th);
3762
3763                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3764                         tcp_ofo_queue(sk);
3765
3766                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3767                          * gap in queue is filled.
3768                          */
3769                         if (skb_queue_empty(&tp->out_of_order_queue))
3770                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3771                 }
3772
3773                 if (tp->rx_opt.num_sacks)
3774                         tcp_sack_remove(tp);
3775
3776                 tcp_fast_path_check(sk);
3777
3778                 if (eaten > 0)
3779                         __kfree_skb(skb);
3780                 else if (!sock_flag(sk, SOCK_DEAD))
3781                         sk->sk_data_ready(sk, 0);
3782                 return;
3783         }
3784
3785         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3786                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3787                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3788                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3789
3790 out_of_window:
3791                 tcp_enter_quickack_mode(sk);
3792                 inet_csk_schedule_ack(sk);
3793 drop:
3794                 __kfree_skb(skb);
3795                 return;
3796         }
3797
3798         /* Out of window. F.e. zero window probe. */
3799         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3800                 goto out_of_window;
3801
3802         tcp_enter_quickack_mode(sk);
3803
3804         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3805                 /* Partial packet, seq < rcv_next < end_seq */
3806                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3807                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3808                            TCP_SKB_CB(skb)->end_seq);
3809
3810                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3811
3812                 /* If window is closed, drop tail of packet. But after
3813                  * remembering D-SACK for its head made in previous line.
3814                  */
3815                 if (!tcp_receive_window(tp))
3816                         goto out_of_window;
3817                 goto queue_and_out;
3818         }
3819
3820         TCP_ECN_check_ce(tp, skb);
3821
3822         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3823             !sk_stream_rmem_schedule(sk, skb)) {
3824                 if (tcp_prune_queue(sk) < 0 ||
3825                     !sk_stream_rmem_schedule(sk, skb))
3826                         goto drop;
3827         }
3828
3829         /* Disable header prediction. */
3830         tp->pred_flags = 0;
3831         inet_csk_schedule_ack(sk);
3832
3833         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3834                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3835
3836         sk_stream_set_owner_r(skb, sk);
3837
3838         if (!skb_peek(&tp->out_of_order_queue)) {
3839                 /* Initial out of order segment, build 1 SACK. */
3840                 if (tcp_is_sack(tp)) {
3841                         tp->rx_opt.num_sacks = 1;
3842                         tp->rx_opt.dsack     = 0;
3843                         tp->rx_opt.eff_sacks = 1;
3844                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3845                         tp->selective_acks[0].end_seq =
3846                                                 TCP_SKB_CB(skb)->end_seq;
3847                 }
3848                 __skb_queue_head(&tp->out_of_order_queue,skb);
3849         } else {
3850                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3851                 u32 seq = TCP_SKB_CB(skb)->seq;
3852                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3853
3854                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3855                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3856
3857                         if (!tp->rx_opt.num_sacks ||
3858                             tp->selective_acks[0].end_seq != seq)
3859                                 goto add_sack;
3860
3861                         /* Common case: data arrive in order after hole. */
3862                         tp->selective_acks[0].end_seq = end_seq;
3863                         return;
3864                 }
3865
3866                 /* Find place to insert this segment. */
3867                 do {
3868                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3869                                 break;
3870                 } while ((skb1 = skb1->prev) !=
3871                          (struct sk_buff*)&tp->out_of_order_queue);
3872
3873                 /* Do skb overlap to previous one? */
3874                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3875                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3876                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3877                                 /* All the bits are present. Drop. */
3878                                 __kfree_skb(skb);
3879                                 tcp_dsack_set(tp, seq, end_seq);
3880                                 goto add_sack;
3881                         }
3882                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3883                                 /* Partial overlap. */
3884                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3885                         } else {
3886                                 skb1 = skb1->prev;
3887                         }
3888                 }
3889                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3890
3891                 /* And clean segments covered by new one as whole. */
3892                 while ((skb1 = skb->next) !=
3893                        (struct sk_buff*)&tp->out_of_order_queue &&
3894                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3895                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3896                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3897                                break;
3898                        }
3899                        __skb_unlink(skb1, &tp->out_of_order_queue);
3900                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3901                        __kfree_skb(skb1);
3902                 }
3903
3904 add_sack:
3905                 if (tcp_is_sack(tp))
3906                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3907         }
3908 }
3909
3910 /* Collapse contiguous sequence of skbs head..tail with
3911  * sequence numbers start..end.
3912  * Segments with FIN/SYN are not collapsed (only because this
3913  * simplifies code)
3914  */
3915 static void
3916 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3917              struct sk_buff *head, struct sk_buff *tail,
3918              u32 start, u32 end)
3919 {
3920         struct sk_buff *skb;
3921
3922         /* First, check that queue is collapsible and find
3923          * the point where collapsing can be useful. */
3924         for (skb = head; skb != tail; ) {
3925                 /* No new bits? It is possible on ofo queue. */
3926                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3927                         struct sk_buff *next = skb->next;
3928                         __skb_unlink(skb, list);
3929                         __kfree_skb(skb);
3930                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3931                         skb = next;
3932                         continue;
3933                 }
3934
3935                 /* The first skb to collapse is:
3936                  * - not SYN/FIN and
3937                  * - bloated or contains data before "start" or
3938                  *   overlaps to the next one.
3939                  */
3940                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3941                     (tcp_win_from_space(skb->truesize) > skb->len ||
3942                      before(TCP_SKB_CB(skb)->seq, start) ||
3943                      (skb->next != tail &&
3944                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3945                         break;
3946
3947                 /* Decided to skip this, advance start seq. */
3948                 start = TCP_SKB_CB(skb)->end_seq;
3949                 skb = skb->next;
3950         }
3951         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3952                 return;
3953
3954         while (before(start, end)) {
3955                 struct sk_buff *nskb;
3956                 unsigned int header = skb_headroom(skb);
3957                 int copy = SKB_MAX_ORDER(header, 0);
3958
3959                 /* Too big header? This can happen with IPv6. */
3960                 if (copy < 0)
3961                         return;
3962                 if (end-start < copy)
3963                         copy = end-start;
3964                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3965                 if (!nskb)
3966                         return;
3967
3968                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3969                 skb_set_network_header(nskb, (skb_network_header(skb) -
3970                                               skb->head));
3971                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3972                                                 skb->head));
3973                 skb_reserve(nskb, header);
3974                 memcpy(nskb->head, skb->head, header);
3975                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3976                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3977                 __skb_insert(nskb, skb->prev, skb, list);
3978                 sk_stream_set_owner_r(nskb, sk);
3979
3980                 /* Copy data, releasing collapsed skbs. */
3981                 while (copy > 0) {
3982                         int offset = start - TCP_SKB_CB(skb)->seq;
3983                         int size = TCP_SKB_CB(skb)->end_seq - start;
3984
3985                         BUG_ON(offset < 0);
3986                         if (size > 0) {
3987                                 size = min(copy, size);
3988                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3989                                         BUG();
3990                                 TCP_SKB_CB(nskb)->end_seq += size;
3991                                 copy -= size;
3992                                 start += size;
3993                         }
3994                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3995                                 struct sk_buff *next = skb->next;
3996                                 __skb_unlink(skb, list);
3997                                 __kfree_skb(skb);
3998                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3999                                 skb = next;
4000                                 if (skb == tail ||
4001                                     tcp_hdr(skb)->syn ||
4002                                     tcp_hdr(skb)->fin)
4003                                         return;
4004                         }
4005                 }
4006         }
4007 }
4008
4009 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4010  * and tcp_collapse() them until all the queue is collapsed.
4011  */
4012 static void tcp_collapse_ofo_queue(struct sock *sk)
4013 {
4014         struct tcp_sock *tp = tcp_sk(sk);
4015         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4016         struct sk_buff *head;
4017         u32 start, end;
4018
4019         if (skb == NULL)
4020                 return;
4021
4022         start = TCP_SKB_CB(skb)->seq;
4023         end = TCP_SKB_CB(skb)->end_seq;
4024         head = skb;
4025
4026         for (;;) {
4027                 skb = skb->next;
4028
4029                 /* Segment is terminated when we see gap or when
4030                  * we are at the end of all the queue. */
4031                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4032                     after(TCP_SKB_CB(skb)->seq, end) ||
4033                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4034                         tcp_collapse(sk, &tp->out_of_order_queue,
4035                                      head, skb, start, end);
4036                         head = skb;
4037                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4038                                 break;
4039                         /* Start new segment */
4040                         start = TCP_SKB_CB(skb)->seq;
4041                         end = TCP_SKB_CB(skb)->end_seq;
4042                 } else {
4043                         if (before(TCP_SKB_CB(skb)->seq, start))
4044                                 start = TCP_SKB_CB(skb)->seq;
4045                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4046                                 end = TCP_SKB_CB(skb)->end_seq;
4047                 }
4048         }
4049 }
4050
4051 /* Reduce allocated memory if we can, trying to get
4052  * the socket within its memory limits again.
4053  *
4054  * Return less than zero if we should start dropping frames
4055  * until the socket owning process reads some of the data
4056  * to stabilize the situation.
4057  */
4058 static int tcp_prune_queue(struct sock *sk)
4059 {
4060         struct tcp_sock *tp = tcp_sk(sk);
4061
4062         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4063
4064         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
4065
4066         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4067                 tcp_clamp_window(sk);
4068         else if (tcp_memory_pressure)
4069                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4070
4071         tcp_collapse_ofo_queue(sk);
4072         tcp_collapse(sk, &sk->sk_receive_queue,
4073                      sk->sk_receive_queue.next,
4074                      (struct sk_buff*)&sk->sk_receive_queue,
4075                      tp->copied_seq, tp->rcv_nxt);
4076         sk_stream_mem_reclaim(sk);
4077
4078         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4079                 return 0;
4080
4081         /* Collapsing did not help, destructive actions follow.
4082          * This must not ever occur. */
4083
4084         /* First, purge the out_of_order queue. */
4085         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4086                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4087                 __skb_queue_purge(&tp->out_of_order_queue);
4088
4089                 /* Reset SACK state.  A conforming SACK implementation will
4090                  * do the same at a timeout based retransmit.  When a connection
4091                  * is in a sad state like this, we care only about integrity
4092                  * of the connection not performance.
4093                  */
4094                 if (tcp_is_sack(tp))
4095                         tcp_sack_reset(&tp->rx_opt);
4096                 sk_stream_mem_reclaim(sk);
4097         }
4098
4099         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4100                 return 0;
4101
4102         /* If we are really being abused, tell the caller to silently
4103          * drop receive data on the floor.  It will get retransmitted
4104          * and hopefully then we'll have sufficient space.
4105          */
4106         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4107
4108         /* Massive buffer overcommit. */
4109         tp->pred_flags = 0;
4110         return -1;
4111 }
4112
4113
4114 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4115  * As additional protections, we do not touch cwnd in retransmission phases,
4116  * and if application hit its sndbuf limit recently.
4117  */
4118 void tcp_cwnd_application_limited(struct sock *sk)
4119 {
4120         struct tcp_sock *tp = tcp_sk(sk);
4121
4122         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4123             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4124                 /* Limited by application or receiver window. */
4125                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4126                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4127                 if (win_used < tp->snd_cwnd) {
4128                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4129                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4130                 }
4131                 tp->snd_cwnd_used = 0;
4132         }
4133         tp->snd_cwnd_stamp = tcp_time_stamp;
4134 }
4135
4136 static int tcp_should_expand_sndbuf(struct sock *sk)
4137 {
4138         struct tcp_sock *tp = tcp_sk(sk);
4139
4140         /* If the user specified a specific send buffer setting, do
4141          * not modify it.
4142          */
4143         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4144                 return 0;
4145
4146         /* If we are under global TCP memory pressure, do not expand.  */
4147         if (tcp_memory_pressure)
4148                 return 0;
4149
4150         /* If we are under soft global TCP memory pressure, do not expand.  */
4151         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4152                 return 0;
4153
4154         /* If we filled the congestion window, do not expand.  */
4155         if (tp->packets_out >= tp->snd_cwnd)
4156                 return 0;
4157
4158         return 1;
4159 }
4160
4161 /* When incoming ACK allowed to free some skb from write_queue,
4162  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4163  * on the exit from tcp input handler.
4164  *
4165  * PROBLEM: sndbuf expansion does not work well with largesend.
4166  */
4167 static void tcp_new_space(struct sock *sk)
4168 {
4169         struct tcp_sock *tp = tcp_sk(sk);
4170
4171         if (tcp_should_expand_sndbuf(sk)) {
4172                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4173                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4174                     demanded = max_t(unsigned int, tp->snd_cwnd,
4175                                                    tp->reordering + 1);
4176                 sndmem *= 2*demanded;
4177                 if (sndmem > sk->sk_sndbuf)
4178                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4179                 tp->snd_cwnd_stamp = tcp_time_stamp;
4180         }
4181
4182         sk->sk_write_space(sk);
4183 }
4184
4185 static void tcp_check_space(struct sock *sk)
4186 {
4187         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4188                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4189                 if (sk->sk_socket &&
4190                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4191                         tcp_new_space(sk);
4192         }
4193 }
4194
4195 static inline void tcp_data_snd_check(struct sock *sk)
4196 {
4197         tcp_push_pending_frames(sk);
4198         tcp_check_space(sk);
4199 }
4200
4201 /*
4202  * Check if sending an ack is needed.
4203  */
4204 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4205 {
4206         struct tcp_sock *tp = tcp_sk(sk);
4207
4208             /* More than one full frame received... */
4209         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4210              /* ... and right edge of window advances far enough.
4211               * (tcp_recvmsg() will send ACK otherwise). Or...
4212               */
4213              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4214             /* We ACK each frame or... */
4215             tcp_in_quickack_mode(sk) ||
4216             /* We have out of order data. */
4217             (ofo_possible &&
4218              skb_peek(&tp->out_of_order_queue))) {
4219                 /* Then ack it now */
4220                 tcp_send_ack(sk);
4221         } else {
4222                 /* Else, send delayed ack. */
4223                 tcp_send_delayed_ack(sk);
4224         }
4225 }
4226
4227 static inline void tcp_ack_snd_check(struct sock *sk)
4228 {
4229         if (!inet_csk_ack_scheduled(sk)) {
4230                 /* We sent a data segment already. */
4231                 return;
4232         }
4233         __tcp_ack_snd_check(sk, 1);
4234 }
4235
4236 /*
4237  *      This routine is only called when we have urgent data
4238  *      signaled. Its the 'slow' part of tcp_urg. It could be
4239  *      moved inline now as tcp_urg is only called from one
4240  *      place. We handle URGent data wrong. We have to - as
4241  *      BSD still doesn't use the correction from RFC961.
4242  *      For 1003.1g we should support a new option TCP_STDURG to permit
4243  *      either form (or just set the sysctl tcp_stdurg).
4244  */
4245
4246 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4247 {
4248         struct tcp_sock *tp = tcp_sk(sk);
4249         u32 ptr = ntohs(th->urg_ptr);
4250
4251         if (ptr && !sysctl_tcp_stdurg)
4252                 ptr--;
4253         ptr += ntohl(th->seq);
4254
4255         /* Ignore urgent data that we've already seen and read. */
4256         if (after(tp->copied_seq, ptr))
4257                 return;
4258
4259         /* Do not replay urg ptr.
4260          *
4261          * NOTE: interesting situation not covered by specs.
4262          * Misbehaving sender may send urg ptr, pointing to segment,
4263          * which we already have in ofo queue. We are not able to fetch
4264          * such data and will stay in TCP_URG_NOTYET until will be eaten
4265          * by recvmsg(). Seems, we are not obliged to handle such wicked
4266          * situations. But it is worth to think about possibility of some
4267          * DoSes using some hypothetical application level deadlock.
4268          */
4269         if (before(ptr, tp->rcv_nxt))
4270                 return;
4271
4272         /* Do we already have a newer (or duplicate) urgent pointer? */
4273         if (tp->urg_data && !after(ptr, tp->urg_seq))
4274                 return;
4275
4276         /* Tell the world about our new urgent pointer. */
4277         sk_send_sigurg(sk);
4278
4279         /* We may be adding urgent data when the last byte read was
4280          * urgent. To do this requires some care. We cannot just ignore
4281          * tp->copied_seq since we would read the last urgent byte again
4282          * as data, nor can we alter copied_seq until this data arrives
4283          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4284          *
4285          * NOTE. Double Dutch. Rendering to plain English: author of comment
4286          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4287          * and expect that both A and B disappear from stream. This is _wrong_.
4288          * Though this happens in BSD with high probability, this is occasional.
4289          * Any application relying on this is buggy. Note also, that fix "works"
4290          * only in this artificial test. Insert some normal data between A and B and we will
4291          * decline of BSD again. Verdict: it is better to remove to trap
4292          * buggy users.
4293          */
4294         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4295             !sock_flag(sk, SOCK_URGINLINE) &&
4296             tp->copied_seq != tp->rcv_nxt) {
4297                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4298                 tp->copied_seq++;
4299                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4300                         __skb_unlink(skb, &sk->sk_receive_queue);
4301                         __kfree_skb(skb);
4302                 }
4303         }
4304
4305         tp->urg_data   = TCP_URG_NOTYET;
4306         tp->urg_seq    = ptr;
4307
4308         /* Disable header prediction. */
4309         tp->pred_flags = 0;
4310 }
4311
4312 /* This is the 'fast' part of urgent handling. */
4313 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4314 {
4315         struct tcp_sock *tp = tcp_sk(sk);
4316
4317         /* Check if we get a new urgent pointer - normally not. */
4318         if (th->urg)
4319                 tcp_check_urg(sk,th);
4320
4321         /* Do we wait for any urgent data? - normally not... */
4322         if (tp->urg_data == TCP_URG_NOTYET) {
4323                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4324                           th->syn;
4325
4326                 /* Is the urgent pointer pointing into this packet? */
4327                 if (ptr < skb->len) {
4328                         u8 tmp;
4329                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4330                                 BUG();
4331                         tp->urg_data = TCP_URG_VALID | tmp;
4332                         if (!sock_flag(sk, SOCK_DEAD))
4333                                 sk->sk_data_ready(sk, 0);
4334                 }
4335         }
4336 }
4337
4338 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4339 {
4340         struct tcp_sock *tp = tcp_sk(sk);
4341         int chunk = skb->len - hlen;
4342         int err;
4343
4344         local_bh_enable();
4345         if (skb_csum_unnecessary(skb))
4346                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4347         else
4348                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4349                                                        tp->ucopy.iov);
4350
4351         if (!err) {
4352                 tp->ucopy.len -= chunk;
4353                 tp->copied_seq += chunk;
4354                 tcp_rcv_space_adjust(sk);
4355         }
4356
4357         local_bh_disable();
4358         return err;
4359 }
4360
4361 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4362 {
4363         __sum16 result;
4364
4365         if (sock_owned_by_user(sk)) {
4366                 local_bh_enable();
4367                 result = __tcp_checksum_complete(skb);
4368                 local_bh_disable();
4369         } else {
4370                 result = __tcp_checksum_complete(skb);
4371         }
4372         return result;
4373 }
4374
4375 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4376 {
4377         return !skb_csum_unnecessary(skb) &&
4378                 __tcp_checksum_complete_user(sk, skb);
4379 }
4380
4381 #ifdef CONFIG_NET_DMA
4382 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4383 {
4384         struct tcp_sock *tp = tcp_sk(sk);
4385         int chunk = skb->len - hlen;
4386         int dma_cookie;
4387         int copied_early = 0;
4388
4389         if (tp->ucopy.wakeup)
4390                 return 0;
4391
4392         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4393                 tp->ucopy.dma_chan = get_softnet_dma();
4394
4395         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4396
4397                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4398                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4399
4400                 if (dma_cookie < 0)
4401                         goto out;
4402
4403                 tp->ucopy.dma_cookie = dma_cookie;
4404                 copied_early = 1;
4405
4406                 tp->ucopy.len -= chunk;
4407                 tp->copied_seq += chunk;
4408                 tcp_rcv_space_adjust(sk);
4409
4410                 if ((tp->ucopy.len == 0) ||
4411                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4412                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4413                         tp->ucopy.wakeup = 1;
4414                         sk->sk_data_ready(sk, 0);
4415                 }
4416         } else if (chunk > 0) {
4417                 tp->ucopy.wakeup = 1;
4418                 sk->sk_data_ready(sk, 0);
4419         }
4420 out:
4421         return copied_early;
4422 }
4423 #endif /* CONFIG_NET_DMA */
4424
4425 /*
4426  *      TCP receive function for the ESTABLISHED state.
4427  *
4428  *      It is split into a fast path and a slow path. The fast path is
4429  *      disabled when:
4430  *      - A zero window was announced from us - zero window probing
4431  *        is only handled properly in the slow path.
4432  *      - Out of order segments arrived.
4433  *      - Urgent data is expected.
4434  *      - There is no buffer space left
4435  *      - Unexpected TCP flags/window values/header lengths are received
4436  *        (detected by checking the TCP header against pred_flags)
4437  *      - Data is sent in both directions. Fast path only supports pure senders
4438  *        or pure receivers (this means either the sequence number or the ack
4439  *        value must stay constant)
4440  *      - Unexpected TCP option.
4441  *
4442  *      When these conditions are not satisfied it drops into a standard
4443  *      receive procedure patterned after RFC793 to handle all cases.
4444  *      The first three cases are guaranteed by proper pred_flags setting,
4445  *      the rest is checked inline. Fast processing is turned on in
4446  *      tcp_data_queue when everything is OK.
4447  */
4448 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4449                         struct tcphdr *th, unsigned len)
4450 {
4451         struct tcp_sock *tp = tcp_sk(sk);
4452
4453         /*
4454          *      Header prediction.
4455          *      The code loosely follows the one in the famous
4456          *      "30 instruction TCP receive" Van Jacobson mail.
4457          *
4458          *      Van's trick is to deposit buffers into socket queue
4459          *      on a device interrupt, to call tcp_recv function
4460          *      on the receive process context and checksum and copy
4461          *      the buffer to user space. smart...
4462          *
4463          *      Our current scheme is not silly either but we take the
4464          *      extra cost of the net_bh soft interrupt processing...
4465          *      We do checksum and copy also but from device to kernel.
4466          */
4467
4468         tp->rx_opt.saw_tstamp = 0;
4469
4470         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4471          *      if header_prediction is to be made
4472          *      'S' will always be tp->tcp_header_len >> 2
4473          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4474          *  turn it off (when there are holes in the receive
4475          *       space for instance)
4476          *      PSH flag is ignored.
4477          */
4478
4479         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4480                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4481                 int tcp_header_len = tp->tcp_header_len;
4482
4483                 /* Timestamp header prediction: tcp_header_len
4484                  * is automatically equal to th->doff*4 due to pred_flags
4485                  * match.
4486                  */
4487
4488                 /* Check timestamp */
4489                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4490                         __be32 *ptr = (__be32 *)(th + 1);
4491
4492                         /* No? Slow path! */
4493                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4494                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4495                                 goto slow_path;
4496
4497                         tp->rx_opt.saw_tstamp = 1;
4498                         ++ptr;
4499                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4500                         ++ptr;
4501                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4502
4503                         /* If PAWS failed, check it more carefully in slow path */
4504                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4505                                 goto slow_path;
4506
4507                         /* DO NOT update ts_recent here, if checksum fails
4508                          * and timestamp was corrupted part, it will result
4509                          * in a hung connection since we will drop all
4510                          * future packets due to the PAWS test.
4511                          */
4512                 }
4513
4514                 if (len <= tcp_header_len) {
4515                         /* Bulk data transfer: sender */
4516                         if (len == tcp_header_len) {
4517                                 /* Predicted packet is in window by definition.
4518                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4519                                  * Hence, check seq<=rcv_wup reduces to:
4520                                  */
4521                                 if (tcp_header_len ==
4522                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4523                                     tp->rcv_nxt == tp->rcv_wup)
4524                                         tcp_store_ts_recent(tp);
4525
4526                                 /* We know that such packets are checksummed
4527                                  * on entry.
4528                                  */
4529                                 tcp_ack(sk, skb, 0);
4530                                 __kfree_skb(skb);
4531                                 tcp_data_snd_check(sk);
4532                                 return 0;
4533                         } else { /* Header too small */
4534                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4535                                 goto discard;
4536                         }
4537                 } else {
4538                         int eaten = 0;
4539                         int copied_early = 0;
4540
4541                         if (tp->copied_seq == tp->rcv_nxt &&
4542                             len - tcp_header_len <= tp->ucopy.len) {
4543 #ifdef CONFIG_NET_DMA
4544                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4545                                         copied_early = 1;
4546                                         eaten = 1;
4547                                 }
4548 #endif
4549                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4550                                         __set_current_state(TASK_RUNNING);
4551
4552                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4553                                                 eaten = 1;
4554                                 }
4555                                 if (eaten) {
4556                                         /* Predicted packet is in window by definition.
4557                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4558                                          * Hence, check seq<=rcv_wup reduces to:
4559                                          */
4560                                         if (tcp_header_len ==
4561                                             (sizeof(struct tcphdr) +
4562                                              TCPOLEN_TSTAMP_ALIGNED) &&
4563                                             tp->rcv_nxt == tp->rcv_wup)
4564                                                 tcp_store_ts_recent(tp);
4565
4566                                         tcp_rcv_rtt_measure_ts(sk, skb);
4567
4568                                         __skb_pull(skb, tcp_header_len);
4569                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4570                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4571                                 }
4572                                 if (copied_early)
4573                                         tcp_cleanup_rbuf(sk, skb->len);
4574                         }
4575                         if (!eaten) {
4576                                 if (tcp_checksum_complete_user(sk, skb))
4577                                         goto csum_error;
4578
4579                                 /* Predicted packet is in window by definition.
4580                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4581                                  * Hence, check seq<=rcv_wup reduces to:
4582                                  */
4583                                 if (tcp_header_len ==
4584                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4585                                     tp->rcv_nxt == tp->rcv_wup)
4586                                         tcp_store_ts_recent(tp);
4587
4588                                 tcp_rcv_rtt_measure_ts(sk, skb);
4589
4590                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4591                                         goto step5;
4592
4593                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4594
4595                                 /* Bulk data transfer: receiver */
4596                                 __skb_pull(skb,tcp_header_len);
4597                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4598                                 sk_stream_set_owner_r(skb, sk);
4599                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4600                         }
4601
4602                         tcp_event_data_recv(sk, skb);
4603
4604                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4605                                 /* Well, only one small jumplet in fast path... */
4606                                 tcp_ack(sk, skb, FLAG_DATA);
4607                                 tcp_data_snd_check(sk);
4608                                 if (!inet_csk_ack_scheduled(sk))
4609                                         goto no_ack;
4610                         }
4611
4612                         __tcp_ack_snd_check(sk, 0);
4613 no_ack:
4614 #ifdef CONFIG_NET_DMA
4615                         if (copied_early)
4616                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4617                         else
4618 #endif
4619                         if (eaten)
4620                                 __kfree_skb(skb);
4621                         else
4622                                 sk->sk_data_ready(sk, 0);
4623                         return 0;
4624                 }
4625         }
4626
4627 slow_path:
4628         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4629                 goto csum_error;
4630
4631         /*
4632          * RFC1323: H1. Apply PAWS check first.
4633          */
4634         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4635             tcp_paws_discard(sk, skb)) {
4636                 if (!th->rst) {
4637                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4638                         tcp_send_dupack(sk, skb);
4639                         goto discard;
4640                 }
4641                 /* Resets are accepted even if PAWS failed.
4642
4643                    ts_recent update must be made after we are sure
4644                    that the packet is in window.
4645                  */
4646         }
4647
4648         /*
4649          *      Standard slow path.
4650          */
4651
4652         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4653                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4654                  * (RST) segments are validated by checking their SEQ-fields."
4655                  * And page 69: "If an incoming segment is not acceptable,
4656                  * an acknowledgment should be sent in reply (unless the RST bit
4657                  * is set, if so drop the segment and return)".
4658                  */
4659                 if (!th->rst)
4660                         tcp_send_dupack(sk, skb);
4661                 goto discard;
4662         }
4663
4664         if (th->rst) {
4665                 tcp_reset(sk);
4666                 goto discard;
4667         }
4668
4669         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4670
4671         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4672                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4673                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4674                 tcp_reset(sk);
4675                 return 1;
4676         }
4677
4678 step5:
4679         if (th->ack)
4680                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4681
4682         tcp_rcv_rtt_measure_ts(sk, skb);
4683
4684         /* Process urgent data. */
4685         tcp_urg(sk, skb, th);
4686
4687         /* step 7: process the segment text */
4688         tcp_data_queue(sk, skb);
4689
4690         tcp_data_snd_check(sk);
4691         tcp_ack_snd_check(sk);
4692         return 0;
4693
4694 csum_error:
4695         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4696
4697 discard:
4698         __kfree_skb(skb);
4699         return 0;
4700 }
4701
4702 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4703                                          struct tcphdr *th, unsigned len)
4704 {
4705         struct tcp_sock *tp = tcp_sk(sk);
4706         struct inet_connection_sock *icsk = inet_csk(sk);
4707         int saved_clamp = tp->rx_opt.mss_clamp;
4708
4709         tcp_parse_options(skb, &tp->rx_opt, 0);
4710
4711         if (th->ack) {
4712                 /* rfc793:
4713                  * "If the state is SYN-SENT then
4714                  *    first check the ACK bit
4715                  *      If the ACK bit is set
4716                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4717                  *        a reset (unless the RST bit is set, if so drop
4718                  *        the segment and return)"
4719                  *
4720                  *  We do not send data with SYN, so that RFC-correct
4721                  *  test reduces to:
4722                  */
4723                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4724                         goto reset_and_undo;
4725
4726                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4727                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4728                              tcp_time_stamp)) {
4729                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4730                         goto reset_and_undo;
4731                 }
4732
4733                 /* Now ACK is acceptable.
4734                  *
4735                  * "If the RST bit is set
4736                  *    If the ACK was acceptable then signal the user "error:
4737                  *    connection reset", drop the segment, enter CLOSED state,
4738                  *    delete TCB, and return."
4739                  */
4740
4741                 if (th->rst) {
4742                         tcp_reset(sk);
4743                         goto discard;
4744                 }
4745
4746                 /* rfc793:
4747                  *   "fifth, if neither of the SYN or RST bits is set then
4748                  *    drop the segment and return."
4749                  *
4750                  *    See note below!
4751                  *                                        --ANK(990513)
4752                  */
4753                 if (!th->syn)
4754                         goto discard_and_undo;
4755
4756                 /* rfc793:
4757                  *   "If the SYN bit is on ...
4758                  *    are acceptable then ...
4759                  *    (our SYN has been ACKed), change the connection
4760                  *    state to ESTABLISHED..."
4761                  */
4762
4763                 TCP_ECN_rcv_synack(tp, th);
4764
4765                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4766                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4767
4768                 /* Ok.. it's good. Set up sequence numbers and
4769                  * move to established.
4770                  */
4771                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4772                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4773
4774                 /* RFC1323: The window in SYN & SYN/ACK segments is
4775                  * never scaled.
4776                  */
4777                 tp->snd_wnd = ntohs(th->window);
4778                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4779
4780                 if (!tp->rx_opt.wscale_ok) {
4781                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4782                         tp->window_clamp = min(tp->window_clamp, 65535U);
4783                 }
4784
4785                 if (tp->rx_opt.saw_tstamp) {
4786                         tp->rx_opt.tstamp_ok       = 1;
4787                         tp->tcp_header_len =
4788                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4789                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4790                         tcp_store_ts_recent(tp);
4791                 } else {
4792                         tp->tcp_header_len = sizeof(struct tcphdr);
4793                 }
4794
4795                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4796                         tcp_enable_fack(tp);
4797
4798                 tcp_mtup_init(sk);
4799                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4800                 tcp_initialize_rcv_mss(sk);
4801
4802                 /* Remember, tcp_poll() does not lock socket!
4803                  * Change state from SYN-SENT only after copied_seq
4804                  * is initialized. */
4805                 tp->copied_seq = tp->rcv_nxt;
4806                 smp_mb();
4807                 tcp_set_state(sk, TCP_ESTABLISHED);
4808
4809                 security_inet_conn_established(sk, skb);
4810
4811                 /* Make sure socket is routed, for correct metrics.  */
4812                 icsk->icsk_af_ops->rebuild_header(sk);
4813
4814                 tcp_init_metrics(sk);
4815
4816                 tcp_init_congestion_control(sk);
4817
4818                 /* Prevent spurious tcp_cwnd_restart() on first data
4819                  * packet.
4820                  */
4821                 tp->lsndtime = tcp_time_stamp;
4822
4823                 tcp_init_buffer_space(sk);
4824
4825                 if (sock_flag(sk, SOCK_KEEPOPEN))
4826                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4827
4828                 if (!tp->rx_opt.snd_wscale)
4829                         __tcp_fast_path_on(tp, tp->snd_wnd);
4830                 else
4831                         tp->pred_flags = 0;
4832
4833                 if (!sock_flag(sk, SOCK_DEAD)) {
4834                         sk->sk_state_change(sk);
4835                         sk_wake_async(sk, 0, POLL_OUT);
4836                 }
4837
4838                 if (sk->sk_write_pending ||
4839                     icsk->icsk_accept_queue.rskq_defer_accept ||
4840                     icsk->icsk_ack.pingpong) {
4841                         /* Save one ACK. Data will be ready after
4842                          * several ticks, if write_pending is set.
4843                          *
4844                          * It may be deleted, but with this feature tcpdumps
4845                          * look so _wonderfully_ clever, that I was not able
4846                          * to stand against the temptation 8)     --ANK
4847                          */
4848                         inet_csk_schedule_ack(sk);
4849                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4850                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4851                         tcp_incr_quickack(sk);
4852                         tcp_enter_quickack_mode(sk);
4853                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4854                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4855
4856 discard:
4857                         __kfree_skb(skb);
4858                         return 0;
4859                 } else {
4860                         tcp_send_ack(sk);
4861                 }
4862                 return -1;
4863         }
4864
4865         /* No ACK in the segment */
4866
4867         if (th->rst) {
4868                 /* rfc793:
4869                  * "If the RST bit is set
4870                  *
4871                  *      Otherwise (no ACK) drop the segment and return."
4872                  */
4873
4874                 goto discard_and_undo;
4875         }
4876
4877         /* PAWS check. */
4878         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4879                 goto discard_and_undo;
4880
4881         if (th->syn) {
4882                 /* We see SYN without ACK. It is attempt of
4883                  * simultaneous connect with crossed SYNs.
4884                  * Particularly, it can be connect to self.
4885                  */
4886                 tcp_set_state(sk, TCP_SYN_RECV);
4887
4888                 if (tp->rx_opt.saw_tstamp) {
4889                         tp->rx_opt.tstamp_ok = 1;
4890                         tcp_store_ts_recent(tp);
4891                         tp->tcp_header_len =
4892                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4893                 } else {
4894                         tp->tcp_header_len = sizeof(struct tcphdr);
4895                 }
4896
4897                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4898                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4899
4900                 /* RFC1323: The window in SYN & SYN/ACK segments is
4901                  * never scaled.
4902                  */
4903                 tp->snd_wnd    = ntohs(th->window);
4904                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4905                 tp->max_window = tp->snd_wnd;
4906
4907                 TCP_ECN_rcv_syn(tp, th);
4908
4909                 tcp_mtup_init(sk);
4910                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4911                 tcp_initialize_rcv_mss(sk);
4912
4913
4914                 tcp_send_synack(sk);
4915 #if 0
4916                 /* Note, we could accept data and URG from this segment.
4917                  * There are no obstacles to make this.
4918                  *
4919                  * However, if we ignore data in ACKless segments sometimes,
4920                  * we have no reasons to accept it sometimes.
4921                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4922                  * is not flawless. So, discard packet for sanity.
4923                  * Uncomment this return to process the data.
4924                  */
4925                 return -1;
4926 #else
4927                 goto discard;
4928 #endif
4929         }
4930         /* "fifth, if neither of the SYN or RST bits is set then
4931          * drop the segment and return."
4932          */
4933
4934 discard_and_undo:
4935         tcp_clear_options(&tp->rx_opt);
4936         tp->rx_opt.mss_clamp = saved_clamp;
4937         goto discard;
4938
4939 reset_and_undo:
4940         tcp_clear_options(&tp->rx_opt);
4941         tp->rx_opt.mss_clamp = saved_clamp;
4942         return 1;
4943 }
4944
4945
4946 /*
4947  *      This function implements the receiving procedure of RFC 793 for
4948  *      all states except ESTABLISHED and TIME_WAIT.
4949  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4950  *      address independent.
4951  */
4952
4953 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4954                           struct tcphdr *th, unsigned len)
4955 {
4956         struct tcp_sock *tp = tcp_sk(sk);
4957         struct inet_connection_sock *icsk = inet_csk(sk);
4958         int queued = 0;
4959
4960         tp->rx_opt.saw_tstamp = 0;
4961
4962         switch (sk->sk_state) {
4963         case TCP_CLOSE:
4964                 goto discard;
4965
4966         case TCP_LISTEN:
4967                 if (th->ack)
4968                         return 1;
4969
4970                 if (th->rst)
4971                         goto discard;
4972
4973                 if (th->syn) {
4974                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4975                                 return 1;
4976
4977                         /* Now we have several options: In theory there is
4978                          * nothing else in the frame. KA9Q has an option to
4979                          * send data with the syn, BSD accepts data with the
4980                          * syn up to the [to be] advertised window and
4981                          * Solaris 2.1 gives you a protocol error. For now
4982                          * we just ignore it, that fits the spec precisely
4983                          * and avoids incompatibilities. It would be nice in
4984                          * future to drop through and process the data.
4985                          *
4986                          * Now that TTCP is starting to be used we ought to
4987                          * queue this data.
4988                          * But, this leaves one open to an easy denial of
4989                          * service attack, and SYN cookies can't defend
4990                          * against this problem. So, we drop the data
4991                          * in the interest of security over speed unless
4992                          * it's still in use.
4993                          */
4994                         kfree_skb(skb);
4995                         return 0;
4996                 }
4997                 goto discard;
4998
4999         case TCP_SYN_SENT:
5000                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5001                 if (queued >= 0)
5002                         return queued;
5003
5004                 /* Do step6 onward by hand. */
5005                 tcp_urg(sk, skb, th);
5006                 __kfree_skb(skb);
5007                 tcp_data_snd_check(sk);
5008                 return 0;
5009         }
5010
5011         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5012             tcp_paws_discard(sk, skb)) {
5013                 if (!th->rst) {
5014                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
5015                         tcp_send_dupack(sk, skb);
5016                         goto discard;
5017                 }
5018                 /* Reset is accepted even if it did not pass PAWS. */
5019         }
5020
5021         /* step 1: check sequence number */
5022         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5023                 if (!th->rst)
5024                         tcp_send_dupack(sk, skb);
5025                 goto discard;
5026         }
5027
5028         /* step 2: check RST bit */
5029         if (th->rst) {
5030                 tcp_reset(sk);
5031                 goto discard;
5032         }
5033
5034         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5035
5036         /* step 3: check security and precedence [ignored] */
5037
5038         /*      step 4:
5039          *
5040          *      Check for a SYN in window.
5041          */
5042         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5043                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
5044                 tcp_reset(sk);
5045                 return 1;
5046         }
5047
5048         /* step 5: check the ACK field */
5049         if (th->ack) {
5050                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5051
5052                 switch (sk->sk_state) {
5053                 case TCP_SYN_RECV:
5054                         if (acceptable) {
5055                                 tp->copied_seq = tp->rcv_nxt;
5056                                 smp_mb();
5057                                 tcp_set_state(sk, TCP_ESTABLISHED);
5058                                 sk->sk_state_change(sk);
5059
5060                                 /* Note, that this wakeup is only for marginal
5061                                  * crossed SYN case. Passively open sockets
5062                                  * are not waked up, because sk->sk_sleep ==
5063                                  * NULL and sk->sk_socket == NULL.
5064                                  */
5065                                 if (sk->sk_socket) {
5066                                         sk_wake_async(sk,0,POLL_OUT);
5067                                 }
5068
5069                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5070                                 tp->snd_wnd = ntohs(th->window) <<
5071                                               tp->rx_opt.snd_wscale;
5072                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5073                                             TCP_SKB_CB(skb)->seq);
5074
5075                                 /* tcp_ack considers this ACK as duplicate
5076                                  * and does not calculate rtt.
5077                                  * Fix it at least with timestamps.
5078                                  */
5079                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5080                                     !tp->srtt)
5081                                         tcp_ack_saw_tstamp(sk, 0);
5082
5083                                 if (tp->rx_opt.tstamp_ok)
5084                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5085
5086                                 /* Make sure socket is routed, for
5087                                  * correct metrics.
5088                                  */
5089                                 icsk->icsk_af_ops->rebuild_header(sk);
5090
5091                                 tcp_init_metrics(sk);
5092
5093                                 tcp_init_congestion_control(sk);
5094
5095                                 /* Prevent spurious tcp_cwnd_restart() on
5096                                  * first data packet.
5097                                  */
5098                                 tp->lsndtime = tcp_time_stamp;
5099
5100                                 tcp_mtup_init(sk);
5101                                 tcp_initialize_rcv_mss(sk);
5102                                 tcp_init_buffer_space(sk);
5103                                 tcp_fast_path_on(tp);
5104                         } else {
5105                                 return 1;
5106                         }
5107                         break;
5108
5109                 case TCP_FIN_WAIT1:
5110                         if (tp->snd_una == tp->write_seq) {
5111                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5112                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5113                                 dst_confirm(sk->sk_dst_cache);
5114
5115                                 if (!sock_flag(sk, SOCK_DEAD))
5116                                         /* Wake up lingering close() */
5117                                         sk->sk_state_change(sk);
5118                                 else {
5119                                         int tmo;
5120
5121                                         if (tp->linger2 < 0 ||
5122                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5123                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5124                                                 tcp_done(sk);
5125                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5126                                                 return 1;
5127                                         }
5128
5129                                         tmo = tcp_fin_time(sk);
5130                                         if (tmo > TCP_TIMEWAIT_LEN) {
5131                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5132                                         } else if (th->fin || sock_owned_by_user(sk)) {
5133                                                 /* Bad case. We could lose such FIN otherwise.
5134                                                  * It is not a big problem, but it looks confusing
5135                                                  * and not so rare event. We still can lose it now,
5136                                                  * if it spins in bh_lock_sock(), but it is really
5137                                                  * marginal case.
5138                                                  */
5139                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5140                                         } else {
5141                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5142                                                 goto discard;
5143                                         }
5144                                 }
5145                         }
5146                         break;
5147
5148                 case TCP_CLOSING:
5149                         if (tp->snd_una == tp->write_seq) {
5150                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5151                                 goto discard;
5152                         }
5153                         break;
5154
5155                 case TCP_LAST_ACK:
5156                         if (tp->snd_una == tp->write_seq) {
5157                                 tcp_update_metrics(sk);
5158                                 tcp_done(sk);
5159                                 goto discard;
5160                         }
5161                         break;
5162                 }
5163         } else
5164                 goto discard;
5165
5166         /* step 6: check the URG bit */
5167         tcp_urg(sk, skb, th);
5168
5169         /* step 7: process the segment text */
5170         switch (sk->sk_state) {
5171         case TCP_CLOSE_WAIT:
5172         case TCP_CLOSING:
5173         case TCP_LAST_ACK:
5174                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5175                         break;
5176         case TCP_FIN_WAIT1:
5177         case TCP_FIN_WAIT2:
5178                 /* RFC 793 says to queue data in these states,
5179                  * RFC 1122 says we MUST send a reset.
5180                  * BSD 4.4 also does reset.
5181                  */
5182                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5183                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5184                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5185                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5186                                 tcp_reset(sk);
5187                                 return 1;
5188                         }
5189                 }
5190                 /* Fall through */
5191         case TCP_ESTABLISHED:
5192                 tcp_data_queue(sk, skb);
5193                 queued = 1;
5194                 break;
5195         }
5196
5197         /* tcp_data could move socket to TIME-WAIT */
5198         if (sk->sk_state != TCP_CLOSE) {
5199                 tcp_data_snd_check(sk);
5200                 tcp_ack_snd_check(sk);
5201         }
5202
5203         if (!queued) {
5204 discard:
5205                 __kfree_skb(skb);
5206         }
5207         return 0;
5208 }
5209
5210 EXPORT_SYMBOL(sysctl_tcp_ecn);
5211 EXPORT_SYMBOL(sysctl_tcp_reordering);
5212 EXPORT_SYMBOL(tcp_parse_options);
5213 EXPORT_SYMBOL(tcp_rcv_established);
5214 EXPORT_SYMBOL(tcp_rcv_state_process);
5215 EXPORT_SYMBOL(tcp_initialize_rcv_mss);