[TCP]: Avoid clearing sacktag hint in trivial situations
[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 DSACK 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 DSACKs */
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 (reports
1031  * SACK reneging when it should advance SND.UNA).
1032  *
1033  * Implements also blockage to start_seq wrap-around. Problem lies in the
1034  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1035  * there's no guarantee that it will be before snd_nxt (n). The problem
1036  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1037  * wrap (s_w):
1038  *
1039  *         <- outs wnd ->                          <- wrapzone ->
1040  *         u     e      n                         u_w   e_w  s n_w
1041  *         |     |      |                          |     |   |  |
1042  * |<------------+------+----- TCP seqno space --------------+---------->|
1043  * ...-- <2^31 ->|                                           |<--------...
1044  * ...---- >2^31 ------>|                                    |<--------...
1045  *
1046  * Current code wouldn't be vulnerable but it's better still to discard such
1047  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1048  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1049  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1050  * equal to the ideal case (infinite seqno space without wrap caused issues).
1051  *
1052  * With D-SACK the lower bound is extended to cover sequence space below
1053  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1054  * again, DSACK block must not to go across snd_una (for the same reason as
1055  * for the normal SACK blocks, explained above). But there all simplicity
1056  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1057  * fully below undo_marker they do not affect behavior in anyway and can
1058  * therefore be safely ignored. In rare cases (which are more or less
1059  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1060  * fragmentation and packet reordering past skb's retransmission. To consider
1061  * them correctly, the acceptable range must be extended even more though
1062  * the exact amount is rather hard to quantify. However, tp->max_window can
1063  * be used as an exaggerated estimate.
1064  */
1065 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1066                                   u32 start_seq, u32 end_seq)
1067 {
1068         /* Too far in future, or reversed (interpretation is ambiguous) */
1069         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1070                 return 0;
1071
1072         /* Nasty start_seq wrap-around check (see comments above) */
1073         if (!before(start_seq, tp->snd_nxt))
1074                 return 0;
1075
1076         /* In outstanding window? ...This is valid exit for DSACKs too.
1077          * start_seq == snd_una is non-sensical (see comments above)
1078          */
1079         if (after(start_seq, tp->snd_una))
1080                 return 1;
1081
1082         if (!is_dsack || !tp->undo_marker)
1083                 return 0;
1084
1085         /* ...Then it's D-SACK, and must reside below snd_una completely */
1086         if (!after(end_seq, tp->snd_una))
1087                 return 0;
1088
1089         if (!before(start_seq, tp->undo_marker))
1090                 return 1;
1091
1092         /* Too old */
1093         if (!after(end_seq, tp->undo_marker))
1094                 return 0;
1095
1096         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1097          *   start_seq < undo_marker and end_seq >= undo_marker.
1098          */
1099         return !before(start_seq, end_seq - tp->max_window);
1100 }
1101
1102
1103 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1104                            struct tcp_sack_block_wire *sp, int num_sacks,
1105                            u32 prior_snd_una)
1106 {
1107         u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1108         u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1109         int dup_sack = 0;
1110
1111         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1112                 dup_sack = 1;
1113                 tcp_dsack_seen(tp);
1114                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1115         } else if (num_sacks > 1) {
1116                 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1117                 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1118
1119                 if (!after(end_seq_0, end_seq_1) &&
1120                     !before(start_seq_0, start_seq_1)) {
1121                         dup_sack = 1;
1122                         tcp_dsack_seen(tp);
1123                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1124                 }
1125         }
1126
1127         /* D-SACK for already forgotten data... Do dumb counting. */
1128         if (dup_sack &&
1129             !after(end_seq_0, prior_snd_una) &&
1130             after(end_seq_0, tp->undo_marker))
1131                 tp->undo_retrans--;
1132
1133         return dup_sack;
1134 }
1135
1136 static int
1137 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1138 {
1139         const struct inet_connection_sock *icsk = inet_csk(sk);
1140         struct tcp_sock *tp = tcp_sk(sk);
1141         unsigned char *ptr = (skb_transport_header(ack_skb) +
1142                               TCP_SKB_CB(ack_skb)->sacked);
1143         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1144         struct sk_buff *cached_skb;
1145         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1146         int reord = tp->packets_out;
1147         int prior_fackets;
1148         u32 lost_retrans = 0;
1149         int flag = 0;
1150         int found_dup_sack = 0;
1151         int cached_fack_count;
1152         int i;
1153         int first_sack_index;
1154
1155         if (!tp->sacked_out) {
1156                 tp->fackets_out = 0;
1157                 tp->highest_sack = tp->snd_una;
1158         }
1159         prior_fackets = tp->fackets_out;
1160
1161         found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1162                                          num_sacks, prior_snd_una);
1163         if (found_dup_sack)
1164                 flag |= FLAG_DSACKING_ACK;
1165
1166         /* Eliminate too old ACKs, but take into
1167          * account more or less fresh ones, they can
1168          * contain valid SACK info.
1169          */
1170         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1171                 return 0;
1172
1173         /* SACK fastpath:
1174          * if the only SACK change is the increase of the end_seq of
1175          * the first block then only apply that SACK block
1176          * and use retrans queue hinting otherwise slowpath */
1177         flag = 1;
1178         for (i = 0; i < num_sacks; i++) {
1179                 __be32 start_seq = sp[i].start_seq;
1180                 __be32 end_seq = sp[i].end_seq;
1181
1182                 if (i == 0) {
1183                         if (tp->recv_sack_cache[i].start_seq != start_seq)
1184                                 flag = 0;
1185                 } else {
1186                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1187                             (tp->recv_sack_cache[i].end_seq != end_seq))
1188                                 flag = 0;
1189                 }
1190                 tp->recv_sack_cache[i].start_seq = start_seq;
1191                 tp->recv_sack_cache[i].end_seq = end_seq;
1192         }
1193         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1194         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1195                 tp->recv_sack_cache[i].start_seq = 0;
1196                 tp->recv_sack_cache[i].end_seq = 0;
1197         }
1198
1199         first_sack_index = 0;
1200         if (flag)
1201                 num_sacks = 1;
1202         else {
1203                 int j;
1204                 tp->fastpath_skb_hint = NULL;
1205
1206                 /* order SACK blocks to allow in order walk of the retrans queue */
1207                 for (i = num_sacks-1; i > 0; i--) {
1208                         for (j = 0; j < i; j++){
1209                                 if (after(ntohl(sp[j].start_seq),
1210                                           ntohl(sp[j+1].start_seq))){
1211                                         struct tcp_sack_block_wire tmp;
1212
1213                                         tmp = sp[j];
1214                                         sp[j] = sp[j+1];
1215                                         sp[j+1] = tmp;
1216
1217                                         /* Track where the first SACK block goes to */
1218                                         if (j == first_sack_index)
1219                                                 first_sack_index = j+1;
1220                                 }
1221
1222                         }
1223                 }
1224         }
1225
1226         /* clear flag as used for different purpose in following code */
1227         flag = 0;
1228
1229         /* Use SACK fastpath hint if valid */
1230         cached_skb = tp->fastpath_skb_hint;
1231         cached_fack_count = tp->fastpath_cnt_hint;
1232         if (!cached_skb) {
1233                 cached_skb = tcp_write_queue_head(sk);
1234                 cached_fack_count = 0;
1235         }
1236
1237         for (i=0; i<num_sacks; i++, sp++) {
1238                 struct sk_buff *skb;
1239                 __u32 start_seq = ntohl(sp->start_seq);
1240                 __u32 end_seq = ntohl(sp->end_seq);
1241                 int fack_count;
1242                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1243
1244                 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
1245                         if (dup_sack) {
1246                                 if (!tp->undo_marker)
1247                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1248                                 else
1249                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1250                         } else
1251                                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1252                         continue;
1253                 }
1254
1255                 skb = cached_skb;
1256                 fack_count = cached_fack_count;
1257
1258                 /* Event "B" in the comment above. */
1259                 if (after(end_seq, tp->high_seq))
1260                         flag |= FLAG_DATA_LOST;
1261
1262                 tcp_for_write_queue_from(skb, sk) {
1263                         int in_sack, pcount;
1264                         u8 sacked;
1265
1266                         if (skb == tcp_send_head(sk))
1267                                 break;
1268
1269                         cached_skb = skb;
1270                         cached_fack_count = fack_count;
1271                         if (i == first_sack_index) {
1272                                 tp->fastpath_skb_hint = skb;
1273                                 tp->fastpath_cnt_hint = fack_count;
1274                         }
1275
1276                         /* The retransmission queue is always in order, so
1277                          * we can short-circuit the walk early.
1278                          */
1279                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1280                                 break;
1281
1282                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1283                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1284
1285                         pcount = tcp_skb_pcount(skb);
1286
1287                         if (pcount > 1 && !in_sack &&
1288                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1289                                 unsigned int pkt_len;
1290
1291                                 in_sack = !after(start_seq,
1292                                                  TCP_SKB_CB(skb)->seq);
1293
1294                                 if (!in_sack)
1295                                         pkt_len = (start_seq -
1296                                                    TCP_SKB_CB(skb)->seq);
1297                                 else
1298                                         pkt_len = (end_seq -
1299                                                    TCP_SKB_CB(skb)->seq);
1300                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1301                                         break;
1302                                 pcount = tcp_skb_pcount(skb);
1303                         }
1304
1305                         fack_count += pcount;
1306
1307                         sacked = TCP_SKB_CB(skb)->sacked;
1308
1309                         /* Account D-SACK for retransmitted packet. */
1310                         if ((dup_sack && in_sack) &&
1311                             (sacked & TCPCB_RETRANS) &&
1312                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1313                                 tp->undo_retrans--;
1314
1315                         /* The frame is ACKed. */
1316                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1317                                 if (sacked&TCPCB_RETRANS) {
1318                                         if ((dup_sack && in_sack) &&
1319                                             (sacked&TCPCB_SACKED_ACKED))
1320                                                 reord = min(fack_count, reord);
1321                                 } else {
1322                                         /* If it was in a hole, we detected reordering. */
1323                                         if (fack_count < prior_fackets &&
1324                                             !(sacked&TCPCB_SACKED_ACKED))
1325                                                 reord = min(fack_count, reord);
1326                                 }
1327
1328                                 /* Nothing to do; acked frame is about to be dropped. */
1329                                 continue;
1330                         }
1331
1332                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1333                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1334                             (!lost_retrans || after(end_seq, lost_retrans)))
1335                                 lost_retrans = end_seq;
1336
1337                         if (!in_sack)
1338                                 continue;
1339
1340                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1341                                 if (sacked & TCPCB_SACKED_RETRANS) {
1342                                         /* If the segment is not tagged as lost,
1343                                          * we do not clear RETRANS, believing
1344                                          * that retransmission is still in flight.
1345                                          */
1346                                         if (sacked & TCPCB_LOST) {
1347                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1348                                                 tp->lost_out -= tcp_skb_pcount(skb);
1349                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1350
1351                                                 /* clear lost hint */
1352                                                 tp->retransmit_skb_hint = NULL;
1353                                         }
1354                                 } else {
1355                                         /* New sack for not retransmitted frame,
1356                                          * which was in hole. It is reordering.
1357                                          */
1358                                         if (!(sacked & TCPCB_RETRANS) &&
1359                                             fack_count < prior_fackets)
1360                                                 reord = min(fack_count, reord);
1361
1362                                         if (sacked & TCPCB_LOST) {
1363                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1364                                                 tp->lost_out -= tcp_skb_pcount(skb);
1365
1366                                                 /* clear lost hint */
1367                                                 tp->retransmit_skb_hint = NULL;
1368                                         }
1369                                         /* SACK enhanced F-RTO detection.
1370                                          * Set flag if and only if non-rexmitted
1371                                          * segments below frto_highmark are
1372                                          * SACKed (RFC4138; Appendix B).
1373                                          * Clearing correct due to in-order walk
1374                                          */
1375                                         if (after(end_seq, tp->frto_highmark)) {
1376                                                 flag &= ~FLAG_ONLY_ORIG_SACKED;
1377                                         } else {
1378                                                 if (!(sacked & TCPCB_RETRANS))
1379                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1380                                         }
1381                                 }
1382
1383                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1384                                 flag |= FLAG_DATA_SACKED;
1385                                 tp->sacked_out += tcp_skb_pcount(skb);
1386
1387                                 if (fack_count > tp->fackets_out)
1388                                         tp->fackets_out = fack_count;
1389
1390                                 if (after(TCP_SKB_CB(skb)->seq,
1391                                     tp->highest_sack))
1392                                         tp->highest_sack = TCP_SKB_CB(skb)->seq;
1393                         } else {
1394                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1395                                         reord = min(fack_count, reord);
1396                         }
1397
1398                         /* D-SACK. We can detect redundant retransmission
1399                          * in S|R and plain R frames and clear it.
1400                          * undo_retrans is decreased above, L|R frames
1401                          * are accounted above as well.
1402                          */
1403                         if (dup_sack &&
1404                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1405                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1406                                 tp->retrans_out -= tcp_skb_pcount(skb);
1407                                 tp->retransmit_skb_hint = NULL;
1408                         }
1409                 }
1410         }
1411
1412         /* Check for lost retransmit. This superb idea is
1413          * borrowed from "ratehalving". Event "C".
1414          * Later note: FACK people cheated me again 8),
1415          * we have to account for reordering! Ugly,
1416          * but should help.
1417          */
1418         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1419                 struct sk_buff *skb;
1420
1421                 tcp_for_write_queue(skb, sk) {
1422                         if (skb == tcp_send_head(sk))
1423                                 break;
1424                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1425                                 break;
1426                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1427                                 continue;
1428                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1429                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1430                             (tcp_is_fack(tp) ||
1431                              !before(lost_retrans,
1432                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1433                                      tp->mss_cache))) {
1434                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1435                                 tp->retrans_out -= tcp_skb_pcount(skb);
1436
1437                                 /* clear lost hint */
1438                                 tp->retransmit_skb_hint = NULL;
1439
1440                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1441                                         tp->lost_out += tcp_skb_pcount(skb);
1442                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1443                                         flag |= FLAG_DATA_SACKED;
1444                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1445                                 }
1446                         }
1447                 }
1448         }
1449
1450         tcp_verify_left_out(tp);
1451
1452         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1453             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1454                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1455
1456 #if FASTRETRANS_DEBUG > 0
1457         BUG_TRAP((int)tp->sacked_out >= 0);
1458         BUG_TRAP((int)tp->lost_out >= 0);
1459         BUG_TRAP((int)tp->retrans_out >= 0);
1460         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1461 #endif
1462         return flag;
1463 }
1464
1465 /* F-RTO can only be used if TCP has never retransmitted anything other than
1466  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1467  */
1468 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1469 {
1470         struct tcp_sock *tp = tcp_sk(sk);
1471         u32 holes;
1472
1473         holes = max(tp->lost_out, 1U);
1474         holes = min(holes, tp->packets_out);
1475
1476         if ((tp->sacked_out + holes) > tp->packets_out) {
1477                 tp->sacked_out = tp->packets_out - holes;
1478                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1479         }
1480 }
1481
1482 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1483
1484 static void tcp_add_reno_sack(struct sock *sk)
1485 {
1486         struct tcp_sock *tp = tcp_sk(sk);
1487         tp->sacked_out++;
1488         tcp_check_reno_reordering(sk, 0);
1489         tcp_verify_left_out(tp);
1490 }
1491
1492 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1493
1494 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1495 {
1496         struct tcp_sock *tp = tcp_sk(sk);
1497
1498         if (acked > 0) {
1499                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1500                 if (acked-1 >= tp->sacked_out)
1501                         tp->sacked_out = 0;
1502                 else
1503                         tp->sacked_out -= acked-1;
1504         }
1505         tcp_check_reno_reordering(sk, acked);
1506         tcp_verify_left_out(tp);
1507 }
1508
1509 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1510 {
1511         tp->sacked_out = 0;
1512 }
1513
1514 int tcp_use_frto(struct sock *sk)
1515 {
1516         const struct tcp_sock *tp = tcp_sk(sk);
1517         struct sk_buff *skb;
1518
1519         if (!sysctl_tcp_frto)
1520                 return 0;
1521
1522         if (IsSackFrto())
1523                 return 1;
1524
1525         /* Avoid expensive walking of rexmit queue if possible */
1526         if (tp->retrans_out > 1)
1527                 return 0;
1528
1529         skb = tcp_write_queue_head(sk);
1530         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1531         tcp_for_write_queue_from(skb, sk) {
1532                 if (skb == tcp_send_head(sk))
1533                         break;
1534                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1535                         return 0;
1536                 /* Short-circuit when first non-SACKed skb has been checked */
1537                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1538                         break;
1539         }
1540         return 1;
1541 }
1542
1543 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1544  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1545  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1546  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1547  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1548  * bits are handled if the Loss state is really to be entered (in
1549  * tcp_enter_frto_loss).
1550  *
1551  * Do like tcp_enter_loss() would; when RTO expires the second time it
1552  * does:
1553  *  "Reduce ssthresh if it has not yet been made inside this window."
1554  */
1555 void tcp_enter_frto(struct sock *sk)
1556 {
1557         const struct inet_connection_sock *icsk = inet_csk(sk);
1558         struct tcp_sock *tp = tcp_sk(sk);
1559         struct sk_buff *skb;
1560
1561         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1562             tp->snd_una == tp->high_seq ||
1563             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1564              !icsk->icsk_retransmits)) {
1565                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1566                 /* Our state is too optimistic in ssthresh() call because cwnd
1567                  * is not reduced until tcp_enter_frto_loss() when previous FRTO
1568                  * recovery has not yet completed. Pattern would be this: RTO,
1569                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1570                  * up here twice).
1571                  * RFC4138 should be more specific on what to do, even though
1572                  * RTO is quite unlikely to occur after the first Cumulative ACK
1573                  * due to back-off and complexity of triggering events ...
1574                  */
1575                 if (tp->frto_counter) {
1576                         u32 stored_cwnd;
1577                         stored_cwnd = tp->snd_cwnd;
1578                         tp->snd_cwnd = 2;
1579                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1580                         tp->snd_cwnd = stored_cwnd;
1581                 } else {
1582                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1583                 }
1584                 /* ... in theory, cong.control module could do "any tricks" in
1585                  * ssthresh(), which means that ca_state, lost bits and lost_out
1586                  * counter would have to be faked before the call occurs. We
1587                  * consider that too expensive, unlikely and hacky, so modules
1588                  * using these in ssthresh() must deal these incompatibility
1589                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1590                  */
1591                 tcp_ca_event(sk, CA_EVENT_FRTO);
1592         }
1593
1594         tp->undo_marker = tp->snd_una;
1595         tp->undo_retrans = 0;
1596
1597         skb = tcp_write_queue_head(sk);
1598         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1599                 tp->undo_marker = 0;
1600         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1601                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1602                 tp->retrans_out -= tcp_skb_pcount(skb);
1603         }
1604         tcp_verify_left_out(tp);
1605
1606         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1607          * The last condition is necessary at least in tp->frto_counter case.
1608          */
1609         if (IsSackFrto() && (tp->frto_counter ||
1610             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1611             after(tp->high_seq, tp->snd_una)) {
1612                 tp->frto_highmark = tp->high_seq;
1613         } else {
1614                 tp->frto_highmark = tp->snd_nxt;
1615         }
1616         tcp_set_ca_state(sk, TCP_CA_Disorder);
1617         tp->high_seq = tp->snd_nxt;
1618         tp->frto_counter = 1;
1619 }
1620
1621 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1622  * which indicates that we should follow the traditional RTO recovery,
1623  * i.e. mark everything lost and do go-back-N retransmission.
1624  */
1625 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1626 {
1627         struct tcp_sock *tp = tcp_sk(sk);
1628         struct sk_buff *skb;
1629
1630         tp->lost_out = 0;
1631         tp->retrans_out = 0;
1632         if (tcp_is_reno(tp))
1633                 tcp_reset_reno_sack(tp);
1634
1635         tcp_for_write_queue(skb, sk) {
1636                 if (skb == tcp_send_head(sk))
1637                         break;
1638                 /*
1639                  * Count the retransmission made on RTO correctly (only when
1640                  * waiting for the first ACK and did not get it)...
1641                  */
1642                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1643                         /* For some reason this R-bit might get cleared? */
1644                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1645                                 tp->retrans_out += tcp_skb_pcount(skb);
1646                         /* ...enter this if branch just for the first segment */
1647                         flag |= FLAG_DATA_ACKED;
1648                 } else {
1649                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1650                                 tp->undo_marker = 0;
1651                         TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1652                 }
1653
1654                 /* Don't lost mark skbs that were fwd transmitted after RTO */
1655                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1656                     !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1657                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1658                         tp->lost_out += tcp_skb_pcount(skb);
1659                 }
1660         }
1661         tcp_verify_left_out(tp);
1662
1663         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1664         tp->snd_cwnd_cnt = 0;
1665         tp->snd_cwnd_stamp = tcp_time_stamp;
1666         tp->frto_counter = 0;
1667
1668         tp->reordering = min_t(unsigned int, tp->reordering,
1669                                              sysctl_tcp_reordering);
1670         tcp_set_ca_state(sk, TCP_CA_Loss);
1671         tp->high_seq = tp->frto_highmark;
1672         TCP_ECN_queue_cwr(tp);
1673
1674         tcp_clear_retrans_hints_partial(tp);
1675 }
1676
1677 void tcp_clear_retrans(struct tcp_sock *tp)
1678 {
1679         tp->retrans_out = 0;
1680
1681         tp->fackets_out = 0;
1682         tp->sacked_out = 0;
1683         tp->lost_out = 0;
1684
1685         tp->undo_marker = 0;
1686         tp->undo_retrans = 0;
1687 }
1688
1689 /* Enter Loss state. If "how" is not zero, forget all SACK information
1690  * and reset tags completely, otherwise preserve SACKs. If receiver
1691  * dropped its ofo queue, we will know this due to reneging detection.
1692  */
1693 void tcp_enter_loss(struct sock *sk, int how)
1694 {
1695         const struct inet_connection_sock *icsk = inet_csk(sk);
1696         struct tcp_sock *tp = tcp_sk(sk);
1697         struct sk_buff *skb;
1698         int cnt = 0;
1699
1700         /* Reduce ssthresh if it has not yet been made inside this window. */
1701         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1702             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1703                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1704                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1705                 tcp_ca_event(sk, CA_EVENT_LOSS);
1706         }
1707         tp->snd_cwnd       = 1;
1708         tp->snd_cwnd_cnt   = 0;
1709         tp->snd_cwnd_stamp = tcp_time_stamp;
1710
1711         tp->bytes_acked = 0;
1712         tcp_clear_retrans(tp);
1713
1714         if (!how) {
1715                 /* Push undo marker, if it was plain RTO and nothing
1716                  * was retransmitted. */
1717                 tp->undo_marker = tp->snd_una;
1718                 tcp_clear_retrans_hints_partial(tp);
1719         } else {
1720                 tcp_clear_all_retrans_hints(tp);
1721         }
1722
1723         tcp_for_write_queue(skb, sk) {
1724                 if (skb == tcp_send_head(sk))
1725                         break;
1726                 cnt += tcp_skb_pcount(skb);
1727                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1728                         tp->undo_marker = 0;
1729                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1730                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1731                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1732                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1733                         tp->lost_out += tcp_skb_pcount(skb);
1734                 } else {
1735                         tp->sacked_out += tcp_skb_pcount(skb);
1736                         tp->fackets_out = cnt;
1737                 }
1738         }
1739         tcp_verify_left_out(tp);
1740
1741         tp->reordering = min_t(unsigned int, tp->reordering,
1742                                              sysctl_tcp_reordering);
1743         tcp_set_ca_state(sk, TCP_CA_Loss);
1744         tp->high_seq = tp->snd_nxt;
1745         TCP_ECN_queue_cwr(tp);
1746         /* Abort FRTO algorithm if one is in progress */
1747         tp->frto_counter = 0;
1748 }
1749
1750 static int tcp_check_sack_reneging(struct sock *sk)
1751 {
1752         struct sk_buff *skb;
1753
1754         /* If ACK arrived pointing to a remembered SACK,
1755          * it means that our remembered SACKs do not reflect
1756          * real state of receiver i.e.
1757          * receiver _host_ is heavily congested (or buggy).
1758          * Do processing similar to RTO timeout.
1759          */
1760         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1761             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1762                 struct inet_connection_sock *icsk = inet_csk(sk);
1763                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1764
1765                 tcp_enter_loss(sk, 1);
1766                 icsk->icsk_retransmits++;
1767                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1768                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1769                                           icsk->icsk_rto, TCP_RTO_MAX);
1770                 return 1;
1771         }
1772         return 0;
1773 }
1774
1775 static inline int tcp_fackets_out(struct tcp_sock *tp)
1776 {
1777         return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1778 }
1779
1780 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1781 {
1782         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1783 }
1784
1785 static inline int tcp_head_timedout(struct sock *sk)
1786 {
1787         struct tcp_sock *tp = tcp_sk(sk);
1788
1789         return tp->packets_out &&
1790                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1791 }
1792
1793 /* Linux NewReno/SACK/FACK/ECN state machine.
1794  * --------------------------------------
1795  *
1796  * "Open"       Normal state, no dubious events, fast path.
1797  * "Disorder"   In all the respects it is "Open",
1798  *              but requires a bit more attention. It is entered when
1799  *              we see some SACKs or dupacks. It is split of "Open"
1800  *              mainly to move some processing from fast path to slow one.
1801  * "CWR"        CWND was reduced due to some Congestion Notification event.
1802  *              It can be ECN, ICMP source quench, local device congestion.
1803  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1804  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1805  *
1806  * tcp_fastretrans_alert() is entered:
1807  * - each incoming ACK, if state is not "Open"
1808  * - when arrived ACK is unusual, namely:
1809  *      * SACK
1810  *      * Duplicate ACK.
1811  *      * ECN ECE.
1812  *
1813  * Counting packets in flight is pretty simple.
1814  *
1815  *      in_flight = packets_out - left_out + retrans_out
1816  *
1817  *      packets_out is SND.NXT-SND.UNA counted in packets.
1818  *
1819  *      retrans_out is number of retransmitted segments.
1820  *
1821  *      left_out is number of segments left network, but not ACKed yet.
1822  *
1823  *              left_out = sacked_out + lost_out
1824  *
1825  *     sacked_out: Packets, which arrived to receiver out of order
1826  *                 and hence not ACKed. With SACKs this number is simply
1827  *                 amount of SACKed data. Even without SACKs
1828  *                 it is easy to give pretty reliable estimate of this number,
1829  *                 counting duplicate ACKs.
1830  *
1831  *       lost_out: Packets lost by network. TCP has no explicit
1832  *                 "loss notification" feedback from network (for now).
1833  *                 It means that this number can be only _guessed_.
1834  *                 Actually, it is the heuristics to predict lossage that
1835  *                 distinguishes different algorithms.
1836  *
1837  *      F.e. after RTO, when all the queue is considered as lost,
1838  *      lost_out = packets_out and in_flight = retrans_out.
1839  *
1840  *              Essentially, we have now two algorithms counting
1841  *              lost packets.
1842  *
1843  *              FACK: It is the simplest heuristics. As soon as we decided
1844  *              that something is lost, we decide that _all_ not SACKed
1845  *              packets until the most forward SACK are lost. I.e.
1846  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1847  *              It is absolutely correct estimate, if network does not reorder
1848  *              packets. And it loses any connection to reality when reordering
1849  *              takes place. We use FACK by default until reordering
1850  *              is suspected on the path to this destination.
1851  *
1852  *              NewReno: when Recovery is entered, we assume that one segment
1853  *              is lost (classic Reno). While we are in Recovery and
1854  *              a partial ACK arrives, we assume that one more packet
1855  *              is lost (NewReno). This heuristics are the same in NewReno
1856  *              and SACK.
1857  *
1858  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1859  *  deflation etc. CWND is real congestion window, never inflated, changes
1860  *  only according to classic VJ rules.
1861  *
1862  * Really tricky (and requiring careful tuning) part of algorithm
1863  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1864  * The first determines the moment _when_ we should reduce CWND and,
1865  * hence, slow down forward transmission. In fact, it determines the moment
1866  * when we decide that hole is caused by loss, rather than by a reorder.
1867  *
1868  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1869  * holes, caused by lost packets.
1870  *
1871  * And the most logically complicated part of algorithm is undo
1872  * heuristics. We detect false retransmits due to both too early
1873  * fast retransmit (reordering) and underestimated RTO, analyzing
1874  * timestamps and D-SACKs. When we detect that some segments were
1875  * retransmitted by mistake and CWND reduction was wrong, we undo
1876  * window reduction and abort recovery phase. This logic is hidden
1877  * inside several functions named tcp_try_undo_<something>.
1878  */
1879
1880 /* This function decides, when we should leave Disordered state
1881  * and enter Recovery phase, reducing congestion window.
1882  *
1883  * Main question: may we further continue forward transmission
1884  * with the same cwnd?
1885  */
1886 static int tcp_time_to_recover(struct sock *sk)
1887 {
1888         struct tcp_sock *tp = tcp_sk(sk);
1889         __u32 packets_out;
1890
1891         /* Do not perform any recovery during FRTO algorithm */
1892         if (tp->frto_counter)
1893                 return 0;
1894
1895         /* Trick#1: The loss is proven. */
1896         if (tp->lost_out)
1897                 return 1;
1898
1899         /* Not-A-Trick#2 : Classic rule... */
1900         if (tcp_fackets_out(tp) > tp->reordering)
1901                 return 1;
1902
1903         /* Trick#3 : when we use RFC2988 timer restart, fast
1904          * retransmit can be triggered by timeout of queue head.
1905          */
1906         if (tcp_head_timedout(sk))
1907                 return 1;
1908
1909         /* Trick#4: It is still not OK... But will it be useful to delay
1910          * recovery more?
1911          */
1912         packets_out = tp->packets_out;
1913         if (packets_out <= tp->reordering &&
1914             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1915             !tcp_may_send_now(sk)) {
1916                 /* We have nothing to send. This connection is limited
1917                  * either by receiver window or by application.
1918                  */
1919                 return 1;
1920         }
1921
1922         return 0;
1923 }
1924
1925 /* RFC: This is from the original, I doubt that this is necessary at all:
1926  * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1927  * retransmitted past LOST markings in the first place? I'm not fully sure
1928  * about undo and end of connection cases, which can cause R without L?
1929  */
1930 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
1931                                        struct sk_buff *skb)
1932 {
1933         if ((tp->retransmit_skb_hint != NULL) &&
1934             before(TCP_SKB_CB(skb)->seq,
1935             TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1936                 tp->retransmit_skb_hint = NULL;
1937 }
1938
1939 /* Mark head of queue up as lost. */
1940 static void tcp_mark_head_lost(struct sock *sk,
1941                                int packets, u32 high_seq)
1942 {
1943         struct tcp_sock *tp = tcp_sk(sk);
1944         struct sk_buff *skb;
1945         int cnt;
1946
1947         BUG_TRAP(packets <= tp->packets_out);
1948         if (tp->lost_skb_hint) {
1949                 skb = tp->lost_skb_hint;
1950                 cnt = tp->lost_cnt_hint;
1951         } else {
1952                 skb = tcp_write_queue_head(sk);
1953                 cnt = 0;
1954         }
1955
1956         tcp_for_write_queue_from(skb, sk) {
1957                 if (skb == tcp_send_head(sk))
1958                         break;
1959                 /* TODO: do this better */
1960                 /* this is not the most efficient way to do this... */
1961                 tp->lost_skb_hint = skb;
1962                 tp->lost_cnt_hint = cnt;
1963                 cnt += tcp_skb_pcount(skb);
1964                 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1965                         break;
1966                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1967                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1968                         tp->lost_out += tcp_skb_pcount(skb);
1969                         tcp_verify_retransmit_hint(tp, skb);
1970                 }
1971         }
1972         tcp_verify_left_out(tp);
1973 }
1974
1975 /* Account newly detected lost packet(s) */
1976
1977 static void tcp_update_scoreboard(struct sock *sk)
1978 {
1979         struct tcp_sock *tp = tcp_sk(sk);
1980
1981         if (tcp_is_fack(tp)) {
1982                 int lost = tp->fackets_out - tp->reordering;
1983                 if (lost <= 0)
1984                         lost = 1;
1985                 tcp_mark_head_lost(sk, lost, tp->high_seq);
1986         } else {
1987                 tcp_mark_head_lost(sk, 1, tp->high_seq);
1988         }
1989
1990         /* New heuristics: it is possible only after we switched
1991          * to restart timer each time when something is ACKed.
1992          * Hence, we can detect timed out packets during fast
1993          * retransmit without falling to slow start.
1994          */
1995         if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) {
1996                 struct sk_buff *skb;
1997
1998                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1999                         : tcp_write_queue_head(sk);
2000
2001                 tcp_for_write_queue_from(skb, sk) {
2002                         if (skb == tcp_send_head(sk))
2003                                 break;
2004                         if (!tcp_skb_timedout(sk, skb))
2005                                 break;
2006
2007                         if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
2008                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2009                                 tp->lost_out += tcp_skb_pcount(skb);
2010                                 tcp_verify_retransmit_hint(tp, skb);
2011                         }
2012                 }
2013
2014                 tp->scoreboard_skb_hint = skb;
2015
2016                 tcp_verify_left_out(tp);
2017         }
2018 }
2019
2020 /* CWND moderation, preventing bursts due to too big ACKs
2021  * in dubious situations.
2022  */
2023 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2024 {
2025         tp->snd_cwnd = min(tp->snd_cwnd,
2026                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2027         tp->snd_cwnd_stamp = tcp_time_stamp;
2028 }
2029
2030 /* Lower bound on congestion window is slow start threshold
2031  * unless congestion avoidance choice decides to overide it.
2032  */
2033 static inline u32 tcp_cwnd_min(const struct sock *sk)
2034 {
2035         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2036
2037         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2038 }
2039
2040 /* Decrease cwnd each second ack. */
2041 static void tcp_cwnd_down(struct sock *sk, int flag)
2042 {
2043         struct tcp_sock *tp = tcp_sk(sk);
2044         int decr = tp->snd_cwnd_cnt + 1;
2045
2046         if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2047             (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2048                 tp->snd_cwnd_cnt = decr&1;
2049                 decr >>= 1;
2050
2051                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2052                         tp->snd_cwnd -= decr;
2053
2054                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2055                 tp->snd_cwnd_stamp = tcp_time_stamp;
2056         }
2057 }
2058
2059 /* Nothing was retransmitted or returned timestamp is less
2060  * than timestamp of the first retransmission.
2061  */
2062 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2063 {
2064         return !tp->retrans_stamp ||
2065                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2066                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2067 }
2068
2069 /* Undo procedures. */
2070
2071 #if FASTRETRANS_DEBUG > 1
2072 static void DBGUNDO(struct sock *sk, const char *msg)
2073 {
2074         struct tcp_sock *tp = tcp_sk(sk);
2075         struct inet_sock *inet = inet_sk(sk);
2076
2077         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2078                msg,
2079                NIPQUAD(inet->daddr), ntohs(inet->dport),
2080                tp->snd_cwnd, tcp_left_out(tp),
2081                tp->snd_ssthresh, tp->prior_ssthresh,
2082                tp->packets_out);
2083 }
2084 #else
2085 #define DBGUNDO(x...) do { } while (0)
2086 #endif
2087
2088 static void tcp_undo_cwr(struct sock *sk, const int undo)
2089 {
2090         struct tcp_sock *tp = tcp_sk(sk);
2091
2092         if (tp->prior_ssthresh) {
2093                 const struct inet_connection_sock *icsk = inet_csk(sk);
2094
2095                 if (icsk->icsk_ca_ops->undo_cwnd)
2096                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2097                 else
2098                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2099
2100                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2101                         tp->snd_ssthresh = tp->prior_ssthresh;
2102                         TCP_ECN_withdraw_cwr(tp);
2103                 }
2104         } else {
2105                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2106         }
2107         tcp_moderate_cwnd(tp);
2108         tp->snd_cwnd_stamp = tcp_time_stamp;
2109
2110         /* There is something screwy going on with the retrans hints after
2111            an undo */
2112         tcp_clear_all_retrans_hints(tp);
2113 }
2114
2115 static inline int tcp_may_undo(struct tcp_sock *tp)
2116 {
2117         return tp->undo_marker &&
2118                 (!tp->undo_retrans || tcp_packet_delayed(tp));
2119 }
2120
2121 /* People celebrate: "We love our President!" */
2122 static int tcp_try_undo_recovery(struct sock *sk)
2123 {
2124         struct tcp_sock *tp = tcp_sk(sk);
2125
2126         if (tcp_may_undo(tp)) {
2127                 /* Happy end! We did not retransmit anything
2128                  * or our original transmission succeeded.
2129                  */
2130                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2131                 tcp_undo_cwr(sk, 1);
2132                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2133                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2134                 else
2135                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2136                 tp->undo_marker = 0;
2137         }
2138         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2139                 /* Hold old state until something *above* high_seq
2140                  * is ACKed. For Reno it is MUST to prevent false
2141                  * fast retransmits (RFC2582). SACK TCP is safe. */
2142                 tcp_moderate_cwnd(tp);
2143                 return 1;
2144         }
2145         tcp_set_ca_state(sk, TCP_CA_Open);
2146         return 0;
2147 }
2148
2149 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2150 static void tcp_try_undo_dsack(struct sock *sk)
2151 {
2152         struct tcp_sock *tp = tcp_sk(sk);
2153
2154         if (tp->undo_marker && !tp->undo_retrans) {
2155                 DBGUNDO(sk, "D-SACK");
2156                 tcp_undo_cwr(sk, 1);
2157                 tp->undo_marker = 0;
2158                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2159         }
2160 }
2161
2162 /* Undo during fast recovery after partial ACK. */
2163
2164 static int tcp_try_undo_partial(struct sock *sk, int acked)
2165 {
2166         struct tcp_sock *tp = tcp_sk(sk);
2167         /* Partial ACK arrived. Force Hoe's retransmit. */
2168         int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering;
2169
2170         if (tcp_may_undo(tp)) {
2171                 /* Plain luck! Hole if filled with delayed
2172                  * packet, rather than with a retransmit.
2173                  */
2174                 if (tp->retrans_out == 0)
2175                         tp->retrans_stamp = 0;
2176
2177                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2178
2179                 DBGUNDO(sk, "Hoe");
2180                 tcp_undo_cwr(sk, 0);
2181                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2182
2183                 /* So... Do not make Hoe's retransmit yet.
2184                  * If the first packet was delayed, the rest
2185                  * ones are most probably delayed as well.
2186                  */
2187                 failed = 0;
2188         }
2189         return failed;
2190 }
2191
2192 /* Undo during loss recovery after partial ACK. */
2193 static int tcp_try_undo_loss(struct sock *sk)
2194 {
2195         struct tcp_sock *tp = tcp_sk(sk);
2196
2197         if (tcp_may_undo(tp)) {
2198                 struct sk_buff *skb;
2199                 tcp_for_write_queue(skb, sk) {
2200                         if (skb == tcp_send_head(sk))
2201                                 break;
2202                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2203                 }
2204
2205                 tcp_clear_all_retrans_hints(tp);
2206
2207                 DBGUNDO(sk, "partial loss");
2208                 tp->lost_out = 0;
2209                 tcp_undo_cwr(sk, 1);
2210                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2211                 inet_csk(sk)->icsk_retransmits = 0;
2212                 tp->undo_marker = 0;
2213                 if (tcp_is_sack(tp))
2214                         tcp_set_ca_state(sk, TCP_CA_Open);
2215                 return 1;
2216         }
2217         return 0;
2218 }
2219
2220 static inline void tcp_complete_cwr(struct sock *sk)
2221 {
2222         struct tcp_sock *tp = tcp_sk(sk);
2223         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2224         tp->snd_cwnd_stamp = tcp_time_stamp;
2225         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2226 }
2227
2228 static void tcp_try_to_open(struct sock *sk, int flag)
2229 {
2230         struct tcp_sock *tp = tcp_sk(sk);
2231
2232         tcp_verify_left_out(tp);
2233
2234         if (tp->retrans_out == 0)
2235                 tp->retrans_stamp = 0;
2236
2237         if (flag&FLAG_ECE)
2238                 tcp_enter_cwr(sk, 1);
2239
2240         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2241                 int state = TCP_CA_Open;
2242
2243                 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2244                         state = TCP_CA_Disorder;
2245
2246                 if (inet_csk(sk)->icsk_ca_state != state) {
2247                         tcp_set_ca_state(sk, state);
2248                         tp->high_seq = tp->snd_nxt;
2249                 }
2250                 tcp_moderate_cwnd(tp);
2251         } else {
2252                 tcp_cwnd_down(sk, flag);
2253         }
2254 }
2255
2256 static void tcp_mtup_probe_failed(struct sock *sk)
2257 {
2258         struct inet_connection_sock *icsk = inet_csk(sk);
2259
2260         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2261         icsk->icsk_mtup.probe_size = 0;
2262 }
2263
2264 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2265 {
2266         struct tcp_sock *tp = tcp_sk(sk);
2267         struct inet_connection_sock *icsk = inet_csk(sk);
2268
2269         /* FIXME: breaks with very large cwnd */
2270         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2271         tp->snd_cwnd = tp->snd_cwnd *
2272                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2273                        icsk->icsk_mtup.probe_size;
2274         tp->snd_cwnd_cnt = 0;
2275         tp->snd_cwnd_stamp = tcp_time_stamp;
2276         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2277
2278         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2279         icsk->icsk_mtup.probe_size = 0;
2280         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2281 }
2282
2283
2284 /* Process an event, which can update packets-in-flight not trivially.
2285  * Main goal of this function is to calculate new estimate for left_out,
2286  * taking into account both packets sitting in receiver's buffer and
2287  * packets lost by network.
2288  *
2289  * Besides that it does CWND reduction, when packet loss is detected
2290  * and changes state of machine.
2291  *
2292  * It does _not_ decide what to send, it is made in function
2293  * tcp_xmit_retransmit_queue().
2294  */
2295 static void
2296 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2297 {
2298         struct inet_connection_sock *icsk = inet_csk(sk);
2299         struct tcp_sock *tp = tcp_sk(sk);
2300         int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2301         int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2302                                     (tp->fackets_out > tp->reordering));
2303
2304         /* Some technical things:
2305          * 1. Reno does not count dupacks (sacked_out) automatically. */
2306         if (!tp->packets_out)
2307                 tp->sacked_out = 0;
2308
2309         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2310                 tp->fackets_out = 0;
2311
2312         /* Now state machine starts.
2313          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2314         if (flag&FLAG_ECE)
2315                 tp->prior_ssthresh = 0;
2316
2317         /* B. In all the states check for reneging SACKs. */
2318         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2319                 return;
2320
2321         /* C. Process data loss notification, provided it is valid. */
2322         if ((flag&FLAG_DATA_LOST) &&
2323             before(tp->snd_una, tp->high_seq) &&
2324             icsk->icsk_ca_state != TCP_CA_Open &&
2325             tp->fackets_out > tp->reordering) {
2326                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2327                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2328         }
2329
2330         /* D. Check consistency of the current state. */
2331         tcp_verify_left_out(tp);
2332
2333         /* E. Check state exit conditions. State can be terminated
2334          *    when high_seq is ACKed. */
2335         if (icsk->icsk_ca_state == TCP_CA_Open) {
2336                 BUG_TRAP(tp->retrans_out == 0);
2337                 tp->retrans_stamp = 0;
2338         } else if (!before(tp->snd_una, tp->high_seq)) {
2339                 switch (icsk->icsk_ca_state) {
2340                 case TCP_CA_Loss:
2341                         icsk->icsk_retransmits = 0;
2342                         if (tcp_try_undo_recovery(sk))
2343                                 return;
2344                         break;
2345
2346                 case TCP_CA_CWR:
2347                         /* CWR is to be held something *above* high_seq
2348                          * is ACKed for CWR bit to reach receiver. */
2349                         if (tp->snd_una != tp->high_seq) {
2350                                 tcp_complete_cwr(sk);
2351                                 tcp_set_ca_state(sk, TCP_CA_Open);
2352                         }
2353                         break;
2354
2355                 case TCP_CA_Disorder:
2356                         tcp_try_undo_dsack(sk);
2357                         if (!tp->undo_marker ||
2358                             /* For SACK case do not Open to allow to undo
2359                              * catching for all duplicate ACKs. */
2360                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2361                                 tp->undo_marker = 0;
2362                                 tcp_set_ca_state(sk, TCP_CA_Open);
2363                         }
2364                         break;
2365
2366                 case TCP_CA_Recovery:
2367                         if (tcp_is_reno(tp))
2368                                 tcp_reset_reno_sack(tp);
2369                         if (tcp_try_undo_recovery(sk))
2370                                 return;
2371                         tcp_complete_cwr(sk);
2372                         break;
2373                 }
2374         }
2375
2376         /* F. Process state. */
2377         switch (icsk->icsk_ca_state) {
2378         case TCP_CA_Recovery:
2379                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2380                         if (tcp_is_reno(tp) && is_dupack)
2381                                 tcp_add_reno_sack(sk);
2382                 } else
2383                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2384                 break;
2385         case TCP_CA_Loss:
2386                 if (flag&FLAG_DATA_ACKED)
2387                         icsk->icsk_retransmits = 0;
2388                 if (!tcp_try_undo_loss(sk)) {
2389                         tcp_moderate_cwnd(tp);
2390                         tcp_xmit_retransmit_queue(sk);
2391                         return;
2392                 }
2393                 if (icsk->icsk_ca_state != TCP_CA_Open)
2394                         return;
2395                 /* Loss is undone; fall through to processing in Open state. */
2396         default:
2397                 if (tcp_is_reno(tp)) {
2398                         if (flag & FLAG_SND_UNA_ADVANCED)
2399                                 tcp_reset_reno_sack(tp);
2400                         if (is_dupack)
2401                                 tcp_add_reno_sack(sk);
2402                 }
2403
2404                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2405                         tcp_try_undo_dsack(sk);
2406
2407                 if (!tcp_time_to_recover(sk)) {
2408                         tcp_try_to_open(sk, flag);
2409                         return;
2410                 }
2411
2412                 /* MTU probe failure: don't reduce cwnd */
2413                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2414                     icsk->icsk_mtup.probe_size &&
2415                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2416                         tcp_mtup_probe_failed(sk);
2417                         /* Restores the reduction we did in tcp_mtup_probe() */
2418                         tp->snd_cwnd++;
2419                         tcp_simple_retransmit(sk);
2420                         return;
2421                 }
2422
2423                 /* Otherwise enter Recovery state */
2424
2425                 if (tcp_is_reno(tp))
2426                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2427                 else
2428                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2429
2430                 tp->high_seq = tp->snd_nxt;
2431                 tp->prior_ssthresh = 0;
2432                 tp->undo_marker = tp->snd_una;
2433                 tp->undo_retrans = tp->retrans_out;
2434
2435                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2436                         if (!(flag&FLAG_ECE))
2437                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2438                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2439                         TCP_ECN_queue_cwr(tp);
2440                 }
2441
2442                 tp->bytes_acked = 0;
2443                 tp->snd_cwnd_cnt = 0;
2444                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2445         }
2446
2447         if (do_lost || tcp_head_timedout(sk))
2448                 tcp_update_scoreboard(sk);
2449         tcp_cwnd_down(sk, flag);
2450         tcp_xmit_retransmit_queue(sk);
2451 }
2452
2453 /* Read draft-ietf-tcplw-high-performance before mucking
2454  * with this code. (Supersedes RFC1323)
2455  */
2456 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2457 {
2458         /* RTTM Rule: A TSecr value received in a segment is used to
2459          * update the averaged RTT measurement only if the segment
2460          * acknowledges some new data, i.e., only if it advances the
2461          * left edge of the send window.
2462          *
2463          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2464          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2465          *
2466          * Changed: reset backoff as soon as we see the first valid sample.
2467          * If we do not, we get strongly overestimated rto. With timestamps
2468          * samples are accepted even from very old segments: f.e., when rtt=1
2469          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2470          * answer arrives rto becomes 120 seconds! If at least one of segments
2471          * in window is lost... Voila.                          --ANK (010210)
2472          */
2473         struct tcp_sock *tp = tcp_sk(sk);
2474         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2475         tcp_rtt_estimator(sk, seq_rtt);
2476         tcp_set_rto(sk);
2477         inet_csk(sk)->icsk_backoff = 0;
2478         tcp_bound_rto(sk);
2479 }
2480
2481 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2482 {
2483         /* We don't have a timestamp. Can only use
2484          * packets that are not retransmitted to determine
2485          * rtt estimates. Also, we must not reset the
2486          * backoff for rto until we get a non-retransmitted
2487          * packet. This allows us to deal with a situation
2488          * where the network delay has increased suddenly.
2489          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2490          */
2491
2492         if (flag & FLAG_RETRANS_DATA_ACKED)
2493                 return;
2494
2495         tcp_rtt_estimator(sk, seq_rtt);
2496         tcp_set_rto(sk);
2497         inet_csk(sk)->icsk_backoff = 0;
2498         tcp_bound_rto(sk);
2499 }
2500
2501 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2502                                       const s32 seq_rtt)
2503 {
2504         const struct tcp_sock *tp = tcp_sk(sk);
2505         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2506         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2507                 tcp_ack_saw_tstamp(sk, flag);
2508         else if (seq_rtt >= 0)
2509                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2510 }
2511
2512 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2513                            u32 in_flight, int good)
2514 {
2515         const struct inet_connection_sock *icsk = inet_csk(sk);
2516         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2517         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2518 }
2519
2520 /* Restart timer after forward progress on connection.
2521  * RFC2988 recommends to restart timer to now+rto.
2522  */
2523 static void tcp_rearm_rto(struct sock *sk)
2524 {
2525         struct tcp_sock *tp = tcp_sk(sk);
2526
2527         if (!tp->packets_out) {
2528                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2529         } else {
2530                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2531         }
2532 }
2533
2534 /* If we get here, the whole TSO packet has not been acked. */
2535 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2536 {
2537         struct tcp_sock *tp = tcp_sk(sk);
2538         u32 packets_acked;
2539
2540         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2541
2542         packets_acked = tcp_skb_pcount(skb);
2543         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2544                 return 0;
2545         packets_acked -= tcp_skb_pcount(skb);
2546
2547         if (packets_acked) {
2548                 BUG_ON(tcp_skb_pcount(skb) == 0);
2549                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2550         }
2551
2552         return packets_acked;
2553 }
2554
2555 /* Remove acknowledged frames from the retransmission queue. If our packet
2556  * is before the ack sequence we can discard it as it's confirmed to have
2557  * arrived at the other end.
2558  */
2559 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
2560 {
2561         struct tcp_sock *tp = tcp_sk(sk);
2562         const struct inet_connection_sock *icsk = inet_csk(sk);
2563         struct sk_buff *skb;
2564         u32 now = tcp_time_stamp;
2565         int fully_acked = 1;
2566         int flag = 0;
2567         int prior_packets = tp->packets_out;
2568         s32 seq_rtt = -1;
2569         ktime_t last_ackt = net_invalid_timestamp();
2570
2571         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2572                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2573                 u32 end_seq;
2574                 u32 packets_acked;
2575                 u8 sacked = scb->sacked;
2576
2577                 if (after(scb->end_seq, tp->snd_una)) {
2578                         if (tcp_skb_pcount(skb) == 1 ||
2579                             !after(tp->snd_una, scb->seq))
2580                                 break;
2581
2582                         packets_acked = tcp_tso_acked(sk, skb);
2583                         if (!packets_acked)
2584                                 break;
2585
2586                         fully_acked = 0;
2587                         end_seq = tp->snd_una;
2588                 } else {
2589                         packets_acked = tcp_skb_pcount(skb);
2590                         end_seq = scb->end_seq;
2591                 }
2592
2593                 /* MTU probing checks */
2594                 if (fully_acked && icsk->icsk_mtup.probe_size &&
2595                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2596                         tcp_mtup_probe_success(sk, skb);
2597                 }
2598
2599                 if (sacked) {
2600                         if (sacked & TCPCB_RETRANS) {
2601                                 if (sacked & TCPCB_SACKED_RETRANS)
2602                                         tp->retrans_out -= packets_acked;
2603                                 flag |= FLAG_RETRANS_DATA_ACKED;
2604                                 seq_rtt = -1;
2605                                 if ((flag & FLAG_DATA_ACKED) ||
2606                                     (packets_acked > 1))
2607                                         flag |= FLAG_NONHEAD_RETRANS_ACKED;
2608                         } else if (seq_rtt < 0) {
2609                                 seq_rtt = now - scb->when;
2610                                 if (fully_acked)
2611                                         last_ackt = skb->tstamp;
2612                         }
2613
2614                         if (sacked & TCPCB_SACKED_ACKED)
2615                                 tp->sacked_out -= packets_acked;
2616                         if (sacked & TCPCB_LOST)
2617                                 tp->lost_out -= packets_acked;
2618
2619                         if ((sacked & TCPCB_URG) && tp->urg_mode &&
2620                             !before(end_seq, tp->snd_up))
2621                                 tp->urg_mode = 0;
2622                 } else if (seq_rtt < 0) {
2623                         seq_rtt = now - scb->when;
2624                         if (fully_acked)
2625                                 last_ackt = skb->tstamp;
2626                 }
2627                 tp->packets_out -= packets_acked;
2628
2629                 /* Initial outgoing SYN's get put onto the write_queue
2630                  * just like anything else we transmit.  It is not
2631                  * true data, and if we misinform our callers that
2632                  * this ACK acks real data, we will erroneously exit
2633                  * connection startup slow start one packet too
2634                  * quickly.  This is severely frowned upon behavior.
2635                  */
2636                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2637                         flag |= FLAG_DATA_ACKED;
2638                 } else {
2639                         flag |= FLAG_SYN_ACKED;
2640                         tp->retrans_stamp = 0;
2641                 }
2642
2643                 if (!fully_acked)
2644                         break;
2645
2646                 tcp_unlink_write_queue(skb, sk);
2647                 sk_stream_free_skb(sk, skb);
2648                 tcp_clear_all_retrans_hints(tp);
2649         }
2650
2651         if (flag & FLAG_ACKED) {
2652                 u32 pkts_acked = prior_packets - tp->packets_out;
2653                 const struct tcp_congestion_ops *ca_ops
2654                         = inet_csk(sk)->icsk_ca_ops;
2655
2656                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2657                 tcp_rearm_rto(sk);
2658
2659                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2660                 /* hint's skb might be NULL but we don't need to care */
2661                 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2662                                                tp->fastpath_cnt_hint);
2663                 if (tcp_is_reno(tp))
2664                         tcp_remove_reno_sacks(sk, pkts_acked);
2665
2666                 if (ca_ops->pkts_acked) {
2667                         s32 rtt_us = -1;
2668
2669                         /* Is the ACK triggering packet unambiguous? */
2670                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2671                                 /* High resolution needed and available? */
2672                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2673                                     !ktime_equal(last_ackt,
2674                                                  net_invalid_timestamp()))
2675                                         rtt_us = ktime_us_delta(ktime_get_real(),
2676                                                                 last_ackt);
2677                                 else if (seq_rtt > 0)
2678                                         rtt_us = jiffies_to_usecs(seq_rtt);
2679                         }
2680
2681                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2682                 }
2683         }
2684
2685 #if FASTRETRANS_DEBUG > 0
2686         BUG_TRAP((int)tp->sacked_out >= 0);
2687         BUG_TRAP((int)tp->lost_out >= 0);
2688         BUG_TRAP((int)tp->retrans_out >= 0);
2689         if (!tp->packets_out && tcp_is_sack(tp)) {
2690                 const struct inet_connection_sock *icsk = inet_csk(sk);
2691                 if (tp->lost_out) {
2692                         printk(KERN_DEBUG "Leak l=%u %d\n",
2693                                tp->lost_out, icsk->icsk_ca_state);
2694                         tp->lost_out = 0;
2695                 }
2696                 if (tp->sacked_out) {
2697                         printk(KERN_DEBUG "Leak s=%u %d\n",
2698                                tp->sacked_out, icsk->icsk_ca_state);
2699                         tp->sacked_out = 0;
2700                 }
2701                 if (tp->retrans_out) {
2702                         printk(KERN_DEBUG "Leak r=%u %d\n",
2703                                tp->retrans_out, icsk->icsk_ca_state);
2704                         tp->retrans_out = 0;
2705                 }
2706         }
2707 #endif
2708         *seq_rtt_p = seq_rtt;
2709         return flag;
2710 }
2711
2712 static void tcp_ack_probe(struct sock *sk)
2713 {
2714         const struct tcp_sock *tp = tcp_sk(sk);
2715         struct inet_connection_sock *icsk = inet_csk(sk);
2716
2717         /* Was it a usable window open? */
2718
2719         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2720                    tp->snd_una + tp->snd_wnd)) {
2721                 icsk->icsk_backoff = 0;
2722                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2723                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2724                  * This function is not for random using!
2725                  */
2726         } else {
2727                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2728                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2729                                           TCP_RTO_MAX);
2730         }
2731 }
2732
2733 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2734 {
2735         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2736                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2737 }
2738
2739 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2740 {
2741         const struct tcp_sock *tp = tcp_sk(sk);
2742         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2743                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2744 }
2745
2746 /* Check that window update is acceptable.
2747  * The function assumes that snd_una<=ack<=snd_next.
2748  */
2749 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2750                                         const u32 ack_seq, const u32 nwin)
2751 {
2752         return (after(ack, tp->snd_una) ||
2753                 after(ack_seq, tp->snd_wl1) ||
2754                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2755 }
2756
2757 /* Update our send window.
2758  *
2759  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2760  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2761  */
2762 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2763                                  u32 ack_seq)
2764 {
2765         struct tcp_sock *tp = tcp_sk(sk);
2766         int flag = 0;
2767         u32 nwin = ntohs(tcp_hdr(skb)->window);
2768
2769         if (likely(!tcp_hdr(skb)->syn))
2770                 nwin <<= tp->rx_opt.snd_wscale;
2771
2772         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2773                 flag |= FLAG_WIN_UPDATE;
2774                 tcp_update_wl(tp, ack, ack_seq);
2775
2776                 if (tp->snd_wnd != nwin) {
2777                         tp->snd_wnd = nwin;
2778
2779                         /* Note, it is the only place, where
2780                          * fast path is recovered for sending TCP.
2781                          */
2782                         tp->pred_flags = 0;
2783                         tcp_fast_path_check(sk);
2784
2785                         if (nwin > tp->max_window) {
2786                                 tp->max_window = nwin;
2787                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2788                         }
2789                 }
2790         }
2791
2792         tp->snd_una = ack;
2793
2794         return flag;
2795 }
2796
2797 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2798  * continue in congestion avoidance.
2799  */
2800 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2801 {
2802         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2803         tp->snd_cwnd_cnt = 0;
2804         TCP_ECN_queue_cwr(tp);
2805         tcp_moderate_cwnd(tp);
2806 }
2807
2808 /* A conservative spurious RTO response algorithm: reduce cwnd using
2809  * rate halving and continue in congestion avoidance.
2810  */
2811 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2812 {
2813         tcp_enter_cwr(sk, 0);
2814 }
2815
2816 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2817 {
2818         if (flag&FLAG_ECE)
2819                 tcp_ratehalving_spur_to_response(sk);
2820         else
2821                 tcp_undo_cwr(sk, 1);
2822 }
2823
2824 /* F-RTO spurious RTO detection algorithm (RFC4138)
2825  *
2826  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2827  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2828  * window (but not to or beyond highest sequence sent before RTO):
2829  *   On First ACK,  send two new segments out.
2830  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2831  *                  algorithm is not part of the F-RTO detection algorithm
2832  *                  given in RFC4138 but can be selected separately).
2833  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2834  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2835  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2836  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2837  *
2838  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2839  * original window even after we transmit two new data segments.
2840  *
2841  * SACK version:
2842  *   on first step, wait until first cumulative ACK arrives, then move to
2843  *   the second step. In second step, the next ACK decides.
2844  *
2845  * F-RTO is implemented (mainly) in four functions:
2846  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2847  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2848  *     called when tcp_use_frto() showed green light
2849  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2850  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2851  *     to prove that the RTO is indeed spurious. It transfers the control
2852  *     from F-RTO to the conventional RTO recovery
2853  */
2854 static int tcp_process_frto(struct sock *sk, int flag)
2855 {
2856         struct tcp_sock *tp = tcp_sk(sk);
2857
2858         tcp_verify_left_out(tp);
2859
2860         /* Duplicate the behavior from Loss state (fastretrans_alert) */
2861         if (flag&FLAG_DATA_ACKED)
2862                 inet_csk(sk)->icsk_retransmits = 0;
2863
2864         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
2865             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
2866                 tp->undo_marker = 0;
2867
2868         if (!before(tp->snd_una, tp->frto_highmark)) {
2869                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2870                 return 1;
2871         }
2872
2873         if (!IsSackFrto() || tcp_is_reno(tp)) {
2874                 /* RFC4138 shortcoming in step 2; should also have case c):
2875                  * ACK isn't duplicate nor advances window, e.g., opposite dir
2876                  * data, winupdate
2877                  */
2878                 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2879                         return 1;
2880
2881                 if (!(flag&FLAG_DATA_ACKED)) {
2882                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2883                                             flag);
2884                         return 1;
2885                 }
2886         } else {
2887                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2888                         /* Prevent sending of new data. */
2889                         tp->snd_cwnd = min(tp->snd_cwnd,
2890                                            tcp_packets_in_flight(tp));
2891                         return 1;
2892                 }
2893
2894                 if ((tp->frto_counter >= 2) &&
2895                     (!(flag&FLAG_FORWARD_PROGRESS) ||
2896                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2897                         /* RFC4138 shortcoming (see comment above) */
2898                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2899                                 return 1;
2900
2901                         tcp_enter_frto_loss(sk, 3, flag);
2902                         return 1;
2903                 }
2904         }
2905
2906         if (tp->frto_counter == 1) {
2907                 /* Sending of the next skb must be allowed or no FRTO */
2908                 if (!tcp_send_head(sk) ||
2909                     after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2910                                      tp->snd_una + tp->snd_wnd)) {
2911                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2912                                             flag);
2913                         return 1;
2914                 }
2915
2916                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2917                 tp->frto_counter = 2;
2918                 return 1;
2919         } else {
2920                 switch (sysctl_tcp_frto_response) {
2921                 case 2:
2922                         tcp_undo_spur_to_response(sk, flag);
2923                         break;
2924                 case 1:
2925                         tcp_conservative_spur_to_response(tp);
2926                         break;
2927                 default:
2928                         tcp_ratehalving_spur_to_response(sk);
2929                         break;
2930                 }
2931                 tp->frto_counter = 0;
2932                 tp->undo_marker = 0;
2933         }
2934         return 0;
2935 }
2936
2937 /* This routine deals with incoming acks, but not outgoing ones. */
2938 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2939 {
2940         struct inet_connection_sock *icsk = inet_csk(sk);
2941         struct tcp_sock *tp = tcp_sk(sk);
2942         u32 prior_snd_una = tp->snd_una;
2943         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2944         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2945         u32 prior_in_flight;
2946         s32 seq_rtt;
2947         int prior_packets;
2948         int frto_cwnd = 0;
2949
2950         /* If the ack is newer than sent or older than previous acks
2951          * then we can probably ignore it.
2952          */
2953         if (after(ack, tp->snd_nxt))
2954                 goto uninteresting_ack;
2955
2956         if (before(ack, prior_snd_una))
2957                 goto old_ack;
2958
2959         if (after(ack, prior_snd_una))
2960                 flag |= FLAG_SND_UNA_ADVANCED;
2961
2962         if (sysctl_tcp_abc) {
2963                 if (icsk->icsk_ca_state < TCP_CA_CWR)
2964                         tp->bytes_acked += ack - prior_snd_una;
2965                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2966                         /* we assume just one segment left network */
2967                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2968         }
2969
2970         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2971                 /* Window is constant, pure forward advance.
2972                  * No more checks are required.
2973                  * Note, we use the fact that SND.UNA>=SND.WL2.
2974                  */
2975                 tcp_update_wl(tp, ack, ack_seq);
2976                 tp->snd_una = ack;
2977                 flag |= FLAG_WIN_UPDATE;
2978
2979                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2980
2981                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2982         } else {
2983                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2984                         flag |= FLAG_DATA;
2985                 else
2986                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2987
2988                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2989
2990                 if (TCP_SKB_CB(skb)->sacked)
2991                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2992
2993                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2994                         flag |= FLAG_ECE;
2995
2996                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2997         }
2998
2999         /* We passed data and got it acked, remove any soft error
3000          * log. Something worked...
3001          */
3002         sk->sk_err_soft = 0;
3003         tp->rcv_tstamp = tcp_time_stamp;
3004         prior_packets = tp->packets_out;
3005         if (!prior_packets)
3006                 goto no_queue;
3007
3008         prior_in_flight = tcp_packets_in_flight(tp);
3009
3010         /* See if we can take anything off of the retransmit queue. */
3011         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
3012
3013         if (tp->frto_counter)
3014                 frto_cwnd = tcp_process_frto(sk, flag);
3015
3016         if (tcp_ack_is_dubious(sk, flag)) {
3017                 /* Advance CWND, if state allows this. */
3018                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3019                     tcp_may_raise_cwnd(sk, flag))
3020                         tcp_cong_avoid(sk, ack, prior_in_flight, 0);
3021                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3022         } else {
3023                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3024                         tcp_cong_avoid(sk, ack, prior_in_flight, 1);
3025         }
3026
3027         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3028                 dst_confirm(sk->sk_dst_cache);
3029
3030         return 1;
3031
3032 no_queue:
3033         icsk->icsk_probes_out = 0;
3034
3035         /* If this ack opens up a zero window, clear backoff.  It was
3036          * being used to time the probes, and is probably far higher than
3037          * it needs to be for normal retransmission.
3038          */
3039         if (tcp_send_head(sk))
3040                 tcp_ack_probe(sk);
3041         return 1;
3042
3043 old_ack:
3044         if (TCP_SKB_CB(skb)->sacked)
3045                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3046
3047 uninteresting_ack:
3048         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3049         return 0;
3050 }
3051
3052
3053 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3054  * But, this can also be called on packets in the established flow when
3055  * the fast version below fails.
3056  */
3057 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3058 {
3059         unsigned char *ptr;
3060         struct tcphdr *th = tcp_hdr(skb);
3061         int length=(th->doff*4)-sizeof(struct tcphdr);
3062
3063         ptr = (unsigned char *)(th + 1);
3064         opt_rx->saw_tstamp = 0;
3065
3066         while (length > 0) {
3067                 int opcode=*ptr++;
3068                 int opsize;
3069
3070                 switch (opcode) {
3071                         case TCPOPT_EOL:
3072                                 return;
3073                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3074                                 length--;
3075                                 continue;
3076                         default:
3077                                 opsize=*ptr++;
3078                                 if (opsize < 2) /* "silly options" */
3079                                         return;
3080                                 if (opsize > length)
3081                                         return; /* don't parse partial options */
3082                                 switch (opcode) {
3083                                 case TCPOPT_MSS:
3084                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3085                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3086                                                 if (in_mss) {
3087                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3088                                                                 in_mss = opt_rx->user_mss;
3089                                                         opt_rx->mss_clamp = in_mss;
3090                                                 }
3091                                         }
3092                                         break;
3093                                 case TCPOPT_WINDOW:
3094                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3095                                                 if (sysctl_tcp_window_scaling) {
3096                                                         __u8 snd_wscale = *(__u8 *) ptr;
3097                                                         opt_rx->wscale_ok = 1;
3098                                                         if (snd_wscale > 14) {
3099                                                                 if (net_ratelimit())
3100                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3101                                                                                "scaling value %d >14 received.\n",
3102                                                                                snd_wscale);
3103                                                                 snd_wscale = 14;
3104                                                         }
3105                                                         opt_rx->snd_wscale = snd_wscale;
3106                                                 }
3107                                         break;
3108                                 case TCPOPT_TIMESTAMP:
3109                                         if (opsize==TCPOLEN_TIMESTAMP) {
3110                                                 if ((estab && opt_rx->tstamp_ok) ||
3111                                                     (!estab && sysctl_tcp_timestamps)) {
3112                                                         opt_rx->saw_tstamp = 1;
3113                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3114                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3115                                                 }
3116                                         }
3117                                         break;
3118                                 case TCPOPT_SACK_PERM:
3119                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3120                                                 if (sysctl_tcp_sack) {
3121                                                         opt_rx->sack_ok = 1;
3122                                                         tcp_sack_reset(opt_rx);
3123                                                 }
3124                                         }
3125                                         break;
3126
3127                                 case TCPOPT_SACK:
3128                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3129                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3130                                            opt_rx->sack_ok) {
3131                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3132                                         }
3133                                         break;
3134 #ifdef CONFIG_TCP_MD5SIG
3135                                 case TCPOPT_MD5SIG:
3136                                         /*
3137                                          * The MD5 Hash has already been
3138                                          * checked (see tcp_v{4,6}_do_rcv()).
3139                                          */
3140                                         break;
3141 #endif
3142                                 }
3143
3144                                 ptr+=opsize-2;
3145                                 length-=opsize;
3146                 }
3147         }
3148 }
3149
3150 /* Fast parse options. This hopes to only see timestamps.
3151  * If it is wrong it falls back on tcp_parse_options().
3152  */
3153 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3154                                   struct tcp_sock *tp)
3155 {
3156         if (th->doff == sizeof(struct tcphdr)>>2) {
3157                 tp->rx_opt.saw_tstamp = 0;
3158                 return 0;
3159         } else if (tp->rx_opt.tstamp_ok &&
3160                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3161                 __be32 *ptr = (__be32 *)(th + 1);
3162                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3163                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3164                         tp->rx_opt.saw_tstamp = 1;
3165                         ++ptr;
3166                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3167                         ++ptr;
3168                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3169                         return 1;
3170                 }
3171         }
3172         tcp_parse_options(skb, &tp->rx_opt, 1);
3173         return 1;
3174 }
3175
3176 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3177 {
3178         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3179         tp->rx_opt.ts_recent_stamp = get_seconds();
3180 }
3181
3182 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3183 {
3184         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3185                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3186                  * extra check below makes sure this can only happen
3187                  * for pure ACK frames.  -DaveM
3188                  *
3189                  * Not only, also it occurs for expired timestamps.
3190                  */
3191
3192                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3193                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3194                         tcp_store_ts_recent(tp);
3195         }
3196 }
3197
3198 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3199  *
3200  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3201  * it can pass through stack. So, the following predicate verifies that
3202  * this segment is not used for anything but congestion avoidance or
3203  * fast retransmit. Moreover, we even are able to eliminate most of such
3204  * second order effects, if we apply some small "replay" window (~RTO)
3205  * to timestamp space.
3206  *
3207  * All these measures still do not guarantee that we reject wrapped ACKs
3208  * on networks with high bandwidth, when sequence space is recycled fastly,
3209  * but it guarantees that such events will be very rare and do not affect
3210  * connection seriously. This doesn't look nice, but alas, PAWS is really
3211  * buggy extension.
3212  *
3213  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3214  * states that events when retransmit arrives after original data are rare.
3215  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3216  * the biggest problem on large power networks even with minor reordering.
3217  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3218  * up to bandwidth of 18Gigabit/sec. 8) ]
3219  */
3220
3221 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3222 {
3223         struct tcp_sock *tp = tcp_sk(sk);
3224         struct tcphdr *th = tcp_hdr(skb);
3225         u32 seq = TCP_SKB_CB(skb)->seq;
3226         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3227
3228         return (/* 1. Pure ACK with correct sequence number. */
3229                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3230
3231                 /* 2. ... and duplicate ACK. */
3232                 ack == tp->snd_una &&
3233
3234                 /* 3. ... and does not update window. */
3235                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3236
3237                 /* 4. ... and sits in replay window. */
3238                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3239 }
3240
3241 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3242 {
3243         const struct tcp_sock *tp = tcp_sk(sk);
3244         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3245                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3246                 !tcp_disordered_ack(sk, skb));
3247 }
3248
3249 /* Check segment sequence number for validity.
3250  *
3251  * Segment controls are considered valid, if the segment
3252  * fits to the window after truncation to the window. Acceptability
3253  * of data (and SYN, FIN, of course) is checked separately.
3254  * See tcp_data_queue(), for example.
3255  *
3256  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3257  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3258  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3259  * (borrowed from freebsd)
3260  */
3261
3262 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3263 {
3264         return  !before(end_seq, tp->rcv_wup) &&
3265                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3266 }
3267
3268 /* When we get a reset we do this. */
3269 static void tcp_reset(struct sock *sk)
3270 {
3271         /* We want the right error as BSD sees it (and indeed as we do). */
3272         switch (sk->sk_state) {
3273                 case TCP_SYN_SENT:
3274                         sk->sk_err = ECONNREFUSED;
3275                         break;
3276                 case TCP_CLOSE_WAIT:
3277                         sk->sk_err = EPIPE;
3278                         break;
3279                 case TCP_CLOSE:
3280                         return;
3281                 default:
3282                         sk->sk_err = ECONNRESET;
3283         }
3284
3285         if (!sock_flag(sk, SOCK_DEAD))
3286                 sk->sk_error_report(sk);
3287
3288         tcp_done(sk);
3289 }
3290
3291 /*
3292  *      Process the FIN bit. This now behaves as it is supposed to work
3293  *      and the FIN takes effect when it is validly part of sequence
3294  *      space. Not before when we get holes.
3295  *
3296  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3297  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3298  *      TIME-WAIT)
3299  *
3300  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3301  *      close and we go into CLOSING (and later onto TIME-WAIT)
3302  *
3303  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3304  */
3305 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3306 {
3307         struct tcp_sock *tp = tcp_sk(sk);
3308
3309         inet_csk_schedule_ack(sk);
3310
3311         sk->sk_shutdown |= RCV_SHUTDOWN;
3312         sock_set_flag(sk, SOCK_DONE);
3313
3314         switch (sk->sk_state) {
3315                 case TCP_SYN_RECV:
3316                 case TCP_ESTABLISHED:
3317                         /* Move to CLOSE_WAIT */
3318                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3319                         inet_csk(sk)->icsk_ack.pingpong = 1;
3320                         break;
3321
3322                 case TCP_CLOSE_WAIT:
3323                 case TCP_CLOSING:
3324                         /* Received a retransmission of the FIN, do
3325                          * nothing.
3326                          */
3327                         break;
3328                 case TCP_LAST_ACK:
3329                         /* RFC793: Remain in the LAST-ACK state. */
3330                         break;
3331
3332                 case TCP_FIN_WAIT1:
3333                         /* This case occurs when a simultaneous close
3334                          * happens, we must ack the received FIN and
3335                          * enter the CLOSING state.
3336                          */
3337                         tcp_send_ack(sk);
3338                         tcp_set_state(sk, TCP_CLOSING);
3339                         break;
3340                 case TCP_FIN_WAIT2:
3341                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3342                         tcp_send_ack(sk);
3343                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3344                         break;
3345                 default:
3346                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3347                          * cases we should never reach this piece of code.
3348                          */
3349                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3350                                __FUNCTION__, sk->sk_state);
3351                         break;
3352         }
3353
3354         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3355          * Probably, we should reset in this case. For now drop them.
3356          */
3357         __skb_queue_purge(&tp->out_of_order_queue);
3358         if (tcp_is_sack(tp))
3359                 tcp_sack_reset(&tp->rx_opt);
3360         sk_stream_mem_reclaim(sk);
3361
3362         if (!sock_flag(sk, SOCK_DEAD)) {
3363                 sk->sk_state_change(sk);
3364
3365                 /* Do not send POLL_HUP for half duplex close. */
3366                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3367                     sk->sk_state == TCP_CLOSE)
3368                         sk_wake_async(sk, 1, POLL_HUP);
3369                 else
3370                         sk_wake_async(sk, 1, POLL_IN);
3371         }
3372 }
3373
3374 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3375 {
3376         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3377                 if (before(seq, sp->start_seq))
3378                         sp->start_seq = seq;
3379                 if (after(end_seq, sp->end_seq))
3380                         sp->end_seq = end_seq;
3381                 return 1;
3382         }
3383         return 0;
3384 }
3385
3386 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3387 {
3388         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3389                 if (before(seq, tp->rcv_nxt))
3390                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3391                 else
3392                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3393
3394                 tp->rx_opt.dsack = 1;
3395                 tp->duplicate_sack[0].start_seq = seq;
3396                 tp->duplicate_sack[0].end_seq = end_seq;
3397                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3398         }
3399 }
3400
3401 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3402 {
3403         if (!tp->rx_opt.dsack)
3404                 tcp_dsack_set(tp, seq, end_seq);
3405         else
3406                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3407 }
3408
3409 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3410 {
3411         struct tcp_sock *tp = tcp_sk(sk);
3412
3413         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3414             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3415                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3416                 tcp_enter_quickack_mode(sk);
3417
3418                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3419                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3420
3421                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3422                                 end_seq = tp->rcv_nxt;
3423                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3424                 }
3425         }
3426
3427         tcp_send_ack(sk);
3428 }
3429
3430 /* These routines update the SACK block as out-of-order packets arrive or
3431  * in-order packets close up the sequence space.
3432  */
3433 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3434 {
3435         int this_sack;
3436         struct tcp_sack_block *sp = &tp->selective_acks[0];
3437         struct tcp_sack_block *swalk = sp+1;
3438
3439         /* See if the recent change to the first SACK eats into
3440          * or hits the sequence space of other SACK blocks, if so coalesce.
3441          */
3442         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3443                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3444                         int i;
3445
3446                         /* Zap SWALK, by moving every further SACK up by one slot.
3447                          * Decrease num_sacks.
3448                          */
3449                         tp->rx_opt.num_sacks--;
3450                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3451                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3452                                 sp[i] = sp[i+1];
3453                         continue;
3454                 }
3455                 this_sack++, swalk++;
3456         }
3457 }
3458
3459 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3460 {
3461         __u32 tmp;
3462
3463         tmp = sack1->start_seq;
3464         sack1->start_seq = sack2->start_seq;
3465         sack2->start_seq = tmp;
3466
3467         tmp = sack1->end_seq;
3468         sack1->end_seq = sack2->end_seq;
3469         sack2->end_seq = tmp;
3470 }
3471
3472 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3473 {
3474         struct tcp_sock *tp = tcp_sk(sk);
3475         struct tcp_sack_block *sp = &tp->selective_acks[0];
3476         int cur_sacks = tp->rx_opt.num_sacks;
3477         int this_sack;
3478
3479         if (!cur_sacks)
3480                 goto new_sack;
3481
3482         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3483                 if (tcp_sack_extend(sp, seq, end_seq)) {
3484                         /* Rotate this_sack to the first one. */
3485                         for (; this_sack>0; this_sack--, sp--)
3486                                 tcp_sack_swap(sp, sp-1);
3487                         if (cur_sacks > 1)
3488                                 tcp_sack_maybe_coalesce(tp);
3489                         return;
3490                 }
3491         }
3492
3493         /* Could not find an adjacent existing SACK, build a new one,
3494          * put it at the front, and shift everyone else down.  We
3495          * always know there is at least one SACK present already here.
3496          *
3497          * If the sack array is full, forget about the last one.
3498          */
3499         if (this_sack >= 4) {
3500                 this_sack--;
3501                 tp->rx_opt.num_sacks--;
3502                 sp--;
3503         }
3504         for (; this_sack > 0; this_sack--, sp--)
3505                 *sp = *(sp-1);
3506
3507 new_sack:
3508         /* Build the new head SACK, and we're done. */
3509         sp->start_seq = seq;
3510         sp->end_seq = end_seq;
3511         tp->rx_opt.num_sacks++;
3512         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3513 }
3514
3515 /* RCV.NXT advances, some SACKs should be eaten. */
3516
3517 static void tcp_sack_remove(struct tcp_sock *tp)
3518 {
3519         struct tcp_sack_block *sp = &tp->selective_acks[0];
3520         int num_sacks = tp->rx_opt.num_sacks;
3521         int this_sack;
3522
3523         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3524         if (skb_queue_empty(&tp->out_of_order_queue)) {
3525                 tp->rx_opt.num_sacks = 0;
3526                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3527                 return;
3528         }
3529
3530         for (this_sack = 0; this_sack < num_sacks; ) {
3531                 /* Check if the start of the sack is covered by RCV.NXT. */
3532                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3533                         int i;
3534
3535                         /* RCV.NXT must cover all the block! */
3536                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3537
3538                         /* Zap this SACK, by moving forward any other SACKS. */
3539                         for (i=this_sack+1; i < num_sacks; i++)
3540                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3541                         num_sacks--;
3542                         continue;
3543                 }
3544                 this_sack++;
3545                 sp++;
3546         }
3547         if (num_sacks != tp->rx_opt.num_sacks) {
3548                 tp->rx_opt.num_sacks = num_sacks;
3549                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3550         }
3551 }
3552
3553 /* This one checks to see if we can put data from the
3554  * out_of_order queue into the receive_queue.
3555  */
3556 static void tcp_ofo_queue(struct sock *sk)
3557 {
3558         struct tcp_sock *tp = tcp_sk(sk);
3559         __u32 dsack_high = tp->rcv_nxt;
3560         struct sk_buff *skb;
3561
3562         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3563                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3564                         break;
3565
3566                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3567                         __u32 dsack = dsack_high;
3568                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3569                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3570                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3571                 }
3572
3573                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3574                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3575                         __skb_unlink(skb, &tp->out_of_order_queue);
3576                         __kfree_skb(skb);
3577                         continue;
3578                 }
3579                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3580                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3581                            TCP_SKB_CB(skb)->end_seq);
3582
3583                 __skb_unlink(skb, &tp->out_of_order_queue);
3584                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3585                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3586                 if (tcp_hdr(skb)->fin)
3587                         tcp_fin(skb, sk, tcp_hdr(skb));
3588         }
3589 }
3590
3591 static int tcp_prune_queue(struct sock *sk);
3592
3593 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3594 {
3595         struct tcphdr *th = tcp_hdr(skb);
3596         struct tcp_sock *tp = tcp_sk(sk);
3597         int eaten = -1;
3598
3599         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3600                 goto drop;
3601
3602         __skb_pull(skb, th->doff*4);
3603
3604         TCP_ECN_accept_cwr(tp, skb);
3605
3606         if (tp->rx_opt.dsack) {
3607                 tp->rx_opt.dsack = 0;
3608                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3609                                                     4 - tp->rx_opt.tstamp_ok);
3610         }
3611
3612         /*  Queue data for delivery to the user.
3613          *  Packets in sequence go to the receive queue.
3614          *  Out of sequence packets to the out_of_order_queue.
3615          */
3616         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3617                 if (tcp_receive_window(tp) == 0)
3618                         goto out_of_window;
3619
3620                 /* Ok. In sequence. In window. */
3621                 if (tp->ucopy.task == current &&
3622                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3623                     sock_owned_by_user(sk) && !tp->urg_data) {
3624                         int chunk = min_t(unsigned int, skb->len,
3625                                                         tp->ucopy.len);
3626
3627                         __set_current_state(TASK_RUNNING);
3628
3629                         local_bh_enable();
3630                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3631                                 tp->ucopy.len -= chunk;
3632                                 tp->copied_seq += chunk;
3633                                 eaten = (chunk == skb->len && !th->fin);
3634                                 tcp_rcv_space_adjust(sk);
3635                         }
3636                         local_bh_disable();
3637                 }
3638
3639                 if (eaten <= 0) {
3640 queue_and_out:
3641                         if (eaten < 0 &&
3642                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3643                              !sk_stream_rmem_schedule(sk, skb))) {
3644                                 if (tcp_prune_queue(sk) < 0 ||
3645                                     !sk_stream_rmem_schedule(sk, skb))
3646                                         goto drop;
3647                         }
3648                         sk_stream_set_owner_r(skb, sk);
3649                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3650                 }
3651                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3652                 if (skb->len)
3653                         tcp_event_data_recv(sk, skb);
3654                 if (th->fin)
3655                         tcp_fin(skb, sk, th);
3656
3657                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3658                         tcp_ofo_queue(sk);
3659
3660                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3661                          * gap in queue is filled.
3662                          */
3663                         if (skb_queue_empty(&tp->out_of_order_queue))
3664                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3665                 }
3666
3667                 if (tp->rx_opt.num_sacks)
3668                         tcp_sack_remove(tp);
3669
3670                 tcp_fast_path_check(sk);
3671
3672                 if (eaten > 0)
3673                         __kfree_skb(skb);
3674                 else if (!sock_flag(sk, SOCK_DEAD))
3675                         sk->sk_data_ready(sk, 0);
3676                 return;
3677         }
3678
3679         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3680                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3681                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3682                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3683
3684 out_of_window:
3685                 tcp_enter_quickack_mode(sk);
3686                 inet_csk_schedule_ack(sk);
3687 drop:
3688                 __kfree_skb(skb);
3689                 return;
3690         }
3691
3692         /* Out of window. F.e. zero window probe. */
3693         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3694                 goto out_of_window;
3695
3696         tcp_enter_quickack_mode(sk);
3697
3698         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3699                 /* Partial packet, seq < rcv_next < end_seq */
3700                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3701                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3702                            TCP_SKB_CB(skb)->end_seq);
3703
3704                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3705
3706                 /* If window is closed, drop tail of packet. But after
3707                  * remembering D-SACK for its head made in previous line.
3708                  */
3709                 if (!tcp_receive_window(tp))
3710                         goto out_of_window;
3711                 goto queue_and_out;
3712         }
3713
3714         TCP_ECN_check_ce(tp, skb);
3715
3716         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3717             !sk_stream_rmem_schedule(sk, skb)) {
3718                 if (tcp_prune_queue(sk) < 0 ||
3719                     !sk_stream_rmem_schedule(sk, skb))
3720                         goto drop;
3721         }
3722
3723         /* Disable header prediction. */
3724         tp->pred_flags = 0;
3725         inet_csk_schedule_ack(sk);
3726
3727         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3728                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3729
3730         sk_stream_set_owner_r(skb, sk);
3731
3732         if (!skb_peek(&tp->out_of_order_queue)) {
3733                 /* Initial out of order segment, build 1 SACK. */
3734                 if (tcp_is_sack(tp)) {
3735                         tp->rx_opt.num_sacks = 1;
3736                         tp->rx_opt.dsack     = 0;
3737                         tp->rx_opt.eff_sacks = 1;
3738                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3739                         tp->selective_acks[0].end_seq =
3740                                                 TCP_SKB_CB(skb)->end_seq;
3741                 }
3742                 __skb_queue_head(&tp->out_of_order_queue,skb);
3743         } else {
3744                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3745                 u32 seq = TCP_SKB_CB(skb)->seq;
3746                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3747
3748                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3749                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3750
3751                         if (!tp->rx_opt.num_sacks ||
3752                             tp->selective_acks[0].end_seq != seq)
3753                                 goto add_sack;
3754
3755                         /* Common case: data arrive in order after hole. */
3756                         tp->selective_acks[0].end_seq = end_seq;
3757                         return;
3758                 }
3759
3760                 /* Find place to insert this segment. */
3761                 do {
3762                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3763                                 break;
3764                 } while ((skb1 = skb1->prev) !=
3765                          (struct sk_buff*)&tp->out_of_order_queue);
3766
3767                 /* Do skb overlap to previous one? */
3768                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3769                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3770                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3771                                 /* All the bits are present. Drop. */
3772                                 __kfree_skb(skb);
3773                                 tcp_dsack_set(tp, seq, end_seq);
3774                                 goto add_sack;
3775                         }
3776                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3777                                 /* Partial overlap. */
3778                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3779                         } else {
3780                                 skb1 = skb1->prev;
3781                         }
3782                 }
3783                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3784
3785                 /* And clean segments covered by new one as whole. */
3786                 while ((skb1 = skb->next) !=
3787                        (struct sk_buff*)&tp->out_of_order_queue &&
3788                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3789                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3790                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3791                                break;
3792                        }
3793                        __skb_unlink(skb1, &tp->out_of_order_queue);
3794                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3795                        __kfree_skb(skb1);
3796                 }
3797
3798 add_sack:
3799                 if (tcp_is_sack(tp))
3800                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3801         }
3802 }
3803
3804 /* Collapse contiguous sequence of skbs head..tail with
3805  * sequence numbers start..end.
3806  * Segments with FIN/SYN are not collapsed (only because this
3807  * simplifies code)
3808  */
3809 static void
3810 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3811              struct sk_buff *head, struct sk_buff *tail,
3812              u32 start, u32 end)
3813 {
3814         struct sk_buff *skb;
3815
3816         /* First, check that queue is collapsible and find
3817          * the point where collapsing can be useful. */
3818         for (skb = head; skb != tail; ) {
3819                 /* No new bits? It is possible on ofo queue. */
3820                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3821                         struct sk_buff *next = skb->next;
3822                         __skb_unlink(skb, list);
3823                         __kfree_skb(skb);
3824                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3825                         skb = next;
3826                         continue;
3827                 }
3828
3829                 /* The first skb to collapse is:
3830                  * - not SYN/FIN and
3831                  * - bloated or contains data before "start" or
3832                  *   overlaps to the next one.
3833                  */
3834                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3835                     (tcp_win_from_space(skb->truesize) > skb->len ||
3836                      before(TCP_SKB_CB(skb)->seq, start) ||
3837                      (skb->next != tail &&
3838                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3839                         break;
3840
3841                 /* Decided to skip this, advance start seq. */
3842                 start = TCP_SKB_CB(skb)->end_seq;
3843                 skb = skb->next;
3844         }
3845         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3846                 return;
3847
3848         while (before(start, end)) {
3849                 struct sk_buff *nskb;
3850                 int header = skb_headroom(skb);
3851                 int copy = SKB_MAX_ORDER(header, 0);
3852
3853                 /* Too big header? This can happen with IPv6. */
3854                 if (copy < 0)
3855                         return;
3856                 if (end-start < copy)
3857                         copy = end-start;
3858                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3859                 if (!nskb)
3860                         return;
3861
3862                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3863                 skb_set_network_header(nskb, (skb_network_header(skb) -
3864                                               skb->head));
3865                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3866                                                 skb->head));
3867                 skb_reserve(nskb, header);
3868                 memcpy(nskb->head, skb->head, header);
3869                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3870                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3871                 __skb_insert(nskb, skb->prev, skb, list);
3872                 sk_stream_set_owner_r(nskb, sk);
3873
3874                 /* Copy data, releasing collapsed skbs. */
3875                 while (copy > 0) {
3876                         int offset = start - TCP_SKB_CB(skb)->seq;
3877                         int size = TCP_SKB_CB(skb)->end_seq - start;
3878
3879                         BUG_ON(offset < 0);
3880                         if (size > 0) {
3881                                 size = min(copy, size);
3882                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3883                                         BUG();
3884                                 TCP_SKB_CB(nskb)->end_seq += size;
3885                                 copy -= size;
3886                                 start += size;
3887                         }
3888                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3889                                 struct sk_buff *next = skb->next;
3890                                 __skb_unlink(skb, list);
3891                                 __kfree_skb(skb);
3892                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3893                                 skb = next;
3894                                 if (skb == tail ||
3895                                     tcp_hdr(skb)->syn ||
3896                                     tcp_hdr(skb)->fin)
3897                                         return;
3898                         }
3899                 }
3900         }
3901 }
3902
3903 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3904  * and tcp_collapse() them until all the queue is collapsed.
3905  */
3906 static void tcp_collapse_ofo_queue(struct sock *sk)
3907 {
3908         struct tcp_sock *tp = tcp_sk(sk);
3909         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3910         struct sk_buff *head;
3911         u32 start, end;
3912
3913         if (skb == NULL)
3914                 return;
3915
3916         start = TCP_SKB_CB(skb)->seq;
3917         end = TCP_SKB_CB(skb)->end_seq;
3918         head = skb;
3919
3920         for (;;) {
3921                 skb = skb->next;
3922
3923                 /* Segment is terminated when we see gap or when
3924                  * we are at the end of all the queue. */
3925                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3926                     after(TCP_SKB_CB(skb)->seq, end) ||
3927                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3928                         tcp_collapse(sk, &tp->out_of_order_queue,
3929                                      head, skb, start, end);
3930                         head = skb;
3931                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3932                                 break;
3933                         /* Start new segment */
3934                         start = TCP_SKB_CB(skb)->seq;
3935                         end = TCP_SKB_CB(skb)->end_seq;
3936                 } else {
3937                         if (before(TCP_SKB_CB(skb)->seq, start))
3938                                 start = TCP_SKB_CB(skb)->seq;
3939                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3940                                 end = TCP_SKB_CB(skb)->end_seq;
3941                 }
3942         }
3943 }
3944
3945 /* Reduce allocated memory if we can, trying to get
3946  * the socket within its memory limits again.
3947  *
3948  * Return less than zero if we should start dropping frames
3949  * until the socket owning process reads some of the data
3950  * to stabilize the situation.
3951  */
3952 static int tcp_prune_queue(struct sock *sk)
3953 {
3954         struct tcp_sock *tp = tcp_sk(sk);
3955
3956         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3957
3958         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3959
3960         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3961                 tcp_clamp_window(sk);
3962         else if (tcp_memory_pressure)
3963                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3964
3965         tcp_collapse_ofo_queue(sk);
3966         tcp_collapse(sk, &sk->sk_receive_queue,
3967                      sk->sk_receive_queue.next,
3968                      (struct sk_buff*)&sk->sk_receive_queue,
3969                      tp->copied_seq, tp->rcv_nxt);
3970         sk_stream_mem_reclaim(sk);
3971
3972         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3973                 return 0;
3974
3975         /* Collapsing did not help, destructive actions follow.
3976          * This must not ever occur. */
3977
3978         /* First, purge the out_of_order queue. */
3979         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3980                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3981                 __skb_queue_purge(&tp->out_of_order_queue);
3982
3983                 /* Reset SACK state.  A conforming SACK implementation will
3984                  * do the same at a timeout based retransmit.  When a connection
3985                  * is in a sad state like this, we care only about integrity
3986                  * of the connection not performance.
3987                  */
3988                 if (tcp_is_sack(tp))
3989                         tcp_sack_reset(&tp->rx_opt);
3990                 sk_stream_mem_reclaim(sk);
3991         }
3992
3993         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3994                 return 0;
3995
3996         /* If we are really being abused, tell the caller to silently
3997          * drop receive data on the floor.  It will get retransmitted
3998          * and hopefully then we'll have sufficient space.
3999          */
4000         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4001
4002         /* Massive buffer overcommit. */
4003         tp->pred_flags = 0;
4004         return -1;
4005 }
4006
4007
4008 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4009  * As additional protections, we do not touch cwnd in retransmission phases,
4010  * and if application hit its sndbuf limit recently.
4011  */
4012 void tcp_cwnd_application_limited(struct sock *sk)
4013 {
4014         struct tcp_sock *tp = tcp_sk(sk);
4015
4016         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4017             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4018                 /* Limited by application or receiver window. */
4019                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4020                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4021                 if (win_used < tp->snd_cwnd) {
4022                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4023                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4024                 }
4025                 tp->snd_cwnd_used = 0;
4026         }
4027         tp->snd_cwnd_stamp = tcp_time_stamp;
4028 }
4029
4030 static int tcp_should_expand_sndbuf(struct sock *sk)
4031 {
4032         struct tcp_sock *tp = tcp_sk(sk);
4033
4034         /* If the user specified a specific send buffer setting, do
4035          * not modify it.
4036          */
4037         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4038                 return 0;
4039
4040         /* If we are under global TCP memory pressure, do not expand.  */
4041         if (tcp_memory_pressure)
4042                 return 0;
4043
4044         /* If we are under soft global TCP memory pressure, do not expand.  */
4045         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4046                 return 0;
4047
4048         /* If we filled the congestion window, do not expand.  */
4049         if (tp->packets_out >= tp->snd_cwnd)
4050                 return 0;
4051
4052         return 1;
4053 }
4054
4055 /* When incoming ACK allowed to free some skb from write_queue,
4056  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4057  * on the exit from tcp input handler.
4058  *
4059  * PROBLEM: sndbuf expansion does not work well with largesend.
4060  */
4061 static void tcp_new_space(struct sock *sk)
4062 {
4063         struct tcp_sock *tp = tcp_sk(sk);
4064
4065         if (tcp_should_expand_sndbuf(sk)) {
4066                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4067                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4068                     demanded = max_t(unsigned int, tp->snd_cwnd,
4069                                                    tp->reordering + 1);
4070                 sndmem *= 2*demanded;
4071                 if (sndmem > sk->sk_sndbuf)
4072                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4073                 tp->snd_cwnd_stamp = tcp_time_stamp;
4074         }
4075
4076         sk->sk_write_space(sk);
4077 }
4078
4079 static void tcp_check_space(struct sock *sk)
4080 {
4081         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4082                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4083                 if (sk->sk_socket &&
4084                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4085                         tcp_new_space(sk);
4086         }
4087 }
4088
4089 static inline void tcp_data_snd_check(struct sock *sk)
4090 {
4091         tcp_push_pending_frames(sk);
4092         tcp_check_space(sk);
4093 }
4094
4095 /*
4096  * Check if sending an ack is needed.
4097  */
4098 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4099 {
4100         struct tcp_sock *tp = tcp_sk(sk);
4101
4102             /* More than one full frame received... */
4103         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4104              /* ... and right edge of window advances far enough.
4105               * (tcp_recvmsg() will send ACK otherwise). Or...
4106               */
4107              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4108             /* We ACK each frame or... */
4109             tcp_in_quickack_mode(sk) ||
4110             /* We have out of order data. */
4111             (ofo_possible &&
4112              skb_peek(&tp->out_of_order_queue))) {
4113                 /* Then ack it now */
4114                 tcp_send_ack(sk);
4115         } else {
4116                 /* Else, send delayed ack. */
4117                 tcp_send_delayed_ack(sk);
4118         }
4119 }
4120
4121 static inline void tcp_ack_snd_check(struct sock *sk)
4122 {
4123         if (!inet_csk_ack_scheduled(sk)) {
4124                 /* We sent a data segment already. */
4125                 return;
4126         }
4127         __tcp_ack_snd_check(sk, 1);
4128 }
4129
4130 /*
4131  *      This routine is only called when we have urgent data
4132  *      signaled. Its the 'slow' part of tcp_urg. It could be
4133  *      moved inline now as tcp_urg is only called from one
4134  *      place. We handle URGent data wrong. We have to - as
4135  *      BSD still doesn't use the correction from RFC961.
4136  *      For 1003.1g we should support a new option TCP_STDURG to permit
4137  *      either form (or just set the sysctl tcp_stdurg).
4138  */
4139
4140 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4141 {
4142         struct tcp_sock *tp = tcp_sk(sk);
4143         u32 ptr = ntohs(th->urg_ptr);
4144
4145         if (ptr && !sysctl_tcp_stdurg)
4146                 ptr--;
4147         ptr += ntohl(th->seq);
4148
4149         /* Ignore urgent data that we've already seen and read. */
4150         if (after(tp->copied_seq, ptr))
4151                 return;
4152
4153         /* Do not replay urg ptr.
4154          *
4155          * NOTE: interesting situation not covered by specs.
4156          * Misbehaving sender may send urg ptr, pointing to segment,
4157          * which we already have in ofo queue. We are not able to fetch
4158          * such data and will stay in TCP_URG_NOTYET until will be eaten
4159          * by recvmsg(). Seems, we are not obliged to handle such wicked
4160          * situations. But it is worth to think about possibility of some
4161          * DoSes using some hypothetical application level deadlock.
4162          */
4163         if (before(ptr, tp->rcv_nxt))
4164                 return;
4165
4166         /* Do we already have a newer (or duplicate) urgent pointer? */
4167         if (tp->urg_data && !after(ptr, tp->urg_seq))
4168                 return;
4169
4170         /* Tell the world about our new urgent pointer. */
4171         sk_send_sigurg(sk);
4172
4173         /* We may be adding urgent data when the last byte read was
4174          * urgent. To do this requires some care. We cannot just ignore
4175          * tp->copied_seq since we would read the last urgent byte again
4176          * as data, nor can we alter copied_seq until this data arrives
4177          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4178          *
4179          * NOTE. Double Dutch. Rendering to plain English: author of comment
4180          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4181          * and expect that both A and B disappear from stream. This is _wrong_.
4182          * Though this happens in BSD with high probability, this is occasional.
4183          * Any application relying on this is buggy. Note also, that fix "works"
4184          * only in this artificial test. Insert some normal data between A and B and we will
4185          * decline of BSD again. Verdict: it is better to remove to trap
4186          * buggy users.
4187          */
4188         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4189             !sock_flag(sk, SOCK_URGINLINE) &&
4190             tp->copied_seq != tp->rcv_nxt) {
4191                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4192                 tp->copied_seq++;
4193                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4194                         __skb_unlink(skb, &sk->sk_receive_queue);
4195                         __kfree_skb(skb);
4196                 }
4197         }
4198
4199         tp->urg_data   = TCP_URG_NOTYET;
4200         tp->urg_seq    = ptr;
4201
4202         /* Disable header prediction. */
4203         tp->pred_flags = 0;
4204 }
4205
4206 /* This is the 'fast' part of urgent handling. */
4207 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4208 {
4209         struct tcp_sock *tp = tcp_sk(sk);
4210
4211         /* Check if we get a new urgent pointer - normally not. */
4212         if (th->urg)
4213                 tcp_check_urg(sk,th);
4214
4215         /* Do we wait for any urgent data? - normally not... */
4216         if (tp->urg_data == TCP_URG_NOTYET) {
4217                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4218                           th->syn;
4219
4220                 /* Is the urgent pointer pointing into this packet? */
4221                 if (ptr < skb->len) {
4222                         u8 tmp;
4223                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4224                                 BUG();
4225                         tp->urg_data = TCP_URG_VALID | tmp;
4226                         if (!sock_flag(sk, SOCK_DEAD))
4227                                 sk->sk_data_ready(sk, 0);
4228                 }
4229         }
4230 }
4231
4232 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4233 {
4234         struct tcp_sock *tp = tcp_sk(sk);
4235         int chunk = skb->len - hlen;
4236         int err;
4237
4238         local_bh_enable();
4239         if (skb_csum_unnecessary(skb))
4240                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4241         else
4242                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4243                                                        tp->ucopy.iov);
4244
4245         if (!err) {
4246                 tp->ucopy.len -= chunk;
4247                 tp->copied_seq += chunk;
4248                 tcp_rcv_space_adjust(sk);
4249         }
4250
4251         local_bh_disable();
4252         return err;
4253 }
4254
4255 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4256 {
4257         __sum16 result;
4258
4259         if (sock_owned_by_user(sk)) {
4260                 local_bh_enable();
4261                 result = __tcp_checksum_complete(skb);
4262                 local_bh_disable();
4263         } else {
4264                 result = __tcp_checksum_complete(skb);
4265         }
4266         return result;
4267 }
4268
4269 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4270 {
4271         return !skb_csum_unnecessary(skb) &&
4272                 __tcp_checksum_complete_user(sk, skb);
4273 }
4274
4275 #ifdef CONFIG_NET_DMA
4276 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4277 {
4278         struct tcp_sock *tp = tcp_sk(sk);
4279         int chunk = skb->len - hlen;
4280         int dma_cookie;
4281         int copied_early = 0;
4282
4283         if (tp->ucopy.wakeup)
4284                 return 0;
4285
4286         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4287                 tp->ucopy.dma_chan = get_softnet_dma();
4288
4289         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4290
4291                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4292                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4293
4294                 if (dma_cookie < 0)
4295                         goto out;
4296
4297                 tp->ucopy.dma_cookie = dma_cookie;
4298                 copied_early = 1;
4299
4300                 tp->ucopy.len -= chunk;
4301                 tp->copied_seq += chunk;
4302                 tcp_rcv_space_adjust(sk);
4303
4304                 if ((tp->ucopy.len == 0) ||
4305                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4306                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4307                         tp->ucopy.wakeup = 1;
4308                         sk->sk_data_ready(sk, 0);
4309                 }
4310         } else if (chunk > 0) {
4311                 tp->ucopy.wakeup = 1;
4312                 sk->sk_data_ready(sk, 0);
4313         }
4314 out:
4315         return copied_early;
4316 }
4317 #endif /* CONFIG_NET_DMA */
4318
4319 /*
4320  *      TCP receive function for the ESTABLISHED state.
4321  *
4322  *      It is split into a fast path and a slow path. The fast path is
4323  *      disabled when:
4324  *      - A zero window was announced from us - zero window probing
4325  *        is only handled properly in the slow path.
4326  *      - Out of order segments arrived.
4327  *      - Urgent data is expected.
4328  *      - There is no buffer space left
4329  *      - Unexpected TCP flags/window values/header lengths are received
4330  *        (detected by checking the TCP header against pred_flags)
4331  *      - Data is sent in both directions. Fast path only supports pure senders
4332  *        or pure receivers (this means either the sequence number or the ack
4333  *        value must stay constant)
4334  *      - Unexpected TCP option.
4335  *
4336  *      When these conditions are not satisfied it drops into a standard
4337  *      receive procedure patterned after RFC793 to handle all cases.
4338  *      The first three cases are guaranteed by proper pred_flags setting,
4339  *      the rest is checked inline. Fast processing is turned on in
4340  *      tcp_data_queue when everything is OK.
4341  */
4342 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4343                         struct tcphdr *th, unsigned len)
4344 {
4345         struct tcp_sock *tp = tcp_sk(sk);
4346
4347         /*
4348          *      Header prediction.
4349          *      The code loosely follows the one in the famous
4350          *      "30 instruction TCP receive" Van Jacobson mail.
4351          *
4352          *      Van's trick is to deposit buffers into socket queue
4353          *      on a device interrupt, to call tcp_recv function
4354          *      on the receive process context and checksum and copy
4355          *      the buffer to user space. smart...
4356          *
4357          *      Our current scheme is not silly either but we take the
4358          *      extra cost of the net_bh soft interrupt processing...
4359          *      We do checksum and copy also but from device to kernel.
4360          */
4361
4362         tp->rx_opt.saw_tstamp = 0;
4363
4364         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4365          *      if header_prediction is to be made
4366          *      'S' will always be tp->tcp_header_len >> 2
4367          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4368          *  turn it off (when there are holes in the receive
4369          *       space for instance)
4370          *      PSH flag is ignored.
4371          */
4372
4373         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4374                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4375                 int tcp_header_len = tp->tcp_header_len;
4376
4377                 /* Timestamp header prediction: tcp_header_len
4378                  * is automatically equal to th->doff*4 due to pred_flags
4379                  * match.
4380                  */
4381
4382                 /* Check timestamp */
4383                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4384                         __be32 *ptr = (__be32 *)(th + 1);
4385
4386                         /* No? Slow path! */
4387                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4388                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4389                                 goto slow_path;
4390
4391                         tp->rx_opt.saw_tstamp = 1;
4392                         ++ptr;
4393                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4394                         ++ptr;
4395                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4396
4397                         /* If PAWS failed, check it more carefully in slow path */
4398                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4399                                 goto slow_path;
4400
4401                         /* DO NOT update ts_recent here, if checksum fails
4402                          * and timestamp was corrupted part, it will result
4403                          * in a hung connection since we will drop all
4404                          * future packets due to the PAWS test.
4405                          */
4406                 }
4407
4408                 if (len <= tcp_header_len) {
4409                         /* Bulk data transfer: sender */
4410                         if (len == tcp_header_len) {
4411                                 /* Predicted packet is in window by definition.
4412                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4413                                  * Hence, check seq<=rcv_wup reduces to:
4414                                  */
4415                                 if (tcp_header_len ==
4416                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4417                                     tp->rcv_nxt == tp->rcv_wup)
4418                                         tcp_store_ts_recent(tp);
4419
4420                                 /* We know that such packets are checksummed
4421                                  * on entry.
4422                                  */
4423                                 tcp_ack(sk, skb, 0);
4424                                 __kfree_skb(skb);
4425                                 tcp_data_snd_check(sk);
4426                                 return 0;
4427                         } else { /* Header too small */
4428                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4429                                 goto discard;
4430                         }
4431                 } else {
4432                         int eaten = 0;
4433                         int copied_early = 0;
4434
4435                         if (tp->copied_seq == tp->rcv_nxt &&
4436                             len - tcp_header_len <= tp->ucopy.len) {
4437 #ifdef CONFIG_NET_DMA
4438                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4439                                         copied_early = 1;
4440                                         eaten = 1;
4441                                 }
4442 #endif
4443                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4444                                         __set_current_state(TASK_RUNNING);
4445
4446                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4447                                                 eaten = 1;
4448                                 }
4449                                 if (eaten) {
4450                                         /* Predicted packet is in window by definition.
4451                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4452                                          * Hence, check seq<=rcv_wup reduces to:
4453                                          */
4454                                         if (tcp_header_len ==
4455                                             (sizeof(struct tcphdr) +
4456                                              TCPOLEN_TSTAMP_ALIGNED) &&
4457                                             tp->rcv_nxt == tp->rcv_wup)
4458                                                 tcp_store_ts_recent(tp);
4459
4460                                         tcp_rcv_rtt_measure_ts(sk, skb);
4461
4462                                         __skb_pull(skb, tcp_header_len);
4463                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4464                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4465                                 }
4466                                 if (copied_early)
4467                                         tcp_cleanup_rbuf(sk, skb->len);
4468                         }
4469                         if (!eaten) {
4470                                 if (tcp_checksum_complete_user(sk, skb))
4471                                         goto csum_error;
4472
4473                                 /* Predicted packet is in window by definition.
4474                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4475                                  * Hence, check seq<=rcv_wup reduces to:
4476                                  */
4477                                 if (tcp_header_len ==
4478                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4479                                     tp->rcv_nxt == tp->rcv_wup)
4480                                         tcp_store_ts_recent(tp);
4481
4482                                 tcp_rcv_rtt_measure_ts(sk, skb);
4483
4484                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4485                                         goto step5;
4486
4487                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4488
4489                                 /* Bulk data transfer: receiver */
4490                                 __skb_pull(skb,tcp_header_len);
4491                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4492                                 sk_stream_set_owner_r(skb, sk);
4493                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4494                         }
4495
4496                         tcp_event_data_recv(sk, skb);
4497
4498                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4499                                 /* Well, only one small jumplet in fast path... */
4500                                 tcp_ack(sk, skb, FLAG_DATA);
4501                                 tcp_data_snd_check(sk);
4502                                 if (!inet_csk_ack_scheduled(sk))
4503                                         goto no_ack;
4504                         }
4505
4506                         __tcp_ack_snd_check(sk, 0);
4507 no_ack:
4508 #ifdef CONFIG_NET_DMA
4509                         if (copied_early)
4510                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4511                         else
4512 #endif
4513                         if (eaten)
4514                                 __kfree_skb(skb);
4515                         else
4516                                 sk->sk_data_ready(sk, 0);
4517                         return 0;
4518                 }
4519         }
4520
4521 slow_path:
4522         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4523                 goto csum_error;
4524
4525         /*
4526          * RFC1323: H1. Apply PAWS check first.
4527          */
4528         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4529             tcp_paws_discard(sk, skb)) {
4530                 if (!th->rst) {
4531                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4532                         tcp_send_dupack(sk, skb);
4533                         goto discard;
4534                 }
4535                 /* Resets are accepted even if PAWS failed.
4536
4537                    ts_recent update must be made after we are sure
4538                    that the packet is in window.
4539                  */
4540         }
4541
4542         /*
4543          *      Standard slow path.
4544          */
4545
4546         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4547                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4548                  * (RST) segments are validated by checking their SEQ-fields."
4549                  * And page 69: "If an incoming segment is not acceptable,
4550                  * an acknowledgment should be sent in reply (unless the RST bit
4551                  * is set, if so drop the segment and return)".
4552                  */
4553                 if (!th->rst)
4554                         tcp_send_dupack(sk, skb);
4555                 goto discard;
4556         }
4557
4558         if (th->rst) {
4559                 tcp_reset(sk);
4560                 goto discard;
4561         }
4562
4563         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4564
4565         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4566                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4567                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4568                 tcp_reset(sk);
4569                 return 1;
4570         }
4571
4572 step5:
4573         if (th->ack)
4574                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4575
4576         tcp_rcv_rtt_measure_ts(sk, skb);
4577
4578         /* Process urgent data. */
4579         tcp_urg(sk, skb, th);
4580
4581         /* step 7: process the segment text */
4582         tcp_data_queue(sk, skb);
4583
4584         tcp_data_snd_check(sk);
4585         tcp_ack_snd_check(sk);
4586         return 0;
4587
4588 csum_error:
4589         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4590
4591 discard:
4592         __kfree_skb(skb);
4593         return 0;
4594 }
4595
4596 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4597                                          struct tcphdr *th, unsigned len)
4598 {
4599         struct tcp_sock *tp = tcp_sk(sk);
4600         struct inet_connection_sock *icsk = inet_csk(sk);
4601         int saved_clamp = tp->rx_opt.mss_clamp;
4602
4603         tcp_parse_options(skb, &tp->rx_opt, 0);
4604
4605         if (th->ack) {
4606                 /* rfc793:
4607                  * "If the state is SYN-SENT then
4608                  *    first check the ACK bit
4609                  *      If the ACK bit is set
4610                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4611                  *        a reset (unless the RST bit is set, if so drop
4612                  *        the segment and return)"
4613                  *
4614                  *  We do not send data with SYN, so that RFC-correct
4615                  *  test reduces to:
4616                  */
4617                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4618                         goto reset_and_undo;
4619
4620                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4621                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4622                              tcp_time_stamp)) {
4623                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4624                         goto reset_and_undo;
4625                 }
4626
4627                 /* Now ACK is acceptable.
4628                  *
4629                  * "If the RST bit is set
4630                  *    If the ACK was acceptable then signal the user "error:
4631                  *    connection reset", drop the segment, enter CLOSED state,
4632                  *    delete TCB, and return."
4633                  */
4634
4635                 if (th->rst) {
4636                         tcp_reset(sk);
4637                         goto discard;
4638                 }
4639
4640                 /* rfc793:
4641                  *   "fifth, if neither of the SYN or RST bits is set then
4642                  *    drop the segment and return."
4643                  *
4644                  *    See note below!
4645                  *                                        --ANK(990513)
4646                  */
4647                 if (!th->syn)
4648                         goto discard_and_undo;
4649
4650                 /* rfc793:
4651                  *   "If the SYN bit is on ...
4652                  *    are acceptable then ...
4653                  *    (our SYN has been ACKed), change the connection
4654                  *    state to ESTABLISHED..."
4655                  */
4656
4657                 TCP_ECN_rcv_synack(tp, th);
4658
4659                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4660                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4661
4662                 /* Ok.. it's good. Set up sequence numbers and
4663                  * move to established.
4664                  */
4665                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4666                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4667
4668                 /* RFC1323: The window in SYN & SYN/ACK segments is
4669                  * never scaled.
4670                  */
4671                 tp->snd_wnd = ntohs(th->window);
4672                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4673
4674                 if (!tp->rx_opt.wscale_ok) {
4675                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4676                         tp->window_clamp = min(tp->window_clamp, 65535U);
4677                 }
4678
4679                 if (tp->rx_opt.saw_tstamp) {
4680                         tp->rx_opt.tstamp_ok       = 1;
4681                         tp->tcp_header_len =
4682                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4683                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4684                         tcp_store_ts_recent(tp);
4685                 } else {
4686                         tp->tcp_header_len = sizeof(struct tcphdr);
4687                 }
4688
4689                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4690                         tcp_enable_fack(tp);
4691
4692                 tcp_mtup_init(sk);
4693                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4694                 tcp_initialize_rcv_mss(sk);
4695
4696                 /* Remember, tcp_poll() does not lock socket!
4697                  * Change state from SYN-SENT only after copied_seq
4698                  * is initialized. */
4699                 tp->copied_seq = tp->rcv_nxt;
4700                 smp_mb();
4701                 tcp_set_state(sk, TCP_ESTABLISHED);
4702
4703                 security_inet_conn_established(sk, skb);
4704
4705                 /* Make sure socket is routed, for correct metrics.  */
4706                 icsk->icsk_af_ops->rebuild_header(sk);
4707
4708                 tcp_init_metrics(sk);
4709
4710                 tcp_init_congestion_control(sk);
4711
4712                 /* Prevent spurious tcp_cwnd_restart() on first data
4713                  * packet.
4714                  */
4715                 tp->lsndtime = tcp_time_stamp;
4716
4717                 tcp_init_buffer_space(sk);
4718
4719                 if (sock_flag(sk, SOCK_KEEPOPEN))
4720                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4721
4722                 if (!tp->rx_opt.snd_wscale)
4723                         __tcp_fast_path_on(tp, tp->snd_wnd);
4724                 else
4725                         tp->pred_flags = 0;
4726
4727                 if (!sock_flag(sk, SOCK_DEAD)) {
4728                         sk->sk_state_change(sk);
4729                         sk_wake_async(sk, 0, POLL_OUT);
4730                 }
4731
4732                 if (sk->sk_write_pending ||
4733                     icsk->icsk_accept_queue.rskq_defer_accept ||
4734                     icsk->icsk_ack.pingpong) {
4735                         /* Save one ACK. Data will be ready after
4736                          * several ticks, if write_pending is set.
4737                          *
4738                          * It may be deleted, but with this feature tcpdumps
4739                          * look so _wonderfully_ clever, that I was not able
4740                          * to stand against the temptation 8)     --ANK
4741                          */
4742                         inet_csk_schedule_ack(sk);
4743                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4744                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4745                         tcp_incr_quickack(sk);
4746                         tcp_enter_quickack_mode(sk);
4747                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4748                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4749
4750 discard:
4751                         __kfree_skb(skb);
4752                         return 0;
4753                 } else {
4754                         tcp_send_ack(sk);
4755                 }
4756                 return -1;
4757         }
4758
4759         /* No ACK in the segment */
4760
4761         if (th->rst) {
4762                 /* rfc793:
4763                  * "If the RST bit is set
4764                  *
4765                  *      Otherwise (no ACK) drop the segment and return."
4766                  */
4767
4768                 goto discard_and_undo;
4769         }
4770
4771         /* PAWS check. */
4772         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4773                 goto discard_and_undo;
4774
4775         if (th->syn) {
4776                 /* We see SYN without ACK. It is attempt of
4777                  * simultaneous connect with crossed SYNs.
4778                  * Particularly, it can be connect to self.
4779                  */
4780                 tcp_set_state(sk, TCP_SYN_RECV);
4781
4782                 if (tp->rx_opt.saw_tstamp) {
4783                         tp->rx_opt.tstamp_ok = 1;
4784                         tcp_store_ts_recent(tp);
4785                         tp->tcp_header_len =
4786                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4787                 } else {
4788                         tp->tcp_header_len = sizeof(struct tcphdr);
4789                 }
4790
4791                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4792                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4793
4794                 /* RFC1323: The window in SYN & SYN/ACK segments is
4795                  * never scaled.
4796                  */
4797                 tp->snd_wnd    = ntohs(th->window);
4798                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4799                 tp->max_window = tp->snd_wnd;
4800
4801                 TCP_ECN_rcv_syn(tp, th);
4802
4803                 tcp_mtup_init(sk);
4804                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4805                 tcp_initialize_rcv_mss(sk);
4806
4807
4808                 tcp_send_synack(sk);
4809 #if 0
4810                 /* Note, we could accept data and URG from this segment.
4811                  * There are no obstacles to make this.
4812                  *
4813                  * However, if we ignore data in ACKless segments sometimes,
4814                  * we have no reasons to accept it sometimes.
4815                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4816                  * is not flawless. So, discard packet for sanity.
4817                  * Uncomment this return to process the data.
4818                  */
4819                 return -1;
4820 #else
4821                 goto discard;
4822 #endif
4823         }
4824         /* "fifth, if neither of the SYN or RST bits is set then
4825          * drop the segment and return."
4826          */
4827
4828 discard_and_undo:
4829         tcp_clear_options(&tp->rx_opt);
4830         tp->rx_opt.mss_clamp = saved_clamp;
4831         goto discard;
4832
4833 reset_and_undo:
4834         tcp_clear_options(&tp->rx_opt);
4835         tp->rx_opt.mss_clamp = saved_clamp;
4836         return 1;
4837 }
4838
4839
4840 /*
4841  *      This function implements the receiving procedure of RFC 793 for
4842  *      all states except ESTABLISHED and TIME_WAIT.
4843  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4844  *      address independent.
4845  */
4846
4847 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4848                           struct tcphdr *th, unsigned len)
4849 {
4850         struct tcp_sock *tp = tcp_sk(sk);
4851         struct inet_connection_sock *icsk = inet_csk(sk);
4852         int queued = 0;
4853
4854         tp->rx_opt.saw_tstamp = 0;
4855
4856         switch (sk->sk_state) {
4857         case TCP_CLOSE:
4858                 goto discard;
4859
4860         case TCP_LISTEN:
4861                 if (th->ack)
4862                         return 1;
4863
4864                 if (th->rst)
4865                         goto discard;
4866
4867                 if (th->syn) {
4868                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4869                                 return 1;
4870
4871                         /* Now we have several options: In theory there is
4872                          * nothing else in the frame. KA9Q has an option to
4873                          * send data with the syn, BSD accepts data with the
4874                          * syn up to the [to be] advertised window and
4875                          * Solaris 2.1 gives you a protocol error. For now
4876                          * we just ignore it, that fits the spec precisely
4877                          * and avoids incompatibilities. It would be nice in
4878                          * future to drop through and process the data.
4879                          *
4880                          * Now that TTCP is starting to be used we ought to
4881                          * queue this data.
4882                          * But, this leaves one open to an easy denial of
4883                          * service attack, and SYN cookies can't defend
4884                          * against this problem. So, we drop the data
4885                          * in the interest of security over speed unless
4886                          * it's still in use.
4887                          */
4888                         kfree_skb(skb);
4889                         return 0;
4890                 }
4891                 goto discard;
4892
4893         case TCP_SYN_SENT:
4894                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4895                 if (queued >= 0)
4896                         return queued;
4897
4898                 /* Do step6 onward by hand. */
4899                 tcp_urg(sk, skb, th);
4900                 __kfree_skb(skb);
4901                 tcp_data_snd_check(sk);
4902                 return 0;
4903         }
4904
4905         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4906             tcp_paws_discard(sk, skb)) {
4907                 if (!th->rst) {
4908                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4909                         tcp_send_dupack(sk, skb);
4910                         goto discard;
4911                 }
4912                 /* Reset is accepted even if it did not pass PAWS. */
4913         }
4914
4915         /* step 1: check sequence number */
4916         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4917                 if (!th->rst)
4918                         tcp_send_dupack(sk, skb);
4919                 goto discard;
4920         }
4921
4922         /* step 2: check RST bit */
4923         if (th->rst) {
4924                 tcp_reset(sk);
4925                 goto discard;
4926         }
4927
4928         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4929
4930         /* step 3: check security and precedence [ignored] */
4931
4932         /*      step 4:
4933          *
4934          *      Check for a SYN in window.
4935          */
4936         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4937                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4938                 tcp_reset(sk);
4939                 return 1;
4940         }
4941
4942         /* step 5: check the ACK field */
4943         if (th->ack) {
4944                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4945
4946                 switch (sk->sk_state) {
4947                 case TCP_SYN_RECV:
4948                         if (acceptable) {
4949                                 tp->copied_seq = tp->rcv_nxt;
4950                                 smp_mb();
4951                                 tcp_set_state(sk, TCP_ESTABLISHED);
4952                                 sk->sk_state_change(sk);
4953
4954                                 /* Note, that this wakeup is only for marginal
4955                                  * crossed SYN case. Passively open sockets
4956                                  * are not waked up, because sk->sk_sleep ==
4957                                  * NULL and sk->sk_socket == NULL.
4958                                  */
4959                                 if (sk->sk_socket) {
4960                                         sk_wake_async(sk,0,POLL_OUT);
4961                                 }
4962
4963                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4964                                 tp->snd_wnd = ntohs(th->window) <<
4965                                               tp->rx_opt.snd_wscale;
4966                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4967                                             TCP_SKB_CB(skb)->seq);
4968
4969                                 /* tcp_ack considers this ACK as duplicate
4970                                  * and does not calculate rtt.
4971                                  * Fix it at least with timestamps.
4972                                  */
4973                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4974                                     !tp->srtt)
4975                                         tcp_ack_saw_tstamp(sk, 0);
4976
4977                                 if (tp->rx_opt.tstamp_ok)
4978                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4979
4980                                 /* Make sure socket is routed, for
4981                                  * correct metrics.
4982                                  */
4983                                 icsk->icsk_af_ops->rebuild_header(sk);
4984
4985                                 tcp_init_metrics(sk);
4986
4987                                 tcp_init_congestion_control(sk);
4988
4989                                 /* Prevent spurious tcp_cwnd_restart() on
4990                                  * first data packet.
4991                                  */
4992                                 tp->lsndtime = tcp_time_stamp;
4993
4994                                 tcp_mtup_init(sk);
4995                                 tcp_initialize_rcv_mss(sk);
4996                                 tcp_init_buffer_space(sk);
4997                                 tcp_fast_path_on(tp);
4998                         } else {
4999                                 return 1;
5000                         }
5001                         break;
5002
5003                 case TCP_FIN_WAIT1:
5004                         if (tp->snd_una == tp->write_seq) {
5005                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5006                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5007                                 dst_confirm(sk->sk_dst_cache);
5008
5009                                 if (!sock_flag(sk, SOCK_DEAD))
5010                                         /* Wake up lingering close() */
5011                                         sk->sk_state_change(sk);
5012                                 else {
5013                                         int tmo;
5014
5015                                         if (tp->linger2 < 0 ||
5016                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5017                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5018                                                 tcp_done(sk);
5019                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5020                                                 return 1;
5021                                         }
5022
5023                                         tmo = tcp_fin_time(sk);
5024                                         if (tmo > TCP_TIMEWAIT_LEN) {
5025                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5026                                         } else if (th->fin || sock_owned_by_user(sk)) {
5027                                                 /* Bad case. We could lose such FIN otherwise.
5028                                                  * It is not a big problem, but it looks confusing
5029                                                  * and not so rare event. We still can lose it now,
5030                                                  * if it spins in bh_lock_sock(), but it is really
5031                                                  * marginal case.
5032                                                  */
5033                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5034                                         } else {
5035                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5036                                                 goto discard;
5037                                         }
5038                                 }
5039                         }
5040                         break;
5041
5042                 case TCP_CLOSING:
5043                         if (tp->snd_una == tp->write_seq) {
5044                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5045                                 goto discard;
5046                         }
5047                         break;
5048
5049                 case TCP_LAST_ACK:
5050                         if (tp->snd_una == tp->write_seq) {
5051                                 tcp_update_metrics(sk);
5052                                 tcp_done(sk);
5053                                 goto discard;
5054                         }
5055                         break;
5056                 }
5057         } else
5058                 goto discard;
5059
5060         /* step 6: check the URG bit */
5061         tcp_urg(sk, skb, th);
5062
5063         /* step 7: process the segment text */
5064         switch (sk->sk_state) {
5065         case TCP_CLOSE_WAIT:
5066         case TCP_CLOSING:
5067         case TCP_LAST_ACK:
5068                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5069                         break;
5070         case TCP_FIN_WAIT1:
5071         case TCP_FIN_WAIT2:
5072                 /* RFC 793 says to queue data in these states,
5073                  * RFC 1122 says we MUST send a reset.
5074                  * BSD 4.4 also does reset.
5075                  */
5076                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5077                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5078                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5079                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5080                                 tcp_reset(sk);
5081                                 return 1;
5082                         }
5083                 }
5084                 /* Fall through */
5085         case TCP_ESTABLISHED:
5086                 tcp_data_queue(sk, skb);
5087                 queued = 1;
5088                 break;
5089         }
5090
5091         /* tcp_data could move socket to TIME-WAIT */
5092         if (sk->sk_state != TCP_CLOSE) {
5093                 tcp_data_snd_check(sk);
5094                 tcp_ack_snd_check(sk);
5095         }
5096
5097         if (!queued) {
5098 discard:
5099                 __kfree_skb(skb);
5100         }
5101         return 0;
5102 }
5103
5104 EXPORT_SYMBOL(sysctl_tcp_ecn);
5105 EXPORT_SYMBOL(sysctl_tcp_reordering);
5106 EXPORT_SYMBOL(tcp_parse_options);
5107 EXPORT_SYMBOL(tcp_rcv_established);
5108 EXPORT_SYMBOL(tcp_rcv_state_process);
5109 EXPORT_SYMBOL(tcp_initialize_rcv_mss);