Merge branch 'oprofile-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[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  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  */
20
21 /*
22  * Changes:
23  *              Pedro Roque     :       Fast Retransmit/Recovery.
24  *                                      Two receive queues.
25  *                                      Retransmit queue handled by TCP.
26  *                                      Better retransmit timer handling.
27  *                                      New congestion avoidance.
28  *                                      Header prediction.
29  *                                      Variable renaming.
30  *
31  *              Eric            :       Fast Retransmit.
32  *              Randy Scott     :       MSS option defines.
33  *              Eric Schenk     :       Fixes to slow start algorithm.
34  *              Eric Schenk     :       Yet another double ACK bug.
35  *              Eric Schenk     :       Delayed ACK bug fixes.
36  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
37  *              David S. Miller :       Don't allow zero congestion window.
38  *              Eric Schenk     :       Fix retransmitter so that it sends
39  *                                      next packet on ack of previous packet.
40  *              Andi Kleen      :       Moved open_request checking here
41  *                                      and process RSTs for open_requests.
42  *              Andi Kleen      :       Better prune_queue, and other fixes.
43  *              Andrey Savochkin:       Fix RTT measurements in the presence of
44  *                                      timestamps.
45  *              Andrey Savochkin:       Check sequence numbers correctly when
46  *                                      removing SACKs due to in sequence incoming
47  *                                      data segments.
48  *              Andi Kleen:             Make sure we never ack data there is not
49  *                                      enough room for. Also make this condition
50  *                                      a fatal error if it might still happen.
51  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
52  *                                      connections with MSS<min(MTU,ann. MSS)
53  *                                      work without delayed acks.
54  *              Andi Kleen:             Process packets with PSH set in the
55  *                                      fast path.
56  *              J Hadi Salim:           ECN support
57  *              Andrei Gurtov,
58  *              Pasi Sarolahti,
59  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
60  *                                      engine. Lots of bugs are found.
61  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
62  */
63
64 #include <linux/mm.h>
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
67 #include <net/dst.h>
68 #include <net/tcp.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
73
74 int sysctl_tcp_timestamps __read_mostly = 1;
75 int sysctl_tcp_window_scaling __read_mostly = 1;
76 int sysctl_tcp_sack __read_mostly = 1;
77 int sysctl_tcp_fack __read_mostly = 1;
78 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
79 int sysctl_tcp_ecn __read_mostly;
80 int sysctl_tcp_dsack __read_mostly = 1;
81 int sysctl_tcp_app_win __read_mostly = 31;
82 int sysctl_tcp_adv_win_scale __read_mostly = 2;
83
84 int sysctl_tcp_stdurg __read_mostly;
85 int sysctl_tcp_rfc1337 __read_mostly;
86 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
87 int sysctl_tcp_frto __read_mostly = 2;
88 int sysctl_tcp_frto_response __read_mostly;
89 int sysctl_tcp_nometrics_save __read_mostly;
90
91 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
92 int sysctl_tcp_abc __read_mostly;
93
94 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
95 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
96 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
98 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
99 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
100 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
101 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
102 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
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 TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
117
118 /* Adapt the MSS value used to make delayed ack decision to the
119  * real world.
120  */
121 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
122 {
123         struct inet_connection_sock *icsk = inet_csk(sk);
124         const unsigned int lss = icsk->icsk_ack.last_seg_size;
125         unsigned int len;
126
127         icsk->icsk_ack.last_seg_size = 0;
128
129         /* skb->len may jitter because of SACKs, even if peer
130          * sends good full-sized frames.
131          */
132         len = skb_shinfo(skb)->gso_size ? : skb->len;
133         if (len >= icsk->icsk_ack.rcv_mss) {
134                 icsk->icsk_ack.rcv_mss = len;
135         } else {
136                 /* Otherwise, we make more careful check taking into account,
137                  * that SACKs block is variable.
138                  *
139                  * "len" is invariant segment length, including TCP header.
140                  */
141                 len += skb->data - skb_transport_header(skb);
142                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
143                     /* If PSH is not set, packet should be
144                      * full sized, provided peer TCP is not badly broken.
145                      * This observation (if it is correct 8)) allows
146                      * to handle super-low mtu links fairly.
147                      */
148                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
149                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
150                         /* Subtract also invariant (if peer is RFC compliant),
151                          * tcp header plus fixed timestamp option length.
152                          * Resulting "len" is MSS free of SACK jitter.
153                          */
154                         len -= tcp_sk(sk)->tcp_header_len;
155                         icsk->icsk_ack.last_seg_size = len;
156                         if (len == lss) {
157                                 icsk->icsk_ack.rcv_mss = len;
158                                 return;
159                         }
160                 }
161                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
162                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
163                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
164         }
165 }
166
167 static void tcp_incr_quickack(struct sock *sk)
168 {
169         struct inet_connection_sock *icsk = inet_csk(sk);
170         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
171
172         if (quickacks == 0)
173                 quickacks = 2;
174         if (quickacks > icsk->icsk_ack.quick)
175                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
176 }
177
178 void tcp_enter_quickack_mode(struct sock *sk)
179 {
180         struct inet_connection_sock *icsk = inet_csk(sk);
181         tcp_incr_quickack(sk);
182         icsk->icsk_ack.pingpong = 0;
183         icsk->icsk_ack.ato = TCP_ATO_MIN;
184 }
185
186 /* Send ACKs quickly, if "quick" count is not exhausted
187  * and the session is not interactive.
188  */
189
190 static inline int tcp_in_quickack_mode(const struct sock *sk)
191 {
192         const struct inet_connection_sock *icsk = inet_csk(sk);
193         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
194 }
195
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
197 {
198         if (tp->ecn_flags & TCP_ECN_OK)
199                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
200 }
201
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
203 {
204         if (tcp_hdr(skb)->cwr)
205                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
206 }
207
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
209 {
210         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
211 }
212
213 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
214 {
215         if (tp->ecn_flags & TCP_ECN_OK) {
216                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
217                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
218                 /* Funny extension: if ECT is not set on a segment,
219                  * it is surely retransmit. It is not in ECN RFC,
220                  * but Linux follows this rule. */
221                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
222                         tcp_enter_quickack_mode((struct sock *)tp);
223         }
224 }
225
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
227 {
228         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
229                 tp->ecn_flags &= ~TCP_ECN_OK;
230 }
231
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
233 {
234         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
235                 tp->ecn_flags &= ~TCP_ECN_OK;
236 }
237
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
239 {
240         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
241                 return 1;
242         return 0;
243 }
244
245 /* Buffer size and advertised window tuning.
246  *
247  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
248  */
249
250 static void tcp_fixup_sndbuf(struct sock *sk)
251 {
252         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
253                      sizeof(struct sk_buff);
254
255         if (sk->sk_sndbuf < 3 * sndmem)
256                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
257 }
258
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
260  *
261  * All tcp_full_space() is split to two parts: "network" buffer, allocated
262  * forward and advertised in receiver window (tp->rcv_wnd) and
263  * "application buffer", required to isolate scheduling/application
264  * latencies from network.
265  * window_clamp is maximal advertised window. It can be less than
266  * tcp_full_space(), in this case tcp_full_space() - window_clamp
267  * is reserved for "application" buffer. The less window_clamp is
268  * the smoother our behaviour from viewpoint of network, but the lower
269  * throughput and the higher sensitivity of the connection to losses. 8)
270  *
271  * rcv_ssthresh is more strict window_clamp used at "slow start"
272  * phase to predict further behaviour of this connection.
273  * It is used for two goals:
274  * - to enforce header prediction at sender, even when application
275  *   requires some significant "application buffer". It is check #1.
276  * - to prevent pruning of receive queue because of misprediction
277  *   of receiver window. Check #2.
278  *
279  * The scheme does not work when sender sends good segments opening
280  * window and then starts to feed us spaghetti. But it should work
281  * in common situations. Otherwise, we have to rely on queue collapsing.
282  */
283
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
286 {
287         struct tcp_sock *tp = tcp_sk(sk);
288         /* Optimize this! */
289         int truesize = tcp_win_from_space(skb->truesize) >> 1;
290         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
291
292         while (tp->rcv_ssthresh <= window) {
293                 if (truesize <= skb->len)
294                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
295
296                 truesize >>= 1;
297                 window >>= 1;
298         }
299         return 0;
300 }
301
302 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
303 {
304         struct tcp_sock *tp = tcp_sk(sk);
305
306         /* Check #1 */
307         if (tp->rcv_ssthresh < tp->window_clamp &&
308             (int)tp->rcv_ssthresh < tcp_space(sk) &&
309             !tcp_memory_pressure) {
310                 int incr;
311
312                 /* Check #2. Increase window, if skb with such overhead
313                  * will fit to rcvbuf in future.
314                  */
315                 if (tcp_win_from_space(skb->truesize) <= skb->len)
316                         incr = 2 * tp->advmss;
317                 else
318                         incr = __tcp_grow_window(sk, skb);
319
320                 if (incr) {
321                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
322                                                tp->window_clamp);
323                         inet_csk(sk)->icsk_ack.quick |= 1;
324                 }
325         }
326 }
327
328 /* 3. Tuning rcvbuf, when connection enters established state. */
329
330 static void tcp_fixup_rcvbuf(struct sock *sk)
331 {
332         struct tcp_sock *tp = tcp_sk(sk);
333         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
334
335         /* Try to select rcvbuf so that 4 mss-sized segments
336          * will fit to window and corresponding skbs will fit to our rcvbuf.
337          * (was 3; 4 is minimum to allow fast retransmit to work.)
338          */
339         while (tcp_win_from_space(rcvmem) < tp->advmss)
340                 rcvmem += 128;
341         if (sk->sk_rcvbuf < 4 * rcvmem)
342                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
343 }
344
345 /* 4. Try to fixup all. It is made immediately after connection enters
346  *    established state.
347  */
348 static void tcp_init_buffer_space(struct sock *sk)
349 {
350         struct tcp_sock *tp = tcp_sk(sk);
351         int maxwin;
352
353         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
354                 tcp_fixup_rcvbuf(sk);
355         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
356                 tcp_fixup_sndbuf(sk);
357
358         tp->rcvq_space.space = tp->rcv_wnd;
359
360         maxwin = tcp_full_space(sk);
361
362         if (tp->window_clamp >= maxwin) {
363                 tp->window_clamp = maxwin;
364
365                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
366                         tp->window_clamp = max(maxwin -
367                                                (maxwin >> sysctl_tcp_app_win),
368                                                4 * tp->advmss);
369         }
370
371         /* Force reservation of one segment. */
372         if (sysctl_tcp_app_win &&
373             tp->window_clamp > 2 * tp->advmss &&
374             tp->window_clamp + tp->advmss > maxwin)
375                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
376
377         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
378         tp->snd_cwnd_stamp = tcp_time_stamp;
379 }
380
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock *sk)
383 {
384         struct tcp_sock *tp = tcp_sk(sk);
385         struct inet_connection_sock *icsk = inet_csk(sk);
386
387         icsk->icsk_ack.quick = 0;
388
389         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
390             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
391             !tcp_memory_pressure &&
392             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
393                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
394                                     sysctl_tcp_rmem[2]);
395         }
396         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
397                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
398 }
399
400 /* Initialize RCV_MSS value.
401  * RCV_MSS is an our guess about MSS used by the peer.
402  * We haven't any direct information about the MSS.
403  * It's better to underestimate the RCV_MSS rather than overestimate.
404  * Overestimations make us ACKing less frequently than needed.
405  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
406  */
407 void tcp_initialize_rcv_mss(struct sock *sk)
408 {
409         struct tcp_sock *tp = tcp_sk(sk);
410         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
411
412         hint = min(hint, tp->rcv_wnd / 2);
413         hint = min(hint, TCP_MIN_RCVMSS);
414         hint = max(hint, TCP_MIN_MSS);
415
416         inet_csk(sk)->icsk_ack.rcv_mss = hint;
417 }
418
419 /* Receiver "autotuning" code.
420  *
421  * The algorithm for RTT estimation w/o timestamps is based on
422  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
424  *
425  * More detail on this code can be found at
426  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427  * though this reference is out of date.  A new paper
428  * is pending.
429  */
430 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
431 {
432         u32 new_sample = tp->rcv_rtt_est.rtt;
433         long m = sample;
434
435         if (m == 0)
436                 m = 1;
437
438         if (new_sample != 0) {
439                 /* If we sample in larger samples in the non-timestamp
440                  * case, we could grossly overestimate the RTT especially
441                  * with chatty applications or bulk transfer apps which
442                  * are stalled on filesystem I/O.
443                  *
444                  * Also, since we are only going for a minimum in the
445                  * non-timestamp case, we do not smooth things out
446                  * else with timestamps disabled convergence takes too
447                  * long.
448                  */
449                 if (!win_dep) {
450                         m -= (new_sample >> 3);
451                         new_sample += m;
452                 } else if (m < new_sample)
453                         new_sample = m << 3;
454         } else {
455                 /* No previous measure. */
456                 new_sample = m << 3;
457         }
458
459         if (tp->rcv_rtt_est.rtt != new_sample)
460                 tp->rcv_rtt_est.rtt = new_sample;
461 }
462
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
464 {
465         if (tp->rcv_rtt_est.time == 0)
466                 goto new_measure;
467         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
468                 return;
469         tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
470
471 new_measure:
472         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
473         tp->rcv_rtt_est.time = tcp_time_stamp;
474 }
475
476 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
477                                           const struct sk_buff *skb)
478 {
479         struct tcp_sock *tp = tcp_sk(sk);
480         if (tp->rx_opt.rcv_tsecr &&
481             (TCP_SKB_CB(skb)->end_seq -
482              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
483                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
484 }
485
486 /*
487  * This function should be called every time data is copied to user space.
488  * It calculates the appropriate TCP receive buffer space.
489  */
490 void tcp_rcv_space_adjust(struct sock *sk)
491 {
492         struct tcp_sock *tp = tcp_sk(sk);
493         int time;
494         int space;
495
496         if (tp->rcvq_space.time == 0)
497                 goto new_measure;
498
499         time = tcp_time_stamp - tp->rcvq_space.time;
500         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
501                 return;
502
503         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
504
505         space = max(tp->rcvq_space.space, space);
506
507         if (tp->rcvq_space.space != space) {
508                 int rcvmem;
509
510                 tp->rcvq_space.space = space;
511
512                 if (sysctl_tcp_moderate_rcvbuf &&
513                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
514                         int new_clamp = space;
515
516                         /* Receive space grows, normalize in order to
517                          * take into account packet headers and sk_buff
518                          * structure overhead.
519                          */
520                         space /= tp->advmss;
521                         if (!space)
522                                 space = 1;
523                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
524                                   16 + sizeof(struct sk_buff));
525                         while (tcp_win_from_space(rcvmem) < tp->advmss)
526                                 rcvmem += 128;
527                         space *= rcvmem;
528                         space = min(space, sysctl_tcp_rmem[2]);
529                         if (space > sk->sk_rcvbuf) {
530                                 sk->sk_rcvbuf = space;
531
532                                 /* Make the window clamp follow along.  */
533                                 tp->window_clamp = new_clamp;
534                         }
535                 }
536         }
537
538 new_measure:
539         tp->rcvq_space.seq = tp->copied_seq;
540         tp->rcvq_space.time = tcp_time_stamp;
541 }
542
543 /* There is something which you must keep in mind when you analyze the
544  * behavior of the tp->ato delayed ack timeout interval.  When a
545  * connection starts up, we want to ack as quickly as possible.  The
546  * problem is that "good" TCP's do slow start at the beginning of data
547  * transmission.  The means that until we send the first few ACK's the
548  * sender will sit on his end and only queue most of his data, because
549  * he can only send snd_cwnd unacked packets at any given time.  For
550  * each ACK we send, he increments snd_cwnd and transmits more of his
551  * queue.  -DaveM
552  */
553 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
554 {
555         struct tcp_sock *tp = tcp_sk(sk);
556         struct inet_connection_sock *icsk = inet_csk(sk);
557         u32 now;
558
559         inet_csk_schedule_ack(sk);
560
561         tcp_measure_rcv_mss(sk, skb);
562
563         tcp_rcv_rtt_measure(tp);
564
565         now = tcp_time_stamp;
566
567         if (!icsk->icsk_ack.ato) {
568                 /* The _first_ data packet received, initialize
569                  * delayed ACK engine.
570                  */
571                 tcp_incr_quickack(sk);
572                 icsk->icsk_ack.ato = TCP_ATO_MIN;
573         } else {
574                 int m = now - icsk->icsk_ack.lrcvtime;
575
576                 if (m <= TCP_ATO_MIN / 2) {
577                         /* The fastest case is the first. */
578                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
579                 } else if (m < icsk->icsk_ack.ato) {
580                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
581                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
582                                 icsk->icsk_ack.ato = icsk->icsk_rto;
583                 } else if (m > icsk->icsk_rto) {
584                         /* Too long gap. Apparently sender failed to
585                          * restart window, so that we send ACKs quickly.
586                          */
587                         tcp_incr_quickack(sk);
588                         sk_mem_reclaim(sk);
589                 }
590         }
591         icsk->icsk_ack.lrcvtime = now;
592
593         TCP_ECN_check_ce(tp, skb);
594
595         if (skb->len >= 128)
596                 tcp_grow_window(sk, skb);
597 }
598
599 static u32 tcp_rto_min(struct sock *sk)
600 {
601         struct dst_entry *dst = __sk_dst_get(sk);
602         u32 rto_min = TCP_RTO_MIN;
603
604         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
605                 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
606         return rto_min;
607 }
608
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610  * routine either comes from timestamps, or from segments that were
611  * known _not_ to have been retransmitted [see Karn/Partridge
612  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613  * piece by Van Jacobson.
614  * NOTE: the next three routines used to be one big routine.
615  * To save cycles in the RFC 1323 implementation it was better to break
616  * it up into three procedures. -- erics
617  */
618 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
619 {
620         struct tcp_sock *tp = tcp_sk(sk);
621         long m = mrtt; /* RTT */
622
623         /*      The following amusing code comes from Jacobson's
624          *      article in SIGCOMM '88.  Note that rtt and mdev
625          *      are scaled versions of rtt and mean deviation.
626          *      This is designed to be as fast as possible
627          *      m stands for "measurement".
628          *
629          *      On a 1990 paper the rto value is changed to:
630          *      RTO = rtt + 4 * mdev
631          *
632          * Funny. This algorithm seems to be very broken.
633          * These formulae increase RTO, when it should be decreased, increase
634          * too slowly, when it should be increased quickly, decrease too quickly
635          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636          * does not matter how to _calculate_ it. Seems, it was trap
637          * that VJ failed to avoid. 8)
638          */
639         if (m == 0)
640                 m = 1;
641         if (tp->srtt != 0) {
642                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
643                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
644                 if (m < 0) {
645                         m = -m;         /* m is now abs(error) */
646                         m -= (tp->mdev >> 2);   /* similar update on mdev */
647                         /* This is similar to one of Eifel findings.
648                          * Eifel blocks mdev updates when rtt decreases.
649                          * This solution is a bit different: we use finer gain
650                          * for mdev in this case (alpha*beta).
651                          * Like Eifel it also prevents growth of rto,
652                          * but also it limits too fast rto decreases,
653                          * happening in pure Eifel.
654                          */
655                         if (m > 0)
656                                 m >>= 3;
657                 } else {
658                         m -= (tp->mdev >> 2);   /* similar update on mdev */
659                 }
660                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
661                 if (tp->mdev > tp->mdev_max) {
662                         tp->mdev_max = tp->mdev;
663                         if (tp->mdev_max > tp->rttvar)
664                                 tp->rttvar = tp->mdev_max;
665                 }
666                 if (after(tp->snd_una, tp->rtt_seq)) {
667                         if (tp->mdev_max < tp->rttvar)
668                                 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
669                         tp->rtt_seq = tp->snd_nxt;
670                         tp->mdev_max = tcp_rto_min(sk);
671                 }
672         } else {
673                 /* no previous measure. */
674                 tp->srtt = m << 3;      /* take the measured time to be rtt */
675                 tp->mdev = m << 1;      /* make sure rto = 3*rtt */
676                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
677                 tp->rtt_seq = tp->snd_nxt;
678         }
679 }
680
681 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
682  * routine referred to above.
683  */
684 static inline void tcp_set_rto(struct sock *sk)
685 {
686         const struct tcp_sock *tp = tcp_sk(sk);
687         /* Old crap is replaced with new one. 8)
688          *
689          * More seriously:
690          * 1. If rtt variance happened to be less 50msec, it is hallucination.
691          *    It cannot be less due to utterly erratic ACK generation made
692          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693          *    to do with delayed acks, because at cwnd>2 true delack timeout
694          *    is invisible. Actually, Linux-2.4 also generates erratic
695          *    ACKs in some circumstances.
696          */
697         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
698
699         /* 2. Fixups made earlier cannot be right.
700          *    If we do not estimate RTO correctly without them,
701          *    all the algo is pure shit and should be replaced
702          *    with correct one. It is exactly, which we pretend to do.
703          */
704
705         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
706          * guarantees that rto is higher.
707          */
708         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
709                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
710 }
711
712 /* Save metrics learned by this TCP session.
713    This function is called only, when TCP finishes successfully
714    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
715  */
716 void tcp_update_metrics(struct sock *sk)
717 {
718         struct tcp_sock *tp = tcp_sk(sk);
719         struct dst_entry *dst = __sk_dst_get(sk);
720
721         if (sysctl_tcp_nometrics_save)
722                 return;
723
724         dst_confirm(dst);
725
726         if (dst && (dst->flags & DST_HOST)) {
727                 const struct inet_connection_sock *icsk = inet_csk(sk);
728                 int m;
729                 unsigned long rtt;
730
731                 if (icsk->icsk_backoff || !tp->srtt) {
732                         /* This session failed to estimate rtt. Why?
733                          * Probably, no packets returned in time.
734                          * Reset our results.
735                          */
736                         if (!(dst_metric_locked(dst, RTAX_RTT)))
737                                 dst->metrics[RTAX_RTT - 1] = 0;
738                         return;
739                 }
740
741                 rtt = dst_metric_rtt(dst, RTAX_RTT);
742                 m = rtt - tp->srtt;
743
744                 /* If newly calculated rtt larger than stored one,
745                  * store new one. Otherwise, use EWMA. Remember,
746                  * rtt overestimation is always better than underestimation.
747                  */
748                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
749                         if (m <= 0)
750                                 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
751                         else
752                                 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
753                 }
754
755                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
756                         unsigned long var;
757                         if (m < 0)
758                                 m = -m;
759
760                         /* Scale deviation to rttvar fixed point */
761                         m >>= 1;
762                         if (m < tp->mdev)
763                                 m = tp->mdev;
764
765                         var = dst_metric_rtt(dst, RTAX_RTTVAR);
766                         if (m >= var)
767                                 var = m;
768                         else
769                                 var -= (var - m) >> 2;
770
771                         set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
772                 }
773
774                 if (tp->snd_ssthresh >= 0xFFFF) {
775                         /* Slow start still did not finish. */
776                         if (dst_metric(dst, RTAX_SSTHRESH) &&
777                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
778                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
779                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
780                         if (!dst_metric_locked(dst, RTAX_CWND) &&
781                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
782                                 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
783                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
784                            icsk->icsk_ca_state == TCP_CA_Open) {
785                         /* Cong. avoidance phase, cwnd is reliable. */
786                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
787                                 dst->metrics[RTAX_SSTHRESH-1] =
788                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
789                         if (!dst_metric_locked(dst, RTAX_CWND))
790                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
791                 } else {
792                         /* Else slow start did not finish, cwnd is non-sense,
793                            ssthresh may be also invalid.
794                          */
795                         if (!dst_metric_locked(dst, RTAX_CWND))
796                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
797                         if (dst_metric(dst, RTAX_SSTHRESH) &&
798                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
799                             tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
800                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
801                 }
802
803                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
804                         if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
805                             tp->reordering != sysctl_tcp_reordering)
806                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
807                 }
808         }
809 }
810
811 /* Numbers are taken from RFC3390.
812  *
813  * John Heffner states:
814  *
815  *      The RFC specifies a window of no more than 4380 bytes
816  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
817  *      is a bit misleading because they use a clamp at 4380 bytes
818  *      rather than use a multiplier in the relevant range.
819  */
820 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
821 {
822         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
823
824         if (!cwnd) {
825                 if (tp->mss_cache > 1460)
826                         cwnd = 2;
827                 else
828                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
829         }
830         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
831 }
832
833 /* Set slow start threshold and cwnd not falling to slow start */
834 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
835 {
836         struct tcp_sock *tp = tcp_sk(sk);
837         const struct inet_connection_sock *icsk = inet_csk(sk);
838
839         tp->prior_ssthresh = 0;
840         tp->bytes_acked = 0;
841         if (icsk->icsk_ca_state < TCP_CA_CWR) {
842                 tp->undo_marker = 0;
843                 if (set_ssthresh)
844                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
845                 tp->snd_cwnd = min(tp->snd_cwnd,
846                                    tcp_packets_in_flight(tp) + 1U);
847                 tp->snd_cwnd_cnt = 0;
848                 tp->high_seq = tp->snd_nxt;
849                 tp->snd_cwnd_stamp = tcp_time_stamp;
850                 TCP_ECN_queue_cwr(tp);
851
852                 tcp_set_ca_state(sk, TCP_CA_CWR);
853         }
854 }
855
856 /*
857  * Packet counting of FACK is based on in-order assumptions, therefore TCP
858  * disables it when reordering is detected
859  */
860 static void tcp_disable_fack(struct tcp_sock *tp)
861 {
862         /* RFC3517 uses different metric in lost marker => reset on change */
863         if (tcp_is_fack(tp))
864                 tp->lost_skb_hint = NULL;
865         tp->rx_opt.sack_ok &= ~2;
866 }
867
868 /* Take a notice that peer is sending D-SACKs */
869 static void tcp_dsack_seen(struct tcp_sock *tp)
870 {
871         tp->rx_opt.sack_ok |= 4;
872 }
873
874 /* Initialize metrics on socket. */
875
876 static void tcp_init_metrics(struct sock *sk)
877 {
878         struct tcp_sock *tp = tcp_sk(sk);
879         struct dst_entry *dst = __sk_dst_get(sk);
880
881         if (dst == NULL)
882                 goto reset;
883
884         dst_confirm(dst);
885
886         if (dst_metric_locked(dst, RTAX_CWND))
887                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
888         if (dst_metric(dst, RTAX_SSTHRESH)) {
889                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
890                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
891                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
892         }
893         if (dst_metric(dst, RTAX_REORDERING) &&
894             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
895                 tcp_disable_fack(tp);
896                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
897         }
898
899         if (dst_metric(dst, RTAX_RTT) == 0)
900                 goto reset;
901
902         if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
903                 goto reset;
904
905         /* Initial rtt is determined from SYN,SYN-ACK.
906          * The segment is small and rtt may appear much
907          * less than real one. Use per-dst memory
908          * to make it more realistic.
909          *
910          * A bit of theory. RTT is time passed after "normal" sized packet
911          * is sent until it is ACKed. In normal circumstances sending small
912          * packets force peer to delay ACKs and calculation is correct too.
913          * The algorithm is adaptive and, provided we follow specs, it
914          * NEVER underestimate RTT. BUT! If peer tries to make some clever
915          * tricks sort of "quick acks" for time long enough to decrease RTT
916          * to low value, and then abruptly stops to do it and starts to delay
917          * ACKs, wait for troubles.
918          */
919         if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
920                 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
921                 tp->rtt_seq = tp->snd_nxt;
922         }
923         if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
924                 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
925                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
926         }
927         tcp_set_rto(sk);
928         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
929                 goto reset;
930         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
931         tp->snd_cwnd_stamp = tcp_time_stamp;
932         return;
933
934 reset:
935         /* Play conservative. If timestamps are not
936          * supported, TCP will fail to recalculate correct
937          * rtt, if initial rto is too small. FORGET ALL AND RESET!
938          */
939         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
940                 tp->srtt = 0;
941                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
942                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
943         }
944 }
945
946 static void tcp_update_reordering(struct sock *sk, const int metric,
947                                   const int ts)
948 {
949         struct tcp_sock *tp = tcp_sk(sk);
950         if (metric > tp->reordering) {
951                 int mib_idx;
952
953                 tp->reordering = min(TCP_MAX_REORDERING, metric);
954
955                 /* This exciting event is worth to be remembered. 8) */
956                 if (ts)
957                         mib_idx = LINUX_MIB_TCPTSREORDER;
958                 else if (tcp_is_reno(tp))
959                         mib_idx = LINUX_MIB_TCPRENOREORDER;
960                 else if (tcp_is_fack(tp))
961                         mib_idx = LINUX_MIB_TCPFACKREORDER;
962                 else
963                         mib_idx = LINUX_MIB_TCPSACKREORDER;
964
965                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
966 #if FASTRETRANS_DEBUG > 1
967                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
968                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
969                        tp->reordering,
970                        tp->fackets_out,
971                        tp->sacked_out,
972                        tp->undo_marker ? tp->undo_retrans : 0);
973 #endif
974                 tcp_disable_fack(tp);
975         }
976 }
977
978 /* This must be called before lost_out is incremented */
979 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
980 {
981         if ((tp->retransmit_skb_hint == NULL) ||
982             before(TCP_SKB_CB(skb)->seq,
983                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
984                 tp->retransmit_skb_hint = skb;
985
986         if (!tp->lost_out ||
987             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
988                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
989 }
990
991 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
992 {
993         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
994                 tcp_verify_retransmit_hint(tp, skb);
995
996                 tp->lost_out += tcp_skb_pcount(skb);
997                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
998         }
999 }
1000
1001 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1002                                             struct sk_buff *skb)
1003 {
1004         tcp_verify_retransmit_hint(tp, skb);
1005
1006         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1007                 tp->lost_out += tcp_skb_pcount(skb);
1008                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1009         }
1010 }
1011
1012 /* This procedure tags the retransmission queue when SACKs arrive.
1013  *
1014  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1015  * Packets in queue with these bits set are counted in variables
1016  * sacked_out, retrans_out and lost_out, correspondingly.
1017  *
1018  * Valid combinations are:
1019  * Tag  InFlight        Description
1020  * 0    1               - orig segment is in flight.
1021  * S    0               - nothing flies, orig reached receiver.
1022  * L    0               - nothing flies, orig lost by net.
1023  * R    2               - both orig and retransmit are in flight.
1024  * L|R  1               - orig is lost, retransmit is in flight.
1025  * S|R  1               - orig reached receiver, retrans is still in flight.
1026  * (L|S|R is logically valid, it could occur when L|R is sacked,
1027  *  but it is equivalent to plain S and code short-curcuits it to S.
1028  *  L|S is logically invalid, it would mean -1 packet in flight 8))
1029  *
1030  * These 6 states form finite state machine, controlled by the following events:
1031  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1032  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1033  * 3. Loss detection event of one of three flavors:
1034  *      A. Scoreboard estimator decided the packet is lost.
1035  *         A'. Reno "three dupacks" marks head of queue lost.
1036  *         A''. Its FACK modfication, head until snd.fack is lost.
1037  *      B. SACK arrives sacking data transmitted after never retransmitted
1038  *         hole was sent out.
1039  *      C. SACK arrives sacking SND.NXT at the moment, when the
1040  *         segment was retransmitted.
1041  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1042  *
1043  * It is pleasant to note, that state diagram turns out to be commutative,
1044  * so that we are allowed not to be bothered by order of our actions,
1045  * when multiple events arrive simultaneously. (see the function below).
1046  *
1047  * Reordering detection.
1048  * --------------------
1049  * Reordering metric is maximal distance, which a packet can be displaced
1050  * in packet stream. With SACKs we can estimate it:
1051  *
1052  * 1. SACK fills old hole and the corresponding segment was not
1053  *    ever retransmitted -> reordering. Alas, we cannot use it
1054  *    when segment was retransmitted.
1055  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1056  *    for retransmitted and already SACKed segment -> reordering..
1057  * Both of these heuristics are not used in Loss state, when we cannot
1058  * account for retransmits accurately.
1059  *
1060  * SACK block validation.
1061  * ----------------------
1062  *
1063  * SACK block range validation checks that the received SACK block fits to
1064  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1065  * Note that SND.UNA is not included to the range though being valid because
1066  * it means that the receiver is rather inconsistent with itself reporting
1067  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1068  * perfectly valid, however, in light of RFC2018 which explicitly states
1069  * that "SACK block MUST reflect the newest segment.  Even if the newest
1070  * segment is going to be discarded ...", not that it looks very clever
1071  * in case of head skb. Due to potentional receiver driven attacks, we
1072  * choose to avoid immediate execution of a walk in write queue due to
1073  * reneging and defer head skb's loss recovery to standard loss recovery
1074  * procedure that will eventually trigger (nothing forbids us doing this).
1075  *
1076  * Implements also blockage to start_seq wrap-around. Problem lies in the
1077  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1078  * there's no guarantee that it will be before snd_nxt (n). The problem
1079  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1080  * wrap (s_w):
1081  *
1082  *         <- outs wnd ->                          <- wrapzone ->
1083  *         u     e      n                         u_w   e_w  s n_w
1084  *         |     |      |                          |     |   |  |
1085  * |<------------+------+----- TCP seqno space --------------+---------->|
1086  * ...-- <2^31 ->|                                           |<--------...
1087  * ...---- >2^31 ------>|                                    |<--------...
1088  *
1089  * Current code wouldn't be vulnerable but it's better still to discard such
1090  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1091  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1092  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1093  * equal to the ideal case (infinite seqno space without wrap caused issues).
1094  *
1095  * With D-SACK the lower bound is extended to cover sequence space below
1096  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1097  * again, D-SACK block must not to go across snd_una (for the same reason as
1098  * for the normal SACK blocks, explained above). But there all simplicity
1099  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1100  * fully below undo_marker they do not affect behavior in anyway and can
1101  * therefore be safely ignored. In rare cases (which are more or less
1102  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1103  * fragmentation and packet reordering past skb's retransmission. To consider
1104  * them correctly, the acceptable range must be extended even more though
1105  * the exact amount is rather hard to quantify. However, tp->max_window can
1106  * be used as an exaggerated estimate.
1107  */
1108 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1109                                   u32 start_seq, u32 end_seq)
1110 {
1111         /* Too far in future, or reversed (interpretation is ambiguous) */
1112         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1113                 return 0;
1114
1115         /* Nasty start_seq wrap-around check (see comments above) */
1116         if (!before(start_seq, tp->snd_nxt))
1117                 return 0;
1118
1119         /* In outstanding window? ...This is valid exit for D-SACKs too.
1120          * start_seq == snd_una is non-sensical (see comments above)
1121          */
1122         if (after(start_seq, tp->snd_una))
1123                 return 1;
1124
1125         if (!is_dsack || !tp->undo_marker)
1126                 return 0;
1127
1128         /* ...Then it's D-SACK, and must reside below snd_una completely */
1129         if (!after(end_seq, tp->snd_una))
1130                 return 0;
1131
1132         if (!before(start_seq, tp->undo_marker))
1133                 return 1;
1134
1135         /* Too old */
1136         if (!after(end_seq, tp->undo_marker))
1137                 return 0;
1138
1139         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1140          *   start_seq < undo_marker and end_seq >= undo_marker.
1141          */
1142         return !before(start_seq, end_seq - tp->max_window);
1143 }
1144
1145 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1146  * Event "C". Later note: FACK people cheated me again 8), we have to account
1147  * for reordering! Ugly, but should help.
1148  *
1149  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1150  * less than what is now known to be received by the other end (derived from
1151  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1152  * retransmitted skbs to avoid some costly processing per ACKs.
1153  */
1154 static void tcp_mark_lost_retrans(struct sock *sk)
1155 {
1156         const struct inet_connection_sock *icsk = inet_csk(sk);
1157         struct tcp_sock *tp = tcp_sk(sk);
1158         struct sk_buff *skb;
1159         int cnt = 0;
1160         u32 new_low_seq = tp->snd_nxt;
1161         u32 received_upto = tcp_highest_sack_seq(tp);
1162
1163         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1164             !after(received_upto, tp->lost_retrans_low) ||
1165             icsk->icsk_ca_state != TCP_CA_Recovery)
1166                 return;
1167
1168         tcp_for_write_queue(skb, sk) {
1169                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1170
1171                 if (skb == tcp_send_head(sk))
1172                         break;
1173                 if (cnt == tp->retrans_out)
1174                         break;
1175                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1176                         continue;
1177
1178                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1179                         continue;
1180
1181                 if (after(received_upto, ack_seq) &&
1182                     (tcp_is_fack(tp) ||
1183                      !before(received_upto,
1184                              ack_seq + tp->reordering * tp->mss_cache))) {
1185                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1186                         tp->retrans_out -= tcp_skb_pcount(skb);
1187
1188                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1189                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1190                 } else {
1191                         if (before(ack_seq, new_low_seq))
1192                                 new_low_seq = ack_seq;
1193                         cnt += tcp_skb_pcount(skb);
1194                 }
1195         }
1196
1197         if (tp->retrans_out)
1198                 tp->lost_retrans_low = new_low_seq;
1199 }
1200
1201 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1202                            struct tcp_sack_block_wire *sp, int num_sacks,
1203                            u32 prior_snd_una)
1204 {
1205         struct tcp_sock *tp = tcp_sk(sk);
1206         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1207         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1208         int dup_sack = 0;
1209
1210         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1211                 dup_sack = 1;
1212                 tcp_dsack_seen(tp);
1213                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1214         } else if (num_sacks > 1) {
1215                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1216                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1217
1218                 if (!after(end_seq_0, end_seq_1) &&
1219                     !before(start_seq_0, start_seq_1)) {
1220                         dup_sack = 1;
1221                         tcp_dsack_seen(tp);
1222                         NET_INC_STATS_BH(sock_net(sk),
1223                                         LINUX_MIB_TCPDSACKOFORECV);
1224                 }
1225         }
1226
1227         /* D-SACK for already forgotten data... Do dumb counting. */
1228         if (dup_sack &&
1229             !after(end_seq_0, prior_snd_una) &&
1230             after(end_seq_0, tp->undo_marker))
1231                 tp->undo_retrans--;
1232
1233         return dup_sack;
1234 }
1235
1236 struct tcp_sacktag_state {
1237         int reord;
1238         int fack_count;
1239         int flag;
1240 };
1241
1242 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1243  * the incoming SACK may not exactly match but we can find smaller MSS
1244  * aligned portion of it that matches. Therefore we might need to fragment
1245  * which may fail and creates some hassle (caller must handle error case
1246  * returns).
1247  *
1248  * FIXME: this could be merged to shift decision code
1249  */
1250 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1251                                  u32 start_seq, u32 end_seq)
1252 {
1253         int in_sack, err;
1254         unsigned int pkt_len;
1255         unsigned int mss;
1256
1257         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1258                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1259
1260         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1261             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1262                 mss = tcp_skb_mss(skb);
1263                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1264
1265                 if (!in_sack) {
1266                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1267                         if (pkt_len < mss)
1268                                 pkt_len = mss;
1269                 } else {
1270                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1271                         if (pkt_len < mss)
1272                                 return -EINVAL;
1273                 }
1274
1275                 /* Round if necessary so that SACKs cover only full MSSes
1276                  * and/or the remaining small portion (if present)
1277                  */
1278                 if (pkt_len > mss) {
1279                         unsigned int new_len = (pkt_len / mss) * mss;
1280                         if (!in_sack && new_len < pkt_len) {
1281                                 new_len += mss;
1282                                 if (new_len > skb->len)
1283                                         return 0;
1284                         }
1285                         pkt_len = new_len;
1286                 }
1287                 err = tcp_fragment(sk, skb, pkt_len, mss);
1288                 if (err < 0)
1289                         return err;
1290         }
1291
1292         return in_sack;
1293 }
1294
1295 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1296                           struct tcp_sacktag_state *state,
1297                           int dup_sack, int pcount)
1298 {
1299         struct tcp_sock *tp = tcp_sk(sk);
1300         u8 sacked = TCP_SKB_CB(skb)->sacked;
1301         int fack_count = state->fack_count;
1302
1303         /* Account D-SACK for retransmitted packet. */
1304         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1305                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1306                         tp->undo_retrans--;
1307                 if (sacked & TCPCB_SACKED_ACKED)
1308                         state->reord = min(fack_count, state->reord);
1309         }
1310
1311         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1312         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1313                 return sacked;
1314
1315         if (!(sacked & TCPCB_SACKED_ACKED)) {
1316                 if (sacked & TCPCB_SACKED_RETRANS) {
1317                         /* If the segment is not tagged as lost,
1318                          * we do not clear RETRANS, believing
1319                          * that retransmission is still in flight.
1320                          */
1321                         if (sacked & TCPCB_LOST) {
1322                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1323                                 tp->lost_out -= pcount;
1324                                 tp->retrans_out -= pcount;
1325                         }
1326                 } else {
1327                         if (!(sacked & TCPCB_RETRANS)) {
1328                                 /* New sack for not retransmitted frame,
1329                                  * which was in hole. It is reordering.
1330                                  */
1331                                 if (before(TCP_SKB_CB(skb)->seq,
1332                                            tcp_highest_sack_seq(tp)))
1333                                         state->reord = min(fack_count,
1334                                                            state->reord);
1335
1336                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1337                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1338                                         state->flag |= FLAG_ONLY_ORIG_SACKED;
1339                         }
1340
1341                         if (sacked & TCPCB_LOST) {
1342                                 sacked &= ~TCPCB_LOST;
1343                                 tp->lost_out -= pcount;
1344                         }
1345                 }
1346
1347                 sacked |= TCPCB_SACKED_ACKED;
1348                 state->flag |= FLAG_DATA_SACKED;
1349                 tp->sacked_out += pcount;
1350
1351                 fack_count += pcount;
1352
1353                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1354                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1355                     before(TCP_SKB_CB(skb)->seq,
1356                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1357                         tp->lost_cnt_hint += pcount;
1358
1359                 if (fack_count > tp->fackets_out)
1360                         tp->fackets_out = fack_count;
1361         }
1362
1363         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1364          * frames and clear it. undo_retrans is decreased above, L|R frames
1365          * are accounted above as well.
1366          */
1367         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1368                 sacked &= ~TCPCB_SACKED_RETRANS;
1369                 tp->retrans_out -= pcount;
1370         }
1371
1372         return sacked;
1373 }
1374
1375 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1376                            struct tcp_sacktag_state *state,
1377                            unsigned int pcount, int shifted, int mss)
1378 {
1379         struct tcp_sock *tp = tcp_sk(sk);
1380         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1381
1382         BUG_ON(!pcount);
1383
1384         /* Tweak before seqno plays */
1385         if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1386             !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1387                 tp->lost_cnt_hint += pcount;
1388
1389         TCP_SKB_CB(prev)->end_seq += shifted;
1390         TCP_SKB_CB(skb)->seq += shifted;
1391
1392         skb_shinfo(prev)->gso_segs += pcount;
1393         BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1394         skb_shinfo(skb)->gso_segs -= pcount;
1395
1396         /* When we're adding to gso_segs == 1, gso_size will be zero,
1397          * in theory this shouldn't be necessary but as long as DSACK
1398          * code can come after this skb later on it's better to keep
1399          * setting gso_size to something.
1400          */
1401         if (!skb_shinfo(prev)->gso_size) {
1402                 skb_shinfo(prev)->gso_size = mss;
1403                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1404         }
1405
1406         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1407         if (skb_shinfo(skb)->gso_segs <= 1) {
1408                 skb_shinfo(skb)->gso_size = 0;
1409                 skb_shinfo(skb)->gso_type = 0;
1410         }
1411
1412         /* We discard results */
1413         tcp_sacktag_one(skb, sk, state, 0, pcount);
1414
1415         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1416         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1417
1418         if (skb->len > 0) {
1419                 BUG_ON(!tcp_skb_pcount(skb));
1420                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1421                 return 0;
1422         }
1423
1424         /* Whole SKB was eaten :-) */
1425
1426         if (skb == tp->retransmit_skb_hint)
1427                 tp->retransmit_skb_hint = prev;
1428         if (skb == tp->scoreboard_skb_hint)
1429                 tp->scoreboard_skb_hint = prev;
1430         if (skb == tp->lost_skb_hint) {
1431                 tp->lost_skb_hint = prev;
1432                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1433         }
1434
1435         TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1436         if (skb == tcp_highest_sack(sk))
1437                 tcp_advance_highest_sack(sk, skb);
1438
1439         tcp_unlink_write_queue(skb, sk);
1440         sk_wmem_free_skb(sk, skb);
1441
1442         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1443
1444         return 1;
1445 }
1446
1447 /* I wish gso_size would have a bit more sane initialization than
1448  * something-or-zero which complicates things
1449  */
1450 static int tcp_skb_seglen(struct sk_buff *skb)
1451 {
1452         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1453 }
1454
1455 /* Shifting pages past head area doesn't work */
1456 static int skb_can_shift(struct sk_buff *skb)
1457 {
1458         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1459 }
1460
1461 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1462  * skb.
1463  */
1464 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1465                                           struct tcp_sacktag_state *state,
1466                                           u32 start_seq, u32 end_seq,
1467                                           int dup_sack)
1468 {
1469         struct tcp_sock *tp = tcp_sk(sk);
1470         struct sk_buff *prev;
1471         int mss;
1472         int pcount = 0;
1473         int len;
1474         int in_sack;
1475
1476         if (!sk_can_gso(sk))
1477                 goto fallback;
1478
1479         /* Normally R but no L won't result in plain S */
1480         if (!dup_sack &&
1481             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1482                 goto fallback;
1483         if (!skb_can_shift(skb))
1484                 goto fallback;
1485         /* This frame is about to be dropped (was ACKed). */
1486         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1487                 goto fallback;
1488
1489         /* Can only happen with delayed DSACK + discard craziness */
1490         if (unlikely(skb == tcp_write_queue_head(sk)))
1491                 goto fallback;
1492         prev = tcp_write_queue_prev(sk, skb);
1493
1494         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1495                 goto fallback;
1496
1497         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1498                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1499
1500         if (in_sack) {
1501                 len = skb->len;
1502                 pcount = tcp_skb_pcount(skb);
1503                 mss = tcp_skb_seglen(skb);
1504
1505                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1506                  * drop this restriction as unnecessary
1507                  */
1508                 if (mss != tcp_skb_seglen(prev))
1509                         goto fallback;
1510         } else {
1511                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1512                         goto noop;
1513                 /* CHECKME: This is non-MSS split case only?, this will
1514                  * cause skipped skbs due to advancing loop btw, original
1515                  * has that feature too
1516                  */
1517                 if (tcp_skb_pcount(skb) <= 1)
1518                         goto noop;
1519
1520                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1521                 if (!in_sack) {
1522                         /* TODO: head merge to next could be attempted here
1523                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1524                          * though it might not be worth of the additional hassle
1525                          *
1526                          * ...we can probably just fallback to what was done
1527                          * previously. We could try merging non-SACKed ones
1528                          * as well but it probably isn't going to buy off
1529                          * because later SACKs might again split them, and
1530                          * it would make skb timestamp tracking considerably
1531                          * harder problem.
1532                          */
1533                         goto fallback;
1534                 }
1535
1536                 len = end_seq - TCP_SKB_CB(skb)->seq;
1537                 BUG_ON(len < 0);
1538                 BUG_ON(len > skb->len);
1539
1540                 /* MSS boundaries should be honoured or else pcount will
1541                  * severely break even though it makes things bit trickier.
1542                  * Optimize common case to avoid most of the divides
1543                  */
1544                 mss = tcp_skb_mss(skb);
1545
1546                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1547                  * drop this restriction as unnecessary
1548                  */
1549                 if (mss != tcp_skb_seglen(prev))
1550                         goto fallback;
1551
1552                 if (len == mss) {
1553                         pcount = 1;
1554                 } else if (len < mss) {
1555                         goto noop;
1556                 } else {
1557                         pcount = len / mss;
1558                         len = pcount * mss;
1559                 }
1560         }
1561
1562         if (!skb_shift(prev, skb, len))
1563                 goto fallback;
1564         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss))
1565                 goto out;
1566
1567         /* Hole filled allows collapsing with the next as well, this is very
1568          * useful when hole on every nth skb pattern happens
1569          */
1570         if (prev == tcp_write_queue_tail(sk))
1571                 goto out;
1572         skb = tcp_write_queue_next(sk, prev);
1573
1574         if (!skb_can_shift(skb) ||
1575             (skb == tcp_send_head(sk)) ||
1576             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1577             (mss != tcp_skb_seglen(skb)))
1578                 goto out;
1579
1580         len = skb->len;
1581         if (skb_shift(prev, skb, len)) {
1582                 pcount += tcp_skb_pcount(skb);
1583                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss);
1584         }
1585
1586 out:
1587         state->fack_count += pcount;
1588         return prev;
1589
1590 noop:
1591         return skb;
1592
1593 fallback:
1594         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1595         return NULL;
1596 }
1597
1598 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1599                                         struct tcp_sack_block *next_dup,
1600                                         struct tcp_sacktag_state *state,
1601                                         u32 start_seq, u32 end_seq,
1602                                         int dup_sack_in)
1603 {
1604         struct tcp_sock *tp = tcp_sk(sk);
1605         struct sk_buff *tmp;
1606
1607         tcp_for_write_queue_from(skb, sk) {
1608                 int in_sack = 0;
1609                 int dup_sack = dup_sack_in;
1610
1611                 if (skb == tcp_send_head(sk))
1612                         break;
1613
1614                 /* queue is in-order => we can short-circuit the walk early */
1615                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1616                         break;
1617
1618                 if ((next_dup != NULL) &&
1619                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1620                         in_sack = tcp_match_skb_to_sack(sk, skb,
1621                                                         next_dup->start_seq,
1622                                                         next_dup->end_seq);
1623                         if (in_sack > 0)
1624                                 dup_sack = 1;
1625                 }
1626
1627                 /* skb reference here is a bit tricky to get right, since
1628                  * shifting can eat and free both this skb and the next,
1629                  * so not even _safe variant of the loop is enough.
1630                  */
1631                 if (in_sack <= 0) {
1632                         tmp = tcp_shift_skb_data(sk, skb, state,
1633                                                  start_seq, end_seq, dup_sack);
1634                         if (tmp != NULL) {
1635                                 if (tmp != skb) {
1636                                         skb = tmp;
1637                                         continue;
1638                                 }
1639
1640                                 in_sack = 0;
1641                         } else {
1642                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1643                                                                 start_seq,
1644                                                                 end_seq);
1645                         }
1646                 }
1647
1648                 if (unlikely(in_sack < 0))
1649                         break;
1650
1651                 if (in_sack) {
1652                         TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1653                                                                   state,
1654                                                                   dup_sack,
1655                                                                   tcp_skb_pcount(skb));
1656
1657                         if (!before(TCP_SKB_CB(skb)->seq,
1658                                     tcp_highest_sack_seq(tp)))
1659                                 tcp_advance_highest_sack(sk, skb);
1660                 }
1661
1662                 state->fack_count += tcp_skb_pcount(skb);
1663         }
1664         return skb;
1665 }
1666
1667 /* Avoid all extra work that is being done by sacktag while walking in
1668  * a normal way
1669  */
1670 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1671                                         struct tcp_sacktag_state *state,
1672                                         u32 skip_to_seq)
1673 {
1674         tcp_for_write_queue_from(skb, sk) {
1675                 if (skb == tcp_send_head(sk))
1676                         break;
1677
1678                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1679                         break;
1680
1681                 state->fack_count += tcp_skb_pcount(skb);
1682         }
1683         return skb;
1684 }
1685
1686 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1687                                                 struct sock *sk,
1688                                                 struct tcp_sack_block *next_dup,
1689                                                 struct tcp_sacktag_state *state,
1690                                                 u32 skip_to_seq)
1691 {
1692         if (next_dup == NULL)
1693                 return skb;
1694
1695         if (before(next_dup->start_seq, skip_to_seq)) {
1696                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1697                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1698                                        next_dup->start_seq, next_dup->end_seq,
1699                                        1);
1700         }
1701
1702         return skb;
1703 }
1704
1705 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1706 {
1707         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1708 }
1709
1710 static int
1711 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1712                         u32 prior_snd_una)
1713 {
1714         const struct inet_connection_sock *icsk = inet_csk(sk);
1715         struct tcp_sock *tp = tcp_sk(sk);
1716         unsigned char *ptr = (skb_transport_header(ack_skb) +
1717                               TCP_SKB_CB(ack_skb)->sacked);
1718         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1719         struct tcp_sack_block sp[TCP_NUM_SACKS];
1720         struct tcp_sack_block *cache;
1721         struct tcp_sacktag_state state;
1722         struct sk_buff *skb;
1723         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1724         int used_sacks;
1725         int found_dup_sack = 0;
1726         int i, j;
1727         int first_sack_index;
1728
1729         state.flag = 0;
1730         state.reord = tp->packets_out;
1731
1732         if (!tp->sacked_out) {
1733                 if (WARN_ON(tp->fackets_out))
1734                         tp->fackets_out = 0;
1735                 tcp_highest_sack_reset(sk);
1736         }
1737
1738         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1739                                          num_sacks, prior_snd_una);
1740         if (found_dup_sack)
1741                 state.flag |= FLAG_DSACKING_ACK;
1742
1743         /* Eliminate too old ACKs, but take into
1744          * account more or less fresh ones, they can
1745          * contain valid SACK info.
1746          */
1747         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1748                 return 0;
1749
1750         if (!tp->packets_out)
1751                 goto out;
1752
1753         used_sacks = 0;
1754         first_sack_index = 0;
1755         for (i = 0; i < num_sacks; i++) {
1756                 int dup_sack = !i && found_dup_sack;
1757
1758                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1759                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1760
1761                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1762                                             sp[used_sacks].start_seq,
1763                                             sp[used_sacks].end_seq)) {
1764                         int mib_idx;
1765
1766                         if (dup_sack) {
1767                                 if (!tp->undo_marker)
1768                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1769                                 else
1770                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1771                         } else {
1772                                 /* Don't count olds caused by ACK reordering */
1773                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1774                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1775                                         continue;
1776                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1777                         }
1778
1779                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1780                         if (i == 0)
1781                                 first_sack_index = -1;
1782                         continue;
1783                 }
1784
1785                 /* Ignore very old stuff early */
1786                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1787                         continue;
1788
1789                 used_sacks++;
1790         }
1791
1792         /* order SACK blocks to allow in order walk of the retrans queue */
1793         for (i = used_sacks - 1; i > 0; i--) {
1794                 for (j = 0; j < i; j++) {
1795                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1796                                 struct tcp_sack_block tmp;
1797
1798                                 tmp = sp[j];
1799                                 sp[j] = sp[j + 1];
1800                                 sp[j + 1] = tmp;
1801
1802                                 /* Track where the first SACK block goes to */
1803                                 if (j == first_sack_index)
1804                                         first_sack_index = j + 1;
1805                         }
1806                 }
1807         }
1808
1809         skb = tcp_write_queue_head(sk);
1810         state.fack_count = 0;
1811         i = 0;
1812
1813         if (!tp->sacked_out) {
1814                 /* It's already past, so skip checking against it */
1815                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1816         } else {
1817                 cache = tp->recv_sack_cache;
1818                 /* Skip empty blocks in at head of the cache */
1819                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1820                        !cache->end_seq)
1821                         cache++;
1822         }
1823
1824         while (i < used_sacks) {
1825                 u32 start_seq = sp[i].start_seq;
1826                 u32 end_seq = sp[i].end_seq;
1827                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1828                 struct tcp_sack_block *next_dup = NULL;
1829
1830                 if (found_dup_sack && ((i + 1) == first_sack_index))
1831                         next_dup = &sp[i + 1];
1832
1833                 /* Event "B" in the comment above. */
1834                 if (after(end_seq, tp->high_seq))
1835                         state.flag |= FLAG_DATA_LOST;
1836
1837                 /* Skip too early cached blocks */
1838                 while (tcp_sack_cache_ok(tp, cache) &&
1839                        !before(start_seq, cache->end_seq))
1840                         cache++;
1841
1842                 /* Can skip some work by looking recv_sack_cache? */
1843                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1844                     after(end_seq, cache->start_seq)) {
1845
1846                         /* Head todo? */
1847                         if (before(start_seq, cache->start_seq)) {
1848                                 skb = tcp_sacktag_skip(skb, sk, &state,
1849                                                        start_seq);
1850                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1851                                                        &state,
1852                                                        start_seq,
1853                                                        cache->start_seq,
1854                                                        dup_sack);
1855                         }
1856
1857                         /* Rest of the block already fully processed? */
1858                         if (!after(end_seq, cache->end_seq))
1859                                 goto advance_sp;
1860
1861                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1862                                                        &state,
1863                                                        cache->end_seq);
1864
1865                         /* ...tail remains todo... */
1866                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1867                                 /* ...but better entrypoint exists! */
1868                                 skb = tcp_highest_sack(sk);
1869                                 if (skb == NULL)
1870                                         break;
1871                                 state.fack_count = tp->fackets_out;
1872                                 cache++;
1873                                 goto walk;
1874                         }
1875
1876                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1877                         /* Check overlap against next cached too (past this one already) */
1878                         cache++;
1879                         continue;
1880                 }
1881
1882                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1883                         skb = tcp_highest_sack(sk);
1884                         if (skb == NULL)
1885                                 break;
1886                         state.fack_count = tp->fackets_out;
1887                 }
1888                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1889
1890 walk:
1891                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1892                                        start_seq, end_seq, dup_sack);
1893
1894 advance_sp:
1895                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1896                  * due to in-order walk
1897                  */
1898                 if (after(end_seq, tp->frto_highmark))
1899                         state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1900
1901                 i++;
1902         }
1903
1904         /* Clear the head of the cache sack blocks so we can skip it next time */
1905         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1906                 tp->recv_sack_cache[i].start_seq = 0;
1907                 tp->recv_sack_cache[i].end_seq = 0;
1908         }
1909         for (j = 0; j < used_sacks; j++)
1910                 tp->recv_sack_cache[i++] = sp[j];
1911
1912         tcp_mark_lost_retrans(sk);
1913
1914         tcp_verify_left_out(tp);
1915
1916         if ((state.reord < tp->fackets_out) &&
1917             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1918             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1919                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1920
1921 out:
1922
1923 #if FASTRETRANS_DEBUG > 0
1924         WARN_ON((int)tp->sacked_out < 0);
1925         WARN_ON((int)tp->lost_out < 0);
1926         WARN_ON((int)tp->retrans_out < 0);
1927         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1928 #endif
1929         return state.flag;
1930 }
1931
1932 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1933  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1934  */
1935 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1936 {
1937         u32 holes;
1938
1939         holes = max(tp->lost_out, 1U);
1940         holes = min(holes, tp->packets_out);
1941
1942         if ((tp->sacked_out + holes) > tp->packets_out) {
1943                 tp->sacked_out = tp->packets_out - holes;
1944                 return 1;
1945         }
1946         return 0;
1947 }
1948
1949 /* If we receive more dupacks than we expected counting segments
1950  * in assumption of absent reordering, interpret this as reordering.
1951  * The only another reason could be bug in receiver TCP.
1952  */
1953 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1954 {
1955         struct tcp_sock *tp = tcp_sk(sk);
1956         if (tcp_limit_reno_sacked(tp))
1957                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1958 }
1959
1960 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1961
1962 static void tcp_add_reno_sack(struct sock *sk)
1963 {
1964         struct tcp_sock *tp = tcp_sk(sk);
1965         tp->sacked_out++;
1966         tcp_check_reno_reordering(sk, 0);
1967         tcp_verify_left_out(tp);
1968 }
1969
1970 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1971
1972 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1973 {
1974         struct tcp_sock *tp = tcp_sk(sk);
1975
1976         if (acked > 0) {
1977                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1978                 if (acked - 1 >= tp->sacked_out)
1979                         tp->sacked_out = 0;
1980                 else
1981                         tp->sacked_out -= acked - 1;
1982         }
1983         tcp_check_reno_reordering(sk, acked);
1984         tcp_verify_left_out(tp);
1985 }
1986
1987 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1988 {
1989         tp->sacked_out = 0;
1990 }
1991
1992 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1993 {
1994         return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1995 }
1996
1997 /* F-RTO can only be used if TCP has never retransmitted anything other than
1998  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1999  */
2000 int tcp_use_frto(struct sock *sk)
2001 {
2002         const struct tcp_sock *tp = tcp_sk(sk);
2003         const struct inet_connection_sock *icsk = inet_csk(sk);
2004         struct sk_buff *skb;
2005
2006         if (!sysctl_tcp_frto)
2007                 return 0;
2008
2009         /* MTU probe and F-RTO won't really play nicely along currently */
2010         if (icsk->icsk_mtup.probe_size)
2011                 return 0;
2012
2013         if (tcp_is_sackfrto(tp))
2014                 return 1;
2015
2016         /* Avoid expensive walking of rexmit queue if possible */
2017         if (tp->retrans_out > 1)
2018                 return 0;
2019
2020         skb = tcp_write_queue_head(sk);
2021         if (tcp_skb_is_last(sk, skb))
2022                 return 1;
2023         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
2024         tcp_for_write_queue_from(skb, sk) {
2025                 if (skb == tcp_send_head(sk))
2026                         break;
2027                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2028                         return 0;
2029                 /* Short-circuit when first non-SACKed skb has been checked */
2030                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2031                         break;
2032         }
2033         return 1;
2034 }
2035
2036 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2037  * recovery a bit and use heuristics in tcp_process_frto() to detect if
2038  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2039  * keep retrans_out counting accurate (with SACK F-RTO, other than head
2040  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2041  * bits are handled if the Loss state is really to be entered (in
2042  * tcp_enter_frto_loss).
2043  *
2044  * Do like tcp_enter_loss() would; when RTO expires the second time it
2045  * does:
2046  *  "Reduce ssthresh if it has not yet been made inside this window."
2047  */
2048 void tcp_enter_frto(struct sock *sk)
2049 {
2050         const struct inet_connection_sock *icsk = inet_csk(sk);
2051         struct tcp_sock *tp = tcp_sk(sk);
2052         struct sk_buff *skb;
2053
2054         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2055             tp->snd_una == tp->high_seq ||
2056             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2057              !icsk->icsk_retransmits)) {
2058                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2059                 /* Our state is too optimistic in ssthresh() call because cwnd
2060                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2061                  * recovery has not yet completed. Pattern would be this: RTO,
2062                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
2063                  * up here twice).
2064                  * RFC4138 should be more specific on what to do, even though
2065                  * RTO is quite unlikely to occur after the first Cumulative ACK
2066                  * due to back-off and complexity of triggering events ...
2067                  */
2068                 if (tp->frto_counter) {
2069                         u32 stored_cwnd;
2070                         stored_cwnd = tp->snd_cwnd;
2071                         tp->snd_cwnd = 2;
2072                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2073                         tp->snd_cwnd = stored_cwnd;
2074                 } else {
2075                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2076                 }
2077                 /* ... in theory, cong.control module could do "any tricks" in
2078                  * ssthresh(), which means that ca_state, lost bits and lost_out
2079                  * counter would have to be faked before the call occurs. We
2080                  * consider that too expensive, unlikely and hacky, so modules
2081                  * using these in ssthresh() must deal these incompatibility
2082                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2083                  */
2084                 tcp_ca_event(sk, CA_EVENT_FRTO);
2085         }
2086
2087         tp->undo_marker = tp->snd_una;
2088         tp->undo_retrans = 0;
2089
2090         skb = tcp_write_queue_head(sk);
2091         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2092                 tp->undo_marker = 0;
2093         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2094                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2095                 tp->retrans_out -= tcp_skb_pcount(skb);
2096         }
2097         tcp_verify_left_out(tp);
2098
2099         /* Too bad if TCP was application limited */
2100         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2101
2102         /* Earlier loss recovery underway (see RFC4138; Appendix B).
2103          * The last condition is necessary at least in tp->frto_counter case.
2104          */
2105         if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2106             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2107             after(tp->high_seq, tp->snd_una)) {
2108                 tp->frto_highmark = tp->high_seq;
2109         } else {
2110                 tp->frto_highmark = tp->snd_nxt;
2111         }
2112         tcp_set_ca_state(sk, TCP_CA_Disorder);
2113         tp->high_seq = tp->snd_nxt;
2114         tp->frto_counter = 1;
2115 }
2116
2117 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2118  * which indicates that we should follow the traditional RTO recovery,
2119  * i.e. mark everything lost and do go-back-N retransmission.
2120  */
2121 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2122 {
2123         struct tcp_sock *tp = tcp_sk(sk);
2124         struct sk_buff *skb;
2125
2126         tp->lost_out = 0;
2127         tp->retrans_out = 0;
2128         if (tcp_is_reno(tp))
2129                 tcp_reset_reno_sack(tp);
2130
2131         tcp_for_write_queue(skb, sk) {
2132                 if (skb == tcp_send_head(sk))
2133                         break;
2134
2135                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2136                 /*
2137                  * Count the retransmission made on RTO correctly (only when
2138                  * waiting for the first ACK and did not get it)...
2139                  */
2140                 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2141                         /* For some reason this R-bit might get cleared? */
2142                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2143                                 tp->retrans_out += tcp_skb_pcount(skb);
2144                         /* ...enter this if branch just for the first segment */
2145                         flag |= FLAG_DATA_ACKED;
2146                 } else {
2147                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2148                                 tp->undo_marker = 0;
2149                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2150                 }
2151
2152                 /* Marking forward transmissions that were made after RTO lost
2153                  * can cause unnecessary retransmissions in some scenarios,
2154                  * SACK blocks will mitigate that in some but not in all cases.
2155                  * We used to not mark them but it was causing break-ups with
2156                  * receivers that do only in-order receival.
2157                  *
2158                  * TODO: we could detect presence of such receiver and select
2159                  * different behavior per flow.
2160                  */
2161                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2162                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2163                         tp->lost_out += tcp_skb_pcount(skb);
2164                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2165                 }
2166         }
2167         tcp_verify_left_out(tp);
2168
2169         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2170         tp->snd_cwnd_cnt = 0;
2171         tp->snd_cwnd_stamp = tcp_time_stamp;
2172         tp->frto_counter = 0;
2173         tp->bytes_acked = 0;
2174
2175         tp->reordering = min_t(unsigned int, tp->reordering,
2176                                sysctl_tcp_reordering);
2177         tcp_set_ca_state(sk, TCP_CA_Loss);
2178         tp->high_seq = tp->snd_nxt;
2179         TCP_ECN_queue_cwr(tp);
2180
2181         tcp_clear_all_retrans_hints(tp);
2182 }
2183
2184 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2185 {
2186         tp->retrans_out = 0;
2187         tp->lost_out = 0;
2188
2189         tp->undo_marker = 0;
2190         tp->undo_retrans = 0;
2191 }
2192
2193 void tcp_clear_retrans(struct tcp_sock *tp)
2194 {
2195         tcp_clear_retrans_partial(tp);
2196
2197         tp->fackets_out = 0;
2198         tp->sacked_out = 0;
2199 }
2200
2201 /* Enter Loss state. If "how" is not zero, forget all SACK information
2202  * and reset tags completely, otherwise preserve SACKs. If receiver
2203  * dropped its ofo queue, we will know this due to reneging detection.
2204  */
2205 void tcp_enter_loss(struct sock *sk, int how)
2206 {
2207         const struct inet_connection_sock *icsk = inet_csk(sk);
2208         struct tcp_sock *tp = tcp_sk(sk);
2209         struct sk_buff *skb;
2210
2211         /* Reduce ssthresh if it has not yet been made inside this window. */
2212         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2213             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2214                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2215                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2216                 tcp_ca_event(sk, CA_EVENT_LOSS);
2217         }
2218         tp->snd_cwnd       = 1;
2219         tp->snd_cwnd_cnt   = 0;
2220         tp->snd_cwnd_stamp = tcp_time_stamp;
2221
2222         tp->bytes_acked = 0;
2223         tcp_clear_retrans_partial(tp);
2224
2225         if (tcp_is_reno(tp))
2226                 tcp_reset_reno_sack(tp);
2227
2228         if (!how) {
2229                 /* Push undo marker, if it was plain RTO and nothing
2230                  * was retransmitted. */
2231                 tp->undo_marker = tp->snd_una;
2232         } else {
2233                 tp->sacked_out = 0;
2234                 tp->fackets_out = 0;
2235         }
2236         tcp_clear_all_retrans_hints(tp);
2237
2238         tcp_for_write_queue(skb, sk) {
2239                 if (skb == tcp_send_head(sk))
2240                         break;
2241
2242                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2243                         tp->undo_marker = 0;
2244                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2245                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2246                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2247                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2248                         tp->lost_out += tcp_skb_pcount(skb);
2249                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2250                 }
2251         }
2252         tcp_verify_left_out(tp);
2253
2254         tp->reordering = min_t(unsigned int, tp->reordering,
2255                                sysctl_tcp_reordering);
2256         tcp_set_ca_state(sk, TCP_CA_Loss);
2257         tp->high_seq = tp->snd_nxt;
2258         TCP_ECN_queue_cwr(tp);
2259         /* Abort F-RTO algorithm if one is in progress */
2260         tp->frto_counter = 0;
2261 }
2262
2263 /* If ACK arrived pointing to a remembered SACK, it means that our
2264  * remembered SACKs do not reflect real state of receiver i.e.
2265  * receiver _host_ is heavily congested (or buggy).
2266  *
2267  * Do processing similar to RTO timeout.
2268  */
2269 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2270 {
2271         if (flag & FLAG_SACK_RENEGING) {
2272                 struct inet_connection_sock *icsk = inet_csk(sk);
2273                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2274
2275                 tcp_enter_loss(sk, 1);
2276                 icsk->icsk_retransmits++;
2277                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2278                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2279                                           icsk->icsk_rto, TCP_RTO_MAX);
2280                 return 1;
2281         }
2282         return 0;
2283 }
2284
2285 static inline int tcp_fackets_out(struct tcp_sock *tp)
2286 {
2287         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2288 }
2289
2290 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2291  * counter when SACK is enabled (without SACK, sacked_out is used for
2292  * that purpose).
2293  *
2294  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2295  * segments up to the highest received SACK block so far and holes in
2296  * between them.
2297  *
2298  * With reordering, holes may still be in flight, so RFC3517 recovery
2299  * uses pure sacked_out (total number of SACKed segments) even though
2300  * it violates the RFC that uses duplicate ACKs, often these are equal
2301  * but when e.g. out-of-window ACKs or packet duplication occurs,
2302  * they differ. Since neither occurs due to loss, TCP should really
2303  * ignore them.
2304  */
2305 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2306 {
2307         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2308 }
2309
2310 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2311 {
2312         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2313 }
2314
2315 static inline int tcp_head_timedout(struct sock *sk)
2316 {
2317         struct tcp_sock *tp = tcp_sk(sk);
2318
2319         return tp->packets_out &&
2320                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2321 }
2322
2323 /* Linux NewReno/SACK/FACK/ECN state machine.
2324  * --------------------------------------
2325  *
2326  * "Open"       Normal state, no dubious events, fast path.
2327  * "Disorder"   In all the respects it is "Open",
2328  *              but requires a bit more attention. It is entered when
2329  *              we see some SACKs or dupacks. It is split of "Open"
2330  *              mainly to move some processing from fast path to slow one.
2331  * "CWR"        CWND was reduced due to some Congestion Notification event.
2332  *              It can be ECN, ICMP source quench, local device congestion.
2333  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2334  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2335  *
2336  * tcp_fastretrans_alert() is entered:
2337  * - each incoming ACK, if state is not "Open"
2338  * - when arrived ACK is unusual, namely:
2339  *      * SACK
2340  *      * Duplicate ACK.
2341  *      * ECN ECE.
2342  *
2343  * Counting packets in flight is pretty simple.
2344  *
2345  *      in_flight = packets_out - left_out + retrans_out
2346  *
2347  *      packets_out is SND.NXT-SND.UNA counted in packets.
2348  *
2349  *      retrans_out is number of retransmitted segments.
2350  *
2351  *      left_out is number of segments left network, but not ACKed yet.
2352  *
2353  *              left_out = sacked_out + lost_out
2354  *
2355  *     sacked_out: Packets, which arrived to receiver out of order
2356  *                 and hence not ACKed. With SACKs this number is simply
2357  *                 amount of SACKed data. Even without SACKs
2358  *                 it is easy to give pretty reliable estimate of this number,
2359  *                 counting duplicate ACKs.
2360  *
2361  *       lost_out: Packets lost by network. TCP has no explicit
2362  *                 "loss notification" feedback from network (for now).
2363  *                 It means that this number can be only _guessed_.
2364  *                 Actually, it is the heuristics to predict lossage that
2365  *                 distinguishes different algorithms.
2366  *
2367  *      F.e. after RTO, when all the queue is considered as lost,
2368  *      lost_out = packets_out and in_flight = retrans_out.
2369  *
2370  *              Essentially, we have now two algorithms counting
2371  *              lost packets.
2372  *
2373  *              FACK: It is the simplest heuristics. As soon as we decided
2374  *              that something is lost, we decide that _all_ not SACKed
2375  *              packets until the most forward SACK are lost. I.e.
2376  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2377  *              It is absolutely correct estimate, if network does not reorder
2378  *              packets. And it loses any connection to reality when reordering
2379  *              takes place. We use FACK by default until reordering
2380  *              is suspected on the path to this destination.
2381  *
2382  *              NewReno: when Recovery is entered, we assume that one segment
2383  *              is lost (classic Reno). While we are in Recovery and
2384  *              a partial ACK arrives, we assume that one more packet
2385  *              is lost (NewReno). This heuristics are the same in NewReno
2386  *              and SACK.
2387  *
2388  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2389  *  deflation etc. CWND is real congestion window, never inflated, changes
2390  *  only according to classic VJ rules.
2391  *
2392  * Really tricky (and requiring careful tuning) part of algorithm
2393  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2394  * The first determines the moment _when_ we should reduce CWND and,
2395  * hence, slow down forward transmission. In fact, it determines the moment
2396  * when we decide that hole is caused by loss, rather than by a reorder.
2397  *
2398  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2399  * holes, caused by lost packets.
2400  *
2401  * And the most logically complicated part of algorithm is undo
2402  * heuristics. We detect false retransmits due to both too early
2403  * fast retransmit (reordering) and underestimated RTO, analyzing
2404  * timestamps and D-SACKs. When we detect that some segments were
2405  * retransmitted by mistake and CWND reduction was wrong, we undo
2406  * window reduction and abort recovery phase. This logic is hidden
2407  * inside several functions named tcp_try_undo_<something>.
2408  */
2409
2410 /* This function decides, when we should leave Disordered state
2411  * and enter Recovery phase, reducing congestion window.
2412  *
2413  * Main question: may we further continue forward transmission
2414  * with the same cwnd?
2415  */
2416 static int tcp_time_to_recover(struct sock *sk)
2417 {
2418         struct tcp_sock *tp = tcp_sk(sk);
2419         __u32 packets_out;
2420
2421         /* Do not perform any recovery during F-RTO algorithm */
2422         if (tp->frto_counter)
2423                 return 0;
2424
2425         /* Trick#1: The loss is proven. */
2426         if (tp->lost_out)
2427                 return 1;
2428
2429         /* Not-A-Trick#2 : Classic rule... */
2430         if (tcp_dupack_heurestics(tp) > tp->reordering)
2431                 return 1;
2432
2433         /* Trick#3 : when we use RFC2988 timer restart, fast
2434          * retransmit can be triggered by timeout of queue head.
2435          */
2436         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2437                 return 1;
2438
2439         /* Trick#4: It is still not OK... But will it be useful to delay
2440          * recovery more?
2441          */
2442         packets_out = tp->packets_out;
2443         if (packets_out <= tp->reordering &&
2444             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2445             !tcp_may_send_now(sk)) {
2446                 /* We have nothing to send. This connection is limited
2447                  * either by receiver window or by application.
2448                  */
2449                 return 1;
2450         }
2451
2452         return 0;
2453 }
2454
2455 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2456  * is against sacked "cnt", otherwise it's against facked "cnt"
2457  */
2458 static void tcp_mark_head_lost(struct sock *sk, int packets)
2459 {
2460         struct tcp_sock *tp = tcp_sk(sk);
2461         struct sk_buff *skb;
2462         int cnt, oldcnt;
2463         int err;
2464         unsigned int mss;
2465
2466         WARN_ON(packets > tp->packets_out);
2467         if (tp->lost_skb_hint) {
2468                 skb = tp->lost_skb_hint;
2469                 cnt = tp->lost_cnt_hint;
2470         } else {
2471                 skb = tcp_write_queue_head(sk);
2472                 cnt = 0;
2473         }
2474
2475         tcp_for_write_queue_from(skb, sk) {
2476                 if (skb == tcp_send_head(sk))
2477                         break;
2478                 /* TODO: do this better */
2479                 /* this is not the most efficient way to do this... */
2480                 tp->lost_skb_hint = skb;
2481                 tp->lost_cnt_hint = cnt;
2482
2483                 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2484                         break;
2485
2486                 oldcnt = cnt;
2487                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2488                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2489                         cnt += tcp_skb_pcount(skb);
2490
2491                 if (cnt > packets) {
2492                         if (tcp_is_sack(tp) || (oldcnt >= packets))
2493                                 break;
2494
2495                         mss = skb_shinfo(skb)->gso_size;
2496                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2497                         if (err < 0)
2498                                 break;
2499                         cnt = packets;
2500                 }
2501
2502                 tcp_skb_mark_lost(tp, skb);
2503         }
2504         tcp_verify_left_out(tp);
2505 }
2506
2507 /* Account newly detected lost packet(s) */
2508
2509 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2510 {
2511         struct tcp_sock *tp = tcp_sk(sk);
2512
2513         if (tcp_is_reno(tp)) {
2514                 tcp_mark_head_lost(sk, 1);
2515         } else if (tcp_is_fack(tp)) {
2516                 int lost = tp->fackets_out - tp->reordering;
2517                 if (lost <= 0)
2518                         lost = 1;
2519                 tcp_mark_head_lost(sk, lost);
2520         } else {
2521                 int sacked_upto = tp->sacked_out - tp->reordering;
2522                 if (sacked_upto < fast_rexmit)
2523                         sacked_upto = fast_rexmit;
2524                 tcp_mark_head_lost(sk, sacked_upto);
2525         }
2526
2527         /* New heuristics: it is possible only after we switched
2528          * to restart timer each time when something is ACKed.
2529          * Hence, we can detect timed out packets during fast
2530          * retransmit without falling to slow start.
2531          */
2532         if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2533                 struct sk_buff *skb;
2534
2535                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2536                         : tcp_write_queue_head(sk);
2537
2538                 tcp_for_write_queue_from(skb, sk) {
2539                         if (skb == tcp_send_head(sk))
2540                                 break;
2541                         if (!tcp_skb_timedout(sk, skb))
2542                                 break;
2543
2544                         tcp_skb_mark_lost(tp, skb);
2545                 }
2546
2547                 tp->scoreboard_skb_hint = skb;
2548
2549                 tcp_verify_left_out(tp);
2550         }
2551 }
2552
2553 /* CWND moderation, preventing bursts due to too big ACKs
2554  * in dubious situations.
2555  */
2556 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2557 {
2558         tp->snd_cwnd = min(tp->snd_cwnd,
2559                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2560         tp->snd_cwnd_stamp = tcp_time_stamp;
2561 }
2562
2563 /* Lower bound on congestion window is slow start threshold
2564  * unless congestion avoidance choice decides to overide it.
2565  */
2566 static inline u32 tcp_cwnd_min(const struct sock *sk)
2567 {
2568         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2569
2570         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2571 }
2572
2573 /* Decrease cwnd each second ack. */
2574 static void tcp_cwnd_down(struct sock *sk, int flag)
2575 {
2576         struct tcp_sock *tp = tcp_sk(sk);
2577         int decr = tp->snd_cwnd_cnt + 1;
2578
2579         if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2580             (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2581                 tp->snd_cwnd_cnt = decr & 1;
2582                 decr >>= 1;
2583
2584                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2585                         tp->snd_cwnd -= decr;
2586
2587                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2588                 tp->snd_cwnd_stamp = tcp_time_stamp;
2589         }
2590 }
2591
2592 /* Nothing was retransmitted or returned timestamp is less
2593  * than timestamp of the first retransmission.
2594  */
2595 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2596 {
2597         return !tp->retrans_stamp ||
2598                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2599                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2600 }
2601
2602 /* Undo procedures. */
2603
2604 #if FASTRETRANS_DEBUG > 1
2605 static void DBGUNDO(struct sock *sk, const char *msg)
2606 {
2607         struct tcp_sock *tp = tcp_sk(sk);
2608         struct inet_sock *inet = inet_sk(sk);
2609
2610         if (sk->sk_family == AF_INET) {
2611                 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2612                        msg,
2613                        &inet->daddr, ntohs(inet->dport),
2614                        tp->snd_cwnd, tcp_left_out(tp),
2615                        tp->snd_ssthresh, tp->prior_ssthresh,
2616                        tp->packets_out);
2617         }
2618 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2619         else if (sk->sk_family == AF_INET6) {
2620                 struct ipv6_pinfo *np = inet6_sk(sk);
2621                 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2622                        msg,
2623                        &np->daddr, ntohs(inet->dport),
2624                        tp->snd_cwnd, tcp_left_out(tp),
2625                        tp->snd_ssthresh, tp->prior_ssthresh,
2626                        tp->packets_out);
2627         }
2628 #endif
2629 }
2630 #else
2631 #define DBGUNDO(x...) do { } while (0)
2632 #endif
2633
2634 static void tcp_undo_cwr(struct sock *sk, const int undo)
2635 {
2636         struct tcp_sock *tp = tcp_sk(sk);
2637
2638         if (tp->prior_ssthresh) {
2639                 const struct inet_connection_sock *icsk = inet_csk(sk);
2640
2641                 if (icsk->icsk_ca_ops->undo_cwnd)
2642                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2643                 else
2644                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2645
2646                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2647                         tp->snd_ssthresh = tp->prior_ssthresh;
2648                         TCP_ECN_withdraw_cwr(tp);
2649                 }
2650         } else {
2651                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2652         }
2653         tcp_moderate_cwnd(tp);
2654         tp->snd_cwnd_stamp = tcp_time_stamp;
2655 }
2656
2657 static inline int tcp_may_undo(struct tcp_sock *tp)
2658 {
2659         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2660 }
2661
2662 /* People celebrate: "We love our President!" */
2663 static int tcp_try_undo_recovery(struct sock *sk)
2664 {
2665         struct tcp_sock *tp = tcp_sk(sk);
2666
2667         if (tcp_may_undo(tp)) {
2668                 int mib_idx;
2669
2670                 /* Happy end! We did not retransmit anything
2671                  * or our original transmission succeeded.
2672                  */
2673                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2674                 tcp_undo_cwr(sk, 1);
2675                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2676                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2677                 else
2678                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2679
2680                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2681                 tp->undo_marker = 0;
2682         }
2683         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2684                 /* Hold old state until something *above* high_seq
2685                  * is ACKed. For Reno it is MUST to prevent false
2686                  * fast retransmits (RFC2582). SACK TCP is safe. */
2687                 tcp_moderate_cwnd(tp);
2688                 return 1;
2689         }
2690         tcp_set_ca_state(sk, TCP_CA_Open);
2691         return 0;
2692 }
2693
2694 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2695 static void tcp_try_undo_dsack(struct sock *sk)
2696 {
2697         struct tcp_sock *tp = tcp_sk(sk);
2698
2699         if (tp->undo_marker && !tp->undo_retrans) {
2700                 DBGUNDO(sk, "D-SACK");
2701                 tcp_undo_cwr(sk, 1);
2702                 tp->undo_marker = 0;
2703                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2704         }
2705 }
2706
2707 /* Undo during fast recovery after partial ACK. */
2708
2709 static int tcp_try_undo_partial(struct sock *sk, int acked)
2710 {
2711         struct tcp_sock *tp = tcp_sk(sk);
2712         /* Partial ACK arrived. Force Hoe's retransmit. */
2713         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2714
2715         if (tcp_may_undo(tp)) {
2716                 /* Plain luck! Hole if filled with delayed
2717                  * packet, rather than with a retransmit.
2718                  */
2719                 if (tp->retrans_out == 0)
2720                         tp->retrans_stamp = 0;
2721
2722                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2723
2724                 DBGUNDO(sk, "Hoe");
2725                 tcp_undo_cwr(sk, 0);
2726                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2727
2728                 /* So... Do not make Hoe's retransmit yet.
2729                  * If the first packet was delayed, the rest
2730                  * ones are most probably delayed as well.
2731                  */
2732                 failed = 0;
2733         }
2734         return failed;
2735 }
2736
2737 /* Undo during loss recovery after partial ACK. */
2738 static int tcp_try_undo_loss(struct sock *sk)
2739 {
2740         struct tcp_sock *tp = tcp_sk(sk);
2741
2742         if (tcp_may_undo(tp)) {
2743                 struct sk_buff *skb;
2744                 tcp_for_write_queue(skb, sk) {
2745                         if (skb == tcp_send_head(sk))
2746                                 break;
2747                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2748                 }
2749
2750                 tcp_clear_all_retrans_hints(tp);
2751
2752                 DBGUNDO(sk, "partial loss");
2753                 tp->lost_out = 0;
2754                 tcp_undo_cwr(sk, 1);
2755                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2756                 inet_csk(sk)->icsk_retransmits = 0;
2757                 tp->undo_marker = 0;
2758                 if (tcp_is_sack(tp))
2759                         tcp_set_ca_state(sk, TCP_CA_Open);
2760                 return 1;
2761         }
2762         return 0;
2763 }
2764
2765 static inline void tcp_complete_cwr(struct sock *sk)
2766 {
2767         struct tcp_sock *tp = tcp_sk(sk);
2768         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2769         tp->snd_cwnd_stamp = tcp_time_stamp;
2770         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2771 }
2772
2773 static void tcp_try_keep_open(struct sock *sk)
2774 {
2775         struct tcp_sock *tp = tcp_sk(sk);
2776         int state = TCP_CA_Open;
2777
2778         if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2779                 state = TCP_CA_Disorder;
2780
2781         if (inet_csk(sk)->icsk_ca_state != state) {
2782                 tcp_set_ca_state(sk, state);
2783                 tp->high_seq = tp->snd_nxt;
2784         }
2785 }
2786
2787 static void tcp_try_to_open(struct sock *sk, int flag)
2788 {
2789         struct tcp_sock *tp = tcp_sk(sk);
2790
2791         tcp_verify_left_out(tp);
2792
2793         if (!tp->frto_counter && tp->retrans_out == 0)
2794                 tp->retrans_stamp = 0;
2795
2796         if (flag & FLAG_ECE)
2797                 tcp_enter_cwr(sk, 1);
2798
2799         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2800                 tcp_try_keep_open(sk);
2801                 tcp_moderate_cwnd(tp);
2802         } else {
2803                 tcp_cwnd_down(sk, flag);
2804         }
2805 }
2806
2807 static void tcp_mtup_probe_failed(struct sock *sk)
2808 {
2809         struct inet_connection_sock *icsk = inet_csk(sk);
2810
2811         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2812         icsk->icsk_mtup.probe_size = 0;
2813 }
2814
2815 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2816 {
2817         struct tcp_sock *tp = tcp_sk(sk);
2818         struct inet_connection_sock *icsk = inet_csk(sk);
2819
2820         /* FIXME: breaks with very large cwnd */
2821         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2822         tp->snd_cwnd = tp->snd_cwnd *
2823                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2824                        icsk->icsk_mtup.probe_size;
2825         tp->snd_cwnd_cnt = 0;
2826         tp->snd_cwnd_stamp = tcp_time_stamp;
2827         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2828
2829         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2830         icsk->icsk_mtup.probe_size = 0;
2831         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2832 }
2833
2834 /* Do a simple retransmit without using the backoff mechanisms in
2835  * tcp_timer. This is used for path mtu discovery.
2836  * The socket is already locked here.
2837  */
2838 void tcp_simple_retransmit(struct sock *sk)
2839 {
2840         const struct inet_connection_sock *icsk = inet_csk(sk);
2841         struct tcp_sock *tp = tcp_sk(sk);
2842         struct sk_buff *skb;
2843         unsigned int mss = tcp_current_mss(sk, 0);
2844         u32 prior_lost = tp->lost_out;
2845
2846         tcp_for_write_queue(skb, sk) {
2847                 if (skb == tcp_send_head(sk))
2848                         break;
2849                 if (tcp_skb_seglen(skb) > mss &&
2850                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2851                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2852                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2853                                 tp->retrans_out -= tcp_skb_pcount(skb);
2854                         }
2855                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2856                 }
2857         }
2858
2859         tcp_clear_retrans_hints_partial(tp);
2860
2861         if (prior_lost == tp->lost_out)
2862                 return;
2863
2864         if (tcp_is_reno(tp))
2865                 tcp_limit_reno_sacked(tp);
2866
2867         tcp_verify_left_out(tp);
2868
2869         /* Don't muck with the congestion window here.
2870          * Reason is that we do not increase amount of _data_
2871          * in network, but units changed and effective
2872          * cwnd/ssthresh really reduced now.
2873          */
2874         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2875                 tp->high_seq = tp->snd_nxt;
2876                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2877                 tp->prior_ssthresh = 0;
2878                 tp->undo_marker = 0;
2879                 tcp_set_ca_state(sk, TCP_CA_Loss);
2880         }
2881         tcp_xmit_retransmit_queue(sk);
2882 }
2883
2884 /* Process an event, which can update packets-in-flight not trivially.
2885  * Main goal of this function is to calculate new estimate for left_out,
2886  * taking into account both packets sitting in receiver's buffer and
2887  * packets lost by network.
2888  *
2889  * Besides that it does CWND reduction, when packet loss is detected
2890  * and changes state of machine.
2891  *
2892  * It does _not_ decide what to send, it is made in function
2893  * tcp_xmit_retransmit_queue().
2894  */
2895 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2896 {
2897         struct inet_connection_sock *icsk = inet_csk(sk);
2898         struct tcp_sock *tp = tcp_sk(sk);
2899         int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2900         int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2901                                     (tcp_fackets_out(tp) > tp->reordering));
2902         int fast_rexmit = 0, mib_idx;
2903
2904         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2905                 tp->sacked_out = 0;
2906         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2907                 tp->fackets_out = 0;
2908
2909         /* Now state machine starts.
2910          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2911         if (flag & FLAG_ECE)
2912                 tp->prior_ssthresh = 0;
2913
2914         /* B. In all the states check for reneging SACKs. */
2915         if (tcp_check_sack_reneging(sk, flag))
2916                 return;
2917
2918         /* C. Process data loss notification, provided it is valid. */
2919         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2920             before(tp->snd_una, tp->high_seq) &&
2921             icsk->icsk_ca_state != TCP_CA_Open &&
2922             tp->fackets_out > tp->reordering) {
2923                 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2924                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2925         }
2926
2927         /* D. Check consistency of the current state. */
2928         tcp_verify_left_out(tp);
2929
2930         /* E. Check state exit conditions. State can be terminated
2931          *    when high_seq is ACKed. */
2932         if (icsk->icsk_ca_state == TCP_CA_Open) {
2933                 WARN_ON(tp->retrans_out != 0);
2934                 tp->retrans_stamp = 0;
2935         } else if (!before(tp->snd_una, tp->high_seq)) {
2936                 switch (icsk->icsk_ca_state) {
2937                 case TCP_CA_Loss:
2938                         icsk->icsk_retransmits = 0;
2939                         if (tcp_try_undo_recovery(sk))
2940                                 return;
2941                         break;
2942
2943                 case TCP_CA_CWR:
2944                         /* CWR is to be held something *above* high_seq
2945                          * is ACKed for CWR bit to reach receiver. */
2946                         if (tp->snd_una != tp->high_seq) {
2947                                 tcp_complete_cwr(sk);
2948                                 tcp_set_ca_state(sk, TCP_CA_Open);
2949                         }
2950                         break;
2951
2952                 case TCP_CA_Disorder:
2953                         tcp_try_undo_dsack(sk);
2954                         if (!tp->undo_marker ||
2955                             /* For SACK case do not Open to allow to undo
2956                              * catching for all duplicate ACKs. */
2957                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2958                                 tp->undo_marker = 0;
2959                                 tcp_set_ca_state(sk, TCP_CA_Open);
2960                         }
2961                         break;
2962
2963                 case TCP_CA_Recovery:
2964                         if (tcp_is_reno(tp))
2965                                 tcp_reset_reno_sack(tp);
2966                         if (tcp_try_undo_recovery(sk))
2967                                 return;
2968                         tcp_complete_cwr(sk);
2969                         break;
2970                 }
2971         }
2972
2973         /* F. Process state. */
2974         switch (icsk->icsk_ca_state) {
2975         case TCP_CA_Recovery:
2976                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2977                         if (tcp_is_reno(tp) && is_dupack)
2978                                 tcp_add_reno_sack(sk);
2979                 } else
2980                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2981                 break;
2982         case TCP_CA_Loss:
2983                 if (flag & FLAG_DATA_ACKED)
2984                         icsk->icsk_retransmits = 0;
2985                 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2986                         tcp_reset_reno_sack(tp);
2987                 if (!tcp_try_undo_loss(sk)) {
2988                         tcp_moderate_cwnd(tp);
2989                         tcp_xmit_retransmit_queue(sk);
2990                         return;
2991                 }
2992                 if (icsk->icsk_ca_state != TCP_CA_Open)
2993                         return;
2994                 /* Loss is undone; fall through to processing in Open state. */
2995         default:
2996                 if (tcp_is_reno(tp)) {
2997                         if (flag & FLAG_SND_UNA_ADVANCED)
2998                                 tcp_reset_reno_sack(tp);
2999                         if (is_dupack)
3000                                 tcp_add_reno_sack(sk);
3001                 }
3002
3003                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3004                         tcp_try_undo_dsack(sk);
3005
3006                 if (!tcp_time_to_recover(sk)) {
3007                         tcp_try_to_open(sk, flag);
3008                         return;
3009                 }
3010
3011                 /* MTU probe failure: don't reduce cwnd */
3012                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3013                     icsk->icsk_mtup.probe_size &&
3014                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
3015                         tcp_mtup_probe_failed(sk);
3016                         /* Restores the reduction we did in tcp_mtup_probe() */
3017                         tp->snd_cwnd++;
3018                         tcp_simple_retransmit(sk);
3019                         return;
3020                 }
3021
3022                 /* Otherwise enter Recovery state */
3023
3024                 if (tcp_is_reno(tp))
3025                         mib_idx = LINUX_MIB_TCPRENORECOVERY;
3026                 else
3027                         mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3028
3029                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3030
3031                 tp->high_seq = tp->snd_nxt;
3032                 tp->prior_ssthresh = 0;
3033                 tp->undo_marker = tp->snd_una;
3034                 tp->undo_retrans = tp->retrans_out;
3035
3036                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3037                         if (!(flag & FLAG_ECE))
3038                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3039                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3040                         TCP_ECN_queue_cwr(tp);
3041                 }
3042
3043                 tp->bytes_acked = 0;
3044                 tp->snd_cwnd_cnt = 0;
3045                 tcp_set_ca_state(sk, TCP_CA_Recovery);
3046                 fast_rexmit = 1;
3047         }
3048
3049         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3050                 tcp_update_scoreboard(sk, fast_rexmit);
3051         tcp_cwnd_down(sk, flag);
3052         tcp_xmit_retransmit_queue(sk);
3053 }
3054
3055 static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3056 {
3057         tcp_rtt_estimator(sk, seq_rtt);
3058         tcp_set_rto(sk);
3059         inet_csk(sk)->icsk_backoff = 0;
3060 }
3061
3062 /* Read draft-ietf-tcplw-high-performance before mucking
3063  * with this code. (Supersedes RFC1323)
3064  */
3065 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3066 {
3067         /* RTTM Rule: A TSecr value received in a segment is used to
3068          * update the averaged RTT measurement only if the segment
3069          * acknowledges some new data, i.e., only if it advances the
3070          * left edge of the send window.
3071          *
3072          * See draft-ietf-tcplw-high-performance-00, section 3.3.
3073          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3074          *
3075          * Changed: reset backoff as soon as we see the first valid sample.
3076          * If we do not, we get strongly overestimated rto. With timestamps
3077          * samples are accepted even from very old segments: f.e., when rtt=1
3078          * increases to 8, we retransmit 5 times and after 8 seconds delayed
3079          * answer arrives rto becomes 120 seconds! If at least one of segments
3080          * in window is lost... Voila.                          --ANK (010210)
3081          */
3082         struct tcp_sock *tp = tcp_sk(sk);
3083
3084         tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3085 }
3086
3087 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3088 {
3089         /* We don't have a timestamp. Can only use
3090          * packets that are not retransmitted to determine
3091          * rtt estimates. Also, we must not reset the
3092          * backoff for rto until we get a non-retransmitted
3093          * packet. This allows us to deal with a situation
3094          * where the network delay has increased suddenly.
3095          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3096          */
3097
3098         if (flag & FLAG_RETRANS_DATA_ACKED)
3099                 return;
3100
3101         tcp_valid_rtt_meas(sk, seq_rtt);
3102 }
3103
3104 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3105                                       const s32 seq_rtt)
3106 {
3107         const struct tcp_sock *tp = tcp_sk(sk);
3108         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3109         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3110                 tcp_ack_saw_tstamp(sk, flag);
3111         else if (seq_rtt >= 0)
3112                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3113 }
3114
3115 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3116 {
3117         const struct inet_connection_sock *icsk = inet_csk(sk);
3118         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3119         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3120 }
3121
3122 /* Restart timer after forward progress on connection.
3123  * RFC2988 recommends to restart timer to now+rto.
3124  */
3125 static void tcp_rearm_rto(struct sock *sk)
3126 {
3127         struct tcp_sock *tp = tcp_sk(sk);
3128
3129         if (!tp->packets_out) {
3130                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3131         } else {
3132                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3133                                           inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3134         }
3135 }
3136
3137 /* If we get here, the whole TSO packet has not been acked. */
3138 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3139 {
3140         struct tcp_sock *tp = tcp_sk(sk);
3141         u32 packets_acked;
3142
3143         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3144
3145         packets_acked = tcp_skb_pcount(skb);
3146         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3147                 return 0;
3148         packets_acked -= tcp_skb_pcount(skb);
3149
3150         if (packets_acked) {
3151                 BUG_ON(tcp_skb_pcount(skb) == 0);
3152                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3153         }
3154
3155         return packets_acked;
3156 }
3157
3158 /* Remove acknowledged frames from the retransmission queue. If our packet
3159  * is before the ack sequence we can discard it as it's confirmed to have
3160  * arrived at the other end.
3161  */
3162 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3163                                u32 prior_snd_una)
3164 {
3165         struct tcp_sock *tp = tcp_sk(sk);
3166         const struct inet_connection_sock *icsk = inet_csk(sk);
3167         struct sk_buff *skb;
3168         u32 now = tcp_time_stamp;
3169         int fully_acked = 1;
3170         int flag = 0;
3171         u32 pkts_acked = 0;
3172         u32 reord = tp->packets_out;
3173         u32 prior_sacked = tp->sacked_out;
3174         s32 seq_rtt = -1;
3175         s32 ca_seq_rtt = -1;
3176         ktime_t last_ackt = net_invalid_timestamp();
3177
3178         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3179                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3180                 u32 end_seq;
3181                 u32 acked_pcount;
3182                 u8 sacked = scb->sacked;
3183
3184                 /* Determine how many packets and what bytes were acked, tso and else */
3185                 if (after(scb->end_seq, tp->snd_una)) {
3186                         if (tcp_skb_pcount(skb) == 1 ||
3187                             !after(tp->snd_una, scb->seq))
3188                                 break;
3189
3190                         acked_pcount = tcp_tso_acked(sk, skb);
3191                         if (!acked_pcount)
3192                                 break;
3193
3194                         fully_acked = 0;
3195                         end_seq = tp->snd_una;
3196                 } else {
3197                         acked_pcount = tcp_skb_pcount(skb);
3198                         end_seq = scb->end_seq;
3199                 }
3200
3201                 /* MTU probing checks */
3202                 if (fully_acked && icsk->icsk_mtup.probe_size &&
3203                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
3204                         tcp_mtup_probe_success(sk, skb);
3205                 }
3206
3207                 if (sacked & TCPCB_RETRANS) {
3208                         if (sacked & TCPCB_SACKED_RETRANS)
3209                                 tp->retrans_out -= acked_pcount;
3210                         flag |= FLAG_RETRANS_DATA_ACKED;
3211                         ca_seq_rtt = -1;
3212                         seq_rtt = -1;
3213                         if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3214                                 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3215                 } else {
3216                         ca_seq_rtt = now - scb->when;
3217                         last_ackt = skb->tstamp;
3218                         if (seq_rtt < 0) {
3219                                 seq_rtt = ca_seq_rtt;
3220                         }
3221                         if (!(sacked & TCPCB_SACKED_ACKED))
3222                                 reord = min(pkts_acked, reord);
3223                 }
3224
3225                 if (sacked & TCPCB_SACKED_ACKED)
3226                         tp->sacked_out -= acked_pcount;
3227                 if (sacked & TCPCB_LOST)
3228                         tp->lost_out -= acked_pcount;
3229
3230                 tp->packets_out -= acked_pcount;
3231                 pkts_acked += acked_pcount;
3232
3233                 /* Initial outgoing SYN's get put onto the write_queue
3234                  * just like anything else we transmit.  It is not
3235                  * true data, and if we misinform our callers that
3236                  * this ACK acks real data, we will erroneously exit
3237                  * connection startup slow start one packet too
3238                  * quickly.  This is severely frowned upon behavior.
3239                  */
3240                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3241                         flag |= FLAG_DATA_ACKED;
3242                 } else {
3243                         flag |= FLAG_SYN_ACKED;
3244                         tp->retrans_stamp = 0;
3245                 }
3246
3247                 if (!fully_acked)
3248                         break;
3249
3250                 tcp_unlink_write_queue(skb, sk);
3251                 sk_wmem_free_skb(sk, skb);
3252                 tp->scoreboard_skb_hint = NULL;
3253                 if (skb == tp->retransmit_skb_hint)
3254                         tp->retransmit_skb_hint = NULL;
3255                 if (skb == tp->lost_skb_hint)
3256                         tp->lost_skb_hint = NULL;
3257         }
3258
3259         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3260                 tp->snd_up = tp->snd_una;
3261
3262         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3263                 flag |= FLAG_SACK_RENEGING;
3264
3265         if (flag & FLAG_ACKED) {
3266                 const struct tcp_congestion_ops *ca_ops
3267                         = inet_csk(sk)->icsk_ca_ops;
3268
3269                 tcp_ack_update_rtt(sk, flag, seq_rtt);
3270                 tcp_rearm_rto(sk);
3271
3272                 if (tcp_is_reno(tp)) {
3273                         tcp_remove_reno_sacks(sk, pkts_acked);
3274                 } else {
3275                         /* Non-retransmitted hole got filled? That's reordering */
3276                         if (reord < prior_fackets)
3277                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3278
3279                         /* No need to care for underflows here because
3280                          * the lost_skb_hint gets NULLed if we're past it
3281                          * (or something non-trivial happened)
3282                          */
3283                         if (tcp_is_fack(tp))
3284                                 tp->lost_cnt_hint -= pkts_acked;
3285                         else
3286                                 tp->lost_cnt_hint -= prior_sacked - tp->sacked_out;
3287                 }
3288
3289                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3290
3291                 if (ca_ops->pkts_acked) {
3292                         s32 rtt_us = -1;
3293
3294                         /* Is the ACK triggering packet unambiguous? */
3295                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3296                                 /* High resolution needed and available? */
3297                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3298                                     !ktime_equal(last_ackt,
3299                                                  net_invalid_timestamp()))
3300                                         rtt_us = ktime_us_delta(ktime_get_real(),
3301                                                                 last_ackt);
3302                                 else if (ca_seq_rtt > 0)
3303                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
3304                         }
3305
3306                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3307                 }
3308         }
3309
3310 #if FASTRETRANS_DEBUG > 0
3311         WARN_ON((int)tp->sacked_out < 0);
3312         WARN_ON((int)tp->lost_out < 0);
3313         WARN_ON((int)tp->retrans_out < 0);
3314         if (!tp->packets_out && tcp_is_sack(tp)) {
3315                 icsk = inet_csk(sk);
3316                 if (tp->lost_out) {
3317                         printk(KERN_DEBUG "Leak l=%u %d\n",
3318                                tp->lost_out, icsk->icsk_ca_state);
3319                         tp->lost_out = 0;
3320                 }
3321                 if (tp->sacked_out) {
3322                         printk(KERN_DEBUG "Leak s=%u %d\n",
3323                                tp->sacked_out, icsk->icsk_ca_state);
3324                         tp->sacked_out = 0;
3325                 }
3326                 if (tp->retrans_out) {
3327                         printk(KERN_DEBUG "Leak r=%u %d\n",
3328                                tp->retrans_out, icsk->icsk_ca_state);
3329                         tp->retrans_out = 0;
3330                 }
3331         }
3332 #endif
3333         return flag;
3334 }
3335
3336 static void tcp_ack_probe(struct sock *sk)
3337 {
3338         const struct tcp_sock *tp = tcp_sk(sk);
3339         struct inet_connection_sock *icsk = inet_csk(sk);
3340
3341         /* Was it a usable window open? */
3342
3343         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3344                 icsk->icsk_backoff = 0;
3345                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3346                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3347                  * This function is not for random using!
3348                  */
3349         } else {
3350                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3351                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3352                                           TCP_RTO_MAX);
3353         }
3354 }
3355
3356 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3357 {
3358         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3359                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3360 }
3361
3362 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3363 {
3364         const struct tcp_sock *tp = tcp_sk(sk);
3365         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3366                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3367 }
3368
3369 /* Check that window update is acceptable.
3370  * The function assumes that snd_una<=ack<=snd_next.
3371  */
3372 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3373                                         const u32 ack, const u32 ack_seq,
3374                                         const u32 nwin)
3375 {
3376         return (after(ack, tp->snd_una) ||
3377                 after(ack_seq, tp->snd_wl1) ||
3378                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3379 }
3380
3381 /* Update our send window.
3382  *
3383  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3384  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3385  */
3386 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3387                                  u32 ack_seq)
3388 {
3389         struct tcp_sock *tp = tcp_sk(sk);
3390         int flag = 0;
3391         u32 nwin = ntohs(tcp_hdr(skb)->window);
3392
3393         if (likely(!tcp_hdr(skb)->syn))
3394                 nwin <<= tp->rx_opt.snd_wscale;
3395
3396         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3397                 flag |= FLAG_WIN_UPDATE;
3398                 tcp_update_wl(tp, ack, ack_seq);
3399
3400                 if (tp->snd_wnd != nwin) {
3401                         tp->snd_wnd = nwin;
3402
3403                         /* Note, it is the only place, where
3404                          * fast path is recovered for sending TCP.
3405                          */
3406                         tp->pred_flags = 0;
3407                         tcp_fast_path_check(sk);
3408
3409                         if (nwin > tp->max_window) {
3410                                 tp->max_window = nwin;
3411                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3412                         }
3413                 }
3414         }
3415
3416         tp->snd_una = ack;
3417
3418         return flag;
3419 }
3420
3421 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3422  * continue in congestion avoidance.
3423  */
3424 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3425 {
3426         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3427         tp->snd_cwnd_cnt = 0;
3428         tp->bytes_acked = 0;
3429         TCP_ECN_queue_cwr(tp);
3430         tcp_moderate_cwnd(tp);
3431 }
3432
3433 /* A conservative spurious RTO response algorithm: reduce cwnd using
3434  * rate halving and continue in congestion avoidance.
3435  */
3436 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3437 {
3438         tcp_enter_cwr(sk, 0);
3439 }
3440
3441 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3442 {
3443         if (flag & FLAG_ECE)
3444                 tcp_ratehalving_spur_to_response(sk);
3445         else
3446                 tcp_undo_cwr(sk, 1);
3447 }
3448
3449 /* F-RTO spurious RTO detection algorithm (RFC4138)
3450  *
3451  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3452  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3453  * window (but not to or beyond highest sequence sent before RTO):
3454  *   On First ACK,  send two new segments out.
3455  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3456  *                  algorithm is not part of the F-RTO detection algorithm
3457  *                  given in RFC4138 but can be selected separately).
3458  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3459  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3460  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3461  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3462  *
3463  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3464  * original window even after we transmit two new data segments.
3465  *
3466  * SACK version:
3467  *   on first step, wait until first cumulative ACK arrives, then move to
3468  *   the second step. In second step, the next ACK decides.
3469  *
3470  * F-RTO is implemented (mainly) in four functions:
3471  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3472  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3473  *     called when tcp_use_frto() showed green light
3474  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3475  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3476  *     to prove that the RTO is indeed spurious. It transfers the control
3477  *     from F-RTO to the conventional RTO recovery
3478  */
3479 static int tcp_process_frto(struct sock *sk, int flag)
3480 {
3481         struct tcp_sock *tp = tcp_sk(sk);
3482
3483         tcp_verify_left_out(tp);
3484
3485         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3486         if (flag & FLAG_DATA_ACKED)
3487                 inet_csk(sk)->icsk_retransmits = 0;
3488
3489         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3490             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3491                 tp->undo_marker = 0;
3492
3493         if (!before(tp->snd_una, tp->frto_highmark)) {
3494                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3495                 return 1;
3496         }
3497
3498         if (!tcp_is_sackfrto(tp)) {
3499                 /* RFC4138 shortcoming in step 2; should also have case c):
3500                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3501                  * data, winupdate
3502                  */
3503                 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3504                         return 1;
3505
3506                 if (!(flag & FLAG_DATA_ACKED)) {
3507                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3508                                             flag);
3509                         return 1;
3510                 }
3511         } else {
3512                 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3513                         /* Prevent sending of new data. */
3514                         tp->snd_cwnd = min(tp->snd_cwnd,
3515                                            tcp_packets_in_flight(tp));
3516                         return 1;
3517                 }
3518
3519                 if ((tp->frto_counter >= 2) &&
3520                     (!(flag & FLAG_FORWARD_PROGRESS) ||
3521                      ((flag & FLAG_DATA_SACKED) &&
3522                       !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3523                         /* RFC4138 shortcoming (see comment above) */
3524                         if (!(flag & FLAG_FORWARD_PROGRESS) &&
3525                             (flag & FLAG_NOT_DUP))
3526                                 return 1;
3527
3528                         tcp_enter_frto_loss(sk, 3, flag);
3529                         return 1;
3530                 }
3531         }
3532
3533         if (tp->frto_counter == 1) {
3534                 /* tcp_may_send_now needs to see updated state */
3535                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3536                 tp->frto_counter = 2;
3537
3538                 if (!tcp_may_send_now(sk))
3539                         tcp_enter_frto_loss(sk, 2, flag);
3540
3541                 return 1;
3542         } else {
3543                 switch (sysctl_tcp_frto_response) {
3544                 case 2:
3545                         tcp_undo_spur_to_response(sk, flag);
3546                         break;
3547                 case 1:
3548                         tcp_conservative_spur_to_response(tp);
3549                         break;
3550                 default:
3551                         tcp_ratehalving_spur_to_response(sk);
3552                         break;
3553                 }
3554                 tp->frto_counter = 0;
3555                 tp->undo_marker = 0;
3556                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3557         }
3558         return 0;
3559 }
3560
3561 /* This routine deals with incoming acks, but not outgoing ones. */
3562 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3563 {
3564         struct inet_connection_sock *icsk = inet_csk(sk);
3565         struct tcp_sock *tp = tcp_sk(sk);
3566         u32 prior_snd_una = tp->snd_una;
3567         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3568         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3569         u32 prior_in_flight;
3570         u32 prior_fackets;
3571         int prior_packets;
3572         int frto_cwnd = 0;
3573
3574         /* If the ack is newer than sent or older than previous acks
3575          * then we can probably ignore it.
3576          */
3577         if (after(ack, tp->snd_nxt))
3578                 goto uninteresting_ack;
3579
3580         if (before(ack, prior_snd_una))
3581                 goto old_ack;
3582
3583         if (after(ack, prior_snd_una))
3584                 flag |= FLAG_SND_UNA_ADVANCED;
3585
3586         if (sysctl_tcp_abc) {
3587                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3588                         tp->bytes_acked += ack - prior_snd_una;
3589                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3590                         /* we assume just one segment left network */
3591                         tp->bytes_acked += min(ack - prior_snd_una,
3592                                                tp->mss_cache);
3593         }
3594
3595         prior_fackets = tp->fackets_out;
3596         prior_in_flight = tcp_packets_in_flight(tp);
3597
3598         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3599                 /* Window is constant, pure forward advance.
3600                  * No more checks are required.
3601                  * Note, we use the fact that SND.UNA>=SND.WL2.
3602                  */
3603                 tcp_update_wl(tp, ack, ack_seq);
3604                 tp->snd_una = ack;
3605                 flag |= FLAG_WIN_UPDATE;
3606
3607                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3608
3609                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3610         } else {
3611                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3612                         flag |= FLAG_DATA;
3613                 else
3614                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3615
3616                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3617
3618                 if (TCP_SKB_CB(skb)->sacked)
3619                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3620
3621                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3622                         flag |= FLAG_ECE;
3623
3624                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3625         }
3626
3627         /* We passed data and got it acked, remove any soft error
3628          * log. Something worked...
3629          */
3630         sk->sk_err_soft = 0;
3631         icsk->icsk_probes_out = 0;
3632         tp->rcv_tstamp = tcp_time_stamp;
3633         prior_packets = tp->packets_out;
3634         if (!prior_packets)
3635                 goto no_queue;
3636
3637         /* See if we can take anything off of the retransmit queue. */
3638         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3639
3640         if (tp->frto_counter)
3641                 frto_cwnd = tcp_process_frto(sk, flag);
3642         /* Guarantee sacktag reordering detection against wrap-arounds */
3643         if (before(tp->frto_highmark, tp->snd_una))
3644                 tp->frto_highmark = 0;
3645
3646         if (tcp_ack_is_dubious(sk, flag)) {
3647                 /* Advance CWND, if state allows this. */
3648                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3649                     tcp_may_raise_cwnd(sk, flag))
3650                         tcp_cong_avoid(sk, ack, prior_in_flight);
3651                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3652                                       flag);
3653         } else {
3654                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3655                         tcp_cong_avoid(sk, ack, prior_in_flight);
3656         }
3657
3658         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3659                 dst_confirm(sk->sk_dst_cache);
3660
3661         return 1;
3662
3663 no_queue:
3664         /* If this ack opens up a zero window, clear backoff.  It was
3665          * being used to time the probes, and is probably far higher than
3666          * it needs to be for normal retransmission.
3667          */
3668         if (tcp_send_head(sk))
3669                 tcp_ack_probe(sk);
3670         return 1;
3671
3672 old_ack:
3673         if (TCP_SKB_CB(skb)->sacked) {
3674                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3675                 if (icsk->icsk_ca_state == TCP_CA_Open)
3676                         tcp_try_keep_open(sk);
3677         }
3678
3679 uninteresting_ack:
3680         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3681         return 0;
3682 }
3683
3684 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3685  * But, this can also be called on packets in the established flow when
3686  * the fast version below fails.
3687  */
3688 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3689                        int estab)
3690 {
3691         unsigned char *ptr;
3692         struct tcphdr *th = tcp_hdr(skb);
3693         int length = (th->doff * 4) - sizeof(struct tcphdr);
3694
3695         ptr = (unsigned char *)(th + 1);
3696         opt_rx->saw_tstamp = 0;
3697
3698         while (length > 0) {
3699                 int opcode = *ptr++;
3700                 int opsize;
3701
3702                 switch (opcode) {
3703                 case TCPOPT_EOL:
3704                         return;
3705                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3706                         length--;
3707                         continue;
3708                 default:
3709                         opsize = *ptr++;
3710                         if (opsize < 2) /* "silly options" */
3711                                 return;
3712                         if (opsize > length)
3713                                 return; /* don't parse partial options */
3714                         switch (opcode) {
3715                         case TCPOPT_MSS:
3716                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3717                                         u16 in_mss = get_unaligned_be16(ptr);
3718                                         if (in_mss) {
3719                                                 if (opt_rx->user_mss &&
3720                                                     opt_rx->user_mss < in_mss)
3721                                                         in_mss = opt_rx->user_mss;
3722                                                 opt_rx->mss_clamp = in_mss;
3723                                         }
3724                                 }
3725                                 break;
3726                         case TCPOPT_WINDOW:
3727                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3728                                     !estab && sysctl_tcp_window_scaling) {
3729                                         __u8 snd_wscale = *(__u8 *)ptr;
3730                                         opt_rx->wscale_ok = 1;
3731                                         if (snd_wscale > 14) {
3732                                                 if (net_ratelimit())
3733                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3734                                                                "scaling value %d >14 received.\n",
3735                                                                snd_wscale);
3736                                                 snd_wscale = 14;
3737                                         }
3738                                         opt_rx->snd_wscale = snd_wscale;
3739                                 }
3740                                 break;
3741                         case TCPOPT_TIMESTAMP:
3742                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3743                                     ((estab && opt_rx->tstamp_ok) ||
3744                                      (!estab && sysctl_tcp_timestamps))) {
3745                                         opt_rx->saw_tstamp = 1;
3746                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3747                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3748                                 }
3749                                 break;
3750                         case TCPOPT_SACK_PERM:
3751                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3752                                     !estab && sysctl_tcp_sack) {
3753                                         opt_rx->sack_ok = 1;
3754                                         tcp_sack_reset(opt_rx);
3755                                 }
3756                                 break;
3757
3758                         case TCPOPT_SACK:
3759                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3760                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3761                                    opt_rx->sack_ok) {
3762                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3763                                 }
3764                                 break;
3765 #ifdef CONFIG_TCP_MD5SIG
3766                         case TCPOPT_MD5SIG:
3767                                 /*
3768                                  * The MD5 Hash has already been
3769                                  * checked (see tcp_v{4,6}_do_rcv()).
3770                                  */
3771                                 break;
3772 #endif
3773                         }
3774
3775                         ptr += opsize-2;
3776                         length -= opsize;
3777                 }
3778         }
3779 }
3780
3781 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3782 {
3783         __be32 *ptr = (__be32 *)(th + 1);
3784
3785         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3786                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3787                 tp->rx_opt.saw_tstamp = 1;
3788                 ++ptr;
3789                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3790                 ++ptr;
3791                 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3792                 return 1;
3793         }
3794         return 0;
3795 }
3796
3797 /* Fast parse options. This hopes to only see timestamps.
3798  * If it is wrong it falls back on tcp_parse_options().
3799  */
3800 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3801                                   struct tcp_sock *tp)
3802 {
3803         if (th->doff == sizeof(struct tcphdr) >> 2) {
3804                 tp->rx_opt.saw_tstamp = 0;
3805                 return 0;
3806         } else if (tp->rx_opt.tstamp_ok &&
3807                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3808                 if (tcp_parse_aligned_timestamp(tp, th))
3809                         return 1;
3810         }
3811         tcp_parse_options(skb, &tp->rx_opt, 1);
3812         return 1;
3813 }
3814
3815 #ifdef CONFIG_TCP_MD5SIG
3816 /*
3817  * Parse MD5 Signature option
3818  */
3819 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3820 {
3821         int length = (th->doff << 2) - sizeof (*th);
3822         u8 *ptr = (u8*)(th + 1);
3823
3824         /* If the TCP option is too short, we can short cut */
3825         if (length < TCPOLEN_MD5SIG)
3826                 return NULL;
3827
3828         while (length > 0) {
3829                 int opcode = *ptr++;
3830                 int opsize;
3831
3832                 switch(opcode) {
3833                 case TCPOPT_EOL:
3834                         return NULL;
3835                 case TCPOPT_NOP:
3836                         length--;
3837                         continue;
3838                 default:
3839                         opsize = *ptr++;
3840                         if (opsize < 2 || opsize > length)
3841                                 return NULL;
3842                         if (opcode == TCPOPT_MD5SIG)
3843                                 return ptr;
3844                 }
3845                 ptr += opsize - 2;
3846                 length -= opsize;
3847         }
3848         return NULL;
3849 }
3850 #endif
3851
3852 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3853 {
3854         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3855         tp->rx_opt.ts_recent_stamp = get_seconds();
3856 }
3857
3858 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3859 {
3860         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3861                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3862                  * extra check below makes sure this can only happen
3863                  * for pure ACK frames.  -DaveM
3864                  *
3865                  * Not only, also it occurs for expired timestamps.
3866                  */
3867
3868                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3869                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3870                         tcp_store_ts_recent(tp);
3871         }
3872 }
3873
3874 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3875  *
3876  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3877  * it can pass through stack. So, the following predicate verifies that
3878  * this segment is not used for anything but congestion avoidance or
3879  * fast retransmit. Moreover, we even are able to eliminate most of such
3880  * second order effects, if we apply some small "replay" window (~RTO)
3881  * to timestamp space.
3882  *
3883  * All these measures still do not guarantee that we reject wrapped ACKs
3884  * on networks with high bandwidth, when sequence space is recycled fastly,
3885  * but it guarantees that such events will be very rare and do not affect
3886  * connection seriously. This doesn't look nice, but alas, PAWS is really
3887  * buggy extension.
3888  *
3889  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3890  * states that events when retransmit arrives after original data are rare.
3891  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3892  * the biggest problem on large power networks even with minor reordering.
3893  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3894  * up to bandwidth of 18Gigabit/sec. 8) ]
3895  */
3896
3897 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3898 {
3899         struct tcp_sock *tp = tcp_sk(sk);
3900         struct tcphdr *th = tcp_hdr(skb);
3901         u32 seq = TCP_SKB_CB(skb)->seq;
3902         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3903
3904         return (/* 1. Pure ACK with correct sequence number. */
3905                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3906
3907                 /* 2. ... and duplicate ACK. */
3908                 ack == tp->snd_una &&
3909
3910                 /* 3. ... and does not update window. */
3911                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3912
3913                 /* 4. ... and sits in replay window. */
3914                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3915 }
3916
3917 static inline int tcp_paws_discard(const struct sock *sk,
3918                                    const struct sk_buff *skb)
3919 {
3920         const struct tcp_sock *tp = tcp_sk(sk);
3921         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3922                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3923                 !tcp_disordered_ack(sk, skb));
3924 }
3925
3926 /* Check segment sequence number for validity.
3927  *
3928  * Segment controls are considered valid, if the segment
3929  * fits to the window after truncation to the window. Acceptability
3930  * of data (and SYN, FIN, of course) is checked separately.
3931  * See tcp_data_queue(), for example.
3932  *
3933  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3934  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3935  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3936  * (borrowed from freebsd)
3937  */
3938
3939 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3940 {
3941         return  !before(end_seq, tp->rcv_wup) &&
3942                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3943 }
3944
3945 /* When we get a reset we do this. */
3946 static void tcp_reset(struct sock *sk)
3947 {
3948         /* We want the right error as BSD sees it (and indeed as we do). */
3949         switch (sk->sk_state) {
3950         case TCP_SYN_SENT:
3951                 sk->sk_err = ECONNREFUSED;
3952                 break;
3953         case TCP_CLOSE_WAIT:
3954                 sk->sk_err = EPIPE;
3955                 break;
3956         case TCP_CLOSE:
3957                 return;
3958         default:
3959                 sk->sk_err = ECONNRESET;
3960         }
3961
3962         if (!sock_flag(sk, SOCK_DEAD))
3963                 sk->sk_error_report(sk);
3964
3965         tcp_done(sk);
3966 }
3967
3968 /*
3969  *      Process the FIN bit. This now behaves as it is supposed to work
3970  *      and the FIN takes effect when it is validly part of sequence
3971  *      space. Not before when we get holes.
3972  *
3973  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3974  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3975  *      TIME-WAIT)
3976  *
3977  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3978  *      close and we go into CLOSING (and later onto TIME-WAIT)
3979  *
3980  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3981  */
3982 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3983 {
3984         struct tcp_sock *tp = tcp_sk(sk);
3985
3986         inet_csk_schedule_ack(sk);
3987
3988         sk->sk_shutdown |= RCV_SHUTDOWN;
3989         sock_set_flag(sk, SOCK_DONE);
3990
3991         switch (sk->sk_state) {
3992         case TCP_SYN_RECV:
3993         case TCP_ESTABLISHED:
3994                 /* Move to CLOSE_WAIT */
3995                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3996                 inet_csk(sk)->icsk_ack.pingpong = 1;
3997                 break;
3998
3999         case TCP_CLOSE_WAIT:
4000         case TCP_CLOSING:
4001                 /* Received a retransmission of the FIN, do
4002                  * nothing.
4003                  */
4004                 break;
4005         case TCP_LAST_ACK:
4006                 /* RFC793: Remain in the LAST-ACK state. */
4007                 break;
4008
4009         case TCP_FIN_WAIT1:
4010                 /* This case occurs when a simultaneous close
4011                  * happens, we must ack the received FIN and
4012                  * enter the CLOSING state.
4013                  */
4014                 tcp_send_ack(sk);
4015                 tcp_set_state(sk, TCP_CLOSING);
4016                 break;
4017         case TCP_FIN_WAIT2:
4018                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4019                 tcp_send_ack(sk);
4020                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4021                 break;
4022         default:
4023                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4024                  * cases we should never reach this piece of code.
4025                  */
4026                 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4027                        __func__, sk->sk_state);
4028                 break;
4029         }
4030
4031         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4032          * Probably, we should reset in this case. For now drop them.
4033          */
4034         __skb_queue_purge(&tp->out_of_order_queue);
4035         if (tcp_is_sack(tp))
4036                 tcp_sack_reset(&tp->rx_opt);
4037         sk_mem_reclaim(sk);
4038
4039         if (!sock_flag(sk, SOCK_DEAD)) {
4040                 sk->sk_state_change(sk);
4041
4042                 /* Do not send POLL_HUP for half duplex close. */
4043                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4044                     sk->sk_state == TCP_CLOSE)
4045                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4046                 else
4047                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4048         }
4049 }
4050
4051 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4052                                   u32 end_seq)
4053 {
4054         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4055                 if (before(seq, sp->start_seq))
4056                         sp->start_seq = seq;
4057                 if (after(end_seq, sp->end_seq))
4058                         sp->end_seq = end_seq;
4059                 return 1;
4060         }
4061         return 0;
4062 }
4063
4064 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4065 {
4066         struct tcp_sock *tp = tcp_sk(sk);
4067
4068         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4069                 int mib_idx;
4070
4071                 if (before(seq, tp->rcv_nxt))
4072                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4073                 else
4074                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4075
4076                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4077
4078                 tp->rx_opt.dsack = 1;
4079                 tp->duplicate_sack[0].start_seq = seq;
4080                 tp->duplicate_sack[0].end_seq = end_seq;
4081                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1;
4082         }
4083 }
4084
4085 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4086 {
4087         struct tcp_sock *tp = tcp_sk(sk);
4088
4089         if (!tp->rx_opt.dsack)
4090                 tcp_dsack_set(sk, seq, end_seq);
4091         else
4092                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4093 }
4094
4095 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4096 {
4097         struct tcp_sock *tp = tcp_sk(sk);
4098
4099         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4100             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4101                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4102                 tcp_enter_quickack_mode(sk);
4103
4104                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4105                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4106
4107                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4108                                 end_seq = tp->rcv_nxt;
4109                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4110                 }
4111         }
4112
4113         tcp_send_ack(sk);
4114 }
4115
4116 /* These routines update the SACK block as out-of-order packets arrive or
4117  * in-order packets close up the sequence space.
4118  */
4119 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4120 {
4121         int this_sack;
4122         struct tcp_sack_block *sp = &tp->selective_acks[0];
4123         struct tcp_sack_block *swalk = sp + 1;
4124
4125         /* See if the recent change to the first SACK eats into
4126          * or hits the sequence space of other SACK blocks, if so coalesce.
4127          */
4128         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4129                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4130                         int i;
4131
4132                         /* Zap SWALK, by moving every further SACK up by one slot.
4133                          * Decrease num_sacks.
4134                          */
4135                         tp->rx_opt.num_sacks--;
4136                         tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
4137                                                tp->rx_opt.dsack;
4138                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4139                                 sp[i] = sp[i + 1];
4140                         continue;
4141                 }
4142                 this_sack++, swalk++;
4143         }
4144 }
4145
4146 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
4147                                  struct tcp_sack_block *sack2)
4148 {
4149         __u32 tmp;
4150
4151         tmp = sack1->start_seq;
4152         sack1->start_seq = sack2->start_seq;
4153         sack2->start_seq = tmp;
4154
4155         tmp = sack1->end_seq;
4156         sack1->end_seq = sack2->end_seq;
4157         sack2->end_seq = tmp;
4158 }
4159
4160 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4161 {
4162         struct tcp_sock *tp = tcp_sk(sk);
4163         struct tcp_sack_block *sp = &tp->selective_acks[0];
4164         int cur_sacks = tp->rx_opt.num_sacks;
4165         int this_sack;
4166
4167         if (!cur_sacks)
4168                 goto new_sack;
4169
4170         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4171                 if (tcp_sack_extend(sp, seq, end_seq)) {
4172                         /* Rotate this_sack to the first one. */
4173                         for (; this_sack > 0; this_sack--, sp--)
4174                                 tcp_sack_swap(sp, sp - 1);
4175                         if (cur_sacks > 1)
4176                                 tcp_sack_maybe_coalesce(tp);
4177                         return;
4178                 }
4179         }
4180
4181         /* Could not find an adjacent existing SACK, build a new one,
4182          * put it at the front, and shift everyone else down.  We
4183          * always know there is at least one SACK present already here.
4184          *
4185          * If the sack array is full, forget about the last one.
4186          */
4187         if (this_sack >= TCP_NUM_SACKS) {
4188                 this_sack--;
4189                 tp->rx_opt.num_sacks--;
4190                 sp--;
4191         }
4192         for (; this_sack > 0; this_sack--, sp--)
4193                 *sp = *(sp - 1);
4194
4195 new_sack:
4196         /* Build the new head SACK, and we're done. */
4197         sp->start_seq = seq;
4198         sp->end_seq = end_seq;
4199         tp->rx_opt.num_sacks++;
4200         tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
4201 }
4202
4203 /* RCV.NXT advances, some SACKs should be eaten. */
4204
4205 static void tcp_sack_remove(struct tcp_sock *tp)
4206 {
4207         struct tcp_sack_block *sp = &tp->selective_acks[0];
4208         int num_sacks = tp->rx_opt.num_sacks;
4209         int this_sack;
4210
4211         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4212         if (skb_queue_empty(&tp->out_of_order_queue)) {
4213                 tp->rx_opt.num_sacks = 0;
4214                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
4215                 return;
4216         }
4217
4218         for (this_sack = 0; this_sack < num_sacks;) {
4219                 /* Check if the start of the sack is covered by RCV.NXT. */
4220                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4221                         int i;
4222
4223                         /* RCV.NXT must cover all the block! */
4224                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4225
4226                         /* Zap this SACK, by moving forward any other SACKS. */
4227                         for (i=this_sack+1; i < num_sacks; i++)
4228                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4229                         num_sacks--;
4230                         continue;
4231                 }
4232                 this_sack++;
4233                 sp++;
4234         }
4235         if (num_sacks != tp->rx_opt.num_sacks) {
4236                 tp->rx_opt.num_sacks = num_sacks;
4237                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
4238                                        tp->rx_opt.dsack;
4239         }
4240 }
4241
4242 /* This one checks to see if we can put data from the
4243  * out_of_order queue into the receive_queue.
4244  */
4245 static void tcp_ofo_queue(struct sock *sk)
4246 {
4247         struct tcp_sock *tp = tcp_sk(sk);
4248         __u32 dsack_high = tp->rcv_nxt;
4249         struct sk_buff *skb;
4250
4251         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4252                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4253                         break;
4254
4255                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4256                         __u32 dsack = dsack_high;
4257                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4258                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4259                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4260                 }
4261
4262                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4263                         SOCK_DEBUG(sk, "ofo packet was already received \n");
4264                         __skb_unlink(skb, &tp->out_of_order_queue);
4265                         __kfree_skb(skb);
4266                         continue;
4267                 }
4268                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4269                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4270                            TCP_SKB_CB(skb)->end_seq);
4271
4272                 __skb_unlink(skb, &tp->out_of_order_queue);
4273                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4274                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4275                 if (tcp_hdr(skb)->fin)
4276                         tcp_fin(skb, sk, tcp_hdr(skb));
4277         }
4278 }
4279
4280 static int tcp_prune_ofo_queue(struct sock *sk);
4281 static int tcp_prune_queue(struct sock *sk);
4282
4283 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4284 {
4285         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4286             !sk_rmem_schedule(sk, size)) {
4287
4288                 if (tcp_prune_queue(sk) < 0)
4289                         return -1;
4290
4291                 if (!sk_rmem_schedule(sk, size)) {
4292                         if (!tcp_prune_ofo_queue(sk))
4293                                 return -1;
4294
4295                         if (!sk_rmem_schedule(sk, size))
4296                                 return -1;
4297                 }
4298         }
4299         return 0;
4300 }
4301
4302 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4303 {
4304         struct tcphdr *th = tcp_hdr(skb);
4305         struct tcp_sock *tp = tcp_sk(sk);
4306         int eaten = -1;
4307
4308         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4309                 goto drop;
4310
4311         __skb_pull(skb, th->doff * 4);
4312
4313         TCP_ECN_accept_cwr(tp, skb);
4314
4315         if (tp->rx_opt.dsack) {
4316                 tp->rx_opt.dsack = 0;
4317                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
4318         }
4319
4320         /*  Queue data for delivery to the user.
4321          *  Packets in sequence go to the receive queue.
4322          *  Out of sequence packets to the out_of_order_queue.
4323          */
4324         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4325                 if (tcp_receive_window(tp) == 0)
4326                         goto out_of_window;
4327
4328                 /* Ok. In sequence. In window. */
4329                 if (tp->ucopy.task == current &&
4330                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4331                     sock_owned_by_user(sk) && !tp->urg_data) {
4332                         int chunk = min_t(unsigned int, skb->len,
4333                                           tp->ucopy.len);
4334
4335                         __set_current_state(TASK_RUNNING);
4336
4337                         local_bh_enable();
4338                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4339                                 tp->ucopy.len -= chunk;
4340                                 tp->copied_seq += chunk;
4341                                 eaten = (chunk == skb->len && !th->fin);
4342                                 tcp_rcv_space_adjust(sk);
4343                         }
4344                         local_bh_disable();
4345                 }
4346
4347                 if (eaten <= 0) {
4348 queue_and_out:
4349                         if (eaten < 0 &&
4350                             tcp_try_rmem_schedule(sk, skb->truesize))
4351                                 goto drop;
4352
4353                         skb_set_owner_r(skb, sk);
4354                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4355                 }
4356                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4357                 if (skb->len)
4358                         tcp_event_data_recv(sk, skb);
4359                 if (th->fin)
4360                         tcp_fin(skb, sk, th);
4361
4362                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4363                         tcp_ofo_queue(sk);
4364
4365                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4366                          * gap in queue is filled.
4367                          */
4368                         if (skb_queue_empty(&tp->out_of_order_queue))
4369                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4370                 }
4371
4372                 if (tp->rx_opt.num_sacks)
4373                         tcp_sack_remove(tp);
4374
4375                 tcp_fast_path_check(sk);
4376
4377                 if (eaten > 0)
4378                         __kfree_skb(skb);
4379                 else if (!sock_flag(sk, SOCK_DEAD))
4380                         sk->sk_data_ready(sk, 0);
4381                 return;
4382         }
4383
4384         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4385                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4386                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4387                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4388
4389 out_of_window:
4390                 tcp_enter_quickack_mode(sk);
4391                 inet_csk_schedule_ack(sk);
4392 drop:
4393                 __kfree_skb(skb);
4394                 return;
4395         }
4396
4397         /* Out of window. F.e. zero window probe. */
4398         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4399                 goto out_of_window;
4400
4401         tcp_enter_quickack_mode(sk);
4402
4403         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4404                 /* Partial packet, seq < rcv_next < end_seq */
4405                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4406                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4407                            TCP_SKB_CB(skb)->end_seq);
4408
4409                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4410
4411                 /* If window is closed, drop tail of packet. But after
4412                  * remembering D-SACK for its head made in previous line.
4413                  */
4414                 if (!tcp_receive_window(tp))
4415                         goto out_of_window;
4416                 goto queue_and_out;
4417         }
4418
4419         TCP_ECN_check_ce(tp, skb);
4420
4421         if (tcp_try_rmem_schedule(sk, skb->truesize))
4422                 goto drop;
4423
4424         /* Disable header prediction. */
4425         tp->pred_flags = 0;
4426         inet_csk_schedule_ack(sk);
4427
4428         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4429                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4430
4431         skb_set_owner_r(skb, sk);
4432
4433         if (!skb_peek(&tp->out_of_order_queue)) {
4434                 /* Initial out of order segment, build 1 SACK. */
4435                 if (tcp_is_sack(tp)) {
4436                         tp->rx_opt.num_sacks = 1;
4437                         tp->rx_opt.dsack     = 0;
4438                         tp->rx_opt.eff_sacks = 1;
4439                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4440                         tp->selective_acks[0].end_seq =
4441                                                 TCP_SKB_CB(skb)->end_seq;
4442                 }
4443                 __skb_queue_head(&tp->out_of_order_queue, skb);
4444         } else {
4445                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4446                 u32 seq = TCP_SKB_CB(skb)->seq;
4447                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4448
4449                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4450                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4451
4452                         if (!tp->rx_opt.num_sacks ||
4453                             tp->selective_acks[0].end_seq != seq)
4454                                 goto add_sack;
4455
4456                         /* Common case: data arrive in order after hole. */
4457                         tp->selective_acks[0].end_seq = end_seq;
4458                         return;
4459                 }
4460
4461                 /* Find place to insert this segment. */
4462                 do {
4463                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
4464                                 break;
4465                 } while ((skb1 = skb1->prev) !=
4466                          (struct sk_buff *)&tp->out_of_order_queue);
4467
4468                 /* Do skb overlap to previous one? */
4469                 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4470                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4471                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4472                                 /* All the bits are present. Drop. */
4473                                 __kfree_skb(skb);
4474                                 tcp_dsack_set(sk, seq, end_seq);
4475                                 goto add_sack;
4476                         }
4477                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4478                                 /* Partial overlap. */
4479                                 tcp_dsack_set(sk, seq,
4480                                               TCP_SKB_CB(skb1)->end_seq);
4481                         } else {
4482                                 skb1 = skb1->prev;
4483                         }
4484                 }
4485                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4486
4487                 /* And clean segments covered by new one as whole. */
4488                 while ((skb1 = skb->next) !=
4489                        (struct sk_buff *)&tp->out_of_order_queue &&
4490                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4491                         if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4492                                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4493                                                  end_seq);
4494                                 break;
4495                         }
4496                         __skb_unlink(skb1, &tp->out_of_order_queue);
4497                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4498                                          TCP_SKB_CB(skb1)->end_seq);
4499                         __kfree_skb(skb1);
4500                 }
4501
4502 add_sack:
4503                 if (tcp_is_sack(tp))
4504                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4505         }
4506 }
4507
4508 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4509                                         struct sk_buff_head *list)
4510 {
4511         struct sk_buff *next = skb->next;
4512
4513         __skb_unlink(skb, list);
4514         __kfree_skb(skb);
4515         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4516
4517         return next;
4518 }
4519
4520 /* Collapse contiguous sequence of skbs head..tail with
4521  * sequence numbers start..end.
4522  * Segments with FIN/SYN are not collapsed (only because this
4523  * simplifies code)
4524  */
4525 static void
4526 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4527              struct sk_buff *head, struct sk_buff *tail,
4528              u32 start, u32 end)
4529 {
4530         struct sk_buff *skb;
4531
4532         /* First, check that queue is collapsible and find
4533          * the point where collapsing can be useful. */
4534         for (skb = head; skb != tail;) {
4535                 /* No new bits? It is possible on ofo queue. */
4536                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4537                         skb = tcp_collapse_one(sk, skb, list);
4538                         continue;
4539                 }
4540
4541                 /* The first skb to collapse is:
4542                  * - not SYN/FIN and
4543                  * - bloated or contains data before "start" or
4544                  *   overlaps to the next one.
4545                  */
4546                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4547                     (tcp_win_from_space(skb->truesize) > skb->len ||
4548                      before(TCP_SKB_CB(skb)->seq, start) ||
4549                      (skb->next != tail &&
4550                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4551                         break;
4552
4553                 /* Decided to skip this, advance start seq. */
4554                 start = TCP_SKB_CB(skb)->end_seq;
4555                 skb = skb->next;
4556         }
4557         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4558                 return;
4559
4560         while (before(start, end)) {
4561                 struct sk_buff *nskb;
4562                 unsigned int header = skb_headroom(skb);
4563                 int copy = SKB_MAX_ORDER(header, 0);
4564
4565                 /* Too big header? This can happen with IPv6. */
4566                 if (copy < 0)
4567                         return;
4568                 if (end - start < copy)
4569                         copy = end - start;
4570                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4571                 if (!nskb)
4572                         return;
4573
4574                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4575                 skb_set_network_header(nskb, (skb_network_header(skb) -
4576                                               skb->head));
4577                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4578                                                 skb->head));
4579                 skb_reserve(nskb, header);
4580                 memcpy(nskb->head, skb->head, header);
4581                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4582                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4583                 __skb_queue_before(list, skb, nskb);
4584                 skb_set_owner_r(nskb, sk);
4585
4586                 /* Copy data, releasing collapsed skbs. */
4587                 while (copy > 0) {
4588                         int offset = start - TCP_SKB_CB(skb)->seq;
4589                         int size = TCP_SKB_CB(skb)->end_seq - start;
4590
4591                         BUG_ON(offset < 0);
4592                         if (size > 0) {
4593                                 size = min(copy, size);
4594                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4595                                         BUG();
4596                                 TCP_SKB_CB(nskb)->end_seq += size;
4597                                 copy -= size;
4598                                 start += size;
4599                         }
4600                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4601                                 skb = tcp_collapse_one(sk, skb, list);
4602                                 if (skb == tail ||
4603                                     tcp_hdr(skb)->syn ||
4604                                     tcp_hdr(skb)->fin)
4605                                         return;
4606                         }
4607                 }
4608         }
4609 }
4610
4611 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4612  * and tcp_collapse() them until all the queue is collapsed.
4613  */
4614 static void tcp_collapse_ofo_queue(struct sock *sk)
4615 {
4616         struct tcp_sock *tp = tcp_sk(sk);
4617         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4618         struct sk_buff *head;
4619         u32 start, end;
4620
4621         if (skb == NULL)
4622                 return;
4623
4624         start = TCP_SKB_CB(skb)->seq;
4625         end = TCP_SKB_CB(skb)->end_seq;
4626         head = skb;
4627
4628         for (;;) {
4629                 skb = skb->next;
4630
4631                 /* Segment is terminated when we see gap or when
4632                  * we are at the end of all the queue. */
4633                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4634                     after(TCP_SKB_CB(skb)->seq, end) ||
4635                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4636                         tcp_collapse(sk, &tp->out_of_order_queue,
4637                                      head, skb, start, end);
4638                         head = skb;
4639                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4640                                 break;
4641                         /* Start new segment */
4642                         start = TCP_SKB_CB(skb)->seq;
4643                         end = TCP_SKB_CB(skb)->end_seq;
4644                 } else {
4645                         if (before(TCP_SKB_CB(skb)->seq, start))
4646                                 start = TCP_SKB_CB(skb)->seq;
4647                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4648                                 end = TCP_SKB_CB(skb)->end_seq;
4649                 }
4650         }
4651 }
4652
4653 /*
4654  * Purge the out-of-order queue.
4655  * Return true if queue was pruned.
4656  */
4657 static int tcp_prune_ofo_queue(struct sock *sk)
4658 {
4659         struct tcp_sock *tp = tcp_sk(sk);
4660         int res = 0;
4661
4662         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4663                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4664                 __skb_queue_purge(&tp->out_of_order_queue);
4665
4666                 /* Reset SACK state.  A conforming SACK implementation will
4667                  * do the same at a timeout based retransmit.  When a connection
4668                  * is in a sad state like this, we care only about integrity
4669                  * of the connection not performance.
4670                  */
4671                 if (tp->rx_opt.sack_ok)
4672                         tcp_sack_reset(&tp->rx_opt);
4673                 sk_mem_reclaim(sk);
4674                 res = 1;
4675         }
4676         return res;
4677 }
4678
4679 /* Reduce allocated memory if we can, trying to get
4680  * the socket within its memory limits again.
4681  *
4682  * Return less than zero if we should start dropping frames
4683  * until the socket owning process reads some of the data
4684  * to stabilize the situation.
4685  */
4686 static int tcp_prune_queue(struct sock *sk)
4687 {
4688         struct tcp_sock *tp = tcp_sk(sk);
4689
4690         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4691
4692         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4693
4694         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4695                 tcp_clamp_window(sk);
4696         else if (tcp_memory_pressure)
4697                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4698
4699         tcp_collapse_ofo_queue(sk);
4700         tcp_collapse(sk, &sk->sk_receive_queue,
4701                      sk->sk_receive_queue.next,
4702                      (struct sk_buff *)&sk->sk_receive_queue,
4703                      tp->copied_seq, tp->rcv_nxt);
4704         sk_mem_reclaim(sk);
4705
4706         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4707                 return 0;
4708
4709         /* Collapsing did not help, destructive actions follow.
4710          * This must not ever occur. */
4711
4712         tcp_prune_ofo_queue(sk);
4713
4714         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4715                 return 0;
4716
4717         /* If we are really being abused, tell the caller to silently
4718          * drop receive data on the floor.  It will get retransmitted
4719          * and hopefully then we'll have sufficient space.
4720          */
4721         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4722
4723         /* Massive buffer overcommit. */
4724         tp->pred_flags = 0;
4725         return -1;
4726 }
4727
4728 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4729  * As additional protections, we do not touch cwnd in retransmission phases,
4730  * and if application hit its sndbuf limit recently.
4731  */
4732 void tcp_cwnd_application_limited(struct sock *sk)
4733 {
4734         struct tcp_sock *tp = tcp_sk(sk);
4735
4736         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4737             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4738                 /* Limited by application or receiver window. */
4739                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4740                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4741                 if (win_used < tp->snd_cwnd) {
4742                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4743                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4744                 }
4745                 tp->snd_cwnd_used = 0;
4746         }
4747         tp->snd_cwnd_stamp = tcp_time_stamp;
4748 }
4749
4750 static int tcp_should_expand_sndbuf(struct sock *sk)
4751 {
4752         struct tcp_sock *tp = tcp_sk(sk);
4753
4754         /* If the user specified a specific send buffer setting, do
4755          * not modify it.
4756          */
4757         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4758                 return 0;
4759
4760         /* If we are under global TCP memory pressure, do not expand.  */
4761         if (tcp_memory_pressure)
4762                 return 0;
4763
4764         /* If we are under soft global TCP memory pressure, do not expand.  */
4765         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4766                 return 0;
4767
4768         /* If we filled the congestion window, do not expand.  */
4769         if (tp->packets_out >= tp->snd_cwnd)
4770                 return 0;
4771
4772         return 1;
4773 }
4774
4775 /* When incoming ACK allowed to free some skb from write_queue,
4776  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4777  * on the exit from tcp input handler.
4778  *
4779  * PROBLEM: sndbuf expansion does not work well with largesend.
4780  */
4781 static void tcp_new_space(struct sock *sk)
4782 {
4783         struct tcp_sock *tp = tcp_sk(sk);
4784
4785         if (tcp_should_expand_sndbuf(sk)) {
4786                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4787                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4788                 int demanded = max_t(unsigned int, tp->snd_cwnd,
4789                                      tp->reordering + 1);
4790                 sndmem *= 2 * demanded;
4791                 if (sndmem > sk->sk_sndbuf)
4792                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4793                 tp->snd_cwnd_stamp = tcp_time_stamp;
4794         }
4795
4796         sk->sk_write_space(sk);
4797 }
4798
4799 static void tcp_check_space(struct sock *sk)
4800 {
4801         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4802                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4803                 if (sk->sk_socket &&
4804                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4805                         tcp_new_space(sk);
4806         }
4807 }
4808
4809 static inline void tcp_data_snd_check(struct sock *sk)
4810 {
4811         tcp_push_pending_frames(sk);
4812         tcp_check_space(sk);
4813 }
4814
4815 /*
4816  * Check if sending an ack is needed.
4817  */
4818 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4819 {
4820         struct tcp_sock *tp = tcp_sk(sk);
4821
4822             /* More than one full frame received... */
4823         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4824              /* ... and right edge of window advances far enough.
4825               * (tcp_recvmsg() will send ACK otherwise). Or...
4826               */
4827              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4828             /* We ACK each frame or... */
4829             tcp_in_quickack_mode(sk) ||
4830             /* We have out of order data. */
4831             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4832                 /* Then ack it now */
4833                 tcp_send_ack(sk);
4834         } else {
4835                 /* Else, send delayed ack. */
4836                 tcp_send_delayed_ack(sk);
4837         }
4838 }
4839
4840 static inline void tcp_ack_snd_check(struct sock *sk)
4841 {
4842         if (!inet_csk_ack_scheduled(sk)) {
4843                 /* We sent a data segment already. */
4844                 return;
4845         }
4846         __tcp_ack_snd_check(sk, 1);
4847 }
4848
4849 /*
4850  *      This routine is only called when we have urgent data
4851  *      signaled. Its the 'slow' part of tcp_urg. It could be
4852  *      moved inline now as tcp_urg is only called from one
4853  *      place. We handle URGent data wrong. We have to - as
4854  *      BSD still doesn't use the correction from RFC961.
4855  *      For 1003.1g we should support a new option TCP_STDURG to permit
4856  *      either form (or just set the sysctl tcp_stdurg).
4857  */
4858
4859 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4860 {
4861         struct tcp_sock *tp = tcp_sk(sk);
4862         u32 ptr = ntohs(th->urg_ptr);
4863
4864         if (ptr && !sysctl_tcp_stdurg)
4865                 ptr--;
4866         ptr += ntohl(th->seq);
4867
4868         /* Ignore urgent data that we've already seen and read. */
4869         if (after(tp->copied_seq, ptr))
4870                 return;
4871
4872         /* Do not replay urg ptr.
4873          *
4874          * NOTE: interesting situation not covered by specs.
4875          * Misbehaving sender may send urg ptr, pointing to segment,
4876          * which we already have in ofo queue. We are not able to fetch
4877          * such data and will stay in TCP_URG_NOTYET until will be eaten
4878          * by recvmsg(). Seems, we are not obliged to handle such wicked
4879          * situations. But it is worth to think about possibility of some
4880          * DoSes using some hypothetical application level deadlock.
4881          */
4882         if (before(ptr, tp->rcv_nxt))
4883                 return;
4884
4885         /* Do we already have a newer (or duplicate) urgent pointer? */
4886         if (tp->urg_data && !after(ptr, tp->urg_seq))
4887                 return;
4888
4889         /* Tell the world about our new urgent pointer. */
4890         sk_send_sigurg(sk);
4891
4892         /* We may be adding urgent data when the last byte read was
4893          * urgent. To do this requires some care. We cannot just ignore
4894          * tp->copied_seq since we would read the last urgent byte again
4895          * as data, nor can we alter copied_seq until this data arrives
4896          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4897          *
4898          * NOTE. Double Dutch. Rendering to plain English: author of comment
4899          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4900          * and expect that both A and B disappear from stream. This is _wrong_.
4901          * Though this happens in BSD with high probability, this is occasional.
4902          * Any application relying on this is buggy. Note also, that fix "works"
4903          * only in this artificial test. Insert some normal data between A and B and we will
4904          * decline of BSD again. Verdict: it is better to remove to trap
4905          * buggy users.
4906          */
4907         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4908             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4909                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4910                 tp->copied_seq++;
4911                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4912                         __skb_unlink(skb, &sk->sk_receive_queue);
4913                         __kfree_skb(skb);
4914                 }
4915         }
4916
4917         tp->urg_data = TCP_URG_NOTYET;
4918         tp->urg_seq = ptr;
4919
4920         /* Disable header prediction. */
4921         tp->pred_flags = 0;
4922 }
4923
4924 /* This is the 'fast' part of urgent handling. */
4925 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4926 {
4927         struct tcp_sock *tp = tcp_sk(sk);
4928
4929         /* Check if we get a new urgent pointer - normally not. */
4930         if (th->urg)
4931                 tcp_check_urg(sk, th);
4932
4933         /* Do we wait for any urgent data? - normally not... */
4934         if (tp->urg_data == TCP_URG_NOTYET) {
4935                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4936                           th->syn;
4937
4938                 /* Is the urgent pointer pointing into this packet? */
4939                 if (ptr < skb->len) {
4940                         u8 tmp;
4941                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4942                                 BUG();
4943                         tp->urg_data = TCP_URG_VALID | tmp;
4944                         if (!sock_flag(sk, SOCK_DEAD))
4945                                 sk->sk_data_ready(sk, 0);
4946                 }
4947         }
4948 }
4949
4950 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4951 {
4952         struct tcp_sock *tp = tcp_sk(sk);
4953         int chunk = skb->len - hlen;
4954         int err;
4955
4956         local_bh_enable();
4957         if (skb_csum_unnecessary(skb))
4958                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4959         else
4960                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4961                                                        tp->ucopy.iov);
4962
4963         if (!err) {
4964                 tp->ucopy.len -= chunk;
4965                 tp->copied_seq += chunk;
4966                 tcp_rcv_space_adjust(sk);
4967         }
4968
4969         local_bh_disable();
4970         return err;
4971 }
4972
4973 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4974                                             struct sk_buff *skb)
4975 {
4976         __sum16 result;
4977
4978         if (sock_owned_by_user(sk)) {
4979                 local_bh_enable();
4980                 result = __tcp_checksum_complete(skb);
4981                 local_bh_disable();
4982         } else {
4983                 result = __tcp_checksum_complete(skb);
4984         }
4985         return result;
4986 }
4987
4988 static inline int tcp_checksum_complete_user(struct sock *sk,
4989                                              struct sk_buff *skb)
4990 {
4991         return !skb_csum_unnecessary(skb) &&
4992                __tcp_checksum_complete_user(sk, skb);
4993 }
4994
4995 #ifdef CONFIG_NET_DMA
4996 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4997                                   int hlen)
4998 {
4999         struct tcp_sock *tp = tcp_sk(sk);
5000         int chunk = skb->len - hlen;
5001         int dma_cookie;
5002         int copied_early = 0;
5003
5004         if (tp->ucopy.wakeup)
5005                 return 0;
5006
5007         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5008                 tp->ucopy.dma_chan = get_softnet_dma();
5009
5010         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5011
5012                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5013                                                          skb, hlen,
5014                                                          tp->ucopy.iov, chunk,
5015                                                          tp->ucopy.pinned_list);
5016
5017                 if (dma_cookie < 0)
5018                         goto out;
5019
5020                 tp->ucopy.dma_cookie = dma_cookie;
5021                 copied_early = 1;
5022
5023                 tp->ucopy.len -= chunk;
5024                 tp->copied_seq += chunk;
5025                 tcp_rcv_space_adjust(sk);
5026
5027                 if ((tp->ucopy.len == 0) ||
5028                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5029                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5030                         tp->ucopy.wakeup = 1;
5031                         sk->sk_data_ready(sk, 0);
5032                 }
5033         } else if (chunk > 0) {
5034                 tp->ucopy.wakeup = 1;
5035                 sk->sk_data_ready(sk, 0);
5036         }
5037 out:
5038         return copied_early;
5039 }
5040 #endif /* CONFIG_NET_DMA */
5041
5042 /* Does PAWS and seqno based validation of an incoming segment, flags will
5043  * play significant role here.
5044  */
5045 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5046                               struct tcphdr *th, int syn_inerr)
5047 {
5048         struct tcp_sock *tp = tcp_sk(sk);
5049
5050         /* RFC1323: H1. Apply PAWS check first. */
5051         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5052             tcp_paws_discard(sk, skb)) {
5053                 if (!th->rst) {
5054                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5055                         tcp_send_dupack(sk, skb);
5056                         goto discard;
5057                 }
5058                 /* Reset is accepted even if it did not pass PAWS. */
5059         }
5060
5061         /* Step 1: check sequence number */
5062         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5063                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5064                  * (RST) segments are validated by checking their SEQ-fields."
5065                  * And page 69: "If an incoming segment is not acceptable,
5066                  * an acknowledgment should be sent in reply (unless the RST
5067                  * bit is set, if so drop the segment and return)".
5068                  */
5069                 if (!th->rst)
5070                         tcp_send_dupack(sk, skb);
5071                 goto discard;
5072         }
5073
5074         /* Step 2: check RST bit */
5075         if (th->rst) {
5076                 tcp_reset(sk);
5077                 goto discard;
5078         }
5079
5080         /* ts_recent update must be made after we are sure that the packet
5081          * is in window.
5082          */
5083         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5084
5085         /* step 3: check security and precedence [ignored] */
5086
5087         /* step 4: Check for a SYN in window. */
5088         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5089                 if (syn_inerr)
5090                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5091                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5092                 tcp_reset(sk);
5093                 return -1;
5094         }
5095
5096         return 1;
5097
5098 discard:
5099         __kfree_skb(skb);
5100         return 0;
5101 }
5102
5103 /*
5104  *      TCP receive function for the ESTABLISHED state.
5105  *
5106  *      It is split into a fast path and a slow path. The fast path is
5107  *      disabled when:
5108  *      - A zero window was announced from us - zero window probing
5109  *        is only handled properly in the slow path.
5110  *      - Out of order segments arrived.
5111  *      - Urgent data is expected.
5112  *      - There is no buffer space left
5113  *      - Unexpected TCP flags/window values/header lengths are received
5114  *        (detected by checking the TCP header against pred_flags)
5115  *      - Data is sent in both directions. Fast path only supports pure senders
5116  *        or pure receivers (this means either the sequence number or the ack
5117  *        value must stay constant)
5118  *      - Unexpected TCP option.
5119  *
5120  *      When these conditions are not satisfied it drops into a standard
5121  *      receive procedure patterned after RFC793 to handle all cases.
5122  *      The first three cases are guaranteed by proper pred_flags setting,
5123  *      the rest is checked inline. Fast processing is turned on in
5124  *      tcp_data_queue when everything is OK.
5125  */
5126 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5127                         struct tcphdr *th, unsigned len)
5128 {
5129         struct tcp_sock *tp = tcp_sk(sk);
5130         int res;
5131
5132         /*
5133          *      Header prediction.
5134          *      The code loosely follows the one in the famous
5135          *      "30 instruction TCP receive" Van Jacobson mail.
5136          *
5137          *      Van's trick is to deposit buffers into socket queue
5138          *      on a device interrupt, to call tcp_recv function
5139          *      on the receive process context and checksum and copy
5140          *      the buffer to user space. smart...
5141          *
5142          *      Our current scheme is not silly either but we take the
5143          *      extra cost of the net_bh soft interrupt processing...
5144          *      We do checksum and copy also but from device to kernel.
5145          */
5146
5147         tp->rx_opt.saw_tstamp = 0;
5148
5149         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5150          *      if header_prediction is to be made
5151          *      'S' will always be tp->tcp_header_len >> 2
5152          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5153          *  turn it off (when there are holes in the receive
5154          *       space for instance)
5155          *      PSH flag is ignored.
5156          */
5157
5158         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5159             TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5160                 int tcp_header_len = tp->tcp_header_len;
5161
5162                 /* Timestamp header prediction: tcp_header_len
5163                  * is automatically equal to th->doff*4 due to pred_flags
5164                  * match.
5165                  */
5166
5167                 /* Check timestamp */
5168                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5169                         /* No? Slow path! */
5170                         if (!tcp_parse_aligned_timestamp(tp, th))
5171                                 goto slow_path;
5172
5173                         /* If PAWS failed, check it more carefully in slow path */
5174                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5175                                 goto slow_path;
5176
5177                         /* DO NOT update ts_recent here, if checksum fails
5178                          * and timestamp was corrupted part, it will result
5179                          * in a hung connection since we will drop all
5180                          * future packets due to the PAWS test.
5181                          */
5182                 }
5183
5184                 if (len <= tcp_header_len) {
5185                         /* Bulk data transfer: sender */
5186                         if (len == tcp_header_len) {
5187                                 /* Predicted packet is in window by definition.
5188                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5189                                  * Hence, check seq<=rcv_wup reduces to:
5190                                  */
5191                                 if (tcp_header_len ==
5192                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5193                                     tp->rcv_nxt == tp->rcv_wup)
5194                                         tcp_store_ts_recent(tp);
5195
5196                                 /* We know that such packets are checksummed
5197                                  * on entry.
5198                                  */
5199                                 tcp_ack(sk, skb, 0);
5200                                 __kfree_skb(skb);
5201                                 tcp_data_snd_check(sk);
5202                                 return 0;
5203                         } else { /* Header too small */
5204                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5205                                 goto discard;
5206                         }
5207                 } else {
5208                         int eaten = 0;
5209                         int copied_early = 0;
5210
5211                         if (tp->copied_seq == tp->rcv_nxt &&
5212                             len - tcp_header_len <= tp->ucopy.len) {
5213 #ifdef CONFIG_NET_DMA
5214                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5215                                         copied_early = 1;
5216                                         eaten = 1;
5217                                 }
5218 #endif
5219                                 if (tp->ucopy.task == current &&
5220                                     sock_owned_by_user(sk) && !copied_early) {
5221                                         __set_current_state(TASK_RUNNING);
5222
5223                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5224                                                 eaten = 1;
5225                                 }
5226                                 if (eaten) {
5227                                         /* Predicted packet is in window by definition.
5228                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5229                                          * Hence, check seq<=rcv_wup reduces to:
5230                                          */
5231                                         if (tcp_header_len ==
5232                                             (sizeof(struct tcphdr) +
5233                                              TCPOLEN_TSTAMP_ALIGNED) &&
5234                                             tp->rcv_nxt == tp->rcv_wup)
5235                                                 tcp_store_ts_recent(tp);
5236
5237                                         tcp_rcv_rtt_measure_ts(sk, skb);
5238
5239                                         __skb_pull(skb, tcp_header_len);
5240                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5241                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5242                                 }
5243                                 if (copied_early)
5244                                         tcp_cleanup_rbuf(sk, skb->len);
5245                         }
5246                         if (!eaten) {
5247                                 if (tcp_checksum_complete_user(sk, skb))
5248                                         goto csum_error;
5249
5250                                 /* Predicted packet is in window by definition.
5251                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5252                                  * Hence, check seq<=rcv_wup reduces to:
5253                                  */
5254                                 if (tcp_header_len ==
5255                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5256                                     tp->rcv_nxt == tp->rcv_wup)
5257                                         tcp_store_ts_recent(tp);
5258
5259                                 tcp_rcv_rtt_measure_ts(sk, skb);
5260
5261                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5262                                         goto step5;
5263
5264                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5265
5266                                 /* Bulk data transfer: receiver */
5267                                 __skb_pull(skb, tcp_header_len);
5268                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
5269                                 skb_set_owner_r(skb, sk);
5270                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5271                         }
5272
5273                         tcp_event_data_recv(sk, skb);
5274
5275                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5276                                 /* Well, only one small jumplet in fast path... */
5277                                 tcp_ack(sk, skb, FLAG_DATA);
5278                                 tcp_data_snd_check(sk);
5279                                 if (!inet_csk_ack_scheduled(sk))
5280                                         goto no_ack;
5281                         }
5282
5283                         if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5284                                 __tcp_ack_snd_check(sk, 0);
5285 no_ack:
5286 #ifdef CONFIG_NET_DMA
5287                         if (copied_early)
5288                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5289                         else
5290 #endif
5291                         if (eaten)
5292                                 __kfree_skb(skb);
5293                         else
5294                                 sk->sk_data_ready(sk, 0);
5295                         return 0;
5296                 }
5297         }
5298
5299 slow_path:
5300         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5301                 goto csum_error;
5302
5303         /*
5304          *      Standard slow path.
5305          */
5306
5307         res = tcp_validate_incoming(sk, skb, th, 1);
5308         if (res <= 0)
5309                 return -res;
5310
5311 step5:
5312         if (th->ack)
5313                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5314
5315         tcp_rcv_rtt_measure_ts(sk, skb);
5316
5317         /* Process urgent data. */
5318         tcp_urg(sk, skb, th);
5319
5320         /* step 7: process the segment text */
5321         tcp_data_queue(sk, skb);
5322
5323         tcp_data_snd_check(sk);
5324         tcp_ack_snd_check(sk);
5325         return 0;
5326
5327 csum_error:
5328         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5329
5330 discard:
5331         __kfree_skb(skb);
5332         return 0;
5333 }
5334
5335 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5336                                          struct tcphdr *th, unsigned len)
5337 {
5338         struct tcp_sock *tp = tcp_sk(sk);
5339         struct inet_connection_sock *icsk = inet_csk(sk);
5340         int saved_clamp = tp->rx_opt.mss_clamp;
5341
5342         tcp_parse_options(skb, &tp->rx_opt, 0);
5343
5344         if (th->ack) {
5345                 /* rfc793:
5346                  * "If the state is SYN-SENT then
5347                  *    first check the ACK bit
5348                  *      If the ACK bit is set
5349                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5350                  *        a reset (unless the RST bit is set, if so drop
5351                  *        the segment and return)"
5352                  *
5353                  *  We do not send data with SYN, so that RFC-correct
5354                  *  test reduces to:
5355                  */
5356                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5357                         goto reset_and_undo;
5358
5359                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5360                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5361                              tcp_time_stamp)) {
5362                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5363                         goto reset_and_undo;
5364                 }
5365
5366                 /* Now ACK is acceptable.
5367                  *
5368                  * "If the RST bit is set
5369                  *    If the ACK was acceptable then signal the user "error:
5370                  *    connection reset", drop the segment, enter CLOSED state,
5371                  *    delete TCB, and return."
5372                  */
5373
5374                 if (th->rst) {
5375                         tcp_reset(sk);
5376                         goto discard;
5377                 }
5378
5379                 /* rfc793:
5380                  *   "fifth, if neither of the SYN or RST bits is set then
5381                  *    drop the segment and return."
5382                  *
5383                  *    See note below!
5384                  *                                        --ANK(990513)
5385                  */
5386                 if (!th->syn)
5387                         goto discard_and_undo;
5388
5389                 /* rfc793:
5390                  *   "If the SYN bit is on ...
5391                  *    are acceptable then ...
5392                  *    (our SYN has been ACKed), change the connection
5393                  *    state to ESTABLISHED..."
5394                  */
5395
5396                 TCP_ECN_rcv_synack(tp, th);
5397
5398                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5399                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5400
5401                 /* Ok.. it's good. Set up sequence numbers and
5402                  * move to established.
5403                  */
5404                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5405                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5406
5407                 /* RFC1323: The window in SYN & SYN/ACK segments is
5408                  * never scaled.
5409                  */
5410                 tp->snd_wnd = ntohs(th->window);
5411                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
5412
5413                 if (!tp->rx_opt.wscale_ok) {
5414                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5415                         tp->window_clamp = min(tp->window_clamp, 65535U);
5416                 }
5417
5418                 if (tp->rx_opt.saw_tstamp) {
5419                         tp->rx_opt.tstamp_ok       = 1;
5420                         tp->tcp_header_len =
5421                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5422                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5423                         tcp_store_ts_recent(tp);
5424                 } else {
5425                         tp->tcp_header_len = sizeof(struct tcphdr);
5426                 }
5427
5428                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5429                         tcp_enable_fack(tp);
5430
5431                 tcp_mtup_init(sk);
5432                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5433                 tcp_initialize_rcv_mss(sk);
5434
5435                 /* Remember, tcp_poll() does not lock socket!
5436                  * Change state from SYN-SENT only after copied_seq
5437                  * is initialized. */
5438                 tp->copied_seq = tp->rcv_nxt;
5439                 smp_mb();
5440                 tcp_set_state(sk, TCP_ESTABLISHED);
5441
5442                 security_inet_conn_established(sk, skb);
5443
5444                 /* Make sure socket is routed, for correct metrics.  */
5445                 icsk->icsk_af_ops->rebuild_header(sk);
5446
5447                 tcp_init_metrics(sk);
5448
5449                 tcp_init_congestion_control(sk);
5450
5451                 /* Prevent spurious tcp_cwnd_restart() on first data
5452                  * packet.
5453                  */
5454                 tp->lsndtime = tcp_time_stamp;
5455
5456                 tcp_init_buffer_space(sk);
5457
5458                 if (sock_flag(sk, SOCK_KEEPOPEN))
5459                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5460
5461                 if (!tp->rx_opt.snd_wscale)
5462                         __tcp_fast_path_on(tp, tp->snd_wnd);
5463                 else
5464                         tp->pred_flags = 0;
5465
5466                 if (!sock_flag(sk, SOCK_DEAD)) {
5467                         sk->sk_state_change(sk);
5468                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5469                 }
5470
5471                 if (sk->sk_write_pending ||
5472                     icsk->icsk_accept_queue.rskq_defer_accept ||
5473                     icsk->icsk_ack.pingpong) {
5474                         /* Save one ACK. Data will be ready after
5475                          * several ticks, if write_pending is set.
5476                          *
5477                          * It may be deleted, but with this feature tcpdumps
5478                          * look so _wonderfully_ clever, that I was not able
5479                          * to stand against the temptation 8)     --ANK
5480                          */
5481                         inet_csk_schedule_ack(sk);
5482                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5483                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
5484                         tcp_incr_quickack(sk);
5485                         tcp_enter_quickack_mode(sk);
5486                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5487                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5488
5489 discard:
5490                         __kfree_skb(skb);
5491                         return 0;
5492                 } else {
5493                         tcp_send_ack(sk);
5494                 }
5495                 return -1;
5496         }
5497
5498         /* No ACK in the segment */
5499
5500         if (th->rst) {
5501                 /* rfc793:
5502                  * "If the RST bit is set
5503                  *
5504                  *      Otherwise (no ACK) drop the segment and return."
5505                  */
5506
5507                 goto discard_and_undo;
5508         }
5509
5510         /* PAWS check. */
5511         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5512             tcp_paws_check(&tp->rx_opt, 0))
5513                 goto discard_and_undo;
5514
5515         if (th->syn) {
5516                 /* We see SYN without ACK. It is attempt of
5517                  * simultaneous connect with crossed SYNs.
5518                  * Particularly, it can be connect to self.
5519                  */
5520                 tcp_set_state(sk, TCP_SYN_RECV);
5521
5522                 if (tp->rx_opt.saw_tstamp) {
5523                         tp->rx_opt.tstamp_ok = 1;
5524                         tcp_store_ts_recent(tp);
5525                         tp->tcp_header_len =
5526                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5527                 } else {
5528                         tp->tcp_header_len = sizeof(struct tcphdr);
5529                 }
5530
5531                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5532                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5533
5534                 /* RFC1323: The window in SYN & SYN/ACK segments is
5535                  * never scaled.
5536                  */
5537                 tp->snd_wnd    = ntohs(th->window);
5538                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5539                 tp->max_window = tp->snd_wnd;
5540
5541                 TCP_ECN_rcv_syn(tp, th);
5542
5543                 tcp_mtup_init(sk);
5544                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5545                 tcp_initialize_rcv_mss(sk);
5546
5547                 tcp_send_synack(sk);
5548 #if 0
5549                 /* Note, we could accept data and URG from this segment.
5550                  * There are no obstacles to make this.
5551                  *
5552                  * However, if we ignore data in ACKless segments sometimes,
5553                  * we have no reasons to accept it sometimes.
5554                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5555                  * is not flawless. So, discard packet for sanity.
5556                  * Uncomment this return to process the data.
5557                  */
5558                 return -1;
5559 #else
5560                 goto discard;
5561 #endif
5562         }
5563         /* "fifth, if neither of the SYN or RST bits is set then
5564          * drop the segment and return."
5565          */
5566
5567 discard_and_undo:
5568         tcp_clear_options(&tp->rx_opt);
5569         tp->rx_opt.mss_clamp = saved_clamp;
5570         goto discard;
5571
5572 reset_and_undo:
5573         tcp_clear_options(&tp->rx_opt);
5574         tp->rx_opt.mss_clamp = saved_clamp;
5575         return 1;
5576 }
5577
5578 /*
5579  *      This function implements the receiving procedure of RFC 793 for
5580  *      all states except ESTABLISHED and TIME_WAIT.
5581  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5582  *      address independent.
5583  */
5584
5585 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5586                           struct tcphdr *th, unsigned len)
5587 {
5588         struct tcp_sock *tp = tcp_sk(sk);
5589         struct inet_connection_sock *icsk = inet_csk(sk);
5590         int queued = 0;
5591         int res;
5592
5593         tp->rx_opt.saw_tstamp = 0;
5594
5595         switch (sk->sk_state) {
5596         case TCP_CLOSE:
5597                 goto discard;
5598
5599         case TCP_LISTEN:
5600                 if (th->ack)
5601                         return 1;
5602
5603                 if (th->rst)
5604                         goto discard;
5605
5606                 if (th->syn) {
5607                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5608                                 return 1;
5609
5610                         /* Now we have several options: In theory there is
5611                          * nothing else in the frame. KA9Q has an option to
5612                          * send data with the syn, BSD accepts data with the
5613                          * syn up to the [to be] advertised window and
5614                          * Solaris 2.1 gives you a protocol error. For now
5615                          * we just ignore it, that fits the spec precisely
5616                          * and avoids incompatibilities. It would be nice in
5617                          * future to drop through and process the data.
5618                          *
5619                          * Now that TTCP is starting to be used we ought to
5620                          * queue this data.
5621                          * But, this leaves one open to an easy denial of
5622                          * service attack, and SYN cookies can't defend
5623                          * against this problem. So, we drop the data
5624                          * in the interest of security over speed unless
5625                          * it's still in use.
5626                          */
5627                         kfree_skb(skb);
5628                         return 0;
5629                 }
5630                 goto discard;
5631
5632         case TCP_SYN_SENT:
5633                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5634                 if (queued >= 0)
5635                         return queued;
5636
5637                 /* Do step6 onward by hand. */
5638                 tcp_urg(sk, skb, th);
5639                 __kfree_skb(skb);
5640                 tcp_data_snd_check(sk);
5641                 return 0;
5642         }
5643
5644         res = tcp_validate_incoming(sk, skb, th, 0);
5645         if (res <= 0)
5646                 return -res;
5647
5648         /* step 5: check the ACK field */
5649         if (th->ack) {
5650                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5651
5652                 switch (sk->sk_state) {
5653                 case TCP_SYN_RECV:
5654                         if (acceptable) {
5655                                 tp->copied_seq = tp->rcv_nxt;
5656                                 smp_mb();
5657                                 tcp_set_state(sk, TCP_ESTABLISHED);
5658                                 sk->sk_state_change(sk);
5659
5660                                 /* Note, that this wakeup is only for marginal
5661                                  * crossed SYN case. Passively open sockets
5662                                  * are not waked up, because sk->sk_sleep ==
5663                                  * NULL and sk->sk_socket == NULL.
5664                                  */
5665                                 if (sk->sk_socket)
5666                                         sk_wake_async(sk,
5667                                                       SOCK_WAKE_IO, POLL_OUT);
5668
5669                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5670                                 tp->snd_wnd = ntohs(th->window) <<
5671                                               tp->rx_opt.snd_wscale;
5672                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5673                                             TCP_SKB_CB(skb)->seq);
5674
5675                                 /* tcp_ack considers this ACK as duplicate
5676                                  * and does not calculate rtt.
5677                                  * Fix it at least with timestamps.
5678                                  */
5679                                 if (tp->rx_opt.saw_tstamp &&
5680                                     tp->rx_opt.rcv_tsecr && !tp->srtt)
5681                                         tcp_ack_saw_tstamp(sk, 0);
5682
5683                                 if (tp->rx_opt.tstamp_ok)
5684                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5685
5686                                 /* Make sure socket is routed, for
5687                                  * correct metrics.
5688                                  */
5689                                 icsk->icsk_af_ops->rebuild_header(sk);
5690
5691                                 tcp_init_metrics(sk);
5692
5693                                 tcp_init_congestion_control(sk);
5694
5695                                 /* Prevent spurious tcp_cwnd_restart() on
5696                                  * first data packet.
5697                                  */
5698                                 tp->lsndtime = tcp_time_stamp;
5699
5700                                 tcp_mtup_init(sk);
5701                                 tcp_initialize_rcv_mss(sk);
5702                                 tcp_init_buffer_space(sk);
5703                                 tcp_fast_path_on(tp);
5704                         } else {
5705                                 return 1;
5706                         }
5707                         break;
5708
5709                 case TCP_FIN_WAIT1:
5710                         if (tp->snd_una == tp->write_seq) {
5711                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5712                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5713                                 dst_confirm(sk->sk_dst_cache);
5714
5715                                 if (!sock_flag(sk, SOCK_DEAD))
5716                                         /* Wake up lingering close() */
5717                                         sk->sk_state_change(sk);
5718                                 else {
5719                                         int tmo;
5720
5721                                         if (tp->linger2 < 0 ||
5722                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5723                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5724                                                 tcp_done(sk);
5725                                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5726                                                 return 1;
5727                                         }
5728
5729                                         tmo = tcp_fin_time(sk);
5730                                         if (tmo > TCP_TIMEWAIT_LEN) {
5731                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5732                                         } else if (th->fin || sock_owned_by_user(sk)) {
5733                                                 /* Bad case. We could lose such FIN otherwise.
5734                                                  * It is not a big problem, but it looks confusing
5735                                                  * and not so rare event. We still can lose it now,
5736                                                  * if it spins in bh_lock_sock(), but it is really
5737                                                  * marginal case.
5738                                                  */
5739                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5740                                         } else {
5741                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5742                                                 goto discard;
5743                                         }
5744                                 }
5745                         }
5746                         break;
5747
5748                 case TCP_CLOSING:
5749                         if (tp->snd_una == tp->write_seq) {
5750                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5751                                 goto discard;
5752                         }
5753                         break;
5754
5755                 case TCP_LAST_ACK:
5756                         if (tp->snd_una == tp->write_seq) {
5757                                 tcp_update_metrics(sk);
5758                                 tcp_done(sk);
5759                                 goto discard;
5760                         }
5761                         break;
5762                 }
5763         } else
5764                 goto discard;
5765
5766         /* step 6: check the URG bit */
5767         tcp_urg(sk, skb, th);
5768
5769         /* step 7: process the segment text */
5770         switch (sk->sk_state) {
5771         case TCP_CLOSE_WAIT:
5772         case TCP_CLOSING:
5773         case TCP_LAST_ACK:
5774                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5775                         break;
5776         case TCP_FIN_WAIT1:
5777         case TCP_FIN_WAIT2:
5778                 /* RFC 793 says to queue data in these states,
5779                  * RFC 1122 says we MUST send a reset.
5780                  * BSD 4.4 also does reset.
5781                  */
5782                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5783                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5784                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5785                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5786                                 tcp_reset(sk);
5787                                 return 1;
5788                         }
5789                 }
5790                 /* Fall through */
5791         case TCP_ESTABLISHED:
5792                 tcp_data_queue(sk, skb);
5793                 queued = 1;
5794                 break;
5795         }
5796
5797         /* tcp_data could move socket to TIME-WAIT */
5798         if (sk->sk_state != TCP_CLOSE) {
5799                 tcp_data_snd_check(sk);
5800                 tcp_ack_snd_check(sk);
5801         }
5802
5803         if (!queued) {
5804 discard:
5805                 __kfree_skb(skb);
5806         }
5807         return 0;
5808 }
5809
5810 EXPORT_SYMBOL(sysctl_tcp_ecn);
5811 EXPORT_SYMBOL(sysctl_tcp_reordering);
5812 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5813 EXPORT_SYMBOL(tcp_parse_options);
5814 #ifdef CONFIG_TCP_MD5SIG
5815 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5816 #endif
5817 EXPORT_SYMBOL(tcp_rcv_established);
5818 EXPORT_SYMBOL(tcp_rcv_state_process);
5819 EXPORT_SYMBOL(tcp_initialize_rcv_mss);