tcp: kill dead end_seq variable in clean_rtx_queue
[pandora-kernel.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * 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                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1182                  * constraint here (see above) but figuring out that at
1183                  * least tp->reordering SACK blocks reside between ack_seq
1184                  * and received_upto is not easy task to do cheaply with
1185                  * the available datastructures.
1186                  *
1187                  * Whether FACK should check here for tp->reordering segs
1188                  * in-between one could argue for either way (it would be
1189                  * rather simple to implement as we could count fack_count
1190                  * during the walk and do tp->fackets_out - fack_count).
1191                  */
1192                 if (after(received_upto, ack_seq)) {
1193                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1194                         tp->retrans_out -= tcp_skb_pcount(skb);
1195
1196                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1197                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1198                 } else {
1199                         if (before(ack_seq, new_low_seq))
1200                                 new_low_seq = ack_seq;
1201                         cnt += tcp_skb_pcount(skb);
1202                 }
1203         }
1204
1205         if (tp->retrans_out)
1206                 tp->lost_retrans_low = new_low_seq;
1207 }
1208
1209 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1210                            struct tcp_sack_block_wire *sp, int num_sacks,
1211                            u32 prior_snd_una)
1212 {
1213         struct tcp_sock *tp = tcp_sk(sk);
1214         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1215         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1216         int dup_sack = 0;
1217
1218         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1219                 dup_sack = 1;
1220                 tcp_dsack_seen(tp);
1221                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1222         } else if (num_sacks > 1) {
1223                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1224                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1225
1226                 if (!after(end_seq_0, end_seq_1) &&
1227                     !before(start_seq_0, start_seq_1)) {
1228                         dup_sack = 1;
1229                         tcp_dsack_seen(tp);
1230                         NET_INC_STATS_BH(sock_net(sk),
1231                                         LINUX_MIB_TCPDSACKOFORECV);
1232                 }
1233         }
1234
1235         /* D-SACK for already forgotten data... Do dumb counting. */
1236         if (dup_sack &&
1237             !after(end_seq_0, prior_snd_una) &&
1238             after(end_seq_0, tp->undo_marker))
1239                 tp->undo_retrans--;
1240
1241         return dup_sack;
1242 }
1243
1244 struct tcp_sacktag_state {
1245         int reord;
1246         int fack_count;
1247         int flag;
1248 };
1249
1250 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1251  * the incoming SACK may not exactly match but we can find smaller MSS
1252  * aligned portion of it that matches. Therefore we might need to fragment
1253  * which may fail and creates some hassle (caller must handle error case
1254  * returns).
1255  *
1256  * FIXME: this could be merged to shift decision code
1257  */
1258 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1259                                  u32 start_seq, u32 end_seq)
1260 {
1261         int in_sack, err;
1262         unsigned int pkt_len;
1263         unsigned int mss;
1264
1265         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1266                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1267
1268         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1269             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1270                 mss = tcp_skb_mss(skb);
1271                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1272
1273                 if (!in_sack) {
1274                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1275                         if (pkt_len < mss)
1276                                 pkt_len = mss;
1277                 } else {
1278                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1279                         if (pkt_len < mss)
1280                                 return -EINVAL;
1281                 }
1282
1283                 /* Round if necessary so that SACKs cover only full MSSes
1284                  * and/or the remaining small portion (if present)
1285                  */
1286                 if (pkt_len > mss) {
1287                         unsigned int new_len = (pkt_len / mss) * mss;
1288                         if (!in_sack && new_len < pkt_len) {
1289                                 new_len += mss;
1290                                 if (new_len > skb->len)
1291                                         return 0;
1292                         }
1293                         pkt_len = new_len;
1294                 }
1295                 err = tcp_fragment(sk, skb, pkt_len, mss);
1296                 if (err < 0)
1297                         return err;
1298         }
1299
1300         return in_sack;
1301 }
1302
1303 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1304                           struct tcp_sacktag_state *state,
1305                           int dup_sack, int pcount)
1306 {
1307         struct tcp_sock *tp = tcp_sk(sk);
1308         u8 sacked = TCP_SKB_CB(skb)->sacked;
1309         int fack_count = state->fack_count;
1310
1311         /* Account D-SACK for retransmitted packet. */
1312         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1313                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1314                         tp->undo_retrans--;
1315                 if (sacked & TCPCB_SACKED_ACKED)
1316                         state->reord = min(fack_count, state->reord);
1317         }
1318
1319         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1320         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1321                 return sacked;
1322
1323         if (!(sacked & TCPCB_SACKED_ACKED)) {
1324                 if (sacked & TCPCB_SACKED_RETRANS) {
1325                         /* If the segment is not tagged as lost,
1326                          * we do not clear RETRANS, believing
1327                          * that retransmission is still in flight.
1328                          */
1329                         if (sacked & TCPCB_LOST) {
1330                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1331                                 tp->lost_out -= pcount;
1332                                 tp->retrans_out -= pcount;
1333                         }
1334                 } else {
1335                         if (!(sacked & TCPCB_RETRANS)) {
1336                                 /* New sack for not retransmitted frame,
1337                                  * which was in hole. It is reordering.
1338                                  */
1339                                 if (before(TCP_SKB_CB(skb)->seq,
1340                                            tcp_highest_sack_seq(tp)))
1341                                         state->reord = min(fack_count,
1342                                                            state->reord);
1343
1344                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1345                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1346                                         state->flag |= FLAG_ONLY_ORIG_SACKED;
1347                         }
1348
1349                         if (sacked & TCPCB_LOST) {
1350                                 sacked &= ~TCPCB_LOST;
1351                                 tp->lost_out -= pcount;
1352                         }
1353                 }
1354
1355                 sacked |= TCPCB_SACKED_ACKED;
1356                 state->flag |= FLAG_DATA_SACKED;
1357                 tp->sacked_out += pcount;
1358
1359                 fack_count += pcount;
1360
1361                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1362                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1363                     before(TCP_SKB_CB(skb)->seq,
1364                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1365                         tp->lost_cnt_hint += pcount;
1366
1367                 if (fack_count > tp->fackets_out)
1368                         tp->fackets_out = fack_count;
1369         }
1370
1371         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1372          * frames and clear it. undo_retrans is decreased above, L|R frames
1373          * are accounted above as well.
1374          */
1375         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1376                 sacked &= ~TCPCB_SACKED_RETRANS;
1377                 tp->retrans_out -= pcount;
1378         }
1379
1380         return sacked;
1381 }
1382
1383 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1384                            struct tcp_sacktag_state *state,
1385                            unsigned int pcount, int shifted, int mss,
1386                            int dup_sack)
1387 {
1388         struct tcp_sock *tp = tcp_sk(sk);
1389         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1390
1391         BUG_ON(!pcount);
1392
1393         /* Tweak before seqno plays */
1394         if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1395             !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1396                 tp->lost_cnt_hint += pcount;
1397
1398         TCP_SKB_CB(prev)->end_seq += shifted;
1399         TCP_SKB_CB(skb)->seq += shifted;
1400
1401         skb_shinfo(prev)->gso_segs += pcount;
1402         BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1403         skb_shinfo(skb)->gso_segs -= pcount;
1404
1405         /* When we're adding to gso_segs == 1, gso_size will be zero,
1406          * in theory this shouldn't be necessary but as long as DSACK
1407          * code can come after this skb later on it's better to keep
1408          * setting gso_size to something.
1409          */
1410         if (!skb_shinfo(prev)->gso_size) {
1411                 skb_shinfo(prev)->gso_size = mss;
1412                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1413         }
1414
1415         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1416         if (skb_shinfo(skb)->gso_segs <= 1) {
1417                 skb_shinfo(skb)->gso_size = 0;
1418                 skb_shinfo(skb)->gso_type = 0;
1419         }
1420
1421         /* We discard results */
1422         tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1423
1424         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1425         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1426
1427         if (skb->len > 0) {
1428                 BUG_ON(!tcp_skb_pcount(skb));
1429                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1430                 return 0;
1431         }
1432
1433         /* Whole SKB was eaten :-) */
1434
1435         if (skb == tp->retransmit_skb_hint)
1436                 tp->retransmit_skb_hint = prev;
1437         if (skb == tp->scoreboard_skb_hint)
1438                 tp->scoreboard_skb_hint = prev;
1439         if (skb == tp->lost_skb_hint) {
1440                 tp->lost_skb_hint = prev;
1441                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1442         }
1443
1444         TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1445         if (skb == tcp_highest_sack(sk))
1446                 tcp_advance_highest_sack(sk, skb);
1447
1448         tcp_unlink_write_queue(skb, sk);
1449         sk_wmem_free_skb(sk, skb);
1450
1451         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1452
1453         return 1;
1454 }
1455
1456 /* I wish gso_size would have a bit more sane initialization than
1457  * something-or-zero which complicates things
1458  */
1459 static int tcp_skb_seglen(struct sk_buff *skb)
1460 {
1461         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1462 }
1463
1464 /* Shifting pages past head area doesn't work */
1465 static int skb_can_shift(struct sk_buff *skb)
1466 {
1467         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1468 }
1469
1470 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1471  * skb.
1472  */
1473 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1474                                           struct tcp_sacktag_state *state,
1475                                           u32 start_seq, u32 end_seq,
1476                                           int dup_sack)
1477 {
1478         struct tcp_sock *tp = tcp_sk(sk);
1479         struct sk_buff *prev;
1480         int mss;
1481         int pcount = 0;
1482         int len;
1483         int in_sack;
1484
1485         if (!sk_can_gso(sk))
1486                 goto fallback;
1487
1488         /* Normally R but no L won't result in plain S */
1489         if (!dup_sack &&
1490             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1491                 goto fallback;
1492         if (!skb_can_shift(skb))
1493                 goto fallback;
1494         /* This frame is about to be dropped (was ACKed). */
1495         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1496                 goto fallback;
1497
1498         /* Can only happen with delayed DSACK + discard craziness */
1499         if (unlikely(skb == tcp_write_queue_head(sk)))
1500                 goto fallback;
1501         prev = tcp_write_queue_prev(sk, skb);
1502
1503         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1504                 goto fallback;
1505
1506         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1507                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1508
1509         if (in_sack) {
1510                 len = skb->len;
1511                 pcount = tcp_skb_pcount(skb);
1512                 mss = tcp_skb_seglen(skb);
1513
1514                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1515                  * drop this restriction as unnecessary
1516                  */
1517                 if (mss != tcp_skb_seglen(prev))
1518                         goto fallback;
1519         } else {
1520                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1521                         goto noop;
1522                 /* CHECKME: This is non-MSS split case only?, this will
1523                  * cause skipped skbs due to advancing loop btw, original
1524                  * has that feature too
1525                  */
1526                 if (tcp_skb_pcount(skb) <= 1)
1527                         goto noop;
1528
1529                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1530                 if (!in_sack) {
1531                         /* TODO: head merge to next could be attempted here
1532                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1533                          * though it might not be worth of the additional hassle
1534                          *
1535                          * ...we can probably just fallback to what was done
1536                          * previously. We could try merging non-SACKed ones
1537                          * as well but it probably isn't going to buy off
1538                          * because later SACKs might again split them, and
1539                          * it would make skb timestamp tracking considerably
1540                          * harder problem.
1541                          */
1542                         goto fallback;
1543                 }
1544
1545                 len = end_seq - TCP_SKB_CB(skb)->seq;
1546                 BUG_ON(len < 0);
1547                 BUG_ON(len > skb->len);
1548
1549                 /* MSS boundaries should be honoured or else pcount will
1550                  * severely break even though it makes things bit trickier.
1551                  * Optimize common case to avoid most of the divides
1552                  */
1553                 mss = tcp_skb_mss(skb);
1554
1555                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1556                  * drop this restriction as unnecessary
1557                  */
1558                 if (mss != tcp_skb_seglen(prev))
1559                         goto fallback;
1560
1561                 if (len == mss) {
1562                         pcount = 1;
1563                 } else if (len < mss) {
1564                         goto noop;
1565                 } else {
1566                         pcount = len / mss;
1567                         len = pcount * mss;
1568                 }
1569         }
1570
1571         if (!skb_shift(prev, skb, len))
1572                 goto fallback;
1573         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1574                 goto out;
1575
1576         /* Hole filled allows collapsing with the next as well, this is very
1577          * useful when hole on every nth skb pattern happens
1578          */
1579         if (prev == tcp_write_queue_tail(sk))
1580                 goto out;
1581         skb = tcp_write_queue_next(sk, prev);
1582
1583         if (!skb_can_shift(skb) ||
1584             (skb == tcp_send_head(sk)) ||
1585             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1586             (mss != tcp_skb_seglen(skb)))
1587                 goto out;
1588
1589         len = skb->len;
1590         if (skb_shift(prev, skb, len)) {
1591                 pcount += tcp_skb_pcount(skb);
1592                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1593         }
1594
1595 out:
1596         state->fack_count += pcount;
1597         return prev;
1598
1599 noop:
1600         return skb;
1601
1602 fallback:
1603         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1604         return NULL;
1605 }
1606
1607 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1608                                         struct tcp_sack_block *next_dup,
1609                                         struct tcp_sacktag_state *state,
1610                                         u32 start_seq, u32 end_seq,
1611                                         int dup_sack_in)
1612 {
1613         struct tcp_sock *tp = tcp_sk(sk);
1614         struct sk_buff *tmp;
1615
1616         tcp_for_write_queue_from(skb, sk) {
1617                 int in_sack = 0;
1618                 int dup_sack = dup_sack_in;
1619
1620                 if (skb == tcp_send_head(sk))
1621                         break;
1622
1623                 /* queue is in-order => we can short-circuit the walk early */
1624                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1625                         break;
1626
1627                 if ((next_dup != NULL) &&
1628                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1629                         in_sack = tcp_match_skb_to_sack(sk, skb,
1630                                                         next_dup->start_seq,
1631                                                         next_dup->end_seq);
1632                         if (in_sack > 0)
1633                                 dup_sack = 1;
1634                 }
1635
1636                 /* skb reference here is a bit tricky to get right, since
1637                  * shifting can eat and free both this skb and the next,
1638                  * so not even _safe variant of the loop is enough.
1639                  */
1640                 if (in_sack <= 0) {
1641                         tmp = tcp_shift_skb_data(sk, skb, state,
1642                                                  start_seq, end_seq, dup_sack);
1643                         if (tmp != NULL) {
1644                                 if (tmp != skb) {
1645                                         skb = tmp;
1646                                         continue;
1647                                 }
1648
1649                                 in_sack = 0;
1650                         } else {
1651                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1652                                                                 start_seq,
1653                                                                 end_seq);
1654                         }
1655                 }
1656
1657                 if (unlikely(in_sack < 0))
1658                         break;
1659
1660                 if (in_sack) {
1661                         TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1662                                                                   state,
1663                                                                   dup_sack,
1664                                                                   tcp_skb_pcount(skb));
1665
1666                         if (!before(TCP_SKB_CB(skb)->seq,
1667                                     tcp_highest_sack_seq(tp)))
1668                                 tcp_advance_highest_sack(sk, skb);
1669                 }
1670
1671                 state->fack_count += tcp_skb_pcount(skb);
1672         }
1673         return skb;
1674 }
1675
1676 /* Avoid all extra work that is being done by sacktag while walking in
1677  * a normal way
1678  */
1679 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1680                                         struct tcp_sacktag_state *state,
1681                                         u32 skip_to_seq)
1682 {
1683         tcp_for_write_queue_from(skb, sk) {
1684                 if (skb == tcp_send_head(sk))
1685                         break;
1686
1687                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1688                         break;
1689
1690                 state->fack_count += tcp_skb_pcount(skb);
1691         }
1692         return skb;
1693 }
1694
1695 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1696                                                 struct sock *sk,
1697                                                 struct tcp_sack_block *next_dup,
1698                                                 struct tcp_sacktag_state *state,
1699                                                 u32 skip_to_seq)
1700 {
1701         if (next_dup == NULL)
1702                 return skb;
1703
1704         if (before(next_dup->start_seq, skip_to_seq)) {
1705                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1706                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1707                                        next_dup->start_seq, next_dup->end_seq,
1708                                        1);
1709         }
1710
1711         return skb;
1712 }
1713
1714 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1715 {
1716         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1717 }
1718
1719 static int
1720 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1721                         u32 prior_snd_una)
1722 {
1723         const struct inet_connection_sock *icsk = inet_csk(sk);
1724         struct tcp_sock *tp = tcp_sk(sk);
1725         unsigned char *ptr = (skb_transport_header(ack_skb) +
1726                               TCP_SKB_CB(ack_skb)->sacked);
1727         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1728         struct tcp_sack_block sp[TCP_NUM_SACKS];
1729         struct tcp_sack_block *cache;
1730         struct tcp_sacktag_state state;
1731         struct sk_buff *skb;
1732         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1733         int used_sacks;
1734         int found_dup_sack = 0;
1735         int i, j;
1736         int first_sack_index;
1737
1738         state.flag = 0;
1739         state.reord = tp->packets_out;
1740
1741         if (!tp->sacked_out) {
1742                 if (WARN_ON(tp->fackets_out))
1743                         tp->fackets_out = 0;
1744                 tcp_highest_sack_reset(sk);
1745         }
1746
1747         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1748                                          num_sacks, prior_snd_una);
1749         if (found_dup_sack)
1750                 state.flag |= FLAG_DSACKING_ACK;
1751
1752         /* Eliminate too old ACKs, but take into
1753          * account more or less fresh ones, they can
1754          * contain valid SACK info.
1755          */
1756         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1757                 return 0;
1758
1759         if (!tp->packets_out)
1760                 goto out;
1761
1762         used_sacks = 0;
1763         first_sack_index = 0;
1764         for (i = 0; i < num_sacks; i++) {
1765                 int dup_sack = !i && found_dup_sack;
1766
1767                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1768                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1769
1770                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1771                                             sp[used_sacks].start_seq,
1772                                             sp[used_sacks].end_seq)) {
1773                         int mib_idx;
1774
1775                         if (dup_sack) {
1776                                 if (!tp->undo_marker)
1777                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1778                                 else
1779                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1780                         } else {
1781                                 /* Don't count olds caused by ACK reordering */
1782                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1783                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1784                                         continue;
1785                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1786                         }
1787
1788                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1789                         if (i == 0)
1790                                 first_sack_index = -1;
1791                         continue;
1792                 }
1793
1794                 /* Ignore very old stuff early */
1795                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1796                         continue;
1797
1798                 used_sacks++;
1799         }
1800
1801         /* order SACK blocks to allow in order walk of the retrans queue */
1802         for (i = used_sacks - 1; i > 0; i--) {
1803                 for (j = 0; j < i; j++) {
1804                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1805                                 struct tcp_sack_block tmp;
1806
1807                                 tmp = sp[j];
1808                                 sp[j] = sp[j + 1];
1809                                 sp[j + 1] = tmp;
1810
1811                                 /* Track where the first SACK block goes to */
1812                                 if (j == first_sack_index)
1813                                         first_sack_index = j + 1;
1814                         }
1815                 }
1816         }
1817
1818         skb = tcp_write_queue_head(sk);
1819         state.fack_count = 0;
1820         i = 0;
1821
1822         if (!tp->sacked_out) {
1823                 /* It's already past, so skip checking against it */
1824                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1825         } else {
1826                 cache = tp->recv_sack_cache;
1827                 /* Skip empty blocks in at head of the cache */
1828                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1829                        !cache->end_seq)
1830                         cache++;
1831         }
1832
1833         while (i < used_sacks) {
1834                 u32 start_seq = sp[i].start_seq;
1835                 u32 end_seq = sp[i].end_seq;
1836                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1837                 struct tcp_sack_block *next_dup = NULL;
1838
1839                 if (found_dup_sack && ((i + 1) == first_sack_index))
1840                         next_dup = &sp[i + 1];
1841
1842                 /* Event "B" in the comment above. */
1843                 if (after(end_seq, tp->high_seq))
1844                         state.flag |= FLAG_DATA_LOST;
1845
1846                 /* Skip too early cached blocks */
1847                 while (tcp_sack_cache_ok(tp, cache) &&
1848                        !before(start_seq, cache->end_seq))
1849                         cache++;
1850
1851                 /* Can skip some work by looking recv_sack_cache? */
1852                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1853                     after(end_seq, cache->start_seq)) {
1854
1855                         /* Head todo? */
1856                         if (before(start_seq, cache->start_seq)) {
1857                                 skb = tcp_sacktag_skip(skb, sk, &state,
1858                                                        start_seq);
1859                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1860                                                        &state,
1861                                                        start_seq,
1862                                                        cache->start_seq,
1863                                                        dup_sack);
1864                         }
1865
1866                         /* Rest of the block already fully processed? */
1867                         if (!after(end_seq, cache->end_seq))
1868                                 goto advance_sp;
1869
1870                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1871                                                        &state,
1872                                                        cache->end_seq);
1873
1874                         /* ...tail remains todo... */
1875                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1876                                 /* ...but better entrypoint exists! */
1877                                 skb = tcp_highest_sack(sk);
1878                                 if (skb == NULL)
1879                                         break;
1880                                 state.fack_count = tp->fackets_out;
1881                                 cache++;
1882                                 goto walk;
1883                         }
1884
1885                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1886                         /* Check overlap against next cached too (past this one already) */
1887                         cache++;
1888                         continue;
1889                 }
1890
1891                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1892                         skb = tcp_highest_sack(sk);
1893                         if (skb == NULL)
1894                                 break;
1895                         state.fack_count = tp->fackets_out;
1896                 }
1897                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1898
1899 walk:
1900                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1901                                        start_seq, end_seq, dup_sack);
1902
1903 advance_sp:
1904                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1905                  * due to in-order walk
1906                  */
1907                 if (after(end_seq, tp->frto_highmark))
1908                         state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1909
1910                 i++;
1911         }
1912
1913         /* Clear the head of the cache sack blocks so we can skip it next time */
1914         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1915                 tp->recv_sack_cache[i].start_seq = 0;
1916                 tp->recv_sack_cache[i].end_seq = 0;
1917         }
1918         for (j = 0; j < used_sacks; j++)
1919                 tp->recv_sack_cache[i++] = sp[j];
1920
1921         tcp_mark_lost_retrans(sk);
1922
1923         tcp_verify_left_out(tp);
1924
1925         if ((state.reord < tp->fackets_out) &&
1926             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1927             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1928                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1929
1930 out:
1931
1932 #if FASTRETRANS_DEBUG > 0
1933         WARN_ON((int)tp->sacked_out < 0);
1934         WARN_ON((int)tp->lost_out < 0);
1935         WARN_ON((int)tp->retrans_out < 0);
1936         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1937 #endif
1938         return state.flag;
1939 }
1940
1941 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1942  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1943  */
1944 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1945 {
1946         u32 holes;
1947
1948         holes = max(tp->lost_out, 1U);
1949         holes = min(holes, tp->packets_out);
1950
1951         if ((tp->sacked_out + holes) > tp->packets_out) {
1952                 tp->sacked_out = tp->packets_out - holes;
1953                 return 1;
1954         }
1955         return 0;
1956 }
1957
1958 /* If we receive more dupacks than we expected counting segments
1959  * in assumption of absent reordering, interpret this as reordering.
1960  * The only another reason could be bug in receiver TCP.
1961  */
1962 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1963 {
1964         struct tcp_sock *tp = tcp_sk(sk);
1965         if (tcp_limit_reno_sacked(tp))
1966                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1967 }
1968
1969 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1970
1971 static void tcp_add_reno_sack(struct sock *sk)
1972 {
1973         struct tcp_sock *tp = tcp_sk(sk);
1974         tp->sacked_out++;
1975         tcp_check_reno_reordering(sk, 0);
1976         tcp_verify_left_out(tp);
1977 }
1978
1979 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1980
1981 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1982 {
1983         struct tcp_sock *tp = tcp_sk(sk);
1984
1985         if (acked > 0) {
1986                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1987                 if (acked - 1 >= tp->sacked_out)
1988                         tp->sacked_out = 0;
1989                 else
1990                         tp->sacked_out -= acked - 1;
1991         }
1992         tcp_check_reno_reordering(sk, acked);
1993         tcp_verify_left_out(tp);
1994 }
1995
1996 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1997 {
1998         tp->sacked_out = 0;
1999 }
2000
2001 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2002 {
2003         return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2004 }
2005
2006 /* F-RTO can only be used if TCP has never retransmitted anything other than
2007  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2008  */
2009 int tcp_use_frto(struct sock *sk)
2010 {
2011         const struct tcp_sock *tp = tcp_sk(sk);
2012         const struct inet_connection_sock *icsk = inet_csk(sk);
2013         struct sk_buff *skb;
2014
2015         if (!sysctl_tcp_frto)
2016                 return 0;
2017
2018         /* MTU probe and F-RTO won't really play nicely along currently */
2019         if (icsk->icsk_mtup.probe_size)
2020                 return 0;
2021
2022         if (tcp_is_sackfrto(tp))
2023                 return 1;
2024
2025         /* Avoid expensive walking of rexmit queue if possible */
2026         if (tp->retrans_out > 1)
2027                 return 0;
2028
2029         skb = tcp_write_queue_head(sk);
2030         if (tcp_skb_is_last(sk, skb))
2031                 return 1;
2032         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
2033         tcp_for_write_queue_from(skb, sk) {
2034                 if (skb == tcp_send_head(sk))
2035                         break;
2036                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2037                         return 0;
2038                 /* Short-circuit when first non-SACKed skb has been checked */
2039                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2040                         break;
2041         }
2042         return 1;
2043 }
2044
2045 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2046  * recovery a bit and use heuristics in tcp_process_frto() to detect if
2047  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2048  * keep retrans_out counting accurate (with SACK F-RTO, other than head
2049  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2050  * bits are handled if the Loss state is really to be entered (in
2051  * tcp_enter_frto_loss).
2052  *
2053  * Do like tcp_enter_loss() would; when RTO expires the second time it
2054  * does:
2055  *  "Reduce ssthresh if it has not yet been made inside this window."
2056  */
2057 void tcp_enter_frto(struct sock *sk)
2058 {
2059         const struct inet_connection_sock *icsk = inet_csk(sk);
2060         struct tcp_sock *tp = tcp_sk(sk);
2061         struct sk_buff *skb;
2062
2063         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2064             tp->snd_una == tp->high_seq ||
2065             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2066              !icsk->icsk_retransmits)) {
2067                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2068                 /* Our state is too optimistic in ssthresh() call because cwnd
2069                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2070                  * recovery has not yet completed. Pattern would be this: RTO,
2071                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
2072                  * up here twice).
2073                  * RFC4138 should be more specific on what to do, even though
2074                  * RTO is quite unlikely to occur after the first Cumulative ACK
2075                  * due to back-off and complexity of triggering events ...
2076                  */
2077                 if (tp->frto_counter) {
2078                         u32 stored_cwnd;
2079                         stored_cwnd = tp->snd_cwnd;
2080                         tp->snd_cwnd = 2;
2081                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2082                         tp->snd_cwnd = stored_cwnd;
2083                 } else {
2084                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2085                 }
2086                 /* ... in theory, cong.control module could do "any tricks" in
2087                  * ssthresh(), which means that ca_state, lost bits and lost_out
2088                  * counter would have to be faked before the call occurs. We
2089                  * consider that too expensive, unlikely and hacky, so modules
2090                  * using these in ssthresh() must deal these incompatibility
2091                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2092                  */
2093                 tcp_ca_event(sk, CA_EVENT_FRTO);
2094         }
2095
2096         tp->undo_marker = tp->snd_una;
2097         tp->undo_retrans = 0;
2098
2099         skb = tcp_write_queue_head(sk);
2100         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2101                 tp->undo_marker = 0;
2102         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2103                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2104                 tp->retrans_out -= tcp_skb_pcount(skb);
2105         }
2106         tcp_verify_left_out(tp);
2107
2108         /* Too bad if TCP was application limited */
2109         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2110
2111         /* Earlier loss recovery underway (see RFC4138; Appendix B).
2112          * The last condition is necessary at least in tp->frto_counter case.
2113          */
2114         if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2115             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2116             after(tp->high_seq, tp->snd_una)) {
2117                 tp->frto_highmark = tp->high_seq;
2118         } else {
2119                 tp->frto_highmark = tp->snd_nxt;
2120         }
2121         tcp_set_ca_state(sk, TCP_CA_Disorder);
2122         tp->high_seq = tp->snd_nxt;
2123         tp->frto_counter = 1;
2124 }
2125
2126 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2127  * which indicates that we should follow the traditional RTO recovery,
2128  * i.e. mark everything lost and do go-back-N retransmission.
2129  */
2130 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2131 {
2132         struct tcp_sock *tp = tcp_sk(sk);
2133         struct sk_buff *skb;
2134
2135         tp->lost_out = 0;
2136         tp->retrans_out = 0;
2137         if (tcp_is_reno(tp))
2138                 tcp_reset_reno_sack(tp);
2139
2140         tcp_for_write_queue(skb, sk) {
2141                 if (skb == tcp_send_head(sk))
2142                         break;
2143
2144                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2145                 /*
2146                  * Count the retransmission made on RTO correctly (only when
2147                  * waiting for the first ACK and did not get it)...
2148                  */
2149                 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2150                         /* For some reason this R-bit might get cleared? */
2151                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2152                                 tp->retrans_out += tcp_skb_pcount(skb);
2153                         /* ...enter this if branch just for the first segment */
2154                         flag |= FLAG_DATA_ACKED;
2155                 } else {
2156                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2157                                 tp->undo_marker = 0;
2158                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2159                 }
2160
2161                 /* Marking forward transmissions that were made after RTO lost
2162                  * can cause unnecessary retransmissions in some scenarios,
2163                  * SACK blocks will mitigate that in some but not in all cases.
2164                  * We used to not mark them but it was causing break-ups with
2165                  * receivers that do only in-order receival.
2166                  *
2167                  * TODO: we could detect presence of such receiver and select
2168                  * different behavior per flow.
2169                  */
2170                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2171                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2172                         tp->lost_out += tcp_skb_pcount(skb);
2173                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2174                 }
2175         }
2176         tcp_verify_left_out(tp);
2177
2178         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2179         tp->snd_cwnd_cnt = 0;
2180         tp->snd_cwnd_stamp = tcp_time_stamp;
2181         tp->frto_counter = 0;
2182         tp->bytes_acked = 0;
2183
2184         tp->reordering = min_t(unsigned int, tp->reordering,
2185                                sysctl_tcp_reordering);
2186         tcp_set_ca_state(sk, TCP_CA_Loss);
2187         tp->high_seq = tp->snd_nxt;
2188         TCP_ECN_queue_cwr(tp);
2189
2190         tcp_clear_all_retrans_hints(tp);
2191 }
2192
2193 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2194 {
2195         tp->retrans_out = 0;
2196         tp->lost_out = 0;
2197
2198         tp->undo_marker = 0;
2199         tp->undo_retrans = 0;
2200 }
2201
2202 void tcp_clear_retrans(struct tcp_sock *tp)
2203 {
2204         tcp_clear_retrans_partial(tp);
2205
2206         tp->fackets_out = 0;
2207         tp->sacked_out = 0;
2208 }
2209
2210 /* Enter Loss state. If "how" is not zero, forget all SACK information
2211  * and reset tags completely, otherwise preserve SACKs. If receiver
2212  * dropped its ofo queue, we will know this due to reneging detection.
2213  */
2214 void tcp_enter_loss(struct sock *sk, int how)
2215 {
2216         const struct inet_connection_sock *icsk = inet_csk(sk);
2217         struct tcp_sock *tp = tcp_sk(sk);
2218         struct sk_buff *skb;
2219
2220         /* Reduce ssthresh if it has not yet been made inside this window. */
2221         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2222             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2223                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2224                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2225                 tcp_ca_event(sk, CA_EVENT_LOSS);
2226         }
2227         tp->snd_cwnd       = 1;
2228         tp->snd_cwnd_cnt   = 0;
2229         tp->snd_cwnd_stamp = tcp_time_stamp;
2230
2231         tp->bytes_acked = 0;
2232         tcp_clear_retrans_partial(tp);
2233
2234         if (tcp_is_reno(tp))
2235                 tcp_reset_reno_sack(tp);
2236
2237         if (!how) {
2238                 /* Push undo marker, if it was plain RTO and nothing
2239                  * was retransmitted. */
2240                 tp->undo_marker = tp->snd_una;
2241         } else {
2242                 tp->sacked_out = 0;
2243                 tp->fackets_out = 0;
2244         }
2245         tcp_clear_all_retrans_hints(tp);
2246
2247         tcp_for_write_queue(skb, sk) {
2248                 if (skb == tcp_send_head(sk))
2249                         break;
2250
2251                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2252                         tp->undo_marker = 0;
2253                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2254                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2255                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2256                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2257                         tp->lost_out += tcp_skb_pcount(skb);
2258                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2259                 }
2260         }
2261         tcp_verify_left_out(tp);
2262
2263         tp->reordering = min_t(unsigned int, tp->reordering,
2264                                sysctl_tcp_reordering);
2265         tcp_set_ca_state(sk, TCP_CA_Loss);
2266         tp->high_seq = tp->snd_nxt;
2267         TCP_ECN_queue_cwr(tp);
2268         /* Abort F-RTO algorithm if one is in progress */
2269         tp->frto_counter = 0;
2270 }
2271
2272 /* If ACK arrived pointing to a remembered SACK, it means that our
2273  * remembered SACKs do not reflect real state of receiver i.e.
2274  * receiver _host_ is heavily congested (or buggy).
2275  *
2276  * Do processing similar to RTO timeout.
2277  */
2278 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2279 {
2280         if (flag & FLAG_SACK_RENEGING) {
2281                 struct inet_connection_sock *icsk = inet_csk(sk);
2282                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2283
2284                 tcp_enter_loss(sk, 1);
2285                 icsk->icsk_retransmits++;
2286                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2287                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2288                                           icsk->icsk_rto, TCP_RTO_MAX);
2289                 return 1;
2290         }
2291         return 0;
2292 }
2293
2294 static inline int tcp_fackets_out(struct tcp_sock *tp)
2295 {
2296         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2297 }
2298
2299 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2300  * counter when SACK is enabled (without SACK, sacked_out is used for
2301  * that purpose).
2302  *
2303  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2304  * segments up to the highest received SACK block so far and holes in
2305  * between them.
2306  *
2307  * With reordering, holes may still be in flight, so RFC3517 recovery
2308  * uses pure sacked_out (total number of SACKed segments) even though
2309  * it violates the RFC that uses duplicate ACKs, often these are equal
2310  * but when e.g. out-of-window ACKs or packet duplication occurs,
2311  * they differ. Since neither occurs due to loss, TCP should really
2312  * ignore them.
2313  */
2314 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2315 {
2316         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2317 }
2318
2319 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2320 {
2321         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2322 }
2323
2324 static inline int tcp_head_timedout(struct sock *sk)
2325 {
2326         struct tcp_sock *tp = tcp_sk(sk);
2327
2328         return tp->packets_out &&
2329                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2330 }
2331
2332 /* Linux NewReno/SACK/FACK/ECN state machine.
2333  * --------------------------------------
2334  *
2335  * "Open"       Normal state, no dubious events, fast path.
2336  * "Disorder"   In all the respects it is "Open",
2337  *              but requires a bit more attention. It is entered when
2338  *              we see some SACKs or dupacks. It is split of "Open"
2339  *              mainly to move some processing from fast path to slow one.
2340  * "CWR"        CWND was reduced due to some Congestion Notification event.
2341  *              It can be ECN, ICMP source quench, local device congestion.
2342  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2343  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2344  *
2345  * tcp_fastretrans_alert() is entered:
2346  * - each incoming ACK, if state is not "Open"
2347  * - when arrived ACK is unusual, namely:
2348  *      * SACK
2349  *      * Duplicate ACK.
2350  *      * ECN ECE.
2351  *
2352  * Counting packets in flight is pretty simple.
2353  *
2354  *      in_flight = packets_out - left_out + retrans_out
2355  *
2356  *      packets_out is SND.NXT-SND.UNA counted in packets.
2357  *
2358  *      retrans_out is number of retransmitted segments.
2359  *
2360  *      left_out is number of segments left network, but not ACKed yet.
2361  *
2362  *              left_out = sacked_out + lost_out
2363  *
2364  *     sacked_out: Packets, which arrived to receiver out of order
2365  *                 and hence not ACKed. With SACKs this number is simply
2366  *                 amount of SACKed data. Even without SACKs
2367  *                 it is easy to give pretty reliable estimate of this number,
2368  *                 counting duplicate ACKs.
2369  *
2370  *       lost_out: Packets lost by network. TCP has no explicit
2371  *                 "loss notification" feedback from network (for now).
2372  *                 It means that this number can be only _guessed_.
2373  *                 Actually, it is the heuristics to predict lossage that
2374  *                 distinguishes different algorithms.
2375  *
2376  *      F.e. after RTO, when all the queue is considered as lost,
2377  *      lost_out = packets_out and in_flight = retrans_out.
2378  *
2379  *              Essentially, we have now two algorithms counting
2380  *              lost packets.
2381  *
2382  *              FACK: It is the simplest heuristics. As soon as we decided
2383  *              that something is lost, we decide that _all_ not SACKed
2384  *              packets until the most forward SACK are lost. I.e.
2385  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2386  *              It is absolutely correct estimate, if network does not reorder
2387  *              packets. And it loses any connection to reality when reordering
2388  *              takes place. We use FACK by default until reordering
2389  *              is suspected on the path to this destination.
2390  *
2391  *              NewReno: when Recovery is entered, we assume that one segment
2392  *              is lost (classic Reno). While we are in Recovery and
2393  *              a partial ACK arrives, we assume that one more packet
2394  *              is lost (NewReno). This heuristics are the same in NewReno
2395  *              and SACK.
2396  *
2397  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2398  *  deflation etc. CWND is real congestion window, never inflated, changes
2399  *  only according to classic VJ rules.
2400  *
2401  * Really tricky (and requiring careful tuning) part of algorithm
2402  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2403  * The first determines the moment _when_ we should reduce CWND and,
2404  * hence, slow down forward transmission. In fact, it determines the moment
2405  * when we decide that hole is caused by loss, rather than by a reorder.
2406  *
2407  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2408  * holes, caused by lost packets.
2409  *
2410  * And the most logically complicated part of algorithm is undo
2411  * heuristics. We detect false retransmits due to both too early
2412  * fast retransmit (reordering) and underestimated RTO, analyzing
2413  * timestamps and D-SACKs. When we detect that some segments were
2414  * retransmitted by mistake and CWND reduction was wrong, we undo
2415  * window reduction and abort recovery phase. This logic is hidden
2416  * inside several functions named tcp_try_undo_<something>.
2417  */
2418
2419 /* This function decides, when we should leave Disordered state
2420  * and enter Recovery phase, reducing congestion window.
2421  *
2422  * Main question: may we further continue forward transmission
2423  * with the same cwnd?
2424  */
2425 static int tcp_time_to_recover(struct sock *sk)
2426 {
2427         struct tcp_sock *tp = tcp_sk(sk);
2428         __u32 packets_out;
2429
2430         /* Do not perform any recovery during F-RTO algorithm */
2431         if (tp->frto_counter)
2432                 return 0;
2433
2434         /* Trick#1: The loss is proven. */
2435         if (tp->lost_out)
2436                 return 1;
2437
2438         /* Not-A-Trick#2 : Classic rule... */
2439         if (tcp_dupack_heurestics(tp) > tp->reordering)
2440                 return 1;
2441
2442         /* Trick#3 : when we use RFC2988 timer restart, fast
2443          * retransmit can be triggered by timeout of queue head.
2444          */
2445         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2446                 return 1;
2447
2448         /* Trick#4: It is still not OK... But will it be useful to delay
2449          * recovery more?
2450          */
2451         packets_out = tp->packets_out;
2452         if (packets_out <= tp->reordering &&
2453             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2454             !tcp_may_send_now(sk)) {
2455                 /* We have nothing to send. This connection is limited
2456                  * either by receiver window or by application.
2457                  */
2458                 return 1;
2459         }
2460
2461         return 0;
2462 }
2463
2464 /* New heuristics: it is possible only after we switched to restart timer
2465  * each time when something is ACKed. Hence, we can detect timed out packets
2466  * during fast retransmit without falling to slow start.
2467  *
2468  * Usefulness of this as is very questionable, since we should know which of
2469  * the segments is the next to timeout which is relatively expensive to find
2470  * in general case unless we add some data structure just for that. The
2471  * current approach certainly won't find the right one too often and when it
2472  * finally does find _something_ it usually marks large part of the window
2473  * right away (because a retransmission with a larger timestamp blocks the
2474  * loop from advancing). -ij
2475  */
2476 static void tcp_timeout_skbs(struct sock *sk)
2477 {
2478         struct tcp_sock *tp = tcp_sk(sk);
2479         struct sk_buff *skb;
2480
2481         if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2482                 return;
2483
2484         skb = tp->scoreboard_skb_hint;
2485         if (tp->scoreboard_skb_hint == NULL)
2486                 skb = tcp_write_queue_head(sk);
2487
2488         tcp_for_write_queue_from(skb, sk) {
2489                 if (skb == tcp_send_head(sk))
2490                         break;
2491                 if (!tcp_skb_timedout(sk, skb))
2492                         break;
2493
2494                 tcp_skb_mark_lost(tp, skb);
2495         }
2496
2497         tp->scoreboard_skb_hint = skb;
2498
2499         tcp_verify_left_out(tp);
2500 }
2501
2502 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2503  * is against sacked "cnt", otherwise it's against facked "cnt"
2504  */
2505 static void tcp_mark_head_lost(struct sock *sk, int packets)
2506 {
2507         struct tcp_sock *tp = tcp_sk(sk);
2508         struct sk_buff *skb;
2509         int cnt, oldcnt;
2510         int err;
2511         unsigned int mss;
2512
2513         WARN_ON(packets > tp->packets_out);
2514         if (tp->lost_skb_hint) {
2515                 skb = tp->lost_skb_hint;
2516                 cnt = tp->lost_cnt_hint;
2517         } else {
2518                 skb = tcp_write_queue_head(sk);
2519                 cnt = 0;
2520         }
2521
2522         tcp_for_write_queue_from(skb, sk) {
2523                 if (skb == tcp_send_head(sk))
2524                         break;
2525                 /* TODO: do this better */
2526                 /* this is not the most efficient way to do this... */
2527                 tp->lost_skb_hint = skb;
2528                 tp->lost_cnt_hint = cnt;
2529
2530                 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2531                         break;
2532
2533                 oldcnt = cnt;
2534                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2535                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2536                         cnt += tcp_skb_pcount(skb);
2537
2538                 if (cnt > packets) {
2539                         if (tcp_is_sack(tp) || (oldcnt >= packets))
2540                                 break;
2541
2542                         mss = skb_shinfo(skb)->gso_size;
2543                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2544                         if (err < 0)
2545                                 break;
2546                         cnt = packets;
2547                 }
2548
2549                 tcp_skb_mark_lost(tp, skb);
2550         }
2551         tcp_verify_left_out(tp);
2552 }
2553
2554 /* Account newly detected lost packet(s) */
2555
2556 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2557 {
2558         struct tcp_sock *tp = tcp_sk(sk);
2559
2560         if (tcp_is_reno(tp)) {
2561                 tcp_mark_head_lost(sk, 1);
2562         } else if (tcp_is_fack(tp)) {
2563                 int lost = tp->fackets_out - tp->reordering;
2564                 if (lost <= 0)
2565                         lost = 1;
2566                 tcp_mark_head_lost(sk, lost);
2567         } else {
2568                 int sacked_upto = tp->sacked_out - tp->reordering;
2569                 if (sacked_upto < fast_rexmit)
2570                         sacked_upto = fast_rexmit;
2571                 tcp_mark_head_lost(sk, sacked_upto);
2572         }
2573
2574         tcp_timeout_skbs(sk);
2575 }
2576
2577 /* CWND moderation, preventing bursts due to too big ACKs
2578  * in dubious situations.
2579  */
2580 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2581 {
2582         tp->snd_cwnd = min(tp->snd_cwnd,
2583                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2584         tp->snd_cwnd_stamp = tcp_time_stamp;
2585 }
2586
2587 /* Lower bound on congestion window is slow start threshold
2588  * unless congestion avoidance choice decides to overide it.
2589  */
2590 static inline u32 tcp_cwnd_min(const struct sock *sk)
2591 {
2592         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2593
2594         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2595 }
2596
2597 /* Decrease cwnd each second ack. */
2598 static void tcp_cwnd_down(struct sock *sk, int flag)
2599 {
2600         struct tcp_sock *tp = tcp_sk(sk);
2601         int decr = tp->snd_cwnd_cnt + 1;
2602
2603         if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2604             (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2605                 tp->snd_cwnd_cnt = decr & 1;
2606                 decr >>= 1;
2607
2608                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2609                         tp->snd_cwnd -= decr;
2610
2611                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2612                 tp->snd_cwnd_stamp = tcp_time_stamp;
2613         }
2614 }
2615
2616 /* Nothing was retransmitted or returned timestamp is less
2617  * than timestamp of the first retransmission.
2618  */
2619 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2620 {
2621         return !tp->retrans_stamp ||
2622                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2623                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2624 }
2625
2626 /* Undo procedures. */
2627
2628 #if FASTRETRANS_DEBUG > 1
2629 static void DBGUNDO(struct sock *sk, const char *msg)
2630 {
2631         struct tcp_sock *tp = tcp_sk(sk);
2632         struct inet_sock *inet = inet_sk(sk);
2633
2634         if (sk->sk_family == AF_INET) {
2635                 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2636                        msg,
2637                        &inet->daddr, ntohs(inet->dport),
2638                        tp->snd_cwnd, tcp_left_out(tp),
2639                        tp->snd_ssthresh, tp->prior_ssthresh,
2640                        tp->packets_out);
2641         }
2642 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2643         else if (sk->sk_family == AF_INET6) {
2644                 struct ipv6_pinfo *np = inet6_sk(sk);
2645                 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2646                        msg,
2647                        &np->daddr, ntohs(inet->dport),
2648                        tp->snd_cwnd, tcp_left_out(tp),
2649                        tp->snd_ssthresh, tp->prior_ssthresh,
2650                        tp->packets_out);
2651         }
2652 #endif
2653 }
2654 #else
2655 #define DBGUNDO(x...) do { } while (0)
2656 #endif
2657
2658 static void tcp_undo_cwr(struct sock *sk, const int undo)
2659 {
2660         struct tcp_sock *tp = tcp_sk(sk);
2661
2662         if (tp->prior_ssthresh) {
2663                 const struct inet_connection_sock *icsk = inet_csk(sk);
2664
2665                 if (icsk->icsk_ca_ops->undo_cwnd)
2666                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2667                 else
2668                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2669
2670                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2671                         tp->snd_ssthresh = tp->prior_ssthresh;
2672                         TCP_ECN_withdraw_cwr(tp);
2673                 }
2674         } else {
2675                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2676         }
2677         tcp_moderate_cwnd(tp);
2678         tp->snd_cwnd_stamp = tcp_time_stamp;
2679 }
2680
2681 static inline int tcp_may_undo(struct tcp_sock *tp)
2682 {
2683         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2684 }
2685
2686 /* People celebrate: "We love our President!" */
2687 static int tcp_try_undo_recovery(struct sock *sk)
2688 {
2689         struct tcp_sock *tp = tcp_sk(sk);
2690
2691         if (tcp_may_undo(tp)) {
2692                 int mib_idx;
2693
2694                 /* Happy end! We did not retransmit anything
2695                  * or our original transmission succeeded.
2696                  */
2697                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2698                 tcp_undo_cwr(sk, 1);
2699                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2700                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2701                 else
2702                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2703
2704                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2705                 tp->undo_marker = 0;
2706         }
2707         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2708                 /* Hold old state until something *above* high_seq
2709                  * is ACKed. For Reno it is MUST to prevent false
2710                  * fast retransmits (RFC2582). SACK TCP is safe. */
2711                 tcp_moderate_cwnd(tp);
2712                 return 1;
2713         }
2714         tcp_set_ca_state(sk, TCP_CA_Open);
2715         return 0;
2716 }
2717
2718 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2719 static void tcp_try_undo_dsack(struct sock *sk)
2720 {
2721         struct tcp_sock *tp = tcp_sk(sk);
2722
2723         if (tp->undo_marker && !tp->undo_retrans) {
2724                 DBGUNDO(sk, "D-SACK");
2725                 tcp_undo_cwr(sk, 1);
2726                 tp->undo_marker = 0;
2727                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2728         }
2729 }
2730
2731 /* Undo during fast recovery after partial ACK. */
2732
2733 static int tcp_try_undo_partial(struct sock *sk, int acked)
2734 {
2735         struct tcp_sock *tp = tcp_sk(sk);
2736         /* Partial ACK arrived. Force Hoe's retransmit. */
2737         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2738
2739         if (tcp_may_undo(tp)) {
2740                 /* Plain luck! Hole if filled with delayed
2741                  * packet, rather than with a retransmit.
2742                  */
2743                 if (tp->retrans_out == 0)
2744                         tp->retrans_stamp = 0;
2745
2746                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2747
2748                 DBGUNDO(sk, "Hoe");
2749                 tcp_undo_cwr(sk, 0);
2750                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2751
2752                 /* So... Do not make Hoe's retransmit yet.
2753                  * If the first packet was delayed, the rest
2754                  * ones are most probably delayed as well.
2755                  */
2756                 failed = 0;
2757         }
2758         return failed;
2759 }
2760
2761 /* Undo during loss recovery after partial ACK. */
2762 static int tcp_try_undo_loss(struct sock *sk)
2763 {
2764         struct tcp_sock *tp = tcp_sk(sk);
2765
2766         if (tcp_may_undo(tp)) {
2767                 struct sk_buff *skb;
2768                 tcp_for_write_queue(skb, sk) {
2769                         if (skb == tcp_send_head(sk))
2770                                 break;
2771                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2772                 }
2773
2774                 tcp_clear_all_retrans_hints(tp);
2775
2776                 DBGUNDO(sk, "partial loss");
2777                 tp->lost_out = 0;
2778                 tcp_undo_cwr(sk, 1);
2779                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2780                 inet_csk(sk)->icsk_retransmits = 0;
2781                 tp->undo_marker = 0;
2782                 if (tcp_is_sack(tp))
2783                         tcp_set_ca_state(sk, TCP_CA_Open);
2784                 return 1;
2785         }
2786         return 0;
2787 }
2788
2789 static inline void tcp_complete_cwr(struct sock *sk)
2790 {
2791         struct tcp_sock *tp = tcp_sk(sk);
2792         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2793         tp->snd_cwnd_stamp = tcp_time_stamp;
2794         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2795 }
2796
2797 static void tcp_try_keep_open(struct sock *sk)
2798 {
2799         struct tcp_sock *tp = tcp_sk(sk);
2800         int state = TCP_CA_Open;
2801
2802         if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2803                 state = TCP_CA_Disorder;
2804
2805         if (inet_csk(sk)->icsk_ca_state != state) {
2806                 tcp_set_ca_state(sk, state);
2807                 tp->high_seq = tp->snd_nxt;
2808         }
2809 }
2810
2811 static void tcp_try_to_open(struct sock *sk, int flag)
2812 {
2813         struct tcp_sock *tp = tcp_sk(sk);
2814
2815         tcp_verify_left_out(tp);
2816
2817         if (!tp->frto_counter && tp->retrans_out == 0)
2818                 tp->retrans_stamp = 0;
2819
2820         if (flag & FLAG_ECE)
2821                 tcp_enter_cwr(sk, 1);
2822
2823         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2824                 tcp_try_keep_open(sk);
2825                 tcp_moderate_cwnd(tp);
2826         } else {
2827                 tcp_cwnd_down(sk, flag);
2828         }
2829 }
2830
2831 static void tcp_mtup_probe_failed(struct sock *sk)
2832 {
2833         struct inet_connection_sock *icsk = inet_csk(sk);
2834
2835         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2836         icsk->icsk_mtup.probe_size = 0;
2837 }
2838
2839 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2840 {
2841         struct tcp_sock *tp = tcp_sk(sk);
2842         struct inet_connection_sock *icsk = inet_csk(sk);
2843
2844         /* FIXME: breaks with very large cwnd */
2845         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2846         tp->snd_cwnd = tp->snd_cwnd *
2847                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2848                        icsk->icsk_mtup.probe_size;
2849         tp->snd_cwnd_cnt = 0;
2850         tp->snd_cwnd_stamp = tcp_time_stamp;
2851         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2852
2853         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2854         icsk->icsk_mtup.probe_size = 0;
2855         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2856 }
2857
2858 /* Do a simple retransmit without using the backoff mechanisms in
2859  * tcp_timer. This is used for path mtu discovery.
2860  * The socket is already locked here.
2861  */
2862 void tcp_simple_retransmit(struct sock *sk)
2863 {
2864         const struct inet_connection_sock *icsk = inet_csk(sk);
2865         struct tcp_sock *tp = tcp_sk(sk);
2866         struct sk_buff *skb;
2867         unsigned int mss = tcp_current_mss(sk, 0);
2868         u32 prior_lost = tp->lost_out;
2869
2870         tcp_for_write_queue(skb, sk) {
2871                 if (skb == tcp_send_head(sk))
2872                         break;
2873                 if (tcp_skb_seglen(skb) > mss &&
2874                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2875                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2876                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2877                                 tp->retrans_out -= tcp_skb_pcount(skb);
2878                         }
2879                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2880                 }
2881         }
2882
2883         tcp_clear_retrans_hints_partial(tp);
2884
2885         if (prior_lost == tp->lost_out)
2886                 return;
2887
2888         if (tcp_is_reno(tp))
2889                 tcp_limit_reno_sacked(tp);
2890
2891         tcp_verify_left_out(tp);
2892
2893         /* Don't muck with the congestion window here.
2894          * Reason is that we do not increase amount of _data_
2895          * in network, but units changed and effective
2896          * cwnd/ssthresh really reduced now.
2897          */
2898         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2899                 tp->high_seq = tp->snd_nxt;
2900                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2901                 tp->prior_ssthresh = 0;
2902                 tp->undo_marker = 0;
2903                 tcp_set_ca_state(sk, TCP_CA_Loss);
2904         }
2905         tcp_xmit_retransmit_queue(sk);
2906 }
2907
2908 /* Process an event, which can update packets-in-flight not trivially.
2909  * Main goal of this function is to calculate new estimate for left_out,
2910  * taking into account both packets sitting in receiver's buffer and
2911  * packets lost by network.
2912  *
2913  * Besides that it does CWND reduction, when packet loss is detected
2914  * and changes state of machine.
2915  *
2916  * It does _not_ decide what to send, it is made in function
2917  * tcp_xmit_retransmit_queue().
2918  */
2919 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2920 {
2921         struct inet_connection_sock *icsk = inet_csk(sk);
2922         struct tcp_sock *tp = tcp_sk(sk);
2923         int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2924         int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2925                                     (tcp_fackets_out(tp) > tp->reordering));
2926         int fast_rexmit = 0, mib_idx;
2927
2928         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2929                 tp->sacked_out = 0;
2930         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2931                 tp->fackets_out = 0;
2932
2933         /* Now state machine starts.
2934          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2935         if (flag & FLAG_ECE)
2936                 tp->prior_ssthresh = 0;
2937
2938         /* B. In all the states check for reneging SACKs. */
2939         if (tcp_check_sack_reneging(sk, flag))
2940                 return;
2941
2942         /* C. Process data loss notification, provided it is valid. */
2943         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2944             before(tp->snd_una, tp->high_seq) &&
2945             icsk->icsk_ca_state != TCP_CA_Open &&
2946             tp->fackets_out > tp->reordering) {
2947                 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2948                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2949         }
2950
2951         /* D. Check consistency of the current state. */
2952         tcp_verify_left_out(tp);
2953
2954         /* E. Check state exit conditions. State can be terminated
2955          *    when high_seq is ACKed. */
2956         if (icsk->icsk_ca_state == TCP_CA_Open) {
2957                 WARN_ON(tp->retrans_out != 0);
2958                 tp->retrans_stamp = 0;
2959         } else if (!before(tp->snd_una, tp->high_seq)) {
2960                 switch (icsk->icsk_ca_state) {
2961                 case TCP_CA_Loss:
2962                         icsk->icsk_retransmits = 0;
2963                         if (tcp_try_undo_recovery(sk))
2964                                 return;
2965                         break;
2966
2967                 case TCP_CA_CWR:
2968                         /* CWR is to be held something *above* high_seq
2969                          * is ACKed for CWR bit to reach receiver. */
2970                         if (tp->snd_una != tp->high_seq) {
2971                                 tcp_complete_cwr(sk);
2972                                 tcp_set_ca_state(sk, TCP_CA_Open);
2973                         }
2974                         break;
2975
2976                 case TCP_CA_Disorder:
2977                         tcp_try_undo_dsack(sk);
2978                         if (!tp->undo_marker ||
2979                             /* For SACK case do not Open to allow to undo
2980                              * catching for all duplicate ACKs. */
2981                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2982                                 tp->undo_marker = 0;
2983                                 tcp_set_ca_state(sk, TCP_CA_Open);
2984                         }
2985                         break;
2986
2987                 case TCP_CA_Recovery:
2988                         if (tcp_is_reno(tp))
2989                                 tcp_reset_reno_sack(tp);
2990                         if (tcp_try_undo_recovery(sk))
2991                                 return;
2992                         tcp_complete_cwr(sk);
2993                         break;
2994                 }
2995         }
2996
2997         /* F. Process state. */
2998         switch (icsk->icsk_ca_state) {
2999         case TCP_CA_Recovery:
3000                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3001                         if (tcp_is_reno(tp) && is_dupack)
3002                                 tcp_add_reno_sack(sk);
3003                 } else
3004                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
3005                 break;
3006         case TCP_CA_Loss:
3007                 if (flag & FLAG_DATA_ACKED)
3008                         icsk->icsk_retransmits = 0;
3009                 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3010                         tcp_reset_reno_sack(tp);
3011                 if (!tcp_try_undo_loss(sk)) {
3012                         tcp_moderate_cwnd(tp);
3013                         tcp_xmit_retransmit_queue(sk);
3014                         return;
3015                 }
3016                 if (icsk->icsk_ca_state != TCP_CA_Open)
3017                         return;
3018                 /* Loss is undone; fall through to processing in Open state. */
3019         default:
3020                 if (tcp_is_reno(tp)) {
3021                         if (flag & FLAG_SND_UNA_ADVANCED)
3022                                 tcp_reset_reno_sack(tp);
3023                         if (is_dupack)
3024                                 tcp_add_reno_sack(sk);
3025                 }
3026
3027                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3028                         tcp_try_undo_dsack(sk);
3029
3030                 if (!tcp_time_to_recover(sk)) {
3031                         tcp_try_to_open(sk, flag);
3032                         return;
3033                 }
3034
3035                 /* MTU probe failure: don't reduce cwnd */
3036                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3037                     icsk->icsk_mtup.probe_size &&
3038                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
3039                         tcp_mtup_probe_failed(sk);
3040                         /* Restores the reduction we did in tcp_mtup_probe() */
3041                         tp->snd_cwnd++;
3042                         tcp_simple_retransmit(sk);
3043                         return;
3044                 }
3045
3046                 /* Otherwise enter Recovery state */
3047
3048                 if (tcp_is_reno(tp))
3049                         mib_idx = LINUX_MIB_TCPRENORECOVERY;
3050                 else
3051                         mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3052
3053                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3054
3055                 tp->high_seq = tp->snd_nxt;
3056                 tp->prior_ssthresh = 0;
3057                 tp->undo_marker = tp->snd_una;
3058                 tp->undo_retrans = tp->retrans_out;
3059
3060                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3061                         if (!(flag & FLAG_ECE))
3062                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3063                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3064                         TCP_ECN_queue_cwr(tp);
3065                 }
3066
3067                 tp->bytes_acked = 0;
3068                 tp->snd_cwnd_cnt = 0;
3069                 tcp_set_ca_state(sk, TCP_CA_Recovery);
3070                 fast_rexmit = 1;
3071         }
3072
3073         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3074                 tcp_update_scoreboard(sk, fast_rexmit);
3075         tcp_cwnd_down(sk, flag);
3076         tcp_xmit_retransmit_queue(sk);
3077 }
3078
3079 static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3080 {
3081         tcp_rtt_estimator(sk, seq_rtt);
3082         tcp_set_rto(sk);
3083         inet_csk(sk)->icsk_backoff = 0;
3084 }
3085
3086 /* Read draft-ietf-tcplw-high-performance before mucking
3087  * with this code. (Supersedes RFC1323)
3088  */
3089 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3090 {
3091         /* RTTM Rule: A TSecr value received in a segment is used to
3092          * update the averaged RTT measurement only if the segment
3093          * acknowledges some new data, i.e., only if it advances the
3094          * left edge of the send window.
3095          *
3096          * See draft-ietf-tcplw-high-performance-00, section 3.3.
3097          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3098          *
3099          * Changed: reset backoff as soon as we see the first valid sample.
3100          * If we do not, we get strongly overestimated rto. With timestamps
3101          * samples are accepted even from very old segments: f.e., when rtt=1
3102          * increases to 8, we retransmit 5 times and after 8 seconds delayed
3103          * answer arrives rto becomes 120 seconds! If at least one of segments
3104          * in window is lost... Voila.                          --ANK (010210)
3105          */
3106         struct tcp_sock *tp = tcp_sk(sk);
3107
3108         tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3109 }
3110
3111 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3112 {
3113         /* We don't have a timestamp. Can only use
3114          * packets that are not retransmitted to determine
3115          * rtt estimates. Also, we must not reset the
3116          * backoff for rto until we get a non-retransmitted
3117          * packet. This allows us to deal with a situation
3118          * where the network delay has increased suddenly.
3119          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3120          */
3121
3122         if (flag & FLAG_RETRANS_DATA_ACKED)
3123                 return;
3124
3125         tcp_valid_rtt_meas(sk, seq_rtt);
3126 }
3127
3128 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3129                                       const s32 seq_rtt)
3130 {
3131         const struct tcp_sock *tp = tcp_sk(sk);
3132         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3133         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3134                 tcp_ack_saw_tstamp(sk, flag);
3135         else if (seq_rtt >= 0)
3136                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3137 }
3138
3139 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3140 {
3141         const struct inet_connection_sock *icsk = inet_csk(sk);
3142         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3143         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3144 }
3145
3146 /* Restart timer after forward progress on connection.
3147  * RFC2988 recommends to restart timer to now+rto.
3148  */
3149 static void tcp_rearm_rto(struct sock *sk)
3150 {
3151         struct tcp_sock *tp = tcp_sk(sk);
3152
3153         if (!tp->packets_out) {
3154                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3155         } else {
3156                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3157                                           inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3158         }
3159 }
3160
3161 /* If we get here, the whole TSO packet has not been acked. */
3162 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3163 {
3164         struct tcp_sock *tp = tcp_sk(sk);
3165         u32 packets_acked;
3166
3167         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3168
3169         packets_acked = tcp_skb_pcount(skb);
3170         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3171                 return 0;
3172         packets_acked -= tcp_skb_pcount(skb);
3173
3174         if (packets_acked) {
3175                 BUG_ON(tcp_skb_pcount(skb) == 0);
3176                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3177         }
3178
3179         return packets_acked;
3180 }
3181
3182 /* Remove acknowledged frames from the retransmission queue. If our packet
3183  * is before the ack sequence we can discard it as it's confirmed to have
3184  * arrived at the other end.
3185  */
3186 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3187                                u32 prior_snd_una)
3188 {
3189         struct tcp_sock *tp = tcp_sk(sk);
3190         const struct inet_connection_sock *icsk = inet_csk(sk);
3191         struct sk_buff *skb;
3192         u32 now = tcp_time_stamp;
3193         int fully_acked = 1;
3194         int flag = 0;
3195         u32 pkts_acked = 0;
3196         u32 reord = tp->packets_out;
3197         u32 prior_sacked = tp->sacked_out;
3198         s32 seq_rtt = -1;
3199         s32 ca_seq_rtt = -1;
3200         ktime_t last_ackt = net_invalid_timestamp();
3201
3202         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3203                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3204                 u32 acked_pcount;
3205                 u8 sacked = scb->sacked;
3206
3207                 /* Determine how many packets and what bytes were acked, tso and else */
3208                 if (after(scb->end_seq, tp->snd_una)) {
3209                         if (tcp_skb_pcount(skb) == 1 ||
3210                             !after(tp->snd_una, scb->seq))
3211                                 break;
3212
3213                         acked_pcount = tcp_tso_acked(sk, skb);
3214                         if (!acked_pcount)
3215                                 break;
3216
3217                         fully_acked = 0;
3218                 } else {
3219                         acked_pcount = tcp_skb_pcount(skb);
3220                 }
3221
3222                 /* MTU probing checks */
3223                 if (fully_acked && icsk->icsk_mtup.probe_size &&
3224                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
3225                         tcp_mtup_probe_success(sk, skb);
3226                 }
3227
3228                 if (sacked & TCPCB_RETRANS) {
3229                         if (sacked & TCPCB_SACKED_RETRANS)
3230                                 tp->retrans_out -= acked_pcount;
3231                         flag |= FLAG_RETRANS_DATA_ACKED;
3232                         ca_seq_rtt = -1;
3233                         seq_rtt = -1;
3234                         if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3235                                 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3236                 } else {
3237                         ca_seq_rtt = now - scb->when;
3238                         last_ackt = skb->tstamp;
3239                         if (seq_rtt < 0) {
3240                                 seq_rtt = ca_seq_rtt;
3241                         }
3242                         if (!(sacked & TCPCB_SACKED_ACKED))
3243                                 reord = min(pkts_acked, reord);
3244                 }
3245
3246                 if (sacked & TCPCB_SACKED_ACKED)
3247                         tp->sacked_out -= acked_pcount;
3248                 if (sacked & TCPCB_LOST)
3249                         tp->lost_out -= acked_pcount;
3250
3251                 tp->packets_out -= acked_pcount;
3252                 pkts_acked += acked_pcount;
3253
3254                 /* Initial outgoing SYN's get put onto the write_queue
3255                  * just like anything else we transmit.  It is not
3256                  * true data, and if we misinform our callers that
3257                  * this ACK acks real data, we will erroneously exit
3258                  * connection startup slow start one packet too
3259                  * quickly.  This is severely frowned upon behavior.
3260                  */
3261                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3262                         flag |= FLAG_DATA_ACKED;
3263                 } else {
3264                         flag |= FLAG_SYN_ACKED;
3265                         tp->retrans_stamp = 0;
3266                 }
3267
3268                 if (!fully_acked)
3269                         break;
3270
3271                 tcp_unlink_write_queue(skb, sk);
3272                 sk_wmem_free_skb(sk, skb);
3273                 tp->scoreboard_skb_hint = NULL;
3274                 if (skb == tp->retransmit_skb_hint)
3275                         tp->retransmit_skb_hint = NULL;
3276                 if (skb == tp->lost_skb_hint)
3277                         tp->lost_skb_hint = NULL;
3278         }
3279
3280         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3281                 tp->snd_up = tp->snd_una;
3282
3283         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3284                 flag |= FLAG_SACK_RENEGING;
3285
3286         if (flag & FLAG_ACKED) {
3287                 const struct tcp_congestion_ops *ca_ops
3288                         = inet_csk(sk)->icsk_ca_ops;
3289
3290                 tcp_ack_update_rtt(sk, flag, seq_rtt);
3291                 tcp_rearm_rto(sk);
3292
3293                 if (tcp_is_reno(tp)) {
3294                         tcp_remove_reno_sacks(sk, pkts_acked);
3295                 } else {
3296                         int delta;
3297
3298                         /* Non-retransmitted hole got filled? That's reordering */
3299                         if (reord < prior_fackets)
3300                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3301
3302                         delta = tcp_is_fack(tp) ? pkts_acked :
3303                                                   prior_sacked - tp->sacked_out;
3304                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3305                 }
3306
3307                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3308
3309                 if (ca_ops->pkts_acked) {
3310                         s32 rtt_us = -1;
3311
3312                         /* Is the ACK triggering packet unambiguous? */
3313                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3314                                 /* High resolution needed and available? */
3315                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3316                                     !ktime_equal(last_ackt,
3317                                                  net_invalid_timestamp()))
3318                                         rtt_us = ktime_us_delta(ktime_get_real(),
3319                                                                 last_ackt);
3320                                 else if (ca_seq_rtt > 0)
3321                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
3322                         }
3323
3324                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3325                 }
3326         }
3327
3328 #if FASTRETRANS_DEBUG > 0
3329         WARN_ON((int)tp->sacked_out < 0);
3330         WARN_ON((int)tp->lost_out < 0);
3331         WARN_ON((int)tp->retrans_out < 0);
3332         if (!tp->packets_out && tcp_is_sack(tp)) {
3333                 icsk = inet_csk(sk);
3334                 if (tp->lost_out) {
3335                         printk(KERN_DEBUG "Leak l=%u %d\n",
3336                                tp->lost_out, icsk->icsk_ca_state);
3337                         tp->lost_out = 0;
3338                 }
3339                 if (tp->sacked_out) {
3340                         printk(KERN_DEBUG "Leak s=%u %d\n",
3341                                tp->sacked_out, icsk->icsk_ca_state);
3342                         tp->sacked_out = 0;
3343                 }
3344                 if (tp->retrans_out) {
3345                         printk(KERN_DEBUG "Leak r=%u %d\n",
3346                                tp->retrans_out, icsk->icsk_ca_state);
3347                         tp->retrans_out = 0;
3348                 }
3349         }
3350 #endif
3351         return flag;
3352 }
3353
3354 static void tcp_ack_probe(struct sock *sk)
3355 {
3356         const struct tcp_sock *tp = tcp_sk(sk);
3357         struct inet_connection_sock *icsk = inet_csk(sk);
3358
3359         /* Was it a usable window open? */
3360
3361         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3362                 icsk->icsk_backoff = 0;
3363                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3364                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3365                  * This function is not for random using!
3366                  */
3367         } else {
3368                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3369                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3370                                           TCP_RTO_MAX);
3371         }
3372 }
3373
3374 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3375 {
3376         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3377                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3378 }
3379
3380 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3381 {
3382         const struct tcp_sock *tp = tcp_sk(sk);
3383         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3384                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3385 }
3386
3387 /* Check that window update is acceptable.
3388  * The function assumes that snd_una<=ack<=snd_next.
3389  */
3390 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3391                                         const u32 ack, const u32 ack_seq,
3392                                         const u32 nwin)
3393 {
3394         return (after(ack, tp->snd_una) ||
3395                 after(ack_seq, tp->snd_wl1) ||
3396                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3397 }
3398
3399 /* Update our send window.
3400  *
3401  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3402  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3403  */
3404 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3405                                  u32 ack_seq)
3406 {
3407         struct tcp_sock *tp = tcp_sk(sk);
3408         int flag = 0;
3409         u32 nwin = ntohs(tcp_hdr(skb)->window);
3410
3411         if (likely(!tcp_hdr(skb)->syn))
3412                 nwin <<= tp->rx_opt.snd_wscale;
3413
3414         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3415                 flag |= FLAG_WIN_UPDATE;
3416                 tcp_update_wl(tp, ack_seq);
3417
3418                 if (tp->snd_wnd != nwin) {
3419                         tp->snd_wnd = nwin;
3420
3421                         /* Note, it is the only place, where
3422                          * fast path is recovered for sending TCP.
3423                          */
3424                         tp->pred_flags = 0;
3425                         tcp_fast_path_check(sk);
3426
3427                         if (nwin > tp->max_window) {
3428                                 tp->max_window = nwin;
3429                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3430                         }
3431                 }
3432         }
3433
3434         tp->snd_una = ack;
3435
3436         return flag;
3437 }
3438
3439 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3440  * continue in congestion avoidance.
3441  */
3442 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3443 {
3444         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3445         tp->snd_cwnd_cnt = 0;
3446         tp->bytes_acked = 0;
3447         TCP_ECN_queue_cwr(tp);
3448         tcp_moderate_cwnd(tp);
3449 }
3450
3451 /* A conservative spurious RTO response algorithm: reduce cwnd using
3452  * rate halving and continue in congestion avoidance.
3453  */
3454 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3455 {
3456         tcp_enter_cwr(sk, 0);
3457 }
3458
3459 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3460 {
3461         if (flag & FLAG_ECE)
3462                 tcp_ratehalving_spur_to_response(sk);
3463         else
3464                 tcp_undo_cwr(sk, 1);
3465 }
3466
3467 /* F-RTO spurious RTO detection algorithm (RFC4138)
3468  *
3469  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3470  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3471  * window (but not to or beyond highest sequence sent before RTO):
3472  *   On First ACK,  send two new segments out.
3473  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3474  *                  algorithm is not part of the F-RTO detection algorithm
3475  *                  given in RFC4138 but can be selected separately).
3476  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3477  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3478  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3479  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3480  *
3481  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3482  * original window even after we transmit two new data segments.
3483  *
3484  * SACK version:
3485  *   on first step, wait until first cumulative ACK arrives, then move to
3486  *   the second step. In second step, the next ACK decides.
3487  *
3488  * F-RTO is implemented (mainly) in four functions:
3489  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3490  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3491  *     called when tcp_use_frto() showed green light
3492  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3493  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3494  *     to prove that the RTO is indeed spurious. It transfers the control
3495  *     from F-RTO to the conventional RTO recovery
3496  */
3497 static int tcp_process_frto(struct sock *sk, int flag)
3498 {
3499         struct tcp_sock *tp = tcp_sk(sk);
3500
3501         tcp_verify_left_out(tp);
3502
3503         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3504         if (flag & FLAG_DATA_ACKED)
3505                 inet_csk(sk)->icsk_retransmits = 0;
3506
3507         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3508             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3509                 tp->undo_marker = 0;
3510
3511         if (!before(tp->snd_una, tp->frto_highmark)) {
3512                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3513                 return 1;
3514         }
3515
3516         if (!tcp_is_sackfrto(tp)) {
3517                 /* RFC4138 shortcoming in step 2; should also have case c):
3518                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3519                  * data, winupdate
3520                  */
3521                 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3522                         return 1;
3523
3524                 if (!(flag & FLAG_DATA_ACKED)) {
3525                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3526                                             flag);
3527                         return 1;
3528                 }
3529         } else {
3530                 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3531                         /* Prevent sending of new data. */
3532                         tp->snd_cwnd = min(tp->snd_cwnd,
3533                                            tcp_packets_in_flight(tp));
3534                         return 1;
3535                 }
3536
3537                 if ((tp->frto_counter >= 2) &&
3538                     (!(flag & FLAG_FORWARD_PROGRESS) ||
3539                      ((flag & FLAG_DATA_SACKED) &&
3540                       !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3541                         /* RFC4138 shortcoming (see comment above) */
3542                         if (!(flag & FLAG_FORWARD_PROGRESS) &&
3543                             (flag & FLAG_NOT_DUP))
3544                                 return 1;
3545
3546                         tcp_enter_frto_loss(sk, 3, flag);
3547                         return 1;
3548                 }
3549         }
3550
3551         if (tp->frto_counter == 1) {
3552                 /* tcp_may_send_now needs to see updated state */
3553                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3554                 tp->frto_counter = 2;
3555
3556                 if (!tcp_may_send_now(sk))
3557                         tcp_enter_frto_loss(sk, 2, flag);
3558
3559                 return 1;
3560         } else {
3561                 switch (sysctl_tcp_frto_response) {
3562                 case 2:
3563                         tcp_undo_spur_to_response(sk, flag);
3564                         break;
3565                 case 1:
3566                         tcp_conservative_spur_to_response(tp);
3567                         break;
3568                 default:
3569                         tcp_ratehalving_spur_to_response(sk);
3570                         break;
3571                 }
3572                 tp->frto_counter = 0;
3573                 tp->undo_marker = 0;
3574                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3575         }
3576         return 0;
3577 }
3578
3579 /* This routine deals with incoming acks, but not outgoing ones. */
3580 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3581 {
3582         struct inet_connection_sock *icsk = inet_csk(sk);
3583         struct tcp_sock *tp = tcp_sk(sk);
3584         u32 prior_snd_una = tp->snd_una;
3585         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3586         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3587         u32 prior_in_flight;
3588         u32 prior_fackets;
3589         int prior_packets;
3590         int frto_cwnd = 0;
3591
3592         /* If the ack is newer than sent or older than previous acks
3593          * then we can probably ignore it.
3594          */
3595         if (after(ack, tp->snd_nxt))
3596                 goto uninteresting_ack;
3597
3598         if (before(ack, prior_snd_una))
3599                 goto old_ack;
3600
3601         if (after(ack, prior_snd_una))
3602                 flag |= FLAG_SND_UNA_ADVANCED;
3603
3604         if (sysctl_tcp_abc) {
3605                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3606                         tp->bytes_acked += ack - prior_snd_una;
3607                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3608                         /* we assume just one segment left network */
3609                         tp->bytes_acked += min(ack - prior_snd_una,
3610                                                tp->mss_cache);
3611         }
3612
3613         prior_fackets = tp->fackets_out;
3614         prior_in_flight = tcp_packets_in_flight(tp);
3615
3616         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3617                 /* Window is constant, pure forward advance.
3618                  * No more checks are required.
3619                  * Note, we use the fact that SND.UNA>=SND.WL2.
3620                  */
3621                 tcp_update_wl(tp, ack_seq);
3622                 tp->snd_una = ack;
3623                 flag |= FLAG_WIN_UPDATE;
3624
3625                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3626
3627                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3628         } else {
3629                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3630                         flag |= FLAG_DATA;
3631                 else
3632                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3633
3634                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3635
3636                 if (TCP_SKB_CB(skb)->sacked)
3637                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3638
3639                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3640                         flag |= FLAG_ECE;
3641
3642                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3643         }
3644
3645         /* We passed data and got it acked, remove any soft error
3646          * log. Something worked...
3647          */
3648         sk->sk_err_soft = 0;
3649         icsk->icsk_probes_out = 0;
3650         tp->rcv_tstamp = tcp_time_stamp;
3651         prior_packets = tp->packets_out;
3652         if (!prior_packets)
3653                 goto no_queue;
3654
3655         /* See if we can take anything off of the retransmit queue. */
3656         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3657
3658         if (tp->frto_counter)
3659                 frto_cwnd = tcp_process_frto(sk, flag);
3660         /* Guarantee sacktag reordering detection against wrap-arounds */
3661         if (before(tp->frto_highmark, tp->snd_una))
3662                 tp->frto_highmark = 0;
3663
3664         if (tcp_ack_is_dubious(sk, flag)) {
3665                 /* Advance CWND, if state allows this. */
3666                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3667                     tcp_may_raise_cwnd(sk, flag))
3668                         tcp_cong_avoid(sk, ack, prior_in_flight);
3669                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3670                                       flag);
3671         } else {
3672                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3673                         tcp_cong_avoid(sk, ack, prior_in_flight);
3674         }
3675
3676         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3677                 dst_confirm(sk->sk_dst_cache);
3678
3679         return 1;
3680
3681 no_queue:
3682         /* If this ack opens up a zero window, clear backoff.  It was
3683          * being used to time the probes, and is probably far higher than
3684          * it needs to be for normal retransmission.
3685          */
3686         if (tcp_send_head(sk))
3687                 tcp_ack_probe(sk);
3688         return 1;
3689
3690 old_ack:
3691         if (TCP_SKB_CB(skb)->sacked) {
3692                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3693                 if (icsk->icsk_ca_state == TCP_CA_Open)
3694                         tcp_try_keep_open(sk);
3695         }
3696
3697 uninteresting_ack:
3698         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3699         return 0;
3700 }
3701
3702 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3703  * But, this can also be called on packets in the established flow when
3704  * the fast version below fails.
3705  */
3706 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3707                        int estab)
3708 {
3709         unsigned char *ptr;
3710         struct tcphdr *th = tcp_hdr(skb);
3711         int length = (th->doff * 4) - sizeof(struct tcphdr);
3712
3713         ptr = (unsigned char *)(th + 1);
3714         opt_rx->saw_tstamp = 0;
3715
3716         while (length > 0) {
3717                 int opcode = *ptr++;
3718                 int opsize;
3719
3720                 switch (opcode) {
3721                 case TCPOPT_EOL:
3722                         return;
3723                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3724                         length--;
3725                         continue;
3726                 default:
3727                         opsize = *ptr++;
3728                         if (opsize < 2) /* "silly options" */
3729                                 return;
3730                         if (opsize > length)
3731                                 return; /* don't parse partial options */
3732                         switch (opcode) {
3733                         case TCPOPT_MSS:
3734                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3735                                         u16 in_mss = get_unaligned_be16(ptr);
3736                                         if (in_mss) {
3737                                                 if (opt_rx->user_mss &&
3738                                                     opt_rx->user_mss < in_mss)
3739                                                         in_mss = opt_rx->user_mss;
3740                                                 opt_rx->mss_clamp = in_mss;
3741                                         }
3742                                 }
3743                                 break;
3744                         case TCPOPT_WINDOW:
3745                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3746                                     !estab && sysctl_tcp_window_scaling) {
3747                                         __u8 snd_wscale = *(__u8 *)ptr;
3748                                         opt_rx->wscale_ok = 1;
3749                                         if (snd_wscale > 14) {
3750                                                 if (net_ratelimit())
3751                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3752                                                                "scaling value %d >14 received.\n",
3753                                                                snd_wscale);
3754                                                 snd_wscale = 14;
3755                                         }
3756                                         opt_rx->snd_wscale = snd_wscale;
3757                                 }
3758                                 break;
3759                         case TCPOPT_TIMESTAMP:
3760                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3761                                     ((estab && opt_rx->tstamp_ok) ||
3762                                      (!estab && sysctl_tcp_timestamps))) {
3763                                         opt_rx->saw_tstamp = 1;
3764                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3765                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3766                                 }
3767                                 break;
3768                         case TCPOPT_SACK_PERM:
3769                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3770                                     !estab && sysctl_tcp_sack) {
3771                                         opt_rx->sack_ok = 1;
3772                                         tcp_sack_reset(opt_rx);
3773                                 }
3774                                 break;
3775
3776                         case TCPOPT_SACK:
3777                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3778                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3779                                    opt_rx->sack_ok) {
3780                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3781                                 }
3782                                 break;
3783 #ifdef CONFIG_TCP_MD5SIG
3784                         case TCPOPT_MD5SIG:
3785                                 /*
3786                                  * The MD5 Hash has already been
3787                                  * checked (see tcp_v{4,6}_do_rcv()).
3788                                  */
3789                                 break;
3790 #endif
3791                         }
3792
3793                         ptr += opsize-2;
3794                         length -= opsize;
3795                 }
3796         }
3797 }
3798
3799 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3800 {
3801         __be32 *ptr = (__be32 *)(th + 1);
3802
3803         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3804                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3805                 tp->rx_opt.saw_tstamp = 1;
3806                 ++ptr;
3807                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3808                 ++ptr;
3809                 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3810                 return 1;
3811         }
3812         return 0;
3813 }
3814
3815 /* Fast parse options. This hopes to only see timestamps.
3816  * If it is wrong it falls back on tcp_parse_options().
3817  */
3818 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3819                                   struct tcp_sock *tp)
3820 {
3821         if (th->doff == sizeof(struct tcphdr) >> 2) {
3822                 tp->rx_opt.saw_tstamp = 0;
3823                 return 0;
3824         } else if (tp->rx_opt.tstamp_ok &&
3825                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3826                 if (tcp_parse_aligned_timestamp(tp, th))
3827                         return 1;
3828         }
3829         tcp_parse_options(skb, &tp->rx_opt, 1);
3830         return 1;
3831 }
3832
3833 #ifdef CONFIG_TCP_MD5SIG
3834 /*
3835  * Parse MD5 Signature option
3836  */
3837 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3838 {
3839         int length = (th->doff << 2) - sizeof (*th);
3840         u8 *ptr = (u8*)(th + 1);
3841
3842         /* If the TCP option is too short, we can short cut */
3843         if (length < TCPOLEN_MD5SIG)
3844                 return NULL;
3845
3846         while (length > 0) {
3847                 int opcode = *ptr++;
3848                 int opsize;
3849
3850                 switch(opcode) {
3851                 case TCPOPT_EOL:
3852                         return NULL;
3853                 case TCPOPT_NOP:
3854                         length--;
3855                         continue;
3856                 default:
3857                         opsize = *ptr++;
3858                         if (opsize < 2 || opsize > length)
3859                                 return NULL;
3860                         if (opcode == TCPOPT_MD5SIG)
3861                                 return ptr;
3862                 }
3863                 ptr += opsize - 2;
3864                 length -= opsize;
3865         }
3866         return NULL;
3867 }
3868 #endif
3869
3870 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3871 {
3872         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3873         tp->rx_opt.ts_recent_stamp = get_seconds();
3874 }
3875
3876 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3877 {
3878         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3879                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3880                  * extra check below makes sure this can only happen
3881                  * for pure ACK frames.  -DaveM
3882                  *
3883                  * Not only, also it occurs for expired timestamps.
3884                  */
3885
3886                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3887                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3888                         tcp_store_ts_recent(tp);
3889         }
3890 }
3891
3892 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3893  *
3894  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3895  * it can pass through stack. So, the following predicate verifies that
3896  * this segment is not used for anything but congestion avoidance or
3897  * fast retransmit. Moreover, we even are able to eliminate most of such
3898  * second order effects, if we apply some small "replay" window (~RTO)
3899  * to timestamp space.
3900  *
3901  * All these measures still do not guarantee that we reject wrapped ACKs
3902  * on networks with high bandwidth, when sequence space is recycled fastly,
3903  * but it guarantees that such events will be very rare and do not affect
3904  * connection seriously. This doesn't look nice, but alas, PAWS is really
3905  * buggy extension.
3906  *
3907  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3908  * states that events when retransmit arrives after original data are rare.
3909  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3910  * the biggest problem on large power networks even with minor reordering.
3911  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3912  * up to bandwidth of 18Gigabit/sec. 8) ]
3913  */
3914
3915 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3916 {
3917         struct tcp_sock *tp = tcp_sk(sk);
3918         struct tcphdr *th = tcp_hdr(skb);
3919         u32 seq = TCP_SKB_CB(skb)->seq;
3920         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3921
3922         return (/* 1. Pure ACK with correct sequence number. */
3923                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3924
3925                 /* 2. ... and duplicate ACK. */
3926                 ack == tp->snd_una &&
3927
3928                 /* 3. ... and does not update window. */
3929                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3930
3931                 /* 4. ... and sits in replay window. */
3932                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3933 }
3934
3935 static inline int tcp_paws_discard(const struct sock *sk,
3936                                    const struct sk_buff *skb)
3937 {
3938         const struct tcp_sock *tp = tcp_sk(sk);
3939         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3940                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3941                 !tcp_disordered_ack(sk, skb));
3942 }
3943
3944 /* Check segment sequence number for validity.
3945  *
3946  * Segment controls are considered valid, if the segment
3947  * fits to the window after truncation to the window. Acceptability
3948  * of data (and SYN, FIN, of course) is checked separately.
3949  * See tcp_data_queue(), for example.
3950  *
3951  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3952  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3953  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3954  * (borrowed from freebsd)
3955  */
3956
3957 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3958 {
3959         return  !before(end_seq, tp->rcv_wup) &&
3960                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3961 }
3962
3963 /* When we get a reset we do this. */
3964 static void tcp_reset(struct sock *sk)
3965 {
3966         /* We want the right error as BSD sees it (and indeed as we do). */
3967         switch (sk->sk_state) {
3968         case TCP_SYN_SENT:
3969                 sk->sk_err = ECONNREFUSED;
3970                 break;
3971         case TCP_CLOSE_WAIT:
3972                 sk->sk_err = EPIPE;
3973                 break;
3974         case TCP_CLOSE:
3975                 return;
3976         default:
3977                 sk->sk_err = ECONNRESET;
3978         }
3979
3980         if (!sock_flag(sk, SOCK_DEAD))
3981                 sk->sk_error_report(sk);
3982
3983         tcp_done(sk);
3984 }
3985
3986 /*
3987  *      Process the FIN bit. This now behaves as it is supposed to work
3988  *      and the FIN takes effect when it is validly part of sequence
3989  *      space. Not before when we get holes.
3990  *
3991  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3992  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3993  *      TIME-WAIT)
3994  *
3995  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3996  *      close and we go into CLOSING (and later onto TIME-WAIT)
3997  *
3998  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3999  */
4000 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4001 {
4002         struct tcp_sock *tp = tcp_sk(sk);
4003
4004         inet_csk_schedule_ack(sk);
4005
4006         sk->sk_shutdown |= RCV_SHUTDOWN;
4007         sock_set_flag(sk, SOCK_DONE);
4008
4009         switch (sk->sk_state) {
4010         case TCP_SYN_RECV:
4011         case TCP_ESTABLISHED:
4012                 /* Move to CLOSE_WAIT */
4013                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4014                 inet_csk(sk)->icsk_ack.pingpong = 1;
4015                 break;
4016
4017         case TCP_CLOSE_WAIT:
4018         case TCP_CLOSING:
4019                 /* Received a retransmission of the FIN, do
4020                  * nothing.
4021                  */
4022                 break;
4023         case TCP_LAST_ACK:
4024                 /* RFC793: Remain in the LAST-ACK state. */
4025                 break;
4026
4027         case TCP_FIN_WAIT1:
4028                 /* This case occurs when a simultaneous close
4029                  * happens, we must ack the received FIN and
4030                  * enter the CLOSING state.
4031                  */
4032                 tcp_send_ack(sk);
4033                 tcp_set_state(sk, TCP_CLOSING);
4034                 break;
4035         case TCP_FIN_WAIT2:
4036                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4037                 tcp_send_ack(sk);
4038                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4039                 break;
4040         default:
4041                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4042                  * cases we should never reach this piece of code.
4043                  */
4044                 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4045                        __func__, sk->sk_state);
4046                 break;
4047         }
4048
4049         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4050          * Probably, we should reset in this case. For now drop them.
4051          */
4052         __skb_queue_purge(&tp->out_of_order_queue);
4053         if (tcp_is_sack(tp))
4054                 tcp_sack_reset(&tp->rx_opt);
4055         sk_mem_reclaim(sk);
4056
4057         if (!sock_flag(sk, SOCK_DEAD)) {
4058                 sk->sk_state_change(sk);
4059
4060                 /* Do not send POLL_HUP for half duplex close. */
4061                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4062                     sk->sk_state == TCP_CLOSE)
4063                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4064                 else
4065                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4066         }
4067 }
4068
4069 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4070                                   u32 end_seq)
4071 {
4072         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4073                 if (before(seq, sp->start_seq))
4074                         sp->start_seq = seq;
4075                 if (after(end_seq, sp->end_seq))
4076                         sp->end_seq = end_seq;
4077                 return 1;
4078         }
4079         return 0;
4080 }
4081
4082 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4083 {
4084         struct tcp_sock *tp = tcp_sk(sk);
4085
4086         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4087                 int mib_idx;
4088
4089                 if (before(seq, tp->rcv_nxt))
4090                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4091                 else
4092                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4093
4094                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4095
4096                 tp->rx_opt.dsack = 1;
4097                 tp->duplicate_sack[0].start_seq = seq;
4098                 tp->duplicate_sack[0].end_seq = end_seq;
4099         }
4100 }
4101
4102 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4103 {
4104         struct tcp_sock *tp = tcp_sk(sk);
4105
4106         if (!tp->rx_opt.dsack)
4107                 tcp_dsack_set(sk, seq, end_seq);
4108         else
4109                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4110 }
4111
4112 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4113 {
4114         struct tcp_sock *tp = tcp_sk(sk);
4115
4116         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4117             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4118                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4119                 tcp_enter_quickack_mode(sk);
4120
4121                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4122                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4123
4124                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4125                                 end_seq = tp->rcv_nxt;
4126                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4127                 }
4128         }
4129
4130         tcp_send_ack(sk);
4131 }
4132
4133 /* These routines update the SACK block as out-of-order packets arrive or
4134  * in-order packets close up the sequence space.
4135  */
4136 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4137 {
4138         int this_sack;
4139         struct tcp_sack_block *sp = &tp->selective_acks[0];
4140         struct tcp_sack_block *swalk = sp + 1;
4141
4142         /* See if the recent change to the first SACK eats into
4143          * or hits the sequence space of other SACK blocks, if so coalesce.
4144          */
4145         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4146                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4147                         int i;
4148
4149                         /* Zap SWALK, by moving every further SACK up by one slot.
4150                          * Decrease num_sacks.
4151                          */
4152                         tp->rx_opt.num_sacks--;
4153                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4154                                 sp[i] = sp[i + 1];
4155                         continue;
4156                 }
4157                 this_sack++, swalk++;
4158         }
4159 }
4160
4161 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
4162                                  struct tcp_sack_block *sack2)
4163 {
4164         __u32 tmp;
4165
4166         tmp = sack1->start_seq;
4167         sack1->start_seq = sack2->start_seq;
4168         sack2->start_seq = tmp;
4169
4170         tmp = sack1->end_seq;
4171         sack1->end_seq = sack2->end_seq;
4172         sack2->end_seq = tmp;
4173 }
4174
4175 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4176 {
4177         struct tcp_sock *tp = tcp_sk(sk);
4178         struct tcp_sack_block *sp = &tp->selective_acks[0];
4179         int cur_sacks = tp->rx_opt.num_sacks;
4180         int this_sack;
4181
4182         if (!cur_sacks)
4183                 goto new_sack;
4184
4185         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4186                 if (tcp_sack_extend(sp, seq, end_seq)) {
4187                         /* Rotate this_sack to the first one. */
4188                         for (; this_sack > 0; this_sack--, sp--)
4189                                 tcp_sack_swap(sp, sp - 1);
4190                         if (cur_sacks > 1)
4191                                 tcp_sack_maybe_coalesce(tp);
4192                         return;
4193                 }
4194         }
4195
4196         /* Could not find an adjacent existing SACK, build a new one,
4197          * put it at the front, and shift everyone else down.  We
4198          * always know there is at least one SACK present already here.
4199          *
4200          * If the sack array is full, forget about the last one.
4201          */
4202         if (this_sack >= TCP_NUM_SACKS) {
4203                 this_sack--;
4204                 tp->rx_opt.num_sacks--;
4205                 sp--;
4206         }
4207         for (; this_sack > 0; this_sack--, sp--)
4208                 *sp = *(sp - 1);
4209
4210 new_sack:
4211         /* Build the new head SACK, and we're done. */
4212         sp->start_seq = seq;
4213         sp->end_seq = end_seq;
4214         tp->rx_opt.num_sacks++;
4215 }
4216
4217 /* RCV.NXT advances, some SACKs should be eaten. */
4218
4219 static void tcp_sack_remove(struct tcp_sock *tp)
4220 {
4221         struct tcp_sack_block *sp = &tp->selective_acks[0];
4222         int num_sacks = tp->rx_opt.num_sacks;
4223         int this_sack;
4224
4225         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4226         if (skb_queue_empty(&tp->out_of_order_queue)) {
4227                 tp->rx_opt.num_sacks = 0;
4228                 return;
4229         }
4230
4231         for (this_sack = 0; this_sack < num_sacks;) {
4232                 /* Check if the start of the sack is covered by RCV.NXT. */
4233                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4234                         int i;
4235
4236                         /* RCV.NXT must cover all the block! */
4237                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4238
4239                         /* Zap this SACK, by moving forward any other SACKS. */
4240                         for (i=this_sack+1; i < num_sacks; i++)
4241                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4242                         num_sacks--;
4243                         continue;
4244                 }
4245                 this_sack++;
4246                 sp++;
4247         }
4248         tp->rx_opt.num_sacks = num_sacks;
4249 }
4250
4251 /* This one checks to see if we can put data from the
4252  * out_of_order queue into the receive_queue.
4253  */
4254 static void tcp_ofo_queue(struct sock *sk)
4255 {
4256         struct tcp_sock *tp = tcp_sk(sk);
4257         __u32 dsack_high = tp->rcv_nxt;
4258         struct sk_buff *skb;
4259
4260         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4261                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4262                         break;
4263
4264                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4265                         __u32 dsack = dsack_high;
4266                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4267                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4268                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4269                 }
4270
4271                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4272                         SOCK_DEBUG(sk, "ofo packet was already received \n");
4273                         __skb_unlink(skb, &tp->out_of_order_queue);
4274                         __kfree_skb(skb);
4275                         continue;
4276                 }
4277                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4278                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4279                            TCP_SKB_CB(skb)->end_seq);
4280
4281                 __skb_unlink(skb, &tp->out_of_order_queue);
4282                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4283                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4284                 if (tcp_hdr(skb)->fin)
4285                         tcp_fin(skb, sk, tcp_hdr(skb));
4286         }
4287 }
4288
4289 static int tcp_prune_ofo_queue(struct sock *sk);
4290 static int tcp_prune_queue(struct sock *sk);
4291
4292 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4293 {
4294         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4295             !sk_rmem_schedule(sk, size)) {
4296
4297                 if (tcp_prune_queue(sk) < 0)
4298                         return -1;
4299
4300                 if (!sk_rmem_schedule(sk, size)) {
4301                         if (!tcp_prune_ofo_queue(sk))
4302                                 return -1;
4303
4304                         if (!sk_rmem_schedule(sk, size))
4305                                 return -1;
4306                 }
4307         }
4308         return 0;
4309 }
4310
4311 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4312 {
4313         struct tcphdr *th = tcp_hdr(skb);
4314         struct tcp_sock *tp = tcp_sk(sk);
4315         int eaten = -1;
4316
4317         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4318                 goto drop;
4319
4320         __skb_pull(skb, th->doff * 4);
4321
4322         TCP_ECN_accept_cwr(tp, skb);
4323
4324         tp->rx_opt.dsack = 0;
4325
4326         /*  Queue data for delivery to the user.
4327          *  Packets in sequence go to the receive queue.
4328          *  Out of sequence packets to the out_of_order_queue.
4329          */
4330         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4331                 if (tcp_receive_window(tp) == 0)
4332                         goto out_of_window;
4333
4334                 /* Ok. In sequence. In window. */
4335                 if (tp->ucopy.task == current &&
4336                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4337                     sock_owned_by_user(sk) && !tp->urg_data) {
4338                         int chunk = min_t(unsigned int, skb->len,
4339                                           tp->ucopy.len);
4340
4341                         __set_current_state(TASK_RUNNING);
4342
4343                         local_bh_enable();
4344                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4345                                 tp->ucopy.len -= chunk;
4346                                 tp->copied_seq += chunk;
4347                                 eaten = (chunk == skb->len && !th->fin);
4348                                 tcp_rcv_space_adjust(sk);
4349                         }
4350                         local_bh_disable();
4351                 }
4352
4353                 if (eaten <= 0) {
4354 queue_and_out:
4355                         if (eaten < 0 &&
4356                             tcp_try_rmem_schedule(sk, skb->truesize))
4357                                 goto drop;
4358
4359                         skb_set_owner_r(skb, sk);
4360                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4361                 }
4362                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4363                 if (skb->len)
4364                         tcp_event_data_recv(sk, skb);
4365                 if (th->fin)
4366                         tcp_fin(skb, sk, th);
4367
4368                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4369                         tcp_ofo_queue(sk);
4370
4371                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4372                          * gap in queue is filled.
4373                          */
4374                         if (skb_queue_empty(&tp->out_of_order_queue))
4375                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4376                 }
4377
4378                 if (tp->rx_opt.num_sacks)
4379                         tcp_sack_remove(tp);
4380
4381                 tcp_fast_path_check(sk);
4382
4383                 if (eaten > 0)
4384                         __kfree_skb(skb);
4385                 else if (!sock_flag(sk, SOCK_DEAD))
4386                         sk->sk_data_ready(sk, 0);
4387                 return;
4388         }
4389
4390         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4391                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4392                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4393                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4394
4395 out_of_window:
4396                 tcp_enter_quickack_mode(sk);
4397                 inet_csk_schedule_ack(sk);
4398 drop:
4399                 __kfree_skb(skb);
4400                 return;
4401         }
4402
4403         /* Out of window. F.e. zero window probe. */
4404         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4405                 goto out_of_window;
4406
4407         tcp_enter_quickack_mode(sk);
4408
4409         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4410                 /* Partial packet, seq < rcv_next < end_seq */
4411                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4412                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4413                            TCP_SKB_CB(skb)->end_seq);
4414
4415                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4416
4417                 /* If window is closed, drop tail of packet. But after
4418                  * remembering D-SACK for its head made in previous line.
4419                  */
4420                 if (!tcp_receive_window(tp))
4421                         goto out_of_window;
4422                 goto queue_and_out;
4423         }
4424
4425         TCP_ECN_check_ce(tp, skb);
4426
4427         if (tcp_try_rmem_schedule(sk, skb->truesize))
4428                 goto drop;
4429
4430         /* Disable header prediction. */
4431         tp->pred_flags = 0;
4432         inet_csk_schedule_ack(sk);
4433
4434         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4435                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4436
4437         skb_set_owner_r(skb, sk);
4438
4439         if (!skb_peek(&tp->out_of_order_queue)) {
4440                 /* Initial out of order segment, build 1 SACK. */
4441                 if (tcp_is_sack(tp)) {
4442                         tp->rx_opt.num_sacks = 1;
4443                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4444                         tp->selective_acks[0].end_seq =
4445                                                 TCP_SKB_CB(skb)->end_seq;
4446                 }
4447                 __skb_queue_head(&tp->out_of_order_queue, skb);
4448         } else {
4449                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4450                 u32 seq = TCP_SKB_CB(skb)->seq;
4451                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4452
4453                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4454                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4455
4456                         if (!tp->rx_opt.num_sacks ||
4457                             tp->selective_acks[0].end_seq != seq)
4458                                 goto add_sack;
4459
4460                         /* Common case: data arrive in order after hole. */
4461                         tp->selective_acks[0].end_seq = end_seq;
4462                         return;
4463                 }
4464
4465                 /* Find place to insert this segment. */
4466                 do {
4467                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
4468                                 break;
4469                 } while ((skb1 = skb1->prev) !=
4470                          (struct sk_buff *)&tp->out_of_order_queue);
4471
4472                 /* Do skb overlap to previous one? */
4473                 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4474                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4475                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4476                                 /* All the bits are present. Drop. */
4477                                 __kfree_skb(skb);
4478                                 tcp_dsack_set(sk, seq, end_seq);
4479                                 goto add_sack;
4480                         }
4481                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4482                                 /* Partial overlap. */
4483                                 tcp_dsack_set(sk, seq,
4484                                               TCP_SKB_CB(skb1)->end_seq);
4485                         } else {
4486                                 skb1 = skb1->prev;
4487                         }
4488                 }
4489                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4490
4491                 /* And clean segments covered by new one as whole. */
4492                 while ((skb1 = skb->next) !=
4493                        (struct sk_buff *)&tp->out_of_order_queue &&
4494                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4495                         if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4496                                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4497                                                  end_seq);
4498                                 break;
4499                         }
4500                         __skb_unlink(skb1, &tp->out_of_order_queue);
4501                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4502                                          TCP_SKB_CB(skb1)->end_seq);
4503                         __kfree_skb(skb1);
4504                 }
4505
4506 add_sack:
4507                 if (tcp_is_sack(tp))
4508                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4509         }
4510 }
4511
4512 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4513                                         struct sk_buff_head *list)
4514 {
4515         struct sk_buff *next = skb->next;
4516
4517         __skb_unlink(skb, list);
4518         __kfree_skb(skb);
4519         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4520
4521         return next;
4522 }
4523
4524 /* Collapse contiguous sequence of skbs head..tail with
4525  * sequence numbers start..end.
4526  * Segments with FIN/SYN are not collapsed (only because this
4527  * simplifies code)
4528  */
4529 static void
4530 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4531              struct sk_buff *head, struct sk_buff *tail,
4532              u32 start, u32 end)
4533 {
4534         struct sk_buff *skb;
4535
4536         /* First, check that queue is collapsible and find
4537          * the point where collapsing can be useful. */
4538         for (skb = head; skb != tail;) {
4539                 /* No new bits? It is possible on ofo queue. */
4540                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4541                         skb = tcp_collapse_one(sk, skb, list);
4542                         continue;
4543                 }
4544
4545                 /* The first skb to collapse is:
4546                  * - not SYN/FIN and
4547                  * - bloated or contains data before "start" or
4548                  *   overlaps to the next one.
4549                  */
4550                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4551                     (tcp_win_from_space(skb->truesize) > skb->len ||
4552                      before(TCP_SKB_CB(skb)->seq, start) ||
4553                      (skb->next != tail &&
4554                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4555                         break;
4556
4557                 /* Decided to skip this, advance start seq. */
4558                 start = TCP_SKB_CB(skb)->end_seq;
4559                 skb = skb->next;
4560         }
4561         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4562                 return;
4563
4564         while (before(start, end)) {
4565                 struct sk_buff *nskb;
4566                 unsigned int header = skb_headroom(skb);
4567                 int copy = SKB_MAX_ORDER(header, 0);
4568
4569                 /* Too big header? This can happen with IPv6. */
4570                 if (copy < 0)
4571                         return;
4572                 if (end - start < copy)
4573                         copy = end - start;
4574                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4575                 if (!nskb)
4576                         return;
4577
4578                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4579                 skb_set_network_header(nskb, (skb_network_header(skb) -
4580                                               skb->head));
4581                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4582                                                 skb->head));
4583                 skb_reserve(nskb, header);
4584                 memcpy(nskb->head, skb->head, header);
4585                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4586                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4587                 __skb_queue_before(list, skb, nskb);
4588                 skb_set_owner_r(nskb, sk);
4589
4590                 /* Copy data, releasing collapsed skbs. */
4591                 while (copy > 0) {
4592                         int offset = start - TCP_SKB_CB(skb)->seq;
4593                         int size = TCP_SKB_CB(skb)->end_seq - start;
4594
4595                         BUG_ON(offset < 0);
4596                         if (size > 0) {
4597                                 size = min(copy, size);
4598                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4599                                         BUG();
4600                                 TCP_SKB_CB(nskb)->end_seq += size;
4601                                 copy -= size;
4602                                 start += size;
4603                         }
4604                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4605                                 skb = tcp_collapse_one(sk, skb, list);
4606                                 if (skb == tail ||
4607                                     tcp_hdr(skb)->syn ||
4608                                     tcp_hdr(skb)->fin)
4609                                         return;
4610                         }
4611                 }
4612         }
4613 }
4614
4615 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4616  * and tcp_collapse() them until all the queue is collapsed.
4617  */
4618 static void tcp_collapse_ofo_queue(struct sock *sk)
4619 {
4620         struct tcp_sock *tp = tcp_sk(sk);
4621         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4622         struct sk_buff *head;
4623         u32 start, end;
4624
4625         if (skb == NULL)
4626                 return;
4627
4628         start = TCP_SKB_CB(skb)->seq;
4629         end = TCP_SKB_CB(skb)->end_seq;
4630         head = skb;
4631
4632         for (;;) {
4633                 skb = skb->next;
4634
4635                 /* Segment is terminated when we see gap or when
4636                  * we are at the end of all the queue. */
4637                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4638                     after(TCP_SKB_CB(skb)->seq, end) ||
4639                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4640                         tcp_collapse(sk, &tp->out_of_order_queue,
4641                                      head, skb, start, end);
4642                         head = skb;
4643                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4644                                 break;
4645                         /* Start new segment */
4646                         start = TCP_SKB_CB(skb)->seq;
4647                         end = TCP_SKB_CB(skb)->end_seq;
4648                 } else {
4649                         if (before(TCP_SKB_CB(skb)->seq, start))
4650                                 start = TCP_SKB_CB(skb)->seq;
4651                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4652                                 end = TCP_SKB_CB(skb)->end_seq;
4653                 }
4654         }
4655 }
4656
4657 /*
4658  * Purge the out-of-order queue.
4659  * Return true if queue was pruned.
4660  */
4661 static int tcp_prune_ofo_queue(struct sock *sk)
4662 {
4663         struct tcp_sock *tp = tcp_sk(sk);
4664         int res = 0;
4665
4666         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4667                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4668                 __skb_queue_purge(&tp->out_of_order_queue);
4669
4670                 /* Reset SACK state.  A conforming SACK implementation will
4671                  * do the same at a timeout based retransmit.  When a connection
4672                  * is in a sad state like this, we care only about integrity
4673                  * of the connection not performance.
4674                  */
4675                 if (tp->rx_opt.sack_ok)
4676                         tcp_sack_reset(&tp->rx_opt);
4677                 sk_mem_reclaim(sk);
4678                 res = 1;
4679         }
4680         return res;
4681 }
4682
4683 /* Reduce allocated memory if we can, trying to get
4684  * the socket within its memory limits again.
4685  *
4686  * Return less than zero if we should start dropping frames
4687  * until the socket owning process reads some of the data
4688  * to stabilize the situation.
4689  */
4690 static int tcp_prune_queue(struct sock *sk)
4691 {
4692         struct tcp_sock *tp = tcp_sk(sk);
4693
4694         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4695
4696         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4697
4698         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4699                 tcp_clamp_window(sk);
4700         else if (tcp_memory_pressure)
4701                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4702
4703         tcp_collapse_ofo_queue(sk);
4704         tcp_collapse(sk, &sk->sk_receive_queue,
4705                      sk->sk_receive_queue.next,
4706                      (struct sk_buff *)&sk->sk_receive_queue,
4707                      tp->copied_seq, tp->rcv_nxt);
4708         sk_mem_reclaim(sk);
4709
4710         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4711                 return 0;
4712
4713         /* Collapsing did not help, destructive actions follow.
4714          * This must not ever occur. */
4715
4716         tcp_prune_ofo_queue(sk);
4717
4718         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4719                 return 0;
4720
4721         /* If we are really being abused, tell the caller to silently
4722          * drop receive data on the floor.  It will get retransmitted
4723          * and hopefully then we'll have sufficient space.
4724          */
4725         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4726
4727         /* Massive buffer overcommit. */
4728         tp->pred_flags = 0;
4729         return -1;
4730 }
4731
4732 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4733  * As additional protections, we do not touch cwnd in retransmission phases,
4734  * and if application hit its sndbuf limit recently.
4735  */
4736 void tcp_cwnd_application_limited(struct sock *sk)
4737 {
4738         struct tcp_sock *tp = tcp_sk(sk);
4739
4740         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4741             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4742                 /* Limited by application or receiver window. */
4743                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4744                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4745                 if (win_used < tp->snd_cwnd) {
4746                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4747                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4748                 }
4749                 tp->snd_cwnd_used = 0;
4750         }
4751         tp->snd_cwnd_stamp = tcp_time_stamp;
4752 }
4753
4754 static int tcp_should_expand_sndbuf(struct sock *sk)
4755 {
4756         struct tcp_sock *tp = tcp_sk(sk);
4757
4758         /* If the user specified a specific send buffer setting, do
4759          * not modify it.
4760          */
4761         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4762                 return 0;
4763
4764         /* If we are under global TCP memory pressure, do not expand.  */
4765         if (tcp_memory_pressure)
4766                 return 0;
4767
4768         /* If we are under soft global TCP memory pressure, do not expand.  */
4769         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4770                 return 0;
4771
4772         /* If we filled the congestion window, do not expand.  */
4773         if (tp->packets_out >= tp->snd_cwnd)
4774                 return 0;
4775
4776         return 1;
4777 }
4778
4779 /* When incoming ACK allowed to free some skb from write_queue,
4780  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4781  * on the exit from tcp input handler.
4782  *
4783  * PROBLEM: sndbuf expansion does not work well with largesend.
4784  */
4785 static void tcp_new_space(struct sock *sk)
4786 {
4787         struct tcp_sock *tp = tcp_sk(sk);
4788
4789         if (tcp_should_expand_sndbuf(sk)) {
4790                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4791                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4792                 int demanded = max_t(unsigned int, tp->snd_cwnd,
4793                                      tp->reordering + 1);
4794                 sndmem *= 2 * demanded;
4795                 if (sndmem > sk->sk_sndbuf)
4796                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4797                 tp->snd_cwnd_stamp = tcp_time_stamp;
4798         }
4799
4800         sk->sk_write_space(sk);
4801 }
4802
4803 static void tcp_check_space(struct sock *sk)
4804 {
4805         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4806                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4807                 if (sk->sk_socket &&
4808                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4809                         tcp_new_space(sk);
4810         }
4811 }
4812
4813 static inline void tcp_data_snd_check(struct sock *sk)
4814 {
4815         tcp_push_pending_frames(sk);
4816         tcp_check_space(sk);
4817 }
4818
4819 /*
4820  * Check if sending an ack is needed.
4821  */
4822 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4823 {
4824         struct tcp_sock *tp = tcp_sk(sk);
4825
4826             /* More than one full frame received... */
4827         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4828              /* ... and right edge of window advances far enough.
4829               * (tcp_recvmsg() will send ACK otherwise). Or...
4830               */
4831              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4832             /* We ACK each frame or... */
4833             tcp_in_quickack_mode(sk) ||
4834             /* We have out of order data. */
4835             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4836                 /* Then ack it now */
4837                 tcp_send_ack(sk);
4838         } else {
4839                 /* Else, send delayed ack. */
4840                 tcp_send_delayed_ack(sk);
4841         }
4842 }
4843
4844 static inline void tcp_ack_snd_check(struct sock *sk)
4845 {
4846         if (!inet_csk_ack_scheduled(sk)) {
4847                 /* We sent a data segment already. */
4848                 return;
4849         }
4850         __tcp_ack_snd_check(sk, 1);
4851 }
4852
4853 /*
4854  *      This routine is only called when we have urgent data
4855  *      signaled. Its the 'slow' part of tcp_urg. It could be
4856  *      moved inline now as tcp_urg is only called from one
4857  *      place. We handle URGent data wrong. We have to - as
4858  *      BSD still doesn't use the correction from RFC961.
4859  *      For 1003.1g we should support a new option TCP_STDURG to permit
4860  *      either form (or just set the sysctl tcp_stdurg).
4861  */
4862
4863 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4864 {
4865         struct tcp_sock *tp = tcp_sk(sk);
4866         u32 ptr = ntohs(th->urg_ptr);
4867
4868         if (ptr && !sysctl_tcp_stdurg)
4869                 ptr--;
4870         ptr += ntohl(th->seq);
4871
4872         /* Ignore urgent data that we've already seen and read. */
4873         if (after(tp->copied_seq, ptr))
4874                 return;
4875
4876         /* Do not replay urg ptr.
4877          *
4878          * NOTE: interesting situation not covered by specs.
4879          * Misbehaving sender may send urg ptr, pointing to segment,
4880          * which we already have in ofo queue. We are not able to fetch
4881          * such data and will stay in TCP_URG_NOTYET until will be eaten
4882          * by recvmsg(). Seems, we are not obliged to handle such wicked
4883          * situations. But it is worth to think about possibility of some
4884          * DoSes using some hypothetical application level deadlock.
4885          */
4886         if (before(ptr, tp->rcv_nxt))
4887                 return;
4888
4889         /* Do we already have a newer (or duplicate) urgent pointer? */
4890         if (tp->urg_data && !after(ptr, tp->urg_seq))
4891                 return;
4892
4893         /* Tell the world about our new urgent pointer. */
4894         sk_send_sigurg(sk);
4895
4896         /* We may be adding urgent data when the last byte read was
4897          * urgent. To do this requires some care. We cannot just ignore
4898          * tp->copied_seq since we would read the last urgent byte again
4899          * as data, nor can we alter copied_seq until this data arrives
4900          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4901          *
4902          * NOTE. Double Dutch. Rendering to plain English: author of comment
4903          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4904          * and expect that both A and B disappear from stream. This is _wrong_.
4905          * Though this happens in BSD with high probability, this is occasional.
4906          * Any application relying on this is buggy. Note also, that fix "works"
4907          * only in this artificial test. Insert some normal data between A and B and we will
4908          * decline of BSD again. Verdict: it is better to remove to trap
4909          * buggy users.
4910          */
4911         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4912             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4913                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4914                 tp->copied_seq++;
4915                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4916                         __skb_unlink(skb, &sk->sk_receive_queue);
4917                         __kfree_skb(skb);
4918                 }
4919         }
4920
4921         tp->urg_data = TCP_URG_NOTYET;
4922         tp->urg_seq = ptr;
4923
4924         /* Disable header prediction. */
4925         tp->pred_flags = 0;
4926 }
4927
4928 /* This is the 'fast' part of urgent handling. */
4929 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4930 {
4931         struct tcp_sock *tp = tcp_sk(sk);
4932
4933         /* Check if we get a new urgent pointer - normally not. */
4934         if (th->urg)
4935                 tcp_check_urg(sk, th);
4936
4937         /* Do we wait for any urgent data? - normally not... */
4938         if (tp->urg_data == TCP_URG_NOTYET) {
4939                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4940                           th->syn;
4941
4942                 /* Is the urgent pointer pointing into this packet? */
4943                 if (ptr < skb->len) {
4944                         u8 tmp;
4945                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4946                                 BUG();
4947                         tp->urg_data = TCP_URG_VALID | tmp;
4948                         if (!sock_flag(sk, SOCK_DEAD))
4949                                 sk->sk_data_ready(sk, 0);
4950                 }
4951         }
4952 }
4953
4954 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4955 {
4956         struct tcp_sock *tp = tcp_sk(sk);
4957         int chunk = skb->len - hlen;
4958         int err;
4959
4960         local_bh_enable();
4961         if (skb_csum_unnecessary(skb))
4962                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4963         else
4964                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4965                                                        tp->ucopy.iov);
4966
4967         if (!err) {
4968                 tp->ucopy.len -= chunk;
4969                 tp->copied_seq += chunk;
4970                 tcp_rcv_space_adjust(sk);
4971         }
4972
4973         local_bh_disable();
4974         return err;
4975 }
4976
4977 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4978                                             struct sk_buff *skb)
4979 {
4980         __sum16 result;
4981
4982         if (sock_owned_by_user(sk)) {
4983                 local_bh_enable();
4984                 result = __tcp_checksum_complete(skb);
4985                 local_bh_disable();
4986         } else {
4987                 result = __tcp_checksum_complete(skb);
4988         }
4989         return result;
4990 }
4991
4992 static inline int tcp_checksum_complete_user(struct sock *sk,
4993                                              struct sk_buff *skb)
4994 {
4995         return !skb_csum_unnecessary(skb) &&
4996                __tcp_checksum_complete_user(sk, skb);
4997 }
4998
4999 #ifdef CONFIG_NET_DMA
5000 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5001                                   int hlen)
5002 {
5003         struct tcp_sock *tp = tcp_sk(sk);
5004         int chunk = skb->len - hlen;
5005         int dma_cookie;
5006         int copied_early = 0;
5007
5008         if (tp->ucopy.wakeup)
5009                 return 0;
5010
5011         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5012                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5013
5014         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5015
5016                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5017                                                          skb, hlen,
5018                                                          tp->ucopy.iov, chunk,
5019                                                          tp->ucopy.pinned_list);
5020
5021                 if (dma_cookie < 0)
5022                         goto out;
5023
5024                 tp->ucopy.dma_cookie = dma_cookie;
5025                 copied_early = 1;
5026
5027                 tp->ucopy.len -= chunk;
5028                 tp->copied_seq += chunk;
5029                 tcp_rcv_space_adjust(sk);
5030
5031                 if ((tp->ucopy.len == 0) ||
5032                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5033                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5034                         tp->ucopy.wakeup = 1;
5035                         sk->sk_data_ready(sk, 0);
5036                 }
5037         } else if (chunk > 0) {
5038                 tp->ucopy.wakeup = 1;
5039                 sk->sk_data_ready(sk, 0);
5040         }
5041 out:
5042         return copied_early;
5043 }
5044 #endif /* CONFIG_NET_DMA */
5045
5046 /* Does PAWS and seqno based validation of an incoming segment, flags will
5047  * play significant role here.
5048  */
5049 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5050                               struct tcphdr *th, int syn_inerr)
5051 {
5052         struct tcp_sock *tp = tcp_sk(sk);
5053
5054         /* RFC1323: H1. Apply PAWS check first. */
5055         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5056             tcp_paws_discard(sk, skb)) {
5057                 if (!th->rst) {
5058                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5059                         tcp_send_dupack(sk, skb);
5060                         goto discard;
5061                 }
5062                 /* Reset is accepted even if it did not pass PAWS. */
5063         }
5064
5065         /* Step 1: check sequence number */
5066         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5067                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5068                  * (RST) segments are validated by checking their SEQ-fields."
5069                  * And page 69: "If an incoming segment is not acceptable,
5070                  * an acknowledgment should be sent in reply (unless the RST
5071                  * bit is set, if so drop the segment and return)".
5072                  */
5073                 if (!th->rst)
5074                         tcp_send_dupack(sk, skb);
5075                 goto discard;
5076         }
5077
5078         /* Step 2: check RST bit */
5079         if (th->rst) {
5080                 tcp_reset(sk);
5081                 goto discard;
5082         }
5083
5084         /* ts_recent update must be made after we are sure that the packet
5085          * is in window.
5086          */
5087         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5088
5089         /* step 3: check security and precedence [ignored] */
5090
5091         /* step 4: Check for a SYN in window. */
5092         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5093                 if (syn_inerr)
5094                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5095                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5096                 tcp_reset(sk);
5097                 return -1;
5098         }
5099
5100         return 1;
5101
5102 discard:
5103         __kfree_skb(skb);
5104         return 0;
5105 }
5106
5107 /*
5108  *      TCP receive function for the ESTABLISHED state.
5109  *
5110  *      It is split into a fast path and a slow path. The fast path is
5111  *      disabled when:
5112  *      - A zero window was announced from us - zero window probing
5113  *        is only handled properly in the slow path.
5114  *      - Out of order segments arrived.
5115  *      - Urgent data is expected.
5116  *      - There is no buffer space left
5117  *      - Unexpected TCP flags/window values/header lengths are received
5118  *        (detected by checking the TCP header against pred_flags)
5119  *      - Data is sent in both directions. Fast path only supports pure senders
5120  *        or pure receivers (this means either the sequence number or the ack
5121  *        value must stay constant)
5122  *      - Unexpected TCP option.
5123  *
5124  *      When these conditions are not satisfied it drops into a standard
5125  *      receive procedure patterned after RFC793 to handle all cases.
5126  *      The first three cases are guaranteed by proper pred_flags setting,
5127  *      the rest is checked inline. Fast processing is turned on in
5128  *      tcp_data_queue when everything is OK.
5129  */
5130 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5131                         struct tcphdr *th, unsigned len)
5132 {
5133         struct tcp_sock *tp = tcp_sk(sk);
5134         int res;
5135
5136         /*
5137          *      Header prediction.
5138          *      The code loosely follows the one in the famous
5139          *      "30 instruction TCP receive" Van Jacobson mail.
5140          *
5141          *      Van's trick is to deposit buffers into socket queue
5142          *      on a device interrupt, to call tcp_recv function
5143          *      on the receive process context and checksum and copy
5144          *      the buffer to user space. smart...
5145          *
5146          *      Our current scheme is not silly either but we take the
5147          *      extra cost of the net_bh soft interrupt processing...
5148          *      We do checksum and copy also but from device to kernel.
5149          */
5150
5151         tp->rx_opt.saw_tstamp = 0;
5152
5153         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5154          *      if header_prediction is to be made
5155          *      'S' will always be tp->tcp_header_len >> 2
5156          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5157          *  turn it off (when there are holes in the receive
5158          *       space for instance)
5159          *      PSH flag is ignored.
5160          */
5161
5162         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5163             TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5164                 int tcp_header_len = tp->tcp_header_len;
5165
5166                 /* Timestamp header prediction: tcp_header_len
5167                  * is automatically equal to th->doff*4 due to pred_flags
5168                  * match.
5169                  */
5170
5171                 /* Check timestamp */
5172                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5173                         /* No? Slow path! */
5174                         if (!tcp_parse_aligned_timestamp(tp, th))
5175                                 goto slow_path;
5176
5177                         /* If PAWS failed, check it more carefully in slow path */
5178                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5179                                 goto slow_path;
5180
5181                         /* DO NOT update ts_recent here, if checksum fails
5182                          * and timestamp was corrupted part, it will result
5183                          * in a hung connection since we will drop all
5184                          * future packets due to the PAWS test.
5185                          */
5186                 }
5187
5188                 if (len <= tcp_header_len) {
5189                         /* Bulk data transfer: sender */
5190                         if (len == tcp_header_len) {
5191                                 /* Predicted packet is in window by definition.
5192                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5193                                  * Hence, check seq<=rcv_wup reduces to:
5194                                  */
5195                                 if (tcp_header_len ==
5196                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5197                                     tp->rcv_nxt == tp->rcv_wup)
5198                                         tcp_store_ts_recent(tp);
5199
5200                                 /* We know that such packets are checksummed
5201                                  * on entry.
5202                                  */
5203                                 tcp_ack(sk, skb, 0);
5204                                 __kfree_skb(skb);
5205                                 tcp_data_snd_check(sk);
5206                                 return 0;
5207                         } else { /* Header too small */
5208                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5209                                 goto discard;
5210                         }
5211                 } else {
5212                         int eaten = 0;
5213                         int copied_early = 0;
5214
5215                         if (tp->copied_seq == tp->rcv_nxt &&
5216                             len - tcp_header_len <= tp->ucopy.len) {
5217 #ifdef CONFIG_NET_DMA
5218                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5219                                         copied_early = 1;
5220                                         eaten = 1;
5221                                 }
5222 #endif
5223                                 if (tp->ucopy.task == current &&
5224                                     sock_owned_by_user(sk) && !copied_early) {
5225                                         __set_current_state(TASK_RUNNING);
5226
5227                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5228                                                 eaten = 1;
5229                                 }
5230                                 if (eaten) {
5231                                         /* Predicted packet is in window by definition.
5232                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5233                                          * Hence, check seq<=rcv_wup reduces to:
5234                                          */
5235                                         if (tcp_header_len ==
5236                                             (sizeof(struct tcphdr) +
5237                                              TCPOLEN_TSTAMP_ALIGNED) &&
5238                                             tp->rcv_nxt == tp->rcv_wup)
5239                                                 tcp_store_ts_recent(tp);
5240
5241                                         tcp_rcv_rtt_measure_ts(sk, skb);
5242
5243                                         __skb_pull(skb, tcp_header_len);
5244                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5245                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5246                                 }
5247                                 if (copied_early)
5248                                         tcp_cleanup_rbuf(sk, skb->len);
5249                         }
5250                         if (!eaten) {
5251                                 if (tcp_checksum_complete_user(sk, skb))
5252                                         goto csum_error;
5253
5254                                 /* Predicted packet is in window by definition.
5255                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5256                                  * Hence, check seq<=rcv_wup reduces to:
5257                                  */
5258                                 if (tcp_header_len ==
5259                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5260                                     tp->rcv_nxt == tp->rcv_wup)
5261                                         tcp_store_ts_recent(tp);
5262
5263                                 tcp_rcv_rtt_measure_ts(sk, skb);
5264
5265                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5266                                         goto step5;
5267
5268                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5269
5270                                 /* Bulk data transfer: receiver */
5271                                 __skb_pull(skb, tcp_header_len);
5272                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
5273                                 skb_set_owner_r(skb, sk);
5274                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5275                         }
5276
5277                         tcp_event_data_recv(sk, skb);
5278
5279                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5280                                 /* Well, only one small jumplet in fast path... */
5281                                 tcp_ack(sk, skb, FLAG_DATA);
5282                                 tcp_data_snd_check(sk);
5283                                 if (!inet_csk_ack_scheduled(sk))
5284                                         goto no_ack;
5285                         }
5286
5287                         if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5288                                 __tcp_ack_snd_check(sk, 0);
5289 no_ack:
5290 #ifdef CONFIG_NET_DMA
5291                         if (copied_early)
5292                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5293                         else
5294 #endif
5295                         if (eaten)
5296                                 __kfree_skb(skb);
5297                         else
5298                                 sk->sk_data_ready(sk, 0);
5299                         return 0;
5300                 }
5301         }
5302
5303 slow_path:
5304         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5305                 goto csum_error;
5306
5307         /*
5308          *      Standard slow path.
5309          */
5310
5311         res = tcp_validate_incoming(sk, skb, th, 1);
5312         if (res <= 0)
5313                 return -res;
5314
5315 step5:
5316         if (th->ack)
5317                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5318
5319         tcp_rcv_rtt_measure_ts(sk, skb);
5320
5321         /* Process urgent data. */
5322         tcp_urg(sk, skb, th);
5323
5324         /* step 7: process the segment text */
5325         tcp_data_queue(sk, skb);
5326
5327         tcp_data_snd_check(sk);
5328         tcp_ack_snd_check(sk);
5329         return 0;
5330
5331 csum_error:
5332         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5333
5334 discard:
5335         __kfree_skb(skb);
5336         return 0;
5337 }
5338
5339 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5340                                          struct tcphdr *th, unsigned len)
5341 {
5342         struct tcp_sock *tp = tcp_sk(sk);
5343         struct inet_connection_sock *icsk = inet_csk(sk);
5344         int saved_clamp = tp->rx_opt.mss_clamp;
5345
5346         tcp_parse_options(skb, &tp->rx_opt, 0);
5347
5348         if (th->ack) {
5349                 /* rfc793:
5350                  * "If the state is SYN-SENT then
5351                  *    first check the ACK bit
5352                  *      If the ACK bit is set
5353                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5354                  *        a reset (unless the RST bit is set, if so drop
5355                  *        the segment and return)"
5356                  *
5357                  *  We do not send data with SYN, so that RFC-correct
5358                  *  test reduces to:
5359                  */
5360                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5361                         goto reset_and_undo;
5362
5363                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5364                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5365                              tcp_time_stamp)) {
5366                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5367                         goto reset_and_undo;
5368                 }
5369
5370                 /* Now ACK is acceptable.
5371                  *
5372                  * "If the RST bit is set
5373                  *    If the ACK was acceptable then signal the user "error:
5374                  *    connection reset", drop the segment, enter CLOSED state,
5375                  *    delete TCB, and return."
5376                  */
5377
5378                 if (th->rst) {
5379                         tcp_reset(sk);
5380                         goto discard;
5381                 }
5382
5383                 /* rfc793:
5384                  *   "fifth, if neither of the SYN or RST bits is set then
5385                  *    drop the segment and return."
5386                  *
5387                  *    See note below!
5388                  *                                        --ANK(990513)
5389                  */
5390                 if (!th->syn)
5391                         goto discard_and_undo;
5392
5393                 /* rfc793:
5394                  *   "If the SYN bit is on ...
5395                  *    are acceptable then ...
5396                  *    (our SYN has been ACKed), change the connection
5397                  *    state to ESTABLISHED..."
5398                  */
5399
5400                 TCP_ECN_rcv_synack(tp, th);
5401
5402                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5403                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5404
5405                 /* Ok.. it's good. Set up sequence numbers and
5406                  * move to established.
5407                  */
5408                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5409                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5410
5411                 /* RFC1323: The window in SYN & SYN/ACK segments is
5412                  * never scaled.
5413                  */
5414                 tp->snd_wnd = ntohs(th->window);
5415                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5416
5417                 if (!tp->rx_opt.wscale_ok) {
5418                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5419                         tp->window_clamp = min(tp->window_clamp, 65535U);
5420                 }
5421
5422                 if (tp->rx_opt.saw_tstamp) {
5423                         tp->rx_opt.tstamp_ok       = 1;
5424                         tp->tcp_header_len =
5425                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5426                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5427                         tcp_store_ts_recent(tp);
5428                 } else {
5429                         tp->tcp_header_len = sizeof(struct tcphdr);
5430                 }
5431
5432                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5433                         tcp_enable_fack(tp);
5434
5435                 tcp_mtup_init(sk);
5436                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5437                 tcp_initialize_rcv_mss(sk);
5438
5439                 /* Remember, tcp_poll() does not lock socket!
5440                  * Change state from SYN-SENT only after copied_seq
5441                  * is initialized. */
5442                 tp->copied_seq = tp->rcv_nxt;
5443                 smp_mb();
5444                 tcp_set_state(sk, TCP_ESTABLISHED);
5445
5446                 security_inet_conn_established(sk, skb);
5447
5448                 /* Make sure socket is routed, for correct metrics.  */
5449                 icsk->icsk_af_ops->rebuild_header(sk);
5450
5451                 tcp_init_metrics(sk);
5452
5453                 tcp_init_congestion_control(sk);
5454
5455                 /* Prevent spurious tcp_cwnd_restart() on first data
5456                  * packet.
5457                  */
5458                 tp->lsndtime = tcp_time_stamp;
5459
5460                 tcp_init_buffer_space(sk);
5461
5462                 if (sock_flag(sk, SOCK_KEEPOPEN))
5463                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5464
5465                 if (!tp->rx_opt.snd_wscale)
5466                         __tcp_fast_path_on(tp, tp->snd_wnd);
5467                 else
5468                         tp->pred_flags = 0;
5469
5470                 if (!sock_flag(sk, SOCK_DEAD)) {
5471                         sk->sk_state_change(sk);
5472                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5473                 }
5474
5475                 if (sk->sk_write_pending ||
5476                     icsk->icsk_accept_queue.rskq_defer_accept ||
5477                     icsk->icsk_ack.pingpong) {
5478                         /* Save one ACK. Data will be ready after
5479                          * several ticks, if write_pending is set.
5480                          *
5481                          * It may be deleted, but with this feature tcpdumps
5482                          * look so _wonderfully_ clever, that I was not able
5483                          * to stand against the temptation 8)     --ANK
5484                          */
5485                         inet_csk_schedule_ack(sk);
5486                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5487                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
5488                         tcp_incr_quickack(sk);
5489                         tcp_enter_quickack_mode(sk);
5490                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5491                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5492
5493 discard:
5494                         __kfree_skb(skb);
5495                         return 0;
5496                 } else {
5497                         tcp_send_ack(sk);
5498                 }
5499                 return -1;
5500         }
5501
5502         /* No ACK in the segment */
5503
5504         if (th->rst) {
5505                 /* rfc793:
5506                  * "If the RST bit is set
5507                  *
5508                  *      Otherwise (no ACK) drop the segment and return."
5509                  */
5510
5511                 goto discard_and_undo;
5512         }
5513
5514         /* PAWS check. */
5515         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5516             tcp_paws_check(&tp->rx_opt, 0))
5517                 goto discard_and_undo;
5518
5519         if (th->syn) {
5520                 /* We see SYN without ACK. It is attempt of
5521                  * simultaneous connect with crossed SYNs.
5522                  * Particularly, it can be connect to self.
5523                  */
5524                 tcp_set_state(sk, TCP_SYN_RECV);
5525
5526                 if (tp->rx_opt.saw_tstamp) {
5527                         tp->rx_opt.tstamp_ok = 1;
5528                         tcp_store_ts_recent(tp);
5529                         tp->tcp_header_len =
5530                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5531                 } else {
5532                         tp->tcp_header_len = sizeof(struct tcphdr);
5533                 }
5534
5535                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5536                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5537
5538                 /* RFC1323: The window in SYN & SYN/ACK segments is
5539                  * never scaled.
5540                  */
5541                 tp->snd_wnd    = ntohs(th->window);
5542                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5543                 tp->max_window = tp->snd_wnd;
5544
5545                 TCP_ECN_rcv_syn(tp, th);
5546
5547                 tcp_mtup_init(sk);
5548                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5549                 tcp_initialize_rcv_mss(sk);
5550
5551                 tcp_send_synack(sk);
5552 #if 0
5553                 /* Note, we could accept data and URG from this segment.
5554                  * There are no obstacles to make this.
5555                  *
5556                  * However, if we ignore data in ACKless segments sometimes,
5557                  * we have no reasons to accept it sometimes.
5558                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5559                  * is not flawless. So, discard packet for sanity.
5560                  * Uncomment this return to process the data.
5561                  */
5562                 return -1;
5563 #else
5564                 goto discard;
5565 #endif
5566         }
5567         /* "fifth, if neither of the SYN or RST bits is set then
5568          * drop the segment and return."
5569          */
5570
5571 discard_and_undo:
5572         tcp_clear_options(&tp->rx_opt);
5573         tp->rx_opt.mss_clamp = saved_clamp;
5574         goto discard;
5575
5576 reset_and_undo:
5577         tcp_clear_options(&tp->rx_opt);
5578         tp->rx_opt.mss_clamp = saved_clamp;
5579         return 1;
5580 }
5581
5582 /*
5583  *      This function implements the receiving procedure of RFC 793 for
5584  *      all states except ESTABLISHED and TIME_WAIT.
5585  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5586  *      address independent.
5587  */
5588
5589 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5590                           struct tcphdr *th, unsigned len)
5591 {
5592         struct tcp_sock *tp = tcp_sk(sk);
5593         struct inet_connection_sock *icsk = inet_csk(sk);
5594         int queued = 0;
5595         int res;
5596
5597         tp->rx_opt.saw_tstamp = 0;
5598
5599         switch (sk->sk_state) {
5600         case TCP_CLOSE:
5601                 goto discard;
5602
5603         case TCP_LISTEN:
5604                 if (th->ack)
5605                         return 1;
5606
5607                 if (th->rst)
5608                         goto discard;
5609
5610                 if (th->syn) {
5611                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5612                                 return 1;
5613
5614                         /* Now we have several options: In theory there is
5615                          * nothing else in the frame. KA9Q has an option to
5616                          * send data with the syn, BSD accepts data with the
5617                          * syn up to the [to be] advertised window and
5618                          * Solaris 2.1 gives you a protocol error. For now
5619                          * we just ignore it, that fits the spec precisely
5620                          * and avoids incompatibilities. It would be nice in
5621                          * future to drop through and process the data.
5622                          *
5623                          * Now that TTCP is starting to be used we ought to
5624                          * queue this data.
5625                          * But, this leaves one open to an easy denial of
5626                          * service attack, and SYN cookies can't defend
5627                          * against this problem. So, we drop the data
5628                          * in the interest of security over speed unless
5629                          * it's still in use.
5630                          */
5631                         kfree_skb(skb);
5632                         return 0;
5633                 }
5634                 goto discard;
5635
5636         case TCP_SYN_SENT:
5637                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5638                 if (queued >= 0)
5639                         return queued;
5640
5641                 /* Do step6 onward by hand. */
5642                 tcp_urg(sk, skb, th);
5643                 __kfree_skb(skb);
5644                 tcp_data_snd_check(sk);
5645                 return 0;
5646         }
5647
5648         res = tcp_validate_incoming(sk, skb, th, 0);
5649         if (res <= 0)
5650                 return -res;
5651
5652         /* step 5: check the ACK field */
5653         if (th->ack) {
5654                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5655
5656                 switch (sk->sk_state) {
5657                 case TCP_SYN_RECV:
5658                         if (acceptable) {
5659                                 tp->copied_seq = tp->rcv_nxt;
5660                                 smp_mb();
5661                                 tcp_set_state(sk, TCP_ESTABLISHED);
5662                                 sk->sk_state_change(sk);
5663
5664                                 /* Note, that this wakeup is only for marginal
5665                                  * crossed SYN case. Passively open sockets
5666                                  * are not waked up, because sk->sk_sleep ==
5667                                  * NULL and sk->sk_socket == NULL.
5668                                  */
5669                                 if (sk->sk_socket)
5670                                         sk_wake_async(sk,
5671                                                       SOCK_WAKE_IO, POLL_OUT);
5672
5673                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5674                                 tp->snd_wnd = ntohs(th->window) <<
5675                                               tp->rx_opt.snd_wscale;
5676                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5677
5678                                 /* tcp_ack considers this ACK as duplicate
5679                                  * and does not calculate rtt.
5680                                  * Fix it at least with timestamps.
5681                                  */
5682                                 if (tp->rx_opt.saw_tstamp &&
5683                                     tp->rx_opt.rcv_tsecr && !tp->srtt)
5684                                         tcp_ack_saw_tstamp(sk, 0);
5685
5686                                 if (tp->rx_opt.tstamp_ok)
5687                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5688
5689                                 /* Make sure socket is routed, for
5690                                  * correct metrics.
5691                                  */
5692                                 icsk->icsk_af_ops->rebuild_header(sk);
5693
5694                                 tcp_init_metrics(sk);
5695
5696                                 tcp_init_congestion_control(sk);
5697
5698                                 /* Prevent spurious tcp_cwnd_restart() on
5699                                  * first data packet.
5700                                  */
5701                                 tp->lsndtime = tcp_time_stamp;
5702
5703                                 tcp_mtup_init(sk);
5704                                 tcp_initialize_rcv_mss(sk);
5705                                 tcp_init_buffer_space(sk);
5706                                 tcp_fast_path_on(tp);
5707                         } else {
5708                                 return 1;
5709                         }
5710                         break;
5711
5712                 case TCP_FIN_WAIT1:
5713                         if (tp->snd_una == tp->write_seq) {
5714                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5715                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5716                                 dst_confirm(sk->sk_dst_cache);
5717
5718                                 if (!sock_flag(sk, SOCK_DEAD))
5719                                         /* Wake up lingering close() */
5720                                         sk->sk_state_change(sk);
5721                                 else {
5722                                         int tmo;
5723
5724                                         if (tp->linger2 < 0 ||
5725                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5726                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5727                                                 tcp_done(sk);
5728                                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5729                                                 return 1;
5730                                         }
5731
5732                                         tmo = tcp_fin_time(sk);
5733                                         if (tmo > TCP_TIMEWAIT_LEN) {
5734                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5735                                         } else if (th->fin || sock_owned_by_user(sk)) {
5736                                                 /* Bad case. We could lose such FIN otherwise.
5737                                                  * It is not a big problem, but it looks confusing
5738                                                  * and not so rare event. We still can lose it now,
5739                                                  * if it spins in bh_lock_sock(), but it is really
5740                                                  * marginal case.
5741                                                  */
5742                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5743                                         } else {
5744                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5745                                                 goto discard;
5746                                         }
5747                                 }
5748                         }
5749                         break;
5750
5751                 case TCP_CLOSING:
5752                         if (tp->snd_una == tp->write_seq) {
5753                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5754                                 goto discard;
5755                         }
5756                         break;
5757
5758                 case TCP_LAST_ACK:
5759                         if (tp->snd_una == tp->write_seq) {
5760                                 tcp_update_metrics(sk);
5761                                 tcp_done(sk);
5762                                 goto discard;
5763                         }
5764                         break;
5765                 }
5766         } else
5767                 goto discard;
5768
5769         /* step 6: check the URG bit */
5770         tcp_urg(sk, skb, th);
5771
5772         /* step 7: process the segment text */
5773         switch (sk->sk_state) {
5774         case TCP_CLOSE_WAIT:
5775         case TCP_CLOSING:
5776         case TCP_LAST_ACK:
5777                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5778                         break;
5779         case TCP_FIN_WAIT1:
5780         case TCP_FIN_WAIT2:
5781                 /* RFC 793 says to queue data in these states,
5782                  * RFC 1122 says we MUST send a reset.
5783                  * BSD 4.4 also does reset.
5784                  */
5785                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5786                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5787                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5788                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5789                                 tcp_reset(sk);
5790                                 return 1;
5791                         }
5792                 }
5793                 /* Fall through */
5794         case TCP_ESTABLISHED:
5795                 tcp_data_queue(sk, skb);
5796                 queued = 1;
5797                 break;
5798         }
5799
5800         /* tcp_data could move socket to TIME-WAIT */
5801         if (sk->sk_state != TCP_CLOSE) {
5802                 tcp_data_snd_check(sk);
5803                 tcp_ack_snd_check(sk);
5804         }
5805
5806         if (!queued) {
5807 discard:
5808                 __kfree_skb(skb);
5809         }
5810         return 0;
5811 }
5812
5813 EXPORT_SYMBOL(sysctl_tcp_ecn);
5814 EXPORT_SYMBOL(sysctl_tcp_reordering);
5815 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5816 EXPORT_SYMBOL(tcp_parse_options);
5817 #ifdef CONFIG_TCP_MD5SIG
5818 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5819 #endif
5820 EXPORT_SYMBOL(tcp_rcv_established);
5821 EXPORT_SYMBOL(tcp_rcv_state_process);
5822 EXPORT_SYMBOL(tcp_initialize_rcv_mss);