Merge branch 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  */
20
21 /*
22  * Changes:
23  *              Pedro Roque     :       Fast Retransmit/Recovery.
24  *                                      Two receive queues.
25  *                                      Retransmit queue handled by TCP.
26  *                                      Better retransmit timer handling.
27  *                                      New congestion avoidance.
28  *                                      Header prediction.
29  *                                      Variable renaming.
30  *
31  *              Eric            :       Fast Retransmit.
32  *              Randy Scott     :       MSS option defines.
33  *              Eric Schenk     :       Fixes to slow start algorithm.
34  *              Eric Schenk     :       Yet another double ACK bug.
35  *              Eric Schenk     :       Delayed ACK bug fixes.
36  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
37  *              David S. Miller :       Don't allow zero congestion window.
38  *              Eric Schenk     :       Fix retransmitter so that it sends
39  *                                      next packet on ack of previous packet.
40  *              Andi Kleen      :       Moved open_request checking here
41  *                                      and process RSTs for open_requests.
42  *              Andi Kleen      :       Better prune_queue, and other fixes.
43  *              Andrey Savochkin:       Fix RTT measurements in the presence of
44  *                                      timestamps.
45  *              Andrey Savochkin:       Check sequence numbers correctly when
46  *                                      removing SACKs due to in sequence incoming
47  *                                      data segments.
48  *              Andi Kleen:             Make sure we never ack data there is not
49  *                                      enough room for. Also make this condition
50  *                                      a fatal error if it might still happen.
51  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
52  *                                      connections with MSS<min(MTU,ann. MSS)
53  *                                      work without delayed acks.
54  *              Andi Kleen:             Process packets with PSH set in the
55  *                                      fast path.
56  *              J Hadi Salim:           ECN support
57  *              Andrei Gurtov,
58  *              Pasi Sarolahti,
59  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
60  *                                      engine. Lots of bugs are found.
61  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
62  */
63
64 #include <linux/mm.h>
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
67 #include <linux/kernel.h>
68 #include <net/dst.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
108 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
109
110 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
111 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
112 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
113 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
114 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115
116 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
117 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118
119 /* Adapt the MSS value used to make delayed ack decision to the
120  * real world.
121  */
122 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
123 {
124         struct inet_connection_sock *icsk = inet_csk(sk);
125         const unsigned int lss = icsk->icsk_ack.last_seg_size;
126         unsigned int len;
127
128         icsk->icsk_ack.last_seg_size = 0;
129
130         /* skb->len may jitter because of SACKs, even if peer
131          * sends good full-sized frames.
132          */
133         len = skb_shinfo(skb)->gso_size ? : skb->len;
134         if (len >= icsk->icsk_ack.rcv_mss) {
135                 icsk->icsk_ack.rcv_mss = len;
136         } else {
137                 /* Otherwise, we make more careful check taking into account,
138                  * that SACKs block is variable.
139                  *
140                  * "len" is invariant segment length, including TCP header.
141                  */
142                 len += skb->data - skb_transport_header(skb);
143                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
144                     /* If PSH is not set, packet should be
145                      * full sized, provided peer TCP is not badly broken.
146                      * This observation (if it is correct 8)) allows
147                      * to handle super-low mtu links fairly.
148                      */
149                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
150                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
151                         /* Subtract also invariant (if peer is RFC compliant),
152                          * tcp header plus fixed timestamp option length.
153                          * Resulting "len" is MSS free of SACK jitter.
154                          */
155                         len -= tcp_sk(sk)->tcp_header_len;
156                         icsk->icsk_ack.last_seg_size = len;
157                         if (len == lss) {
158                                 icsk->icsk_ack.rcv_mss = len;
159                                 return;
160                         }
161                 }
162                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
163                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
164                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
165         }
166 }
167
168 static void tcp_incr_quickack(struct sock *sk)
169 {
170         struct inet_connection_sock *icsk = inet_csk(sk);
171         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
172
173         if (quickacks == 0)
174                 quickacks = 2;
175         if (quickacks > icsk->icsk_ack.quick)
176                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
177 }
178
179 void tcp_enter_quickack_mode(struct sock *sk)
180 {
181         struct inet_connection_sock *icsk = inet_csk(sk);
182         tcp_incr_quickack(sk);
183         icsk->icsk_ack.pingpong = 0;
184         icsk->icsk_ack.ato = TCP_ATO_MIN;
185 }
186
187 /* Send ACKs quickly, if "quick" count is not exhausted
188  * and the session is not interactive.
189  */
190
191 static inline int tcp_in_quickack_mode(const struct sock *sk)
192 {
193         const struct inet_connection_sock *icsk = inet_csk(sk);
194         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
195 }
196
197 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
198 {
199         if (tp->ecn_flags & TCP_ECN_OK)
200                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
201 }
202
203 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
204 {
205         if (tcp_hdr(skb)->cwr)
206                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
207 }
208
209 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
210 {
211         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
212 }
213
214 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
215 {
216         if (tp->ecn_flags & TCP_ECN_OK) {
217                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
218                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
219                 /* Funny extension: if ECT is not set on a segment,
220                  * it is surely retransmit. It is not in ECN RFC,
221                  * but Linux follows this rule. */
222                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
223                         tcp_enter_quickack_mode((struct sock *)tp);
224         }
225 }
226
227 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
228 {
229         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
230                 tp->ecn_flags &= ~TCP_ECN_OK;
231 }
232
233 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
234 {
235         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
236                 tp->ecn_flags &= ~TCP_ECN_OK;
237 }
238
239 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
240 {
241         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
242                 return 1;
243         return 0;
244 }
245
246 /* Buffer size and advertised window tuning.
247  *
248  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
249  */
250
251 static void tcp_fixup_sndbuf(struct sock *sk)
252 {
253         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
254                      sizeof(struct sk_buff);
255
256         if (sk->sk_sndbuf < 3 * sndmem)
257                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
258 }
259
260 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261  *
262  * All tcp_full_space() is split to two parts: "network" buffer, allocated
263  * forward and advertised in receiver window (tp->rcv_wnd) and
264  * "application buffer", required to isolate scheduling/application
265  * latencies from network.
266  * window_clamp is maximal advertised window. It can be less than
267  * tcp_full_space(), in this case tcp_full_space() - window_clamp
268  * is reserved for "application" buffer. The less window_clamp is
269  * the smoother our behaviour from viewpoint of network, but the lower
270  * throughput and the higher sensitivity of the connection to losses. 8)
271  *
272  * rcv_ssthresh is more strict window_clamp used at "slow start"
273  * phase to predict further behaviour of this connection.
274  * It is used for two goals:
275  * - to enforce header prediction at sender, even when application
276  *   requires some significant "application buffer". It is check #1.
277  * - to prevent pruning of receive queue because of misprediction
278  *   of receiver window. Check #2.
279  *
280  * The scheme does not work when sender sends good segments opening
281  * window and then starts to feed us spaghetti. But it should work
282  * in common situations. Otherwise, we have to rely on queue collapsing.
283  */
284
285 /* Slow part of check#2. */
286 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
287 {
288         struct tcp_sock *tp = tcp_sk(sk);
289         /* Optimize this! */
290         int truesize = tcp_win_from_space(skb->truesize) >> 1;
291         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
292
293         while (tp->rcv_ssthresh <= window) {
294                 if (truesize <= skb->len)
295                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
296
297                 truesize >>= 1;
298                 window >>= 1;
299         }
300         return 0;
301 }
302
303 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
304 {
305         struct tcp_sock *tp = tcp_sk(sk);
306
307         /* Check #1 */
308         if (tp->rcv_ssthresh < tp->window_clamp &&
309             (int)tp->rcv_ssthresh < tcp_space(sk) &&
310             !tcp_memory_pressure) {
311                 int incr;
312
313                 /* Check #2. Increase window, if skb with such overhead
314                  * will fit to rcvbuf in future.
315                  */
316                 if (tcp_win_from_space(skb->truesize) <= skb->len)
317                         incr = 2 * tp->advmss;
318                 else
319                         incr = __tcp_grow_window(sk, skb);
320
321                 if (incr) {
322                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
323                                                tp->window_clamp);
324                         inet_csk(sk)->icsk_ack.quick |= 1;
325                 }
326         }
327 }
328
329 /* 3. Tuning rcvbuf, when connection enters established state. */
330
331 static void tcp_fixup_rcvbuf(struct sock *sk)
332 {
333         struct tcp_sock *tp = tcp_sk(sk);
334         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
335
336         /* Try to select rcvbuf so that 4 mss-sized segments
337          * will fit to window and corresponding skbs will fit to our rcvbuf.
338          * (was 3; 4 is minimum to allow fast retransmit to work.)
339          */
340         while (tcp_win_from_space(rcvmem) < tp->advmss)
341                 rcvmem += 128;
342         if (sk->sk_rcvbuf < 4 * rcvmem)
343                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
344 }
345
346 /* 4. Try to fixup all. It is made immediately after connection enters
347  *    established state.
348  */
349 static void tcp_init_buffer_space(struct sock *sk)
350 {
351         struct tcp_sock *tp = tcp_sk(sk);
352         int maxwin;
353
354         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
355                 tcp_fixup_rcvbuf(sk);
356         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
357                 tcp_fixup_sndbuf(sk);
358
359         tp->rcvq_space.space = tp->rcv_wnd;
360
361         maxwin = tcp_full_space(sk);
362
363         if (tp->window_clamp >= maxwin) {
364                 tp->window_clamp = maxwin;
365
366                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
367                         tp->window_clamp = max(maxwin -
368                                                (maxwin >> sysctl_tcp_app_win),
369                                                4 * tp->advmss);
370         }
371
372         /* Force reservation of one segment. */
373         if (sysctl_tcp_app_win &&
374             tp->window_clamp > 2 * tp->advmss &&
375             tp->window_clamp + tp->advmss > maxwin)
376                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
377
378         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
379         tp->snd_cwnd_stamp = tcp_time_stamp;
380 }
381
382 /* 5. Recalculate window clamp after socket hit its memory bounds. */
383 static void tcp_clamp_window(struct sock *sk)
384 {
385         struct tcp_sock *tp = tcp_sk(sk);
386         struct inet_connection_sock *icsk = inet_csk(sk);
387
388         icsk->icsk_ack.quick = 0;
389
390         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
391             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
392             !tcp_memory_pressure &&
393             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
394                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
395                                     sysctl_tcp_rmem[2]);
396         }
397         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
398                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
399 }
400
401 /* Initialize RCV_MSS value.
402  * RCV_MSS is an our guess about MSS used by the peer.
403  * We haven't any direct information about the MSS.
404  * It's better to underestimate the RCV_MSS rather than overestimate.
405  * Overestimations make us ACKing less frequently than needed.
406  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407  */
408 void tcp_initialize_rcv_mss(struct sock *sk)
409 {
410         struct tcp_sock *tp = tcp_sk(sk);
411         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
412
413         hint = min(hint, tp->rcv_wnd / 2);
414         hint = min(hint, TCP_MIN_RCVMSS);
415         hint = max(hint, TCP_MIN_MSS);
416
417         inet_csk(sk)->icsk_ack.rcv_mss = hint;
418 }
419
420 /* Receiver "autotuning" code.
421  *
422  * The algorithm for RTT estimation w/o timestamps is based on
423  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
424  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425  *
426  * More detail on this code can be found at
427  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
428  * though this reference is out of date.  A new paper
429  * is pending.
430  */
431 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
432 {
433         u32 new_sample = tp->rcv_rtt_est.rtt;
434         long m = sample;
435
436         if (m == 0)
437                 m = 1;
438
439         if (new_sample != 0) {
440                 /* If we sample in larger samples in the non-timestamp
441                  * case, we could grossly overestimate the RTT especially
442                  * with chatty applications or bulk transfer apps which
443                  * are stalled on filesystem I/O.
444                  *
445                  * Also, since we are only going for a minimum in the
446                  * non-timestamp case, we do not smooth things out
447                  * else with timestamps disabled convergence takes too
448                  * long.
449                  */
450                 if (!win_dep) {
451                         m -= (new_sample >> 3);
452                         new_sample += m;
453                 } else if (m < new_sample)
454                         new_sample = m << 3;
455         } else {
456                 /* No previous measure. */
457                 new_sample = m << 3;
458         }
459
460         if (tp->rcv_rtt_est.rtt != new_sample)
461                 tp->rcv_rtt_est.rtt = new_sample;
462 }
463
464 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
465 {
466         if (tp->rcv_rtt_est.time == 0)
467                 goto new_measure;
468         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
469                 return;
470         tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
471
472 new_measure:
473         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
474         tp->rcv_rtt_est.time = tcp_time_stamp;
475 }
476
477 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
478                                           const struct sk_buff *skb)
479 {
480         struct tcp_sock *tp = tcp_sk(sk);
481         if (tp->rx_opt.rcv_tsecr &&
482             (TCP_SKB_CB(skb)->end_seq -
483              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
484                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
485 }
486
487 /*
488  * This function should be called every time data is copied to user space.
489  * It calculates the appropriate TCP receive buffer space.
490  */
491 void tcp_rcv_space_adjust(struct sock *sk)
492 {
493         struct tcp_sock *tp = tcp_sk(sk);
494         int time;
495         int space;
496
497         if (tp->rcvq_space.time == 0)
498                 goto new_measure;
499
500         time = tcp_time_stamp - tp->rcvq_space.time;
501         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
502                 return;
503
504         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
505
506         space = max(tp->rcvq_space.space, space);
507
508         if (tp->rcvq_space.space != space) {
509                 int rcvmem;
510
511                 tp->rcvq_space.space = space;
512
513                 if (sysctl_tcp_moderate_rcvbuf &&
514                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
515                         int new_clamp = space;
516
517                         /* Receive space grows, normalize in order to
518                          * take into account packet headers and sk_buff
519                          * structure overhead.
520                          */
521                         space /= tp->advmss;
522                         if (!space)
523                                 space = 1;
524                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
525                                   16 + sizeof(struct sk_buff));
526                         while (tcp_win_from_space(rcvmem) < tp->advmss)
527                                 rcvmem += 128;
528                         space *= rcvmem;
529                         space = min(space, sysctl_tcp_rmem[2]);
530                         if (space > sk->sk_rcvbuf) {
531                                 sk->sk_rcvbuf = space;
532
533                                 /* Make the window clamp follow along.  */
534                                 tp->window_clamp = new_clamp;
535                         }
536                 }
537         }
538
539 new_measure:
540         tp->rcvq_space.seq = tp->copied_seq;
541         tp->rcvq_space.time = tcp_time_stamp;
542 }
543
544 /* There is something which you must keep in mind when you analyze the
545  * behavior of the tp->ato delayed ack timeout interval.  When a
546  * connection starts up, we want to ack as quickly as possible.  The
547  * problem is that "good" TCP's do slow start at the beginning of data
548  * transmission.  The means that until we send the first few ACK's the
549  * sender will sit on his end and only queue most of his data, because
550  * he can only send snd_cwnd unacked packets at any given time.  For
551  * each ACK we send, he increments snd_cwnd and transmits more of his
552  * queue.  -DaveM
553  */
554 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
555 {
556         struct tcp_sock *tp = tcp_sk(sk);
557         struct inet_connection_sock *icsk = inet_csk(sk);
558         u32 now;
559
560         inet_csk_schedule_ack(sk);
561
562         tcp_measure_rcv_mss(sk, skb);
563
564         tcp_rcv_rtt_measure(tp);
565
566         now = tcp_time_stamp;
567
568         if (!icsk->icsk_ack.ato) {
569                 /* The _first_ data packet received, initialize
570                  * delayed ACK engine.
571                  */
572                 tcp_incr_quickack(sk);
573                 icsk->icsk_ack.ato = TCP_ATO_MIN;
574         } else {
575                 int m = now - icsk->icsk_ack.lrcvtime;
576
577                 if (m <= TCP_ATO_MIN / 2) {
578                         /* The fastest case is the first. */
579                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
580                 } else if (m < icsk->icsk_ack.ato) {
581                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
582                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
583                                 icsk->icsk_ack.ato = icsk->icsk_rto;
584                 } else if (m > icsk->icsk_rto) {
585                         /* Too long gap. Apparently sender failed to
586                          * restart window, so that we send ACKs quickly.
587                          */
588                         tcp_incr_quickack(sk);
589                         sk_mem_reclaim(sk);
590                 }
591         }
592         icsk->icsk_ack.lrcvtime = now;
593
594         TCP_ECN_check_ce(tp, skb);
595
596         if (skb->len >= 128)
597                 tcp_grow_window(sk, skb);
598 }
599
600 static u32 tcp_rto_min(struct sock *sk)
601 {
602         struct dst_entry *dst = __sk_dst_get(sk);
603         u32 rto_min = TCP_RTO_MIN;
604
605         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
606                 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
607         return rto_min;
608 }
609
610 /* Called to compute a smoothed rtt estimate. The data fed to this
611  * routine either comes from timestamps, or from segments that were
612  * known _not_ to have been retransmitted [see Karn/Partridge
613  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
614  * piece by Van Jacobson.
615  * NOTE: the next three routines used to be one big routine.
616  * To save cycles in the RFC 1323 implementation it was better to break
617  * it up into three procedures. -- erics
618  */
619 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
620 {
621         struct tcp_sock *tp = tcp_sk(sk);
622         long m = mrtt; /* RTT */
623
624         /*      The following amusing code comes from Jacobson's
625          *      article in SIGCOMM '88.  Note that rtt and mdev
626          *      are scaled versions of rtt and mean deviation.
627          *      This is designed to be as fast as possible
628          *      m stands for "measurement".
629          *
630          *      On a 1990 paper the rto value is changed to:
631          *      RTO = rtt + 4 * mdev
632          *
633          * Funny. This algorithm seems to be very broken.
634          * These formulae increase RTO, when it should be decreased, increase
635          * too slowly, when it should be increased quickly, decrease too quickly
636          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
637          * does not matter how to _calculate_ it. Seems, it was trap
638          * that VJ failed to avoid. 8)
639          */
640         if (m == 0)
641                 m = 1;
642         if (tp->srtt != 0) {
643                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
644                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
645                 if (m < 0) {
646                         m = -m;         /* m is now abs(error) */
647                         m -= (tp->mdev >> 2);   /* similar update on mdev */
648                         /* This is similar to one of Eifel findings.
649                          * Eifel blocks mdev updates when rtt decreases.
650                          * This solution is a bit different: we use finer gain
651                          * for mdev in this case (alpha*beta).
652                          * Like Eifel it also prevents growth of rto,
653                          * but also it limits too fast rto decreases,
654                          * happening in pure Eifel.
655                          */
656                         if (m > 0)
657                                 m >>= 3;
658                 } else {
659                         m -= (tp->mdev >> 2);   /* similar update on mdev */
660                 }
661                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
662                 if (tp->mdev > tp->mdev_max) {
663                         tp->mdev_max = tp->mdev;
664                         if (tp->mdev_max > tp->rttvar)
665                                 tp->rttvar = tp->mdev_max;
666                 }
667                 if (after(tp->snd_una, tp->rtt_seq)) {
668                         if (tp->mdev_max < tp->rttvar)
669                                 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
670                         tp->rtt_seq = tp->snd_nxt;
671                         tp->mdev_max = tcp_rto_min(sk);
672                 }
673         } else {
674                 /* no previous measure. */
675                 tp->srtt = m << 3;      /* take the measured time to be rtt */
676                 tp->mdev = m << 1;      /* make sure rto = 3*rtt */
677                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
678                 tp->rtt_seq = tp->snd_nxt;
679         }
680 }
681
682 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
683  * routine referred to above.
684  */
685 static inline void tcp_set_rto(struct sock *sk)
686 {
687         const struct tcp_sock *tp = tcp_sk(sk);
688         /* Old crap is replaced with new one. 8)
689          *
690          * More seriously:
691          * 1. If rtt variance happened to be less 50msec, it is hallucination.
692          *    It cannot be less due to utterly erratic ACK generation made
693          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
694          *    to do with delayed acks, because at cwnd>2 true delack timeout
695          *    is invisible. Actually, Linux-2.4 also generates erratic
696          *    ACKs in some circumstances.
697          */
698         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
699
700         /* 2. Fixups made earlier cannot be right.
701          *    If we do not estimate RTO correctly without them,
702          *    all the algo is pure shit and should be replaced
703          *    with correct one. It is exactly, which we pretend to do.
704          */
705
706         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707          * guarantees that rto is higher.
708          */
709         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
710                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
711 }
712
713 /* Save metrics learned by this TCP session.
714    This function is called only, when TCP finishes successfully
715    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
716  */
717 void tcp_update_metrics(struct sock *sk)
718 {
719         struct tcp_sock *tp = tcp_sk(sk);
720         struct dst_entry *dst = __sk_dst_get(sk);
721
722         if (sysctl_tcp_nometrics_save)
723                 return;
724
725         dst_confirm(dst);
726
727         if (dst && (dst->flags & DST_HOST)) {
728                 const struct inet_connection_sock *icsk = inet_csk(sk);
729                 int m;
730                 unsigned long rtt;
731
732                 if (icsk->icsk_backoff || !tp->srtt) {
733                         /* This session failed to estimate rtt. Why?
734                          * Probably, no packets returned in time.
735                          * Reset our results.
736                          */
737                         if (!(dst_metric_locked(dst, RTAX_RTT)))
738                                 dst->metrics[RTAX_RTT - 1] = 0;
739                         return;
740                 }
741
742                 rtt = dst_metric_rtt(dst, RTAX_RTT);
743                 m = rtt - tp->srtt;
744
745                 /* If newly calculated rtt larger than stored one,
746                  * store new one. Otherwise, use EWMA. Remember,
747                  * rtt overestimation is always better than underestimation.
748                  */
749                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
750                         if (m <= 0)
751                                 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
752                         else
753                                 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
754                 }
755
756                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
757                         unsigned long var;
758                         if (m < 0)
759                                 m = -m;
760
761                         /* Scale deviation to rttvar fixed point */
762                         m >>= 1;
763                         if (m < tp->mdev)
764                                 m = tp->mdev;
765
766                         var = dst_metric_rtt(dst, RTAX_RTTVAR);
767                         if (m >= var)
768                                 var = m;
769                         else
770                                 var -= (var - m) >> 2;
771
772                         set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
773                 }
774
775                 if (tp->snd_ssthresh >= 0xFFFF) {
776                         /* Slow start still did not finish. */
777                         if (dst_metric(dst, RTAX_SSTHRESH) &&
778                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
779                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
780                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
781                         if (!dst_metric_locked(dst, RTAX_CWND) &&
782                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
783                                 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
784                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
785                            icsk->icsk_ca_state == TCP_CA_Open) {
786                         /* Cong. avoidance phase, cwnd is reliable. */
787                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
788                                 dst->metrics[RTAX_SSTHRESH-1] =
789                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
790                         if (!dst_metric_locked(dst, RTAX_CWND))
791                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
792                 } else {
793                         /* Else slow start did not finish, cwnd is non-sense,
794                            ssthresh may be also invalid.
795                          */
796                         if (!dst_metric_locked(dst, RTAX_CWND))
797                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
798                         if (dst_metric(dst, RTAX_SSTHRESH) &&
799                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
800                             tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
801                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
802                 }
803
804                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
805                         if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
806                             tp->reordering != sysctl_tcp_reordering)
807                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
808                 }
809         }
810 }
811
812 /* Numbers are taken from RFC3390.
813  *
814  * John Heffner states:
815  *
816  *      The RFC specifies a window of no more than 4380 bytes
817  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
818  *      is a bit misleading because they use a clamp at 4380 bytes
819  *      rather than use a multiplier in the relevant range.
820  */
821 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
822 {
823         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
824
825         if (!cwnd) {
826                 if (tp->mss_cache > 1460)
827                         cwnd = 2;
828                 else
829                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
830         }
831         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
832 }
833
834 /* Set slow start threshold and cwnd not falling to slow start */
835 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
836 {
837         struct tcp_sock *tp = tcp_sk(sk);
838         const struct inet_connection_sock *icsk = inet_csk(sk);
839
840         tp->prior_ssthresh = 0;
841         tp->bytes_acked = 0;
842         if (icsk->icsk_ca_state < TCP_CA_CWR) {
843                 tp->undo_marker = 0;
844                 if (set_ssthresh)
845                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
846                 tp->snd_cwnd = min(tp->snd_cwnd,
847                                    tcp_packets_in_flight(tp) + 1U);
848                 tp->snd_cwnd_cnt = 0;
849                 tp->high_seq = tp->snd_nxt;
850                 tp->snd_cwnd_stamp = tcp_time_stamp;
851                 TCP_ECN_queue_cwr(tp);
852
853                 tcp_set_ca_state(sk, TCP_CA_CWR);
854         }
855 }
856
857 /*
858  * Packet counting of FACK is based on in-order assumptions, therefore TCP
859  * disables it when reordering is detected
860  */
861 static void tcp_disable_fack(struct tcp_sock *tp)
862 {
863         /* RFC3517 uses different metric in lost marker => reset on change */
864         if (tcp_is_fack(tp))
865                 tp->lost_skb_hint = NULL;
866         tp->rx_opt.sack_ok &= ~2;
867 }
868
869 /* Take a notice that peer is sending D-SACKs */
870 static void tcp_dsack_seen(struct tcp_sock *tp)
871 {
872         tp->rx_opt.sack_ok |= 4;
873 }
874
875 /* Initialize metrics on socket. */
876
877 static void tcp_init_metrics(struct sock *sk)
878 {
879         struct tcp_sock *tp = tcp_sk(sk);
880         struct dst_entry *dst = __sk_dst_get(sk);
881
882         if (dst == NULL)
883                 goto reset;
884
885         dst_confirm(dst);
886
887         if (dst_metric_locked(dst, RTAX_CWND))
888                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889         if (dst_metric(dst, RTAX_SSTHRESH)) {
890                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
893         }
894         if (dst_metric(dst, RTAX_REORDERING) &&
895             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
896                 tcp_disable_fack(tp);
897                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
898         }
899
900         if (dst_metric(dst, RTAX_RTT) == 0)
901                 goto reset;
902
903         if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
904                 goto reset;
905
906         /* Initial rtt is determined from SYN,SYN-ACK.
907          * The segment is small and rtt may appear much
908          * less than real one. Use per-dst memory
909          * to make it more realistic.
910          *
911          * A bit of theory. RTT is time passed after "normal" sized packet
912          * is sent until it is ACKed. In normal circumstances sending small
913          * packets force peer to delay ACKs and calculation is correct too.
914          * The algorithm is adaptive and, provided we follow specs, it
915          * NEVER underestimate RTT. BUT! If peer tries to make some clever
916          * tricks sort of "quick acks" for time long enough to decrease RTT
917          * to low value, and then abruptly stops to do it and starts to delay
918          * ACKs, wait for troubles.
919          */
920         if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
921                 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
922                 tp->rtt_seq = tp->snd_nxt;
923         }
924         if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
925                 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
926                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
927         }
928         tcp_set_rto(sk);
929         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
930                 goto reset;
931
932 cwnd:
933         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
934         tp->snd_cwnd_stamp = tcp_time_stamp;
935         return;
936
937 reset:
938         /* Play conservative. If timestamps are not
939          * supported, TCP will fail to recalculate correct
940          * rtt, if initial rto is too small. FORGET ALL AND RESET!
941          */
942         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
943                 tp->srtt = 0;
944                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
945                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
946         }
947         goto cwnd;
948 }
949
950 static void tcp_update_reordering(struct sock *sk, const int metric,
951                                   const int ts)
952 {
953         struct tcp_sock *tp = tcp_sk(sk);
954         if (metric > tp->reordering) {
955                 int mib_idx;
956
957                 tp->reordering = min(TCP_MAX_REORDERING, metric);
958
959                 /* This exciting event is worth to be remembered. 8) */
960                 if (ts)
961                         mib_idx = LINUX_MIB_TCPTSREORDER;
962                 else if (tcp_is_reno(tp))
963                         mib_idx = LINUX_MIB_TCPRENOREORDER;
964                 else if (tcp_is_fack(tp))
965                         mib_idx = LINUX_MIB_TCPFACKREORDER;
966                 else
967                         mib_idx = LINUX_MIB_TCPSACKREORDER;
968
969                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
970 #if FASTRETRANS_DEBUG > 1
971                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
972                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
973                        tp->reordering,
974                        tp->fackets_out,
975                        tp->sacked_out,
976                        tp->undo_marker ? tp->undo_retrans : 0);
977 #endif
978                 tcp_disable_fack(tp);
979         }
980 }
981
982 /* This must be called before lost_out is incremented */
983 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
984 {
985         if ((tp->retransmit_skb_hint == NULL) ||
986             before(TCP_SKB_CB(skb)->seq,
987                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
988                 tp->retransmit_skb_hint = skb;
989
990         if (!tp->lost_out ||
991             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
992                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
993 }
994
995 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
996 {
997         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
998                 tcp_verify_retransmit_hint(tp, skb);
999
1000                 tp->lost_out += tcp_skb_pcount(skb);
1001                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1002         }
1003 }
1004
1005 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1006                                             struct sk_buff *skb)
1007 {
1008         tcp_verify_retransmit_hint(tp, skb);
1009
1010         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1011                 tp->lost_out += tcp_skb_pcount(skb);
1012                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1013         }
1014 }
1015
1016 /* This procedure tags the retransmission queue when SACKs arrive.
1017  *
1018  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1019  * Packets in queue with these bits set are counted in variables
1020  * sacked_out, retrans_out and lost_out, correspondingly.
1021  *
1022  * Valid combinations are:
1023  * Tag  InFlight        Description
1024  * 0    1               - orig segment is in flight.
1025  * S    0               - nothing flies, orig reached receiver.
1026  * L    0               - nothing flies, orig lost by net.
1027  * R    2               - both orig and retransmit are in flight.
1028  * L|R  1               - orig is lost, retransmit is in flight.
1029  * S|R  1               - orig reached receiver, retrans is still in flight.
1030  * (L|S|R is logically valid, it could occur when L|R is sacked,
1031  *  but it is equivalent to plain S and code short-curcuits it to S.
1032  *  L|S is logically invalid, it would mean -1 packet in flight 8))
1033  *
1034  * These 6 states form finite state machine, controlled by the following events:
1035  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1036  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1037  * 3. Loss detection event of one of three flavors:
1038  *      A. Scoreboard estimator decided the packet is lost.
1039  *         A'. Reno "three dupacks" marks head of queue lost.
1040  *         A''. Its FACK modfication, head until snd.fack is lost.
1041  *      B. SACK arrives sacking data transmitted after never retransmitted
1042  *         hole was sent out.
1043  *      C. SACK arrives sacking SND.NXT at the moment, when the
1044  *         segment was retransmitted.
1045  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1046  *
1047  * It is pleasant to note, that state diagram turns out to be commutative,
1048  * so that we are allowed not to be bothered by order of our actions,
1049  * when multiple events arrive simultaneously. (see the function below).
1050  *
1051  * Reordering detection.
1052  * --------------------
1053  * Reordering metric is maximal distance, which a packet can be displaced
1054  * in packet stream. With SACKs we can estimate it:
1055  *
1056  * 1. SACK fills old hole and the corresponding segment was not
1057  *    ever retransmitted -> reordering. Alas, we cannot use it
1058  *    when segment was retransmitted.
1059  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1060  *    for retransmitted and already SACKed segment -> reordering..
1061  * Both of these heuristics are not used in Loss state, when we cannot
1062  * account for retransmits accurately.
1063  *
1064  * SACK block validation.
1065  * ----------------------
1066  *
1067  * SACK block range validation checks that the received SACK block fits to
1068  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1069  * Note that SND.UNA is not included to the range though being valid because
1070  * it means that the receiver is rather inconsistent with itself reporting
1071  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1072  * perfectly valid, however, in light of RFC2018 which explicitly states
1073  * that "SACK block MUST reflect the newest segment.  Even if the newest
1074  * segment is going to be discarded ...", not that it looks very clever
1075  * in case of head skb. Due to potentional receiver driven attacks, we
1076  * choose to avoid immediate execution of a walk in write queue due to
1077  * reneging and defer head skb's loss recovery to standard loss recovery
1078  * procedure that will eventually trigger (nothing forbids us doing this).
1079  *
1080  * Implements also blockage to start_seq wrap-around. Problem lies in the
1081  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1082  * there's no guarantee that it will be before snd_nxt (n). The problem
1083  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1084  * wrap (s_w):
1085  *
1086  *         <- outs wnd ->                          <- wrapzone ->
1087  *         u     e      n                         u_w   e_w  s n_w
1088  *         |     |      |                          |     |   |  |
1089  * |<------------+------+----- TCP seqno space --------------+---------->|
1090  * ...-- <2^31 ->|                                           |<--------...
1091  * ...---- >2^31 ------>|                                    |<--------...
1092  *
1093  * Current code wouldn't be vulnerable but it's better still to discard such
1094  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1095  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1096  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1097  * equal to the ideal case (infinite seqno space without wrap caused issues).
1098  *
1099  * With D-SACK the lower bound is extended to cover sequence space below
1100  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1101  * again, D-SACK block must not to go across snd_una (for the same reason as
1102  * for the normal SACK blocks, explained above). But there all simplicity
1103  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1104  * fully below undo_marker they do not affect behavior in anyway and can
1105  * therefore be safely ignored. In rare cases (which are more or less
1106  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1107  * fragmentation and packet reordering past skb's retransmission. To consider
1108  * them correctly, the acceptable range must be extended even more though
1109  * the exact amount is rather hard to quantify. However, tp->max_window can
1110  * be used as an exaggerated estimate.
1111  */
1112 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1113                                   u32 start_seq, u32 end_seq)
1114 {
1115         /* Too far in future, or reversed (interpretation is ambiguous) */
1116         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1117                 return 0;
1118
1119         /* Nasty start_seq wrap-around check (see comments above) */
1120         if (!before(start_seq, tp->snd_nxt))
1121                 return 0;
1122
1123         /* In outstanding window? ...This is valid exit for D-SACKs too.
1124          * start_seq == snd_una is non-sensical (see comments above)
1125          */
1126         if (after(start_seq, tp->snd_una))
1127                 return 1;
1128
1129         if (!is_dsack || !tp->undo_marker)
1130                 return 0;
1131
1132         /* ...Then it's D-SACK, and must reside below snd_una completely */
1133         if (!after(end_seq, tp->snd_una))
1134                 return 0;
1135
1136         if (!before(start_seq, tp->undo_marker))
1137                 return 1;
1138
1139         /* Too old */
1140         if (!after(end_seq, tp->undo_marker))
1141                 return 0;
1142
1143         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1144          *   start_seq < undo_marker and end_seq >= undo_marker.
1145          */
1146         return !before(start_seq, end_seq - tp->max_window);
1147 }
1148
1149 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1150  * Event "C". Later note: FACK people cheated me again 8), we have to account
1151  * for reordering! Ugly, but should help.
1152  *
1153  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1154  * less than what is now known to be received by the other end (derived from
1155  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1156  * retransmitted skbs to avoid some costly processing per ACKs.
1157  */
1158 static void tcp_mark_lost_retrans(struct sock *sk)
1159 {
1160         const struct inet_connection_sock *icsk = inet_csk(sk);
1161         struct tcp_sock *tp = tcp_sk(sk);
1162         struct sk_buff *skb;
1163         int cnt = 0;
1164         u32 new_low_seq = tp->snd_nxt;
1165         u32 received_upto = tcp_highest_sack_seq(tp);
1166
1167         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1168             !after(received_upto, tp->lost_retrans_low) ||
1169             icsk->icsk_ca_state != TCP_CA_Recovery)
1170                 return;
1171
1172         tcp_for_write_queue(skb, sk) {
1173                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1174
1175                 if (skb == tcp_send_head(sk))
1176                         break;
1177                 if (cnt == tp->retrans_out)
1178                         break;
1179                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1180                         continue;
1181
1182                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1183                         continue;
1184
1185                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1186                  * constraint here (see above) but figuring out that at
1187                  * least tp->reordering SACK blocks reside between ack_seq
1188                  * and received_upto is not easy task to do cheaply with
1189                  * the available datastructures.
1190                  *
1191                  * Whether FACK should check here for tp->reordering segs
1192                  * in-between one could argue for either way (it would be
1193                  * rather simple to implement as we could count fack_count
1194                  * during the walk and do tp->fackets_out - fack_count).
1195                  */
1196                 if (after(received_upto, ack_seq)) {
1197                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1198                         tp->retrans_out -= tcp_skb_pcount(skb);
1199
1200                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1201                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1202                 } else {
1203                         if (before(ack_seq, new_low_seq))
1204                                 new_low_seq = ack_seq;
1205                         cnt += tcp_skb_pcount(skb);
1206                 }
1207         }
1208
1209         if (tp->retrans_out)
1210                 tp->lost_retrans_low = new_low_seq;
1211 }
1212
1213 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1214                            struct tcp_sack_block_wire *sp, int num_sacks,
1215                            u32 prior_snd_una)
1216 {
1217         struct tcp_sock *tp = tcp_sk(sk);
1218         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1219         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1220         int dup_sack = 0;
1221
1222         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1223                 dup_sack = 1;
1224                 tcp_dsack_seen(tp);
1225                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1226         } else if (num_sacks > 1) {
1227                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1228                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1229
1230                 if (!after(end_seq_0, end_seq_1) &&
1231                     !before(start_seq_0, start_seq_1)) {
1232                         dup_sack = 1;
1233                         tcp_dsack_seen(tp);
1234                         NET_INC_STATS_BH(sock_net(sk),
1235                                         LINUX_MIB_TCPDSACKOFORECV);
1236                 }
1237         }
1238
1239         /* D-SACK for already forgotten data... Do dumb counting. */
1240         if (dup_sack &&
1241             !after(end_seq_0, prior_snd_una) &&
1242             after(end_seq_0, tp->undo_marker))
1243                 tp->undo_retrans--;
1244
1245         return dup_sack;
1246 }
1247
1248 struct tcp_sacktag_state {
1249         int reord;
1250         int fack_count;
1251         int flag;
1252 };
1253
1254 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1255  * the incoming SACK may not exactly match but we can find smaller MSS
1256  * aligned portion of it that matches. Therefore we might need to fragment
1257  * which may fail and creates some hassle (caller must handle error case
1258  * returns).
1259  *
1260  * FIXME: this could be merged to shift decision code
1261  */
1262 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1263                                  u32 start_seq, u32 end_seq)
1264 {
1265         int in_sack, err;
1266         unsigned int pkt_len;
1267         unsigned int mss;
1268
1269         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1270                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1271
1272         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1273             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1274                 mss = tcp_skb_mss(skb);
1275                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1276
1277                 if (!in_sack) {
1278                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1279                         if (pkt_len < mss)
1280                                 pkt_len = mss;
1281                 } else {
1282                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1283                         if (pkt_len < mss)
1284                                 return -EINVAL;
1285                 }
1286
1287                 /* Round if necessary so that SACKs cover only full MSSes
1288                  * and/or the remaining small portion (if present)
1289                  */
1290                 if (pkt_len > mss) {
1291                         unsigned int new_len = (pkt_len / mss) * mss;
1292                         if (!in_sack && new_len < pkt_len) {
1293                                 new_len += mss;
1294                                 if (new_len > skb->len)
1295                                         return 0;
1296                         }
1297                         pkt_len = new_len;
1298                 }
1299                 err = tcp_fragment(sk, skb, pkt_len, mss);
1300                 if (err < 0)
1301                         return err;
1302         }
1303
1304         return in_sack;
1305 }
1306
1307 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1308                           struct tcp_sacktag_state *state,
1309                           int dup_sack, int pcount)
1310 {
1311         struct tcp_sock *tp = tcp_sk(sk);
1312         u8 sacked = TCP_SKB_CB(skb)->sacked;
1313         int fack_count = state->fack_count;
1314
1315         /* Account D-SACK for retransmitted packet. */
1316         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1317                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1318                         tp->undo_retrans--;
1319                 if (sacked & TCPCB_SACKED_ACKED)
1320                         state->reord = min(fack_count, state->reord);
1321         }
1322
1323         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1324         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1325                 return sacked;
1326
1327         if (!(sacked & TCPCB_SACKED_ACKED)) {
1328                 if (sacked & TCPCB_SACKED_RETRANS) {
1329                         /* If the segment is not tagged as lost,
1330                          * we do not clear RETRANS, believing
1331                          * that retransmission is still in flight.
1332                          */
1333                         if (sacked & TCPCB_LOST) {
1334                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1335                                 tp->lost_out -= pcount;
1336                                 tp->retrans_out -= pcount;
1337                         }
1338                 } else {
1339                         if (!(sacked & TCPCB_RETRANS)) {
1340                                 /* New sack for not retransmitted frame,
1341                                  * which was in hole. It is reordering.
1342                                  */
1343                                 if (before(TCP_SKB_CB(skb)->seq,
1344                                            tcp_highest_sack_seq(tp)))
1345                                         state->reord = min(fack_count,
1346                                                            state->reord);
1347
1348                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1349                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1350                                         state->flag |= FLAG_ONLY_ORIG_SACKED;
1351                         }
1352
1353                         if (sacked & TCPCB_LOST) {
1354                                 sacked &= ~TCPCB_LOST;
1355                                 tp->lost_out -= pcount;
1356                         }
1357                 }
1358
1359                 sacked |= TCPCB_SACKED_ACKED;
1360                 state->flag |= FLAG_DATA_SACKED;
1361                 tp->sacked_out += pcount;
1362
1363                 fack_count += pcount;
1364
1365                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1366                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1367                     before(TCP_SKB_CB(skb)->seq,
1368                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1369                         tp->lost_cnt_hint += pcount;
1370
1371                 if (fack_count > tp->fackets_out)
1372                         tp->fackets_out = fack_count;
1373         }
1374
1375         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1376          * frames and clear it. undo_retrans is decreased above, L|R frames
1377          * are accounted above as well.
1378          */
1379         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1380                 sacked &= ~TCPCB_SACKED_RETRANS;
1381                 tp->retrans_out -= pcount;
1382         }
1383
1384         return sacked;
1385 }
1386
1387 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1388                            struct tcp_sacktag_state *state,
1389                            unsigned int pcount, int shifted, int mss,
1390                            int dup_sack)
1391 {
1392         struct tcp_sock *tp = tcp_sk(sk);
1393         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1394
1395         BUG_ON(!pcount);
1396
1397         /* Tweak before seqno plays */
1398         if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1399             !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1400                 tp->lost_cnt_hint += pcount;
1401
1402         TCP_SKB_CB(prev)->end_seq += shifted;
1403         TCP_SKB_CB(skb)->seq += shifted;
1404
1405         skb_shinfo(prev)->gso_segs += pcount;
1406         BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1407         skb_shinfo(skb)->gso_segs -= pcount;
1408
1409         /* When we're adding to gso_segs == 1, gso_size will be zero,
1410          * in theory this shouldn't be necessary but as long as DSACK
1411          * code can come after this skb later on it's better to keep
1412          * setting gso_size to something.
1413          */
1414         if (!skb_shinfo(prev)->gso_size) {
1415                 skb_shinfo(prev)->gso_size = mss;
1416                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1417         }
1418
1419         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1420         if (skb_shinfo(skb)->gso_segs <= 1) {
1421                 skb_shinfo(skb)->gso_size = 0;
1422                 skb_shinfo(skb)->gso_type = 0;
1423         }
1424
1425         /* We discard results */
1426         tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1427
1428         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1429         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1430
1431         if (skb->len > 0) {
1432                 BUG_ON(!tcp_skb_pcount(skb));
1433                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1434                 return 0;
1435         }
1436
1437         /* Whole SKB was eaten :-) */
1438
1439         if (skb == tp->retransmit_skb_hint)
1440                 tp->retransmit_skb_hint = prev;
1441         if (skb == tp->scoreboard_skb_hint)
1442                 tp->scoreboard_skb_hint = prev;
1443         if (skb == tp->lost_skb_hint) {
1444                 tp->lost_skb_hint = prev;
1445                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1446         }
1447
1448         TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1449         if (skb == tcp_highest_sack(sk))
1450                 tcp_advance_highest_sack(sk, skb);
1451
1452         tcp_unlink_write_queue(skb, sk);
1453         sk_wmem_free_skb(sk, skb);
1454
1455         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1456
1457         return 1;
1458 }
1459
1460 /* I wish gso_size would have a bit more sane initialization than
1461  * something-or-zero which complicates things
1462  */
1463 static int tcp_skb_seglen(struct sk_buff *skb)
1464 {
1465         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1466 }
1467
1468 /* Shifting pages past head area doesn't work */
1469 static int skb_can_shift(struct sk_buff *skb)
1470 {
1471         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1472 }
1473
1474 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1475  * skb.
1476  */
1477 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1478                                           struct tcp_sacktag_state *state,
1479                                           u32 start_seq, u32 end_seq,
1480                                           int dup_sack)
1481 {
1482         struct tcp_sock *tp = tcp_sk(sk);
1483         struct sk_buff *prev;
1484         int mss;
1485         int pcount = 0;
1486         int len;
1487         int in_sack;
1488
1489         if (!sk_can_gso(sk))
1490                 goto fallback;
1491
1492         /* Normally R but no L won't result in plain S */
1493         if (!dup_sack &&
1494             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1495                 goto fallback;
1496         if (!skb_can_shift(skb))
1497                 goto fallback;
1498         /* This frame is about to be dropped (was ACKed). */
1499         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1500                 goto fallback;
1501
1502         /* Can only happen with delayed DSACK + discard craziness */
1503         if (unlikely(skb == tcp_write_queue_head(sk)))
1504                 goto fallback;
1505         prev = tcp_write_queue_prev(sk, skb);
1506
1507         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1508                 goto fallback;
1509
1510         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1511                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1512
1513         if (in_sack) {
1514                 len = skb->len;
1515                 pcount = tcp_skb_pcount(skb);
1516                 mss = tcp_skb_seglen(skb);
1517
1518                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1519                  * drop this restriction as unnecessary
1520                  */
1521                 if (mss != tcp_skb_seglen(prev))
1522                         goto fallback;
1523         } else {
1524                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1525                         goto noop;
1526                 /* CHECKME: This is non-MSS split case only?, this will
1527                  * cause skipped skbs due to advancing loop btw, original
1528                  * has that feature too
1529                  */
1530                 if (tcp_skb_pcount(skb) <= 1)
1531                         goto noop;
1532
1533                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1534                 if (!in_sack) {
1535                         /* TODO: head merge to next could be attempted here
1536                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1537                          * though it might not be worth of the additional hassle
1538                          *
1539                          * ...we can probably just fallback to what was done
1540                          * previously. We could try merging non-SACKed ones
1541                          * as well but it probably isn't going to buy off
1542                          * because later SACKs might again split them, and
1543                          * it would make skb timestamp tracking considerably
1544                          * harder problem.
1545                          */
1546                         goto fallback;
1547                 }
1548
1549                 len = end_seq - TCP_SKB_CB(skb)->seq;
1550                 BUG_ON(len < 0);
1551                 BUG_ON(len > skb->len);
1552
1553                 /* MSS boundaries should be honoured or else pcount will
1554                  * severely break even though it makes things bit trickier.
1555                  * Optimize common case to avoid most of the divides
1556                  */
1557                 mss = tcp_skb_mss(skb);
1558
1559                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1560                  * drop this restriction as unnecessary
1561                  */
1562                 if (mss != tcp_skb_seglen(prev))
1563                         goto fallback;
1564
1565                 if (len == mss) {
1566                         pcount = 1;
1567                 } else if (len < mss) {
1568                         goto noop;
1569                 } else {
1570                         pcount = len / mss;
1571                         len = pcount * mss;
1572                 }
1573         }
1574
1575         if (!skb_shift(prev, skb, len))
1576                 goto fallback;
1577         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1578                 goto out;
1579
1580         /* Hole filled allows collapsing with the next as well, this is very
1581          * useful when hole on every nth skb pattern happens
1582          */
1583         if (prev == tcp_write_queue_tail(sk))
1584                 goto out;
1585         skb = tcp_write_queue_next(sk, prev);
1586
1587         if (!skb_can_shift(skb) ||
1588             (skb == tcp_send_head(sk)) ||
1589             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1590             (mss != tcp_skb_seglen(skb)))
1591                 goto out;
1592
1593         len = skb->len;
1594         if (skb_shift(prev, skb, len)) {
1595                 pcount += tcp_skb_pcount(skb);
1596                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1597         }
1598
1599 out:
1600         state->fack_count += pcount;
1601         return prev;
1602
1603 noop:
1604         return skb;
1605
1606 fallback:
1607         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1608         return NULL;
1609 }
1610
1611 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1612                                         struct tcp_sack_block *next_dup,
1613                                         struct tcp_sacktag_state *state,
1614                                         u32 start_seq, u32 end_seq,
1615                                         int dup_sack_in)
1616 {
1617         struct tcp_sock *tp = tcp_sk(sk);
1618         struct sk_buff *tmp;
1619
1620         tcp_for_write_queue_from(skb, sk) {
1621                 int in_sack = 0;
1622                 int dup_sack = dup_sack_in;
1623
1624                 if (skb == tcp_send_head(sk))
1625                         break;
1626
1627                 /* queue is in-order => we can short-circuit the walk early */
1628                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1629                         break;
1630
1631                 if ((next_dup != NULL) &&
1632                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1633                         in_sack = tcp_match_skb_to_sack(sk, skb,
1634                                                         next_dup->start_seq,
1635                                                         next_dup->end_seq);
1636                         if (in_sack > 0)
1637                                 dup_sack = 1;
1638                 }
1639
1640                 /* skb reference here is a bit tricky to get right, since
1641                  * shifting can eat and free both this skb and the next,
1642                  * so not even _safe variant of the loop is enough.
1643                  */
1644                 if (in_sack <= 0) {
1645                         tmp = tcp_shift_skb_data(sk, skb, state,
1646                                                  start_seq, end_seq, dup_sack);
1647                         if (tmp != NULL) {
1648                                 if (tmp != skb) {
1649                                         skb = tmp;
1650                                         continue;
1651                                 }
1652
1653                                 in_sack = 0;
1654                         } else {
1655                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1656                                                                 start_seq,
1657                                                                 end_seq);
1658                         }
1659                 }
1660
1661                 if (unlikely(in_sack < 0))
1662                         break;
1663
1664                 if (in_sack) {
1665                         TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1666                                                                   state,
1667                                                                   dup_sack,
1668                                                                   tcp_skb_pcount(skb));
1669
1670                         if (!before(TCP_SKB_CB(skb)->seq,
1671                                     tcp_highest_sack_seq(tp)))
1672                                 tcp_advance_highest_sack(sk, skb);
1673                 }
1674
1675                 state->fack_count += tcp_skb_pcount(skb);
1676         }
1677         return skb;
1678 }
1679
1680 /* Avoid all extra work that is being done by sacktag while walking in
1681  * a normal way
1682  */
1683 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1684                                         struct tcp_sacktag_state *state,
1685                                         u32 skip_to_seq)
1686 {
1687         tcp_for_write_queue_from(skb, sk) {
1688                 if (skb == tcp_send_head(sk))
1689                         break;
1690
1691                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1692                         break;
1693
1694                 state->fack_count += tcp_skb_pcount(skb);
1695         }
1696         return skb;
1697 }
1698
1699 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1700                                                 struct sock *sk,
1701                                                 struct tcp_sack_block *next_dup,
1702                                                 struct tcp_sacktag_state *state,
1703                                                 u32 skip_to_seq)
1704 {
1705         if (next_dup == NULL)
1706                 return skb;
1707
1708         if (before(next_dup->start_seq, skip_to_seq)) {
1709                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1710                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1711                                        next_dup->start_seq, next_dup->end_seq,
1712                                        1);
1713         }
1714
1715         return skb;
1716 }
1717
1718 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1719 {
1720         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1721 }
1722
1723 static int
1724 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1725                         u32 prior_snd_una)
1726 {
1727         const struct inet_connection_sock *icsk = inet_csk(sk);
1728         struct tcp_sock *tp = tcp_sk(sk);
1729         unsigned char *ptr = (skb_transport_header(ack_skb) +
1730                               TCP_SKB_CB(ack_skb)->sacked);
1731         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1732         struct tcp_sack_block sp[TCP_NUM_SACKS];
1733         struct tcp_sack_block *cache;
1734         struct tcp_sacktag_state state;
1735         struct sk_buff *skb;
1736         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1737         int used_sacks;
1738         int found_dup_sack = 0;
1739         int i, j;
1740         int first_sack_index;
1741
1742         state.flag = 0;
1743         state.reord = tp->packets_out;
1744
1745         if (!tp->sacked_out) {
1746                 if (WARN_ON(tp->fackets_out))
1747                         tp->fackets_out = 0;
1748                 tcp_highest_sack_reset(sk);
1749         }
1750
1751         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1752                                          num_sacks, prior_snd_una);
1753         if (found_dup_sack)
1754                 state.flag |= FLAG_DSACKING_ACK;
1755
1756         /* Eliminate too old ACKs, but take into
1757          * account more or less fresh ones, they can
1758          * contain valid SACK info.
1759          */
1760         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1761                 return 0;
1762
1763         if (!tp->packets_out)
1764                 goto out;
1765
1766         used_sacks = 0;
1767         first_sack_index = 0;
1768         for (i = 0; i < num_sacks; i++) {
1769                 int dup_sack = !i && found_dup_sack;
1770
1771                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1772                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1773
1774                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1775                                             sp[used_sacks].start_seq,
1776                                             sp[used_sacks].end_seq)) {
1777                         int mib_idx;
1778
1779                         if (dup_sack) {
1780                                 if (!tp->undo_marker)
1781                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1782                                 else
1783                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1784                         } else {
1785                                 /* Don't count olds caused by ACK reordering */
1786                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1787                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1788                                         continue;
1789                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1790                         }
1791
1792                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1793                         if (i == 0)
1794                                 first_sack_index = -1;
1795                         continue;
1796                 }
1797
1798                 /* Ignore very old stuff early */
1799                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1800                         continue;
1801
1802                 used_sacks++;
1803         }
1804
1805         /* order SACK blocks to allow in order walk of the retrans queue */
1806         for (i = used_sacks - 1; i > 0; i--) {
1807                 for (j = 0; j < i; j++) {
1808                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1809                                 swap(sp[j], sp[j + 1]);
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)
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);
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                 if (sacked & TCPCB_RETRANS) {
3223                         if (sacked & TCPCB_SACKED_RETRANS)
3224                                 tp->retrans_out -= acked_pcount;
3225                         flag |= FLAG_RETRANS_DATA_ACKED;
3226                         ca_seq_rtt = -1;
3227                         seq_rtt = -1;
3228                         if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3229                                 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3230                 } else {
3231                         ca_seq_rtt = now - scb->when;
3232                         last_ackt = skb->tstamp;
3233                         if (seq_rtt < 0) {
3234                                 seq_rtt = ca_seq_rtt;
3235                         }
3236                         if (!(sacked & TCPCB_SACKED_ACKED))
3237                                 reord = min(pkts_acked, reord);
3238                 }
3239
3240                 if (sacked & TCPCB_SACKED_ACKED)
3241                         tp->sacked_out -= acked_pcount;
3242                 if (sacked & TCPCB_LOST)
3243                         tp->lost_out -= acked_pcount;
3244
3245                 tp->packets_out -= acked_pcount;
3246                 pkts_acked += acked_pcount;
3247
3248                 /* Initial outgoing SYN's get put onto the write_queue
3249                  * just like anything else we transmit.  It is not
3250                  * true data, and if we misinform our callers that
3251                  * this ACK acks real data, we will erroneously exit
3252                  * connection startup slow start one packet too
3253                  * quickly.  This is severely frowned upon behavior.
3254                  */
3255                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3256                         flag |= FLAG_DATA_ACKED;
3257                 } else {
3258                         flag |= FLAG_SYN_ACKED;
3259                         tp->retrans_stamp = 0;
3260                 }
3261
3262                 if (!fully_acked)
3263                         break;
3264
3265                 tcp_unlink_write_queue(skb, sk);
3266                 sk_wmem_free_skb(sk, skb);
3267                 tp->scoreboard_skb_hint = NULL;
3268                 if (skb == tp->retransmit_skb_hint)
3269                         tp->retransmit_skb_hint = NULL;
3270                 if (skb == tp->lost_skb_hint)
3271                         tp->lost_skb_hint = NULL;
3272         }
3273
3274         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3275                 tp->snd_up = tp->snd_una;
3276
3277         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3278                 flag |= FLAG_SACK_RENEGING;
3279
3280         if (flag & FLAG_ACKED) {
3281                 const struct tcp_congestion_ops *ca_ops
3282                         = inet_csk(sk)->icsk_ca_ops;
3283
3284                 if (unlikely(icsk->icsk_mtup.probe_size &&
3285                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3286                         tcp_mtup_probe_success(sk);
3287                 }
3288
3289                 tcp_ack_update_rtt(sk, flag, seq_rtt);
3290                 tcp_rearm_rto(sk);
3291
3292                 if (tcp_is_reno(tp)) {
3293                         tcp_remove_reno_sacks(sk, pkts_acked);
3294                 } else {
3295                         int delta;
3296
3297                         /* Non-retransmitted hole got filled? That's reordering */
3298                         if (reord < prior_fackets)
3299                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3300
3301                         delta = tcp_is_fack(tp) ? pkts_acked :
3302                                                   prior_sacked - tp->sacked_out;
3303                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3304                 }
3305
3306                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3307
3308                 if (ca_ops->pkts_acked) {
3309                         s32 rtt_us = -1;
3310
3311                         /* Is the ACK triggering packet unambiguous? */
3312                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3313                                 /* High resolution needed and available? */
3314                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3315                                     !ktime_equal(last_ackt,
3316                                                  net_invalid_timestamp()))
3317                                         rtt_us = ktime_us_delta(ktime_get_real(),
3318                                                                 last_ackt);
3319                                 else if (ca_seq_rtt > 0)
3320                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
3321                         }
3322
3323                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3324                 }
3325         }
3326
3327 #if FASTRETRANS_DEBUG > 0
3328         WARN_ON((int)tp->sacked_out < 0);
3329         WARN_ON((int)tp->lost_out < 0);
3330         WARN_ON((int)tp->retrans_out < 0);
3331         if (!tp->packets_out && tcp_is_sack(tp)) {
3332                 icsk = inet_csk(sk);
3333                 if (tp->lost_out) {
3334                         printk(KERN_DEBUG "Leak l=%u %d\n",
3335                                tp->lost_out, icsk->icsk_ca_state);
3336                         tp->lost_out = 0;
3337                 }
3338                 if (tp->sacked_out) {
3339                         printk(KERN_DEBUG "Leak s=%u %d\n",
3340                                tp->sacked_out, icsk->icsk_ca_state);
3341                         tp->sacked_out = 0;
3342                 }
3343                 if (tp->retrans_out) {
3344                         printk(KERN_DEBUG "Leak r=%u %d\n",
3345                                tp->retrans_out, icsk->icsk_ca_state);
3346                         tp->retrans_out = 0;
3347                 }
3348         }
3349 #endif
3350         return flag;
3351 }
3352
3353 static void tcp_ack_probe(struct sock *sk)
3354 {
3355         const struct tcp_sock *tp = tcp_sk(sk);
3356         struct inet_connection_sock *icsk = inet_csk(sk);
3357
3358         /* Was it a usable window open? */
3359
3360         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3361                 icsk->icsk_backoff = 0;
3362                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3363                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3364                  * This function is not for random using!
3365                  */
3366         } else {
3367                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3368                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3369                                           TCP_RTO_MAX);
3370         }
3371 }
3372
3373 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3374 {
3375         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3376                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3377 }
3378
3379 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3380 {
3381         const struct tcp_sock *tp = tcp_sk(sk);
3382         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3383                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3384 }
3385
3386 /* Check that window update is acceptable.
3387  * The function assumes that snd_una<=ack<=snd_next.
3388  */
3389 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3390                                         const u32 ack, const u32 ack_seq,
3391                                         const u32 nwin)
3392 {
3393         return (after(ack, tp->snd_una) ||
3394                 after(ack_seq, tp->snd_wl1) ||
3395                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3396 }
3397
3398 /* Update our send window.
3399  *
3400  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3401  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3402  */
3403 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3404                                  u32 ack_seq)
3405 {
3406         struct tcp_sock *tp = tcp_sk(sk);
3407         int flag = 0;
3408         u32 nwin = ntohs(tcp_hdr(skb)->window);
3409
3410         if (likely(!tcp_hdr(skb)->syn))
3411                 nwin <<= tp->rx_opt.snd_wscale;
3412
3413         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3414                 flag |= FLAG_WIN_UPDATE;
3415                 tcp_update_wl(tp, ack_seq);
3416
3417                 if (tp->snd_wnd != nwin) {
3418                         tp->snd_wnd = nwin;
3419
3420                         /* Note, it is the only place, where
3421                          * fast path is recovered for sending TCP.
3422                          */
3423                         tp->pred_flags = 0;
3424                         tcp_fast_path_check(sk);
3425
3426                         if (nwin > tp->max_window) {
3427                                 tp->max_window = nwin;
3428                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3429                         }
3430                 }
3431         }
3432
3433         tp->snd_una = ack;
3434
3435         return flag;
3436 }
3437
3438 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3439  * continue in congestion avoidance.
3440  */
3441 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3442 {
3443         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3444         tp->snd_cwnd_cnt = 0;
3445         tp->bytes_acked = 0;
3446         TCP_ECN_queue_cwr(tp);
3447         tcp_moderate_cwnd(tp);
3448 }
3449
3450 /* A conservative spurious RTO response algorithm: reduce cwnd using
3451  * rate halving and continue in congestion avoidance.
3452  */
3453 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3454 {
3455         tcp_enter_cwr(sk, 0);
3456 }
3457
3458 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3459 {
3460         if (flag & FLAG_ECE)
3461                 tcp_ratehalving_spur_to_response(sk);
3462         else
3463                 tcp_undo_cwr(sk, 1);
3464 }
3465
3466 /* F-RTO spurious RTO detection algorithm (RFC4138)
3467  *
3468  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3469  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3470  * window (but not to or beyond highest sequence sent before RTO):
3471  *   On First ACK,  send two new segments out.
3472  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3473  *                  algorithm is not part of the F-RTO detection algorithm
3474  *                  given in RFC4138 but can be selected separately).
3475  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3476  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3477  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3478  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3479  *
3480  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3481  * original window even after we transmit two new data segments.
3482  *
3483  * SACK version:
3484  *   on first step, wait until first cumulative ACK arrives, then move to
3485  *   the second step. In second step, the next ACK decides.
3486  *
3487  * F-RTO is implemented (mainly) in four functions:
3488  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3489  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3490  *     called when tcp_use_frto() showed green light
3491  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3492  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3493  *     to prove that the RTO is indeed spurious. It transfers the control
3494  *     from F-RTO to the conventional RTO recovery
3495  */
3496 static int tcp_process_frto(struct sock *sk, int flag)
3497 {
3498         struct tcp_sock *tp = tcp_sk(sk);
3499
3500         tcp_verify_left_out(tp);
3501
3502         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3503         if (flag & FLAG_DATA_ACKED)
3504                 inet_csk(sk)->icsk_retransmits = 0;
3505
3506         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3507             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3508                 tp->undo_marker = 0;
3509
3510         if (!before(tp->snd_una, tp->frto_highmark)) {
3511                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3512                 return 1;
3513         }
3514
3515         if (!tcp_is_sackfrto(tp)) {
3516                 /* RFC4138 shortcoming in step 2; should also have case c):
3517                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3518                  * data, winupdate
3519                  */
3520                 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3521                         return 1;
3522
3523                 if (!(flag & FLAG_DATA_ACKED)) {
3524                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3525                                             flag);
3526                         return 1;
3527                 }
3528         } else {
3529                 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3530                         /* Prevent sending of new data. */
3531                         tp->snd_cwnd = min(tp->snd_cwnd,
3532                                            tcp_packets_in_flight(tp));
3533                         return 1;
3534                 }
3535
3536                 if ((tp->frto_counter >= 2) &&
3537                     (!(flag & FLAG_FORWARD_PROGRESS) ||
3538                      ((flag & FLAG_DATA_SACKED) &&
3539                       !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3540                         /* RFC4138 shortcoming (see comment above) */
3541                         if (!(flag & FLAG_FORWARD_PROGRESS) &&
3542                             (flag & FLAG_NOT_DUP))
3543                                 return 1;
3544
3545                         tcp_enter_frto_loss(sk, 3, flag);
3546                         return 1;
3547                 }
3548         }
3549
3550         if (tp->frto_counter == 1) {
3551                 /* tcp_may_send_now needs to see updated state */
3552                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3553                 tp->frto_counter = 2;
3554
3555                 if (!tcp_may_send_now(sk))
3556                         tcp_enter_frto_loss(sk, 2, flag);
3557
3558                 return 1;
3559         } else {
3560                 switch (sysctl_tcp_frto_response) {
3561                 case 2:
3562                         tcp_undo_spur_to_response(sk, flag);
3563                         break;
3564                 case 1:
3565                         tcp_conservative_spur_to_response(tp);
3566                         break;
3567                 default:
3568                         tcp_ratehalving_spur_to_response(sk);
3569                         break;
3570                 }
3571                 tp->frto_counter = 0;
3572                 tp->undo_marker = 0;
3573                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3574         }
3575         return 0;
3576 }
3577
3578 /* This routine deals with incoming acks, but not outgoing ones. */
3579 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3580 {
3581         struct inet_connection_sock *icsk = inet_csk(sk);
3582         struct tcp_sock *tp = tcp_sk(sk);
3583         u32 prior_snd_una = tp->snd_una;
3584         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3585         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3586         u32 prior_in_flight;
3587         u32 prior_fackets;
3588         int prior_packets;
3589         int frto_cwnd = 0;
3590
3591         /* If the ack is older than previous acks
3592          * then we can probably ignore it.
3593          */
3594         if (before(ack, prior_snd_una))
3595                 goto old_ack;
3596
3597         /* If the ack includes data we haven't sent yet, discard
3598          * this segment (RFC793 Section 3.9).
3599          */
3600         if (after(ack, tp->snd_nxt))
3601                 goto invalid_ack;
3602
3603         if (after(ack, prior_snd_una))
3604                 flag |= FLAG_SND_UNA_ADVANCED;
3605
3606         if (sysctl_tcp_abc) {
3607                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3608                         tp->bytes_acked += ack - prior_snd_una;
3609                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3610                         /* we assume just one segment left network */
3611                         tp->bytes_acked += min(ack - prior_snd_una,
3612                                                tp->mss_cache);
3613         }
3614
3615         prior_fackets = tp->fackets_out;
3616         prior_in_flight = tcp_packets_in_flight(tp);
3617
3618         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3619                 /* Window is constant, pure forward advance.
3620                  * No more checks are required.
3621                  * Note, we use the fact that SND.UNA>=SND.WL2.
3622                  */
3623                 tcp_update_wl(tp, ack_seq);
3624                 tp->snd_una = ack;
3625                 flag |= FLAG_WIN_UPDATE;
3626
3627                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3628
3629                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3630         } else {
3631                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3632                         flag |= FLAG_DATA;
3633                 else
3634                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3635
3636                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3637
3638                 if (TCP_SKB_CB(skb)->sacked)
3639                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3640
3641                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3642                         flag |= FLAG_ECE;
3643
3644                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3645         }
3646
3647         /* We passed data and got it acked, remove any soft error
3648          * log. Something worked...
3649          */
3650         sk->sk_err_soft = 0;
3651         icsk->icsk_probes_out = 0;
3652         tp->rcv_tstamp = tcp_time_stamp;
3653         prior_packets = tp->packets_out;
3654         if (!prior_packets)
3655                 goto no_queue;
3656
3657         /* See if we can take anything off of the retransmit queue. */
3658         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3659
3660         if (tp->frto_counter)
3661                 frto_cwnd = tcp_process_frto(sk, flag);
3662         /* Guarantee sacktag reordering detection against wrap-arounds */
3663         if (before(tp->frto_highmark, tp->snd_una))
3664                 tp->frto_highmark = 0;
3665
3666         if (tcp_ack_is_dubious(sk, flag)) {
3667                 /* Advance CWND, if state allows this. */
3668                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3669                     tcp_may_raise_cwnd(sk, flag))
3670                         tcp_cong_avoid(sk, ack, prior_in_flight);
3671                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3672                                       flag);
3673         } else {
3674                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3675                         tcp_cong_avoid(sk, ack, prior_in_flight);
3676         }
3677
3678         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3679                 dst_confirm(sk->sk_dst_cache);
3680
3681         return 1;
3682
3683 no_queue:
3684         /* If this ack opens up a zero window, clear backoff.  It was
3685          * being used to time the probes, and is probably far higher than
3686          * it needs to be for normal retransmission.
3687          */
3688         if (tcp_send_head(sk))
3689                 tcp_ack_probe(sk);
3690         return 1;
3691
3692 invalid_ack:
3693         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3694         return -1;
3695
3696 old_ack:
3697         if (TCP_SKB_CB(skb)->sacked) {
3698                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3699                 if (icsk->icsk_ca_state == TCP_CA_Open)
3700                         tcp_try_keep_open(sk);
3701         }
3702
3703         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3704         return 0;
3705 }
3706
3707 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3708  * But, this can also be called on packets in the established flow when
3709  * the fast version below fails.
3710  */
3711 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3712                        int estab)
3713 {
3714         unsigned char *ptr;
3715         struct tcphdr *th = tcp_hdr(skb);
3716         int length = (th->doff * 4) - sizeof(struct tcphdr);
3717
3718         ptr = (unsigned char *)(th + 1);
3719         opt_rx->saw_tstamp = 0;
3720
3721         while (length > 0) {
3722                 int opcode = *ptr++;
3723                 int opsize;
3724
3725                 switch (opcode) {
3726                 case TCPOPT_EOL:
3727                         return;
3728                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3729                         length--;
3730                         continue;
3731                 default:
3732                         opsize = *ptr++;
3733                         if (opsize < 2) /* "silly options" */
3734                                 return;
3735                         if (opsize > length)
3736                                 return; /* don't parse partial options */
3737                         switch (opcode) {
3738                         case TCPOPT_MSS:
3739                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3740                                         u16 in_mss = get_unaligned_be16(ptr);
3741                                         if (in_mss) {
3742                                                 if (opt_rx->user_mss &&
3743                                                     opt_rx->user_mss < in_mss)
3744                                                         in_mss = opt_rx->user_mss;
3745                                                 opt_rx->mss_clamp = in_mss;
3746                                         }
3747                                 }
3748                                 break;
3749                         case TCPOPT_WINDOW:
3750                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3751                                     !estab && sysctl_tcp_window_scaling) {
3752                                         __u8 snd_wscale = *(__u8 *)ptr;
3753                                         opt_rx->wscale_ok = 1;
3754                                         if (snd_wscale > 14) {
3755                                                 if (net_ratelimit())
3756                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3757                                                                "scaling value %d >14 received.\n",
3758                                                                snd_wscale);
3759                                                 snd_wscale = 14;
3760                                         }
3761                                         opt_rx->snd_wscale = snd_wscale;
3762                                 }
3763                                 break;
3764                         case TCPOPT_TIMESTAMP:
3765                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3766                                     ((estab && opt_rx->tstamp_ok) ||
3767                                      (!estab && sysctl_tcp_timestamps))) {
3768                                         opt_rx->saw_tstamp = 1;
3769                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3770                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3771                                 }
3772                                 break;
3773                         case TCPOPT_SACK_PERM:
3774                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3775                                     !estab && sysctl_tcp_sack) {
3776                                         opt_rx->sack_ok = 1;
3777                                         tcp_sack_reset(opt_rx);
3778                                 }
3779                                 break;
3780
3781                         case TCPOPT_SACK:
3782                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3783                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3784                                    opt_rx->sack_ok) {
3785                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3786                                 }
3787                                 break;
3788 #ifdef CONFIG_TCP_MD5SIG
3789                         case TCPOPT_MD5SIG:
3790                                 /*
3791                                  * The MD5 Hash has already been
3792                                  * checked (see tcp_v{4,6}_do_rcv()).
3793                                  */
3794                                 break;
3795 #endif
3796                         }
3797
3798                         ptr += opsize-2;
3799                         length -= opsize;
3800                 }
3801         }
3802 }
3803
3804 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3805 {
3806         __be32 *ptr = (__be32 *)(th + 1);
3807
3808         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3809                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3810                 tp->rx_opt.saw_tstamp = 1;
3811                 ++ptr;
3812                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3813                 ++ptr;
3814                 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3815                 return 1;
3816         }
3817         return 0;
3818 }
3819
3820 /* Fast parse options. This hopes to only see timestamps.
3821  * If it is wrong it falls back on tcp_parse_options().
3822  */
3823 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3824                                   struct tcp_sock *tp)
3825 {
3826         if (th->doff == sizeof(struct tcphdr) >> 2) {
3827                 tp->rx_opt.saw_tstamp = 0;
3828                 return 0;
3829         } else if (tp->rx_opt.tstamp_ok &&
3830                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3831                 if (tcp_parse_aligned_timestamp(tp, th))
3832                         return 1;
3833         }
3834         tcp_parse_options(skb, &tp->rx_opt, 1);
3835         return 1;
3836 }
3837
3838 #ifdef CONFIG_TCP_MD5SIG
3839 /*
3840  * Parse MD5 Signature option
3841  */
3842 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3843 {
3844         int length = (th->doff << 2) - sizeof (*th);
3845         u8 *ptr = (u8*)(th + 1);
3846
3847         /* If the TCP option is too short, we can short cut */
3848         if (length < TCPOLEN_MD5SIG)
3849                 return NULL;
3850
3851         while (length > 0) {
3852                 int opcode = *ptr++;
3853                 int opsize;
3854
3855                 switch(opcode) {
3856                 case TCPOPT_EOL:
3857                         return NULL;
3858                 case TCPOPT_NOP:
3859                         length--;
3860                         continue;
3861                 default:
3862                         opsize = *ptr++;
3863                         if (opsize < 2 || opsize > length)
3864                                 return NULL;
3865                         if (opcode == TCPOPT_MD5SIG)
3866                                 return ptr;
3867                 }
3868                 ptr += opsize - 2;
3869                 length -= opsize;
3870         }
3871         return NULL;
3872 }
3873 #endif
3874
3875 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3876 {
3877         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3878         tp->rx_opt.ts_recent_stamp = get_seconds();
3879 }
3880
3881 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3882 {
3883         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3884                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3885                  * extra check below makes sure this can only happen
3886                  * for pure ACK frames.  -DaveM
3887                  *
3888                  * Not only, also it occurs for expired timestamps.
3889                  */
3890
3891                 if (tcp_paws_check(&tp->rx_opt, 0))
3892                         tcp_store_ts_recent(tp);
3893         }
3894 }
3895
3896 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3897  *
3898  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3899  * it can pass through stack. So, the following predicate verifies that
3900  * this segment is not used for anything but congestion avoidance or
3901  * fast retransmit. Moreover, we even are able to eliminate most of such
3902  * second order effects, if we apply some small "replay" window (~RTO)
3903  * to timestamp space.
3904  *
3905  * All these measures still do not guarantee that we reject wrapped ACKs
3906  * on networks with high bandwidth, when sequence space is recycled fastly,
3907  * but it guarantees that such events will be very rare and do not affect
3908  * connection seriously. This doesn't look nice, but alas, PAWS is really
3909  * buggy extension.
3910  *
3911  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3912  * states that events when retransmit arrives after original data are rare.
3913  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3914  * the biggest problem on large power networks even with minor reordering.
3915  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3916  * up to bandwidth of 18Gigabit/sec. 8) ]
3917  */
3918
3919 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3920 {
3921         struct tcp_sock *tp = tcp_sk(sk);
3922         struct tcphdr *th = tcp_hdr(skb);
3923         u32 seq = TCP_SKB_CB(skb)->seq;
3924         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3925
3926         return (/* 1. Pure ACK with correct sequence number. */
3927                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3928
3929                 /* 2. ... and duplicate ACK. */
3930                 ack == tp->snd_una &&
3931
3932                 /* 3. ... and does not update window. */
3933                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3934
3935                 /* 4. ... and sits in replay window. */
3936                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3937 }
3938
3939 static inline int tcp_paws_discard(const struct sock *sk,
3940                                    const struct sk_buff *skb)
3941 {
3942         const struct tcp_sock *tp = tcp_sk(sk);
3943
3944         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3945                !tcp_disordered_ack(sk, skb);
3946 }
3947
3948 /* Check segment sequence number for validity.
3949  *
3950  * Segment controls are considered valid, if the segment
3951  * fits to the window after truncation to the window. Acceptability
3952  * of data (and SYN, FIN, of course) is checked separately.
3953  * See tcp_data_queue(), for example.
3954  *
3955  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3956  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3957  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3958  * (borrowed from freebsd)
3959  */
3960
3961 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3962 {
3963         return  !before(end_seq, tp->rcv_wup) &&
3964                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3965 }
3966
3967 /* When we get a reset we do this. */
3968 static void tcp_reset(struct sock *sk)
3969 {
3970         /* We want the right error as BSD sees it (and indeed as we do). */
3971         switch (sk->sk_state) {
3972         case TCP_SYN_SENT:
3973                 sk->sk_err = ECONNREFUSED;
3974                 break;
3975         case TCP_CLOSE_WAIT:
3976                 sk->sk_err = EPIPE;
3977                 break;
3978         case TCP_CLOSE:
3979                 return;
3980         default:
3981                 sk->sk_err = ECONNRESET;
3982         }
3983
3984         if (!sock_flag(sk, SOCK_DEAD))
3985                 sk->sk_error_report(sk);
3986
3987         tcp_done(sk);
3988 }
3989
3990 /*
3991  *      Process the FIN bit. This now behaves as it is supposed to work
3992  *      and the FIN takes effect when it is validly part of sequence
3993  *      space. Not before when we get holes.
3994  *
3995  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3996  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3997  *      TIME-WAIT)
3998  *
3999  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4000  *      close and we go into CLOSING (and later onto TIME-WAIT)
4001  *
4002  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4003  */
4004 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4005 {
4006         struct tcp_sock *tp = tcp_sk(sk);
4007
4008         inet_csk_schedule_ack(sk);
4009
4010         sk->sk_shutdown |= RCV_SHUTDOWN;
4011         sock_set_flag(sk, SOCK_DONE);
4012
4013         switch (sk->sk_state) {
4014         case TCP_SYN_RECV:
4015         case TCP_ESTABLISHED:
4016                 /* Move to CLOSE_WAIT */
4017                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4018                 inet_csk(sk)->icsk_ack.pingpong = 1;
4019                 break;
4020
4021         case TCP_CLOSE_WAIT:
4022         case TCP_CLOSING:
4023                 /* Received a retransmission of the FIN, do
4024                  * nothing.
4025                  */
4026                 break;
4027         case TCP_LAST_ACK:
4028                 /* RFC793: Remain in the LAST-ACK state. */
4029                 break;
4030
4031         case TCP_FIN_WAIT1:
4032                 /* This case occurs when a simultaneous close
4033                  * happens, we must ack the received FIN and
4034                  * enter the CLOSING state.
4035                  */
4036                 tcp_send_ack(sk);
4037                 tcp_set_state(sk, TCP_CLOSING);
4038                 break;
4039         case TCP_FIN_WAIT2:
4040                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4041                 tcp_send_ack(sk);
4042                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4043                 break;
4044         default:
4045                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4046                  * cases we should never reach this piece of code.
4047                  */
4048                 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4049                        __func__, sk->sk_state);
4050                 break;
4051         }
4052
4053         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4054          * Probably, we should reset in this case. For now drop them.
4055          */
4056         __skb_queue_purge(&tp->out_of_order_queue);
4057         if (tcp_is_sack(tp))
4058                 tcp_sack_reset(&tp->rx_opt);
4059         sk_mem_reclaim(sk);
4060
4061         if (!sock_flag(sk, SOCK_DEAD)) {
4062                 sk->sk_state_change(sk);
4063
4064                 /* Do not send POLL_HUP for half duplex close. */
4065                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4066                     sk->sk_state == TCP_CLOSE)
4067                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4068                 else
4069                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4070         }
4071 }
4072
4073 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4074                                   u32 end_seq)
4075 {
4076         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4077                 if (before(seq, sp->start_seq))
4078                         sp->start_seq = seq;
4079                 if (after(end_seq, sp->end_seq))
4080                         sp->end_seq = end_seq;
4081                 return 1;
4082         }
4083         return 0;
4084 }
4085
4086 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4087 {
4088         struct tcp_sock *tp = tcp_sk(sk);
4089
4090         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4091                 int mib_idx;
4092
4093                 if (before(seq, tp->rcv_nxt))
4094                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4095                 else
4096                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4097
4098                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4099
4100                 tp->rx_opt.dsack = 1;
4101                 tp->duplicate_sack[0].start_seq = seq;
4102                 tp->duplicate_sack[0].end_seq = end_seq;
4103         }
4104 }
4105
4106 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4107 {
4108         struct tcp_sock *tp = tcp_sk(sk);
4109
4110         if (!tp->rx_opt.dsack)
4111                 tcp_dsack_set(sk, seq, end_seq);
4112         else
4113                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4114 }
4115
4116 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4117 {
4118         struct tcp_sock *tp = tcp_sk(sk);
4119
4120         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4121             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4122                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4123                 tcp_enter_quickack_mode(sk);
4124
4125                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4126                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4127
4128                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4129                                 end_seq = tp->rcv_nxt;
4130                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4131                 }
4132         }
4133
4134         tcp_send_ack(sk);
4135 }
4136
4137 /* These routines update the SACK block as out-of-order packets arrive or
4138  * in-order packets close up the sequence space.
4139  */
4140 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4141 {
4142         int this_sack;
4143         struct tcp_sack_block *sp = &tp->selective_acks[0];
4144         struct tcp_sack_block *swalk = sp + 1;
4145
4146         /* See if the recent change to the first SACK eats into
4147          * or hits the sequence space of other SACK blocks, if so coalesce.
4148          */
4149         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4150                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4151                         int i;
4152
4153                         /* Zap SWALK, by moving every further SACK up by one slot.
4154                          * Decrease num_sacks.
4155                          */
4156                         tp->rx_opt.num_sacks--;
4157                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4158                                 sp[i] = sp[i + 1];
4159                         continue;
4160                 }
4161                 this_sack++, swalk++;
4162         }
4163 }
4164
4165 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4166 {
4167         struct tcp_sock *tp = tcp_sk(sk);
4168         struct tcp_sack_block *sp = &tp->selective_acks[0];
4169         int cur_sacks = tp->rx_opt.num_sacks;
4170         int this_sack;
4171
4172         if (!cur_sacks)
4173                 goto new_sack;
4174
4175         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4176                 if (tcp_sack_extend(sp, seq, end_seq)) {
4177                         /* Rotate this_sack to the first one. */
4178                         for (; this_sack > 0; this_sack--, sp--)
4179                                 swap(*sp, *(sp - 1));
4180                         if (cur_sacks > 1)
4181                                 tcp_sack_maybe_coalesce(tp);
4182                         return;
4183                 }
4184         }
4185
4186         /* Could not find an adjacent existing SACK, build a new one,
4187          * put it at the front, and shift everyone else down.  We
4188          * always know there is at least one SACK present already here.
4189          *
4190          * If the sack array is full, forget about the last one.
4191          */
4192         if (this_sack >= TCP_NUM_SACKS) {
4193                 this_sack--;
4194                 tp->rx_opt.num_sacks--;
4195                 sp--;
4196         }
4197         for (; this_sack > 0; this_sack--, sp--)
4198                 *sp = *(sp - 1);
4199
4200 new_sack:
4201         /* Build the new head SACK, and we're done. */
4202         sp->start_seq = seq;
4203         sp->end_seq = end_seq;
4204         tp->rx_opt.num_sacks++;
4205 }
4206
4207 /* RCV.NXT advances, some SACKs should be eaten. */
4208
4209 static void tcp_sack_remove(struct tcp_sock *tp)
4210 {
4211         struct tcp_sack_block *sp = &tp->selective_acks[0];
4212         int num_sacks = tp->rx_opt.num_sacks;
4213         int this_sack;
4214
4215         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4216         if (skb_queue_empty(&tp->out_of_order_queue)) {
4217                 tp->rx_opt.num_sacks = 0;
4218                 return;
4219         }
4220
4221         for (this_sack = 0; this_sack < num_sacks;) {
4222                 /* Check if the start of the sack is covered by RCV.NXT. */
4223                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4224                         int i;
4225
4226                         /* RCV.NXT must cover all the block! */
4227                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4228
4229                         /* Zap this SACK, by moving forward any other SACKS. */
4230                         for (i=this_sack+1; i < num_sacks; i++)
4231                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4232                         num_sacks--;
4233                         continue;
4234                 }
4235                 this_sack++;
4236                 sp++;
4237         }
4238         tp->rx_opt.num_sacks = num_sacks;
4239 }
4240
4241 /* This one checks to see if we can put data from the
4242  * out_of_order queue into the receive_queue.
4243  */
4244 static void tcp_ofo_queue(struct sock *sk)
4245 {
4246         struct tcp_sock *tp = tcp_sk(sk);
4247         __u32 dsack_high = tp->rcv_nxt;
4248         struct sk_buff *skb;
4249
4250         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4251                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4252                         break;
4253
4254                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4255                         __u32 dsack = dsack_high;
4256                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4257                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4258                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4259                 }
4260
4261                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4262                         SOCK_DEBUG(sk, "ofo packet was already received \n");
4263                         __skb_unlink(skb, &tp->out_of_order_queue);
4264                         __kfree_skb(skb);
4265                         continue;
4266                 }
4267                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4268                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4269                            TCP_SKB_CB(skb)->end_seq);
4270
4271                 __skb_unlink(skb, &tp->out_of_order_queue);
4272                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4273                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4274                 if (tcp_hdr(skb)->fin)
4275                         tcp_fin(skb, sk, tcp_hdr(skb));
4276         }
4277 }
4278
4279 static int tcp_prune_ofo_queue(struct sock *sk);
4280 static int tcp_prune_queue(struct sock *sk);
4281
4282 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4283 {
4284         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4285             !sk_rmem_schedule(sk, size)) {
4286
4287                 if (tcp_prune_queue(sk) < 0)
4288                         return -1;
4289
4290                 if (!sk_rmem_schedule(sk, size)) {
4291                         if (!tcp_prune_ofo_queue(sk))
4292                                 return -1;
4293
4294                         if (!sk_rmem_schedule(sk, size))
4295                                 return -1;
4296                 }
4297         }
4298         return 0;
4299 }
4300
4301 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4302 {
4303         struct tcphdr *th = tcp_hdr(skb);
4304         struct tcp_sock *tp = tcp_sk(sk);
4305         int eaten = -1;
4306
4307         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4308                 goto drop;
4309
4310         __skb_pull(skb, th->doff * 4);
4311
4312         TCP_ECN_accept_cwr(tp, skb);
4313
4314         tp->rx_opt.dsack = 0;
4315
4316         /*  Queue data for delivery to the user.
4317          *  Packets in sequence go to the receive queue.
4318          *  Out of sequence packets to the out_of_order_queue.
4319          */
4320         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4321                 if (tcp_receive_window(tp) == 0)
4322                         goto out_of_window;
4323
4324                 /* Ok. In sequence. In window. */
4325                 if (tp->ucopy.task == current &&
4326                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4327                     sock_owned_by_user(sk) && !tp->urg_data) {
4328                         int chunk = min_t(unsigned int, skb->len,
4329                                           tp->ucopy.len);
4330
4331                         __set_current_state(TASK_RUNNING);
4332
4333                         local_bh_enable();
4334                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4335                                 tp->ucopy.len -= chunk;
4336                                 tp->copied_seq += chunk;
4337                                 eaten = (chunk == skb->len && !th->fin);
4338                                 tcp_rcv_space_adjust(sk);
4339                         }
4340                         local_bh_disable();
4341                 }
4342
4343                 if (eaten <= 0) {
4344 queue_and_out:
4345                         if (eaten < 0 &&
4346                             tcp_try_rmem_schedule(sk, skb->truesize))
4347                                 goto drop;
4348
4349                         skb_set_owner_r(skb, sk);
4350                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4351                 }
4352                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4353                 if (skb->len)
4354                         tcp_event_data_recv(sk, skb);
4355                 if (th->fin)
4356                         tcp_fin(skb, sk, th);
4357
4358                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4359                         tcp_ofo_queue(sk);
4360
4361                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4362                          * gap in queue is filled.
4363                          */
4364                         if (skb_queue_empty(&tp->out_of_order_queue))
4365                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4366                 }
4367
4368                 if (tp->rx_opt.num_sacks)
4369                         tcp_sack_remove(tp);
4370
4371                 tcp_fast_path_check(sk);
4372
4373                 if (eaten > 0)
4374                         __kfree_skb(skb);
4375                 else if (!sock_flag(sk, SOCK_DEAD))
4376                         sk->sk_data_ready(sk, 0);
4377                 return;
4378         }
4379
4380         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4381                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4382                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4383                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4384
4385 out_of_window:
4386                 tcp_enter_quickack_mode(sk);
4387                 inet_csk_schedule_ack(sk);
4388 drop:
4389                 __kfree_skb(skb);
4390                 return;
4391         }
4392
4393         /* Out of window. F.e. zero window probe. */
4394         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4395                 goto out_of_window;
4396
4397         tcp_enter_quickack_mode(sk);
4398
4399         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4400                 /* Partial packet, seq < rcv_next < end_seq */
4401                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4402                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4403                            TCP_SKB_CB(skb)->end_seq);
4404
4405                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4406
4407                 /* If window is closed, drop tail of packet. But after
4408                  * remembering D-SACK for its head made in previous line.
4409                  */
4410                 if (!tcp_receive_window(tp))
4411                         goto out_of_window;
4412                 goto queue_and_out;
4413         }
4414
4415         TCP_ECN_check_ce(tp, skb);
4416
4417         if (tcp_try_rmem_schedule(sk, skb->truesize))
4418                 goto drop;
4419
4420         /* Disable header prediction. */
4421         tp->pred_flags = 0;
4422         inet_csk_schedule_ack(sk);
4423
4424         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4425                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4426
4427         skb_set_owner_r(skb, sk);
4428
4429         if (!skb_peek(&tp->out_of_order_queue)) {
4430                 /* Initial out of order segment, build 1 SACK. */
4431                 if (tcp_is_sack(tp)) {
4432                         tp->rx_opt.num_sacks = 1;
4433                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4434                         tp->selective_acks[0].end_seq =
4435                                                 TCP_SKB_CB(skb)->end_seq;
4436                 }
4437                 __skb_queue_head(&tp->out_of_order_queue, skb);
4438         } else {
4439                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4440                 u32 seq = TCP_SKB_CB(skb)->seq;
4441                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4442
4443                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4444                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4445
4446                         if (!tp->rx_opt.num_sacks ||
4447                             tp->selective_acks[0].end_seq != seq)
4448                                 goto add_sack;
4449
4450                         /* Common case: data arrive in order after hole. */
4451                         tp->selective_acks[0].end_seq = end_seq;
4452                         return;
4453                 }
4454
4455                 /* Find place to insert this segment. */
4456                 do {
4457                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
4458                                 break;
4459                 } while ((skb1 = skb1->prev) !=
4460                          (struct sk_buff *)&tp->out_of_order_queue);
4461
4462                 /* Do skb overlap to previous one? */
4463                 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4464                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4465                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4466                                 /* All the bits are present. Drop. */
4467                                 __kfree_skb(skb);
4468                                 tcp_dsack_set(sk, seq, end_seq);
4469                                 goto add_sack;
4470                         }
4471                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4472                                 /* Partial overlap. */
4473                                 tcp_dsack_set(sk, seq,
4474                                               TCP_SKB_CB(skb1)->end_seq);
4475                         } else {
4476                                 skb1 = skb1->prev;
4477                         }
4478                 }
4479                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4480
4481                 /* And clean segments covered by new one as whole. */
4482                 while ((skb1 = skb->next) !=
4483                        (struct sk_buff *)&tp->out_of_order_queue &&
4484                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4485                         if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4486                                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4487                                                  end_seq);
4488                                 break;
4489                         }
4490                         __skb_unlink(skb1, &tp->out_of_order_queue);
4491                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4492                                          TCP_SKB_CB(skb1)->end_seq);
4493                         __kfree_skb(skb1);
4494                 }
4495
4496 add_sack:
4497                 if (tcp_is_sack(tp))
4498                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4499         }
4500 }
4501
4502 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4503                                         struct sk_buff_head *list)
4504 {
4505         struct sk_buff *next = skb->next;
4506
4507         __skb_unlink(skb, list);
4508         __kfree_skb(skb);
4509         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4510
4511         return next;
4512 }
4513
4514 /* Collapse contiguous sequence of skbs head..tail with
4515  * sequence numbers start..end.
4516  * Segments with FIN/SYN are not collapsed (only because this
4517  * simplifies code)
4518  */
4519 static void
4520 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4521              struct sk_buff *head, struct sk_buff *tail,
4522              u32 start, u32 end)
4523 {
4524         struct sk_buff *skb;
4525
4526         /* First, check that queue is collapsible and find
4527          * the point where collapsing can be useful. */
4528         for (skb = head; skb != tail;) {
4529                 /* No new bits? It is possible on ofo queue. */
4530                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4531                         skb = tcp_collapse_one(sk, skb, list);
4532                         continue;
4533                 }
4534
4535                 /* The first skb to collapse is:
4536                  * - not SYN/FIN and
4537                  * - bloated or contains data before "start" or
4538                  *   overlaps to the next one.
4539                  */
4540                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4541                     (tcp_win_from_space(skb->truesize) > skb->len ||
4542                      before(TCP_SKB_CB(skb)->seq, start) ||
4543                      (skb->next != tail &&
4544                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4545                         break;
4546
4547                 /* Decided to skip this, advance start seq. */
4548                 start = TCP_SKB_CB(skb)->end_seq;
4549                 skb = skb->next;
4550         }
4551         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4552                 return;
4553
4554         while (before(start, end)) {
4555                 struct sk_buff *nskb;
4556                 unsigned int header = skb_headroom(skb);
4557                 int copy = SKB_MAX_ORDER(header, 0);
4558
4559                 /* Too big header? This can happen with IPv6. */
4560                 if (copy < 0)
4561                         return;
4562                 if (end - start < copy)
4563                         copy = end - start;
4564                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4565                 if (!nskb)
4566                         return;
4567
4568                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4569                 skb_set_network_header(nskb, (skb_network_header(skb) -
4570                                               skb->head));
4571                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4572                                                 skb->head));
4573                 skb_reserve(nskb, header);
4574                 memcpy(nskb->head, skb->head, header);
4575                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4576                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4577                 __skb_queue_before(list, skb, nskb);
4578                 skb_set_owner_r(nskb, sk);
4579
4580                 /* Copy data, releasing collapsed skbs. */
4581                 while (copy > 0) {
4582                         int offset = start - TCP_SKB_CB(skb)->seq;
4583                         int size = TCP_SKB_CB(skb)->end_seq - start;
4584
4585                         BUG_ON(offset < 0);
4586                         if (size > 0) {
4587                                 size = min(copy, size);
4588                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4589                                         BUG();
4590                                 TCP_SKB_CB(nskb)->end_seq += size;
4591                                 copy -= size;
4592                                 start += size;
4593                         }
4594                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4595                                 skb = tcp_collapse_one(sk, skb, list);
4596                                 if (skb == tail ||
4597                                     tcp_hdr(skb)->syn ||
4598                                     tcp_hdr(skb)->fin)
4599                                         return;
4600                         }
4601                 }
4602         }
4603 }
4604
4605 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4606  * and tcp_collapse() them until all the queue is collapsed.
4607  */
4608 static void tcp_collapse_ofo_queue(struct sock *sk)
4609 {
4610         struct tcp_sock *tp = tcp_sk(sk);
4611         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4612         struct sk_buff *head;
4613         u32 start, end;
4614
4615         if (skb == NULL)
4616                 return;
4617
4618         start = TCP_SKB_CB(skb)->seq;
4619         end = TCP_SKB_CB(skb)->end_seq;
4620         head = skb;
4621
4622         for (;;) {
4623                 skb = skb->next;
4624
4625                 /* Segment is terminated when we see gap or when
4626                  * we are at the end of all the queue. */
4627                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4628                     after(TCP_SKB_CB(skb)->seq, end) ||
4629                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4630                         tcp_collapse(sk, &tp->out_of_order_queue,
4631                                      head, skb, start, end);
4632                         head = skb;
4633                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4634                                 break;
4635                         /* Start new segment */
4636                         start = TCP_SKB_CB(skb)->seq;
4637                         end = TCP_SKB_CB(skb)->end_seq;
4638                 } else {
4639                         if (before(TCP_SKB_CB(skb)->seq, start))
4640                                 start = TCP_SKB_CB(skb)->seq;
4641                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4642                                 end = TCP_SKB_CB(skb)->end_seq;
4643                 }
4644         }
4645 }
4646
4647 /*
4648  * Purge the out-of-order queue.
4649  * Return true if queue was pruned.
4650  */
4651 static int tcp_prune_ofo_queue(struct sock *sk)
4652 {
4653         struct tcp_sock *tp = tcp_sk(sk);
4654         int res = 0;
4655
4656         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4657                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4658                 __skb_queue_purge(&tp->out_of_order_queue);
4659
4660                 /* Reset SACK state.  A conforming SACK implementation will
4661                  * do the same at a timeout based retransmit.  When a connection
4662                  * is in a sad state like this, we care only about integrity
4663                  * of the connection not performance.
4664                  */
4665                 if (tp->rx_opt.sack_ok)
4666                         tcp_sack_reset(&tp->rx_opt);
4667                 sk_mem_reclaim(sk);
4668                 res = 1;
4669         }
4670         return res;
4671 }
4672
4673 /* Reduce allocated memory if we can, trying to get
4674  * the socket within its memory limits again.
4675  *
4676  * Return less than zero if we should start dropping frames
4677  * until the socket owning process reads some of the data
4678  * to stabilize the situation.
4679  */
4680 static int tcp_prune_queue(struct sock *sk)
4681 {
4682         struct tcp_sock *tp = tcp_sk(sk);
4683
4684         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4685
4686         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4687
4688         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4689                 tcp_clamp_window(sk);
4690         else if (tcp_memory_pressure)
4691                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4692
4693         tcp_collapse_ofo_queue(sk);
4694         tcp_collapse(sk, &sk->sk_receive_queue,
4695                      sk->sk_receive_queue.next,
4696                      (struct sk_buff *)&sk->sk_receive_queue,
4697                      tp->copied_seq, tp->rcv_nxt);
4698         sk_mem_reclaim(sk);
4699
4700         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4701                 return 0;
4702
4703         /* Collapsing did not help, destructive actions follow.
4704          * This must not ever occur. */
4705
4706         tcp_prune_ofo_queue(sk);
4707
4708         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4709                 return 0;
4710
4711         /* If we are really being abused, tell the caller to silently
4712          * drop receive data on the floor.  It will get retransmitted
4713          * and hopefully then we'll have sufficient space.
4714          */
4715         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4716
4717         /* Massive buffer overcommit. */
4718         tp->pred_flags = 0;
4719         return -1;
4720 }
4721
4722 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4723  * As additional protections, we do not touch cwnd in retransmission phases,
4724  * and if application hit its sndbuf limit recently.
4725  */
4726 void tcp_cwnd_application_limited(struct sock *sk)
4727 {
4728         struct tcp_sock *tp = tcp_sk(sk);
4729
4730         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4731             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4732                 /* Limited by application or receiver window. */
4733                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4734                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4735                 if (win_used < tp->snd_cwnd) {
4736                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4737                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4738                 }
4739                 tp->snd_cwnd_used = 0;
4740         }
4741         tp->snd_cwnd_stamp = tcp_time_stamp;
4742 }
4743
4744 static int tcp_should_expand_sndbuf(struct sock *sk)
4745 {
4746         struct tcp_sock *tp = tcp_sk(sk);
4747
4748         /* If the user specified a specific send buffer setting, do
4749          * not modify it.
4750          */
4751         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4752                 return 0;
4753
4754         /* If we are under global TCP memory pressure, do not expand.  */
4755         if (tcp_memory_pressure)
4756                 return 0;
4757
4758         /* If we are under soft global TCP memory pressure, do not expand.  */
4759         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4760                 return 0;
4761
4762         /* If we filled the congestion window, do not expand.  */
4763         if (tp->packets_out >= tp->snd_cwnd)
4764                 return 0;
4765
4766         return 1;
4767 }
4768
4769 /* When incoming ACK allowed to free some skb from write_queue,
4770  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4771  * on the exit from tcp input handler.
4772  *
4773  * PROBLEM: sndbuf expansion does not work well with largesend.
4774  */
4775 static void tcp_new_space(struct sock *sk)
4776 {
4777         struct tcp_sock *tp = tcp_sk(sk);
4778
4779         if (tcp_should_expand_sndbuf(sk)) {
4780                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4781                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4782                 int demanded = max_t(unsigned int, tp->snd_cwnd,
4783                                      tp->reordering + 1);
4784                 sndmem *= 2 * demanded;
4785                 if (sndmem > sk->sk_sndbuf)
4786                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4787                 tp->snd_cwnd_stamp = tcp_time_stamp;
4788         }
4789
4790         sk->sk_write_space(sk);
4791 }
4792
4793 static void tcp_check_space(struct sock *sk)
4794 {
4795         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4796                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4797                 if (sk->sk_socket &&
4798                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4799                         tcp_new_space(sk);
4800         }
4801 }
4802
4803 static inline void tcp_data_snd_check(struct sock *sk)
4804 {
4805         tcp_push_pending_frames(sk);
4806         tcp_check_space(sk);
4807 }
4808
4809 /*
4810  * Check if sending an ack is needed.
4811  */
4812 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4813 {
4814         struct tcp_sock *tp = tcp_sk(sk);
4815
4816             /* More than one full frame received... */
4817         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4818              /* ... and right edge of window advances far enough.
4819               * (tcp_recvmsg() will send ACK otherwise). Or...
4820               */
4821              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4822             /* We ACK each frame or... */
4823             tcp_in_quickack_mode(sk) ||
4824             /* We have out of order data. */
4825             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4826                 /* Then ack it now */
4827                 tcp_send_ack(sk);
4828         } else {
4829                 /* Else, send delayed ack. */
4830                 tcp_send_delayed_ack(sk);
4831         }
4832 }
4833
4834 static inline void tcp_ack_snd_check(struct sock *sk)
4835 {
4836         if (!inet_csk_ack_scheduled(sk)) {
4837                 /* We sent a data segment already. */
4838                 return;
4839         }
4840         __tcp_ack_snd_check(sk, 1);
4841 }
4842
4843 /*
4844  *      This routine is only called when we have urgent data
4845  *      signaled. Its the 'slow' part of tcp_urg. It could be
4846  *      moved inline now as tcp_urg is only called from one
4847  *      place. We handle URGent data wrong. We have to - as
4848  *      BSD still doesn't use the correction from RFC961.
4849  *      For 1003.1g we should support a new option TCP_STDURG to permit
4850  *      either form (or just set the sysctl tcp_stdurg).
4851  */
4852
4853 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4854 {
4855         struct tcp_sock *tp = tcp_sk(sk);
4856         u32 ptr = ntohs(th->urg_ptr);
4857
4858         if (ptr && !sysctl_tcp_stdurg)
4859                 ptr--;
4860         ptr += ntohl(th->seq);
4861
4862         /* Ignore urgent data that we've already seen and read. */
4863         if (after(tp->copied_seq, ptr))
4864                 return;
4865
4866         /* Do not replay urg ptr.
4867          *
4868          * NOTE: interesting situation not covered by specs.
4869          * Misbehaving sender may send urg ptr, pointing to segment,
4870          * which we already have in ofo queue. We are not able to fetch
4871          * such data and will stay in TCP_URG_NOTYET until will be eaten
4872          * by recvmsg(). Seems, we are not obliged to handle such wicked
4873          * situations. But it is worth to think about possibility of some
4874          * DoSes using some hypothetical application level deadlock.
4875          */
4876         if (before(ptr, tp->rcv_nxt))
4877                 return;
4878
4879         /* Do we already have a newer (or duplicate) urgent pointer? */
4880         if (tp->urg_data && !after(ptr, tp->urg_seq))
4881                 return;
4882
4883         /* Tell the world about our new urgent pointer. */
4884         sk_send_sigurg(sk);
4885
4886         /* We may be adding urgent data when the last byte read was
4887          * urgent. To do this requires some care. We cannot just ignore
4888          * tp->copied_seq since we would read the last urgent byte again
4889          * as data, nor can we alter copied_seq until this data arrives
4890          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4891          *
4892          * NOTE. Double Dutch. Rendering to plain English: author of comment
4893          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4894          * and expect that both A and B disappear from stream. This is _wrong_.
4895          * Though this happens in BSD with high probability, this is occasional.
4896          * Any application relying on this is buggy. Note also, that fix "works"
4897          * only in this artificial test. Insert some normal data between A and B and we will
4898          * decline of BSD again. Verdict: it is better to remove to trap
4899          * buggy users.
4900          */
4901         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4902             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4903                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4904                 tp->copied_seq++;
4905                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4906                         __skb_unlink(skb, &sk->sk_receive_queue);
4907                         __kfree_skb(skb);
4908                 }
4909         }
4910
4911         tp->urg_data = TCP_URG_NOTYET;
4912         tp->urg_seq = ptr;
4913
4914         /* Disable header prediction. */
4915         tp->pred_flags = 0;
4916 }
4917
4918 /* This is the 'fast' part of urgent handling. */
4919 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4920 {
4921         struct tcp_sock *tp = tcp_sk(sk);
4922
4923         /* Check if we get a new urgent pointer - normally not. */
4924         if (th->urg)
4925                 tcp_check_urg(sk, th);
4926
4927         /* Do we wait for any urgent data? - normally not... */
4928         if (tp->urg_data == TCP_URG_NOTYET) {
4929                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4930                           th->syn;
4931
4932                 /* Is the urgent pointer pointing into this packet? */
4933                 if (ptr < skb->len) {
4934                         u8 tmp;
4935                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4936                                 BUG();
4937                         tp->urg_data = TCP_URG_VALID | tmp;
4938                         if (!sock_flag(sk, SOCK_DEAD))
4939                                 sk->sk_data_ready(sk, 0);
4940                 }
4941         }
4942 }
4943
4944 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4945 {
4946         struct tcp_sock *tp = tcp_sk(sk);
4947         int chunk = skb->len - hlen;
4948         int err;
4949
4950         local_bh_enable();
4951         if (skb_csum_unnecessary(skb))
4952                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4953         else
4954                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4955                                                        tp->ucopy.iov);
4956
4957         if (!err) {
4958                 tp->ucopy.len -= chunk;
4959                 tp->copied_seq += chunk;
4960                 tcp_rcv_space_adjust(sk);
4961         }
4962
4963         local_bh_disable();
4964         return err;
4965 }
4966
4967 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4968                                             struct sk_buff *skb)
4969 {
4970         __sum16 result;
4971
4972         if (sock_owned_by_user(sk)) {
4973                 local_bh_enable();
4974                 result = __tcp_checksum_complete(skb);
4975                 local_bh_disable();
4976         } else {
4977                 result = __tcp_checksum_complete(skb);
4978         }
4979         return result;
4980 }
4981
4982 static inline int tcp_checksum_complete_user(struct sock *sk,
4983                                              struct sk_buff *skb)
4984 {
4985         return !skb_csum_unnecessary(skb) &&
4986                __tcp_checksum_complete_user(sk, skb);
4987 }
4988
4989 #ifdef CONFIG_NET_DMA
4990 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4991                                   int hlen)
4992 {
4993         struct tcp_sock *tp = tcp_sk(sk);
4994         int chunk = skb->len - hlen;
4995         int dma_cookie;
4996         int copied_early = 0;
4997
4998         if (tp->ucopy.wakeup)
4999                 return 0;
5000
5001         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5002                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5003
5004         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5005
5006                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5007                                                          skb, hlen,
5008                                                          tp->ucopy.iov, chunk,
5009                                                          tp->ucopy.pinned_list);
5010
5011                 if (dma_cookie < 0)
5012                         goto out;
5013
5014                 tp->ucopy.dma_cookie = dma_cookie;
5015                 copied_early = 1;
5016
5017                 tp->ucopy.len -= chunk;
5018                 tp->copied_seq += chunk;
5019                 tcp_rcv_space_adjust(sk);
5020
5021                 if ((tp->ucopy.len == 0) ||
5022                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5023                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5024                         tp->ucopy.wakeup = 1;
5025                         sk->sk_data_ready(sk, 0);
5026                 }
5027         } else if (chunk > 0) {
5028                 tp->ucopy.wakeup = 1;
5029                 sk->sk_data_ready(sk, 0);
5030         }
5031 out:
5032         return copied_early;
5033 }
5034 #endif /* CONFIG_NET_DMA */
5035
5036 /* Does PAWS and seqno based validation of an incoming segment, flags will
5037  * play significant role here.
5038  */
5039 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5040                               struct tcphdr *th, int syn_inerr)
5041 {
5042         struct tcp_sock *tp = tcp_sk(sk);
5043
5044         /* RFC1323: H1. Apply PAWS check first. */
5045         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5046             tcp_paws_discard(sk, skb)) {
5047                 if (!th->rst) {
5048                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5049                         tcp_send_dupack(sk, skb);
5050                         goto discard;
5051                 }
5052                 /* Reset is accepted even if it did not pass PAWS. */
5053         }
5054
5055         /* Step 1: check sequence number */
5056         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5057                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5058                  * (RST) segments are validated by checking their SEQ-fields."
5059                  * And page 69: "If an incoming segment is not acceptable,
5060                  * an acknowledgment should be sent in reply (unless the RST
5061                  * bit is set, if so drop the segment and return)".
5062                  */
5063                 if (!th->rst)
5064                         tcp_send_dupack(sk, skb);
5065                 goto discard;
5066         }
5067
5068         /* Step 2: check RST bit */
5069         if (th->rst) {
5070                 tcp_reset(sk);
5071                 goto discard;
5072         }
5073
5074         /* ts_recent update must be made after we are sure that the packet
5075          * is in window.
5076          */
5077         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5078
5079         /* step 3: check security and precedence [ignored] */
5080
5081         /* step 4: Check for a SYN in window. */
5082         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5083                 if (syn_inerr)
5084                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5085                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5086                 tcp_reset(sk);
5087                 return -1;
5088         }
5089
5090         return 1;
5091
5092 discard:
5093         __kfree_skb(skb);
5094         return 0;
5095 }
5096
5097 /*
5098  *      TCP receive function for the ESTABLISHED state.
5099  *
5100  *      It is split into a fast path and a slow path. The fast path is
5101  *      disabled when:
5102  *      - A zero window was announced from us - zero window probing
5103  *        is only handled properly in the slow path.
5104  *      - Out of order segments arrived.
5105  *      - Urgent data is expected.
5106  *      - There is no buffer space left
5107  *      - Unexpected TCP flags/window values/header lengths are received
5108  *        (detected by checking the TCP header against pred_flags)
5109  *      - Data is sent in both directions. Fast path only supports pure senders
5110  *        or pure receivers (this means either the sequence number or the ack
5111  *        value must stay constant)
5112  *      - Unexpected TCP option.
5113  *
5114  *      When these conditions are not satisfied it drops into a standard
5115  *      receive procedure patterned after RFC793 to handle all cases.
5116  *      The first three cases are guaranteed by proper pred_flags setting,
5117  *      the rest is checked inline. Fast processing is turned on in
5118  *      tcp_data_queue when everything is OK.
5119  */
5120 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5121                         struct tcphdr *th, unsigned len)
5122 {
5123         struct tcp_sock *tp = tcp_sk(sk);
5124         int res;
5125
5126         /*
5127          *      Header prediction.
5128          *      The code loosely follows the one in the famous
5129          *      "30 instruction TCP receive" Van Jacobson mail.
5130          *
5131          *      Van's trick is to deposit buffers into socket queue
5132          *      on a device interrupt, to call tcp_recv function
5133          *      on the receive process context and checksum and copy
5134          *      the buffer to user space. smart...
5135          *
5136          *      Our current scheme is not silly either but we take the
5137          *      extra cost of the net_bh soft interrupt processing...
5138          *      We do checksum and copy also but from device to kernel.
5139          */
5140
5141         tp->rx_opt.saw_tstamp = 0;
5142
5143         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5144          *      if header_prediction is to be made
5145          *      'S' will always be tp->tcp_header_len >> 2
5146          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5147          *  turn it off (when there are holes in the receive
5148          *       space for instance)
5149          *      PSH flag is ignored.
5150          */
5151
5152         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5153             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5154             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5155                 int tcp_header_len = tp->tcp_header_len;
5156
5157                 /* Timestamp header prediction: tcp_header_len
5158                  * is automatically equal to th->doff*4 due to pred_flags
5159                  * match.
5160                  */
5161
5162                 /* Check timestamp */
5163                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5164                         /* No? Slow path! */
5165                         if (!tcp_parse_aligned_timestamp(tp, th))
5166                                 goto slow_path;
5167
5168                         /* If PAWS failed, check it more carefully in slow path */
5169                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5170                                 goto slow_path;
5171
5172                         /* DO NOT update ts_recent here, if checksum fails
5173                          * and timestamp was corrupted part, it will result
5174                          * in a hung connection since we will drop all
5175                          * future packets due to the PAWS test.
5176                          */
5177                 }
5178
5179                 if (len <= tcp_header_len) {
5180                         /* Bulk data transfer: sender */
5181                         if (len == tcp_header_len) {
5182                                 /* Predicted packet is in window by definition.
5183                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5184                                  * Hence, check seq<=rcv_wup reduces to:
5185                                  */
5186                                 if (tcp_header_len ==
5187                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5188                                     tp->rcv_nxt == tp->rcv_wup)
5189                                         tcp_store_ts_recent(tp);
5190
5191                                 /* We know that such packets are checksummed
5192                                  * on entry.
5193                                  */
5194                                 tcp_ack(sk, skb, 0);
5195                                 __kfree_skb(skb);
5196                                 tcp_data_snd_check(sk);
5197                                 return 0;
5198                         } else { /* Header too small */
5199                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5200                                 goto discard;
5201                         }
5202                 } else {
5203                         int eaten = 0;
5204                         int copied_early = 0;
5205
5206                         if (tp->copied_seq == tp->rcv_nxt &&
5207                             len - tcp_header_len <= tp->ucopy.len) {
5208 #ifdef CONFIG_NET_DMA
5209                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5210                                         copied_early = 1;
5211                                         eaten = 1;
5212                                 }
5213 #endif
5214                                 if (tp->ucopy.task == current &&
5215                                     sock_owned_by_user(sk) && !copied_early) {
5216                                         __set_current_state(TASK_RUNNING);
5217
5218                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5219                                                 eaten = 1;
5220                                 }
5221                                 if (eaten) {
5222                                         /* Predicted packet is in window by definition.
5223                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5224                                          * Hence, check seq<=rcv_wup reduces to:
5225                                          */
5226                                         if (tcp_header_len ==
5227                                             (sizeof(struct tcphdr) +
5228                                              TCPOLEN_TSTAMP_ALIGNED) &&
5229                                             tp->rcv_nxt == tp->rcv_wup)
5230                                                 tcp_store_ts_recent(tp);
5231
5232                                         tcp_rcv_rtt_measure_ts(sk, skb);
5233
5234                                         __skb_pull(skb, tcp_header_len);
5235                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5236                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5237                                 }
5238                                 if (copied_early)
5239                                         tcp_cleanup_rbuf(sk, skb->len);
5240                         }
5241                         if (!eaten) {
5242                                 if (tcp_checksum_complete_user(sk, skb))
5243                                         goto csum_error;
5244
5245                                 /* Predicted packet is in window by definition.
5246                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5247                                  * Hence, check seq<=rcv_wup reduces to:
5248                                  */
5249                                 if (tcp_header_len ==
5250                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5251                                     tp->rcv_nxt == tp->rcv_wup)
5252                                         tcp_store_ts_recent(tp);
5253
5254                                 tcp_rcv_rtt_measure_ts(sk, skb);
5255
5256                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5257                                         goto step5;
5258
5259                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5260
5261                                 /* Bulk data transfer: receiver */
5262                                 __skb_pull(skb, tcp_header_len);
5263                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
5264                                 skb_set_owner_r(skb, sk);
5265                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5266                         }
5267
5268                         tcp_event_data_recv(sk, skb);
5269
5270                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5271                                 /* Well, only one small jumplet in fast path... */
5272                                 tcp_ack(sk, skb, FLAG_DATA);
5273                                 tcp_data_snd_check(sk);
5274                                 if (!inet_csk_ack_scheduled(sk))
5275                                         goto no_ack;
5276                         }
5277
5278                         if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5279                                 __tcp_ack_snd_check(sk, 0);
5280 no_ack:
5281 #ifdef CONFIG_NET_DMA
5282                         if (copied_early)
5283                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5284                         else
5285 #endif
5286                         if (eaten)
5287                                 __kfree_skb(skb);
5288                         else
5289                                 sk->sk_data_ready(sk, 0);
5290                         return 0;
5291                 }
5292         }
5293
5294 slow_path:
5295         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5296                 goto csum_error;
5297
5298         /*
5299          *      Standard slow path.
5300          */
5301
5302         res = tcp_validate_incoming(sk, skb, th, 1);
5303         if (res <= 0)
5304                 return -res;
5305
5306 step5:
5307         if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5308                 goto discard;
5309
5310         tcp_rcv_rtt_measure_ts(sk, skb);
5311
5312         /* Process urgent data. */
5313         tcp_urg(sk, skb, th);
5314
5315         /* step 7: process the segment text */
5316         tcp_data_queue(sk, skb);
5317
5318         tcp_data_snd_check(sk);
5319         tcp_ack_snd_check(sk);
5320         return 0;
5321
5322 csum_error:
5323         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5324
5325 discard:
5326         __kfree_skb(skb);
5327         return 0;
5328 }
5329
5330 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5331                                          struct tcphdr *th, unsigned len)
5332 {
5333         struct tcp_sock *tp = tcp_sk(sk);
5334         struct inet_connection_sock *icsk = inet_csk(sk);
5335         int saved_clamp = tp->rx_opt.mss_clamp;
5336
5337         tcp_parse_options(skb, &tp->rx_opt, 0);
5338
5339         if (th->ack) {
5340                 /* rfc793:
5341                  * "If the state is SYN-SENT then
5342                  *    first check the ACK bit
5343                  *      If the ACK bit is set
5344                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5345                  *        a reset (unless the RST bit is set, if so drop
5346                  *        the segment and return)"
5347                  *
5348                  *  We do not send data with SYN, so that RFC-correct
5349                  *  test reduces to:
5350                  */
5351                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5352                         goto reset_and_undo;
5353
5354                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5355                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5356                              tcp_time_stamp)) {
5357                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5358                         goto reset_and_undo;
5359                 }
5360
5361                 /* Now ACK is acceptable.
5362                  *
5363                  * "If the RST bit is set
5364                  *    If the ACK was acceptable then signal the user "error:
5365                  *    connection reset", drop the segment, enter CLOSED state,
5366                  *    delete TCB, and return."
5367                  */
5368
5369                 if (th->rst) {
5370                         tcp_reset(sk);
5371                         goto discard;
5372                 }
5373
5374                 /* rfc793:
5375                  *   "fifth, if neither of the SYN or RST bits is set then
5376                  *    drop the segment and return."
5377                  *
5378                  *    See note below!
5379                  *                                        --ANK(990513)
5380                  */
5381                 if (!th->syn)
5382                         goto discard_and_undo;
5383
5384                 /* rfc793:
5385                  *   "If the SYN bit is on ...
5386                  *    are acceptable then ...
5387                  *    (our SYN has been ACKed), change the connection
5388                  *    state to ESTABLISHED..."
5389                  */
5390
5391                 TCP_ECN_rcv_synack(tp, th);
5392
5393                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5394                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5395
5396                 /* Ok.. it's good. Set up sequence numbers and
5397                  * move to established.
5398                  */
5399                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5400                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5401
5402                 /* RFC1323: The window in SYN & SYN/ACK segments is
5403                  * never scaled.
5404                  */
5405                 tp->snd_wnd = ntohs(th->window);
5406                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5407
5408                 if (!tp->rx_opt.wscale_ok) {
5409                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5410                         tp->window_clamp = min(tp->window_clamp, 65535U);
5411                 }
5412
5413                 if (tp->rx_opt.saw_tstamp) {
5414                         tp->rx_opt.tstamp_ok       = 1;
5415                         tp->tcp_header_len =
5416                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5417                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5418                         tcp_store_ts_recent(tp);
5419                 } else {
5420                         tp->tcp_header_len = sizeof(struct tcphdr);
5421                 }
5422
5423                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5424                         tcp_enable_fack(tp);
5425
5426                 tcp_mtup_init(sk);
5427                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5428                 tcp_initialize_rcv_mss(sk);
5429
5430                 /* Remember, tcp_poll() does not lock socket!
5431                  * Change state from SYN-SENT only after copied_seq
5432                  * is initialized. */
5433                 tp->copied_seq = tp->rcv_nxt;
5434                 smp_mb();
5435                 tcp_set_state(sk, TCP_ESTABLISHED);
5436
5437                 security_inet_conn_established(sk, skb);
5438
5439                 /* Make sure socket is routed, for correct metrics.  */
5440                 icsk->icsk_af_ops->rebuild_header(sk);
5441
5442                 tcp_init_metrics(sk);
5443
5444                 tcp_init_congestion_control(sk);
5445
5446                 /* Prevent spurious tcp_cwnd_restart() on first data
5447                  * packet.
5448                  */
5449                 tp->lsndtime = tcp_time_stamp;
5450
5451                 tcp_init_buffer_space(sk);
5452
5453                 if (sock_flag(sk, SOCK_KEEPOPEN))
5454                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5455
5456                 if (!tp->rx_opt.snd_wscale)
5457                         __tcp_fast_path_on(tp, tp->snd_wnd);
5458                 else
5459                         tp->pred_flags = 0;
5460
5461                 if (!sock_flag(sk, SOCK_DEAD)) {
5462                         sk->sk_state_change(sk);
5463                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5464                 }
5465
5466                 if (sk->sk_write_pending ||
5467                     icsk->icsk_accept_queue.rskq_defer_accept ||
5468                     icsk->icsk_ack.pingpong) {
5469                         /* Save one ACK. Data will be ready after
5470                          * several ticks, if write_pending is set.
5471                          *
5472                          * It may be deleted, but with this feature tcpdumps
5473                          * look so _wonderfully_ clever, that I was not able
5474                          * to stand against the temptation 8)     --ANK
5475                          */
5476                         inet_csk_schedule_ack(sk);
5477                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5478                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
5479                         tcp_incr_quickack(sk);
5480                         tcp_enter_quickack_mode(sk);
5481                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5482                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5483
5484 discard:
5485                         __kfree_skb(skb);
5486                         return 0;
5487                 } else {
5488                         tcp_send_ack(sk);
5489                 }
5490                 return -1;
5491         }
5492
5493         /* No ACK in the segment */
5494
5495         if (th->rst) {
5496                 /* rfc793:
5497                  * "If the RST bit is set
5498                  *
5499                  *      Otherwise (no ACK) drop the segment and return."
5500                  */
5501
5502                 goto discard_and_undo;
5503         }
5504
5505         /* PAWS check. */
5506         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5507             tcp_paws_reject(&tp->rx_opt, 0))
5508                 goto discard_and_undo;
5509
5510         if (th->syn) {
5511                 /* We see SYN without ACK. It is attempt of
5512                  * simultaneous connect with crossed SYNs.
5513                  * Particularly, it can be connect to self.
5514                  */
5515                 tcp_set_state(sk, TCP_SYN_RECV);
5516
5517                 if (tp->rx_opt.saw_tstamp) {
5518                         tp->rx_opt.tstamp_ok = 1;
5519                         tcp_store_ts_recent(tp);
5520                         tp->tcp_header_len =
5521                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5522                 } else {
5523                         tp->tcp_header_len = sizeof(struct tcphdr);
5524                 }
5525
5526                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5527                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5528
5529                 /* RFC1323: The window in SYN & SYN/ACK segments is
5530                  * never scaled.
5531                  */
5532                 tp->snd_wnd    = ntohs(th->window);
5533                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5534                 tp->max_window = tp->snd_wnd;
5535
5536                 TCP_ECN_rcv_syn(tp, th);
5537
5538                 tcp_mtup_init(sk);
5539                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5540                 tcp_initialize_rcv_mss(sk);
5541
5542                 tcp_send_synack(sk);
5543 #if 0
5544                 /* Note, we could accept data and URG from this segment.
5545                  * There are no obstacles to make this.
5546                  *
5547                  * However, if we ignore data in ACKless segments sometimes,
5548                  * we have no reasons to accept it sometimes.
5549                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5550                  * is not flawless. So, discard packet for sanity.
5551                  * Uncomment this return to process the data.
5552                  */
5553                 return -1;
5554 #else
5555                 goto discard;
5556 #endif
5557         }
5558         /* "fifth, if neither of the SYN or RST bits is set then
5559          * drop the segment and return."
5560          */
5561
5562 discard_and_undo:
5563         tcp_clear_options(&tp->rx_opt);
5564         tp->rx_opt.mss_clamp = saved_clamp;
5565         goto discard;
5566
5567 reset_and_undo:
5568         tcp_clear_options(&tp->rx_opt);
5569         tp->rx_opt.mss_clamp = saved_clamp;
5570         return 1;
5571 }
5572
5573 /*
5574  *      This function implements the receiving procedure of RFC 793 for
5575  *      all states except ESTABLISHED and TIME_WAIT.
5576  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5577  *      address independent.
5578  */
5579
5580 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5581                           struct tcphdr *th, unsigned len)
5582 {
5583         struct tcp_sock *tp = tcp_sk(sk);
5584         struct inet_connection_sock *icsk = inet_csk(sk);
5585         int queued = 0;
5586         int res;
5587
5588         tp->rx_opt.saw_tstamp = 0;
5589
5590         switch (sk->sk_state) {
5591         case TCP_CLOSE:
5592                 goto discard;
5593
5594         case TCP_LISTEN:
5595                 if (th->ack)
5596                         return 1;
5597
5598                 if (th->rst)
5599                         goto discard;
5600
5601                 if (th->syn) {
5602                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5603                                 return 1;
5604
5605                         /* Now we have several options: In theory there is
5606                          * nothing else in the frame. KA9Q has an option to
5607                          * send data with the syn, BSD accepts data with the
5608                          * syn up to the [to be] advertised window and
5609                          * Solaris 2.1 gives you a protocol error. For now
5610                          * we just ignore it, that fits the spec precisely
5611                          * and avoids incompatibilities. It would be nice in
5612                          * future to drop through and process the data.
5613                          *
5614                          * Now that TTCP is starting to be used we ought to
5615                          * queue this data.
5616                          * But, this leaves one open to an easy denial of
5617                          * service attack, and SYN cookies can't defend
5618                          * against this problem. So, we drop the data
5619                          * in the interest of security over speed unless
5620                          * it's still in use.
5621                          */
5622                         kfree_skb(skb);
5623                         return 0;
5624                 }
5625                 goto discard;
5626
5627         case TCP_SYN_SENT:
5628                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5629                 if (queued >= 0)
5630                         return queued;
5631
5632                 /* Do step6 onward by hand. */
5633                 tcp_urg(sk, skb, th);
5634                 __kfree_skb(skb);
5635                 tcp_data_snd_check(sk);
5636                 return 0;
5637         }
5638
5639         res = tcp_validate_incoming(sk, skb, th, 0);
5640         if (res <= 0)
5641                 return -res;
5642
5643         /* step 5: check the ACK field */
5644         if (th->ack) {
5645                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5646
5647                 switch (sk->sk_state) {
5648                 case TCP_SYN_RECV:
5649                         if (acceptable) {
5650                                 tp->copied_seq = tp->rcv_nxt;
5651                                 smp_mb();
5652                                 tcp_set_state(sk, TCP_ESTABLISHED);
5653                                 sk->sk_state_change(sk);
5654
5655                                 /* Note, that this wakeup is only for marginal
5656                                  * crossed SYN case. Passively open sockets
5657                                  * are not waked up, because sk->sk_sleep ==
5658                                  * NULL and sk->sk_socket == NULL.
5659                                  */
5660                                 if (sk->sk_socket)
5661                                         sk_wake_async(sk,
5662                                                       SOCK_WAKE_IO, POLL_OUT);
5663
5664                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5665                                 tp->snd_wnd = ntohs(th->window) <<
5666                                               tp->rx_opt.snd_wscale;
5667                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5668
5669                                 /* tcp_ack considers this ACK as duplicate
5670                                  * and does not calculate rtt.
5671                                  * Fix it at least with timestamps.
5672                                  */
5673                                 if (tp->rx_opt.saw_tstamp &&
5674                                     tp->rx_opt.rcv_tsecr && !tp->srtt)
5675                                         tcp_ack_saw_tstamp(sk, 0);
5676
5677                                 if (tp->rx_opt.tstamp_ok)
5678                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5679
5680                                 /* Make sure socket is routed, for
5681                                  * correct metrics.
5682                                  */
5683                                 icsk->icsk_af_ops->rebuild_header(sk);
5684
5685                                 tcp_init_metrics(sk);
5686
5687                                 tcp_init_congestion_control(sk);
5688
5689                                 /* Prevent spurious tcp_cwnd_restart() on
5690                                  * first data packet.
5691                                  */
5692                                 tp->lsndtime = tcp_time_stamp;
5693
5694                                 tcp_mtup_init(sk);
5695                                 tcp_initialize_rcv_mss(sk);
5696                                 tcp_init_buffer_space(sk);
5697                                 tcp_fast_path_on(tp);
5698                         } else {
5699                                 return 1;
5700                         }
5701                         break;
5702
5703                 case TCP_FIN_WAIT1:
5704                         if (tp->snd_una == tp->write_seq) {
5705                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5706                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5707                                 dst_confirm(sk->sk_dst_cache);
5708
5709                                 if (!sock_flag(sk, SOCK_DEAD))
5710                                         /* Wake up lingering close() */
5711                                         sk->sk_state_change(sk);
5712                                 else {
5713                                         int tmo;
5714
5715                                         if (tp->linger2 < 0 ||
5716                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5717                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5718                                                 tcp_done(sk);
5719                                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5720                                                 return 1;
5721                                         }
5722
5723                                         tmo = tcp_fin_time(sk);
5724                                         if (tmo > TCP_TIMEWAIT_LEN) {
5725                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5726                                         } else if (th->fin || sock_owned_by_user(sk)) {
5727                                                 /* Bad case. We could lose such FIN otherwise.
5728                                                  * It is not a big problem, but it looks confusing
5729                                                  * and not so rare event. We still can lose it now,
5730                                                  * if it spins in bh_lock_sock(), but it is really
5731                                                  * marginal case.
5732                                                  */
5733                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5734                                         } else {
5735                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5736                                                 goto discard;
5737                                         }
5738                                 }
5739                         }
5740                         break;
5741
5742                 case TCP_CLOSING:
5743                         if (tp->snd_una == tp->write_seq) {
5744                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5745                                 goto discard;
5746                         }
5747                         break;
5748
5749                 case TCP_LAST_ACK:
5750                         if (tp->snd_una == tp->write_seq) {
5751                                 tcp_update_metrics(sk);
5752                                 tcp_done(sk);
5753                                 goto discard;
5754                         }
5755                         break;
5756                 }
5757         } else
5758                 goto discard;
5759
5760         /* step 6: check the URG bit */
5761         tcp_urg(sk, skb, th);
5762
5763         /* step 7: process the segment text */
5764         switch (sk->sk_state) {
5765         case TCP_CLOSE_WAIT:
5766         case TCP_CLOSING:
5767         case TCP_LAST_ACK:
5768                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5769                         break;
5770         case TCP_FIN_WAIT1:
5771         case TCP_FIN_WAIT2:
5772                 /* RFC 793 says to queue data in these states,
5773                  * RFC 1122 says we MUST send a reset.
5774                  * BSD 4.4 also does reset.
5775                  */
5776                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5777                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5778                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5779                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5780                                 tcp_reset(sk);
5781                                 return 1;
5782                         }
5783                 }
5784                 /* Fall through */
5785         case TCP_ESTABLISHED:
5786                 tcp_data_queue(sk, skb);
5787                 queued = 1;
5788                 break;
5789         }
5790
5791         /* tcp_data could move socket to TIME-WAIT */
5792         if (sk->sk_state != TCP_CLOSE) {
5793                 tcp_data_snd_check(sk);
5794                 tcp_ack_snd_check(sk);
5795         }
5796
5797         if (!queued) {
5798 discard:
5799                 __kfree_skb(skb);
5800         }
5801         return 0;
5802 }
5803
5804 EXPORT_SYMBOL(sysctl_tcp_ecn);
5805 EXPORT_SYMBOL(sysctl_tcp_reordering);
5806 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5807 EXPORT_SYMBOL(tcp_parse_options);
5808 #ifdef CONFIG_TCP_MD5SIG
5809 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5810 #endif
5811 EXPORT_SYMBOL(tcp_rcv_established);
5812 EXPORT_SYMBOL(tcp_rcv_state_process);
5813 EXPORT_SYMBOL(tcp_initialize_rcv_mss);