[NET] CORE: Introducing new memory accounting interface.
[linux-2.6.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
108
109 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
116
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
119
120 /* Adapt the MSS value used to make delayed ack decision to the
121  * real world.
122  */
123 static void tcp_measure_rcv_mss(struct sock *sk,
124                                 const struct sk_buff *skb)
125 {
126         struct inet_connection_sock *icsk = inet_csk(sk);
127         const unsigned int lss = icsk->icsk_ack.last_seg_size;
128         unsigned int len;
129
130         icsk->icsk_ack.last_seg_size = 0;
131
132         /* skb->len may jitter because of SACKs, even if peer
133          * sends good full-sized frames.
134          */
135         len = skb_shinfo(skb)->gso_size ?: skb->len;
136         if (len >= icsk->icsk_ack.rcv_mss) {
137                 icsk->icsk_ack.rcv_mss = len;
138         } else {
139                 /* Otherwise, we make more careful check taking into account,
140                  * that SACKs block is variable.
141                  *
142                  * "len" is invariant segment length, including TCP header.
143                  */
144                 len += skb->data - skb_transport_header(skb);
145                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
146                     /* If PSH is not set, packet should be
147                      * full sized, provided peer TCP is not badly broken.
148                      * This observation (if it is correct 8)) allows
149                      * to handle super-low mtu links fairly.
150                      */
151                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
152                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
153                         /* Subtract also invariant (if peer is RFC compliant),
154                          * tcp header plus fixed timestamp option length.
155                          * Resulting "len" is MSS free of SACK jitter.
156                          */
157                         len -= tcp_sk(sk)->tcp_header_len;
158                         icsk->icsk_ack.last_seg_size = len;
159                         if (len == lss) {
160                                 icsk->icsk_ack.rcv_mss = len;
161                                 return;
162                         }
163                 }
164                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
166                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
167         }
168 }
169
170 static void tcp_incr_quickack(struct sock *sk)
171 {
172         struct inet_connection_sock *icsk = inet_csk(sk);
173         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
174
175         if (quickacks==0)
176                 quickacks=2;
177         if (quickacks > icsk->icsk_ack.quick)
178                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
179 }
180
181 void tcp_enter_quickack_mode(struct sock *sk)
182 {
183         struct inet_connection_sock *icsk = inet_csk(sk);
184         tcp_incr_quickack(sk);
185         icsk->icsk_ack.pingpong = 0;
186         icsk->icsk_ack.ato = TCP_ATO_MIN;
187 }
188
189 /* Send ACKs quickly, if "quick" count is not exhausted
190  * and the session is not interactive.
191  */
192
193 static inline int tcp_in_quickack_mode(const struct sock *sk)
194 {
195         const struct inet_connection_sock *icsk = inet_csk(sk);
196         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
197 }
198
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
200 {
201         if (tp->ecn_flags&TCP_ECN_OK)
202                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
203 }
204
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
206 {
207         if (tcp_hdr(skb)->cwr)
208                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
209 }
210
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
212 {
213         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 }
215
216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
217 {
218         if (tp->ecn_flags&TCP_ECN_OK) {
219                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221                 /* Funny extension: if ECT is not set on a segment,
222                  * it is surely retransmit. It is not in ECN RFC,
223                  * but Linux follows this rule. */
224                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225                         tcp_enter_quickack_mode((struct sock *)tp);
226         }
227 }
228
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
230 {
231         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
232                 tp->ecn_flags &= ~TCP_ECN_OK;
233 }
234
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
236 {
237         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
238                 tp->ecn_flags &= ~TCP_ECN_OK;
239 }
240
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
242 {
243         if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
244                 return 1;
245         return 0;
246 }
247
248 /* Buffer size and advertised window tuning.
249  *
250  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
251  */
252
253 static void tcp_fixup_sndbuf(struct sock *sk)
254 {
255         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256                      sizeof(struct sk_buff);
257
258         if (sk->sk_sndbuf < 3 * sndmem)
259                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
260 }
261
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263  *
264  * All tcp_full_space() is split to two parts: "network" buffer, allocated
265  * forward and advertised in receiver window (tp->rcv_wnd) and
266  * "application buffer", required to isolate scheduling/application
267  * latencies from network.
268  * window_clamp is maximal advertised window. It can be less than
269  * tcp_full_space(), in this case tcp_full_space() - window_clamp
270  * is reserved for "application" buffer. The less window_clamp is
271  * the smoother our behaviour from viewpoint of network, but the lower
272  * throughput and the higher sensitivity of the connection to losses. 8)
273  *
274  * rcv_ssthresh is more strict window_clamp used at "slow start"
275  * phase to predict further behaviour of this connection.
276  * It is used for two goals:
277  * - to enforce header prediction at sender, even when application
278  *   requires some significant "application buffer". It is check #1.
279  * - to prevent pruning of receive queue because of misprediction
280  *   of receiver window. Check #2.
281  *
282  * The scheme does not work when sender sends good segments opening
283  * window and then starts to feed us spaghetti. But it should work
284  * in common situations. Otherwise, we have to rely on queue collapsing.
285  */
286
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
289 {
290         struct tcp_sock *tp = tcp_sk(sk);
291         /* Optimize this! */
292         int truesize = tcp_win_from_space(skb->truesize) >> 1;
293         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
294
295         while (tp->rcv_ssthresh <= window) {
296                 if (truesize <= skb->len)
297                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
298
299                 truesize >>= 1;
300                 window >>= 1;
301         }
302         return 0;
303 }
304
305 static void tcp_grow_window(struct sock *sk,
306                             struct sk_buff *skb)
307 {
308         struct tcp_sock *tp = tcp_sk(sk);
309
310         /* Check #1 */
311         if (tp->rcv_ssthresh < tp->window_clamp &&
312             (int)tp->rcv_ssthresh < tcp_space(sk) &&
313             !tcp_memory_pressure) {
314                 int incr;
315
316                 /* Check #2. Increase window, if skb with such overhead
317                  * will fit to rcvbuf in future.
318                  */
319                 if (tcp_win_from_space(skb->truesize) <= skb->len)
320                         incr = 2*tp->advmss;
321                 else
322                         incr = __tcp_grow_window(sk, skb);
323
324                 if (incr) {
325                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
326                         inet_csk(sk)->icsk_ack.quick |= 1;
327                 }
328         }
329 }
330
331 /* 3. Tuning rcvbuf, when connection enters established state. */
332
333 static void tcp_fixup_rcvbuf(struct sock *sk)
334 {
335         struct tcp_sock *tp = tcp_sk(sk);
336         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
337
338         /* Try to select rcvbuf so that 4 mss-sized segments
339          * will fit to window and corresponding skbs will fit to our rcvbuf.
340          * (was 3; 4 is minimum to allow fast retransmit to work.)
341          */
342         while (tcp_win_from_space(rcvmem) < tp->advmss)
343                 rcvmem += 128;
344         if (sk->sk_rcvbuf < 4 * rcvmem)
345                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
346 }
347
348 /* 4. Try to fixup all. It is made immediately after connection enters
349  *    established state.
350  */
351 static void tcp_init_buffer_space(struct sock *sk)
352 {
353         struct tcp_sock *tp = tcp_sk(sk);
354         int maxwin;
355
356         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357                 tcp_fixup_rcvbuf(sk);
358         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359                 tcp_fixup_sndbuf(sk);
360
361         tp->rcvq_space.space = tp->rcv_wnd;
362
363         maxwin = tcp_full_space(sk);
364
365         if (tp->window_clamp >= maxwin) {
366                 tp->window_clamp = maxwin;
367
368                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369                         tp->window_clamp = max(maxwin -
370                                                (maxwin >> sysctl_tcp_app_win),
371                                                4 * tp->advmss);
372         }
373
374         /* Force reservation of one segment. */
375         if (sysctl_tcp_app_win &&
376             tp->window_clamp > 2 * tp->advmss &&
377             tp->window_clamp + tp->advmss > maxwin)
378                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
379
380         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381         tp->snd_cwnd_stamp = tcp_time_stamp;
382 }
383
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock *sk)
386 {
387         struct tcp_sock *tp = tcp_sk(sk);
388         struct inet_connection_sock *icsk = inet_csk(sk);
389
390         icsk->icsk_ack.quick = 0;
391
392         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394             !tcp_memory_pressure &&
395             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
397                                     sysctl_tcp_rmem[2]);
398         }
399         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
400                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
401 }
402
403
404 /* Initialize RCV_MSS value.
405  * RCV_MSS is an our guess about MSS used by the peer.
406  * We haven't any direct information about the MSS.
407  * It's better to underestimate the RCV_MSS rather than overestimate.
408  * Overestimations make us ACKing less frequently than needed.
409  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
410  */
411 void tcp_initialize_rcv_mss(struct sock *sk)
412 {
413         struct tcp_sock *tp = tcp_sk(sk);
414         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
415
416         hint = min(hint, tp->rcv_wnd/2);
417         hint = min(hint, TCP_MIN_RCVMSS);
418         hint = max(hint, TCP_MIN_MSS);
419
420         inet_csk(sk)->icsk_ack.rcv_mss = hint;
421 }
422
423 /* Receiver "autotuning" code.
424  *
425  * The algorithm for RTT estimation w/o timestamps is based on
426  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
427  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
428  *
429  * More detail on this code can be found at
430  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
431  * though this reference is out of date.  A new paper
432  * is pending.
433  */
434 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
435 {
436         u32 new_sample = tp->rcv_rtt_est.rtt;
437         long m = sample;
438
439         if (m == 0)
440                 m = 1;
441
442         if (new_sample != 0) {
443                 /* If we sample in larger samples in the non-timestamp
444                  * case, we could grossly overestimate the RTT especially
445                  * with chatty applications or bulk transfer apps which
446                  * are stalled on filesystem I/O.
447                  *
448                  * Also, since we are only going for a minimum in the
449                  * non-timestamp case, we do not smooth things out
450                  * else with timestamps disabled convergence takes too
451                  * long.
452                  */
453                 if (!win_dep) {
454                         m -= (new_sample >> 3);
455                         new_sample += m;
456                 } else if (m < new_sample)
457                         new_sample = m << 3;
458         } else {
459                 /* No previous measure. */
460                 new_sample = m << 3;
461         }
462
463         if (tp->rcv_rtt_est.rtt != new_sample)
464                 tp->rcv_rtt_est.rtt = new_sample;
465 }
466
467 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
468 {
469         if (tp->rcv_rtt_est.time == 0)
470                 goto new_measure;
471         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
472                 return;
473         tcp_rcv_rtt_update(tp,
474                            jiffies - tp->rcv_rtt_est.time,
475                            1);
476
477 new_measure:
478         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
479         tp->rcv_rtt_est.time = tcp_time_stamp;
480 }
481
482 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
483 {
484         struct tcp_sock *tp = tcp_sk(sk);
485         if (tp->rx_opt.rcv_tsecr &&
486             (TCP_SKB_CB(skb)->end_seq -
487              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
488                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
489 }
490
491 /*
492  * This function should be called every time data is copied to user space.
493  * It calculates the appropriate TCP receive buffer space.
494  */
495 void tcp_rcv_space_adjust(struct sock *sk)
496 {
497         struct tcp_sock *tp = tcp_sk(sk);
498         int time;
499         int space;
500
501         if (tp->rcvq_space.time == 0)
502                 goto new_measure;
503
504         time = tcp_time_stamp - tp->rcvq_space.time;
505         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
506             tp->rcv_rtt_est.rtt == 0)
507                 return;
508
509         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
510
511         space = max(tp->rcvq_space.space, space);
512
513         if (tp->rcvq_space.space != space) {
514                 int rcvmem;
515
516                 tp->rcvq_space.space = space;
517
518                 if (sysctl_tcp_moderate_rcvbuf &&
519                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
520                         int new_clamp = space;
521
522                         /* Receive space grows, normalize in order to
523                          * take into account packet headers and sk_buff
524                          * structure overhead.
525                          */
526                         space /= tp->advmss;
527                         if (!space)
528                                 space = 1;
529                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
530                                   16 + sizeof(struct sk_buff));
531                         while (tcp_win_from_space(rcvmem) < tp->advmss)
532                                 rcvmem += 128;
533                         space *= rcvmem;
534                         space = min(space, sysctl_tcp_rmem[2]);
535                         if (space > sk->sk_rcvbuf) {
536                                 sk->sk_rcvbuf = space;
537
538                                 /* Make the window clamp follow along.  */
539                                 tp->window_clamp = new_clamp;
540                         }
541                 }
542         }
543
544 new_measure:
545         tp->rcvq_space.seq = tp->copied_seq;
546         tp->rcvq_space.time = tcp_time_stamp;
547 }
548
549 /* There is something which you must keep in mind when you analyze the
550  * behavior of the tp->ato delayed ack timeout interval.  When a
551  * connection starts up, we want to ack as quickly as possible.  The
552  * problem is that "good" TCP's do slow start at the beginning of data
553  * transmission.  The means that until we send the first few ACK's the
554  * sender will sit on his end and only queue most of his data, because
555  * he can only send snd_cwnd unacked packets at any given time.  For
556  * each ACK we send, he increments snd_cwnd and transmits more of his
557  * queue.  -DaveM
558  */
559 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
560 {
561         struct tcp_sock *tp = tcp_sk(sk);
562         struct inet_connection_sock *icsk = inet_csk(sk);
563         u32 now;
564
565         inet_csk_schedule_ack(sk);
566
567         tcp_measure_rcv_mss(sk, skb);
568
569         tcp_rcv_rtt_measure(tp);
570
571         now = tcp_time_stamp;
572
573         if (!icsk->icsk_ack.ato) {
574                 /* The _first_ data packet received, initialize
575                  * delayed ACK engine.
576                  */
577                 tcp_incr_quickack(sk);
578                 icsk->icsk_ack.ato = TCP_ATO_MIN;
579         } else {
580                 int m = now - icsk->icsk_ack.lrcvtime;
581
582                 if (m <= TCP_ATO_MIN/2) {
583                         /* The fastest case is the first. */
584                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
585                 } else if (m < icsk->icsk_ack.ato) {
586                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
587                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
588                                 icsk->icsk_ack.ato = icsk->icsk_rto;
589                 } else if (m > icsk->icsk_rto) {
590                         /* Too long gap. Apparently sender failed to
591                          * restart window, so that we send ACKs quickly.
592                          */
593                         tcp_incr_quickack(sk);
594                         sk_mem_reclaim(sk);
595                 }
596         }
597         icsk->icsk_ack.lrcvtime = now;
598
599         TCP_ECN_check_ce(tp, skb);
600
601         if (skb->len >= 128)
602                 tcp_grow_window(sk, skb);
603 }
604
605 static u32 tcp_rto_min(struct sock *sk)
606 {
607         struct dst_entry *dst = __sk_dst_get(sk);
608         u32 rto_min = TCP_RTO_MIN;
609
610         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
611                 rto_min = dst->metrics[RTAX_RTO_MIN-1];
612         return rto_min;
613 }
614
615 /* Called to compute a smoothed rtt estimate. The data fed to this
616  * routine either comes from timestamps, or from segments that were
617  * known _not_ to have been retransmitted [see Karn/Partridge
618  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
619  * piece by Van Jacobson.
620  * NOTE: the next three routines used to be one big routine.
621  * To save cycles in the RFC 1323 implementation it was better to break
622  * it up into three procedures. -- erics
623  */
624 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
625 {
626         struct tcp_sock *tp = tcp_sk(sk);
627         long m = mrtt; /* RTT */
628
629         /*      The following amusing code comes from Jacobson's
630          *      article in SIGCOMM '88.  Note that rtt and mdev
631          *      are scaled versions of rtt and mean deviation.
632          *      This is designed to be as fast as possible
633          *      m stands for "measurement".
634          *
635          *      On a 1990 paper the rto value is changed to:
636          *      RTO = rtt + 4 * mdev
637          *
638          * Funny. This algorithm seems to be very broken.
639          * These formulae increase RTO, when it should be decreased, increase
640          * too slowly, when it should be increased quickly, decrease too quickly
641          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
642          * does not matter how to _calculate_ it. Seems, it was trap
643          * that VJ failed to avoid. 8)
644          */
645         if (m == 0)
646                 m = 1;
647         if (tp->srtt != 0) {
648                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
649                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
650                 if (m < 0) {
651                         m = -m;         /* m is now abs(error) */
652                         m -= (tp->mdev >> 2);   /* similar update on mdev */
653                         /* This is similar to one of Eifel findings.
654                          * Eifel blocks mdev updates when rtt decreases.
655                          * This solution is a bit different: we use finer gain
656                          * for mdev in this case (alpha*beta).
657                          * Like Eifel it also prevents growth of rto,
658                          * but also it limits too fast rto decreases,
659                          * happening in pure Eifel.
660                          */
661                         if (m > 0)
662                                 m >>= 3;
663                 } else {
664                         m -= (tp->mdev >> 2);   /* similar update on mdev */
665                 }
666                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
667                 if (tp->mdev > tp->mdev_max) {
668                         tp->mdev_max = tp->mdev;
669                         if (tp->mdev_max > tp->rttvar)
670                                 tp->rttvar = tp->mdev_max;
671                 }
672                 if (after(tp->snd_una, tp->rtt_seq)) {
673                         if (tp->mdev_max < tp->rttvar)
674                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
675                         tp->rtt_seq = tp->snd_nxt;
676                         tp->mdev_max = tcp_rto_min(sk);
677                 }
678         } else {
679                 /* no previous measure. */
680                 tp->srtt = m<<3;        /* take the measured time to be rtt */
681                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
682                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
683                 tp->rtt_seq = tp->snd_nxt;
684         }
685 }
686
687 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
688  * routine referred to above.
689  */
690 static inline void tcp_set_rto(struct sock *sk)
691 {
692         const struct tcp_sock *tp = tcp_sk(sk);
693         /* Old crap is replaced with new one. 8)
694          *
695          * More seriously:
696          * 1. If rtt variance happened to be less 50msec, it is hallucination.
697          *    It cannot be less due to utterly erratic ACK generation made
698          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
699          *    to do with delayed acks, because at cwnd>2 true delack timeout
700          *    is invisible. Actually, Linux-2.4 also generates erratic
701          *    ACKs in some circumstances.
702          */
703         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
704
705         /* 2. Fixups made earlier cannot be right.
706          *    If we do not estimate RTO correctly without them,
707          *    all the algo is pure shit and should be replaced
708          *    with correct one. It is exactly, which we pretend to do.
709          */
710 }
711
712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
713  * guarantees that rto is higher.
714  */
715 static inline void tcp_bound_rto(struct sock *sk)
716 {
717         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
718                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
719 }
720
721 /* Save metrics learned by this TCP session.
722    This function is called only, when TCP finishes successfully
723    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
724  */
725 void tcp_update_metrics(struct sock *sk)
726 {
727         struct tcp_sock *tp = tcp_sk(sk);
728         struct dst_entry *dst = __sk_dst_get(sk);
729
730         if (sysctl_tcp_nometrics_save)
731                 return;
732
733         dst_confirm(dst);
734
735         if (dst && (dst->flags&DST_HOST)) {
736                 const struct inet_connection_sock *icsk = inet_csk(sk);
737                 int m;
738
739                 if (icsk->icsk_backoff || !tp->srtt) {
740                         /* This session failed to estimate rtt. Why?
741                          * Probably, no packets returned in time.
742                          * Reset our results.
743                          */
744                         if (!(dst_metric_locked(dst, RTAX_RTT)))
745                                 dst->metrics[RTAX_RTT-1] = 0;
746                         return;
747                 }
748
749                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
750
751                 /* If newly calculated rtt larger than stored one,
752                  * store new one. Otherwise, use EWMA. Remember,
753                  * rtt overestimation is always better than underestimation.
754                  */
755                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
756                         if (m <= 0)
757                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
758                         else
759                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
760                 }
761
762                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
763                         if (m < 0)
764                                 m = -m;
765
766                         /* Scale deviation to rttvar fixed point */
767                         m >>= 1;
768                         if (m < tp->mdev)
769                                 m = tp->mdev;
770
771                         if (m >= dst_metric(dst, RTAX_RTTVAR))
772                                 dst->metrics[RTAX_RTTVAR-1] = m;
773                         else
774                                 dst->metrics[RTAX_RTTVAR-1] -=
775                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
776                 }
777
778                 if (tp->snd_ssthresh >= 0xFFFF) {
779                         /* Slow start still did not finish. */
780                         if (dst_metric(dst, RTAX_SSTHRESH) &&
781                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
782                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
783                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
784                         if (!dst_metric_locked(dst, RTAX_CWND) &&
785                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
786                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
787                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
788                            icsk->icsk_ca_state == TCP_CA_Open) {
789                         /* Cong. avoidance phase, cwnd is reliable. */
790                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
791                                 dst->metrics[RTAX_SSTHRESH-1] =
792                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
793                         if (!dst_metric_locked(dst, RTAX_CWND))
794                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
795                 } else {
796                         /* Else slow start did not finish, cwnd is non-sense,
797                            ssthresh may be also invalid.
798                          */
799                         if (!dst_metric_locked(dst, RTAX_CWND))
800                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
801                         if (dst->metrics[RTAX_SSTHRESH-1] &&
802                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
803                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
804                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
805                 }
806
807                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
808                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
809                             tp->reordering != sysctl_tcp_reordering)
810                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
811                 }
812         }
813 }
814
815 /* Numbers are taken from RFC3390.
816  *
817  * John Heffner states:
818  *
819  *      The RFC specifies a window of no more than 4380 bytes
820  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
821  *      is a bit misleading because they use a clamp at 4380 bytes
822  *      rather than use a multiplier in the relevant range.
823  */
824 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
825 {
826         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
827
828         if (!cwnd) {
829                 if (tp->mss_cache > 1460)
830                         cwnd = 2;
831                 else
832                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
833         }
834         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
835 }
836
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
839 {
840         struct tcp_sock *tp = tcp_sk(sk);
841         const struct inet_connection_sock *icsk = inet_csk(sk);
842
843         tp->prior_ssthresh = 0;
844         tp->bytes_acked = 0;
845         if (icsk->icsk_ca_state < TCP_CA_CWR) {
846                 tp->undo_marker = 0;
847                 if (set_ssthresh)
848                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
849                 tp->snd_cwnd = min(tp->snd_cwnd,
850                                    tcp_packets_in_flight(tp) + 1U);
851                 tp->snd_cwnd_cnt = 0;
852                 tp->high_seq = tp->snd_nxt;
853                 tp->snd_cwnd_stamp = tcp_time_stamp;
854                 TCP_ECN_queue_cwr(tp);
855
856                 tcp_set_ca_state(sk, TCP_CA_CWR);
857         }
858 }
859
860 /*
861  * Packet counting of FACK is based on in-order assumptions, therefore TCP
862  * disables it when reordering is detected
863  */
864 static void tcp_disable_fack(struct tcp_sock *tp)
865 {
866         /* RFC3517 uses different metric in lost marker => reset on change */
867         if (tcp_is_fack(tp))
868                 tp->lost_skb_hint = NULL;
869         tp->rx_opt.sack_ok &= ~2;
870 }
871
872 /* Take a notice that peer is sending D-SACKs */
873 static void tcp_dsack_seen(struct tcp_sock *tp)
874 {
875         tp->rx_opt.sack_ok |= 4;
876 }
877
878 /* Initialize metrics on socket. */
879
880 static void tcp_init_metrics(struct sock *sk)
881 {
882         struct tcp_sock *tp = tcp_sk(sk);
883         struct dst_entry *dst = __sk_dst_get(sk);
884
885         if (dst == NULL)
886                 goto reset;
887
888         dst_confirm(dst);
889
890         if (dst_metric_locked(dst, RTAX_CWND))
891                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
892         if (dst_metric(dst, RTAX_SSTHRESH)) {
893                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
894                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
895                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
896         }
897         if (dst_metric(dst, RTAX_REORDERING) &&
898             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
899                 tcp_disable_fack(tp);
900                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
901         }
902
903         if (dst_metric(dst, RTAX_RTT) == 0)
904                 goto reset;
905
906         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
907                 goto reset;
908
909         /* Initial rtt is determined from SYN,SYN-ACK.
910          * The segment is small and rtt may appear much
911          * less than real one. Use per-dst memory
912          * to make it more realistic.
913          *
914          * A bit of theory. RTT is time passed after "normal" sized packet
915          * is sent until it is ACKed. In normal circumstances sending small
916          * packets force peer to delay ACKs and calculation is correct too.
917          * The algorithm is adaptive and, provided we follow specs, it
918          * NEVER underestimate RTT. BUT! If peer tries to make some clever
919          * tricks sort of "quick acks" for time long enough to decrease RTT
920          * to low value, and then abruptly stops to do it and starts to delay
921          * ACKs, wait for troubles.
922          */
923         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
924                 tp->srtt = dst_metric(dst, RTAX_RTT);
925                 tp->rtt_seq = tp->snd_nxt;
926         }
927         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
928                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
929                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
930         }
931         tcp_set_rto(sk);
932         tcp_bound_rto(sk);
933         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
934                 goto reset;
935         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
936         tp->snd_cwnd_stamp = tcp_time_stamp;
937         return;
938
939 reset:
940         /* Play conservative. If timestamps are not
941          * supported, TCP will fail to recalculate correct
942          * rtt, if initial rto is too small. FORGET ALL AND RESET!
943          */
944         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
945                 tp->srtt = 0;
946                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
947                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
948         }
949 }
950
951 static void tcp_update_reordering(struct sock *sk, const int metric,
952                                   const int ts)
953 {
954         struct tcp_sock *tp = tcp_sk(sk);
955         if (metric > tp->reordering) {
956                 tp->reordering = min(TCP_MAX_REORDERING, metric);
957
958                 /* This exciting event is worth to be remembered. 8) */
959                 if (ts)
960                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
961                 else if (tcp_is_reno(tp))
962                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
963                 else if (tcp_is_fack(tp))
964                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
965                 else
966                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
967 #if FASTRETRANS_DEBUG > 1
968                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
969                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
970                        tp->reordering,
971                        tp->fackets_out,
972                        tp->sacked_out,
973                        tp->undo_marker ? tp->undo_retrans : 0);
974 #endif
975                 tcp_disable_fack(tp);
976         }
977 }
978
979 /* This procedure tags the retransmission queue when SACKs arrive.
980  *
981  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
982  * Packets in queue with these bits set are counted in variables
983  * sacked_out, retrans_out and lost_out, correspondingly.
984  *
985  * Valid combinations are:
986  * Tag  InFlight        Description
987  * 0    1               - orig segment is in flight.
988  * S    0               - nothing flies, orig reached receiver.
989  * L    0               - nothing flies, orig lost by net.
990  * R    2               - both orig and retransmit are in flight.
991  * L|R  1               - orig is lost, retransmit is in flight.
992  * S|R  1               - orig reached receiver, retrans is still in flight.
993  * (L|S|R is logically valid, it could occur when L|R is sacked,
994  *  but it is equivalent to plain S and code short-curcuits it to S.
995  *  L|S is logically invalid, it would mean -1 packet in flight 8))
996  *
997  * These 6 states form finite state machine, controlled by the following events:
998  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
999  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1000  * 3. Loss detection event of one of three flavors:
1001  *      A. Scoreboard estimator decided the packet is lost.
1002  *         A'. Reno "three dupacks" marks head of queue lost.
1003  *         A''. Its FACK modfication, head until snd.fack is lost.
1004  *      B. SACK arrives sacking data transmitted after never retransmitted
1005  *         hole was sent out.
1006  *      C. SACK arrives sacking SND.NXT at the moment, when the
1007  *         segment was retransmitted.
1008  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1009  *
1010  * It is pleasant to note, that state diagram turns out to be commutative,
1011  * so that we are allowed not to be bothered by order of our actions,
1012  * when multiple events arrive simultaneously. (see the function below).
1013  *
1014  * Reordering detection.
1015  * --------------------
1016  * Reordering metric is maximal distance, which a packet can be displaced
1017  * in packet stream. With SACKs we can estimate it:
1018  *
1019  * 1. SACK fills old hole and the corresponding segment was not
1020  *    ever retransmitted -> reordering. Alas, we cannot use it
1021  *    when segment was retransmitted.
1022  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1023  *    for retransmitted and already SACKed segment -> reordering..
1024  * Both of these heuristics are not used in Loss state, when we cannot
1025  * account for retransmits accurately.
1026  *
1027  * SACK block validation.
1028  * ----------------------
1029  *
1030  * SACK block range validation checks that the received SACK block fits to
1031  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1032  * Note that SND.UNA is not included to the range though being valid because
1033  * it means that the receiver is rather inconsistent with itself reporting
1034  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1035  * perfectly valid, however, in light of RFC2018 which explicitly states
1036  * that "SACK block MUST reflect the newest segment.  Even if the newest
1037  * segment is going to be discarded ...", not that it looks very clever
1038  * in case of head skb. Due to potentional receiver driven attacks, we
1039  * choose to avoid immediate execution of a walk in write queue due to
1040  * reneging and defer head skb's loss recovery to standard loss recovery
1041  * procedure that will eventually trigger (nothing forbids us doing this).
1042  *
1043  * Implements also blockage to start_seq wrap-around. Problem lies in the
1044  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1045  * there's no guarantee that it will be before snd_nxt (n). The problem
1046  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1047  * wrap (s_w):
1048  *
1049  *         <- outs wnd ->                          <- wrapzone ->
1050  *         u     e      n                         u_w   e_w  s n_w
1051  *         |     |      |                          |     |   |  |
1052  * |<------------+------+----- TCP seqno space --------------+---------->|
1053  * ...-- <2^31 ->|                                           |<--------...
1054  * ...---- >2^31 ------>|                                    |<--------...
1055  *
1056  * Current code wouldn't be vulnerable but it's better still to discard such
1057  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1058  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1059  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1060  * equal to the ideal case (infinite seqno space without wrap caused issues).
1061  *
1062  * With D-SACK the lower bound is extended to cover sequence space below
1063  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1064  * again, D-SACK block must not to go across snd_una (for the same reason as
1065  * for the normal SACK blocks, explained above). But there all simplicity
1066  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1067  * fully below undo_marker they do not affect behavior in anyway and can
1068  * therefore be safely ignored. In rare cases (which are more or less
1069  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1070  * fragmentation and packet reordering past skb's retransmission. To consider
1071  * them correctly, the acceptable range must be extended even more though
1072  * the exact amount is rather hard to quantify. However, tp->max_window can
1073  * be used as an exaggerated estimate.
1074  */
1075 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1076                                   u32 start_seq, u32 end_seq)
1077 {
1078         /* Too far in future, or reversed (interpretation is ambiguous) */
1079         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1080                 return 0;
1081
1082         /* Nasty start_seq wrap-around check (see comments above) */
1083         if (!before(start_seq, tp->snd_nxt))
1084                 return 0;
1085
1086         /* In outstanding window? ...This is valid exit for D-SACKs too.
1087          * start_seq == snd_una is non-sensical (see comments above)
1088          */
1089         if (after(start_seq, tp->snd_una))
1090                 return 1;
1091
1092         if (!is_dsack || !tp->undo_marker)
1093                 return 0;
1094
1095         /* ...Then it's D-SACK, and must reside below snd_una completely */
1096         if (!after(end_seq, tp->snd_una))
1097                 return 0;
1098
1099         if (!before(start_seq, tp->undo_marker))
1100                 return 1;
1101
1102         /* Too old */
1103         if (!after(end_seq, tp->undo_marker))
1104                 return 0;
1105
1106         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1107          *   start_seq < undo_marker and end_seq >= undo_marker.
1108          */
1109         return !before(start_seq, end_seq - tp->max_window);
1110 }
1111
1112 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1113  * Event "C". Later note: FACK people cheated me again 8), we have to account
1114  * for reordering! Ugly, but should help.
1115  *
1116  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1117  * less than what is now known to be received by the other end (derived from
1118  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1119  * retransmitted skbs to avoid some costly processing per ACKs.
1120  */
1121 static void tcp_mark_lost_retrans(struct sock *sk)
1122 {
1123         const struct inet_connection_sock *icsk = inet_csk(sk);
1124         struct tcp_sock *tp = tcp_sk(sk);
1125         struct sk_buff *skb;
1126         int cnt = 0;
1127         u32 new_low_seq = tp->snd_nxt;
1128         u32 received_upto = tcp_highest_sack_seq(tp);
1129
1130         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1131             !after(received_upto, tp->lost_retrans_low) ||
1132             icsk->icsk_ca_state != TCP_CA_Recovery)
1133                 return;
1134
1135         tcp_for_write_queue(skb, sk) {
1136                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1137
1138                 if (skb == tcp_send_head(sk))
1139                         break;
1140                 if (cnt == tp->retrans_out)
1141                         break;
1142                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1143                         continue;
1144
1145                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1146                         continue;
1147
1148                 if (after(received_upto, ack_seq) &&
1149                     (tcp_is_fack(tp) ||
1150                      !before(received_upto,
1151                              ack_seq + tp->reordering * tp->mss_cache))) {
1152                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1153                         tp->retrans_out -= tcp_skb_pcount(skb);
1154
1155                         /* clear lost hint */
1156                         tp->retransmit_skb_hint = NULL;
1157
1158                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1159                                 tp->lost_out += tcp_skb_pcount(skb);
1160                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1161                         }
1162                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1163                 } else {
1164                         if (before(ack_seq, new_low_seq))
1165                                 new_low_seq = ack_seq;
1166                         cnt += tcp_skb_pcount(skb);
1167                 }
1168         }
1169
1170         if (tp->retrans_out)
1171                 tp->lost_retrans_low = new_low_seq;
1172 }
1173
1174 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1175                            struct tcp_sack_block_wire *sp, int num_sacks,
1176                            u32 prior_snd_una)
1177 {
1178         u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1179         u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1180         int dup_sack = 0;
1181
1182         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1183                 dup_sack = 1;
1184                 tcp_dsack_seen(tp);
1185                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1186         } else if (num_sacks > 1) {
1187                 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1188                 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1189
1190                 if (!after(end_seq_0, end_seq_1) &&
1191                     !before(start_seq_0, start_seq_1)) {
1192                         dup_sack = 1;
1193                         tcp_dsack_seen(tp);
1194                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1195                 }
1196         }
1197
1198         /* D-SACK for already forgotten data... Do dumb counting. */
1199         if (dup_sack &&
1200             !after(end_seq_0, prior_snd_una) &&
1201             after(end_seq_0, tp->undo_marker))
1202                 tp->undo_retrans--;
1203
1204         return dup_sack;
1205 }
1206
1207 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1208  * the incoming SACK may not exactly match but we can find smaller MSS
1209  * aligned portion of it that matches. Therefore we might need to fragment
1210  * which may fail and creates some hassle (caller must handle error case
1211  * returns).
1212  */
1213 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1214                                  u32 start_seq, u32 end_seq)
1215 {
1216         int in_sack, err;
1217         unsigned int pkt_len;
1218
1219         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1220                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1221
1222         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1223             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1224
1225                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1226
1227                 if (!in_sack)
1228                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1229                 else
1230                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1231                 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1232                 if (err < 0)
1233                         return err;
1234         }
1235
1236         return in_sack;
1237 }
1238
1239 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1240                            int *reord, int dup_sack, int fack_count)
1241 {
1242         struct tcp_sock *tp = tcp_sk(sk);
1243         u8 sacked = TCP_SKB_CB(skb)->sacked;
1244         int flag = 0;
1245
1246         /* Account D-SACK for retransmitted packet. */
1247         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1248                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1249                         tp->undo_retrans--;
1250                 if (sacked & TCPCB_SACKED_ACKED)
1251                         *reord = min(fack_count, *reord);
1252         }
1253
1254         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1255         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1256                 return flag;
1257
1258         if (!(sacked & TCPCB_SACKED_ACKED)) {
1259                 if (sacked & TCPCB_SACKED_RETRANS) {
1260                         /* If the segment is not tagged as lost,
1261                          * we do not clear RETRANS, believing
1262                          * that retransmission is still in flight.
1263                          */
1264                         if (sacked & TCPCB_LOST) {
1265                                 TCP_SKB_CB(skb)->sacked &=
1266                                         ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1267                                 tp->lost_out -= tcp_skb_pcount(skb);
1268                                 tp->retrans_out -= tcp_skb_pcount(skb);
1269
1270                                 /* clear lost hint */
1271                                 tp->retransmit_skb_hint = NULL;
1272                         }
1273                 } else {
1274                         if (!(sacked & TCPCB_RETRANS)) {
1275                                 /* New sack for not retransmitted frame,
1276                                  * which was in hole. It is reordering.
1277                                  */
1278                                 if (before(TCP_SKB_CB(skb)->seq,
1279                                            tcp_highest_sack_seq(tp)))
1280                                         *reord = min(fack_count, *reord);
1281
1282                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1283                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1284                                         flag |= FLAG_ONLY_ORIG_SACKED;
1285                         }
1286
1287                         if (sacked & TCPCB_LOST) {
1288                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1289                                 tp->lost_out -= tcp_skb_pcount(skb);
1290
1291                                 /* clear lost hint */
1292                                 tp->retransmit_skb_hint = NULL;
1293                         }
1294                 }
1295
1296                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1297                 flag |= FLAG_DATA_SACKED;
1298                 tp->sacked_out += tcp_skb_pcount(skb);
1299
1300                 fack_count += tcp_skb_pcount(skb);
1301
1302                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1303                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1304                     before(TCP_SKB_CB(skb)->seq,
1305                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1306                         tp->lost_cnt_hint += tcp_skb_pcount(skb);
1307
1308                 if (fack_count > tp->fackets_out)
1309                         tp->fackets_out = fack_count;
1310
1311                 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1312                         tcp_advance_highest_sack(sk, skb);
1313         }
1314
1315         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1316          * frames and clear it. undo_retrans is decreased above, L|R frames
1317          * are accounted above as well.
1318          */
1319         if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1320                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1321                 tp->retrans_out -= tcp_skb_pcount(skb);
1322                 tp->retransmit_skb_hint = NULL;
1323         }
1324
1325         return flag;
1326 }
1327
1328 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1329                                         struct tcp_sack_block *next_dup,
1330                                         u32 start_seq, u32 end_seq,
1331                                         int dup_sack_in, int *fack_count,
1332                                         int *reord, int *flag)
1333 {
1334         tcp_for_write_queue_from(skb, sk) {
1335                 int in_sack = 0;
1336                 int dup_sack = dup_sack_in;
1337
1338                 if (skb == tcp_send_head(sk))
1339                         break;
1340
1341                 /* queue is in-order => we can short-circuit the walk early */
1342                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1343                         break;
1344
1345                 if ((next_dup != NULL) &&
1346                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1347                         in_sack = tcp_match_skb_to_sack(sk, skb,
1348                                                         next_dup->start_seq,
1349                                                         next_dup->end_seq);
1350                         if (in_sack > 0)
1351                                 dup_sack = 1;
1352                 }
1353
1354                 if (in_sack <= 0)
1355                         in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq);
1356                 if (unlikely(in_sack < 0))
1357                         break;
1358
1359                 if (in_sack)
1360                         *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack, *fack_count);
1361
1362                 *fack_count += tcp_skb_pcount(skb);
1363         }
1364         return skb;
1365 }
1366
1367 /* Avoid all extra work that is being done by sacktag while walking in
1368  * a normal way
1369  */
1370 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1371                                         u32 skip_to_seq)
1372 {
1373         tcp_for_write_queue_from(skb, sk) {
1374                 if (skb == tcp_send_head(sk))
1375                         break;
1376
1377                 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1378                         break;
1379         }
1380         return skb;
1381 }
1382
1383 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1384                                                 struct sock *sk,
1385                                                 struct tcp_sack_block *next_dup,
1386                                                 u32 skip_to_seq,
1387                                                 int *fack_count, int *reord,
1388                                                 int *flag)
1389 {
1390         if (next_dup == NULL)
1391                 return skb;
1392
1393         if (before(next_dup->start_seq, skip_to_seq)) {
1394                 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1395                 tcp_sacktag_walk(skb, sk, NULL,
1396                                  next_dup->start_seq, next_dup->end_seq,
1397                                  1, fack_count, reord, flag);
1398         }
1399
1400         return skb;
1401 }
1402
1403 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1404 {
1405         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1406 }
1407
1408 static int
1409 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1410 {
1411         const struct inet_connection_sock *icsk = inet_csk(sk);
1412         struct tcp_sock *tp = tcp_sk(sk);
1413         unsigned char *ptr = (skb_transport_header(ack_skb) +
1414                               TCP_SKB_CB(ack_skb)->sacked);
1415         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1416         struct tcp_sack_block sp[4];
1417         struct tcp_sack_block *cache;
1418         struct sk_buff *skb;
1419         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1420         int used_sacks;
1421         int reord = tp->packets_out;
1422         int flag = 0;
1423         int found_dup_sack = 0;
1424         int fack_count;
1425         int i, j;
1426         int first_sack_index;
1427
1428         if (!tp->sacked_out) {
1429                 if (WARN_ON(tp->fackets_out))
1430                         tp->fackets_out = 0;
1431                 tcp_highest_sack_reset(sk);
1432         }
1433
1434         found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire,
1435                                          num_sacks, prior_snd_una);
1436         if (found_dup_sack)
1437                 flag |= FLAG_DSACKING_ACK;
1438
1439         /* Eliminate too old ACKs, but take into
1440          * account more or less fresh ones, they can
1441          * contain valid SACK info.
1442          */
1443         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1444                 return 0;
1445
1446         if (!tp->packets_out)
1447                 goto out;
1448
1449         used_sacks = 0;
1450         first_sack_index = 0;
1451         for (i = 0; i < num_sacks; i++) {
1452                 int dup_sack = !i && found_dup_sack;
1453
1454                 sp[used_sacks].start_seq = ntohl(get_unaligned(&sp_wire[i].start_seq));
1455                 sp[used_sacks].end_seq = ntohl(get_unaligned(&sp_wire[i].end_seq));
1456
1457                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1458                                             sp[used_sacks].start_seq,
1459                                             sp[used_sacks].end_seq)) {
1460                         if (dup_sack) {
1461                                 if (!tp->undo_marker)
1462                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1463                                 else
1464                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1465                         } else {
1466                                 /* Don't count olds caused by ACK reordering */
1467                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1468                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1469                                         continue;
1470                                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1471                         }
1472                         if (i == 0)
1473                                 first_sack_index = -1;
1474                         continue;
1475                 }
1476
1477                 /* Ignore very old stuff early */
1478                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1479                         continue;
1480
1481                 used_sacks++;
1482         }
1483
1484         /* order SACK blocks to allow in order walk of the retrans queue */
1485         for (i = used_sacks - 1; i > 0; i--) {
1486                 for (j = 0; j < i; j++){
1487                         if (after(sp[j].start_seq, sp[j+1].start_seq)) {
1488                                 struct tcp_sack_block tmp;
1489
1490                                 tmp = sp[j];
1491                                 sp[j] = sp[j+1];
1492                                 sp[j+1] = tmp;
1493
1494                                 /* Track where the first SACK block goes to */
1495                                 if (j == first_sack_index)
1496                                         first_sack_index = j+1;
1497                         }
1498                 }
1499         }
1500
1501         skb = tcp_write_queue_head(sk);
1502         fack_count = 0;
1503         i = 0;
1504
1505         if (!tp->sacked_out) {
1506                 /* It's already past, so skip checking against it */
1507                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1508         } else {
1509                 cache = tp->recv_sack_cache;
1510                 /* Skip empty blocks in at head of the cache */
1511                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1512                        !cache->end_seq)
1513                         cache++;
1514         }
1515
1516         while (i < used_sacks) {
1517                 u32 start_seq = sp[i].start_seq;
1518                 u32 end_seq = sp[i].end_seq;
1519                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1520                 struct tcp_sack_block *next_dup = NULL;
1521
1522                 if (found_dup_sack && ((i + 1) == first_sack_index))
1523                         next_dup = &sp[i + 1];
1524
1525                 /* Event "B" in the comment above. */
1526                 if (after(end_seq, tp->high_seq))
1527                         flag |= FLAG_DATA_LOST;
1528
1529                 /* Skip too early cached blocks */
1530                 while (tcp_sack_cache_ok(tp, cache) &&
1531                        !before(start_seq, cache->end_seq))
1532                         cache++;
1533
1534                 /* Can skip some work by looking recv_sack_cache? */
1535                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1536                     after(end_seq, cache->start_seq)) {
1537
1538                         /* Head todo? */
1539                         if (before(start_seq, cache->start_seq)) {
1540                                 skb = tcp_sacktag_skip(skb, sk, start_seq);
1541                                 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq,
1542                                                        cache->start_seq, dup_sack,
1543                                                        &fack_count, &reord, &flag);
1544                         }
1545
1546                         /* Rest of the block already fully processed? */
1547                         if (!after(end_seq, cache->end_seq))
1548                                 goto advance_sp;
1549
1550                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, cache->end_seq,
1551                                                        &fack_count, &reord, &flag);
1552
1553                         /* ...tail remains todo... */
1554                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1555                                 /* ...but better entrypoint exists! */
1556                                 skb = tcp_highest_sack(sk);
1557                                 if (skb == NULL)
1558                                         break;
1559                                 fack_count = tp->fackets_out;
1560                                 cache++;
1561                                 goto walk;
1562                         }
1563
1564                         skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1565                         /* Check overlap against next cached too (past this one already) */
1566                         cache++;
1567                         continue;
1568                 }
1569
1570                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1571                         skb = tcp_highest_sack(sk);
1572                         if (skb == NULL)
1573                                 break;
1574                         fack_count = tp->fackets_out;
1575                 }
1576                 skb = tcp_sacktag_skip(skb, sk, start_seq);
1577
1578 walk:
1579                 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1580                                        dup_sack, &fack_count, &reord, &flag);
1581
1582 advance_sp:
1583                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1584                  * due to in-order walk
1585                  */
1586                 if (after(end_seq, tp->frto_highmark))
1587                         flag &= ~FLAG_ONLY_ORIG_SACKED;
1588
1589                 i++;
1590         }
1591
1592         /* Clear the head of the cache sack blocks so we can skip it next time */
1593         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1594                 tp->recv_sack_cache[i].start_seq = 0;
1595                 tp->recv_sack_cache[i].end_seq = 0;
1596         }
1597         for (j = 0; j < used_sacks; j++)
1598                 tp->recv_sack_cache[i++] = sp[j];
1599
1600         tcp_mark_lost_retrans(sk);
1601
1602         tcp_verify_left_out(tp);
1603
1604         if ((reord < tp->fackets_out) &&
1605             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1606             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1607                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1608
1609 out:
1610
1611 #if FASTRETRANS_DEBUG > 0
1612         BUG_TRAP((int)tp->sacked_out >= 0);
1613         BUG_TRAP((int)tp->lost_out >= 0);
1614         BUG_TRAP((int)tp->retrans_out >= 0);
1615         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1616 #endif
1617         return flag;
1618 }
1619
1620 /* If we receive more dupacks than we expected counting segments
1621  * in assumption of absent reordering, interpret this as reordering.
1622  * The only another reason could be bug in receiver TCP.
1623  */
1624 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1625 {
1626         struct tcp_sock *tp = tcp_sk(sk);
1627         u32 holes;
1628
1629         holes = max(tp->lost_out, 1U);
1630         holes = min(holes, tp->packets_out);
1631
1632         if ((tp->sacked_out + holes) > tp->packets_out) {
1633                 tp->sacked_out = tp->packets_out - holes;
1634                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1635         }
1636 }
1637
1638 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1639
1640 static void tcp_add_reno_sack(struct sock *sk)
1641 {
1642         struct tcp_sock *tp = tcp_sk(sk);
1643         tp->sacked_out++;
1644         tcp_check_reno_reordering(sk, 0);
1645         tcp_verify_left_out(tp);
1646 }
1647
1648 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1649
1650 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1651 {
1652         struct tcp_sock *tp = tcp_sk(sk);
1653
1654         if (acked > 0) {
1655                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1656                 if (acked-1 >= tp->sacked_out)
1657                         tp->sacked_out = 0;
1658                 else
1659                         tp->sacked_out -= acked-1;
1660         }
1661         tcp_check_reno_reordering(sk, acked);
1662         tcp_verify_left_out(tp);
1663 }
1664
1665 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1666 {
1667         tp->sacked_out = 0;
1668 }
1669
1670 /* F-RTO can only be used if TCP has never retransmitted anything other than
1671  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1672  */
1673 int tcp_use_frto(struct sock *sk)
1674 {
1675         const struct tcp_sock *tp = tcp_sk(sk);
1676         struct sk_buff *skb;
1677
1678         if (!sysctl_tcp_frto)
1679                 return 0;
1680
1681         if (IsSackFrto())
1682                 return 1;
1683
1684         /* Avoid expensive walking of rexmit queue if possible */
1685         if (tp->retrans_out > 1)
1686                 return 0;
1687
1688         skb = tcp_write_queue_head(sk);
1689         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1690         tcp_for_write_queue_from(skb, sk) {
1691                 if (skb == tcp_send_head(sk))
1692                         break;
1693                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1694                         return 0;
1695                 /* Short-circuit when first non-SACKed skb has been checked */
1696                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1697                         break;
1698         }
1699         return 1;
1700 }
1701
1702 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1703  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1704  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1705  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1706  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1707  * bits are handled if the Loss state is really to be entered (in
1708  * tcp_enter_frto_loss).
1709  *
1710  * Do like tcp_enter_loss() would; when RTO expires the second time it
1711  * does:
1712  *  "Reduce ssthresh if it has not yet been made inside this window."
1713  */
1714 void tcp_enter_frto(struct sock *sk)
1715 {
1716         const struct inet_connection_sock *icsk = inet_csk(sk);
1717         struct tcp_sock *tp = tcp_sk(sk);
1718         struct sk_buff *skb;
1719
1720         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1721             tp->snd_una == tp->high_seq ||
1722             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1723              !icsk->icsk_retransmits)) {
1724                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1725                 /* Our state is too optimistic in ssthresh() call because cwnd
1726                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1727                  * recovery has not yet completed. Pattern would be this: RTO,
1728                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1729                  * up here twice).
1730                  * RFC4138 should be more specific on what to do, even though
1731                  * RTO is quite unlikely to occur after the first Cumulative ACK
1732                  * due to back-off and complexity of triggering events ...
1733                  */
1734                 if (tp->frto_counter) {
1735                         u32 stored_cwnd;
1736                         stored_cwnd = tp->snd_cwnd;
1737                         tp->snd_cwnd = 2;
1738                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1739                         tp->snd_cwnd = stored_cwnd;
1740                 } else {
1741                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1742                 }
1743                 /* ... in theory, cong.control module could do "any tricks" in
1744                  * ssthresh(), which means that ca_state, lost bits and lost_out
1745                  * counter would have to be faked before the call occurs. We
1746                  * consider that too expensive, unlikely and hacky, so modules
1747                  * using these in ssthresh() must deal these incompatibility
1748                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1749                  */
1750                 tcp_ca_event(sk, CA_EVENT_FRTO);
1751         }
1752
1753         tp->undo_marker = tp->snd_una;
1754         tp->undo_retrans = 0;
1755
1756         skb = tcp_write_queue_head(sk);
1757         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1758                 tp->undo_marker = 0;
1759         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1760                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1761                 tp->retrans_out -= tcp_skb_pcount(skb);
1762         }
1763         tcp_verify_left_out(tp);
1764
1765         /* Too bad if TCP was application limited */
1766         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1767
1768         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1769          * The last condition is necessary at least in tp->frto_counter case.
1770          */
1771         if (IsSackFrto() && (tp->frto_counter ||
1772             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1773             after(tp->high_seq, tp->snd_una)) {
1774                 tp->frto_highmark = tp->high_seq;
1775         } else {
1776                 tp->frto_highmark = tp->snd_nxt;
1777         }
1778         tcp_set_ca_state(sk, TCP_CA_Disorder);
1779         tp->high_seq = tp->snd_nxt;
1780         tp->frto_counter = 1;
1781 }
1782
1783 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1784  * which indicates that we should follow the traditional RTO recovery,
1785  * i.e. mark everything lost and do go-back-N retransmission.
1786  */
1787 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1788 {
1789         struct tcp_sock *tp = tcp_sk(sk);
1790         struct sk_buff *skb;
1791
1792         tp->lost_out = 0;
1793         tp->retrans_out = 0;
1794         if (tcp_is_reno(tp))
1795                 tcp_reset_reno_sack(tp);
1796
1797         tcp_for_write_queue(skb, sk) {
1798                 if (skb == tcp_send_head(sk))
1799                         break;
1800
1801                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1802                 /*
1803                  * Count the retransmission made on RTO correctly (only when
1804                  * waiting for the first ACK and did not get it)...
1805                  */
1806                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1807                         /* For some reason this R-bit might get cleared? */
1808                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1809                                 tp->retrans_out += tcp_skb_pcount(skb);
1810                         /* ...enter this if branch just for the first segment */
1811                         flag |= FLAG_DATA_ACKED;
1812                 } else {
1813                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1814                                 tp->undo_marker = 0;
1815                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1816                 }
1817
1818                 /* Don't lost mark skbs that were fwd transmitted after RTO */
1819                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1820                     !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1821                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1822                         tp->lost_out += tcp_skb_pcount(skb);
1823                 }
1824         }
1825         tcp_verify_left_out(tp);
1826
1827         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1828         tp->snd_cwnd_cnt = 0;
1829         tp->snd_cwnd_stamp = tcp_time_stamp;
1830         tp->frto_counter = 0;
1831         tp->bytes_acked = 0;
1832
1833         tp->reordering = min_t(unsigned int, tp->reordering,
1834                                              sysctl_tcp_reordering);
1835         tcp_set_ca_state(sk, TCP_CA_Loss);
1836         tp->high_seq = tp->frto_highmark;
1837         TCP_ECN_queue_cwr(tp);
1838
1839         tcp_clear_retrans_hints_partial(tp);
1840 }
1841
1842 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1843 {
1844         tp->retrans_out = 0;
1845         tp->lost_out = 0;
1846
1847         tp->undo_marker = 0;
1848         tp->undo_retrans = 0;
1849 }
1850
1851 void tcp_clear_retrans(struct tcp_sock *tp)
1852 {
1853         tcp_clear_retrans_partial(tp);
1854
1855         tp->fackets_out = 0;
1856         tp->sacked_out = 0;
1857 }
1858
1859 /* Enter Loss state. If "how" is not zero, forget all SACK information
1860  * and reset tags completely, otherwise preserve SACKs. If receiver
1861  * dropped its ofo queue, we will know this due to reneging detection.
1862  */
1863 void tcp_enter_loss(struct sock *sk, int how)
1864 {
1865         const struct inet_connection_sock *icsk = inet_csk(sk);
1866         struct tcp_sock *tp = tcp_sk(sk);
1867         struct sk_buff *skb;
1868
1869         /* Reduce ssthresh if it has not yet been made inside this window. */
1870         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1871             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1872                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1873                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1874                 tcp_ca_event(sk, CA_EVENT_LOSS);
1875         }
1876         tp->snd_cwnd       = 1;
1877         tp->snd_cwnd_cnt   = 0;
1878         tp->snd_cwnd_stamp = tcp_time_stamp;
1879
1880         tp->bytes_acked = 0;
1881         tcp_clear_retrans_partial(tp);
1882
1883         if (tcp_is_reno(tp))
1884                 tcp_reset_reno_sack(tp);
1885
1886         if (!how) {
1887                 /* Push undo marker, if it was plain RTO and nothing
1888                  * was retransmitted. */
1889                 tp->undo_marker = tp->snd_una;
1890                 tcp_clear_retrans_hints_partial(tp);
1891         } else {
1892                 tp->sacked_out = 0;
1893                 tp->fackets_out = 0;
1894                 tcp_clear_all_retrans_hints(tp);
1895         }
1896
1897         tcp_for_write_queue(skb, sk) {
1898                 if (skb == tcp_send_head(sk))
1899                         break;
1900
1901                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1902                         tp->undo_marker = 0;
1903                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1904                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1905                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1906                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1907                         tp->lost_out += tcp_skb_pcount(skb);
1908                 }
1909         }
1910         tcp_verify_left_out(tp);
1911
1912         tp->reordering = min_t(unsigned int, tp->reordering,
1913                                              sysctl_tcp_reordering);
1914         tcp_set_ca_state(sk, TCP_CA_Loss);
1915         tp->high_seq = tp->snd_nxt;
1916         TCP_ECN_queue_cwr(tp);
1917         /* Abort F-RTO algorithm if one is in progress */
1918         tp->frto_counter = 0;
1919 }
1920
1921 static int tcp_check_sack_reneging(struct sock *sk)
1922 {
1923         struct sk_buff *skb;
1924
1925         /* If ACK arrived pointing to a remembered SACK,
1926          * it means that our remembered SACKs do not reflect
1927          * real state of receiver i.e.
1928          * receiver _host_ is heavily congested (or buggy).
1929          * Do processing similar to RTO timeout.
1930          */
1931         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1932             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1933                 struct inet_connection_sock *icsk = inet_csk(sk);
1934                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1935
1936                 tcp_enter_loss(sk, 1);
1937                 icsk->icsk_retransmits++;
1938                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1939                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1940                                           icsk->icsk_rto, TCP_RTO_MAX);
1941                 return 1;
1942         }
1943         return 0;
1944 }
1945
1946 static inline int tcp_fackets_out(struct tcp_sock *tp)
1947 {
1948         return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1949 }
1950
1951 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1952  * counter when SACK is enabled (without SACK, sacked_out is used for
1953  * that purpose).
1954  *
1955  * Instead, with FACK TCP uses fackets_out that includes both SACKed
1956  * segments up to the highest received SACK block so far and holes in
1957  * between them.
1958  *
1959  * With reordering, holes may still be in flight, so RFC3517 recovery
1960  * uses pure sacked_out (total number of SACKed segments) even though
1961  * it violates the RFC that uses duplicate ACKs, often these are equal
1962  * but when e.g. out-of-window ACKs or packet duplication occurs,
1963  * they differ. Since neither occurs due to loss, TCP should really
1964  * ignore them.
1965  */
1966 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
1967 {
1968         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1969 }
1970
1971 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1972 {
1973         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1974 }
1975
1976 static inline int tcp_head_timedout(struct sock *sk)
1977 {
1978         struct tcp_sock *tp = tcp_sk(sk);
1979
1980         return tp->packets_out &&
1981                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1982 }
1983
1984 /* Linux NewReno/SACK/FACK/ECN state machine.
1985  * --------------------------------------
1986  *
1987  * "Open"       Normal state, no dubious events, fast path.
1988  * "Disorder"   In all the respects it is "Open",
1989  *              but requires a bit more attention. It is entered when
1990  *              we see some SACKs or dupacks. It is split of "Open"
1991  *              mainly to move some processing from fast path to slow one.
1992  * "CWR"        CWND was reduced due to some Congestion Notification event.
1993  *              It can be ECN, ICMP source quench, local device congestion.
1994  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1995  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1996  *
1997  * tcp_fastretrans_alert() is entered:
1998  * - each incoming ACK, if state is not "Open"
1999  * - when arrived ACK is unusual, namely:
2000  *      * SACK
2001  *      * Duplicate ACK.
2002  *      * ECN ECE.
2003  *
2004  * Counting packets in flight is pretty simple.
2005  *
2006  *      in_flight = packets_out - left_out + retrans_out
2007  *
2008  *      packets_out is SND.NXT-SND.UNA counted in packets.
2009  *
2010  *      retrans_out is number of retransmitted segments.
2011  *
2012  *      left_out is number of segments left network, but not ACKed yet.
2013  *
2014  *              left_out = sacked_out + lost_out
2015  *
2016  *     sacked_out: Packets, which arrived to receiver out of order
2017  *                 and hence not ACKed. With SACKs this number is simply
2018  *                 amount of SACKed data. Even without SACKs
2019  *                 it is easy to give pretty reliable estimate of this number,
2020  *                 counting duplicate ACKs.
2021  *
2022  *       lost_out: Packets lost by network. TCP has no explicit
2023  *                 "loss notification" feedback from network (for now).
2024  *                 It means that this number can be only _guessed_.
2025  *                 Actually, it is the heuristics to predict lossage that
2026  *                 distinguishes different algorithms.
2027  *
2028  *      F.e. after RTO, when all the queue is considered as lost,
2029  *      lost_out = packets_out and in_flight = retrans_out.
2030  *
2031  *              Essentially, we have now two algorithms counting
2032  *              lost packets.
2033  *
2034  *              FACK: It is the simplest heuristics. As soon as we decided
2035  *              that something is lost, we decide that _all_ not SACKed
2036  *              packets until the most forward SACK are lost. I.e.
2037  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2038  *              It is absolutely correct estimate, if network does not reorder
2039  *              packets. And it loses any connection to reality when reordering
2040  *              takes place. We use FACK by default until reordering
2041  *              is suspected on the path to this destination.
2042  *
2043  *              NewReno: when Recovery is entered, we assume that one segment
2044  *              is lost (classic Reno). While we are in Recovery and
2045  *              a partial ACK arrives, we assume that one more packet
2046  *              is lost (NewReno). This heuristics are the same in NewReno
2047  *              and SACK.
2048  *
2049  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2050  *  deflation etc. CWND is real congestion window, never inflated, changes
2051  *  only according to classic VJ rules.
2052  *
2053  * Really tricky (and requiring careful tuning) part of algorithm
2054  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2055  * The first determines the moment _when_ we should reduce CWND and,
2056  * hence, slow down forward transmission. In fact, it determines the moment
2057  * when we decide that hole is caused by loss, rather than by a reorder.
2058  *
2059  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2060  * holes, caused by lost packets.
2061  *
2062  * And the most logically complicated part of algorithm is undo
2063  * heuristics. We detect false retransmits due to both too early
2064  * fast retransmit (reordering) and underestimated RTO, analyzing
2065  * timestamps and D-SACKs. When we detect that some segments were
2066  * retransmitted by mistake and CWND reduction was wrong, we undo
2067  * window reduction and abort recovery phase. This logic is hidden
2068  * inside several functions named tcp_try_undo_<something>.
2069  */
2070
2071 /* This function decides, when we should leave Disordered state
2072  * and enter Recovery phase, reducing congestion window.
2073  *
2074  * Main question: may we further continue forward transmission
2075  * with the same cwnd?
2076  */
2077 static int tcp_time_to_recover(struct sock *sk)
2078 {
2079         struct tcp_sock *tp = tcp_sk(sk);
2080         __u32 packets_out;
2081
2082         /* Do not perform any recovery during F-RTO algorithm */
2083         if (tp->frto_counter)
2084                 return 0;
2085
2086         /* Trick#1: The loss is proven. */
2087         if (tp->lost_out)
2088                 return 1;
2089
2090         /* Not-A-Trick#2 : Classic rule... */
2091         if (tcp_dupack_heurestics(tp) > tp->reordering)
2092                 return 1;
2093
2094         /* Trick#3 : when we use RFC2988 timer restart, fast
2095          * retransmit can be triggered by timeout of queue head.
2096          */
2097         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2098                 return 1;
2099
2100         /* Trick#4: It is still not OK... But will it be useful to delay
2101          * recovery more?
2102          */
2103         packets_out = tp->packets_out;
2104         if (packets_out <= tp->reordering &&
2105             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2106             !tcp_may_send_now(sk)) {
2107                 /* We have nothing to send. This connection is limited
2108                  * either by receiver window or by application.
2109                  */
2110                 return 1;
2111         }
2112
2113         return 0;
2114 }
2115
2116 /* RFC: This is from the original, I doubt that this is necessary at all:
2117  * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2118  * retransmitted past LOST markings in the first place? I'm not fully sure
2119  * about undo and end of connection cases, which can cause R without L?
2120  */
2121 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
2122                                        struct sk_buff *skb)
2123 {
2124         if ((tp->retransmit_skb_hint != NULL) &&
2125             before(TCP_SKB_CB(skb)->seq,
2126             TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
2127                 tp->retransmit_skb_hint = NULL;
2128 }
2129
2130 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2131  * is against sacked "cnt", otherwise it's against facked "cnt"
2132  */
2133 static void tcp_mark_head_lost(struct sock *sk, int packets, int fast_rexmit)
2134 {
2135         struct tcp_sock *tp = tcp_sk(sk);
2136         struct sk_buff *skb;
2137         int cnt;
2138
2139         BUG_TRAP(packets <= tp->packets_out);
2140         if (tp->lost_skb_hint) {
2141                 skb = tp->lost_skb_hint;
2142                 cnt = tp->lost_cnt_hint;
2143         } else {
2144                 skb = tcp_write_queue_head(sk);
2145                 cnt = 0;
2146         }
2147
2148         tcp_for_write_queue_from(skb, sk) {
2149                 if (skb == tcp_send_head(sk))
2150                         break;
2151                 /* TODO: do this better */
2152                 /* this is not the most efficient way to do this... */
2153                 tp->lost_skb_hint = skb;
2154                 tp->lost_cnt_hint = cnt;
2155
2156                 if (tcp_is_fack(tp) ||
2157                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2158                         cnt += tcp_skb_pcount(skb);
2159
2160                 if (((!fast_rexmit || (tp->lost_out > 0)) && (cnt > packets)) ||
2161                      after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2162                         break;
2163                 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2164                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2165                         tp->lost_out += tcp_skb_pcount(skb);
2166                         tcp_verify_retransmit_hint(tp, skb);
2167                 }
2168         }
2169         tcp_verify_left_out(tp);
2170 }
2171
2172 /* Account newly detected lost packet(s) */
2173
2174 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2175 {
2176         struct tcp_sock *tp = tcp_sk(sk);
2177
2178         if (tcp_is_reno(tp)) {
2179                 tcp_mark_head_lost(sk, 1, fast_rexmit);
2180         } else if (tcp_is_fack(tp)) {
2181                 int lost = tp->fackets_out - tp->reordering;
2182                 if (lost <= 0)
2183                         lost = 1;
2184                 tcp_mark_head_lost(sk, lost, fast_rexmit);
2185         } else {
2186                 int sacked_upto = tp->sacked_out - tp->reordering;
2187                 if (sacked_upto < 0)
2188                         sacked_upto = 0;
2189                 tcp_mark_head_lost(sk, sacked_upto, fast_rexmit);
2190         }
2191
2192         /* New heuristics: it is possible only after we switched
2193          * to restart timer each time when something is ACKed.
2194          * Hence, we can detect timed out packets during fast
2195          * retransmit without falling to slow start.
2196          */
2197         if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2198                 struct sk_buff *skb;
2199
2200                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2201                         : tcp_write_queue_head(sk);
2202
2203                 tcp_for_write_queue_from(skb, sk) {
2204                         if (skb == tcp_send_head(sk))
2205                                 break;
2206                         if (!tcp_skb_timedout(sk, skb))
2207                                 break;
2208
2209                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2210                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2211                                 tp->lost_out += tcp_skb_pcount(skb);
2212                                 tcp_verify_retransmit_hint(tp, skb);
2213                         }
2214                 }
2215
2216                 tp->scoreboard_skb_hint = skb;
2217
2218                 tcp_verify_left_out(tp);
2219         }
2220 }
2221
2222 /* CWND moderation, preventing bursts due to too big ACKs
2223  * in dubious situations.
2224  */
2225 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2226 {
2227         tp->snd_cwnd = min(tp->snd_cwnd,
2228                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2229         tp->snd_cwnd_stamp = tcp_time_stamp;
2230 }
2231
2232 /* Lower bound on congestion window is slow start threshold
2233  * unless congestion avoidance choice decides to overide it.
2234  */
2235 static inline u32 tcp_cwnd_min(const struct sock *sk)
2236 {
2237         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2238
2239         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2240 }
2241
2242 /* Decrease cwnd each second ack. */
2243 static void tcp_cwnd_down(struct sock *sk, int flag)
2244 {
2245         struct tcp_sock *tp = tcp_sk(sk);
2246         int decr = tp->snd_cwnd_cnt + 1;
2247
2248         if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2249             (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2250                 tp->snd_cwnd_cnt = decr&1;
2251                 decr >>= 1;
2252
2253                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2254                         tp->snd_cwnd -= decr;
2255
2256                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2257                 tp->snd_cwnd_stamp = tcp_time_stamp;
2258         }
2259 }
2260
2261 /* Nothing was retransmitted or returned timestamp is less
2262  * than timestamp of the first retransmission.
2263  */
2264 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2265 {
2266         return !tp->retrans_stamp ||
2267                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2268                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2269 }
2270
2271 /* Undo procedures. */
2272
2273 #if FASTRETRANS_DEBUG > 1
2274 static void DBGUNDO(struct sock *sk, const char *msg)
2275 {
2276         struct tcp_sock *tp = tcp_sk(sk);
2277         struct inet_sock *inet = inet_sk(sk);
2278
2279         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2280                msg,
2281                NIPQUAD(inet->daddr), ntohs(inet->dport),
2282                tp->snd_cwnd, tcp_left_out(tp),
2283                tp->snd_ssthresh, tp->prior_ssthresh,
2284                tp->packets_out);
2285 }
2286 #else
2287 #define DBGUNDO(x...) do { } while (0)
2288 #endif
2289
2290 static void tcp_undo_cwr(struct sock *sk, const int undo)
2291 {
2292         struct tcp_sock *tp = tcp_sk(sk);
2293
2294         if (tp->prior_ssthresh) {
2295                 const struct inet_connection_sock *icsk = inet_csk(sk);
2296
2297                 if (icsk->icsk_ca_ops->undo_cwnd)
2298                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2299                 else
2300                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2301
2302                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2303                         tp->snd_ssthresh = tp->prior_ssthresh;
2304                         TCP_ECN_withdraw_cwr(tp);
2305                 }
2306         } else {
2307                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2308         }
2309         tcp_moderate_cwnd(tp);
2310         tp->snd_cwnd_stamp = tcp_time_stamp;
2311
2312         /* There is something screwy going on with the retrans hints after
2313            an undo */
2314         tcp_clear_all_retrans_hints(tp);
2315 }
2316
2317 static inline int tcp_may_undo(struct tcp_sock *tp)
2318 {
2319         return tp->undo_marker &&
2320                 (!tp->undo_retrans || tcp_packet_delayed(tp));
2321 }
2322
2323 /* People celebrate: "We love our President!" */
2324 static int tcp_try_undo_recovery(struct sock *sk)
2325 {
2326         struct tcp_sock *tp = tcp_sk(sk);
2327
2328         if (tcp_may_undo(tp)) {
2329                 /* Happy end! We did not retransmit anything
2330                  * or our original transmission succeeded.
2331                  */
2332                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2333                 tcp_undo_cwr(sk, 1);
2334                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2335                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2336                 else
2337                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2338                 tp->undo_marker = 0;
2339         }
2340         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2341                 /* Hold old state until something *above* high_seq
2342                  * is ACKed. For Reno it is MUST to prevent false
2343                  * fast retransmits (RFC2582). SACK TCP is safe. */
2344                 tcp_moderate_cwnd(tp);
2345                 return 1;
2346         }
2347         tcp_set_ca_state(sk, TCP_CA_Open);
2348         return 0;
2349 }
2350
2351 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2352 static void tcp_try_undo_dsack(struct sock *sk)
2353 {
2354         struct tcp_sock *tp = tcp_sk(sk);
2355
2356         if (tp->undo_marker && !tp->undo_retrans) {
2357                 DBGUNDO(sk, "D-SACK");
2358                 tcp_undo_cwr(sk, 1);
2359                 tp->undo_marker = 0;
2360                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2361         }
2362 }
2363
2364 /* Undo during fast recovery after partial ACK. */
2365
2366 static int tcp_try_undo_partial(struct sock *sk, int acked)
2367 {
2368         struct tcp_sock *tp = tcp_sk(sk);
2369         /* Partial ACK arrived. Force Hoe's retransmit. */
2370         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2371
2372         if (tcp_may_undo(tp)) {
2373                 /* Plain luck! Hole if filled with delayed
2374                  * packet, rather than with a retransmit.
2375                  */
2376                 if (tp->retrans_out == 0)
2377                         tp->retrans_stamp = 0;
2378
2379                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2380
2381                 DBGUNDO(sk, "Hoe");
2382                 tcp_undo_cwr(sk, 0);
2383                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2384
2385                 /* So... Do not make Hoe's retransmit yet.
2386                  * If the first packet was delayed, the rest
2387                  * ones are most probably delayed as well.
2388                  */
2389                 failed = 0;
2390         }
2391         return failed;
2392 }
2393
2394 /* Undo during loss recovery after partial ACK. */
2395 static int tcp_try_undo_loss(struct sock *sk)
2396 {
2397         struct tcp_sock *tp = tcp_sk(sk);
2398
2399         if (tcp_may_undo(tp)) {
2400                 struct sk_buff *skb;
2401                 tcp_for_write_queue(skb, sk) {
2402                         if (skb == tcp_send_head(sk))
2403                                 break;
2404                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2405                 }
2406
2407                 tcp_clear_all_retrans_hints(tp);
2408
2409                 DBGUNDO(sk, "partial loss");
2410                 tp->lost_out = 0;
2411                 tcp_undo_cwr(sk, 1);
2412                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2413                 inet_csk(sk)->icsk_retransmits = 0;
2414                 tp->undo_marker = 0;
2415                 if (tcp_is_sack(tp))
2416                         tcp_set_ca_state(sk, TCP_CA_Open);
2417                 return 1;
2418         }
2419         return 0;
2420 }
2421
2422 static inline void tcp_complete_cwr(struct sock *sk)
2423 {
2424         struct tcp_sock *tp = tcp_sk(sk);
2425         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2426         tp->snd_cwnd_stamp = tcp_time_stamp;
2427         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2428 }
2429
2430 static void tcp_try_to_open(struct sock *sk, int flag)
2431 {
2432         struct tcp_sock *tp = tcp_sk(sk);
2433
2434         tcp_verify_left_out(tp);
2435
2436         if (tp->retrans_out == 0)
2437                 tp->retrans_stamp = 0;
2438
2439         if (flag&FLAG_ECE)
2440                 tcp_enter_cwr(sk, 1);
2441
2442         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2443                 int state = TCP_CA_Open;
2444
2445                 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2446                         state = TCP_CA_Disorder;
2447
2448                 if (inet_csk(sk)->icsk_ca_state != state) {
2449                         tcp_set_ca_state(sk, state);
2450                         tp->high_seq = tp->snd_nxt;
2451                 }
2452                 tcp_moderate_cwnd(tp);
2453         } else {
2454                 tcp_cwnd_down(sk, flag);
2455         }
2456 }
2457
2458 static void tcp_mtup_probe_failed(struct sock *sk)
2459 {
2460         struct inet_connection_sock *icsk = inet_csk(sk);
2461
2462         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2463         icsk->icsk_mtup.probe_size = 0;
2464 }
2465
2466 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2467 {
2468         struct tcp_sock *tp = tcp_sk(sk);
2469         struct inet_connection_sock *icsk = inet_csk(sk);
2470
2471         /* FIXME: breaks with very large cwnd */
2472         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2473         tp->snd_cwnd = tp->snd_cwnd *
2474                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2475                        icsk->icsk_mtup.probe_size;
2476         tp->snd_cwnd_cnt = 0;
2477         tp->snd_cwnd_stamp = tcp_time_stamp;
2478         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2479
2480         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2481         icsk->icsk_mtup.probe_size = 0;
2482         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2483 }
2484
2485
2486 /* Process an event, which can update packets-in-flight not trivially.
2487  * Main goal of this function is to calculate new estimate for left_out,
2488  * taking into account both packets sitting in receiver's buffer and
2489  * packets lost by network.
2490  *
2491  * Besides that it does CWND reduction, when packet loss is detected
2492  * and changes state of machine.
2493  *
2494  * It does _not_ decide what to send, it is made in function
2495  * tcp_xmit_retransmit_queue().
2496  */
2497 static void
2498 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2499 {
2500         struct inet_connection_sock *icsk = inet_csk(sk);
2501         struct tcp_sock *tp = tcp_sk(sk);
2502         int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2503         int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2504                                     (tcp_fackets_out(tp) > tp->reordering));
2505         int fast_rexmit = 0;
2506
2507         /* Some technical things:
2508          * 1. Reno does not count dupacks (sacked_out) automatically. */
2509         if (!tp->packets_out)
2510                 tp->sacked_out = 0;
2511
2512         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2513                 tp->fackets_out = 0;
2514
2515         /* Now state machine starts.
2516          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2517         if (flag&FLAG_ECE)
2518                 tp->prior_ssthresh = 0;
2519
2520         /* B. In all the states check for reneging SACKs. */
2521         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2522                 return;
2523
2524         /* C. Process data loss notification, provided it is valid. */
2525         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2526             before(tp->snd_una, tp->high_seq) &&
2527             icsk->icsk_ca_state != TCP_CA_Open &&
2528             tp->fackets_out > tp->reordering) {
2529                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, 0);
2530                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2531         }
2532
2533         /* D. Check consistency of the current state. */
2534         tcp_verify_left_out(tp);
2535
2536         /* E. Check state exit conditions. State can be terminated
2537          *    when high_seq is ACKed. */
2538         if (icsk->icsk_ca_state == TCP_CA_Open) {
2539                 BUG_TRAP(tp->retrans_out == 0);
2540                 tp->retrans_stamp = 0;
2541         } else if (!before(tp->snd_una, tp->high_seq)) {
2542                 switch (icsk->icsk_ca_state) {
2543                 case TCP_CA_Loss:
2544                         icsk->icsk_retransmits = 0;
2545                         if (tcp_try_undo_recovery(sk))
2546                                 return;
2547                         break;
2548
2549                 case TCP_CA_CWR:
2550                         /* CWR is to be held something *above* high_seq
2551                          * is ACKed for CWR bit to reach receiver. */
2552                         if (tp->snd_una != tp->high_seq) {
2553                                 tcp_complete_cwr(sk);
2554                                 tcp_set_ca_state(sk, TCP_CA_Open);
2555                         }
2556                         break;
2557
2558                 case TCP_CA_Disorder:
2559                         tcp_try_undo_dsack(sk);
2560                         if (!tp->undo_marker ||
2561                             /* For SACK case do not Open to allow to undo
2562                              * catching for all duplicate ACKs. */
2563                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2564                                 tp->undo_marker = 0;
2565                                 tcp_set_ca_state(sk, TCP_CA_Open);
2566                         }
2567                         break;
2568
2569                 case TCP_CA_Recovery:
2570                         if (tcp_is_reno(tp))
2571                                 tcp_reset_reno_sack(tp);
2572                         if (tcp_try_undo_recovery(sk))
2573                                 return;
2574                         tcp_complete_cwr(sk);
2575                         break;
2576                 }
2577         }
2578
2579         /* F. Process state. */
2580         switch (icsk->icsk_ca_state) {
2581         case TCP_CA_Recovery:
2582                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2583                         if (tcp_is_reno(tp) && is_dupack)
2584                                 tcp_add_reno_sack(sk);
2585                 } else
2586                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2587                 break;
2588         case TCP_CA_Loss:
2589                 if (flag&FLAG_DATA_ACKED)
2590                         icsk->icsk_retransmits = 0;
2591                 if (!tcp_try_undo_loss(sk)) {
2592                         tcp_moderate_cwnd(tp);
2593                         tcp_xmit_retransmit_queue(sk);
2594                         return;
2595                 }
2596                 if (icsk->icsk_ca_state != TCP_CA_Open)
2597                         return;
2598                 /* Loss is undone; fall through to processing in Open state. */
2599         default:
2600                 if (tcp_is_reno(tp)) {
2601                         if (flag & FLAG_SND_UNA_ADVANCED)
2602                                 tcp_reset_reno_sack(tp);
2603                         if (is_dupack)
2604                                 tcp_add_reno_sack(sk);
2605                 }
2606
2607                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2608                         tcp_try_undo_dsack(sk);
2609
2610                 if (!tcp_time_to_recover(sk)) {
2611                         tcp_try_to_open(sk, flag);
2612                         return;
2613                 }
2614
2615                 /* MTU probe failure: don't reduce cwnd */
2616                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2617                     icsk->icsk_mtup.probe_size &&
2618                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2619                         tcp_mtup_probe_failed(sk);
2620                         /* Restores the reduction we did in tcp_mtup_probe() */
2621                         tp->snd_cwnd++;
2622                         tcp_simple_retransmit(sk);
2623                         return;
2624                 }
2625
2626                 /* Otherwise enter Recovery state */
2627
2628                 if (tcp_is_reno(tp))
2629                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2630                 else
2631                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2632
2633                 tp->high_seq = tp->snd_nxt;
2634                 tp->prior_ssthresh = 0;
2635                 tp->undo_marker = tp->snd_una;
2636                 tp->undo_retrans = tp->retrans_out;
2637
2638                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2639                         if (!(flag&FLAG_ECE))
2640                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2641                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2642                         TCP_ECN_queue_cwr(tp);
2643                 }
2644
2645                 tp->bytes_acked = 0;
2646                 tp->snd_cwnd_cnt = 0;
2647                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2648                 fast_rexmit = 1;
2649         }
2650
2651         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2652                 tcp_update_scoreboard(sk, fast_rexmit);
2653         tcp_cwnd_down(sk, flag);
2654         tcp_xmit_retransmit_queue(sk);
2655 }
2656
2657 /* Read draft-ietf-tcplw-high-performance before mucking
2658  * with this code. (Supersedes RFC1323)
2659  */
2660 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2661 {
2662         /* RTTM Rule: A TSecr value received in a segment is used to
2663          * update the averaged RTT measurement only if the segment
2664          * acknowledges some new data, i.e., only if it advances the
2665          * left edge of the send window.
2666          *
2667          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2668          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2669          *
2670          * Changed: reset backoff as soon as we see the first valid sample.
2671          * If we do not, we get strongly overestimated rto. With timestamps
2672          * samples are accepted even from very old segments: f.e., when rtt=1
2673          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2674          * answer arrives rto becomes 120 seconds! If at least one of segments
2675          * in window is lost... Voila.                          --ANK (010210)
2676          */
2677         struct tcp_sock *tp = tcp_sk(sk);
2678         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2679         tcp_rtt_estimator(sk, seq_rtt);
2680         tcp_set_rto(sk);
2681         inet_csk(sk)->icsk_backoff = 0;
2682         tcp_bound_rto(sk);
2683 }
2684
2685 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2686 {
2687         /* We don't have a timestamp. Can only use
2688          * packets that are not retransmitted to determine
2689          * rtt estimates. Also, we must not reset the
2690          * backoff for rto until we get a non-retransmitted
2691          * packet. This allows us to deal with a situation
2692          * where the network delay has increased suddenly.
2693          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2694          */
2695
2696         if (flag & FLAG_RETRANS_DATA_ACKED)
2697                 return;
2698
2699         tcp_rtt_estimator(sk, seq_rtt);
2700         tcp_set_rto(sk);
2701         inet_csk(sk)->icsk_backoff = 0;
2702         tcp_bound_rto(sk);
2703 }
2704
2705 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2706                                       const s32 seq_rtt)
2707 {
2708         const struct tcp_sock *tp = tcp_sk(sk);
2709         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2710         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2711                 tcp_ack_saw_tstamp(sk, flag);
2712         else if (seq_rtt >= 0)
2713                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2714 }
2715
2716 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2717 {
2718         const struct inet_connection_sock *icsk = inet_csk(sk);
2719         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2720         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2721 }
2722
2723 /* Restart timer after forward progress on connection.
2724  * RFC2988 recommends to restart timer to now+rto.
2725  */
2726 static void tcp_rearm_rto(struct sock *sk)
2727 {
2728         struct tcp_sock *tp = tcp_sk(sk);
2729
2730         if (!tp->packets_out) {
2731                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2732         } else {
2733                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2734         }
2735 }
2736
2737 /* If we get here, the whole TSO packet has not been acked. */
2738 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2739 {
2740         struct tcp_sock *tp = tcp_sk(sk);
2741         u32 packets_acked;
2742
2743         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2744
2745         packets_acked = tcp_skb_pcount(skb);
2746         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2747                 return 0;
2748         packets_acked -= tcp_skb_pcount(skb);
2749
2750         if (packets_acked) {
2751                 BUG_ON(tcp_skb_pcount(skb) == 0);
2752                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2753         }
2754
2755         return packets_acked;
2756 }
2757
2758 /* Remove acknowledged frames from the retransmission queue. If our packet
2759  * is before the ack sequence we can discard it as it's confirmed to have
2760  * arrived at the other end.
2761  */
2762 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
2763 {
2764         struct tcp_sock *tp = tcp_sk(sk);
2765         const struct inet_connection_sock *icsk = inet_csk(sk);
2766         struct sk_buff *skb;
2767         u32 now = tcp_time_stamp;
2768         int fully_acked = 1;
2769         int flag = 0;
2770         u32 pkts_acked = 0;
2771         u32 reord = tp->packets_out;
2772         s32 seq_rtt = -1;
2773         s32 ca_seq_rtt = -1;
2774         ktime_t last_ackt = net_invalid_timestamp();
2775
2776         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2777                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2778                 u32 end_seq;
2779                 u32 acked_pcount;
2780                 u8 sacked = scb->sacked;
2781
2782                 /* Determine how many packets and what bytes were acked, tso and else */
2783                 if (after(scb->end_seq, tp->snd_una)) {
2784                         if (tcp_skb_pcount(skb) == 1 ||
2785                             !after(tp->snd_una, scb->seq))
2786                                 break;
2787
2788                         acked_pcount = tcp_tso_acked(sk, skb);
2789                         if (!acked_pcount)
2790                                 break;
2791
2792                         fully_acked = 0;
2793                         end_seq = tp->snd_una;
2794                 } else {
2795                         acked_pcount = tcp_skb_pcount(skb);
2796                         end_seq = scb->end_seq;
2797                 }
2798
2799                 /* MTU probing checks */
2800                 if (fully_acked && icsk->icsk_mtup.probe_size &&
2801                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2802                         tcp_mtup_probe_success(sk, skb);
2803                 }
2804
2805                 if (sacked & TCPCB_RETRANS) {
2806                         if (sacked & TCPCB_SACKED_RETRANS)
2807                                 tp->retrans_out -= acked_pcount;
2808                         flag |= FLAG_RETRANS_DATA_ACKED;
2809                         ca_seq_rtt = -1;
2810                         seq_rtt = -1;
2811                         if ((flag & FLAG_DATA_ACKED) ||
2812                             (acked_pcount > 1))
2813                                 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2814                 } else {
2815                         ca_seq_rtt = now - scb->when;
2816                         last_ackt = skb->tstamp;
2817                         if (seq_rtt < 0) {
2818                                 seq_rtt = ca_seq_rtt;
2819                         }
2820                         if (!(sacked & TCPCB_SACKED_ACKED))
2821                                 reord = min(pkts_acked, reord);
2822                 }
2823
2824                 if (sacked & TCPCB_SACKED_ACKED)
2825                         tp->sacked_out -= acked_pcount;
2826                 if (sacked & TCPCB_LOST)
2827                         tp->lost_out -= acked_pcount;
2828
2829                 if (unlikely((sacked & TCPCB_URG) && tp->urg_mode &&
2830                              !before(end_seq, tp->snd_up)))
2831                         tp->urg_mode = 0;
2832
2833                 tp->packets_out -= acked_pcount;
2834                 pkts_acked += acked_pcount;
2835
2836                 /* Initial outgoing SYN's get put onto the write_queue
2837                  * just like anything else we transmit.  It is not
2838                  * true data, and if we misinform our callers that
2839                  * this ACK acks real data, we will erroneously exit
2840                  * connection startup slow start one packet too
2841                  * quickly.  This is severely frowned upon behavior.
2842                  */
2843                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2844                         flag |= FLAG_DATA_ACKED;
2845                 } else {
2846                         flag |= FLAG_SYN_ACKED;
2847                         tp->retrans_stamp = 0;
2848                 }
2849
2850                 if (!fully_acked)
2851                         break;
2852
2853                 tcp_unlink_write_queue(skb, sk);
2854                 sk_wmem_free_skb(sk, skb);
2855                 tcp_clear_all_retrans_hints(tp);
2856         }
2857
2858         if (flag & FLAG_ACKED) {
2859                 const struct tcp_congestion_ops *ca_ops
2860                         = inet_csk(sk)->icsk_ca_ops;
2861
2862                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2863                 tcp_rearm_rto(sk);
2864
2865                 if (tcp_is_reno(tp)) {
2866                         tcp_remove_reno_sacks(sk, pkts_acked);
2867                 } else {
2868                         /* Non-retransmitted hole got filled? That's reordering */
2869                         if (reord < prior_fackets)
2870                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2871                 }
2872
2873                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2874
2875                 if (ca_ops->pkts_acked) {
2876                         s32 rtt_us = -1;
2877
2878                         /* Is the ACK triggering packet unambiguous? */
2879                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2880                                 /* High resolution needed and available? */
2881                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2882                                     !ktime_equal(last_ackt,
2883                                                  net_invalid_timestamp()))
2884                                         rtt_us = ktime_us_delta(ktime_get_real(),
2885                                                                 last_ackt);
2886                                 else if (ca_seq_rtt > 0)
2887                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
2888                         }
2889
2890                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2891                 }
2892         }
2893
2894 #if FASTRETRANS_DEBUG > 0
2895         BUG_TRAP((int)tp->sacked_out >= 0);
2896         BUG_TRAP((int)tp->lost_out >= 0);
2897         BUG_TRAP((int)tp->retrans_out >= 0);
2898         if (!tp->packets_out && tcp_is_sack(tp)) {
2899                 icsk = inet_csk(sk);
2900                 if (tp->lost_out) {
2901                         printk(KERN_DEBUG "Leak l=%u %d\n",
2902                                tp->lost_out, icsk->icsk_ca_state);
2903                         tp->lost_out = 0;
2904                 }
2905                 if (tp->sacked_out) {
2906                         printk(KERN_DEBUG "Leak s=%u %d\n",
2907                                tp->sacked_out, icsk->icsk_ca_state);
2908                         tp->sacked_out = 0;
2909                 }
2910                 if (tp->retrans_out) {
2911                         printk(KERN_DEBUG "Leak r=%u %d\n",
2912                                tp->retrans_out, icsk->icsk_ca_state);
2913                         tp->retrans_out = 0;
2914                 }
2915         }
2916 #endif
2917         return flag;
2918 }
2919
2920 static void tcp_ack_probe(struct sock *sk)
2921 {
2922         const struct tcp_sock *tp = tcp_sk(sk);
2923         struct inet_connection_sock *icsk = inet_csk(sk);
2924
2925         /* Was it a usable window open? */
2926
2927         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2928                    tp->snd_una + tp->snd_wnd)) {
2929                 icsk->icsk_backoff = 0;
2930                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2931                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2932                  * This function is not for random using!
2933                  */
2934         } else {
2935                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2936                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2937                                           TCP_RTO_MAX);
2938         }
2939 }
2940
2941 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2942 {
2943         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2944                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2945 }
2946
2947 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2948 {
2949         const struct tcp_sock *tp = tcp_sk(sk);
2950         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2951                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2952 }
2953
2954 /* Check that window update is acceptable.
2955  * The function assumes that snd_una<=ack<=snd_next.
2956  */
2957 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2958                                         const u32 ack_seq, const u32 nwin)
2959 {
2960         return (after(ack, tp->snd_una) ||
2961                 after(ack_seq, tp->snd_wl1) ||
2962                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2963 }
2964
2965 /* Update our send window.
2966  *
2967  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2968  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2969  */
2970 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2971                                  u32 ack_seq)
2972 {
2973         struct tcp_sock *tp = tcp_sk(sk);
2974         int flag = 0;
2975         u32 nwin = ntohs(tcp_hdr(skb)->window);
2976
2977         if (likely(!tcp_hdr(skb)->syn))
2978                 nwin <<= tp->rx_opt.snd_wscale;
2979
2980         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2981                 flag |= FLAG_WIN_UPDATE;
2982                 tcp_update_wl(tp, ack, ack_seq);
2983
2984                 if (tp->snd_wnd != nwin) {
2985                         tp->snd_wnd = nwin;
2986
2987                         /* Note, it is the only place, where
2988                          * fast path is recovered for sending TCP.
2989                          */
2990                         tp->pred_flags = 0;
2991                         tcp_fast_path_check(sk);
2992
2993                         if (nwin > tp->max_window) {
2994                                 tp->max_window = nwin;
2995                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2996                         }
2997                 }
2998         }
2999
3000         tp->snd_una = ack;
3001
3002         return flag;
3003 }
3004
3005 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3006  * continue in congestion avoidance.
3007  */
3008 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3009 {
3010         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3011         tp->snd_cwnd_cnt = 0;
3012         tp->bytes_acked = 0;
3013         TCP_ECN_queue_cwr(tp);
3014         tcp_moderate_cwnd(tp);
3015 }
3016
3017 /* A conservative spurious RTO response algorithm: reduce cwnd using
3018  * rate halving and continue in congestion avoidance.
3019  */
3020 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3021 {
3022         tcp_enter_cwr(sk, 0);
3023 }
3024
3025 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3026 {
3027         if (flag&FLAG_ECE)
3028                 tcp_ratehalving_spur_to_response(sk);
3029         else
3030                 tcp_undo_cwr(sk, 1);
3031 }
3032
3033 /* F-RTO spurious RTO detection algorithm (RFC4138)
3034  *
3035  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3036  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3037  * window (but not to or beyond highest sequence sent before RTO):
3038  *   On First ACK,  send two new segments out.
3039  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3040  *                  algorithm is not part of the F-RTO detection algorithm
3041  *                  given in RFC4138 but can be selected separately).
3042  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3043  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3044  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3045  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3046  *
3047  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3048  * original window even after we transmit two new data segments.
3049  *
3050  * SACK version:
3051  *   on first step, wait until first cumulative ACK arrives, then move to
3052  *   the second step. In second step, the next ACK decides.
3053  *
3054  * F-RTO is implemented (mainly) in four functions:
3055  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3056  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3057  *     called when tcp_use_frto() showed green light
3058  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3059  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3060  *     to prove that the RTO is indeed spurious. It transfers the control
3061  *     from F-RTO to the conventional RTO recovery
3062  */
3063 static int tcp_process_frto(struct sock *sk, int flag)
3064 {
3065         struct tcp_sock *tp = tcp_sk(sk);
3066
3067         tcp_verify_left_out(tp);
3068
3069         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3070         if (flag&FLAG_DATA_ACKED)
3071                 inet_csk(sk)->icsk_retransmits = 0;
3072
3073         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3074             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3075                 tp->undo_marker = 0;
3076
3077         if (!before(tp->snd_una, tp->frto_highmark)) {
3078                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3079                 return 1;
3080         }
3081
3082         if (!IsSackFrto() || tcp_is_reno(tp)) {
3083                 /* RFC4138 shortcoming in step 2; should also have case c):
3084                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3085                  * data, winupdate
3086                  */
3087                 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
3088                         return 1;
3089
3090                 if (!(flag&FLAG_DATA_ACKED)) {
3091                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3092                                             flag);
3093                         return 1;
3094                 }
3095         } else {
3096                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3097                         /* Prevent sending of new data. */
3098                         tp->snd_cwnd = min(tp->snd_cwnd,
3099                                            tcp_packets_in_flight(tp));
3100                         return 1;
3101                 }
3102
3103                 if ((tp->frto_counter >= 2) &&
3104                     (!(flag&FLAG_FORWARD_PROGRESS) ||
3105                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
3106                         /* RFC4138 shortcoming (see comment above) */
3107                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
3108                                 return 1;
3109
3110                         tcp_enter_frto_loss(sk, 3, flag);
3111                         return 1;
3112                 }
3113         }
3114
3115         if (tp->frto_counter == 1) {
3116                 /* tcp_may_send_now needs to see updated state */
3117                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3118                 tp->frto_counter = 2;
3119
3120                 if (!tcp_may_send_now(sk))
3121                         tcp_enter_frto_loss(sk, 2, flag);
3122
3123                 return 1;
3124         } else {
3125                 switch (sysctl_tcp_frto_response) {
3126                 case 2:
3127                         tcp_undo_spur_to_response(sk, flag);
3128                         break;
3129                 case 1:
3130                         tcp_conservative_spur_to_response(tp);
3131                         break;
3132                 default:
3133                         tcp_ratehalving_spur_to_response(sk);
3134                         break;
3135                 }
3136                 tp->frto_counter = 0;
3137                 tp->undo_marker = 0;
3138                 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
3139         }
3140         return 0;
3141 }
3142
3143 /* This routine deals with incoming acks, but not outgoing ones. */
3144 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3145 {
3146         struct inet_connection_sock *icsk = inet_csk(sk);
3147         struct tcp_sock *tp = tcp_sk(sk);
3148         u32 prior_snd_una = tp->snd_una;
3149         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3150         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3151         u32 prior_in_flight;
3152         u32 prior_fackets;
3153         int prior_packets;
3154         int frto_cwnd = 0;
3155
3156         /* If the ack is newer than sent or older than previous acks
3157          * then we can probably ignore it.
3158          */
3159         if (after(ack, tp->snd_nxt))
3160                 goto uninteresting_ack;
3161
3162         if (before(ack, prior_snd_una))
3163                 goto old_ack;
3164
3165         if (after(ack, prior_snd_una))
3166                 flag |= FLAG_SND_UNA_ADVANCED;
3167
3168         if (sysctl_tcp_abc) {
3169                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3170                         tp->bytes_acked += ack - prior_snd_una;
3171                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3172                         /* we assume just one segment left network */
3173                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
3174         }
3175
3176         prior_fackets = tp->fackets_out;
3177         prior_in_flight = tcp_packets_in_flight(tp);
3178
3179         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3180                 /* Window is constant, pure forward advance.
3181                  * No more checks are required.
3182                  * Note, we use the fact that SND.UNA>=SND.WL2.
3183                  */
3184                 tcp_update_wl(tp, ack, ack_seq);
3185                 tp->snd_una = ack;
3186                 flag |= FLAG_WIN_UPDATE;
3187
3188                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3189
3190                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3191         } else {
3192                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3193                         flag |= FLAG_DATA;
3194                 else
3195                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3196
3197                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3198
3199                 if (TCP_SKB_CB(skb)->sacked)
3200                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3201
3202                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3203                         flag |= FLAG_ECE;
3204
3205                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3206         }
3207
3208         /* We passed data and got it acked, remove any soft error
3209          * log. Something worked...
3210          */
3211         sk->sk_err_soft = 0;
3212         tp->rcv_tstamp = tcp_time_stamp;
3213         prior_packets = tp->packets_out;
3214         if (!prior_packets)
3215                 goto no_queue;
3216
3217         /* See if we can take anything off of the retransmit queue. */
3218         flag |= tcp_clean_rtx_queue(sk, prior_fackets);
3219
3220         if (tp->frto_counter)
3221                 frto_cwnd = tcp_process_frto(sk, flag);
3222         /* Guarantee sacktag reordering detection against wrap-arounds */
3223         if (before(tp->frto_highmark, tp->snd_una))
3224                 tp->frto_highmark = 0;
3225
3226         if (tcp_ack_is_dubious(sk, flag)) {
3227                 /* Advance CWND, if state allows this. */
3228                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3229                     tcp_may_raise_cwnd(sk, flag))
3230                         tcp_cong_avoid(sk, ack, prior_in_flight);
3231                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3232         } else {
3233                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3234                         tcp_cong_avoid(sk, ack, prior_in_flight);
3235         }
3236
3237         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3238                 dst_confirm(sk->sk_dst_cache);
3239
3240         return 1;
3241
3242 no_queue:
3243         icsk->icsk_probes_out = 0;
3244
3245         /* If this ack opens up a zero window, clear backoff.  It was
3246          * being used to time the probes, and is probably far higher than
3247          * it needs to be for normal retransmission.
3248          */
3249         if (tcp_send_head(sk))
3250                 tcp_ack_probe(sk);
3251         return 1;
3252
3253 old_ack:
3254         if (TCP_SKB_CB(skb)->sacked)
3255                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3256
3257 uninteresting_ack:
3258         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3259         return 0;
3260 }
3261
3262
3263 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3264  * But, this can also be called on packets in the established flow when
3265  * the fast version below fails.
3266  */
3267 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3268 {
3269         unsigned char *ptr;
3270         struct tcphdr *th = tcp_hdr(skb);
3271         int length=(th->doff*4)-sizeof(struct tcphdr);
3272
3273         ptr = (unsigned char *)(th + 1);
3274         opt_rx->saw_tstamp = 0;
3275
3276         while (length > 0) {
3277                 int opcode=*ptr++;
3278                 int opsize;
3279
3280                 switch (opcode) {
3281                         case TCPOPT_EOL:
3282                                 return;
3283                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3284                                 length--;
3285                                 continue;
3286                         default:
3287                                 opsize=*ptr++;
3288                                 if (opsize < 2) /* "silly options" */
3289                                         return;
3290                                 if (opsize > length)
3291                                         return; /* don't parse partial options */
3292                                 switch (opcode) {
3293                                 case TCPOPT_MSS:
3294                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3295                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3296                                                 if (in_mss) {
3297                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3298                                                                 in_mss = opt_rx->user_mss;
3299                                                         opt_rx->mss_clamp = in_mss;
3300                                                 }
3301                                         }
3302                                         break;
3303                                 case TCPOPT_WINDOW:
3304                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3305                                                 if (sysctl_tcp_window_scaling) {
3306                                                         __u8 snd_wscale = *(__u8 *) ptr;
3307                                                         opt_rx->wscale_ok = 1;
3308                                                         if (snd_wscale > 14) {
3309                                                                 if (net_ratelimit())
3310                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3311                                                                                "scaling value %d >14 received.\n",
3312                                                                                snd_wscale);
3313                                                                 snd_wscale = 14;
3314                                                         }
3315                                                         opt_rx->snd_wscale = snd_wscale;
3316                                                 }
3317                                         break;
3318                                 case TCPOPT_TIMESTAMP:
3319                                         if (opsize==TCPOLEN_TIMESTAMP) {
3320                                                 if ((estab && opt_rx->tstamp_ok) ||
3321                                                     (!estab && sysctl_tcp_timestamps)) {
3322                                                         opt_rx->saw_tstamp = 1;
3323                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3324                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3325                                                 }
3326                                         }
3327                                         break;
3328                                 case TCPOPT_SACK_PERM:
3329                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3330                                                 if (sysctl_tcp_sack) {
3331                                                         opt_rx->sack_ok = 1;
3332                                                         tcp_sack_reset(opt_rx);
3333                                                 }
3334                                         }
3335                                         break;
3336
3337                                 case TCPOPT_SACK:
3338                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3339                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3340                                            opt_rx->sack_ok) {
3341                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3342                                         }
3343                                         break;
3344 #ifdef CONFIG_TCP_MD5SIG
3345                                 case TCPOPT_MD5SIG:
3346                                         /*
3347                                          * The MD5 Hash has already been
3348                                          * checked (see tcp_v{4,6}_do_rcv()).
3349                                          */
3350                                         break;
3351 #endif
3352                                 }
3353
3354                                 ptr+=opsize-2;
3355                                 length-=opsize;
3356                 }
3357         }
3358 }
3359
3360 /* Fast parse options. This hopes to only see timestamps.
3361  * If it is wrong it falls back on tcp_parse_options().
3362  */
3363 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3364                                   struct tcp_sock *tp)
3365 {
3366         if (th->doff == sizeof(struct tcphdr)>>2) {
3367                 tp->rx_opt.saw_tstamp = 0;
3368                 return 0;
3369         } else if (tp->rx_opt.tstamp_ok &&
3370                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3371                 __be32 *ptr = (__be32 *)(th + 1);
3372                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3373                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3374                         tp->rx_opt.saw_tstamp = 1;
3375                         ++ptr;
3376                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3377                         ++ptr;
3378                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3379                         return 1;
3380                 }
3381         }
3382         tcp_parse_options(skb, &tp->rx_opt, 1);
3383         return 1;
3384 }
3385
3386 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3387 {
3388         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3389         tp->rx_opt.ts_recent_stamp = get_seconds();
3390 }
3391
3392 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3393 {
3394         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3395                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3396                  * extra check below makes sure this can only happen
3397                  * for pure ACK frames.  -DaveM
3398                  *
3399                  * Not only, also it occurs for expired timestamps.
3400                  */
3401
3402                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3403                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3404                         tcp_store_ts_recent(tp);
3405         }
3406 }
3407
3408 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3409  *
3410  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3411  * it can pass through stack. So, the following predicate verifies that
3412  * this segment is not used for anything but congestion avoidance or
3413  * fast retransmit. Moreover, we even are able to eliminate most of such
3414  * second order effects, if we apply some small "replay" window (~RTO)
3415  * to timestamp space.
3416  *
3417  * All these measures still do not guarantee that we reject wrapped ACKs
3418  * on networks with high bandwidth, when sequence space is recycled fastly,
3419  * but it guarantees that such events will be very rare and do not affect
3420  * connection seriously. This doesn't look nice, but alas, PAWS is really
3421  * buggy extension.
3422  *
3423  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3424  * states that events when retransmit arrives after original data are rare.
3425  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3426  * the biggest problem on large power networks even with minor reordering.
3427  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3428  * up to bandwidth of 18Gigabit/sec. 8) ]
3429  */
3430
3431 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3432 {
3433         struct tcp_sock *tp = tcp_sk(sk);
3434         struct tcphdr *th = tcp_hdr(skb);
3435         u32 seq = TCP_SKB_CB(skb)->seq;
3436         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3437
3438         return (/* 1. Pure ACK with correct sequence number. */
3439                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3440
3441                 /* 2. ... and duplicate ACK. */
3442                 ack == tp->snd_una &&
3443
3444                 /* 3. ... and does not update window. */
3445                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3446
3447                 /* 4. ... and sits in replay window. */
3448                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3449 }
3450
3451 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3452 {
3453         const struct tcp_sock *tp = tcp_sk(sk);
3454         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3455                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3456                 !tcp_disordered_ack(sk, skb));
3457 }
3458
3459 /* Check segment sequence number for validity.
3460  *
3461  * Segment controls are considered valid, if the segment
3462  * fits to the window after truncation to the window. Acceptability
3463  * of data (and SYN, FIN, of course) is checked separately.
3464  * See tcp_data_queue(), for example.
3465  *
3466  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3467  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3468  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3469  * (borrowed from freebsd)
3470  */
3471
3472 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3473 {
3474         return  !before(end_seq, tp->rcv_wup) &&
3475                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3476 }
3477
3478 /* When we get a reset we do this. */
3479 static void tcp_reset(struct sock *sk)
3480 {
3481         /* We want the right error as BSD sees it (and indeed as we do). */
3482         switch (sk->sk_state) {
3483                 case TCP_SYN_SENT:
3484                         sk->sk_err = ECONNREFUSED;
3485                         break;
3486                 case TCP_CLOSE_WAIT:
3487                         sk->sk_err = EPIPE;
3488                         break;
3489                 case TCP_CLOSE:
3490                         return;
3491                 default:
3492                         sk->sk_err = ECONNRESET;
3493         }
3494
3495         if (!sock_flag(sk, SOCK_DEAD))
3496                 sk->sk_error_report(sk);
3497
3498         tcp_done(sk);
3499 }
3500
3501 /*
3502  *      Process the FIN bit. This now behaves as it is supposed to work
3503  *      and the FIN takes effect when it is validly part of sequence
3504  *      space. Not before when we get holes.
3505  *
3506  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3507  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3508  *      TIME-WAIT)
3509  *
3510  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3511  *      close and we go into CLOSING (and later onto TIME-WAIT)
3512  *
3513  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3514  */
3515 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3516 {
3517         struct tcp_sock *tp = tcp_sk(sk);
3518
3519         inet_csk_schedule_ack(sk);
3520
3521         sk->sk_shutdown |= RCV_SHUTDOWN;
3522         sock_set_flag(sk, SOCK_DONE);
3523
3524         switch (sk->sk_state) {
3525                 case TCP_SYN_RECV:
3526                 case TCP_ESTABLISHED:
3527                         /* Move to CLOSE_WAIT */
3528                         tcp_set_state(sk, TCP_CLOSE_WAIT);
3529                         inet_csk(sk)->icsk_ack.pingpong = 1;
3530                         break;
3531
3532                 case TCP_CLOSE_WAIT:
3533                 case TCP_CLOSING:
3534                         /* Received a retransmission of the FIN, do
3535                          * nothing.
3536                          */
3537                         break;
3538                 case TCP_LAST_ACK:
3539                         /* RFC793: Remain in the LAST-ACK state. */
3540                         break;
3541
3542                 case TCP_FIN_WAIT1:
3543                         /* This case occurs when a simultaneous close
3544                          * happens, we must ack the received FIN and
3545                          * enter the CLOSING state.
3546                          */
3547                         tcp_send_ack(sk);
3548                         tcp_set_state(sk, TCP_CLOSING);
3549                         break;
3550                 case TCP_FIN_WAIT2:
3551                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
3552                         tcp_send_ack(sk);
3553                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3554                         break;
3555                 default:
3556                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3557                          * cases we should never reach this piece of code.
3558                          */
3559                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3560                                __FUNCTION__, sk->sk_state);
3561                         break;
3562         }
3563
3564         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3565          * Probably, we should reset in this case. For now drop them.
3566          */
3567         __skb_queue_purge(&tp->out_of_order_queue);
3568         if (tcp_is_sack(tp))
3569                 tcp_sack_reset(&tp->rx_opt);
3570         sk_mem_reclaim(sk);
3571
3572         if (!sock_flag(sk, SOCK_DEAD)) {
3573                 sk->sk_state_change(sk);
3574
3575                 /* Do not send POLL_HUP for half duplex close. */
3576                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3577                     sk->sk_state == TCP_CLOSE)
3578                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3579                 else
3580                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3581         }
3582 }
3583
3584 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3585 {
3586         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3587                 if (before(seq, sp->start_seq))
3588                         sp->start_seq = seq;
3589                 if (after(end_seq, sp->end_seq))
3590                         sp->end_seq = end_seq;
3591                 return 1;
3592         }
3593         return 0;
3594 }
3595
3596 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3597 {
3598         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3599                 if (before(seq, tp->rcv_nxt))
3600                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3601                 else
3602                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3603
3604                 tp->rx_opt.dsack = 1;
3605                 tp->duplicate_sack[0].start_seq = seq;
3606                 tp->duplicate_sack[0].end_seq = end_seq;
3607                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3608         }
3609 }
3610
3611 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3612 {
3613         if (!tp->rx_opt.dsack)
3614                 tcp_dsack_set(tp, seq, end_seq);
3615         else
3616                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3617 }
3618
3619 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3620 {
3621         struct tcp_sock *tp = tcp_sk(sk);
3622
3623         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3624             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3625                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3626                 tcp_enter_quickack_mode(sk);
3627
3628                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3629                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3630
3631                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3632                                 end_seq = tp->rcv_nxt;
3633                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3634                 }
3635         }
3636
3637         tcp_send_ack(sk);
3638 }
3639
3640 /* These routines update the SACK block as out-of-order packets arrive or
3641  * in-order packets close up the sequence space.
3642  */
3643 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3644 {
3645         int this_sack;
3646         struct tcp_sack_block *sp = &tp->selective_acks[0];
3647         struct tcp_sack_block *swalk = sp+1;
3648
3649         /* See if the recent change to the first SACK eats into
3650          * or hits the sequence space of other SACK blocks, if so coalesce.
3651          */
3652         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3653                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3654                         int i;
3655
3656                         /* Zap SWALK, by moving every further SACK up by one slot.
3657                          * Decrease num_sacks.
3658                          */
3659                         tp->rx_opt.num_sacks--;
3660                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3661                         for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3662                                 sp[i] = sp[i+1];
3663                         continue;
3664                 }
3665                 this_sack++, swalk++;
3666         }
3667 }
3668
3669 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3670 {
3671         __u32 tmp;
3672
3673         tmp = sack1->start_seq;
3674         sack1->start_seq = sack2->start_seq;
3675         sack2->start_seq = tmp;
3676
3677         tmp = sack1->end_seq;
3678         sack1->end_seq = sack2->end_seq;
3679         sack2->end_seq = tmp;
3680 }
3681
3682 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3683 {
3684         struct tcp_sock *tp = tcp_sk(sk);
3685         struct tcp_sack_block *sp = &tp->selective_acks[0];
3686         int cur_sacks = tp->rx_opt.num_sacks;
3687         int this_sack;
3688
3689         if (!cur_sacks)
3690                 goto new_sack;
3691
3692         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3693                 if (tcp_sack_extend(sp, seq, end_seq)) {
3694                         /* Rotate this_sack to the first one. */
3695                         for (; this_sack>0; this_sack--, sp--)
3696                                 tcp_sack_swap(sp, sp-1);
3697                         if (cur_sacks > 1)
3698                                 tcp_sack_maybe_coalesce(tp);
3699                         return;
3700                 }
3701         }
3702
3703         /* Could not find an adjacent existing SACK, build a new one,
3704          * put it at the front, and shift everyone else down.  We
3705          * always know there is at least one SACK present already here.
3706          *
3707          * If the sack array is full, forget about the last one.
3708          */
3709         if (this_sack >= 4) {
3710                 this_sack--;
3711                 tp->rx_opt.num_sacks--;
3712                 sp--;
3713         }
3714         for (; this_sack > 0; this_sack--, sp--)
3715                 *sp = *(sp-1);
3716
3717 new_sack:
3718         /* Build the new head SACK, and we're done. */
3719         sp->start_seq = seq;
3720         sp->end_seq = end_seq;
3721         tp->rx_opt.num_sacks++;
3722         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3723 }
3724
3725 /* RCV.NXT advances, some SACKs should be eaten. */
3726
3727 static void tcp_sack_remove(struct tcp_sock *tp)
3728 {
3729         struct tcp_sack_block *sp = &tp->selective_acks[0];
3730         int num_sacks = tp->rx_opt.num_sacks;
3731         int this_sack;
3732
3733         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3734         if (skb_queue_empty(&tp->out_of_order_queue)) {
3735                 tp->rx_opt.num_sacks = 0;
3736                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3737                 return;
3738         }
3739
3740         for (this_sack = 0; this_sack < num_sacks; ) {
3741                 /* Check if the start of the sack is covered by RCV.NXT. */
3742                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3743                         int i;
3744
3745                         /* RCV.NXT must cover all the block! */
3746                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3747
3748                         /* Zap this SACK, by moving forward any other SACKS. */
3749                         for (i=this_sack+1; i < num_sacks; i++)
3750                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3751                         num_sacks--;
3752                         continue;
3753                 }
3754                 this_sack++;
3755                 sp++;
3756         }
3757         if (num_sacks != tp->rx_opt.num_sacks) {
3758                 tp->rx_opt.num_sacks = num_sacks;
3759                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3760         }
3761 }
3762
3763 /* This one checks to see if we can put data from the
3764  * out_of_order queue into the receive_queue.
3765  */
3766 static void tcp_ofo_queue(struct sock *sk)
3767 {
3768         struct tcp_sock *tp = tcp_sk(sk);
3769         __u32 dsack_high = tp->rcv_nxt;
3770         struct sk_buff *skb;
3771
3772         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3773                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3774                         break;
3775
3776                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3777                         __u32 dsack = dsack_high;
3778                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3779                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3780                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3781                 }
3782
3783                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3784                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3785                         __skb_unlink(skb, &tp->out_of_order_queue);
3786                         __kfree_skb(skb);
3787                         continue;
3788                 }
3789                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3790                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3791                            TCP_SKB_CB(skb)->end_seq);
3792
3793                 __skb_unlink(skb, &tp->out_of_order_queue);
3794                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3795                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3796                 if (tcp_hdr(skb)->fin)
3797                         tcp_fin(skb, sk, tcp_hdr(skb));
3798         }
3799 }
3800
3801 static int tcp_prune_queue(struct sock *sk);
3802
3803 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3804 {
3805         struct tcphdr *th = tcp_hdr(skb);
3806         struct tcp_sock *tp = tcp_sk(sk);
3807         int eaten = -1;
3808
3809         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3810                 goto drop;
3811
3812         __skb_pull(skb, th->doff*4);
3813
3814         TCP_ECN_accept_cwr(tp, skb);
3815
3816         if (tp->rx_opt.dsack) {
3817                 tp->rx_opt.dsack = 0;
3818                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3819                                                     4 - tp->rx_opt.tstamp_ok);
3820         }
3821
3822         /*  Queue data for delivery to the user.
3823          *  Packets in sequence go to the receive queue.
3824          *  Out of sequence packets to the out_of_order_queue.
3825          */
3826         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3827                 if (tcp_receive_window(tp) == 0)
3828                         goto out_of_window;
3829
3830                 /* Ok. In sequence. In window. */
3831                 if (tp->ucopy.task == current &&
3832                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3833                     sock_owned_by_user(sk) && !tp->urg_data) {
3834                         int chunk = min_t(unsigned int, skb->len,
3835                                                         tp->ucopy.len);
3836
3837                         __set_current_state(TASK_RUNNING);
3838
3839                         local_bh_enable();
3840                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3841                                 tp->ucopy.len -= chunk;
3842                                 tp->copied_seq += chunk;
3843                                 eaten = (chunk == skb->len && !th->fin);
3844                                 tcp_rcv_space_adjust(sk);
3845                         }
3846                         local_bh_disable();
3847                 }
3848
3849                 if (eaten <= 0) {
3850 queue_and_out:
3851                         if (eaten < 0 &&
3852                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3853                              !sk_rmem_schedule(sk, skb->truesize))) {
3854                                 if (tcp_prune_queue(sk) < 0 ||
3855                                     !sk_rmem_schedule(sk, skb->truesize))
3856                                         goto drop;
3857                         }
3858                         skb_set_owner_r(skb, sk);
3859                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3860                 }
3861                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3862                 if (skb->len)
3863                         tcp_event_data_recv(sk, skb);
3864                 if (th->fin)
3865                         tcp_fin(skb, sk, th);
3866
3867                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3868                         tcp_ofo_queue(sk);
3869
3870                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3871                          * gap in queue is filled.
3872                          */
3873                         if (skb_queue_empty(&tp->out_of_order_queue))
3874                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3875                 }
3876
3877                 if (tp->rx_opt.num_sacks)
3878                         tcp_sack_remove(tp);
3879
3880                 tcp_fast_path_check(sk);
3881
3882                 if (eaten > 0)
3883                         __kfree_skb(skb);
3884                 else if (!sock_flag(sk, SOCK_DEAD))
3885                         sk->sk_data_ready(sk, 0);
3886                 return;
3887         }
3888
3889         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3890                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3891                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3892                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3893
3894 out_of_window:
3895                 tcp_enter_quickack_mode(sk);
3896                 inet_csk_schedule_ack(sk);
3897 drop:
3898                 __kfree_skb(skb);
3899                 return;
3900         }
3901
3902         /* Out of window. F.e. zero window probe. */
3903         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3904                 goto out_of_window;
3905
3906         tcp_enter_quickack_mode(sk);
3907
3908         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3909                 /* Partial packet, seq < rcv_next < end_seq */
3910                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3911                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3912                            TCP_SKB_CB(skb)->end_seq);
3913
3914                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3915
3916                 /* If window is closed, drop tail of packet. But after
3917                  * remembering D-SACK for its head made in previous line.
3918                  */
3919                 if (!tcp_receive_window(tp))
3920                         goto out_of_window;
3921                 goto queue_and_out;
3922         }
3923
3924         TCP_ECN_check_ce(tp, skb);
3925
3926         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3927             !sk_rmem_schedule(sk, skb->truesize)) {
3928                 if (tcp_prune_queue(sk) < 0 ||
3929                     !sk_rmem_schedule(sk, skb->truesize))
3930                         goto drop;
3931         }
3932
3933         /* Disable header prediction. */
3934         tp->pred_flags = 0;
3935         inet_csk_schedule_ack(sk);
3936
3937         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3938                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3939
3940         skb_set_owner_r(skb, sk);
3941
3942         if (!skb_peek(&tp->out_of_order_queue)) {
3943                 /* Initial out of order segment, build 1 SACK. */
3944                 if (tcp_is_sack(tp)) {
3945                         tp->rx_opt.num_sacks = 1;
3946                         tp->rx_opt.dsack     = 0;
3947                         tp->rx_opt.eff_sacks = 1;
3948                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3949                         tp->selective_acks[0].end_seq =
3950                                                 TCP_SKB_CB(skb)->end_seq;
3951                 }
3952                 __skb_queue_head(&tp->out_of_order_queue,skb);
3953         } else {
3954                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3955                 u32 seq = TCP_SKB_CB(skb)->seq;
3956                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3957
3958                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3959                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3960
3961                         if (!tp->rx_opt.num_sacks ||
3962                             tp->selective_acks[0].end_seq != seq)
3963                                 goto add_sack;
3964
3965                         /* Common case: data arrive in order after hole. */
3966                         tp->selective_acks[0].end_seq = end_seq;
3967                         return;
3968                 }
3969
3970                 /* Find place to insert this segment. */
3971                 do {
3972                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3973                                 break;
3974                 } while ((skb1 = skb1->prev) !=
3975                          (struct sk_buff*)&tp->out_of_order_queue);
3976
3977                 /* Do skb overlap to previous one? */
3978                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3979                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3980                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3981                                 /* All the bits are present. Drop. */
3982                                 __kfree_skb(skb);
3983                                 tcp_dsack_set(tp, seq, end_seq);
3984                                 goto add_sack;
3985                         }
3986                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3987                                 /* Partial overlap. */
3988                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3989                         } else {
3990                                 skb1 = skb1->prev;
3991                         }
3992                 }
3993                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3994
3995                 /* And clean segments covered by new one as whole. */
3996                 while ((skb1 = skb->next) !=
3997                        (struct sk_buff*)&tp->out_of_order_queue &&
3998                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3999                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4000                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
4001                                break;
4002                        }
4003                        __skb_unlink(skb1, &tp->out_of_order_queue);
4004                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
4005                        __kfree_skb(skb1);
4006                 }
4007
4008 add_sack:
4009                 if (tcp_is_sack(tp))
4010                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4011         }
4012 }
4013
4014 /* Collapse contiguous sequence of skbs head..tail with
4015  * sequence numbers start..end.
4016  * Segments with FIN/SYN are not collapsed (only because this
4017  * simplifies code)
4018  */
4019 static void
4020 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4021              struct sk_buff *head, struct sk_buff *tail,
4022              u32 start, u32 end)
4023 {
4024         struct sk_buff *skb;
4025
4026         /* First, check that queue is collapsible and find
4027          * the point where collapsing can be useful. */
4028         for (skb = head; skb != tail; ) {
4029                 /* No new bits? It is possible on ofo queue. */
4030                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4031                         struct sk_buff *next = skb->next;
4032                         __skb_unlink(skb, list);
4033                         __kfree_skb(skb);
4034                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4035                         skb = next;
4036                         continue;
4037                 }
4038
4039                 /* The first skb to collapse is:
4040                  * - not SYN/FIN and
4041                  * - bloated or contains data before "start" or
4042                  *   overlaps to the next one.
4043                  */
4044                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4045                     (tcp_win_from_space(skb->truesize) > skb->len ||
4046                      before(TCP_SKB_CB(skb)->seq, start) ||
4047                      (skb->next != tail &&
4048                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4049                         break;
4050
4051                 /* Decided to skip this, advance start seq. */
4052                 start = TCP_SKB_CB(skb)->end_seq;
4053                 skb = skb->next;
4054         }
4055         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4056                 return;
4057
4058         while (before(start, end)) {
4059                 struct sk_buff *nskb;
4060                 unsigned int header = skb_headroom(skb);
4061                 int copy = SKB_MAX_ORDER(header, 0);
4062
4063                 /* Too big header? This can happen with IPv6. */
4064                 if (copy < 0)
4065                         return;
4066                 if (end-start < copy)
4067                         copy = end-start;
4068                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
4069                 if (!nskb)
4070                         return;
4071
4072                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4073                 skb_set_network_header(nskb, (skb_network_header(skb) -
4074                                               skb->head));
4075                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4076                                                 skb->head));
4077                 skb_reserve(nskb, header);
4078                 memcpy(nskb->head, skb->head, header);
4079                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4080                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4081                 __skb_insert(nskb, skb->prev, skb, list);
4082                 skb_set_owner_r(nskb, sk);
4083
4084                 /* Copy data, releasing collapsed skbs. */
4085                 while (copy > 0) {
4086                         int offset = start - TCP_SKB_CB(skb)->seq;
4087                         int size = TCP_SKB_CB(skb)->end_seq - start;
4088
4089                         BUG_ON(offset < 0);
4090                         if (size > 0) {
4091                                 size = min(copy, size);
4092                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4093                                         BUG();
4094                                 TCP_SKB_CB(nskb)->end_seq += size;
4095                                 copy -= size;
4096                                 start += size;
4097                         }
4098                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4099                                 struct sk_buff *next = skb->next;
4100                                 __skb_unlink(skb, list);
4101                                 __kfree_skb(skb);
4102                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4103                                 skb = next;
4104                                 if (skb == tail ||
4105                                     tcp_hdr(skb)->syn ||
4106                                     tcp_hdr(skb)->fin)
4107                                         return;
4108                         }
4109                 }
4110         }
4111 }
4112
4113 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4114  * and tcp_collapse() them until all the queue is collapsed.
4115  */
4116 static void tcp_collapse_ofo_queue(struct sock *sk)
4117 {
4118         struct tcp_sock *tp = tcp_sk(sk);
4119         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4120         struct sk_buff *head;
4121         u32 start, end;
4122
4123         if (skb == NULL)
4124                 return;
4125
4126         start = TCP_SKB_CB(skb)->seq;
4127         end = TCP_SKB_CB(skb)->end_seq;
4128         head = skb;
4129
4130         for (;;) {
4131                 skb = skb->next;
4132
4133                 /* Segment is terminated when we see gap or when
4134                  * we are at the end of all the queue. */
4135                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4136                     after(TCP_SKB_CB(skb)->seq, end) ||
4137                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4138                         tcp_collapse(sk, &tp->out_of_order_queue,
4139                                      head, skb, start, end);
4140                         head = skb;
4141                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4142                                 break;
4143                         /* Start new segment */
4144                         start = TCP_SKB_CB(skb)->seq;
4145                         end = TCP_SKB_CB(skb)->end_seq;
4146                 } else {
4147                         if (before(TCP_SKB_CB(skb)->seq, start))
4148                                 start = TCP_SKB_CB(skb)->seq;
4149                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4150                                 end = TCP_SKB_CB(skb)->end_seq;
4151                 }
4152         }
4153 }
4154
4155 /* Reduce allocated memory if we can, trying to get
4156  * the socket within its memory limits again.
4157  *
4158  * Return less than zero if we should start dropping frames
4159  * until the socket owning process reads some of the data
4160  * to stabilize the situation.
4161  */
4162 static int tcp_prune_queue(struct sock *sk)
4163 {
4164         struct tcp_sock *tp = tcp_sk(sk);
4165
4166         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4167
4168         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
4169
4170         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4171                 tcp_clamp_window(sk);
4172         else if (tcp_memory_pressure)
4173                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4174
4175         tcp_collapse_ofo_queue(sk);
4176         tcp_collapse(sk, &sk->sk_receive_queue,
4177                      sk->sk_receive_queue.next,
4178                      (struct sk_buff*)&sk->sk_receive_queue,
4179                      tp->copied_seq, tp->rcv_nxt);
4180         sk_mem_reclaim(sk);
4181
4182         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4183                 return 0;
4184
4185         /* Collapsing did not help, destructive actions follow.
4186          * This must not ever occur. */
4187
4188         /* First, purge the out_of_order queue. */
4189         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4190                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4191                 __skb_queue_purge(&tp->out_of_order_queue);
4192
4193                 /* Reset SACK state.  A conforming SACK implementation will
4194                  * do the same at a timeout based retransmit.  When a connection
4195                  * is in a sad state like this, we care only about integrity
4196                  * of the connection not performance.
4197                  */
4198                 if (tcp_is_sack(tp))
4199                         tcp_sack_reset(&tp->rx_opt);
4200                 sk_mem_reclaim(sk);
4201         }
4202
4203         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4204                 return 0;
4205
4206         /* If we are really being abused, tell the caller to silently
4207          * drop receive data on the floor.  It will get retransmitted
4208          * and hopefully then we'll have sufficient space.
4209          */
4210         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4211
4212         /* Massive buffer overcommit. */
4213         tp->pred_flags = 0;
4214         return -1;
4215 }
4216
4217
4218 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4219  * As additional protections, we do not touch cwnd in retransmission phases,
4220  * and if application hit its sndbuf limit recently.
4221  */
4222 void tcp_cwnd_application_limited(struct sock *sk)
4223 {
4224         struct tcp_sock *tp = tcp_sk(sk);
4225
4226         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4227             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4228                 /* Limited by application or receiver window. */
4229                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4230                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4231                 if (win_used < tp->snd_cwnd) {
4232                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4233                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4234                 }
4235                 tp->snd_cwnd_used = 0;
4236         }
4237         tp->snd_cwnd_stamp = tcp_time_stamp;
4238 }
4239
4240 static int tcp_should_expand_sndbuf(struct sock *sk)
4241 {
4242         struct tcp_sock *tp = tcp_sk(sk);
4243
4244         /* If the user specified a specific send buffer setting, do
4245          * not modify it.
4246          */
4247         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4248                 return 0;
4249
4250         /* If we are under global TCP memory pressure, do not expand.  */
4251         if (tcp_memory_pressure)
4252                 return 0;
4253
4254         /* If we are under soft global TCP memory pressure, do not expand.  */
4255         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4256                 return 0;
4257
4258         /* If we filled the congestion window, do not expand.  */
4259         if (tp->packets_out >= tp->snd_cwnd)
4260                 return 0;
4261
4262         return 1;
4263 }
4264
4265 /* When incoming ACK allowed to free some skb from write_queue,
4266  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4267  * on the exit from tcp input handler.
4268  *
4269  * PROBLEM: sndbuf expansion does not work well with largesend.
4270  */
4271 static void tcp_new_space(struct sock *sk)
4272 {
4273         struct tcp_sock *tp = tcp_sk(sk);
4274
4275         if (tcp_should_expand_sndbuf(sk)) {
4276                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4277                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4278                     demanded = max_t(unsigned int, tp->snd_cwnd,
4279                                                    tp->reordering + 1);
4280                 sndmem *= 2*demanded;
4281                 if (sndmem > sk->sk_sndbuf)
4282                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4283                 tp->snd_cwnd_stamp = tcp_time_stamp;
4284         }
4285
4286         sk->sk_write_space(sk);
4287 }
4288
4289 static void tcp_check_space(struct sock *sk)
4290 {
4291         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4292                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4293                 if (sk->sk_socket &&
4294                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4295                         tcp_new_space(sk);
4296         }
4297 }
4298
4299 static inline void tcp_data_snd_check(struct sock *sk)
4300 {
4301         tcp_push_pending_frames(sk);
4302         tcp_check_space(sk);
4303 }
4304
4305 /*
4306  * Check if sending an ack is needed.
4307  */
4308 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4309 {
4310         struct tcp_sock *tp = tcp_sk(sk);
4311
4312             /* More than one full frame received... */
4313         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4314              /* ... and right edge of window advances far enough.
4315               * (tcp_recvmsg() will send ACK otherwise). Or...
4316               */
4317              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4318             /* We ACK each frame or... */
4319             tcp_in_quickack_mode(sk) ||
4320             /* We have out of order data. */
4321             (ofo_possible &&
4322              skb_peek(&tp->out_of_order_queue))) {
4323                 /* Then ack it now */
4324                 tcp_send_ack(sk);
4325         } else {
4326                 /* Else, send delayed ack. */
4327                 tcp_send_delayed_ack(sk);
4328         }
4329 }
4330
4331 static inline void tcp_ack_snd_check(struct sock *sk)
4332 {
4333         if (!inet_csk_ack_scheduled(sk)) {
4334                 /* We sent a data segment already. */
4335                 return;
4336         }
4337         __tcp_ack_snd_check(sk, 1);
4338 }
4339
4340 /*
4341  *      This routine is only called when we have urgent data
4342  *      signaled. Its the 'slow' part of tcp_urg. It could be
4343  *      moved inline now as tcp_urg is only called from one
4344  *      place. We handle URGent data wrong. We have to - as
4345  *      BSD still doesn't use the correction from RFC961.
4346  *      For 1003.1g we should support a new option TCP_STDURG to permit
4347  *      either form (or just set the sysctl tcp_stdurg).
4348  */
4349
4350 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4351 {
4352         struct tcp_sock *tp = tcp_sk(sk);
4353         u32 ptr = ntohs(th->urg_ptr);
4354
4355         if (ptr && !sysctl_tcp_stdurg)
4356                 ptr--;
4357         ptr += ntohl(th->seq);
4358
4359         /* Ignore urgent data that we've already seen and read. */
4360         if (after(tp->copied_seq, ptr))
4361                 return;
4362
4363         /* Do not replay urg ptr.
4364          *
4365          * NOTE: interesting situation not covered by specs.
4366          * Misbehaving sender may send urg ptr, pointing to segment,
4367          * which we already have in ofo queue. We are not able to fetch
4368          * such data and will stay in TCP_URG_NOTYET until will be eaten
4369          * by recvmsg(). Seems, we are not obliged to handle such wicked
4370          * situations. But it is worth to think about possibility of some
4371          * DoSes using some hypothetical application level deadlock.
4372          */
4373         if (before(ptr, tp->rcv_nxt))
4374                 return;
4375
4376         /* Do we already have a newer (or duplicate) urgent pointer? */
4377         if (tp->urg_data && !after(ptr, tp->urg_seq))
4378                 return;
4379
4380         /* Tell the world about our new urgent pointer. */
4381         sk_send_sigurg(sk);
4382
4383         /* We may be adding urgent data when the last byte read was
4384          * urgent. To do this requires some care. We cannot just ignore
4385          * tp->copied_seq since we would read the last urgent byte again
4386          * as data, nor can we alter copied_seq until this data arrives
4387          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4388          *
4389          * NOTE. Double Dutch. Rendering to plain English: author of comment
4390          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4391          * and expect that both A and B disappear from stream. This is _wrong_.
4392          * Though this happens in BSD with high probability, this is occasional.
4393          * Any application relying on this is buggy. Note also, that fix "works"
4394          * only in this artificial test. Insert some normal data between A and B and we will
4395          * decline of BSD again. Verdict: it is better to remove to trap
4396          * buggy users.
4397          */
4398         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4399             !sock_flag(sk, SOCK_URGINLINE) &&
4400             tp->copied_seq != tp->rcv_nxt) {
4401                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4402                 tp->copied_seq++;
4403                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4404                         __skb_unlink(skb, &sk->sk_receive_queue);
4405                         __kfree_skb(skb);
4406                 }
4407         }
4408
4409         tp->urg_data   = TCP_URG_NOTYET;
4410         tp->urg_seq    = ptr;
4411
4412         /* Disable header prediction. */
4413         tp->pred_flags = 0;
4414 }
4415
4416 /* This is the 'fast' part of urgent handling. */
4417 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4418 {
4419         struct tcp_sock *tp = tcp_sk(sk);
4420
4421         /* Check if we get a new urgent pointer - normally not. */
4422         if (th->urg)
4423                 tcp_check_urg(sk,th);
4424
4425         /* Do we wait for any urgent data? - normally not... */
4426         if (tp->urg_data == TCP_URG_NOTYET) {
4427                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4428                           th->syn;
4429
4430                 /* Is the urgent pointer pointing into this packet? */
4431                 if (ptr < skb->len) {
4432                         u8 tmp;
4433                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4434                                 BUG();
4435                         tp->urg_data = TCP_URG_VALID | tmp;
4436                         if (!sock_flag(sk, SOCK_DEAD))
4437                                 sk->sk_data_ready(sk, 0);
4438                 }
4439         }
4440 }
4441
4442 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4443 {
4444         struct tcp_sock *tp = tcp_sk(sk);
4445         int chunk = skb->len - hlen;
4446         int err;
4447
4448         local_bh_enable();
4449         if (skb_csum_unnecessary(skb))
4450                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4451         else
4452                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4453                                                        tp->ucopy.iov);
4454
4455         if (!err) {
4456                 tp->ucopy.len -= chunk;
4457                 tp->copied_seq += chunk;
4458                 tcp_rcv_space_adjust(sk);
4459         }
4460
4461         local_bh_disable();
4462         return err;
4463 }
4464
4465 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4466 {
4467         __sum16 result;
4468
4469         if (sock_owned_by_user(sk)) {
4470                 local_bh_enable();
4471                 result = __tcp_checksum_complete(skb);
4472                 local_bh_disable();
4473         } else {
4474                 result = __tcp_checksum_complete(skb);
4475         }
4476         return result;
4477 }
4478
4479 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4480 {
4481         return !skb_csum_unnecessary(skb) &&
4482                 __tcp_checksum_complete_user(sk, skb);
4483 }
4484
4485 #ifdef CONFIG_NET_DMA
4486 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4487 {
4488         struct tcp_sock *tp = tcp_sk(sk);
4489         int chunk = skb->len - hlen;
4490         int dma_cookie;
4491         int copied_early = 0;
4492
4493         if (tp->ucopy.wakeup)
4494                 return 0;
4495
4496         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4497                 tp->ucopy.dma_chan = get_softnet_dma();
4498
4499         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4500
4501                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4502                         skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4503
4504                 if (dma_cookie < 0)
4505                         goto out;
4506
4507                 tp->ucopy.dma_cookie = dma_cookie;
4508                 copied_early = 1;
4509
4510                 tp->ucopy.len -= chunk;
4511                 tp->copied_seq += chunk;
4512                 tcp_rcv_space_adjust(sk);
4513
4514                 if ((tp->ucopy.len == 0) ||
4515                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4516                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4517                         tp->ucopy.wakeup = 1;
4518                         sk->sk_data_ready(sk, 0);
4519                 }
4520         } else if (chunk > 0) {
4521                 tp->ucopy.wakeup = 1;
4522                 sk->sk_data_ready(sk, 0);
4523         }
4524 out:
4525         return copied_early;
4526 }
4527 #endif /* CONFIG_NET_DMA */
4528
4529 /*
4530  *      TCP receive function for the ESTABLISHED state.
4531  *
4532  *      It is split into a fast path and a slow path. The fast path is
4533  *      disabled when:
4534  *      - A zero window was announced from us - zero window probing
4535  *        is only handled properly in the slow path.
4536  *      - Out of order segments arrived.
4537  *      - Urgent data is expected.
4538  *      - There is no buffer space left
4539  *      - Unexpected TCP flags/window values/header lengths are received
4540  *        (detected by checking the TCP header against pred_flags)
4541  *      - Data is sent in both directions. Fast path only supports pure senders
4542  *        or pure receivers (this means either the sequence number or the ack
4543  *        value must stay constant)
4544  *      - Unexpected TCP option.
4545  *
4546  *      When these conditions are not satisfied it drops into a standard
4547  *      receive procedure patterned after RFC793 to handle all cases.
4548  *      The first three cases are guaranteed by proper pred_flags setting,
4549  *      the rest is checked inline. Fast processing is turned on in
4550  *      tcp_data_queue when everything is OK.
4551  */
4552 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4553                         struct tcphdr *th, unsigned len)
4554 {
4555         struct tcp_sock *tp = tcp_sk(sk);
4556
4557         /*
4558          *      Header prediction.
4559          *      The code loosely follows the one in the famous
4560          *      "30 instruction TCP receive" Van Jacobson mail.
4561          *
4562          *      Van's trick is to deposit buffers into socket queue
4563          *      on a device interrupt, to call tcp_recv function
4564          *      on the receive process context and checksum and copy
4565          *      the buffer to user space. smart...
4566          *
4567          *      Our current scheme is not silly either but we take the
4568          *      extra cost of the net_bh soft interrupt processing...
4569          *      We do checksum and copy also but from device to kernel.
4570          */
4571
4572         tp->rx_opt.saw_tstamp = 0;
4573
4574         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4575          *      if header_prediction is to be made
4576          *      'S' will always be tp->tcp_header_len >> 2
4577          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4578          *  turn it off (when there are holes in the receive
4579          *       space for instance)
4580          *      PSH flag is ignored.
4581          */
4582
4583         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4584                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4585                 int tcp_header_len = tp->tcp_header_len;
4586
4587                 /* Timestamp header prediction: tcp_header_len
4588                  * is automatically equal to th->doff*4 due to pred_flags
4589                  * match.
4590                  */
4591
4592                 /* Check timestamp */
4593                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4594                         __be32 *ptr = (__be32 *)(th + 1);
4595
4596                         /* No? Slow path! */
4597                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4598                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4599                                 goto slow_path;
4600
4601                         tp->rx_opt.saw_tstamp = 1;
4602                         ++ptr;
4603                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4604                         ++ptr;
4605                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4606
4607                         /* If PAWS failed, check it more carefully in slow path */
4608                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4609                                 goto slow_path;
4610
4611                         /* DO NOT update ts_recent here, if checksum fails
4612                          * and timestamp was corrupted part, it will result
4613                          * in a hung connection since we will drop all
4614                          * future packets due to the PAWS test.
4615                          */
4616                 }
4617
4618                 if (len <= tcp_header_len) {
4619                         /* Bulk data transfer: sender */
4620                         if (len == tcp_header_len) {
4621                                 /* Predicted packet is in window by definition.
4622                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4623                                  * Hence, check seq<=rcv_wup reduces to:
4624                                  */
4625                                 if (tcp_header_len ==
4626                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4627                                     tp->rcv_nxt == tp->rcv_wup)
4628                                         tcp_store_ts_recent(tp);
4629
4630                                 /* We know that such packets are checksummed
4631                                  * on entry.
4632                                  */
4633                                 tcp_ack(sk, skb, 0);
4634                                 __kfree_skb(skb);
4635                                 tcp_data_snd_check(sk);
4636                                 return 0;
4637                         } else { /* Header too small */
4638                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4639                                 goto discard;
4640                         }
4641                 } else {
4642                         int eaten = 0;
4643                         int copied_early = 0;
4644
4645                         if (tp->copied_seq == tp->rcv_nxt &&
4646                             len - tcp_header_len <= tp->ucopy.len) {
4647 #ifdef CONFIG_NET_DMA
4648                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4649                                         copied_early = 1;
4650                                         eaten = 1;
4651                                 }
4652 #endif
4653                                 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4654                                         __set_current_state(TASK_RUNNING);
4655
4656                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4657                                                 eaten = 1;
4658                                 }
4659                                 if (eaten) {
4660                                         /* Predicted packet is in window by definition.
4661                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4662                                          * Hence, check seq<=rcv_wup reduces to:
4663                                          */
4664                                         if (tcp_header_len ==
4665                                             (sizeof(struct tcphdr) +
4666                                              TCPOLEN_TSTAMP_ALIGNED) &&
4667                                             tp->rcv_nxt == tp->rcv_wup)
4668                                                 tcp_store_ts_recent(tp);
4669
4670                                         tcp_rcv_rtt_measure_ts(sk, skb);
4671
4672                                         __skb_pull(skb, tcp_header_len);
4673                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4674                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4675                                 }
4676                                 if (copied_early)
4677                                         tcp_cleanup_rbuf(sk, skb->len);
4678                         }
4679                         if (!eaten) {
4680                                 if (tcp_checksum_complete_user(sk, skb))
4681                                         goto csum_error;
4682
4683                                 /* Predicted packet is in window by definition.
4684                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4685                                  * Hence, check seq<=rcv_wup reduces to:
4686                                  */
4687                                 if (tcp_header_len ==
4688                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4689                                     tp->rcv_nxt == tp->rcv_wup)
4690                                         tcp_store_ts_recent(tp);
4691
4692                                 tcp_rcv_rtt_measure_ts(sk, skb);
4693
4694                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4695                                         goto step5;
4696
4697                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4698
4699                                 /* Bulk data transfer: receiver */
4700                                 __skb_pull(skb,tcp_header_len);
4701                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4702                                 skb_set_owner_r(skb, sk);
4703                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4704                         }
4705
4706                         tcp_event_data_recv(sk, skb);
4707
4708                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4709                                 /* Well, only one small jumplet in fast path... */
4710                                 tcp_ack(sk, skb, FLAG_DATA);
4711                                 tcp_data_snd_check(sk);
4712                                 if (!inet_csk_ack_scheduled(sk))
4713                                         goto no_ack;
4714                         }
4715
4716                         __tcp_ack_snd_check(sk, 0);
4717 no_ack:
4718 #ifdef CONFIG_NET_DMA
4719                         if (copied_early)
4720                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4721                         else
4722 #endif
4723                         if (eaten)
4724                                 __kfree_skb(skb);
4725                         else
4726                                 sk->sk_data_ready(sk, 0);
4727                         return 0;
4728                 }
4729         }
4730
4731 slow_path:
4732         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4733                 goto csum_error;
4734
4735         /*
4736          * RFC1323: H1. Apply PAWS check first.
4737          */
4738         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4739             tcp_paws_discard(sk, skb)) {
4740                 if (!th->rst) {
4741                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4742                         tcp_send_dupack(sk, skb);
4743                         goto discard;
4744                 }
4745                 /* Resets are accepted even if PAWS failed.
4746
4747                    ts_recent update must be made after we are sure
4748                    that the packet is in window.
4749                  */
4750         }
4751
4752         /*
4753          *      Standard slow path.
4754          */
4755
4756         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4757                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4758                  * (RST) segments are validated by checking their SEQ-fields."
4759                  * And page 69: "If an incoming segment is not acceptable,
4760                  * an acknowledgment should be sent in reply (unless the RST bit
4761                  * is set, if so drop the segment and return)".
4762                  */
4763                 if (!th->rst)
4764                         tcp_send_dupack(sk, skb);
4765                 goto discard;
4766         }
4767
4768         if (th->rst) {
4769                 tcp_reset(sk);
4770                 goto discard;
4771         }
4772
4773         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4774
4775         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4776                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4777                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4778                 tcp_reset(sk);
4779                 return 1;
4780         }
4781
4782 step5:
4783         if (th->ack)
4784                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4785
4786         tcp_rcv_rtt_measure_ts(sk, skb);
4787
4788         /* Process urgent data. */
4789         tcp_urg(sk, skb, th);
4790
4791         /* step 7: process the segment text */
4792         tcp_data_queue(sk, skb);
4793
4794         tcp_data_snd_check(sk);
4795         tcp_ack_snd_check(sk);
4796         return 0;
4797
4798 csum_error:
4799         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4800
4801 discard:
4802         __kfree_skb(skb);
4803         return 0;
4804 }
4805
4806 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4807                                          struct tcphdr *th, unsigned len)
4808 {
4809         struct tcp_sock *tp = tcp_sk(sk);
4810         struct inet_connection_sock *icsk = inet_csk(sk);
4811         int saved_clamp = tp->rx_opt.mss_clamp;
4812
4813         tcp_parse_options(skb, &tp->rx_opt, 0);
4814
4815         if (th->ack) {
4816                 /* rfc793:
4817                  * "If the state is SYN-SENT then
4818                  *    first check the ACK bit
4819                  *      If the ACK bit is set
4820                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4821                  *        a reset (unless the RST bit is set, if so drop
4822                  *        the segment and return)"
4823                  *
4824                  *  We do not send data with SYN, so that RFC-correct
4825                  *  test reduces to:
4826                  */
4827                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4828                         goto reset_and_undo;
4829
4830                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4831                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4832                              tcp_time_stamp)) {
4833                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4834                         goto reset_and_undo;
4835                 }
4836
4837                 /* Now ACK is acceptable.
4838                  *
4839                  * "If the RST bit is set
4840                  *    If the ACK was acceptable then signal the user "error:
4841                  *    connection reset", drop the segment, enter CLOSED state,
4842                  *    delete TCB, and return."
4843                  */
4844
4845                 if (th->rst) {
4846                         tcp_reset(sk);
4847                         goto discard;
4848                 }
4849
4850                 /* rfc793:
4851                  *   "fifth, if neither of the SYN or RST bits is set then
4852                  *    drop the segment and return."
4853                  *
4854                  *    See note below!
4855                  *                                        --ANK(990513)
4856                  */
4857                 if (!th->syn)
4858                         goto discard_and_undo;
4859
4860                 /* rfc793:
4861                  *   "If the SYN bit is on ...
4862                  *    are acceptable then ...
4863                  *    (our SYN has been ACKed), change the connection
4864                  *    state to ESTABLISHED..."
4865                  */
4866
4867                 TCP_ECN_rcv_synack(tp, th);
4868
4869                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4870                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4871
4872                 /* Ok.. it's good. Set up sequence numbers and
4873                  * move to established.
4874                  */
4875                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4876                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4877
4878                 /* RFC1323: The window in SYN & SYN/ACK segments is
4879                  * never scaled.
4880                  */
4881                 tp->snd_wnd = ntohs(th->window);
4882                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4883
4884                 if (!tp->rx_opt.wscale_ok) {
4885                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4886                         tp->window_clamp = min(tp->window_clamp, 65535U);
4887                 }
4888
4889                 if (tp->rx_opt.saw_tstamp) {
4890                         tp->rx_opt.tstamp_ok       = 1;
4891                         tp->tcp_header_len =
4892                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4893                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4894                         tcp_store_ts_recent(tp);
4895                 } else {
4896                         tp->tcp_header_len = sizeof(struct tcphdr);
4897                 }
4898
4899                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4900                         tcp_enable_fack(tp);
4901
4902                 tcp_mtup_init(sk);
4903                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4904                 tcp_initialize_rcv_mss(sk);
4905
4906                 /* Remember, tcp_poll() does not lock socket!
4907                  * Change state from SYN-SENT only after copied_seq
4908                  * is initialized. */
4909                 tp->copied_seq = tp->rcv_nxt;
4910                 smp_mb();
4911                 tcp_set_state(sk, TCP_ESTABLISHED);
4912
4913                 security_inet_conn_established(sk, skb);
4914
4915                 /* Make sure socket is routed, for correct metrics.  */
4916                 icsk->icsk_af_ops->rebuild_header(sk);
4917
4918                 tcp_init_metrics(sk);
4919
4920                 tcp_init_congestion_control(sk);
4921
4922                 /* Prevent spurious tcp_cwnd_restart() on first data
4923                  * packet.
4924                  */
4925                 tp->lsndtime = tcp_time_stamp;
4926
4927                 tcp_init_buffer_space(sk);
4928
4929                 if (sock_flag(sk, SOCK_KEEPOPEN))
4930                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4931
4932                 if (!tp->rx_opt.snd_wscale)
4933                         __tcp_fast_path_on(tp, tp->snd_wnd);
4934                 else
4935                         tp->pred_flags = 0;
4936
4937                 if (!sock_flag(sk, SOCK_DEAD)) {
4938                         sk->sk_state_change(sk);
4939                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
4940                 }
4941
4942                 if (sk->sk_write_pending ||
4943                     icsk->icsk_accept_queue.rskq_defer_accept ||
4944                     icsk->icsk_ack.pingpong) {
4945                         /* Save one ACK. Data will be ready after
4946                          * several ticks, if write_pending is set.
4947                          *
4948                          * It may be deleted, but with this feature tcpdumps
4949                          * look so _wonderfully_ clever, that I was not able
4950                          * to stand against the temptation 8)     --ANK
4951                          */
4952                         inet_csk_schedule_ack(sk);
4953                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4954                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4955                         tcp_incr_quickack(sk);
4956                         tcp_enter_quickack_mode(sk);
4957                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4958                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4959
4960 discard:
4961                         __kfree_skb(skb);
4962                         return 0;
4963                 } else {
4964                         tcp_send_ack(sk);
4965                 }
4966                 return -1;
4967         }
4968
4969         /* No ACK in the segment */
4970
4971         if (th->rst) {
4972                 /* rfc793:
4973                  * "If the RST bit is set
4974                  *
4975                  *      Otherwise (no ACK) drop the segment and return."
4976                  */
4977
4978                 goto discard_and_undo;
4979         }
4980
4981         /* PAWS check. */
4982         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4983                 goto discard_and_undo;
4984
4985         if (th->syn) {
4986                 /* We see SYN without ACK. It is attempt of
4987                  * simultaneous connect with crossed SYNs.
4988                  * Particularly, it can be connect to self.
4989                  */
4990                 tcp_set_state(sk, TCP_SYN_RECV);
4991
4992                 if (tp->rx_opt.saw_tstamp) {
4993                         tp->rx_opt.tstamp_ok = 1;
4994                         tcp_store_ts_recent(tp);
4995                         tp->tcp_header_len =
4996                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4997                 } else {
4998                         tp->tcp_header_len = sizeof(struct tcphdr);
4999                 }
5000
5001                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5002                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5003
5004                 /* RFC1323: The window in SYN & SYN/ACK segments is
5005                  * never scaled.
5006                  */
5007                 tp->snd_wnd    = ntohs(th->window);
5008                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5009                 tp->max_window = tp->snd_wnd;
5010
5011                 TCP_ECN_rcv_syn(tp, th);
5012
5013                 tcp_mtup_init(sk);
5014                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5015                 tcp_initialize_rcv_mss(sk);
5016
5017
5018                 tcp_send_synack(sk);
5019 #if 0
5020                 /* Note, we could accept data and URG from this segment.
5021                  * There are no obstacles to make this.
5022                  *
5023                  * However, if we ignore data in ACKless segments sometimes,
5024                  * we have no reasons to accept it sometimes.
5025                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5026                  * is not flawless. So, discard packet for sanity.
5027                  * Uncomment this return to process the data.
5028                  */
5029                 return -1;
5030 #else
5031                 goto discard;
5032 #endif
5033         }
5034         /* "fifth, if neither of the SYN or RST bits is set then
5035          * drop the segment and return."
5036          */
5037
5038 discard_and_undo:
5039         tcp_clear_options(&tp->rx_opt);
5040         tp->rx_opt.mss_clamp = saved_clamp;
5041         goto discard;
5042
5043 reset_and_undo:
5044         tcp_clear_options(&tp->rx_opt);
5045         tp->rx_opt.mss_clamp = saved_clamp;
5046         return 1;
5047 }
5048
5049
5050 /*
5051  *      This function implements the receiving procedure of RFC 793 for
5052  *      all states except ESTABLISHED and TIME_WAIT.
5053  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5054  *      address independent.
5055  */
5056
5057 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5058                           struct tcphdr *th, unsigned len)
5059 {
5060         struct tcp_sock *tp = tcp_sk(sk);
5061         struct inet_connection_sock *icsk = inet_csk(sk);
5062         int queued = 0;
5063
5064         tp->rx_opt.saw_tstamp = 0;
5065
5066         switch (sk->sk_state) {
5067         case TCP_CLOSE:
5068                 goto discard;
5069
5070         case TCP_LISTEN:
5071                 if (th->ack)
5072                         return 1;
5073
5074                 if (th->rst)
5075                         goto discard;
5076
5077                 if (th->syn) {
5078                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5079                                 return 1;
5080
5081                         /* Now we have several options: In theory there is
5082                          * nothing else in the frame. KA9Q has an option to
5083                          * send data with the syn, BSD accepts data with the
5084                          * syn up to the [to be] advertised window and
5085                          * Solaris 2.1 gives you a protocol error. For now
5086                          * we just ignore it, that fits the spec precisely
5087                          * and avoids incompatibilities. It would be nice in
5088                          * future to drop through and process the data.
5089                          *
5090                          * Now that TTCP is starting to be used we ought to
5091                          * queue this data.
5092                          * But, this leaves one open to an easy denial of
5093                          * service attack, and SYN cookies can't defend
5094                          * against this problem. So, we drop the data
5095                          * in the interest of security over speed unless
5096                          * it's still in use.
5097                          */
5098                         kfree_skb(skb);
5099                         return 0;
5100                 }
5101                 goto discard;
5102
5103         case TCP_SYN_SENT:
5104                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5105                 if (queued >= 0)
5106                         return queued;
5107
5108                 /* Do step6 onward by hand. */
5109                 tcp_urg(sk, skb, th);
5110                 __kfree_skb(skb);
5111                 tcp_data_snd_check(sk);
5112                 return 0;
5113         }
5114
5115         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5116             tcp_paws_discard(sk, skb)) {
5117                 if (!th->rst) {
5118                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
5119                         tcp_send_dupack(sk, skb);
5120                         goto discard;
5121                 }
5122                 /* Reset is accepted even if it did not pass PAWS. */
5123         }
5124
5125         /* step 1: check sequence number */
5126         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5127                 if (!th->rst)
5128                         tcp_send_dupack(sk, skb);
5129                 goto discard;
5130         }
5131
5132         /* step 2: check RST bit */
5133         if (th->rst) {
5134                 tcp_reset(sk);
5135                 goto discard;
5136         }
5137
5138         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5139
5140         /* step 3: check security and precedence [ignored] */
5141
5142         /*      step 4:
5143          *
5144          *      Check for a SYN in window.
5145          */
5146         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5147                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
5148                 tcp_reset(sk);
5149                 return 1;
5150         }
5151
5152         /* step 5: check the ACK field */
5153         if (th->ack) {
5154                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5155
5156                 switch (sk->sk_state) {
5157                 case TCP_SYN_RECV:
5158                         if (acceptable) {
5159                                 tp->copied_seq = tp->rcv_nxt;
5160                                 smp_mb();
5161                                 tcp_set_state(sk, TCP_ESTABLISHED);
5162                                 sk->sk_state_change(sk);
5163
5164                                 /* Note, that this wakeup is only for marginal
5165                                  * crossed SYN case. Passively open sockets
5166                                  * are not waked up, because sk->sk_sleep ==
5167                                  * NULL and sk->sk_socket == NULL.
5168                                  */
5169                                 if (sk->sk_socket)
5170                                         sk_wake_async(sk,
5171                                                         SOCK_WAKE_IO, POLL_OUT);
5172
5173                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5174                                 tp->snd_wnd = ntohs(th->window) <<
5175                                               tp->rx_opt.snd_wscale;
5176                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5177                                             TCP_SKB_CB(skb)->seq);
5178
5179                                 /* tcp_ack considers this ACK as duplicate
5180                                  * and does not calculate rtt.
5181                                  * Fix it at least with timestamps.
5182                                  */
5183                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5184                                     !tp->srtt)
5185                                         tcp_ack_saw_tstamp(sk, 0);
5186
5187                                 if (tp->rx_opt.tstamp_ok)
5188                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5189
5190                                 /* Make sure socket is routed, for
5191                                  * correct metrics.
5192                                  */
5193                                 icsk->icsk_af_ops->rebuild_header(sk);
5194
5195                                 tcp_init_metrics(sk);
5196
5197                                 tcp_init_congestion_control(sk);
5198
5199                                 /* Prevent spurious tcp_cwnd_restart() on
5200                                  * first data packet.
5201                                  */
5202                                 tp->lsndtime = tcp_time_stamp;
5203
5204                                 tcp_mtup_init(sk);
5205                                 tcp_initialize_rcv_mss(sk);
5206                                 tcp_init_buffer_space(sk);
5207                                 tcp_fast_path_on(tp);
5208                         } else {
5209                                 return 1;
5210                         }
5211                         break;
5212
5213                 case TCP_FIN_WAIT1:
5214                         if (tp->snd_una == tp->write_seq) {
5215                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5216                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5217                                 dst_confirm(sk->sk_dst_cache);
5218
5219                                 if (!sock_flag(sk, SOCK_DEAD))
5220                                         /* Wake up lingering close() */
5221                                         sk->sk_state_change(sk);
5222                                 else {
5223                                         int tmo;
5224
5225                                         if (tp->linger2 < 0 ||
5226                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5227                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5228                                                 tcp_done(sk);
5229                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5230                                                 return 1;
5231                                         }
5232
5233                                         tmo = tcp_fin_time(sk);
5234                                         if (tmo > TCP_TIMEWAIT_LEN) {
5235                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5236                                         } else if (th->fin || sock_owned_by_user(sk)) {
5237                                                 /* Bad case. We could lose such FIN otherwise.
5238                                                  * It is not a big problem, but it looks confusing
5239                                                  * and not so rare event. We still can lose it now,
5240                                                  * if it spins in bh_lock_sock(), but it is really
5241                                                  * marginal case.
5242                                                  */
5243                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5244                                         } else {
5245                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5246                                                 goto discard;
5247                                         }
5248                                 }
5249                         }
5250                         break;
5251
5252                 case TCP_CLOSING:
5253                         if (tp->snd_una == tp->write_seq) {
5254                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5255                                 goto discard;
5256                         }
5257                         break;
5258
5259                 case TCP_LAST_ACK:
5260                         if (tp->snd_una == tp->write_seq) {
5261                                 tcp_update_metrics(sk);
5262                                 tcp_done(sk);
5263                                 goto discard;
5264                         }
5265                         break;
5266                 }
5267         } else
5268                 goto discard;
5269
5270         /* step 6: check the URG bit */
5271         tcp_urg(sk, skb, th);
5272
5273         /* step 7: process the segment text */
5274         switch (sk->sk_state) {
5275         case TCP_CLOSE_WAIT:
5276         case TCP_CLOSING:
5277         case TCP_LAST_ACK:
5278                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5279                         break;
5280         case TCP_FIN_WAIT1:
5281         case TCP_FIN_WAIT2:
5282                 /* RFC 793 says to queue data in these states,
5283                  * RFC 1122 says we MUST send a reset.
5284                  * BSD 4.4 also does reset.
5285                  */
5286                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5287                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5288                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5289                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5290                                 tcp_reset(sk);
5291                                 return 1;
5292                         }
5293                 }
5294                 /* Fall through */
5295         case TCP_ESTABLISHED:
5296                 tcp_data_queue(sk, skb);
5297                 queued = 1;
5298                 break;
5299         }
5300
5301         /* tcp_data could move socket to TIME-WAIT */
5302         if (sk->sk_state != TCP_CLOSE) {
5303                 tcp_data_snd_check(sk);
5304                 tcp_ack_snd_check(sk);
5305         }
5306
5307         if (!queued) {
5308 discard:
5309                 __kfree_skb(skb);
5310         }
5311         return 0;
5312 }
5313
5314 EXPORT_SYMBOL(sysctl_tcp_ecn);
5315 EXPORT_SYMBOL(sysctl_tcp_reordering);
5316 EXPORT_SYMBOL(tcp_parse_options);
5317 EXPORT_SYMBOL(tcp_rcv_established);
5318 EXPORT_SYMBOL(tcp_rcv_state_process);
5319 EXPORT_SYMBOL(tcp_initialize_rcv_mss);