[IPV4]: Fix byte value boundary check in do_ip_getsockopt().
[linux-2.6] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
108 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
109
110 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
111 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
112 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
113 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
114 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115
116 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
117
118 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
119 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
120
121 /* Adapt the MSS value used to make delayed ack decision to the
122  * real world.
123  */
124 static void tcp_measure_rcv_mss(struct sock *sk, 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, struct sk_buff *skb)
306 {
307         struct tcp_sock *tp = tcp_sk(sk);
308
309         /* Check #1 */
310         if (tp->rcv_ssthresh < tp->window_clamp &&
311             (int)tp->rcv_ssthresh < tcp_space(sk) &&
312             !tcp_memory_pressure) {
313                 int incr;
314
315                 /* Check #2. Increase window, if skb with such overhead
316                  * will fit to rcvbuf in future.
317                  */
318                 if (tcp_win_from_space(skb->truesize) <= skb->len)
319                         incr = 2 * tp->advmss;
320                 else
321                         incr = __tcp_grow_window(sk, skb);
322
323                 if (incr) {
324                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
325                                                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 /* Initialize RCV_MSS value.
404  * RCV_MSS is an our guess about MSS used by the peer.
405  * We haven't any direct information about the MSS.
406  * It's better to underestimate the RCV_MSS rather than overestimate.
407  * Overestimations make us ACKing less frequently than needed.
408  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
409  */
410 void tcp_initialize_rcv_mss(struct sock *sk)
411 {
412         struct tcp_sock *tp = tcp_sk(sk);
413         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
414
415         hint = min(hint, tp->rcv_wnd / 2);
416         hint = min(hint, TCP_MIN_RCVMSS);
417         hint = max(hint, TCP_MIN_MSS);
418
419         inet_csk(sk)->icsk_ack.rcv_mss = hint;
420 }
421
422 /* Receiver "autotuning" code.
423  *
424  * The algorithm for RTT estimation w/o timestamps is based on
425  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
426  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
427  *
428  * More detail on this code can be found at
429  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
430  * though this reference is out of date.  A new paper
431  * is pending.
432  */
433 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
434 {
435         u32 new_sample = tp->rcv_rtt_est.rtt;
436         long m = sample;
437
438         if (m == 0)
439                 m = 1;
440
441         if (new_sample != 0) {
442                 /* If we sample in larger samples in the non-timestamp
443                  * case, we could grossly overestimate the RTT especially
444                  * with chatty applications or bulk transfer apps which
445                  * are stalled on filesystem I/O.
446                  *
447                  * Also, since we are only going for a minimum in the
448                  * non-timestamp case, we do not smooth things out
449                  * else with timestamps disabled convergence takes too
450                  * long.
451                  */
452                 if (!win_dep) {
453                         m -= (new_sample >> 3);
454                         new_sample += m;
455                 } else if (m < new_sample)
456                         new_sample = m << 3;
457         } else {
458                 /* No previous measure. */
459                 new_sample = m << 3;
460         }
461
462         if (tp->rcv_rtt_est.rtt != new_sample)
463                 tp->rcv_rtt_est.rtt = new_sample;
464 }
465
466 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
467 {
468         if (tp->rcv_rtt_est.time == 0)
469                 goto new_measure;
470         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
471                 return;
472         tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
473
474 new_measure:
475         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
476         tp->rcv_rtt_est.time = tcp_time_stamp;
477 }
478
479 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
480                                           const struct sk_buff *skb)
481 {
482         struct tcp_sock *tp = tcp_sk(sk);
483         if (tp->rx_opt.rcv_tsecr &&
484             (TCP_SKB_CB(skb)->end_seq -
485              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
486                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
487 }
488
489 /*
490  * This function should be called every time data is copied to user space.
491  * It calculates the appropriate TCP receive buffer space.
492  */
493 void tcp_rcv_space_adjust(struct sock *sk)
494 {
495         struct tcp_sock *tp = tcp_sk(sk);
496         int time;
497         int space;
498
499         if (tp->rcvq_space.time == 0)
500                 goto new_measure;
501
502         time = tcp_time_stamp - tp->rcvq_space.time;
503         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
504                 return;
505
506         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
507
508         space = max(tp->rcvq_space.space, space);
509
510         if (tp->rcvq_space.space != space) {
511                 int rcvmem;
512
513                 tp->rcvq_space.space = space;
514
515                 if (sysctl_tcp_moderate_rcvbuf &&
516                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
517                         int new_clamp = space;
518
519                         /* Receive space grows, normalize in order to
520                          * take into account packet headers and sk_buff
521                          * structure overhead.
522                          */
523                         space /= tp->advmss;
524                         if (!space)
525                                 space = 1;
526                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
527                                   16 + sizeof(struct sk_buff));
528                         while (tcp_win_from_space(rcvmem) < tp->advmss)
529                                 rcvmem += 128;
530                         space *= rcvmem;
531                         space = min(space, sysctl_tcp_rmem[2]);
532                         if (space > sk->sk_rcvbuf) {
533                                 sk->sk_rcvbuf = space;
534
535                                 /* Make the window clamp follow along.  */
536                                 tp->window_clamp = new_clamp;
537                         }
538                 }
539         }
540
541 new_measure:
542         tp->rcvq_space.seq = tp->copied_seq;
543         tp->rcvq_space.time = tcp_time_stamp;
544 }
545
546 /* There is something which you must keep in mind when you analyze the
547  * behavior of the tp->ato delayed ack timeout interval.  When a
548  * connection starts up, we want to ack as quickly as possible.  The
549  * problem is that "good" TCP's do slow start at the beginning of data
550  * transmission.  The means that until we send the first few ACK's the
551  * sender will sit on his end and only queue most of his data, because
552  * he can only send snd_cwnd unacked packets at any given time.  For
553  * each ACK we send, he increments snd_cwnd and transmits more of his
554  * queue.  -DaveM
555  */
556 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
557 {
558         struct tcp_sock *tp = tcp_sk(sk);
559         struct inet_connection_sock *icsk = inet_csk(sk);
560         u32 now;
561
562         inet_csk_schedule_ack(sk);
563
564         tcp_measure_rcv_mss(sk, skb);
565
566         tcp_rcv_rtt_measure(tp);
567
568         now = tcp_time_stamp;
569
570         if (!icsk->icsk_ack.ato) {
571                 /* The _first_ data packet received, initialize
572                  * delayed ACK engine.
573                  */
574                 tcp_incr_quickack(sk);
575                 icsk->icsk_ack.ato = TCP_ATO_MIN;
576         } else {
577                 int m = now - icsk->icsk_ack.lrcvtime;
578
579                 if (m <= TCP_ATO_MIN / 2) {
580                         /* The fastest case is the first. */
581                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
582                 } else if (m < icsk->icsk_ack.ato) {
583                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
584                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
585                                 icsk->icsk_ack.ato = icsk->icsk_rto;
586                 } else if (m > icsk->icsk_rto) {
587                         /* Too long gap. Apparently sender failed to
588                          * restart window, so that we send ACKs quickly.
589                          */
590                         tcp_incr_quickack(sk);
591                         sk_mem_reclaim(sk);
592                 }
593         }
594         icsk->icsk_ack.lrcvtime = now;
595
596         TCP_ECN_check_ce(tp, skb);
597
598         if (skb->len >= 128)
599                 tcp_grow_window(sk, skb);
600 }
601
602 static u32 tcp_rto_min(struct sock *sk)
603 {
604         struct dst_entry *dst = __sk_dst_get(sk);
605         u32 rto_min = TCP_RTO_MIN;
606
607         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
608                 rto_min = dst->metrics[RTAX_RTO_MIN - 1];
609         return rto_min;
610 }
611
612 /* Called to compute a smoothed rtt estimate. The data fed to this
613  * routine either comes from timestamps, or from segments that were
614  * known _not_ to have been retransmitted [see Karn/Partridge
615  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
616  * piece by Van Jacobson.
617  * NOTE: the next three routines used to be one big routine.
618  * To save cycles in the RFC 1323 implementation it was better to break
619  * it up into three procedures. -- erics
620  */
621 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
622 {
623         struct tcp_sock *tp = tcp_sk(sk);
624         long m = mrtt; /* RTT */
625
626         /*      The following amusing code comes from Jacobson's
627          *      article in SIGCOMM '88.  Note that rtt and mdev
628          *      are scaled versions of rtt and mean deviation.
629          *      This is designed to be as fast as possible
630          *      m stands for "measurement".
631          *
632          *      On a 1990 paper the rto value is changed to:
633          *      RTO = rtt + 4 * mdev
634          *
635          * Funny. This algorithm seems to be very broken.
636          * These formulae increase RTO, when it should be decreased, increase
637          * too slowly, when it should be increased quickly, decrease too quickly
638          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
639          * does not matter how to _calculate_ it. Seems, it was trap
640          * that VJ failed to avoid. 8)
641          */
642         if (m == 0)
643                 m = 1;
644         if (tp->srtt != 0) {
645                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
646                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
647                 if (m < 0) {
648                         m = -m;         /* m is now abs(error) */
649                         m -= (tp->mdev >> 2);   /* similar update on mdev */
650                         /* This is similar to one of Eifel findings.
651                          * Eifel blocks mdev updates when rtt decreases.
652                          * This solution is a bit different: we use finer gain
653                          * for mdev in this case (alpha*beta).
654                          * Like Eifel it also prevents growth of rto,
655                          * but also it limits too fast rto decreases,
656                          * happening in pure Eifel.
657                          */
658                         if (m > 0)
659                                 m >>= 3;
660                 } else {
661                         m -= (tp->mdev >> 2);   /* similar update on mdev */
662                 }
663                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
664                 if (tp->mdev > tp->mdev_max) {
665                         tp->mdev_max = tp->mdev;
666                         if (tp->mdev_max > tp->rttvar)
667                                 tp->rttvar = tp->mdev_max;
668                 }
669                 if (after(tp->snd_una, tp->rtt_seq)) {
670                         if (tp->mdev_max < tp->rttvar)
671                                 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
672                         tp->rtt_seq = tp->snd_nxt;
673                         tp->mdev_max = tcp_rto_min(sk);
674                 }
675         } else {
676                 /* no previous measure. */
677                 tp->srtt = m << 3;      /* take the measured time to be rtt */
678                 tp->mdev = m << 1;      /* make sure rto = 3*rtt */
679                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
680                 tp->rtt_seq = tp->snd_nxt;
681         }
682 }
683
684 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
685  * routine referred to above.
686  */
687 static inline void tcp_set_rto(struct sock *sk)
688 {
689         const struct tcp_sock *tp = tcp_sk(sk);
690         /* Old crap is replaced with new one. 8)
691          *
692          * More seriously:
693          * 1. If rtt variance happened to be less 50msec, it is hallucination.
694          *    It cannot be less due to utterly erratic ACK generation made
695          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
696          *    to do with delayed acks, because at cwnd>2 true delack timeout
697          *    is invisible. Actually, Linux-2.4 also generates erratic
698          *    ACKs in some circumstances.
699          */
700         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
701
702         /* 2. Fixups made earlier cannot be right.
703          *    If we do not estimate RTO correctly without them,
704          *    all the algo is pure shit and should be replaced
705          *    with correct one. It is exactly, which we pretend to do.
706          */
707 }
708
709 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
710  * guarantees that rto is higher.
711  */
712 static inline void tcp_bound_rto(struct sock *sk)
713 {
714         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
715                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
716 }
717
718 /* Save metrics learned by this TCP session.
719    This function is called only, when TCP finishes successfully
720    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
721  */
722 void tcp_update_metrics(struct sock *sk)
723 {
724         struct tcp_sock *tp = tcp_sk(sk);
725         struct dst_entry *dst = __sk_dst_get(sk);
726
727         if (sysctl_tcp_nometrics_save)
728                 return;
729
730         dst_confirm(dst);
731
732         if (dst && (dst->flags & DST_HOST)) {
733                 const struct inet_connection_sock *icsk = inet_csk(sk);
734                 int m;
735
736                 if (icsk->icsk_backoff || !tp->srtt) {
737                         /* This session failed to estimate rtt. Why?
738                          * Probably, no packets returned in time.
739                          * Reset our results.
740                          */
741                         if (!(dst_metric_locked(dst, RTAX_RTT)))
742                                 dst->metrics[RTAX_RTT - 1] = 0;
743                         return;
744                 }
745
746                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
747
748                 /* If newly calculated rtt larger than stored one,
749                  * store new one. Otherwise, use EWMA. Remember,
750                  * rtt overestimation is always better than underestimation.
751                  */
752                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
753                         if (m <= 0)
754                                 dst->metrics[RTAX_RTT - 1] = tp->srtt;
755                         else
756                                 dst->metrics[RTAX_RTT - 1] -= (m >> 3);
757                 }
758
759                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
760                         if (m < 0)
761                                 m = -m;
762
763                         /* Scale deviation to rttvar fixed point */
764                         m >>= 1;
765                         if (m < tp->mdev)
766                                 m = tp->mdev;
767
768                         if (m >= dst_metric(dst, RTAX_RTTVAR))
769                                 dst->metrics[RTAX_RTTVAR - 1] = m;
770                         else
771                                 dst->metrics[RTAX_RTTVAR-1] -=
772                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
773                 }
774
775                 if (tp->snd_ssthresh >= 0xFFFF) {
776                         /* Slow start still did not finish. */
777                         if (dst_metric(dst, RTAX_SSTHRESH) &&
778                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
779                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
780                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
781                         if (!dst_metric_locked(dst, RTAX_CWND) &&
782                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
783                                 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
784                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
785                            icsk->icsk_ca_state == TCP_CA_Open) {
786                         /* Cong. avoidance phase, cwnd is reliable. */
787                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
788                                 dst->metrics[RTAX_SSTHRESH-1] =
789                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
790                         if (!dst_metric_locked(dst, RTAX_CWND))
791                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
792                 } else {
793                         /* Else slow start did not finish, cwnd is non-sense,
794                            ssthresh may be also invalid.
795                          */
796                         if (!dst_metric_locked(dst, RTAX_CWND))
797                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
798                         if (dst->metrics[RTAX_SSTHRESH-1] &&
799                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
800                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
801                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
802                 }
803
804                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
805                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
806                             tp->reordering != sysctl_tcp_reordering)
807                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
808                 }
809         }
810 }
811
812 /* Numbers are taken from RFC3390.
813  *
814  * John Heffner states:
815  *
816  *      The RFC specifies a window of no more than 4380 bytes
817  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
818  *      is a bit misleading because they use a clamp at 4380 bytes
819  *      rather than use a multiplier in the relevant range.
820  */
821 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
822 {
823         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
824
825         if (!cwnd) {
826                 if (tp->mss_cache > 1460)
827                         cwnd = 2;
828                 else
829                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
830         }
831         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
832 }
833
834 /* Set slow start threshold and cwnd not falling to slow start */
835 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
836 {
837         struct tcp_sock *tp = tcp_sk(sk);
838         const struct inet_connection_sock *icsk = inet_csk(sk);
839
840         tp->prior_ssthresh = 0;
841         tp->bytes_acked = 0;
842         if (icsk->icsk_ca_state < TCP_CA_CWR) {
843                 tp->undo_marker = 0;
844                 if (set_ssthresh)
845                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
846                 tp->snd_cwnd = min(tp->snd_cwnd,
847                                    tcp_packets_in_flight(tp) + 1U);
848                 tp->snd_cwnd_cnt = 0;
849                 tp->high_seq = tp->snd_nxt;
850                 tp->snd_cwnd_stamp = tcp_time_stamp;
851                 TCP_ECN_queue_cwr(tp);
852
853                 tcp_set_ca_state(sk, TCP_CA_CWR);
854         }
855 }
856
857 /*
858  * Packet counting of FACK is based on in-order assumptions, therefore TCP
859  * disables it when reordering is detected
860  */
861 static void tcp_disable_fack(struct tcp_sock *tp)
862 {
863         /* RFC3517 uses different metric in lost marker => reset on change */
864         if (tcp_is_fack(tp))
865                 tp->lost_skb_hint = NULL;
866         tp->rx_opt.sack_ok &= ~2;
867 }
868
869 /* Take a notice that peer is sending D-SACKs */
870 static void tcp_dsack_seen(struct tcp_sock *tp)
871 {
872         tp->rx_opt.sack_ok |= 4;
873 }
874
875 /* Initialize metrics on socket. */
876
877 static void tcp_init_metrics(struct sock *sk)
878 {
879         struct tcp_sock *tp = tcp_sk(sk);
880         struct dst_entry *dst = __sk_dst_get(sk);
881
882         if (dst == NULL)
883                 goto reset;
884
885         dst_confirm(dst);
886
887         if (dst_metric_locked(dst, RTAX_CWND))
888                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889         if (dst_metric(dst, RTAX_SSTHRESH)) {
890                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
893         }
894         if (dst_metric(dst, RTAX_REORDERING) &&
895             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
896                 tcp_disable_fack(tp);
897                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
898         }
899
900         if (dst_metric(dst, RTAX_RTT) == 0)
901                 goto reset;
902
903         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
904                 goto reset;
905
906         /* Initial rtt is determined from SYN,SYN-ACK.
907          * The segment is small and rtt may appear much
908          * less than real one. Use per-dst memory
909          * to make it more realistic.
910          *
911          * A bit of theory. RTT is time passed after "normal" sized packet
912          * is sent until it is ACKed. In normal circumstances sending small
913          * packets force peer to delay ACKs and calculation is correct too.
914          * The algorithm is adaptive and, provided we follow specs, it
915          * NEVER underestimate RTT. BUT! If peer tries to make some clever
916          * tricks sort of "quick acks" for time long enough to decrease RTT
917          * to low value, and then abruptly stops to do it and starts to delay
918          * ACKs, wait for troubles.
919          */
920         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
921                 tp->srtt = dst_metric(dst, RTAX_RTT);
922                 tp->rtt_seq = tp->snd_nxt;
923         }
924         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
925                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
926                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
927         }
928         tcp_set_rto(sk);
929         tcp_bound_rto(sk);
930         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
931                 goto reset;
932         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
933         tp->snd_cwnd_stamp = tcp_time_stamp;
934         return;
935
936 reset:
937         /* Play conservative. If timestamps are not
938          * supported, TCP will fail to recalculate correct
939          * rtt, if initial rto is too small. FORGET ALL AND RESET!
940          */
941         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
942                 tp->srtt = 0;
943                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
944                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
945         }
946 }
947
948 static void tcp_update_reordering(struct sock *sk, const int metric,
949                                   const int ts)
950 {
951         struct tcp_sock *tp = tcp_sk(sk);
952         if (metric > tp->reordering) {
953                 tp->reordering = min(TCP_MAX_REORDERING, metric);
954
955                 /* This exciting event is worth to be remembered. 8) */
956                 if (ts)
957                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
958                 else if (tcp_is_reno(tp))
959                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
960                 else if (tcp_is_fack(tp))
961                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
962                 else
963                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
964 #if FASTRETRANS_DEBUG > 1
965                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
966                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
967                        tp->reordering,
968                        tp->fackets_out,
969                        tp->sacked_out,
970                        tp->undo_marker ? tp->undo_retrans : 0);
971 #endif
972                 tcp_disable_fack(tp);
973         }
974 }
975
976 /* This procedure tags the retransmission queue when SACKs arrive.
977  *
978  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
979  * Packets in queue with these bits set are counted in variables
980  * sacked_out, retrans_out and lost_out, correspondingly.
981  *
982  * Valid combinations are:
983  * Tag  InFlight        Description
984  * 0    1               - orig segment is in flight.
985  * S    0               - nothing flies, orig reached receiver.
986  * L    0               - nothing flies, orig lost by net.
987  * R    2               - both orig and retransmit are in flight.
988  * L|R  1               - orig is lost, retransmit is in flight.
989  * S|R  1               - orig reached receiver, retrans is still in flight.
990  * (L|S|R is logically valid, it could occur when L|R is sacked,
991  *  but it is equivalent to plain S and code short-curcuits it to S.
992  *  L|S is logically invalid, it would mean -1 packet in flight 8))
993  *
994  * These 6 states form finite state machine, controlled by the following events:
995  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
996  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
997  * 3. Loss detection event of one of three flavors:
998  *      A. Scoreboard estimator decided the packet is lost.
999  *         A'. Reno "three dupacks" marks head of queue lost.
1000  *         A''. Its FACK modfication, head until snd.fack is lost.
1001  *      B. SACK arrives sacking data transmitted after never retransmitted
1002  *         hole was sent out.
1003  *      C. SACK arrives sacking SND.NXT at the moment, when the
1004  *         segment was retransmitted.
1005  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1006  *
1007  * It is pleasant to note, that state diagram turns out to be commutative,
1008  * so that we are allowed not to be bothered by order of our actions,
1009  * when multiple events arrive simultaneously. (see the function below).
1010  *
1011  * Reordering detection.
1012  * --------------------
1013  * Reordering metric is maximal distance, which a packet can be displaced
1014  * in packet stream. With SACKs we can estimate it:
1015  *
1016  * 1. SACK fills old hole and the corresponding segment was not
1017  *    ever retransmitted -> reordering. Alas, we cannot use it
1018  *    when segment was retransmitted.
1019  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1020  *    for retransmitted and already SACKed segment -> reordering..
1021  * Both of these heuristics are not used in Loss state, when we cannot
1022  * account for retransmits accurately.
1023  *
1024  * SACK block validation.
1025  * ----------------------
1026  *
1027  * SACK block range validation checks that the received SACK block fits to
1028  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1029  * Note that SND.UNA is not included to the range though being valid because
1030  * it means that the receiver is rather inconsistent with itself reporting
1031  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1032  * perfectly valid, however, in light of RFC2018 which explicitly states
1033  * that "SACK block MUST reflect the newest segment.  Even if the newest
1034  * segment is going to be discarded ...", not that it looks very clever
1035  * in case of head skb. Due to potentional receiver driven attacks, we
1036  * choose to avoid immediate execution of a walk in write queue due to
1037  * reneging and defer head skb's loss recovery to standard loss recovery
1038  * procedure that will eventually trigger (nothing forbids us doing this).
1039  *
1040  * Implements also blockage to start_seq wrap-around. Problem lies in the
1041  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1042  * there's no guarantee that it will be before snd_nxt (n). The problem
1043  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1044  * wrap (s_w):
1045  *
1046  *         <- outs wnd ->                          <- wrapzone ->
1047  *         u     e      n                         u_w   e_w  s n_w
1048  *         |     |      |                          |     |   |  |
1049  * |<------------+------+----- TCP seqno space --------------+---------->|
1050  * ...-- <2^31 ->|                                           |<--------...
1051  * ...---- >2^31 ------>|                                    |<--------...
1052  *
1053  * Current code wouldn't be vulnerable but it's better still to discard such
1054  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1055  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1056  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1057  * equal to the ideal case (infinite seqno space without wrap caused issues).
1058  *
1059  * With D-SACK the lower bound is extended to cover sequence space below
1060  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1061  * again, D-SACK block must not to go across snd_una (for the same reason as
1062  * for the normal SACK blocks, explained above). But there all simplicity
1063  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1064  * fully below undo_marker they do not affect behavior in anyway and can
1065  * therefore be safely ignored. In rare cases (which are more or less
1066  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1067  * fragmentation and packet reordering past skb's retransmission. To consider
1068  * them correctly, the acceptable range must be extended even more though
1069  * the exact amount is rather hard to quantify. However, tp->max_window can
1070  * be used as an exaggerated estimate.
1071  */
1072 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1073                                   u32 start_seq, u32 end_seq)
1074 {
1075         /* Too far in future, or reversed (interpretation is ambiguous) */
1076         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1077                 return 0;
1078
1079         /* Nasty start_seq wrap-around check (see comments above) */
1080         if (!before(start_seq, tp->snd_nxt))
1081                 return 0;
1082
1083         /* In outstanding window? ...This is valid exit for D-SACKs too.
1084          * start_seq == snd_una is non-sensical (see comments above)
1085          */
1086         if (after(start_seq, tp->snd_una))
1087                 return 1;
1088
1089         if (!is_dsack || !tp->undo_marker)
1090                 return 0;
1091
1092         /* ...Then it's D-SACK, and must reside below snd_una completely */
1093         if (!after(end_seq, tp->snd_una))
1094                 return 0;
1095
1096         if (!before(start_seq, tp->undo_marker))
1097                 return 1;
1098
1099         /* Too old */
1100         if (!after(end_seq, tp->undo_marker))
1101                 return 0;
1102
1103         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1104          *   start_seq < undo_marker and end_seq >= undo_marker.
1105          */
1106         return !before(start_seq, end_seq - tp->max_window);
1107 }
1108
1109 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1110  * Event "C". Later note: FACK people cheated me again 8), we have to account
1111  * for reordering! Ugly, but should help.
1112  *
1113  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1114  * less than what is now known to be received by the other end (derived from
1115  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1116  * retransmitted skbs to avoid some costly processing per ACKs.
1117  */
1118 static void tcp_mark_lost_retrans(struct sock *sk)
1119 {
1120         const struct inet_connection_sock *icsk = inet_csk(sk);
1121         struct tcp_sock *tp = tcp_sk(sk);
1122         struct sk_buff *skb;
1123         int cnt = 0;
1124         u32 new_low_seq = tp->snd_nxt;
1125         u32 received_upto = tcp_highest_sack_seq(tp);
1126
1127         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1128             !after(received_upto, tp->lost_retrans_low) ||
1129             icsk->icsk_ca_state != TCP_CA_Recovery)
1130                 return;
1131
1132         tcp_for_write_queue(skb, sk) {
1133                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1134
1135                 if (skb == tcp_send_head(sk))
1136                         break;
1137                 if (cnt == tp->retrans_out)
1138                         break;
1139                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1140                         continue;
1141
1142                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1143                         continue;
1144
1145                 if (after(received_upto, ack_seq) &&
1146                     (tcp_is_fack(tp) ||
1147                      !before(received_upto,
1148                              ack_seq + tp->reordering * tp->mss_cache))) {
1149                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1150                         tp->retrans_out -= tcp_skb_pcount(skb);
1151
1152                         /* clear lost hint */
1153                         tp->retransmit_skb_hint = NULL;
1154
1155                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1156                                 tp->lost_out += tcp_skb_pcount(skb);
1157                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1158                         }
1159                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1160                 } else {
1161                         if (before(ack_seq, new_low_seq))
1162                                 new_low_seq = ack_seq;
1163                         cnt += tcp_skb_pcount(skb);
1164                 }
1165         }
1166
1167         if (tp->retrans_out)
1168                 tp->lost_retrans_low = new_low_seq;
1169 }
1170
1171 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1172                            struct tcp_sack_block_wire *sp, int num_sacks,
1173                            u32 prior_snd_una)
1174 {
1175         u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1176         u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1177         int dup_sack = 0;
1178
1179         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1180                 dup_sack = 1;
1181                 tcp_dsack_seen(tp);
1182                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1183         } else if (num_sacks > 1) {
1184                 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1185                 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1186
1187                 if (!after(end_seq_0, end_seq_1) &&
1188                     !before(start_seq_0, start_seq_1)) {
1189                         dup_sack = 1;
1190                         tcp_dsack_seen(tp);
1191                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1192                 }
1193         }
1194
1195         /* D-SACK for already forgotten data... Do dumb counting. */
1196         if (dup_sack &&
1197             !after(end_seq_0, prior_snd_una) &&
1198             after(end_seq_0, tp->undo_marker))
1199                 tp->undo_retrans--;
1200
1201         return dup_sack;
1202 }
1203
1204 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1205  * the incoming SACK may not exactly match but we can find smaller MSS
1206  * aligned portion of it that matches. Therefore we might need to fragment
1207  * which may fail and creates some hassle (caller must handle error case
1208  * returns).
1209  */
1210 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1211                                  u32 start_seq, u32 end_seq)
1212 {
1213         int in_sack, err;
1214         unsigned int pkt_len;
1215
1216         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1217                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1218
1219         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1220             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1221
1222                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1223
1224                 if (!in_sack)
1225                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1226                 else
1227                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1228                 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1229                 if (err < 0)
1230                         return err;
1231         }
1232
1233         return in_sack;
1234 }
1235
1236 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1237                            int *reord, int dup_sack, int fack_count)
1238 {
1239         struct tcp_sock *tp = tcp_sk(sk);
1240         u8 sacked = TCP_SKB_CB(skb)->sacked;
1241         int flag = 0;
1242
1243         /* Account D-SACK for retransmitted packet. */
1244         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1245                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1246                         tp->undo_retrans--;
1247                 if (sacked & TCPCB_SACKED_ACKED)
1248                         *reord = min(fack_count, *reord);
1249         }
1250
1251         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1252         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1253                 return flag;
1254
1255         if (!(sacked & TCPCB_SACKED_ACKED)) {
1256                 if (sacked & TCPCB_SACKED_RETRANS) {
1257                         /* If the segment is not tagged as lost,
1258                          * we do not clear RETRANS, believing
1259                          * that retransmission is still in flight.
1260                          */
1261                         if (sacked & TCPCB_LOST) {
1262                                 TCP_SKB_CB(skb)->sacked &=
1263                                         ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1264                                 tp->lost_out -= tcp_skb_pcount(skb);
1265                                 tp->retrans_out -= tcp_skb_pcount(skb);
1266
1267                                 /* clear lost hint */
1268                                 tp->retransmit_skb_hint = NULL;
1269                         }
1270                 } else {
1271                         if (!(sacked & TCPCB_RETRANS)) {
1272                                 /* New sack for not retransmitted frame,
1273                                  * which was in hole. It is reordering.
1274                                  */
1275                                 if (before(TCP_SKB_CB(skb)->seq,
1276                                            tcp_highest_sack_seq(tp)))
1277                                         *reord = min(fack_count, *reord);
1278
1279                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1280                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1281                                         flag |= FLAG_ONLY_ORIG_SACKED;
1282                         }
1283
1284                         if (sacked & TCPCB_LOST) {
1285                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1286                                 tp->lost_out -= tcp_skb_pcount(skb);
1287
1288                                 /* clear lost hint */
1289                                 tp->retransmit_skb_hint = NULL;
1290                         }
1291                 }
1292
1293                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1294                 flag |= FLAG_DATA_SACKED;
1295                 tp->sacked_out += tcp_skb_pcount(skb);
1296
1297                 fack_count += tcp_skb_pcount(skb);
1298
1299                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1300                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1301                     before(TCP_SKB_CB(skb)->seq,
1302                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1303                         tp->lost_cnt_hint += tcp_skb_pcount(skb);
1304
1305                 if (fack_count > tp->fackets_out)
1306                         tp->fackets_out = fack_count;
1307
1308                 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1309                         tcp_advance_highest_sack(sk, skb);
1310         }
1311
1312         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1313          * frames and clear it. undo_retrans is decreased above, L|R frames
1314          * are accounted above as well.
1315          */
1316         if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1317                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1318                 tp->retrans_out -= tcp_skb_pcount(skb);
1319                 tp->retransmit_skb_hint = NULL;
1320         }
1321
1322         return flag;
1323 }
1324
1325 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1326                                         struct tcp_sack_block *next_dup,
1327                                         u32 start_seq, u32 end_seq,
1328                                         int dup_sack_in, int *fack_count,
1329                                         int *reord, int *flag)
1330 {
1331         tcp_for_write_queue_from(skb, sk) {
1332                 int in_sack = 0;
1333                 int dup_sack = dup_sack_in;
1334
1335                 if (skb == tcp_send_head(sk))
1336                         break;
1337
1338                 /* queue is in-order => we can short-circuit the walk early */
1339                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1340                         break;
1341
1342                 if ((next_dup != NULL) &&
1343                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1344                         in_sack = tcp_match_skb_to_sack(sk, skb,
1345                                                         next_dup->start_seq,
1346                                                         next_dup->end_seq);
1347                         if (in_sack > 0)
1348                                 dup_sack = 1;
1349                 }
1350
1351                 if (in_sack <= 0)
1352                         in_sack = tcp_match_skb_to_sack(sk, skb, start_seq,
1353                                                         end_seq);
1354                 if (unlikely(in_sack < 0))
1355                         break;
1356
1357                 if (in_sack)
1358                         *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack,
1359                                                  *fack_count);
1360
1361                 *fack_count += tcp_skb_pcount(skb);
1362         }
1363         return skb;
1364 }
1365
1366 /* Avoid all extra work that is being done by sacktag while walking in
1367  * a normal way
1368  */
1369 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1370                                         u32 skip_to_seq, int *fack_count)
1371 {
1372         tcp_for_write_queue_from(skb, sk) {
1373                 if (skb == tcp_send_head(sk))
1374                         break;
1375
1376                 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1377                         break;
1378
1379                 *fack_count += tcp_skb_pcount(skb);
1380         }
1381         return skb;
1382 }
1383
1384 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1385                                                 struct sock *sk,
1386                                                 struct tcp_sack_block *next_dup,
1387                                                 u32 skip_to_seq,
1388                                                 int *fack_count, int *reord,
1389                                                 int *flag)
1390 {
1391         if (next_dup == NULL)
1392                 return skb;
1393
1394         if (before(next_dup->start_seq, skip_to_seq)) {
1395                 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
1396                 tcp_sacktag_walk(skb, sk, NULL,
1397                                  next_dup->start_seq, next_dup->end_seq,
1398                                  1, fack_count, reord, flag);
1399         }
1400
1401         return skb;
1402 }
1403
1404 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1405 {
1406         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1407 }
1408
1409 static int
1410 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1411                         u32 prior_snd_una)
1412 {
1413         const struct inet_connection_sock *icsk = inet_csk(sk);
1414         struct tcp_sock *tp = tcp_sk(sk);
1415         unsigned char *ptr = (skb_transport_header(ack_skb) +
1416                               TCP_SKB_CB(ack_skb)->sacked);
1417         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1418         struct tcp_sack_block sp[4];
1419         struct tcp_sack_block *cache;
1420         struct sk_buff *skb;
1421         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE) >> 3;
1422         int used_sacks;
1423         int reord = tp->packets_out;
1424         int flag = 0;
1425         int found_dup_sack = 0;
1426         int fack_count;
1427         int i, j;
1428         int first_sack_index;
1429
1430         if (!tp->sacked_out) {
1431                 if (WARN_ON(tp->fackets_out))
1432                         tp->fackets_out = 0;
1433                 tcp_highest_sack_reset(sk);
1434         }
1435
1436         found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire,
1437                                          num_sacks, prior_snd_una);
1438         if (found_dup_sack)
1439                 flag |= FLAG_DSACKING_ACK;
1440
1441         /* Eliminate too old ACKs, but take into
1442          * account more or less fresh ones, they can
1443          * contain valid SACK info.
1444          */
1445         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1446                 return 0;
1447
1448         if (!tp->packets_out)
1449                 goto out;
1450
1451         used_sacks = 0;
1452         first_sack_index = 0;
1453         for (i = 0; i < num_sacks; i++) {
1454                 int dup_sack = !i && found_dup_sack;
1455
1456                 sp[used_sacks].start_seq = ntohl(get_unaligned(&sp_wire[i].start_seq));
1457                 sp[used_sacks].end_seq = ntohl(get_unaligned(&sp_wire[i].end_seq));
1458
1459                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1460                                             sp[used_sacks].start_seq,
1461                                             sp[used_sacks].end_seq)) {
1462                         if (dup_sack) {
1463                                 if (!tp->undo_marker)
1464                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1465                                 else
1466                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1467                         } else {
1468                                 /* Don't count olds caused by ACK reordering */
1469                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1470                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1471                                         continue;
1472                                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1473                         }
1474                         if (i == 0)
1475                                 first_sack_index = -1;
1476                         continue;
1477                 }
1478
1479                 /* Ignore very old stuff early */
1480                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1481                         continue;
1482
1483                 used_sacks++;
1484         }
1485
1486         /* order SACK blocks to allow in order walk of the retrans queue */
1487         for (i = used_sacks - 1; i > 0; i--) {
1488                 for (j = 0; j < i; j++) {
1489                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1490                                 struct tcp_sack_block tmp;
1491
1492                                 tmp = sp[j];
1493                                 sp[j] = sp[j + 1];
1494                                 sp[j + 1] = tmp;
1495
1496                                 /* Track where the first SACK block goes to */
1497                                 if (j == first_sack_index)
1498                                         first_sack_index = j + 1;
1499                         }
1500                 }
1501         }
1502
1503         skb = tcp_write_queue_head(sk);
1504         fack_count = 0;
1505         i = 0;
1506
1507         if (!tp->sacked_out) {
1508                 /* It's already past, so skip checking against it */
1509                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1510         } else {
1511                 cache = tp->recv_sack_cache;
1512                 /* Skip empty blocks in at head of the cache */
1513                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1514                        !cache->end_seq)
1515                         cache++;
1516         }
1517
1518         while (i < used_sacks) {
1519                 u32 start_seq = sp[i].start_seq;
1520                 u32 end_seq = sp[i].end_seq;
1521                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1522                 struct tcp_sack_block *next_dup = NULL;
1523
1524                 if (found_dup_sack && ((i + 1) == first_sack_index))
1525                         next_dup = &sp[i + 1];
1526
1527                 /* Event "B" in the comment above. */
1528                 if (after(end_seq, tp->high_seq))
1529                         flag |= FLAG_DATA_LOST;
1530
1531                 /* Skip too early cached blocks */
1532                 while (tcp_sack_cache_ok(tp, cache) &&
1533                        !before(start_seq, cache->end_seq))
1534                         cache++;
1535
1536                 /* Can skip some work by looking recv_sack_cache? */
1537                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1538                     after(end_seq, cache->start_seq)) {
1539
1540                         /* Head todo? */
1541                         if (before(start_seq, cache->start_seq)) {
1542                                 skb = tcp_sacktag_skip(skb, sk, start_seq,
1543                                                        &fack_count);
1544                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1545                                                        start_seq,
1546                                                        cache->start_seq,
1547                                                        dup_sack, &fack_count,
1548                                                        &reord, &flag);
1549                         }
1550
1551                         /* Rest of the block already fully processed? */
1552                         if (!after(end_seq, cache->end_seq))
1553                                 goto advance_sp;
1554
1555                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1556                                                        cache->end_seq,
1557                                                        &fack_count, &reord,
1558                                                        &flag);
1559
1560                         /* ...tail remains todo... */
1561                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1562                                 /* ...but better entrypoint exists! */
1563                                 skb = tcp_highest_sack(sk);
1564                                 if (skb == NULL)
1565                                         break;
1566                                 fack_count = tp->fackets_out;
1567                                 cache++;
1568                                 goto walk;
1569                         }
1570
1571                         skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
1572                                                &fack_count);
1573                         /* Check overlap against next cached too (past this one already) */
1574                         cache++;
1575                         continue;
1576                 }
1577
1578                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1579                         skb = tcp_highest_sack(sk);
1580                         if (skb == NULL)
1581                                 break;
1582                         fack_count = tp->fackets_out;
1583                 }
1584                 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
1585
1586 walk:
1587                 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1588                                        dup_sack, &fack_count, &reord, &flag);
1589
1590 advance_sp:
1591                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1592                  * due to in-order walk
1593                  */
1594                 if (after(end_seq, tp->frto_highmark))
1595                         flag &= ~FLAG_ONLY_ORIG_SACKED;
1596
1597                 i++;
1598         }
1599
1600         /* Clear the head of the cache sack blocks so we can skip it next time */
1601         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1602                 tp->recv_sack_cache[i].start_seq = 0;
1603                 tp->recv_sack_cache[i].end_seq = 0;
1604         }
1605         for (j = 0; j < used_sacks; j++)
1606                 tp->recv_sack_cache[i++] = sp[j];
1607
1608         tcp_mark_lost_retrans(sk);
1609
1610         tcp_verify_left_out(tp);
1611
1612         if ((reord < tp->fackets_out) &&
1613             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1614             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1615                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1616
1617 out:
1618
1619 #if FASTRETRANS_DEBUG > 0
1620         BUG_TRAP((int)tp->sacked_out >= 0);
1621         BUG_TRAP((int)tp->lost_out >= 0);
1622         BUG_TRAP((int)tp->retrans_out >= 0);
1623         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1624 #endif
1625         return flag;
1626 }
1627
1628 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1629  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1630  */
1631 int tcp_limit_reno_sacked(struct tcp_sock *tp)
1632 {
1633         u32 holes;
1634
1635         holes = max(tp->lost_out, 1U);
1636         holes = min(holes, tp->packets_out);
1637
1638         if ((tp->sacked_out + holes) > tp->packets_out) {
1639                 tp->sacked_out = tp->packets_out - holes;
1640                 return 1;
1641         }
1642         return 0;
1643 }
1644
1645 /* If we receive more dupacks than we expected counting segments
1646  * in assumption of absent reordering, interpret this as reordering.
1647  * The only another reason could be bug in receiver TCP.
1648  */
1649 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1650 {
1651         struct tcp_sock *tp = tcp_sk(sk);
1652         if (tcp_limit_reno_sacked(tp))
1653                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1654 }
1655
1656 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1657
1658 static void tcp_add_reno_sack(struct sock *sk)
1659 {
1660         struct tcp_sock *tp = tcp_sk(sk);
1661         tp->sacked_out++;
1662         tcp_check_reno_reordering(sk, 0);
1663         tcp_verify_left_out(tp);
1664 }
1665
1666 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1667
1668 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1669 {
1670         struct tcp_sock *tp = tcp_sk(sk);
1671
1672         if (acked > 0) {
1673                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1674                 if (acked - 1 >= tp->sacked_out)
1675                         tp->sacked_out = 0;
1676                 else
1677                         tp->sacked_out -= acked - 1;
1678         }
1679         tcp_check_reno_reordering(sk, acked);
1680         tcp_verify_left_out(tp);
1681 }
1682
1683 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1684 {
1685         tp->sacked_out = 0;
1686 }
1687
1688 /* F-RTO can only be used if TCP has never retransmitted anything other than
1689  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1690  */
1691 int tcp_use_frto(struct sock *sk)
1692 {
1693         const struct tcp_sock *tp = tcp_sk(sk);
1694         const struct inet_connection_sock *icsk = inet_csk(sk);
1695         struct sk_buff *skb;
1696
1697         if (!sysctl_tcp_frto)
1698                 return 0;
1699
1700         /* MTU probe and F-RTO won't really play nicely along currently */
1701         if (icsk->icsk_mtup.probe_size)
1702                 return 0;
1703
1704         if (IsSackFrto())
1705                 return 1;
1706
1707         /* Avoid expensive walking of rexmit queue if possible */
1708         if (tp->retrans_out > 1)
1709                 return 0;
1710
1711         skb = tcp_write_queue_head(sk);
1712         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1713         tcp_for_write_queue_from(skb, sk) {
1714                 if (skb == tcp_send_head(sk))
1715                         break;
1716                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1717                         return 0;
1718                 /* Short-circuit when first non-SACKed skb has been checked */
1719                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1720                         break;
1721         }
1722         return 1;
1723 }
1724
1725 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1726  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1727  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1728  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1729  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1730  * bits are handled if the Loss state is really to be entered (in
1731  * tcp_enter_frto_loss).
1732  *
1733  * Do like tcp_enter_loss() would; when RTO expires the second time it
1734  * does:
1735  *  "Reduce ssthresh if it has not yet been made inside this window."
1736  */
1737 void tcp_enter_frto(struct sock *sk)
1738 {
1739         const struct inet_connection_sock *icsk = inet_csk(sk);
1740         struct tcp_sock *tp = tcp_sk(sk);
1741         struct sk_buff *skb;
1742
1743         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1744             tp->snd_una == tp->high_seq ||
1745             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1746              !icsk->icsk_retransmits)) {
1747                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1748                 /* Our state is too optimistic in ssthresh() call because cwnd
1749                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1750                  * recovery has not yet completed. Pattern would be this: RTO,
1751                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1752                  * up here twice).
1753                  * RFC4138 should be more specific on what to do, even though
1754                  * RTO is quite unlikely to occur after the first Cumulative ACK
1755                  * due to back-off and complexity of triggering events ...
1756                  */
1757                 if (tp->frto_counter) {
1758                         u32 stored_cwnd;
1759                         stored_cwnd = tp->snd_cwnd;
1760                         tp->snd_cwnd = 2;
1761                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1762                         tp->snd_cwnd = stored_cwnd;
1763                 } else {
1764                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1765                 }
1766                 /* ... in theory, cong.control module could do "any tricks" in
1767                  * ssthresh(), which means that ca_state, lost bits and lost_out
1768                  * counter would have to be faked before the call occurs. We
1769                  * consider that too expensive, unlikely and hacky, so modules
1770                  * using these in ssthresh() must deal these incompatibility
1771                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1772                  */
1773                 tcp_ca_event(sk, CA_EVENT_FRTO);
1774         }
1775
1776         tp->undo_marker = tp->snd_una;
1777         tp->undo_retrans = 0;
1778
1779         skb = tcp_write_queue_head(sk);
1780         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1781                 tp->undo_marker = 0;
1782         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1783                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1784                 tp->retrans_out -= tcp_skb_pcount(skb);
1785         }
1786         tcp_verify_left_out(tp);
1787
1788         /* Too bad if TCP was application limited */
1789         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1790
1791         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1792          * The last condition is necessary at least in tp->frto_counter case.
1793          */
1794         if (IsSackFrto() && (tp->frto_counter ||
1795             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1796             after(tp->high_seq, tp->snd_una)) {
1797                 tp->frto_highmark = tp->high_seq;
1798         } else {
1799                 tp->frto_highmark = tp->snd_nxt;
1800         }
1801         tcp_set_ca_state(sk, TCP_CA_Disorder);
1802         tp->high_seq = tp->snd_nxt;
1803         tp->frto_counter = 1;
1804 }
1805
1806 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1807  * which indicates that we should follow the traditional RTO recovery,
1808  * i.e. mark everything lost and do go-back-N retransmission.
1809  */
1810 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1811 {
1812         struct tcp_sock *tp = tcp_sk(sk);
1813         struct sk_buff *skb;
1814
1815         tp->lost_out = 0;
1816         tp->retrans_out = 0;
1817         if (tcp_is_reno(tp))
1818                 tcp_reset_reno_sack(tp);
1819
1820         tcp_for_write_queue(skb, sk) {
1821                 if (skb == tcp_send_head(sk))
1822                         break;
1823
1824                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1825                 /*
1826                  * Count the retransmission made on RTO correctly (only when
1827                  * waiting for the first ACK and did not get it)...
1828                  */
1829                 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1830                         /* For some reason this R-bit might get cleared? */
1831                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1832                                 tp->retrans_out += tcp_skb_pcount(skb);
1833                         /* ...enter this if branch just for the first segment */
1834                         flag |= FLAG_DATA_ACKED;
1835                 } else {
1836                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1837                                 tp->undo_marker = 0;
1838                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1839                 }
1840
1841                 /* Don't lost mark skbs that were fwd transmitted after RTO */
1842                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) &&
1843                     !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1844                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1845                         tp->lost_out += tcp_skb_pcount(skb);
1846                 }
1847         }
1848         tcp_verify_left_out(tp);
1849
1850         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1851         tp->snd_cwnd_cnt = 0;
1852         tp->snd_cwnd_stamp = tcp_time_stamp;
1853         tp->frto_counter = 0;
1854         tp->bytes_acked = 0;
1855
1856         tp->reordering = min_t(unsigned int, tp->reordering,
1857                                sysctl_tcp_reordering);
1858         tcp_set_ca_state(sk, TCP_CA_Loss);
1859         tp->high_seq = tp->frto_highmark;
1860         TCP_ECN_queue_cwr(tp);
1861
1862         tcp_clear_retrans_hints_partial(tp);
1863 }
1864
1865 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1866 {
1867         tp->retrans_out = 0;
1868         tp->lost_out = 0;
1869
1870         tp->undo_marker = 0;
1871         tp->undo_retrans = 0;
1872 }
1873
1874 void tcp_clear_retrans(struct tcp_sock *tp)
1875 {
1876         tcp_clear_retrans_partial(tp);
1877
1878         tp->fackets_out = 0;
1879         tp->sacked_out = 0;
1880 }
1881
1882 /* Enter Loss state. If "how" is not zero, forget all SACK information
1883  * and reset tags completely, otherwise preserve SACKs. If receiver
1884  * dropped its ofo queue, we will know this due to reneging detection.
1885  */
1886 void tcp_enter_loss(struct sock *sk, int how)
1887 {
1888         const struct inet_connection_sock *icsk = inet_csk(sk);
1889         struct tcp_sock *tp = tcp_sk(sk);
1890         struct sk_buff *skb;
1891
1892         /* Reduce ssthresh if it has not yet been made inside this window. */
1893         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1894             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1895                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1896                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1897                 tcp_ca_event(sk, CA_EVENT_LOSS);
1898         }
1899         tp->snd_cwnd       = 1;
1900         tp->snd_cwnd_cnt   = 0;
1901         tp->snd_cwnd_stamp = tcp_time_stamp;
1902
1903         tp->bytes_acked = 0;
1904         tcp_clear_retrans_partial(tp);
1905
1906         if (tcp_is_reno(tp))
1907                 tcp_reset_reno_sack(tp);
1908
1909         if (!how) {
1910                 /* Push undo marker, if it was plain RTO and nothing
1911                  * was retransmitted. */
1912                 tp->undo_marker = tp->snd_una;
1913                 tcp_clear_retrans_hints_partial(tp);
1914         } else {
1915                 tp->sacked_out = 0;
1916                 tp->fackets_out = 0;
1917                 tcp_clear_all_retrans_hints(tp);
1918         }
1919
1920         tcp_for_write_queue(skb, sk) {
1921                 if (skb == tcp_send_head(sk))
1922                         break;
1923
1924                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1925                         tp->undo_marker = 0;
1926                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1927                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1928                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1929                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1930                         tp->lost_out += tcp_skb_pcount(skb);
1931                 }
1932         }
1933         tcp_verify_left_out(tp);
1934
1935         tp->reordering = min_t(unsigned int, tp->reordering,
1936                                sysctl_tcp_reordering);
1937         tcp_set_ca_state(sk, TCP_CA_Loss);
1938         tp->high_seq = tp->snd_nxt;
1939         TCP_ECN_queue_cwr(tp);
1940         /* Abort F-RTO algorithm if one is in progress */
1941         tp->frto_counter = 0;
1942 }
1943
1944 /* If ACK arrived pointing to a remembered SACK, it means that our
1945  * remembered SACKs do not reflect real state of receiver i.e.
1946  * receiver _host_ is heavily congested (or buggy).
1947  *
1948  * Do processing similar to RTO timeout.
1949  */
1950 static int tcp_check_sack_reneging(struct sock *sk, int flag)
1951 {
1952         if (flag & FLAG_SACK_RENEGING) {
1953                 struct inet_connection_sock *icsk = inet_csk(sk);
1954                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1955
1956                 tcp_enter_loss(sk, 1);
1957                 icsk->icsk_retransmits++;
1958                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1959                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1960                                           icsk->icsk_rto, TCP_RTO_MAX);
1961                 return 1;
1962         }
1963         return 0;
1964 }
1965
1966 static inline int tcp_fackets_out(struct tcp_sock *tp)
1967 {
1968         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1969 }
1970
1971 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1972  * counter when SACK is enabled (without SACK, sacked_out is used for
1973  * that purpose).
1974  *
1975  * Instead, with FACK TCP uses fackets_out that includes both SACKed
1976  * segments up to the highest received SACK block so far and holes in
1977  * between them.
1978  *
1979  * With reordering, holes may still be in flight, so RFC3517 recovery
1980  * uses pure sacked_out (total number of SACKed segments) even though
1981  * it violates the RFC that uses duplicate ACKs, often these are equal
1982  * but when e.g. out-of-window ACKs or packet duplication occurs,
1983  * they differ. Since neither occurs due to loss, TCP should really
1984  * ignore them.
1985  */
1986 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
1987 {
1988         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1989 }
1990
1991 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1992 {
1993         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1994 }
1995
1996 static inline int tcp_head_timedout(struct sock *sk)
1997 {
1998         struct tcp_sock *tp = tcp_sk(sk);
1999
2000         return tp->packets_out &&
2001                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2002 }
2003
2004 /* Linux NewReno/SACK/FACK/ECN state machine.
2005  * --------------------------------------
2006  *
2007  * "Open"       Normal state, no dubious events, fast path.
2008  * "Disorder"   In all the respects it is "Open",
2009  *              but requires a bit more attention. It is entered when
2010  *              we see some SACKs or dupacks. It is split of "Open"
2011  *              mainly to move some processing from fast path to slow one.
2012  * "CWR"        CWND was reduced due to some Congestion Notification event.
2013  *              It can be ECN, ICMP source quench, local device congestion.
2014  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2015  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2016  *
2017  * tcp_fastretrans_alert() is entered:
2018  * - each incoming ACK, if state is not "Open"
2019  * - when arrived ACK is unusual, namely:
2020  *      * SACK
2021  *      * Duplicate ACK.
2022  *      * ECN ECE.
2023  *
2024  * Counting packets in flight is pretty simple.
2025  *
2026  *      in_flight = packets_out - left_out + retrans_out
2027  *
2028  *      packets_out is SND.NXT-SND.UNA counted in packets.
2029  *
2030  *      retrans_out is number of retransmitted segments.
2031  *
2032  *      left_out is number of segments left network, but not ACKed yet.
2033  *
2034  *              left_out = sacked_out + lost_out
2035  *
2036  *     sacked_out: Packets, which arrived to receiver out of order
2037  *                 and hence not ACKed. With SACKs this number is simply
2038  *                 amount of SACKed data. Even without SACKs
2039  *                 it is easy to give pretty reliable estimate of this number,
2040  *                 counting duplicate ACKs.
2041  *
2042  *       lost_out: Packets lost by network. TCP has no explicit
2043  *                 "loss notification" feedback from network (for now).
2044  *                 It means that this number can be only _guessed_.
2045  *                 Actually, it is the heuristics to predict lossage that
2046  *                 distinguishes different algorithms.
2047  *
2048  *      F.e. after RTO, when all the queue is considered as lost,
2049  *      lost_out = packets_out and in_flight = retrans_out.
2050  *
2051  *              Essentially, we have now two algorithms counting
2052  *              lost packets.
2053  *
2054  *              FACK: It is the simplest heuristics. As soon as we decided
2055  *              that something is lost, we decide that _all_ not SACKed
2056  *              packets until the most forward SACK are lost. I.e.
2057  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2058  *              It is absolutely correct estimate, if network does not reorder
2059  *              packets. And it loses any connection to reality when reordering
2060  *              takes place. We use FACK by default until reordering
2061  *              is suspected on the path to this destination.
2062  *
2063  *              NewReno: when Recovery is entered, we assume that one segment
2064  *              is lost (classic Reno). While we are in Recovery and
2065  *              a partial ACK arrives, we assume that one more packet
2066  *              is lost (NewReno). This heuristics are the same in NewReno
2067  *              and SACK.
2068  *
2069  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2070  *  deflation etc. CWND is real congestion window, never inflated, changes
2071  *  only according to classic VJ rules.
2072  *
2073  * Really tricky (and requiring careful tuning) part of algorithm
2074  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2075  * The first determines the moment _when_ we should reduce CWND and,
2076  * hence, slow down forward transmission. In fact, it determines the moment
2077  * when we decide that hole is caused by loss, rather than by a reorder.
2078  *
2079  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2080  * holes, caused by lost packets.
2081  *
2082  * And the most logically complicated part of algorithm is undo
2083  * heuristics. We detect false retransmits due to both too early
2084  * fast retransmit (reordering) and underestimated RTO, analyzing
2085  * timestamps and D-SACKs. When we detect that some segments were
2086  * retransmitted by mistake and CWND reduction was wrong, we undo
2087  * window reduction and abort recovery phase. This logic is hidden
2088  * inside several functions named tcp_try_undo_<something>.
2089  */
2090
2091 /* This function decides, when we should leave Disordered state
2092  * and enter Recovery phase, reducing congestion window.
2093  *
2094  * Main question: may we further continue forward transmission
2095  * with the same cwnd?
2096  */
2097 static int tcp_time_to_recover(struct sock *sk)
2098 {
2099         struct tcp_sock *tp = tcp_sk(sk);
2100         __u32 packets_out;
2101
2102         /* Do not perform any recovery during F-RTO algorithm */
2103         if (tp->frto_counter)
2104                 return 0;
2105
2106         /* Trick#1: The loss is proven. */
2107         if (tp->lost_out)
2108                 return 1;
2109
2110         /* Not-A-Trick#2 : Classic rule... */
2111         if (tcp_dupack_heurestics(tp) > tp->reordering)
2112                 return 1;
2113
2114         /* Trick#3 : when we use RFC2988 timer restart, fast
2115          * retransmit can be triggered by timeout of queue head.
2116          */
2117         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2118                 return 1;
2119
2120         /* Trick#4: It is still not OK... But will it be useful to delay
2121          * recovery more?
2122          */
2123         packets_out = tp->packets_out;
2124         if (packets_out <= tp->reordering &&
2125             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2126             !tcp_may_send_now(sk)) {
2127                 /* We have nothing to send. This connection is limited
2128                  * either by receiver window or by application.
2129                  */
2130                 return 1;
2131         }
2132
2133         return 0;
2134 }
2135
2136 /* RFC: This is from the original, I doubt that this is necessary at all:
2137  * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2138  * retransmitted past LOST markings in the first place? I'm not fully sure
2139  * about undo and end of connection cases, which can cause R without L?
2140  */
2141 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
2142 {
2143         if ((tp->retransmit_skb_hint != NULL) &&
2144             before(TCP_SKB_CB(skb)->seq,
2145                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
2146                 tp->retransmit_skb_hint = NULL;
2147 }
2148
2149 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2150  * is against sacked "cnt", otherwise it's against facked "cnt"
2151  */
2152 static void tcp_mark_head_lost(struct sock *sk, int packets)
2153 {
2154         struct tcp_sock *tp = tcp_sk(sk);
2155         struct sk_buff *skb;
2156         int cnt, oldcnt;
2157         int err;
2158         unsigned int mss;
2159
2160         BUG_TRAP(packets <= tp->packets_out);
2161         if (tp->lost_skb_hint) {
2162                 skb = tp->lost_skb_hint;
2163                 cnt = tp->lost_cnt_hint;
2164         } else {
2165                 skb = tcp_write_queue_head(sk);
2166                 cnt = 0;
2167         }
2168
2169         tcp_for_write_queue_from(skb, sk) {
2170                 if (skb == tcp_send_head(sk))
2171                         break;
2172                 /* TODO: do this better */
2173                 /* this is not the most efficient way to do this... */
2174                 tp->lost_skb_hint = skb;
2175                 tp->lost_cnt_hint = cnt;
2176
2177                 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2178                         break;
2179
2180                 oldcnt = cnt;
2181                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2182                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2183                         cnt += tcp_skb_pcount(skb);
2184
2185                 if (cnt > packets) {
2186                         if (tcp_is_sack(tp) || (oldcnt >= packets))
2187                                 break;
2188
2189                         mss = skb_shinfo(skb)->gso_size;
2190                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2191                         if (err < 0)
2192                                 break;
2193                         cnt = packets;
2194                 }
2195
2196                 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2197                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2198                         tp->lost_out += tcp_skb_pcount(skb);
2199                         tcp_verify_retransmit_hint(tp, skb);
2200                 }
2201         }
2202         tcp_verify_left_out(tp);
2203 }
2204
2205 /* Account newly detected lost packet(s) */
2206
2207 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2208 {
2209         struct tcp_sock *tp = tcp_sk(sk);
2210
2211         if (tcp_is_reno(tp)) {
2212                 tcp_mark_head_lost(sk, 1);
2213         } else if (tcp_is_fack(tp)) {
2214                 int lost = tp->fackets_out - tp->reordering;
2215                 if (lost <= 0)
2216                         lost = 1;
2217                 tcp_mark_head_lost(sk, lost);
2218         } else {
2219                 int sacked_upto = tp->sacked_out - tp->reordering;
2220                 if (sacked_upto < fast_rexmit)
2221                         sacked_upto = fast_rexmit;
2222                 tcp_mark_head_lost(sk, sacked_upto);
2223         }
2224
2225         /* New heuristics: it is possible only after we switched
2226          * to restart timer each time when something is ACKed.
2227          * Hence, we can detect timed out packets during fast
2228          * retransmit without falling to slow start.
2229          */
2230         if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2231                 struct sk_buff *skb;
2232
2233                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2234                         : tcp_write_queue_head(sk);
2235
2236                 tcp_for_write_queue_from(skb, sk) {
2237                         if (skb == tcp_send_head(sk))
2238                                 break;
2239                         if (!tcp_skb_timedout(sk, skb))
2240                                 break;
2241
2242                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2243                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2244                                 tp->lost_out += tcp_skb_pcount(skb);
2245                                 tcp_verify_retransmit_hint(tp, skb);
2246                         }
2247                 }
2248
2249                 tp->scoreboard_skb_hint = skb;
2250
2251                 tcp_verify_left_out(tp);
2252         }
2253 }
2254
2255 /* CWND moderation, preventing bursts due to too big ACKs
2256  * in dubious situations.
2257  */
2258 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2259 {
2260         tp->snd_cwnd = min(tp->snd_cwnd,
2261                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2262         tp->snd_cwnd_stamp = tcp_time_stamp;
2263 }
2264
2265 /* Lower bound on congestion window is slow start threshold
2266  * unless congestion avoidance choice decides to overide it.
2267  */
2268 static inline u32 tcp_cwnd_min(const struct sock *sk)
2269 {
2270         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2271
2272         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2273 }
2274
2275 /* Decrease cwnd each second ack. */
2276 static void tcp_cwnd_down(struct sock *sk, int flag)
2277 {
2278         struct tcp_sock *tp = tcp_sk(sk);
2279         int decr = tp->snd_cwnd_cnt + 1;
2280
2281         if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2282             (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2283                 tp->snd_cwnd_cnt = decr & 1;
2284                 decr >>= 1;
2285
2286                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2287                         tp->snd_cwnd -= decr;
2288
2289                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2290                 tp->snd_cwnd_stamp = tcp_time_stamp;
2291         }
2292 }
2293
2294 /* Nothing was retransmitted or returned timestamp is less
2295  * than timestamp of the first retransmission.
2296  */
2297 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2298 {
2299         return !tp->retrans_stamp ||
2300                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2301                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2302 }
2303
2304 /* Undo procedures. */
2305
2306 #if FASTRETRANS_DEBUG > 1
2307 static void DBGUNDO(struct sock *sk, const char *msg)
2308 {
2309         struct tcp_sock *tp = tcp_sk(sk);
2310         struct inet_sock *inet = inet_sk(sk);
2311
2312         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2313                msg,
2314                NIPQUAD(inet->daddr), ntohs(inet->dport),
2315                tp->snd_cwnd, tcp_left_out(tp),
2316                tp->snd_ssthresh, tp->prior_ssthresh,
2317                tp->packets_out);
2318 }
2319 #else
2320 #define DBGUNDO(x...) do { } while (0)
2321 #endif
2322
2323 static void tcp_undo_cwr(struct sock *sk, const int undo)
2324 {
2325         struct tcp_sock *tp = tcp_sk(sk);
2326
2327         if (tp->prior_ssthresh) {
2328                 const struct inet_connection_sock *icsk = inet_csk(sk);
2329
2330                 if (icsk->icsk_ca_ops->undo_cwnd)
2331                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2332                 else
2333                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2334
2335                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2336                         tp->snd_ssthresh = tp->prior_ssthresh;
2337                         TCP_ECN_withdraw_cwr(tp);
2338                 }
2339         } else {
2340                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2341         }
2342         tcp_moderate_cwnd(tp);
2343         tp->snd_cwnd_stamp = tcp_time_stamp;
2344
2345         /* There is something screwy going on with the retrans hints after
2346            an undo */
2347         tcp_clear_all_retrans_hints(tp);
2348 }
2349
2350 static inline int tcp_may_undo(struct tcp_sock *tp)
2351 {
2352         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2353 }
2354
2355 /* People celebrate: "We love our President!" */
2356 static int tcp_try_undo_recovery(struct sock *sk)
2357 {
2358         struct tcp_sock *tp = tcp_sk(sk);
2359
2360         if (tcp_may_undo(tp)) {
2361                 /* Happy end! We did not retransmit anything
2362                  * or our original transmission succeeded.
2363                  */
2364                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2365                 tcp_undo_cwr(sk, 1);
2366                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2367                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2368                 else
2369                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2370                 tp->undo_marker = 0;
2371         }
2372         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2373                 /* Hold old state until something *above* high_seq
2374                  * is ACKed. For Reno it is MUST to prevent false
2375                  * fast retransmits (RFC2582). SACK TCP is safe. */
2376                 tcp_moderate_cwnd(tp);
2377                 return 1;
2378         }
2379         tcp_set_ca_state(sk, TCP_CA_Open);
2380         return 0;
2381 }
2382
2383 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2384 static void tcp_try_undo_dsack(struct sock *sk)
2385 {
2386         struct tcp_sock *tp = tcp_sk(sk);
2387
2388         if (tp->undo_marker && !tp->undo_retrans) {
2389                 DBGUNDO(sk, "D-SACK");
2390                 tcp_undo_cwr(sk, 1);
2391                 tp->undo_marker = 0;
2392                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2393         }
2394 }
2395
2396 /* Undo during fast recovery after partial ACK. */
2397
2398 static int tcp_try_undo_partial(struct sock *sk, int acked)
2399 {
2400         struct tcp_sock *tp = tcp_sk(sk);
2401         /* Partial ACK arrived. Force Hoe's retransmit. */
2402         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2403
2404         if (tcp_may_undo(tp)) {
2405                 /* Plain luck! Hole if filled with delayed
2406                  * packet, rather than with a retransmit.
2407                  */
2408                 if (tp->retrans_out == 0)
2409                         tp->retrans_stamp = 0;
2410
2411                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2412
2413                 DBGUNDO(sk, "Hoe");
2414                 tcp_undo_cwr(sk, 0);
2415                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2416
2417                 /* So... Do not make Hoe's retransmit yet.
2418                  * If the first packet was delayed, the rest
2419                  * ones are most probably delayed as well.
2420                  */
2421                 failed = 0;
2422         }
2423         return failed;
2424 }
2425
2426 /* Undo during loss recovery after partial ACK. */
2427 static int tcp_try_undo_loss(struct sock *sk)
2428 {
2429         struct tcp_sock *tp = tcp_sk(sk);
2430
2431         if (tcp_may_undo(tp)) {
2432                 struct sk_buff *skb;
2433                 tcp_for_write_queue(skb, sk) {
2434                         if (skb == tcp_send_head(sk))
2435                                 break;
2436                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2437                 }
2438
2439                 tcp_clear_all_retrans_hints(tp);
2440
2441                 DBGUNDO(sk, "partial loss");
2442                 tp->lost_out = 0;
2443                 tcp_undo_cwr(sk, 1);
2444                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2445                 inet_csk(sk)->icsk_retransmits = 0;
2446                 tp->undo_marker = 0;
2447                 if (tcp_is_sack(tp))
2448                         tcp_set_ca_state(sk, TCP_CA_Open);
2449                 return 1;
2450         }
2451         return 0;
2452 }
2453
2454 static inline void tcp_complete_cwr(struct sock *sk)
2455 {
2456         struct tcp_sock *tp = tcp_sk(sk);
2457         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2458         tp->snd_cwnd_stamp = tcp_time_stamp;
2459         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2460 }
2461
2462 static void tcp_try_to_open(struct sock *sk, int flag)
2463 {
2464         struct tcp_sock *tp = tcp_sk(sk);
2465
2466         tcp_verify_left_out(tp);
2467
2468         if (tp->retrans_out == 0)
2469                 tp->retrans_stamp = 0;
2470
2471         if (flag & FLAG_ECE)
2472                 tcp_enter_cwr(sk, 1);
2473
2474         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2475                 int state = TCP_CA_Open;
2476
2477                 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2478                         state = TCP_CA_Disorder;
2479
2480                 if (inet_csk(sk)->icsk_ca_state != state) {
2481                         tcp_set_ca_state(sk, state);
2482                         tp->high_seq = tp->snd_nxt;
2483                 }
2484                 tcp_moderate_cwnd(tp);
2485         } else {
2486                 tcp_cwnd_down(sk, flag);
2487         }
2488 }
2489
2490 static void tcp_mtup_probe_failed(struct sock *sk)
2491 {
2492         struct inet_connection_sock *icsk = inet_csk(sk);
2493
2494         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2495         icsk->icsk_mtup.probe_size = 0;
2496 }
2497
2498 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2499 {
2500         struct tcp_sock *tp = tcp_sk(sk);
2501         struct inet_connection_sock *icsk = inet_csk(sk);
2502
2503         /* FIXME: breaks with very large cwnd */
2504         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2505         tp->snd_cwnd = tp->snd_cwnd *
2506                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2507                        icsk->icsk_mtup.probe_size;
2508         tp->snd_cwnd_cnt = 0;
2509         tp->snd_cwnd_stamp = tcp_time_stamp;
2510         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2511
2512         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2513         icsk->icsk_mtup.probe_size = 0;
2514         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2515 }
2516
2517 /* Process an event, which can update packets-in-flight not trivially.
2518  * Main goal of this function is to calculate new estimate for left_out,
2519  * taking into account both packets sitting in receiver's buffer and
2520  * packets lost by network.
2521  *
2522  * Besides that it does CWND reduction, when packet loss is detected
2523  * and changes state of machine.
2524  *
2525  * It does _not_ decide what to send, it is made in function
2526  * tcp_xmit_retransmit_queue().
2527  */
2528 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2529 {
2530         struct inet_connection_sock *icsk = inet_csk(sk);
2531         struct tcp_sock *tp = tcp_sk(sk);
2532         int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2533         int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2534                                     (tcp_fackets_out(tp) > tp->reordering));
2535         int fast_rexmit = 0;
2536
2537         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2538                 tp->sacked_out = 0;
2539         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2540                 tp->fackets_out = 0;
2541
2542         /* Now state machine starts.
2543          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2544         if (flag & FLAG_ECE)
2545                 tp->prior_ssthresh = 0;
2546
2547         /* B. In all the states check for reneging SACKs. */
2548         if (tcp_check_sack_reneging(sk, flag))
2549                 return;
2550
2551         /* C. Process data loss notification, provided it is valid. */
2552         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2553             before(tp->snd_una, tp->high_seq) &&
2554             icsk->icsk_ca_state != TCP_CA_Open &&
2555             tp->fackets_out > tp->reordering) {
2556                 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2557                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2558         }
2559
2560         /* D. Check consistency of the current state. */
2561         tcp_verify_left_out(tp);
2562
2563         /* E. Check state exit conditions. State can be terminated
2564          *    when high_seq is ACKed. */
2565         if (icsk->icsk_ca_state == TCP_CA_Open) {
2566                 BUG_TRAP(tp->retrans_out == 0);
2567                 tp->retrans_stamp = 0;
2568         } else if (!before(tp->snd_una, tp->high_seq)) {
2569                 switch (icsk->icsk_ca_state) {
2570                 case TCP_CA_Loss:
2571                         icsk->icsk_retransmits = 0;
2572                         if (tcp_try_undo_recovery(sk))
2573                                 return;
2574                         break;
2575
2576                 case TCP_CA_CWR:
2577                         /* CWR is to be held something *above* high_seq
2578                          * is ACKed for CWR bit to reach receiver. */
2579                         if (tp->snd_una != tp->high_seq) {
2580                                 tcp_complete_cwr(sk);
2581                                 tcp_set_ca_state(sk, TCP_CA_Open);
2582                         }
2583                         break;
2584
2585                 case TCP_CA_Disorder:
2586                         tcp_try_undo_dsack(sk);
2587                         if (!tp->undo_marker ||
2588                             /* For SACK case do not Open to allow to undo
2589                              * catching for all duplicate ACKs. */
2590                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2591                                 tp->undo_marker = 0;
2592                                 tcp_set_ca_state(sk, TCP_CA_Open);
2593                         }
2594                         break;
2595
2596                 case TCP_CA_Recovery:
2597                         if (tcp_is_reno(tp))
2598                                 tcp_reset_reno_sack(tp);
2599                         if (tcp_try_undo_recovery(sk))
2600                                 return;
2601                         tcp_complete_cwr(sk);
2602                         break;
2603                 }
2604         }
2605
2606         /* F. Process state. */
2607         switch (icsk->icsk_ca_state) {
2608         case TCP_CA_Recovery:
2609                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2610                         if (tcp_is_reno(tp) && is_dupack)
2611                                 tcp_add_reno_sack(sk);
2612                 } else
2613                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2614                 break;
2615         case TCP_CA_Loss:
2616                 if (flag & FLAG_DATA_ACKED)
2617                         icsk->icsk_retransmits = 0;
2618                 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2619                         tcp_reset_reno_sack(tp);
2620                 if (!tcp_try_undo_loss(sk)) {
2621                         tcp_moderate_cwnd(tp);
2622                         tcp_xmit_retransmit_queue(sk);
2623                         return;
2624                 }
2625                 if (icsk->icsk_ca_state != TCP_CA_Open)
2626                         return;
2627                 /* Loss is undone; fall through to processing in Open state. */
2628         default:
2629                 if (tcp_is_reno(tp)) {
2630                         if (flag & FLAG_SND_UNA_ADVANCED)
2631                                 tcp_reset_reno_sack(tp);
2632                         if (is_dupack)
2633                                 tcp_add_reno_sack(sk);
2634                 }
2635
2636                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2637                         tcp_try_undo_dsack(sk);
2638
2639                 if (!tcp_time_to_recover(sk)) {
2640                         tcp_try_to_open(sk, flag);
2641                         return;
2642                 }
2643
2644                 /* MTU probe failure: don't reduce cwnd */
2645                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2646                     icsk->icsk_mtup.probe_size &&
2647                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2648                         tcp_mtup_probe_failed(sk);
2649                         /* Restores the reduction we did in tcp_mtup_probe() */
2650                         tp->snd_cwnd++;
2651                         tcp_simple_retransmit(sk);
2652                         return;
2653                 }
2654
2655                 /* Otherwise enter Recovery state */
2656
2657                 if (tcp_is_reno(tp))
2658                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2659                 else
2660                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2661
2662                 tp->high_seq = tp->snd_nxt;
2663                 tp->prior_ssthresh = 0;
2664                 tp->undo_marker = tp->snd_una;
2665                 tp->undo_retrans = tp->retrans_out;
2666
2667                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2668                         if (!(flag & FLAG_ECE))
2669                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2670                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2671                         TCP_ECN_queue_cwr(tp);
2672                 }
2673
2674                 tp->bytes_acked = 0;
2675                 tp->snd_cwnd_cnt = 0;
2676                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2677                 fast_rexmit = 1;
2678         }
2679
2680         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2681                 tcp_update_scoreboard(sk, fast_rexmit);
2682         tcp_cwnd_down(sk, flag);
2683         tcp_xmit_retransmit_queue(sk);
2684 }
2685
2686 /* Read draft-ietf-tcplw-high-performance before mucking
2687  * with this code. (Supersedes RFC1323)
2688  */
2689 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2690 {
2691         /* RTTM Rule: A TSecr value received in a segment is used to
2692          * update the averaged RTT measurement only if the segment
2693          * acknowledges some new data, i.e., only if it advances the
2694          * left edge of the send window.
2695          *
2696          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2697          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2698          *
2699          * Changed: reset backoff as soon as we see the first valid sample.
2700          * If we do not, we get strongly overestimated rto. With timestamps
2701          * samples are accepted even from very old segments: f.e., when rtt=1
2702          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2703          * answer arrives rto becomes 120 seconds! If at least one of segments
2704          * in window is lost... Voila.                          --ANK (010210)
2705          */
2706         struct tcp_sock *tp = tcp_sk(sk);
2707         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2708         tcp_rtt_estimator(sk, seq_rtt);
2709         tcp_set_rto(sk);
2710         inet_csk(sk)->icsk_backoff = 0;
2711         tcp_bound_rto(sk);
2712 }
2713
2714 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2715 {
2716         /* We don't have a timestamp. Can only use
2717          * packets that are not retransmitted to determine
2718          * rtt estimates. Also, we must not reset the
2719          * backoff for rto until we get a non-retransmitted
2720          * packet. This allows us to deal with a situation
2721          * where the network delay has increased suddenly.
2722          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2723          */
2724
2725         if (flag & FLAG_RETRANS_DATA_ACKED)
2726                 return;
2727
2728         tcp_rtt_estimator(sk, seq_rtt);
2729         tcp_set_rto(sk);
2730         inet_csk(sk)->icsk_backoff = 0;
2731         tcp_bound_rto(sk);
2732 }
2733
2734 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2735                                       const s32 seq_rtt)
2736 {
2737         const struct tcp_sock *tp = tcp_sk(sk);
2738         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2739         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2740                 tcp_ack_saw_tstamp(sk, flag);
2741         else if (seq_rtt >= 0)
2742                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2743 }
2744
2745 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2746 {
2747         const struct inet_connection_sock *icsk = inet_csk(sk);
2748         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2749         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2750 }
2751
2752 /* Restart timer after forward progress on connection.
2753  * RFC2988 recommends to restart timer to now+rto.
2754  */
2755 static void tcp_rearm_rto(struct sock *sk)
2756 {
2757         struct tcp_sock *tp = tcp_sk(sk);
2758
2759         if (!tp->packets_out) {
2760                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2761         } else {
2762                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2763                                           inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2764         }
2765 }
2766
2767 /* If we get here, the whole TSO packet has not been acked. */
2768 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2769 {
2770         struct tcp_sock *tp = tcp_sk(sk);
2771         u32 packets_acked;
2772
2773         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2774
2775         packets_acked = tcp_skb_pcount(skb);
2776         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2777                 return 0;
2778         packets_acked -= tcp_skb_pcount(skb);
2779
2780         if (packets_acked) {
2781                 BUG_ON(tcp_skb_pcount(skb) == 0);
2782                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2783         }
2784
2785         return packets_acked;
2786 }
2787
2788 /* Remove acknowledged frames from the retransmission queue. If our packet
2789  * is before the ack sequence we can discard it as it's confirmed to have
2790  * arrived at the other end.
2791  */
2792 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
2793 {
2794         struct tcp_sock *tp = tcp_sk(sk);
2795         const struct inet_connection_sock *icsk = inet_csk(sk);
2796         struct sk_buff *skb;
2797         u32 now = tcp_time_stamp;
2798         int fully_acked = 1;
2799         int flag = 0;
2800         u32 pkts_acked = 0;
2801         u32 reord = tp->packets_out;
2802         s32 seq_rtt = -1;
2803         s32 ca_seq_rtt = -1;
2804         ktime_t last_ackt = net_invalid_timestamp();
2805
2806         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2807                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2808                 u32 end_seq;
2809                 u32 acked_pcount;
2810                 u8 sacked = scb->sacked;
2811
2812                 /* Determine how many packets and what bytes were acked, tso and else */
2813                 if (after(scb->end_seq, tp->snd_una)) {
2814                         if (tcp_skb_pcount(skb) == 1 ||
2815                             !after(tp->snd_una, scb->seq))
2816                                 break;
2817
2818                         acked_pcount = tcp_tso_acked(sk, skb);
2819                         if (!acked_pcount)
2820                                 break;
2821
2822                         fully_acked = 0;
2823                         end_seq = tp->snd_una;
2824                 } else {
2825                         acked_pcount = tcp_skb_pcount(skb);
2826                         end_seq = scb->end_seq;
2827                 }
2828
2829                 /* MTU probing checks */
2830                 if (fully_acked && icsk->icsk_mtup.probe_size &&
2831                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2832                         tcp_mtup_probe_success(sk, skb);
2833                 }
2834
2835                 if (sacked & TCPCB_RETRANS) {
2836                         if (sacked & TCPCB_SACKED_RETRANS)
2837                                 tp->retrans_out -= acked_pcount;
2838                         flag |= FLAG_RETRANS_DATA_ACKED;
2839                         ca_seq_rtt = -1;
2840                         seq_rtt = -1;
2841                         if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
2842                                 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2843                 } else {
2844                         ca_seq_rtt = now - scb->when;
2845                         last_ackt = skb->tstamp;
2846                         if (seq_rtt < 0) {
2847                                 seq_rtt = ca_seq_rtt;
2848                         }
2849                         if (!(sacked & TCPCB_SACKED_ACKED))
2850                                 reord = min(pkts_acked, reord);
2851                 }
2852
2853                 if (sacked & TCPCB_SACKED_ACKED)
2854                         tp->sacked_out -= acked_pcount;
2855                 if (sacked & TCPCB_LOST)
2856                         tp->lost_out -= acked_pcount;
2857
2858                 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up)))
2859                         tp->urg_mode = 0;
2860
2861                 tp->packets_out -= acked_pcount;
2862                 pkts_acked += acked_pcount;
2863
2864                 /* Initial outgoing SYN's get put onto the write_queue
2865                  * just like anything else we transmit.  It is not
2866                  * true data, and if we misinform our callers that
2867                  * this ACK acks real data, we will erroneously exit
2868                  * connection startup slow start one packet too
2869                  * quickly.  This is severely frowned upon behavior.
2870                  */
2871                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2872                         flag |= FLAG_DATA_ACKED;
2873                 } else {
2874                         flag |= FLAG_SYN_ACKED;
2875                         tp->retrans_stamp = 0;
2876                 }
2877
2878                 if (!fully_acked)
2879                         break;
2880
2881                 tcp_unlink_write_queue(skb, sk);
2882                 sk_wmem_free_skb(sk, skb);
2883                 tcp_clear_all_retrans_hints(tp);
2884         }
2885
2886         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2887                 flag |= FLAG_SACK_RENEGING;
2888
2889         if (flag & FLAG_ACKED) {
2890                 const struct tcp_congestion_ops *ca_ops
2891                         = inet_csk(sk)->icsk_ca_ops;
2892
2893                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2894                 tcp_rearm_rto(sk);
2895
2896                 if (tcp_is_reno(tp)) {
2897                         tcp_remove_reno_sacks(sk, pkts_acked);
2898                 } else {
2899                         /* Non-retransmitted hole got filled? That's reordering */
2900                         if (reord < prior_fackets)
2901                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2902                 }
2903
2904                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2905
2906                 if (ca_ops->pkts_acked) {
2907                         s32 rtt_us = -1;
2908
2909                         /* Is the ACK triggering packet unambiguous? */
2910                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2911                                 /* High resolution needed and available? */
2912                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2913                                     !ktime_equal(last_ackt,
2914                                                  net_invalid_timestamp()))
2915                                         rtt_us = ktime_us_delta(ktime_get_real(),
2916                                                                 last_ackt);
2917                                 else if (ca_seq_rtt > 0)
2918                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
2919                         }
2920
2921                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2922                 }
2923         }
2924
2925 #if FASTRETRANS_DEBUG > 0
2926         BUG_TRAP((int)tp->sacked_out >= 0);
2927         BUG_TRAP((int)tp->lost_out >= 0);
2928         BUG_TRAP((int)tp->retrans_out >= 0);
2929         if (!tp->packets_out && tcp_is_sack(tp)) {
2930                 icsk = inet_csk(sk);
2931                 if (tp->lost_out) {
2932                         printk(KERN_DEBUG "Leak l=%u %d\n",
2933                                tp->lost_out, icsk->icsk_ca_state);
2934                         tp->lost_out = 0;
2935                 }
2936                 if (tp->sacked_out) {
2937                         printk(KERN_DEBUG "Leak s=%u %d\n",
2938                                tp->sacked_out, icsk->icsk_ca_state);
2939                         tp->sacked_out = 0;
2940                 }
2941                 if (tp->retrans_out) {
2942                         printk(KERN_DEBUG "Leak r=%u %d\n",
2943                                tp->retrans_out, icsk->icsk_ca_state);
2944                         tp->retrans_out = 0;
2945                 }
2946         }
2947 #endif
2948         return flag;
2949 }
2950
2951 static void tcp_ack_probe(struct sock *sk)
2952 {
2953         const struct tcp_sock *tp = tcp_sk(sk);
2954         struct inet_connection_sock *icsk = inet_csk(sk);
2955
2956         /* Was it a usable window open? */
2957
2958         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
2959                 icsk->icsk_backoff = 0;
2960                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2961                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2962                  * This function is not for random using!
2963                  */
2964         } else {
2965                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2966                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2967                                           TCP_RTO_MAX);
2968         }
2969 }
2970
2971 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2972 {
2973         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2974                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2975 }
2976
2977 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2978 {
2979         const struct tcp_sock *tp = tcp_sk(sk);
2980         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2981                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2982 }
2983
2984 /* Check that window update is acceptable.
2985  * The function assumes that snd_una<=ack<=snd_next.
2986  */
2987 static inline int tcp_may_update_window(const struct tcp_sock *tp,
2988                                         const u32 ack, const u32 ack_seq,
2989                                         const u32 nwin)
2990 {
2991         return (after(ack, tp->snd_una) ||
2992                 after(ack_seq, tp->snd_wl1) ||
2993                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2994 }
2995
2996 /* Update our send window.
2997  *
2998  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2999  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3000  */
3001 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3002                                  u32 ack_seq)
3003 {
3004         struct tcp_sock *tp = tcp_sk(sk);
3005         int flag = 0;
3006         u32 nwin = ntohs(tcp_hdr(skb)->window);
3007
3008         if (likely(!tcp_hdr(skb)->syn))
3009                 nwin <<= tp->rx_opt.snd_wscale;
3010
3011         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3012                 flag |= FLAG_WIN_UPDATE;
3013                 tcp_update_wl(tp, ack, ack_seq);
3014
3015                 if (tp->snd_wnd != nwin) {
3016                         tp->snd_wnd = nwin;
3017
3018                         /* Note, it is the only place, where
3019                          * fast path is recovered for sending TCP.
3020                          */
3021                         tp->pred_flags = 0;
3022                         tcp_fast_path_check(sk);
3023
3024                         if (nwin > tp->max_window) {
3025                                 tp->max_window = nwin;
3026                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3027                         }
3028                 }
3029         }
3030
3031         tp->snd_una = ack;
3032
3033         return flag;
3034 }
3035
3036 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3037  * continue in congestion avoidance.
3038  */
3039 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3040 {
3041         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3042         tp->snd_cwnd_cnt = 0;
3043         tp->bytes_acked = 0;
3044         TCP_ECN_queue_cwr(tp);
3045         tcp_moderate_cwnd(tp);
3046 }
3047
3048 /* A conservative spurious RTO response algorithm: reduce cwnd using
3049  * rate halving and continue in congestion avoidance.
3050  */
3051 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3052 {
3053         tcp_enter_cwr(sk, 0);
3054 }
3055
3056 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3057 {
3058         if (flag & FLAG_ECE)
3059                 tcp_ratehalving_spur_to_response(sk);
3060         else
3061                 tcp_undo_cwr(sk, 1);
3062 }
3063
3064 /* F-RTO spurious RTO detection algorithm (RFC4138)
3065  *
3066  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3067  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3068  * window (but not to or beyond highest sequence sent before RTO):
3069  *   On First ACK,  send two new segments out.
3070  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3071  *                  algorithm is not part of the F-RTO detection algorithm
3072  *                  given in RFC4138 but can be selected separately).
3073  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3074  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3075  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3076  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3077  *
3078  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3079  * original window even after we transmit two new data segments.
3080  *
3081  * SACK version:
3082  *   on first step, wait until first cumulative ACK arrives, then move to
3083  *   the second step. In second step, the next ACK decides.
3084  *
3085  * F-RTO is implemented (mainly) in four functions:
3086  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3087  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3088  *     called when tcp_use_frto() showed green light
3089  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3090  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3091  *     to prove that the RTO is indeed spurious. It transfers the control
3092  *     from F-RTO to the conventional RTO recovery
3093  */
3094 static int tcp_process_frto(struct sock *sk, int flag)
3095 {
3096         struct tcp_sock *tp = tcp_sk(sk);
3097
3098         tcp_verify_left_out(tp);
3099
3100         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3101         if (flag & FLAG_DATA_ACKED)
3102                 inet_csk(sk)->icsk_retransmits = 0;
3103
3104         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3105             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3106                 tp->undo_marker = 0;
3107
3108         if (!before(tp->snd_una, tp->frto_highmark)) {
3109                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3110                 return 1;
3111         }
3112
3113         if (!IsSackFrto() || tcp_is_reno(tp)) {
3114                 /* RFC4138 shortcoming in step 2; should also have case c):
3115                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3116                  * data, winupdate
3117                  */
3118                 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3119                         return 1;
3120
3121                 if (!(flag & FLAG_DATA_ACKED)) {
3122                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3123                                             flag);
3124                         return 1;
3125                 }
3126         } else {
3127                 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3128                         /* Prevent sending of new data. */
3129                         tp->snd_cwnd = min(tp->snd_cwnd,
3130                                            tcp_packets_in_flight(tp));
3131                         return 1;
3132                 }
3133
3134                 if ((tp->frto_counter >= 2) &&
3135                     (!(flag & FLAG_FORWARD_PROGRESS) ||
3136                      ((flag & FLAG_DATA_SACKED) &&
3137                       !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3138                         /* RFC4138 shortcoming (see comment above) */
3139                         if (!(flag & FLAG_FORWARD_PROGRESS) &&
3140                             (flag & FLAG_NOT_DUP))
3141                                 return 1;
3142
3143                         tcp_enter_frto_loss(sk, 3, flag);
3144                         return 1;
3145                 }
3146         }
3147
3148         if (tp->frto_counter == 1) {
3149                 /* tcp_may_send_now needs to see updated state */
3150                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3151                 tp->frto_counter = 2;
3152
3153                 if (!tcp_may_send_now(sk))
3154                         tcp_enter_frto_loss(sk, 2, flag);
3155
3156                 return 1;
3157         } else {
3158                 switch (sysctl_tcp_frto_response) {
3159                 case 2:
3160                         tcp_undo_spur_to_response(sk, flag);
3161                         break;
3162                 case 1:
3163                         tcp_conservative_spur_to_response(tp);
3164                         break;
3165                 default:
3166                         tcp_ratehalving_spur_to_response(sk);
3167                         break;
3168                 }
3169                 tp->frto_counter = 0;
3170                 tp->undo_marker = 0;
3171                 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
3172         }
3173         return 0;
3174 }
3175
3176 /* This routine deals with incoming acks, but not outgoing ones. */
3177 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3178 {
3179         struct inet_connection_sock *icsk = inet_csk(sk);
3180         struct tcp_sock *tp = tcp_sk(sk);
3181         u32 prior_snd_una = tp->snd_una;
3182         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3183         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3184         u32 prior_in_flight;
3185         u32 prior_fackets;
3186         int prior_packets;
3187         int frto_cwnd = 0;
3188
3189         /* If the ack is newer than sent or older than previous acks
3190          * then we can probably ignore it.
3191          */
3192         if (after(ack, tp->snd_nxt))
3193                 goto uninteresting_ack;
3194
3195         if (before(ack, prior_snd_una))
3196                 goto old_ack;
3197
3198         if (after(ack, prior_snd_una))
3199                 flag |= FLAG_SND_UNA_ADVANCED;
3200
3201         if (sysctl_tcp_abc) {
3202                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3203                         tp->bytes_acked += ack - prior_snd_una;
3204                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3205                         /* we assume just one segment left network */
3206                         tp->bytes_acked += min(ack - prior_snd_una,
3207                                                tp->mss_cache);
3208         }
3209
3210         prior_fackets = tp->fackets_out;
3211         prior_in_flight = tcp_packets_in_flight(tp);
3212
3213         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3214                 /* Window is constant, pure forward advance.
3215                  * No more checks are required.
3216                  * Note, we use the fact that SND.UNA>=SND.WL2.
3217                  */
3218                 tcp_update_wl(tp, ack, ack_seq);
3219                 tp->snd_una = ack;
3220                 flag |= FLAG_WIN_UPDATE;
3221
3222                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3223
3224                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3225         } else {
3226                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3227                         flag |= FLAG_DATA;
3228                 else
3229                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3230
3231                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3232
3233                 if (TCP_SKB_CB(skb)->sacked)
3234                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3235
3236                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3237                         flag |= FLAG_ECE;
3238
3239                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3240         }
3241
3242         /* We passed data and got it acked, remove any soft error
3243          * log. Something worked...
3244          */
3245         sk->sk_err_soft = 0;
3246         tp->rcv_tstamp = tcp_time_stamp;
3247         prior_packets = tp->packets_out;
3248         if (!prior_packets)
3249                 goto no_queue;
3250
3251         /* See if we can take anything off of the retransmit queue. */
3252         flag |= tcp_clean_rtx_queue(sk, prior_fackets);
3253
3254         if (tp->frto_counter)
3255                 frto_cwnd = tcp_process_frto(sk, flag);
3256         /* Guarantee sacktag reordering detection against wrap-arounds */
3257         if (before(tp->frto_highmark, tp->snd_una))
3258                 tp->frto_highmark = 0;
3259
3260         if (tcp_ack_is_dubious(sk, flag)) {
3261                 /* Advance CWND, if state allows this. */
3262                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3263                     tcp_may_raise_cwnd(sk, flag))
3264                         tcp_cong_avoid(sk, ack, prior_in_flight);
3265                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3266                                       flag);
3267         } else {
3268                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3269                         tcp_cong_avoid(sk, ack, prior_in_flight);
3270         }
3271
3272         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3273                 dst_confirm(sk->sk_dst_cache);
3274
3275         return 1;
3276
3277 no_queue:
3278         icsk->icsk_probes_out = 0;
3279
3280         /* If this ack opens up a zero window, clear backoff.  It was
3281          * being used to time the probes, and is probably far higher than
3282          * it needs to be for normal retransmission.
3283          */
3284         if (tcp_send_head(sk))
3285                 tcp_ack_probe(sk);
3286         return 1;
3287
3288 old_ack:
3289         if (TCP_SKB_CB(skb)->sacked)
3290                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3291
3292 uninteresting_ack:
3293         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3294         return 0;
3295 }
3296
3297 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3298  * But, this can also be called on packets in the established flow when
3299  * the fast version below fails.
3300  */
3301 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3302                        int estab)
3303 {
3304         unsigned char *ptr;
3305         struct tcphdr *th = tcp_hdr(skb);
3306         int length = (th->doff * 4) - sizeof(struct tcphdr);
3307
3308         ptr = (unsigned char *)(th + 1);
3309         opt_rx->saw_tstamp = 0;
3310
3311         while (length > 0) {
3312                 int opcode = *ptr++;
3313                 int opsize;
3314
3315                 switch (opcode) {
3316                 case TCPOPT_EOL:
3317                         return;
3318                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3319                         length--;
3320                         continue;
3321                 default:
3322                         opsize = *ptr++;
3323                         if (opsize < 2) /* "silly options" */
3324                                 return;
3325                         if (opsize > length)
3326                                 return; /* don't parse partial options */
3327                         switch (opcode) {
3328                         case TCPOPT_MSS:
3329                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3330                                         u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3331                                         if (in_mss) {
3332                                                 if (opt_rx->user_mss &&
3333                                                     opt_rx->user_mss < in_mss)
3334                                                         in_mss = opt_rx->user_mss;
3335                                                 opt_rx->mss_clamp = in_mss;
3336                                         }
3337                                 }
3338                                 break;
3339                         case TCPOPT_WINDOW:
3340                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3341                                     !estab && sysctl_tcp_window_scaling) {
3342                                         __u8 snd_wscale = *(__u8 *)ptr;
3343                                         opt_rx->wscale_ok = 1;
3344                                         if (snd_wscale > 14) {
3345                                                 if (net_ratelimit())
3346                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3347                                                                "scaling value %d >14 received.\n",
3348                                                                snd_wscale);
3349                                                 snd_wscale = 14;
3350                                         }
3351                                         opt_rx->snd_wscale = snd_wscale;
3352                                 }
3353                                 break;
3354                         case TCPOPT_TIMESTAMP:
3355                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3356                                     ((estab && opt_rx->tstamp_ok) ||
3357                                      (!estab && sysctl_tcp_timestamps))) {
3358                                         opt_rx->saw_tstamp = 1;
3359                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3360                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3361                                 }
3362                                 break;
3363                         case TCPOPT_SACK_PERM:
3364                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3365                                     !estab && sysctl_tcp_sack) {
3366                                         opt_rx->sack_ok = 1;
3367                                         tcp_sack_reset(opt_rx);
3368                                 }
3369                                 break;
3370
3371                         case TCPOPT_SACK:
3372                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3373                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3374                                    opt_rx->sack_ok) {
3375                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3376                                 }
3377                                 break;
3378 #ifdef CONFIG_TCP_MD5SIG
3379                         case TCPOPT_MD5SIG:
3380                                 /*
3381                                  * The MD5 Hash has already been
3382                                  * checked (see tcp_v{4,6}_do_rcv()).
3383                                  */
3384                                 break;
3385 #endif
3386                         }
3387
3388                         ptr += opsize-2;
3389                         length -= opsize;
3390                 }
3391         }
3392 }
3393
3394 /* Fast parse options. This hopes to only see timestamps.
3395  * If it is wrong it falls back on tcp_parse_options().
3396  */
3397 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3398                                   struct tcp_sock *tp)
3399 {
3400         if (th->doff == sizeof(struct tcphdr) >> 2) {
3401                 tp->rx_opt.saw_tstamp = 0;
3402                 return 0;
3403         } else if (tp->rx_opt.tstamp_ok &&
3404                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3405                 __be32 *ptr = (__be32 *)(th + 1);
3406                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3407                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3408                         tp->rx_opt.saw_tstamp = 1;
3409                         ++ptr;
3410                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3411                         ++ptr;
3412                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3413                         return 1;
3414                 }
3415         }
3416         tcp_parse_options(skb, &tp->rx_opt, 1);
3417         return 1;
3418 }
3419
3420 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3421 {
3422         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3423         tp->rx_opt.ts_recent_stamp = get_seconds();
3424 }
3425
3426 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3427 {
3428         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3429                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3430                  * extra check below makes sure this can only happen
3431                  * for pure ACK frames.  -DaveM
3432                  *
3433                  * Not only, also it occurs for expired timestamps.
3434                  */
3435
3436                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3437                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3438                         tcp_store_ts_recent(tp);
3439         }
3440 }
3441
3442 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3443  *
3444  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3445  * it can pass through stack. So, the following predicate verifies that
3446  * this segment is not used for anything but congestion avoidance or
3447  * fast retransmit. Moreover, we even are able to eliminate most of such
3448  * second order effects, if we apply some small "replay" window (~RTO)
3449  * to timestamp space.
3450  *
3451  * All these measures still do not guarantee that we reject wrapped ACKs
3452  * on networks with high bandwidth, when sequence space is recycled fastly,
3453  * but it guarantees that such events will be very rare and do not affect
3454  * connection seriously. This doesn't look nice, but alas, PAWS is really
3455  * buggy extension.
3456  *
3457  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3458  * states that events when retransmit arrives after original data are rare.
3459  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3460  * the biggest problem on large power networks even with minor reordering.
3461  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3462  * up to bandwidth of 18Gigabit/sec. 8) ]
3463  */
3464
3465 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3466 {
3467         struct tcp_sock *tp = tcp_sk(sk);
3468         struct tcphdr *th = tcp_hdr(skb);
3469         u32 seq = TCP_SKB_CB(skb)->seq;
3470         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3471
3472         return (/* 1. Pure ACK with correct sequence number. */
3473                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3474
3475                 /* 2. ... and duplicate ACK. */
3476                 ack == tp->snd_una &&
3477
3478                 /* 3. ... and does not update window. */
3479                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3480
3481                 /* 4. ... and sits in replay window. */
3482                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3483 }
3484
3485 static inline int tcp_paws_discard(const struct sock *sk,
3486                                    const struct sk_buff *skb)
3487 {
3488         const struct tcp_sock *tp = tcp_sk(sk);
3489         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3490                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3491                 !tcp_disordered_ack(sk, skb));
3492 }
3493
3494 /* Check segment sequence number for validity.
3495  *
3496  * Segment controls are considered valid, if the segment
3497  * fits to the window after truncation to the window. Acceptability
3498  * of data (and SYN, FIN, of course) is checked separately.
3499  * See tcp_data_queue(), for example.
3500  *
3501  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3502  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3503  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3504  * (borrowed from freebsd)
3505  */
3506
3507 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3508 {
3509         return  !before(end_seq, tp->rcv_wup) &&
3510                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3511 }
3512
3513 /* When we get a reset we do this. */
3514 static void tcp_reset(struct sock *sk)
3515 {
3516         /* We want the right error as BSD sees it (and indeed as we do). */
3517         switch (sk->sk_state) {
3518         case TCP_SYN_SENT:
3519                 sk->sk_err = ECONNREFUSED;
3520                 break;
3521         case TCP_CLOSE_WAIT:
3522                 sk->sk_err = EPIPE;
3523                 break;
3524         case TCP_CLOSE:
3525                 return;
3526         default:
3527                 sk->sk_err = ECONNRESET;
3528         }
3529
3530         if (!sock_flag(sk, SOCK_DEAD))
3531                 sk->sk_error_report(sk);
3532
3533         tcp_done(sk);
3534 }
3535
3536 /*
3537  *      Process the FIN bit. This now behaves as it is supposed to work
3538  *      and the FIN takes effect when it is validly part of sequence
3539  *      space. Not before when we get holes.
3540  *
3541  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3542  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3543  *      TIME-WAIT)
3544  *
3545  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3546  *      close and we go into CLOSING (and later onto TIME-WAIT)
3547  *
3548  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3549  */
3550 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3551 {
3552         struct tcp_sock *tp = tcp_sk(sk);
3553
3554         inet_csk_schedule_ack(sk);
3555
3556         sk->sk_shutdown |= RCV_SHUTDOWN;
3557         sock_set_flag(sk, SOCK_DONE);
3558
3559         switch (sk->sk_state) {
3560         case TCP_SYN_RECV:
3561         case TCP_ESTABLISHED:
3562                 /* Move to CLOSE_WAIT */
3563                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3564                 inet_csk(sk)->icsk_ack.pingpong = 1;
3565                 break;
3566
3567         case TCP_CLOSE_WAIT:
3568         case TCP_CLOSING:
3569                 /* Received a retransmission of the FIN, do
3570                  * nothing.
3571                  */
3572                 break;
3573         case TCP_LAST_ACK:
3574                 /* RFC793: Remain in the LAST-ACK state. */
3575                 break;
3576
3577         case TCP_FIN_WAIT1:
3578                 /* This case occurs when a simultaneous close
3579                  * happens, we must ack the received FIN and
3580                  * enter the CLOSING state.
3581                  */
3582                 tcp_send_ack(sk);
3583                 tcp_set_state(sk, TCP_CLOSING);
3584                 break;
3585         case TCP_FIN_WAIT2:
3586                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3587                 tcp_send_ack(sk);
3588                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3589                 break;
3590         default:
3591                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3592                  * cases we should never reach this piece of code.
3593                  */
3594                 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3595                        __FUNCTION__, sk->sk_state);
3596                 break;
3597         }
3598
3599         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3600          * Probably, we should reset in this case. For now drop them.
3601          */
3602         __skb_queue_purge(&tp->out_of_order_queue);
3603         if (tcp_is_sack(tp))
3604                 tcp_sack_reset(&tp->rx_opt);
3605         sk_mem_reclaim(sk);
3606
3607         if (!sock_flag(sk, SOCK_DEAD)) {
3608                 sk->sk_state_change(sk);
3609
3610                 /* Do not send POLL_HUP for half duplex close. */
3611                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3612                     sk->sk_state == TCP_CLOSE)
3613                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3614                 else
3615                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3616         }
3617 }
3618
3619 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3620                                   u32 end_seq)
3621 {
3622         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3623                 if (before(seq, sp->start_seq))
3624                         sp->start_seq = seq;
3625                 if (after(end_seq, sp->end_seq))
3626                         sp->end_seq = end_seq;
3627                 return 1;
3628         }
3629         return 0;
3630 }
3631
3632 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3633 {
3634         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3635                 if (before(seq, tp->rcv_nxt))
3636                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3637                 else
3638                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3639
3640                 tp->rx_opt.dsack = 1;
3641                 tp->duplicate_sack[0].start_seq = seq;
3642                 tp->duplicate_sack[0].end_seq = end_seq;
3643                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1,
3644                                            4 - tp->rx_opt.tstamp_ok);
3645         }
3646 }
3647
3648 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3649 {
3650         if (!tp->rx_opt.dsack)
3651                 tcp_dsack_set(tp, seq, end_seq);
3652         else
3653                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3654 }
3655
3656 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3657 {
3658         struct tcp_sock *tp = tcp_sk(sk);
3659
3660         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3661             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3662                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3663                 tcp_enter_quickack_mode(sk);
3664
3665                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3666                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3667
3668                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3669                                 end_seq = tp->rcv_nxt;
3670                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3671                 }
3672         }
3673
3674         tcp_send_ack(sk);
3675 }
3676
3677 /* These routines update the SACK block as out-of-order packets arrive or
3678  * in-order packets close up the sequence space.
3679  */
3680 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3681 {
3682         int this_sack;
3683         struct tcp_sack_block *sp = &tp->selective_acks[0];
3684         struct tcp_sack_block *swalk = sp + 1;
3685
3686         /* See if the recent change to the first SACK eats into
3687          * or hits the sequence space of other SACK blocks, if so coalesce.
3688          */
3689         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3690                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3691                         int i;
3692
3693                         /* Zap SWALK, by moving every further SACK up by one slot.
3694                          * Decrease num_sacks.
3695                          */
3696                         tp->rx_opt.num_sacks--;
3697                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3698                                                    tp->rx_opt.dsack,
3699                                                    4 - tp->rx_opt.tstamp_ok);
3700                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3701                                 sp[i] = sp[i + 1];
3702                         continue;
3703                 }
3704                 this_sack++, swalk++;
3705         }
3706 }
3707
3708 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
3709                                  struct tcp_sack_block *sack2)
3710 {
3711         __u32 tmp;
3712
3713         tmp = sack1->start_seq;
3714         sack1->start_seq = sack2->start_seq;
3715         sack2->start_seq = tmp;
3716
3717         tmp = sack1->end_seq;
3718         sack1->end_seq = sack2->end_seq;
3719         sack2->end_seq = tmp;
3720 }
3721
3722 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3723 {
3724         struct tcp_sock *tp = tcp_sk(sk);
3725         struct tcp_sack_block *sp = &tp->selective_acks[0];
3726         int cur_sacks = tp->rx_opt.num_sacks;
3727         int this_sack;
3728
3729         if (!cur_sacks)
3730                 goto new_sack;
3731
3732         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3733                 if (tcp_sack_extend(sp, seq, end_seq)) {
3734                         /* Rotate this_sack to the first one. */
3735                         for (; this_sack > 0; this_sack--, sp--)
3736                                 tcp_sack_swap(sp, sp - 1);
3737                         if (cur_sacks > 1)
3738                                 tcp_sack_maybe_coalesce(tp);
3739                         return;
3740                 }
3741         }
3742
3743         /* Could not find an adjacent existing SACK, build a new one,
3744          * put it at the front, and shift everyone else down.  We
3745          * always know there is at least one SACK present already here.
3746          *
3747          * If the sack array is full, forget about the last one.
3748          */
3749         if (this_sack >= 4) {
3750                 this_sack--;
3751                 tp->rx_opt.num_sacks--;
3752                 sp--;
3753         }
3754         for (; this_sack > 0; this_sack--, sp--)
3755                 *sp = *(sp - 1);
3756
3757 new_sack:
3758         /* Build the new head SACK, and we're done. */
3759         sp->start_seq = seq;
3760         sp->end_seq = end_seq;
3761         tp->rx_opt.num_sacks++;
3762         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack,
3763                                    4 - tp->rx_opt.tstamp_ok);
3764 }
3765
3766 /* RCV.NXT advances, some SACKs should be eaten. */
3767
3768 static void tcp_sack_remove(struct tcp_sock *tp)
3769 {
3770         struct tcp_sack_block *sp = &tp->selective_acks[0];
3771         int num_sacks = tp->rx_opt.num_sacks;
3772         int this_sack;
3773
3774         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3775         if (skb_queue_empty(&tp->out_of_order_queue)) {
3776                 tp->rx_opt.num_sacks = 0;
3777                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3778                 return;
3779         }
3780
3781         for (this_sack = 0; this_sack < num_sacks;) {
3782                 /* Check if the start of the sack is covered by RCV.NXT. */
3783                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3784                         int i;
3785
3786                         /* RCV.NXT must cover all the block! */
3787                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3788
3789                         /* Zap this SACK, by moving forward any other SACKS. */
3790                         for (i=this_sack+1; i < num_sacks; i++)
3791                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3792                         num_sacks--;
3793                         continue;
3794                 }
3795                 this_sack++;
3796                 sp++;
3797         }
3798         if (num_sacks != tp->rx_opt.num_sacks) {
3799                 tp->rx_opt.num_sacks = num_sacks;
3800                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3801                                            tp->rx_opt.dsack,
3802                                            4 - tp->rx_opt.tstamp_ok);
3803         }
3804 }
3805
3806 /* This one checks to see if we can put data from the
3807  * out_of_order queue into the receive_queue.
3808  */
3809 static void tcp_ofo_queue(struct sock *sk)
3810 {
3811         struct tcp_sock *tp = tcp_sk(sk);
3812         __u32 dsack_high = tp->rcv_nxt;
3813         struct sk_buff *skb;
3814
3815         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3816                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3817                         break;
3818
3819                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3820                         __u32 dsack = dsack_high;
3821                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3822                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3823                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3824                 }
3825
3826                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3827                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3828                         __skb_unlink(skb, &tp->out_of_order_queue);
3829                         __kfree_skb(skb);
3830                         continue;
3831                 }
3832                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3833                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3834                            TCP_SKB_CB(skb)->end_seq);
3835
3836                 __skb_unlink(skb, &tp->out_of_order_queue);
3837                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3838                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3839                 if (tcp_hdr(skb)->fin)
3840                         tcp_fin(skb, sk, tcp_hdr(skb));
3841         }
3842 }
3843
3844 static int tcp_prune_queue(struct sock *sk);
3845
3846 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3847 {
3848         struct tcphdr *th = tcp_hdr(skb);
3849         struct tcp_sock *tp = tcp_sk(sk);
3850         int eaten = -1;
3851
3852         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3853                 goto drop;
3854
3855         __skb_pull(skb, th->doff * 4);
3856
3857         TCP_ECN_accept_cwr(tp, skb);
3858
3859         if (tp->rx_opt.dsack) {
3860                 tp->rx_opt.dsack = 0;
3861                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3862                                              4 - tp->rx_opt.tstamp_ok);
3863         }
3864
3865         /*  Queue data for delivery to the user.
3866          *  Packets in sequence go to the receive queue.
3867          *  Out of sequence packets to the out_of_order_queue.
3868          */
3869         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3870                 if (tcp_receive_window(tp) == 0)
3871                         goto out_of_window;
3872
3873                 /* Ok. In sequence. In window. */
3874                 if (tp->ucopy.task == current &&
3875                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3876                     sock_owned_by_user(sk) && !tp->urg_data) {
3877                         int chunk = min_t(unsigned int, skb->len,
3878                                           tp->ucopy.len);
3879
3880                         __set_current_state(TASK_RUNNING);
3881
3882                         local_bh_enable();
3883                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3884                                 tp->ucopy.len -= chunk;
3885                                 tp->copied_seq += chunk;
3886                                 eaten = (chunk == skb->len && !th->fin);
3887                                 tcp_rcv_space_adjust(sk);
3888                         }
3889                         local_bh_disable();
3890                 }
3891
3892                 if (eaten <= 0) {
3893 queue_and_out:
3894                         if (eaten < 0 &&
3895                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3896                              !sk_rmem_schedule(sk, skb->truesize))) {
3897                                 if (tcp_prune_queue(sk) < 0 ||
3898                                     !sk_rmem_schedule(sk, skb->truesize))
3899                                         goto drop;
3900                         }
3901                         skb_set_owner_r(skb, sk);
3902                         __skb_queue_tail(&sk->sk_receive_queue, skb);
3903                 }
3904                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3905                 if (skb->len)
3906                         tcp_event_data_recv(sk, skb);
3907                 if (th->fin)
3908                         tcp_fin(skb, sk, th);
3909
3910                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3911                         tcp_ofo_queue(sk);
3912
3913                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
3914                          * gap in queue is filled.
3915                          */
3916                         if (skb_queue_empty(&tp->out_of_order_queue))
3917                                 inet_csk(sk)->icsk_ack.pingpong = 0;
3918                 }
3919
3920                 if (tp->rx_opt.num_sacks)
3921                         tcp_sack_remove(tp);
3922
3923                 tcp_fast_path_check(sk);
3924
3925                 if (eaten > 0)
3926                         __kfree_skb(skb);
3927                 else if (!sock_flag(sk, SOCK_DEAD))
3928                         sk->sk_data_ready(sk, 0);
3929                 return;
3930         }
3931
3932         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3933                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3934                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3935                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3936
3937 out_of_window:
3938                 tcp_enter_quickack_mode(sk);
3939                 inet_csk_schedule_ack(sk);
3940 drop:
3941                 __kfree_skb(skb);
3942                 return;
3943         }
3944
3945         /* Out of window. F.e. zero window probe. */
3946         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3947                 goto out_of_window;
3948
3949         tcp_enter_quickack_mode(sk);
3950
3951         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3952                 /* Partial packet, seq < rcv_next < end_seq */
3953                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3954                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3955                            TCP_SKB_CB(skb)->end_seq);
3956
3957                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3958
3959                 /* If window is closed, drop tail of packet. But after
3960                  * remembering D-SACK for its head made in previous line.
3961                  */
3962                 if (!tcp_receive_window(tp))
3963                         goto out_of_window;
3964                 goto queue_and_out;
3965         }
3966
3967         TCP_ECN_check_ce(tp, skb);
3968
3969         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3970             !sk_rmem_schedule(sk, skb->truesize)) {
3971                 if (tcp_prune_queue(sk) < 0 ||
3972                     !sk_rmem_schedule(sk, skb->truesize))
3973                         goto drop;
3974         }
3975
3976         /* Disable header prediction. */
3977         tp->pred_flags = 0;
3978         inet_csk_schedule_ack(sk);
3979
3980         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3981                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3982
3983         skb_set_owner_r(skb, sk);
3984
3985         if (!skb_peek(&tp->out_of_order_queue)) {
3986                 /* Initial out of order segment, build 1 SACK. */
3987                 if (tcp_is_sack(tp)) {
3988                         tp->rx_opt.num_sacks = 1;
3989                         tp->rx_opt.dsack     = 0;
3990                         tp->rx_opt.eff_sacks = 1;
3991                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3992                         tp->selective_acks[0].end_seq =
3993                                                 TCP_SKB_CB(skb)->end_seq;
3994                 }
3995                 __skb_queue_head(&tp->out_of_order_queue, skb);
3996         } else {
3997                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3998                 u32 seq = TCP_SKB_CB(skb)->seq;
3999                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4000
4001                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4002                         __skb_append(skb1, skb, &tp->out_of_order_queue);
4003
4004                         if (!tp->rx_opt.num_sacks ||
4005                             tp->selective_acks[0].end_seq != seq)
4006                                 goto add_sack;
4007
4008                         /* Common case: data arrive in order after hole. */
4009                         tp->selective_acks[0].end_seq = end_seq;
4010                         return;
4011                 }
4012
4013                 /* Find place to insert this segment. */
4014                 do {
4015                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
4016                                 break;
4017                 } while ((skb1 = skb1->prev) !=
4018                          (struct sk_buff *)&tp->out_of_order_queue);
4019
4020                 /* Do skb overlap to previous one? */
4021                 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4022                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4023                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4024                                 /* All the bits are present. Drop. */
4025                                 __kfree_skb(skb);
4026                                 tcp_dsack_set(tp, seq, end_seq);
4027                                 goto add_sack;
4028                         }
4029                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4030                                 /* Partial overlap. */
4031                                 tcp_dsack_set(tp, seq,
4032                                               TCP_SKB_CB(skb1)->end_seq);
4033                         } else {
4034                                 skb1 = skb1->prev;
4035                         }
4036                 }
4037                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
4038
4039                 /* And clean segments covered by new one as whole. */
4040                 while ((skb1 = skb->next) !=
4041                        (struct sk_buff *)&tp->out_of_order_queue &&
4042                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4043                         if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4044                                 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq,
4045                                                  end_seq);
4046                                 break;
4047                         }
4048                         __skb_unlink(skb1, &tp->out_of_order_queue);
4049                         tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq,
4050                                          TCP_SKB_CB(skb1)->end_seq);
4051                         __kfree_skb(skb1);
4052                 }
4053
4054 add_sack:
4055                 if (tcp_is_sack(tp))
4056                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4057         }
4058 }
4059
4060 /* Collapse contiguous sequence of skbs head..tail with
4061  * sequence numbers start..end.
4062  * Segments with FIN/SYN are not collapsed (only because this
4063  * simplifies code)
4064  */
4065 static void
4066 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4067              struct sk_buff *head, struct sk_buff *tail,
4068              u32 start, u32 end)
4069 {
4070         struct sk_buff *skb;
4071
4072         /* First, check that queue is collapsible and find
4073          * the point where collapsing can be useful. */
4074         for (skb = head; skb != tail;) {
4075                 /* No new bits? It is possible on ofo queue. */
4076                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4077                         struct sk_buff *next = skb->next;
4078                         __skb_unlink(skb, list);
4079                         __kfree_skb(skb);
4080                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4081                         skb = next;
4082                         continue;
4083                 }
4084
4085                 /* The first skb to collapse is:
4086                  * - not SYN/FIN and
4087                  * - bloated or contains data before "start" or
4088                  *   overlaps to the next one.
4089                  */
4090                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4091                     (tcp_win_from_space(skb->truesize) > skb->len ||
4092                      before(TCP_SKB_CB(skb)->seq, start) ||
4093                      (skb->next != tail &&
4094                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4095                         break;
4096
4097                 /* Decided to skip this, advance start seq. */
4098                 start = TCP_SKB_CB(skb)->end_seq;
4099                 skb = skb->next;
4100         }
4101         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4102                 return;
4103
4104         while (before(start, end)) {
4105                 struct sk_buff *nskb;
4106                 unsigned int header = skb_headroom(skb);
4107                 int copy = SKB_MAX_ORDER(header, 0);
4108
4109                 /* Too big header? This can happen with IPv6. */
4110                 if (copy < 0)
4111                         return;
4112                 if (end - start < copy)
4113                         copy = end - start;
4114                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4115                 if (!nskb)
4116                         return;
4117
4118                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4119                 skb_set_network_header(nskb, (skb_network_header(skb) -
4120                                               skb->head));
4121                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4122                                                 skb->head));
4123                 skb_reserve(nskb, header);
4124                 memcpy(nskb->head, skb->head, header);
4125                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4126                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4127                 __skb_insert(nskb, skb->prev, skb, list);
4128                 skb_set_owner_r(nskb, sk);
4129
4130                 /* Copy data, releasing collapsed skbs. */
4131                 while (copy > 0) {
4132                         int offset = start - TCP_SKB_CB(skb)->seq;
4133                         int size = TCP_SKB_CB(skb)->end_seq - start;
4134
4135                         BUG_ON(offset < 0);
4136                         if (size > 0) {
4137                                 size = min(copy, size);
4138                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4139                                         BUG();
4140                                 TCP_SKB_CB(nskb)->end_seq += size;
4141                                 copy -= size;
4142                                 start += size;
4143                         }
4144                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4145                                 struct sk_buff *next = skb->next;
4146                                 __skb_unlink(skb, list);
4147                                 __kfree_skb(skb);
4148                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4149                                 skb = next;
4150                                 if (skb == tail ||
4151                                     tcp_hdr(skb)->syn ||
4152                                     tcp_hdr(skb)->fin)
4153                                         return;
4154                         }
4155                 }
4156         }
4157 }
4158
4159 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4160  * and tcp_collapse() them until all the queue is collapsed.
4161  */
4162 static void tcp_collapse_ofo_queue(struct sock *sk)
4163 {
4164         struct tcp_sock *tp = tcp_sk(sk);
4165         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4166         struct sk_buff *head;
4167         u32 start, end;
4168
4169         if (skb == NULL)
4170                 return;
4171
4172         start = TCP_SKB_CB(skb)->seq;
4173         end = TCP_SKB_CB(skb)->end_seq;
4174         head = skb;
4175
4176         for (;;) {
4177                 skb = skb->next;
4178
4179                 /* Segment is terminated when we see gap or when
4180                  * we are at the end of all the queue. */
4181                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4182                     after(TCP_SKB_CB(skb)->seq, end) ||
4183                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4184                         tcp_collapse(sk, &tp->out_of_order_queue,
4185                                      head, skb, start, end);
4186                         head = skb;
4187                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4188                                 break;
4189                         /* Start new segment */
4190                         start = TCP_SKB_CB(skb)->seq;
4191                         end = TCP_SKB_CB(skb)->end_seq;
4192                 } else {
4193                         if (before(TCP_SKB_CB(skb)->seq, start))
4194                                 start = TCP_SKB_CB(skb)->seq;
4195                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4196                                 end = TCP_SKB_CB(skb)->end_seq;
4197                 }
4198         }
4199 }
4200
4201 /* Reduce allocated memory if we can, trying to get
4202  * the socket within its memory limits again.
4203  *
4204  * Return less than zero if we should start dropping frames
4205  * until the socket owning process reads some of the data
4206  * to stabilize the situation.
4207  */
4208 static int tcp_prune_queue(struct sock *sk)
4209 {
4210         struct tcp_sock *tp = tcp_sk(sk);
4211
4212         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4213
4214         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
4215
4216         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4217                 tcp_clamp_window(sk);
4218         else if (tcp_memory_pressure)
4219                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4220
4221         tcp_collapse_ofo_queue(sk);
4222         tcp_collapse(sk, &sk->sk_receive_queue,
4223                      sk->sk_receive_queue.next,
4224                      (struct sk_buff *)&sk->sk_receive_queue,
4225                      tp->copied_seq, tp->rcv_nxt);
4226         sk_mem_reclaim(sk);
4227
4228         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4229                 return 0;
4230
4231         /* Collapsing did not help, destructive actions follow.
4232          * This must not ever occur. */
4233
4234         /* First, purge the out_of_order queue. */
4235         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4236                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4237                 __skb_queue_purge(&tp->out_of_order_queue);
4238
4239                 /* Reset SACK state.  A conforming SACK implementation will
4240                  * do the same at a timeout based retransmit.  When a connection
4241                  * is in a sad state like this, we care only about integrity
4242                  * of the connection not performance.
4243                  */
4244                 if (tcp_is_sack(tp))
4245                         tcp_sack_reset(&tp->rx_opt);
4246                 sk_mem_reclaim(sk);
4247         }
4248
4249         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4250                 return 0;
4251
4252         /* If we are really being abused, tell the caller to silently
4253          * drop receive data on the floor.  It will get retransmitted
4254          * and hopefully then we'll have sufficient space.
4255          */
4256         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4257
4258         /* Massive buffer overcommit. */
4259         tp->pred_flags = 0;
4260         return -1;
4261 }
4262
4263 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4264  * As additional protections, we do not touch cwnd in retransmission phases,
4265  * and if application hit its sndbuf limit recently.
4266  */
4267 void tcp_cwnd_application_limited(struct sock *sk)
4268 {
4269         struct tcp_sock *tp = tcp_sk(sk);
4270
4271         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4272             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4273                 /* Limited by application or receiver window. */
4274                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4275                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4276                 if (win_used < tp->snd_cwnd) {
4277                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4278                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4279                 }
4280                 tp->snd_cwnd_used = 0;
4281         }
4282         tp->snd_cwnd_stamp = tcp_time_stamp;
4283 }
4284
4285 static int tcp_should_expand_sndbuf(struct sock *sk)
4286 {
4287         struct tcp_sock *tp = tcp_sk(sk);
4288
4289         /* If the user specified a specific send buffer setting, do
4290          * not modify it.
4291          */
4292         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4293                 return 0;
4294
4295         /* If we are under global TCP memory pressure, do not expand.  */
4296         if (tcp_memory_pressure)
4297                 return 0;
4298
4299         /* If we are under soft global TCP memory pressure, do not expand.  */
4300         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4301                 return 0;
4302
4303         /* If we filled the congestion window, do not expand.  */
4304         if (tp->packets_out >= tp->snd_cwnd)
4305                 return 0;
4306
4307         return 1;
4308 }
4309
4310 /* When incoming ACK allowed to free some skb from write_queue,
4311  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4312  * on the exit from tcp input handler.
4313  *
4314  * PROBLEM: sndbuf expansion does not work well with largesend.
4315  */
4316 static void tcp_new_space(struct sock *sk)
4317 {
4318         struct tcp_sock *tp = tcp_sk(sk);
4319
4320         if (tcp_should_expand_sndbuf(sk)) {
4321                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4322                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4323                     demanded = max_t(unsigned int, tp->snd_cwnd,
4324                                      tp->reordering + 1);
4325                 sndmem *= 2 * demanded;
4326                 if (sndmem > sk->sk_sndbuf)
4327                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4328                 tp->snd_cwnd_stamp = tcp_time_stamp;
4329         }
4330
4331         sk->sk_write_space(sk);
4332 }
4333
4334 static void tcp_check_space(struct sock *sk)
4335 {
4336         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4337                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4338                 if (sk->sk_socket &&
4339                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4340                         tcp_new_space(sk);
4341         }
4342 }
4343
4344 static inline void tcp_data_snd_check(struct sock *sk)
4345 {
4346         tcp_push_pending_frames(sk);
4347         tcp_check_space(sk);
4348 }
4349
4350 /*
4351  * Check if sending an ack is needed.
4352  */
4353 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4354 {
4355         struct tcp_sock *tp = tcp_sk(sk);
4356
4357             /* More than one full frame received... */
4358         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4359              /* ... and right edge of window advances far enough.
4360               * (tcp_recvmsg() will send ACK otherwise). Or...
4361               */
4362              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4363             /* We ACK each frame or... */
4364             tcp_in_quickack_mode(sk) ||
4365             /* We have out of order data. */
4366             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4367                 /* Then ack it now */
4368                 tcp_send_ack(sk);
4369         } else {
4370                 /* Else, send delayed ack. */
4371                 tcp_send_delayed_ack(sk);
4372         }
4373 }
4374
4375 static inline void tcp_ack_snd_check(struct sock *sk)
4376 {
4377         if (!inet_csk_ack_scheduled(sk)) {
4378                 /* We sent a data segment already. */
4379                 return;
4380         }
4381         __tcp_ack_snd_check(sk, 1);
4382 }
4383
4384 /*
4385  *      This routine is only called when we have urgent data
4386  *      signaled. Its the 'slow' part of tcp_urg. It could be
4387  *      moved inline now as tcp_urg is only called from one
4388  *      place. We handle URGent data wrong. We have to - as
4389  *      BSD still doesn't use the correction from RFC961.
4390  *      For 1003.1g we should support a new option TCP_STDURG to permit
4391  *      either form (or just set the sysctl tcp_stdurg).
4392  */
4393
4394 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4395 {
4396         struct tcp_sock *tp = tcp_sk(sk);
4397         u32 ptr = ntohs(th->urg_ptr);
4398
4399         if (ptr && !sysctl_tcp_stdurg)
4400                 ptr--;
4401         ptr += ntohl(th->seq);
4402
4403         /* Ignore urgent data that we've already seen and read. */
4404         if (after(tp->copied_seq, ptr))
4405                 return;
4406
4407         /* Do not replay urg ptr.
4408          *
4409          * NOTE: interesting situation not covered by specs.
4410          * Misbehaving sender may send urg ptr, pointing to segment,
4411          * which we already have in ofo queue. We are not able to fetch
4412          * such data and will stay in TCP_URG_NOTYET until will be eaten
4413          * by recvmsg(). Seems, we are not obliged to handle such wicked
4414          * situations. But it is worth to think about possibility of some
4415          * DoSes using some hypothetical application level deadlock.
4416          */
4417         if (before(ptr, tp->rcv_nxt))
4418                 return;
4419
4420         /* Do we already have a newer (or duplicate) urgent pointer? */
4421         if (tp->urg_data && !after(ptr, tp->urg_seq))
4422                 return;
4423
4424         /* Tell the world about our new urgent pointer. */
4425         sk_send_sigurg(sk);
4426
4427         /* We may be adding urgent data when the last byte read was
4428          * urgent. To do this requires some care. We cannot just ignore
4429          * tp->copied_seq since we would read the last urgent byte again
4430          * as data, nor can we alter copied_seq until this data arrives
4431          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4432          *
4433          * NOTE. Double Dutch. Rendering to plain English: author of comment
4434          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4435          * and expect that both A and B disappear from stream. This is _wrong_.
4436          * Though this happens in BSD with high probability, this is occasional.
4437          * Any application relying on this is buggy. Note also, that fix "works"
4438          * only in this artificial test. Insert some normal data between A and B and we will
4439          * decline of BSD again. Verdict: it is better to remove to trap
4440          * buggy users.
4441          */
4442         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4443             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4444                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4445                 tp->copied_seq++;
4446                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4447                         __skb_unlink(skb, &sk->sk_receive_queue);
4448                         __kfree_skb(skb);
4449                 }
4450         }
4451
4452         tp->urg_data = TCP_URG_NOTYET;
4453         tp->urg_seq = ptr;
4454
4455         /* Disable header prediction. */
4456         tp->pred_flags = 0;
4457 }
4458
4459 /* This is the 'fast' part of urgent handling. */
4460 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4461 {
4462         struct tcp_sock *tp = tcp_sk(sk);
4463
4464         /* Check if we get a new urgent pointer - normally not. */
4465         if (th->urg)
4466                 tcp_check_urg(sk, th);
4467
4468         /* Do we wait for any urgent data? - normally not... */
4469         if (tp->urg_data == TCP_URG_NOTYET) {
4470                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4471                           th->syn;
4472
4473                 /* Is the urgent pointer pointing into this packet? */
4474                 if (ptr < skb->len) {
4475                         u8 tmp;
4476                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4477                                 BUG();
4478                         tp->urg_data = TCP_URG_VALID | tmp;
4479                         if (!sock_flag(sk, SOCK_DEAD))
4480                                 sk->sk_data_ready(sk, 0);
4481                 }
4482         }
4483 }
4484
4485 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4486 {
4487         struct tcp_sock *tp = tcp_sk(sk);
4488         int chunk = skb->len - hlen;
4489         int err;
4490
4491         local_bh_enable();
4492         if (skb_csum_unnecessary(skb))
4493                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4494         else
4495                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4496                                                        tp->ucopy.iov);
4497
4498         if (!err) {
4499                 tp->ucopy.len -= chunk;
4500                 tp->copied_seq += chunk;
4501                 tcp_rcv_space_adjust(sk);
4502         }
4503
4504         local_bh_disable();
4505         return err;
4506 }
4507
4508 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4509                                             struct sk_buff *skb)
4510 {
4511         __sum16 result;
4512
4513         if (sock_owned_by_user(sk)) {
4514                 local_bh_enable();
4515                 result = __tcp_checksum_complete(skb);
4516                 local_bh_disable();
4517         } else {
4518                 result = __tcp_checksum_complete(skb);
4519         }
4520         return result;
4521 }
4522
4523 static inline int tcp_checksum_complete_user(struct sock *sk,
4524                                              struct sk_buff *skb)
4525 {
4526         return !skb_csum_unnecessary(skb) &&
4527                __tcp_checksum_complete_user(sk, skb);
4528 }
4529
4530 #ifdef CONFIG_NET_DMA
4531 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4532                                   int hlen)
4533 {
4534         struct tcp_sock *tp = tcp_sk(sk);
4535         int chunk = skb->len - hlen;
4536         int dma_cookie;
4537         int copied_early = 0;
4538
4539         if (tp->ucopy.wakeup)
4540                 return 0;
4541
4542         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4543                 tp->ucopy.dma_chan = get_softnet_dma();
4544
4545         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4546
4547                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4548                                                          skb, hlen,
4549                                                          tp->ucopy.iov, chunk,
4550                                                          tp->ucopy.pinned_list);
4551
4552                 if (dma_cookie < 0)
4553                         goto out;
4554
4555                 tp->ucopy.dma_cookie = dma_cookie;
4556                 copied_early = 1;
4557
4558                 tp->ucopy.len -= chunk;
4559                 tp->copied_seq += chunk;
4560                 tcp_rcv_space_adjust(sk);
4561
4562                 if ((tp->ucopy.len == 0) ||
4563                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4564                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4565                         tp->ucopy.wakeup = 1;
4566                         sk->sk_data_ready(sk, 0);
4567                 }
4568         } else if (chunk > 0) {
4569                 tp->ucopy.wakeup = 1;
4570                 sk->sk_data_ready(sk, 0);
4571         }
4572 out:
4573         return copied_early;
4574 }
4575 #endif /* CONFIG_NET_DMA */
4576
4577 /*
4578  *      TCP receive function for the ESTABLISHED state.
4579  *
4580  *      It is split into a fast path and a slow path. The fast path is
4581  *      disabled when:
4582  *      - A zero window was announced from us - zero window probing
4583  *        is only handled properly in the slow path.
4584  *      - Out of order segments arrived.
4585  *      - Urgent data is expected.
4586  *      - There is no buffer space left
4587  *      - Unexpected TCP flags/window values/header lengths are received
4588  *        (detected by checking the TCP header against pred_flags)
4589  *      - Data is sent in both directions. Fast path only supports pure senders
4590  *        or pure receivers (this means either the sequence number or the ack
4591  *        value must stay constant)
4592  *      - Unexpected TCP option.
4593  *
4594  *      When these conditions are not satisfied it drops into a standard
4595  *      receive procedure patterned after RFC793 to handle all cases.
4596  *      The first three cases are guaranteed by proper pred_flags setting,
4597  *      the rest is checked inline. Fast processing is turned on in
4598  *      tcp_data_queue when everything is OK.
4599  */
4600 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4601                         struct tcphdr *th, unsigned len)
4602 {
4603         struct tcp_sock *tp = tcp_sk(sk);
4604
4605         /*
4606          *      Header prediction.
4607          *      The code loosely follows the one in the famous
4608          *      "30 instruction TCP receive" Van Jacobson mail.
4609          *
4610          *      Van's trick is to deposit buffers into socket queue
4611          *      on a device interrupt, to call tcp_recv function
4612          *      on the receive process context and checksum and copy
4613          *      the buffer to user space. smart...
4614          *
4615          *      Our current scheme is not silly either but we take the
4616          *      extra cost of the net_bh soft interrupt processing...
4617          *      We do checksum and copy also but from device to kernel.
4618          */
4619
4620         tp->rx_opt.saw_tstamp = 0;
4621
4622         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4623          *      if header_prediction is to be made
4624          *      'S' will always be tp->tcp_header_len >> 2
4625          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4626          *  turn it off (when there are holes in the receive
4627          *       space for instance)
4628          *      PSH flag is ignored.
4629          */
4630
4631         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4632             TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4633                 int tcp_header_len = tp->tcp_header_len;
4634
4635                 /* Timestamp header prediction: tcp_header_len
4636                  * is automatically equal to th->doff*4 due to pred_flags
4637                  * match.
4638                  */
4639
4640                 /* Check timestamp */
4641                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4642                         __be32 *ptr = (__be32 *)(th + 1);
4643
4644                         /* No? Slow path! */
4645                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4646                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4647                                 goto slow_path;
4648
4649                         tp->rx_opt.saw_tstamp = 1;
4650                         ++ptr;
4651                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4652                         ++ptr;
4653                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4654
4655                         /* If PAWS failed, check it more carefully in slow path */
4656                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4657                                 goto slow_path;
4658
4659                         /* DO NOT update ts_recent here, if checksum fails
4660                          * and timestamp was corrupted part, it will result
4661                          * in a hung connection since we will drop all
4662                          * future packets due to the PAWS test.
4663                          */
4664                 }
4665
4666                 if (len <= tcp_header_len) {
4667                         /* Bulk data transfer: sender */
4668                         if (len == tcp_header_len) {
4669                                 /* Predicted packet is in window by definition.
4670                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4671                                  * Hence, check seq<=rcv_wup reduces to:
4672                                  */
4673                                 if (tcp_header_len ==
4674                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4675                                     tp->rcv_nxt == tp->rcv_wup)
4676                                         tcp_store_ts_recent(tp);
4677
4678                                 /* We know that such packets are checksummed
4679                                  * on entry.
4680                                  */
4681                                 tcp_ack(sk, skb, 0);
4682                                 __kfree_skb(skb);
4683                                 tcp_data_snd_check(sk);
4684                                 return 0;
4685                         } else { /* Header too small */
4686                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4687                                 goto discard;
4688                         }
4689                 } else {
4690                         int eaten = 0;
4691                         int copied_early = 0;
4692
4693                         if (tp->copied_seq == tp->rcv_nxt &&
4694                             len - tcp_header_len <= tp->ucopy.len) {
4695 #ifdef CONFIG_NET_DMA
4696                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4697                                         copied_early = 1;
4698                                         eaten = 1;
4699                                 }
4700 #endif
4701                                 if (tp->ucopy.task == current &&
4702                                     sock_owned_by_user(sk) && !copied_early) {
4703                                         __set_current_state(TASK_RUNNING);
4704
4705                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4706                                                 eaten = 1;
4707                                 }
4708                                 if (eaten) {
4709                                         /* Predicted packet is in window by definition.
4710                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4711                                          * Hence, check seq<=rcv_wup reduces to:
4712                                          */
4713                                         if (tcp_header_len ==
4714                                             (sizeof(struct tcphdr) +
4715                                              TCPOLEN_TSTAMP_ALIGNED) &&
4716                                             tp->rcv_nxt == tp->rcv_wup)
4717                                                 tcp_store_ts_recent(tp);
4718
4719                                         tcp_rcv_rtt_measure_ts(sk, skb);
4720
4721                                         __skb_pull(skb, tcp_header_len);
4722                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4723                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4724                                 }
4725                                 if (copied_early)
4726                                         tcp_cleanup_rbuf(sk, skb->len);
4727                         }
4728                         if (!eaten) {
4729                                 if (tcp_checksum_complete_user(sk, skb))
4730                                         goto csum_error;
4731
4732                                 /* Predicted packet is in window by definition.
4733                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4734                                  * Hence, check seq<=rcv_wup reduces to:
4735                                  */
4736                                 if (tcp_header_len ==
4737                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4738                                     tp->rcv_nxt == tp->rcv_wup)
4739                                         tcp_store_ts_recent(tp);
4740
4741                                 tcp_rcv_rtt_measure_ts(sk, skb);
4742
4743                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4744                                         goto step5;
4745
4746                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4747
4748                                 /* Bulk data transfer: receiver */
4749                                 __skb_pull(skb, tcp_header_len);
4750                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4751                                 skb_set_owner_r(skb, sk);
4752                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4753                         }
4754
4755                         tcp_event_data_recv(sk, skb);
4756
4757                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4758                                 /* Well, only one small jumplet in fast path... */
4759                                 tcp_ack(sk, skb, FLAG_DATA);
4760                                 tcp_data_snd_check(sk);
4761                                 if (!inet_csk_ack_scheduled(sk))
4762                                         goto no_ack;
4763                         }
4764
4765                         __tcp_ack_snd_check(sk, 0);
4766 no_ack:
4767 #ifdef CONFIG_NET_DMA
4768                         if (copied_early)
4769                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4770                         else
4771 #endif
4772                         if (eaten)
4773                                 __kfree_skb(skb);
4774                         else
4775                                 sk->sk_data_ready(sk, 0);
4776                         return 0;
4777                 }
4778         }
4779
4780 slow_path:
4781         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
4782                 goto csum_error;
4783
4784         /*
4785          * RFC1323: H1. Apply PAWS check first.
4786          */
4787         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4788             tcp_paws_discard(sk, skb)) {
4789                 if (!th->rst) {
4790                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4791                         tcp_send_dupack(sk, skb);
4792                         goto discard;
4793                 }
4794                 /* Resets are accepted even if PAWS failed.
4795
4796                    ts_recent update must be made after we are sure
4797                    that the packet is in window.
4798                  */
4799         }
4800
4801         /*
4802          *      Standard slow path.
4803          */
4804
4805         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4806                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4807                  * (RST) segments are validated by checking their SEQ-fields."
4808                  * And page 69: "If an incoming segment is not acceptable,
4809                  * an acknowledgment should be sent in reply (unless the RST bit
4810                  * is set, if so drop the segment and return)".
4811                  */
4812                 if (!th->rst)
4813                         tcp_send_dupack(sk, skb);
4814                 goto discard;
4815         }
4816
4817         if (th->rst) {
4818                 tcp_reset(sk);
4819                 goto discard;
4820         }
4821
4822         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4823
4824         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4825                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4826                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4827                 tcp_reset(sk);
4828                 return 1;
4829         }
4830
4831 step5:
4832         if (th->ack)
4833                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4834
4835         tcp_rcv_rtt_measure_ts(sk, skb);
4836
4837         /* Process urgent data. */
4838         tcp_urg(sk, skb, th);
4839
4840         /* step 7: process the segment text */
4841         tcp_data_queue(sk, skb);
4842
4843         tcp_data_snd_check(sk);
4844         tcp_ack_snd_check(sk);
4845         return 0;
4846
4847 csum_error:
4848         TCP_INC_STATS_BH(TCP_MIB_INERRS);
4849
4850 discard:
4851         __kfree_skb(skb);
4852         return 0;
4853 }
4854
4855 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4856                                          struct tcphdr *th, unsigned len)
4857 {
4858         struct tcp_sock *tp = tcp_sk(sk);
4859         struct inet_connection_sock *icsk = inet_csk(sk);
4860         int saved_clamp = tp->rx_opt.mss_clamp;
4861
4862         tcp_parse_options(skb, &tp->rx_opt, 0);
4863
4864         if (th->ack) {
4865                 /* rfc793:
4866                  * "If the state is SYN-SENT then
4867                  *    first check the ACK bit
4868                  *      If the ACK bit is set
4869                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4870                  *        a reset (unless the RST bit is set, if so drop
4871                  *        the segment and return)"
4872                  *
4873                  *  We do not send data with SYN, so that RFC-correct
4874                  *  test reduces to:
4875                  */
4876                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4877                         goto reset_and_undo;
4878
4879                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4880                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4881                              tcp_time_stamp)) {
4882                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4883                         goto reset_and_undo;
4884                 }
4885
4886                 /* Now ACK is acceptable.
4887                  *
4888                  * "If the RST bit is set
4889                  *    If the ACK was acceptable then signal the user "error:
4890                  *    connection reset", drop the segment, enter CLOSED state,
4891                  *    delete TCB, and return."
4892                  */
4893
4894                 if (th->rst) {
4895                         tcp_reset(sk);
4896                         goto discard;
4897                 }
4898
4899                 /* rfc793:
4900                  *   "fifth, if neither of the SYN or RST bits is set then
4901                  *    drop the segment and return."
4902                  *
4903                  *    See note below!
4904                  *                                        --ANK(990513)
4905                  */
4906                 if (!th->syn)
4907                         goto discard_and_undo;
4908
4909                 /* rfc793:
4910                  *   "If the SYN bit is on ...
4911                  *    are acceptable then ...
4912                  *    (our SYN has been ACKed), change the connection
4913                  *    state to ESTABLISHED..."
4914                  */
4915
4916                 TCP_ECN_rcv_synack(tp, th);
4917
4918                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4919                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4920
4921                 /* Ok.. it's good. Set up sequence numbers and
4922                  * move to established.
4923                  */
4924                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4925                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4926
4927                 /* RFC1323: The window in SYN & SYN/ACK segments is
4928                  * never scaled.
4929                  */
4930                 tp->snd_wnd = ntohs(th->window);
4931                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4932
4933                 if (!tp->rx_opt.wscale_ok) {
4934                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4935                         tp->window_clamp = min(tp->window_clamp, 65535U);
4936                 }
4937
4938                 if (tp->rx_opt.saw_tstamp) {
4939                         tp->rx_opt.tstamp_ok       = 1;
4940                         tp->tcp_header_len =
4941                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4942                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
4943                         tcp_store_ts_recent(tp);
4944                 } else {
4945                         tp->tcp_header_len = sizeof(struct tcphdr);
4946                 }
4947
4948                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4949                         tcp_enable_fack(tp);
4950
4951                 tcp_mtup_init(sk);
4952                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4953                 tcp_initialize_rcv_mss(sk);
4954
4955                 /* Remember, tcp_poll() does not lock socket!
4956                  * Change state from SYN-SENT only after copied_seq
4957                  * is initialized. */
4958                 tp->copied_seq = tp->rcv_nxt;
4959                 smp_mb();
4960                 tcp_set_state(sk, TCP_ESTABLISHED);
4961
4962                 security_inet_conn_established(sk, skb);
4963
4964                 /* Make sure socket is routed, for correct metrics.  */
4965                 icsk->icsk_af_ops->rebuild_header(sk);
4966
4967                 tcp_init_metrics(sk);
4968
4969                 tcp_init_congestion_control(sk);
4970
4971                 /* Prevent spurious tcp_cwnd_restart() on first data
4972                  * packet.
4973                  */
4974                 tp->lsndtime = tcp_time_stamp;
4975
4976                 tcp_init_buffer_space(sk);
4977
4978                 if (sock_flag(sk, SOCK_KEEPOPEN))
4979                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4980
4981                 if (!tp->rx_opt.snd_wscale)
4982                         __tcp_fast_path_on(tp, tp->snd_wnd);
4983                 else
4984                         tp->pred_flags = 0;
4985
4986                 if (!sock_flag(sk, SOCK_DEAD)) {
4987                         sk->sk_state_change(sk);
4988                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
4989                 }
4990
4991                 if (sk->sk_write_pending ||
4992                     icsk->icsk_accept_queue.rskq_defer_accept ||
4993                     icsk->icsk_ack.pingpong) {
4994                         /* Save one ACK. Data will be ready after
4995                          * several ticks, if write_pending is set.
4996                          *
4997                          * It may be deleted, but with this feature tcpdumps
4998                          * look so _wonderfully_ clever, that I was not able
4999                          * to stand against the temptation 8)     --ANK
5000                          */
5001                         inet_csk_schedule_ack(sk);
5002                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5003                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
5004                         tcp_incr_quickack(sk);
5005                         tcp_enter_quickack_mode(sk);
5006                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5007                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5008
5009 discard:
5010                         __kfree_skb(skb);
5011                         return 0;
5012                 } else {
5013                         tcp_send_ack(sk);
5014                 }
5015                 return -1;
5016         }
5017
5018         /* No ACK in the segment */
5019
5020         if (th->rst) {
5021                 /* rfc793:
5022                  * "If the RST bit is set
5023                  *
5024                  *      Otherwise (no ACK) drop the segment and return."
5025                  */
5026
5027                 goto discard_and_undo;
5028         }
5029
5030         /* PAWS check. */
5031         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5032             tcp_paws_check(&tp->rx_opt, 0))
5033                 goto discard_and_undo;
5034
5035         if (th->syn) {
5036                 /* We see SYN without ACK. It is attempt of
5037                  * simultaneous connect with crossed SYNs.
5038                  * Particularly, it can be connect to self.
5039                  */
5040                 tcp_set_state(sk, TCP_SYN_RECV);
5041
5042                 if (tp->rx_opt.saw_tstamp) {
5043                         tp->rx_opt.tstamp_ok = 1;
5044                         tcp_store_ts_recent(tp);
5045                         tp->tcp_header_len =
5046                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5047                 } else {
5048                         tp->tcp_header_len = sizeof(struct tcphdr);
5049                 }
5050
5051                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5052                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5053
5054                 /* RFC1323: The window in SYN & SYN/ACK segments is
5055                  * never scaled.
5056                  */
5057                 tp->snd_wnd    = ntohs(th->window);
5058                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5059                 tp->max_window = tp->snd_wnd;
5060
5061                 TCP_ECN_rcv_syn(tp, th);
5062
5063                 tcp_mtup_init(sk);
5064                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5065                 tcp_initialize_rcv_mss(sk);
5066
5067                 tcp_send_synack(sk);
5068 #if 0
5069                 /* Note, we could accept data and URG from this segment.
5070                  * There are no obstacles to make this.
5071                  *
5072                  * However, if we ignore data in ACKless segments sometimes,
5073                  * we have no reasons to accept it sometimes.
5074                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5075                  * is not flawless. So, discard packet for sanity.
5076                  * Uncomment this return to process the data.
5077                  */
5078                 return -1;
5079 #else
5080                 goto discard;
5081 #endif
5082         }
5083         /* "fifth, if neither of the SYN or RST bits is set then
5084          * drop the segment and return."
5085          */
5086
5087 discard_and_undo:
5088         tcp_clear_options(&tp->rx_opt);
5089         tp->rx_opt.mss_clamp = saved_clamp;
5090         goto discard;
5091
5092 reset_and_undo:
5093         tcp_clear_options(&tp->rx_opt);
5094         tp->rx_opt.mss_clamp = saved_clamp;
5095         return 1;
5096 }
5097
5098 /*
5099  *      This function implements the receiving procedure of RFC 793 for
5100  *      all states except ESTABLISHED and TIME_WAIT.
5101  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5102  *      address independent.
5103  */
5104
5105 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5106                           struct tcphdr *th, unsigned len)
5107 {
5108         struct tcp_sock *tp = tcp_sk(sk);
5109         struct inet_connection_sock *icsk = inet_csk(sk);
5110         int queued = 0;
5111
5112         tp->rx_opt.saw_tstamp = 0;
5113
5114         switch (sk->sk_state) {
5115         case TCP_CLOSE:
5116                 goto discard;
5117
5118         case TCP_LISTEN:
5119                 if (th->ack)
5120                         return 1;
5121
5122                 if (th->rst)
5123                         goto discard;
5124
5125                 if (th->syn) {
5126                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5127                                 return 1;
5128
5129                         /* Now we have several options: In theory there is
5130                          * nothing else in the frame. KA9Q has an option to
5131                          * send data with the syn, BSD accepts data with the
5132                          * syn up to the [to be] advertised window and
5133                          * Solaris 2.1 gives you a protocol error. For now
5134                          * we just ignore it, that fits the spec precisely
5135                          * and avoids incompatibilities. It would be nice in
5136                          * future to drop through and process the data.
5137                          *
5138                          * Now that TTCP is starting to be used we ought to
5139                          * queue this data.
5140                          * But, this leaves one open to an easy denial of
5141                          * service attack, and SYN cookies can't defend
5142                          * against this problem. So, we drop the data
5143                          * in the interest of security over speed unless
5144                          * it's still in use.
5145                          */
5146                         kfree_skb(skb);
5147                         return 0;
5148                 }
5149                 goto discard;
5150
5151         case TCP_SYN_SENT:
5152                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5153                 if (queued >= 0)
5154                         return queued;
5155
5156                 /* Do step6 onward by hand. */
5157                 tcp_urg(sk, skb, th);
5158                 __kfree_skb(skb);
5159                 tcp_data_snd_check(sk);
5160                 return 0;
5161         }
5162
5163         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5164             tcp_paws_discard(sk, skb)) {
5165                 if (!th->rst) {
5166                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
5167                         tcp_send_dupack(sk, skb);
5168                         goto discard;
5169                 }
5170                 /* Reset is accepted even if it did not pass PAWS. */
5171         }
5172
5173         /* step 1: check sequence number */
5174         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5175                 if (!th->rst)
5176                         tcp_send_dupack(sk, skb);
5177                 goto discard;
5178         }
5179
5180         /* step 2: check RST bit */
5181         if (th->rst) {
5182                 tcp_reset(sk);
5183                 goto discard;
5184         }
5185
5186         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5187
5188         /* step 3: check security and precedence [ignored] */
5189
5190         /*      step 4:
5191          *
5192          *      Check for a SYN in window.
5193          */
5194         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5195                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
5196                 tcp_reset(sk);
5197                 return 1;
5198         }
5199
5200         /* step 5: check the ACK field */
5201         if (th->ack) {
5202                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5203
5204                 switch (sk->sk_state) {
5205                 case TCP_SYN_RECV:
5206                         if (acceptable) {
5207                                 tp->copied_seq = tp->rcv_nxt;
5208                                 smp_mb();
5209                                 tcp_set_state(sk, TCP_ESTABLISHED);
5210                                 sk->sk_state_change(sk);
5211
5212                                 /* Note, that this wakeup is only for marginal
5213                                  * crossed SYN case. Passively open sockets
5214                                  * are not waked up, because sk->sk_sleep ==
5215                                  * NULL and sk->sk_socket == NULL.
5216                                  */
5217                                 if (sk->sk_socket)
5218                                         sk_wake_async(sk,
5219                                                       SOCK_WAKE_IO, POLL_OUT);
5220
5221                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5222                                 tp->snd_wnd = ntohs(th->window) <<
5223                                               tp->rx_opt.snd_wscale;
5224                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5225                                             TCP_SKB_CB(skb)->seq);
5226
5227                                 /* tcp_ack considers this ACK as duplicate
5228                                  * and does not calculate rtt.
5229                                  * Fix it at least with timestamps.
5230                                  */
5231                                 if (tp->rx_opt.saw_tstamp &&
5232                                     tp->rx_opt.rcv_tsecr && !tp->srtt)
5233                                         tcp_ack_saw_tstamp(sk, 0);
5234
5235                                 if (tp->rx_opt.tstamp_ok)
5236                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5237
5238                                 /* Make sure socket is routed, for
5239                                  * correct metrics.
5240                                  */
5241                                 icsk->icsk_af_ops->rebuild_header(sk);
5242
5243                                 tcp_init_metrics(sk);
5244
5245                                 tcp_init_congestion_control(sk);
5246
5247                                 /* Prevent spurious tcp_cwnd_restart() on
5248                                  * first data packet.
5249                                  */
5250                                 tp->lsndtime = tcp_time_stamp;
5251
5252                                 tcp_mtup_init(sk);
5253                                 tcp_initialize_rcv_mss(sk);
5254                                 tcp_init_buffer_space(sk);
5255                                 tcp_fast_path_on(tp);
5256                         } else {
5257                                 return 1;
5258                         }
5259                         break;
5260
5261                 case TCP_FIN_WAIT1:
5262                         if (tp->snd_una == tp->write_seq) {
5263                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5264                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5265                                 dst_confirm(sk->sk_dst_cache);
5266
5267                                 if (!sock_flag(sk, SOCK_DEAD))
5268                                         /* Wake up lingering close() */
5269                                         sk->sk_state_change(sk);
5270                                 else {
5271                                         int tmo;
5272
5273                                         if (tp->linger2 < 0 ||
5274                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5275                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5276                                                 tcp_done(sk);
5277                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5278                                                 return 1;
5279                                         }
5280
5281                                         tmo = tcp_fin_time(sk);
5282                                         if (tmo > TCP_TIMEWAIT_LEN) {
5283                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5284                                         } else if (th->fin || sock_owned_by_user(sk)) {
5285                                                 /* Bad case. We could lose such FIN otherwise.
5286                                                  * It is not a big problem, but it looks confusing
5287                                                  * and not so rare event. We still can lose it now,
5288                                                  * if it spins in bh_lock_sock(), but it is really
5289                                                  * marginal case.
5290                                                  */
5291                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5292                                         } else {
5293                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5294                                                 goto discard;
5295                                         }
5296                                 }
5297                         }
5298                         break;
5299
5300                 case TCP_CLOSING:
5301                         if (tp->snd_una == tp->write_seq) {
5302                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5303                                 goto discard;
5304                         }
5305                         break;
5306
5307                 case TCP_LAST_ACK:
5308                         if (tp->snd_una == tp->write_seq) {
5309                                 tcp_update_metrics(sk);
5310                                 tcp_done(sk);
5311                                 goto discard;
5312                         }
5313                         break;
5314                 }
5315         } else
5316                 goto discard;
5317
5318         /* step 6: check the URG bit */
5319         tcp_urg(sk, skb, th);
5320
5321         /* step 7: process the segment text */
5322         switch (sk->sk_state) {
5323         case TCP_CLOSE_WAIT:
5324         case TCP_CLOSING:
5325         case TCP_LAST_ACK:
5326                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5327                         break;
5328         case TCP_FIN_WAIT1:
5329         case TCP_FIN_WAIT2:
5330                 /* RFC 793 says to queue data in these states,
5331                  * RFC 1122 says we MUST send a reset.
5332                  * BSD 4.4 also does reset.
5333                  */
5334                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5335                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5336                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5337                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5338                                 tcp_reset(sk);
5339                                 return 1;
5340                         }
5341                 }
5342                 /* Fall through */
5343         case TCP_ESTABLISHED:
5344                 tcp_data_queue(sk, skb);
5345                 queued = 1;
5346                 break;
5347         }
5348
5349         /* tcp_data could move socket to TIME-WAIT */
5350         if (sk->sk_state != TCP_CLOSE) {
5351                 tcp_data_snd_check(sk);
5352                 tcp_ack_snd_check(sk);
5353         }
5354
5355         if (!queued) {
5356 discard:
5357                 __kfree_skb(skb);
5358         }
5359         return 0;
5360 }
5361
5362 EXPORT_SYMBOL(sysctl_tcp_ecn);
5363 EXPORT_SYMBOL(sysctl_tcp_reordering);
5364 EXPORT_SYMBOL(tcp_parse_options);
5365 EXPORT_SYMBOL(tcp_rcv_established);
5366 EXPORT_SYMBOL(tcp_rcv_state_process);
5367 EXPORT_SYMBOL(tcp_initialize_rcv_mss);