[TCP]: Convert highest_sack to sk_buff to allow direct access
[pandora-kernel.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Version:     $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9  *
10  * Authors:     Ross Biro
11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
14  *              Florian La Roche, <flla@stud.uni-sb.de>
15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20  *              Jorge Cwik, <jorge@laser.satlink.net>
21  */
22
23 /*
24  * Changes:
25  *              Pedro Roque     :       Fast Retransmit/Recovery.
26  *                                      Two receive queues.
27  *                                      Retransmit queue handled by TCP.
28  *                                      Better retransmit timer handling.
29  *                                      New congestion avoidance.
30  *                                      Header prediction.
31  *                                      Variable renaming.
32  *
33  *              Eric            :       Fast Retransmit.
34  *              Randy Scott     :       MSS option defines.
35  *              Eric Schenk     :       Fixes to slow start algorithm.
36  *              Eric Schenk     :       Yet another double ACK bug.
37  *              Eric Schenk     :       Delayed ACK bug fixes.
38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
39  *              David S. Miller :       Don't allow zero congestion window.
40  *              Eric Schenk     :       Fix retransmitter so that it sends
41  *                                      next packet on ack of previous packet.
42  *              Andi Kleen      :       Moved open_request checking here
43  *                                      and process RSTs for open_requests.
44  *              Andi Kleen      :       Better prune_queue, and other fixes.
45  *              Andrey Savochkin:       Fix RTT measurements in the presence of
46  *                                      timestamps.
47  *              Andrey Savochkin:       Check sequence numbers correctly when
48  *                                      removing SACKs due to in sequence incoming
49  *                                      data segments.
50  *              Andi Kleen:             Make sure we never ack data there is not
51  *                                      enough room for. Also make this condition
52  *                                      a fatal error if it might still happen.
53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
54  *                                      connections with MSS<min(MTU,ann. MSS)
55  *                                      work without delayed acks.
56  *              Andi Kleen:             Process packets with PSH set in the
57  *                                      fast path.
58  *              J Hadi Salim:           ECN support
59  *              Andrei Gurtov,
60  *              Pasi Sarolahti,
61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
62  *                                      engine. Lots of bugs are found.
63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
64  */
65
66 #include <linux/mm.h>
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
108
109 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
116
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
119
120 /* Adapt the MSS value used to make delayed ack decision to the
121  * real world.
122  */
123 static void tcp_measure_rcv_mss(struct sock *sk,
124                                 const struct sk_buff *skb)
125 {
126         struct inet_connection_sock *icsk = inet_csk(sk);
127         const unsigned int lss = icsk->icsk_ack.last_seg_size;
128         unsigned int len;
129
130         icsk->icsk_ack.last_seg_size = 0;
131
132         /* skb->len may jitter because of SACKs, even if peer
133          * sends good full-sized frames.
134          */
135         len = skb_shinfo(skb)->gso_size ?: skb->len;
136         if (len >= icsk->icsk_ack.rcv_mss) {
137                 icsk->icsk_ack.rcv_mss = len;
138         } else {
139                 /* Otherwise, we make more careful check taking into account,
140                  * that SACKs block is variable.
141                  *
142                  * "len" is invariant segment length, including TCP header.
143                  */
144                 len += skb->data - skb_transport_header(skb);
145                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
146                     /* If PSH is not set, packet should be
147                      * full sized, provided peer TCP is not badly broken.
148                      * This observation (if it is correct 8)) allows
149                      * to handle super-low mtu links fairly.
150                      */
151                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
152                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
153                         /* Subtract also invariant (if peer is RFC compliant),
154                          * tcp header plus fixed timestamp option length.
155                          * Resulting "len" is MSS free of SACK jitter.
156                          */
157                         len -= tcp_sk(sk)->tcp_header_len;
158                         icsk->icsk_ack.last_seg_size = len;
159                         if (len == lss) {
160                                 icsk->icsk_ack.rcv_mss = len;
161                                 return;
162                         }
163                 }
164                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
166                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
167         }
168 }
169
170 static void tcp_incr_quickack(struct sock *sk)
171 {
172         struct inet_connection_sock *icsk = inet_csk(sk);
173         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
174
175         if (quickacks==0)
176                 quickacks=2;
177         if (quickacks > icsk->icsk_ack.quick)
178                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
179 }
180
181 void tcp_enter_quickack_mode(struct sock *sk)
182 {
183         struct inet_connection_sock *icsk = inet_csk(sk);
184         tcp_incr_quickack(sk);
185         icsk->icsk_ack.pingpong = 0;
186         icsk->icsk_ack.ato = TCP_ATO_MIN;
187 }
188
189 /* Send ACKs quickly, if "quick" count is not exhausted
190  * and the session is not interactive.
191  */
192
193 static inline int tcp_in_quickack_mode(const struct sock *sk)
194 {
195         const struct inet_connection_sock *icsk = inet_csk(sk);
196         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
197 }
198
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
200 {
201         if (tp->ecn_flags&TCP_ECN_OK)
202                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
203 }
204
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
206 {
207         if (tcp_hdr(skb)->cwr)
208                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
209 }
210
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
212 {
213         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 }
215
216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
217 {
218         if (tp->ecn_flags&TCP_ECN_OK) {
219                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221                 /* Funny extension: if ECT is not set on a segment,
222                  * it is surely retransmit. It is not in ECN RFC,
223                  * but Linux follows this rule. */
224                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225                         tcp_enter_quickack_mode((struct sock *)tp);
226         }
227 }
228
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
230 {
231         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
232                 tp->ecn_flags &= ~TCP_ECN_OK;
233 }
234
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
236 {
237         if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
238                 tp->ecn_flags &= ~TCP_ECN_OK;
239 }
240
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
242 {
243         if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
244                 return 1;
245         return 0;
246 }
247
248 /* Buffer size and advertised window tuning.
249  *
250  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
251  */
252
253 static void tcp_fixup_sndbuf(struct sock *sk)
254 {
255         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256                      sizeof(struct sk_buff);
257
258         if (sk->sk_sndbuf < 3 * sndmem)
259                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
260 }
261
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263  *
264  * All tcp_full_space() is split to two parts: "network" buffer, allocated
265  * forward and advertised in receiver window (tp->rcv_wnd) and
266  * "application buffer", required to isolate scheduling/application
267  * latencies from network.
268  * window_clamp is maximal advertised window. It can be less than
269  * tcp_full_space(), in this case tcp_full_space() - window_clamp
270  * is reserved for "application" buffer. The less window_clamp is
271  * the smoother our behaviour from viewpoint of network, but the lower
272  * throughput and the higher sensitivity of the connection to losses. 8)
273  *
274  * rcv_ssthresh is more strict window_clamp used at "slow start"
275  * phase to predict further behaviour of this connection.
276  * It is used for two goals:
277  * - to enforce header prediction at sender, even when application
278  *   requires some significant "application buffer". It is check #1.
279  * - to prevent pruning of receive queue because of misprediction
280  *   of receiver window. Check #2.
281  *
282  * The scheme does not work when sender sends good segments opening
283  * window and then starts to feed us spaghetti. But it should work
284  * in common situations. Otherwise, we have to rely on queue collapsing.
285  */
286
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
289 {
290         struct tcp_sock *tp = tcp_sk(sk);
291         /* Optimize this! */
292         int truesize = tcp_win_from_space(skb->truesize)/2;
293         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
294
295         while (tp->rcv_ssthresh <= window) {
296                 if (truesize <= skb->len)
297                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
298
299                 truesize >>= 1;
300                 window >>= 1;
301         }
302         return 0;
303 }
304
305 static void tcp_grow_window(struct sock *sk,
306                             struct sk_buff *skb)
307 {
308         struct tcp_sock *tp = tcp_sk(sk);
309
310         /* Check #1 */
311         if (tp->rcv_ssthresh < tp->window_clamp &&
312             (int)tp->rcv_ssthresh < tcp_space(sk) &&
313             !tcp_memory_pressure) {
314                 int incr;
315
316                 /* Check #2. Increase window, if skb with such overhead
317                  * will fit to rcvbuf in future.
318                  */
319                 if (tcp_win_from_space(skb->truesize) <= skb->len)
320                         incr = 2*tp->advmss;
321                 else
322                         incr = __tcp_grow_window(sk, skb);
323
324                 if (incr) {
325                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
326                         inet_csk(sk)->icsk_ack.quick |= 1;
327                 }
328         }
329 }
330
331 /* 3. Tuning rcvbuf, when connection enters established state. */
332
333 static void tcp_fixup_rcvbuf(struct sock *sk)
334 {
335         struct tcp_sock *tp = tcp_sk(sk);
336         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
337
338         /* Try to select rcvbuf so that 4 mss-sized segments
339          * will fit to window and corresponding skbs will fit to our rcvbuf.
340          * (was 3; 4 is minimum to allow fast retransmit to work.)
341          */
342         while (tcp_win_from_space(rcvmem) < tp->advmss)
343                 rcvmem += 128;
344         if (sk->sk_rcvbuf < 4 * rcvmem)
345                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
346 }
347
348 /* 4. Try to fixup all. It is made immediately after connection enters
349  *    established state.
350  */
351 static void tcp_init_buffer_space(struct sock *sk)
352 {
353         struct tcp_sock *tp = tcp_sk(sk);
354         int maxwin;
355
356         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357                 tcp_fixup_rcvbuf(sk);
358         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359                 tcp_fixup_sndbuf(sk);
360
361         tp->rcvq_space.space = tp->rcv_wnd;
362
363         maxwin = tcp_full_space(sk);
364
365         if (tp->window_clamp >= maxwin) {
366                 tp->window_clamp = maxwin;
367
368                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369                         tp->window_clamp = max(maxwin -
370                                                (maxwin >> sysctl_tcp_app_win),
371                                                4 * tp->advmss);
372         }
373
374         /* Force reservation of one segment. */
375         if (sysctl_tcp_app_win &&
376             tp->window_clamp > 2 * tp->advmss &&
377             tp->window_clamp + tp->advmss > maxwin)
378                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
379
380         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381         tp->snd_cwnd_stamp = tcp_time_stamp;
382 }
383
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock *sk)
386 {
387         struct tcp_sock *tp = tcp_sk(sk);
388         struct inet_connection_sock *icsk = inet_csk(sk);
389
390         icsk->icsk_ack.quick = 0;
391
392         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394             !tcp_memory_pressure &&
395             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
397                                     sysctl_tcp_rmem[2]);
398         }
399         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
400                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
401 }
402
403
404 /* Initialize RCV_MSS value.
405  * RCV_MSS is an our guess about MSS used by the peer.
406  * We haven't any direct information about the MSS.
407  * It's better to underestimate the RCV_MSS rather than overestimate.
408  * Overestimations make us ACKing less frequently than needed.
409  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
410  */
411 void tcp_initialize_rcv_mss(struct sock *sk)
412 {
413         struct tcp_sock *tp = tcp_sk(sk);
414         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
415
416         hint = min(hint, tp->rcv_wnd/2);
417         hint = min(hint, TCP_MIN_RCVMSS);
418         hint = max(hint, TCP_MIN_MSS);
419
420         inet_csk(sk)->icsk_ack.rcv_mss = hint;
421 }
422
423 /* Receiver "autotuning" code.
424  *
425  * The algorithm for RTT estimation w/o timestamps is based on
426  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
427  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
428  *
429  * More detail on this code can be found at
430  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
431  * though this reference is out of date.  A new paper
432  * is pending.
433  */
434 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
435 {
436         u32 new_sample = tp->rcv_rtt_est.rtt;
437         long m = sample;
438
439         if (m == 0)
440                 m = 1;
441
442         if (new_sample != 0) {
443                 /* If we sample in larger samples in the non-timestamp
444                  * case, we could grossly overestimate the RTT especially
445                  * with chatty applications or bulk transfer apps which
446                  * are stalled on filesystem I/O.
447                  *
448                  * Also, since we are only going for a minimum in the
449                  * non-timestamp case, we do not smooth things out
450                  * else with timestamps disabled convergence takes too
451                  * long.
452                  */
453                 if (!win_dep) {
454                         m -= (new_sample >> 3);
455                         new_sample += m;
456                 } else if (m < new_sample)
457                         new_sample = m << 3;
458         } else {
459                 /* No previous measure. */
460                 new_sample = m << 3;
461         }
462
463         if (tp->rcv_rtt_est.rtt != new_sample)
464                 tp->rcv_rtt_est.rtt = new_sample;
465 }
466
467 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
468 {
469         if (tp->rcv_rtt_est.time == 0)
470                 goto new_measure;
471         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
472                 return;
473         tcp_rcv_rtt_update(tp,
474                            jiffies - tp->rcv_rtt_est.time,
475                            1);
476
477 new_measure:
478         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
479         tp->rcv_rtt_est.time = tcp_time_stamp;
480 }
481
482 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
483 {
484         struct tcp_sock *tp = tcp_sk(sk);
485         if (tp->rx_opt.rcv_tsecr &&
486             (TCP_SKB_CB(skb)->end_seq -
487              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
488                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
489 }
490
491 /*
492  * This function should be called every time data is copied to user space.
493  * It calculates the appropriate TCP receive buffer space.
494  */
495 void tcp_rcv_space_adjust(struct sock *sk)
496 {
497         struct tcp_sock *tp = tcp_sk(sk);
498         int time;
499         int space;
500
501         if (tp->rcvq_space.time == 0)
502                 goto new_measure;
503
504         time = tcp_time_stamp - tp->rcvq_space.time;
505         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
506             tp->rcv_rtt_est.rtt == 0)
507                 return;
508
509         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
510
511         space = max(tp->rcvq_space.space, space);
512
513         if (tp->rcvq_space.space != space) {
514                 int rcvmem;
515
516                 tp->rcvq_space.space = space;
517
518                 if (sysctl_tcp_moderate_rcvbuf &&
519                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
520                         int new_clamp = space;
521
522                         /* Receive space grows, normalize in order to
523                          * take into account packet headers and sk_buff
524                          * structure overhead.
525                          */
526                         space /= tp->advmss;
527                         if (!space)
528                                 space = 1;
529                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
530                                   16 + sizeof(struct sk_buff));
531                         while (tcp_win_from_space(rcvmem) < tp->advmss)
532                                 rcvmem += 128;
533                         space *= rcvmem;
534                         space = min(space, sysctl_tcp_rmem[2]);
535                         if (space > sk->sk_rcvbuf) {
536                                 sk->sk_rcvbuf = space;
537
538                                 /* Make the window clamp follow along.  */
539                                 tp->window_clamp = new_clamp;
540                         }
541                 }
542         }
543
544 new_measure:
545         tp->rcvq_space.seq = tp->copied_seq;
546         tp->rcvq_space.time = tcp_time_stamp;
547 }
548
549 /* There is something which you must keep in mind when you analyze the
550  * behavior of the tp->ato delayed ack timeout interval.  When a
551  * connection starts up, we want to ack as quickly as possible.  The
552  * problem is that "good" TCP's do slow start at the beginning of data
553  * transmission.  The means that until we send the first few ACK's the
554  * sender will sit on his end and only queue most of his data, because
555  * he can only send snd_cwnd unacked packets at any given time.  For
556  * each ACK we send, he increments snd_cwnd and transmits more of his
557  * queue.  -DaveM
558  */
559 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
560 {
561         struct tcp_sock *tp = tcp_sk(sk);
562         struct inet_connection_sock *icsk = inet_csk(sk);
563         u32 now;
564
565         inet_csk_schedule_ack(sk);
566
567         tcp_measure_rcv_mss(sk, skb);
568
569         tcp_rcv_rtt_measure(tp);
570
571         now = tcp_time_stamp;
572
573         if (!icsk->icsk_ack.ato) {
574                 /* The _first_ data packet received, initialize
575                  * delayed ACK engine.
576                  */
577                 tcp_incr_quickack(sk);
578                 icsk->icsk_ack.ato = TCP_ATO_MIN;
579         } else {
580                 int m = now - icsk->icsk_ack.lrcvtime;
581
582                 if (m <= TCP_ATO_MIN/2) {
583                         /* The fastest case is the first. */
584                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
585                 } else if (m < icsk->icsk_ack.ato) {
586                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
587                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
588                                 icsk->icsk_ack.ato = icsk->icsk_rto;
589                 } else if (m > icsk->icsk_rto) {
590                         /* Too long gap. Apparently sender failed to
591                          * restart window, so that we send ACKs quickly.
592                          */
593                         tcp_incr_quickack(sk);
594                         sk_stream_mem_reclaim(sk);
595                 }
596         }
597         icsk->icsk_ack.lrcvtime = now;
598
599         TCP_ECN_check_ce(tp, skb);
600
601         if (skb->len >= 128)
602                 tcp_grow_window(sk, skb);
603 }
604
605 static u32 tcp_rto_min(struct sock *sk)
606 {
607         struct dst_entry *dst = __sk_dst_get(sk);
608         u32 rto_min = TCP_RTO_MIN;
609
610         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
611                 rto_min = dst->metrics[RTAX_RTO_MIN-1];
612         return rto_min;
613 }
614
615 /* Called to compute a smoothed rtt estimate. The data fed to this
616  * routine either comes from timestamps, or from segments that were
617  * known _not_ to have been retransmitted [see Karn/Partridge
618  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
619  * piece by Van Jacobson.
620  * NOTE: the next three routines used to be one big routine.
621  * To save cycles in the RFC 1323 implementation it was better to break
622  * it up into three procedures. -- erics
623  */
624 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
625 {
626         struct tcp_sock *tp = tcp_sk(sk);
627         long m = mrtt; /* RTT */
628
629         /*      The following amusing code comes from Jacobson's
630          *      article in SIGCOMM '88.  Note that rtt and mdev
631          *      are scaled versions of rtt and mean deviation.
632          *      This is designed to be as fast as possible
633          *      m stands for "measurement".
634          *
635          *      On a 1990 paper the rto value is changed to:
636          *      RTO = rtt + 4 * mdev
637          *
638          * Funny. This algorithm seems to be very broken.
639          * These formulae increase RTO, when it should be decreased, increase
640          * too slowly, when it should be increased quickly, decrease too quickly
641          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
642          * does not matter how to _calculate_ it. Seems, it was trap
643          * that VJ failed to avoid. 8)
644          */
645         if (m == 0)
646                 m = 1;
647         if (tp->srtt != 0) {
648                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
649                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
650                 if (m < 0) {
651                         m = -m;         /* m is now abs(error) */
652                         m -= (tp->mdev >> 2);   /* similar update on mdev */
653                         /* This is similar to one of Eifel findings.
654                          * Eifel blocks mdev updates when rtt decreases.
655                          * This solution is a bit different: we use finer gain
656                          * for mdev in this case (alpha*beta).
657                          * Like Eifel it also prevents growth of rto,
658                          * but also it limits too fast rto decreases,
659                          * happening in pure Eifel.
660                          */
661                         if (m > 0)
662                                 m >>= 3;
663                 } else {
664                         m -= (tp->mdev >> 2);   /* similar update on mdev */
665                 }
666                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
667                 if (tp->mdev > tp->mdev_max) {
668                         tp->mdev_max = tp->mdev;
669                         if (tp->mdev_max > tp->rttvar)
670                                 tp->rttvar = tp->mdev_max;
671                 }
672                 if (after(tp->snd_una, tp->rtt_seq)) {
673                         if (tp->mdev_max < tp->rttvar)
674                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
675                         tp->rtt_seq = tp->snd_nxt;
676                         tp->mdev_max = tcp_rto_min(sk);
677                 }
678         } else {
679                 /* no previous measure. */
680                 tp->srtt = m<<3;        /* take the measured time to be rtt */
681                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
682                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
683                 tp->rtt_seq = tp->snd_nxt;
684         }
685 }
686
687 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
688  * routine referred to above.
689  */
690 static inline void tcp_set_rto(struct sock *sk)
691 {
692         const struct tcp_sock *tp = tcp_sk(sk);
693         /* Old crap is replaced with new one. 8)
694          *
695          * More seriously:
696          * 1. If rtt variance happened to be less 50msec, it is hallucination.
697          *    It cannot be less due to utterly erratic ACK generation made
698          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
699          *    to do with delayed acks, because at cwnd>2 true delack timeout
700          *    is invisible. Actually, Linux-2.4 also generates erratic
701          *    ACKs in some circumstances.
702          */
703         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
704
705         /* 2. Fixups made earlier cannot be right.
706          *    If we do not estimate RTO correctly without them,
707          *    all the algo is pure shit and should be replaced
708          *    with correct one. It is exactly, which we pretend to do.
709          */
710 }
711
712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
713  * guarantees that rto is higher.
714  */
715 static inline void tcp_bound_rto(struct sock *sk)
716 {
717         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
718                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
719 }
720
721 /* Save metrics learned by this TCP session.
722    This function is called only, when TCP finishes successfully
723    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
724  */
725 void tcp_update_metrics(struct sock *sk)
726 {
727         struct tcp_sock *tp = tcp_sk(sk);
728         struct dst_entry *dst = __sk_dst_get(sk);
729
730         if (sysctl_tcp_nometrics_save)
731                 return;
732
733         dst_confirm(dst);
734
735         if (dst && (dst->flags&DST_HOST)) {
736                 const struct inet_connection_sock *icsk = inet_csk(sk);
737                 int m;
738
739                 if (icsk->icsk_backoff || !tp->srtt) {
740                         /* This session failed to estimate rtt. Why?
741                          * Probably, no packets returned in time.
742                          * Reset our results.
743                          */
744                         if (!(dst_metric_locked(dst, RTAX_RTT)))
745                                 dst->metrics[RTAX_RTT-1] = 0;
746                         return;
747                 }
748
749                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
750
751                 /* If newly calculated rtt larger than stored one,
752                  * store new one. Otherwise, use EWMA. Remember,
753                  * rtt overestimation is always better than underestimation.
754                  */
755                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
756                         if (m <= 0)
757                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
758                         else
759                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
760                 }
761
762                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
763                         if (m < 0)
764                                 m = -m;
765
766                         /* Scale deviation to rttvar fixed point */
767                         m >>= 1;
768                         if (m < tp->mdev)
769                                 m = tp->mdev;
770
771                         if (m >= dst_metric(dst, RTAX_RTTVAR))
772                                 dst->metrics[RTAX_RTTVAR-1] = m;
773                         else
774                                 dst->metrics[RTAX_RTTVAR-1] -=
775                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
776                 }
777
778                 if (tp->snd_ssthresh >= 0xFFFF) {
779                         /* Slow start still did not finish. */
780                         if (dst_metric(dst, RTAX_SSTHRESH) &&
781                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
782                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
783                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
784                         if (!dst_metric_locked(dst, RTAX_CWND) &&
785                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
786                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
787                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
788                            icsk->icsk_ca_state == TCP_CA_Open) {
789                         /* Cong. avoidance phase, cwnd is reliable. */
790                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
791                                 dst->metrics[RTAX_SSTHRESH-1] =
792                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
793                         if (!dst_metric_locked(dst, RTAX_CWND))
794                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
795                 } else {
796                         /* Else slow start did not finish, cwnd is non-sense,
797                            ssthresh may be also invalid.
798                          */
799                         if (!dst_metric_locked(dst, RTAX_CWND))
800                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
801                         if (dst->metrics[RTAX_SSTHRESH-1] &&
802                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
803                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
804                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
805                 }
806
807                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
808                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
809                             tp->reordering != sysctl_tcp_reordering)
810                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
811                 }
812         }
813 }
814
815 /* Numbers are taken from RFC3390.
816  *
817  * John Heffner states:
818  *
819  *      The RFC specifies a window of no more than 4380 bytes
820  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
821  *      is a bit misleading because they use a clamp at 4380 bytes
822  *      rather than use a multiplier in the relevant range.
823  */
824 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
825 {
826         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
827
828         if (!cwnd) {
829                 if (tp->mss_cache > 1460)
830                         cwnd = 2;
831                 else
832                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
833         }
834         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
835 }
836
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
839 {
840         struct tcp_sock *tp = tcp_sk(sk);
841         const struct inet_connection_sock *icsk = inet_csk(sk);
842
843         tp->prior_ssthresh = 0;
844         tp->bytes_acked = 0;
845         if (icsk->icsk_ca_state < TCP_CA_CWR) {
846                 tp->undo_marker = 0;
847                 if (set_ssthresh)
848                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
849                 tp->snd_cwnd = min(tp->snd_cwnd,
850                                    tcp_packets_in_flight(tp) + 1U);
851                 tp->snd_cwnd_cnt = 0;
852                 tp->high_seq = tp->snd_nxt;
853                 tp->snd_cwnd_stamp = tcp_time_stamp;
854                 TCP_ECN_queue_cwr(tp);
855
856                 tcp_set_ca_state(sk, TCP_CA_CWR);
857         }
858 }
859
860 /*
861  * Packet counting of FACK is based on in-order assumptions, therefore TCP
862  * disables it when reordering is detected
863  */
864 static void tcp_disable_fack(struct tcp_sock *tp)
865 {
866         /* RFC3517 uses different metric in lost marker => reset on change */
867         if (tcp_is_fack(tp))
868                 tp->lost_skb_hint = NULL;
869         tp->rx_opt.sack_ok &= ~2;
870 }
871
872 /* Take a notice that peer is sending D-SACKs */
873 static void tcp_dsack_seen(struct tcp_sock *tp)
874 {
875         tp->rx_opt.sack_ok |= 4;
876 }
877
878 /* Initialize metrics on socket. */
879
880 static void tcp_init_metrics(struct sock *sk)
881 {
882         struct tcp_sock *tp = tcp_sk(sk);
883         struct dst_entry *dst = __sk_dst_get(sk);
884
885         if (dst == NULL)
886                 goto reset;
887
888         dst_confirm(dst);
889
890         if (dst_metric_locked(dst, RTAX_CWND))
891                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
892         if (dst_metric(dst, RTAX_SSTHRESH)) {
893                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
894                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
895                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
896         }
897         if (dst_metric(dst, RTAX_REORDERING) &&
898             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
899                 tcp_disable_fack(tp);
900                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
901         }
902
903         if (dst_metric(dst, RTAX_RTT) == 0)
904                 goto reset;
905
906         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
907                 goto reset;
908
909         /* Initial rtt is determined from SYN,SYN-ACK.
910          * The segment is small and rtt may appear much
911          * less than real one. Use per-dst memory
912          * to make it more realistic.
913          *
914          * A bit of theory. RTT is time passed after "normal" sized packet
915          * is sent until it is ACKed. In normal circumstances sending small
916          * packets force peer to delay ACKs and calculation is correct too.
917          * The algorithm is adaptive and, provided we follow specs, it
918          * NEVER underestimate RTT. BUT! If peer tries to make some clever
919          * tricks sort of "quick acks" for time long enough to decrease RTT
920          * to low value, and then abruptly stops to do it and starts to delay
921          * ACKs, wait for troubles.
922          */
923         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
924                 tp->srtt = dst_metric(dst, RTAX_RTT);
925                 tp->rtt_seq = tp->snd_nxt;
926         }
927         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
928                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
929                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
930         }
931         tcp_set_rto(sk);
932         tcp_bound_rto(sk);
933         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
934                 goto reset;
935         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
936         tp->snd_cwnd_stamp = tcp_time_stamp;
937         return;
938
939 reset:
940         /* Play conservative. If timestamps are not
941          * supported, TCP will fail to recalculate correct
942          * rtt, if initial rto is too small. FORGET ALL AND RESET!
943          */
944         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
945                 tp->srtt = 0;
946                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
947                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
948         }
949 }
950
951 static void tcp_update_reordering(struct sock *sk, const int metric,
952                                   const int ts)
953 {
954         struct tcp_sock *tp = tcp_sk(sk);
955         if (metric > tp->reordering) {
956                 tp->reordering = min(TCP_MAX_REORDERING, metric);
957
958                 /* This exciting event is worth to be remembered. 8) */
959                 if (ts)
960                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
961                 else if (tcp_is_reno(tp))
962                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
963                 else if (tcp_is_fack(tp))
964                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
965                 else
966                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
967 #if FASTRETRANS_DEBUG > 1
968                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
969                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
970                        tp->reordering,
971                        tp->fackets_out,
972                        tp->sacked_out,
973                        tp->undo_marker ? tp->undo_retrans : 0);
974 #endif
975                 tcp_disable_fack(tp);
976         }
977 }
978
979 /* This procedure tags the retransmission queue when SACKs arrive.
980  *
981  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
982  * Packets in queue with these bits set are counted in variables
983  * sacked_out, retrans_out and lost_out, correspondingly.
984  *
985  * Valid combinations are:
986  * Tag  InFlight        Description
987  * 0    1               - orig segment is in flight.
988  * S    0               - nothing flies, orig reached receiver.
989  * L    0               - nothing flies, orig lost by net.
990  * R    2               - both orig and retransmit are in flight.
991  * L|R  1               - orig is lost, retransmit is in flight.
992  * S|R  1               - orig reached receiver, retrans is still in flight.
993  * (L|S|R is logically valid, it could occur when L|R is sacked,
994  *  but it is equivalent to plain S and code short-curcuits it to S.
995  *  L|S is logically invalid, it would mean -1 packet in flight 8))
996  *
997  * These 6 states form finite state machine, controlled by the following events:
998  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
999  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1000  * 3. Loss detection event of one of three flavors:
1001  *      A. Scoreboard estimator decided the packet is lost.
1002  *         A'. Reno "three dupacks" marks head of queue lost.
1003  *         A''. Its FACK modfication, head until snd.fack is lost.
1004  *      B. SACK arrives sacking data transmitted after never retransmitted
1005  *         hole was sent out.
1006  *      C. SACK arrives sacking SND.NXT at the moment, when the
1007  *         segment was retransmitted.
1008  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1009  *
1010  * It is pleasant to note, that state diagram turns out to be commutative,
1011  * so that we are allowed not to be bothered by order of our actions,
1012  * when multiple events arrive simultaneously. (see the function below).
1013  *
1014  * Reordering detection.
1015  * --------------------
1016  * Reordering metric is maximal distance, which a packet can be displaced
1017  * in packet stream. With SACKs we can estimate it:
1018  *
1019  * 1. SACK fills old hole and the corresponding segment was not
1020  *    ever retransmitted -> reordering. Alas, we cannot use it
1021  *    when segment was retransmitted.
1022  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1023  *    for retransmitted and already SACKed segment -> reordering..
1024  * Both of these heuristics are not used in Loss state, when we cannot
1025  * account for retransmits accurately.
1026  *
1027  * SACK block validation.
1028  * ----------------------
1029  *
1030  * SACK block range validation checks that the received SACK block fits to
1031  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1032  * Note that SND.UNA is not included to the range though being valid because
1033  * it means that the receiver is rather inconsistent with itself reporting
1034  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1035  * perfectly valid, however, in light of RFC2018 which explicitly states
1036  * that "SACK block MUST reflect the newest segment.  Even if the newest
1037  * segment is going to be discarded ...", not that it looks very clever
1038  * in case of head skb. Due to potentional receiver driven attacks, we
1039  * choose to avoid immediate execution of a walk in write queue due to
1040  * reneging and defer head skb's loss recovery to standard loss recovery
1041  * procedure that will eventually trigger (nothing forbids us doing this).
1042  *
1043  * Implements also blockage to start_seq wrap-around. Problem lies in the
1044  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1045  * there's no guarantee that it will be before snd_nxt (n). The problem
1046  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1047  * wrap (s_w):
1048  *
1049  *         <- outs wnd ->                          <- wrapzone ->
1050  *         u     e      n                         u_w   e_w  s n_w
1051  *         |     |      |                          |     |   |  |
1052  * |<------------+------+----- TCP seqno space --------------+---------->|
1053  * ...-- <2^31 ->|                                           |<--------...
1054  * ...---- >2^31 ------>|                                    |<--------...
1055  *
1056  * Current code wouldn't be vulnerable but it's better still to discard such
1057  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1058  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1059  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1060  * equal to the ideal case (infinite seqno space without wrap caused issues).
1061  *
1062  * With D-SACK the lower bound is extended to cover sequence space below
1063  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1064  * again, D-SACK block must not to go across snd_una (for the same reason as
1065  * for the normal SACK blocks, explained above). But there all simplicity
1066  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1067  * fully below undo_marker they do not affect behavior in anyway and can
1068  * therefore be safely ignored. In rare cases (which are more or less
1069  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1070  * fragmentation and packet reordering past skb's retransmission. To consider
1071  * them correctly, the acceptable range must be extended even more though
1072  * the exact amount is rather hard to quantify. However, tp->max_window can
1073  * be used as an exaggerated estimate.
1074  */
1075 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1076                                   u32 start_seq, u32 end_seq)
1077 {
1078         /* Too far in future, or reversed (interpretation is ambiguous) */
1079         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1080                 return 0;
1081
1082         /* Nasty start_seq wrap-around check (see comments above) */
1083         if (!before(start_seq, tp->snd_nxt))
1084                 return 0;
1085
1086         /* In outstanding window? ...This is valid exit for D-SACKs too.
1087          * start_seq == snd_una is non-sensical (see comments above)
1088          */
1089         if (after(start_seq, tp->snd_una))
1090                 return 1;
1091
1092         if (!is_dsack || !tp->undo_marker)
1093                 return 0;
1094
1095         /* ...Then it's D-SACK, and must reside below snd_una completely */
1096         if (!after(end_seq, tp->snd_una))
1097                 return 0;
1098
1099         if (!before(start_seq, tp->undo_marker))
1100                 return 1;
1101
1102         /* Too old */
1103         if (!after(end_seq, tp->undo_marker))
1104                 return 0;
1105
1106         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1107          *   start_seq < undo_marker and end_seq >= undo_marker.
1108          */
1109         return !before(start_seq, end_seq - tp->max_window);
1110 }
1111
1112 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1113  * Event "C". Later note: FACK people cheated me again 8), we have to account
1114  * for reordering! Ugly, but should help.
1115  *
1116  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1117  * less than what is now known to be received by the other end (derived from
1118  * SACK blocks by the caller). Also calculate the lowest snd_nxt among the
1119  * remaining retransmitted skbs to avoid some costly processing per ACKs.
1120  */
1121 static int tcp_mark_lost_retrans(struct sock *sk, u32 received_upto)
1122 {
1123         struct tcp_sock *tp = tcp_sk(sk);
1124         struct sk_buff *skb;
1125         int flag = 0;
1126         int cnt = 0;
1127         u32 new_low_seq = tp->snd_nxt;
1128
1129         tcp_for_write_queue(skb, sk) {
1130                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1131
1132                 if (skb == tcp_send_head(sk))
1133                         break;
1134                 if (cnt == tp->retrans_out)
1135                         break;
1136                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1137                         continue;
1138
1139                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1140                         continue;
1141
1142                 if (after(received_upto, ack_seq) &&
1143                     (tcp_is_fack(tp) ||
1144                      !before(received_upto,
1145                              ack_seq + tp->reordering * tp->mss_cache))) {
1146                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1147                         tp->retrans_out -= tcp_skb_pcount(skb);
1148
1149                         /* clear lost hint */
1150                         tp->retransmit_skb_hint = NULL;
1151
1152                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1153                                 tp->lost_out += tcp_skb_pcount(skb);
1154                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1155                                 flag |= FLAG_DATA_SACKED;
1156                                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1157                         }
1158                 } else {
1159                         if (before(ack_seq, new_low_seq))
1160                                 new_low_seq = ack_seq;
1161                         cnt += tcp_skb_pcount(skb);
1162                 }
1163         }
1164
1165         if (tp->retrans_out)
1166                 tp->lost_retrans_low = new_low_seq;
1167
1168         return flag;
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
1237 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1238 {
1239         const struct inet_connection_sock *icsk = inet_csk(sk);
1240         struct tcp_sock *tp = tcp_sk(sk);
1241         unsigned char *ptr = (skb_transport_header(ack_skb) +
1242                               TCP_SKB_CB(ack_skb)->sacked);
1243         struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1244         struct sk_buff *cached_skb;
1245         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1246         int reord = tp->packets_out;
1247         int prior_fackets;
1248         u32 highest_sack_end_seq;
1249         int flag = 0;
1250         int found_dup_sack = 0;
1251         int cached_fack_count;
1252         int i;
1253         int first_sack_index;
1254         int force_one_sack;
1255
1256         if (!tp->sacked_out) {
1257                 if (WARN_ON(tp->fackets_out))
1258                         tp->fackets_out = 0;
1259                 tp->highest_sack = tcp_write_queue_head(sk);
1260         }
1261         prior_fackets = tp->fackets_out;
1262
1263         found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1264                                          num_sacks, prior_snd_una);
1265         if (found_dup_sack)
1266                 flag |= FLAG_DSACKING_ACK;
1267
1268         /* Eliminate too old ACKs, but take into
1269          * account more or less fresh ones, they can
1270          * contain valid SACK info.
1271          */
1272         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1273                 return 0;
1274
1275         if (!tp->packets_out)
1276                 goto out;
1277
1278         /* SACK fastpath:
1279          * if the only SACK change is the increase of the end_seq of
1280          * the first block then only apply that SACK block
1281          * and use retrans queue hinting otherwise slowpath */
1282         force_one_sack = 1;
1283         for (i = 0; i < num_sacks; i++) {
1284                 __be32 start_seq = sp[i].start_seq;
1285                 __be32 end_seq = sp[i].end_seq;
1286
1287                 if (i == 0) {
1288                         if (tp->recv_sack_cache[i].start_seq != start_seq)
1289                                 force_one_sack = 0;
1290                 } else {
1291                         if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1292                             (tp->recv_sack_cache[i].end_seq != end_seq))
1293                                 force_one_sack = 0;
1294                 }
1295                 tp->recv_sack_cache[i].start_seq = start_seq;
1296                 tp->recv_sack_cache[i].end_seq = end_seq;
1297         }
1298         /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1299         for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1300                 tp->recv_sack_cache[i].start_seq = 0;
1301                 tp->recv_sack_cache[i].end_seq = 0;
1302         }
1303
1304         first_sack_index = 0;
1305         if (force_one_sack)
1306                 num_sacks = 1;
1307         else {
1308                 int j;
1309                 tp->fastpath_skb_hint = NULL;
1310
1311                 /* order SACK blocks to allow in order walk of the retrans queue */
1312                 for (i = num_sacks-1; i > 0; i--) {
1313                         for (j = 0; j < i; j++){
1314                                 if (after(ntohl(sp[j].start_seq),
1315                                           ntohl(sp[j+1].start_seq))){
1316                                         struct tcp_sack_block_wire tmp;
1317
1318                                         tmp = sp[j];
1319                                         sp[j] = sp[j+1];
1320                                         sp[j+1] = tmp;
1321
1322                                         /* Track where the first SACK block goes to */
1323                                         if (j == first_sack_index)
1324                                                 first_sack_index = j+1;
1325                                 }
1326
1327                         }
1328                 }
1329         }
1330
1331         /* Use SACK fastpath hint if valid */
1332         cached_skb = tp->fastpath_skb_hint;
1333         cached_fack_count = tp->fastpath_cnt_hint;
1334         if (!cached_skb) {
1335                 cached_skb = tcp_write_queue_head(sk);
1336                 cached_fack_count = 0;
1337         }
1338
1339         for (i = 0; i < num_sacks; i++) {
1340                 struct sk_buff *skb;
1341                 __u32 start_seq = ntohl(sp->start_seq);
1342                 __u32 end_seq = ntohl(sp->end_seq);
1343                 int fack_count;
1344                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1345                 int next_dup = (found_dup_sack && (i+1 == first_sack_index));
1346
1347                 sp++;
1348
1349                 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
1350                         if (dup_sack) {
1351                                 if (!tp->undo_marker)
1352                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1353                                 else
1354                                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1355                         } else {
1356                                 /* Don't count olds caused by ACK reordering */
1357                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1358                                     !after(end_seq, tp->snd_una))
1359                                         continue;
1360                                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1361                         }
1362                         continue;
1363                 }
1364
1365                 skb = cached_skb;
1366                 fack_count = cached_fack_count;
1367
1368                 /* Event "B" in the comment above. */
1369                 if (after(end_seq, tp->high_seq))
1370                         flag |= FLAG_DATA_LOST;
1371
1372                 tcp_for_write_queue_from(skb, sk) {
1373                         int in_sack = 0;
1374                         u8 sacked;
1375
1376                         if (skb == tcp_send_head(sk))
1377                                 break;
1378
1379                         cached_skb = skb;
1380                         cached_fack_count = fack_count;
1381                         if (i == first_sack_index) {
1382                                 tp->fastpath_skb_hint = skb;
1383                                 tp->fastpath_cnt_hint = fack_count;
1384                         }
1385
1386                         /* The retransmission queue is always in order, so
1387                          * we can short-circuit the walk early.
1388                          */
1389                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1390                                 break;
1391
1392                         dup_sack = (found_dup_sack && (i == first_sack_index));
1393
1394                         /* Due to sorting DSACK may reside within this SACK block! */
1395                         if (next_dup) {
1396                                 u32 dup_start = ntohl(sp->start_seq);
1397                                 u32 dup_end = ntohl(sp->end_seq);
1398
1399                                 if (before(TCP_SKB_CB(skb)->seq, dup_end)) {
1400                                         in_sack = tcp_match_skb_to_sack(sk, skb, dup_start, dup_end);
1401                                         if (in_sack > 0)
1402                                                 dup_sack = 1;
1403                                 }
1404                         }
1405
1406                         /* DSACK info lost if out-of-mem, try SACK still */
1407                         if (in_sack <= 0)
1408                                 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq);
1409                         if (unlikely(in_sack < 0))
1410                                 break;
1411
1412                         if (!in_sack) {
1413                                 fack_count += tcp_skb_pcount(skb);
1414                                 continue;
1415                         }
1416
1417                         sacked = TCP_SKB_CB(skb)->sacked;
1418
1419                         /* Account D-SACK for retransmitted packet. */
1420                         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1421                                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1422                                         tp->undo_retrans--;
1423                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una) &&
1424                                     (sacked & TCPCB_SACKED_ACKED))
1425                                         reord = min(fack_count, reord);
1426                         }
1427
1428
1429                         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1430                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1431                                 fack_count += tcp_skb_pcount(skb);
1432                                 continue;
1433                         }
1434
1435                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1436                                 if (sacked & TCPCB_SACKED_RETRANS) {
1437                                         /* If the segment is not tagged as lost,
1438                                          * we do not clear RETRANS, believing
1439                                          * that retransmission is still in flight.
1440                                          */
1441                                         if (sacked & TCPCB_LOST) {
1442                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1443                                                 tp->lost_out -= tcp_skb_pcount(skb);
1444                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1445
1446                                                 /* clear lost hint */
1447                                                 tp->retransmit_skb_hint = NULL;
1448                                         }
1449                                 } else {
1450                                         if (!(sacked & TCPCB_RETRANS)) {
1451                                                 /* New sack for not retransmitted frame,
1452                                                  * which was in hole. It is reordering.
1453                                                  */
1454                                                 if (fack_count < prior_fackets)
1455                                                         reord = min(fack_count, reord);
1456
1457                                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1458                                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1459                                                         flag |= FLAG_ONLY_ORIG_SACKED;
1460                                         }
1461
1462                                         if (sacked & TCPCB_LOST) {
1463                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1464                                                 tp->lost_out -= tcp_skb_pcount(skb);
1465
1466                                                 /* clear lost hint */
1467                                                 tp->retransmit_skb_hint = NULL;
1468                                         }
1469                                 }
1470
1471                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1472                                 flag |= FLAG_DATA_SACKED;
1473                                 tp->sacked_out += tcp_skb_pcount(skb);
1474
1475                                 fack_count += tcp_skb_pcount(skb);
1476
1477                                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1478                                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1479                                     before(TCP_SKB_CB(skb)->seq,
1480                                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1481                                         tp->lost_cnt_hint += tcp_skb_pcount(skb);
1482
1483                                 if (fack_count > tp->fackets_out)
1484                                         tp->fackets_out = fack_count;
1485
1486                                 if (after(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1487                                         tp->highest_sack = skb;
1488
1489                         } else {
1490                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1491                                         reord = min(fack_count, reord);
1492
1493                                 fack_count += tcp_skb_pcount(skb);
1494                         }
1495
1496                         /* D-SACK. We can detect redundant retransmission
1497                          * in S|R and plain R frames and clear it.
1498                          * undo_retrans is decreased above, L|R frames
1499                          * are accounted above as well.
1500                          */
1501                         if (dup_sack &&
1502                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1503                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1504                                 tp->retrans_out -= tcp_skb_pcount(skb);
1505                                 tp->retransmit_skb_hint = NULL;
1506                         }
1507                 }
1508
1509                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1510                  * due to in-order walk
1511                  */
1512                 if (after(end_seq, tp->frto_highmark))
1513                         flag &= ~FLAG_ONLY_ORIG_SACKED;
1514         }
1515
1516         highest_sack_end_seq = TCP_SKB_CB(tp->highest_sack)->end_seq;
1517         if (tcp_is_fack(tp) && tp->retrans_out &&
1518             after(highest_sack_end_seq, tp->lost_retrans_low) &&
1519             icsk->icsk_ca_state == TCP_CA_Recovery)
1520                 flag |= tcp_mark_lost_retrans(sk, highest_sack_end_seq);
1521
1522         tcp_verify_left_out(tp);
1523
1524         if ((reord < tp->fackets_out) &&
1525             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1526             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1527                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1528
1529 out:
1530
1531 #if FASTRETRANS_DEBUG > 0
1532         BUG_TRAP((int)tp->sacked_out >= 0);
1533         BUG_TRAP((int)tp->lost_out >= 0);
1534         BUG_TRAP((int)tp->retrans_out >= 0);
1535         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1536 #endif
1537         return flag;
1538 }
1539
1540 /* If we receive more dupacks than we expected counting segments
1541  * in assumption of absent reordering, interpret this as reordering.
1542  * The only another reason could be bug in receiver TCP.
1543  */
1544 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1545 {
1546         struct tcp_sock *tp = tcp_sk(sk);
1547         u32 holes;
1548
1549         holes = max(tp->lost_out, 1U);
1550         holes = min(holes, tp->packets_out);
1551
1552         if ((tp->sacked_out + holes) > tp->packets_out) {
1553                 tp->sacked_out = tp->packets_out - holes;
1554                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1555         }
1556 }
1557
1558 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1559
1560 static void tcp_add_reno_sack(struct sock *sk)
1561 {
1562         struct tcp_sock *tp = tcp_sk(sk);
1563         tp->sacked_out++;
1564         tcp_check_reno_reordering(sk, 0);
1565         tcp_verify_left_out(tp);
1566 }
1567
1568 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1569
1570 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1571 {
1572         struct tcp_sock *tp = tcp_sk(sk);
1573
1574         if (acked > 0) {
1575                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1576                 if (acked-1 >= tp->sacked_out)
1577                         tp->sacked_out = 0;
1578                 else
1579                         tp->sacked_out -= acked-1;
1580         }
1581         tcp_check_reno_reordering(sk, acked);
1582         tcp_verify_left_out(tp);
1583 }
1584
1585 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1586 {
1587         tp->sacked_out = 0;
1588 }
1589
1590 /* F-RTO can only be used if TCP has never retransmitted anything other than
1591  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1592  */
1593 int tcp_use_frto(struct sock *sk)
1594 {
1595         const struct tcp_sock *tp = tcp_sk(sk);
1596         struct sk_buff *skb;
1597
1598         if (!sysctl_tcp_frto)
1599                 return 0;
1600
1601         if (IsSackFrto())
1602                 return 1;
1603
1604         /* Avoid expensive walking of rexmit queue if possible */
1605         if (tp->retrans_out > 1)
1606                 return 0;
1607
1608         skb = tcp_write_queue_head(sk);
1609         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1610         tcp_for_write_queue_from(skb, sk) {
1611                 if (skb == tcp_send_head(sk))
1612                         break;
1613                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1614                         return 0;
1615                 /* Short-circuit when first non-SACKed skb has been checked */
1616                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1617                         break;
1618         }
1619         return 1;
1620 }
1621
1622 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1623  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1624  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1625  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1626  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1627  * bits are handled if the Loss state is really to be entered (in
1628  * tcp_enter_frto_loss).
1629  *
1630  * Do like tcp_enter_loss() would; when RTO expires the second time it
1631  * does:
1632  *  "Reduce ssthresh if it has not yet been made inside this window."
1633  */
1634 void tcp_enter_frto(struct sock *sk)
1635 {
1636         const struct inet_connection_sock *icsk = inet_csk(sk);
1637         struct tcp_sock *tp = tcp_sk(sk);
1638         struct sk_buff *skb;
1639
1640         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1641             tp->snd_una == tp->high_seq ||
1642             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1643              !icsk->icsk_retransmits)) {
1644                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1645                 /* Our state is too optimistic in ssthresh() call because cwnd
1646                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1647                  * recovery has not yet completed. Pattern would be this: RTO,
1648                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1649                  * up here twice).
1650                  * RFC4138 should be more specific on what to do, even though
1651                  * RTO is quite unlikely to occur after the first Cumulative ACK
1652                  * due to back-off and complexity of triggering events ...
1653                  */
1654                 if (tp->frto_counter) {
1655                         u32 stored_cwnd;
1656                         stored_cwnd = tp->snd_cwnd;
1657                         tp->snd_cwnd = 2;
1658                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1659                         tp->snd_cwnd = stored_cwnd;
1660                 } else {
1661                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1662                 }
1663                 /* ... in theory, cong.control module could do "any tricks" in
1664                  * ssthresh(), which means that ca_state, lost bits and lost_out
1665                  * counter would have to be faked before the call occurs. We
1666                  * consider that too expensive, unlikely and hacky, so modules
1667                  * using these in ssthresh() must deal these incompatibility
1668                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1669                  */
1670                 tcp_ca_event(sk, CA_EVENT_FRTO);
1671         }
1672
1673         tp->undo_marker = tp->snd_una;
1674         tp->undo_retrans = 0;
1675
1676         skb = tcp_write_queue_head(sk);
1677         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1678                 tp->undo_marker = 0;
1679         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1680                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1681                 tp->retrans_out -= tcp_skb_pcount(skb);
1682         }
1683         tcp_verify_left_out(tp);
1684
1685         /* Too bad if TCP was application limited */
1686         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1687
1688         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1689          * The last condition is necessary at least in tp->frto_counter case.
1690          */
1691         if (IsSackFrto() && (tp->frto_counter ||
1692             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1693             after(tp->high_seq, tp->snd_una)) {
1694                 tp->frto_highmark = tp->high_seq;
1695         } else {
1696                 tp->frto_highmark = tp->snd_nxt;
1697         }
1698         tcp_set_ca_state(sk, TCP_CA_Disorder);
1699         tp->high_seq = tp->snd_nxt;
1700         tp->frto_counter = 1;
1701 }
1702
1703 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1704  * which indicates that we should follow the traditional RTO recovery,
1705  * i.e. mark everything lost and do go-back-N retransmission.
1706  */
1707 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1708 {
1709         struct tcp_sock *tp = tcp_sk(sk);
1710         struct sk_buff *skb;
1711
1712         tp->lost_out = 0;
1713         tp->retrans_out = 0;
1714         if (tcp_is_reno(tp))
1715                 tcp_reset_reno_sack(tp);
1716
1717         tcp_for_write_queue(skb, sk) {
1718                 if (skb == tcp_send_head(sk))
1719                         break;
1720
1721                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1722                 /*
1723                  * Count the retransmission made on RTO correctly (only when
1724                  * waiting for the first ACK and did not get it)...
1725                  */
1726                 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1727                         /* For some reason this R-bit might get cleared? */
1728                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1729                                 tp->retrans_out += tcp_skb_pcount(skb);
1730                         /* ...enter this if branch just for the first segment */
1731                         flag |= FLAG_DATA_ACKED;
1732                 } else {
1733                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1734                                 tp->undo_marker = 0;
1735                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1736                 }
1737
1738                 /* Don't lost mark skbs that were fwd transmitted after RTO */
1739                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1740                     !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1741                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1742                         tp->lost_out += tcp_skb_pcount(skb);
1743                 }
1744         }
1745         tcp_verify_left_out(tp);
1746
1747         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1748         tp->snd_cwnd_cnt = 0;
1749         tp->snd_cwnd_stamp = tcp_time_stamp;
1750         tp->frto_counter = 0;
1751         tp->bytes_acked = 0;
1752
1753         tp->reordering = min_t(unsigned int, tp->reordering,
1754                                              sysctl_tcp_reordering);
1755         tcp_set_ca_state(sk, TCP_CA_Loss);
1756         tp->high_seq = tp->frto_highmark;
1757         TCP_ECN_queue_cwr(tp);
1758
1759         tcp_clear_retrans_hints_partial(tp);
1760 }
1761
1762 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1763 {
1764         tp->retrans_out = 0;
1765         tp->lost_out = 0;
1766
1767         tp->undo_marker = 0;
1768         tp->undo_retrans = 0;
1769 }
1770
1771 void tcp_clear_retrans(struct tcp_sock *tp)
1772 {
1773         tcp_clear_retrans_partial(tp);
1774
1775         tp->fackets_out = 0;
1776         tp->sacked_out = 0;
1777 }
1778
1779 /* Enter Loss state. If "how" is not zero, forget all SACK information
1780  * and reset tags completely, otherwise preserve SACKs. If receiver
1781  * dropped its ofo queue, we will know this due to reneging detection.
1782  */
1783 void tcp_enter_loss(struct sock *sk, int how)
1784 {
1785         const struct inet_connection_sock *icsk = inet_csk(sk);
1786         struct tcp_sock *tp = tcp_sk(sk);
1787         struct sk_buff *skb;
1788
1789         /* Reduce ssthresh if it has not yet been made inside this window. */
1790         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1791             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1792                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1793                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1794                 tcp_ca_event(sk, CA_EVENT_LOSS);
1795         }
1796         tp->snd_cwnd       = 1;
1797         tp->snd_cwnd_cnt   = 0;
1798         tp->snd_cwnd_stamp = tcp_time_stamp;
1799
1800         tp->bytes_acked = 0;
1801         tcp_clear_retrans_partial(tp);
1802
1803         if (tcp_is_reno(tp))
1804                 tcp_reset_reno_sack(tp);
1805
1806         if (!how) {
1807                 /* Push undo marker, if it was plain RTO and nothing
1808                  * was retransmitted. */
1809                 tp->undo_marker = tp->snd_una;
1810                 tcp_clear_retrans_hints_partial(tp);
1811         } else {
1812                 tp->sacked_out = 0;
1813                 tp->fackets_out = 0;
1814                 tcp_clear_all_retrans_hints(tp);
1815         }
1816
1817         tcp_for_write_queue(skb, sk) {
1818                 if (skb == tcp_send_head(sk))
1819                         break;
1820
1821                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1822                         tp->undo_marker = 0;
1823                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1824                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1825                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1826                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1827                         tp->lost_out += tcp_skb_pcount(skb);
1828                 }
1829         }
1830         tcp_verify_left_out(tp);
1831
1832         tp->reordering = min_t(unsigned int, tp->reordering,
1833                                              sysctl_tcp_reordering);
1834         tcp_set_ca_state(sk, TCP_CA_Loss);
1835         tp->high_seq = tp->snd_nxt;
1836         TCP_ECN_queue_cwr(tp);
1837         /* Abort F-RTO algorithm if one is in progress */
1838         tp->frto_counter = 0;
1839 }
1840
1841 static int tcp_check_sack_reneging(struct sock *sk)
1842 {
1843         struct sk_buff *skb;
1844
1845         /* If ACK arrived pointing to a remembered SACK,
1846          * it means that our remembered SACKs do not reflect
1847          * real state of receiver i.e.
1848          * receiver _host_ is heavily congested (or buggy).
1849          * Do processing similar to RTO timeout.
1850          */
1851         if ((skb = tcp_write_queue_head(sk)) != NULL &&
1852             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1853                 struct inet_connection_sock *icsk = inet_csk(sk);
1854                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1855
1856                 tcp_enter_loss(sk, 1);
1857                 icsk->icsk_retransmits++;
1858                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1859                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1860                                           icsk->icsk_rto, TCP_RTO_MAX);
1861                 return 1;
1862         }
1863         return 0;
1864 }
1865
1866 static inline int tcp_fackets_out(struct tcp_sock *tp)
1867 {
1868         return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1869 }
1870
1871 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1872  * counter when SACK is enabled (without SACK, sacked_out is used for
1873  * that purpose).
1874  *
1875  * Instead, with FACK TCP uses fackets_out that includes both SACKed
1876  * segments up to the highest received SACK block so far and holes in
1877  * between them.
1878  *
1879  * With reordering, holes may still be in flight, so RFC3517 recovery
1880  * uses pure sacked_out (total number of SACKed segments) even though
1881  * it violates the RFC that uses duplicate ACKs, often these are equal
1882  * but when e.g. out-of-window ACKs or packet duplication occurs,
1883  * they differ. Since neither occurs due to loss, TCP should really
1884  * ignore them.
1885  */
1886 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
1887 {
1888         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1889 }
1890
1891 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1892 {
1893         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1894 }
1895
1896 static inline int tcp_head_timedout(struct sock *sk)
1897 {
1898         struct tcp_sock *tp = tcp_sk(sk);
1899
1900         return tp->packets_out &&
1901                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1902 }
1903
1904 /* Linux NewReno/SACK/FACK/ECN state machine.
1905  * --------------------------------------
1906  *
1907  * "Open"       Normal state, no dubious events, fast path.
1908  * "Disorder"   In all the respects it is "Open",
1909  *              but requires a bit more attention. It is entered when
1910  *              we see some SACKs or dupacks. It is split of "Open"
1911  *              mainly to move some processing from fast path to slow one.
1912  * "CWR"        CWND was reduced due to some Congestion Notification event.
1913  *              It can be ECN, ICMP source quench, local device congestion.
1914  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1915  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1916  *
1917  * tcp_fastretrans_alert() is entered:
1918  * - each incoming ACK, if state is not "Open"
1919  * - when arrived ACK is unusual, namely:
1920  *      * SACK
1921  *      * Duplicate ACK.
1922  *      * ECN ECE.
1923  *
1924  * Counting packets in flight is pretty simple.
1925  *
1926  *      in_flight = packets_out - left_out + retrans_out
1927  *
1928  *      packets_out is SND.NXT-SND.UNA counted in packets.
1929  *
1930  *      retrans_out is number of retransmitted segments.
1931  *
1932  *      left_out is number of segments left network, but not ACKed yet.
1933  *
1934  *              left_out = sacked_out + lost_out
1935  *
1936  *     sacked_out: Packets, which arrived to receiver out of order
1937  *                 and hence not ACKed. With SACKs this number is simply
1938  *                 amount of SACKed data. Even without SACKs
1939  *                 it is easy to give pretty reliable estimate of this number,
1940  *                 counting duplicate ACKs.
1941  *
1942  *       lost_out: Packets lost by network. TCP has no explicit
1943  *                 "loss notification" feedback from network (for now).
1944  *                 It means that this number can be only _guessed_.
1945  *                 Actually, it is the heuristics to predict lossage that
1946  *                 distinguishes different algorithms.
1947  *
1948  *      F.e. after RTO, when all the queue is considered as lost,
1949  *      lost_out = packets_out and in_flight = retrans_out.
1950  *
1951  *              Essentially, we have now two algorithms counting
1952  *              lost packets.
1953  *
1954  *              FACK: It is the simplest heuristics. As soon as we decided
1955  *              that something is lost, we decide that _all_ not SACKed
1956  *              packets until the most forward SACK are lost. I.e.
1957  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1958  *              It is absolutely correct estimate, if network does not reorder
1959  *              packets. And it loses any connection to reality when reordering
1960  *              takes place. We use FACK by default until reordering
1961  *              is suspected on the path to this destination.
1962  *
1963  *              NewReno: when Recovery is entered, we assume that one segment
1964  *              is lost (classic Reno). While we are in Recovery and
1965  *              a partial ACK arrives, we assume that one more packet
1966  *              is lost (NewReno). This heuristics are the same in NewReno
1967  *              and SACK.
1968  *
1969  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1970  *  deflation etc. CWND is real congestion window, never inflated, changes
1971  *  only according to classic VJ rules.
1972  *
1973  * Really tricky (and requiring careful tuning) part of algorithm
1974  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1975  * The first determines the moment _when_ we should reduce CWND and,
1976  * hence, slow down forward transmission. In fact, it determines the moment
1977  * when we decide that hole is caused by loss, rather than by a reorder.
1978  *
1979  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1980  * holes, caused by lost packets.
1981  *
1982  * And the most logically complicated part of algorithm is undo
1983  * heuristics. We detect false retransmits due to both too early
1984  * fast retransmit (reordering) and underestimated RTO, analyzing
1985  * timestamps and D-SACKs. When we detect that some segments were
1986  * retransmitted by mistake and CWND reduction was wrong, we undo
1987  * window reduction and abort recovery phase. This logic is hidden
1988  * inside several functions named tcp_try_undo_<something>.
1989  */
1990
1991 /* This function decides, when we should leave Disordered state
1992  * and enter Recovery phase, reducing congestion window.
1993  *
1994  * Main question: may we further continue forward transmission
1995  * with the same cwnd?
1996  */
1997 static int tcp_time_to_recover(struct sock *sk)
1998 {
1999         struct tcp_sock *tp = tcp_sk(sk);
2000         __u32 packets_out;
2001
2002         /* Do not perform any recovery during F-RTO algorithm */
2003         if (tp->frto_counter)
2004                 return 0;
2005
2006         /* Trick#1: The loss is proven. */
2007         if (tp->lost_out)
2008                 return 1;
2009
2010         /* Not-A-Trick#2 : Classic rule... */
2011         if (tcp_dupack_heurestics(tp) > tp->reordering)
2012                 return 1;
2013
2014         /* Trick#3 : when we use RFC2988 timer restart, fast
2015          * retransmit can be triggered by timeout of queue head.
2016          */
2017         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2018                 return 1;
2019
2020         /* Trick#4: It is still not OK... But will it be useful to delay
2021          * recovery more?
2022          */
2023         packets_out = tp->packets_out;
2024         if (packets_out <= tp->reordering &&
2025             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2026             !tcp_may_send_now(sk)) {
2027                 /* We have nothing to send. This connection is limited
2028                  * either by receiver window or by application.
2029                  */
2030                 return 1;
2031         }
2032
2033         return 0;
2034 }
2035
2036 /* RFC: This is from the original, I doubt that this is necessary at all:
2037  * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2038  * retransmitted past LOST markings in the first place? I'm not fully sure
2039  * about undo and end of connection cases, which can cause R without L?
2040  */
2041 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
2042                                        struct sk_buff *skb)
2043 {
2044         if ((tp->retransmit_skb_hint != NULL) &&
2045             before(TCP_SKB_CB(skb)->seq,
2046             TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
2047                 tp->retransmit_skb_hint = NULL;
2048 }
2049
2050 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2051  * is against sacked "cnt", otherwise it's against facked "cnt"
2052  */
2053 static void tcp_mark_head_lost(struct sock *sk, int packets, int fast_rexmit)
2054 {
2055         struct tcp_sock *tp = tcp_sk(sk);
2056         struct sk_buff *skb;
2057         int cnt;
2058
2059         BUG_TRAP(packets <= tp->packets_out);
2060         if (tp->lost_skb_hint) {
2061                 skb = tp->lost_skb_hint;
2062                 cnt = tp->lost_cnt_hint;
2063         } else {
2064                 skb = tcp_write_queue_head(sk);
2065                 cnt = 0;
2066         }
2067
2068         tcp_for_write_queue_from(skb, sk) {
2069                 if (skb == tcp_send_head(sk))
2070                         break;
2071                 /* TODO: do this better */
2072                 /* this is not the most efficient way to do this... */
2073                 tp->lost_skb_hint = skb;
2074                 tp->lost_cnt_hint = cnt;
2075
2076                 if (tcp_is_fack(tp) ||
2077                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2078                         cnt += tcp_skb_pcount(skb);
2079
2080                 if (((!fast_rexmit || (tp->lost_out > 0)) && (cnt > packets)) ||
2081                      after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2082                         break;
2083                 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2084                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2085                         tp->lost_out += tcp_skb_pcount(skb);
2086                         tcp_verify_retransmit_hint(tp, skb);
2087                 }
2088         }
2089         tcp_verify_left_out(tp);
2090 }
2091
2092 /* Account newly detected lost packet(s) */
2093
2094 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2095 {
2096         struct tcp_sock *tp = tcp_sk(sk);
2097
2098         if (tcp_is_reno(tp)) {
2099                 tcp_mark_head_lost(sk, 1, fast_rexmit);
2100         } else if (tcp_is_fack(tp)) {
2101                 int lost = tp->fackets_out - tp->reordering;
2102                 if (lost <= 0)
2103                         lost = 1;
2104                 tcp_mark_head_lost(sk, lost, fast_rexmit);
2105         } else {
2106                 int sacked_upto = tp->sacked_out - tp->reordering;
2107                 if (sacked_upto < 0)
2108                         sacked_upto = 0;
2109                 tcp_mark_head_lost(sk, sacked_upto, fast_rexmit);
2110         }
2111
2112         /* New heuristics: it is possible only after we switched
2113          * to restart timer each time when something is ACKed.
2114          * Hence, we can detect timed out packets during fast
2115          * retransmit without falling to slow start.
2116          */
2117         if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2118                 struct sk_buff *skb;
2119
2120                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2121                         : tcp_write_queue_head(sk);
2122
2123                 tcp_for_write_queue_from(skb, sk) {
2124                         if (skb == tcp_send_head(sk))
2125                                 break;
2126                         if (!tcp_skb_timedout(sk, skb))
2127                                 break;
2128
2129                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2130                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2131                                 tp->lost_out += tcp_skb_pcount(skb);
2132                                 tcp_verify_retransmit_hint(tp, skb);
2133                         }
2134                 }
2135
2136                 tp->scoreboard_skb_hint = skb;
2137
2138                 tcp_verify_left_out(tp);
2139         }
2140 }
2141
2142 /* CWND moderation, preventing bursts due to too big ACKs
2143  * in dubious situations.
2144  */
2145 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2146 {
2147         tp->snd_cwnd = min(tp->snd_cwnd,
2148                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2149         tp->snd_cwnd_stamp = tcp_time_stamp;
2150 }
2151
2152 /* Lower bound on congestion window is slow start threshold
2153  * unless congestion avoidance choice decides to overide it.
2154  */
2155 static inline u32 tcp_cwnd_min(const struct sock *sk)
2156 {
2157         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2158
2159         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2160 }
2161
2162 /* Decrease cwnd each second ack. */
2163 static void tcp_cwnd_down(struct sock *sk, int flag)
2164 {
2165         struct tcp_sock *tp = tcp_sk(sk);
2166         int decr = tp->snd_cwnd_cnt + 1;
2167
2168         if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2169             (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2170                 tp->snd_cwnd_cnt = decr&1;
2171                 decr >>= 1;
2172
2173                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2174                         tp->snd_cwnd -= decr;
2175
2176                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2177                 tp->snd_cwnd_stamp = tcp_time_stamp;
2178         }
2179 }
2180
2181 /* Nothing was retransmitted or returned timestamp is less
2182  * than timestamp of the first retransmission.
2183  */
2184 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2185 {
2186         return !tp->retrans_stamp ||
2187                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2188                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2189 }
2190
2191 /* Undo procedures. */
2192
2193 #if FASTRETRANS_DEBUG > 1
2194 static void DBGUNDO(struct sock *sk, const char *msg)
2195 {
2196         struct tcp_sock *tp = tcp_sk(sk);
2197         struct inet_sock *inet = inet_sk(sk);
2198
2199         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2200                msg,
2201                NIPQUAD(inet->daddr), ntohs(inet->dport),
2202                tp->snd_cwnd, tcp_left_out(tp),
2203                tp->snd_ssthresh, tp->prior_ssthresh,
2204                tp->packets_out);
2205 }
2206 #else
2207 #define DBGUNDO(x...) do { } while (0)
2208 #endif
2209
2210 static void tcp_undo_cwr(struct sock *sk, const int undo)
2211 {
2212         struct tcp_sock *tp = tcp_sk(sk);
2213
2214         if (tp->prior_ssthresh) {
2215                 const struct inet_connection_sock *icsk = inet_csk(sk);
2216
2217                 if (icsk->icsk_ca_ops->undo_cwnd)
2218                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2219                 else
2220                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2221
2222                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2223                         tp->snd_ssthresh = tp->prior_ssthresh;
2224                         TCP_ECN_withdraw_cwr(tp);
2225                 }
2226         } else {
2227                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2228         }
2229         tcp_moderate_cwnd(tp);
2230         tp->snd_cwnd_stamp = tcp_time_stamp;
2231
2232         /* There is something screwy going on with the retrans hints after
2233            an undo */
2234         tcp_clear_all_retrans_hints(tp);
2235 }
2236
2237 static inline int tcp_may_undo(struct tcp_sock *tp)
2238 {
2239         return tp->undo_marker &&
2240                 (!tp->undo_retrans || tcp_packet_delayed(tp));
2241 }
2242
2243 /* People celebrate: "We love our President!" */
2244 static int tcp_try_undo_recovery(struct sock *sk)
2245 {
2246         struct tcp_sock *tp = tcp_sk(sk);
2247
2248         if (tcp_may_undo(tp)) {
2249                 /* Happy end! We did not retransmit anything
2250                  * or our original transmission succeeded.
2251                  */
2252                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2253                 tcp_undo_cwr(sk, 1);
2254                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2255                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2256                 else
2257                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2258                 tp->undo_marker = 0;
2259         }
2260         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2261                 /* Hold old state until something *above* high_seq
2262                  * is ACKed. For Reno it is MUST to prevent false
2263                  * fast retransmits (RFC2582). SACK TCP is safe. */
2264                 tcp_moderate_cwnd(tp);
2265                 return 1;
2266         }
2267         tcp_set_ca_state(sk, TCP_CA_Open);
2268         return 0;
2269 }
2270
2271 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2272 static void tcp_try_undo_dsack(struct sock *sk)
2273 {
2274         struct tcp_sock *tp = tcp_sk(sk);
2275
2276         if (tp->undo_marker && !tp->undo_retrans) {
2277                 DBGUNDO(sk, "D-SACK");
2278                 tcp_undo_cwr(sk, 1);
2279                 tp->undo_marker = 0;
2280                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2281         }
2282 }
2283
2284 /* Undo during fast recovery after partial ACK. */
2285
2286 static int tcp_try_undo_partial(struct sock *sk, int acked)
2287 {
2288         struct tcp_sock *tp = tcp_sk(sk);
2289         /* Partial ACK arrived. Force Hoe's retransmit. */
2290         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2291
2292         if (tcp_may_undo(tp)) {
2293                 /* Plain luck! Hole if filled with delayed
2294                  * packet, rather than with a retransmit.
2295                  */
2296                 if (tp->retrans_out == 0)
2297                         tp->retrans_stamp = 0;
2298
2299                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2300
2301                 DBGUNDO(sk, "Hoe");
2302                 tcp_undo_cwr(sk, 0);
2303                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2304
2305                 /* So... Do not make Hoe's retransmit yet.
2306                  * If the first packet was delayed, the rest
2307                  * ones are most probably delayed as well.
2308                  */
2309                 failed = 0;
2310         }
2311         return failed;
2312 }
2313
2314 /* Undo during loss recovery after partial ACK. */
2315 static int tcp_try_undo_loss(struct sock *sk)
2316 {
2317         struct tcp_sock *tp = tcp_sk(sk);
2318
2319         if (tcp_may_undo(tp)) {
2320                 struct sk_buff *skb;
2321                 tcp_for_write_queue(skb, sk) {
2322                         if (skb == tcp_send_head(sk))
2323                                 break;
2324                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2325                 }
2326
2327                 tcp_clear_all_retrans_hints(tp);
2328
2329                 DBGUNDO(sk, "partial loss");
2330                 tp->lost_out = 0;
2331                 tcp_undo_cwr(sk, 1);
2332                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2333                 inet_csk(sk)->icsk_retransmits = 0;
2334                 tp->undo_marker = 0;
2335                 if (tcp_is_sack(tp))
2336                         tcp_set_ca_state(sk, TCP_CA_Open);
2337                 return 1;
2338         }
2339         return 0;
2340 }
2341
2342 static inline void tcp_complete_cwr(struct sock *sk)
2343 {
2344         struct tcp_sock *tp = tcp_sk(sk);
2345         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2346         tp->snd_cwnd_stamp = tcp_time_stamp;
2347         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2348 }
2349
2350 static void tcp_try_to_open(struct sock *sk, int flag)
2351 {
2352         struct tcp_sock *tp = tcp_sk(sk);
2353
2354         tcp_verify_left_out(tp);
2355
2356         if (tp->retrans_out == 0)
2357                 tp->retrans_stamp = 0;
2358
2359         if (flag&FLAG_ECE)
2360                 tcp_enter_cwr(sk, 1);
2361
2362         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2363                 int state = TCP_CA_Open;
2364
2365                 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2366                         state = TCP_CA_Disorder;
2367
2368                 if (inet_csk(sk)->icsk_ca_state != state) {
2369                         tcp_set_ca_state(sk, state);
2370                         tp->high_seq = tp->snd_nxt;
2371                 }
2372                 tcp_moderate_cwnd(tp);
2373         } else {
2374                 tcp_cwnd_down(sk, flag);
2375         }
2376 }
2377
2378 static void tcp_mtup_probe_failed(struct sock *sk)
2379 {
2380         struct inet_connection_sock *icsk = inet_csk(sk);
2381
2382         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2383         icsk->icsk_mtup.probe_size = 0;
2384 }
2385
2386 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2387 {
2388         struct tcp_sock *tp = tcp_sk(sk);
2389         struct inet_connection_sock *icsk = inet_csk(sk);
2390
2391         /* FIXME: breaks with very large cwnd */
2392         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2393         tp->snd_cwnd = tp->snd_cwnd *
2394                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2395                        icsk->icsk_mtup.probe_size;
2396         tp->snd_cwnd_cnt = 0;
2397         tp->snd_cwnd_stamp = tcp_time_stamp;
2398         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2399
2400         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2401         icsk->icsk_mtup.probe_size = 0;
2402         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2403 }
2404
2405
2406 /* Process an event, which can update packets-in-flight not trivially.
2407  * Main goal of this function is to calculate new estimate for left_out,
2408  * taking into account both packets sitting in receiver's buffer and
2409  * packets lost by network.
2410  *
2411  * Besides that it does CWND reduction, when packet loss is detected
2412  * and changes state of machine.
2413  *
2414  * It does _not_ decide what to send, it is made in function
2415  * tcp_xmit_retransmit_queue().
2416  */
2417 static void
2418 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2419 {
2420         struct inet_connection_sock *icsk = inet_csk(sk);
2421         struct tcp_sock *tp = tcp_sk(sk);
2422         int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2423         int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2424                                     (tcp_fackets_out(tp) > tp->reordering));
2425         int fast_rexmit = 0;
2426
2427         /* Some technical things:
2428          * 1. Reno does not count dupacks (sacked_out) automatically. */
2429         if (!tp->packets_out)
2430                 tp->sacked_out = 0;
2431
2432         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2433                 tp->fackets_out = 0;
2434
2435         /* Now state machine starts.
2436          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2437         if (flag&FLAG_ECE)
2438                 tp->prior_ssthresh = 0;
2439
2440         /* B. In all the states check for reneging SACKs. */
2441         if (tp->sacked_out && tcp_check_sack_reneging(sk))
2442                 return;
2443
2444         /* C. Process data loss notification, provided it is valid. */
2445         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2446             before(tp->snd_una, tp->high_seq) &&
2447             icsk->icsk_ca_state != TCP_CA_Open &&
2448             tp->fackets_out > tp->reordering) {
2449                 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, 0);
2450                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2451         }
2452
2453         /* D. Check consistency of the current state. */
2454         tcp_verify_left_out(tp);
2455
2456         /* E. Check state exit conditions. State can be terminated
2457          *    when high_seq is ACKed. */
2458         if (icsk->icsk_ca_state == TCP_CA_Open) {
2459                 BUG_TRAP(tp->retrans_out == 0);
2460                 tp->retrans_stamp = 0;
2461         } else if (!before(tp->snd_una, tp->high_seq)) {
2462                 switch (icsk->icsk_ca_state) {
2463                 case TCP_CA_Loss:
2464                         icsk->icsk_retransmits = 0;
2465                         if (tcp_try_undo_recovery(sk))
2466                                 return;
2467                         break;
2468
2469                 case TCP_CA_CWR:
2470                         /* CWR is to be held something *above* high_seq
2471                          * is ACKed for CWR bit to reach receiver. */
2472                         if (tp->snd_una != tp->high_seq) {
2473                                 tcp_complete_cwr(sk);
2474                                 tcp_set_ca_state(sk, TCP_CA_Open);
2475                         }
2476                         break;
2477
2478                 case TCP_CA_Disorder:
2479                         tcp_try_undo_dsack(sk);
2480                         if (!tp->undo_marker ||
2481                             /* For SACK case do not Open to allow to undo
2482                              * catching for all duplicate ACKs. */
2483                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2484                                 tp->undo_marker = 0;
2485                                 tcp_set_ca_state(sk, TCP_CA_Open);
2486                         }
2487                         break;
2488
2489                 case TCP_CA_Recovery:
2490                         if (tcp_is_reno(tp))
2491                                 tcp_reset_reno_sack(tp);
2492                         if (tcp_try_undo_recovery(sk))
2493                                 return;
2494                         tcp_complete_cwr(sk);
2495                         break;
2496                 }
2497         }
2498
2499         /* F. Process state. */
2500         switch (icsk->icsk_ca_state) {
2501         case TCP_CA_Recovery:
2502                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2503                         if (tcp_is_reno(tp) && is_dupack)
2504                                 tcp_add_reno_sack(sk);
2505                 } else
2506                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2507                 break;
2508         case TCP_CA_Loss:
2509                 if (flag&FLAG_DATA_ACKED)
2510                         icsk->icsk_retransmits = 0;
2511                 if (!tcp_try_undo_loss(sk)) {
2512                         tcp_moderate_cwnd(tp);
2513                         tcp_xmit_retransmit_queue(sk);
2514                         return;
2515                 }
2516                 if (icsk->icsk_ca_state != TCP_CA_Open)
2517                         return;
2518                 /* Loss is undone; fall through to processing in Open state. */
2519         default:
2520                 if (tcp_is_reno(tp)) {
2521                         if (flag & FLAG_SND_UNA_ADVANCED)
2522                                 tcp_reset_reno_sack(tp);
2523                         if (is_dupack)
2524                                 tcp_add_reno_sack(sk);
2525                 }
2526
2527                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2528                         tcp_try_undo_dsack(sk);
2529
2530                 if (!tcp_time_to_recover(sk)) {
2531                         tcp_try_to_open(sk, flag);
2532                         return;
2533                 }
2534
2535                 /* MTU probe failure: don't reduce cwnd */
2536                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2537                     icsk->icsk_mtup.probe_size &&
2538                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2539                         tcp_mtup_probe_failed(sk);
2540                         /* Restores the reduction we did in tcp_mtup_probe() */
2541                         tp->snd_cwnd++;
2542                         tcp_simple_retransmit(sk);
2543                         return;
2544                 }
2545
2546                 /* Otherwise enter Recovery state */
2547
2548                 if (tcp_is_reno(tp))
2549                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2550                 else
2551                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2552
2553                 tp->high_seq = tp->snd_nxt;
2554                 tp->prior_ssthresh = 0;
2555                 tp->undo_marker = tp->snd_una;
2556                 tp->undo_retrans = tp->retrans_out;
2557
2558                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2559                         if (!(flag&FLAG_ECE))
2560                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2561                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2562                         TCP_ECN_queue_cwr(tp);
2563                 }
2564
2565                 tp->bytes_acked = 0;
2566                 tp->snd_cwnd_cnt = 0;
2567                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2568                 fast_rexmit = 1;
2569         }
2570
2571         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2572                 tcp_update_scoreboard(sk, fast_rexmit);
2573         tcp_cwnd_down(sk, flag);
2574         tcp_xmit_retransmit_queue(sk);
2575 }
2576
2577 /* Read draft-ietf-tcplw-high-performance before mucking
2578  * with this code. (Supersedes RFC1323)
2579  */
2580 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2581 {
2582         /* RTTM Rule: A TSecr value received in a segment is used to
2583          * update the averaged RTT measurement only if the segment
2584          * acknowledges some new data, i.e., only if it advances the
2585          * left edge of the send window.
2586          *
2587          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2588          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2589          *
2590          * Changed: reset backoff as soon as we see the first valid sample.
2591          * If we do not, we get strongly overestimated rto. With timestamps
2592          * samples are accepted even from very old segments: f.e., when rtt=1
2593          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2594          * answer arrives rto becomes 120 seconds! If at least one of segments
2595          * in window is lost... Voila.                          --ANK (010210)
2596          */
2597         struct tcp_sock *tp = tcp_sk(sk);
2598         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2599         tcp_rtt_estimator(sk, seq_rtt);
2600         tcp_set_rto(sk);
2601         inet_csk(sk)->icsk_backoff = 0;
2602         tcp_bound_rto(sk);
2603 }
2604
2605 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2606 {
2607         /* We don't have a timestamp. Can only use
2608          * packets that are not retransmitted to determine
2609          * rtt estimates. Also, we must not reset the
2610          * backoff for rto until we get a non-retransmitted
2611          * packet. This allows us to deal with a situation
2612          * where the network delay has increased suddenly.
2613          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2614          */
2615
2616         if (flag & FLAG_RETRANS_DATA_ACKED)
2617                 return;
2618
2619         tcp_rtt_estimator(sk, seq_rtt);
2620         tcp_set_rto(sk);
2621         inet_csk(sk)->icsk_backoff = 0;
2622         tcp_bound_rto(sk);
2623 }
2624
2625 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2626                                       const s32 seq_rtt)
2627 {
2628         const struct tcp_sock *tp = tcp_sk(sk);
2629         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2630         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2631                 tcp_ack_saw_tstamp(sk, flag);
2632         else if (seq_rtt >= 0)
2633                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2634 }
2635
2636 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2637                            u32 in_flight, int good)
2638 {
2639         const struct inet_connection_sock *icsk = inet_csk(sk);
2640         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2641         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2642 }
2643
2644 /* Restart timer after forward progress on connection.
2645  * RFC2988 recommends to restart timer to now+rto.
2646  */
2647 static void tcp_rearm_rto(struct sock *sk)
2648 {
2649         struct tcp_sock *tp = tcp_sk(sk);
2650
2651         if (!tp->packets_out) {
2652                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2653         } else {
2654                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2655         }
2656 }
2657
2658 /* If we get here, the whole TSO packet has not been acked. */
2659 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2660 {
2661         struct tcp_sock *tp = tcp_sk(sk);
2662         u32 packets_acked;
2663
2664         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2665
2666         packets_acked = tcp_skb_pcount(skb);
2667         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2668                 return 0;
2669         packets_acked -= tcp_skb_pcount(skb);
2670
2671         if (packets_acked) {
2672                 BUG_ON(tcp_skb_pcount(skb) == 0);
2673                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2674         }
2675
2676         return packets_acked;
2677 }
2678
2679 /* Remove acknowledged frames from the retransmission queue. If our packet
2680  * is before the ack sequence we can discard it as it's confirmed to have
2681  * arrived at the other end.
2682  */
2683 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p,
2684                                int prior_fackets)
2685 {
2686         struct tcp_sock *tp = tcp_sk(sk);
2687         const struct inet_connection_sock *icsk = inet_csk(sk);
2688         struct sk_buff *skb;
2689         u32 now = tcp_time_stamp;
2690         int fully_acked = 1;
2691         int flag = 0;
2692         int prior_packets = tp->packets_out;
2693         u32 cnt = 0;
2694         u32 reord = tp->packets_out;
2695         s32 seq_rtt = -1;
2696         s32 ca_seq_rtt = -1;
2697         ktime_t last_ackt = net_invalid_timestamp();
2698
2699         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2700                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2701                 u32 end_seq;
2702                 u32 packets_acked;
2703                 u8 sacked = scb->sacked;
2704
2705                 /* Determine how many packets and what bytes were acked, tso and else */
2706                 if (after(scb->end_seq, tp->snd_una)) {
2707                         if (tcp_skb_pcount(skb) == 1 ||
2708                             !after(tp->snd_una, scb->seq))
2709                                 break;
2710
2711                         packets_acked = tcp_tso_acked(sk, skb);
2712                         if (!packets_acked)
2713                                 break;
2714
2715                         fully_acked = 0;
2716                         end_seq = tp->snd_una;
2717                 } else {
2718                         packets_acked = tcp_skb_pcount(skb);
2719                         end_seq = scb->end_seq;
2720                 }
2721
2722                 /* MTU probing checks */
2723                 if (fully_acked && icsk->icsk_mtup.probe_size &&
2724                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2725                         tcp_mtup_probe_success(sk, skb);
2726                 }
2727
2728                 if (sacked) {
2729                         if (sacked & TCPCB_RETRANS) {
2730                                 if (sacked & TCPCB_SACKED_RETRANS)
2731                                         tp->retrans_out -= packets_acked;
2732                                 flag |= FLAG_RETRANS_DATA_ACKED;
2733                                 ca_seq_rtt = -1;
2734                                 seq_rtt = -1;
2735                                 if ((flag & FLAG_DATA_ACKED) ||
2736                                     (packets_acked > 1))
2737                                         flag |= FLAG_NONHEAD_RETRANS_ACKED;
2738                         } else {
2739                                 ca_seq_rtt = now - scb->when;
2740                                 last_ackt = skb->tstamp;
2741                                 if (seq_rtt < 0) {
2742                                         seq_rtt = ca_seq_rtt;
2743                                 }
2744                                 if (!(sacked & TCPCB_SACKED_ACKED))
2745                                         reord = min(cnt, reord);
2746                         }
2747
2748                         if (sacked & TCPCB_SACKED_ACKED)
2749                                 tp->sacked_out -= packets_acked;
2750                         if (sacked & TCPCB_LOST)
2751                                 tp->lost_out -= packets_acked;
2752
2753                         if ((sacked & TCPCB_URG) && tp->urg_mode &&
2754                             !before(end_seq, tp->snd_up))
2755                                 tp->urg_mode = 0;
2756                 } else {
2757                         ca_seq_rtt = now - scb->when;
2758                         last_ackt = skb->tstamp;
2759                         if (seq_rtt < 0) {
2760                                 seq_rtt = ca_seq_rtt;
2761                         }
2762                         reord = min(cnt, reord);
2763                 }
2764                 tp->packets_out -= packets_acked;
2765                 cnt += packets_acked;
2766
2767                 /* Initial outgoing SYN's get put onto the write_queue
2768                  * just like anything else we transmit.  It is not
2769                  * true data, and if we misinform our callers that
2770                  * this ACK acks real data, we will erroneously exit
2771                  * connection startup slow start one packet too
2772                  * quickly.  This is severely frowned upon behavior.
2773                  */
2774                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2775                         flag |= FLAG_DATA_ACKED;
2776                 } else {
2777                         flag |= FLAG_SYN_ACKED;
2778                         tp->retrans_stamp = 0;
2779                 }
2780
2781                 if (!fully_acked)
2782                         break;
2783
2784                 tcp_unlink_write_queue(skb, sk);
2785                 sk_stream_free_skb(sk, skb);
2786                 tcp_clear_all_retrans_hints(tp);
2787         }
2788
2789         if (flag & FLAG_ACKED) {
2790                 u32 pkts_acked = prior_packets - tp->packets_out;
2791                 const struct tcp_congestion_ops *ca_ops
2792                         = inet_csk(sk)->icsk_ca_ops;
2793
2794                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2795                 tcp_rearm_rto(sk);
2796
2797                 if (tcp_is_reno(tp)) {
2798                         tcp_remove_reno_sacks(sk, pkts_acked);
2799                 } else {
2800                         /* Non-retransmitted hole got filled? That's reordering */
2801                         if (reord < prior_fackets)
2802                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2803                 }
2804
2805                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2806                 /* hint's skb might be NULL but we don't need to care */
2807                 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2808                                                tp->fastpath_cnt_hint);
2809                 if (ca_ops->pkts_acked) {
2810                         s32 rtt_us = -1;
2811
2812                         /* Is the ACK triggering packet unambiguous? */
2813                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2814                                 /* High resolution needed and available? */
2815                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2816                                     !ktime_equal(last_ackt,
2817                                                  net_invalid_timestamp()))
2818                                         rtt_us = ktime_us_delta(ktime_get_real(),
2819                                                                 last_ackt);
2820                                 else if (ca_seq_rtt > 0)
2821                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
2822                         }
2823
2824                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2825                 }
2826         }
2827
2828 #if FASTRETRANS_DEBUG > 0
2829         BUG_TRAP((int)tp->sacked_out >= 0);
2830         BUG_TRAP((int)tp->lost_out >= 0);
2831         BUG_TRAP((int)tp->retrans_out >= 0);
2832         if (!tp->packets_out && tcp_is_sack(tp)) {
2833                 icsk = inet_csk(sk);
2834                 if (tp->lost_out) {
2835                         printk(KERN_DEBUG "Leak l=%u %d\n",
2836                                tp->lost_out, icsk->icsk_ca_state);
2837                         tp->lost_out = 0;
2838                 }
2839                 if (tp->sacked_out) {
2840                         printk(KERN_DEBUG "Leak s=%u %d\n",
2841                                tp->sacked_out, icsk->icsk_ca_state);
2842                         tp->sacked_out = 0;
2843                 }
2844                 if (tp->retrans_out) {
2845                         printk(KERN_DEBUG "Leak r=%u %d\n",
2846                                tp->retrans_out, icsk->icsk_ca_state);
2847                         tp->retrans_out = 0;
2848                 }
2849         }
2850 #endif
2851         *seq_rtt_p = seq_rtt;
2852         return flag;
2853 }
2854
2855 static void tcp_ack_probe(struct sock *sk)
2856 {
2857         const struct tcp_sock *tp = tcp_sk(sk);
2858         struct inet_connection_sock *icsk = inet_csk(sk);
2859
2860         /* Was it a usable window open? */
2861
2862         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2863                    tp->snd_una + tp->snd_wnd)) {
2864                 icsk->icsk_backoff = 0;
2865                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2866                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2867                  * This function is not for random using!
2868                  */
2869         } else {
2870                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2871                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2872                                           TCP_RTO_MAX);
2873         }
2874 }
2875
2876 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2877 {
2878         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2879                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2880 }
2881
2882 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2883 {
2884         const struct tcp_sock *tp = tcp_sk(sk);
2885         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2886                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2887 }
2888
2889 /* Check that window update is acceptable.
2890  * The function assumes that snd_una<=ack<=snd_next.
2891  */
2892 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2893                                         const u32 ack_seq, const u32 nwin)
2894 {
2895         return (after(ack, tp->snd_una) ||
2896                 after(ack_seq, tp->snd_wl1) ||
2897                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2898 }
2899
2900 /* Update our send window.
2901  *
2902  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2903  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2904  */
2905 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2906                                  u32 ack_seq)
2907 {
2908         struct tcp_sock *tp = tcp_sk(sk);
2909         int flag = 0;
2910         u32 nwin = ntohs(tcp_hdr(skb)->window);
2911
2912         if (likely(!tcp_hdr(skb)->syn))
2913                 nwin <<= tp->rx_opt.snd_wscale;
2914
2915         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2916                 flag |= FLAG_WIN_UPDATE;
2917                 tcp_update_wl(tp, ack, ack_seq);
2918
2919                 if (tp->snd_wnd != nwin) {
2920                         tp->snd_wnd = nwin;
2921
2922                         /* Note, it is the only place, where
2923                          * fast path is recovered for sending TCP.
2924                          */
2925                         tp->pred_flags = 0;
2926                         tcp_fast_path_check(sk);
2927
2928                         if (nwin > tp->max_window) {
2929                                 tp->max_window = nwin;
2930                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2931                         }
2932                 }
2933         }
2934
2935         tp->snd_una = ack;
2936
2937         return flag;
2938 }
2939
2940 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2941  * continue in congestion avoidance.
2942  */
2943 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2944 {
2945         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2946         tp->snd_cwnd_cnt = 0;
2947         tp->bytes_acked = 0;
2948         TCP_ECN_queue_cwr(tp);
2949         tcp_moderate_cwnd(tp);
2950 }
2951
2952 /* A conservative spurious RTO response algorithm: reduce cwnd using
2953  * rate halving and continue in congestion avoidance.
2954  */
2955 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2956 {
2957         tcp_enter_cwr(sk, 0);
2958 }
2959
2960 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2961 {
2962         if (flag&FLAG_ECE)
2963                 tcp_ratehalving_spur_to_response(sk);
2964         else
2965                 tcp_undo_cwr(sk, 1);
2966 }
2967
2968 /* F-RTO spurious RTO detection algorithm (RFC4138)
2969  *
2970  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2971  * comments). State (ACK number) is kept in frto_counter. When ACK advances
2972  * window (but not to or beyond highest sequence sent before RTO):
2973  *   On First ACK,  send two new segments out.
2974  *   On Second ACK, RTO was likely spurious. Do spurious response (response
2975  *                  algorithm is not part of the F-RTO detection algorithm
2976  *                  given in RFC4138 but can be selected separately).
2977  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2978  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2979  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2980  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2981  *
2982  * Rationale: if the RTO was spurious, new ACKs should arrive from the
2983  * original window even after we transmit two new data segments.
2984  *
2985  * SACK version:
2986  *   on first step, wait until first cumulative ACK arrives, then move to
2987  *   the second step. In second step, the next ACK decides.
2988  *
2989  * F-RTO is implemented (mainly) in four functions:
2990  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
2991  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2992  *     called when tcp_use_frto() showed green light
2993  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2994  *   - tcp_enter_frto_loss() is called if there is not enough evidence
2995  *     to prove that the RTO is indeed spurious. It transfers the control
2996  *     from F-RTO to the conventional RTO recovery
2997  */
2998 static int tcp_process_frto(struct sock *sk, int flag)
2999 {
3000         struct tcp_sock *tp = tcp_sk(sk);
3001
3002         tcp_verify_left_out(tp);
3003
3004         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3005         if (flag&FLAG_DATA_ACKED)
3006                 inet_csk(sk)->icsk_retransmits = 0;
3007
3008         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3009             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3010                 tp->undo_marker = 0;
3011
3012         if (!before(tp->snd_una, tp->frto_highmark)) {
3013                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3014                 return 1;
3015         }
3016
3017         if (!IsSackFrto() || tcp_is_reno(tp)) {
3018                 /* RFC4138 shortcoming in step 2; should also have case c):
3019                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3020                  * data, winupdate
3021                  */
3022                 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
3023                         return 1;
3024
3025                 if (!(flag&FLAG_DATA_ACKED)) {
3026                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3027                                             flag);
3028                         return 1;
3029                 }
3030         } else {
3031                 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3032                         /* Prevent sending of new data. */
3033                         tp->snd_cwnd = min(tp->snd_cwnd,
3034                                            tcp_packets_in_flight(tp));
3035                         return 1;
3036                 }
3037
3038                 if ((tp->frto_counter >= 2) &&
3039                     (!(flag&FLAG_FORWARD_PROGRESS) ||
3040                      ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
3041                         /* RFC4138 shortcoming (see comment above) */
3042                         if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
3043                                 return 1;
3044
3045                         tcp_enter_frto_loss(sk, 3, flag);
3046                         return 1;
3047                 }
3048         }
3049
3050         if (tp->frto_counter == 1) {
3051                 /* tcp_may_send_now needs to see updated state */
3052                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3053                 tp->frto_counter = 2;
3054
3055                 if (!tcp_may_send_now(sk))
3056                         tcp_enter_frto_loss(sk, 2, flag);
3057
3058                 return 1;
3059         } else {
3060                 switch (sysctl_tcp_frto_response) {
3061                 case 2:
3062                         tcp_undo_spur_to_response(sk, flag);
3063                         break;
3064                 case 1:
3065                         tcp_conservative_spur_to_response(tp);
3066                         break;
3067                 default:
3068                         tcp_ratehalving_spur_to_response(sk);
3069                         break;
3070                 }
3071                 tp->frto_counter = 0;
3072                 tp->undo_marker = 0;
3073                 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
3074         }
3075         return 0;
3076 }
3077
3078 /* This routine deals with incoming acks, but not outgoing ones. */
3079 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3080 {
3081         struct inet_connection_sock *icsk = inet_csk(sk);
3082         struct tcp_sock *tp = tcp_sk(sk);
3083         u32 prior_snd_una = tp->snd_una;
3084         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3085         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3086         u32 prior_in_flight;
3087         u32 prior_fackets;
3088         s32 seq_rtt;
3089         int prior_packets;
3090         int frto_cwnd = 0;
3091
3092         /* If the ack is newer than sent or older than previous acks
3093          * then we can probably ignore it.
3094          */
3095         if (after(ack, tp->snd_nxt))
3096                 goto uninteresting_ack;
3097
3098         if (before(ack, prior_snd_una))
3099                 goto old_ack;
3100
3101         if (after(ack, prior_snd_una))
3102                 flag |= FLAG_SND_UNA_ADVANCED;
3103
3104         if (sysctl_tcp_abc) {
3105                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3106                         tp->bytes_acked += ack - prior_snd_una;
3107                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3108                         /* we assume just one segment left network */
3109                         tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
3110         }
3111
3112         prior_fackets = tp->fackets_out;
3113         prior_in_flight = tcp_packets_in_flight(tp);
3114
3115         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3116                 /* Window is constant, pure forward advance.
3117                  * No more checks are required.
3118                  * Note, we use the fact that SND.UNA>=SND.WL2.
3119                  */
3120                 tcp_update_wl(tp, ack, ack_seq);
3121                 tp->snd_una = ack;
3122                 flag |= FLAG_WIN_UPDATE;
3123
3124                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3125
3126                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3127         } else {
3128                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3129                         flag |= FLAG_DATA;
3130                 else
3131                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3132
3133                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3134
3135                 if (TCP_SKB_CB(skb)->sacked)
3136                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3137
3138                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3139                         flag |= FLAG_ECE;
3140
3141                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3142         }
3143
3144         /* We passed data and got it acked, remove any soft error
3145          * log. Something worked...
3146          */
3147         sk->sk_err_soft = 0;
3148         tp->rcv_tstamp = tcp_time_stamp;
3149         prior_packets = tp->packets_out;
3150         if (!prior_packets)
3151                 goto no_queue;
3152
3153         /* See if we can take anything off of the retransmit queue. */
3154         flag |= tcp_clean_rtx_queue(sk, &seq_rtt, prior_fackets);
3155
3156         if (tp->frto_counter)
3157                 frto_cwnd = tcp_process_frto(sk, flag);
3158         /* Guarantee sacktag reordering detection against wrap-arounds */
3159         if (before(tp->frto_highmark, tp->snd_una))
3160                 tp->frto_highmark = 0;
3161
3162         if (tcp_ack_is_dubious(sk, flag)) {
3163                 /* Advance CWND, if state allows this. */
3164                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3165                     tcp_may_raise_cwnd(sk, flag))
3166                         tcp_cong_avoid(sk, ack, prior_in_flight, 0);
3167                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3168         } else {
3169                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3170                         tcp_cong_avoid(sk, ack, prior_in_flight, 1);
3171         }
3172
3173         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3174                 dst_confirm(sk->sk_dst_cache);
3175
3176         return 1;
3177
3178 no_queue:
3179         icsk->icsk_probes_out = 0;
3180
3181         /* If this ack opens up a zero window, clear backoff.  It was
3182          * being used to time the probes, and is probably far higher than
3183          * it needs to be for normal retransmission.
3184          */
3185         if (tcp_send_head(sk))
3186                 tcp_ack_probe(sk);
3187         return 1;
3188
3189 old_ack:
3190         if (TCP_SKB_CB(skb)->sacked)
3191                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3192
3193 uninteresting_ack:
3194         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3195         return 0;
3196 }
3197
3198
3199 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3200  * But, this can also be called on packets in the established flow when
3201  * the fast version below fails.
3202  */
3203 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3204 {
3205         unsigned char *ptr;
3206         struct tcphdr *th = tcp_hdr(skb);
3207         int length=(th->doff*4)-sizeof(struct tcphdr);
3208
3209         ptr = (unsigned char *)(th + 1);
3210         opt_rx->saw_tstamp = 0;
3211
3212         while (length > 0) {
3213                 int opcode=*ptr++;
3214                 int opsize;
3215
3216                 switch (opcode) {
3217                         case TCPOPT_EOL:
3218                                 return;
3219                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3220                                 length--;
3221                                 continue;
3222                         default:
3223                                 opsize=*ptr++;
3224                                 if (opsize < 2) /* "silly options" */
3225                                         return;
3226                                 if (opsize > length)
3227                                         return; /* don't parse partial options */
3228                                 switch (opcode) {
3229                                 case TCPOPT_MSS:
3230                                         if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3231                                                 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3232                                                 if (in_mss) {
3233                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3234                                                                 in_mss = opt_rx->user_mss;
3235                                                         opt_rx->mss_clamp = in_mss;
3236                                                 }
3237                                         }
3238                                         break;
3239                                 case TCPOPT_WINDOW:
3240                                         if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3241                                                 if (sysctl_tcp_window_scaling) {
3242                                                         __u8 snd_wscale = *(__u8 *) ptr;
3243                                                         opt_rx->wscale_ok = 1;
3244                                                         if (snd_wscale > 14) {
3245                                                                 if (net_ratelimit())
3246                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3247                                                                                "scaling value %d >14 received.\n",
3248                                                                                snd_wscale);
3249                                                                 snd_wscale = 14;
3250                                                         }
3251                                                         opt_rx->snd_wscale = snd_wscale;
3252                                                 }
3253                                         break;
3254                                 case TCPOPT_TIMESTAMP:
3255                                         if (opsize==TCPOLEN_TIMESTAMP) {
3256                                                 if ((estab && opt_rx->tstamp_ok) ||
3257                                                     (!estab && sysctl_tcp_timestamps)) {
3258                                                         opt_rx->saw_tstamp = 1;
3259                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3260                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3261                                                 }
3262                                         }
3263                                         break;
3264                                 case TCPOPT_SACK_PERM:
3265                                         if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3266                                                 if (sysctl_tcp_sack) {
3267                                                         opt_rx->sack_ok = 1;
3268                                                         tcp_sack_reset(opt_rx);
3269                                                 }
3270                                         }
3271                                         break;
3272
3273                                 case TCPOPT_SACK:
3274                                         if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3275                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3276                                            opt_rx->sack_ok) {
3277                           &nbs